]>
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 | ||
37356e78 | 112 | static void tid_rdma_trigger_resume(struct work_struct *work); |
838b6fd2 KW |
113 | static void hfi1_kern_exp_rcv_free_flows(struct tid_rdma_request *req); |
114 | static int hfi1_kern_exp_rcv_alloc_flows(struct tid_rdma_request *req, | |
115 | gfp_t gfp); | |
116 | static void hfi1_init_trdma_req(struct rvt_qp *qp, | |
117 | struct tid_rdma_request *req); | |
37356e78 | 118 | |
d22a207d KW |
119 | static u64 tid_rdma_opfn_encode(struct tid_rdma_params *p) |
120 | { | |
121 | return | |
122 | (((u64)p->qp & TID_OPFN_QP_CTXT_MASK) << | |
123 | TID_OPFN_QP_CTXT_SHIFT) | | |
124 | ((((u64)p->qp >> 16) & TID_OPFN_QP_KDETH_MASK) << | |
125 | TID_OPFN_QP_KDETH_SHIFT) | | |
126 | (((u64)((p->max_len >> PAGE_SHIFT) - 1) & | |
127 | TID_OPFN_MAX_LEN_MASK) << TID_OPFN_MAX_LEN_SHIFT) | | |
128 | (((u64)p->timeout & TID_OPFN_TIMEOUT_MASK) << | |
129 | TID_OPFN_TIMEOUT_SHIFT) | | |
130 | (((u64)p->urg & TID_OPFN_URG_MASK) << TID_OPFN_URG_SHIFT) | | |
131 | (((u64)p->jkey & TID_OPFN_JKEY_MASK) << TID_OPFN_JKEY_SHIFT) | | |
132 | (((u64)p->max_read & TID_OPFN_MAX_READ_MASK) << | |
133 | TID_OPFN_MAX_READ_SHIFT) | | |
134 | (((u64)p->max_write & TID_OPFN_MAX_WRITE_MASK) << | |
135 | TID_OPFN_MAX_WRITE_SHIFT); | |
136 | } | |
137 | ||
138 | static void tid_rdma_opfn_decode(struct tid_rdma_params *p, u64 data) | |
139 | { | |
140 | p->max_len = (((data >> TID_OPFN_MAX_LEN_SHIFT) & | |
141 | TID_OPFN_MAX_LEN_MASK) + 1) << PAGE_SHIFT; | |
142 | p->jkey = (data >> TID_OPFN_JKEY_SHIFT) & TID_OPFN_JKEY_MASK; | |
143 | p->max_write = (data >> TID_OPFN_MAX_WRITE_SHIFT) & | |
144 | TID_OPFN_MAX_WRITE_MASK; | |
145 | p->max_read = (data >> TID_OPFN_MAX_READ_SHIFT) & | |
146 | TID_OPFN_MAX_READ_MASK; | |
147 | p->qp = | |
148 | ((((data >> TID_OPFN_QP_KDETH_SHIFT) & TID_OPFN_QP_KDETH_MASK) | |
149 | << 16) | | |
150 | ((data >> TID_OPFN_QP_CTXT_SHIFT) & TID_OPFN_QP_CTXT_MASK)); | |
151 | p->urg = (data >> TID_OPFN_URG_SHIFT) & TID_OPFN_URG_MASK; | |
152 | p->timeout = (data >> TID_OPFN_TIMEOUT_SHIFT) & TID_OPFN_TIMEOUT_MASK; | |
153 | } | |
154 | ||
155 | void tid_rdma_opfn_init(struct rvt_qp *qp, struct tid_rdma_params *p) | |
156 | { | |
157 | struct hfi1_qp_priv *priv = qp->priv; | |
158 | ||
159 | p->qp = (kdeth_qp << 16) | priv->rcd->ctxt; | |
160 | p->max_len = TID_RDMA_MAX_SEGMENT_SIZE; | |
161 | p->jkey = priv->rcd->jkey; | |
162 | p->max_read = TID_RDMA_MAX_READ_SEGS_PER_REQ; | |
163 | p->max_write = TID_RDMA_MAX_WRITE_SEGS_PER_REQ; | |
164 | p->timeout = qp->timeout; | |
165 | p->urg = is_urg_masked(priv->rcd); | |
166 | } | |
167 | ||
168 | bool tid_rdma_conn_req(struct rvt_qp *qp, u64 *data) | |
169 | { | |
170 | struct hfi1_qp_priv *priv = qp->priv; | |
171 | ||
172 | *data = tid_rdma_opfn_encode(&priv->tid_rdma.local); | |
173 | return true; | |
174 | } | |
175 | ||
176 | bool tid_rdma_conn_reply(struct rvt_qp *qp, u64 data) | |
177 | { | |
178 | struct hfi1_qp_priv *priv = qp->priv; | |
179 | struct tid_rdma_params *remote, *old; | |
180 | bool ret = true; | |
181 | ||
182 | old = rcu_dereference_protected(priv->tid_rdma.remote, | |
183 | lockdep_is_held(&priv->opfn.lock)); | |
184 | data &= ~0xfULL; | |
185 | /* | |
186 | * If data passed in is zero, return true so as not to continue the | |
187 | * negotiation process | |
188 | */ | |
189 | if (!data || !HFI1_CAP_IS_KSET(TID_RDMA)) | |
190 | goto null; | |
191 | /* | |
192 | * If kzalloc fails, return false. This will result in: | |
193 | * * at the requester a new OPFN request being generated to retry | |
194 | * the negotiation | |
195 | * * at the responder, 0 being returned to the requester so as to | |
196 | * disable TID RDMA at both the requester and the responder | |
197 | */ | |
198 | remote = kzalloc(sizeof(*remote), GFP_ATOMIC); | |
199 | if (!remote) { | |
200 | ret = false; | |
201 | goto null; | |
202 | } | |
203 | ||
204 | tid_rdma_opfn_decode(remote, data); | |
205 | priv->tid_timer_timeout_jiffies = | |
206 | usecs_to_jiffies((((4096UL * (1UL << remote->timeout)) / | |
207 | 1000UL) << 3) * 7); | |
a131d164 KW |
208 | trace_hfi1_opfn_param(qp, 0, &priv->tid_rdma.local); |
209 | trace_hfi1_opfn_param(qp, 1, remote); | |
d22a207d KW |
210 | rcu_assign_pointer(priv->tid_rdma.remote, remote); |
211 | /* | |
212 | * A TID RDMA READ request's segment size is not equal to | |
213 | * remote->max_len only when the request's data length is smaller | |
214 | * than remote->max_len. In that case, there will be only one segment. | |
215 | * Therefore, when priv->pkts_ps is used to calculate req->cur_seg | |
216 | * during retry, it will lead to req->cur_seg = 0, which is exactly | |
217 | * what is expected. | |
218 | */ | |
219 | priv->pkts_ps = (u16)rvt_div_mtu(qp, remote->max_len); | |
220 | priv->timeout_shift = ilog2(priv->pkts_ps - 1) + 1; | |
221 | goto free; | |
222 | null: | |
223 | RCU_INIT_POINTER(priv->tid_rdma.remote, NULL); | |
224 | priv->timeout_shift = 0; | |
225 | free: | |
226 | if (old) | |
227 | kfree_rcu(old, rcu_head); | |
228 | return ret; | |
229 | } | |
230 | ||
231 | bool tid_rdma_conn_resp(struct rvt_qp *qp, u64 *data) | |
232 | { | |
233 | bool ret; | |
234 | ||
235 | ret = tid_rdma_conn_reply(qp, *data); | |
236 | *data = 0; | |
237 | /* | |
238 | * If tid_rdma_conn_reply() returns error, set *data as 0 to indicate | |
239 | * TID RDMA could not be enabled. This will result in TID RDMA being | |
240 | * disabled at the requester too. | |
241 | */ | |
242 | if (ret) | |
243 | (void)tid_rdma_conn_req(qp, data); | |
244 | return ret; | |
245 | } | |
246 | ||
247 | void tid_rdma_conn_error(struct rvt_qp *qp) | |
248 | { | |
249 | struct hfi1_qp_priv *priv = qp->priv; | |
250 | struct tid_rdma_params *old; | |
251 | ||
252 | old = rcu_dereference_protected(priv->tid_rdma.remote, | |
253 | lockdep_is_held(&priv->opfn.lock)); | |
254 | RCU_INIT_POINTER(priv->tid_rdma.remote, NULL); | |
255 | if (old) | |
256 | kfree_rcu(old, rcu_head); | |
257 | } | |
258 | ||
259 | /* This is called at context initialization time */ | |
260 | int hfi1_kern_exp_rcv_init(struct hfi1_ctxtdata *rcd, int reinit) | |
261 | { | |
262 | if (reinit) | |
263 | return 0; | |
264 | ||
265 | BUILD_BUG_ON(TID_RDMA_JKEY < HFI1_KERNEL_MIN_JKEY); | |
266 | BUILD_BUG_ON(TID_RDMA_JKEY > HFI1_KERNEL_MAX_JKEY); | |
267 | rcd->jkey = TID_RDMA_JKEY; | |
268 | hfi1_set_ctxt_jkey(rcd->dd, rcd, rcd->jkey); | |
838b6fd2 | 269 | return hfi1_alloc_ctxt_rcv_groups(rcd); |
d22a207d KW |
270 | } |
271 | ||
5190f052 MM |
272 | /** |
273 | * qp_to_rcd - determine the receive context used by a qp | |
274 | * @qp - the qp | |
275 | * | |
276 | * This routine returns the receive context associated | |
277 | * with a a qp's qpn. | |
278 | * | |
279 | * Returns the context. | |
280 | */ | |
281 | static struct hfi1_ctxtdata *qp_to_rcd(struct rvt_dev_info *rdi, | |
282 | struct rvt_qp *qp) | |
283 | { | |
284 | struct hfi1_ibdev *verbs_dev = container_of(rdi, | |
285 | struct hfi1_ibdev, | |
286 | rdi); | |
287 | struct hfi1_devdata *dd = container_of(verbs_dev, | |
288 | struct hfi1_devdata, | |
289 | verbs_dev); | |
290 | unsigned int ctxt; | |
291 | ||
292 | if (qp->ibqp.qp_num == 0) | |
293 | ctxt = 0; | |
294 | else | |
295 | ctxt = ((qp->ibqp.qp_num >> dd->qos_shift) % | |
296 | (dd->n_krcv_queues - 1)) + 1; | |
297 | ||
298 | return dd->rcd[ctxt]; | |
299 | } | |
300 | ||
301 | int hfi1_qp_priv_init(struct rvt_dev_info *rdi, struct rvt_qp *qp, | |
302 | struct ib_qp_init_attr *init_attr) | |
303 | { | |
304 | struct hfi1_qp_priv *qpriv = qp->priv; | |
838b6fd2 | 305 | int i, ret; |
5190f052 MM |
306 | |
307 | qpriv->rcd = qp_to_rcd(rdi, qp); | |
308 | ||
48a615dc KW |
309 | spin_lock_init(&qpriv->opfn.lock); |
310 | INIT_WORK(&qpriv->opfn.opfn_work, opfn_send_conn_request); | |
37356e78 KW |
311 | INIT_WORK(&qpriv->tid_rdma.trigger_work, tid_rdma_trigger_resume); |
312 | qpriv->flow_state.psn = 0; | |
313 | qpriv->flow_state.index = RXE_NUM_TID_FLOWS; | |
314 | qpriv->flow_state.last_index = RXE_NUM_TID_FLOWS; | |
315 | qpriv->flow_state.generation = KERN_GENERATION_RESERVED; | |
316 | INIT_LIST_HEAD(&qpriv->tid_wait); | |
48a615dc | 317 | |
838b6fd2 KW |
318 | if (init_attr->qp_type == IB_QPT_RC && HFI1_CAP_IS_KSET(TID_RDMA)) { |
319 | struct hfi1_devdata *dd = qpriv->rcd->dd; | |
320 | ||
321 | qpriv->pages = kzalloc_node(TID_RDMA_MAX_PAGES * | |
322 | sizeof(*qpriv->pages), | |
323 | GFP_KERNEL, dd->node); | |
324 | if (!qpriv->pages) | |
325 | return -ENOMEM; | |
326 | for (i = 0; i < qp->s_size; i++) { | |
327 | struct hfi1_swqe_priv *priv; | |
328 | struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, i); | |
329 | ||
330 | priv = kzalloc_node(sizeof(*priv), GFP_KERNEL, | |
331 | dd->node); | |
332 | if (!priv) | |
333 | return -ENOMEM; | |
334 | ||
335 | hfi1_init_trdma_req(qp, &priv->tid_req); | |
336 | priv->tid_req.e.swqe = wqe; | |
337 | wqe->priv = priv; | |
338 | } | |
339 | for (i = 0; i < rvt_max_atomic(rdi); i++) { | |
340 | struct hfi1_ack_priv *priv; | |
341 | ||
342 | priv = kzalloc_node(sizeof(*priv), GFP_KERNEL, | |
343 | dd->node); | |
344 | if (!priv) | |
345 | return -ENOMEM; | |
346 | ||
347 | hfi1_init_trdma_req(qp, &priv->tid_req); | |
348 | priv->tid_req.e.ack = &qp->s_ack_queue[i]; | |
349 | ||
350 | ret = hfi1_kern_exp_rcv_alloc_flows(&priv->tid_req, | |
351 | GFP_KERNEL); | |
352 | if (ret) { | |
353 | kfree(priv); | |
354 | return ret; | |
355 | } | |
356 | qp->s_ack_queue[i].priv = priv; | |
357 | } | |
358 | } | |
359 | ||
5190f052 MM |
360 | return 0; |
361 | } | |
48a615dc KW |
362 | |
363 | void hfi1_qp_priv_tid_free(struct rvt_dev_info *rdi, struct rvt_qp *qp) | |
364 | { | |
838b6fd2 KW |
365 | struct hfi1_qp_priv *qpriv = qp->priv; |
366 | struct rvt_swqe *wqe; | |
367 | u32 i; | |
368 | ||
369 | if (qp->ibqp.qp_type == IB_QPT_RC && HFI1_CAP_IS_KSET(TID_RDMA)) { | |
370 | for (i = 0; i < qp->s_size; i++) { | |
371 | wqe = rvt_get_swqe_ptr(qp, i); | |
372 | kfree(wqe->priv); | |
373 | wqe->priv = NULL; | |
374 | } | |
375 | for (i = 0; i < rvt_max_atomic(rdi); i++) { | |
376 | struct hfi1_ack_priv *priv = qp->s_ack_queue[i].priv; | |
377 | ||
378 | if (priv) | |
379 | hfi1_kern_exp_rcv_free_flows(&priv->tid_req); | |
380 | kfree(priv); | |
381 | qp->s_ack_queue[i].priv = NULL; | |
382 | } | |
383 | cancel_work_sync(&qpriv->opfn.opfn_work); | |
384 | kfree(qpriv->pages); | |
385 | qpriv->pages = NULL; | |
386 | } | |
48a615dc | 387 | } |
37356e78 KW |
388 | |
389 | /* Flow and tid waiter functions */ | |
390 | /** | |
391 | * DOC: lock ordering | |
392 | * | |
393 | * There are two locks involved with the queuing | |
394 | * routines: the qp s_lock and the exp_lock. | |
395 | * | |
396 | * Since the tid space allocation is called from | |
397 | * the send engine, the qp s_lock is already held. | |
398 | * | |
399 | * The allocation routines will get the exp_lock. | |
400 | * | |
401 | * The first_qp() call is provided to allow the head of | |
402 | * the rcd wait queue to be fetched under the exp_lock and | |
403 | * followed by a drop of the exp_lock. | |
404 | * | |
405 | * Any qp in the wait list will have the qp reference count held | |
406 | * to hold the qp in memory. | |
407 | */ | |
408 | ||
409 | /* | |
410 | * return head of rcd wait list | |
411 | * | |
412 | * Must hold the exp_lock. | |
413 | * | |
414 | * Get a reference to the QP to hold the QP in memory. | |
415 | * | |
416 | * The caller must release the reference when the local | |
417 | * is no longer being used. | |
418 | */ | |
419 | static struct rvt_qp *first_qp(struct hfi1_ctxtdata *rcd, | |
420 | struct tid_queue *queue) | |
421 | __must_hold(&rcd->exp_lock) | |
422 | { | |
423 | struct hfi1_qp_priv *priv; | |
424 | ||
425 | lockdep_assert_held(&rcd->exp_lock); | |
426 | priv = list_first_entry_or_null(&queue->queue_head, | |
427 | struct hfi1_qp_priv, | |
428 | tid_wait); | |
429 | if (!priv) | |
430 | return NULL; | |
431 | rvt_get_qp(priv->owner); | |
432 | return priv->owner; | |
433 | } | |
434 | ||
435 | /** | |
436 | * kernel_tid_waiters - determine rcd wait | |
437 | * @rcd: the receive context | |
438 | * @qp: the head of the qp being processed | |
439 | * | |
440 | * This routine will return false IFF | |
441 | * the list is NULL or the head of the | |
442 | * list is the indicated qp. | |
443 | * | |
444 | * Must hold the qp s_lock and the exp_lock. | |
445 | * | |
446 | * Return: | |
447 | * false if either of the conditions below are statisfied: | |
448 | * 1. The list is empty or | |
449 | * 2. The indicated qp is at the head of the list and the | |
450 | * HFI1_S_WAIT_TID_SPACE bit is set in qp->s_flags. | |
451 | * true is returned otherwise. | |
452 | */ | |
453 | static bool kernel_tid_waiters(struct hfi1_ctxtdata *rcd, | |
454 | struct tid_queue *queue, struct rvt_qp *qp) | |
455 | __must_hold(&rcd->exp_lock) __must_hold(&qp->s_lock) | |
456 | { | |
457 | struct rvt_qp *fqp; | |
458 | bool ret = true; | |
459 | ||
460 | lockdep_assert_held(&qp->s_lock); | |
461 | lockdep_assert_held(&rcd->exp_lock); | |
462 | fqp = first_qp(rcd, queue); | |
463 | if (!fqp || (fqp == qp && (qp->s_flags & HFI1_S_WAIT_TID_SPACE))) | |
464 | ret = false; | |
465 | rvt_put_qp(fqp); | |
466 | return ret; | |
467 | } | |
468 | ||
469 | /** | |
470 | * dequeue_tid_waiter - dequeue the qp from the list | |
471 | * @qp - the qp to remove the wait list | |
