2 * Broadcom NetXtreme-E RoCE driver.
4 * Copyright (c) 2016 - 2017, Broadcom. All rights reserved. The term
5 * Broadcom refers to Broadcom Limited and/or its subsidiaries.
7 * This software is available to you under a choice of one of two
8 * licenses. You may choose to be licensed under the terms of the GNU
9 * General Public License (GPL) Version 2, available from the file
10 * COPYING in the main directory of this source tree, or the
13 * Redistribution and use in source and binary forms, with or without
14 * modification, are permitted provided that the following conditions
17 * 1. Redistributions of source code must retain the above copyright
18 * notice, this list of conditions and the following disclaimer.
19 * 2. Redistributions in binary form must reproduce the above copyright
20 * notice, this list of conditions and the following disclaimer in
21 * the documentation and/or other materials provided with the
24 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS''
25 * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
26 * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
27 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS
28 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
29 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
30 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR
31 * BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
32 * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE
33 * OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN
34 * IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
36 * Description: RDMA Controller HW interface
39 #define dev_fmt(fmt) "QPLIB: " fmt
41 #include <linux/interrupt.h>
42 #include <linux/spinlock.h>
43 #include <linux/pci.h>
44 #include <linux/prefetch.h>
45 #include <linux/delay.h>
48 #include "qplib_res.h"
49 #include "qplib_rcfw.h"
52 #include "qplib_tlv.h"
54 static void bnxt_qplib_service_creq(struct tasklet_struct *t);
57 * bnxt_qplib_map_rc - map return type based on opcode
58 * @opcode: roce slow path opcode
61 * Firmware initiated error recovery is a safe state machine and
62 * driver can consider all the underlying rdma resources are free.
63 * In this state, it is safe to return success for opcodes related to
64 * destroying rdma resources (like destroy qp, destroy cq etc.).
67 * If driver detect potential firmware stall, it is not safe state machine
68 * and the driver can not consider all the underlying rdma resources are
70 * In this state, it is not safe to return success for opcodes related to
71 * destroying rdma resources (like destroy qp, destroy cq etc.).
73 * Scope of this helper function is only for case #1.
76 * 0 to communicate success to caller.
77 * Non zero error code to communicate failure to caller.
79 static int bnxt_qplib_map_rc(u8 opcode)
82 case CMDQ_BASE_OPCODE_DESTROY_QP:
83 case CMDQ_BASE_OPCODE_DESTROY_SRQ:
84 case CMDQ_BASE_OPCODE_DESTROY_CQ:
85 case CMDQ_BASE_OPCODE_DEALLOCATE_KEY:
86 case CMDQ_BASE_OPCODE_DEREGISTER_MR:
87 case CMDQ_BASE_OPCODE_DELETE_GID:
88 case CMDQ_BASE_OPCODE_DESTROY_QP1:
89 case CMDQ_BASE_OPCODE_DESTROY_AH:
90 case CMDQ_BASE_OPCODE_DEINITIALIZE_FW:
91 case CMDQ_BASE_OPCODE_MODIFY_ROCE_CC:
92 case CMDQ_BASE_OPCODE_SET_LINK_AGGR_MODE:
100 * bnxt_re_is_fw_stalled - Check firmware health
101 * @rcfw: rcfw channel instance of rdev
102 * @cookie: cookie to track the command
104 * If firmware has not responded any rcfw command within
105 * rcfw->max_timeout, consider firmware as stalled.
108 * 0 if firmware is responding
109 * -ENODEV if firmware is not responding
111 static int bnxt_re_is_fw_stalled(struct bnxt_qplib_rcfw *rcfw,
114 struct bnxt_qplib_cmdq_ctx *cmdq;
115 struct bnxt_qplib_crsqe *crsqe;
117 crsqe = &rcfw->crsqe_tbl[cookie];
120 if (time_after(jiffies, cmdq->last_seen +
121 (rcfw->max_timeout * HZ))) {
122 dev_warn_ratelimited(&rcfw->pdev->dev,
123 "%s: FW STALL Detected. cmdq[%#x]=%#x waited (%d > %d) msec active %d ",
124 __func__, cookie, crsqe->opcode,
125 jiffies_to_msecs(jiffies - cmdq->last_seen),
126 rcfw->max_timeout * 1000,
135 * __wait_for_resp - Don't hold the cpu context and wait for response
136 * @rcfw: rcfw channel instance of rdev
137 * @cookie: cookie to track the command
139 * Wait for command completion in sleepable context.
142 * 0 if command is completed by firmware.
143 * Non zero error code for rest of the case.
145 static int __wait_for_resp(struct bnxt_qplib_rcfw *rcfw, u16 cookie)
147 struct bnxt_qplib_cmdq_ctx *cmdq;
148 struct bnxt_qplib_crsqe *crsqe;
152 crsqe = &rcfw->crsqe_tbl[cookie];
155 if (test_bit(ERR_DEVICE_DETACHED, &cmdq->flags))
156 return bnxt_qplib_map_rc(crsqe->opcode);
157 if (test_bit(FIRMWARE_STALL_DETECTED, &cmdq->flags))
160 wait_event_timeout(cmdq->waitq,
161 !crsqe->is_in_used ||
162 test_bit(ERR_DEVICE_DETACHED, &cmdq->flags),
163 msecs_to_jiffies(rcfw->max_timeout * 1000));
165 if (!crsqe->is_in_used)
168 bnxt_qplib_service_creq(&rcfw->creq.creq_tasklet);
170 if (!crsqe->is_in_used)
173 ret = bnxt_re_is_fw_stalled(rcfw, cookie);
181 * __block_for_resp - hold the cpu context and wait for response
182 * @rcfw: rcfw channel instance of rdev
183 * @cookie: cookie to track the command
185 * This function will hold the cpu (non-sleepable context) and
186 * wait for command completion. Maximum holding interval is 8 second.
