2 * Copyright(c) 2016, 2017 Intel Corporation.
4 * This file is provided under a dual BSD/GPLv2 license. When using or
5 * redistributing this file, you may do so under either license.
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of version 2 of the GNU General Public License as
11 * published by the Free Software Foundation.
13 * This program is distributed in the hope that it will be useful, but
14 * WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * General Public License for more details.
20 * Redistribution and use in source and binary forms, with or without
21 * modification, are permitted provided that the following conditions
24 * - Redistributions of source code must retain the above copyright
25 * notice, this list of conditions and the following disclaimer.
26 * - Redistributions in binary form must reproduce the above copyright
27 * notice, this list of conditions and the following disclaimer in
28 * the documentation and/or other materials provided with the
30 * - Neither the name of Intel Corporation nor the names of its
31 * contributors may be used to endorse or promote products derived
32 * from this software without specific prior written permission.
34 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
48 #include <linux/hash.h>
49 #include <linux/bitops.h>
50 #include <linux/lockdep.h>
51 #include <linux/vmalloc.h>
52 #include <linux/slab.h>
53 #include <rdma/ib_verbs.h>
54 #include <rdma/ib_hdrs.h>
55 #include <rdma/opa_addr.h>
60 static void rvt_rc_timeout(struct timer_list
*t
);
63 * Convert the AETH RNR timeout code into the number of microseconds.
65 static const u32 ib_rvt_rnr_table
[32] = {
66 655360, /* 00: 655.36 */
86 10240, /* 14: 10.24 */
87 15360, /* 15: 15.36 */
88 20480, /* 16: 20.48 */
89 30720, /* 17: 30.72 */
90 40960, /* 18: 40.96 */
91 61440, /* 19: 61.44 */
92 81920, /* 1A: 81.92 */
93 122880, /* 1B: 122.88 */
94 163840, /* 1C: 163.84 */
95 245760, /* 1D: 245.76 */
96 327680, /* 1E: 327.68 */
97 491520 /* 1F: 491.52 */
101 * Note that it is OK to post send work requests in the SQE and ERR
102 * states; rvt_do_send() will process them and generate error
103 * completions as per IB 1.2 C10-96.
105 const int ib_rvt_state_ops
[IB_QPS_ERR
+ 1] = {
107 [IB_QPS_INIT
] = RVT_POST_RECV_OK
,
108 [IB_QPS_RTR
] = RVT_POST_RECV_OK
| RVT_PROCESS_RECV_OK
,
109 [IB_QPS_RTS
] = RVT_POST_RECV_OK
| RVT_PROCESS_RECV_OK
|
110 RVT_POST_SEND_OK
| RVT_PROCESS_SEND_OK
|
111 RVT_PROCESS_NEXT_SEND_OK
,
112 [IB_QPS_SQD
] = RVT_POST_RECV_OK
| RVT_PROCESS_RECV_OK
|
113 RVT_POST_SEND_OK
| RVT_PROCESS_SEND_OK
,
114 [IB_QPS_SQE
] = RVT_POST_RECV_OK
| RVT_PROCESS_RECV_OK
|
115 RVT_POST_SEND_OK
| RVT_FLUSH_SEND
,
116 [IB_QPS_ERR
] = RVT_POST_RECV_OK
| RVT_FLUSH_RECV
|
117 RVT_POST_SEND_OK
| RVT_FLUSH_SEND
,
119 EXPORT_SYMBOL(ib_rvt_state_ops
);
121 static void get_map_page(struct rvt_qpn_table
*qpt
,
122 struct rvt_qpn_map
*map
)
124 unsigned long page
= get_zeroed_page(GFP_KERNEL
);
127 * Free the page if someone raced with us installing it.
130 spin_lock(&qpt
->lock
);
134 map
->page
= (void *)page
;
135 spin_unlock(&qpt
->lock
);
139 * init_qpn_table - initialize the QP number table for a device
140 * @qpt: the QPN table
142 static int init_qpn_table(struct rvt_dev_info
*rdi
, struct rvt_qpn_table
*qpt
)
145 struct rvt_qpn_map
*map
;
148 if (!(rdi
->dparms
.qpn_res_end
>= rdi
->dparms
.qpn_res_start
))
151 spin_lock_init(&qpt
->lock
);
153 qpt
->last
= rdi
->dparms
.qpn_start
;
154 qpt
->incr
= rdi
->dparms
.qpn_inc
<< rdi
->dparms
.qos_shift
;
157 * Drivers may want some QPs beyond what we need for verbs let them use
158 * our qpn table. No need for two. Lets go ahead and mark the bitmaps
159 * for those. The reserved range must be *after* the range which verbs
163 /* Figure out number of bit maps needed before reserved range */
164 qpt
->nmaps
= rdi
->dparms
.qpn_res_start
/ RVT_BITS_PER_PAGE
;
166 /* This should always be zero */
167 offset
= rdi
->dparms
.qpn_res_start
& RVT_BITS_PER_PAGE_MASK
;
169 /* Starting with the first reserved bit map */
170 map
= &qpt
->map
[qpt
->nmaps
];
172 rvt_pr_info(rdi
, "Reserving QPNs from 0x%x to 0x%x for non-verbs use\n",
173 rdi
->dparms
.qpn_res_start
, rdi
->dparms
.qpn_res_end
);
174 for (i
= rdi
->dparms
.qpn_res_start
; i
<= rdi
->dparms
.qpn_res_end
; i
++) {
176 get_map_page(qpt
, map
);
182 set_bit(offset
, map
->page
);
184 if (offset
== RVT_BITS_PER_PAGE
) {
195 * free_qpn_table - free the QP number table for a device
196 * @qpt: the QPN table
198 static void free_qpn_table(struct rvt_qpn_table
*qpt
)
202 for (i
= 0; i
< ARRAY_SIZE(qpt
->map
); i
++)
203 free_page((unsigned long)qpt
->map
[i
].page
);
207 * rvt_driver_qp_init - Init driver qp resources
208 * @rdi: rvt dev strucutre
210 * Return: 0 on success
212 int rvt_driver_qp_init(struct rvt_dev_info
*rdi
)
217 if (!rdi
->dparms
.qp_table_size
)
221 * If driver is not doing any QP allocation then make sure it is
222 * providing the necessary QP functions.
224 if (!rdi
->driver_f
.free_all_qps
||
225 !rdi
->driver_f
.qp_priv_alloc
||
226 !rdi
->driver_f
.qp_priv_free
||
227 !rdi
->driver_f
.notify_qp_reset
||
228 !rdi
->driver_f
.notify_restart_rc
)
231 /* allocate parent object */
232 rdi
->qp_dev
= kzalloc_node(sizeof(*rdi
->qp_dev
), GFP_KERNEL
,
237 /* allocate hash table */
238 rdi
->qp_dev
->qp_table_size
= rdi
->dparms
.qp_table_size
;
239 rdi
->qp_dev
->qp_table_bits
= ilog2(rdi
->dparms
.qp_table_size
);
240 rdi
->qp_dev
->qp_table
=
241 kmalloc_array_node(rdi
->qp_dev
->qp_table_size
,
242 sizeof(*rdi
->qp_dev
->qp_table
),
243 GFP_KERNEL
, rdi
->dparms
.node
);
244 if (!rdi
->qp_dev
->qp_table
)
247 for (i
= 0; i
< rdi
->qp_dev
->qp_table_size
; i
++)
248 RCU_INIT_POINTER(rdi
->qp_dev
->qp_table
[i
], NULL
);
250 spin_lock_init(&rdi
->qp_dev
->qpt_lock
);
252 /* initialize qpn map */
253 if (init_qpn_table(rdi
, &rdi
->qp_dev
->qpn_table
))
256 spin_lock_init(&rdi
->n_qps_lock
);
261 kfree(rdi
->qp_dev
->qp_table
);
262 free_qpn_table(&rdi
->qp_dev
->qpn_table
);
271 * free_all_qps - check for QPs still in use
272 * @rdi: rvt device info structure
274 * There should not be any QPs still in use.
275 * Free memory for table.
277 static unsigned rvt_free_all_qps(struct rvt_dev_info
*rdi
)
281 unsigned n
, qp_inuse
= 0;
282 spinlock_t
*ql
; /* work around too long line below */
284 if (rdi
->driver_f
.free_all_qps
)
285 qp_inuse
= rdi
->driver_f
.free_all_qps(rdi
);
287 qp_inuse
+= rvt_mcast_tree_empty(rdi
);
292 ql
= &rdi
->qp_dev
->qpt_lock
;
293 spin_lock_irqsave(ql
, flags
);
294 for (n
= 0; n
< rdi
->qp_dev
->qp_table_size
; n
++) {
295 qp
= rcu_dereference_protected(rdi
->qp_dev
->qp_table
[n
],
296 lockdep_is_held(ql
));
297 RCU_INIT_POINTER(rdi
->qp_dev
->qp_table
[n
], NULL
);
299 for (; qp
; qp
= rcu_dereference_protected(qp
->next
,
300 lockdep_is_held(ql
)))
303 spin_unlock_irqrestore(ql
, flags
);
309 * rvt_qp_exit - clean up qps on device exit
310 * @rdi: rvt dev structure
312 * Check for qp leaks and free resources.
314 void rvt_qp_exit(struct rvt_dev_info
*rdi
)
316 u32 qps_inuse
= rvt_free_all_qps(rdi
);
319 rvt_pr_err(rdi
, "QP memory leak! %u still in use\n",
324 kfree(rdi
->qp_dev
->qp_table
);
325 free_qpn_table(&rdi
->qp_dev
->qpn_table
);
329 static inline unsigned mk_qpn(struct rvt_qpn_table
*qpt
,
330 struct rvt_qpn_map
*map
, unsigned off
)
332 return (map
- qpt
->map
) * RVT_BITS_PER_PAGE
+ off
;
336 * alloc_qpn - Allocate the next available qpn or zero/one for QP type
337 * IB_QPT_SMI/IB_QPT_GSI
338 * @rdi: rvt device info structure
339 * @qpt: queue pair number table pointer
340 * @port_num: IB port number, 1 based, comes from core
342 * Return: The queue pair number
344 static int alloc_qpn(struct rvt_dev_info
*rdi
, struct rvt_qpn_table
*qpt
,
345 enum ib_qp_type type
, u8 port_num
)
347 u32 i
, offset
, max_scan
, qpn
;
348 struct rvt_qpn_map
*map
;
351 if (rdi
->driver_f
.alloc_qpn
)
352 return rdi
->driver_f
.alloc_qpn(rdi
, qpt
, type
, port_num
);
354 if (type
== IB_QPT_SMI
|| type
== IB_QPT_GSI
) {
357 ret
= type
== IB_QPT_GSI
;
358 n
= 1 << (ret
+ 2 * (port_num
- 1));
359 spin_lock(&qpt
->lock
);
364 spin_unlock(&qpt
->lock
);
368 qpn
= qpt
->last
+ qpt
->incr
;
369 if (qpn
>= RVT_QPN_MAX
)
370 qpn
= qpt
->incr
| ((qpt
->last
& 1) ^ 1);
371 /* offset carries bit 0 */
372 offset
= qpn
& RVT_BITS_PER_PAGE_MASK
;
373 map
= &qpt
->map
[qpn
/ RVT_BITS_PER_PAGE
];
374 max_scan
= qpt
->nmaps
- !offset
;
376 if (unlikely(!map
->page
)) {
377 get_map_page(qpt
, map
);
378 if (unlikely(!map
->page
))
382 if (!test_and_set_bit(offset
, map
->page
)) {
389 * This qpn might be bogus if offset >= BITS_PER_PAGE.