472 | * | |
473 | * This routine removes the indicated qp from the | |
474 | * wait list if it is there. | |
475 | * | |
476 | * This should be done after the hardware flow and | |
477 | * tid array resources have been allocated. | |
478 | * | |
479 | * Must hold the qp s_lock and the rcd exp_lock. | |
480 | * | |
481 | * It assumes the s_lock to protect the s_flags | |
482 | * field and to reliably test the HFI1_S_WAIT_TID_SPACE flag. | |
483 | */ | |
484 | static void dequeue_tid_waiter(struct hfi1_ctxtdata *rcd, | |
485 | struct tid_queue *queue, struct rvt_qp *qp) | |
486 | __must_hold(&rcd->exp_lock) __must_hold(&qp->s_lock) | |
487 | { | |
488 | struct hfi1_qp_priv *priv = qp->priv; | |
489 | ||
490 | lockdep_assert_held(&qp->s_lock); | |
491 | lockdep_assert_held(&rcd->exp_lock); | |
492 | if (list_empty(&priv->tid_wait)) | |
493 | return; | |
494 | list_del_init(&priv->tid_wait); | |
495 | qp->s_flags &= ~HFI1_S_WAIT_TID_SPACE; | |
496 | queue->dequeue++; | |
497 | rvt_put_qp(qp); | |
498 | } | |
499 | ||
500 | /** | |
501 | * queue_qp_for_tid_wait - suspend QP on tid space | |
502 | * @rcd: the receive context | |
503 | * @qp: the qp | |
504 | * | |
505 | * The qp is inserted at the tail of the rcd | |
506 | * wait queue and the HFI1_S_WAIT_TID_SPACE s_flag is set. | |
507 | * | |
508 | * Must hold the qp s_lock and the exp_lock. | |
509 | */ | |
510 | static void queue_qp_for_tid_wait(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 | qp->s_flags |= HFI1_S_WAIT_TID_SPACE; | |
520 | list_add_tail(&priv->tid_wait, &queue->queue_head); | |
521 | priv->tid_enqueue = ++queue->enqueue; | |
2f16a696 | 522 | rcd->dd->verbs_dev.n_tidwait++; |
37356e78 KW |
523 | trace_hfi1_qpsleep(qp, HFI1_S_WAIT_TID_SPACE); |
524 | rvt_get_qp(qp); | |
525 | } | |
526 | } | |
527 | ||
528 | /** | |
529 | * __trigger_tid_waiter - trigger tid waiter | |
530 | * @qp: the qp | |
531 | * | |
532 | * This is a private entrance to schedule the qp | |
533 | * assuming the caller is holding the qp->s_lock. | |
534 | */ | |
535 | static void __trigger_tid_waiter(struct rvt_qp *qp) | |
536 | __must_hold(&qp->s_lock) | |
537 | { | |
538 | lockdep_assert_held(&qp->s_lock); | |
539 | if (!(qp->s_flags & HFI1_S_WAIT_TID_SPACE)) | |
540 | return; | |
541 | trace_hfi1_qpwakeup(qp, HFI1_S_WAIT_TID_SPACE); | |
542 | hfi1_schedule_send(qp); | |
543 | } | |
544 | ||
545 | /** | |
546 | * tid_rdma_schedule_tid_wakeup - schedule wakeup for a qp | |
547 | * @qp - the qp | |
548 | * | |
549 | * trigger a schedule or a waiting qp in a deadlock | |
550 | * safe manner. The qp reference is held prior | |
551 | * to this call via first_qp(). | |
552 | * | |
553 | * If the qp trigger was already scheduled (!rval) | |
554 | * the the reference is dropped, otherwise the resume | |
555 | * or the destroy cancel will dispatch the reference. | |
556 | */ | |
557 | static void tid_rdma_schedule_tid_wakeup(struct rvt_qp *qp) | |
558 | { | |
559 | struct hfi1_qp_priv *priv; | |
560 | struct hfi1_ibport *ibp; | |
561 | struct hfi1_pportdata *ppd; | |
562 | struct hfi1_devdata *dd; | |
563 | bool rval; | |
564 | ||
565 | if (!qp) | |
566 | return; | |
567 | ||
568 | priv = qp->priv; | |
569 | ibp = to_iport(qp->ibqp.device, qp->port_num); | |
570 | ppd = ppd_from_ibp(ibp); | |
571 | dd = dd_from_ibdev(qp->ibqp.device); | |
572 | ||
573 | rval = queue_work_on(priv->s_sde ? | |
574 | priv->s_sde->cpu : | |
575 | cpumask_first(cpumask_of_node(dd->node)), | |
576 | ppd->hfi1_wq, | |
577 | &priv->tid_rdma.trigger_work); | |
578 | if (!rval) | |
579 | rvt_put_qp(qp); | |
580 | } | |
581 | ||
582 | /** | |
583 | * tid_rdma_trigger_resume - field a trigger work request | |
584 | * @work - the work item | |
585 | * | |
586 | * Complete the off qp trigger processing by directly | |
587 | * calling the progress routine. | |
588 | */ | |
589 | static void tid_rdma_trigger_resume(struct work_struct *work) | |
590 | { | |
591 | struct tid_rdma_qp_params *tr; | |
592 | struct hfi1_qp_priv *priv; | |
593 | struct rvt_qp *qp; | |
594 | ||
595 | tr = container_of(work, struct tid_rdma_qp_params, trigger_work); | |
596 | priv = container_of(tr, struct hfi1_qp_priv, tid_rdma); | |
597 | qp = priv->owner; | |
598 | spin_lock_irq(&qp->s_lock); | |
599 | if (qp->s_flags & HFI1_S_WAIT_TID_SPACE) { | |
600 | spin_unlock_irq(&qp->s_lock); | |
601 | hfi1_do_send(priv->owner, true); | |
602 | } else { | |
603 | spin_unlock_irq(&qp->s_lock); | |
604 | } | |
605 | rvt_put_qp(qp); | |
606 | } | |
607 | ||
608 | /** | |
609 | * tid_rdma_flush_wait - unwind any tid space wait | |
610 | * | |
611 | * This is called when resetting a qp to | |
612 | * allow a destroy or reset to get rid | |
613 | * of any tid space linkage and reference counts. | |
614 | */ | |
615 | static void _tid_rdma_flush_wait(struct rvt_qp *qp, struct tid_queue *queue) | |
616 | __must_hold(&qp->s_lock) | |
617 | { | |
618 | struct hfi1_qp_priv *priv; | |
619 | ||
620 | if (!qp) | |
621 | return; | |
622 | lockdep_assert_held(&qp->s_lock); | |
623 | priv = qp->priv; | |
624 | qp->s_flags &= ~HFI1_S_WAIT_TID_SPACE; | |
625 | spin_lock(&priv->rcd->exp_lock); | |
626 | if (!list_empty(&priv->tid_wait)) { | |
627 | list_del_init(&priv->tid_wait); | |
628 | qp->s_flags &= ~HFI1_S_WAIT_TID_SPACE; | |
629 | queue->dequeue++; | |
630 | rvt_put_qp(qp); | |
631 | } | |
632 | spin_unlock(&priv->rcd->exp_lock); | |
633 | } | |
634 | ||
635 | void hfi1_tid_rdma_flush_wait(struct rvt_qp *qp) | |
636 | __must_hold(&qp->s_lock) | |
637 | { | |
638 | struct hfi1_qp_priv *priv = qp->priv; | |
639 | ||
640 | _tid_rdma_flush_wait(qp, &priv->rcd->flow_queue); | |
838b6fd2 | 641 | _tid_rdma_flush_wait(qp, &priv->rcd->rarr_queue); |
37356e78 KW |
642 | } |
643 | ||
644 | /* Flow functions */ | |
645 | /** | |
646 | * kern_reserve_flow - allocate a hardware flow | |
647 | * @rcd - the context to use for allocation | |
648 | * @last - the index of the preferred flow. Use RXE_NUM_TID_FLOWS to | |
649 | * signify "don't care". | |
650 | * | |
651 | * Use a bit mask based allocation to reserve a hardware | |
652 | * flow for use in receiving KDETH data packets. If a preferred flow is | |
653 | * specified the function will attempt to reserve that flow again, if | |
654 | * available. | |
655 | * | |
656 | * The exp_lock must be held. | |
657 | * | |
658 | * Return: | |
659 | * On success: a value postive value between 0 and RXE_NUM_TID_FLOWS - 1 | |
660 | * On failure: -EAGAIN | |
661 | */ | |
662 | static int kern_reserve_flow(struct hfi1_ctxtdata *rcd, int last) | |
663 | __must_hold(&rcd->exp_lock) | |
664 | { | |
665 | int nr; | |
666 | ||
667 | /* Attempt to reserve the preferred flow index */ | |
668 | if (last >= 0 && last < RXE_NUM_TID_FLOWS && | |
669 | !test_and_set_bit(last, &rcd->flow_mask)) | |
670 | return last; | |
671 | ||
672 | nr = ffz(rcd->flow_mask); | |
673 | BUILD_BUG_ON(RXE_NUM_TID_FLOWS >= | |
674 | (sizeof(rcd->flow_mask) * BITS_PER_BYTE)); | |
675 | if (nr > (RXE_NUM_TID_FLOWS - 1)) | |
676 | return -EAGAIN; | |
677 | set_bit(nr, &rcd->flow_mask); | |
678 | return nr; | |
679 | } | |
680 | ||
681 | static void kern_set_hw_flow(struct hfi1_ctxtdata *rcd, u32 generation, | |
682 | u32 flow_idx) | |
683 | { | |
684 | u64 reg; | |
685 | ||
686 | reg = ((u64)generation << HFI1_KDETH_BTH_SEQ_SHIFT) | | |
687 | RCV_TID_FLOW_TABLE_CTRL_FLOW_VALID_SMASK | | |
688 | RCV_TID_FLOW_TABLE_CTRL_KEEP_AFTER_SEQ_ERR_SMASK | | |
689 | RCV_TID_FLOW_TABLE_CTRL_KEEP_ON_GEN_ERR_SMASK | | |
690 | RCV_TID_FLOW_TABLE_STATUS_SEQ_MISMATCH_SMASK | | |
691 | RCV_TID_FLOW_TABLE_STATUS_GEN_MISMATCH_SMASK; | |
692 | ||
693 | if (generation != KERN_GENERATION_RESERVED) | |
694 | reg |= RCV_TID_FLOW_TABLE_CTRL_HDR_SUPP_EN_SMASK; | |
695 | ||
696 | write_uctxt_csr(rcd->dd, rcd->ctxt, | |
697 | RCV_TID_FLOW_TABLE + 8 * flow_idx, reg); | |
698 | } | |
699 | ||
700 | static u32 kern_setup_hw_flow(struct hfi1_ctxtdata *rcd, u32 flow_idx) | |
701 | __must_hold(&rcd->exp_lock) | |
702 | { | |
703 | u32 generation = rcd->flows[flow_idx].generation; | |
704 | ||
705 | kern_set_hw_flow(rcd, generation, flow_idx); | |
706 | return generation; | |
707 | } | |
708 | ||
709 | static u32 kern_flow_generation_next(u32 gen) | |
710 | { | |
711 | u32 generation = mask_generation(gen + 1); | |
712 | ||
713 | if (generation == KERN_GENERATION_RESERVED) | |
714 | generation = mask_generation(generation + 1); | |
715 | return generation; | |
716 | } | |
717 | ||
718 | static void kern_clear_hw_flow(struct hfi1_ctxtdata *rcd, u32 flow_idx) | |
719 | __must_hold(&rcd->exp_lock) | |
720 | { | |
721 | rcd->flows[flow_idx].generation = | |
722 | kern_flow_generation_next(rcd->flows[flow_idx].generation); | |
723 | kern_set_hw_flow(rcd, KERN_GENERATION_RESERVED, flow_idx); | |
724 | } | |
725 | ||
726 | int hfi1_kern_setup_hw_flow(struct hfi1_ctxtdata *rcd, struct rvt_qp *qp) | |
727 | { | |
728 | struct hfi1_qp_priv *qpriv = (struct hfi1_qp_priv *)qp->priv; | |
729 | struct tid_flow_state *fs = &qpriv->flow_state; | |
730 | struct rvt_qp *fqp; | |
731 | unsigned long flags; | |
732 | int ret = 0; | |
733 | ||
734 | /* The QP already has an allocated flow */ | |
735 | if (fs->index != RXE_NUM_TID_FLOWS) | |
736 | return ret; | |
737 | ||
738 | spin_lock_irqsave(&rcd->exp_lock, flags); | |
739 | if (kernel_tid_waiters(rcd, &rcd->flow_queue, qp)) | |
740 | goto queue; | |
741 | ||
742 | ret = kern_reserve_flow(rcd, fs->last_index); | |
743 | if (ret < 0) | |
744 | goto queue; | |
745 | fs->index = ret; | |
746 | fs->last_index = fs->index; | |
747 | ||
748 | /* Generation received in a RESYNC overrides default flow generation */ | |
749 | if (fs->generation != KERN_GENERATION_RESERVED) | |
750 | rcd->flows[fs->index].generation = fs->generation; | |
751 | fs->generation = kern_setup_hw_flow(rcd, fs->index); | |
752 | fs->psn = 0; | |
753 | fs->flags = 0; | |
754 | dequeue_tid_waiter(rcd, &rcd->flow_queue, qp); | |
755 | /* get head before dropping lock */ | |
756 | fqp = first_qp(rcd, &rcd->flow_queue); | |
757 | spin_unlock_irqrestore(&rcd->exp_lock, flags); | |
758 | ||
759 | tid_rdma_schedule_tid_wakeup(fqp); | |
760 | return 0; | |
761 | queue: | |
762 | queue_qp_for_tid_wait(rcd, &rcd->flow_queue, qp); | |
763 | spin_unlock_irqrestore(&rcd->exp_lock, flags); | |
764 | return -EAGAIN; | |
765 | } | |
766 | ||
767 | void hfi1_kern_clear_hw_flow(struct hfi1_ctxtdata *rcd, struct rvt_qp *qp) | |
768 | { | |
769 | struct hfi1_qp_priv *qpriv = (struct hfi1_qp_priv *)qp->priv; | |
770 | struct tid_flow_state *fs = &qpriv->flow_state; | |
771 | struct rvt_qp *fqp; | |
772 | unsigned long flags; | |
773 | ||
774 | if (fs->index >= RXE_NUM_TID_FLOWS) | |
775 | return; | |
776 | spin_lock_irqsave(&rcd->exp_lock, flags); | |
777 | kern_clear_hw_flow(rcd, fs->index); | |
778 | clear_bit(fs->index, &rcd->flow_mask); | |
779 | fs->index = RXE_NUM_TID_FLOWS; | |
780 | fs->psn = 0; | |
781 | fs->generation = KERN_GENERATION_RESERVED; | |
782 | ||
783 | /* get head before dropping lock */ | |
784 | fqp = first_qp(rcd, &rcd->flow_queue); | |
785 | spin_unlock_irqrestore(&rcd->exp_lock, flags); | |
786 | ||
787 | if (fqp == qp) { | |
788 | __trigger_tid_waiter(fqp); | |
789 | rvt_put_qp(fqp); | |
790 | } else { | |
791 | tid_rdma_schedule_tid_wakeup(fqp); | |
792 | } | |
793 | } | |
794 | ||
795 | void hfi1_kern_init_ctxt_generations(struct hfi1_ctxtdata *rcd) | |
796 | { | |
797 | int i; | |
798 | ||
799 | for (i = 0; i < RXE_NUM_TID_FLOWS; i++) { | |
800 | rcd->flows[i].generation = mask_generation(prandom_u32()); | |
801 | kern_set_hw_flow(rcd, KERN_GENERATION_RESERVED, i); | |
802 | } | |
803 | } | |
838b6fd2 KW |
804 | |
805 | /* TID allocation functions */ | |
806 | static u8 trdma_pset_order(struct tid_rdma_pageset *s) | |
807 | { | |
808 | u8 count = s->count; | |
809 | ||
810 | return ilog2(count) + 1; | |
811 | } | |
812 | ||
813 | /** | |
814 | * tid_rdma_find_phys_blocks_4k - get groups base on mr info | |
815 | * @npages - number of pages | |
816 | * @pages - pointer to an array of page structs | |
817 | * @list - page set array to return | |
818 | * | |
819 | * This routine returns the number of groups associated with | |
820 | * the current sge information. This implementation is based | |
821 | * on the expected receive find_phys_blocks() adjusted to | |
822 | * use the MR information vs. the pfn. | |
823 | * | |
824 | * Return: | |
825 | * the number of RcvArray entries | |
826 | */ | |
827 | static u32 tid_rdma_find_phys_blocks_4k(struct tid_rdma_flow *flow, | |
828 | struct page **pages, | |
829 | u32 npages, | |
830 | struct tid_rdma_pageset *list) | |
831 | { | |
832 | u32 pagecount, pageidx, setcount = 0, i; | |
833 | void *vaddr, *this_vaddr; | |
834 | ||
835 | if (!npages) | |
836 | return 0; | |
837 | ||
838 | /* | |
839 | * Look for sets of physically contiguous pages in the user buffer. | |
840 | * This will allow us to optimize Expected RcvArray entry usage by | |
841 | * using the bigger supported sizes. | |
842 | */ | |
843 | vaddr = page_address(pages[0]); | |
84f4a40d | 844 | trace_hfi1_tid_flow_page(flow->req->qp, flow, 0, 0, 0, vaddr); |
838b6fd2 KW |
845 | for (pageidx = 0, pagecount = 1, i = 1; i <= npages; i++) { |
846 | this_vaddr = i < npages ? page_address(pages[i]) : NULL; | |
84f4a40d KW |
847 | trace_hfi1_tid_flow_page(flow->req->qp, flow, i, 0, 0, |
848 | this_vaddr); | |
838b6fd2 KW |
849 | /* |