189 * -ETIMEOUT if command is not completed in specific time interval.
190 * 0 if command is completed by firmware.
192 static int __block_for_resp(struct bnxt_qplib_rcfw *rcfw, u16 cookie)
194 struct bnxt_qplib_cmdq_ctx *cmdq = &rcfw->cmdq;
195 struct bnxt_qplib_crsqe *crsqe;
196 unsigned long issue_time = 0;
198 issue_time = jiffies;
199 crsqe = &rcfw->crsqe_tbl[cookie];
202 if (test_bit(ERR_DEVICE_DETACHED, &cmdq->flags))
203 return bnxt_qplib_map_rc(crsqe->opcode);
204 if (test_bit(FIRMWARE_STALL_DETECTED, &cmdq->flags))
209 bnxt_qplib_service_creq(&rcfw->creq.creq_tasklet);
210 if (!crsqe->is_in_used)
213 } while (time_before(jiffies, issue_time + (8 * HZ)));
218 /* __send_message_no_waiter - get cookie and post the message.
219 * @rcfw: rcfw channel instance of rdev
220 * @msg: qplib message internal
222 * This function will just post and don't bother about completion.
223 * Current design of this function is -
224 * user must hold the completion queue hwq->lock.
225 * user must have used existing completion and free the resources.
226 * this function will not check queue full condition.
227 * this function will explicitly set is_waiter_alive=false.
228 * current use case is - send destroy_ah if create_ah is return
229 * after waiter of create_ah is lost. It can be extended for other
235 static void __send_message_no_waiter(struct bnxt_qplib_rcfw *rcfw,
236 struct bnxt_qplib_cmdqmsg *msg)
238 struct bnxt_qplib_cmdq_ctx *cmdq = &rcfw->cmdq;
239 struct bnxt_qplib_hwq *hwq = &cmdq->hwq;
240 struct bnxt_qplib_crsqe *crsqe;
241 struct bnxt_qplib_cmdqe *cmdqe;
242 u32 sw_prod, cmdq_prod;
247 cookie = cmdq->seq_num & RCFW_MAX_COOKIE_VALUE;
248 __set_cmdq_base_cookie(msg->req, msg->req_sz, cpu_to_le16(cookie));
249 crsqe = &rcfw->crsqe_tbl[cookie];
251 /* Set cmd_size in terms of 16B slots in req. */
252 bsize = bnxt_qplib_set_cmd_slots(msg->req);
253 /* GET_CMD_SIZE would return number of slots in either case of tlv
254 * and non-tlv commands after call to bnxt_qplib_set_cmd_slots()
256 crsqe->is_internal_cmd = true;
257 crsqe->is_waiter_alive = false;
258 crsqe->is_in_used = true;
259 crsqe->req_size = __get_cmdq_base_cmd_size(msg->req, msg->req_sz);
261 preq = (u8 *)msg->req;
263 /* Locate the next cmdq slot */
264 sw_prod = HWQ_CMP(hwq->prod, hwq);
265 cmdqe = bnxt_qplib_get_qe(hwq, sw_prod, NULL);
266 /* Copy a segment of the req cmd to the cmdq */
267 memset(cmdqe, 0, sizeof(*cmdqe));
268 memcpy(cmdqe, preq, min_t(u32, bsize, sizeof(*cmdqe)));
269 preq += min_t(u32, bsize, sizeof(*cmdqe));
270 bsize -= min_t(u32, bsize, sizeof(*cmdqe));
275 cmdq_prod = hwq->prod;
276 atomic_inc(&rcfw->timeout_send);
279 writel(cmdq_prod, cmdq->cmdq_mbox.prod);
280 writel(RCFW_CMDQ_TRIG_VAL, cmdq->cmdq_mbox.db);
283 static int __send_message(struct bnxt_qplib_rcfw *rcfw,
284 struct bnxt_qplib_cmdqmsg *msg, u8 opcode)
286 u32 bsize, free_slots, required_slots;
287 struct bnxt_qplib_cmdq_ctx *cmdq;
288 struct bnxt_qplib_crsqe *crsqe;
289 struct bnxt_qplib_cmdqe *cmdqe;
290 struct bnxt_qplib_hwq *hwq;
291 u32 sw_prod, cmdq_prod;
292 struct pci_dev *pdev;
301 /* Cmdq are in 16-byte units, each request can consume 1 or more
304 spin_lock_irqsave(&hwq->lock, flags);
305 required_slots = bnxt_qplib_get_cmd_slots(msg->req);
306 free_slots = HWQ_FREE_SLOTS(hwq);
307 cookie = cmdq->seq_num & RCFW_MAX_COOKIE_VALUE;
308 crsqe = &rcfw->crsqe_tbl[cookie];
310 if (required_slots >= free_slots) {
311 dev_info_ratelimited(&pdev->dev,
312 "CMDQ is full req/free %d/%d!",
313 required_slots, free_slots);
314 spin_unlock_irqrestore(&hwq->lock, flags);
318 cookie |= RCFW_CMD_IS_BLOCKING;
319 __set_cmdq_base_cookie(msg->req, msg->req_sz, cpu_to_le16(cookie));
321 bsize = bnxt_qplib_set_cmd_slots(msg->req);
322 crsqe->free_slots = free_slots;
323 crsqe->resp = (struct creq_qp_event *)msg->resp;
324 crsqe->resp->cookie = cpu_to_le16(cookie);
325 crsqe->is_internal_cmd = false;
326 crsqe->is_waiter_alive = true;
327 crsqe->is_in_used = true;
328 crsqe->opcode = opcode;
