]>
Commit | Line | Data |
---|---|---|
5190f052 MM |
1 | // SPDX-License-Identifier: (GPL-2.0 OR BSD-3-Clause) |
2 | /* | |
3 | * Copyright(c) 2018 Intel Corporation. | |
4 | * | |
5 | */ | |
6 | ||
7 | #include "hfi.h" | |
37356e78 | 8 | #include "qp.h" |
742a3826 | 9 | #include "rc.h" |
5190f052 MM |
10 | #include "verbs.h" |
11 | #include "tid_rdma.h" | |
838b6fd2 | 12 | #include "exp_rcv.h" |
a131d164 | 13 | #include "trace.h" |
5190f052 | 14 | |
742a3826 KW |
15 | /** |
16 | * DOC: TID RDMA READ protocol | |
17 | * | |
18 | * This is an end-to-end protocol at the hfi1 level between two nodes that | |
19 | * improves performance by avoiding data copy on the requester side. It | |
20 | * converts a qualified RDMA READ request into a TID RDMA READ request on | |
21 | * the requester side and thereafter handles the request and response | |
22 | * differently. To be qualified, the RDMA READ request should meet the | |
23 | * following: | |
24 | * -- The total data length should be greater than 256K; | |
25 | * -- The total data length should be a multiple of 4K page size; | |
26 | * -- Each local scatter-gather entry should be 4K page aligned; | |
27 | * -- Each local scatter-gather entry should be a multiple of 4K page size; | |
28 | */ | |
29 | ||
37356e78 KW |
30 | #define RCV_TID_FLOW_TABLE_CTRL_FLOW_VALID_SMASK BIT_ULL(32) |
31 | #define RCV_TID_FLOW_TABLE_CTRL_HDR_SUPP_EN_SMASK BIT_ULL(33) | |
32 | #define RCV_TID_FLOW_TABLE_CTRL_KEEP_AFTER_SEQ_ERR_SMASK BIT_ULL(34) | |
33 | #define RCV_TID_FLOW_TABLE_CTRL_KEEP_ON_GEN_ERR_SMASK BIT_ULL(35) | |
34 | #define RCV_TID_FLOW_TABLE_STATUS_SEQ_MISMATCH_SMASK BIT_ULL(37) | |
35 | #define RCV_TID_FLOW_TABLE_STATUS_GEN_MISMATCH_SMASK BIT_ULL(38) | |
36 | ||
742a3826 KW |
37 | /* Maximum number of packets within a flow generation. */ |
38 | #define MAX_TID_FLOW_PSN BIT(HFI1_KDETH_BTH_SEQ_SHIFT) | |
39 | ||
37356e78 KW |
40 | #define GENERATION_MASK 0xFFFFF |
41 | ||
42 | static u32 mask_generation(u32 a) | |
43 | { | |
44 | return a & GENERATION_MASK; | |
45 | } | |
46 | ||
47 | /* Reserved generation value to set to unused flows for kernel contexts */ | |
48 | #define KERN_GENERATION_RESERVED mask_generation(U32_MAX) | |
49 | ||
d22a207d KW |
50 | /* |
51 | * J_KEY for kernel contexts when TID RDMA is used. | |
52 | * See generate_jkey() in hfi.h for more information. | |
53 | */ | |
54 | #define TID_RDMA_JKEY 32 | |
55 | #define HFI1_KERNEL_MIN_JKEY HFI1_ADMIN_JKEY_RANGE | |
56 | #define HFI1_KERNEL_MAX_JKEY (2 * HFI1_ADMIN_JKEY_RANGE - 1) | |
57 | ||
838b6fd2 | 58 | /* Maximum number of segments in flight per QP request. */ |
d22a207d KW |
59 | #define TID_RDMA_MAX_READ_SEGS_PER_REQ 6 |
60 | #define TID_RDMA_MAX_WRITE_SEGS_PER_REQ 4 | |
838b6fd2 KW |
61 | #define MAX_REQ max_t(u16, TID_RDMA_MAX_READ_SEGS_PER_REQ, \ |
62 | TID_RDMA_MAX_WRITE_SEGS_PER_REQ) | |
63 | #define MAX_FLOWS roundup_pow_of_two(MAX_REQ + 1) | |
64 | ||
65 | #define MAX_EXPECTED_PAGES (MAX_EXPECTED_BUFFER / PAGE_SIZE) | |
d22a207d | 66 | |
742a3826 KW |
67 | #define TID_RDMA_DESTQP_FLOW_SHIFT 11 |
68 | #define TID_RDMA_DESTQP_FLOW_MASK 0x1f | |
69 | ||
9905bf06 KW |
70 | #define TID_FLOW_SW_PSN BIT(0) |
71 | ||
d22a207d KW |
72 | #define TID_OPFN_QP_CTXT_MASK 0xff |
73 | #define TID_OPFN_QP_CTXT_SHIFT 56 | |
74 | #define TID_OPFN_QP_KDETH_MASK 0xff | |
75 | #define TID_OPFN_QP_KDETH_SHIFT 48 | |
76 | #define TID_OPFN_MAX_LEN_MASK 0x7ff | |
77 | #define TID_OPFN_MAX_LEN_SHIFT 37 | |
78 | #define TID_OPFN_TIMEOUT_MASK 0x1f | |
79 | #define TID_OPFN_TIMEOUT_SHIFT 32 | |
80 | #define TID_OPFN_RESERVED_MASK 0x3f | |
81 | #define TID_OPFN_RESERVED_SHIFT 26 | |
82 | #define TID_OPFN_URG_MASK 0x1 | |
83 | #define TID_OPFN_URG_SHIFT 25 | |
84 | #define TID_OPFN_VER_MASK 0x7 | |
85 | #define TID_OPFN_VER_SHIFT 22 | |
86 | #define TID_OPFN_JKEY_MASK 0x3f | |
87 | #define TID_OPFN_JKEY_SHIFT 16 | |
88 | #define TID_OPFN_MAX_READ_MASK 0x3f | |
89 | #define TID_OPFN_MAX_READ_SHIFT 10 | |
90 | #define TID_OPFN_MAX_WRITE_MASK 0x3f | |
91 | #define TID_OPFN_MAX_WRITE_SHIFT 4 | |
92 | ||
93 | /* | |
94 | * OPFN TID layout | |
95 | * | |
96 | * 63 47 31 15 | |
97 | * NNNNNNNNKKKKKKKK MMMMMMMMMMMTTTTT DDDDDDUVVVJJJJJJ RRRRRRWWWWWWCCCC | |
98 | * 3210987654321098 7654321098765432 1098765432109876 5432109876543210 | |
99 | * N - the context Number | |
100 | * K - the Kdeth_qp | |
101 | * M - Max_len | |
102 | * T - Timeout | |
103 | * D - reserveD | |
104 | * V - version | |
105 | * U - Urg capable | |
106 | * J - Jkey | |
107 | * R - max_Read | |
108 | * W - max_Write | |
109 | * C - Capcode | |
110 | */ | |
111 | ||
07b92370 KW |
112 | static u32 tid_rdma_flow_wt; |
113 | ||
37356e78 | 114 | static void tid_rdma_trigger_resume(struct work_struct *work); |
838b6fd2 KW |
115 | static void hfi1_kern_exp_rcv_free_flows(struct tid_rdma_request *req); |
116 | static int hfi1_kern_exp_rcv_alloc_flows(struct tid_rdma_request *req, | |
117 | gfp_t gfp); | |
118 | static void hfi1_init_trdma_req(struct rvt_qp *qp, | |
119 | struct tid_rdma_request *req); | |
07b92370 | 120 | static void hfi1_tid_write_alloc_resources(struct rvt_qp *qp, bool intr_ctx); |
3c759e00 KW |
121 | static void hfi1_tid_timeout(struct timer_list *t); |
122 | static void hfi1_add_tid_reap_timer(struct rvt_qp *qp); | |
123 | static void hfi1_mod_tid_reap_timer(struct rvt_qp *qp); | |
829eaee5 KW |
124 | static void hfi1_mod_tid_retry_timer(struct rvt_qp *qp); |
125 | static int hfi1_stop_tid_retry_timer(struct rvt_qp *qp); | |
126 | static void hfi1_tid_retry_timeout(struct timer_list *t); | |
24c5bfea KW |
127 | static int make_tid_rdma_ack(struct rvt_qp *qp, |
128 | struct ib_other_headers *ohdr, | |
129 | struct hfi1_pkt_state *ps); | |
572f0c33 | 130 | static void hfi1_do_tid_send(struct rvt_qp *qp); |
37356e78 | 131 | |
d22a207d KW |
132 | static u64 tid_rdma_opfn_encode(struct tid_rdma_params *p) |
133 | { | |
134 | return | |
135 | (((u64)p->qp & TID_OPFN_QP_CTXT_MASK) << | |
136 | TID_OPFN_QP_CTXT_SHIFT) | | |
137 | ((((u64)p->qp >> 16) & TID_OPFN_QP_KDETH_MASK) << | |
138 | TID_OPFN_QP_KDETH_SHIFT) | | |
139 | (((u64)((p->max_len >> PAGE_SHIFT) - 1) & | |
140 | TID_OPFN_MAX_LEN_MASK) << TID_OPFN_MAX_LEN_SHIFT) | | |
141 | (((u64)p->timeout & TID_OPFN_TIMEOUT_MASK) << | |
142 | TID_OPFN_TIMEOUT_SHIFT) | | |
143 | (((u64)p->urg & TID_OPFN_URG_MASK) << TID_OPFN_URG_SHIFT) | | |
144 | (((u64)p->jkey & TID_OPFN_JKEY_MASK) << TID_OPFN_JKEY_SHIFT) | | |
145 | (((u64)p->max_read & TID_OPFN_MAX_READ_MASK) << | |
146 | TID_OPFN_MAX_READ_SHIFT) | | |
147 | (((u64)p->max_write & TID_OPFN_MAX_WRITE_MASK) << | |
148 | TID_OPFN_MAX_WRITE_SHIFT); | |
149 | } | |
150 | ||
151 | static void tid_rdma_opfn_decode(struct tid_rdma_params *p, u64 data) | |
152 | { | |
153 | p->max_len = (((data >> TID_OPFN_MAX_LEN_SHIFT) & | |
154 | TID_OPFN_MAX_LEN_MASK) + 1) << PAGE_SHIFT; | |
155 | p->jkey = (data >> TID_OPFN_JKEY_SHIFT) & TID_OPFN_JKEY_MASK; | |
156 | p->max_write = (data >> TID_OPFN_MAX_WRITE_SHIFT) & | |
157 | TID_OPFN_MAX_WRITE_MASK; | |
158 | p->max_read = (data >> TID_OPFN_MAX_READ_SHIFT) & | |
159 | TID_OPFN_MAX_READ_MASK; | |
160 | p->qp = | |
161 | ((((data >> TID_OPFN_QP_KDETH_SHIFT) & TID_OPFN_QP_KDETH_MASK) | |
162 | << 16) | | |
163 | ((data >> TID_OPFN_QP_CTXT_SHIFT) & TID_OPFN_QP_CTXT_MASK)); | |
164 | p->urg = (data >> TID_OPFN_URG_SHIFT) & TID_OPFN_URG_MASK; | |
165 | p->timeout = (data >> TID_OPFN_TIMEOUT_SHIFT) & TID_OPFN_TIMEOUT_MASK; | |
166 | } | |
167 | ||
168 | void tid_rdma_opfn_init(struct rvt_qp *qp, struct tid_rdma_params *p) | |
169 | { | |
170 | struct hfi1_qp_priv *priv = qp->priv; | |
171 | ||
172 | p->qp = (kdeth_qp << 16) | priv->rcd->ctxt; | |
173 | p->max_len = TID_RDMA_MAX_SEGMENT_SIZE; | |
174 | p->jkey = priv->rcd->jkey; | |
175 | p->max_read = TID_RDMA_MAX_READ_SEGS_PER_REQ; | |
176 | p->max_write = TID_RDMA_MAX_WRITE_SEGS_PER_REQ; | |
177 | p->timeout = qp->timeout; | |
178 | p->urg = is_urg_masked(priv->rcd); | |
179 | } | |
180 | ||
181 | bool tid_rdma_conn_req(struct rvt_qp *qp, u64 *data) | |
182 | { | |
183 | struct hfi1_qp_priv *priv = qp->priv; | |
184 | ||
185 | *data = tid_rdma_opfn_encode(&priv->tid_rdma.local); | |
186 | return true; | |
187 | } | |
188 | ||
189 | bool tid_rdma_conn_reply(struct rvt_qp *qp, u64 data) | |
190 | { | |
191 | struct hfi1_qp_priv *priv = qp->priv; | |
192 | struct tid_rdma_params *remote, *old; | |
193 | bool ret = true; | |
194 | ||
195 | old = rcu_dereference_protected(priv->tid_rdma.remote, | |
196 | lockdep_is_held(&priv->opfn.lock)); | |
197 | data &= ~0xfULL; | |
198 | /* | |
199 | * If data passed in is zero, return true so as not to continue the | |
200 | * negotiation process | |
201 | */ | |
202 | if (!data || !HFI1_CAP_IS_KSET(TID_RDMA)) | |
203 | goto null; | |
204 | /* | |
205 | * If kzalloc fails, return false. This will result in: | |
206 | * * at the requester a new OPFN request being generated to retry | |
207 | * the negotiation | |
208 | * * at the responder, 0 being returned to the requester so as to | |
209 | * disable TID RDMA at both the requester and the responder | |
210 | */ | |
211 | remote = kzalloc(sizeof(*remote), GFP_ATOMIC); | |
212 | if (!remote) { | |
213 | ret = false; | |
214 | goto null; | |
215 | } | |
216 | ||
217 | tid_rdma_opfn_decode(remote, data); | |
218 | priv->tid_timer_timeout_jiffies = | |
219 | usecs_to_jiffies((((4096UL * (1UL << remote->timeout)) / | |
220 | 1000UL) << 3) * 7); | |
a131d164 KW |
221 | trace_hfi1_opfn_param(qp, 0, &priv->tid_rdma.local); |
222 | trace_hfi1_opfn_param(qp, 1, remote); | |
d22a207d KW |
223 | rcu_assign_pointer(priv->tid_rdma.remote, remote); |
224 | /* | |
225 | * A TID RDMA READ request's segment size is not equal to | |
226 | * remote->max_len only when the request's data length is smaller | |
227 | * than remote->max_len. In that case, there will be only one segment. | |
228 | * Therefore, when priv->pkts_ps is used to calculate req->cur_seg | |
229 | * during retry, it will lead to req->cur_seg = 0, which is exactly | |
230 | * what is expected. | |
231 | */ | |
232 | priv->pkts_ps = (u16)rvt_div_mtu(qp, remote->max_len); | |
233 | priv->timeout_shift = ilog2(priv->pkts_ps - 1) + 1; | |
234 | goto free; | |
235 | null: | |
236 | RCU_INIT_POINTER(priv->tid_rdma.remote, NULL); | |
237 | priv->timeout_shift = 0; | |
238 | free: | |
239 | if (old) | |
240 | kfree_rcu(old, rcu_head); | |
241 | return ret; | |
242 | } | |
243 | ||
244 | bool tid_rdma_conn_resp(struct rvt_qp *qp, u64 *data) | |
245 | { | |
246 | bool ret; | |
247 | ||
248 | ret = tid_rdma_conn_reply(qp, *data); | |
249 | *data = 0; | |
250 | /* | |
251 | * If tid_rdma_conn_reply() returns error, set *data as 0 to indicate | |
252 | * TID RDMA could not be enabled. This will result in TID RDMA being | |
253 | * disabled at the requester too. | |
254 | */ | |
255 | if (ret) | |
256 | (void)tid_rdma_conn_req(qp, data); | |
257 | return ret; | |
258 | } | |
259 | ||
260 | void tid_rdma_conn_error(struct rvt_qp *qp) | |
261 | { | |
262 | struct hfi1_qp_priv *priv = qp->priv; | |
263 | struct tid_rdma_params *old; | |
264 | ||
265 | old = rcu_dereference_protected(priv->tid_rdma.remote, | |
266 | lockdep_is_held(&priv->opfn.lock)); | |
267 | RCU_INIT_POINTER(priv->tid_rdma.remote, NULL); | |
268 | if (old) | |
269 | kfree_rcu(old, rcu_head); | |
270 | } | |
271 | ||
272 | /* This is called at context initialization time */ | |
273 | int hfi1_kern_exp_rcv_init(struct hfi1_ctxtdata *rcd, int reinit) | |
274 | { | |
275 | if (reinit) | |
276 | return 0; | |
277 | ||
278 | BUILD_BUG_ON(TID_RDMA_JKEY < HFI1_KERNEL_MIN_JKEY); | |
279 | BUILD_BUG_ON(TID_RDMA_JKEY > HFI1_KERNEL_MAX_JKEY); | |
280 | rcd->jkey = TID_RDMA_JKEY; | |
281 | hfi1_set_ctxt_jkey(rcd->dd, rcd, rcd->jkey); | |
838b6fd2 | 282 | return hfi1_alloc_ctxt_rcv_groups(rcd); |
d22a207d KW |
283 | } |
284 | ||
5190f052 MM |
285 | /** |
286 | * qp_to_rcd - determine the receive context used by a qp | |
287 | * @qp - the qp | |
288 | * | |
289 | * This routine returns the receive context associated | |
290 | * with a a qp's qpn. | |
291 | * | |
292 | * Returns the context. | |
293 | */ | |
294 | static struct hfi1_ctxtdata *qp_to_rcd(struct rvt_dev_info *rdi, | |
295 | struct rvt_qp *qp) | |
296 | { | |
297 | struct hfi1_ibdev *verbs_dev = container_of(rdi, | |
298 | struct hfi1_ibdev, | |
299 | rdi); | |
300 | struct hfi1_devdata *dd = container_of(verbs_dev, | |
301 | struct hfi1_devdata, | |
302 | verbs_dev); | |
303 | unsigned int ctxt; | |
304 | ||
305 | if (qp->ibqp.qp_num == 0) | |
306 | ctxt = 0; | |
307 | else | |
308 | ctxt = ((qp->ibqp.qp_num >> dd->qos_shift) % | |
309 | (dd->n_krcv_queues - 1)) + 1; | |
310 | ||
311 | return dd->rcd[ctxt]; | |
312 | } | |
313 | ||
314 | int hfi1_qp_priv_init(struct rvt_dev_info *rdi, struct rvt_qp *qp, | |
315 | struct ib_qp_init_attr *init_attr) | |
316 | { | |
317 | struct hfi1_qp_priv *qpriv = qp->priv; | |
838b6fd2 | 318 | int i, ret; |
5190f052 MM |
319 | |
320 | qpriv->rcd = qp_to_rcd(rdi, qp); | |
321 | ||
48a615dc KW |
322 | spin_lock_init(&qpriv->opfn.lock); |
323 | INIT_WORK(&qpriv->opfn.opfn_work, opfn_send_conn_request); | |
37356e78 KW |
324 | INIT_WORK(&qpriv->tid_rdma.trigger_work, tid_rdma_trigger_resume); |
325 | qpriv->flow_state.psn = 0; | |
326 | qpriv->flow_state.index = RXE_NUM_TID_FLOWS; | |
327 | qpriv->flow_state.last_index = RXE_NUM_TID_FLOWS; | |
328 | qpriv->flow_state.generation = KERN_GENERATION_RESERVED; | |
9e93e967 | 329 | qpriv->s_state = TID_OP(WRITE_RESP); |
72a0ea99 KW |
330 | qpriv->s_tid_cur = HFI1_QP_WQE_INVALID; |
331 | qpriv->s_tid_head = HFI1_QP_WQE_INVALID; | |
332 | qpriv->s_tid_tail = HFI1_QP_WQE_INVALID; | |
07b92370 KW |
333 | qpriv->rnr_nak_state = TID_RNR_NAK_INIT; |
334 | qpriv->r_tid_head = HFI1_QP_WQE_INVALID; | |
335 | qpriv->r_tid_tail = HFI1_QP_WQE_INVALID; | |
336 | qpriv->r_tid_ack = HFI1_QP_WQE_INVALID; | |
337 | qpriv->r_tid_alloc = HFI1_QP_WQE_INVALID; | |
70dcb2e3 | 338 | atomic_set(&qpriv->n_requests, 0); |
9e93e967 | 339 | atomic_set(&qpriv->n_tid_requests, 0); |
3c759e00 | 340 | timer_setup(&qpriv->s_tid_timer, hfi1_tid_timeout, 0); |
829eaee5 | 341 | timer_setup(&qpriv->s_tid_retry_timer, hfi1_tid_retry_timeout, 0); |
37356e78 | 342 | INIT_LIST_HEAD(&qpriv->tid_wait); |
48a615dc | 343 | |
838b6fd2 KW |
344 | if (init_attr->qp_type == IB_QPT_RC && HFI1_CAP_IS_KSET(TID_RDMA)) { |
345 | struct hfi1_devdata *dd = qpriv->rcd->dd; | |
346 | ||
347 | qpriv->pages = kzalloc_node(TID_RDMA_MAX_PAGES * | |
348 | sizeof(*qpriv->pages), | |
349 | GFP_KERNEL, dd->node); | |
350 | if (!qpriv->pages) | |
351 | return -ENOMEM; | |
352 | for (i = 0; i < qp->s_size; i++) { | |
353 | struct hfi1_swqe_priv *priv; | |
354 | struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, i); | |
355 | ||
356 | priv = kzalloc_node(sizeof(*priv), GFP_KERNEL, | |
357 | dd->node); | |
358 | if (!priv) | |
359 | return -ENOMEM; | |
360 | ||
361 | hfi1_init_trdma_req(qp, &priv->tid_req); | |
362 | priv->tid_req.e.swqe = wqe; | |
363 | wqe->priv = priv; | |
364 | } | |
365 | for (i = 0; i < rvt_max_atomic(rdi); i++) { | |
366 | struct hfi1_ack_priv *priv; | |
367 | ||
368 | priv = kzalloc_node(sizeof(*priv), GFP_KERNEL, | |
369 | dd->node); | |
370 | if (!priv) | |
371 | return -ENOMEM; | |
372 | ||
373 | hfi1_init_trdma_req(qp, &priv->tid_req); | |
374 | priv->tid_req.e.ack = &qp->s_ack_queue[i]; | |
375 | ||
376 | ret = hfi1_kern_exp_rcv_alloc_flows(&priv->tid_req, | |
377 | GFP_KERNEL); | |
378 | if (ret) { | |
379 | kfree(priv); | |
380 | return ret; | |
381 | } | |
382 | qp->s_ack_queue[i].priv = priv; | |
383 | } | |
384 | } | |
385 | ||
5190f052 MM |
386 | return 0; |
387 | } | |
48a615dc KW |
388 | |
389 | void hfi1_qp_priv_tid_free(struct rvt_dev_info *rdi, struct rvt_qp *qp) | |
390 | { | |
838b6fd2 KW |
391 | struct hfi1_qp_priv *qpriv = qp->priv; |
392 | struct rvt_swqe *wqe; | |
393 | u32 i; | |
394 | ||
395 | if (qp->ibqp.qp_type == IB_QPT_RC && HFI1_CAP_IS_KSET(TID_RDMA)) { | |
396 | for (i = 0; i < qp->s_size; i++) { | |
397 | wqe = rvt_get_swqe_ptr(qp, i); | |
398 | kfree(wqe->priv); | |
399 | wqe->priv = NULL; | |
400 | } | |
401 | for (i = 0; i < rvt_max_atomic(rdi); i++) { | |
402 | struct hfi1_ack_priv *priv = qp->s_ack_queue[i].priv; | |
403 | ||
404 | if (priv) | |
405 | hfi1_kern_exp_rcv_free_flows(&priv->tid_req); | |
406 | kfree(priv); | |
407 | qp->s_ack_queue[i].priv = NULL; | |
408 | } | |
409 | cancel_work_sync(&qpriv->opfn.opfn_work); | |
410 | kfree(qpriv->pages); | |
411 | qpriv->pages = NULL; | |
412 | } | |
48a615dc | 413 | } |
37356e78 KW |
414 | |
415 | /* Flow and tid waiter functions */ | |
416 | /** | |
417 | * DOC: lock ordering | |
418 | * | |
419 | * There are two locks involved with the queuing | |
420 | * routines: the qp s_lock and the exp_lock. | |
421 | * | |
422 | * Since the tid space allocation is called from | |
423 | * the send engine, the qp s_lock is already held. | |
424 | * | |
425 | * The allocation routines will get the exp_lock. | |
426 | * | |
427 | * The first_qp() call is provided to allow the head of | |
428 | * the rcd wait queue to be fetched under the exp_lock and | |
429 | * followed by a drop of the exp_lock. | |
430 | * | |
431 | * Any qp in the wait list will have the qp reference count held | |
432 | * to hold the qp in memory. | |
433 | */ | |
434 | ||
435 | /* | |
436 | * return head of rcd wait list | |
437 | * | |
438 | * Must hold the exp_lock. | |
439 | * | |
440 | * Get a reference to the QP to hold the QP in memory. | |
441 | * | |
442 | * The caller must release the reference when the local | |
443 | * is no longer being used. | |
444 | */ | |
445 | static struct rvt_qp *first_qp(struct hfi1_ctxtdata *rcd, | |
446 | struct tid_queue *queue) | |
447 | __must_hold(&rcd->exp_lock) | |
448 | { | |
449 | struct hfi1_qp_priv *priv; | |
450 | ||
451 | lockdep_assert_held(&rcd->exp_lock); | |
452 | priv = list_first_entry_or_null(&queue->queue_head, | |
453 | struct hfi1_qp_priv, | |
454 | tid_wait); | |
455 | if (!priv) | |
456 | return NULL; | |
457 | rvt_get_qp(priv->owner); | |
458 | return priv->owner; | |
459 | } | |
460 | ||
461 | /** | |
462 | * kernel_tid_waiters - determine rcd wait | |
463 | * @rcd: the receive context | |
464 | * @qp: the head of the qp being processed | |
465 | * | |
466 | * This routine will return false IFF | |
467 | * the list is NULL or the head of the | |
468 | * list is the indicated qp. | |
469 | * | |
470 | * Must hold the qp s_lock and the exp_lock. | |
471 | * | |
472 | * Return: | |
473 | * false if either of the conditions below are statisfied: | |
474 | * 1. The list is empty or | |
475 | * 2. The indicated qp is at the head of the list and the | |
476 | * HFI1_S_WAIT_TID_SPACE bit is set in qp->s_flags. | |
477 | * true is returned otherwise. | |
478 | */ | |
479 | static bool kernel_tid_waiters(struct hfi1_ctxtdata *rcd, | |
480 | struct tid_queue *queue, struct rvt_qp *qp) | |
481 | __must_hold(&rcd->exp_lock) __must_hold(&qp->s_lock) | |
482 | { | |
483 | struct rvt_qp *fqp; | |
484 | bool ret = true; | |
485 | ||
486 | lockdep_assert_held(&qp->s_lock); | |
487 | lockdep_assert_held(&rcd->exp_lock); | |
488 | fqp = first_qp(rcd, queue); | |
489 | if (!fqp || (fqp == qp && (qp->s_flags & HFI1_S_WAIT_TID_SPACE))) | |
490 | ret = false; | |
491 | rvt_put_qp(fqp); | |
492 | return ret; | |
493 | } | |
494 | ||
495 | /** | |
496 | * dequeue_tid_waiter - dequeue the qp from the list | |
497 | * @qp - the qp to remove the wait list | |
498 | * | |
499 | * This routine removes the indicated qp from the | |
500 | * wait list if it is there. | |
501 | * | |
502 | * This should be done after the hardware flow and | |
503 | * tid array resources have been allocated. | |
504 | * | |
505 | * Must hold the qp s_lock and the rcd exp_lock. | |
506 | * | |
507 | * It assumes the s_lock to protect the s_flags | |
508 | * field and to reliably test the HFI1_S_WAIT_TID_SPACE flag. | |
509 | */ | |
510 | static void dequeue_tid_waiter(struct hfi1_ctxtdata *rcd, | |
511 | struct tid_queue *queue, struct rvt_qp *qp) | |
512 | __must_hold(&rcd->exp_lock) __must_hold(&qp->s_lock) | |
513 | { | |
514 | struct hfi1_qp_priv *priv = qp->priv; | |
515 | ||
516 | lockdep_assert_held(&qp->s_lock); | |
517 | lockdep_assert_held(&rcd->exp_lock); | |
518 | if (list_empty(&priv->tid_wait)) | |
519 | return; | |
520 | list_del_init(&priv->tid_wait); | |
521 | qp->s_flags &= ~HFI1_S_WAIT_TID_SPACE; | |
522 | queue->dequeue++; | |
523 | rvt_put_qp(qp); | |
524 | } | |
525 | ||
526 | /** | |
527 | * queue_qp_for_tid_wait - suspend QP on tid space | |
528 | * @rcd: the receive context | |
529 | * @qp: the qp | |
530 | * | |
531 | * The qp is inserted at the tail of the rcd | |
532 | * wait queue and the HFI1_S_WAIT_TID_SPACE s_flag is set. | |
533 | * | |
534 | * Must hold the qp s_lock and the exp_lock. | |
535 | */ | |
536 | static void queue_qp_for_tid_wait(struct hfi1_ctxtdata *rcd, | |
537 | struct tid_queue *queue, struct rvt_qp *qp) | |
538 | __must_hold(&rcd->exp_lock) __must_hold(&qp->s_lock) | |
539 | { | |
540 | struct hfi1_qp_priv *priv = qp->priv; | |
541 | ||
542 | lockdep_assert_held(&qp->s_lock); | |
543 | lockdep_assert_held(&rcd->exp_lock); | |
544 | if (list_empty(&priv->tid_wait)) { | |
545 | qp->s_flags |= HFI1_S_WAIT_TID_SPACE; | |
546 | list_add_tail(&priv->tid_wait, &queue->queue_head); | |
547 | priv->tid_enqueue = ++queue->enqueue; | |
2f16a696 | 548 | rcd->dd->verbs_dev.n_tidwait++; |
37356e78 KW |
549 | trace_hfi1_qpsleep(qp, HFI1_S_WAIT_TID_SPACE); |
550 | rvt_get_qp(qp); | |
551 | } | |
552 | } | |
553 | ||
554 | /** | |
555 | * __trigger_tid_waiter - trigger tid waiter | |
556 | * @qp: the qp | |
557 | * | |
558 | * This is a private entrance to schedule the qp | |
559 | * assuming the caller is holding the qp->s_lock. | |
560 | */ | |
561 | static void __trigger_tid_waiter(struct rvt_qp *qp) | |
562 | __must_hold(&qp->s_lock) | |
563 | { | |
564 | lockdep_assert_held(&qp->s_lock); | |
565 | if (!(qp->s_flags & HFI1_S_WAIT_TID_SPACE)) | |
566 | return; | |
567 | trace_hfi1_qpwakeup(qp, HFI1_S_WAIT_TID_SPACE); | |
568 | hfi1_schedule_send(qp); | |
569 | } | |
570 | ||
571 | /** | |
572 | * tid_rdma_schedule_tid_wakeup - schedule wakeup for a qp | |
573 | * @qp - the qp | |
574 | * | |
575 | * trigger a schedule or a waiting qp in a deadlock | |
576 | * safe manner. The qp reference is held prior | |
577 | * to this call via first_qp(). | |
578 | * | |
579 | * If the qp trigger was already scheduled (!rval) | |
580 | * the the reference is dropped, otherwise the resume | |
581 | * or the destroy cancel will dispatch the reference. | |
582 | */ | |
583 | static void tid_rdma_schedule_tid_wakeup(struct rvt_qp *qp) | |
584 | { | |
585 | struct hfi1_qp_priv *priv; | |
586 | struct hfi1_ibport *ibp; | |
587 | struct hfi1_pportdata *ppd; | |
588 | struct hfi1_devdata *dd; | |
589 | bool rval; | |
590 | ||
591 | if (!qp) | |
592 | return; | |
593 | ||
594 | priv = qp->priv; | |
595 | ibp = to_iport(qp->ibqp.device, qp->port_num); | |
596 | ppd = ppd_from_ibp(ibp); | |
597 | dd = dd_from_ibdev(qp->ibqp.device); | |
598 | ||
599 | rval = queue_work_on(priv->s_sde ? | |
600 | priv->s_sde->cpu : | |
601 | cpumask_first(cpumask_of_node(dd->node)), | |
602 | ppd->hfi1_wq, | |
603 | &priv->tid_rdma.trigger_work); | |
604 | if (!rval) | |
605 | rvt_put_qp(qp); | |
606 | } | |
607 | ||
608 | /** | |
609 | * tid_rdma_trigger_resume - field a trigger work request | |
610 | * @work - the work item | |
611 | * | |
612 | * Complete the off qp trigger processing by directly | |
613 | * calling the progress routine. | |
614 | */ | |
615 | static void tid_rdma_trigger_resume(struct work_struct *work) | |
616 | { | |
617 | struct tid_rdma_qp_params *tr; | |
618 | struct hfi1_qp_priv *priv; | |
619 | struct rvt_qp *qp; | |
620 | ||
621 | tr = container_of(work, struct tid_rdma_qp_params, trigger_work); | |
622 | priv = container_of(tr, struct hfi1_qp_priv, tid_rdma); | |
623 | qp = priv->owner; | |
624 | spin_lock_irq(&qp->s_lock); | |
625 | if (qp->s_flags & HFI1_S_WAIT_TID_SPACE) { | |
626 | spin_unlock_irq(&qp->s_lock); | |
627 | hfi1_do_send(priv->owner, true); | |
628 | } else { | |
629 | spin_unlock_irq(&qp->s_lock); | |
630 | } | |
631 | rvt_put_qp(qp); | |
632 | } | |
633 | ||
634 | /** | |
635 | * tid_rdma_flush_wait - unwind any tid space wait | |
636 | * | |
637 | * This is called when resetting a qp to | |
638 | * allow a destroy or reset to get rid | |
639 | * of any tid space linkage and reference counts. | |
640 | */ | |
641 | static void _tid_rdma_flush_wait(struct rvt_qp *qp, struct tid_queue *queue) | |
642 | __must_hold(&qp->s_lock) | |
643 | { | |
644 | struct hfi1_qp_priv *priv; | |
645 | ||
646 | if (!qp) | |
647 | return; | |
648 | lockdep_assert_held(&qp->s_lock); | |
649 | priv = qp->priv; | |
650 | qp->s_flags &= ~HFI1_S_WAIT_TID_SPACE; | |
651 | spin_lock(&priv->rcd->exp_lock); | |
652 | if (!list_empty(&priv->tid_wait)) { | |
653 | list_del_init(&priv->tid_wait); | |
654 | qp->s_flags &= ~HFI1_S_WAIT_TID_SPACE; | |
655 | queue->dequeue++; | |
656 | rvt_put_qp(qp); | |
657 | } | |