850 | * If the vaddr's are not sequential, pages are not physically | |
851 | * contiguous. | |
852 | */ | |
853 | if (this_vaddr != (vaddr + PAGE_SIZE)) { | |
854 | /* | |
855 | * At this point we have to loop over the set of | |
856 | * physically contiguous pages and break them down it | |
857 | * sizes supported by the HW. | |
858 | * There are two main constraints: | |
859 | * 1. The max buffer size is MAX_EXPECTED_BUFFER. | |
860 | * If the total set size is bigger than that | |
861 | * program only a MAX_EXPECTED_BUFFER chunk. | |
862 | * 2. The buffer size has to be a power of two. If | |
863 | * it is not, round down to the closes power of | |
864 | * 2 and program that size. | |
865 | */ | |
866 | while (pagecount) { | |
867 | int maxpages = pagecount; | |
868 | u32 bufsize = pagecount * PAGE_SIZE; | |
869 | ||
870 | if (bufsize > MAX_EXPECTED_BUFFER) | |
871 | maxpages = | |
872 | MAX_EXPECTED_BUFFER >> | |
873 | PAGE_SHIFT; | |
874 | else if (!is_power_of_2(bufsize)) | |
875 | maxpages = | |
876 | rounddown_pow_of_two(bufsize) >> | |
877 | PAGE_SHIFT; | |
878 | ||
879 | list[setcount].idx = pageidx; | |
880 | list[setcount].count = maxpages; | |
84f4a40d KW |
881 | trace_hfi1_tid_pageset(flow->req->qp, setcount, |
882 | list[setcount].idx, | |
883 | list[setcount].count); | |
838b6fd2 KW |
884 | pagecount -= maxpages; |
885 | pageidx += maxpages; | |
886 | setcount++; | |
887 | } | |
888 | pageidx = i; | |
889 | pagecount = 1; | |
890 | vaddr = this_vaddr; | |
891 | } else { | |
892 | vaddr += PAGE_SIZE; | |
893 | pagecount++; | |
894 | } | |
895 | } | |
896 | /* insure we always return an even number of sets */ | |
897 | if (setcount & 1) | |
898 | list[setcount++].count = 0; | |
899 | return setcount; | |
900 | } | |
901 | ||
902 | /** | |
903 | * tid_flush_pages - dump out pages into pagesets | |
904 | * @list - list of pagesets | |
905 | * @idx - pointer to current page index | |
906 | * @pages - number of pages to dump | |
907 | * @sets - current number of pagesset | |
908 | * | |
909 | * This routine flushes out accumuated pages. | |
910 | * | |
911 | * To insure an even number of sets the | |
912 | * code may add a filler. | |
913 | * | |
914 | * This can happen with when pages is not | |
915 | * a power of 2 or pages is a power of 2 | |
916 | * less than the maximum pages. | |
917 | * | |
918 | * Return: | |
919 | * The new number of sets | |
920 | */ | |
921 | ||
922 | static u32 tid_flush_pages(struct tid_rdma_pageset *list, | |
923 | u32 *idx, u32 pages, u32 sets) | |
924 | { | |
925 | while (pages) { | |
926 | u32 maxpages = pages; | |
927 | ||
928 | if (maxpages > MAX_EXPECTED_PAGES) | |
929 | maxpages = MAX_EXPECTED_PAGES; | |
930 | else if (!is_power_of_2(maxpages)) | |
931 | maxpages = rounddown_pow_of_two(maxpages); | |
932 | list[sets].idx = *idx; | |
933 | list[sets++].count = maxpages; | |
934 | *idx += maxpages; | |
935 | pages -= maxpages; | |
936 | } | |
937 | /* might need a filler */ | |
938 | if (sets & 1) | |
939 | list[sets++].count = 0; | |
940 | return sets; | |
941 | } | |
942 | ||
943 | /** | |
944 | * tid_rdma_find_phys_blocks_8k - get groups base on mr info | |
945 | * @pages - pointer to an array of page structs | |
946 | * @npages - number of pages | |
947 | * @list - page set array to return | |
948 | * | |
949 | * This routine parses an array of pages to compute pagesets | |
950 | * in an 8k compatible way. | |
951 | * | |
952 | * pages are tested two at a time, i, i + 1 for contiguous | |
953 | * pages and i - 1 and i contiguous pages. | |
954 | * | |
955 | * If any condition is false, any accumlated pages are flushed and | |
956 | * v0,v1 are emitted as separate PAGE_SIZE pagesets | |
957 | * | |
958 | * Otherwise, the current 8k is totaled for a future flush. | |
959 | * | |
960 | * Return: | |
961 | * The number of pagesets | |
962 | * list set with the returned number of pagesets | |
963 | * | |
964 | */ | |
965 | static u32 tid_rdma_find_phys_blocks_8k(struct tid_rdma_flow *flow, | |
966 | struct page **pages, | |
967 | u32 npages, | |
968 | struct tid_rdma_pageset *list) | |
969 | { | |
970 | u32 idx, sets = 0, i; | |
971 | u32 pagecnt = 0; | |
972 | void *v0, *v1, *vm1; | |
973 | ||
974 | if (!npages) | |
975 | return 0; | |
976 | for (idx = 0, i = 0, vm1 = NULL; i < npages; i += 2) { | |
977 | /* get a new v0 */ | |
978 | v0 = page_address(pages[i]); | |
84f4a40d | 979 | trace_hfi1_tid_flow_page(flow->req->qp, flow, i, 1, 0, v0); |
838b6fd2 KW |
980 | v1 = i + 1 < npages ? |
981 | page_address(pages[i + 1]) : NULL; | |
84f4a40d | 982 | trace_hfi1_tid_flow_page(flow->req->qp, flow, i, 1, 1, v1); |
838b6fd2 KW |
983 | /* compare i, i + 1 vaddr */ |
984 | if (v1 != (v0 + PAGE_SIZE)) { | |
985 | /* flush out pages */ | |
986 | sets = tid_flush_pages(list, &idx, pagecnt, sets); | |
987 | /* output v0,v1 as two pagesets */ | |
988 | list[sets].idx = idx++; | |
989 | list[sets++].count = 1; | |
990 | if (v1) { | |
991 | list[sets].count = 1; | |
992 | list[sets++].idx = idx++; | |
993 | } else { | |
994 | list[sets++].count = 0; | |
995 | } | |
996 | vm1 = NULL; | |
997 | pagecnt = 0; | |
998 | continue; | |
999 | } | |
1000 | /* i,i+1 consecutive, look at i-1,i */ | |
1001 | if (vm1 && v0 != (vm1 + PAGE_SIZE)) { | |
1002 | /* flush out pages */ | |
1003 | sets = tid_flush_pages(list, &idx, pagecnt, sets); | |
1004 | pagecnt = 0; | |
1005 | } | |
1006 | /* pages will always be a multiple of 8k */ | |
1007 | pagecnt += 2; | |
1008 | /* save i-1 */ | |
1009 | vm1 = v1; | |
1010 | /* move to next pair */ | |
1011 | } | |
1012 | /* dump residual pages at end */ | |
1013 | sets = tid_flush_pages(list, &idx, npages - idx, sets); | |
1014 | /* by design cannot be odd sets */ | |
1015 | WARN_ON(sets & 1); | |
1016 | return sets; | |
1017 | } | |
1018 | ||
1019 | /** | |
1020 | * Find pages for one segment of a sge array represented by @ss. The function | |
1021 | * does not check the sge, the sge must have been checked for alignment with a | |
1022 | * prior call to hfi1_kern_trdma_ok. Other sge checking is done as part of | |
1023 | * rvt_lkey_ok and rvt_rkey_ok. Also, the function only modifies the local sge | |
1024 | * copy maintained in @ss->sge, the original sge is not modified. | |
1025 | * | |
1026 | * Unlike IB RDMA WRITE, we can't decrement ss->num_sge here because we are not | |
1027 | * releasing the MR reference count at the same time. Otherwise, we'll "leak" | |
1028 | * references to the MR. This difference requires that we keep track of progress | |
1029 | * into the sg_list. This is done by the cur_seg cursor in the tid_rdma_request | |
1030 | * structure. | |
1031 | */ | |
1032 | static u32 kern_find_pages(struct tid_rdma_flow *flow, | |
1033 | struct page **pages, | |
1034 | struct rvt_sge_state *ss, bool *last) | |
1035 | { | |
1036 | struct tid_rdma_request *req = flow->req; | |
1037 | struct rvt_sge *sge = &ss->sge; | |
1038 | u32 length = flow->req->seg_len; | |
1039 | u32 len = PAGE_SIZE; | |
1040 | u32 i = 0; | |
1041 | ||
1042 | while (length && req->isge < ss->num_sge) { | |
1043 | pages[i++] = virt_to_page(sge->vaddr); | |
1044 | ||
1045 | sge->vaddr += len; | |
1046 | sge->length -= len; | |
1047 | sge->sge_length -= len; | |
1048 | if (!sge->sge_length) { | |
1049 | if (++req->isge < ss->num_sge) | |
1050 | *sge = ss->sg_list[req->isge - 1]; | |
1051 | } else if (sge->length == 0 && sge->mr->lkey) { | |
1052 | if (++sge->n >= RVT_SEGSZ) { | |
1053 | ++sge->m; | |
1054 | sge->n = 0; | |
1055 | } | |
1056 | sge->vaddr = sge->mr->map[sge->m]->segs[sge->n].vaddr; | |
1057 | sge->length = sge->mr->map[sge->m]->segs[sge->n].length; | |
1058 | } | |
1059 | length -= len; | |
1060 | } | |
1061 | ||
1062 | flow->length = flow->req->seg_len - length; | |
1063 | *last = req->isge == ss->num_sge ? false : true; | |
1064 | return i; | |
1065 | } | |
1066 | ||
1067 | static void dma_unmap_flow(struct tid_rdma_flow *flow) | |
1068 | { | |
1069 | struct hfi1_devdata *dd; | |
1070 | int i; | |
1071 | struct tid_rdma_pageset *pset; | |
1072 | ||
1073 | dd = flow->req->rcd->dd; | |
1074 | for (i = 0, pset = &flow->pagesets[0]; i < flow->npagesets; | |
1075 | i++, pset++) { | |
1076 | if (pset->count && pset->addr) { | |
1077 | dma_unmap_page(&dd->pcidev->dev, | |
1078 | pset->addr, | |
1079 | PAGE_SIZE * pset->count, | |
1080 | DMA_FROM_DEVICE); | |
1081 | pset->mapped = 0; | |
1082 | } | |
1083 | } | |
1084 | } | |
1085 | ||
1086 | static int dma_map_flow(struct tid_rdma_flow *flow, struct page **pages) | |
1087 | { | |
1088 | int i; | |
1089 | struct hfi1_devdata *dd = flow->req->rcd->dd; | |
1090 | struct tid_rdma_pageset *pset; | |
1091 | ||
1092 | for (i = 0, pset = &flow->pagesets[0]; i < flow->npagesets; | |
1093 | i++, pset++) { | |
1094 | if (pset->count) { | |
1095 | pset->addr = dma_map_page(&dd->pcidev->dev, | |
1096 | pages[pset->idx], | |
1097 | 0, | |
1098 | PAGE_SIZE * pset->count, | |
1099 | DMA_FROM_DEVICE); | |
1100 | ||
1101 | if (dma_mapping_error(&dd->pcidev->dev, pset->addr)) { | |
1102 | dma_unmap_flow(flow); | |
1103 | return -ENOMEM; | |
1104 | } | |
1105 | pset->mapped = 1; | |
1106 | } | |
1107 | } | |
1108 | return 0; | |
1109 | } | |
1110 | ||
1111 | static inline bool dma_mapped(struct tid_rdma_flow *flow) | |
1112 | { | |
1113 | return !!flow->pagesets[0].mapped; | |
1114 | } | |
1115 | ||
1116 | /* | |
1117 | * Get pages pointers and identify contiguous physical memory chunks for a | |
1118 | * segment. All segments are of length flow->req->seg_len. | |
1119 | */ | |
1120 | static int kern_get_phys_blocks(struct tid_rdma_flow *flow, | |
1121 | struct page **pages, | |
1122 | struct rvt_sge_state *ss, bool *last) | |
1123 | { | |
1124 | u8 npages; | |
1125 | ||
1126 | /* Reuse previously computed pagesets, if any */ | |
1127 | if (flow->npagesets) { | |
84f4a40d KW |
1128 | trace_hfi1_tid_flow_alloc(flow->req->qp, flow->req->setup_head, |
1129 | flow); | |
838b6fd2 KW |
1130 | if (!dma_mapped(flow)) |
1131 | return dma_map_flow(flow, pages); | |
1132 | return 0; | |
1133 | } | |
1134 | ||
1135 | npages = kern_find_pages(flow, pages, ss, last); | |
1136 | ||
1137 | if (flow->req->qp->pmtu == enum_to_mtu(OPA_MTU_4096)) | |
1138 | flow->npagesets = | |
1139 | tid_rdma_find_phys_blocks_4k(flow, pages, npages, | |
1140 | flow->pagesets); | |
1141 | else | |
1142 | flow->npagesets = | |
1143 | tid_rdma_find_phys_blocks_8k(flow, pages, npages, | |
1144 | flow->pagesets); | |
1145 | ||
1146 | return dma_map_flow(flow, pages); | |
1147 | } | |
1148 | ||
1149 | static inline void kern_add_tid_node(struct tid_rdma_flow *flow, | |
1150 | struct hfi1_ctxtdata *rcd, char *s, | |
1151 | struct tid_group *grp, u8 cnt) | |
1152 | { | |
1153 | struct kern_tid_node *node = &flow->tnode[flow->tnode_cnt++]; | |
1154 | ||
1155 | WARN_ON_ONCE(flow->tnode_cnt >= | |
1156 | (TID_RDMA_MAX_SEGMENT_SIZE >> PAGE_SHIFT)); | |
1157 | if (WARN_ON_ONCE(cnt & 1)) | |
1158 | dd_dev_err(rcd->dd, | |
1159 | "unexpected odd allocation cnt %u map 0x%x used %u", | |
1160 | cnt, grp->map, grp->used); | |
1161 | ||
1162 | node->grp = grp; | |
1163 | node->map = grp->map; | |
1164 | node->cnt = cnt; | |
84f4a40d KW |
1165 | trace_hfi1_tid_node_add(flow->req->qp, s, flow->tnode_cnt - 1, |
1166 | grp->base, grp->map, grp->used, cnt); | |
838b6fd2 KW |
1167 | } |
1168 | ||
1169 | /* | |
1170 | * Try to allocate pageset_count TID's from TID groups for a context | |
1171 | * | |
1172 | * This function allocates TID's without moving groups between lists or | |
1173 | * modifying grp->map. This is done as follows, being cogizant of the lists | |
1174 | * between which the TID groups will move: | |
1175 | * 1. First allocate complete groups of 8 TID's since this is more efficient, | |
1176 | * these groups will move from group->full without affecting used | |
1177 | * 2. If more TID's are needed allocate from used (will move from used->full or | |
1178 | * stay in used) | |
1179 | * 3. If we still don't have the required number of TID's go back and look again | |
1180 | * at a complete group (will move from group->used) | |
1181 | */ | |
1182 | static int kern_alloc_tids(struct tid_rdma_flow *flow) | |
1183 | { | |
1184 | struct hfi1_ctxtdata *rcd = flow->req->rcd; | |
1185 | struct hfi1_devdata *dd = rcd->dd; | |
1186 | u32 ngroups, pageidx = 0; | |
1187 | struct tid_group *group = NULL, *used; | |
1188 | u8 use; | |
1189 | ||
1190 | flow->tnode_cnt = 0; | |
1191 | ngroups = flow->npagesets / dd->rcv_entries.group_size; | |
1192 | if (!ngroups) | |
1193 | goto used_list; | |
1194 | ||
1195 | /* First look at complete groups */ | |
1196 | list_for_each_entry(group, &rcd->tid_group_list.list, list) { | |
1197 | kern_add_tid_node(flow, rcd, "complete groups", group, | |
1198 | group->size); | |
1199 | ||
1200 | pageidx += group->size; | |
1201 | if (!--ngroups) | |
1202 | break; | |
1203 | } | |
1204 | ||
1205 | if (pageidx >= flow->npagesets) | |
1206 | goto ok; | |
1207 | ||
1208 | used_list: | |
1209 | /* Now look at partially used groups */ | |
1210 | list_for_each_entry(used, &rcd->tid_used_list.list, list) { | |
1211 | use = min_t(u32, flow->npagesets - pageidx, | |
1212 | used->size - used->used); | |
1213 | kern_add_tid_node(flow, rcd, "used groups", used, use); | |
1214 | ||
1215 | pageidx += use; | |
1216 | if (pageidx >= flow->npagesets) | |
1217 | goto ok; | |
1218 | } | |
1219 | ||
1220 | /* | |
1221 | * Look again at a complete group, continuing from where we left. | |