330 crsqe->req_size = __get_cmdq_base_cmd_size(msg->req, msg->req_sz);
331 if (__get_cmdq_base_resp_size(msg->req, msg->req_sz) && msg->sb) {
332 struct bnxt_qplib_rcfw_sbuf *sbuf = msg->sb;
334 __set_cmdq_base_resp_addr(msg->req, msg->req_sz,
335 cpu_to_le64(sbuf->dma_addr));
336 __set_cmdq_base_resp_size(msg->req, msg->req_sz,
338 BNXT_QPLIB_CMDQE_UNITS) /
339 BNXT_QPLIB_CMDQE_UNITS);
342 preq = (u8 *)msg->req;
344 /* Locate the next cmdq slot */
345 sw_prod = HWQ_CMP(hwq->prod, hwq);
346 cmdqe = bnxt_qplib_get_qe(hwq, sw_prod, NULL);
347 /* Copy a segment of the req cmd to the cmdq */
348 memset(cmdqe, 0, sizeof(*cmdqe));
349 memcpy(cmdqe, preq, min_t(u32, bsize, sizeof(*cmdqe)));
350 preq += min_t(u32, bsize, sizeof(*cmdqe));
351 bsize -= min_t(u32, bsize, sizeof(*cmdqe));
356 cmdq_prod = hwq->prod & 0xFFFF;
357 if (test_bit(FIRMWARE_FIRST_FLAG, &cmdq->flags)) {
358 /* The very first doorbell write
359 * is required to set this flag
360 * which prompts the FW to reset
361 * its internal pointers
363 cmdq_prod |= BIT(FIRMWARE_FIRST_FLAG);
364 clear_bit(FIRMWARE_FIRST_FLAG, &cmdq->flags);
368 writel(cmdq_prod, cmdq->cmdq_mbox.prod);
369 writel(RCFW_CMDQ_TRIG_VAL, cmdq->cmdq_mbox.db);
370 spin_unlock_irqrestore(&hwq->lock, flags);
371 /* Return the CREQ response pointer */
376 * __poll_for_resp - self poll completion for rcfw command
377 * @rcfw: rcfw channel instance of rdev
378 * @cookie: cookie to track the command
380 * It works same as __wait_for_resp except this function will
381 * do self polling in sort interval since interrupt is disabled.
382 * This function can not be called from non-sleepable context.
385 * -ETIMEOUT if command is not completed in specific time interval.
386 * 0 if command is completed by firmware.
388 static int __poll_for_resp(struct bnxt_qplib_rcfw *rcfw, u16 cookie)
390 struct bnxt_qplib_cmdq_ctx *cmdq = &rcfw->cmdq;
391 struct bnxt_qplib_crsqe *crsqe;
392 unsigned long issue_time;
395 issue_time = jiffies;
396 crsqe = &rcfw->crsqe_tbl[cookie];
399 if (test_bit(ERR_DEVICE_DETACHED, &cmdq->flags))
400 return bnxt_qplib_map_rc(crsqe->opcode);
401 if (test_bit(FIRMWARE_STALL_DETECTED, &cmdq->flags))
404 usleep_range(1000, 1001);
406 bnxt_qplib_service_creq(&rcfw->creq.creq_tasklet);
407 if (!crsqe->is_in_used)
409 if (jiffies_to_msecs(jiffies - issue_time) >
410 (rcfw->max_timeout * 1000)) {
411 ret = bnxt_re_is_fw_stalled(rcfw, cookie);
418 static int __send_message_basic_sanity(struct bnxt_qplib_rcfw *rcfw,
419 struct bnxt_qplib_cmdqmsg *msg,
422 struct bnxt_qplib_cmdq_ctx *cmdq;
426 /* Prevent posting if f/w is not in a state to process */
427 if (test_bit(ERR_DEVICE_DETACHED, &rcfw->cmdq.flags))
428 return bnxt_qplib_map_rc(opcode);
429 if (test_bit(FIRMWARE_STALL_DETECTED, &cmdq->flags))
432 if (test_bit(FIRMWARE_INITIALIZED_FLAG, &cmdq->flags) &&
433 opcode == CMDQ_BASE_OPCODE_INITIALIZE_FW) {
434 dev_err(&rcfw->pdev->dev, "QPLIB: RCFW already initialized!");
438 if (!test_bit(FIRMWARE_INITIALIZED_FLAG, &cmdq->flags) &&
439 (opcode != CMDQ_BASE_OPCODE_QUERY_FUNC &&
440 opcode != CMDQ_BASE_OPCODE_INITIALIZE_FW &&
441 opcode != CMDQ_BASE_OPCODE_QUERY_VERSION)) {
442 dev_err(&rcfw->pdev->dev,
443 "QPLIB: RCFW not initialized, reject opcode 0x%x",
451 /* This function will just post and do not bother about completion */
452 static void __destroy_timedout_ah(struct bnxt_qplib_rcfw *rcfw,
453 struct creq_create_ah_resp *create_ah_resp)
455 struct bnxt_qplib_cmdqmsg msg = {};
456 struct cmdq_destroy_ah req = {};
458 bnxt_qplib_rcfw_cmd_prep((struct cmdq_base *)&req,
459 CMDQ_BASE_OPCODE_DESTROY_AH,
461 req.ah_cid = create_ah_resp->xid;
462 msg.req = (struct cmdq_base *)&req;
463 msg.req_sz = sizeof(req);
464 __send_message_no_waiter(rcfw, &msg);
465 dev_info_ratelimited(&rcfw->pdev->dev,
466 "From %s: ah_cid = %d timeout_send %d\n",
467 __func__, req.ah_cid,
468 atomic_read(&rcfw->timeout_send));
472 * __bnxt_qplib_rcfw_send_message - qplib interface to send
473 * and complete rcfw command.