390 * That is OK. It gets re-assigned below
392 qpn
= mk_qpn(qpt
, map
, offset
);
393 } while (offset
< RVT_BITS_PER_PAGE
&& qpn
< RVT_QPN_MAX
);
395 * In order to keep the number of pages allocated to a
396 * minimum, we scan the all existing pages before increasing
397 * the size of the bitmap table.
399 if (++i
> max_scan
) {
400 if (qpt
->nmaps
== RVT_QPNMAP_ENTRIES
)
402 map
= &qpt
->map
[qpt
->nmaps
++];
403 /* start at incr with current bit 0 */
404 offset
= qpt
->incr
| (offset
& 1);
405 } else if (map
< &qpt
->map
[qpt
->nmaps
]) {
407 /* start at incr with current bit 0 */
408 offset
= qpt
->incr
| (offset
& 1);
411 /* wrap to first map page, invert bit 0 */
412 offset
= qpt
->incr
| ((offset
& 1) ^ 1);
414 /* there can be no set bits in low-order QoS bits */
415 WARN_ON(offset
& (BIT(rdi
->dparms
.qos_shift
) - 1));
416 qpn
= mk_qpn(qpt
, map
, offset
);
426 * rvt_clear_mr_refs - Drop help mr refs
427 * @qp: rvt qp data structure
428 * @clr_sends: If shoudl clear send side or not
430 static void rvt_clear_mr_refs(struct rvt_qp
*qp
, int clr_sends
)
433 struct rvt_dev_info
*rdi
= ib_to_rvt(qp
->ibqp
.device
);
435 if (test_and_clear_bit(RVT_R_REWIND_SGE
, &qp
->r_aflags
))
436 rvt_put_ss(&qp
->s_rdma_read_sge
);
438 rvt_put_ss(&qp
->r_sge
);
441 while (qp
->s_last
!= qp
->s_head
) {
442 struct rvt_swqe
*wqe
= rvt_get_swqe_ptr(qp
, qp
->s_last
);
446 if (qp
->ibqp
.qp_type
== IB_QPT_UD
||
447 qp
->ibqp
.qp_type
== IB_QPT_SMI
||
448 qp
->ibqp
.qp_type
== IB_QPT_GSI
)
449 atomic_dec(&ibah_to_rvtah(
450 wqe
->ud_wr
.ah
)->refcount
);
451 if (++qp
->s_last
>= qp
->s_size
)
453 smp_wmb(); /* see qp_set_savail */
456 rvt_put_mr(qp
->s_rdma_mr
);
457 qp
->s_rdma_mr
= NULL
;
461 for (n
= 0; qp
->s_ack_queue
&& n
< rvt_max_atomic(rdi
); n
++) {
462 struct rvt_ack_entry
*e
= &qp
->s_ack_queue
[n
];
464 if (e
->rdma_sge
.mr
) {
465 rvt_put_mr(e
->rdma_sge
.mr
);
466 e
->rdma_sge
.mr
= NULL
;
472 * rvt_swqe_has_lkey - return true if lkey is used by swqe
473 * @wqe - the send wqe
476 * Test the swqe for using lkey
478 static bool rvt_swqe_has_lkey(struct rvt_swqe
*wqe
, u32 lkey
)
482 for (i
= 0; i
< wqe
->wr
.num_sge
; i
++) {
483 struct rvt_sge
*sge
= &wqe
->sg_list
[i
];
485 if (rvt_mr_has_lkey(sge
->mr
, lkey
))
492 * rvt_qp_sends_has_lkey - return true is qp sends use lkey
496 static bool rvt_qp_sends_has_lkey(struct rvt_qp
*qp
, u32 lkey
)
498 u32 s_last
= qp
->s_last
;
500 while (s_last
!= qp
->s_head
) {
501 struct rvt_swqe
*wqe
= rvt_get_swqe_ptr(qp
, s_last
);
503 if (rvt_swqe_has_lkey(wqe
, lkey
))
506 if (++s_last
>= qp
->s_size
)
510 if (rvt_mr_has_lkey(qp
->s_rdma_mr
, lkey
))
516 * rvt_qp_acks_has_lkey - return true if acks have lkey
520 static bool rvt_qp_acks_has_lkey(struct rvt_qp
*qp
, u32 lkey
)
523 struct rvt_dev_info
*rdi
= ib_to_rvt(qp
->ibqp
.device
);
525 for (i
= 0; qp
->s_ack_queue
&& i
< rvt_max_atomic(rdi
); i
++) {
526 struct rvt_ack_entry
*e
= &qp
->s_ack_queue
[i
];
528 if (rvt_mr_has_lkey(e
->rdma_sge
.mr
, lkey
))
535 * rvt_qp_mr_clean - clean up remote ops for lkey
537 * @lkey - the lkey that is being de-registered
539 * This routine checks if the lkey is being used by
542 * If so, the qp is put into an error state to elminate
543 * any references from the qp.
545 void rvt_qp_mr_clean(struct rvt_qp
*qp
, u32 lkey
)
547 bool lastwqe
= false;
549 if (qp
->ibqp
.qp_type
== IB_QPT_SMI
||
550 qp
->ibqp
.qp_type
== IB_QPT_GSI
)
551 /* avoid special QPs */
553 spin_lock_irq(&qp
->r_lock
);
554 spin_lock(&qp
->s_hlock
);
555 spin_lock(&qp
->s_lock
);
557 if (qp
->state
== IB_QPS_ERR
|| qp
->state
== IB_QPS_RESET
)
560 if (rvt_ss_has_lkey(&qp
->r_sge
, lkey
) ||
561 rvt_qp_sends_has_lkey(qp
, lkey
) ||
562 rvt_qp_acks_has_lkey(qp
, lkey
))
563 lastwqe
= rvt_error_qp(qp
, IB_WC_LOC_PROT_ERR
);
565 spin_unlock(&qp
->s_lock
);
566 spin_unlock(&qp
->s_hlock
);
567 spin_unlock_irq(&qp
->r_lock
);
571 ev
.device
= qp
->ibqp
.device
;
572 ev
.element
.qp
= &qp
->ibqp
;
573 ev
.event
= IB_EVENT_QP_LAST_WQE_REACHED
;
574 qp
->ibqp
.event_handler(&ev
, qp
->ibqp
.qp_context
);
579 * rvt_remove_qp - remove qp form table
580 * @rdi: rvt dev struct
583 * Remove the QP from the table so it can't be found asynchronously by
584 * the receive routine.