1222 | * However, if we are at the head, we have reached the end of the | |
1223 | * complete groups list from the first loop above | |
1224 | */ | |
1225 | if (group && &group->list == &rcd->tid_group_list.list) | |
1226 | goto bail_eagain; | |
1227 | group = list_prepare_entry(group, &rcd->tid_group_list.list, | |
1228 | list); | |
1229 | if (list_is_last(&group->list, &rcd->tid_group_list.list)) | |
1230 | goto bail_eagain; | |
1231 | group = list_next_entry(group, list); | |
1232 | use = min_t(u32, flow->npagesets - pageidx, group->size); | |
1233 | kern_add_tid_node(flow, rcd, "complete continue", group, use); | |
1234 | pageidx += use; | |
1235 | if (pageidx >= flow->npagesets) | |
1236 | goto ok; | |
1237 | bail_eagain: | |
84f4a40d KW |
1238 | trace_hfi1_msg_alloc_tids(flow->req->qp, " insufficient tids: needed ", |
1239 | (u64)flow->npagesets); | |
838b6fd2 KW |
1240 | return -EAGAIN; |
1241 | ok: | |
1242 | return 0; | |
1243 | } | |
1244 | ||
1245 | static void kern_program_rcv_group(struct tid_rdma_flow *flow, int grp_num, | |
1246 | u32 *pset_idx) | |
1247 | { | |
1248 | struct hfi1_ctxtdata *rcd = flow->req->rcd; | |
1249 | struct hfi1_devdata *dd = rcd->dd; | |
1250 | struct kern_tid_node *node = &flow->tnode[grp_num]; | |
1251 | struct tid_group *grp = node->grp; | |
1252 | struct tid_rdma_pageset *pset; | |
1253 | u32 pmtu_pg = flow->req->qp->pmtu >> PAGE_SHIFT; | |
1254 | u32 rcventry, npages = 0, pair = 0, tidctrl; | |
1255 | u8 i, cnt = 0; | |
1256 | ||
1257 | for (i = 0; i < grp->size; i++) { | |
1258 | rcventry = grp->base + i; | |
1259 | ||
1260 | if (node->map & BIT(i) || cnt >= node->cnt) { | |
1261 | rcv_array_wc_fill(dd, rcventry); | |
1262 | continue; | |
1263 | } | |
1264 | pset = &flow->pagesets[(*pset_idx)++]; | |
1265 | if (pset->count) { | |
1266 | hfi1_put_tid(dd, rcventry, PT_EXPECTED, | |
1267 | pset->addr, trdma_pset_order(pset)); | |
1268 | } else { | |
1269 | hfi1_put_tid(dd, rcventry, PT_INVALID, 0, 0); | |
1270 | } | |
1271 | npages += pset->count; | |
1272 | ||
1273 | rcventry -= rcd->expected_base; | |
1274 | tidctrl = pair ? 0x3 : rcventry & 0x1 ? 0x2 : 0x1; | |
1275 | /* | |
1276 | * A single TID entry will be used to use a rcvarr pair (with | |
1277 | * tidctrl 0x3), if ALL these are true (a) the bit pos is even | |
1278 | * (b) the group map shows current and the next bits as free | |
1279 | * indicating two consecutive rcvarry entries are available (c) | |
1280 | * we actually need 2 more entries | |
1281 | */ | |
1282 | pair = !(i & 0x1) && !((node->map >> i) & 0x3) && | |
1283 | node->cnt >= cnt + 2; | |
1284 | if (!pair) { | |
1285 | if (!pset->count) | |
1286 | tidctrl = 0x1; | |
1287 | flow->tid_entry[flow->tidcnt++] = | |
1288 | EXP_TID_SET(IDX, rcventry >> 1) | | |
1289 | EXP_TID_SET(CTRL, tidctrl) | | |
1290 | EXP_TID_SET(LEN, npages); | |
84f4a40d KW |
1291 | trace_hfi1_tid_entry_alloc(/* entry */ |
1292 | flow->req->qp, flow->tidcnt - 1, | |
1293 | flow->tid_entry[flow->tidcnt - 1]); | |
1294 | ||
838b6fd2 KW |
1295 | /* Efficient DIV_ROUND_UP(npages, pmtu_pg) */ |
1296 | flow->npkts += (npages + pmtu_pg - 1) >> ilog2(pmtu_pg); | |
1297 | npages = 0; | |
1298 | } | |
1299 | ||
1300 | if (grp->used == grp->size - 1) | |
1301 | tid_group_move(grp, &rcd->tid_used_list, | |
1302 | &rcd->tid_full_list); | |
1303 | else if (!grp->used) | |
1304 | tid_group_move(grp, &rcd->tid_group_list, | |
1305 | &rcd->tid_used_list); | |
1306 | ||
1307 | grp->used++; | |
1308 | grp->map |= BIT(i); | |
1309 | cnt++; | |
1310 | } | |
1311 | } | |
1312 | ||
1313 | static void kern_unprogram_rcv_group(struct tid_rdma_flow *flow, int grp_num) | |
1314 | { | |
1315 | struct hfi1_ctxtdata *rcd = flow->req->rcd; | |
1316 | struct hfi1_devdata *dd = rcd->dd; | |
1317 | struct kern_tid_node *node = &flow->tnode[grp_num]; | |
1318 | struct tid_group *grp = node->grp; | |
1319 | u32 rcventry; | |
1320 | u8 i, cnt = 0; | |
1321 | ||
1322 | for (i = 0; i < grp->size; i++) { | |
1323 | rcventry = grp->base + i; | |
1324 | ||
1325 | if (node->map & BIT(i) || cnt >= node->cnt) { | |
1326 | rcv_array_wc_fill(dd, rcventry); | |
1327 | continue; | |
1328 | } | |
1329 | ||
1330 | hfi1_put_tid(dd, rcventry, PT_INVALID, 0, 0); | |
1331 | ||
1332 | grp->used--; | |
1333 | grp->map &= ~BIT(i); | |
1334 | cnt++; | |
1335 | ||
1336 | if (grp->used == grp->size - 1) | |
1337 | tid_group_move(grp, &rcd->tid_full_list, | |
1338 | &rcd->tid_used_list); | |
1339 | else if (!grp->used) | |
1340 | tid_group_move(grp, &rcd->tid_used_list, | |
1341 | &rcd->tid_group_list); | |
1342 | } | |
1343 | if (WARN_ON_ONCE(cnt & 1)) { | |
1344 | struct hfi1_ctxtdata *rcd = flow->req->rcd; | |
1345 | struct hfi1_devdata *dd = rcd->dd; | |
1346 | ||
1347 | dd_dev_err(dd, "unexpected odd free cnt %u map 0x%x used %u", | |
1348 | cnt, grp->map, grp->used); | |
1349 | } | |
1350 | } | |
1351 | ||
1352 | static void kern_program_rcvarray(struct tid_rdma_flow *flow) | |
1353 | { | |
1354 | u32 pset_idx = 0; | |
1355 | int i; | |
1356 | ||
1357 | flow->npkts = 0; | |
1358 | flow->tidcnt = 0; | |
1359 | for (i = 0; i < flow->tnode_cnt; i++) | |
1360 | kern_program_rcv_group(flow, i, &pset_idx); | |
84f4a40d | 1361 | trace_hfi1_tid_flow_alloc(flow->req->qp, flow->req->setup_head, flow); |
838b6fd2 KW |
1362 | } |
1363 | ||
1364 | /** | |
1365 | * hfi1_kern_exp_rcv_setup() - setup TID's and flow for one segment of a | |
1366 | * TID RDMA request | |
1367 | * | |
1368 | * @req: TID RDMA request for which the segment/flow is being set up | |
1369 | * @ss: sge state, maintains state across successive segments of a sge | |
1370 | * @last: set to true after the last sge segment has been processed | |
1371 | * | |
1372 | * This function | |
1373 | * (1) finds a free flow entry in the flow circular buffer | |
1374 | * (2) finds pages and continuous physical chunks constituing one segment | |
1375 | * of an sge | |
1376 | * (3) allocates TID group entries for those chunks | |
1377 | * (4) programs rcvarray entries in the hardware corresponding to those | |
1378 | * TID's | |
1379 | * (5) computes a tidarray with formatted TID entries which can be sent | |
1380 | * to the sender | |
1381 | * (6) Reserves and programs HW flows. | |
1382 | * (7) It also manages queing the QP when TID/flow resources are not | |
1383 | * available. | |
1384 | * | |
1385 | * @req points to struct tid_rdma_request of which the segments are a part. The | |
1386 | * function uses qp, rcd and seg_len members of @req. In the absence of errors, | |
1387 | * req->flow_idx is the index of the flow which has been prepared in this | |
1388 | * invocation of function call. With flow = &req->flows[req->flow_idx], | |
1389 | * flow->tid_entry contains the TID array which the sender can use for TID RDMA | |
1390 | * sends and flow->npkts contains number of packets required to send the | |
1391 | * segment. | |
1392 | * | |
1393 | * hfi1_check_sge_align should be called prior to calling this function and if | |
1394 | * it signals error TID RDMA cannot be used for this sge and this function | |
1395 | * should not be called. | |
1396 | * | |
1397 | * For the queuing, caller must hold the flow->req->qp s_lock from the send | |
1398 | * engine and the function will procure the exp_lock. | |
1399 | * | |
1400 | * Return: | |
1401 | * The function returns -EAGAIN if sufficient number of TID/flow resources to | |
1402 | * map the segment could not be allocated. In this case the function should be | |
1403 | * called again with previous arguments to retry the TID allocation. There are | |
1404 | * no other error returns. The function returns 0 on success. | |
1405 | */ | |
1406 | int hfi1_kern_exp_rcv_setup(struct tid_rdma_request *req, | |
1407 | struct rvt_sge_state *ss, bool *last) | |
1408 | __must_hold(&req->qp->s_lock) | |
1409 | { | |
1410 | struct tid_rdma_flow *flow = &req->flows[req->setup_head]; | |
1411 | struct hfi1_ctxtdata *rcd = req->rcd; | |
1412 | struct hfi1_qp_priv *qpriv = req->qp->priv; | |
1413 | unsigned long flags; | |
1414 | struct rvt_qp *fqp; | |
1415 | u16 clear_tail = req->clear_tail; | |
1416 | ||
1417 | lockdep_assert_held(&req->qp->s_lock); | |
1418 | /* | |
1419 | * We return error if either (a) we don't have space in the flow | |
1420 | * circular buffer, or (b) we already have max entries in the buffer. | |
1421 | * Max entries depend on the type of request we are processing and the | |
1422 | * negotiated TID RDMA parameters. | |
1423 | */ | |
1424 | if (!CIRC_SPACE(req->setup_head, clear_tail, MAX_FLOWS) || | |
1425 | CIRC_CNT(req->setup_head, clear_tail, MAX_FLOWS) >= | |
1426 | req->n_flows) | |
1427 | return -EINVAL; | |
1428 | ||
1429 | /* | |
1430 | * Get pages, identify contiguous physical memory chunks for the segment | |
1431 | * If we can not determine a DMA address mapping we will treat it just | |
1432 | * like if we ran out of space above. | |
1433 | */ | |
1434 | if (kern_get_phys_blocks(flow, qpriv->pages, ss, last)) { | |
1435 | hfi1_wait_kmem(flow->req->qp); | |
1436 | return -ENOMEM; | |
1437 | } | |
1438 | ||
1439 | spin_lock_irqsave(&rcd->exp_lock, flags); | |
1440 | if (kernel_tid_waiters(rcd, &rcd->rarr_queue, flow->req->qp)) | |
1441 | goto queue; | |
1442 | ||
1443 | /* | |
1444 | * At this point we know the number of pagesets and hence the number of | |
1445 | * TID's to map the segment. Allocate the TID's from the TID groups. If | |
1446 | * we cannot allocate the required number we exit and try again later | |
1447 | */ | |
1448 | if (kern_alloc_tids(flow)) | |
1449 | goto queue; | |
1450 | /* | |
1451 | * Finally program the TID entries with the pagesets, compute the | |
1452 | * tidarray and enable the HW flow | |
1453 | */ | |
1454 | kern_program_rcvarray(flow); | |
1455 | ||
1456 | /* | |
1457 | * Setup the flow state with relevant information. | |
1458 | * This information is used for tracking the sequence of data packets | |
1459 | * for the segment. | |
1460 | * The flow is setup here as this is the most accurate time and place | |
1461 | * to do so. Doing at a later time runs the risk of the flow data in | |
1462 | * qpriv getting out of sync. | |
1463 | */ | |
1464 | memset(&flow->flow_state, 0x0, sizeof(flow->flow_state)); | |
1465 | flow->idx = qpriv->flow_state.index; | |
1466 | flow->flow_state.generation = qpriv->flow_state.generation; | |
1467 | flow->flow_state.spsn = qpriv->flow_state.psn; | |
1468 | flow->flow_state.lpsn = flow->flow_state.spsn + flow->npkts - 1; | |
1469 | flow->flow_state.r_next_psn = | |
1470 | full_flow_psn(flow, flow->flow_state.spsn); | |
1471 | qpriv->flow_state.psn += flow->npkts; | |
1472 | ||
1473 | dequeue_tid_waiter(rcd, &rcd->rarr_queue, flow->req->qp); | |
1474 | /* get head before dropping lock */ | |
1475 | fqp = first_qp(rcd, &rcd->rarr_queue); | |
1476 | spin_unlock_irqrestore(&rcd->exp_lock, flags); | |
1477 | tid_rdma_schedule_tid_wakeup(fqp); | |
1478 | ||
1479 | req->setup_head = (req->setup_head + 1) & (MAX_FLOWS - 1); | |
1480 | return 0; | |
1481 | queue: | |
1482 | queue_qp_for_tid_wait(rcd, &rcd->rarr_queue, flow->req->qp); | |
1483 | spin_unlock_irqrestore(&rcd->exp_lock, flags); | |
1484 | return -EAGAIN; | |
1485 | } | |
1486 | ||
1487 | static void hfi1_tid_rdma_reset_flow(struct tid_rdma_flow *flow) | |
1488 | { | |
1489 | flow->npagesets = 0; | |
1490 | } | |
1491 | ||
1492 | /* | |
1493 | * This function is called after one segment has been successfully sent to | |
1494 | * release the flow and TID HW/SW resources for that segment. The segments for a | |
1495 | * TID RDMA request are setup and cleared in FIFO order which is managed using a | |
1496 | * circular buffer. | |
1497 | */ | |
1498 | int hfi1_kern_exp_rcv_clear(struct tid_rdma_request *req) | |
1499 | __must_hold(&req->qp->s_lock) | |
1500 | { | |
1501 | struct tid_rdma_flow *flow = &req->flows[req->clear_tail]; | |
1502 | struct hfi1_ctxtdata *rcd = req->rcd; | |
1503 | unsigned long flags; | |
1504 | int i; | |
1505 | struct rvt_qp *fqp; | |
1506 | ||
1507 | lockdep_assert_held(&req->qp->s_lock); | |
1508 | /* Exit if we have nothing in the flow circular buffer */ | |
1509 | if (!CIRC_CNT(req->setup_head, req->clear_tail, MAX_FLOWS)) | |
1510 | return -EINVAL; | |
1511 | ||
1512 | spin_lock_irqsave(&rcd->exp_lock, flags); | |
1513 | ||
1514 | for (i = 0; i < flow->tnode_cnt; i++) | |
1515 | kern_unprogram_rcv_group(flow, i); | |
1516 | /* To prevent double unprogramming */ | |
1517 | flow->tnode_cnt = 0; | |
1518 | /* get head before dropping lock */ | |
1519 | fqp = first_qp(rcd, &rcd->rarr_queue); | |
1520 | spin_unlock_irqrestore(&rcd->exp_lock, flags); | |
1521 | ||
1522 | dma_unmap_flow(flow); | |
1523 | ||
1524 | hfi1_tid_rdma_reset_flow(flow); | |
1525 | req->clear_tail = (req->clear_tail + 1) & (MAX_FLOWS - 1); | |
1526 | ||
1527 | if (fqp == req->qp) { | |
1528 | __trigger_tid_waiter(fqp); | |
1529 | rvt_put_qp(fqp); | |
1530 | } else { | |
1531 | tid_rdma_schedule_tid_wakeup(fqp); | |
1532 | } | |
1533 | ||
1534 | return 0; | |
1535 | } | |
1536 | ||
1537 | /* | |
1538 | * This function is called to release all the tid entries for | |
1539 | * a request. | |
1540 | */ | |
1541 | void hfi1_kern_exp_rcv_clear_all(struct tid_rdma_request *req) | |
1542 | __must_hold(&req->qp->s_lock) | |
1543 | { | |
1544 | /* Use memory barrier for proper ordering */ | |
1545 | while (CIRC_CNT(req->setup_head, req->clear_tail, MAX_FLOWS)) { | |