658 | spin_unlock(&priv->rcd->exp_lock); | |
659 | } | |
660 | ||
661 | void hfi1_tid_rdma_flush_wait(struct rvt_qp *qp) | |
662 | __must_hold(&qp->s_lock) | |
663 | { | |
664 | struct hfi1_qp_priv *priv = qp->priv; | |
665 | ||
666 | _tid_rdma_flush_wait(qp, &priv->rcd->flow_queue); | |
838b6fd2 | 667 | _tid_rdma_flush_wait(qp, &priv->rcd->rarr_queue); |
37356e78 KW |
668 | } |
669 | ||
670 | /* Flow functions */ | |
671 | /** | |
672 | * kern_reserve_flow - allocate a hardware flow | |
673 | * @rcd - the context to use for allocation | |
674 | * @last - the index of the preferred flow. Use RXE_NUM_TID_FLOWS to | |
675 | * signify "don't care". | |
676 | * | |
677 | * Use a bit mask based allocation to reserve a hardware | |
678 | * flow for use in receiving KDETH data packets. If a preferred flow is | |
679 | * specified the function will attempt to reserve that flow again, if | |
680 | * available. | |
681 | * | |
682 | * The exp_lock must be held. | |
683 | * | |
684 | * Return: | |
685 | * On success: a value postive value between 0 and RXE_NUM_TID_FLOWS - 1 | |
686 | * On failure: -EAGAIN | |
687 | */ | |
688 | static int kern_reserve_flow(struct hfi1_ctxtdata *rcd, int last) | |
689 | __must_hold(&rcd->exp_lock) | |
690 | { | |
691 | int nr; | |
692 | ||
693 | /* Attempt to reserve the preferred flow index */ | |
694 | if (last >= 0 && last < RXE_NUM_TID_FLOWS && | |
695 | !test_and_set_bit(last, &rcd->flow_mask)) | |
696 | return last; | |
697 | ||
698 | nr = ffz(rcd->flow_mask); | |
699 | BUILD_BUG_ON(RXE_NUM_TID_FLOWS >= | |
700 | (sizeof(rcd->flow_mask) * BITS_PER_BYTE)); | |
701 | if (nr > (RXE_NUM_TID_FLOWS - 1)) | |
702 | return -EAGAIN; | |
703 | set_bit(nr, &rcd->flow_mask); | |
704 | return nr; | |
705 | } | |
706 | ||
707 | static void kern_set_hw_flow(struct hfi1_ctxtdata *rcd, u32 generation, | |
708 | u32 flow_idx) | |
709 | { | |
710 | u64 reg; | |
711 | ||
712 | reg = ((u64)generation << HFI1_KDETH_BTH_SEQ_SHIFT) | | |
713 | RCV_TID_FLOW_TABLE_CTRL_FLOW_VALID_SMASK | | |
714 | RCV_TID_FLOW_TABLE_CTRL_KEEP_AFTER_SEQ_ERR_SMASK | | |
715 | RCV_TID_FLOW_TABLE_CTRL_KEEP_ON_GEN_ERR_SMASK | | |
716 | RCV_TID_FLOW_TABLE_STATUS_SEQ_MISMATCH_SMASK | | |
717 | RCV_TID_FLOW_TABLE_STATUS_GEN_MISMATCH_SMASK; | |
718 | ||
719 | if (generation != KERN_GENERATION_RESERVED) | |
720 | reg |= RCV_TID_FLOW_TABLE_CTRL_HDR_SUPP_EN_SMASK; | |
721 | ||
722 | write_uctxt_csr(rcd->dd, rcd->ctxt, | |
723 | RCV_TID_FLOW_TABLE + 8 * flow_idx, reg); | |
724 | } | |
725 | ||
726 | static u32 kern_setup_hw_flow(struct hfi1_ctxtdata *rcd, u32 flow_idx) | |
727 | __must_hold(&rcd->exp_lock) | |
728 | { | |
729 | u32 generation = rcd->flows[flow_idx].generation; | |
730 | ||
731 | kern_set_hw_flow(rcd, generation, flow_idx); | |
732 | return generation; | |
733 | } | |
734 | ||
735 | static u32 kern_flow_generation_next(u32 gen) | |
736 | { | |
737 | u32 generation = mask_generation(gen + 1); | |
738 | ||
739 | if (generation == KERN_GENERATION_RESERVED) | |
740 | generation = mask_generation(generation + 1); | |
741 | return generation; | |
742 | } | |
743 | ||
744 | static void kern_clear_hw_flow(struct hfi1_ctxtdata *rcd, u32 flow_idx) | |
745 | __must_hold(&rcd->exp_lock) | |
746 | { | |
747 | rcd->flows[flow_idx].generation = | |
748 | kern_flow_generation_next(rcd->flows[flow_idx].generation); | |
749 | kern_set_hw_flow(rcd, KERN_GENERATION_RESERVED, flow_idx); | |
750 | } | |
751 | ||
752 | int hfi1_kern_setup_hw_flow(struct hfi1_ctxtdata *rcd, struct rvt_qp *qp) | |
753 | { | |
754 | struct hfi1_qp_priv *qpriv = (struct hfi1_qp_priv *)qp->priv; | |
755 | struct tid_flow_state *fs = &qpriv->flow_state; | |
756 | struct rvt_qp *fqp; | |
757 | unsigned long flags; | |
758 | int ret = 0; | |
759 | ||
760 | /* The QP already has an allocated flow */ | |
761 | if (fs->index != RXE_NUM_TID_FLOWS) | |
762 | return ret; | |
763 | ||
764 | spin_lock_irqsave(&rcd->exp_lock, flags); | |
765 | if (kernel_tid_waiters(rcd, &rcd->flow_queue, qp)) | |
766 | goto queue; | |
767 | ||
768 | ret = kern_reserve_flow(rcd, fs->last_index); | |
769 | if (ret < 0) | |
770 | goto queue; | |
771 | fs->index = ret; | |
772 | fs->last_index = fs->index; | |
773 | ||
774 | /* Generation received in a RESYNC overrides default flow generation */ | |
775 | if (fs->generation != KERN_GENERATION_RESERVED) | |
776 | rcd->flows[fs->index].generation = fs->generation; | |
777 | fs->generation = kern_setup_hw_flow(rcd, fs->index); | |
778 | fs->psn = 0; | |
779 | fs->flags = 0; | |
780 | dequeue_tid_waiter(rcd, &rcd->flow_queue, qp); | |
781 | /* get head before dropping lock */ | |
782 | fqp = first_qp(rcd, &rcd->flow_queue); | |
783 | spin_unlock_irqrestore(&rcd->exp_lock, flags); | |
784 | ||
785 | tid_rdma_schedule_tid_wakeup(fqp); | |
786 | return 0; | |
787 | queue: | |
788 | queue_qp_for_tid_wait(rcd, &rcd->flow_queue, qp); | |
789 | spin_unlock_irqrestore(&rcd->exp_lock, flags); | |
790 | return -EAGAIN; | |
791 | } | |
792 | ||
793 | void hfi1_kern_clear_hw_flow(struct hfi1_ctxtdata *rcd, struct rvt_qp *qp) | |
794 | { | |
795 | struct hfi1_qp_priv *qpriv = (struct hfi1_qp_priv *)qp->priv; | |
796 | struct tid_flow_state *fs = &qpriv->flow_state; | |
797 | struct rvt_qp *fqp; | |
798 | unsigned long flags; | |
799 | ||
800 | if (fs->index >= RXE_NUM_TID_FLOWS) | |
801 | return; | |
802 | spin_lock_irqsave(&rcd->exp_lock, flags); | |
803 | kern_clear_hw_flow(rcd, fs->index); | |
804 | clear_bit(fs->index, &rcd->flow_mask); | |
805 | fs->index = RXE_NUM_TID_FLOWS; | |
806 | fs->psn = 0; | |
807 | fs->generation = KERN_GENERATION_RESERVED; | |
808 | ||
809 | /* get head before dropping lock */ | |
810 | fqp = first_qp(rcd, &rcd->flow_queue); | |
811 | spin_unlock_irqrestore(&rcd->exp_lock, flags); | |
812 | ||
813 | if (fqp == qp) { | |
814 | __trigger_tid_waiter(fqp); | |
815 | rvt_put_qp(fqp); | |
816 | } else { | |
817 | tid_rdma_schedule_tid_wakeup(fqp); | |
818 | } | |
819 | } | |
820 | ||
821 | void hfi1_kern_init_ctxt_generations(struct hfi1_ctxtdata *rcd) | |
822 | { | |
823 | int i; | |
824 | ||
825 | for (i = 0; i < RXE_NUM_TID_FLOWS; i++) { | |
826 | rcd->flows[i].generation = mask_generation(prandom_u32()); | |
827 | kern_set_hw_flow(rcd, KERN_GENERATION_RESERVED, i); | |
828 | } | |
829 | } | |
838b6fd2 KW |
830 | |
831 | /* TID allocation functions */ | |
832 | static u8 trdma_pset_order(struct tid_rdma_pageset *s) | |
833 | { | |
834 | u8 count = s->count; | |
835 | ||
836 | return ilog2(count) + 1; | |
837 | } | |
838 | ||
839 | /** | |
840 | * tid_rdma_find_phys_blocks_4k - get groups base on mr info | |
841 | * @npages - number of pages | |
842 | * @pages - pointer to an array of page structs | |
843 | * @list - page set array to return | |
844 | * | |
845 | * This routine returns the number of groups associated with | |
846 | * the current sge information. This implementation is based | |
847 | * on the expected receive find_phys_blocks() adjusted to | |
848 | * use the MR information vs. the pfn. | |
849 | * | |
850 | * Return: | |
851 | * the number of RcvArray entries | |
852 | */ | |
853 | static u32 tid_rdma_find_phys_blocks_4k(struct tid_rdma_flow *flow, | |
854 | struct page **pages, | |
855 | u32 npages, | |
856 | struct tid_rdma_pageset *list) | |
857 | { | |
858 | u32 pagecount, pageidx, setcount = 0, i; | |
859 | void *vaddr, *this_vaddr; | |
860 | ||
861 | if (!npages) | |
862 | return 0; | |
863 | ||
864 | /* | |
865 | * Look for sets of physically contiguous pages in the user buffer. | |
866 | * This will allow us to optimize Expected RcvArray entry usage by | |
867 | * using the bigger supported sizes. | |
868 | */ | |
869 | vaddr = page_address(pages[0]); | |
84f4a40d | 870 | trace_hfi1_tid_flow_page(flow->req->qp, flow, 0, 0, 0, vaddr); |
838b6fd2 KW |
871 | for (pageidx = 0, pagecount = 1, i = 1; i <= npages; i++) { |
872 | this_vaddr = i < npages ? page_address(pages[i]) : NULL; | |
84f4a40d KW |
873 | trace_hfi1_tid_flow_page(flow->req->qp, flow, i, 0, 0, |
874 | this_vaddr); | |
838b6fd2 KW |
875 | /* |
876 | * If the vaddr's are not sequential, pages are not physically | |
877 | * contiguous. | |
878 | */ | |
879 | if (this_vaddr != (vaddr + PAGE_SIZE)) { | |
880 | /* | |
881 | * At this point we have to loop over the set of | |
882 | * physically contiguous pages and break them down it | |
883 | * sizes supported by the HW. | |
884 | * There are two main constraints: | |
885 | * 1. The max buffer size is MAX_EXPECTED_BUFFER. | |
886 | * If the total set size is bigger than that | |
887 | * program only a MAX_EXPECTED_BUFFER chunk. | |
888 | * 2. The buffer size has to be a power of two. If | |
889 | * it is not, round down to the closes power of | |
890 | * 2 and program that size. | |
891 | */ | |
892 | while (pagecount) { | |
893 | int maxpages = pagecount; | |
894 | u32 bufsize = pagecount * PAGE_SIZE; | |
895 | ||
896 | if (bufsize > MAX_EXPECTED_BUFFER) | |
897 | maxpages = | |
898 | MAX_EXPECTED_BUFFER >> | |
899 | PAGE_SHIFT; | |
900 | else if (!is_power_of_2(bufsize)) | |
901 | maxpages = | |
902 | rounddown_pow_of_two(bufsize) >> | |
903 | PAGE_SHIFT; | |
904 | ||
905 | list[setcount].idx = pageidx; | |
906 | list[setcount].count = maxpages; | |
84f4a40d KW |
907 | trace_hfi1_tid_pageset(flow->req->qp, setcount, |
908 | list[setcount].idx, | |
909 | list[setcount].count); | |
838b6fd2 KW |
910 | pagecount -= maxpages; |
911 | pageidx += maxpages; | |
912 | setcount++; | |
913 | } | |
914 | pageidx = i; | |
915 | pagecount = 1; | |
916 | vaddr = this_vaddr; | |
917 | } else { | |
918 | vaddr += PAGE_SIZE; | |
919 | pagecount++; | |
920 | } | |
921 | } | |
922 | /* insure we always return an even number of sets */ | |
923 | if (setcount & 1) | |
924 | list[setcount++].count = 0; | |
925 | return setcount; | |
926 | } | |
927 | ||
928 | /** | |
929 | * tid_flush_pages - dump out pages into pagesets | |
930 | * @list - list of pagesets | |
931 | * @idx - pointer to current page index | |
932 | * @pages - number of pages to dump | |
933 | * @sets - current number of pagesset | |
934 | * | |
935 | * This routine flushes out accumuated pages. | |
936 | * | |
937 | * To insure an even number of sets the | |
938 | * code may add a filler. | |
939 | * | |
940 | * This can happen with when pages is not | |
941 | * a power of 2 or pages is a power of 2 | |
942 | * less than the maximum pages. | |
943 | * | |
944 | * Return: | |
945 | * The new number of sets | |
946 | */ | |
947 | ||
948 | static u32 tid_flush_pages(struct tid_rdma_pageset *list, | |
949 | u32 *idx, u32 pages, u32 sets) | |
950 | { | |
951 | while (pages) { | |
952 | u32 maxpages = pages; | |
953 | ||
954 | if (maxpages > MAX_EXPECTED_PAGES) | |
955 | maxpages = MAX_EXPECTED_PAGES; | |
956 | else if (!is_power_of_2(maxpages)) | |
957 | maxpages = rounddown_pow_of_two(maxpages); | |
958 | list[sets].idx = *idx; | |
959 | list[sets++].count = maxpages; | |
960 | *idx += maxpages; | |
961 | pages -= maxpages; | |
962 | } | |
963 | /* might need a filler */ | |
964 | if (sets & 1) | |
965 | list[sets++].count = 0; | |
966 | return sets; | |
967 | } | |
968 | ||
969 | /** | |
970 | * tid_rdma_find_phys_blocks_8k - get groups base on mr info | |
971 | * @pages - pointer to an array of page structs | |
972 | * @npages - number of pages | |
973 | * @list - page set array to return | |
974 | * | |
975 | * This routine parses an array of pages to compute pagesets | |
976 | * in an 8k compatible way. | |
977 | * | |
978 | * pages are tested two at a time, i, i + 1 for contiguous | |
979 | * pages and i - 1 and i contiguous pages. | |
980 | * | |
981 | * If any condition is false, any accumlated pages are flushed and | |
982 | * v0,v1 are emitted as separate PAGE_SIZE pagesets | |
983 | * | |
984 | * Otherwise, the current 8k is totaled for a future flush. | |
985 | * | |
986 | * Return: | |
987 | * The number of pagesets | |
988 | * list set with the returned number of pagesets | |
989 | * | |
990 | */ | |
991 | static u32 tid_rdma_find_phys_blocks_8k(struct tid_rdma_flow *flow, | |
992 | struct page **pages, | |
993 | u32 npages, | |
994 | struct tid_rdma_pageset *list) | |
995 | { | |
996 | u32 idx, sets = 0, i; | |
997 | u32 pagecnt = 0; | |
998 | void *v0, *v1, *vm1; | |
999 | ||
1000 | if (!npages) | |
1001 | return 0; | |
1002 | for (idx = 0, i = 0, vm1 = NULL; i < npages; i += 2) { | |
1003 | /* get a new v0 */ | |
1004 | v0 = page_address(pages[i]); | |
84f4a40d | 1005 | trace_hfi1_tid_flow_page(flow->req->qp, flow, i, 1, 0, v0); |
838b6fd2 KW |
1006 | v1 = i + 1 < npages ? |
1007 | page_address(pages[i + 1]) : NULL; | |
84f4a40d | 1008 | trace_hfi1_tid_flow_page(flow->req->qp, flow, i, 1, 1, v1); |
838b6fd2 KW |
1009 | /* compare i, i + 1 vaddr */ |
1010 | if (v1 != (v0 + PAGE_SIZE)) { | |
1011 | /* flush out pages */ | |
1012 | sets = tid_flush_pages(list, &idx, pagecnt, sets); | |
1013 | /* output v0,v1 as two pagesets */ | |
1014 | list[sets].idx = idx++; | |
1015 | list[sets++].count = 1; | |
1016 | if (v1) { | |
1017 | list[sets].count = 1; | |
1018 | list[sets++].idx = idx++; | |
1019 | } else { | |
1020 | list[sets++].count = 0; | |
1021 | } | |
1022 | vm1 = NULL; | |
1023 | pagecnt = 0; | |
1024 | continue; | |
1025 | } | |
1026 | /* i,i+1 consecutive, look at i-1,i */ | |
1027 | if (vm1 && v0 != (vm1 + PAGE_SIZE)) { | |
1028 | /* flush out pages */ | |
1029 | sets = tid_flush_pages(list, &idx, pagecnt, sets); | |
1030 | pagecnt = 0; | |
1031 | } | |
1032 | /* pages will always be a multiple of 8k */ | |
1033 | pagecnt += 2; | |
1034 | /* save i-1 */ | |
1035 | vm1 = v1; | |
1036 | /* move to next pair */ | |
1037 | } | |
1038 | /* dump residual pages at end */ | |
1039 | sets = tid_flush_pages(list, &idx, npages - idx, sets); | |
1040 | /* by design cannot be odd sets */ | |
1041 | WARN_ON(sets & 1); | |
1042 | return sets; | |
1043 | } | |
1044 | ||
1045 | /** | |
1046 | * Find pages for one segment of a sge array represented by @ss. The function | |
1047 | * does not check the sge, the sge must have been checked for alignment with a | |
1048 | * prior call to hfi1_kern_trdma_ok. Other sge checking is done as part of | |
1049 | * rvt_lkey_ok and rvt_rkey_ok. Also, the function only modifies the local sge | |
1050 | * copy maintained in @ss->sge, the original sge is not modified. | |
1051 | * | |
1052 | * Unlike IB RDMA WRITE, we can't decrement ss->num_sge here because we are not | |
1053 | * releasing the MR reference count at the same time. Otherwise, we'll "leak" | |
1054 | * references to the MR. This difference requires that we keep track of progress | |
1055 | * into the sg_list. This is done by the cur_seg cursor in the tid_rdma_request | |
1056 | * structure. | |
1057 | */ | |
1058 | static u32 kern_find_pages(struct tid_rdma_flow *flow, | |
1059 | struct page **pages, | |
1060 | struct rvt_sge_state *ss, bool *last) | |
1061 | { | |
1062 | struct tid_rdma_request *req = flow->req; | |
1063 | struct rvt_sge *sge = &ss->sge; | |
1064 | u32 length = flow->req->seg_len; | |
1065 | u32 len = PAGE_SIZE; | |
1066 | u32 i = 0; | |
1067 | ||
1068 | while (length && req->isge < ss->num_sge) { | |
1069 | pages[i++] = virt_to_page(sge->vaddr); | |
1070 | ||
1071 | sge->vaddr += len; | |
1072 | sge->length -= len; | |
1073 | sge->sge_length -= len; | |
1074 | if (!sge->sge_length) { | |
1075 | if (++req->isge < ss->num_sge) | |
1076 | *sge = ss->sg_list[req->isge - 1]; | |
1077 | } else if (sge->length == 0 && sge->mr->lkey) { | |
1078 | if (++sge->n >= RVT_SEGSZ) { | |
1079 | ++sge->m; | |
1080 | sge->n = 0; | |
1081 | } | |
1082 | sge->vaddr = sge->mr->map[sge->m]->segs[sge->n].vaddr; | |
1083 | sge->length = sge->mr->map[sge->m]->segs[sge->n].length; | |
1084 | } | |
1085 | length -= len; | |
1086 | } | |
1087 | ||
1088 | flow->length = flow->req->seg_len - length; | |
1089 | *last = req->isge == ss->num_sge ? false : true; | |
1090 | return i; | |
1091 | } | |
1092 | ||
1093 | static void dma_unmap_flow(struct tid_rdma_flow *flow) | |
1094 | { | |
1095 | struct hfi1_devdata *dd; | |
1096 | int i; | |
1097 | struct tid_rdma_pageset *pset; | |
1098 | ||
1099 | dd = flow->req->rcd->dd; | |
1100 | for (i = 0, pset = &flow->pagesets[0]; i < flow->npagesets; | |
1101 | i++, pset++) { | |
1102 | if (pset->count && pset->addr) { | |
1103 | dma_unmap_page(&dd->pcidev->dev, | |
1104 | pset->addr, | |
1105 | PAGE_SIZE * pset->count, | |
1106 | DMA_FROM_DEVICE); | |
1107 | pset->mapped = 0; | |
1108 | } | |
1109 | } | |
1110 | } | |
1111 | ||
1112 | static int dma_map_flow(struct tid_rdma_flow *flow, struct page **pages) | |
1113 | { | |
1114 | int i; | |
1115 | struct hfi1_devdata *dd = flow->req->rcd->dd; | |
1116 | struct tid_rdma_pageset *pset; | |
1117 | ||
1118 | for (i = 0, pset = &flow->pagesets[0]; i < flow->npagesets; | |
1119 | i++, pset++) { | |
1120 | if (pset->count) { | |
1121 | pset->addr = dma_map_page(&dd->pcidev->dev, | |
1122 | pages[pset->idx], | |
1123 | 0, | |
1124 | PAGE_SIZE * pset->count, | |
1125 | DMA_FROM_DEVICE); | |
1126 | ||
1127 | if (dma_mapping_error(&dd->pcidev->dev, pset->addr)) { | |
1128 | dma_unmap_flow(flow); | |
1129 | return -ENOMEM; | |
1130 | } | |
1131 | pset->mapped = 1; | |
1132 | } | |
1133 | } | |
1134 | return 0; | |
1135 | } | |
1136 | ||
1137 | static inline bool dma_mapped(struct tid_rdma_flow *flow) | |
1138 | { | |
1139 | return !!flow->pagesets[0].mapped; | |
1140 | } | |
1141 | ||
1142 | /* | |
1143 | * Get pages pointers and identify contiguous physical memory chunks for a | |
1144 | * segment. All segments are of length flow->req->seg_len. | |
1145 | */ | |
1146 | static int kern_get_phys_blocks(struct tid_rdma_flow *flow, | |
1147 | struct page **pages, | |
1148 | struct rvt_sge_state *ss, bool *last) | |
1149 | { | |
1150 | u8 npages; | |
1151 | ||
1152 | /* Reuse previously computed pagesets, if any */ | |
1153 | if (flow->npagesets) { | |
84f4a40d KW |
1154 | trace_hfi1_tid_flow_alloc(flow->req->qp, flow->req->setup_head, |
1155 | flow); | |
838b6fd2 KW |
1156 | if (!dma_mapped(flow)) |
1157 | return dma_map_flow(flow, pages); | |
1158 | return 0; | |
1159 | } | |
1160 | ||
1161 | npages = kern_find_pages(flow, pages, ss, last); | |
1162 | ||
1163 | if (flow->req->qp->pmtu == enum_to_mtu(OPA_MTU_4096)) | |
1164 | flow->npagesets = | |
1165 | tid_rdma_find_phys_blocks_4k(flow, pages, npages, | |
1166 | flow->pagesets); | |
1167 | else | |
1168 | flow->npagesets = | |
1169 | tid_rdma_find_phys_blocks_8k(flow, pages, npages, | |
1170 | flow->pagesets); | |
1171 | ||
1172 | return dma_map_flow(flow, pages); | |
1173 | } | |
1174 | ||
1175 | static inline void kern_add_tid_node(struct tid_rdma_flow *flow, | |
1176 | struct hfi1_ctxtdata *rcd, char *s, | |
1177 | struct tid_group *grp, u8 cnt) | |
1178 | { | |
1179 | struct kern_tid_node *node = &flow->tnode[flow->tnode_cnt++]; | |
1180 | ||
1181 | WARN_ON_ONCE(flow->tnode_cnt >= | |
1182 | (TID_RDMA_MAX_SEGMENT_SIZE >> PAGE_SHIFT)); | |
1183 | if (WARN_ON_ONCE(cnt & 1)) | |
1184 | dd_dev_err(rcd->dd, | |
1185 | "unexpected odd allocation cnt %u map 0x%x used %u", | |
1186 | cnt, grp->map, grp->used); | |
1187 | ||
1188 | node->grp = grp; | |
1189 | node->map = grp->map; | |
1190 | node->cnt = cnt; | |
84f4a40d KW |
1191 | trace_hfi1_tid_node_add(flow->req->qp, s, flow->tnode_cnt - 1, |
1192 | grp->base, grp->map, grp->used, cnt); | |
838b6fd2 KW |
1193 | } |
1194 | ||
1195 | /* | |
1196 | * Try to allocate pageset_count TID's from TID groups for a context | |
1197 | * | |
1198 | * This function allocates TID's without moving groups between lists or | |
1199 | * modifying grp->map. This is done as follows, being cogizant of the lists | |
1200 | * between which the TID groups will move: | |
1201 | * 1. First allocate complete groups of 8 TID's since this is more efficient, | |
1202 | * these groups will move from group->full without affecting used | |
1203 | * 2. If more TID's are needed allocate from used (will move from used->full or | |
1204 | * stay in used) | |
1205 | * 3. If we still don't have the required number of TID's go back and look again | |
1206 | * at a complete group (will move from group->used) | |
1207 | */ | |
1208 | static int kern_alloc_tids(struct tid_rdma_flow *flow) | |
1209 | { | |
1210 | struct hfi1_ctxtdata *rcd = flow->req->rcd; | |
1211 | struct hfi1_devdata *dd = rcd->dd; | |
1212 | u32 ngroups, pageidx = 0; | |
1213 | struct tid_group *group = NULL, *used; | |
1214 | u8 use; | |
1215 | ||
1216 | flow->tnode_cnt = 0; | |
1217 | ngroups = flow->npagesets / dd->rcv_entries.group_size; | |
1218 | if (!ngroups) | |
1219 | goto used_list; | |
1220 | ||
1221 | /* First look at complete groups */ | |
1222 | list_for_each_entry(group, &rcd->tid_group_list.list, list) { | |
1223 | kern_add_tid_node(flow, rcd, "complete groups", group, | |
1224 | group->size); | |
1225 | ||
1226 | pageidx += group->size; | |
1227 | if (!--ngroups) | |
1228 | break; | |
1229 | } | |
1230 | ||
1231 | if (pageidx >= flow->npagesets) | |
1232 | goto ok; | |
1233 | ||
1234 | used_list: | |
1235 | /* Now look at partially used groups */ | |
1236 | list_for_each_entry(used, &rcd->tid_used_list.list, list) { | |
1237 | use = min_t(u32, flow->npagesets - pageidx, | |
1238 | used->size - used->used); | |
1239 | kern_add_tid_node(flow, rcd, "used groups", used, use); | |
1240 | ||
1241 | pageidx += use; | |
1242 | if (pageidx >= flow->npagesets) | |
1243 | goto ok; | |
1244 | } | |
1245 | ||
1246 | /* | |
1247 | * Look again at a complete group, continuing from where we left. | |
1248 | * However, if we are at the head, we have reached the end of the | |
1249 | * complete groups list from the first loop above | |
1250 | */ | |
1251 | if (group && &group->list == &rcd->tid_group_list.list) | |
1252 | goto bail_eagain; | |
1253 | group = list_prepare_entry(group, &rcd->tid_group_list.list, | |
1254 | list); | |
1255 | if (list_is_last(&group->list, &rcd->tid_group_list.list)) | |
1256 | goto bail_eagain; | |
1257 | group = list_next_entry(group, list); | |
1258 | use = min_t(u32, flow->npagesets - pageidx, group->size); | |
1259 | kern_add_tid_node(flow, rcd, "complete continue", group, use); | |
1260 | pageidx += use; | |
1261 | if (pageidx >= flow->npagesets) | |
1262 | goto ok; | |
1263 | bail_eagain: | |
84f4a40d KW |
1264 | trace_hfi1_msg_alloc_tids(flow->req->qp, " insufficient tids: needed ", |
1265 | (u64)flow->npagesets); | |
838b6fd2 KW |
1266 | return -EAGAIN; |
1267 | ok: | |
1268 | return 0; | |
1269 | } | |
1270 | ||
1271 | static void kern_program_rcv_group(struct tid_rdma_flow *flow, int grp_num, | |
1272 | u32 *pset_idx) | |
1273 | { | |
1274 | struct hfi1_ctxtdata *rcd = flow->req->rcd; | |
1275 | struct hfi1_devdata *dd = rcd->dd; | |
1276 | struct kern_tid_node *node = &flow->tnode[grp_num]; | |
1277 | struct tid_group *grp = node->grp; | |
1278 | struct tid_rdma_pageset *pset; | |
1279 | u32 pmtu_pg = flow->req->qp->pmtu >> PAGE_SHIFT; | |
1280 | u32 rcventry, npages = 0, pair = 0, tidctrl; | |
1281 | u8 i, cnt = 0; | |
1282 | ||
1283 | for (i = 0; i < grp->size; i++) { | |
1284 | rcventry = grp->base + i; | |
1285 | ||
1286 | if (node->map & BIT(i) || cnt >= node->cnt) { | |
1287 | rcv_array_wc_fill(dd, rcventry); | |
1288 | continue; | |
1289 | } | |
1290 | pset = &flow->pagesets[(*pset_idx)++]; | |
1291 | if (pset->count) { | |
1292 | hfi1_put_tid(dd, rcventry, PT_EXPECTED, | |
1293 | pset->addr, trdma_pset_order(pset)); | |
1294 | } else { | |
1295 | hfi1_put_tid(dd, rcventry, PT_INVALID, 0, 0); | |
1296 | } | |
1297 | npages += pset->count; | |
1298 | ||
1299 | rcventry -= rcd->expected_base; | |
1300 | tidctrl = pair ? 0x3 : rcventry & 0x1 ? 0x2 : 0x1; | |
1301 | /* | |
1302 | * A single TID entry will be used to use a rcvarr pair (with | |
1303 | * tidctrl 0x3), if ALL these are true (a) the bit pos is even | |
1304 | * (b) the group map shows current and the next bits as free | |
1305 | * indicating two consecutive rcvarry entries are available (c) | |
1306 | * we actually need 2 more entries | |
1307 | */ | |
1308 | pair = !(i & 0x1) && !((node->map >> i) & 0x3) && | |
1309 | node->cnt >= cnt + 2; | |
1310 | if (!pair) { | |
1311 | if (!pset->count) | |
1312 | tidctrl = 0x1; | |
1313 | flow->tid_entry[flow->tidcnt++] = | |
1314 | EXP_TID_SET(IDX, rcventry >> 1) | | |
1315 | EXP_TID_SET(CTRL, tidctrl) | | |
1316 | EXP_TID_SET(LEN, npages); | |
84f4a40d KW |
1317 | trace_hfi1_tid_entry_alloc(/* entry */ |
1318 | flow->req->qp, flow->tidcnt - 1, | |
1319 | flow->tid_entry[flow->tidcnt - 1]); | |
1320 | ||
838b6fd2 KW |
1321 | /* Efficient DIV_ROUND_UP(npages, pmtu_pg) */ |
1322 | flow->npkts += (npages + pmtu_pg - 1) >> ilog2(pmtu_pg); | |
1323 | npages = 0; | |
1324 | } | |
1325 | ||
1326 | if (grp->used == grp->size - 1) | |
1327 | tid_group_move(grp, &rcd->tid_used_list, | |
1328 | &rcd->tid_full_list); | |
1329 | else if (!grp->used) | |
1330 | tid_group_move(grp, &rcd->tid_group_list, | |
1331 | &rcd->tid_used_list); | |
1332 | ||
1333 | grp->used++; | |
1334 | grp->map |= BIT(i); | |
1335 | cnt++; | |
1336 | } | |
1337 | } | |
1338 | ||
1339 | static void kern_unprogram_rcv_group(struct tid_rdma_flow *flow, int grp_num) | |