586 static void rvt_remove_qp(struct rvt_dev_info
*rdi
, struct rvt_qp
*qp
)
588 struct rvt_ibport
*rvp
= rdi
->ports
[qp
->port_num
- 1];
589 u32 n
= hash_32(qp
->ibqp
.qp_num
, rdi
->qp_dev
->qp_table_bits
);
593 spin_lock_irqsave(&rdi
->qp_dev
->qpt_lock
, flags
);
595 if (rcu_dereference_protected(rvp
->qp
[0],
596 lockdep_is_held(&rdi
->qp_dev
->qpt_lock
)) == qp
) {
597 RCU_INIT_POINTER(rvp
->qp
[0], NULL
);
598 } else if (rcu_dereference_protected(rvp
->qp
[1],
599 lockdep_is_held(&rdi
->qp_dev
->qpt_lock
)) == qp
) {
600 RCU_INIT_POINTER(rvp
->qp
[1], NULL
);
603 struct rvt_qp __rcu
**qpp
;
606 qpp
= &rdi
->qp_dev
->qp_table
[n
];
607 for (; (q
= rcu_dereference_protected(*qpp
,
608 lockdep_is_held(&rdi
->qp_dev
->qpt_lock
))) != NULL
;
611 RCU_INIT_POINTER(*qpp
,
612 rcu_dereference_protected(qp
->next
,
613 lockdep_is_held(&rdi
->qp_dev
->qpt_lock
)));
615 trace_rvt_qpremove(qp
, n
);
621 spin_unlock_irqrestore(&rdi
->qp_dev
->qpt_lock
, flags
);
629 * rvt_init_qp - initialize the QP state to the reset state
630 * @qp: the QP to init or reinit
633 * This function is called from both rvt_create_qp() and
634 * rvt_reset_qp(). The difference is that the reset
635 * patch the necessary locks to protect against concurent
638 static void rvt_init_qp(struct rvt_dev_info
*rdi
, struct rvt_qp
*qp
,
639 enum ib_qp_type type
)
643 qp
->qp_access_flags
= 0;
644 qp
->s_flags
&= RVT_S_SIGNAL_REQ_WR
;
650 qp
->s_sending_psn
= 0;
651 qp
->s_sending_hpsn
= 0;
655 if (type
== IB_QPT_RC
) {
656 qp
->s_state
= IB_OPCODE_RC_SEND_LAST
;
657 qp
->r_state
= IB_OPCODE_RC_SEND_LAST
;
659 qp
->s_state
= IB_OPCODE_UC_SEND_LAST
;
660 qp
->r_state
= IB_OPCODE_UC_SEND_LAST
;
662 qp
->s_ack_state
= IB_OPCODE_RC_ACKNOWLEDGE
;
673 qp
->s_mig_state
= IB_MIG_MIGRATED
;
674 qp
->r_head_ack_queue
= 0;
675 qp
->s_tail_ack_queue
= 0;
676 qp
->s_num_rd_atomic
= 0;
678 qp
->r_rq
.wq
->head
= 0;
679 qp
->r_rq
.wq
->tail
= 0;
681 qp
->r_sge
.num_sge
= 0;
682 atomic_set(&qp
->s_reserved_used
, 0);
686 * rvt_reset_qp - initialize the QP state to the reset state
687 * @qp: the QP to reset
690 * r_lock, s_hlock, and s_lock are required to be held by the caller
692 static void rvt_reset_qp(struct rvt_dev_info
*rdi
, struct rvt_qp
*qp
,
693 enum ib_qp_type type
)
694 __must_hold(&qp
->s_lock
)
695 __must_hold(&qp
->s_hlock
)
696 __must_hold(&qp
->r_lock
)
698 lockdep_assert_held(&qp
->r_lock
);
699 lockdep_assert_held(&qp
->s_hlock
);
700 lockdep_assert_held(&qp
->s_lock
);
701 if (qp
->state
!= IB_QPS_RESET
) {
702 qp
->state
= IB_QPS_RESET
;
704 /* Let drivers flush their waitlist */
705 rdi
->driver_f
.flush_qp_waiters(qp
);
706 rvt_stop_rc_timers(qp
);
707 qp
->s_flags
&= ~(RVT_S_TIMER
| RVT_S_ANY_WAIT
);
708 spin_unlock(&qp
->s_lock
);
709 spin_unlock(&qp
->s_hlock
);
710 spin_unlock_irq(&qp
->r_lock
);
712 /* Stop the send queue and the retry timer */
713 rdi
->driver_f
.stop_send_queue(qp
);
714 rvt_del_timers_sync(qp
);
715 /* Wait for things to stop */
716 rdi
->driver_f
.quiesce_qp(qp
);
718 /* take qp out the hash and wait for it to be unused */
719 rvt_remove_qp(rdi
, qp
);
721 /* grab the lock b/c it was locked at call time */
722 spin_lock_irq(&qp
->r_lock
);
723 spin_lock(&qp
->s_hlock
);
724 spin_lock(&qp
->s_lock
);
726 rvt_clear_mr_refs(qp
, 1);
728 * Let the driver do any tear down or re-init it needs to for
729 * a qp that has been reset
731 rdi
->driver_f
.notify_qp_reset(qp
);
733 rvt_init_qp(rdi
, qp
, type
);
734 lockdep_assert_held(&qp
->r_lock
);
735 lockdep_assert_held(&qp
->s_hlock
);
736 lockdep_assert_held(&qp
->s_lock
);
739 /** rvt_free_qpn - Free a qpn from the bit map
741 * @qpn: queue pair number to free
743 static void rvt_free_qpn(struct rvt_qpn_table
*qpt
, u32 qpn
)
745 struct rvt_qpn_map
*map
;
747 map
= qpt
->map
+ (qpn
& RVT_QPN_MASK
) / RVT_BITS_PER_PAGE
;
749 clear_bit(qpn
& RVT_BITS_PER_PAGE_MASK
, map
->page
);
753 * rvt_create_qp - create a queue pair for a device
754 * @ibpd: the protection domain who's device we create the queue pair for
755 * @init_attr: the attributes of the queue pair
756 * @udata: user data for libibverbs.so
758 * Queue pair creation is mostly an rvt issue. However, drivers have their own
759 * unique idea of what queue pair numbers mean. For instance there is a reserved
762 * Return: the queue pair on success, otherwise returns an errno.
764 * Called by the ib_create_qp() core verbs function.
766 struct ib_qp
*rvt_create_qp(struct ib_pd
*ibpd
,
767 struct ib_qp_init_attr
*init_attr
,
768 struct ib_udata
*udata
)
772 struct rvt_swqe
*swq
= NULL
;
775 struct ib_qp
*ret
= ERR_PTR(-ENOMEM
);
776 struct rvt_dev_info
*rdi
= ib_to_rvt(ibpd
->device
);
781 return ERR_PTR(-EINVAL
);
783 if (init_attr
->cap
.max_send_sge
> rdi
->dparms
.props
.max_sge
||
784 init_attr
->cap
.max_send_wr
> rdi
->dparms
.props
.max_qp_wr
||
785 init_attr
->create_flags
)
786 return ERR_PTR(-EINVAL
);
788 /* Check receive queue parameters if no SRQ is specified. */
789 if (!init_attr
->srq
) {
790 if (init_attr
->cap
.max_recv_sge
> rdi
->dparms
.props
.max_sge
||
791 init_attr
->cap
.max_recv_wr
> rdi
->dparms
.props
.max_qp_wr
)
792 return ERR_PTR(-EINVAL
);
794 if (init_attr
->cap
.max_send_sge
+
795 init_attr
->cap
.max_send_wr
+
796 init_attr
->cap
.max_recv_sge
+
797 init_attr
->cap
.max_recv_wr
== 0)
798 return ERR_PTR(-EINVAL
);
801 init_attr
->cap
.max_send_wr
+ 1 +
802 rdi
->dparms
.reserved_operations
;
803 switch (init_attr
->qp_type
) {
806 if (init_attr
->port_num
== 0 ||
807 init_attr
->port_num
> ibpd
->device
->phys_port_cnt
)
808 return ERR_PTR(-EINVAL
);
813 sz
= sizeof(struct rvt_sge
) *
814 init_attr
->cap
.max_send_sge
+
815 sizeof(struct rvt_swqe
);
816 swq
= vzalloc_node(array_size(sz
, sqsize
), rdi
->dparms
.node
);
818 return ERR_PTR(-ENOMEM
);
822 if (init_attr
->srq
) {
823 struct rvt_srq
*srq
= ibsrq_to_rvtsrq(init_attr
->srq
);
825 if (srq
->rq
.max_sge
> 1)
826 sg_list_sz
= sizeof(*qp
->r_sg_list
) *
827 (srq
->rq
.max_sge
- 1);
828 } else if (init_attr
->cap
.max_recv_sge
> 1)
829 sg_list_sz
= sizeof(*qp
->r_sg_list
) *
830 (init_attr
->cap
.max_recv_sge
- 1);
831 qp
= kzalloc_node(sz
+ sg_list_sz
, GFP_KERNEL
,
836 RCU_INIT_POINTER(qp
->next
, NULL
);
837 if (init_attr
->qp_type
== IB_QPT_RC
) {
839 kcalloc_node(rvt_max_atomic(rdi
),
840 sizeof(*qp
->s_ack_queue
),
843 if (!qp
->s_ack_queue
)
846 /* initialize timers needed for rc qp */
847 timer_setup(&qp
->s_timer
, rvt_rc_timeout
, 0);
848 hrtimer_init(&qp
->s_rnr_timer
, CLOCK_MONOTONIC
,
850 qp
->s_rnr_timer
.function
= rvt_rc_rnr_retry
;
853 * Driver needs to set up it's private QP structure and do any
854 * initialization that is needed.
856 priv
= rdi
->driver_f
.qp_priv_alloc(rdi
, qp
);
862 qp
->timeout_jiffies
=
863 usecs_to_jiffies((4096UL * (1UL << qp
->timeout
)) /
865 if (init_attr
->srq
) {
868 qp
->r_rq
.size
= init_attr
->cap
.max_recv_wr
+ 1;
869 qp
->r_rq
.max_sge
= init_attr
->cap
.max_recv_sge
;
870 sz
= (sizeof(struct ib_sge
) * qp
->r_rq
.max_sge
) +
871 sizeof(struct rvt_rwqe
);
873 qp
->r_rq
.wq
= vmalloc_user(
874 sizeof(struct rvt_rwq
) +
877 qp
->r_rq
.wq
= vzalloc_node(
878 sizeof(struct rvt_rwq
) +
882 goto bail_driver_priv
;
886 * ib_create_qp() will initialize qp->ibqp
887 * except for qp->ibqp.qp_num.
889 spin_lock_init(&qp
->r_lock
);
890 spin_lock_init(&qp
->s_hlock
);
891 spin_lock_init(&qp
->s_lock
);
892 spin_lock_init(&qp
->r_rq
.lock
);
893 atomic_set(&qp
->refcount
, 0);
894 atomic_set(&qp
->local_ops_pending
, 0);
895 init_waitqueue_head(&qp
->wait
);
896 INIT_LIST_HEAD(&qp
->rspwait
);
897 qp
->state
= IB_QPS_RESET
;
900 qp
->s_avail
= init_attr
->cap
.max_send_wr
;
901 qp
->s_max_sge
= init_attr
->cap
.max_send_sge
;
902 if (init_attr
->sq_sig_type
== IB_SIGNAL_REQ_WR
)
903 qp
->s_flags
= RVT_S_SIGNAL_REQ_WR
;
905 err
= alloc_qpn(rdi
, &rdi
->qp_dev
->qpn_table
,
907 init_attr
->port_num
);
912 qp
->ibqp
.qp_num
= err
;
913 qp
->port_num
= init_attr
->port_num
;
914 rvt_init_qp(rdi
, qp
, init_attr
->qp_type
);
918 /* Don't support raw QPs */
919 return ERR_PTR(-EINVAL
);
922 init_attr
->cap
.max_inline_data
= 0;
925 * Return the address of the RWQ as the offset to mmap.
926 * See rvt_mmap() for details.
928 if (udata
&& udata
->outlen
>= sizeof(__u64
)) {
932 err
= ib_copy_to_udata(udata
, &offset
,
939 u32 s
= sizeof(struct rvt_rwq
) + qp
->r_rq
.size
* sz
;
941 qp
->ip
= rvt_create_mmap_info(rdi
, s
,
942 ibpd
->uobject
->context
,
945 ret
= ERR_PTR(-ENOMEM
);
949 err
= ib_copy_to_udata(udata
, &qp
->ip
->offset
,
950 sizeof(qp
->ip
->offset
));
956 qp
->pid
= current
->pid
;
959 spin_lock(&rdi
->n_qps_lock
);
960 if (rdi
->n_qps_allocated
== rdi
->dparms
.props
.max_qp
) {
961 spin_unlock(&rdi
->n_qps_lock
);
962 ret
= ERR_PTR(-ENOMEM
);
966 rdi
->n_qps_allocated
++;
968 * Maintain a busy_jiffies variable that will be added to the timeout
969 * period in mod_retry_timer and add_retry_timer. This busy jiffies
970 * is scaled by the number of rc qps created for the device to reduce
971 * the number of timeouts occurring when there is a large number of
972 * qps. busy_jiffies is incremented every rc qp scaling interval.