1546 | if (hfi1_kern_exp_rcv_clear(req)) | |
1547 | break; | |
1548 | } | |
1549 | } | |
1550 | ||
1551 | /** | |
1552 | * hfi1_kern_exp_rcv_free_flows - free priviously allocated flow information | |
1553 | * @req - the tid rdma request to be cleaned | |
1554 | */ | |
1555 | static void hfi1_kern_exp_rcv_free_flows(struct tid_rdma_request *req) | |
1556 | { | |
1557 | kfree(req->flows); | |
1558 | req->flows = NULL; | |
1559 | } | |
1560 | ||
1561 | /** | |
1562 | * __trdma_clean_swqe - clean up for large sized QPs | |
1563 | * @qp: the queue patch | |
1564 | * @wqe: the send wqe | |
1565 | */ | |
1566 | void __trdma_clean_swqe(struct rvt_qp *qp, struct rvt_swqe *wqe) | |
1567 | { | |
1568 | struct hfi1_swqe_priv *p = wqe->priv; | |
1569 | ||
1570 | hfi1_kern_exp_rcv_free_flows(&p->tid_req); | |
1571 | } | |
1572 | ||
1573 | /* | |
1574 | * This can be called at QP create time or in the data path. | |
1575 | */ | |
1576 | static int hfi1_kern_exp_rcv_alloc_flows(struct tid_rdma_request *req, | |
1577 | gfp_t gfp) | |
1578 | { | |
1579 | struct tid_rdma_flow *flows; | |
1580 | int i; | |
1581 | ||
1582 | if (likely(req->flows)) | |
1583 | return 0; | |
1584 | flows = kmalloc_node(MAX_FLOWS * sizeof(*flows), gfp, | |
1585 | req->rcd->numa_id); | |
1586 | if (!flows) | |
1587 | return -ENOMEM; | |
1588 | /* mini init */ | |
1589 | for (i = 0; i < MAX_FLOWS; i++) { | |
1590 | flows[i].req = req; | |
1591 | flows[i].npagesets = 0; | |
1592 | flows[i].pagesets[0].mapped = 0; | |
1593 | } | |
1594 | req->flows = flows; | |
1595 | return 0; | |
1596 | } | |
1597 | ||
1598 | static void hfi1_init_trdma_req(struct rvt_qp *qp, | |
1599 | struct tid_rdma_request *req) | |
1600 | { | |
1601 | struct hfi1_qp_priv *qpriv = qp->priv; | |
1602 | ||
1603 | /* | |
1604 | * Initialize various TID RDMA request variables. | |
1605 | * These variables are "static", which is why they | |
1606 | * can be pre-initialized here before the WRs has | |
1607 | * even been submitted. | |
1608 | * However, non-NULL values for these variables do not | |
1609 | * imply that this WQE has been enabled for TID RDMA. | |
1610 | * Drivers should check the WQE's opcode to determine | |
1611 | * if a request is a TID RDMA one or not. | |
1612 | */ | |
1613 | req->qp = qp; | |
1614 | req->rcd = qpriv->rcd; | |
1615 | } | |
2f16a696 KW |
1616 | |
1617 | u64 hfi1_access_sw_tid_wait(const struct cntr_entry *entry, | |
1618 | void *context, int vl, int mode, u64 data) | |
1619 | { | |
1620 | struct hfi1_devdata *dd = context; | |
1621 | ||
1622 | return dd->verbs_dev.n_tidwait; | |
1623 | } | |
742a3826 | 1624 | |
b126078e KW |
1625 | static struct tid_rdma_flow *find_flow_ib(struct tid_rdma_request *req, |
1626 | u32 psn, u16 *fidx) | |
1627 | { | |
1628 | u16 head, tail; | |
1629 | struct tid_rdma_flow *flow; | |
1630 | ||
1631 | head = req->setup_head; | |
1632 | tail = req->clear_tail; | |
1633 | for ( ; CIRC_CNT(head, tail, MAX_FLOWS); | |
1634 | tail = CIRC_NEXT(tail, MAX_FLOWS)) { | |
1635 | flow = &req->flows[tail]; | |
1636 | if (cmp_psn(psn, flow->flow_state.ib_spsn) >= 0 && | |
1637 | cmp_psn(psn, flow->flow_state.ib_lpsn) <= 0) { | |
1638 | if (fidx) | |
1639 | *fidx = tail; | |
1640 | return flow; | |
1641 | } | |
1642 | } | |
1643 | return NULL; | |
1644 | } | |
1645 | ||
9905bf06 KW |
1646 | static struct tid_rdma_flow * |
1647 | __find_flow_ranged(struct tid_rdma_request *req, u16 head, u16 tail, | |
1648 | u32 psn, u16 *fidx) | |
1649 | { | |
1650 | for ( ; CIRC_CNT(head, tail, MAX_FLOWS); | |
1651 | tail = CIRC_NEXT(tail, MAX_FLOWS)) { | |
1652 | struct tid_rdma_flow *flow = &req->flows[tail]; | |
1653 | u32 spsn, lpsn; | |
1654 | ||
1655 | spsn = full_flow_psn(flow, flow->flow_state.spsn); | |
1656 | lpsn = full_flow_psn(flow, flow->flow_state.lpsn); | |
1657 | ||
1658 | if (cmp_psn(psn, spsn) >= 0 && cmp_psn(psn, lpsn) <= 0) { | |
1659 | if (fidx) | |
1660 | *fidx = tail; | |
1661 | return flow; | |
1662 | } | |
1663 | } | |
1664 | return NULL; | |
1665 | } | |
1666 | ||
1667 | static struct tid_rdma_flow *find_flow(struct tid_rdma_request *req, | |
1668 | u32 psn, u16 *fidx) | |
1669 | { | |
1670 | return __find_flow_ranged(req, req->setup_head, req->clear_tail, psn, | |
1671 | fidx); | |
1672 | } | |
1673 | ||
742a3826 KW |
1674 | /* TID RDMA READ functions */ |
1675 | u32 hfi1_build_tid_rdma_read_packet(struct rvt_swqe *wqe, | |
1676 | struct ib_other_headers *ohdr, u32 *bth1, | |
1677 | u32 *bth2, u32 *len) | |
1678 | { | |
1679 | struct tid_rdma_request *req = wqe_to_tid_req(wqe); | |
1680 | struct tid_rdma_flow *flow = &req->flows[req->flow_idx]; | |
1681 | struct rvt_qp *qp = req->qp; | |
1682 | struct hfi1_qp_priv *qpriv = qp->priv; | |
1683 | struct hfi1_swqe_priv *wpriv = wqe->priv; | |
1684 | struct tid_rdma_read_req *rreq = &ohdr->u.tid_rdma.r_req; | |
1685 | struct tid_rdma_params *remote; | |
1686 | u32 req_len = 0; | |
1687 | void *req_addr = NULL; | |
1688 | ||
1689 | /* This is the IB psn used to send the request */ | |
1690 | *bth2 = mask_psn(flow->flow_state.ib_spsn + flow->pkt); | |
3ce5daa2 | 1691 | trace_hfi1_tid_flow_build_read_pkt(qp, req->flow_idx, flow); |
742a3826 KW |
1692 | |
1693 | /* TID Entries for TID RDMA READ payload */ | |
1694 | req_addr = &flow->tid_entry[flow->tid_idx]; | |
1695 | req_len = sizeof(*flow->tid_entry) * | |
1696 | (flow->tidcnt - flow->tid_idx); | |
1697 | ||
1698 | memset(&ohdr->u.tid_rdma.r_req, 0, sizeof(ohdr->u.tid_rdma.r_req)); | |
1699 | wpriv->ss.sge.vaddr = req_addr; | |
1700 | wpriv->ss.sge.sge_length = req_len; | |
1701 | wpriv->ss.sge.length = wpriv->ss.sge.sge_length; | |
1702 | /* | |
1703 | * We can safely zero these out. Since the first SGE covers the | |
1704 | * entire packet, nothing else should even look at the MR. | |
1705 | */ | |
1706 | wpriv->ss.sge.mr = NULL; | |
1707 | wpriv->ss.sge.m = 0; | |
1708 | wpriv->ss.sge.n = 0; | |
1709 | ||
1710 | wpriv->ss.sg_list = NULL; | |
1711 | wpriv->ss.total_len = wpriv->ss.sge.sge_length; | |
1712 | wpriv->ss.num_sge = 1; | |
1713 | ||
1714 | /* Construct the TID RDMA READ REQ packet header */ | |
1715 | rcu_read_lock(); | |
1716 | remote = rcu_dereference(qpriv->tid_rdma.remote); | |
1717 | ||
1718 | KDETH_RESET(rreq->kdeth0, KVER, 0x1); | |
1719 | KDETH_RESET(rreq->kdeth1, JKEY, remote->jkey); | |
1720 | rreq->reth.vaddr = cpu_to_be64(wqe->rdma_wr.remote_addr + | |
1721 | req->cur_seg * req->seg_len + flow->sent); | |
1722 | rreq->reth.rkey = cpu_to_be32(wqe->rdma_wr.rkey); | |
1723 | rreq->reth.length = cpu_to_be32(*len); | |
1724 | rreq->tid_flow_psn = | |
1725 | cpu_to_be32((flow->flow_state.generation << | |
1726 | HFI1_KDETH_BTH_SEQ_SHIFT) | | |
1727 | ((flow->flow_state.spsn + flow->pkt) & | |
1728 | HFI1_KDETH_BTH_SEQ_MASK)); | |
1729 | rreq->tid_flow_qp = | |
1730 | cpu_to_be32(qpriv->tid_rdma.local.qp | | |
1731 | ((flow->idx & TID_RDMA_DESTQP_FLOW_MASK) << | |
1732 | TID_RDMA_DESTQP_FLOW_SHIFT) | | |
1733 | qpriv->rcd->ctxt); | |
1734 | rreq->verbs_qp = cpu_to_be32(qp->remote_qpn); | |
1735 | *bth1 &= ~RVT_QPN_MASK; | |
1736 | *bth1 |= remote->qp; | |
1737 | *bth2 |= IB_BTH_REQ_ACK; | |
1738 | rcu_read_unlock(); | |
1739 | ||
1740 | /* We are done with this segment */ | |
1741 | flow->sent += *len; | |
1742 | req->cur_seg++; | |
1743 | qp->s_state = TID_OP(READ_REQ); | |
1744 | req->ack_pending++; | |
1745 | req->flow_idx = (req->flow_idx + 1) & (MAX_FLOWS - 1); | |
1746 | qpriv->pending_tid_r_segs++; | |
1747 | qp->s_num_rd_atomic++; | |
1748 | ||
1749 | /* Set the TID RDMA READ request payload size */ | |
1750 | *len = req_len; | |
1751 | ||
1752 | return sizeof(ohdr->u.tid_rdma.r_req) / sizeof(u32); | |
1753 | } | |
1754 | ||
1755 | /* | |
1756 | * @len: contains the data length to read upon entry and the read request | |
1757 | * payload length upon exit. | |
1758 | */ | |
1759 | u32 hfi1_build_tid_rdma_read_req(struct rvt_qp *qp, struct rvt_swqe *wqe, | |
1760 | struct ib_other_headers *ohdr, u32 *bth1, | |
1761 | u32 *bth2, u32 *len) | |
1762 | __must_hold(&qp->s_lock) | |
1763 | { | |
1764 | struct hfi1_qp_priv *qpriv = qp->priv; | |
1765 | struct tid_rdma_request *req = wqe_to_tid_req(wqe); | |
1766 | struct tid_rdma_flow *flow = NULL; | |
1767 | u32 hdwords = 0; | |
1768 | bool last; | |
1769 | bool retry = true; | |
1770 | u32 npkts = rvt_div_round_up_mtu(qp, *len); | |
1771 | ||
3ce5daa2 KW |
1772 | trace_hfi1_tid_req_build_read_req(qp, 0, wqe->wr.opcode, wqe->psn, |
1773 | wqe->lpsn, req); | |
742a3826 KW |
1774 | /* |
1775 | * Check sync conditions. Make sure that there are no pending | |
1776 | * segments before freeing the flow. | |
1777 | */ | |
1778 | sync_check: | |
1779 | if (req->state == TID_REQUEST_SYNC) { | |
1780 | if (qpriv->pending_tid_r_segs) | |
1781 | goto done; | |
1782 | ||
1783 | hfi1_kern_clear_hw_flow(req->rcd, qp); | |
1784 | req->state = TID_REQUEST_ACTIVE; | |
1785 | } | |
1786 | ||
1787 | /* | |
1788 | * If the request for this segment is resent, the tid resources should | |
1789 | * have been allocated before. In this case, req->flow_idx should | |
1790 | * fall behind req->setup_head. | |
1791 | */ | |
1792 | if (req->flow_idx == req->setup_head) { | |
1793 | retry = false; | |
1794 | if (req->state == TID_REQUEST_RESEND) { | |
1795 | /* | |
1796 | * This is the first new segment for a request whose | |
1797 | * earlier segments have been re-sent. We need to | |
1798 | * set up the sge pointer correctly. | |
1799 | */ | |
1800 | restart_sge(&qp->s_sge, wqe, req->s_next_psn, | |
1801 | qp->pmtu); | |
1802 | req->isge = 0; | |
1803 | req->state = TID_REQUEST_ACTIVE; | |
1804 | } | |
1805 | ||
1806 | /* | |
1807 | * Check sync. The last PSN of each generation is reserved for | |
1808 | * RESYNC. | |
1809 | */ | |
1810 | if ((qpriv->flow_state.psn + npkts) > MAX_TID_FLOW_PSN - 1) { | |
1811 | req->state = TID_REQUEST_SYNC; | |
1812 | goto sync_check; | |
1813 | } | |
1814 | ||
1815 | /* Allocate the flow if not yet */ | |
1816 | if (hfi1_kern_setup_hw_flow(qpriv->rcd, qp)) | |
1817 | goto done; | |
1818 | ||
1819 | /* | |
1820 | * The following call will advance req->setup_head after | |
1821 | * allocating the tid entries. | |
1822 | */ | |
1823 | if (hfi1_kern_exp_rcv_setup(req, &qp->s_sge, &last)) { | |
1824 | req->state = TID_REQUEST_QUEUED; | |
1825 | ||
1826 | /* | |
1827 | * We don't have resources for this segment. The QP has | |
1828 | * already been queued. | |
1829 | */ | |
1830 | goto done; | |
1831 | } | |
1832 | } | |
1833 | ||
1834 | /* req->flow_idx should only be one slot behind req->setup_head */ | |
1835 | flow = &req->flows[req->flow_idx]; | |
1836 | flow->pkt = 0; | |
1837 | flow->tid_idx = 0; | |
1838 | flow->sent = 0; | |
1839 | if (!retry) { | |
1840 | /* Set the first and last IB PSN for the flow in use.*/ | |
1841 | flow->flow_state.ib_spsn = req->s_next_psn; | |
1842 | flow->flow_state.ib_lpsn = | |
1843 | flow->flow_state.ib_spsn + flow->npkts - 1; | |
1844 | } | |
1845 | ||
1846 | /* Calculate the next segment start psn.*/ | |
1847 | req->s_next_psn += flow->npkts; | |
1848 | ||
1849 | /* Build the packet header */ | |
1850 | hdwords = hfi1_build_tid_rdma_read_packet(wqe, ohdr, bth1, bth2, len); | |
1851 | done: | |
1852 | return hdwords; | |
1853 | } | |
d0d564a1 KW |
1854 | |
1855 | /* | |
1856 | * Validate and accept the TID RDMA READ request parameters. | |
1857 | * Return 0 if the request is accepted successfully; | |
1858 | * Return 1 otherwise. | |
1859 | */ | |
1860 | static int tid_rdma_rcv_read_request(struct rvt_qp *qp, | |
1861 | struct rvt_ack_entry *e, | |
1862 | struct hfi1_packet *packet, | |
1863 | struct ib_other_headers *ohdr, | |
1864 | u32 bth0, u32 psn, u64 vaddr, u32 len) | |
1865 | { | |
1866 | struct hfi1_qp_priv *qpriv = qp->priv; | |
1867 | struct tid_rdma_request *req; | |
1868 | struct tid_rdma_flow *flow; | |
1869 | u32 flow_psn, i, tidlen = 0, pktlen, tlen; | |
1870 | ||
1871 | req = ack_to_tid_req(e); | |
1872 | ||
1873 | /* Validate the payload first */ | |
1874 | flow = &req->flows[req->setup_head]; | |
1875 | ||
1876 | /* payload length = packet length - (header length + ICRC length) */ | |
1877 | pktlen = packet->tlen - (packet->hlen + 4); | |
1878 | if (pktlen > sizeof(flow->tid_entry)) | |
1879 | return 1; | |
1880 | memcpy(flow->tid_entry, packet->ebuf, pktlen); | |
1881 | flow->tidcnt = pktlen / sizeof(*flow->tid_entry); | |
1882 | ||
1883 | /* | |
1884 | * Walk the TID_ENTRY list to make sure we have enough space for a | |
1885 | * complete segment. Also calculate the number of required packets. | |
1886 | */ | |
1887 | flow->npkts = rvt_div_round_up_mtu(qp, len); | |
1888 | for (i = 0; i < flow->tidcnt; i++) { | |
3ce5daa2 KW |
1889 | trace_hfi1_tid_entry_rcv_read_req(qp, i, |
1890 | flow->tid_entry[i]); | |
d0d564a1 KW |
1891 | tlen = EXP_TID_GET(flow->tid_entry[i], LEN); |
1892 | if (!tlen) | |
1893 | return 1; | |
1894 | ||
1895 | /* | |
1896 | * For tid pair (tidctr == 3), the buffer size of the pair | |
1897 | * should be the sum of the buffer size described by each | |
1898 | * tid entry. However, only the first entry needs to be | |
1899 | * specified in the request (see WFR HAS Section 8.5.7.1). | |
1900 | */ | |
1901 | tidlen += tlen; | |
1902 | } | |
1903 | if (tidlen * PAGE_SIZE < len) | |
1904 | return 1; | |
1905 | ||
1906 | /* Empty the flow array */ | |
1907 | req->clear_tail = req->setup_head; | |
1908 | flow->pkt = 0; | |