1340 | { | |
1341 | struct hfi1_ctxtdata *rcd = flow->req->rcd; | |
1342 | struct hfi1_devdata *dd = rcd->dd; | |
1343 | struct kern_tid_node *node = &flow->tnode[grp_num]; | |
1344 | struct tid_group *grp = node->grp; | |
1345 | u32 rcventry; | |
1346 | u8 i, cnt = 0; | |
1347 | ||
1348 | for (i = 0; i < grp->size; i++) { | |
1349 | rcventry = grp->base + i; | |
1350 | ||
1351 | if (node->map & BIT(i) || cnt >= node->cnt) { | |
1352 | rcv_array_wc_fill(dd, rcventry); | |
1353 | continue; | |
1354 | } | |
1355 | ||
1356 | hfi1_put_tid(dd, rcventry, PT_INVALID, 0, 0); | |
1357 | ||
1358 | grp->used--; | |
1359 | grp->map &= ~BIT(i); | |
1360 | cnt++; | |
1361 | ||
1362 | if (grp->used == grp->size - 1) | |
1363 | tid_group_move(grp, &rcd->tid_full_list, | |
1364 | &rcd->tid_used_list); | |
1365 | else if (!grp->used) | |
1366 | tid_group_move(grp, &rcd->tid_used_list, | |
1367 | &rcd->tid_group_list); | |
1368 | } | |
1369 | if (WARN_ON_ONCE(cnt & 1)) { | |
1370 | struct hfi1_ctxtdata *rcd = flow->req->rcd; | |
1371 | struct hfi1_devdata *dd = rcd->dd; | |
1372 | ||
1373 | dd_dev_err(dd, "unexpected odd free cnt %u map 0x%x used %u", | |
1374 | cnt, grp->map, grp->used); | |
1375 | } | |
1376 | } | |
1377 | ||
1378 | static void kern_program_rcvarray(struct tid_rdma_flow *flow) | |
1379 | { | |
1380 | u32 pset_idx = 0; | |
1381 | int i; | |
1382 | ||
1383 | flow->npkts = 0; | |
1384 | flow->tidcnt = 0; | |
1385 | for (i = 0; i < flow->tnode_cnt; i++) | |
1386 | kern_program_rcv_group(flow, i, &pset_idx); | |
84f4a40d | 1387 | trace_hfi1_tid_flow_alloc(flow->req->qp, flow->req->setup_head, flow); |
838b6fd2 KW |
1388 | } |
1389 | ||
1390 | /** | |
1391 | * hfi1_kern_exp_rcv_setup() - setup TID's and flow for one segment of a | |
1392 | * TID RDMA request | |
1393 | * | |
1394 | * @req: TID RDMA request for which the segment/flow is being set up | |
1395 | * @ss: sge state, maintains state across successive segments of a sge | |
1396 | * @last: set to true after the last sge segment has been processed | |
1397 | * | |
1398 | * This function | |
1399 | * (1) finds a free flow entry in the flow circular buffer | |
1400 | * (2) finds pages and continuous physical chunks constituing one segment | |
1401 | * of an sge | |
1402 | * (3) allocates TID group entries for those chunks | |
1403 | * (4) programs rcvarray entries in the hardware corresponding to those | |
1404 | * TID's | |
1405 | * (5) computes a tidarray with formatted TID entries which can be sent | |
1406 | * to the sender | |
1407 | * (6) Reserves and programs HW flows. | |
1408 | * (7) It also manages queing the QP when TID/flow resources are not | |
1409 | * available. | |
1410 | * | |
1411 | * @req points to struct tid_rdma_request of which the segments are a part. The | |
1412 | * function uses qp, rcd and seg_len members of @req. In the absence of errors, | |
1413 | * req->flow_idx is the index of the flow which has been prepared in this | |
1414 | * invocation of function call. With flow = &req->flows[req->flow_idx], | |
1415 | * flow->tid_entry contains the TID array which the sender can use for TID RDMA | |
1416 | * sends and flow->npkts contains number of packets required to send the | |
1417 | * segment. | |
1418 | * | |
1419 | * hfi1_check_sge_align should be called prior to calling this function and if | |
1420 | * it signals error TID RDMA cannot be used for this sge and this function | |
1421 | * should not be called. | |
1422 | * | |
1423 | * For the queuing, caller must hold the flow->req->qp s_lock from the send | |
1424 | * engine and the function will procure the exp_lock. | |
1425 | * | |
1426 | * Return: | |
1427 | * The function returns -EAGAIN if sufficient number of TID/flow resources to | |
1428 | * map the segment could not be allocated. In this case the function should be | |
1429 | * called again with previous arguments to retry the TID allocation. There are | |
1430 | * no other error returns. The function returns 0 on success. | |
1431 | */ | |
1432 | int hfi1_kern_exp_rcv_setup(struct tid_rdma_request *req, | |
1433 | struct rvt_sge_state *ss, bool *last) | |
1434 | __must_hold(&req->qp->s_lock) | |
1435 | { | |
1436 | struct tid_rdma_flow *flow = &req->flows[req->setup_head]; | |
1437 | struct hfi1_ctxtdata *rcd = req->rcd; | |
1438 | struct hfi1_qp_priv *qpriv = req->qp->priv; | |
1439 | unsigned long flags; | |
1440 | struct rvt_qp *fqp; | |
1441 | u16 clear_tail = req->clear_tail; | |
1442 | ||
1443 | lockdep_assert_held(&req->qp->s_lock); | |
1444 | /* | |
1445 | * We return error if either (a) we don't have space in the flow | |
1446 | * circular buffer, or (b) we already have max entries in the buffer. | |
1447 | * Max entries depend on the type of request we are processing and the | |
1448 | * negotiated TID RDMA parameters. | |
1449 | */ | |
1450 | if (!CIRC_SPACE(req->setup_head, clear_tail, MAX_FLOWS) || | |
1451 | CIRC_CNT(req->setup_head, clear_tail, MAX_FLOWS) >= | |
1452 | req->n_flows) | |
1453 | return -EINVAL; | |
1454 | ||
1455 | /* | |
1456 | * Get pages, identify contiguous physical memory chunks for the segment | |
1457 | * If we can not determine a DMA address mapping we will treat it just | |
1458 | * like if we ran out of space above. | |
1459 | */ | |
1460 | if (kern_get_phys_blocks(flow, qpriv->pages, ss, last)) { | |
1461 | hfi1_wait_kmem(flow->req->qp); | |
1462 | return -ENOMEM; | |
1463 | } | |
1464 | ||
1465 | spin_lock_irqsave(&rcd->exp_lock, flags); | |
1466 | if (kernel_tid_waiters(rcd, &rcd->rarr_queue, flow->req->qp)) | |
1467 | goto queue; | |
1468 | ||
1469 | /* | |
1470 | * At this point we know the number of pagesets and hence the number of | |
1471 | * TID's to map the segment. Allocate the TID's from the TID groups. If | |
1472 | * we cannot allocate the required number we exit and try again later | |
1473 | */ | |
1474 | if (kern_alloc_tids(flow)) | |
1475 | goto queue; | |
1476 | /* | |
1477 | * Finally program the TID entries with the pagesets, compute the | |
1478 | * tidarray and enable the HW flow | |
1479 | */ | |
1480 | kern_program_rcvarray(flow); | |
1481 | ||
1482 | /* | |
1483 | * Setup the flow state with relevant information. | |
1484 | * This information is used for tracking the sequence of data packets | |
1485 | * for the segment. | |
1486 | * The flow is setup here as this is the most accurate time and place | |
1487 | * to do so. Doing at a later time runs the risk of the flow data in | |
1488 | * qpriv getting out of sync. | |
1489 | */ | |
1490 | memset(&flow->flow_state, 0x0, sizeof(flow->flow_state)); | |
1491 | flow->idx = qpriv->flow_state.index; | |
1492 | flow->flow_state.generation = qpriv->flow_state.generation; | |
1493 | flow->flow_state.spsn = qpriv->flow_state.psn; | |
1494 | flow->flow_state.lpsn = flow->flow_state.spsn + flow->npkts - 1; | |
1495 | flow->flow_state.r_next_psn = | |
1496 | full_flow_psn(flow, flow->flow_state.spsn); | |
1497 | qpriv->flow_state.psn += flow->npkts; | |
1498 | ||
1499 | dequeue_tid_waiter(rcd, &rcd->rarr_queue, flow->req->qp); | |
1500 | /* get head before dropping lock */ | |
1501 | fqp = first_qp(rcd, &rcd->rarr_queue); | |
1502 | spin_unlock_irqrestore(&rcd->exp_lock, flags); | |
1503 | tid_rdma_schedule_tid_wakeup(fqp); | |
1504 | ||
1505 | req->setup_head = (req->setup_head + 1) & (MAX_FLOWS - 1); | |
1506 | return 0; | |
1507 | queue: | |
1508 | queue_qp_for_tid_wait(rcd, &rcd->rarr_queue, flow->req->qp); | |
1509 | spin_unlock_irqrestore(&rcd->exp_lock, flags); | |
1510 | return -EAGAIN; | |
1511 | } | |
1512 | ||
1513 | static void hfi1_tid_rdma_reset_flow(struct tid_rdma_flow *flow) | |
1514 | { | |
1515 | flow->npagesets = 0; | |
1516 | } | |
1517 | ||
1518 | /* | |
1519 | * This function is called after one segment has been successfully sent to | |
1520 | * release the flow and TID HW/SW resources for that segment. The segments for a | |
1521 | * TID RDMA request are setup and cleared in FIFO order which is managed using a | |
1522 | * circular buffer. | |
1523 | */ | |
1524 | int hfi1_kern_exp_rcv_clear(struct tid_rdma_request *req) | |
1525 | __must_hold(&req->qp->s_lock) | |
1526 | { | |
1527 | struct tid_rdma_flow *flow = &req->flows[req->clear_tail]; | |
1528 | struct hfi1_ctxtdata *rcd = req->rcd; | |
1529 | unsigned long flags; | |
1530 | int i; | |
1531 | struct rvt_qp *fqp; | |
1532 | ||
1533 | lockdep_assert_held(&req->qp->s_lock); | |
1534 | /* Exit if we have nothing in the flow circular buffer */ | |
1535 | if (!CIRC_CNT(req->setup_head, req->clear_tail, MAX_FLOWS)) | |
1536 | return -EINVAL; | |
1537 | ||
1538 | spin_lock_irqsave(&rcd->exp_lock, flags); | |
1539 | ||
1540 | for (i = 0; i < flow->tnode_cnt; i++) | |
1541 | kern_unprogram_rcv_group(flow, i); | |
1542 | /* To prevent double unprogramming */ | |
1543 | flow->tnode_cnt = 0; | |
1544 | /* get head before dropping lock */ | |
1545 | fqp = first_qp(rcd, &rcd->rarr_queue); | |
1546 | spin_unlock_irqrestore(&rcd->exp_lock, flags); | |
1547 | ||
1548 | dma_unmap_flow(flow); | |
1549 | ||
1550 | hfi1_tid_rdma_reset_flow(flow); | |
1551 | req->clear_tail = (req->clear_tail + 1) & (MAX_FLOWS - 1); | |
1552 | ||
1553 | if (fqp == req->qp) { | |
1554 | __trigger_tid_waiter(fqp); | |
1555 | rvt_put_qp(fqp); | |
1556 | } else { | |
1557 | tid_rdma_schedule_tid_wakeup(fqp); | |
1558 | } | |
1559 | ||
1560 | return 0; | |
1561 | } | |
1562 | ||
1563 | /* | |
1564 | * This function is called to release all the tid entries for | |
1565 | * a request. | |
1566 | */ | |
1567 | void hfi1_kern_exp_rcv_clear_all(struct tid_rdma_request *req) | |
1568 | __must_hold(&req->qp->s_lock) | |
1569 | { | |
1570 | /* Use memory barrier for proper ordering */ | |
1571 | while (CIRC_CNT(req->setup_head, req->clear_tail, MAX_FLOWS)) { | |
1572 | if (hfi1_kern_exp_rcv_clear(req)) | |
1573 | break; | |
1574 | } | |
1575 | } | |
1576 | ||
1577 | /** | |
1578 | * hfi1_kern_exp_rcv_free_flows - free priviously allocated flow information | |
1579 | * @req - the tid rdma request to be cleaned | |
1580 | */ | |
1581 | static void hfi1_kern_exp_rcv_free_flows(struct tid_rdma_request *req) | |
1582 | { | |
1583 | kfree(req->flows); | |
1584 | req->flows = NULL; | |
1585 | } | |
1586 | ||
1587 | /** | |
1588 | * __trdma_clean_swqe - clean up for large sized QPs | |
1589 | * @qp: the queue patch | |
1590 | * @wqe: the send wqe | |
1591 | */ | |
1592 | void __trdma_clean_swqe(struct rvt_qp *qp, struct rvt_swqe *wqe) | |
1593 | { | |
1594 | struct hfi1_swqe_priv *p = wqe->priv; | |
1595 | ||
1596 | hfi1_kern_exp_rcv_free_flows(&p->tid_req); | |
1597 | } | |
1598 | ||
1599 | /* | |
1600 | * This can be called at QP create time or in the data path. | |
1601 | */ | |
1602 | static int hfi1_kern_exp_rcv_alloc_flows(struct tid_rdma_request *req, | |
1603 | gfp_t gfp) | |
1604 | { | |
1605 | struct tid_rdma_flow *flows; | |
1606 | int i; | |
1607 | ||
1608 | if (likely(req->flows)) | |
1609 | return 0; | |
1610 | flows = kmalloc_node(MAX_FLOWS * sizeof(*flows), gfp, | |
1611 | req->rcd->numa_id); | |
1612 | if (!flows) | |
1613 | return -ENOMEM; | |
1614 | /* mini init */ | |
1615 | for (i = 0; i < MAX_FLOWS; i++) { | |
1616 | flows[i].req = req; | |
1617 | flows[i].npagesets = 0; | |
1618 | flows[i].pagesets[0].mapped = 0; | |
1619 | } | |
1620 | req->flows = flows; | |
1621 | return 0; | |
1622 | } | |
1623 | ||
1624 | static void hfi1_init_trdma_req(struct rvt_qp *qp, | |
1625 | struct tid_rdma_request *req) | |
1626 | { | |
1627 | struct hfi1_qp_priv *qpriv = qp->priv; | |
1628 | ||
1629 | /* | |
1630 | * Initialize various TID RDMA request variables. | |
1631 | * These variables are "static", which is why they | |
1632 | * can be pre-initialized here before the WRs has | |
1633 | * even been submitted. | |
1634 | * However, non-NULL values for these variables do not | |
1635 | * imply that this WQE has been enabled for TID RDMA. | |
1636 | * Drivers should check the WQE's opcode to determine | |
1637 | * if a request is a TID RDMA one or not. | |
1638 | */ | |
1639 | req->qp = qp; | |
1640 | req->rcd = qpriv->rcd; | |
1641 | } | |
2f16a696 KW |
1642 | |
1643 | u64 hfi1_access_sw_tid_wait(const struct cntr_entry *entry, | |
1644 | void *context, int vl, int mode, u64 data) | |
1645 | { | |
1646 | struct hfi1_devdata *dd = context; | |
1647 | ||
1648 | return dd->verbs_dev.n_tidwait; | |
1649 | } | |
742a3826 | 1650 | |
b126078e KW |
1651 | static struct tid_rdma_flow *find_flow_ib(struct tid_rdma_request *req, |
1652 | u32 psn, u16 *fidx) | |
1653 | { | |
1654 | u16 head, tail; | |
1655 | struct tid_rdma_flow *flow; | |
1656 | ||
1657 | head = req->setup_head; | |
1658 | tail = req->clear_tail; | |
1659 | for ( ; CIRC_CNT(head, tail, MAX_FLOWS); | |
1660 | tail = CIRC_NEXT(tail, MAX_FLOWS)) { | |
1661 | flow = &req->flows[tail]; | |
1662 | if (cmp_psn(psn, flow->flow_state.ib_spsn) >= 0 && | |
1663 | cmp_psn(psn, flow->flow_state.ib_lpsn) <= 0) { | |
1664 | if (fidx) | |
1665 | *fidx = tail; | |
1666 | return flow; | |
1667 | } | |
1668 | } | |
1669 | return NULL; | |
1670 | } | |
1671 | ||
9905bf06 KW |
1672 | static struct tid_rdma_flow * |
1673 | __find_flow_ranged(struct tid_rdma_request *req, u16 head, u16 tail, | |
1674 | u32 psn, u16 *fidx) | |
1675 | { | |
1676 | for ( ; CIRC_CNT(head, tail, MAX_FLOWS); | |
1677 | tail = CIRC_NEXT(tail, MAX_FLOWS)) { | |
1678 | struct tid_rdma_flow *flow = &req->flows[tail]; | |
1679 | u32 spsn, lpsn; | |
1680 | ||
1681 | spsn = full_flow_psn(flow, flow->flow_state.spsn); | |
1682 | lpsn = full_flow_psn(flow, flow->flow_state.lpsn); | |
1683 | ||
1684 | if (cmp_psn(psn, spsn) >= 0 && cmp_psn(psn, lpsn) <= 0) { | |
1685 | if (fidx) | |
1686 | *fidx = tail; | |
1687 | return flow; | |
1688 | } | |
1689 | } | |
1690 | return NULL; | |
1691 | } | |
1692 | ||
1693 | static struct tid_rdma_flow *find_flow(struct tid_rdma_request *req, | |
1694 | u32 psn, u16 *fidx) | |
1695 | { | |
1696 | return __find_flow_ranged(req, req->setup_head, req->clear_tail, psn, | |
1697 | fidx); | |
1698 | } | |
1699 | ||
742a3826 KW |
1700 | /* TID RDMA READ functions */ |
1701 | u32 hfi1_build_tid_rdma_read_packet(struct rvt_swqe *wqe, | |
1702 | struct ib_other_headers *ohdr, u32 *bth1, | |
1703 | u32 *bth2, u32 *len) | |
1704 | { | |
1705 | struct tid_rdma_request *req = wqe_to_tid_req(wqe); | |
1706 | struct tid_rdma_flow *flow = &req->flows[req->flow_idx]; | |
1707 | struct rvt_qp *qp = req->qp; | |
1708 | struct hfi1_qp_priv *qpriv = qp->priv; | |
1709 | struct hfi1_swqe_priv *wpriv = wqe->priv; | |
1710 | struct tid_rdma_read_req *rreq = &ohdr->u.tid_rdma.r_req; | |
1711 | struct tid_rdma_params *remote; | |
1712 | u32 req_len = 0; | |
1713 | void *req_addr = NULL; | |
1714 | ||
1715 | /* This is the IB psn used to send the request */ | |
1716 | *bth2 = mask_psn(flow->flow_state.ib_spsn + flow->pkt); | |
3ce5daa2 | 1717 | trace_hfi1_tid_flow_build_read_pkt(qp, req->flow_idx, flow); |
742a3826 KW |
1718 | |
1719 | /* TID Entries for TID RDMA READ payload */ | |
1720 | req_addr = &flow->tid_entry[flow->tid_idx]; | |
1721 | req_len = sizeof(*flow->tid_entry) * | |
1722 | (flow->tidcnt - flow->tid_idx); | |
1723 | ||
1724 | memset(&ohdr->u.tid_rdma.r_req, 0, sizeof(ohdr->u.tid_rdma.r_req)); | |
1725 | wpriv->ss.sge.vaddr = req_addr; | |
1726 | wpriv->ss.sge.sge_length = req_len; | |
1727 | wpriv->ss.sge.length = wpriv->ss.sge.sge_length; | |
1728 | /* | |
1729 | * We can safely zero these out. Since the first SGE covers the | |
1730 | * entire packet, nothing else should even look at the MR. | |
1731 | */ | |
1732 | wpriv->ss.sge.mr = NULL; | |
1733 | wpriv->ss.sge.m = 0; | |
1734 | wpriv->ss.sge.n = 0; | |
1735 | ||
1736 | wpriv->ss.sg_list = NULL; | |
1737 | wpriv->ss.total_len = wpriv->ss.sge.sge_length; | |
1738 | wpriv->ss.num_sge = 1; | |
1739 | ||
1740 | /* Construct the TID RDMA READ REQ packet header */ | |
1741 | rcu_read_lock(); | |
1742 | remote = rcu_dereference(qpriv->tid_rdma.remote); | |
1743 | ||
1744 | KDETH_RESET(rreq->kdeth0, KVER, 0x1); | |
1745 | KDETH_RESET(rreq->kdeth1, JKEY, remote->jkey); | |
1746 | rreq->reth.vaddr = cpu_to_be64(wqe->rdma_wr.remote_addr + | |
1747 | req->cur_seg * req->seg_len + flow->sent); | |
1748 | rreq->reth.rkey = cpu_to_be32(wqe->rdma_wr.rkey); | |
1749 | rreq->reth.length = cpu_to_be32(*len); | |
1750 | rreq->tid_flow_psn = | |
1751 | cpu_to_be32((flow->flow_state.generation << | |
1752 | HFI1_KDETH_BTH_SEQ_SHIFT) | | |
1753 | ((flow->flow_state.spsn + flow->pkt) & | |
1754 | HFI1_KDETH_BTH_SEQ_MASK)); | |
1755 | rreq->tid_flow_qp = | |
1756 | cpu_to_be32(qpriv->tid_rdma.local.qp | | |
1757 | ((flow->idx & TID_RDMA_DESTQP_FLOW_MASK) << | |
1758 | TID_RDMA_DESTQP_FLOW_SHIFT) | | |
1759 | qpriv->rcd->ctxt); | |
1760 | rreq->verbs_qp = cpu_to_be32(qp->remote_qpn); | |
1761 | *bth1 &= ~RVT_QPN_MASK; | |
1762 | *bth1 |= remote->qp; | |
1763 | *bth2 |= IB_BTH_REQ_ACK; | |
1764 | rcu_read_unlock(); | |
1765 | ||
1766 | /* We are done with this segment */ | |
1767 | flow->sent += *len; | |
1768 | req->cur_seg++; | |
1769 | qp->s_state = TID_OP(READ_REQ); | |
1770 | req->ack_pending++; | |
1771 | req->flow_idx = (req->flow_idx + 1) & (MAX_FLOWS - 1); | |
1772 | qpriv->pending_tid_r_segs++; | |
1773 | qp->s_num_rd_atomic++; | |
1774 | ||
1775 | /* Set the TID RDMA READ request payload size */ | |
1776 | *len = req_len; | |
1777 | ||
1778 | return sizeof(ohdr->u.tid_rdma.r_req) / sizeof(u32); | |
1779 | } | |
1780 | ||
1781 | /* | |
1782 | * @len: contains the data length to read upon entry and the read request | |
1783 | * payload length upon exit. | |
1784 | */ | |
1785 | u32 hfi1_build_tid_rdma_read_req(struct rvt_qp *qp, struct rvt_swqe *wqe, | |
1786 | struct ib_other_headers *ohdr, u32 *bth1, | |
1787 | u32 *bth2, u32 *len) | |
1788 | __must_hold(&qp->s_lock) | |
1789 | { | |
1790 | struct hfi1_qp_priv *qpriv = qp->priv; | |
1791 | struct tid_rdma_request *req = wqe_to_tid_req(wqe); | |
1792 | struct tid_rdma_flow *flow = NULL; | |
1793 | u32 hdwords = 0; | |
1794 | bool last; | |
1795 | bool retry = true; | |
1796 | u32 npkts = rvt_div_round_up_mtu(qp, *len); | |
1797 | ||
3ce5daa2 KW |
1798 | trace_hfi1_tid_req_build_read_req(qp, 0, wqe->wr.opcode, wqe->psn, |
1799 | wqe->lpsn, req); | |
742a3826 KW |
1800 | /* |
1801 | * Check sync conditions. Make sure that there are no pending | |
1802 | * segments before freeing the flow. | |
1803 | */ | |
1804 | sync_check: | |
1805 | if (req->state == TID_REQUEST_SYNC) { | |
1806 | if (qpriv->pending_tid_r_segs) | |
1807 | goto done; | |
1808 | ||
1809 | hfi1_kern_clear_hw_flow(req->rcd, qp); | |
1810 | req->state = TID_REQUEST_ACTIVE; | |
1811 | } | |
1812 | ||
1813 | /* | |
1814 | * If the request for this segment is resent, the tid resources should | |
1815 | * have been allocated before. In this case, req->flow_idx should | |
1816 | * fall behind req->setup_head. | |
1817 | */ | |
1818 | if (req->flow_idx == req->setup_head) { | |
1819 | retry = false; | |
1820 | if (req->state == TID_REQUEST_RESEND) { | |
1821 | /* | |
1822 | * This is the first new segment for a request whose | |
1823 | * earlier segments have been re-sent. We need to | |
1824 | * set up the sge pointer correctly. | |
1825 | */ | |
1826 | restart_sge(&qp->s_sge, wqe, req->s_next_psn, | |
1827 | qp->pmtu); | |
1828 | req->isge = 0; | |
1829 | req->state = TID_REQUEST_ACTIVE; | |
1830 | } | |
1831 | ||
1832 | /* | |
1833 | * Check sync. The last PSN of each generation is reserved for | |
1834 | * RESYNC. | |
1835 | */ | |
1836 | if ((qpriv->flow_state.psn + npkts) > MAX_TID_FLOW_PSN - 1) { | |
1837 | req->state = TID_REQUEST_SYNC; | |
1838 | goto sync_check; | |
1839 | } | |
1840 | ||
1841 | /* Allocate the flow if not yet */ | |
1842 | if (hfi1_kern_setup_hw_flow(qpriv->rcd, qp)) | |
1843 | goto done; | |
1844 | ||
1845 | /* | |
1846 | * The following call will advance req->setup_head after | |
1847 | * allocating the tid entries. | |
1848 | */ | |
1849 | if (hfi1_kern_exp_rcv_setup(req, &qp->s_sge, &last)) { | |
1850 | req->state = TID_REQUEST_QUEUED; | |
1851 | ||
1852 | /* | |
1853 | * We don't have resources for this segment. The QP has | |
1854 | * already been queued. | |
1855 | */ | |
1856 | goto done; | |
1857 | } | |
1858 | } | |
1859 | ||
1860 | /* req->flow_idx should only be one slot behind req->setup_head */ | |
1861 | flow = &req->flows[req->flow_idx]; | |
1862 | flow->pkt = 0; | |
1863 | flow->tid_idx = 0; | |
1864 | flow->sent = 0; | |
1865 | if (!retry) { | |
1866 | /* Set the first and last IB PSN for the flow in use.*/ | |
1867 | flow->flow_state.ib_spsn = req->s_next_psn; | |
1868 | flow->flow_state.ib_lpsn = | |
1869 | flow->flow_state.ib_spsn + flow->npkts - 1; | |
1870 | } | |
1871 | ||
1872 | /* Calculate the next segment start psn.*/ | |
1873 | req->s_next_psn += flow->npkts; | |
1874 | ||
1875 | /* Build the packet header */ | |
1876 | hdwords = hfi1_build_tid_rdma_read_packet(wqe, ohdr, bth1, bth2, len); | |
1877 | done: | |
1878 | return hdwords; | |
1879 | } | |
d0d564a1 KW |
1880 | |
1881 | /* | |
1882 | * Validate and accept the TID RDMA READ request parameters. | |
1883 | * Return 0 if the request is accepted successfully; | |
1884 | * Return 1 otherwise. | |
1885 | */ | |
1886 | static int tid_rdma_rcv_read_request(struct rvt_qp *qp, | |
1887 | struct rvt_ack_entry *e, | |
1888 | struct hfi1_packet *packet, | |
1889 | struct ib_other_headers *ohdr, | |
1890 | u32 bth0, u32 psn, u64 vaddr, u32 len) | |
1891 | { | |
1892 | struct hfi1_qp_priv *qpriv = qp->priv; | |
1893 | struct tid_rdma_request *req; | |
1894 | struct tid_rdma_flow *flow; | |
1895 | u32 flow_psn, i, tidlen = 0, pktlen, tlen; | |
1896 | ||
1897 | req = ack_to_tid_req(e); | |
1898 | ||
1899 | /* Validate the payload first */ | |
1900 | flow = &req->flows[req->setup_head]; | |
1901 | ||
1902 | /* payload length = packet length - (header length + ICRC length) */ | |
1903 | pktlen = packet->tlen - (packet->hlen + 4); | |
1904 | if (pktlen > sizeof(flow->tid_entry)) | |
1905 | return 1; | |
1906 | memcpy(flow->tid_entry, packet->ebuf, pktlen); | |
1907 | flow->tidcnt = pktlen / sizeof(*flow->tid_entry); | |
1908 | ||
1909 | /* | |
1910 | * Walk the TID_ENTRY list to make sure we have enough space for a | |
1911 | * complete segment. Also calculate the number of required packets. | |
1912 | */ | |
1913 | flow->npkts = rvt_div_round_up_mtu(qp, len); | |
1914 | for (i = 0; i < flow->tidcnt; i++) { | |
3ce5daa2 KW |
1915 | trace_hfi1_tid_entry_rcv_read_req(qp, i, |
1916 | flow->tid_entry[i]); | |
d0d564a1 KW |
1917 | tlen = EXP_TID_GET(flow->tid_entry[i], LEN); |
1918 | if (!tlen) | |
1919 | return 1; | |
1920 | ||
1921 | /* | |
1922 | * For tid pair (tidctr == 3), the buffer size of the pair | |
1923 | * should be the sum of the buffer size described by each | |
1924 | * tid entry. However, only the first entry needs to be | |
1925 | * specified in the request (see WFR HAS Section 8.5.7.1). | |
1926 | */ | |
1927 | tidlen += tlen; | |
1928 | } | |
1929 | if (tidlen * PAGE_SIZE < len) | |
1930 | return 1; | |
1931 | ||
1932 | /* Empty the flow array */ | |
1933 | req->clear_tail = req->setup_head; | |
1934 | flow->pkt = 0; | |
1935 | flow->tid_idx = 0; | |
1936 | flow->tid_offset = 0; | |
1937 | flow->sent = 0; | |
1938 | flow->tid_qpn = be32_to_cpu(ohdr->u.tid_rdma.r_req.tid_flow_qp); | |
1939 | flow->idx = (flow->tid_qpn >> TID_RDMA_DESTQP_FLOW_SHIFT) & | |
1940 | TID_RDMA_DESTQP_FLOW_MASK; | |
1941 | flow_psn = mask_psn(be32_to_cpu(ohdr->u.tid_rdma.r_req.tid_flow_psn)); | |
1942 | flow->flow_state.generation = flow_psn >> HFI1_KDETH_BTH_SEQ_SHIFT; | |
1943 | flow->flow_state.spsn = flow_psn & HFI1_KDETH_BTH_SEQ_MASK; | |
1944 | flow->length = len; | |
1945 | ||
1946 | flow->flow_state.lpsn = flow->flow_state.spsn + | |
1947 | flow->npkts - 1; | |
1948 | flow->flow_state.ib_spsn = psn; | |
1949 | flow->flow_state.ib_lpsn = flow->flow_state.ib_spsn + flow->npkts - 1; | |
1950 | ||
3ce5daa2 | 1951 | trace_hfi1_tid_flow_rcv_read_req(qp, req->setup_head, flow); |
d0d564a1 KW |
1952 | /* Set the initial flow index to the current flow. */ |
1953 | req->flow_idx = req->setup_head; | |
1954 | ||
1955 | /* advance circular buffer head */ | |
1956 | req->setup_head = (req->setup_head + 1) & (MAX_FLOWS - 1); | |
1957 | ||
1958 | /* | |
1959 | * Compute last PSN for request. | |
1960 | */ | |
1961 | e->opcode = (bth0 >> 24) & 0xff; | |
1962 | e->psn = psn; | |
1963 | e->lpsn = psn + flow->npkts - 1; | |
1964 | e->sent = 0; | |
1965 | ||
1966 | req->n_flows = qpriv->tid_rdma.local.max_read; | |
1967 | req->state = TID_REQUEST_ACTIVE; | |
1968 | req->cur_seg = 0; | |
1969 | req->comp_seg = 0; | |
1970 | req->ack_seg = 0; | |
1971 | req->isge = 0; | |
1972 | req->seg_len = qpriv->tid_rdma.local.max_len; | |
1973 | req->total_len = len; | |
1974 | req->total_segs = 1; | |
1975 | req->r_flow_psn = e->psn; | |
1976 | ||
3ce5daa2 KW |
1977 | trace_hfi1_tid_req_rcv_read_req(qp, 0, e->opcode, e->psn, e->lpsn, |
1978 | req); | |
d0d564a1 KW |
1979 | return 0; |
1980 | } | |
1981 | ||
1982 | static int tid_rdma_rcv_error(struct hfi1_packet *packet, | |
1983 | struct ib_other_headers *ohdr, | |