973 * The scaling interval is selected based on extensive performance
974 * evaluation of targeted workloads.
976 if (init_attr
->qp_type
== IB_QPT_RC
) {
978 rdi
->busy_jiffies
= rdi
->n_rc_qps
/ RC_QP_SCALING_INTERVAL
;
980 spin_unlock(&rdi
->n_qps_lock
);
983 spin_lock_irq(&rdi
->pending_lock
);
984 list_add(&qp
->ip
->pending_mmaps
, &rdi
->pending_mmaps
);
985 spin_unlock_irq(&rdi
->pending_lock
);
991 * We have our QP and its good, now keep track of what types of opcodes
992 * can be processed on this QP. We do this by keeping track of what the
993 * 3 high order bits of the opcode are.
995 switch (init_attr
->qp_type
) {
999 qp
->allowed_ops
= IB_OPCODE_UD
;
1002 qp
->allowed_ops
= IB_OPCODE_RC
;
1005 qp
->allowed_ops
= IB_OPCODE_UC
;
1008 ret
= ERR_PTR(-EINVAL
);
1016 kref_put(&qp
->ip
->ref
, rvt_release_mmap_info
);
1019 rvt_free_qpn(&rdi
->qp_dev
->qpn_table
, qp
->ibqp
.qp_num
);
1026 rdi
->driver_f
.qp_priv_free(rdi
, qp
);
1029 kfree(qp
->s_ack_queue
);
1039 * rvt_error_qp - put a QP into the error state
1040 * @qp: the QP to put into the error state
1041 * @err: the receive completion error to signal if a RWQE is active
1043 * Flushes both send and receive work queues.
1045 * Return: true if last WQE event should be generated.
1046 * The QP r_lock and s_lock should be held and interrupts disabled.
1047 * If we are already in error state, just return.
1049 int rvt_error_qp(struct rvt_qp
*qp
, enum ib_wc_status err
)
1053 struct rvt_dev_info
*rdi
= ib_to_rvt(qp
->ibqp
.device
);
1055 lockdep_assert_held(&qp
->r_lock
);
1056 lockdep_assert_held(&qp
->s_lock
);
1057 if (qp
->state
== IB_QPS_ERR
|| qp
->state
== IB_QPS_RESET
)
1060 qp
->state
= IB_QPS_ERR
;
1062 if (qp
->s_flags
& (RVT_S_TIMER
| RVT_S_WAIT_RNR
)) {
1063 qp
->s_flags
&= ~(RVT_S_TIMER
| RVT_S_WAIT_RNR
);
1064 del_timer(&qp
->s_timer
);
1067 if (qp
->s_flags
& RVT_S_ANY_WAIT_SEND
)
1068 qp
->s_flags
&= ~RVT_S_ANY_WAIT_SEND
;
1070 rdi
->driver_f
.notify_error_qp(qp
);
1072 /* Schedule the sending tasklet to drain the send work queue. */
1073 if (READ_ONCE(qp
->s_last
) != qp
->s_head
)
1074 rdi
->driver_f
.schedule_send(qp
);
1076 rvt_clear_mr_refs(qp
, 0);
1078 memset(&wc
, 0, sizeof(wc
));
1080 wc
.opcode
= IB_WC_RECV
;
1082 if (test_and_clear_bit(RVT_R_WRID_VALID
, &qp
->r_aflags
)) {
1083 wc
.wr_id
= qp
->r_wr_id
;
1085 rvt_cq_enter(ibcq_to_rvtcq(qp
->ibqp
.recv_cq
), &wc
, 1);
1087 wc
.status
= IB_WC_WR_FLUSH_ERR
;
1094 spin_lock(&qp
->r_rq
.lock
);
1096 /* sanity check pointers before trusting them */
1099 if (head
>= qp
->r_rq
.size
)
1102 if (tail
>= qp
->r_rq
.size
)
1104 while (tail
!= head
) {
1105 wc
.wr_id
= rvt_get_rwqe_ptr(&qp
->r_rq
, tail
)->wr_id
;
1106 if (++tail
>= qp
->r_rq
.size
)
1108 rvt_cq_enter(ibcq_to_rvtcq(qp
->ibqp
.recv_cq
), &wc
, 1);
1112 spin_unlock(&qp
->r_rq
.lock
);
1113 } else if (qp
->ibqp
.event_handler
) {
1120 EXPORT_SYMBOL(rvt_error_qp
);
1123 * Put the QP into the hash table.
1124 * The hash table holds a reference to the QP.
1126 static void rvt_insert_qp(struct rvt_dev_info
*rdi
, struct rvt_qp
*qp
)
1128 struct rvt_ibport
*rvp
= rdi
->ports
[qp
->port_num
- 1];
1129 unsigned long flags
;
1132 spin_lock_irqsave(&rdi
->qp_dev
->qpt_lock
, flags
);
1134 if (qp
->ibqp
.qp_num
<= 1) {
1135 rcu_assign_pointer(rvp
->qp
[qp
->ibqp
.qp_num
], qp
);
1137 u32 n
= hash_32(qp
->ibqp
.qp_num
, rdi
->qp_dev
->qp_table_bits
);
1139 qp
->next
= rdi
->qp_dev
->qp_table
[n
];
1140 rcu_assign_pointer(rdi
->qp_dev
->qp_table
[n
], qp
);
1141 trace_rvt_qpinsert(qp
, n
);
1144 spin_unlock_irqrestore(&rdi
->qp_dev
->qpt_lock
, flags
);
1148 * rvt_modify_qp - modify the attributes of a queue pair
1149 * @ibqp: the queue pair who's attributes we're modifying
1150 * @attr: the new attributes
1151 * @attr_mask: the mask of attributes to modify
1152 * @udata: user data for libibverbs.so
1154 * Return: 0 on success, otherwise returns an errno.
1156 int rvt_modify_qp(struct ib_qp
*ibqp
, struct ib_qp_attr
*attr
,
1157 int attr_mask
, struct ib_udata
*udata
)
1159 struct rvt_dev_info
*rdi
= ib_to_rvt(ibqp
->device
);
1160 struct rvt_qp
*qp
= ibqp_to_rvtqp(ibqp
);
1161 enum ib_qp_state cur_state
, new_state
;
1165 int pmtu
= 0; /* for gcc warning only */
1166 enum rdma_link_layer link
;
1169 link
= rdma_port_get_link_layer(ibqp
->device
, qp
->port_num
);
1171 spin_lock_irq(&qp
->r_lock
);
1172 spin_lock(&qp
->s_hlock
);
1173 spin_lock(&qp
->s_lock
);
1175 cur_state
= attr_mask
& IB_QP_CUR_STATE
?
1176 attr
->cur_qp_state
: qp
->state
;
1177 new_state
= attr_mask
& IB_QP_STATE
? attr
->qp_state
: cur_state
;
1178 opa_ah
= rdma_cap_opa_ah(ibqp
->device
, qp
->port_num
);
1180 if (!ib_modify_qp_is_ok(cur_state
, new_state
, ibqp
->qp_type
,
1184 if (rdi
->driver_f
.check_modify_qp
&&
1185 rdi
->driver_f
.check_modify_qp(qp
, attr
, attr_mask
, udata
))
1188 if (attr_mask
& IB_QP_AV
) {
1190 if (rdma_ah_get_dlid(&attr
->ah_attr
) >=
1191 opa_get_mcast_base(OPA_MCAST_NR
))
1194 if (rdma_ah_get_dlid(&attr
->ah_attr
) >=
1195 be16_to_cpu(IB_MULTICAST_LID_BASE
))
1199 if (rvt_check_ah(qp
->ibqp
.device
, &attr
->ah_attr
))
1203 if (attr_mask
& IB_QP_ALT_PATH
) {
1205 if (rdma_ah_get_dlid(&attr
->alt_ah_attr
) >=
1206 opa_get_mcast_base(OPA_MCAST_NR
))
1209 if (rdma_ah_get_dlid(&attr
->alt_ah_attr
) >=
1210 be16_to_cpu(IB_MULTICAST_LID_BASE
))
1214 if (rvt_check_ah(qp
->ibqp
.device
, &attr
->alt_ah_attr
))
1216 if (attr
->alt_pkey_index
>= rvt_get_npkeys(rdi
))
1220 if (attr_mask
& IB_QP_PKEY_INDEX
)
1221 if (attr
->pkey_index
>= rvt_get_npkeys(rdi
))
1224 if (attr_mask
& IB_QP_MIN_RNR_TIMER
)
1225 if (attr
->min_rnr_timer
> 31)
1228 if (attr_mask
& IB_QP_PORT
)
1229 if (qp
->ibqp
.qp_type
== IB_QPT_SMI
||
1230 qp
->ibqp
.qp_type
== IB_QPT_GSI
||
1231 attr
->port_num
== 0 ||
1232 attr
->port_num
> ibqp
->device
->phys_port_cnt
)
1235 if (attr_mask
& IB_QP_DEST_QPN
)
1236 if (attr
->dest_qp_num
> RVT_QPN_MASK
)
1239 if (attr_mask
& IB_QP_RETRY_CNT
)
1240 if (attr
->retry_cnt
> 7)
1243 if (attr_mask
& IB_QP_RNR_RETRY
)
1244 if (attr
->rnr_retry
> 7)
1248 * Don't allow invalid path_mtu values. OK to set greater
1249 * than the active mtu (or even the max_cap, if we have tuned
1250 * that to a small mtu. We'll set qp->path_mtu
1251 * to the lesser of requested attribute mtu and active,
1252 * for packetizing messages.
1253 * Note that the QP port has to be set in INIT and MTU in RTR.