1909 | flow->tid_idx = 0; | |
1910 | flow->tid_offset = 0; | |
1911 | flow->sent = 0; | |
1912 | flow->tid_qpn = be32_to_cpu(ohdr->u.tid_rdma.r_req.tid_flow_qp); | |
1913 | flow->idx = (flow->tid_qpn >> TID_RDMA_DESTQP_FLOW_SHIFT) & | |
1914 | TID_RDMA_DESTQP_FLOW_MASK; | |
1915 | flow_psn = mask_psn(be32_to_cpu(ohdr->u.tid_rdma.r_req.tid_flow_psn)); | |
1916 | flow->flow_state.generation = flow_psn >> HFI1_KDETH_BTH_SEQ_SHIFT; | |
1917 | flow->flow_state.spsn = flow_psn & HFI1_KDETH_BTH_SEQ_MASK; | |
1918 | flow->length = len; | |
1919 | ||
1920 | flow->flow_state.lpsn = flow->flow_state.spsn + | |
1921 | flow->npkts - 1; | |
1922 | flow->flow_state.ib_spsn = psn; | |
1923 | flow->flow_state.ib_lpsn = flow->flow_state.ib_spsn + flow->npkts - 1; | |
1924 | ||
3ce5daa2 | 1925 | trace_hfi1_tid_flow_rcv_read_req(qp, req->setup_head, flow); |
d0d564a1 KW |
1926 | /* Set the initial flow index to the current flow. */ |
1927 | req->flow_idx = req->setup_head; | |
1928 | ||
1929 | /* advance circular buffer head */ | |
1930 | req->setup_head = (req->setup_head + 1) & (MAX_FLOWS - 1); | |
1931 | ||
1932 | /* | |
1933 | * Compute last PSN for request. | |
1934 | */ | |
1935 | e->opcode = (bth0 >> 24) & 0xff; | |
1936 | e->psn = psn; | |
1937 | e->lpsn = psn + flow->npkts - 1; | |
1938 | e->sent = 0; | |
1939 | ||
1940 | req->n_flows = qpriv->tid_rdma.local.max_read; | |
1941 | req->state = TID_REQUEST_ACTIVE; | |
1942 | req->cur_seg = 0; | |
1943 | req->comp_seg = 0; | |
1944 | req->ack_seg = 0; | |
1945 | req->isge = 0; | |
1946 | req->seg_len = qpriv->tid_rdma.local.max_len; | |
1947 | req->total_len = len; | |
1948 | req->total_segs = 1; | |
1949 | req->r_flow_psn = e->psn; | |
1950 | ||
3ce5daa2 KW |
1951 | trace_hfi1_tid_req_rcv_read_req(qp, 0, e->opcode, e->psn, e->lpsn, |
1952 | req); | |
d0d564a1 KW |
1953 | return 0; |
1954 | } | |
1955 | ||
1956 | static int tid_rdma_rcv_error(struct hfi1_packet *packet, | |
1957 | struct ib_other_headers *ohdr, | |
1958 | struct rvt_qp *qp, u32 psn, int diff) | |
1959 | { | |
1960 | struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num); | |
1961 | struct hfi1_ctxtdata *rcd = ((struct hfi1_qp_priv *)qp->priv)->rcd; | |
1962 | struct rvt_ack_entry *e; | |
1963 | struct tid_rdma_request *req; | |
1964 | unsigned long flags; | |
1965 | u8 prev; | |
1966 | bool old_req; | |
1967 | ||
3ce5daa2 KW |
1968 | trace_hfi1_rsp_tid_rcv_error(qp, psn); |
1969 | trace_hfi1_tid_rdma_rcv_err(qp, 0, psn, diff); | |
d0d564a1 KW |
1970 | if (diff > 0) { |
1971 | /* sequence error */ | |
1972 | if (!qp->r_nak_state) { | |
1973 | ibp->rvp.n_rc_seqnak++; | |
1974 | qp->r_nak_state = IB_NAK_PSN_ERROR; | |
1975 | qp->r_ack_psn = qp->r_psn; | |
1976 | rc_defered_ack(rcd, qp); | |
1977 | } | |
1978 | goto done; | |
1979 | } | |
1980 | ||
1981 | ibp->rvp.n_rc_dupreq++; | |
1982 | ||
1983 | spin_lock_irqsave(&qp->s_lock, flags); | |
1984 | e = find_prev_entry(qp, psn, &prev, NULL, &old_req); | |
1985 | if (!e || e->opcode != TID_OP(READ_REQ)) | |
1986 | goto unlock; | |
1987 | ||
1988 | req = ack_to_tid_req(e); | |
1989 | req->r_flow_psn = psn; | |
3ce5daa2 | 1990 | trace_hfi1_tid_req_rcv_err(qp, 0, e->opcode, e->psn, e->lpsn, req); |
d0d564a1 KW |
1991 | if (e->opcode == TID_OP(READ_REQ)) { |
1992 | struct ib_reth *reth; | |
1993 | u32 offset; | |
1994 | u32 len; | |
1995 | u32 rkey; | |
1996 | u64 vaddr; | |
1997 | int ok; | |
1998 | u32 bth0; | |
1999 | ||
2000 | reth = &ohdr->u.tid_rdma.r_req.reth; | |
2001 | /* | |
2002 | * The requester always restarts from the start of the original | |
2003 | * request. | |
2004 | */ | |
2005 | offset = delta_psn(psn, e->psn) * qp->pmtu; | |
2006 | len = be32_to_cpu(reth->length); | |
2007 | if (psn != e->psn || len != req->total_len) | |
2008 | goto unlock; | |
2009 | ||
2010 | if (e->rdma_sge.mr) { | |
2011 | rvt_put_mr(e->rdma_sge.mr); | |
2012 | e->rdma_sge.mr = NULL; | |
2013 | } | |
2014 | ||
2015 | rkey = be32_to_cpu(reth->rkey); | |
2016 | vaddr = get_ib_reth_vaddr(reth); | |
2017 | ||
2018 | qp->r_len = len; | |
2019 | ok = rvt_rkey_ok(qp, &e->rdma_sge, len, vaddr, rkey, | |
2020 | IB_ACCESS_REMOTE_READ); | |
2021 | if (unlikely(!ok)) | |
2022 | goto unlock; | |
2023 | ||
2024 | /* | |
2025 | * If all the response packets for the current request have | |
2026 | * been sent out and this request is complete (old_request | |
2027 | * == false) and the TID flow may be unusable (the | |
2028 | * req->clear_tail is advanced). However, when an earlier | |
2029 | * request is received, this request will not be complete any | |
2030 | * more (qp->s_tail_ack_queue is moved back, see below). | |
2031 | * Consequently, we need to update the TID flow info everytime | |
2032 | * a duplicate request is received. | |
2033 | */ | |
2034 | bth0 = be32_to_cpu(ohdr->bth[0]); | |
2035 | if (tid_rdma_rcv_read_request(qp, e, packet, ohdr, bth0, psn, | |
2036 | vaddr, len)) | |
2037 | goto unlock; | |
2038 | ||
2039 | /* | |
2040 | * True if the request is already scheduled (between | |
2041 | * qp->s_tail_ack_queue and qp->r_head_ack_queue); | |
2042 | */ | |
2043 | if (old_req) | |
2044 | goto unlock; | |
2045 | } | |
2046 | /* Re-process old requests.*/ | |
4f9264d1 KW |
2047 | if (qp->s_acked_ack_queue == qp->s_tail_ack_queue) |
2048 | qp->s_acked_ack_queue = prev; | |
d0d564a1 KW |
2049 | qp->s_tail_ack_queue = prev; |
2050 | /* | |
2051 | * Since the qp->s_tail_ack_queue is modified, the | |
2052 | * qp->s_ack_state must be changed to re-initialize | |
2053 | * qp->s_ack_rdma_sge; Otherwise, we will end up in | |
2054 | * wrong memory region. | |
2055 | */ | |
2056 | qp->s_ack_state = OP(ACKNOWLEDGE); | |
2057 | qp->r_state = e->opcode; | |
2058 | qp->r_nak_state = 0; | |
2059 | qp->s_flags |= RVT_S_RESP_PENDING; | |
2060 | hfi1_schedule_send(qp); | |
2061 | unlock: | |
2062 | spin_unlock_irqrestore(&qp->s_lock, flags); | |
2063 | done: | |
2064 | return 1; | |
2065 | } | |
2066 | ||
2067 | void hfi1_rc_rcv_tid_rdma_read_req(struct hfi1_packet *packet) | |
2068 | { | |
2069 | /* HANDLER FOR TID RDMA READ REQUEST packet (Responder side)*/ | |
2070 | ||
2071 | /* | |
2072 | * 1. Verify TID RDMA READ REQ as per IB_OPCODE_RC_RDMA_READ | |
2073 | * (see hfi1_rc_rcv()) | |
2074 | * 2. Put TID RDMA READ REQ into the response queueu (s_ack_queue) | |
2075 | * - Setup struct tid_rdma_req with request info | |
2076 | * - Initialize struct tid_rdma_flow info; | |
2077 | * - Copy TID entries; | |
2078 | * 3. Set the qp->s_ack_state. | |
2079 | * 4. Set RVT_S_RESP_PENDING in s_flags. | |
2080 | * 5. Kick the send engine (hfi1_schedule_send()) | |
2081 | */ | |
2082 | struct hfi1_ctxtdata *rcd = packet->rcd; | |
2083 | struct rvt_qp *qp = packet->qp; | |
2084 | struct hfi1_ibport *ibp = to_iport(qp->ibqp.device, qp->port_num); | |
2085 | struct ib_other_headers *ohdr = packet->ohdr; | |
2086 | struct rvt_ack_entry *e; | |
2087 | unsigned long flags; | |
2088 | struct ib_reth *reth; | |
2089 | struct hfi1_qp_priv *qpriv = qp->priv; | |
2090 | u32 bth0, psn, len, rkey; | |
2091 | bool is_fecn; | |
2092 | u8 next; | |
2093 | u64 vaddr; | |
2094 | int diff; | |
2095 | u8 nack_state = IB_NAK_INVALID_REQUEST; | |
2096 | ||
2097 | bth0 = be32_to_cpu(ohdr->bth[0]); | |
2098 | if (hfi1_ruc_check_hdr(ibp, packet)) | |
2099 | return; | |
2100 | ||
2101 | is_fecn = process_ecn(qp, packet); | |
2102 | psn = mask_psn(be32_to_cpu(ohdr->bth[2])); | |
3ce5daa2 | 2103 | trace_hfi1_rsp_rcv_tid_read_req(qp, psn); |
d0d564a1 KW |
2104 | |
2105 | if (qp->state == IB_QPS_RTR && !(qp->r_flags & RVT_R_COMM_EST)) | |
2106 | rvt_comm_est(qp); | |
2107 | ||
2108 | if (unlikely(!(qp->qp_access_flags & IB_ACCESS_REMOTE_READ))) | |
2109 | goto nack_inv; | |
2110 | ||
2111 | reth = &ohdr->u.tid_rdma.r_req.reth; | |
2112 | vaddr = be64_to_cpu(reth->vaddr); | |
2113 | len = be32_to_cpu(reth->length); | |
2114 | /* The length needs to be in multiples of PAGE_SIZE */ | |
2115 | if (!len || len & ~PAGE_MASK || len > qpriv->tid_rdma.local.max_len) | |
2116 | goto nack_inv; | |
2117 | ||
2118 | diff = delta_psn(psn, qp->r_psn); | |
2119 | if (unlikely(diff)) { | |
2120 | if (tid_rdma_rcv_error(packet, ohdr, qp, psn, diff)) | |
2121 | return; | |
2122 | goto send_ack; | |
2123 | } | |
2124 | ||
2125 | /* We've verified the request, insert it into the ack queue. */ | |
2126 | next = qp->r_head_ack_queue + 1; | |
2127 | if (next > rvt_size_atomic(ib_to_rvt(qp->ibqp.device))) | |
2128 | next = 0; | |
2129 | spin_lock_irqsave(&qp->s_lock, flags); | |
2130 | if (unlikely(next == qp->s_tail_ack_queue)) { | |
2131 | if (!qp->s_ack_queue[next].sent) { | |
2132 | nack_state = IB_NAK_REMOTE_OPERATIONAL_ERROR; | |
2133 | goto nack_inv_unlock; | |
2134 | } | |
2135 | update_ack_queue(qp, next); | |
2136 | } | |
2137 | e = &qp->s_ack_queue[qp->r_head_ack_queue]; | |
2138 | if (e->rdma_sge.mr) { | |
2139 | rvt_put_mr(e->rdma_sge.mr); | |
2140 | e->rdma_sge.mr = NULL; | |
2141 | } | |
2142 | ||
2143 | rkey = be32_to_cpu(reth->rkey); | |
2144 | qp->r_len = len; | |
2145 | ||
2146 | if (unlikely(!rvt_rkey_ok(qp, &e->rdma_sge, qp->r_len, vaddr, | |
2147 | rkey, IB_ACCESS_REMOTE_READ))) | |
2148 | goto nack_acc; | |
2149 | ||
2150 | /* Accept the request parameters */ | |
2151 | if (tid_rdma_rcv_read_request(qp, e, packet, ohdr, bth0, psn, vaddr, | |
2152 | len)) | |
2153 | goto nack_inv_unlock; | |
2154 | ||
2155 | qp->r_state = e->opcode; | |
2156 | qp->r_nak_state = 0; | |
2157 | /* | |
2158 | * We need to increment the MSN here instead of when we | |
2159 | * finish sending the result since a duplicate request would | |
2160 | * increment it more than once. | |
2161 | */ | |
2162 | qp->r_msn++; | |
2163 | qp->r_psn += e->lpsn - e->psn + 1; | |
2164 | ||
2165 | qp->r_head_ack_queue = next; | |
2166 | ||
2167 | /* Schedule the send tasklet. */ | |
2168 | qp->s_flags |= RVT_S_RESP_PENDING; | |
2169 | hfi1_schedule_send(qp); | |
2170 | ||
2171 | spin_unlock_irqrestore(&qp->s_lock, flags); | |
2172 | if (is_fecn) | |
2173 | goto send_ack; | |
2174 | return; | |
2175 | ||
2176 | nack_inv_unlock: | |
2177 | spin_unlock_irqrestore(&qp->s_lock, flags); | |
2178 | nack_inv: | |
2179 | rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR); | |
2180 | qp->r_nak_state = nack_state; | |
2181 | qp->r_ack_psn = qp->r_psn; | |
2182 | /* Queue NAK for later */ | |
2183 | rc_defered_ack(rcd, qp); | |
2184 | return; | |
2185 | nack_acc: | |
2186 | spin_unlock_irqrestore(&qp->s_lock, flags); | |
2187 | rvt_rc_error(qp, IB_WC_LOC_PROT_ERR); | |
2188 | qp->r_nak_state = IB_NAK_REMOTE_ACCESS_ERROR; | |
2189 | qp->r_ack_psn = qp->r_psn; | |
2190 | send_ack: | |
2191 | hfi1_send_rc_ack(packet, is_fecn); | |
2192 | } | |
1db21b50 KW |
2193 | |
2194 | u32 hfi1_build_tid_rdma_read_resp(struct rvt_qp *qp, struct rvt_ack_entry *e, | |
2195 | struct ib_other_headers *ohdr, u32 *bth0, | |
2196 | u32 *bth1, u32 *bth2, u32 *len, bool *last) | |
2197 | { | |
2198 | struct hfi1_ack_priv *epriv = e->priv; | |
2199 | struct tid_rdma_request *req = &epriv->tid_req; | |
2200 | struct hfi1_qp_priv *qpriv = qp->priv; | |
2201 | struct tid_rdma_flow *flow = &req->flows[req->clear_tail]; | |
2202 | u32 tidentry = flow->tid_entry[flow->tid_idx]; | |
2203 | u32 tidlen = EXP_TID_GET(tidentry, LEN) << PAGE_SHIFT; | |
2204 | struct tid_rdma_read_resp *resp = &ohdr->u.tid_rdma.r_rsp; | |
2205 | u32 next_offset, om = KDETH_OM_LARGE; | |
2206 | bool last_pkt; | |
2207 | u32 hdwords = 0; | |
2208 | struct tid_rdma_params *remote; | |
2209 | ||
2210 | *len = min_t(u32, qp->pmtu, tidlen - flow->tid_offset); | |
2211 | flow->sent += *len; | |
2212 | next_offset = flow->tid_offset + *len; | |
2213 | last_pkt = (flow->sent >= flow->length); | |
2214 | ||
3ce5daa2 KW |
2215 | trace_hfi1_tid_entry_build_read_resp(qp, flow->tid_idx, tidentry); |
2216 | trace_hfi1_tid_flow_build_read_resp(qp, req->clear_tail, flow); | |
2217 | ||
1db21b50 KW |
2218 | rcu_read_lock(); |
2219 | remote = rcu_dereference(qpriv->tid_rdma.remote); | |
2220 | if (!remote) { | |
2221 | rcu_read_unlock(); | |
2222 | goto done; | |
2223 | } | |
2224 | KDETH_RESET(resp->kdeth0, KVER, 0x1); | |
2225 | KDETH_SET(resp->kdeth0, SH, !last_pkt); | |
2226 | KDETH_SET(resp->kdeth0, INTR, !!(!last_pkt && remote->urg)); | |
2227 | KDETH_SET(resp->kdeth0, TIDCTRL, EXP_TID_GET(tidentry, CTRL)); | |
2228 | KDETH_SET(resp->kdeth0, TID, EXP_TID_GET(tidentry, IDX)); | |
2229 | KDETH_SET(resp->kdeth0, OM, om == KDETH_OM_LARGE); | |
2230 | KDETH_SET(resp->kdeth0, OFFSET, flow->tid_offset / om); | |
2231 | KDETH_RESET(resp->kdeth1, JKEY, remote->jkey); | |
2232 | resp->verbs_qp = cpu_to_be32(qp->remote_qpn); | |
2233 | rcu_read_unlock(); | |
2234 | ||
2235 | resp->aeth = rvt_compute_aeth(qp); | |
2236 | resp->verbs_psn = cpu_to_be32(mask_psn(flow->flow_state.ib_spsn + | |
2237 | flow->pkt)); | |
2238 | ||
2239 | *bth0 = TID_OP(READ_RESP) << 24; | |
2240 | *bth1 = flow->tid_qpn; | |
2241 | *bth2 = mask_psn(((flow->flow_state.spsn + flow->pkt++) & | |
2242 | HFI1_KDETH_BTH_SEQ_MASK) | | |
2243 | (flow->flow_state.generation << | |
2244 | HFI1_KDETH_BTH_SEQ_SHIFT)); | |
2245 | *last = last_pkt; | |
2246 | if (last_pkt) | |
2247 | /* Advance to next flow */ | |
2248 | req->clear_tail = (req->clear_tail + 1) & | |
2249 | (MAX_FLOWS - 1); | |
2250 | ||
2251 | if (next_offset >= tidlen) { | |
2252 | flow->tid_offset = 0; | |
2253 | flow->tid_idx++; | |
2254 | } else { | |
2255 | flow->tid_offset = next_offset; | |