1984 | struct rvt_qp *qp, u32 psn, int diff) | |
1985 | { | |
1986 | struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num); | |
1987 | struct hfi1_ctxtdata *rcd = ((struct hfi1_qp_priv *)qp->priv)->rcd; | |
07b92370 KW |
1988 | struct hfi1_ibdev *dev = to_idev(qp->ibqp.device); |
1989 | struct hfi1_qp_priv *qpriv = qp->priv; | |
d0d564a1 KW |
1990 | struct rvt_ack_entry *e; |
1991 | struct tid_rdma_request *req; | |
1992 | unsigned long flags; | |
1993 | u8 prev; | |
1994 | bool old_req; | |
1995 | ||
3ce5daa2 KW |
1996 | trace_hfi1_rsp_tid_rcv_error(qp, psn); |
1997 | trace_hfi1_tid_rdma_rcv_err(qp, 0, psn, diff); | |
d0d564a1 KW |
1998 | if (diff > 0) { |
1999 | /* sequence error */ | |
2000 | if (!qp->r_nak_state) { | |
2001 | ibp->rvp.n_rc_seqnak++; | |
2002 | qp->r_nak_state = IB_NAK_PSN_ERROR; | |
2003 | qp->r_ack_psn = qp->r_psn; | |
2004 | rc_defered_ack(rcd, qp); | |
2005 | } | |
2006 | goto done; | |
2007 | } | |
2008 | ||
2009 | ibp->rvp.n_rc_dupreq++; | |
2010 | ||
2011 | spin_lock_irqsave(&qp->s_lock, flags); | |
2012 | e = find_prev_entry(qp, psn, &prev, NULL, &old_req); | |
07b92370 KW |
2013 | if (!e || (e->opcode != TID_OP(READ_REQ) && |
2014 | e->opcode != TID_OP(WRITE_REQ))) | |
d0d564a1 KW |
2015 | goto unlock; |
2016 | ||
2017 | req = ack_to_tid_req(e); | |
2018 | req->r_flow_psn = psn; | |
3ce5daa2 | 2019 | trace_hfi1_tid_req_rcv_err(qp, 0, e->opcode, e->psn, e->lpsn, req); |
d0d564a1 KW |
2020 | if (e->opcode == TID_OP(READ_REQ)) { |
2021 | struct ib_reth *reth; | |
2022 | u32 offset; | |
2023 | u32 len; | |
2024 | u32 rkey; | |
2025 | u64 vaddr; | |
2026 | int ok; | |
2027 | u32 bth0; | |
2028 | ||
2029 | reth = &ohdr->u.tid_rdma.r_req.reth; | |
2030 | /* | |
2031 | * The requester always restarts from the start of the original | |
2032 | * request. | |
2033 | */ | |
2034 | offset = delta_psn(psn, e->psn) * qp->pmtu; | |
2035 | len = be32_to_cpu(reth->length); | |
2036 | if (psn != e->psn || len != req->total_len) | |
2037 | goto unlock; | |
2038 | ||
2039 | if (e->rdma_sge.mr) { | |
2040 | rvt_put_mr(e->rdma_sge.mr); | |
2041 | e->rdma_sge.mr = NULL; | |
2042 | } | |
2043 | ||
2044 | rkey = be32_to_cpu(reth->rkey); | |
2045 | vaddr = get_ib_reth_vaddr(reth); | |
2046 | ||
2047 | qp->r_len = len; | |
2048 | ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr, rkey, | |
2049 | IB_ACCESS_REMOTE_READ); | |
2050 | if (unlikely(!ok)) | |
2051 | goto unlock; | |
2052 | ||
2053 | /* | |
2054 | * If all the response packets for the current request have | |
2055 | * been sent out and this request is complete (old_request | |
2056 | * == false) and the TID flow may be unusable (the | |
2057 | * req->clear_tail is advanced). However, when an earlier | |
2058 | * request is received, this request will not be complete any | |
2059 | * more (qp->s_tail_ack_queue is moved back, see below). | |
2060 | * Consequently, we need to update the TID flow info everytime | |
2061 | * a duplicate request is received. | |
2062 | */ | |
2063 | bth0 = be32_to_cpu(ohdr->bth[0]); | |
2064 | if (tid_rdma_rcv_read_request(qp, e, packet, ohdr, bth0, psn, | |
2065 | vaddr, len)) | |
2066 | goto unlock; | |
2067 | ||
2068 | /* | |
2069 | * True if the request is already scheduled (between | |
2070 | * qp->s_tail_ack_queue and qp->r_head_ack_queue); | |
2071 | */ | |
2072 | if (old_req) | |
2073 | goto unlock; | |
07b92370 KW |
2074 | } else { |
2075 | struct flow_state *fstate; | |
2076 | bool schedule = false; | |
2077 | u8 i; | |
2078 | ||
2079 | if (req->state == TID_REQUEST_RESEND) { | |
2080 | req->state = TID_REQUEST_RESEND_ACTIVE; | |
2081 | } else if (req->state == TID_REQUEST_INIT_RESEND) { | |
2082 | req->state = TID_REQUEST_INIT; | |
2083 | schedule = true; | |
2084 | } | |
2085 | ||
2086 | /* | |
2087 | * True if the request is already scheduled (between | |
2088 | * qp->s_tail_ack_queue and qp->r_head_ack_queue). | |
2089 | * Also, don't change requests, which are at the SYNC | |
2090 | * point and haven't generated any responses yet. | |
2091 | * There is nothing to retransmit for them yet. | |
2092 | */ | |
2093 | if (old_req || req->state == TID_REQUEST_INIT || | |
2094 | (req->state == TID_REQUEST_SYNC && !req->cur_seg)) { | |
2095 | for (i = prev + 1; ; i++) { | |
2096 | if (i > rvt_size_atomic(&dev->rdi)) | |
2097 | i = 0; | |
2098 | if (i == qp->r_head_ack_queue) | |
2099 | break; | |
2100 | e = &qp->s_ack_queue[i]; | |
2101 | req = ack_to_tid_req(e); | |
2102 | if (e->opcode == TID_OP(WRITE_REQ) && | |
2103 | req->state == TID_REQUEST_INIT) | |
2104 | req->state = TID_REQUEST_INIT_RESEND; | |
2105 | } | |
2106 | /* | |
2107 | * If the state of the request has been changed, | |
2108 | * the first leg needs to get scheduled in order to | |
2109 | * pick up the change. Otherwise, normal response | |
2110 | * processing should take care of it. | |
2111 | */ | |
2112 | if (!schedule) | |
2113 | goto unlock; | |
2114 | } | |
2115 | ||
2116 | /* | |
2117 | * If there is no more allocated segment, just schedule the qp | |
2118 | * without changing any state. | |
2119 | */ | |
2120 | if (req->clear_tail == req->setup_head) | |
2121 | goto schedule; | |
2122 | /* | |
2123 | * If this request has sent responses for segments, which have | |
2124 | * not received data yet (flow_idx != clear_tail), the flow_idx | |
2125 | * pointer needs to be adjusted so the same responses can be | |
2126 | * re-sent. | |
2127 | */ | |
2128 | if (CIRC_CNT(req->flow_idx, req->clear_tail, MAX_FLOWS)) { | |
2129 | fstate = &req->flows[req->clear_tail].flow_state; | |
2130 | qpriv->pending_tid_w_segs -= | |
2131 | CIRC_CNT(req->flow_idx, req->clear_tail, | |
2132 | MAX_FLOWS); | |
2133 | req->flow_idx = | |
2134 | CIRC_ADD(req->clear_tail, | |
2135 | delta_psn(psn, fstate->resp_ib_psn), | |
2136 | MAX_FLOWS); | |
2137 | qpriv->pending_tid_w_segs += | |
2138 | delta_psn(psn, fstate->resp_ib_psn); | |
2139 | /* | |
2140 | * When flow_idx == setup_head, we've gotten a duplicate | |
2141 | * request for a segment, which has not been allocated | |
2142 | * yet. In that case, don't adjust this request. | |
2143 | * However, we still want to go through the loop below | |
2144 | * to adjust all subsequent requests. | |
2145 | */ | |
2146 | if (CIRC_CNT(req->setup_head, req->flow_idx, | |
2147 | MAX_FLOWS)) { | |
2148 | req->cur_seg = delta_psn(psn, e->psn); | |
2149 | req->state = TID_REQUEST_RESEND_ACTIVE; | |
2150 | } | |
2151 | } | |
2152 | ||
2153 | for (i = prev + 1; ; i++) { | |
2154 | /* | |
2155 | * Look at everything up to and including | |
2156 | * s_tail_ack_queue | |
2157 | */ | |
2158 | if (i > rvt_size_atomic(&dev->rdi)) | |
2159 | i = 0; | |
2160 | if (i == qp->r_head_ack_queue) | |
2161 | break; | |
2162 | e = &qp->s_ack_queue[i]; | |
2163 | req = ack_to_tid_req(e); | |
2164 | trace_hfi1_tid_req_rcv_err(qp, 0, e->opcode, e->psn, | |
2165 | e->lpsn, req); | |
2166 | if (e->opcode != TID_OP(WRITE_REQ) || | |
2167 | req->cur_seg == req->comp_seg || | |
2168 | req->state == TID_REQUEST_INIT || | |
2169 | req->state == TID_REQUEST_INIT_RESEND) { | |
2170 | if (req->state == TID_REQUEST_INIT) | |
2171 | req->state = TID_REQUEST_INIT_RESEND; | |
2172 | continue; | |
2173 | } | |
2174 | qpriv->pending_tid_w_segs -= | |
2175 | CIRC_CNT(req->flow_idx, | |
2176 | req->clear_tail, | |
2177 | MAX_FLOWS); | |
2178 | req->flow_idx = req->clear_tail; | |
2179 | req->state = TID_REQUEST_RESEND; | |
2180 | req->cur_seg = req->comp_seg; | |
2181 | } | |
d0d564a1 KW |
2182 | } |
2183 | /* Re-process old requests.*/ | |
4f9264d1 KW |
2184 | if (qp->s_acked_ack_queue == qp->s_tail_ack_queue) |
2185 | qp->s_acked_ack_queue = prev; | |
d0d564a1 KW |
2186 | qp->s_tail_ack_queue = prev; |
2187 | /* | |
2188 | * Since the qp->s_tail_ack_queue is modified, the | |
2189 | * qp->s_ack_state must be changed to re-initialize | |
2190 | * qp->s_ack_rdma_sge; Otherwise, we will end up in | |
2191 | * wrong memory region. | |
2192 | */ | |
2193 | qp->s_ack_state = OP(ACKNOWLEDGE); | |
07b92370 KW |
2194 | schedule: |
2195 | /* | |
2196 | * It's possible to receive a retry psn that is earlier than an RNRNAK | |
2197 | * psn. In this case, the rnrnak state should be cleared. | |
2198 | */ | |
2199 | if (qpriv->rnr_nak_state) { | |
2200 | qp->s_nak_state = 0; | |
2201 | qpriv->rnr_nak_state = TID_RNR_NAK_INIT; | |
2202 | qp->r_psn = e->lpsn + 1; | |
2203 | hfi1_tid_write_alloc_resources(qp, true); | |
2204 | } | |
2205 | ||
d0d564a1 KW |
2206 | qp->r_state = e->opcode; |
2207 | qp->r_nak_state = 0; | |
2208 | qp->s_flags |= RVT_S_RESP_PENDING; | |
2209 | hfi1_schedule_send(qp); | |
2210 | unlock: | |
2211 | spin_unlock_irqrestore(&qp->s_lock, flags); | |
2212 | done: | |
2213 | return 1; | |
2214 | } | |
2215 | ||
2216 | void hfi1_rc_rcv_tid_rdma_read_req(struct hfi1_packet *packet) | |
2217 | { | |
2218 | /* HANDLER FOR TID RDMA READ REQUEST packet (Responder side)*/ | |
2219 | ||
2220 | /* | |
2221 | * 1. Verify TID RDMA READ REQ as per IB_OPCODE_RC_RDMA_READ | |
2222 | * (see hfi1_rc_rcv()) | |
2223 | * 2. Put TID RDMA READ REQ into the response queueu (s_ack_queue) | |
2224 | * - Setup struct tid_rdma_req with request info | |
2225 | * - Initialize struct tid_rdma_flow info; | |
2226 | * - Copy TID entries; | |
2227 | * 3. Set the qp->s_ack_state. | |
2228 | * 4. Set RVT_S_RESP_PENDING in s_flags. | |
2229 | * 5. Kick the send engine (hfi1_schedule_send()) | |
2230 | */ | |
2231 | struct hfi1_ctxtdata *rcd = packet->rcd; | |
2232 | struct rvt_qp *qp = packet->qp; | |
2233 | struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num); | |
2234 | struct ib_other_headers *ohdr = packet->ohdr; | |
2235 | struct rvt_ack_entry *e; | |
2236 | unsigned long flags; | |
2237 | struct ib_reth *reth; | |
2238 | struct hfi1_qp_priv *qpriv = qp->priv; | |
2239 | u32 bth0, psn, len, rkey; | |
2240 | bool is_fecn; | |
2241 | u8 next; | |
2242 | u64 vaddr; | |
2243 | int diff; | |
2244 | u8 nack_state = IB_NAK_INVALID_REQUEST; | |
2245 | ||
2246 | bth0 = be32_to_cpu(ohdr->bth[0]); | |
2247 | if (hfi1_ruc_check_hdr(ibp, packet)) | |
2248 | return; | |
2249 | ||
2250 | is_fecn = process_ecn(qp, packet); | |
2251 | psn = mask_psn(be32_to_cpu(ohdr->bth[2])); | |
3ce5daa2 | 2252 | trace_hfi1_rsp_rcv_tid_read_req(qp, psn); |
d0d564a1 KW |
2253 | |
2254 | if (qp->state == IB_QPS_RTR && !(qp->r_flags & RVT_R_COMM_EST)) | |
2255 | rvt_comm_est(qp); | |
2256 | ||
2257 | if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ))) | |
2258 | goto nack_inv; | |
2259 | ||
2260 | reth = &ohdr->u.tid_rdma.r_req.reth; | |
2261 | vaddr = be64_to_cpu(reth->vaddr); | |
2262 | len = be32_to_cpu(reth->length); | |
2263 | /* The length needs to be in multiples of PAGE_SIZE */ | |
2264 | if (!len || len & ~PAGE_MASK || len > qpriv->tid_rdma.local.max_len) | |
2265 | goto nack_inv; | |
2266 | ||
2267 | diff = delta_psn(psn, qp->r_psn); | |
2268 | if (unlikely(diff)) { | |
2269 | if (tid_rdma_rcv_error(packet, ohdr, qp, psn, diff)) | |
2270 | return; | |
2271 | goto send_ack; | |
2272 | } | |
2273 | ||
2274 | /* We've verified the request, insert it into the ack queue. */ | |
2275 | next = qp->r_head_ack_queue + 1; | |
2276 | if (next > rvt_size_atomic(ib_to_rvt(qp->ibqp.device))) | |
2277 | next = 0; | |
2278 | spin_lock_irqsave(&qp->s_lock, flags); | |
2279 | if (unlikely(next == qp->s_tail_ack_queue)) { | |
2280 | if (!qp->s_ack_queue[next].sent) { | |
2281 | nack_state = IB_NAK_REMOTE_OPERATIONAL_ERROR; | |
2282 | goto nack_inv_unlock; | |
2283 | } | |
2284 | update_ack_queue(qp, next); | |
2285 | } | |
2286 | e = &qp->s_ack_queue[qp->r_head_ack_queue]; | |
2287 | if (e->rdma_sge.mr) { | |
2288 | rvt_put_mr(e->rdma_sge.mr); | |
2289 | e->rdma_sge.mr = NULL; | |
2290 | } | |
2291 | ||
2292 | rkey = be32_to_cpu(reth->rkey); | |
2293 | qp->r_len = len; | |
2294 | ||
2295 | if (unlikely(!rvt_rkey_ok(qp, &e->rdma_sge, qp->r_len, vaddr, | |
2296 | rkey, IB_ACCESS_REMOTE_READ))) | |
2297 | goto nack_acc; | |
2298 | ||
2299 | /* Accept the request parameters */ | |
2300 | if (tid_rdma_rcv_read_request(qp, e, packet, ohdr, bth0, psn, vaddr, | |
2301 | len)) | |
2302 | goto nack_inv_unlock; | |
2303 | ||
2304 | qp->r_state = e->opcode; | |
2305 | qp->r_nak_state = 0; | |
2306 | /* | |
2307 | * We need to increment the MSN here instead of when we | |
2308 | * finish sending the result since a duplicate request would | |
2309 | * increment it more than once. | |
2310 | */ | |
2311 | qp->r_msn++; | |
2312 | qp->r_psn += e->lpsn - e->psn + 1; | |
2313 | ||
2314 | qp->r_head_ack_queue = next; | |
2315 | ||
07b92370 KW |
2316 | /* |
2317 | * For all requests other than TID WRITE which are added to the ack | |
2318 | * queue, qpriv->r_tid_alloc follows qp->r_head_ack_queue. It is ok to | |
2319 | * do this because of interlocks between these and TID WRITE | |
2320 | * requests. The same change has also been made in hfi1_rc_rcv(). | |
2321 | */ | |
2322 | qpriv->r_tid_alloc = qp->r_head_ack_queue; | |
2323 | ||
d0d564a1 KW |
2324 | /* Schedule the send tasklet. */ |
2325 | qp->s_flags |= RVT_S_RESP_PENDING; | |
2326 | hfi1_schedule_send(qp); | |
2327 | ||
2328 | spin_unlock_irqrestore(&qp->s_lock, flags); | |
2329 | if (is_fecn) | |
2330 | goto send_ack; | |
2331 | return; | |
2332 | ||
2333 | nack_inv_unlock: | |
2334 | spin_unlock_irqrestore(&qp->s_lock, flags); | |
2335 | nack_inv: | |
2336 | rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR); | |
2337 | qp->r_nak_state = nack_state; | |
2338 | qp->r_ack_psn = qp->r_psn; | |
2339 | /* Queue NAK for later */ | |
2340 | rc_defered_ack(rcd, qp); | |
2341 | return; | |
2342 | nack_acc: | |
2343 | spin_unlock_irqrestore(&qp->s_lock, flags); | |
2344 | rvt_rc_error(qp, IB_WC_LOC_PROT_ERR); | |
2345 | qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR; | |
2346 | qp->r_ack_psn = qp->r_psn; | |
2347 | send_ack: | |
2348 | hfi1_send_rc_ack(packet, is_fecn); | |
2349 | } | |
1db21b50 KW |
2350 | |
2351 | u32 hfi1_build_tid_rdma_read_resp(struct rvt_qp *qp, struct rvt_ack_entry *e, | |
2352 | struct ib_other_headers *ohdr, u32 *bth0, | |
2353 | u32 *bth1, u32 *bth2, u32 *len, bool *last) | |
2354 | { | |
2355 | struct hfi1_ack_priv *epriv = e->priv; | |
2356 | struct tid_rdma_request *req = &epriv->tid_req; | |
2357 | struct hfi1_qp_priv *qpriv = qp->priv; | |
2358 | struct tid_rdma_flow *flow = &req->flows[req->clear_tail]; | |
2359 | u32 tidentry = flow->tid_entry[flow->tid_idx]; | |
2360 | u32 tidlen = EXP_TID_GET(tidentry, LEN) << PAGE_SHIFT; | |
2361 | struct tid_rdma_read_resp *resp = &ohdr->u.tid_rdma.r_rsp; | |
2362 | u32 next_offset, om = KDETH_OM_LARGE; | |
2363 | bool last_pkt; | |
2364 | u32 hdwords = 0; | |
2365 | struct tid_rdma_params *remote; | |
2366 | ||
2367 | *len = min_t(u32, qp->pmtu, tidlen - flow->tid_offset); | |
2368 | flow->sent += *len; | |
2369 | next_offset = flow->tid_offset + *len; | |
2370 | last_pkt = (flow->sent >= flow->length); | |
2371 | ||
3ce5daa2 KW |
2372 | trace_hfi1_tid_entry_build_read_resp(qp, flow->tid_idx, tidentry); |
2373 | trace_hfi1_tid_flow_build_read_resp(qp, req->clear_tail, flow); | |
2374 | ||
1db21b50 KW |
2375 | rcu_read_lock(); |
2376 | remote = rcu_dereference(qpriv->tid_rdma.remote); | |
2377 | if (!remote) { | |
2378 | rcu_read_unlock(); | |
2379 | goto done; | |
2380 | } | |
2381 | KDETH_RESET(resp->kdeth0, KVER, 0x1); | |
2382 | KDETH_SET(resp->kdeth0, SH, !last_pkt); | |
2383 | KDETH_SET(resp->kdeth0, INTR, !!(!last_pkt && remote->urg)); | |
2384 | KDETH_SET(resp->kdeth0, TIDCTRL, EXP_TID_GET(tidentry, CTRL)); | |
2385 | KDETH_SET(resp->kdeth0, TID, EXP_TID_GET(tidentry, IDX)); | |
2386 | KDETH_SET(resp->kdeth0, OM, om == KDETH_OM_LARGE); | |
2387 | KDETH_SET(resp->kdeth0, OFFSET, flow->tid_offset / om); | |
2388 | KDETH_RESET(resp->kdeth1, JKEY, remote->jkey); | |
2389 | resp->verbs_qp = cpu_to_be32(qp->remote_qpn); | |
2390 | rcu_read_unlock(); | |
2391 | ||
2392 | resp->aeth = rvt_compute_aeth(qp); | |
2393 | resp->verbs_psn = cpu_to_be32(mask_psn(flow->flow_state.ib_spsn + | |
2394 | flow->pkt)); | |
2395 | ||
2396 | *bth0 = TID_OP(READ_RESP) << 24; | |
2397 | *bth1 = flow->tid_qpn; | |
2398 | *bth2 = mask_psn(((flow->flow_state.spsn + flow->pkt++) & | |
2399 | HFI1_KDETH_BTH_SEQ_MASK) | | |
2400 | (flow->flow_state.generation << | |
2401 | HFI1_KDETH_BTH_SEQ_SHIFT)); | |
2402 | *last = last_pkt; | |
2403 | if (last_pkt) | |
2404 | /* Advance to next flow */ | |
2405 | req->clear_tail = (req->clear_tail + 1) & | |
2406 | (MAX_FLOWS - 1); | |
2407 | ||
2408 | if (next_offset >= tidlen) { | |
2409 | flow->tid_offset = 0; | |
2410 | flow->tid_idx++; | |
2411 | } else { | |
2412 | flow->tid_offset = next_offset; | |
2413 | } | |
2414 | ||
2415 | hdwords = sizeof(ohdr->u.tid_rdma.r_rsp) / sizeof(u32); | |
2416 | ||
2417 | done: | |
2418 | return hdwords; | |
2419 | } | |
9905bf06 KW |
2420 | |
2421 | static inline struct tid_rdma_request * | |
2422 | find_tid_request(struct rvt_qp *qp, u32 psn, enum ib_wr_opcode opcode) | |
2423 | __must_hold(&qp->s_lock) | |
2424 | { | |
2425 | struct rvt_swqe *wqe; | |
2426 | struct tid_rdma_request *req = NULL; | |
2427 | u32 i, end; | |
2428 | ||
2429 | end = qp->s_cur + 1; | |
2430 | if (end == qp->s_size) | |
2431 | end = 0; | |
2432 | for (i = qp->s_acked; i != end;) { | |
2433 | wqe = rvt_get_swqe_ptr(qp, i); | |
2434 | if (cmp_psn(psn, wqe->psn) >= 0 && | |
2435 | cmp_psn(psn, wqe->lpsn) <= 0) { | |
2436 | if (wqe->wr.opcode == opcode) | |
2437 | req = wqe_to_tid_req(wqe); | |
2438 | break; | |
2439 | } | |
2440 | if (++i == qp->s_size) | |
2441 | i = 0; | |
2442 | } | |
2443 | ||
2444 | return req; | |
2445 | } | |
2446 | ||
2447 | void hfi1_rc_rcv_tid_rdma_read_resp(struct hfi1_packet *packet) | |
2448 | { | |
2449 | /* HANDLER FOR TID RDMA READ RESPONSE packet (Requestor side */ | |
2450 | ||
2451 | /* | |
2452 | * 1. Find matching SWQE | |
2453 | * 2. Check that the entire segment has been read. | |
2454 | * 3. Remove HFI1_S_WAIT_TID_RESP from s_flags. | |
2455 | * 4. Free the TID flow resources. | |
2456 | * 5. Kick the send engine (hfi1_schedule_send()) | |
2457 | */ | |
2458 | struct ib_other_headers *ohdr = packet->ohdr; | |
2459 | struct rvt_qp *qp = packet->qp; | |
2460 | struct hfi1_qp_priv *priv = qp->priv; | |
2461 | struct hfi1_ctxtdata *rcd = packet->rcd; | |
2462 | struct tid_rdma_request *req; | |
2463 | struct tid_rdma_flow *flow; | |
2464 | u32 opcode, aeth; | |
2465 | bool is_fecn; | |
2466 | unsigned long flags; | |
2467 | u32 kpsn, ipsn; | |
2468 | ||
3ce5daa2 | 2469 | trace_hfi1_sender_rcv_tid_read_resp(qp); |
9905bf06 KW |
2470 | is_fecn = process_ecn(qp, packet); |
2471 | kpsn = mask_psn(be32_to_cpu(ohdr->bth[2])); | |
2472 | aeth = be32_to_cpu(ohdr->u.tid_rdma.r_rsp.aeth); | |
2473 | opcode = (be32_to_cpu(ohdr->bth[0]) >> 24) & 0xff; | |
2474 | ||
2475 | spin_lock_irqsave(&qp->s_lock, flags); | |
2476 | ipsn = mask_psn(be32_to_cpu(ohdr->u.tid_rdma.r_rsp.verbs_psn)); | |
2477 | req = find_tid_request(qp, ipsn, IB_WR_TID_RDMA_READ); | |
2478 | if (unlikely(!req)) | |
2479 | goto ack_op_err; | |
2480 | ||
2481 | flow = &req->flows[req->clear_tail]; | |
2482 | /* When header suppression is disabled */ | |
2483 | if (cmp_psn(ipsn, flow->flow_state.ib_lpsn)) | |
2484 | goto ack_done; | |
2485 | req->ack_pending--; | |
2486 | priv->pending_tid_r_segs--; | |
2487 | qp->s_num_rd_atomic--; | |
2488 | if ((qp->s_flags & RVT_S_WAIT_FENCE) && | |
2489 | !qp->s_num_rd_atomic) { | |
2490 | qp->s_flags &= ~(RVT_S_WAIT_FENCE | | |
2491 | RVT_S_WAIT_ACK); | |
2492 | hfi1_schedule_send(qp); | |
2493 | } | |
2494 | if (qp->s_flags & RVT_S_WAIT_RDMAR) { | |
2495 | qp->s_flags &= ~(RVT_S_WAIT_RDMAR | RVT_S_WAIT_ACK); | |
2496 | hfi1_schedule_send(qp); | |
2497 | } | |
2498 | ||
3ce5daa2 KW |
2499 | trace_hfi1_ack(qp, ipsn); |
2500 | trace_hfi1_tid_req_rcv_read_resp(qp, 0, req->e.swqe->wr.opcode, | |
2501 | req->e.swqe->psn, req->e.swqe->lpsn, | |
2502 | req); | |
2503 | trace_hfi1_tid_flow_rcv_read_resp(qp, req->clear_tail, flow); | |
2504 | ||
9905bf06 KW |
2505 | /* Release the tid resources */ |
2506 | hfi1_kern_exp_rcv_clear(req); | |
2507 | ||
2508 | if (!do_rc_ack(qp, aeth, ipsn, opcode, 0, rcd)) | |
2509 | goto ack_done; | |
2510 | ||
2511 | /* If not done yet, build next read request */ | |
2512 | if (++req->comp_seg >= req->total_segs) { | |
2513 | priv->tid_r_comp++; | |
2514 | req->state = TID_REQUEST_COMPLETE; | |
2515 | } | |
2516 | ||
2517 | /* | |
2518 | * Clear the hw flow under two conditions: | |
2519 | * 1. This request is a sync point and it is complete; | |
2520 | * 2. Current request is completed and there are no more requests. | |
2521 | */ | |
2522 | if ((req->state == TID_REQUEST_SYNC && | |
2523 | req->comp_seg == req->cur_seg) || | |
2524 | priv->tid_r_comp == priv->tid_r_reqs) { | |
2525 | hfi1_kern_clear_hw_flow(priv->rcd, qp); | |
2526 | if (req->state == TID_REQUEST_SYNC) | |
2527 | req->state = TID_REQUEST_ACTIVE; | |
2528 | } | |
2529 | ||
2530 | hfi1_schedule_send(qp); | |
2531 | goto ack_done; | |
2532 | ||
2533 | ack_op_err: | |
2534 | /* | |
2535 | * The test indicates that the send engine has finished its cleanup | |
2536 | * after sending the request and it's now safe to put the QP into error | |
2537 | * state. However, if the wqe queue is empty (qp->s_acked == qp->s_tail | |
2538 | * == qp->s_head), it would be unsafe to complete the wqe pointed by | |
2539 | * qp->s_acked here. Putting the qp into error state will safely flush | |
2540 | * all remaining requests. | |
2541 | */ | |
2542 | if (qp->s_last == qp->s_acked) | |
2543 | rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR); | |
2544 | ||
2545 | ack_done: | |
2546 | spin_unlock_irqrestore(&qp->s_lock, flags); | |
2547 | if (is_fecn) | |
2548 | hfi1_send_rc_ack(packet, is_fecn); | |
2549 | } | |
2550 | ||
2551 | void hfi1_kern_read_tid_flow_free(struct rvt_qp *qp) | |
2552 | __must_hold(&qp->s_lock) | |
2553 | { | |
2554 | u32 n = qp->s_acked; | |
2555 | struct rvt_swqe *wqe; | |
2556 | struct tid_rdma_request *req; | |
2557 | struct hfi1_qp_priv *priv = qp->priv; | |
2558 | ||
2559 | lockdep_assert_held(&qp->s_lock); | |
2560 | /* Free any TID entries */ | |
2561 | while (n != qp->s_tail) { | |
2562 | wqe = rvt_get_swqe_ptr(qp, n); | |
2563 | if (wqe->wr.opcode == IB_WR_TID_RDMA_READ) { | |
2564 | req = wqe_to_tid_req(wqe); | |
2565 | hfi1_kern_exp_rcv_clear_all(req); | |
2566 | } | |
2567 | ||
2568 | if (++n == qp->s_size) | |
2569 | n = 0; | |
2570 | } | |
2571 | /* Free flow */ | |
2572 | hfi1_kern_clear_hw_flow(priv->rcd, qp); | |
2573 | } | |
2574 | ||
2575 | static bool tid_rdma_tid_err(struct hfi1_ctxtdata *rcd, | |
2576 | struct hfi1_packet *packet, u8 rcv_type, | |
2577 | u8 opcode) | |
2578 | { | |
2579 | struct rvt_qp *qp = packet->qp; | |
d72fe7d5 | 2580 | struct hfi1_qp_priv *qpriv = qp->priv; |
9905bf06 KW |
2581 | u32 ipsn; |
2582 | struct ib_other_headers *ohdr = packet->ohdr; | |
d72fe7d5 KW |
2583 | struct rvt_ack_entry *e; |
2584 | struct tid_rdma_request *req; | |
2585 | struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device); | |
2586 | u32 i; | |
9905bf06 KW |
2587 | |
2588 | if (rcv_type >= RHF_RCV_TYPE_IB) | |
2589 | goto done; | |
2590 | ||
2591 | spin_lock(&qp->s_lock); | |
d72fe7d5 KW |
2592 | |
2593 | /* | |
2594 | * We've ran out of space in the eager buffer. | |
2595 | * Eagerly received KDETH packets which require space in the | |
2596 | * Eager buffer (packet that have payload) are TID RDMA WRITE | |
2597 | * response packets. In this case, we have to re-transmit the | |
2598 | * TID RDMA WRITE request. | |
2599 | */ | |
2600 | if (rcv_type == RHF_RCV_TYPE_EAGER) { | |
2601 | hfi1_restart_rc(qp, qp->s_last_psn + 1, 1); | |
2602 | hfi1_schedule_send(qp); | |
2603 | goto done_unlock; | |
2604 | } | |
2605 | ||
9905bf06 KW |
2606 | /* |
2607 | * For TID READ response, error out QP after freeing the tid | |
2608 | * resources. | |
2609 | */ | |
2610 | if (opcode == TID_OP(READ_RESP)) { | |
2611 | ipsn = mask_psn(be32_to_cpu(ohdr->u.tid_rdma.r_rsp.verbs_psn)); | |
2612 | if (cmp_psn(ipsn, qp->s_last_psn) > 0 && | |
2613 | cmp_psn(ipsn, qp->s_psn) < 0) { | |
2614 | hfi1_kern_read_tid_flow_free(qp); | |
2615 | spin_unlock(&qp->s_lock); | |
2616 | rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR); | |
2617 | goto done; | |
2618 | } | |
d72fe7d5 KW |
2619 | goto done_unlock; |
2620 | } | |
2621 | ||
2622 | /* | |
2623 | * Error out the qp for TID RDMA WRITE | |
2624 | */ | |
2625 | hfi1_kern_clear_hw_flow(qpriv->rcd, qp); | |
2626 | for (i = 0; i < rvt_max_atomic(rdi); i++) { | |
2627 | e = &qp->s_ack_queue[i]; | |
2628 | if (e->opcode == TID_OP(WRITE_REQ)) { | |
2629 | req = ack_to_tid_req(e); | |
2630 | hfi1_kern_exp_rcv_clear_all(req); | |
2631 | } | |
9905bf06 | 2632 | } |
d72fe7d5 KW |
2633 | spin_unlock(&qp->s_lock); |
2634 | rvt_rc_error(qp, IB_WC_LOC_LEN_ERR); | |
2635 | goto done; | |
9905bf06 | 2636 | |
d72fe7d5 | 2637 | done_unlock: |
9905bf06 KW |
2638 | spin_unlock(&qp->s_lock); |
2639 | done: | |
2640 | return true; | |
2641 | } | |
2642 | ||
2643 | static void restart_tid_rdma_read_req(struct hfi1_ctxtdata *rcd, | |
2644 | struct rvt_qp *qp, struct rvt_swqe *wqe) | |