474 * @rcfw: rcfw channel instance of rdev
475 * @msg: qplib message internal
477 * This function does not account shadow queue depth. It will send
478 * all the command unconditionally as long as send queue is not full.
481 * 0 if command completed by firmware.
482 * Non zero if the command is not completed by firmware.
484 static int __bnxt_qplib_rcfw_send_message(struct bnxt_qplib_rcfw *rcfw,
485 struct bnxt_qplib_cmdqmsg *msg)
487 struct creq_qp_event *evnt = (struct creq_qp_event *)msg->resp;
488 struct bnxt_qplib_crsqe *crsqe;
494 opcode = __get_cmdq_base_opcode(msg->req, msg->req_sz);
496 rc = __send_message_basic_sanity(rcfw, msg, opcode);
500 rc = __send_message(rcfw, msg, opcode);
504 cookie = le16_to_cpu(__get_cmdq_base_cookie(msg->req, msg->req_sz))
505 & RCFW_MAX_COOKIE_VALUE;
508 rc = __block_for_resp(rcfw, cookie);
509 else if (atomic_read(&rcfw->rcfw_intr_enabled))
510 rc = __wait_for_resp(rcfw, cookie);
512 rc = __poll_for_resp(rcfw, cookie);
515 spin_lock_irqsave(&rcfw->cmdq.hwq.lock, flags);
516 crsqe = &rcfw->crsqe_tbl[cookie];
517 crsqe->is_waiter_alive = false;
519 set_bit(FIRMWARE_STALL_DETECTED, &rcfw->cmdq.flags);
520 spin_unlock_irqrestore(&rcfw->cmdq.hwq.lock, flags);
525 /* failed with status */
526 dev_err(&rcfw->pdev->dev, "cmdq[%#x]=%#x status %#x\n",
527 cookie, opcode, evnt->status);
535 * bnxt_qplib_rcfw_send_message - qplib interface to send
536 * and complete rcfw command.
537 * @rcfw: rcfw channel instance of rdev
538 * @msg: qplib message internal
540 * Driver interact with Firmware through rcfw channel/slow path in two ways.
541 * a. Blocking rcfw command send. In this path, driver cannot hold
542 * the context for longer period since it is holding cpu until
543 * command is not completed.
544 * b. Non-blocking rcfw command send. In this path, driver can hold the
545 * context for longer period. There may be many pending command waiting
546 * for completion because of non-blocking nature.
548 * Driver will use shadow queue depth. Current queue depth of 8K
549 * (due to size of rcfw message there can be actual ~4K rcfw outstanding)
550 * is not optimal for rcfw command processing in firmware.
552 * Restrict at max #RCFW_CMD_NON_BLOCKING_SHADOW_QD Non-Blocking rcfw commands.
553 * Allow all blocking commands until there is no queue full.
556 * 0 if command completed by firmware.
557 * Non zero if the command is not completed by firmware.
559 int bnxt_qplib_rcfw_send_message(struct bnxt_qplib_rcfw *rcfw,
560 struct bnxt_qplib_cmdqmsg *msg)
565 down(&rcfw->rcfw_inflight);
566 ret = __bnxt_qplib_rcfw_send_message(rcfw, msg);
567 up(&rcfw->rcfw_inflight);
569 ret = __bnxt_qplib_rcfw_send_message(rcfw, msg);
576 static int bnxt_qplib_process_func_event(struct bnxt_qplib_rcfw *rcfw,
577 struct creq_func_event *func_event)
581 switch (func_event->event) {
582 case CREQ_FUNC_EVENT_EVENT_TX_WQE_ERROR:
584 case CREQ_FUNC_EVENT_EVENT_TX_DATA_ERROR:
586 case CREQ_FUNC_EVENT_EVENT_RX_WQE_ERROR:
588 case CREQ_FUNC_EVENT_EVENT_RX_DATA_ERROR:
590 case CREQ_FUNC_EVENT_EVENT_CQ_ERROR:
592 case CREQ_FUNC_EVENT_EVENT_TQM_ERROR:
594 case CREQ_FUNC_EVENT_EVENT_CFCQ_ERROR:
596 case CREQ_FUNC_EVENT_EVENT_CFCS_ERROR:
597 /* SRQ ctx error, call srq_handler??
598 * But there's no SRQ handle!
601 case CREQ_FUNC_EVENT_EVENT_CFCC_ERROR:
603 case CREQ_FUNC_EVENT_EVENT_CFCM_ERROR:
605 case CREQ_FUNC_EVENT_EVENT_TIM_ERROR:
607 case CREQ_FUNC_EVENT_EVENT_VF_COMM_REQUEST:
609 case CREQ_FUNC_EVENT_EVENT_RESOURCE_EXHAUSTED:
615 rc = rcfw->creq.aeq_handler(rcfw, (void *)func_event, NULL);
619 static int bnxt_qplib_process_qp_event(struct bnxt_qplib_rcfw *rcfw,
620 struct creq_qp_event *qp_event,
623 struct creq_qp_error_notification *err_event;
624 struct bnxt_qplib_hwq *hwq = &rcfw->cmdq.hwq;
625 struct bnxt_qplib_crsqe *crsqe;
626 u32 qp_id, tbl_indx, req_size;
627 struct bnxt_qplib_qp *qp;
628 u16 cookie, blocked = 0;
629 bool is_waiter_alive;
630 struct pci_dev *pdev;
636 switch (qp_event->event) {
637 case CREQ_QP_EVENT_EVENT_QP_ERROR_NOTIFICATION:
638 err_event = (struct creq_qp_error_notification *)qp_event;
639 qp_id = le32_to_cpu(err_event->xid);
640 tbl_indx = map_qp_id_to_tbl_indx(qp_id, rcfw);
641 qp = rcfw->qp_tbl[tbl_indx].qp_handle;
642 dev_dbg(&pdev->dev, "Received QP error notification\n");
644 "qpid 0x%x, req_err=0x%x, resp_err=0x%x\n",
645 qp_id, err_event->req_err_state_reason,
646 err_event->res_err_state_reason);
649 bnxt_qplib_mark_qp_error(qp);
650 rc = rcfw->creq.aeq_handler(rcfw, qp_event, qp);
655 * cmdq->lock needs to be acquired to synchronie
656 * the command send and completion reaping. This function
657 * is always called with creq->lock held. Using
658 * the nested variant of spin_lock.