1255 if (attr_mask
& IB_QP_PATH_MTU
) {
1256 pmtu
= rdi
->driver_f
.get_pmtu_from_attr(rdi
, qp
, attr
);
1261 if (attr_mask
& IB_QP_PATH_MIG_STATE
) {
1262 if (attr
->path_mig_state
== IB_MIG_REARM
) {
1263 if (qp
->s_mig_state
== IB_MIG_ARMED
)
1265 if (new_state
!= IB_QPS_RTS
)
1267 } else if (attr
->path_mig_state
== IB_MIG_MIGRATED
) {
1268 if (qp
->s_mig_state
== IB_MIG_REARM
)
1270 if (new_state
!= IB_QPS_RTS
&& new_state
!= IB_QPS_SQD
)
1272 if (qp
->s_mig_state
== IB_MIG_ARMED
)
1279 if (attr_mask
& IB_QP_MAX_DEST_RD_ATOMIC
)
1280 if (attr
->max_dest_rd_atomic
> rdi
->dparms
.max_rdma_atomic
)
1283 switch (new_state
) {
1285 if (qp
->state
!= IB_QPS_RESET
)
1286 rvt_reset_qp(rdi
, qp
, ibqp
->qp_type
);
1290 /* Allow event to re-trigger if QP set to RTR more than once */
1291 qp
->r_flags
&= ~RVT_R_COMM_EST
;
1292 qp
->state
= new_state
;
1296 qp
->s_draining
= qp
->s_last
!= qp
->s_cur
;
1297 qp
->state
= new_state
;
1301 if (qp
->ibqp
.qp_type
== IB_QPT_RC
)
1303 qp
->state
= new_state
;
1307 lastwqe
= rvt_error_qp(qp
, IB_WC_WR_FLUSH_ERR
);
1311 qp
->state
= new_state
;
1315 if (attr_mask
& IB_QP_PKEY_INDEX
)
1316 qp
->s_pkey_index
= attr
->pkey_index
;
1318 if (attr_mask
& IB_QP_PORT
)
1319 qp
->port_num
= attr
->port_num
;
1321 if (attr_mask
& IB_QP_DEST_QPN
)
1322 qp
->remote_qpn
= attr
->dest_qp_num
;
1324 if (attr_mask
& IB_QP_SQ_PSN
) {
1325 qp
->s_next_psn
= attr
->sq_psn
& rdi
->dparms
.psn_modify_mask
;
1326 qp
->s_psn
= qp
->s_next_psn
;
1327 qp
->s_sending_psn
= qp
->s_next_psn
;
1328 qp
->s_last_psn
= qp
->s_next_psn
- 1;
1329 qp
->s_sending_hpsn
= qp
->s_last_psn
;
1332 if (attr_mask
& IB_QP_RQ_PSN
)
1333 qp
->r_psn
= attr
->rq_psn
& rdi
->dparms
.psn_modify_mask
;
1335 if (attr_mask
& IB_QP_ACCESS_FLAGS
)
1336 qp
->qp_access_flags
= attr
->qp_access_flags
;
1338 if (attr_mask
& IB_QP_AV
) {
1339 qp
->remote_ah_attr
= attr
->ah_attr
;
1340 qp
->s_srate
= rdma_ah_get_static_rate(&attr
->ah_attr
);
1341 qp
->srate_mbps
= ib_rate_to_mbps(qp
->s_srate
);
1344 if (attr_mask
& IB_QP_ALT_PATH
) {
1345 qp
->alt_ah_attr
= attr
->alt_ah_attr
;
1346 qp
->s_alt_pkey_index
= attr
->alt_pkey_index
;
1349 if (attr_mask
& IB_QP_PATH_MIG_STATE
) {
1350 qp
->s_mig_state
= attr
->path_mig_state
;
1352 qp
->remote_ah_attr
= qp
->alt_ah_attr
;
1353 qp
->port_num
= rdma_ah_get_port_num(&qp
->alt_ah_attr
);
1354 qp
->s_pkey_index
= qp
->s_alt_pkey_index
;
1358 if (attr_mask
& IB_QP_PATH_MTU
) {
1359 qp
->pmtu
= rdi
->driver_f
.mtu_from_qp(rdi
, qp
, pmtu
);
1360 qp
->log_pmtu
= ilog2(qp
->pmtu
);
1363 if (attr_mask
& IB_QP_RETRY_CNT
) {
1364 qp
->s_retry_cnt
= attr
->retry_cnt
;
1365 qp
->s_retry
= attr
->retry_cnt
;
1368 if (attr_mask
& IB_QP_RNR_RETRY
) {
1369 qp
->s_rnr_retry_cnt
= attr
->rnr_retry
;
1370 qp
->s_rnr_retry
= attr
->rnr_retry
;
1373 if (attr_mask
& IB_QP_MIN_RNR_TIMER
)
1374 qp
->r_min_rnr_timer
= attr
->min_rnr_timer
;
1376 if (attr_mask
& IB_QP_TIMEOUT
) {
1377 qp
->timeout
= attr
->timeout
;
1378 qp
->timeout_jiffies
= rvt_timeout_to_jiffies(qp
->timeout
);
1381 if (attr_mask
& IB_QP_QKEY
)
1382 qp
->qkey
= attr
->qkey
;
1384 if (attr_mask
& IB_QP_MAX_DEST_RD_ATOMIC
)
1385 qp
->r_max_rd_atomic
= attr
->max_dest_rd_atomic
;
1387 if (attr_mask
& IB_QP_MAX_QP_RD_ATOMIC
)
1388 qp
->s_max_rd_atomic
= attr
->max_rd_atomic
;
1390 if (rdi
->driver_f
.modify_qp
)
1391 rdi
->driver_f
.modify_qp(qp
, attr
, attr_mask
, udata
);
1393 spin_unlock(&qp
->s_lock
);
1394 spin_unlock(&qp
->s_hlock
);
1395 spin_unlock_irq(&qp
->r_lock
);
1397 if (cur_state
== IB_QPS_RESET
&& new_state
== IB_QPS_INIT
)
1398 rvt_insert_qp(rdi
, qp
);
1401 ev
.device
= qp
->ibqp
.device
;
1402 ev
.element
.qp
= &qp
->ibqp
;
1403 ev
.event
= IB_EVENT_QP_LAST_WQE_REACHED
;
1404 qp
->ibqp
.event_handler(&ev
, qp
->ibqp
.qp_context
);
1407 ev
.device
= qp
->ibqp
.device
;
1408 ev
.element
.qp
= &qp
->ibqp
;
1409 ev
.event
= IB_EVENT_PATH_MIG
;
1410 qp
->ibqp
.event_handler(&ev
, qp
->ibqp
.qp_context
);
1415 spin_unlock(&qp
->s_lock
);
1416 spin_unlock(&qp
->s_hlock
);
1417 spin_unlock_irq(&qp
->r_lock
);
1422 * rvt_destroy_qp - destroy a queue pair
1423 * @ibqp: the queue pair to destroy
1425 * Note that this can be called while the QP is actively sending or
1428 * Return: 0 on success.
1430 int rvt_destroy_qp(struct ib_qp
*ibqp
)
1432 struct rvt_qp
*qp
= ibqp_to_rvtqp(ibqp
);
1433 struct rvt_dev_info
*rdi
= ib_to_rvt(ibqp
->device
);
1435 spin_lock_irq(&qp
->r_lock
);
1436 spin_lock(&qp
->s_hlock
);
1437 spin_lock(&qp
->s_lock
);
1438 rvt_reset_qp(rdi
, qp
, ibqp
->qp_type
);
1439 spin_unlock(&qp
->s_lock
);
1440 spin_unlock(&qp
->s_hlock
);
1441 spin_unlock_irq(&qp
->r_lock
);
1443 wait_event(qp
->wait
, !atomic_read(&qp
->refcount
));
1444 /* qpn is now available for use again */
1445 rvt_free_qpn(&rdi
->qp_dev
->qpn_table
, qp
->ibqp
.qp_num
);
1447 spin_lock(&rdi
->n_qps_lock
);
1448 rdi
->n_qps_allocated
--;
1449 if (qp
->ibqp
.qp_type
== IB_QPT_RC
) {
1451 rdi
->busy_jiffies
= rdi
->n_rc_qps
/ RC_QP_SCALING_INTERVAL
;
1453 spin_unlock(&rdi
->n_qps_lock
);
1456 kref_put(&qp
->ip
->ref
, rvt_release_mmap_info
);
1460 rdi
->driver_f
.qp_priv_free(rdi
, qp
);
1461 kfree(qp
->s_ack_queue
);
1467 * rvt_query_qp - query an ipbq
1468 * @ibqp: IB qp to query
1469 * @attr: attr struct to fill in
1470 * @attr_mask: attr mask ignored
1471 * @init_attr: struct to fill in
1475 int rvt_query_qp(struct ib_qp
*ibqp
, struct ib_qp_attr
*attr
,
1476 int attr_mask
, struct ib_qp_init_attr
*init_attr
)
1478 struct rvt_qp
*qp
= ibqp_to_rvtqp(ibqp
);
1479 struct rvt_dev_info
*rdi
= ib_to_rvt(ibqp
->device
);
1481 attr
->qp_state
= qp
->state
;
1482 attr
->cur_qp_state
= attr
->qp_state
;
1483 attr
->path_mtu
= rdi
->driver_f
.mtu_to_path_mtu(qp
->pmtu
);
1484 attr
->path_mig_state
= qp
->s_mig_state
;
1485 attr
->qkey
= qp
->qkey
;
1486 attr
->rq_psn
= qp
->r_psn
& rdi
->dparms
.psn_mask
;
1487 attr
->sq_psn
= qp
->s_next_psn
& rdi
->dparms
.psn_mask
;
1488 attr
->dest_qp_num
= qp
->remote_qpn
;
1489 attr
->qp_access_flags
= qp
->qp_access_flags
;
1490 attr
->cap
.max_send_wr
= qp
->s_size
- 1 -
1491 rdi
->dparms
.reserved_operations
;
1492 attr
->cap
.max_recv_wr
= qp
->ibqp
.srq
? 0 : qp
->r_rq
.size
- 1;
1493 attr
->cap
.max_send_sge
= qp
->s_max_sge
;
1494 attr
->cap
.max_recv_sge
= qp
->r_rq
.max_sge
;
1495 attr
->cap
.max_inline_data
= 0;
1496 attr
->ah_attr
= qp
->remote_ah_attr
;
1497 attr
->alt_ah_attr
= qp
->alt_ah_attr
;
1498 attr
->pkey_index
= qp
->s_pkey_index
;
1499 attr
->alt_pkey_index
= qp
->s_alt_pkey_index
;
1500 attr
->en_sqd_async_notify
= 0;
1501 attr
->sq_draining
= qp
->s_draining
;
1502 attr
->max_rd_atomic
= qp
->s_max_rd_atomic
;
1503 attr
->max_dest_rd_atomic
= qp
->r_max_rd_atomic
;
1504 attr
->min_rnr_timer
= qp
->r_min_rnr_timer
;
1505 attr
->port_num
= qp
->port_num
;
1506 attr
->timeout
= qp
->timeout
;
1507 attr
->retry_cnt
= qp
->s_retry_cnt
;
1508 attr
->rnr_retry
= qp
->s_rnr_retry_cnt
;
1509 attr
->alt_port_num
=
1510 rdma_ah_get_port_num(&qp
->alt_ah_attr
);
1511 attr
->alt_timeout
= qp
->alt_timeout
;
1513 init_attr
->event_handler
= qp
->ibqp
.event_handler
;
1514 init_attr
->qp_context
= qp
->ibqp
.qp_context
;
1515 init_attr
->send_cq
= qp
->ibqp
.send_cq
;
1516 init_attr
->recv_cq
= qp
->ibqp
.recv_cq
;
1517 init_attr
->srq
= qp
->ibqp
.srq
;
1518 init_attr
->cap
= attr
->cap
;
1519 if (qp
->s_flags
& RVT_S_SIGNAL_REQ_WR
)
1520 init_attr
->sq_sig_type
= IB_SIGNAL_REQ_WR
;
1522 init_attr
->sq_sig_type
= IB_SIGNAL_ALL_WR
;
1523 init_attr
->qp_type
= qp
->ibqp
.qp_type
;
1524 init_attr
->port_num
= qp
->port_num
;
1529 * rvt_post_receive - post a receive on a QP
1530 * @ibqp: the QP to post the receive on
1531 * @wr: the WR to post
1532 * @bad_wr: the first bad WR is put here
1534 * This may be called from interrupt context.