2645 | { | |
2646 | struct tid_rdma_request *req; | |
2647 | struct tid_rdma_flow *flow; | |
2648 | ||
2649 | /* Start from the right segment */ | |
2650 | qp->r_flags |= RVT_R_RDMAR_SEQ; | |
2651 | req = wqe_to_tid_req(wqe); | |
2652 | flow = &req->flows[req->clear_tail]; | |
2653 | hfi1_restart_rc(qp, flow->flow_state.ib_spsn, 0); | |
2654 | if (list_empty(&qp->rspwait)) { | |
2655 | qp->r_flags |= RVT_R_RSP_SEND; | |
2656 | rvt_get_qp(qp); | |
2657 | list_add_tail(&qp->rspwait, &rcd->qp_wait_list); | |
2658 | } | |
2659 | } | |
2660 | ||
2661 | /* | |
2662 | * Handle the KDETH eflags for TID RDMA READ response. | |
2663 | * | |
2664 | * Return true if the last packet for a segment has been received and it is | |
2665 | * time to process the response normally; otherwise, return true. | |
2666 | * | |
2667 | * The caller must hold the packet->qp->r_lock and the rcu_read_lock. | |
2668 | */ | |
2669 | static bool handle_read_kdeth_eflags(struct hfi1_ctxtdata *rcd, | |
2670 | struct hfi1_packet *packet, u8 rcv_type, | |
2671 | u8 rte, u32 psn, u32 ibpsn) | |
2672 | __must_hold(&packet->qp->r_lock) __must_hold(RCU) | |
2673 | { | |
2674 | struct hfi1_pportdata *ppd = rcd->ppd; | |
2675 | struct hfi1_devdata *dd = ppd->dd; | |
2676 | struct hfi1_ibport *ibp; | |
2677 | struct rvt_swqe *wqe; | |
2678 | struct tid_rdma_request *req; | |
2679 | struct tid_rdma_flow *flow; | |
2680 | u32 ack_psn; | |
2681 | struct rvt_qp *qp = packet->qp; | |
2682 | struct hfi1_qp_priv *priv = qp->priv; | |
2683 | bool ret = true; | |
2684 | int diff = 0; | |
2685 | u32 fpsn; | |
2686 | ||
2687 | lockdep_assert_held(&qp->r_lock); | |
2688 | /* If the psn is out of valid range, drop the packet */ | |
2689 | if (cmp_psn(ibpsn, qp->s_last_psn) < 0 || | |
2690 | cmp_psn(ibpsn, qp->s_psn) > 0) | |
2691 | return ret; | |
2692 | ||
2693 | spin_lock(&qp->s_lock); | |
2694 | /* | |
2695 | * Note that NAKs implicitly ACK outstanding SEND and RDMA write | |
2696 | * requests and implicitly NAK RDMA read and atomic requests issued | |
2697 | * before the NAK'ed request. | |
2698 | */ | |
2699 | ack_psn = ibpsn - 1; | |
2700 | wqe = rvt_get_swqe_ptr(qp, qp->s_acked); | |
2701 | ibp = to_iport(qp->ibqp.device, qp->port_num); | |
2702 | ||
2703 | /* Complete WQEs that the PSN finishes. */ | |
2704 | while ((int)delta_psn(ack_psn, wqe->lpsn) >= 0) { | |
2705 | /* | |
2706 | * If this request is a RDMA read or atomic, and the NACK is | |
2707 | * for a later operation, this NACK NAKs the RDMA read or | |
2708 | * atomic. | |
2709 | */ | |
2710 | if (wqe->wr.opcode == IB_WR_RDMA_READ || | |
2711 | wqe->wr.opcode == IB_WR_TID_RDMA_READ || | |
2712 | wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP || | |
2713 | wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) { | |
2714 | /* Retry this request. */ | |
2715 | if (!(qp->r_flags & RVT_R_RDMAR_SEQ)) { | |
2716 | qp->r_flags |= RVT_R_RDMAR_SEQ; | |
2717 | if (wqe->wr.opcode == IB_WR_TID_RDMA_READ) { | |
2718 | restart_tid_rdma_read_req(rcd, qp, | |
2719 | wqe); | |
2720 | } else { | |
2721 | hfi1_restart_rc(qp, qp->s_last_psn + 1, | |
2722 | 0); | |
2723 | if (list_empty(&qp->rspwait)) { | |
2724 | qp->r_flags |= RVT_R_RSP_SEND; | |
2725 | rvt_get_qp(qp); | |
2726 | list_add_tail(/* wait */ | |
2727 | &qp->rspwait, | |
2728 | &rcd->qp_wait_list); | |
2729 | } | |
2730 | } | |
2731 | } | |
2732 | /* | |
2733 | * No need to process the NAK since we are | |
2734 | * restarting an earlier request. | |
2735 | */ | |
2736 | break; | |
2737 | } | |
2738 | ||
2739 | wqe = do_rc_completion(qp, wqe, ibp); | |
2740 | if (qp->s_acked == qp->s_tail) | |
2741 | break; | |
2742 | } | |
2743 | ||
2744 | /* Handle the eflags for the request */ | |
2745 | if (wqe->wr.opcode != IB_WR_TID_RDMA_READ) | |
2746 | goto s_unlock; | |
2747 | ||
2748 | req = wqe_to_tid_req(wqe); | |
2749 | switch (rcv_type) { | |
2750 | case RHF_RCV_TYPE_EXPECTED: | |
2751 | switch (rte) { | |
2752 | case RHF_RTE_EXPECTED_FLOW_SEQ_ERR: | |
2753 | /* | |
2754 | * On the first occurrence of a Flow Sequence error, | |
2755 | * the flag TID_FLOW_SW_PSN is set. | |
2756 | * | |
2757 | * After that, the flow is *not* reprogrammed and the | |
2758 | * protocol falls back to SW PSN checking. This is done | |
2759 | * to prevent continuous Flow Sequence errors for any | |
2760 | * packets that could be still in the fabric. | |
2761 | */ | |
2762 | flow = find_flow(req, psn, NULL); | |
2763 | if (!flow) { | |
2764 | /* | |
2765 | * We can't find the IB PSN matching the | |
2766 | * received KDETH PSN. The only thing we can | |
2767 | * do at this point is report the error to | |
2768 | * the QP. | |
2769 | */ | |
2770 | hfi1_kern_read_tid_flow_free(qp); | |
2771 | spin_unlock(&qp->s_lock); | |
2772 | rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR); | |
2773 | return ret; | |
2774 | } | |
2775 | if (priv->flow_state.flags & TID_FLOW_SW_PSN) { | |
2776 | diff = cmp_psn(psn, | |
2777 | priv->flow_state.r_next_psn); | |
2778 | if (diff > 0) { | |
2779 | if (!(qp->r_flags & RVT_R_RDMAR_SEQ)) | |
2780 | restart_tid_rdma_read_req(rcd, | |
2781 | qp, | |
2782 | wqe); | |
2783 | ||
2784 | /* Drop the packet.*/ | |
2785 | goto s_unlock; | |
2786 | } else if (diff < 0) { | |
2787 | /* | |
2788 | * If a response packet for a restarted | |
2789 | * request has come back, reset the | |
2790 | * restart flag. | |
2791 | */ | |
2792 | if (qp->r_flags & RVT_R_RDMAR_SEQ) | |
2793 | qp->r_flags &= | |
2794 | ~RVT_R_RDMAR_SEQ; | |
2795 | ||
2796 | /* Drop the packet.*/ | |
2797 | goto s_unlock; | |
2798 | } | |
2799 | ||
2800 | /* | |
2801 | * If SW PSN verification is successful and | |
2802 | * this is the last packet in the segment, tell | |
2803 | * the caller to process it as a normal packet. | |
2804 | */ | |
2805 | fpsn = full_flow_psn(flow, | |
2806 | flow->flow_state.lpsn); | |
2807 | if (cmp_psn(fpsn, psn) == 0) { | |
2808 | ret = false; | |
2809 | if (qp->r_flags & RVT_R_RDMAR_SEQ) | |
2810 | qp->r_flags &= | |
2811 | ~RVT_R_RDMAR_SEQ; | |
2812 | } | |
2813 | priv->flow_state.r_next_psn++; | |
2814 | } else { | |
2815 | u64 reg; | |
2816 | u32 last_psn; | |
2817 | ||
2818 | /* | |
2819 | * The only sane way to get the amount of | |
2820 | * progress is to read the HW flow state. | |
2821 | */ | |
2822 | reg = read_uctxt_csr(dd, rcd->ctxt, | |
2823 | RCV_TID_FLOW_TABLE + | |
2824 | (8 * flow->idx)); | |
2825 | last_psn = mask_psn(reg); | |
2826 | ||
2827 | priv->flow_state.r_next_psn = last_psn; | |
2828 | priv->flow_state.flags |= TID_FLOW_SW_PSN; | |
2829 | /* | |
2830 | * If no request has been restarted yet, | |
2831 | * restart the current one. | |
2832 | */ | |
2833 | if (!(qp->r_flags & RVT_R_RDMAR_SEQ)) | |
2834 | restart_tid_rdma_read_req(rcd, qp, | |
2835 | wqe); | |
2836 | } | |
2837 | ||
2838 | break; | |
2839 | ||
2840 | case RHF_RTE_EXPECTED_FLOW_GEN_ERR: | |
2841 | /* | |
2842 | * Since the TID flow is able to ride through | |
2843 | * generation mismatch, drop this stale packet. | |
2844 | */ | |
2845 | break; | |
2846 | ||
2847 | default: | |
2848 | break; | |
2849 | } | |
2850 | break; | |
2851 | ||
2852 | case RHF_RCV_TYPE_ERROR: | |
2853 | switch (rte) { | |
2854 | case RHF_RTE_ERROR_OP_CODE_ERR: | |
2855 | case RHF_RTE_ERROR_KHDR_MIN_LEN_ERR: | |
2856 | case RHF_RTE_ERROR_KHDR_HCRC_ERR: | |
2857 | case RHF_RTE_ERROR_KHDR_KVER_ERR: | |
2858 | case RHF_RTE_ERROR_CONTEXT_ERR: | |
2859 | case RHF_RTE_ERROR_KHDR_TID_ERR: | |
2860 | default: | |
2861 | break; | |
2862 | } | |
2863 | default: | |
2864 | break; | |
2865 | } | |
2866 | s_unlock: | |
2867 | spin_unlock(&qp->s_lock); | |
2868 | return ret; | |
2869 | } | |
2870 | ||
2871 | bool hfi1_handle_kdeth_eflags(struct hfi1_ctxtdata *rcd, | |
2872 | struct hfi1_pportdata *ppd, | |
2873 | struct hfi1_packet *packet) | |
2874 | { | |
2875 | struct hfi1_ibport *ibp = &ppd->ibport_data; | |
2876 | struct hfi1_devdata *dd = ppd->dd; | |
2877 | struct rvt_dev_info *rdi = &dd->verbs_dev.rdi; | |
2878 | u8 rcv_type = rhf_rcv_type(packet->rhf); | |
2879 | u8 rte = rhf_rcv_type_err(packet->rhf); | |
2880 | struct ib_header *hdr = packet->hdr; | |
2881 | struct ib_other_headers *ohdr = NULL; | |
2882 | int lnh = be16_to_cpu(hdr->lrh[0]) & 3; | |
2883 | u16 lid = be16_to_cpu(hdr->lrh[1]); | |
2884 | u8 opcode; | |
2885 | u32 qp_num, psn, ibpsn; | |
2886 | struct rvt_qp *qp; | |
d72fe7d5 | 2887 | struct hfi1_qp_priv *qpriv; |
9905bf06 KW |
2888 | unsigned long flags; |
2889 | bool ret = true; | |
d72fe7d5 KW |
2890 | struct rvt_ack_entry *e; |
2891 | struct tid_rdma_request *req; | |
2892 | struct tid_rdma_flow *flow; | |
9905bf06 | 2893 | |
3ce5daa2 KW |
2894 | trace_hfi1_msg_handle_kdeth_eflags(NULL, "Kdeth error: rhf ", |
2895 | packet->rhf); | |
9905bf06 KW |
2896 | if (packet->rhf & (RHF_VCRC_ERR | RHF_ICRC_ERR)) |
2897 | return ret; | |
2898 | ||
2899 | packet->ohdr = &hdr->u.oth; | |
2900 | ohdr = packet->ohdr; | |
2901 | trace_input_ibhdr(rcd->dd, packet, !!(rhf_dc_info(packet->rhf))); | |
2902 | ||
2903 | /* Get the destination QP number. */ | |
2904 | qp_num = be32_to_cpu(ohdr->u.tid_rdma.r_rsp.verbs_qp) & | |
2905 | RVT_QPN_MASK; | |
2906 | if (lid >= be16_to_cpu(IB_MULTICAST_LID_BASE)) | |
2907 | goto drop; | |
2908 | ||
2909 | psn = mask_psn(be32_to_cpu(ohdr->bth[2])); | |
2910 | opcode = (be32_to_cpu(ohdr->bth[0]) >> 24) & 0xff; | |
2911 | ||
2912 | rcu_read_lock(); | |
2913 | qp = rvt_lookup_qpn(rdi, &ibp->rvp, qp_num); | |
2914 | if (!qp) | |
2915 | goto rcu_unlock; | |
2916 | ||
2917 | packet->qp = qp; | |
2918 | ||
2919 | /* Check for valid receive state. */ | |
2920 | spin_lock_irqsave(&qp->r_lock, flags); | |
2921 | if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) { | |
2922 | ibp->rvp.n_pkt_drops++; | |
2923 | goto r_unlock; | |
2924 | } | |
2925 | ||
2926 | if (packet->rhf & RHF_TID_ERR) { | |
2927 | /* For TIDERR and RC QPs preemptively schedule a NAK */ | |
2928 | u32 tlen = rhf_pkt_len(packet->rhf); /* in bytes */ | |
2929 | ||
2930 | /* Sanity check packet */ | |
2931 | if (tlen < 24) | |
2932 | goto r_unlock; | |
2933 | ||
2934 | /* | |
2935 | * Check for GRH. We should never get packets with GRH in this | |
2936 | * path. | |
2937 | */ | |
2938 | if (lnh == HFI1_LRH_GRH) | |
2939 | goto r_unlock; | |
2940 | ||
2941 | if (tid_rdma_tid_err(rcd, packet, rcv_type, opcode)) | |
2942 | goto r_unlock; | |
2943 | } | |
2944 | ||
2945 | /* handle TID RDMA READ */ | |
2946 | if (opcode == TID_OP(READ_RESP)) { | |
2947 | ibpsn = be32_to_cpu(ohdr->u.tid_rdma.r_rsp.verbs_psn); | |
2948 | ibpsn = mask_psn(ibpsn); | |
2949 | ret = handle_read_kdeth_eflags(rcd, packet, rcv_type, rte, psn, | |
2950 | ibpsn); | |
d72fe7d5 KW |
2951 | goto r_unlock; |
2952 | } | |
2953 | ||
2954 | /* | |
2955 | * qp->s_tail_ack_queue points to the rvt_ack_entry currently being | |
2956 | * processed. These a completed sequentially so we can be sure that | |
2957 | * the pointer will not change until the entire request has completed. | |
2958 | */ | |
2959 | spin_lock(&qp->s_lock); | |
2960 | qpriv = qp->priv; | |
2961 | e = &qp->s_ack_queue[qpriv->r_tid_tail]; | |
2962 | req = ack_to_tid_req(e); | |
2963 | flow = &req->flows[req->clear_tail]; | |
2964 | ||
2965 | switch (rcv_type) { | |
2966 | case RHF_RCV_TYPE_EXPECTED: | |
2967 | switch (rte) { | |
2968 | case RHF_RTE_EXPECTED_FLOW_SEQ_ERR: | |
2969 | if (!(qpriv->s_flags & HFI1_R_TID_SW_PSN)) { | |
2970 | u64 reg; | |
2971 | ||
2972 | qpriv->s_flags |= HFI1_R_TID_SW_PSN; | |
2973 | /* | |
2974 | * The only sane way to get the amount of | |
2975 | * progress is to read the HW flow state. | |
2976 | */ | |
2977 | reg = read_uctxt_csr(dd, rcd->ctxt, | |
2978 | RCV_TID_FLOW_TABLE + | |
2979 | (8 * flow->idx)); | |
2980 | flow->flow_state.r_next_psn = mask_psn(reg); | |
2981 | qpriv->r_next_psn_kdeth = | |
2982 | flow->flow_state.r_next_psn; | |
2983 | goto nak_psn; | |
2984 | } else { | |
2985 | /* | |
2986 | * If the received PSN does not match the next | |
2987 | * expected PSN, NAK the packet. | |
2988 | * However, only do that if we know that the a | |
2989 | * NAK has already been sent. Otherwise, this | |
2990 | * mismatch could be due to packets that were | |
2991 | * already in flight. | |
2992 | */ | |
2993 | if (psn != flow->flow_state.r_next_psn) { | |
2994 | psn = flow->flow_state.r_next_psn; | |
2995 | goto nak_psn; | |
2996 | } | |
2997 | ||
2998 | qpriv->s_nak_state = 0; | |
2999 | /* | |
3000 | * If SW PSN verification is successful and this | |
3001 | * is the last packet in the segment, tell the | |
3002 | * caller to process it as a normal packet. | |
3003 | */ | |
3004 | if (psn == full_flow_psn(flow, | |
3005 | flow->flow_state.lpsn)) | |
3006 | ret = false; | |
3007 | qpriv->r_next_psn_kdeth = | |
3008 | ++flow->flow_state.r_next_psn; | |
3009 | } | |
3010 | break; | |
3011 | ||
3012 | case RHF_RTE_EXPECTED_FLOW_GEN_ERR: | |
3013 | goto nak_psn; | |
3014 | ||
3015 | default: | |
3016 | break; | |
3017 | } | |
3018 | break; | |
3019 | ||
3020 | case RHF_RCV_TYPE_ERROR: | |
3021 | switch (rte) { | |
3022 | case RHF_RTE_ERROR_OP_CODE_ERR: | |
3023 | case RHF_RTE_ERROR_KHDR_MIN_LEN_ERR: | |
3024 | case RHF_RTE_ERROR_KHDR_HCRC_ERR: | |
3025 | case RHF_RTE_ERROR_KHDR_KVER_ERR: | |
3026 | case RHF_RTE_ERROR_CONTEXT_ERR: | |
3027 | case RHF_RTE_ERROR_KHDR_TID_ERR: | |
3028 | default: | |
3029 | break; | |
3030 | } | |
3031 | default: | |
3032 | break; | |
9905bf06 KW |
3033 | } |
3034 | ||
d72fe7d5 KW |
3035 | unlock: |
3036 | spin_unlock(&qp->s_lock); | |
9905bf06 KW |
3037 | r_unlock: |
3038 | spin_unlock_irqrestore(&qp->r_lock, flags); | |
3039 | rcu_unlock: | |
3040 | rcu_read_unlock(); | |
3041 | drop: | |
3042 | return ret; | |
d72fe7d5 KW |
3043 | nak_psn: |
3044 | ibp->rvp.n_rc_seqnak++; | |
3045 | if (!qpriv->s_nak_state) { | |
3046 | qpriv->s_nak_state = IB_NAK_PSN_ERROR; | |
3047 | /* We are NAK'ing the next expected PSN */ | |
3048 | qpriv->s_nak_psn = mask_psn(flow->flow_state.r_next_psn); | |
3049 | qpriv->s_flags |= RVT_S_ACK_PENDING; | |
3050 | if (qpriv->r_tid_ack == HFI1_QP_WQE_INVALID) | |
3051 | qpriv->r_tid_ack = qpriv->r_tid_tail; | |
572f0c33 | 3052 | hfi1_schedule_tid_send(qp); |
d72fe7d5 KW |
3053 | } |
3054 | goto unlock; | |
9905bf06 | 3055 | } |
b126078e KW |
3056 | |
3057 | /* | |
3058 | * "Rewind" the TID request information. | |
3059 | * This means that we reset the state back to ACTIVE, | |
3060 | * find the proper flow, set the flow index to that flow, | |
3061 | * and reset the flow information. | |
3062 | */ | |
3063 | void hfi1_tid_rdma_restart_req(struct rvt_qp *qp, struct rvt_swqe *wqe, | |
3064 | u32 *bth2) | |
3065 | { | |
3066 | struct tid_rdma_request *req = wqe_to_tid_req(wqe); | |
3067 | struct tid_rdma_flow *flow; | |
6e38fca6 KW |
3068 | struct hfi1_qp_priv *qpriv = qp->priv; |
3069 | int diff, delta_pkts; | |
3070 | u32 tididx = 0, i; | |
b126078e KW |
3071 | u16 fidx; |
3072 | ||
3073 | if (wqe->wr.opcode == IB_WR_TID_RDMA_READ) { | |
3074 | *bth2 = mask_psn(qp->s_psn); | |
3075 | flow = find_flow_ib(req, *bth2, &fidx); | |
3ce5daa2 KW |
3076 | if (!flow) { |
3077 | trace_hfi1_msg_tid_restart_req(/* msg */ | |
3078 | qp, "!!!!!! Could not find flow to restart: bth2 ", | |
3079 | (u64)*bth2); | |
3080 | trace_hfi1_tid_req_restart_req(qp, 0, wqe->wr.opcode, | |
3081 | wqe->psn, wqe->lpsn, | |
3082 | req); | |
b126078e | 3083 | return; |
3ce5daa2 | 3084 | } |
b126078e | 3085 | } else { |
6e38fca6 KW |
3086 | fidx = req->acked_tail; |
3087 | flow = &req->flows[fidx]; | |
3088 | *bth2 = mask_psn(req->r_ack_psn); | |
b126078e KW |
3089 | } |
3090 | ||
6e38fca6 KW |
3091 | if (wqe->wr.opcode == IB_WR_TID_RDMA_READ) |
3092 | delta_pkts = delta_psn(*bth2, flow->flow_state.ib_spsn); | |
3093 | else | |
3094 | delta_pkts = delta_psn(*bth2, | |
3095 | full_flow_psn(flow, | |
3096 | flow->flow_state.spsn)); | |
3097 | ||
3ce5daa2 | 3098 | trace_hfi1_tid_flow_restart_req(qp, fidx, flow); |
6e38fca6 | 3099 | diff = delta_pkts + flow->resync_npkts; |
b126078e KW |
3100 | |
3101 | flow->sent = 0; | |
3102 | flow->pkt = 0; | |
3103 | flow->tid_idx = 0; | |
3104 | flow->tid_offset = 0; | |
3105 | if (diff) { | |
3106 | for (tididx = 0; tididx < flow->tidcnt; tididx++) { | |
3107 | u32 tidentry = flow->tid_entry[tididx], tidlen, | |
3108 | tidnpkts, npkts; | |
3109 | ||
3110 | flow->tid_offset = 0; | |
3111 | tidlen = EXP_TID_GET(tidentry, LEN) * PAGE_SIZE; | |
3112 | tidnpkts = rvt_div_round_up_mtu(qp, tidlen); | |
3113 | npkts = min_t(u32, diff, tidnpkts); | |
3114 | flow->pkt += npkts; | |
3115 | flow->sent += (npkts == tidnpkts ? tidlen : | |
3116 | npkts * qp->pmtu); | |
3117 | flow->tid_offset += npkts * qp->pmtu; | |
3118 | diff -= npkts; | |
3119 | if (!diff) | |
3120 | break; | |
3121 | } | |
3122 | } | |
6e38fca6 KW |
3123 | if (wqe->wr.opcode == IB_WR_TID_RDMA_WRITE) { |
3124 | rvt_skip_sge(&qpriv->tid_ss, (req->cur_seg * req->seg_len) + | |
3125 | flow->sent, 0); | |
3126 | /* | |
3127 | * Packet PSN is based on flow_state.spsn + flow->pkt. However, | |
3128 | * during a RESYNC, the generation is incremented and the | |
3129 | * sequence is reset to 0. Since we've adjusted the npkts in the | |
3130 | * flow and the SGE has been sufficiently advanced, we have to | |
3131 | * adjust flow->pkt in order to calculate the correct PSN. | |
3132 | */ | |
3133 | flow->pkt -= flow->resync_npkts; | |
3134 | } | |
b126078e KW |
3135 | |
3136 | if (flow->tid_offset == | |
3137 | EXP_TID_GET(flow->tid_entry[tididx], LEN) * PAGE_SIZE) { | |
3138 | tididx++; | |
3139 | flow->tid_offset = 0; | |
3140 | } | |
3141 | flow->tid_idx = tididx; | |
6e38fca6 KW |
3142 | if (wqe->wr.opcode == IB_WR_TID_RDMA_READ) |
3143 | /* Move flow_idx to correct index */ | |
3144 | req->flow_idx = fidx; | |
3145 | else | |
3146 | req->clear_tail = fidx; | |
b126078e | 3147 | |
3ce5daa2 KW |
3148 | trace_hfi1_tid_flow_restart_req(qp, fidx, flow); |
3149 | trace_hfi1_tid_req_restart_req(qp, 0, wqe->wr.opcode, wqe->psn, | |
3150 | wqe->lpsn, req); | |
b126078e | 3151 | req->state = TID_REQUEST_ACTIVE; |
6e38fca6 KW |
3152 | if (wqe->wr.opcode == IB_WR_TID_RDMA_WRITE) { |
3153 | /* Reset all the flows that we are going to resend */ | |
3154 | fidx = CIRC_NEXT(fidx, MAX_FLOWS); | |
3155 | i = qpriv->s_tid_tail; | |
3156 | do { | |
3157 | for (; CIRC_CNT(req->setup_head, fidx, MAX_FLOWS); | |
3158 | fidx = CIRC_NEXT(fidx, MAX_FLOWS)) { | |
3159 | req->flows[fidx].sent = 0; | |
3160 | req->flows[fidx].pkt = 0; | |
3161 | req->flows[fidx].tid_idx = 0; | |
3162 | req->flows[fidx].tid_offset = 0; | |
3163 | req->flows[fidx].resync_npkts = 0; | |
3164 | } | |
3165 | if (i == qpriv->s_tid_cur) | |
3166 | break; | |
3167 | do { | |
3168 | i = (++i == qp->s_size ? 0 : i); | |
3169 | wqe = rvt_get_swqe_ptr(qp, i); | |
3170 | } while (wqe->wr.opcode != IB_WR_TID_RDMA_WRITE); | |
3171 | req = wqe_to_tid_req(wqe); | |
3172 | req->cur_seg = req->ack_seg; | |
3173 | fidx = req->acked_tail; | |
3174 | /* Pull req->clear_tail back */ | |
3175 | req->clear_tail = fidx; | |
3176 | } while (1); | |
3177 | } | |
b126078e | 3178 | } |
24b11923 KW |
3179 | |
3180 | void hfi1_qp_kern_exp_rcv_clear_all(struct rvt_qp *qp) | |
3181 | { | |
3182 | int i, ret; | |
3183 | struct hfi1_qp_priv *qpriv = qp->priv; | |
3184 | struct tid_flow_state *fs; | |
3185 | ||
3186 | if (qp->ibqp.qp_type != IB_QPT_RC || !HFI1_CAP_IS_KSET(TID_RDMA)) | |
3187 | return; | |
3188 | ||
3189 | /* | |
3190 | * First, clear the flow to help prevent any delayed packets from | |
3191 | * being delivered. | |
3192 | */ | |
3193 | fs = &qpriv->flow_state; | |
3194 | if (fs->index != RXE_NUM_TID_FLOWS) | |
3195 | hfi1_kern_clear_hw_flow(qpriv->rcd, qp); | |
3196 | ||
3197 | for (i = qp->s_acked; i != qp->s_head;) { | |
3198 | struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, i); | |
3199 | ||
3200 | if (++i == qp->s_size) | |
3201 | i = 0; | |
3202 | /* Free only locally allocated TID entries */ | |
3203 | if (wqe->wr.opcode != IB_WR_TID_RDMA_READ) | |
3204 | continue; | |
3205 | do { | |
3206 | struct hfi1_swqe_priv *priv = wqe->priv; | |
3207 | ||
3208 | ret = hfi1_kern_exp_rcv_clear(&priv->tid_req); | |
3209 | } while (!ret); | |
3210 | } | |
3211 | } | |
a0b34f75 KW |
3212 | |
3213 | bool hfi1_tid_rdma_wqe_interlock(struct rvt_qp *qp, struct rvt_swqe *wqe) | |
3214 | { | |
3215 | struct rvt_swqe *prev; | |
3216 | struct hfi1_qp_priv *priv = qp->priv; | |
3217 | u32 s_prev; | |
3218 | ||
3219 | s_prev = (qp->s_cur == 0 ? qp->s_size : qp->s_cur) - 1; | |
3220 | prev = rvt_get_swqe_ptr(qp, s_prev); | |
3221 | ||
3222 | switch (wqe->wr.opcode) { | |
3223 | case IB_WR_SEND: | |
3224 | case IB_WR_SEND_WITH_IMM: | |
3225 | case IB_WR_SEND_WITH_INV: | |
3226 | case IB_WR_ATOMIC_CMP_AND_SWP: | |
3227 | case IB_WR_ATOMIC_FETCH_AND_ADD: | |
3228 | case IB_WR_RDMA_WRITE: | |
3229 | case IB_WR_RDMA_READ: | |
3230 | break; | |
3231 | case IB_WR_TID_RDMA_READ: | |
3232 | switch (prev->wr.opcode) { | |
3233 | case IB_WR_RDMA_READ: | |
3234 | if (qp->s_acked != qp->s_cur) | |
3235 | goto interlock; | |
3236 | break; | |
3237 | default: | |
3238 | break; | |
3239 | } | |
3240 | default: | |
3241 | break; | |
3242 | } | |
3243 | return false; | |
3244 | ||
3245 | interlock: | |
3246 | priv->s_flags |= HFI1_S_TID_WAIT_INTERLCK; | |
3247 | return true; | |
3248 | } | |
f1ab4efa KW |
3249 | |
3250 | /* Does @sge meet the alignment requirements for tid rdma? */ | |
3ce5daa2 KW |
3251 | static inline bool hfi1_check_sge_align(struct rvt_qp *qp, |
3252 | struct rvt_sge *sge, int num_sge) | |
f1ab4efa KW |
3253 | { |
3254 | int i; | |
3255 | ||
3ce5daa2 KW |
3256 | for (i = 0; i < num_sge; i++, sge++) { |
3257 | trace_hfi1_sge_check_align(qp, i, sge); | |
f1ab4efa KW |
3258 | if ((u64)sge->vaddr & ~PAGE_MASK || |
3259 | sge->sge_length & ~PAGE_MASK) | |
3260 | return false; | |
3ce5daa2 | 3261 | } |
f1ab4efa KW |
3262 | return true; |
3263 | } | |
3264 | ||
3265 | void setup_tid_rdma_wqe(struct rvt_qp *qp, struct rvt_swqe *wqe) | |
3266 | { | |
3267 | struct hfi1_qp_priv *qpriv = (struct hfi1_qp_priv *)qp->priv; | |
3268 | struct hfi1_swqe_priv *priv = wqe->priv; | |
3269 | struct tid_rdma_params *remote; | |
3270 | enum ib_wr_opcode new_opcode; | |
3271 | bool do_tid_rdma = false; | |
3272 | struct hfi1_pportdata *ppd = qpriv->rcd->ppd; | |
3273 | ||
3274 | if ((rdma_ah_get_dlid(&qp->remote_ah_attr) & ~((1 << ppd->lmc) - 1)) == | |
3275 | ppd->lid) | |
3276 | return; | |
3277 | if (qpriv->hdr_type != HFI1_PKT_TYPE_9B) | |
3278 | return; | |
3279 | ||
3280 | rcu_read_lock(); | |
3281 | remote = rcu_dereference(qpriv->tid_rdma.remote); | |
3282 | /* | |
3283 | * If TID RDMA is disabled by the negotiation, don't | |
3284 | * use it. | |
3285 | */ | |
3286 | if (!remote) | |
3287 | goto exit; | |
3288 | ||
3289 | if (wqe->wr.opcode == IB_WR_RDMA_READ) { | |
3ce5daa2 KW |
3290 | if (hfi1_check_sge_align(qp, &wqe->sg_list[0], |
3291 | wqe->wr.num_sge)) { | |
f1ab4efa KW |
3292 | new_opcode = IB_WR_TID_RDMA_READ; |
3293 | do_tid_rdma = true; | |
3294 | } | |
3295 | } | |
3296 | ||
3297 | if (do_tid_rdma) { | |
3298 | if (hfi1_kern_exp_rcv_alloc_flows(&priv->tid_req, GFP_ATOMIC)) | |
3299 | goto exit; | |
3300 | wqe->wr.opcode = new_opcode; | |
3301 | priv->tid_req.seg_len = | |
3302 | min_t(u32, remote->max_len, wqe->length); | |
3303 | priv->tid_req.total_segs = | |
3304 | DIV_ROUND_UP(wqe->length, priv->tid_req.seg_len); | |
3305 | /* Compute the last PSN of the request */ | |
3306 | wqe->lpsn = wqe->psn; | |
3307 | if (wqe->wr.opcode == IB_WR_TID_RDMA_READ) { | |
3308 | priv->tid_req.n_flows = remote->max_read; | |
3309 | qpriv->tid_r_reqs++; | |
3310 | wqe->lpsn += rvt_div_round_up_mtu(qp, wqe->length) - 1; | |
3311 | } | |
3312 | ||
3313 | priv->tid_req.cur_seg = 0; | |
3314 | priv->tid_req.comp_seg = 0; | |
3315 | priv->tid_req.ack_seg = 0; | |
3316 | priv->tid_req.state = TID_REQUEST_INACTIVE; | |
3ce5daa2 KW |
3317 | trace_hfi1_tid_req_setup_tid_wqe(qp, 1, wqe->wr.opcode, |
3318 | wqe->psn, wqe->lpsn, | |
3319 | &priv->tid_req); | |
f1ab4efa KW |
3320 | } |
3321 | exit: | |
3322 | rcu_read_unlock(); | |
3323 | } | |
c098bbb0 KW |
3324 | |
3325 | /* TID RDMA WRITE functions */ | |
3326 | ||
3327 | u32 hfi1_build_tid_rdma_write_req(struct rvt_qp *qp, struct rvt_swqe *wqe, | |
3328 | struct ib_other_headers *ohdr, | |
3329 | u32 *bth1, u32 *bth2, u32 *len) | |
3330 | { | |
3331 | struct hfi1_qp_priv *qpriv = qp->priv; | |
3332 | struct tid_rdma_request *req = wqe_to_tid_req(wqe); | |
3333 | struct tid_rdma_params *remote; | |
3334 | ||
3335 | rcu_read_lock(); | |
3336 | remote = rcu_dereference(qpriv->tid_rdma.remote); | |
3337 | /* | |
3338 | * Set the number of flow to be used based on negotiated | |
3339 | * parameters. | |
3340 | */ | |
3341 | req->n_flows = remote->max_write; | |
3342 | req->state = TID_REQUEST_ACTIVE; | |
3343 | ||
3344 | KDETH_RESET(ohdr->u.tid_rdma.w_req.kdeth0, KVER, 0x1); | |
3345 | KDETH_RESET(ohdr->u.tid_rdma.w_req.kdeth1, JKEY, remote->jkey); | |
3346 | ohdr->u.tid_rdma.w_req.reth.vaddr = | |
3347 | cpu_to_be64(wqe->rdma_wr.remote_addr + (wqe->length - *len)); | |
3348 | ohdr->u.tid_rdma.w_req.reth.rkey = | |
3349 | cpu_to_be32(wqe->rdma_wr.rkey); | |