662 spin_lock_irqsave_nested(&hwq->lock, flags,
663 SINGLE_DEPTH_NESTING);
664 cookie = le16_to_cpu(qp_event->cookie);
665 blocked = cookie & RCFW_CMD_IS_BLOCKING;
666 cookie &= RCFW_MAX_COOKIE_VALUE;
667 crsqe = &rcfw->crsqe_tbl[cookie];
669 if (WARN_ONCE(test_bit(FIRMWARE_STALL_DETECTED,
671 "QPLIB: Unreponsive rcfw channel detected.!!")) {
673 "rcfw timedout: cookie = %#x, free_slots = %d",
674 cookie, crsqe->free_slots);
675 spin_unlock_irqrestore(&hwq->lock, flags);
679 if (crsqe->is_internal_cmd && !qp_event->status)
680 atomic_dec(&rcfw->timeout_send);
682 if (crsqe->is_waiter_alive) {
684 memcpy(crsqe->resp, qp_event, sizeof(*qp_event));
685 /* Insert write memory barrier to ensure that
686 * response data is copied before clearing the
695 req_size = crsqe->req_size;
696 is_waiter_alive = crsqe->is_waiter_alive;
699 if (!is_waiter_alive)
702 crsqe->is_in_used = false;
704 hwq->cons += req_size;
706 /* This is a case to handle below scenario -
707 * Create AH is completed successfully by firmware,
708 * but completion took more time and driver already lost
709 * the context of create_ah from caller.
710 * We have already return failure for create_ah verbs,
711 * so let's destroy the same address vector since it is
712 * no more used in stack. We don't care about completion
713 * in __send_message_no_waiter.
714 * If destroy_ah is failued by firmware, there will be AH
715 * resource leak and relatively not critical + unlikely
716 * scenario. Current design is not to handle such case.
718 if (!is_waiter_alive && !qp_event->status &&
719 qp_event->event == CREQ_QP_EVENT_EVENT_CREATE_AH)
720 __destroy_timedout_ah(rcfw,
721 (struct creq_create_ah_resp *)
723 spin_unlock_irqrestore(&hwq->lock, flags);
725 *num_wait += wait_cmds;
729 /* SP - CREQ Completion handlers */
730 static void bnxt_qplib_service_creq(struct tasklet_struct *t)
732 struct bnxt_qplib_rcfw *rcfw = from_tasklet(rcfw, t, creq.creq_tasklet);
733 struct bnxt_qplib_creq_ctx *creq = &rcfw->creq;
734 u32 type, budget = CREQ_ENTRY_POLL_BUDGET;
735 struct bnxt_qplib_hwq *hwq = &creq->hwq;
736 struct creq_base *creqe;
737 u32 sw_cons, raw_cons;
741 /* Service the CREQ until budget is over */
742 spin_lock_irqsave(&hwq->lock, flags);
743 raw_cons = hwq->cons;
745 sw_cons = HWQ_CMP(raw_cons, hwq);
746 creqe = bnxt_qplib_get_qe(hwq, sw_cons, NULL);
747 if (!CREQ_CMP_VALID(creqe, raw_cons, hwq->max_elements))
749 /* The valid test of the entry must be done first before
750 * reading any further.