1536 * Return: 0 on success otherwise errno
1538 int rvt_post_recv(struct ib_qp
*ibqp
, struct ib_recv_wr
*wr
,
1539 struct ib_recv_wr
**bad_wr
)
1541 struct rvt_qp
*qp
= ibqp_to_rvtqp(ibqp
);
1542 struct rvt_rwq
*wq
= qp
->r_rq
.wq
;
1543 unsigned long flags
;
1544 int qp_err_flush
= (ib_rvt_state_ops
[qp
->state
] & RVT_FLUSH_RECV
) &&
1547 /* Check that state is OK to post receive. */
1548 if (!(ib_rvt_state_ops
[qp
->state
] & RVT_POST_RECV_OK
) || !wq
) {
1553 for (; wr
; wr
= wr
->next
) {
1554 struct rvt_rwqe
*wqe
;
1558 if ((unsigned)wr
->num_sge
> qp
->r_rq
.max_sge
) {
1563 spin_lock_irqsave(&qp
->r_rq
.lock
, flags
);
1564 next
= wq
->head
+ 1;
1565 if (next
>= qp
->r_rq
.size
)
1567 if (next
== wq
->tail
) {
1568 spin_unlock_irqrestore(&qp
->r_rq
.lock
, flags
);
1572 if (unlikely(qp_err_flush
)) {
1575 memset(&wc
, 0, sizeof(wc
));
1577 wc
.opcode
= IB_WC_RECV
;
1578 wc
.wr_id
= wr
->wr_id
;
1579 wc
.status
= IB_WC_WR_FLUSH_ERR
;
1580 rvt_cq_enter(ibcq_to_rvtcq(qp
->ibqp
.recv_cq
), &wc
, 1);
1582 wqe
= rvt_get_rwqe_ptr(&qp
->r_rq
, wq
->head
);
1583 wqe
->wr_id
= wr
->wr_id
;
1584 wqe
->num_sge
= wr
->num_sge
;
1585 for (i
= 0; i
< wr
->num_sge
; i
++)
1586 wqe
->sg_list
[i
] = wr
->sg_list
[i
];
1588 * Make sure queue entry is written
1589 * before the head index.
1594 spin_unlock_irqrestore(&qp
->r_rq
.lock
, flags
);
1600 * rvt_qp_valid_operation - validate post send wr request
1602 * @post-parms - the post send table for the driver
1603 * @wr - the work request
1605 * The routine validates the operation based on the
1606 * validation table an returns the length of the operation
1607 * which can extend beyond the ib_send_bw. Operation
1608 * dependent flags key atomic operation validation.
1610 * There is an exception for UD qps that validates the pd and
1611 * overrides the length to include the additional UD specific
1614 * Returns a negative error or the length of the work request
1615 * for building the swqe.
1617 static inline int rvt_qp_valid_operation(
1619 const struct rvt_operation_params
*post_parms
,
1620 struct ib_send_wr
*wr
)
1624 if (wr
->opcode
>= RVT_OPERATION_MAX
|| !post_parms
[wr
->opcode
].length
)
1626 if (!(post_parms
[wr
->opcode
].qpt_support
& BIT(qp
->ibqp
.qp_type
)))
1628 if ((post_parms
[wr
->opcode
].flags
& RVT_OPERATION_PRIV
) &&
1629 ibpd_to_rvtpd(qp
->ibqp
.pd
)->user
)
1631 if (post_parms
[wr
->opcode
].flags
& RVT_OPERATION_ATOMIC_SGE
&&
1632 (wr
->num_sge
== 0 ||
1633 wr
->sg_list
[0].length
< sizeof(u64
) ||
1634 wr
->sg_list
[0].addr
& (sizeof(u64
) - 1)))
1636 if (post_parms
[wr
->opcode
].flags
& RVT_OPERATION_ATOMIC
&&
1637 !qp
->s_max_rd_atomic
)
1639 len
= post_parms
[wr
->opcode
].length
;
1641 if (qp
->ibqp
.qp_type
!= IB_QPT_UC
&&
1642 qp
->ibqp
.qp_type
!= IB_QPT_RC
) {
1643 if (qp
->ibqp
.pd
!= ud_wr(wr
)->ah
->pd
)
1645 len
= sizeof(struct ib_ud_wr
);
1651 * rvt_qp_is_avail - determine queue capacity
1653 * @rdi: the rdmavt device
1654 * @reserved_op: is reserved operation
1656 * This assumes the s_hlock is held but the s_last
1657 * qp variable is uncontrolled.
1659 * For non reserved operations, the qp->s_avail
1662 * The return value is zero or a -ENOMEM.
1664 static inline int rvt_qp_is_avail(
1666 struct rvt_dev_info
*rdi
,
1673 /* see rvt_qp_wqe_unreserve() */
1674 smp_mb__before_atomic();
1675 reserved_used
= atomic_read(&qp
->s_reserved_used
);
1676 if (unlikely(reserved_op
)) {
1677 /* see rvt_qp_wqe_unreserve() */
1678 smp_mb__before_atomic();
1679 if (reserved_used
>= rdi
->dparms
.reserved_operations
)
1683 /* non-reserved operations */
1684 if (likely(qp
->s_avail
))
1686 slast
= READ_ONCE(qp
->s_last
);
1687 if (qp
->s_head
>= slast
)
1688 avail
= qp
->s_size
- (qp
->s_head
- slast
);
1690 avail
= slast
- qp
->s_head
;
1692 /* see rvt_qp_wqe_unreserve() */
1693 smp_mb__before_atomic();
1694 reserved_used
= atomic_read(&qp
->s_reserved_used
);
1696 (rdi
->dparms
.reserved_operations
- reserved_used
);
1697 /* insure we don't assign a negative s_avail */
1698 if ((s32
)avail
<= 0)
1700 qp
->s_avail
= avail
;
1701 if (WARN_ON(qp
->s_avail
>
1702 (qp
->s_size
- 1 - rdi
->dparms
.reserved_operations
)))
1704 "More avail entries than QP RB size.\nQP: %u, size: %u, avail: %u\nhead: %u, tail: %u, cur: %u, acked: %u, last: %u",
1705 qp
->ibqp
.qp_num
, qp
->s_size
, qp
->s_avail
,
1706 qp
->s_head
, qp
->s_tail
, qp
->s_cur
,
1707 qp
->s_acked
, qp
->s_last
);
1712 * rvt_post_one_wr - post one RC, UC, or UD send work request
1713 * @qp: the QP to post on
1714 * @wr: the work request to send
1716 static int rvt_post_one_wr(struct rvt_qp
*qp
,
1717 struct ib_send_wr
*wr
,
1720 struct rvt_swqe
*wqe
;
1725 struct rvt_lkey_table
*rkt
;
1727 struct rvt_dev_info
*rdi
= ib_to_rvt(qp
->ibqp
.device
);
1732 int local_ops_delayed
= 0;
1734 BUILD_BUG_ON(IB_QPT_MAX
>= (sizeof(u32
) * BITS_PER_BYTE
));
1736 /* IB spec says that num_sge == 0 is OK. */
1737 if (unlikely(wr
->num_sge
> qp
->s_max_sge
))
1740 ret
= rvt_qp_valid_operation(qp
, rdi
->post_parms
, wr
);
1746 * Local operations include fast register and local invalidate.
1747 * Fast register needs to be processed immediately because the
1748 * registered lkey may be used by following work requests and the
1749 * lkey needs to be valid at the time those requests are posted.
1750 * Local invalidate can be processed immediately if fencing is
1751 * not required and no previous local invalidate ops are pending.
1752 * Signaled local operations that have been processed immediately
1753 * need to have requests with "completion only" flags set posted
1754 * to the send queue in order to generate completions.
1756 if ((rdi
->post_parms
[wr
->opcode
].flags
& RVT_OPERATION_LOCAL
)) {
1757 switch (wr
->opcode
) {
1759 ret
= rvt_fast_reg_mr(qp
,
1762 reg_wr(wr
)->access
);
1763 if (ret
|| !(wr
->send_flags
& IB_SEND_SIGNALED
))
1766 case IB_WR_LOCAL_INV
:
1767 if ((wr
->send_flags
& IB_SEND_FENCE
) ||
1768 atomic_read(&qp
->local_ops_pending
)) {
1769 local_ops_delayed
= 1;
1771 ret
= rvt_invalidate_rkey(
1772 qp
, wr
->ex
.invalidate_rkey
);
1773 if (ret
|| !(wr
->send_flags
& IB_SEND_SIGNALED
))
1782 reserved_op
= rdi
->post_parms
[wr
->opcode
].flags
&
1783 RVT_OPERATION_USE_RESERVE
;
1784 /* check for avail */
1785 ret
= rvt_qp_is_avail(qp
, rdi
, reserved_op
);
1788 next
= qp
->s_head
+ 1;
1789 if (next
>= qp
->s_size
)
1792 rkt
= &rdi
->lkey_table
;
1793 pd
= ibpd_to_rvtpd(qp
->ibqp
.pd
);
1794 wqe
= rvt_get_swqe_ptr(qp
, qp
->s_head
);
1796 /* cplen has length from above */
1797 memcpy(&wqe
->wr
, wr
, cplen
);
1802 struct rvt_sge
*last_sge
= NULL
;
1804 acc
= wr
->opcode
>= IB_WR_RDMA_READ
?