2256 | } | |
2257 | ||
2258 | hdwords = sizeof(ohdr->u.tid_rdma.r_rsp) / sizeof(u32); | |
2259 | ||
2260 | done: | |
2261 | return hdwords; | |
2262 | } | |
9905bf06 KW |
2263 | |
2264 | static inline struct tid_rdma_request * | |
2265 | find_tid_request(struct rvt_qp *qp, u32 psn, enum ib_wr_opcode opcode) | |
2266 | __must_hold(&qp->s_lock) | |
2267 | { | |
2268 | struct rvt_swqe *wqe; | |
2269 | struct tid_rdma_request *req = NULL; | |
2270 | u32 i, end; | |
2271 | ||
2272 | end = qp->s_cur + 1; | |
2273 | if (end == qp->s_size) | |
2274 | end = 0; | |
2275 | for (i = qp->s_acked; i != end;) { | |
2276 | wqe = rvt_get_swqe_ptr(qp, i); | |
2277 | if (cmp_psn(psn, wqe->psn) >= 0 && | |
2278 | cmp_psn(psn, wqe->lpsn) <= 0) { | |
2279 | if (wqe->wr.opcode == opcode) | |
2280 | req = wqe_to_tid_req(wqe); | |
2281 | break; | |
2282 | } | |
2283 | if (++i == qp->s_size) | |
2284 | i = 0; | |
2285 | } | |
2286 | ||
2287 | return req; | |
2288 | } | |
2289 | ||
2290 | void hfi1_rc_rcv_tid_rdma_read_resp(struct hfi1_packet *packet) | |
2291 | { | |
2292 | /* HANDLER FOR TID RDMA READ RESPONSE packet (Requestor side */ | |
2293 | ||
2294 | /* | |
2295 | * 1. Find matching SWQE | |
2296 | * 2. Check that the entire segment has been read. | |
2297 | * 3. Remove HFI1_S_WAIT_TID_RESP from s_flags. | |
2298 | * 4. Free the TID flow resources. | |
2299 | * 5. Kick the send engine (hfi1_schedule_send()) | |
2300 | */ | |
2301 | struct ib_other_headers *ohdr = packet->ohdr; | |
2302 | struct rvt_qp *qp = packet->qp; | |
2303 | struct hfi1_qp_priv *priv = qp->priv; | |
2304 | struct hfi1_ctxtdata *rcd = packet->rcd; | |
2305 | struct tid_rdma_request *req; | |
2306 | struct tid_rdma_flow *flow; | |
2307 | u32 opcode, aeth; | |
2308 | bool is_fecn; | |
2309 | unsigned long flags; | |
2310 | u32 kpsn, ipsn; | |
2311 | ||
3ce5daa2 | 2312 | trace_hfi1_sender_rcv_tid_read_resp(qp); |
9905bf06 KW |
2313 | is_fecn = process_ecn(qp, packet); |
2314 | kpsn = mask_psn(be32_to_cpu(ohdr->bth[2])); | |
2315 | aeth = be32_to_cpu(ohdr->u.tid_rdma.r_rsp.aeth); | |
2316 | opcode = (be32_to_cpu(ohdr->bth[0]) >> 24) & 0xff; | |
2317 | ||
2318 | spin_lock_irqsave(&qp->s_lock, flags); | |
2319 | ipsn = mask_psn(be32_to_cpu(ohdr->u.tid_rdma.r_rsp.verbs_psn)); | |
2320 | req = find_tid_request(qp, ipsn, IB_WR_TID_RDMA_READ); | |
2321 | if (unlikely(!req)) | |
2322 | goto ack_op_err; | |
2323 | ||
2324 | flow = &req->flows[req->clear_tail]; | |
2325 | /* When header suppression is disabled */ | |
2326 | if (cmp_psn(ipsn, flow->flow_state.ib_lpsn)) | |
2327 | goto ack_done; | |
2328 | req->ack_pending--; | |
2329 | priv->pending_tid_r_segs--; | |
2330 | qp->s_num_rd_atomic--; | |
2331 | if ((qp->s_flags & RVT_S_WAIT_FENCE) && | |
2332 | !qp->s_num_rd_atomic) { | |
2333 | qp->s_flags &= ~(RVT_S_WAIT_FENCE | | |
2334 | RVT_S_WAIT_ACK); | |
2335 | hfi1_schedule_send(qp); | |
2336 | } | |
2337 | if (qp->s_flags & RVT_S_WAIT_RDMAR) { | |
2338 | qp->s_flags &= ~(RVT_S_WAIT_RDMAR | RVT_S_WAIT_ACK); | |
2339 | hfi1_schedule_send(qp); | |
2340 | } | |
2341 | ||
3ce5daa2 KW |
2342 | trace_hfi1_ack(qp, ipsn); |
2343 | trace_hfi1_tid_req_rcv_read_resp(qp, 0, req->e.swqe->wr.opcode, | |
2344 | req->e.swqe->psn, req->e.swqe->lpsn, | |
2345 | req); | |
2346 | trace_hfi1_tid_flow_rcv_read_resp(qp, req->clear_tail, flow); | |
2347 | ||
9905bf06 KW |
2348 | /* Release the tid resources */ |
2349 | hfi1_kern_exp_rcv_clear(req); | |
2350 | ||
2351 | if (!do_rc_ack(qp, aeth, ipsn, opcode, 0, rcd)) | |
2352 | goto ack_done; | |
2353 | ||
2354 | /* If not done yet, build next read request */ | |
2355 | if (++req->comp_seg >= req->total_segs) { | |
2356 | priv->tid_r_comp++; | |
2357 | req->state = TID_REQUEST_COMPLETE; | |
2358 | } | |
2359 | ||
2360 | /* | |
2361 | * Clear the hw flow under two conditions: | |
2362 | * 1. This request is a sync point and it is complete; | |
2363 | * 2. Current request is completed and there are no more requests. | |
2364 | */ | |
2365 | if ((req->state == TID_REQUEST_SYNC && | |
2366 | req->comp_seg == req->cur_seg) || | |
2367 | priv->tid_r_comp == priv->tid_r_reqs) { | |
2368 | hfi1_kern_clear_hw_flow(priv->rcd, qp); | |
2369 | if (req->state == TID_REQUEST_SYNC) | |
2370 | req->state = TID_REQUEST_ACTIVE; | |
2371 | } | |
2372 | ||
2373 | hfi1_schedule_send(qp); | |
2374 | goto ack_done; | |
2375 | ||
2376 | ack_op_err: | |
2377 | /* | |
2378 | * The test indicates that the send engine has finished its cleanup | |
2379 | * after sending the request and it's now safe to put the QP into error | |
2380 | * state. However, if the wqe queue is empty (qp->s_acked == qp->s_tail | |
2381 | * == qp->s_head), it would be unsafe to complete the wqe pointed by | |
2382 | * qp->s_acked here. Putting the qp into error state will safely flush | |
2383 | * all remaining requests. | |
2384 | */ | |
2385 | if (qp->s_last == qp->s_acked) | |
2386 | rvt_error_qp(qp, IB_WC_WR_FLUSH_ERR); | |
2387 | ||
2388 | ack_done: | |
2389 | spin_unlock_irqrestore(&qp->s_lock, flags); | |
2390 | if (is_fecn) | |
2391 | hfi1_send_rc_ack(packet, is_fecn); | |
2392 | } | |
2393 | ||
2394 | void hfi1_kern_read_tid_flow_free(struct rvt_qp *qp) | |
2395 | __must_hold(&qp->s_lock) | |
2396 | { | |
2397 | u32 n = qp->s_acked; | |
2398 | struct rvt_swqe *wqe; | |
2399 | struct tid_rdma_request *req; | |
2400 | struct hfi1_qp_priv *priv = qp->priv; | |
2401 | ||
2402 | lockdep_assert_held(&qp->s_lock); | |
2403 | /* Free any TID entries */ | |
2404 | while (n != qp->s_tail) { | |
2405 | wqe = rvt_get_swqe_ptr(qp, n); | |
2406 | if (wqe->wr.opcode == IB_WR_TID_RDMA_READ) { | |
2407 | req = wqe_to_tid_req(wqe); | |
2408 | hfi1_kern_exp_rcv_clear_all(req); | |
2409 | } | |
2410 | ||
2411 | if (++n == qp->s_size) | |
2412 | n = 0; | |
2413 | } | |
2414 | /* Free flow */ | |
2415 | hfi1_kern_clear_hw_flow(priv->rcd, qp); | |
2416 | } | |
2417 | ||
2418 | static bool tid_rdma_tid_err(struct hfi1_ctxtdata *rcd, | |
2419 | struct hfi1_packet *packet, u8 rcv_type, | |
2420 | u8 opcode) | |
2421 | { | |
2422 | struct rvt_qp *qp = packet->qp; | |
2423 | u32 ipsn; | |
2424 | struct ib_other_headers *ohdr = packet->ohdr; | |
2425 | ||
2426 | if (rcv_type >= RHF_RCV_TYPE_IB) | |
2427 | goto done; | |
2428 | ||
2429 | spin_lock(&qp->s_lock); | |
2430 | /* | |
2431 | * For TID READ response, error out QP after freeing the tid | |
2432 | * resources. | |
2433 | */ | |
2434 | if (opcode == TID_OP(READ_RESP)) { | |
2435 | ipsn = mask_psn(be32_to_cpu(ohdr->u.tid_rdma.r_rsp.verbs_psn)); | |
2436 | if (cmp_psn(ipsn, qp->s_last_psn) > 0 && | |
2437 | cmp_psn(ipsn, qp->s_psn) < 0) { | |
2438 | hfi1_kern_read_tid_flow_free(qp); | |
2439 | spin_unlock(&qp->s_lock); | |
2440 | rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR); | |
2441 | goto done; | |
2442 | } | |
2443 | } | |
2444 | ||
2445 | spin_unlock(&qp->s_lock); | |
2446 | done: | |
2447 | return true; | |
2448 | } | |
2449 | ||
2450 | static void restart_tid_rdma_read_req(struct hfi1_ctxtdata *rcd, | |
2451 | struct rvt_qp *qp, struct rvt_swqe *wqe) | |
2452 | { | |
2453 | struct tid_rdma_request *req; | |
2454 | struct tid_rdma_flow *flow; | |
2455 | ||
2456 | /* Start from the right segment */ | |
2457 | qp->r_flags |= RVT_R_RDMAR_SEQ; | |
2458 | req = wqe_to_tid_req(wqe); | |
2459 | flow = &req->flows[req->clear_tail]; | |
2460 | hfi1_restart_rc(qp, flow->flow_state.ib_spsn, 0); | |
2461 | if (list_empty(&qp->rspwait)) { | |
2462 | qp->r_flags |= RVT_R_RSP_SEND; | |
2463 | rvt_get_qp(qp); | |
2464 | list_add_tail(&qp->rspwait, &rcd->qp_wait_list); | |
2465 | } | |
2466 | } | |
2467 | ||
2468 | /* | |
2469 | * Handle the KDETH eflags for TID RDMA READ response. | |
2470 | * | |
2471 | * Return true if the last packet for a segment has been received and it is | |
2472 | * time to process the response normally; otherwise, return true. | |
2473 | * | |
2474 | * The caller must hold the packet->qp->r_lock and the rcu_read_lock. | |
2475 | */ | |
2476 | static bool handle_read_kdeth_eflags(struct hfi1_ctxtdata *rcd, | |
2477 | struct hfi1_packet *packet, u8 rcv_type, | |
2478 | u8 rte, u32 psn, u32 ibpsn) | |
2479 | __must_hold(&packet->qp->r_lock) __must_hold(RCU) | |
2480 | { | |
2481 | struct hfi1_pportdata *ppd = rcd->ppd; | |
2482 | struct hfi1_devdata *dd = ppd->dd; | |
2483 | struct hfi1_ibport *ibp; | |
2484 | struct rvt_swqe *wqe; | |
2485 | struct tid_rdma_request *req; | |
2486 | struct tid_rdma_flow *flow; | |
2487 | u32 ack_psn; | |
2488 | struct rvt_qp *qp = packet->qp; | |
2489 | struct hfi1_qp_priv *priv = qp->priv; | |
2490 | bool ret = true; | |
2491 | int diff = 0; | |
2492 | u32 fpsn; | |
2493 | ||
2494 | lockdep_assert_held(&qp->r_lock); | |
2495 | /* If the psn is out of valid range, drop the packet */ | |
2496 | if (cmp_psn(ibpsn, qp->s_last_psn) < 0 || | |
2497 | cmp_psn(ibpsn, qp->s_psn) > 0) | |
2498 | return ret; | |
2499 | ||
2500 | spin_lock(&qp->s_lock); | |
2501 | /* | |
2502 | * Note that NAKs implicitly ACK outstanding SEND and RDMA write | |
2503 | * requests and implicitly NAK RDMA read and atomic requests issued | |
2504 | * before the NAK'ed request. | |
2505 | */ | |
2506 | ack_psn = ibpsn - 1; | |
2507 | wqe = rvt_get_swqe_ptr(qp, qp->s_acked); | |
2508 | ibp = to_iport(qp->ibqp.device, qp->port_num); | |
2509 | ||
2510 | /* Complete WQEs that the PSN finishes. */ | |
2511 | while ((int)delta_psn(ack_psn, wqe->lpsn) >= 0) { | |
2512 | /* | |
2513 | * If this request is a RDMA read or atomic, and the NACK is | |
2514 | * for a later operation, this NACK NAKs the RDMA read or | |
2515 | * atomic. | |
2516 | */ | |
2517 | if (wqe->wr.opcode == IB_WR_RDMA_READ || | |
2518 | wqe->wr.opcode == IB_WR_TID_RDMA_READ || | |
2519 | wqe->wr.opcode == IB_WR_ATOMIC_CMP_AND_SWP || | |
2520 | wqe->wr.opcode == IB_WR_ATOMIC_FETCH_AND_ADD) { | |
2521 | /* Retry this request. */ | |
2522 | if (!(qp->r_flags & RVT_R_RDMAR_SEQ)) { | |
2523 | qp->r_flags |= RVT_R_RDMAR_SEQ; | |
2524 | if (wqe->wr.opcode == IB_WR_TID_RDMA_READ) { | |
2525 | restart_tid_rdma_read_req(rcd, qp, | |
2526 | wqe); | |
2527 | } else { | |
2528 | hfi1_restart_rc(qp, qp->s_last_psn + 1, | |
2529 | 0); | |
2530 | if (list_empty(&qp->rspwait)) { | |
2531 | qp->r_flags |= RVT_R_RSP_SEND; | |
2532 | rvt_get_qp(qp); | |
2533 | list_add_tail(/* wait */ | |
2534 | &qp->rspwait, | |
2535 | &rcd->qp_wait_list); | |
2536 | } | |
2537 | } | |
2538 | } | |
2539 | /* | |
2540 | * No need to process the NAK since we are | |
2541 | * restarting an earlier request. | |
2542 | */ | |
2543 | break; | |
2544 | } | |
2545 | ||
2546 | wqe = do_rc_completion(qp, wqe, ibp); | |
2547 | if (qp->s_acked == qp->s_tail) | |
2548 | break; | |
2549 | } | |
2550 | ||
2551 | /* Handle the eflags for the request */ | |
2552 | if (wqe->wr.opcode != IB_WR_TID_RDMA_READ) | |
2553 | goto s_unlock; | |
2554 | ||
2555 | req = wqe_to_tid_req(wqe); | |
2556 | switch (rcv_type) { | |
2557 | case RHF_RCV_TYPE_EXPECTED: | |
2558 | switch (rte) { | |
2559 | case RHF_RTE_EXPECTED_FLOW_SEQ_ERR: | |
2560 | /* | |
2561 | * On the first occurrence of a Flow Sequence error, | |
2562 | * the flag TID_FLOW_SW_PSN is set. | |
2563 | * | |
2564 | * After that, the flow is *not* reprogrammed and the | |
2565 | * protocol falls back to SW PSN checking. This is done | |
2566 | * to prevent continuous Flow Sequence errors for any | |
2567 | * packets that could be still in the fabric. | |
2568 | */ | |
2569 | flow = find_flow(req, psn, NULL); | |
2570 | if (!flow) { | |
2571 | /* | |
2572 | * We can't find the IB PSN matching the | |
2573 | * received KDETH PSN. The only thing we can | |
2574 | * do at this point is report the error to | |
2575 | * the QP. | |
2576 | */ | |
2577 | hfi1_kern_read_tid_flow_free(qp); | |
2578 | spin_unlock(&qp->s_lock); | |
2579 | rvt_rc_error(qp, IB_WC_LOC_QP_OP_ERR); | |
2580 | return ret; | |
2581 | } | |
2582 | if (priv->flow_state.flags & TID_FLOW_SW_PSN) { | |
2583 | diff = cmp_psn(psn, | |
2584 | priv->flow_state.r_next_psn); | |
2585 | if (diff > 0) { | |
2586 | if (!(qp->r_flags & RVT_R_RDMAR_SEQ)) | |
2587 | restart_tid_rdma_read_req(rcd, | |
2588 | qp, | |
2589 | wqe); | |
2590 | ||
2591 | /* Drop the packet.*/ | |
2592 | goto s_unlock; | |
2593 | } else if (diff < 0) { | |
2594 | /* | |
2595 | * If a response packet for a restarted | |
2596 | * request has come back, reset the | |
2597 | * restart flag. | |
2598 | */ | |
2599 | if (qp->r_flags & RVT_R_RDMAR_SEQ) | |
2600 | qp->r_flags &= | |
2601 | ~RVT_R_RDMAR_SEQ; | |
2602 | ||
2603 | /* Drop the packet.*/ | |
2604 | goto s_unlock; | |
2605 | } | |
2606 | ||
2607 | /* | |
2608 | * If SW PSN verification is successful and | |
2609 | * this is the last packet in the segment, tell | |
2610 | * the caller to process it as a normal packet. | |
2611 | */ | |
2612 | fpsn = full_flow_psn(flow, | |
2613 | flow->flow_state.lpsn); | |
2614 | if (cmp_psn(fpsn, psn) == 0) { | |
2615 | ret = false; | |
2616 | if (qp->r_flags & RVT_R_RDMAR_SEQ) | |
2617 | qp->r_flags &= | |
2618 | ~RVT_R_RDMAR_SEQ; | |
2619 | } | |
2620 | priv->flow_state.r_next_psn++; | |
2621 | } else { | |
2622 | u64 reg; | |
2623 | u32 last_psn; | |
2624 | ||
2625 | /* | |
2626 | * The only sane way to get the amount of | |
2627 | * progress is to read the HW flow state. | |
2628 | */ | |
2629 | reg = read_uctxt_csr(dd, rcd->ctxt, | |
2630 | RCV_TID_FLOW_TABLE + | |
2631 | (8 * flow->idx)); | |