3350 | ohdr->u.tid_rdma.w_req.reth.length = cpu_to_be32(*len); | |
3351 | ohdr->u.tid_rdma.w_req.verbs_qp = cpu_to_be32(qp->remote_qpn); | |
3352 | *bth1 &= ~RVT_QPN_MASK; | |
3353 | *bth1 |= remote->qp; | |
3354 | qp->s_state = TID_OP(WRITE_REQ); | |
3355 | qp->s_flags |= HFI1_S_WAIT_TID_RESP; | |
3356 | *bth2 |= IB_BTH_REQ_ACK; | |
3357 | *len = 0; | |
3358 | ||
3359 | rcu_read_unlock(); | |
3360 | return sizeof(ohdr->u.tid_rdma.w_req) / sizeof(u32); | |
3361 | } | |
07b92370 KW |
3362 | |
3363 | void hfi1_compute_tid_rdma_flow_wt(void) | |
3364 | { | |
3365 | /* | |
3366 | * Heuristic for computing the RNR timeout when waiting on the flow | |
3367 | * queue. Rather than a computationaly expensive exact estimate of when | |
3368 | * a flow will be available, we assume that if a QP is at position N in | |
3369 | * the flow queue it has to wait approximately (N + 1) * (number of | |
3370 | * segments between two sync points), assuming PMTU of 4K. The rationale | |
3371 | * for this is that flows are released and recycled at each sync point. | |
3372 | */ | |
3373 | tid_rdma_flow_wt = MAX_TID_FLOW_PSN * enum_to_mtu(OPA_MTU_4096) / | |
3374 | TID_RDMA_MAX_SEGMENT_SIZE; | |
3375 | } | |
3376 | ||
3377 | static u32 position_in_queue(struct hfi1_qp_priv *qpriv, | |
3378 | struct tid_queue *queue) | |
3379 | { | |
3380 | return qpriv->tid_enqueue - queue->dequeue; | |
3381 | } | |
3382 | ||
3383 | /* | |
3384 | * @qp: points to rvt_qp context. | |
3385 | * @to_seg: desired RNR timeout in segments. | |
3386 | * Return: index of the next highest timeout in the ib_hfi1_rnr_table[] | |
3387 | */ | |
3388 | static u32 hfi1_compute_tid_rnr_timeout(struct rvt_qp *qp, u32 to_seg) | |
3389 | { | |
3390 | struct hfi1_qp_priv *qpriv = qp->priv; | |
3391 | u64 timeout; | |
3392 | u32 bytes_per_us; | |
3393 | u8 i; | |
3394 | ||
3395 | bytes_per_us = active_egress_rate(qpriv->rcd->ppd) / 8; | |
3396 | timeout = (to_seg * TID_RDMA_MAX_SEGMENT_SIZE) / bytes_per_us; | |
3397 | /* | |
3398 | * Find the next highest value in the RNR table to the required | |
3399 | * timeout. This gives the responder some padding. | |
3400 | */ | |
3401 | for (i = 1; i <= IB_AETH_CREDIT_MASK; i++) | |
3402 | if (rvt_rnr_tbl_to_usec(i) >= timeout) | |
3403 | return i; | |
3404 | return 0; | |
3405 | } | |
3406 | ||
3407 | /** | |
3408 | * Central place for resource allocation at TID write responder, | |
3409 | * is called from write_req and write_data interrupt handlers as | |
3410 | * well as the send thread when a queued QP is scheduled for | |
3411 | * resource allocation. | |
3412 | * | |
3413 | * Iterates over (a) segments of a request and then (b) queued requests | |
3414 | * themselves to allocate resources for up to local->max_write | |
3415 | * segments across multiple requests. Stop allocating when we | |
3416 | * hit a sync point, resume allocating after data packets at | |
3417 | * sync point have been received. | |
3418 | * | |
3419 | * Resource allocation and sending of responses is decoupled. The | |
3420 | * request/segment which are being allocated and sent are as follows. | |
3421 | * Resources are allocated for: | |
3422 | * [request: qpriv->r_tid_alloc, segment: req->alloc_seg] | |
3423 | * The send thread sends: | |
3424 | * [request: qp->s_tail_ack_queue, segment:req->cur_seg] | |
3425 | */ | |
3426 | static void hfi1_tid_write_alloc_resources(struct rvt_qp *qp, bool intr_ctx) | |
3427 | { | |
3428 | struct tid_rdma_request *req; | |
3429 | struct hfi1_qp_priv *qpriv = qp->priv; | |
3430 | struct hfi1_ctxtdata *rcd = qpriv->rcd; | |
3431 | struct tid_rdma_params *local = &qpriv->tid_rdma.local; | |
3432 | struct rvt_ack_entry *e; | |
3433 | u32 npkts, to_seg; | |
3434 | bool last; | |
3435 | int ret = 0; | |
3436 | ||
3437 | lockdep_assert_held(&qp->s_lock); | |
3438 | ||
3439 | while (1) { | |
3440 | /* | |
3441 | * Don't allocate more segments if a RNR NAK has already been | |
3442 | * scheduled to avoid messing up qp->r_psn: the RNR NAK will | |
3443 | * be sent only when all allocated segments have been sent. | |
3444 | * However, if more segments are allocated before that, TID RDMA | |
3445 | * WRITE RESP packets will be sent out for these new segments | |
3446 | * before the RNR NAK packet. When the requester receives the | |
3447 | * RNR NAK packet, it will restart with qp->s_last_psn + 1, | |
3448 | * which does not match qp->r_psn and will be dropped. | |
3449 | * Consequently, the requester will exhaust its retries and | |
3450 | * put the qp into error state. | |
3451 | */ | |
3452 | if (qpriv->rnr_nak_state == TID_RNR_NAK_SEND) | |
3453 | break; | |
3454 | ||
3455 | /* No requests left to process */ | |
3456 | if (qpriv->r_tid_alloc == qpriv->r_tid_head) { | |
3457 | /* If all data has been received, clear the flow */ | |
3458 | if (qpriv->flow_state.index < RXE_NUM_TID_FLOWS && | |
3459 | !qpriv->alloc_w_segs) | |
3460 | hfi1_kern_clear_hw_flow(rcd, qp); | |
3461 | break; | |
3462 | } | |
3463 | ||
3464 | e = &qp->s_ack_queue[qpriv->r_tid_alloc]; | |
3465 | if (e->opcode != TID_OP(WRITE_REQ)) | |
3466 | goto next_req; | |
3467 | req = ack_to_tid_req(e); | |
3468 | /* Finished allocating for all segments of this request */ | |
3469 | if (req->alloc_seg >= req->total_segs) | |
3470 | goto next_req; | |
3471 | ||
3472 | /* Can allocate only a maximum of local->max_write for a QP */ | |
3473 | if (qpriv->alloc_w_segs >= local->max_write) | |
3474 | break; | |
3475 | ||
3476 | /* Don't allocate at a sync point with data packets pending */ | |
3477 | if (qpriv->sync_pt && qpriv->alloc_w_segs) | |
3478 | break; | |
3479 | ||
3480 | /* All data received at the sync point, continue */ | |
3481 | if (qpriv->sync_pt && !qpriv->alloc_w_segs) { | |
3482 | hfi1_kern_clear_hw_flow(rcd, qp); | |
3483 | qpriv->sync_pt = false; | |
3484 | if (qpriv->s_flags & HFI1_R_TID_SW_PSN) | |
3485 | qpriv->s_flags &= ~HFI1_R_TID_SW_PSN; | |
3486 | } | |
3487 | ||
3488 | /* Allocate flow if we don't have one */ | |
3489 | if (qpriv->flow_state.index >= RXE_NUM_TID_FLOWS) { | |
3490 | ret = hfi1_kern_setup_hw_flow(qpriv->rcd, qp); | |
3491 | if (ret) { | |
3492 | to_seg = tid_rdma_flow_wt * | |
3493 | position_in_queue(qpriv, | |
3494 | &rcd->flow_queue); | |
3495 | break; | |
3496 | } | |
3497 | } | |
3498 | ||
3499 | npkts = rvt_div_round_up_mtu(qp, req->seg_len); | |
3500 | ||
3501 | /* | |
3502 | * We are at a sync point if we run out of KDETH PSN space. | |
3503 | * Last PSN of every generation is reserved for RESYNC. | |
3504 | */ | |
3505 | if (qpriv->flow_state.psn + npkts > MAX_TID_FLOW_PSN - 1) { | |
3506 | qpriv->sync_pt = true; | |
3507 | break; | |
3508 | } | |
3509 | ||
3510 | /* | |
3511 | * If overtaking req->acked_tail, send an RNR NAK. Because the | |
3512 | * QP is not queued in this case, and the issue can only be | |
3513 | * caused due a delay in scheduling the second leg which we | |
3514 | * cannot estimate, we use a rather arbitrary RNR timeout of | |
3515 | * (MAX_FLOWS / 2) segments | |
3516 | */ | |
3517 | if (!CIRC_SPACE(req->setup_head, req->acked_tail, | |
3518 | MAX_FLOWS)) { | |
3519 | ret = -EAGAIN; | |
3520 | to_seg = MAX_FLOWS >> 1; | |
3521 | qpriv->s_flags |= RVT_S_ACK_PENDING; | |
572f0c33 | 3522 | hfi1_schedule_tid_send(qp); |
07b92370 KW |
3523 | break; |
3524 | } | |
3525 | ||
3526 | /* Try to allocate rcv array / TID entries */ | |
3527 | ret = hfi1_kern_exp_rcv_setup(req, &req->ss, &last); | |
3528 | if (ret == -EAGAIN) | |
3529 | to_seg = position_in_queue(qpriv, &rcd->rarr_queue); | |
3530 | if (ret) | |
3531 | break; | |
3532 | ||
3533 | qpriv->alloc_w_segs++; | |
3534 | req->alloc_seg++; | |
3535 | continue; | |
3536 | next_req: | |
3537 | /* Begin processing the next request */ | |
3538 | if (++qpriv->r_tid_alloc > | |
3539 | rvt_size_atomic(ib_to_rvt(qp->ibqp.device))) | |
3540 | qpriv->r_tid_alloc = 0; | |
3541 | } | |
3542 | ||
3543 | /* | |
3544 | * Schedule an RNR NAK to be sent if (a) flow or rcv array allocation | |
3545 | * has failed (b) we are called from the rcv handler interrupt context | |
3546 | * (c) an RNR NAK has not already been scheduled | |
3547 | */ | |
3548 | if (ret == -EAGAIN && intr_ctx && !qp->r_nak_state) | |
3549 | goto send_rnr_nak; | |
3550 | ||
3551 | return; | |
3552 | ||
3553 | send_rnr_nak: | |
3554 | lockdep_assert_held(&qp->r_lock); | |
3555 | ||
3556 | /* Set r_nak_state to prevent unrelated events from generating NAK's */ | |
3557 | qp->r_nak_state = hfi1_compute_tid_rnr_timeout(qp, to_seg) | IB_RNR_NAK; | |
3558 | ||
3559 | /* Pull back r_psn to the segment being RNR NAK'd */ | |
3560 | qp->r_psn = e->psn + req->alloc_seg; | |
3561 | qp->r_ack_psn = qp->r_psn; | |
3562 | /* | |
3563 | * Pull back r_head_ack_queue to the ack entry following the request | |
3564 | * being RNR NAK'd. This allows resources to be allocated to the request | |
3565 | * if the queued QP is scheduled. | |
3566 | */ | |
3567 | qp->r_head_ack_queue = qpriv->r_tid_alloc + 1; | |
3568 | if (qp->r_head_ack_queue > rvt_size_atomic(ib_to_rvt(qp->ibqp.device))) | |
3569 | qp->r_head_ack_queue = 0; | |
3570 | qpriv->r_tid_head = qp->r_head_ack_queue; | |
3571 | /* | |
3572 | * These send side fields are used in make_rc_ack(). They are set in | |
3573 | * hfi1_send_rc_ack() but must be set here before dropping qp->s_lock | |
3574 | * for consistency | |
3575 | */ | |
3576 | qp->s_nak_state = qp->r_nak_state; | |
3577 | qp->s_ack_psn = qp->r_ack_psn; | |
3578 | /* | |
3579 | * Clear the ACK PENDING flag to prevent unwanted ACK because we | |
3580 | * have modified qp->s_ack_psn here. | |
3581 | */ | |
3582 | qp->s_flags &= ~(RVT_S_ACK_PENDING); | |
3583 | ||
3584 | /* | |
3585 | * qpriv->rnr_nak_state is used to determine when the scheduled RNR NAK | |
3586 | * has actually been sent. qp->s_flags RVT_S_ACK_PENDING bit cannot be | |
3587 | * used for this because qp->s_lock is dropped before calling | |
3588 | * hfi1_send_rc_ack() leading to inconsistency between the receive | |
3589 | * interrupt handlers and the send thread in make_rc_ack() | |
3590 | */ | |
3591 | qpriv->rnr_nak_state = TID_RNR_NAK_SEND; | |
3592 | ||
3593 | /* | |
3594 | * Schedule RNR NAK to be sent. RNR NAK's are scheduled from the receive | |
3595 | * interrupt handlers but will be sent from the send engine behind any | |
3596 | * previous responses that may have been scheduled | |
3597 | */ | |
3598 | rc_defered_ack(rcd, qp); | |
3599 | } | |
3600 | ||
3601 | void hfi1_rc_rcv_tid_rdma_write_req(struct hfi1_packet *packet) | |
3602 | { | |
3603 | /* HANDLER FOR TID RDMA WRITE REQUEST packet (Responder side)*/ | |
3604 | ||
3605 | /* | |
3606 | * 1. Verify TID RDMA WRITE REQ as per IB_OPCODE_RC_RDMA_WRITE_FIRST | |
3607 | * (see hfi1_rc_rcv()) | |
3608 | * - Don't allow 0-length requests. | |
3609 | * 2. Put TID RDMA WRITE REQ into the response queueu (s_ack_queue) | |
3610 | * - Setup struct tid_rdma_req with request info | |
3611 | * - Prepare struct tid_rdma_flow array? | |
3612 | * 3. Set the qp->s_ack_state as state diagram in design doc. | |
3613 | * 4. Set RVT_S_RESP_PENDING in s_flags. | |
3614 | * 5. Kick the send engine (hfi1_schedule_send()) | |
3615 | */ | |
3616 | struct hfi1_ctxtdata *rcd = packet->rcd; | |
3617 | struct rvt_qp *qp = packet->qp; | |
3618 | struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num); | |
3619 | struct ib_other_headers *ohdr = packet->ohdr; | |
3620 | struct rvt_ack_entry *e; | |
3621 | unsigned long flags; | |
3622 | struct ib_reth *reth; | |
3623 | struct hfi1_qp_priv *qpriv = qp->priv; | |
3624 | struct tid_rdma_request *req; | |
3625 | u32 bth0, psn, len, rkey, num_segs; | |
3626 | bool is_fecn; | |
3627 | u8 next; | |
3628 | u64 vaddr; | |
3629 | int diff; | |
3630 | ||
3631 | bth0 = be32_to_cpu(ohdr->bth[0]); | |
3632 | if (hfi1_ruc_check_hdr(ibp, packet)) | |
3633 | return; | |
3634 | ||
3635 | is_fecn = process_ecn(qp, packet); | |
3636 | psn = mask_psn(be32_to_cpu(ohdr->bth[2])); | |
3637 | ||
3638 | if (qp->state == IB_QPS_RTR && !(qp->r_flags & RVT_R_COMM_EST)) | |
3639 | rvt_comm_est(qp); | |
3640 | ||
3641 | if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_WRITE))) | |
3642 | goto nack_inv; | |
3643 | ||
3644 | reth = &ohdr->u.tid_rdma.w_req.reth; | |
3645 | vaddr = be64_to_cpu(reth->vaddr); | |
3646 | len = be32_to_cpu(reth->length); | |
3647 | ||
3648 | num_segs = DIV_ROUND_UP(len, qpriv->tid_rdma.local.max_len); | |
3649 | diff = delta_psn(psn, qp->r_psn); | |
3650 | if (unlikely(diff)) { | |
3651 | if (tid_rdma_rcv_error(packet, ohdr, qp, psn, diff)) | |
3652 | return; | |
3653 | goto send_ack; | |
3654 | } | |
3655 | ||
3656 | /* | |
3657 | * The resent request which was previously RNR NAK'd is inserted at the | |
3658 | * location of the original request, which is one entry behind | |
3659 | * r_head_ack_queue | |
3660 | */ | |
3661 | if (qpriv->rnr_nak_state) | |
3662 | qp->r_head_ack_queue = qp->r_head_ack_queue ? | |
3663 | qp->r_head_ack_queue - 1 : | |
3664 | rvt_size_atomic(ib_to_rvt(qp->ibqp.device)); | |
3665 | ||
3666 | /* We've verified the request, insert it into the ack queue. */ | |
3667 | next = qp->r_head_ack_queue + 1; | |
3668 | if (next > rvt_size_atomic(ib_to_rvt(qp->ibqp.device))) | |
3669 | next = 0; | |
3670 | spin_lock_irqsave(&qp->s_lock, flags); | |
3671 | if (unlikely(next == qp->s_acked_ack_queue)) { | |
3672 | if (!qp->s_ack_queue[next].sent) | |
3673 | goto nack_inv_unlock; | |
3674 | update_ack_queue(qp, next); | |
3675 | } | |
3676 | e = &qp->s_ack_queue[qp->r_head_ack_queue]; | |
3677 | req = ack_to_tid_req(e); | |
3678 | ||
3679 | /* Bring previously RNR NAK'd request back to life */ | |
3680 | if (qpriv->rnr_nak_state) { | |
3681 | qp->r_nak_state = 0; | |
3682 | qp->s_nak_state = 0; | |
3683 | qpriv->rnr_nak_state = TID_RNR_NAK_INIT; | |
3684 | qp->r_psn = e->lpsn + 1; | |
3685 | req->state = TID_REQUEST_INIT; | |
3686 | goto update_head; | |
3687 | } | |
3688 | ||
3689 | if (e->rdma_sge.mr) { | |
3690 | rvt_put_mr(e->rdma_sge.mr); | |
3691 | e->rdma_sge.mr = NULL; | |
3692 | } | |
3693 | ||
3694 | /* The length needs to be in multiples of PAGE_SIZE */ | |
3695 | if (!len || len & ~PAGE_MASK) | |
3696 | goto nack_inv_unlock; | |
3697 | ||
3698 | rkey = be32_to_cpu(reth->rkey); | |
3699 | qp->r_len = len; | |
3700 | ||
3701 | if (e->opcode == TID_OP(WRITE_REQ) && | |
3702 | (req->setup_head != req->clear_tail || | |
3703 | req->clear_tail != req->acked_tail)) | |
3704 | goto nack_inv_unlock; | |
3705 | ||
3706 | if (unlikely(!rvt_rkey_ok(qp, &e->rdma_sge, qp->r_len, vaddr, | |
3707 | rkey, IB_ACCESS_REMOTE_WRITE))) | |
3708 | goto nack_acc; | |
3709 | ||
3710 | qp->r_psn += num_segs - 1; | |
3711 | ||
3712 | e->opcode = (bth0 >> 24) & 0xff; | |
3713 | e->psn = psn; | |
3714 | e->lpsn = qp->r_psn; | |
3715 | e->sent = 0; | |
3716 | ||
3717 | req->n_flows = min_t(u16, num_segs, qpriv->tid_rdma.local.max_write); | |
3718 | req->state = TID_REQUEST_INIT; | |
3719 | req->cur_seg = 0; | |
3720 | req->comp_seg = 0; | |
3721 | req->ack_seg = 0; | |
3722 | req->alloc_seg = 0; | |
3723 | req->isge = 0; | |
3724 | req->seg_len = qpriv->tid_rdma.local.max_len; | |
3725 | req->total_len = len; | |
3726 | req->total_segs = num_segs; | |
3727 | req->r_flow_psn = e->psn; | |
3728 | req->ss.sge = e->rdma_sge; | |
3729 | req->ss.num_sge = 1; | |
3730 | ||
3731 | req->flow_idx = req->setup_head; | |
3732 | req->clear_tail = req->setup_head; | |
3733 | req->acked_tail = req->setup_head; | |
3734 | ||
3735 | qp->r_state = e->opcode; | |
3736 | qp->r_nak_state = 0; | |
3737 | /* | |
3738 | * We need to increment the MSN here instead of when we | |
3739 | * finish sending the result since a duplicate request would | |
3740 | * increment it more than once. | |
3741 | */ | |
3742 | qp->r_msn++; | |
3743 | qp->r_psn++; | |
3744 | ||
3745 | if (qpriv->r_tid_tail == HFI1_QP_WQE_INVALID) { | |
3746 | qpriv->r_tid_tail = qp->r_head_ack_queue; | |
3747 | } else if (qpriv->r_tid_tail == qpriv->r_tid_head) { | |
3748 | struct tid_rdma_request *ptr; | |
3749 | ||
3750 | e = &qp->s_ack_queue[qpriv->r_tid_tail]; | |
3751 | ptr = ack_to_tid_req(e); | |
3752 | ||
3753 | if (e->opcode != TID_OP(WRITE_REQ) || | |
3754 | ptr->comp_seg == ptr->total_segs) { | |
3755 | if (qpriv->r_tid_tail == qpriv->r_tid_ack) | |
3756 | qpriv->r_tid_ack = qp->r_head_ack_queue; | |
3757 | qpriv->r_tid_tail = qp->r_head_ack_queue; | |
3758 | } | |
3759 | } | |
3760 | update_head: | |
3761 | qp->r_head_ack_queue = next; | |
3762 | qpriv->r_tid_head = qp->r_head_ack_queue; | |
3763 | ||
3764 | hfi1_tid_write_alloc_resources(qp, true); | |
3765 | ||
3766 | /* Schedule the send tasklet. */ | |
3767 | qp->s_flags |= RVT_S_RESP_PENDING; | |
3768 | hfi1_schedule_send(qp); | |
3769 | ||
3770 | spin_unlock_irqrestore(&qp->s_lock, flags); | |
3771 | if (is_fecn) | |
3772 | goto send_ack; | |
3773 | return; | |
3774 | ||
3775 | nack_inv_unlock: | |
3776 | spin_unlock_irqrestore(&qp->s_lock, flags); | |
3777 | nack_inv: | |
3778 | rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR); | |
3779 | qp->r_nak_state = IB_NAK_INVALID_REQUEST; | |
3780 | qp->r_ack_psn = qp->r_psn; | |
3781 | /* Queue NAK for later */ | |
3782 | rc_defered_ack(rcd, qp); | |
3783 | return; | |
3784 | nack_acc: | |
3785 | spin_unlock_irqrestore(&qp->s_lock, flags); | |
3786 | rvt_rc_error(qp, IB_WC_LOC_PROT_ERR); | |
3787 | qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR; | |
3788 | qp->r_ack_psn = qp->r_psn; | |
3789 | send_ack: | |
3790 | hfi1_send_rc_ack(packet, is_fecn); | |
3791 | } | |
38d46d36 KW |
3792 | |
3793 | u32 hfi1_build_tid_rdma_write_resp(struct rvt_qp *qp, struct rvt_ack_entry *e, | |
3794 | struct ib_other_headers *ohdr, u32 *bth1, | |
3795 | u32 bth2, u32 *len, | |
3796 | struct rvt_sge_state **ss) | |
3797 | { | |
3798 | struct hfi1_ack_priv *epriv = e->priv; | |
3799 | struct tid_rdma_request *req = &epriv->tid_req; | |
3800 | struct hfi1_qp_priv *qpriv = qp->priv; | |
3801 | struct tid_rdma_flow *flow = NULL; | |
3802 | u32 resp_len = 0, hdwords = 0; | |
3803 | void *resp_addr = NULL; | |
3804 | struct tid_rdma_params *remote; | |
3805 | ||
3806 | flow = &req->flows[req->flow_idx]; | |
3807 | switch (req->state) { | |
3808 | default: | |
3809 | /* | |
3810 | * Try to allocate resources here in case QP was queued and was | |
3811 | * later scheduled when resources became available | |
3812 | */ | |
3813 | hfi1_tid_write_alloc_resources(qp, false); | |
3814 | ||
3815 | /* We've already sent everything which is ready */ | |
3816 | if (req->cur_seg >= req->alloc_seg) | |
3817 | goto done; | |
3818 | ||
3819 | /* | |
3820 | * Resources can be assigned but responses cannot be sent in | |
3821 | * rnr_nak state, till the resent request is received | |
3822 | */ | |
3823 | if (qpriv->rnr_nak_state == TID_RNR_NAK_SENT) | |
3824 | goto done; | |
3825 | ||
3826 | req->state = TID_REQUEST_ACTIVE; | |
3827 | req->flow_idx = CIRC_NEXT(req->flow_idx, MAX_FLOWS); | |
3c759e00 | 3828 | hfi1_add_tid_reap_timer(qp); |
38d46d36 KW |
3829 | break; |
3830 | ||
3831 | case TID_REQUEST_RESEND_ACTIVE: | |
3832 | case TID_REQUEST_RESEND: | |
3833 | req->flow_idx = CIRC_NEXT(req->flow_idx, MAX_FLOWS); | |
3834 | if (!CIRC_CNT(req->setup_head, req->flow_idx, MAX_FLOWS)) | |
3835 | req->state = TID_REQUEST_ACTIVE; | |
3836 | ||
3c759e00 | 3837 | hfi1_mod_tid_reap_timer(qp); |
38d46d36 KW |
3838 | break; |
3839 | } | |
3840 | flow->flow_state.resp_ib_psn = bth2; | |
3841 | resp_addr = (void *)flow->tid_entry; | |
3842 | resp_len = sizeof(*flow->tid_entry) * flow->tidcnt; | |
3843 | req->cur_seg++; | |
3844 | ||
3845 | memset(&ohdr->u.tid_rdma.w_rsp, 0, sizeof(ohdr->u.tid_rdma.w_rsp)); | |
3846 | epriv->ss.sge.vaddr = resp_addr; | |
3847 | epriv->ss.sge.sge_length = resp_len; | |
3848 | epriv->ss.sge.length = epriv->ss.sge.sge_length; | |
3849 | /* | |
3850 | * We can safely zero these out. Since the first SGE covers the | |
3851 | * entire packet, nothing else should even look at the MR. | |
3852 | */ | |
3853 | epriv->ss.sge.mr = NULL; | |
3854 | epriv->ss.sge.m = 0; | |
3855 | epriv->ss.sge.n = 0; | |
3856 | ||
3857 | epriv->ss.sg_list = NULL; | |
3858 | epriv->ss.total_len = epriv->ss.sge.sge_length; | |
3859 | epriv->ss.num_sge = 1; | |
3860 | ||
3861 | *ss = &epriv->ss; | |
3862 | *len = epriv->ss.total_len; | |
3863 | ||
3864 | /* Construct the TID RDMA WRITE RESP packet header */ | |
3865 | rcu_read_lock(); | |
3866 | remote = rcu_dereference(qpriv->tid_rdma.remote); | |
3867 | ||
3868 | KDETH_RESET(ohdr->u.tid_rdma.w_rsp.kdeth0, KVER, 0x1); | |
3869 | KDETH_RESET(ohdr->u.tid_rdma.w_rsp.kdeth1, JKEY, remote->jkey); | |
3870 | ohdr->u.tid_rdma.w_rsp.aeth = rvt_compute_aeth(qp); | |
3871 | ohdr->u.tid_rdma.w_rsp.tid_flow_psn = | |
3872 | cpu_to_be32((flow->flow_state.generation << | |
3873 | HFI1_KDETH_BTH_SEQ_SHIFT) | | |
3874 | (flow->flow_state.spsn & | |
3875 | HFI1_KDETH_BTH_SEQ_MASK)); | |
3876 | ohdr->u.tid_rdma.w_rsp.tid_flow_qp = | |
3877 | cpu_to_be32(qpriv->tid_rdma.local.qp | | |
3878 | ((flow->idx & TID_RDMA_DESTQP_FLOW_MASK) << | |
3879 | TID_RDMA_DESTQP_FLOW_SHIFT) | | |
3880 | qpriv->rcd->ctxt); | |
3881 | ohdr->u.tid_rdma.w_rsp.verbs_qp = cpu_to_be32(qp->remote_qpn); | |
3882 | *bth1 = remote->qp; | |
3883 | rcu_read_unlock(); | |
3884 | hdwords = sizeof(ohdr->u.tid_rdma.w_rsp) / sizeof(u32); | |
3885 | qpriv->pending_tid_w_segs++; | |
3886 | done: | |
3887 | return hdwords; | |
3888 | } | |
3c759e00 KW |
3889 | |
3890 | static void hfi1_add_tid_reap_timer(struct rvt_qp *qp) | |
3891 | { | |
3892 | struct hfi1_qp_priv *qpriv = qp->priv; | |
3893 | ||
3894 | lockdep_assert_held(&qp->s_lock); | |
3895 | if (!(qpriv->s_flags & HFI1_R_TID_RSC_TIMER)) { | |
3896 | qpriv->s_flags |= HFI1_R_TID_RSC_TIMER; | |
3897 | qpriv->s_tid_timer.expires = jiffies + | |
3898 | qpriv->tid_timer_timeout_jiffies; | |
3899 | add_timer(&qpriv->s_tid_timer); | |
3900 | } | |
3901 | } | |
3902 | ||
3903 | static void hfi1_mod_tid_reap_timer(struct rvt_qp *qp) | |
3904 | { | |
3905 | struct hfi1_qp_priv *qpriv = qp->priv; | |
3906 | ||
3907 | lockdep_assert_held(&qp->s_lock); | |
3908 | qpriv->s_flags |= HFI1_R_TID_RSC_TIMER; | |
3909 | mod_timer(&qpriv->s_tid_timer, jiffies + | |
3910 | qpriv->tid_timer_timeout_jiffies); | |
3911 | } | |
3912 | ||
3913 | static int hfi1_stop_tid_reap_timer(struct rvt_qp *qp) | |
3914 | { | |
3915 | struct hfi1_qp_priv *qpriv = qp->priv; | |
3916 | int rval = 0; | |
3917 | ||
3918 | lockdep_assert_held(&qp->s_lock); | |
3919 | if (qpriv->s_flags & HFI1_R_TID_RSC_TIMER) { | |
3920 | rval = del_timer(&qpriv->s_tid_timer); | |
3921 | qpriv->s_flags &= ~HFI1_R_TID_RSC_TIMER; | |
3922 | } | |
3923 | return rval; | |
3924 | } | |
3925 | ||
3926 | void hfi1_del_tid_reap_timer(struct rvt_qp *qp) | |
3927 | { | |
3928 | struct hfi1_qp_priv *qpriv = qp->priv; | |
3929 | ||
3930 | del_timer_sync(&qpriv->s_tid_timer); | |
3931 | qpriv->s_flags &= ~HFI1_R_TID_RSC_TIMER; | |
3932 | } | |
3933 | ||
3934 | static void hfi1_tid_timeout(struct timer_list *t) | |
3935 | { | |
3936 | struct hfi1_qp_priv *qpriv = from_timer(qpriv, t, s_tid_timer); | |
3937 | struct rvt_qp *qp = qpriv->owner; | |
3938 | struct rvt_dev_info *rdi = ib_to_rvt(qp->ibqp.device); | |
3939 | unsigned long flags; | |
3940 | u32 i; | |
3941 | ||
3942 | spin_lock_irqsave(&qp->r_lock, flags); | |
3943 | spin_lock(&qp->s_lock); | |
3944 | if (qpriv->s_flags & HFI1_R_TID_RSC_TIMER) { | |
3945 | dd_dev_warn(dd_from_ibdev(qp->ibqp.device), "[QP%u] %s %d\n", | |
3946 | qp->ibqp.qp_num, __func__, __LINE__); | |
3947 | hfi1_stop_tid_reap_timer(qp); | |
3948 | /* | |
3949 | * Go though the entire ack queue and clear any outstanding | |
3950 | * HW flow and RcvArray resources. | |
3951 | */ | |
3952 | hfi1_kern_clear_hw_flow(qpriv->rcd, qp); | |
3953 | for (i = 0; i < rvt_max_atomic(rdi); i++) { | |
3954 | struct tid_rdma_request *req = | |
3955 | ack_to_tid_req(&qp->s_ack_queue[i]); | |
3956 | ||
3957 | hfi1_kern_exp_rcv_clear_all(req); | |
3958 | } | |
3959 | spin_unlock(&qp->s_lock); | |
3960 | if (qp->ibqp.event_handler) { | |
3961 | struct ib_event ev; | |
3962 | ||
3963 | ev.device = qp->ibqp.device; | |
3964 | ev.element.qp = &qp->ibqp; | |
3965 | ev.event = IB_EVENT_QP_FATAL; | |
3966 | qp->ibqp.event_handler(&ev, qp->ibqp.qp_context); | |
3967 | } | |
3968 | rvt_rc_error(qp, IB_WC_RESP_TIMEOUT_ERR); | |
3969 | goto unlock_r_lock; | |
3970 | } | |
3971 | spin_unlock(&qp->s_lock); | |
3972 | unlock_r_lock: | |
3973 | spin_unlock_irqrestore(&qp->r_lock, flags); | |
3974 | } | |
72a0ea99 KW |
3975 | |
3976 | void hfi1_rc_rcv_tid_rdma_write_resp(struct hfi1_packet *packet) | |
3977 | { | |
3978 | /* HANDLER FOR TID RDMA WRITE RESPONSE packet (Requestor side */ | |
3979 | ||
3980 | /* | |
3981 | * 1. Find matching SWQE | |
3982 | * 2. Check that TIDENTRY array has enough space for a complete | |
3983 | * segment. If not, put QP in error state. | |
3984 | * 3. Save response data in struct tid_rdma_req and struct tid_rdma_flow | |
3985 | * 4. Remove HFI1_S_WAIT_TID_RESP from s_flags. | |
3986 | * 5. Set qp->s_state | |
3987 | * 6. Kick the send engine (hfi1_schedule_send()) | |
3988 | */ | |