1805 IB_ACCESS_LOCAL_WRITE
: 0;
1806 for (i
= 0; i
< wr
->num_sge
; i
++) {
1807 u32 length
= wr
->sg_list
[i
].length
;
1811 ret
= rvt_lkey_ok(rkt
, pd
, &wqe
->sg_list
[j
], last_sge
,
1812 &wr
->sg_list
[i
], acc
);
1813 if (unlikely(ret
< 0))
1814 goto bail_inval_free
;
1815 wqe
->length
+= length
;
1817 last_sge
= &wqe
->sg_list
[j
];
1820 wqe
->wr
.num_sge
= j
;
1823 /* general part of wqe valid - allow for driver checks */
1824 if (rdi
->driver_f
.check_send_wqe
) {
1825 ret
= rdi
->driver_f
.check_send_wqe(qp
, wqe
);
1827 goto bail_inval_free
;
1832 log_pmtu
= qp
->log_pmtu
;
1833 if (qp
->ibqp
.qp_type
!= IB_QPT_UC
&&
1834 qp
->ibqp
.qp_type
!= IB_QPT_RC
) {
1835 struct rvt_ah
*ah
= ibah_to_rvtah(wqe
->ud_wr
.ah
);
1837 log_pmtu
= ah
->log_pmtu
;
1838 atomic_inc(&ibah_to_rvtah(ud_wr(wr
)->ah
)->refcount
);
1841 if (rdi
->post_parms
[wr
->opcode
].flags
& RVT_OPERATION_LOCAL
) {
1842 if (local_ops_delayed
)
1843 atomic_inc(&qp
->local_ops_pending
);
1845 wqe
->wr
.send_flags
|= RVT_SEND_COMPLETION_ONLY
;
1850 wqe
->ssn
= qp
->s_ssn
++;
1851 wqe
->psn
= qp
->s_next_psn
;
1852 wqe
->lpsn
= wqe
->psn
+
1854 ((wqe
->length
- 1) >> log_pmtu
) :
1856 qp
->s_next_psn
= wqe
->lpsn
+ 1;
1858 if (unlikely(reserved_op
)) {
1859 wqe
->wr
.send_flags
|= RVT_SEND_RESERVE_USED
;
1860 rvt_qp_wqe_reserve(qp
, wqe
);
1862 wqe
->wr
.send_flags
&= ~RVT_SEND_RESERVE_USED
;
1865 trace_rvt_post_one_wr(qp
, wqe
, wr
->num_sge
);
1866 smp_wmb(); /* see request builders */
1872 /* release mr holds */
1874 struct rvt_sge
*sge
= &wqe
->sg_list
[--j
];
1876 rvt_put_mr(sge
->mr
);
1882 * rvt_post_send - post a send on a QP
1883 * @ibqp: the QP to post the send on
1884 * @wr: the list of work requests to post
1885 * @bad_wr: the first bad WR is put here
1887 * This may be called from interrupt context.
1889 * Return: 0 on success else errno
1891 int rvt_post_send(struct ib_qp
*ibqp
, struct ib_send_wr
*wr
,
1892 struct ib_send_wr
**bad_wr
)
1894 struct rvt_qp
*qp
= ibqp_to_rvtqp(ibqp
);
1895 struct rvt_dev_info
*rdi
= ib_to_rvt(ibqp
->device
);
1896 unsigned long flags
= 0;
1901 spin_lock_irqsave(&qp
->s_hlock
, flags
);
1904 * Ensure QP state is such that we can send. If not bail out early,
1905 * there is no need to do this every time we post a send.
1907 if (unlikely(!(ib_rvt_state_ops
[qp
->state
] & RVT_POST_SEND_OK
))) {
1908 spin_unlock_irqrestore(&qp
->s_hlock
, flags
);
1913 * If the send queue is empty, and we only have a single WR then just go
1914 * ahead and kick the send engine into gear. Otherwise we will always
1915 * just schedule the send to happen later.
1917 call_send
= qp
->s_head
== READ_ONCE(qp
->s_last
) && !wr
->next
;
1919 for (; wr
; wr
= wr
->next
) {
1920 err
= rvt_post_one_wr(qp
, wr
, &call_send
);
1921 if (unlikely(err
)) {
1928 spin_unlock_irqrestore(&qp
->s_hlock
, flags
);
1931 rdi
->driver_f
.do_send(qp
);
1933 rdi
->driver_f
.schedule_send_no_lock(qp
);
1939 * rvt_post_srq_receive - post a receive on a shared receive queue
1940 * @ibsrq: the SRQ to post the receive on
1941 * @wr: the list of work requests to post
1942 * @bad_wr: A pointer to the first WR to cause a problem is put here
1944 * This may be called from interrupt context.
1946 * Return: 0 on success else errno
1948 int rvt_post_srq_recv(struct ib_srq
*ibsrq
, struct ib_recv_wr
*wr
,
1949 struct ib_recv_wr
**bad_wr
)
1951 struct rvt_srq
*srq
= ibsrq_to_rvtsrq(ibsrq
);
1953 unsigned long flags
;
1955 for (; wr
; wr
= wr
->next
) {
1956 struct rvt_rwqe
*wqe
;
1960 if ((unsigned)wr
->num_sge
> srq
->rq
.max_sge
) {
1965 spin_lock_irqsave(&srq
->rq
.lock
, flags
);
1967 next
= wq
->head
+ 1;
1968 if (next
>= srq
->rq
.size
)
1970 if (next
== wq
->tail
) {
1971 spin_unlock_irqrestore(&srq
->rq
.lock
, flags
);
1976 wqe
= rvt_get_rwqe_ptr(&srq
->rq
, wq
->head
);
1977 wqe
->wr_id
= wr
->wr_id
;
1978 wqe
->num_sge
= wr
->num_sge
;
1979 for (i
= 0; i
< wr
->num_sge
; i
++)
1980 wqe
->sg_list
[i
] = wr
->sg_list
[i
];
1981 /* Make sure queue entry is written before the head index. */
1984 spin_unlock_irqrestore(&srq
->rq
.lock
, flags
);
1990 * Validate a RWQE and fill in the SGE state.
1993 static int init_sge(struct rvt_qp
*qp
, struct rvt_rwqe
*wqe
)
1997 struct rvt_lkey_table
*rkt
;
1999 struct rvt_sge_state
*ss
;
2000 struct rvt_dev_info
*rdi
= ib_to_rvt(qp
->ibqp
.device
);
2002 rkt
= &rdi
->lkey_table
;
2003 pd
= ibpd_to_rvtpd(qp
->ibqp
.srq
? qp
->ibqp
.srq
->pd
: qp
->ibqp
.pd
);
2005 ss
->sg_list
= qp
->r_sg_list
;
2007 for (i
= j
= 0; i
< wqe
->num_sge
; i
++) {
2008 if (wqe
->sg_list
[i
].length
== 0)
2011 ret
= rvt_lkey_ok(rkt
, pd
, j
? &ss
->sg_list
[j
- 1] : &ss
->sge
,
2012 NULL
, &wqe
->sg_list
[i
],
2013 IB_ACCESS_LOCAL_WRITE
);
2014 if (unlikely(ret
<= 0))
2016 qp
->r_len
+= wqe
->sg_list
[i
].length
;
2020 ss
->total_len
= qp
->r_len
;
2025 struct rvt_sge
*sge
= --j
? &ss
->sg_list
[j
- 1] : &ss
->sge
;
2027 rvt_put_mr(sge
->mr
);
2030 memset(&wc
, 0, sizeof(wc
));
2031 wc
.wr_id
= wqe
->wr_id
;
2032 wc
.status
= IB_WC_LOC_PROT_ERR
;
2033 wc
.opcode
= IB_WC_RECV
;
2035 /* Signal solicited completion event. */
2036 rvt_cq_enter(ibcq_to_rvtcq(qp
->ibqp
.recv_cq
), &wc
, 1);
2041 * rvt_get_rwqe - copy the next RWQE into the QP's RWQE
2043 * @wr_id_only: update qp->r_wr_id only, not qp->r_sge
2045 * Return -1 if there is a local error, 0 if no RWQE is available,
2046 * otherwise return 1.
2048 * Can be called from interrupt level.
2050 int rvt_get_rwqe(struct rvt_qp
*qp
, bool wr_id_only
)
2052 unsigned long flags
;
2055 struct rvt_srq
*srq
;
2056 struct rvt_rwqe
*wqe
;
2057 void (*handler
)(struct ib_event
*, void *);
2062 srq
= ibsrq_to_rvtsrq(qp
->ibqp
.srq
);
2063 handler
= srq
->ibsrq
.event_handler
;
2071 spin_lock_irqsave(&rq
->lock
, flags
);
2072 if (!(ib_rvt_state_ops
[qp
->state
] & RVT_PROCESS_RECV_OK
)) {
2079 /* Validate tail before using it since it is user writable. */
2080 if (tail
>= rq
->size
)
2082 if (unlikely(tail
== wq
->head
)) {
2086 /* Make sure entry is read after head index is read. */
2088 wqe
= rvt_get_rwqe_ptr(rq
, tail
);
2090 * Even though we update the tail index in memory, the verbs
2091 * consumer is not supposed to post more entries until a
2092 * completion is generated.
2094 if (++tail
>= rq
->size
)
2097 if (!wr_id_only
&& !init_sge(qp
, wqe
)) {
2101 qp
->r_wr_id
= wqe
->wr_id
;
2104 set_bit(RVT_R_WRID_VALID
, &qp
->r_aflags
);
2109 * Validate head pointer value and compute
2110 * the number of remaining WQEs.