753 rcfw->cmdq.last_seen = jiffies;
755 type = creqe->type & CREQ_BASE_TYPE_MASK;
757 case CREQ_BASE_TYPE_QP_EVENT:
758 bnxt_qplib_process_qp_event
759 (rcfw, (struct creq_qp_event *)creqe,
761 creq->stats.creq_qp_event_processed++;
763 case CREQ_BASE_TYPE_FUNC_EVENT:
764 if (!bnxt_qplib_process_func_event
765 (rcfw, (struct creq_func_event *)creqe))
766 creq->stats.creq_func_event_processed++;
768 dev_warn(&rcfw->pdev->dev,
769 "aeqe:%#x Not handled\n", type);
772 if (type != ASYNC_EVENT_CMPL_TYPE_HWRM_ASYNC_EVENT)
773 dev_warn(&rcfw->pdev->dev,
774 "creqe with event 0x%x not handled\n",
782 if (hwq->cons != raw_cons) {
783 hwq->cons = raw_cons;
784 bnxt_qplib_ring_nq_db(&creq->creq_db.dbinfo,
785 rcfw->res->cctx, true);
787 spin_unlock_irqrestore(&hwq->lock, flags);
789 wake_up_nr(&rcfw->cmdq.waitq, num_wakeup);
792 static irqreturn_t bnxt_qplib_creq_irq(int irq, void *dev_instance)
794 struct bnxt_qplib_rcfw *rcfw = dev_instance;
795 struct bnxt_qplib_creq_ctx *creq;
796 struct bnxt_qplib_hwq *hwq;
801 /* Prefetch the CREQ element */
802 sw_cons = HWQ_CMP(hwq->cons, hwq);
803 prefetch(bnxt_qplib_get_qe(hwq, sw_cons, NULL));
805 tasklet_schedule(&creq->creq_tasklet);
811 int bnxt_qplib_deinit_rcfw(struct bnxt_qplib_rcfw *rcfw)
813 struct creq_deinitialize_fw_resp resp = {};
814 struct cmdq_deinitialize_fw req = {};
815 struct bnxt_qplib_cmdqmsg msg = {};
818 bnxt_qplib_rcfw_cmd_prep((struct cmdq_base *)&req,
819 CMDQ_BASE_OPCODE_DEINITIALIZE_FW,
821 bnxt_qplib_fill_cmdqmsg(&msg, &req, &resp, NULL,
822 sizeof(req), sizeof(resp), 0);
823 rc = bnxt_qplib_rcfw_send_message(rcfw, &msg);
827 clear_bit(FIRMWARE_INITIALIZED_FLAG, &rcfw->cmdq.flags);
831 int bnxt_qplib_init_rcfw(struct bnxt_qplib_rcfw *rcfw,
832 struct bnxt_qplib_ctx *ctx, int is_virtfn)
834 struct creq_initialize_fw_resp resp = {};
835 struct cmdq_initialize_fw req = {};
836 struct bnxt_qplib_cmdqmsg msg = {};
840 bnxt_qplib_rcfw_cmd_prep((struct cmdq_base *)&req,
841 CMDQ_BASE_OPCODE_INITIALIZE_FW,
843 /* Supply (log-base-2-of-host-page-size - base-page-shift)
844 * to bono to adjust the doorbell page sizes.
846 req.log2_dbr_pg_size = cpu_to_le16(PAGE_SHIFT -
847 RCFW_DBR_BASE_PAGE_SHIFT);
849 * Gen P5 devices doesn't require this allocation
850 * as the L2 driver does the same for RoCE also.
851 * Also, VFs need not setup the HW context area, PF
852 * shall setup this area for VF. Skipping the
857 if (bnxt_qplib_is_chip_gen_p5(rcfw->res->cctx))
860 lvl = ctx->qpc_tbl.level;
861 pgsz = bnxt_qplib_base_pg_size(&ctx->qpc_tbl);
862 req.qpc_pg_size_qpc_lvl = (pgsz << CMDQ_INITIALIZE_FW_QPC_PG_SIZE_SFT) |
864 lvl = ctx->mrw_tbl.level;
865 pgsz = bnxt_qplib_base_pg_size(&ctx->mrw_tbl);
866 req.mrw_pg_size_mrw_lvl = (pgsz << CMDQ_INITIALIZE_FW_QPC_PG_SIZE_SFT) |
868 lvl = ctx->srqc_tbl.level;
869 pgsz = bnxt_qplib_base_pg_size(&ctx->srqc_tbl);
870 req.srq_pg_size_srq_lvl = (pgsz << CMDQ_INITIALIZE_FW_QPC_PG_SIZE_SFT) |
872 lvl = ctx->cq_tbl.level;
873 pgsz = bnxt_qplib_base_pg_size(&ctx->cq_tbl);
874 req.cq_pg_size_cq_lvl = (pgsz << CMDQ_INITIALIZE_FW_QPC_PG_SIZE_SFT) |
876 lvl = ctx->tim_tbl.level;
877 pgsz = bnxt_qplib_base_pg_size(&ctx->tim_tbl);
878 req.tim_pg_size_tim_lvl = (pgsz << CMDQ_INITIALIZE_FW_QPC_PG_SIZE_SFT) |
880 lvl = ctx->tqm_ctx.pde.level;
881 pgsz = bnxt_qplib_base_pg_size(&ctx->tqm_ctx.pde);
882 req.tqm_pg_size_tqm_lvl = (pgsz << CMDQ_INITIALIZE_FW_QPC_PG_SIZE_SFT) |
885 cpu_to_le64(ctx->qpc_tbl.pbl[PBL_LVL_0].pg_map_arr[0]);
887 cpu_to_le64(ctx->mrw_tbl.pbl[PBL_LVL_0].pg_map_arr[0]);
889 cpu_to_le64(ctx->srqc_tbl.pbl[PBL_LVL_0].pg_map_arr[0]);
891 cpu_to_le64(ctx->cq_tbl.pbl[PBL_LVL_0].pg_map_arr[0]);
893 cpu_to_le64(ctx->tim_tbl.pbl[PBL_LVL_0].pg_map_arr[0]);
895 cpu_to_le64(ctx->tqm_ctx.pde.pbl[PBL_LVL_0].pg_map_arr[0]);