3989 | struct ib_other_headers *ohdr = packet->ohdr; | |
3990 | struct rvt_qp *qp = packet->qp; | |
3991 | struct hfi1_qp_priv *qpriv = qp->priv; | |
3992 | struct hfi1_ctxtdata *rcd = packet->rcd; | |
3993 | struct rvt_swqe *wqe; | |
3994 | struct tid_rdma_request *req; | |
3995 | struct tid_rdma_flow *flow; | |
3996 | enum ib_wc_status status; | |
3997 | u32 opcode, aeth, psn, flow_psn, i, tidlen = 0, pktlen; | |
3998 | bool is_fecn; | |
3999 | unsigned long flags; | |
4000 | ||
4001 | is_fecn = process_ecn(qp, packet); | |
4002 | psn = mask_psn(be32_to_cpu(ohdr->bth[2])); | |
4003 | aeth = be32_to_cpu(ohdr->u.tid_rdma.w_rsp.aeth); | |
4004 | opcode = (be32_to_cpu(ohdr->bth[0]) >> 24) & 0xff; | |
4005 | ||
4006 | spin_lock_irqsave(&qp->s_lock, flags); | |
4007 | ||
4008 | /* Ignore invalid responses */ | |
4009 | if (cmp_psn(psn, qp->s_next_psn) >= 0) | |
4010 | goto ack_done; | |
4011 | ||
4012 | /* Ignore duplicate responses. */ | |
4013 | if (unlikely(cmp_psn(psn, qp->s_last_psn) <= 0)) | |
4014 | goto ack_done; | |
4015 | ||
4016 | if (unlikely(qp->s_acked == qp->s_tail)) | |
4017 | goto ack_done; | |
4018 | ||
4019 | /* | |
4020 | * If we are waiting for a particular packet sequence number | |
4021 | * due to a request being resent, check for it. Otherwise, | |
4022 | * ensure that we haven't missed anything. | |
4023 | */ | |
4024 | if (qp->r_flags & RVT_R_RDMAR_SEQ) { | |
4025 | if (cmp_psn(psn, qp->s_last_psn + 1) != 0) | |
4026 | goto ack_done; | |
4027 | qp->r_flags &= ~RVT_R_RDMAR_SEQ; | |
4028 | } | |
4029 | ||
4030 | wqe = rvt_get_swqe_ptr(qp, qpriv->s_tid_cur); | |
4031 | if (unlikely(wqe->wr.opcode != IB_WR_TID_RDMA_WRITE)) | |
4032 | goto ack_op_err; | |
4033 | ||
4034 | req = wqe_to_tid_req(wqe); | |
4035 | /* | |
4036 | * If we've lost ACKs and our acked_tail pointer is too far | |
4037 | * behind, don't overwrite segments. Just drop the packet and | |
4038 | * let the reliability protocol take care of it. | |
4039 | */ | |
4040 | if (!CIRC_SPACE(req->setup_head, req->acked_tail, MAX_FLOWS)) | |
4041 | goto ack_done; | |
4042 | ||
4043 | /* | |
4044 | * The call to do_rc_ack() should be last in the chain of | |
4045 | * packet checks because it will end up updating the QP state. | |
4046 | * Therefore, anything that would prevent the packet from | |
4047 | * being accepted as a successful response should be prior | |
4048 | * to it. | |
4049 | */ | |
4050 | if (!do_rc_ack(qp, aeth, psn, opcode, 0, rcd)) | |
4051 | goto ack_done; | |
4052 | ||
4053 | flow = &req->flows[req->setup_head]; | |
4054 | flow->pkt = 0; | |
4055 | flow->tid_idx = 0; | |
4056 | flow->tid_offset = 0; | |
4057 | flow->sent = 0; | |
4058 | flow->resync_npkts = 0; | |
4059 | flow->tid_qpn = be32_to_cpu(ohdr->u.tid_rdma.w_rsp.tid_flow_qp); | |
4060 | flow->idx = (flow->tid_qpn >> TID_RDMA_DESTQP_FLOW_SHIFT) & | |
4061 | TID_RDMA_DESTQP_FLOW_MASK; | |
4062 | flow_psn = mask_psn(be32_to_cpu(ohdr->u.tid_rdma.w_rsp.tid_flow_psn)); | |
4063 | flow->flow_state.generation = flow_psn >> HFI1_KDETH_BTH_SEQ_SHIFT; | |
4064 | flow->flow_state.spsn = flow_psn & HFI1_KDETH_BTH_SEQ_MASK; | |
4065 | flow->flow_state.resp_ib_psn = psn; | |
4066 | flow->length = min_t(u32, req->seg_len, | |
4067 | (wqe->length - (req->comp_seg * req->seg_len))); | |
4068 | ||
4069 | flow->npkts = rvt_div_round_up_mtu(qp, flow->length); | |
4070 | flow->flow_state.lpsn = flow->flow_state.spsn + | |
4071 | flow->npkts - 1; | |
4072 | /* payload length = packet length - (header length + ICRC length) */ | |
4073 | pktlen = packet->tlen - (packet->hlen + 4); | |
4074 | if (pktlen > sizeof(flow->tid_entry)) { | |
4075 | status = IB_WC_LOC_LEN_ERR; | |
4076 | goto ack_err; | |
4077 | } | |
4078 | memcpy(flow->tid_entry, packet->ebuf, pktlen); | |
4079 | flow->tidcnt = pktlen / sizeof(*flow->tid_entry); | |
4080 | ||
4081 | req->comp_seg++; | |
4082 | /* | |
4083 | * Walk the TID_ENTRY list to make sure we have enough space for a | |
4084 | * complete segment. | |
4085 | */ | |
4086 | for (i = 0; i < flow->tidcnt; i++) { | |
4087 | if (!EXP_TID_GET(flow->tid_entry[i], LEN)) { | |
4088 | status = IB_WC_LOC_LEN_ERR; | |
4089 | goto ack_err; | |
4090 | } | |
4091 | tidlen += EXP_TID_GET(flow->tid_entry[i], LEN); | |
4092 | } | |
4093 | if (tidlen * PAGE_SIZE < flow->length) { | |
4094 | status = IB_WC_LOC_LEN_ERR; | |
4095 | goto ack_err; | |
4096 | } | |
4097 | ||
4098 | /* | |
4099 | * If this is the first response for this request, set the initial | |
4100 | * flow index to the current flow. | |
4101 | */ | |
4102 | if (!cmp_psn(psn, wqe->psn)) { | |
4103 | req->r_last_acked = mask_psn(wqe->psn - 1); | |
4104 | /* Set acked flow index to head index */ | |
4105 | req->acked_tail = req->setup_head; | |
4106 | } | |
4107 | ||
4108 | /* advance circular buffer head */ | |
4109 | req->setup_head = CIRC_NEXT(req->setup_head, MAX_FLOWS); | |
4110 | req->state = TID_REQUEST_ACTIVE; | |
4111 | ||
4112 | /* | |
4113 | * If all responses for this TID RDMA WRITE request have been received | |
4114 | * advance the pointer to the next one. | |
4115 | * Since TID RDMA requests could be mixed in with regular IB requests, | |
4116 | * they might not appear sequentially in the queue. Therefore, the | |
4117 | * next request needs to be "found". | |
4118 | */ | |
4119 | if (qpriv->s_tid_cur != qpriv->s_tid_head && | |
4120 | req->comp_seg == req->total_segs) { | |
4121 | for (i = qpriv->s_tid_cur + 1; ; i++) { | |
4122 | if (i == qp->s_size) | |
4123 | i = 0; | |
4124 | wqe = rvt_get_swqe_ptr(qp, i); | |
4125 | if (i == qpriv->s_tid_head) | |
4126 | break; | |
4127 | if (wqe->wr.opcode == IB_WR_TID_RDMA_WRITE) | |
4128 | break; | |
4129 | } | |
4130 | qpriv->s_tid_cur = i; | |
4131 | } | |
4132 | qp->s_flags &= ~HFI1_S_WAIT_TID_RESP; | |
4133 | ||
572f0c33 | 4134 | hfi1_schedule_tid_send(qp); |
72a0ea99 KW |
4135 | goto ack_done; |
4136 | ||
4137 | ack_op_err: | |
4138 | status = IB_WC_LOC_QP_OP_ERR; | |
4139 | ack_err: | |
4140 | rvt_error_qp(qp, status); | |
4141 | ack_done: | |
4142 | spin_unlock_irqrestore(&qp->s_lock, flags); | |
4143 | if (is_fecn) | |
4144 | hfi1_send_rc_ack(packet, is_fecn); | |
4145 | } | |
539e1908 KW |
4146 | |
4147 | bool hfi1_build_tid_rdma_packet(struct rvt_swqe *wqe, | |
4148 | struct ib_other_headers *ohdr, | |
4149 | u32 *bth1, u32 *bth2, u32 *len) | |
4150 | { | |
4151 | struct tid_rdma_request *req = wqe_to_tid_req(wqe); | |
4152 | struct tid_rdma_flow *flow = &req->flows[req->clear_tail]; | |
4153 | struct tid_rdma_params *remote; | |
4154 | struct rvt_qp *qp = req->qp; | |
4155 | struct hfi1_qp_priv *qpriv = qp->priv; | |
4156 | u32 tidentry = flow->tid_entry[flow->tid_idx]; | |
4157 | u32 tidlen = EXP_TID_GET(tidentry, LEN) << PAGE_SHIFT; | |
4158 | struct tid_rdma_write_data *wd = &ohdr->u.tid_rdma.w_data; | |
4159 | u32 next_offset, om = KDETH_OM_LARGE; | |
4160 | bool last_pkt; | |
4161 | ||
4162 | if (!tidlen) { | |
4163 | hfi1_trdma_send_complete(qp, wqe, IB_WC_REM_INV_RD_REQ_ERR); | |
4164 | rvt_error_qp(qp, IB_WC_REM_INV_RD_REQ_ERR); | |
4165 | } | |
4166 | ||
4167 | *len = min_t(u32, qp->pmtu, tidlen - flow->tid_offset); | |
4168 | flow->sent += *len; | |
4169 | next_offset = flow->tid_offset + *len; | |
4170 | last_pkt = (flow->tid_idx == (flow->tidcnt - 1) && | |
4171 | next_offset >= tidlen) || (flow->sent >= flow->length); | |
4172 | ||
4173 | rcu_read_lock(); | |
4174 | remote = rcu_dereference(qpriv->tid_rdma.remote); | |
4175 | KDETH_RESET(wd->kdeth0, KVER, 0x1); | |
4176 | KDETH_SET(wd->kdeth0, SH, !last_pkt); | |
4177 | KDETH_SET(wd->kdeth0, INTR, !!(!last_pkt && remote->urg)); | |
4178 | KDETH_SET(wd->kdeth0, TIDCTRL, EXP_TID_GET(tidentry, CTRL)); | |
4179 | KDETH_SET(wd->kdeth0, TID, EXP_TID_GET(tidentry, IDX)); | |
4180 | KDETH_SET(wd->kdeth0, OM, om == KDETH_OM_LARGE); | |
4181 | KDETH_SET(wd->kdeth0, OFFSET, flow->tid_offset / om); | |
4182 | KDETH_RESET(wd->kdeth1, JKEY, remote->jkey); | |
4183 | wd->verbs_qp = cpu_to_be32(qp->remote_qpn); | |
4184 | rcu_read_unlock(); | |
4185 | ||
4186 | *bth1 = flow->tid_qpn; | |
4187 | *bth2 = mask_psn(((flow->flow_state.spsn + flow->pkt++) & | |
4188 | HFI1_KDETH_BTH_SEQ_MASK) | | |
4189 | (flow->flow_state.generation << | |
4190 | HFI1_KDETH_BTH_SEQ_SHIFT)); | |
4191 | if (last_pkt) { | |
4192 | /* PSNs are zero-based, so +1 to count number of packets */ | |
4193 | if (flow->flow_state.lpsn + 1 + | |
4194 | rvt_div_round_up_mtu(qp, req->seg_len) > | |
4195 | MAX_TID_FLOW_PSN) | |
4196 | req->state = TID_REQUEST_SYNC; | |
4197 | *bth2 |= IB_BTH_REQ_ACK; | |
4198 | } | |
4199 | ||
4200 | if (next_offset >= tidlen) { | |
4201 | flow->tid_offset = 0; | |
4202 | flow->tid_idx++; | |
4203 | } else { | |
4204 | flow->tid_offset = next_offset; | |
4205 | } | |
4206 | return last_pkt; | |
4207 | } | |
d72fe7d5 KW |
4208 | |
4209 | void hfi1_rc_rcv_tid_rdma_write_data(struct hfi1_packet *packet) | |
4210 | { | |
4211 | struct rvt_qp *qp = packet->qp; | |
4212 | struct hfi1_qp_priv *priv = qp->priv; | |
4213 | struct hfi1_ctxtdata *rcd = priv->rcd; | |
4214 | struct ib_other_headers *ohdr = packet->ohdr; | |
4215 | struct rvt_ack_entry *e; | |
4216 | struct tid_rdma_request *req; | |
4217 | struct tid_rdma_flow *flow; | |
4218 | struct hfi1_ibdev *dev = to_idev(qp->ibqp.device); | |
4219 | unsigned long flags; | |
4220 | u32 psn, next; | |
4221 | u8 opcode; | |
4222 | ||
4223 | psn = mask_psn(be32_to_cpu(ohdr->bth[2])); | |
4224 | opcode = (be32_to_cpu(ohdr->bth[0]) >> 24) & 0xff; | |
4225 | ||
4226 | /* | |
4227 | * All error handling should be done by now. If we are here, the packet | |
4228 | * is either good or been accepted by the error handler. | |
4229 | */ | |
4230 | spin_lock_irqsave(&qp->s_lock, flags); | |
4231 | e = &qp->s_ack_queue[priv->r_tid_tail]; | |
4232 | req = ack_to_tid_req(e); | |
4233 | flow = &req->flows[req->clear_tail]; | |
4234 | if (cmp_psn(psn, full_flow_psn(flow, flow->flow_state.lpsn))) { | |
4235 | if (cmp_psn(psn, flow->flow_state.r_next_psn)) | |
4236 | goto send_nak; | |
4237 | flow->flow_state.r_next_psn++; | |
4238 | goto exit; | |
4239 | } | |
4240 | flow->flow_state.r_next_psn = mask_psn(psn + 1); | |
4241 | hfi1_kern_exp_rcv_clear(req); | |
4242 | priv->alloc_w_segs--; | |
4243 | rcd->flows[flow->idx].psn = psn & HFI1_KDETH_BTH_SEQ_MASK; | |
4244 | req->comp_seg++; | |
4245 | priv->s_nak_state = 0; | |
4246 | ||
4247 | /* | |
4248 | * Release the flow if one of the following conditions has been met: | |
4249 | * - The request has reached a sync point AND all outstanding | |
4250 | * segments have been completed, or | |
4251 | * - The entire request is complete and there are no more requests | |
4252 | * (of any kind) in the queue. | |
4253 | */ | |
4254 | if (priv->r_tid_ack == HFI1_QP_WQE_INVALID) | |
4255 | priv->r_tid_ack = priv->r_tid_tail; | |
4256 | ||
4257 | if (opcode == TID_OP(WRITE_DATA_LAST)) { | |
4258 | for (next = priv->r_tid_tail + 1; ; next++) { | |
4259 | if (next > rvt_size_atomic(&dev->rdi)) | |
4260 | next = 0; | |
4261 | if (next == priv->r_tid_head) | |
4262 | break; | |
4263 | e = &qp->s_ack_queue[next]; | |
4264 | if (e->opcode == TID_OP(WRITE_REQ)) | |
4265 | break; | |
4266 | } | |
4267 | priv->r_tid_tail = next; | |
4268 | if (++qp->s_acked_ack_queue > rvt_size_atomic(&dev->rdi)) | |
4269 | qp->s_acked_ack_queue = 0; | |
4270 | } | |
4271 | ||
4272 | hfi1_tid_write_alloc_resources(qp, true); | |
4273 | ||
4274 | /* | |
4275 | * If we need to generate more responses, schedule the | |
4276 | * send engine. | |
4277 | */ | |
4278 | if (req->cur_seg < req->total_segs || | |
4279 | qp->s_tail_ack_queue != qp->r_head_ack_queue) { | |
4280 | qp->s_flags |= RVT_S_RESP_PENDING; | |
4281 | hfi1_schedule_send(qp); | |
4282 | } | |
4283 | ||
4284 | priv->pending_tid_w_segs--; | |
4285 | if (priv->s_flags & HFI1_R_TID_RSC_TIMER) { | |
4286 | if (priv->pending_tid_w_segs) | |
4287 | hfi1_mod_tid_reap_timer(req->qp); | |
4288 | else | |
4289 | hfi1_stop_tid_reap_timer(req->qp); | |
4290 | } | |
4291 | ||
4292 | done: | |
4293 | priv->s_flags |= RVT_S_ACK_PENDING; | |
572f0c33 | 4294 | hfi1_schedule_tid_send(qp); |
d72fe7d5 KW |
4295 | exit: |
4296 | priv->r_next_psn_kdeth = flow->flow_state.r_next_psn; | |
4297 | spin_unlock_irqrestore(&qp->s_lock, flags); | |
4298 | return; | |
4299 | ||
4300 | send_nak: | |
4301 | if (!priv->s_nak_state) { | |
4302 | priv->s_nak_state = IB_NAK_PSN_ERROR; | |
4303 | priv->s_nak_psn = flow->flow_state.r_next_psn; | |
4304 | priv->s_flags |= RVT_S_ACK_PENDING; | |
4305 | if (priv->r_tid_ack == HFI1_QP_WQE_INVALID) | |
4306 | priv->r_tid_ack = priv->r_tid_tail; | |
572f0c33 | 4307 | hfi1_schedule_tid_send(qp); |
d72fe7d5 KW |
4308 | } |
4309 | goto done; | |
4310 | } | |
0f75e325 KW |
4311 | |
4312 | static bool hfi1_tid_rdma_is_resync_psn(u32 psn) | |
4313 | { | |
4314 | return (bool)((psn & HFI1_KDETH_BTH_SEQ_MASK) == | |
4315 | HFI1_KDETH_BTH_SEQ_MASK); | |
4316 | } | |
4317 | ||
4318 | u32 hfi1_build_tid_rdma_write_ack(struct rvt_qp *qp, struct rvt_ack_entry *e, | |
4319 | struct ib_other_headers *ohdr, u16 iflow, | |
4320 | u32 *bth1, u32 *bth2) | |
4321 | { | |
4322 | struct hfi1_qp_priv *qpriv = qp->priv; | |
4323 | struct tid_flow_state *fs = &qpriv->flow_state; | |
4324 | struct tid_rdma_request *req = ack_to_tid_req(e); | |
4325 | struct tid_rdma_flow *flow = &req->flows[iflow]; | |
4326 | struct tid_rdma_params *remote; | |
4327 | ||
4328 | rcu_read_lock(); | |
4329 | remote = rcu_dereference(qpriv->tid_rdma.remote); | |
4330 | KDETH_RESET(ohdr->u.tid_rdma.ack.kdeth1, JKEY, remote->jkey); | |
4331 | ohdr->u.tid_rdma.ack.verbs_qp = cpu_to_be32(qp->remote_qpn); | |
4332 | *bth1 = remote->qp; | |
4333 | rcu_read_unlock(); | |
4334 | ||
4335 | if (qpriv->resync) { | |
4336 | *bth2 = mask_psn((fs->generation << | |
4337 | HFI1_KDETH_BTH_SEQ_SHIFT) - 1); | |
4338 | ohdr->u.tid_rdma.ack.aeth = rvt_compute_aeth(qp); | |
4339 | } else if (qpriv->s_nak_state) { | |
4340 | *bth2 = mask_psn(qpriv->s_nak_psn); | |
4341 | ohdr->u.tid_rdma.ack.aeth = | |
4342 | cpu_to_be32((qp->r_msn & IB_MSN_MASK) | | |
4343 | (qpriv->s_nak_state << | |
4344 | IB_AETH_CREDIT_SHIFT)); | |
4345 | } else { | |
4346 | *bth2 = full_flow_psn(flow, flow->flow_state.lpsn); | |
4347 | ohdr->u.tid_rdma.ack.aeth = rvt_compute_aeth(qp); | |
4348 | } | |
4349 | KDETH_RESET(ohdr->u.tid_rdma.ack.kdeth0, KVER, 0x1); | |
4350 | ohdr->u.tid_rdma.ack.tid_flow_qp = | |
4351 | cpu_to_be32(qpriv->tid_rdma.local.qp | | |
4352 | ((flow->idx & TID_RDMA_DESTQP_FLOW_MASK) << | |
4353 | TID_RDMA_DESTQP_FLOW_SHIFT) | | |
4354 | qpriv->rcd->ctxt); | |
4355 | ||
4356 | ohdr->u.tid_rdma.ack.tid_flow_psn = 0; | |
4357 | ohdr->u.tid_rdma.ack.verbs_psn = | |
4358 | cpu_to_be32(flow->flow_state.resp_ib_psn); | |
4359 | ||
4360 | if (qpriv->resync) { | |
4361 | /* | |
4362 | * If the PSN before the current expect KDETH PSN is the | |
4363 | * RESYNC PSN, then we never received a good TID RDMA WRITE | |
4364 | * DATA packet after a previous RESYNC. | |
4365 | * In this case, the next expected KDETH PSN stays the same. | |
4366 | */ | |
4367 | if (hfi1_tid_rdma_is_resync_psn(qpriv->r_next_psn_kdeth - 1)) { | |
4368 | ohdr->u.tid_rdma.ack.tid_flow_psn = | |
4369 | cpu_to_be32(qpriv->r_next_psn_kdeth_save); | |
4370 | } else { | |
4371 | /* | |
4372 | * Because the KDETH PSNs jump during a RESYNC, it's | |
4373 | * not possible to infer (or compute) the previous value | |
4374 | * of r_next_psn_kdeth in the case of back-to-back | |
4375 | * RESYNC packets. Therefore, we save it. | |
4376 | */ | |
4377 | qpriv->r_next_psn_kdeth_save = | |
4378 | qpriv->r_next_psn_kdeth - 1; | |
4379 | ohdr->u.tid_rdma.ack.tid_flow_psn = | |
4380 | cpu_to_be32(qpriv->r_next_psn_kdeth_save); | |
4381 | qpriv->r_next_psn_kdeth = mask_psn(*bth2 + 1); | |
4382 | } | |
4383 | qpriv->resync = false; | |
4384 | } | |
4385 | ||
4386 | return sizeof(ohdr->u.tid_rdma.ack) / sizeof(u32); | |
4387 | } | |
9e93e967 KW |
4388 | |
4389 | void hfi1_rc_rcv_tid_rdma_ack(struct hfi1_packet *packet) | |
4390 | { | |
4391 | struct ib_other_headers *ohdr = packet->ohdr; | |
4392 | struct rvt_qp *qp = packet->qp; | |
4393 | struct hfi1_qp_priv *qpriv = qp->priv; | |
4394 | struct rvt_swqe *wqe; | |
4395 | struct tid_rdma_request *req; | |
4396 | struct tid_rdma_flow *flow; | |
4397 | u32 aeth, psn, req_psn, ack_psn, fspsn, resync_psn, ack_kpsn; | |
4398 | bool is_fecn; | |
4399 | unsigned long flags; | |
4400 | u16 fidx; | |
4401 | ||
4402 | is_fecn = process_ecn(qp, packet); | |
4403 | psn = mask_psn(be32_to_cpu(ohdr->bth[2])); | |
4404 | aeth = be32_to_cpu(ohdr->u.tid_rdma.ack.aeth); | |
4405 | req_psn = mask_psn(be32_to_cpu(ohdr->u.tid_rdma.ack.verbs_psn)); | |
4406 | resync_psn = mask_psn(be32_to_cpu(ohdr->u.tid_rdma.ack.tid_flow_psn)); | |
4407 | ||
4408 | spin_lock_irqsave(&qp->s_lock, flags); | |
4409 | ||
4410 | /* If we are waiting for an ACK to RESYNC, drop any other packets */ | |
4411 | if ((qp->s_flags & HFI1_S_WAIT_HALT) && | |
4412 | cmp_psn(psn, qpriv->s_resync_psn)) | |
4413 | goto ack_op_err; | |
4414 | ||
4415 | ack_psn = req_psn; | |
4416 | if (hfi1_tid_rdma_is_resync_psn(psn)) | |
4417 | ack_kpsn = resync_psn; | |
4418 | else | |
4419 | ack_kpsn = psn; | |
4420 | if (aeth >> 29) { | |
4421 | ack_psn--; | |
4422 | ack_kpsn--; | |
4423 | } | |
4424 | ||
4425 | wqe = rvt_get_swqe_ptr(qp, qp->s_acked); | |
4426 | ||
4427 | if (wqe->wr.opcode != IB_WR_TID_RDMA_WRITE) | |
4428 | goto ack_op_err; | |
4429 | ||
4430 | req = wqe_to_tid_req(wqe); | |
4431 | flow = &req->flows[req->acked_tail]; | |
4432 | ||
4433 | /* Drop stale ACK/NAK */ | |
4434 | if (cmp_psn(psn, full_flow_psn(flow, flow->flow_state.spsn)) < 0) | |
4435 | goto ack_op_err; | |
4436 | ||
4437 | while (cmp_psn(ack_kpsn, | |
4438 | full_flow_psn(flow, flow->flow_state.lpsn)) >= 0 && | |
4439 | req->ack_seg < req->cur_seg) { | |
4440 | req->ack_seg++; | |
4441 | /* advance acked segment pointer */ | |
4442 | req->acked_tail = CIRC_NEXT(req->acked_tail, MAX_FLOWS); | |
4443 | req->r_last_acked = flow->flow_state.resp_ib_psn; | |
4444 | if (req->ack_seg == req->total_segs) { | |
4445 | req->state = TID_REQUEST_COMPLETE; | |
4446 | wqe = do_rc_completion(qp, wqe, | |
4447 | to_iport(qp->ibqp.device, | |
4448 | qp->port_num)); | |
4449 | atomic_dec(&qpriv->n_tid_requests); | |
4450 | if (qp->s_acked == qp->s_tail) | |
4451 | break; | |
4452 | if (wqe->wr.opcode != IB_WR_TID_RDMA_WRITE) | |
4453 | break; | |
4454 | req = wqe_to_tid_req(wqe); | |
4455 | } | |
4456 | flow = &req->flows[req->acked_tail]; | |
4457 | } | |
4458 | ||
4459 | switch (aeth >> 29) { | |
4460 | case 0: /* ACK */ | |
4461 | if (qpriv->s_flags & RVT_S_WAIT_ACK) | |
4462 | qpriv->s_flags &= ~RVT_S_WAIT_ACK; | |
4463 | if (!hfi1_tid_rdma_is_resync_psn(psn)) { | |
829eaee5 KW |
4464 | /* Check if there is any pending TID ACK */ |
4465 | if (wqe->wr.opcode == IB_WR_TID_RDMA_WRITE && | |
4466 | req->ack_seg < req->cur_seg) | |
4467 | hfi1_mod_tid_retry_timer(qp); | |
4468 | else | |
4469 | hfi1_stop_tid_retry_timer(qp); | |
9e93e967 KW |
4470 | hfi1_schedule_send(qp); |
4471 | } else { | |
4472 | u32 spsn, fpsn, last_acked, generation; | |
4473 | struct tid_rdma_request *rptr; | |
4474 | ||
829eaee5 KW |
4475 | /* ACK(RESYNC) */ |
4476 | hfi1_stop_tid_retry_timer(qp); | |
9e93e967 KW |
4477 | /* Allow new requests (see hfi1_make_tid_rdma_pkt) */ |
4478 | qp->s_flags &= ~HFI1_S_WAIT_HALT; | |
4479 | /* | |
4480 | * Clear RVT_S_SEND_ONE flag in case that the TID RDMA | |
4481 | * ACK is received after the TID retry timer is fired | |
4482 | * again. In this case, do not send any more TID | |
4483 | * RESYNC request or wait for any more TID ACK packet. | |
4484 | */ | |
4485 | qpriv->s_flags &= ~RVT_S_SEND_ONE; | |
4486 | hfi1_schedule_send(qp); | |
4487 | ||
4488 | if ((qp->s_acked == qpriv->s_tid_tail && | |
4489 | req->ack_seg == req->total_segs) || | |
4490 | qp->s_acked == qp->s_tail) { | |
4491 | qpriv->s_state = TID_OP(WRITE_DATA_LAST); | |
4492 | goto done; | |
4493 | } | |
4494 | ||
4495 | if (req->ack_seg == req->comp_seg) { | |
4496 | qpriv->s_state = TID_OP(WRITE_DATA); | |
4497 | goto done; | |
4498 | } | |
4499 | ||
4500 | /* | |
4501 | * The PSN to start with is the next PSN after the | |
4502 | * RESYNC PSN. | |
4503 | */ | |
4504 | psn = mask_psn(psn + 1); | |
4505 | generation = psn >> HFI1_KDETH_BTH_SEQ_SHIFT; | |
4506 | spsn = 0; | |
4507 | ||
4508 | /* | |
4509 | * Update to the correct WQE when we get an ACK(RESYNC) | |
4510 | * in the middle of a request. | |
4511 | */ | |
4512 | if (delta_psn(ack_psn, wqe->lpsn)) | |
4513 | wqe = rvt_get_swqe_ptr(qp, qp->s_acked); | |
4514 | req = wqe_to_tid_req(wqe); | |
4515 | flow = &req->flows[req->acked_tail]; | |
4516 | /* | |
4517 | * RESYNC re-numbers the PSN ranges of all remaining | |
4518 | * segments. Also, PSN's start from 0 in the middle of a | |
4519 | * segment and the first segment size is less than the | |
4520 | * default number of packets. flow->resync_npkts is used | |
4521 | * to track the number of packets from the start of the | |
4522 | * real segment to the point of 0 PSN after the RESYNC | |
4523 | * in order to later correctly rewind the SGE. | |
4524 | */ | |
4525 | fpsn = full_flow_psn(flow, flow->flow_state.spsn); | |
4526 | req->r_ack_psn = psn; | |
4527 | flow->resync_npkts += | |
4528 | delta_psn(mask_psn(resync_psn + 1), fpsn); | |
4529 | /* | |
4530 | * Renumber all packet sequence number ranges | |
4531 | * based on the new generation. | |
4532 | */ | |
4533 | last_acked = qp->s_acked; | |
4534 | rptr = req; | |
4535 | while (1) { | |
4536 | /* start from last acked segment */ | |
4537 | for (fidx = rptr->acked_tail; | |
4538 | CIRC_CNT(rptr->setup_head, fidx, | |
4539 | MAX_FLOWS); | |
4540 | fidx = CIRC_NEXT(fidx, MAX_FLOWS)) { | |
4541 | u32 lpsn; | |
4542 | u32 gen; | |
4543 | ||
4544 | flow = &rptr->flows[fidx]; | |
4545 | gen = flow->flow_state.generation; | |
4546 | if (WARN_ON(gen == generation && | |
4547 | flow->flow_state.spsn != | |
4548 | spsn)) | |
4549 | continue; | |
4550 | lpsn = flow->flow_state.lpsn; | |
4551 | lpsn = full_flow_psn(flow, lpsn); | |
4552 | flow->npkts = | |
4553 | delta_psn(lpsn, | |
4554 | mask_psn(resync_psn) | |
4555 | ); | |
4556 | flow->flow_state.generation = | |
4557 | generation; | |
4558 | flow->flow_state.spsn = spsn; | |
4559 | flow->flow_state.lpsn = | |
4560 | flow->flow_state.spsn + | |
4561 | flow->npkts - 1; | |
4562 | flow->pkt = 0; | |
4563 | spsn += flow->npkts; | |
4564 | resync_psn += flow->npkts; | |
4565 | } | |
4566 | if (++last_acked == qpriv->s_tid_cur + 1) | |
4567 | break; | |
4568 | if (last_acked == qp->s_size) | |
4569 | last_acked = 0; | |
4570 | wqe = rvt_get_swqe_ptr(qp, last_acked); | |
4571 | rptr = wqe_to_tid_req(wqe); | |
4572 | } | |
4573 | req->cur_seg = req->ack_seg; | |
4574 | qpriv->s_tid_tail = qp->s_acked; | |
4575 | qpriv->s_state = TID_OP(WRITE_REQ); | |
572f0c33 | 4576 | hfi1_schedule_tid_send(qp); |
9e93e967 KW |
4577 | } |
4578 | done: | |
4579 | qpriv->s_retry = qp->s_retry_cnt; | |
4580 | break; | |
4581 | ||
4582 | case 3: /* NAK */ | |
829eaee5 | 4583 | hfi1_stop_tid_retry_timer(qp); |
9e93e967 KW |
4584 | switch ((aeth >> IB_AETH_CREDIT_SHIFT) & |
4585 | IB_AETH_CREDIT_MASK) { | |
4586 | case 0: /* PSN sequence error */ | |
4587 | flow = &req->flows[req->acked_tail]; | |
4588 | fspsn = full_flow_psn(flow, flow->flow_state.spsn); | |
4589 | req->r_ack_psn = mask_psn(be32_to_cpu(ohdr->bth[2])); | |
4590 | req->cur_seg = req->ack_seg; | |
4591 | qpriv->s_tid_tail = qp->s_acked; | |
4592 | qpriv->s_state = TID_OP(WRITE_REQ); | |
4593 | qpriv->s_retry = qp->s_retry_cnt; | |
572f0c33 | 4594 | hfi1_schedule_tid_send(qp); |
9e93e967 KW |
4595 | break; |
4596 | ||
4597 | default: | |
4598 | break; | |
4599 | } | |
4600 | break; | |
4601 | ||
4602 | default: | |
4603 | break; | |
4604 | } | |
4605 | ||
4606 | ack_op_err: | |
4607 | spin_unlock_irqrestore(&qp->s_lock, flags); | |
4608 | } | |
829eaee5 KW |
4609 | |
4610 | void hfi1_add_tid_retry_timer(struct rvt_qp *qp) | |
4611 | { | |
4612 | struct hfi1_qp_priv *priv = qp->priv; | |
4613 | struct ib_qp *ibqp = &qp->ibqp; | |
4614 | struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device); | |
4615 | ||
4616 | lockdep_assert_held(&qp->s_lock); | |
4617 | if (!(priv->s_flags & HFI1_S_TID_RETRY_TIMER)) { | |
4618 | priv->s_flags |= HFI1_S_TID_RETRY_TIMER; | |
4619 | priv->s_tid_retry_timer.expires = jiffies + | |
4620 | priv->tid_retry_timeout_jiffies + rdi->busy_jiffies; | |
4621 | add_timer(&priv->s_tid_retry_timer); | |
4622 | } | |
4623 | } | |
4624 | ||
4625 | static void hfi1_mod_tid_retry_timer(struct rvt_qp *qp) | |
4626 | { | |
4627 | struct hfi1_qp_priv *priv = qp->priv; | |
4628 | struct ib_qp *ibqp = &qp->ibqp; | |
4629 | struct rvt_dev_info *rdi = ib_to_rvt(ibqp->device); | |
4630 | ||
4631 | lockdep_assert_held(&qp->s_lock); | |
4632 | priv->s_flags |= HFI1_S_TID_RETRY_TIMER; | |
4633 | mod_timer(&priv->s_tid_retry_timer, jiffies + | |
4634 | priv->tid_retry_timeout_jiffies + rdi->busy_jiffies); | |
4635 | } | |
4636 | ||
4637 | static int hfi1_stop_tid_retry_timer(struct rvt_qp *qp) | |
4638 | { | |
4639 | struct hfi1_qp_priv *priv = qp->priv; | |