2116 n
+= rq
->size
- tail
;
2119 if (n
< srq
->limit
) {
2123 spin_unlock_irqrestore(&rq
->lock
, flags
);
2124 ev
.device
= qp
->ibqp
.device
;
2125 ev
.element
.srq
= qp
->ibqp
.srq
;
2126 ev
.event
= IB_EVENT_SRQ_LIMIT_REACHED
;
2127 handler(&ev
, srq
->ibsrq
.srq_context
);
2132 spin_unlock_irqrestore(&rq
->lock
, flags
);
2136 EXPORT_SYMBOL(rvt_get_rwqe
);
2139 * qp_comm_est - handle trap with QP established
2142 void rvt_comm_est(struct rvt_qp
*qp
)
2144 qp
->r_flags
|= RVT_R_COMM_EST
;
2145 if (qp
->ibqp
.event_handler
) {
2148 ev
.device
= qp
->ibqp
.device
;
2149 ev
.element
.qp
= &qp
->ibqp
;
2150 ev
.event
= IB_EVENT_COMM_EST
;
2151 qp
->ibqp
.event_handler(&ev
, qp
->ibqp
.qp_context
);
2154 EXPORT_SYMBOL(rvt_comm_est
);
2156 void rvt_rc_error(struct rvt_qp
*qp
, enum ib_wc_status err
)
2158 unsigned long flags
;
2161 spin_lock_irqsave(&qp
->s_lock
, flags
);
2162 lastwqe
= rvt_error_qp(qp
, err
);
2163 spin_unlock_irqrestore(&qp
->s_lock
, flags
);
2168 ev
.device
= qp
->ibqp
.device
;
2169 ev
.element
.qp
= &qp
->ibqp
;
2170 ev
.event
= IB_EVENT_QP_LAST_WQE_REACHED
;
2171 qp
->ibqp
.event_handler(&ev
, qp
->ibqp
.qp_context
);
2174 EXPORT_SYMBOL(rvt_rc_error
);
2177 * rvt_rnr_tbl_to_usec - return index into ib_rvt_rnr_table
2178 * @index - the index
2179 * return usec from an index into ib_rvt_rnr_table
2181 unsigned long rvt_rnr_tbl_to_usec(u32 index
)
2183 return ib_rvt_rnr_table
[(index
& IB_AETH_CREDIT_MASK
)];
2185 EXPORT_SYMBOL(rvt_rnr_tbl_to_usec
);
2187 static inline unsigned long rvt_aeth_to_usec(u32 aeth
)
2189 return ib_rvt_rnr_table
[(aeth
>> IB_AETH_CREDIT_SHIFT
) &
2190 IB_AETH_CREDIT_MASK
];
2194 * rvt_add_retry_timer - add/start a retry timer
2196 * add a retry timer on the QP
2198 void rvt_add_retry_timer(struct rvt_qp
*qp
)
2200 struct ib_qp
*ibqp
= &qp
->ibqp
;
2201 struct rvt_dev_info
*rdi
= ib_to_rvt(ibqp
->device
);
2203 lockdep_assert_held(&qp
->s_lock
);
2204 qp
->s_flags
|= RVT_S_TIMER
;
2205 /* 4.096 usec. * (1 << qp->timeout) */
2206 qp
->s_timer
.expires
= jiffies
+ qp
->timeout_jiffies
+
2208 add_timer(&qp
->s_timer
);
2210 EXPORT_SYMBOL(rvt_add_retry_timer
);
2213 * rvt_add_rnr_timer - add/start an rnr timer
2215 * @aeth - aeth of RNR timeout, simulated aeth for loopback
2216 * add an rnr timer on the QP
2218 void rvt_add_rnr_timer(struct rvt_qp
*qp
, u32 aeth
)
2222 lockdep_assert_held(&qp
->s_lock
);
2223 qp
->s_flags
|= RVT_S_WAIT_RNR
;
2224 to
= rvt_aeth_to_usec(aeth
);
2225 trace_rvt_rnrnak_add(qp
, to
);
2226 hrtimer_start(&qp
->s_rnr_timer
,
2227 ns_to_ktime(1000 * to
), HRTIMER_MODE_REL_PINNED
);
2229 EXPORT_SYMBOL(rvt_add_rnr_timer
);
2232 * rvt_stop_rc_timers - stop all timers
2234 * stop any pending timers
2236 void rvt_stop_rc_timers(struct rvt_qp
*qp
)
2238 lockdep_assert_held(&qp
->s_lock
);
2239 /* Remove QP from all timers */
2240 if (qp
->s_flags
& (RVT_S_TIMER
| RVT_S_WAIT_RNR
)) {
2241 qp
->s_flags
&= ~(RVT_S_TIMER
| RVT_S_WAIT_RNR
);
2242 del_timer(&qp
->s_timer
);
2243 hrtimer_try_to_cancel(&qp
->s_rnr_timer
);
2246 EXPORT_SYMBOL(rvt_stop_rc_timers
);
2249 * rvt_stop_rnr_timer - stop an rnr timer
2252 * stop an rnr timer and return if the timer
2255 static void rvt_stop_rnr_timer(struct rvt_qp
*qp
)
2257 lockdep_assert_held(&qp
->s_lock
);
2258 /* Remove QP from rnr timer */
2259 if (qp
->s_flags
& RVT_S_WAIT_RNR
) {
2260 qp
->s_flags
&= ~RVT_S_WAIT_RNR
;
2261 trace_rvt_rnrnak_stop(qp
, 0);
2266 * rvt_del_timers_sync - wait for any timeout routines to exit
2269 void rvt_del_timers_sync(struct rvt_qp
*qp
)
2271 del_timer_sync(&qp
->s_timer
);
2272 hrtimer_cancel(&qp
->s_rnr_timer
);
2274 EXPORT_SYMBOL(rvt_del_timers_sync
);
2277 * This is called from s_timer for missing responses.
2279 static void rvt_rc_timeout(struct timer_list
*t
)
2281 struct rvt_qp
*qp
= from_timer(qp
, t
, s_timer
);
2282 struct rvt_dev_info
*rdi
= ib_to_rvt(qp
->ibqp
.device
);
2283 unsigned long flags
;
2285 spin_lock_irqsave(&qp
->r_lock
, flags
);
2286 spin_lock(&qp
->s_lock
);
2287 if (qp
->s_flags
& RVT_S_TIMER
) {
2288 struct rvt_ibport
*rvp
= rdi
->ports
[qp
->port_num
- 1];
2290 qp
->s_flags
&= ~RVT_S_TIMER
;
2291 rvp
->n_rc_timeouts
++;
2292 del_timer(&qp
->s_timer
);
2293 trace_rvt_rc_timeout(qp
, qp
->s_last_psn
+ 1);
2294 if (rdi
->driver_f
.notify_restart_rc
)
2295 rdi
->driver_f
.notify_restart_rc(qp
,
2298 rdi
->driver_f
.schedule_send(qp
);
2300 spin_unlock(&qp
->s_lock
);
2301 spin_unlock_irqrestore(&qp
->r_lock
, flags
);
2305 * This is called from s_timer for RNR timeouts.
2307 enum hrtimer_restart
rvt_rc_rnr_retry(struct hrtimer
*t
)
2309 struct rvt_qp
*qp
= container_of(t
, struct rvt_qp
, s_rnr_timer
);
2310 struct rvt_dev_info
*rdi
= ib_to_rvt(qp
->ibqp
.device
);
2311 unsigned long flags
;
2313 spin_lock_irqsave(&qp
->s_lock
, flags
);
2314 rvt_stop_rnr_timer(qp
);
2315 trace_rvt_rnrnak_timeout(qp
, 0);
2316 rdi
->driver_f
.schedule_send(qp
);
2317 spin_unlock_irqrestore(&qp
->s_lock
, flags
);
2318 return HRTIMER_NORESTART
;
2320 EXPORT_SYMBOL(rvt_rc_rnr_retry
);
2323 * rvt_qp_iter_init - initial for QP iteration
2327 * This returns an iterator suitable for iterating QPs
2330 * The @cb is a user defined callback and @v is a 64
2331 * bit value passed to and relevant for processing in the
2332 * @cb. An example use case would be to alter QP processing
2333 * based on criteria not part of the rvt_qp.
2335 * Use cases that require memory allocation to succeed
2336 * must preallocate appropriately.
2338 * Return: a pointer to an rvt_qp_iter or NULL
2340 struct rvt_qp_iter
*rvt_qp_iter_init(struct rvt_dev_info
*rdi
,
2342 void (*cb
)(struct rvt_qp
*qp
, u64 v
))
2344 struct rvt_qp_iter
*i
;
2346 i
= kzalloc(sizeof(*i
), GFP_KERNEL
);
2351 /* number of special QPs (SMI/GSI) for device */
2352 i
->specials
= rdi
->ibdev
.phys_port_cnt
* 2;
2358 EXPORT_SYMBOL(rvt_qp_iter_init
);
2361 * rvt_qp_iter_next - return the next QP in iter
2362 * @iter - the iterator
2364 * Fine grained QP iterator suitable for use
2365 * with debugfs seq_file mechanisms.
2367 * Updates iter->qp with the current QP when the return
2370 * Return: 0 - iter->qp is valid 1 - no more QPs
2372 int rvt_qp_iter_next(struct rvt_qp_iter
*iter
)
2377 struct rvt_qp
*pqp
= iter
->qp
;
2379 struct rvt_dev_info
*rdi
= iter
->rdi
;
2382 * The approach is to consider the special qps
2383 * as additional table entries before the
2384 * real hash table. Since the qp code sets
2385 * the qp->next hash link to NULL, this works just fine.
2387 * iter->specials is 2 * # ports
2389 * n = 0..iter->specials is the special qp indices
2391 * n = iter->specials..rdi->qp_dev->qp_table_size+iter->specials are
2392 * the potential hash bucket entries
2395 for (; n
< rdi
->qp_dev
->qp_table_size
+ iter
->specials
; n
++) {
2397 qp
= rcu_dereference(pqp
->next
);
2399 if (n
< iter
->specials
) {
2400 struct rvt_ibport
*rvp
;
2403 pidx
= n
% rdi
->ibdev
.phys_port_cnt
;
2404 rvp
= rdi
->ports
[pidx
];
2405 qp
= rcu_dereference(rvp
->qp
[n
& 1]);
2407 qp
= rcu_dereference(
2408 rdi
->qp_dev
->qp_table
[
2409 (n
- iter
->specials
)]);
2421 EXPORT_SYMBOL(rvt_qp_iter_next
);
2424 * rvt_qp_iter - iterate all QPs
2425 * @rdi - rvt devinfo
2426 * @v - a 64 bit value
2429 * This provides a way for iterating all QPs.
2431 * The @cb is a user defined callback and @v is a 64
2432 * bit value passed to and relevant for processing in the
2433 * cb. An example use case would be to alter QP processing
2434 * based on criteria not part of the rvt_qp.
2436 * The code has an internal iterator to simplify
2437 * non seq_file use cases.
2439 void rvt_qp_iter(struct rvt_dev_info
*rdi
,
2441 void (*cb
)(struct rvt_qp
*qp
, u64 v
))
2444 struct rvt_qp_iter i
= {
2446 .specials
= rdi
->ibdev
.phys_port_cnt
* 2,
2453 ret
= rvt_qp_iter_next(&i
);
2464 EXPORT_SYMBOL(rvt_qp_iter
);