897 req.number_of_qp = cpu_to_le32(ctx->qpc_tbl.max_elements);
898 req.number_of_mrw = cpu_to_le32(ctx->mrw_tbl.max_elements);
899 req.number_of_srq = cpu_to_le32(ctx->srqc_tbl.max_elements);
900 req.number_of_cq = cpu_to_le32(ctx->cq_tbl.max_elements);
903 req.max_qp_per_vf = cpu_to_le32(ctx->vf_res.max_qp_per_vf);
904 req.max_mrw_per_vf = cpu_to_le32(ctx->vf_res.max_mrw_per_vf);
905 req.max_srq_per_vf = cpu_to_le32(ctx->vf_res.max_srq_per_vf);
906 req.max_cq_per_vf = cpu_to_le32(ctx->vf_res.max_cq_per_vf);
907 req.max_gid_per_vf = cpu_to_le32(ctx->vf_res.max_gid_per_vf);
910 req.stat_ctx_id = cpu_to_le32(ctx->stats.fw_id);
911 bnxt_qplib_fill_cmdqmsg(&msg, &req, &resp, NULL, sizeof(req), sizeof(resp), 0);
912 rc = bnxt_qplib_rcfw_send_message(rcfw, &msg);
915 set_bit(FIRMWARE_INITIALIZED_FLAG, &rcfw->cmdq.flags);
919 void bnxt_qplib_free_rcfw_channel(struct bnxt_qplib_rcfw *rcfw)
922 kfree(rcfw->crsqe_tbl);
923 bnxt_qplib_free_hwq(rcfw->res, &rcfw->cmdq.hwq);
924 bnxt_qplib_free_hwq(rcfw->res, &rcfw->creq.hwq);
928 int bnxt_qplib_alloc_rcfw_channel(struct bnxt_qplib_res *res,
929 struct bnxt_qplib_rcfw *rcfw,
930 struct bnxt_qplib_ctx *ctx,
933 struct bnxt_qplib_hwq_attr hwq_attr = {};
934 struct bnxt_qplib_sg_info sginfo = {};
935 struct bnxt_qplib_cmdq_ctx *cmdq;
936 struct bnxt_qplib_creq_ctx *creq;
938 rcfw->pdev = res->pdev;
943 sginfo.pgsize = PAGE_SIZE;
944 sginfo.pgshft = PAGE_SHIFT;
946 hwq_attr.sginfo = &sginfo;
947 hwq_attr.res = rcfw->res;
948 hwq_attr.depth = BNXT_QPLIB_CREQE_MAX_CNT;
949 hwq_attr.stride = BNXT_QPLIB_CREQE_UNITS;
950 hwq_attr.type = bnxt_qplib_get_hwq_type(res);
952 if (bnxt_qplib_alloc_init_hwq(&creq->hwq, &hwq_attr)) {
953 dev_err(&rcfw->pdev->dev,
954 "HW channel CREQ allocation failed\n");
958 rcfw->cmdq_depth = BNXT_QPLIB_CMDQE_MAX_CNT;
960 sginfo.pgsize = bnxt_qplib_cmdqe_page_size(rcfw->cmdq_depth);
961 hwq_attr.depth = rcfw->cmdq_depth & 0x7FFFFFFF;
962 hwq_attr.stride = BNXT_QPLIB_CMDQE_UNITS;
963 hwq_attr.type = HWQ_TYPE_CTX;
964 if (bnxt_qplib_alloc_init_hwq(&cmdq->hwq, &hwq_attr)) {
965 dev_err(&rcfw->pdev->dev,
966 "HW channel CMDQ allocation failed\n");
970 rcfw->crsqe_tbl = kcalloc(cmdq->hwq.max_elements,
971 sizeof(*rcfw->crsqe_tbl), GFP_KERNEL);
972 if (!rcfw->crsqe_tbl)
975 /* Allocate one extra to hold the QP1 entries */
976 rcfw->qp_tbl_size = qp_tbl_sz + 1;
977 rcfw->qp_tbl = kcalloc(rcfw->qp_tbl_size, sizeof(struct bnxt_qplib_qp_node),
982 rcfw->max_timeout = res->cctx->hwrm_cmd_max_timeout;
987 bnxt_qplib_free_rcfw_channel(rcfw);
991 void bnxt_qplib_rcfw_stop_irq(struct bnxt_qplib_rcfw *rcfw, bool kill)
993 struct bnxt_qplib_creq_ctx *creq;
997 if (!creq->requested)
1000 creq->requested = false;
1001 /* Mask h/w interrupts */
1002 bnxt_qplib_ring_nq_db(&creq->creq_db.dbinfo, rcfw->res->cctx, false);
1003 /* Sync with last running IRQ-handler */
1004 synchronize_irq(creq->msix_vec);
1005 free_irq(creq->msix_vec, rcfw);
1006 kfree(creq->irq_name);
1007 creq->irq_name = NULL;
1008 atomic_set(&rcfw->rcfw_intr_enabled, 0);
1010 tasklet_kill(&creq->creq_tasklet);
1011 tasklet_disable(&creq->creq_tasklet);
1014 void bnxt_qplib_disable_rcfw_channel(struct bnxt_qplib_rcfw *rcfw)
1016 struct bnxt_qplib_creq_ctx *creq;
1017 struct bnxt_qplib_cmdq_ctx *cmdq;
1021 /* Make sure the HW channel is stopped! */
1022 bnxt_qplib_rcfw_stop_irq(rcfw, true);
1024 iounmap(cmdq->cmdq_mbox.reg.bar_reg);
1025 iounmap(creq->creq_db.reg.bar_reg);
1027 cmdq->cmdq_mbox.reg.bar_reg = NULL;
1028 creq->creq_db.reg.bar_reg = NULL;
1029 creq->aeq_handler = NULL;
1033 int bnxt_qplib_rcfw_start_irq(struct bnxt_qplib_rcfw *rcfw, int msix_vector,
1036 struct bnxt_qplib_creq_ctx *creq;
1037 struct bnxt_qplib_res *res;
1043 if (creq->requested)
1046 creq->msix_vec = msix_vector;
1048 tasklet_setup(&creq->creq_tasklet, bnxt_qplib_service_creq);