4640 | int rval = 0; | |
4641 | ||
4642 | lockdep_assert_held(&qp->s_lock); | |
4643 | if (priv->s_flags & HFI1_S_TID_RETRY_TIMER) { | |
4644 | rval = del_timer(&priv->s_tid_retry_timer); | |
4645 | priv->s_flags &= ~HFI1_S_TID_RETRY_TIMER; | |
4646 | } | |
4647 | return rval; | |
4648 | } | |
4649 | ||
4650 | void hfi1_del_tid_retry_timer(struct rvt_qp *qp) | |
4651 | { | |
4652 | struct hfi1_qp_priv *priv = qp->priv; | |
4653 | ||
4654 | del_timer_sync(&priv->s_tid_retry_timer); | |
4655 | priv->s_flags &= ~HFI1_S_TID_RETRY_TIMER; | |
4656 | } | |
4657 | ||
4658 | static void hfi1_tid_retry_timeout(struct timer_list *t) | |
4659 | { | |
4660 | struct hfi1_qp_priv *priv = from_timer(priv, t, s_tid_retry_timer); | |
4661 | struct rvt_qp *qp = priv->owner; | |
4662 | struct rvt_swqe *wqe; | |
4663 | unsigned long flags; | |
4664 | ||
4665 | spin_lock_irqsave(&qp->r_lock, flags); | |
4666 | spin_lock(&qp->s_lock); | |
4667 | if (priv->s_flags & HFI1_S_TID_RETRY_TIMER) { | |
4668 | hfi1_stop_tid_retry_timer(qp); | |
4669 | if (!priv->s_retry) { | |
4670 | wqe = rvt_get_swqe_ptr(qp, qp->s_acked); | |
4671 | hfi1_trdma_send_complete(qp, wqe, IB_WC_RETRY_EXC_ERR); | |
4672 | rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR); | |
4673 | } else { | |
4674 | priv->s_flags &= ~RVT_S_WAIT_ACK; | |
4675 | /* Only send one packet (the RESYNC) */ | |
4676 | priv->s_flags |= RVT_S_SEND_ONE; | |
4677 | /* | |
4678 | * No additional request shall be made by this QP until | |
4679 | * the RESYNC has been complete. | |
4680 | */ | |
4681 | qp->s_flags |= HFI1_S_WAIT_HALT; | |
4682 | priv->s_state = TID_OP(RESYNC); | |
4683 | priv->s_retry--; | |
572f0c33 | 4684 | hfi1_schedule_tid_send(qp); |
829eaee5 KW |
4685 | } |
4686 | } | |
4687 | spin_unlock(&qp->s_lock); | |
4688 | spin_unlock_irqrestore(&qp->r_lock, flags); | |
4689 | } | |
6e391c6a KW |
4690 | |
4691 | u32 hfi1_build_tid_rdma_resync(struct rvt_qp *qp, struct rvt_swqe *wqe, | |
4692 | struct ib_other_headers *ohdr, u32 *bth1, | |
4693 | u32 *bth2, u16 fidx) | |
4694 | { | |
4695 | struct hfi1_qp_priv *qpriv = qp->priv; | |
4696 | struct tid_rdma_params *remote; | |
4697 | struct tid_rdma_request *req = wqe_to_tid_req(wqe); | |
4698 | struct tid_rdma_flow *flow = &req->flows[fidx]; | |
4699 | u32 generation; | |
4700 | ||
4701 | rcu_read_lock(); | |
4702 | remote = rcu_dereference(qpriv->tid_rdma.remote); | |
4703 | KDETH_RESET(ohdr->u.tid_rdma.ack.kdeth1, JKEY, remote->jkey); | |
4704 | ohdr->u.tid_rdma.ack.verbs_qp = cpu_to_be32(qp->remote_qpn); | |
4705 | *bth1 = remote->qp; | |
4706 | rcu_read_unlock(); | |
4707 | ||
4708 | generation = kern_flow_generation_next(flow->flow_state.generation); | |
4709 | *bth2 = mask_psn((generation << HFI1_KDETH_BTH_SEQ_SHIFT) - 1); | |
4710 | qpriv->s_resync_psn = *bth2; | |
4711 | *bth2 |= IB_BTH_REQ_ACK; | |
4712 | KDETH_RESET(ohdr->u.tid_rdma.ack.kdeth0, KVER, 0x1); | |
4713 | ||
4714 | return sizeof(ohdr->u.tid_rdma.resync) / sizeof(u32); | |
4715 | } | |
7cf0ad67 KW |
4716 | |
4717 | void hfi1_rc_rcv_tid_rdma_resync(struct hfi1_packet *packet) | |
4718 | { | |
4719 | struct ib_other_headers *ohdr = packet->ohdr; | |
4720 | struct rvt_qp *qp = packet->qp; | |
4721 | struct hfi1_qp_priv *qpriv = qp->priv; | |
4722 | struct hfi1_ctxtdata *rcd = qpriv->rcd; | |
4723 | struct hfi1_ibdev *dev = to_idev(qp->ibqp.device); | |
4724 | struct rvt_ack_entry *e; | |
4725 | struct tid_rdma_request *req; | |
4726 | struct tid_rdma_flow *flow; | |
4727 | struct tid_flow_state *fs = &qpriv->flow_state; | |
4728 | u32 psn, generation, idx, gen_next; | |
4729 | bool is_fecn; | |
4730 | unsigned long flags; | |
4731 | ||
4732 | is_fecn = process_ecn(qp, packet); | |
4733 | psn = mask_psn(be32_to_cpu(ohdr->bth[2])); | |
4734 | ||
4735 | generation = mask_psn(psn + 1) >> HFI1_KDETH_BTH_SEQ_SHIFT; | |
4736 | spin_lock_irqsave(&qp->s_lock, flags); | |
4737 | ||
4738 | gen_next = (fs->generation == KERN_GENERATION_RESERVED) ? | |
4739 | generation : kern_flow_generation_next(fs->generation); | |
4740 | /* | |
4741 | * RESYNC packet contains the "next" generation and can only be | |
4742 | * from the current or previous generations | |
4743 | */ | |
4744 | if (generation != mask_generation(gen_next - 1) && | |
4745 | generation != gen_next) | |
4746 | goto bail; | |
4747 | /* Already processing a resync */ | |
4748 | if (qpriv->resync) | |
4749 | goto bail; | |
4750 | ||
4751 | spin_lock(&rcd->exp_lock); | |
4752 | if (fs->index >= RXE_NUM_TID_FLOWS) { | |
4753 | /* | |
4754 | * If we don't have a flow, save the generation so it can be | |
4755 | * applied when a new flow is allocated | |
4756 | */ | |
4757 | fs->generation = generation; | |
4758 | } else { | |
4759 | /* Reprogram the QP flow with new generation */ | |
4760 | rcd->flows[fs->index].generation = generation; | |
4761 | fs->generation = kern_setup_hw_flow(rcd, fs->index); | |
4762 | } | |
4763 | fs->psn = 0; | |
4764 | /* | |
4765 | * Disable SW PSN checking since a RESYNC is equivalent to a | |
4766 | * sync point and the flow has/will be reprogrammed | |
4767 | */ | |
4768 | qpriv->s_flags &= ~HFI1_R_TID_SW_PSN; | |
4769 | ||
4770 | /* | |
4771 | * Reset all TID flow information with the new generation. | |
4772 | * This is done for all requests and segments after the | |
4773 | * last received segment | |
4774 | */ | |
4775 | for (idx = qpriv->r_tid_tail; ; idx++) { | |
4776 | u16 flow_idx; | |
4777 | ||
4778 | if (idx > rvt_size_atomic(&dev->rdi)) | |
4779 | idx = 0; | |
4780 | e = &qp->s_ack_queue[idx]; | |
4781 | if (e->opcode == TID_OP(WRITE_REQ)) { | |
4782 | req = ack_to_tid_req(e); | |
4783 | ||
4784 | /* start from last unacked segment */ | |
4785 | for (flow_idx = req->clear_tail; | |
4786 | CIRC_CNT(req->setup_head, flow_idx, | |
4787 | MAX_FLOWS); | |
4788 | flow_idx = CIRC_NEXT(flow_idx, MAX_FLOWS)) { | |
4789 | u32 lpsn; | |
4790 | u32 next; | |
4791 | ||
4792 | flow = &req->flows[flow_idx]; | |
4793 | lpsn = full_flow_psn(flow, | |
4794 | flow->flow_state.lpsn); | |
4795 | next = flow->flow_state.r_next_psn; | |
4796 | flow->npkts = delta_psn(lpsn, next - 1); | |
4797 | flow->flow_state.generation = fs->generation; | |
4798 | flow->flow_state.spsn = fs->psn; | |
4799 | flow->flow_state.lpsn = | |
4800 | flow->flow_state.spsn + flow->npkts - 1; | |
4801 | flow->flow_state.r_next_psn = | |
4802 | full_flow_psn(flow, | |
4803 | flow->flow_state.spsn); | |
4804 | fs->psn += flow->npkts; | |
4805 | } | |
4806 | } | |
4807 | if (idx == qp->s_tail_ack_queue) | |
4808 | break; | |
4809 | } | |
4810 | ||
4811 | spin_unlock(&rcd->exp_lock); | |
4812 | qpriv->resync = true; | |
4813 | /* RESYNC request always gets a TID RDMA ACK. */ | |
4814 | qpriv->s_nak_state = 0; | |
4815 | qpriv->s_flags |= RVT_S_ACK_PENDING; | |
572f0c33 | 4816 | hfi1_schedule_tid_send(qp); |
7cf0ad67 KW |
4817 | bail: |
4818 | spin_unlock_irqrestore(&qp->s_lock, flags); | |
4819 | } | |
70dcb2e3 KW |
4820 | |
4821 | /* | |
4822 | * Call this function when the last TID RDMA WRITE DATA packet for a request | |
4823 | * is built. | |
4824 | */ | |
4825 | static void update_tid_tail(struct rvt_qp *qp) | |
4826 | __must_hold(&qp->s_lock) | |
4827 | { | |
4828 | struct hfi1_qp_priv *priv = qp->priv; | |
4829 | u32 i; | |
4830 | struct rvt_swqe *wqe; | |
4831 | ||
4832 | lockdep_assert_held(&qp->s_lock); | |
4833 | /* Can't move beyond s_tid_cur */ | |
4834 | if (priv->s_tid_tail == priv->s_tid_cur) | |
4835 | return; | |
4836 | for (i = priv->s_tid_tail + 1; ; i++) { | |
4837 | if (i == qp->s_size) | |
4838 | i = 0; | |
4839 | ||
4840 | if (i == priv->s_tid_cur) | |
4841 | break; | |
4842 | wqe = rvt_get_swqe_ptr(qp, i); | |
4843 | if (wqe->wr.opcode == IB_WR_TID_RDMA_WRITE) | |
4844 | break; | |
4845 | } | |
4846 | priv->s_tid_tail = i; | |
4847 | priv->s_state = TID_OP(WRITE_RESP); | |
4848 | } | |
4849 | ||
4850 | int hfi1_make_tid_rdma_pkt(struct rvt_qp *qp, struct hfi1_pkt_state *ps) | |
4851 | __must_hold(&qp->s_lock) | |
4852 | { | |
4853 | struct hfi1_qp_priv *priv = qp->priv; | |
4854 | struct rvt_swqe *wqe; | |
4855 | u32 bth1 = 0, bth2 = 0, hwords = 5, len, middle = 0; | |
4856 | struct ib_other_headers *ohdr; | |
4857 | struct rvt_sge_state *ss = &qp->s_sge; | |
4858 | struct rvt_ack_entry *e = &qp->s_ack_queue[qp->s_tail_ack_queue]; | |
4859 | struct tid_rdma_request *req = ack_to_tid_req(e); | |
4860 | bool last = false; | |
4861 | u8 opcode = TID_OP(WRITE_DATA); | |
4862 | ||
4863 | lockdep_assert_held(&qp->s_lock); | |
4864 | /* | |
4865 | * Prioritize the sending of the requests and responses over the | |
4866 | * sending of the TID RDMA data packets. | |
4867 | */ | |
4868 | if (((atomic_read(&priv->n_tid_requests) < HFI1_TID_RDMA_WRITE_CNT) && | |
4869 | atomic_read(&priv->n_requests) && | |
4870 | !(qp->s_flags & (RVT_S_BUSY | RVT_S_WAIT_ACK | | |
4871 | HFI1_S_ANY_WAIT_IO))) || | |
4872 | (e->opcode == TID_OP(WRITE_REQ) && req->cur_seg < req->alloc_seg && | |
4873 | !(qp->s_flags & (RVT_S_BUSY | HFI1_S_ANY_WAIT_IO)))) { | |
4874 | struct iowait_work *iowork; | |
4875 | ||
4876 | iowork = iowait_get_ib_work(&priv->s_iowait); | |
4877 | ps->s_txreq = get_waiting_verbs_txreq(iowork); | |
4878 | if (ps->s_txreq || hfi1_make_rc_req(qp, ps)) { | |
4879 | priv->s_flags |= HFI1_S_TID_BUSY_SET; | |
4880 | return 1; | |
4881 | } | |
4882 | } | |
4883 | ||
4884 | ps->s_txreq = get_txreq(ps->dev, qp); | |
4885 | if (!ps->s_txreq) | |
4886 | goto bail_no_tx; | |
4887 | ||
4888 | ohdr = &ps->s_txreq->phdr.hdr.ibh.u.oth; | |
4889 | ||
24c5bfea KW |
4890 | if ((priv->s_flags & RVT_S_ACK_PENDING) && |
4891 | make_tid_rdma_ack(qp, ohdr, ps)) | |
4892 | return 1; | |
4893 | ||
70dcb2e3 KW |
4894 | if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_SEND_OK)) { |
4895 | if (!(ib_rvt_state_ops[qp->state] & RVT_FLUSH_SEND)) | |
4896 | goto bail; | |
4897 | /* We are in the error state, flush the work request. */ | |
4898 | if (qp->s_last == READ_ONCE(qp->s_head)) | |
4899 | goto bail; | |
4900 | /* If DMAs are in progress, we can't flush immediately. */ | |
4901 | if (iowait_sdma_pending(&priv->s_iowait)) { | |
4902 | qp->s_flags |= RVT_S_WAIT_DMA; | |
4903 | goto bail; | |
4904 | } | |
4905 | clear_ahg(qp); | |
4906 | wqe = rvt_get_swqe_ptr(qp, qp->s_last); | |
4907 | hfi1_trdma_send_complete(qp, wqe, qp->s_last != qp->s_acked ? | |
4908 | IB_WC_SUCCESS : IB_WC_WR_FLUSH_ERR); | |
4909 | /* will get called again */ | |
4910 | goto done_free_tx; | |
4911 | } | |
4912 | ||
4913 | if (priv->s_flags & RVT_S_WAIT_ACK) | |
4914 | goto bail; | |
4915 | ||
4916 | /* Check whether there is anything to do. */ | |
4917 | if (priv->s_tid_tail == HFI1_QP_WQE_INVALID) | |
4918 | goto bail; | |
4919 | wqe = rvt_get_swqe_ptr(qp, priv->s_tid_tail); | |
4920 | req = wqe_to_tid_req(wqe); | |
4921 | switch (priv->s_state) { | |
4922 | case TID_OP(WRITE_REQ): | |
4923 | case TID_OP(WRITE_RESP): | |
4924 | priv->tid_ss.sge = wqe->sg_list[0]; | |
4925 | priv->tid_ss.sg_list = wqe->sg_list + 1; | |
4926 | priv->tid_ss.num_sge = wqe->wr.num_sge; | |
4927 | priv->tid_ss.total_len = wqe->length; | |
4928 | ||
4929 | if (priv->s_state == TID_OP(WRITE_REQ)) | |
4930 | hfi1_tid_rdma_restart_req(qp, wqe, &bth2); | |
4931 | priv->s_state = TID_OP(WRITE_DATA); | |
4932 | /* fall through */ | |
4933 | ||
4934 | case TID_OP(WRITE_DATA): | |
4935 | /* | |
4936 | * 1. Check whether TID RDMA WRITE RESP available. | |
4937 | * 2. If no: | |
4938 | * 2.1 If have more segments and no TID RDMA WRITE RESP, | |
4939 | * set HFI1_S_WAIT_TID_RESP | |
4940 | * 2.2 Return indicating no progress made. | |
4941 | * 3. If yes: | |
4942 | * 3.1 Build TID RDMA WRITE DATA packet. | |
4943 | * 3.2 If last packet in segment: | |
4944 | * 3.2.1 Change KDETH header bits | |
4945 | * 3.2.2 Advance RESP pointers. | |
4946 | * 3.3 Return indicating progress made. | |
4947 | */ | |
4948 | wqe = rvt_get_swqe_ptr(qp, priv->s_tid_tail); | |
4949 | req = wqe_to_tid_req(wqe); | |
4950 | len = wqe->length; | |
4951 | ||
4952 | if (!req->comp_seg || req->cur_seg == req->comp_seg) | |
4953 | goto bail; | |
4954 | ||
4955 | last = hfi1_build_tid_rdma_packet(wqe, ohdr, &bth1, &bth2, | |
4956 | &len); | |
4957 | ||
4958 | if (last) { | |
4959 | /* move pointer to next flow */ | |
4960 | req->clear_tail = CIRC_NEXT(req->clear_tail, | |
4961 | MAX_FLOWS); | |
4962 | if (++req->cur_seg < req->total_segs) { | |
4963 | if (!CIRC_CNT(req->setup_head, req->clear_tail, | |
4964 | MAX_FLOWS)) | |
4965 | qp->s_flags |= HFI1_S_WAIT_TID_RESP; | |
4966 | } else { | |
4967 | priv->s_state = TID_OP(WRITE_DATA_LAST); | |
4968 | opcode = TID_OP(WRITE_DATA_LAST); | |
4969 | ||
4970 | /* Advance the s_tid_tail now */ | |
4971 | update_tid_tail(qp); | |
4972 | } | |
4973 | } | |
4974 | hwords += sizeof(ohdr->u.tid_rdma.w_data) / sizeof(u32); | |
4975 | ss = &priv->tid_ss; | |
4976 | break; | |
4977 | ||
4978 | case TID_OP(RESYNC): | |
4979 | /* Use generation from the most recently received response */ | |
4980 | wqe = rvt_get_swqe_ptr(qp, priv->s_tid_cur); | |
4981 | req = wqe_to_tid_req(wqe); | |
4982 | /* If no responses for this WQE look at the previous one */ | |
4983 | if (!req->comp_seg) { | |
4984 | wqe = rvt_get_swqe_ptr(qp, | |
4985 | (!priv->s_tid_cur ? qp->s_size : | |
4986 | priv->s_tid_cur) - 1); | |
4987 | req = wqe_to_tid_req(wqe); | |
4988 | } | |
4989 | hwords += hfi1_build_tid_rdma_resync(qp, wqe, ohdr, &bth1, | |
4990 | &bth2, | |
4991 | CIRC_PREV(req->setup_head, | |
4992 | MAX_FLOWS)); | |
4993 | ss = NULL; | |
4994 | len = 0; | |
4995 | opcode = TID_OP(RESYNC); | |
4996 | break; | |
4997 | ||
4998 | default: | |
4999 | goto bail; | |
5000 | } | |
5001 | if (priv->s_flags & RVT_S_SEND_ONE) { | |
5002 | priv->s_flags &= ~RVT_S_SEND_ONE; | |
5003 | priv->s_flags |= RVT_S_WAIT_ACK; | |
5004 | bth2 |= IB_BTH_REQ_ACK; | |
5005 | } | |
5006 | qp->s_len -= len; | |
5007 | ps->s_txreq->hdr_dwords = hwords; | |
5008 | ps->s_txreq->sde = priv->s_sde; | |
5009 | ps->s_txreq->ss = ss; | |
5010 | ps->s_txreq->s_cur_size = len; | |
5011 | hfi1_make_ruc_header(qp, ohdr, (opcode << 24), bth1, bth2, | |
5012 | middle, ps); | |
5013 | return 1; | |
5014 | done_free_tx: | |
5015 | hfi1_put_txreq(ps->s_txreq); | |
5016 | ps->s_txreq = NULL; | |
5017 | return 1; | |
5018 | ||
5019 | bail: | |
5020 | hfi1_put_txreq(ps->s_txreq); | |
5021 | bail_no_tx: | |
5022 | ps->s_txreq = NULL; | |
5023 | priv->s_flags &= ~RVT_S_BUSY; | |
5024 | /* | |
5025 | * If we didn't get a txreq, the QP will be woken up later to try | |
5026 | * again, set the flags to the the wake up which work item to wake | |
5027 | * up. | |
5028 | * (A better algorithm should be found to do this and generalize the | |
5029 | * sleep/wakeup flags.) | |
5030 | */ | |
5031 | iowait_set_flag(&priv->s_iowait, IOWAIT_PENDING_TID); | |
5032 | return 0; | |
5033 | } | |
24c5bfea KW |
5034 | |
5035 | static int make_tid_rdma_ack(struct rvt_qp *qp, | |
5036 | struct ib_other_headers *ohdr, | |
5037 | struct hfi1_pkt_state *ps) | |
5038 | { | |
5039 | struct rvt_ack_entry *e; | |
5040 | struct hfi1_qp_priv *qpriv = qp->priv; | |
5041 | struct hfi1_ibdev *dev = to_idev(qp->ibqp.device); | |
5042 | u32 hwords, next; | |
5043 | u32 len = 0; | |
5044 | u32 bth1 = 0, bth2 = 0; | |
5045 | int middle = 0; | |
5046 | u16 flow; | |
5047 | struct tid_rdma_request *req, *nreq; | |
5048 | ||
5049 | /* Don't send an ACK if we aren't supposed to. */ | |
5050 | if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) | |
5051 | goto bail; | |
5052 | ||
5053 | /* header size in 32-bit words LRH+BTH = (8+12)/4. */ | |
5054 | hwords = 5; | |
5055 | ||
5056 | e = &qp->s_ack_queue[qpriv->r_tid_ack]; | |
5057 | req = ack_to_tid_req(e); | |
5058 | /* | |
5059 | * In the RESYNC case, we are exactly one segment past the | |
5060 | * previously sent ack or at the previously sent NAK. So to send | |
5061 | * the resync ack, we go back one segment (which might be part of | |
5062 | * the previous request) and let the do-while loop execute again. | |
5063 | * The advantage of executing the do-while loop is that any data | |
5064 | * received after the previous ack is automatically acked in the | |
5065 | * RESYNC ack. It turns out that for the do-while loop we only need | |
5066 | * to pull back qpriv->r_tid_ack, not the segment | |
5067 | * indices/counters. The scheme works even if the previous request | |
5068 | * was not a TID WRITE request. | |
5069 | */ | |
5070 | if (qpriv->resync) { | |
5071 | if (!req->ack_seg || req->ack_seg == req->total_segs) | |
5072 | qpriv->r_tid_ack = !qpriv->r_tid_ack ? | |
5073 | rvt_size_atomic(&dev->rdi) : | |
5074 | qpriv->r_tid_ack - 1; | |
5075 | e = &qp->s_ack_queue[qpriv->r_tid_ack]; | |
5076 | req = ack_to_tid_req(e); | |
5077 | } | |
5078 | ||
5079 | /* | |
5080 | * If we've sent all the ACKs that we can, we are done | |
5081 | * until we get more segments... | |
5082 | */ | |
5083 | if (!qpriv->s_nak_state && !qpriv->resync && | |
5084 | req->ack_seg == req->comp_seg) | |
5085 | goto bail; | |
5086 | ||
5087 | do { | |
5088 | /* | |
5089 | * To deal with coalesced ACKs, the acked_tail pointer | |
5090 | * into the flow array is used. The distance between it | |
5091 | * and the clear_tail is the number of flows that are | |
5092 | * being ACK'ed. | |
5093 | */ | |
5094 | req->ack_seg += | |
5095 | /* Get up-to-date value */ | |
5096 | CIRC_CNT(req->clear_tail, req->acked_tail, | |
5097 | MAX_FLOWS); | |
5098 | /* Advance acked index */ | |
5099 | req->acked_tail = req->clear_tail; | |
5100 | ||
5101 | /* | |
5102 | * req->clear_tail points to the segment currently being | |
5103 | * received. So, when sending an ACK, the previous | |
5104 | * segment is being ACK'ed. | |
5105 | */ | |
5106 | flow = CIRC_PREV(req->acked_tail, MAX_FLOWS); | |
5107 | if (req->ack_seg != req->total_segs) | |
5108 | break; | |
5109 | req->state = TID_REQUEST_COMPLETE; | |
5110 | ||
5111 | next = qpriv->r_tid_ack + 1; | |
5112 | if (next > rvt_size_atomic(&dev->rdi)) | |
5113 | next = 0; | |
5114 | qpriv->r_tid_ack = next; | |
5115 | if (qp->s_ack_queue[next].opcode != TID_OP(WRITE_REQ)) | |
5116 | break; | |
5117 | nreq = ack_to_tid_req(&qp->s_ack_queue[next]); | |
5118 | if (!nreq->comp_seg || nreq->ack_seg == nreq->comp_seg) | |
5119 | break; | |
5120 | ||
5121 | /* Move to the next ack entry now */ | |
5122 | e = &qp->s_ack_queue[qpriv->r_tid_ack]; | |
5123 | req = ack_to_tid_req(e); | |
5124 | } while (1); | |
5125 | ||
5126 | /* | |
5127 | * At this point qpriv->r_tid_ack == qpriv->r_tid_tail but e and | |
5128 | * req could be pointing at the previous ack queue entry | |
5129 | */ | |
5130 | if (qpriv->s_nak_state || | |
5131 | (qpriv->resync && | |
5132 | !hfi1_tid_rdma_is_resync_psn(qpriv->r_next_psn_kdeth - 1) && | |
5133 | (cmp_psn(qpriv->r_next_psn_kdeth - 1, | |
5134 | full_flow_psn(&req->flows[flow], | |
5135 | req->flows[flow].flow_state.lpsn)) > 0))) { | |
5136 | /* | |
5137 | * A NAK will implicitly acknowledge all previous TID RDMA | |
5138 | * requests. Therefore, we NAK with the req->acked_tail | |
5139 | * segment for the request at qpriv->r_tid_ack (same at | |
5140 | * this point as the req->clear_tail segment for the | |
5141 | * qpriv->r_tid_tail request) | |
5142 | */ | |
5143 | e = &qp->s_ack_queue[qpriv->r_tid_ack]; | |
5144 | req = ack_to_tid_req(e); | |
5145 | flow = req->acked_tail; | |
5146 | } | |
5147 | ||
5148 | hwords += hfi1_build_tid_rdma_write_ack(qp, e, ohdr, flow, &bth1, | |
5149 | &bth2); | |
5150 | len = 0; | |
5151 | qpriv->s_flags &= ~RVT_S_ACK_PENDING; | |
5152 | ps->s_txreq->hdr_dwords = hwords; | |
5153 | ps->s_txreq->sde = qpriv->s_sde; | |
5154 | ps->s_txreq->s_cur_size = len; | |
5155 | ps->s_txreq->ss = NULL; | |
5156 | hfi1_make_ruc_header(qp, ohdr, (TID_OP(ACK) << 24), bth1, bth2, middle, | |
5157 | ps); | |
5158 | return 1; | |
5159 | bail: | |
5160 | /* | |
5161 | * Ensure s_rdma_ack_cnt changes are committed prior to resetting | |
5162 | * RVT_S_RESP_PENDING | |
5163 | */ | |
5164 | smp_wmb(); | |
5165 | qpriv->s_flags &= ~RVT_S_ACK_PENDING; | |
5166 | return 0; | |
5167 | } | |
572f0c33 KW |
5168 | |
5169 | static int hfi1_send_tid_ok(struct rvt_qp *qp) | |
5170 | { | |
5171 | struct hfi1_qp_priv *priv = qp->priv; | |
5172 | ||
5173 | return !(priv->s_flags & RVT_S_BUSY || | |
5174 | qp->s_flags & HFI1_S_ANY_WAIT_IO) && | |
5175 | (verbs_txreq_queued(iowait_get_tid_work(&priv->s_iowait)) || | |
5176 | (priv->s_flags & RVT_S_RESP_PENDING) || | |
5177 | !(qp->s_flags & HFI1_S_ANY_TID_WAIT_SEND)); | |
5178 | } | |
5179 | ||
5180 | void _hfi1_do_tid_send(struct work_struct *work) | |
5181 | { | |
5182 | struct iowait_work *w = container_of(work, struct iowait_work, iowork); | |
5183 | struct rvt_qp *qp = iowait_to_qp(w->iow); | |
5184 | ||
5185 | hfi1_do_tid_send(qp); | |
5186 | } | |
5187 | ||
5188 | static void hfi1_do_tid_send(struct rvt_qp *qp) | |
5189 | { | |
5190 | struct hfi1_pkt_state ps; | |
5191 | struct hfi1_qp_priv *priv = qp->priv; | |
5192 | ||
5193 | ps.dev = to_idev(qp->ibqp.device); | |
5194 | ps.ibp = to_iport(qp->ibqp.device, qp->port_num); | |
5195 | ps.ppd = ppd_from_ibp(ps.ibp); | |
5196 | ps.wait = iowait_get_tid_work(&priv->s_iowait); | |
5197 | ps.in_thread = false; | |
5198 | ps.timeout_int = qp->timeout_jiffies / 8; | |
5199 | ||
5200 | spin_lock_irqsave(&qp->s_lock, ps.flags); | |
5201 | ||
5202 | /* Return if we are already busy processing a work request. */ | |
5203 | if (!hfi1_send_tid_ok(qp)) { | |
5204 | if (qp->s_flags & HFI1_S_ANY_WAIT_IO) | |
5205 | iowait_set_flag(&priv->s_iowait, IOWAIT_PENDING_TID); | |
5206 | spin_unlock_irqrestore(&qp->s_lock, ps.flags); | |
5207 | return; | |
5208 | } | |
5209 | ||
5210 | priv->s_flags |= RVT_S_BUSY; | |
5211 | ||
5212 | ps.timeout = jiffies + ps.timeout_int; | |
5213 | ps.cpu = priv->s_sde ? priv->s_sde->cpu : | |
5214 | cpumask_first(cpumask_of_node(ps.ppd->dd->node)); | |
5215 | ps.pkts_sent = false; | |
5216 | ||
5217 | /* insure a pre-built packet is handled */ | |
5218 | ps.s_txreq = get_waiting_verbs_txreq(ps.wait); | |
5219 | do { | |
5220 | /* Check for a constructed packet to be sent. */ | |
5221 | if (ps.s_txreq) { | |
5222 | if (priv->s_flags & HFI1_S_TID_BUSY_SET) { | |
5223 | qp->s_flags |= RVT_S_BUSY; | |
5224 | ps.wait = iowait_get_ib_work(&priv->s_iowait); | |
5225 | } | |
5226 | spin_unlock_irqrestore(&qp->s_lock, ps.flags); | |
5227 | ||
5228 | /* | |
5229 | * If the packet cannot be sent now, return and | |
5230 | * the send tasklet will be woken up later. | |
5231 | */ | |
5232 | if (hfi1_verbs_send(qp, &ps)) | |
5233 | return; | |
5234 | ||
5235 | /* allow other tasks to run */ | |
5236 | if (hfi1_schedule_send_yield(qp, &ps, true)) | |
5237 | return; | |
5238 | ||
5239 | spin_lock_irqsave(&qp->s_lock, ps.flags); | |
5240 | if (priv->s_flags & HFI1_S_TID_BUSY_SET) { | |
5241 | qp->s_flags &= ~RVT_S_BUSY; | |
5242 | priv->s_flags &= ~HFI1_S_TID_BUSY_SET; | |
5243 | ps.wait = iowait_get_tid_work(&priv->s_iowait); | |
5244 | if (iowait_flag_set(&priv->s_iowait, | |
5245 | IOWAIT_PENDING_IB)) | |
5246 | hfi1_schedule_send(qp); | |
5247 | } | |
5248 | } | |
5249 | } while (hfi1_make_tid_rdma_pkt(qp, &ps)); | |
5250 | iowait_starve_clear(ps.pkts_sent, &priv->s_iowait); | |
5251 | spin_unlock_irqrestore(&qp->s_lock, ps.flags); | |
5252 | } | |
5253 | ||
5254 | static bool _hfi1_schedule_tid_send(struct rvt_qp *qp) | |
5255 | { | |
5256 | struct hfi1_qp_priv *priv = qp->priv; | |
5257 | struct hfi1_ibport *ibp = | |
5258 | to_iport(qp->ibqp.device, qp->port_num); | |
5259 | struct hfi1_pportdata *ppd = ppd_from_ibp(ibp); | |
5260 | struct hfi1_devdata *dd = dd_from_ibdev(qp->ibqp.device); | |
5261 | ||
5262 | return iowait_tid_schedule(&priv->s_iowait, ppd->hfi1_wq, | |
5263 | priv->s_sde ? | |
5264 | priv->s_sde->cpu : | |
5265 | cpumask_first(cpumask_of_node(dd->node))); | |
5266 | } | |
5267 | ||
5268 | /** | |
5269 | * hfi1_schedule_tid_send - schedule progress on TID RDMA state machine | |
5270 | * @qp: the QP | |
5271 | * | |
5272 | * This schedules qp progress on the TID RDMA state machine. Caller | |
5273 | * should hold the s_lock. | |
5274 | * Unlike hfi1_schedule_send(), this cannot use hfi1_send_ok() because | |
5275 | * the two state machines can step on each other with respect to the | |
5276 | * RVT_S_BUSY flag. | |
5277 | * Therefore, a modified test is used. | |
5278 | * @return true if the second leg is scheduled; | |
5279 | * false if the second leg is not scheduled. | |
5280 | */ | |
5281 | bool hfi1_schedule_tid_send(struct rvt_qp *qp) | |
5282 | { | |
5283 | lockdep_assert_held(&qp->s_lock); | |
5284 | if (hfi1_send_tid_ok(qp)) { | |
5285 | /* | |
5286 | * The following call returns true if the qp is not on the | |
5287 | * queue and false if the qp is already on the queue before | |
5288 | * this call. Either way, the qp will be on the queue when the | |
5289 | * call returns. | |
5290 | */ | |
5291 | _hfi1_schedule_tid_send(qp); | |
5292 | return true; | |
5293 | } | |
5294 | if (qp->s_flags & HFI1_S_ANY_WAIT_IO) | |
5295 | iowait_set_flag(&((struct hfi1_qp_priv *)qp->priv)->s_iowait, | |
5296 | IOWAIT_PENDING_TID); | |
5297 | return false; | |
5298 | } |