2632 | last_psn = mask_psn(reg); | |
2633 | ||
2634 | priv->flow_state.r_next_psn = last_psn; | |
2635 | priv->flow_state.flags |= TID_FLOW_SW_PSN; | |
2636 | /* | |
2637 | * If no request has been restarted yet, | |
2638 | * restart the current one. | |
2639 | */ | |
2640 | if (!(qp->r_flags & RVT_R_RDMAR_SEQ)) | |
2641 | restart_tid_rdma_read_req(rcd, qp, | |
2642 | wqe); | |
2643 | } | |
2644 | ||
2645 | break; | |
2646 | ||
2647 | case RHF_RTE_EXPECTED_FLOW_GEN_ERR: | |
2648 | /* | |
2649 | * Since the TID flow is able to ride through | |
2650 | * generation mismatch, drop this stale packet. | |
2651 | */ | |
2652 | break; | |
2653 | ||
2654 | default: | |
2655 | break; | |
2656 | } | |
2657 | break; | |
2658 | ||
2659 | case RHF_RCV_TYPE_ERROR: | |
2660 | switch (rte) { | |
2661 | case RHF_RTE_ERROR_OP_CODE_ERR: | |
2662 | case RHF_RTE_ERROR_KHDR_MIN_LEN_ERR: | |
2663 | case RHF_RTE_ERROR_KHDR_HCRC_ERR: | |
2664 | case RHF_RTE_ERROR_KHDR_KVER_ERR: | |
2665 | case RHF_RTE_ERROR_CONTEXT_ERR: | |
2666 | case RHF_RTE_ERROR_KHDR_TID_ERR: | |
2667 | default: | |
2668 | break; | |
2669 | } | |
2670 | default: | |
2671 | break; | |
2672 | } | |
2673 | s_unlock: | |
2674 | spin_unlock(&qp->s_lock); | |
2675 | return ret; | |
2676 | } | |
2677 | ||
2678 | bool hfi1_handle_kdeth_eflags(struct hfi1_ctxtdata *rcd, | |
2679 | struct hfi1_pportdata *ppd, | |
2680 | struct hfi1_packet *packet) | |
2681 | { | |
2682 | struct hfi1_ibport *ibp = &ppd->ibport_data; | |
2683 | struct hfi1_devdata *dd = ppd->dd; | |
2684 | struct rvt_dev_info *rdi = &dd->verbs_dev.rdi; | |
2685 | u8 rcv_type = rhf_rcv_type(packet->rhf); | |
2686 | u8 rte = rhf_rcv_type_err(packet->rhf); | |
2687 | struct ib_header *hdr = packet->hdr; | |
2688 | struct ib_other_headers *ohdr = NULL; | |
2689 | int lnh = be16_to_cpu(hdr->lrh[0]) & 3; | |
2690 | u16 lid = be16_to_cpu(hdr->lrh[1]); | |
2691 | u8 opcode; | |
2692 | u32 qp_num, psn, ibpsn; | |
2693 | struct rvt_qp *qp; | |
2694 | unsigned long flags; | |
2695 | bool ret = true; | |
2696 | ||
3ce5daa2 KW |
2697 | trace_hfi1_msg_handle_kdeth_eflags(NULL, "Kdeth error: rhf ", |
2698 | packet->rhf); | |
9905bf06 KW |
2699 | if (packet->rhf & (RHF_VCRC_ERR | RHF_ICRC_ERR)) |
2700 | return ret; | |
2701 | ||
2702 | packet->ohdr = &hdr->u.oth; | |
2703 | ohdr = packet->ohdr; | |
2704 | trace_input_ibhdr(rcd->dd, packet, !!(rhf_dc_info(packet->rhf))); | |
2705 | ||
2706 | /* Get the destination QP number. */ | |
2707 | qp_num = be32_to_cpu(ohdr->u.tid_rdma.r_rsp.verbs_qp) & | |
2708 | RVT_QPN_MASK; | |
2709 | if (lid >= be16_to_cpu(IB_MULTICAST_LID_BASE)) | |
2710 | goto drop; | |
2711 | ||
2712 | psn = mask_psn(be32_to_cpu(ohdr->bth[2])); | |
2713 | opcode = (be32_to_cpu(ohdr->bth[0]) >> 24) & 0xff; | |
2714 | ||
2715 | rcu_read_lock(); | |
2716 | qp = rvt_lookup_qpn(rdi, &ibp->rvp, qp_num); | |
2717 | if (!qp) | |
2718 | goto rcu_unlock; | |
2719 | ||
2720 | packet->qp = qp; | |
2721 | ||
2722 | /* Check for valid receive state. */ | |
2723 | spin_lock_irqsave(&qp->r_lock, flags); | |
2724 | if (!(ib_rvt_state_ops[qp->state] & RVT_PROCESS_RECV_OK)) { | |
2725 | ibp->rvp.n_pkt_drops++; | |
2726 | goto r_unlock; | |
2727 | } | |
2728 | ||
2729 | if (packet->rhf & RHF_TID_ERR) { | |
2730 | /* For TIDERR and RC QPs preemptively schedule a NAK */ | |
2731 | u32 tlen = rhf_pkt_len(packet->rhf); /* in bytes */ | |
2732 | ||
2733 | /* Sanity check packet */ | |
2734 | if (tlen < 24) | |
2735 | goto r_unlock; | |
2736 | ||
2737 | /* | |
2738 | * Check for GRH. We should never get packets with GRH in this | |
2739 | * path. | |
2740 | */ | |
2741 | if (lnh == HFI1_LRH_GRH) | |
2742 | goto r_unlock; | |
2743 | ||
2744 | if (tid_rdma_tid_err(rcd, packet, rcv_type, opcode)) | |
2745 | goto r_unlock; | |
2746 | } | |
2747 | ||
2748 | /* handle TID RDMA READ */ | |
2749 | if (opcode == TID_OP(READ_RESP)) { | |
2750 | ibpsn = be32_to_cpu(ohdr->u.tid_rdma.r_rsp.verbs_psn); | |
2751 | ibpsn = mask_psn(ibpsn); | |
2752 | ret = handle_read_kdeth_eflags(rcd, packet, rcv_type, rte, psn, | |
2753 | ibpsn); | |
2754 | } | |
2755 | ||
2756 | r_unlock: | |
2757 | spin_unlock_irqrestore(&qp->r_lock, flags); | |
2758 | rcu_unlock: | |
2759 | rcu_read_unlock(); | |
2760 | drop: | |
2761 | return ret; | |
2762 | } | |
b126078e KW |
2763 | |
2764 | /* | |
2765 | * "Rewind" the TID request information. | |
2766 | * This means that we reset the state back to ACTIVE, | |
2767 | * find the proper flow, set the flow index to that flow, | |
2768 | * and reset the flow information. | |
2769 | */ | |
2770 | void hfi1_tid_rdma_restart_req(struct rvt_qp *qp, struct rvt_swqe *wqe, | |
2771 | u32 *bth2) | |
2772 | { | |
2773 | struct tid_rdma_request *req = wqe_to_tid_req(wqe); | |
2774 | struct tid_rdma_flow *flow; | |
2775 | int diff; | |
2776 | u32 tididx = 0; | |
2777 | u16 fidx; | |
2778 | ||
2779 | if (wqe->wr.opcode == IB_WR_TID_RDMA_READ) { | |
2780 | *bth2 = mask_psn(qp->s_psn); | |
2781 | flow = find_flow_ib(req, *bth2, &fidx); | |
3ce5daa2 KW |
2782 | if (!flow) { |
2783 | trace_hfi1_msg_tid_restart_req(/* msg */ | |
2784 | qp, "!!!!!! Could not find flow to restart: bth2 ", | |
2785 | (u64)*bth2); | |
2786 | trace_hfi1_tid_req_restart_req(qp, 0, wqe->wr.opcode, | |
2787 | wqe->psn, wqe->lpsn, | |
2788 | req); | |
b126078e | 2789 | return; |
3ce5daa2 | 2790 | } |
b126078e KW |
2791 | } else { |
2792 | return; | |
2793 | } | |
2794 | ||
3ce5daa2 | 2795 | trace_hfi1_tid_flow_restart_req(qp, fidx, flow); |
b126078e KW |
2796 | diff = delta_psn(*bth2, flow->flow_state.ib_spsn); |
2797 | ||
2798 | flow->sent = 0; | |
2799 | flow->pkt = 0; | |
2800 | flow->tid_idx = 0; | |
2801 | flow->tid_offset = 0; | |
2802 | if (diff) { | |
2803 | for (tididx = 0; tididx < flow->tidcnt; tididx++) { | |
2804 | u32 tidentry = flow->tid_entry[tididx], tidlen, | |
2805 | tidnpkts, npkts; | |
2806 | ||
2807 | flow->tid_offset = 0; | |
2808 | tidlen = EXP_TID_GET(tidentry, LEN) * PAGE_SIZE; | |
2809 | tidnpkts = rvt_div_round_up_mtu(qp, tidlen); | |
2810 | npkts = min_t(u32, diff, tidnpkts); | |
2811 | flow->pkt += npkts; | |
2812 | flow->sent += (npkts == tidnpkts ? tidlen : | |
2813 | npkts * qp->pmtu); | |
2814 | flow->tid_offset += npkts * qp->pmtu; | |
2815 | diff -= npkts; | |
2816 | if (!diff) | |
2817 | break; | |
2818 | } | |
2819 | } | |
2820 | ||
2821 | if (flow->tid_offset == | |
2822 | EXP_TID_GET(flow->tid_entry[tididx], LEN) * PAGE_SIZE) { | |
2823 | tididx++; | |
2824 | flow->tid_offset = 0; | |
2825 | } | |
2826 | flow->tid_idx = tididx; | |
2827 | /* Move flow_idx to correct index */ | |
2828 | req->flow_idx = fidx; | |
2829 | ||
3ce5daa2 KW |
2830 | trace_hfi1_tid_flow_restart_req(qp, fidx, flow); |
2831 | trace_hfi1_tid_req_restart_req(qp, 0, wqe->wr.opcode, wqe->psn, | |
2832 | wqe->lpsn, req); | |
b126078e KW |
2833 | req->state = TID_REQUEST_ACTIVE; |
2834 | } | |
24b11923 KW |
2835 | |
2836 | void hfi1_qp_kern_exp_rcv_clear_all(struct rvt_qp *qp) | |
2837 | { | |
2838 | int i, ret; | |
2839 | struct hfi1_qp_priv *qpriv = qp->priv; | |
2840 | struct tid_flow_state *fs; | |
2841 | ||
2842 | if (qp->ibqp.qp_type != IB_QPT_RC || !HFI1_CAP_IS_KSET(TID_RDMA)) | |
2843 | return; | |
2844 | ||
2845 | /* | |
2846 | * First, clear the flow to help prevent any delayed packets from | |
2847 | * being delivered. | |
2848 | */ | |
2849 | fs = &qpriv->flow_state; | |
2850 | if (fs->index != RXE_NUM_TID_FLOWS) | |
2851 | hfi1_kern_clear_hw_flow(qpriv->rcd, qp); | |
2852 | ||
2853 | for (i = qp->s_acked; i != qp->s_head;) { | |
2854 | struct rvt_swqe *wqe = rvt_get_swqe_ptr(qp, i); | |
2855 | ||
2856 | if (++i == qp->s_size) | |
2857 | i = 0; | |
2858 | /* Free only locally allocated TID entries */ | |
2859 | if (wqe->wr.opcode != IB_WR_TID_RDMA_READ) | |
2860 | continue; | |
2861 | do { | |
2862 | struct hfi1_swqe_priv *priv = wqe->priv; | |
2863 | ||
2864 | ret = hfi1_kern_exp_rcv_clear(&priv->tid_req); | |
2865 | } while (!ret); | |
2866 | } | |
2867 | } | |
a0b34f75 KW |
2868 | |
2869 | bool hfi1_tid_rdma_wqe_interlock(struct rvt_qp *qp, struct rvt_swqe *wqe) | |
2870 | { | |
2871 | struct rvt_swqe *prev; | |
2872 | struct hfi1_qp_priv *priv = qp->priv; | |
2873 | u32 s_prev; | |
2874 | ||
2875 | s_prev = (qp->s_cur == 0 ? qp->s_size : qp->s_cur) - 1; | |
2876 | prev = rvt_get_swqe_ptr(qp, s_prev); | |
2877 | ||
2878 | switch (wqe->wr.opcode) { | |
2879 | case IB_WR_SEND: | |
2880 | case IB_WR_SEND_WITH_IMM: | |
2881 | case IB_WR_SEND_WITH_INV: | |
2882 | case IB_WR_ATOMIC_CMP_AND_SWP: | |
2883 | case IB_WR_ATOMIC_FETCH_AND_ADD: | |
2884 | case IB_WR_RDMA_WRITE: | |
2885 | case IB_WR_RDMA_READ: | |
2886 | break; | |
2887 | case IB_WR_TID_RDMA_READ: | |
2888 | switch (prev->wr.opcode) { | |
2889 | case IB_WR_RDMA_READ: | |
2890 | if (qp->s_acked != qp->s_cur) | |
2891 | goto interlock; | |
2892 | break; | |
2893 | default: | |
2894 | break; | |
2895 | } | |
2896 | default: | |
2897 | break; | |
2898 | } | |
2899 | return false; | |
2900 | ||
2901 | interlock: | |
2902 | priv->s_flags |= HFI1_S_TID_WAIT_INTERLCK; | |
2903 | return true; | |
2904 | } | |
f1ab4efa KW |
2905 | |
2906 | /* Does @sge meet the alignment requirements for tid rdma? */ | |
3ce5daa2 KW |
2907 | static inline bool hfi1_check_sge_align(struct rvt_qp *qp, |
2908 | struct rvt_sge *sge, int num_sge) | |
f1ab4efa KW |
2909 | { |
2910 | int i; | |
2911 | ||
3ce5daa2 KW |
2912 | for (i = 0; i < num_sge; i++, sge++) { |
2913 | trace_hfi1_sge_check_align(qp, i, sge); | |
f1ab4efa KW |
2914 | if ((u64)sge->vaddr & ~PAGE_MASK || |
2915 | sge->sge_length & ~PAGE_MASK) | |
2916 | return false; | |
3ce5daa2 | 2917 | } |
f1ab4efa KW |
2918 | return true; |
2919 | } | |
2920 | ||
2921 | void setup_tid_rdma_wqe(struct rvt_qp *qp, struct rvt_swqe *wqe) | |
2922 | { | |
2923 | struct hfi1_qp_priv *qpriv = (struct hfi1_qp_priv *)qp->priv; | |
2924 | struct hfi1_swqe_priv *priv = wqe->priv; | |
2925 | struct tid_rdma_params *remote; | |
2926 | enum ib_wr_opcode new_opcode; | |
2927 | bool do_tid_rdma = false; | |
2928 | struct hfi1_pportdata *ppd = qpriv->rcd->ppd; | |
2929 | ||
2930 | if ((rdma_ah_get_dlid(&qp->remote_ah_attr) & ~((1 << ppd->lmc) - 1)) == | |
2931 | ppd->lid) | |
2932 | return; | |
2933 | if (qpriv->hdr_type != HFI1_PKT_TYPE_9B) | |
2934 | return; | |
2935 | ||
2936 | rcu_read_lock(); | |
2937 | remote = rcu_dereference(qpriv->tid_rdma.remote); | |
2938 | /* | |
2939 | * If TID RDMA is disabled by the negotiation, don't | |
2940 | * use it. | |
2941 | */ | |
2942 | if (!remote) | |
2943 | goto exit; | |
2944 | ||
2945 | if (wqe->wr.opcode == IB_WR_RDMA_READ) { | |
3ce5daa2 KW |
2946 | if (hfi1_check_sge_align(qp, &wqe->sg_list[0], |
2947 | wqe->wr.num_sge)) { | |
f1ab4efa KW |
2948 | new_opcode = IB_WR_TID_RDMA_READ; |
2949 | do_tid_rdma = true; | |
2950 | } | |
2951 | } | |
2952 | ||
2953 | if (do_tid_rdma) { | |
2954 | if (hfi1_kern_exp_rcv_alloc_flows(&priv->tid_req, GFP_ATOMIC)) | |
2955 | goto exit; | |
2956 | wqe->wr.opcode = new_opcode; | |
2957 | priv->tid_req.seg_len = | |
2958 | min_t(u32, remote->max_len, wqe->length); | |
2959 | priv->tid_req.total_segs = | |
2960 | DIV_ROUND_UP(wqe->length, priv->tid_req.seg_len); | |
2961 | /* Compute the last PSN of the request */ | |
2962 | wqe->lpsn = wqe->psn; | |
2963 | if (wqe->wr.opcode == IB_WR_TID_RDMA_READ) { | |
2964 | priv->tid_req.n_flows = remote->max_read; | |
2965 | qpriv->tid_r_reqs++; | |
2966 | wqe->lpsn += rvt_div_round_up_mtu(qp, wqe->length) - 1; | |
2967 | } | |
2968 | ||
2969 | priv->tid_req.cur_seg = 0; | |
2970 | priv->tid_req.comp_seg = 0; | |
2971 | priv->tid_req.ack_seg = 0; | |
2972 | priv->tid_req.state = TID_REQUEST_INACTIVE; | |
3ce5daa2 KW |
2973 | trace_hfi1_tid_req_setup_tid_wqe(qp, 1, wqe->wr.opcode, |
2974 | wqe->psn, wqe->lpsn, | |
2975 | &priv->tid_req); | |
f1ab4efa KW |
2976 | } |
2977 | exit: | |
2978 | rcu_read_unlock(); | |
2979 | } | |
c098bbb0 KW |
2980 | |
2981 | /* TID RDMA WRITE functions */ | |
2982 | ||
2983 | u32 hfi1_build_tid_rdma_write_req(struct rvt_qp *qp, struct rvt_swqe *wqe, | |
2984 | struct ib_other_headers *ohdr, | |
2985 | u32 *bth1, u32 *bth2, u32 *len) | |
2986 | { | |
2987 | struct hfi1_qp_priv *qpriv = qp->priv; | |
2988 | struct tid_rdma_request *req = wqe_to_tid_req(wqe); | |
2989 | struct tid_rdma_params *remote; | |
2990 | ||
2991 | rcu_read_lock(); | |
2992 | remote = rcu_dereference(qpriv->tid_rdma.remote); | |
2993 | /* | |
2994 | * Set the number of flow to be used based on negotiated | |
2995 | * parameters. | |
2996 | */ | |
2997 | req->n_flows = remote->max_write; | |
2998 | req->state = TID_REQUEST_ACTIVE; | |
2999 | ||
3000 | KDETH_RESET(ohdr->u.tid_rdma.w_req.kdeth0, KVER, 0x1); | |
3001 | KDETH_RESET(ohdr->u.tid_rdma.w_req.kdeth1, JKEY, remote->jkey); | |
3002 | ohdr->u.tid_rdma.w_req.reth.vaddr = | |
3003 | cpu_to_be64(wqe->rdma_wr.remote_addr + (wqe->length - *len)); | |
3004 | ohdr->u.tid_rdma.w_req.reth.rkey = | |
3005 | cpu_to_be32(wqe->rdma_wr.rkey); | |
3006 | ohdr->u.tid_rdma.w_req.reth.length = cpu_to_be32(*len); | |
3007 | ohdr->u.tid_rdma.w_req.verbs_qp = cpu_to_be32(qp->remote_qpn); | |
3008 | *bth1 &= ~RVT_QPN_MASK; | |
3009 | *bth1 |= remote->qp; | |
3010 | qp->s_state = TID_OP(WRITE_REQ); | |
3011 | qp->s_flags |= HFI1_S_WAIT_TID_RESP; | |
3012 | *bth2 |= IB_BTH_REQ_ACK; | |
3013 | *len = 0; | |
3014 | ||
3015 | rcu_read_unlock(); | |
3016 | return sizeof(ohdr->u.tid_rdma.w_req) / sizeof(u32); | |
3017 | } |