1050 tasklet_enable(&creq->creq_tasklet);
1052 creq->irq_name = kasprintf(GFP_KERNEL, "bnxt_re-creq@pci:%s",
1053 pci_name(res->pdev));
1054 if (!creq->irq_name)
1056 rc = request_irq(creq->msix_vec, bnxt_qplib_creq_irq, 0,
1057 creq->irq_name, rcfw);
1059 kfree(creq->irq_name);
1060 creq->irq_name = NULL;
1061 tasklet_disable(&creq->creq_tasklet);
1064 creq->requested = true;
1066 bnxt_qplib_ring_nq_db(&creq->creq_db.dbinfo, res->cctx, true);
1067 atomic_inc(&rcfw->rcfw_intr_enabled);
1072 static int bnxt_qplib_map_cmdq_mbox(struct bnxt_qplib_rcfw *rcfw)
1074 struct bnxt_qplib_cmdq_mbox *mbox;
1075 resource_size_t bar_reg;
1076 struct pci_dev *pdev;
1079 mbox = &rcfw->cmdq.cmdq_mbox;
1081 mbox->reg.bar_id = RCFW_COMM_PCI_BAR_REGION;
1082 mbox->reg.len = RCFW_COMM_SIZE;
1083 mbox->reg.bar_base = pci_resource_start(pdev, mbox->reg.bar_id);
1084 if (!mbox->reg.bar_base) {
1086 "QPLIB: CMDQ BAR region %d resc start is 0!\n",
1091 bar_reg = mbox->reg.bar_base + RCFW_COMM_BASE_OFFSET;
1092 mbox->reg.len = RCFW_COMM_SIZE;
1093 mbox->reg.bar_reg = ioremap(bar_reg, mbox->reg.len);
1094 if (!mbox->reg.bar_reg) {
1096 "QPLIB: CMDQ BAR region %d mapping failed\n",
1101 mbox->prod = (void __iomem *)(mbox->reg.bar_reg +
1102 RCFW_PF_VF_COMM_PROD_OFFSET);
1103 mbox->db = (void __iomem *)(mbox->reg.bar_reg + RCFW_COMM_TRIG_OFFSET);
1107 static int bnxt_qplib_map_creq_db(struct bnxt_qplib_rcfw *rcfw, u32 reg_offt)
1109 struct bnxt_qplib_creq_db *creq_db;
1110 resource_size_t bar_reg;
1111 struct pci_dev *pdev;
1114 creq_db = &rcfw->creq.creq_db;
1116 creq_db->reg.bar_id = RCFW_COMM_CONS_PCI_BAR_REGION;
1117 creq_db->reg.bar_base = pci_resource_start(pdev, creq_db->reg.bar_id);
1118 if (!creq_db->reg.bar_id)
1120 "QPLIB: CREQ BAR region %d resc start is 0!",
1121 creq_db->reg.bar_id);
1123 bar_reg = creq_db->reg.bar_base + reg_offt;
1124 /* Unconditionally map 8 bytes to support 57500 series */
1125 creq_db->reg.len = 8;
1126 creq_db->reg.bar_reg = ioremap(bar_reg, creq_db->reg.len);
1127 if (!creq_db->reg.bar_reg) {
1129 "QPLIB: CREQ BAR region %d mapping failed",
1130 creq_db->reg.bar_id);
1133 creq_db->dbinfo.db = creq_db->reg.bar_reg;
1134 creq_db->dbinfo.hwq = &rcfw->creq.hwq;
1135 creq_db->dbinfo.xid = rcfw->creq.ring_id;
1139 static void bnxt_qplib_start_rcfw(struct bnxt_qplib_rcfw *rcfw)
1141 struct bnxt_qplib_cmdq_ctx *cmdq;
1142 struct bnxt_qplib_creq_ctx *creq;
1143 struct bnxt_qplib_cmdq_mbox *mbox;
1144 struct cmdq_init init = {0};
1148 mbox = &cmdq->cmdq_mbox;
1150 init.cmdq_pbl = cpu_to_le64(cmdq->hwq.pbl[PBL_LVL_0].pg_map_arr[0]);
1151 init.cmdq_size_cmdq_lvl =
1152 cpu_to_le16(((rcfw->cmdq_depth <<
1153 CMDQ_INIT_CMDQ_SIZE_SFT) &
1154 CMDQ_INIT_CMDQ_SIZE_MASK) |
1155 ((cmdq->hwq.level <<
1156 CMDQ_INIT_CMDQ_LVL_SFT) &
1157 CMDQ_INIT_CMDQ_LVL_MASK));
1158 init.creq_ring_id = cpu_to_le16(creq->ring_id);
1159 /* Write to the Bono mailbox register */
1160 __iowrite32_copy(mbox->reg.bar_reg, &init, sizeof(init) / 4);
1163 int bnxt_qplib_enable_rcfw_channel(struct bnxt_qplib_rcfw *rcfw,
1166 aeq_handler_t aeq_handler)
1168 struct bnxt_qplib_cmdq_ctx *cmdq;
1169 struct bnxt_qplib_creq_ctx *creq;
1175 /* Clear to defaults */
1178 set_bit(FIRMWARE_FIRST_FLAG, &cmdq->flags);
1179 init_waitqueue_head(&cmdq->waitq);
1181 creq->stats.creq_qp_event_processed = 0;
1182 creq->stats.creq_func_event_processed = 0;
1183 creq->aeq_handler = aeq_handler;
1185 rc = bnxt_qplib_map_cmdq_mbox(rcfw);
1189 rc = bnxt_qplib_map_creq_db(rcfw, cp_bar_reg_off);
1193 rc = bnxt_qplib_rcfw_start_irq(rcfw, msix_vector, true);
1195 dev_err(&rcfw->pdev->dev,
1196 "Failed to request IRQ for CREQ rc = 0x%x\n", rc);
1197 bnxt_qplib_disable_rcfw_channel(rcfw);
1201 sema_init(&rcfw->rcfw_inflight, RCFW_CMD_NON_BLOCKING_SHADOW_QD);
1202 bnxt_qplib_start_rcfw(rcfw);