1 // SPDX-License-Identifier: GPL-2.0
4 * Copyright 2016-2019 HabanaLabs, Ltd.
8 #include "habanalabs.h"
10 #include <linux/slab.h>
13 * hl_queue_add_ptr - add to pi or ci and checks if it wraps around
15 * @ptr: the current pi/ci value
16 * @val: the amount to add
18 * Add val to ptr. It can go until twice the queue length.
20 inline u32 hl_hw_queue_add_ptr(u32 ptr, u16 val)
23 ptr &= ((HL_QUEUE_LENGTH << 1) - 1);
26 static inline int queue_ci_get(atomic_t *ci, u32 queue_len)
28 return atomic_read(ci) & ((queue_len << 1) - 1);
31 static inline int queue_free_slots(struct hl_hw_queue *q, u32 queue_len)
33 int delta = (q->pi - queue_ci_get(&q->ci, queue_len));
36 return (queue_len - delta);
38 return (abs(delta) - queue_len);
41 void hl_hw_queue_update_ci(struct hl_cs *cs)
43 struct hl_device *hdev = cs->ctx->hdev;
44 struct hl_hw_queue *q;
50 q = &hdev->kernel_queues[0];
52 /* There are no internal queues if H/W queues are being used */
53 if (!hdev->asic_prop.max_queues || q->queue_type == QUEUE_TYPE_HW)
56 /* We must increment CI for every queue that will never get a
57 * completion, there are 2 scenarios this can happen:
58 * 1. All queues of a non completion CS will never get a completion.
59 * 2. Internal queues never gets completion.
61 for (i = 0 ; i < hdev->asic_prop.max_queues ; i++, q++) {
62 if (!cs_needs_completion(cs) || q->queue_type == QUEUE_TYPE_INT)
63 atomic_add(cs->jobs_in_queue_cnt[i], &q->ci);
68 * hl_hw_queue_submit_bd() - Submit a buffer descriptor to an external or a
70 * @hdev: pointer to habanalabs device structure
71 * @q: pointer to habanalabs queue structure
72 * @ctl: BD's control word
76 * This function assumes there is enough space on the queue to submit a new
77 * BD to it. It initializes the next BD and calls the device specific
78 * function to set the pi (and doorbell)
80 * This function must be called when the scheduler mutex is taken
83 void hl_hw_queue_submit_bd(struct hl_device *hdev, struct hl_hw_queue *q,
84 u32 ctl, u32 len, u64 ptr)
88 bd = q->kernel_address;
89 bd += hl_pi_2_offset(q->pi);
90 bd->ctl = cpu_to_le32(ctl);
91 bd->len = cpu_to_le32(len);
92 bd->ptr = cpu_to_le64(ptr);
94 q->pi = hl_queue_inc_ptr(q->pi);
95 hdev->asic_funcs->ring_doorbell(hdev, q->hw_queue_id, q->pi);
99 * ext_queue_sanity_checks - perform some sanity checks on external queue
101 * @hdev : pointer to hl_device structure
102 * @q : pointer to hl_hw_queue structure
103 * @num_of_entries : how many entries to check for space
104 * @reserve_cq_entry : whether to reserve an entry in the cq
106 * H/W queues spinlock should be taken before calling this function
108 * Perform the following:
109 * - Make sure we have enough space in the h/w queue
110 * - Make sure we have enough space in the completion queue
111 * - Reserve space in the completion queue (needs to be reversed if there
112 * is a failure down the road before the actual submission of work). Only
113 * do this action if reserve_cq_entry is true
116 static int ext_queue_sanity_checks(struct hl_device *hdev,
117 struct hl_hw_queue *q, int num_of_entries,
118 bool reserve_cq_entry)
120 atomic_t *free_slots =
121 &hdev->completion_queue[q->cq_id].free_slots_cnt;
124 /* Check we have enough space in the queue */
125 free_slots_cnt = queue_free_slots(q, HL_QUEUE_LENGTH);
127 if (free_slots_cnt < num_of_entries) {
128 dev_dbg(hdev->dev, "Queue %d doesn't have room for %d CBs\n",
129 q->hw_queue_id, num_of_entries);
133 if (reserve_cq_entry) {
135 * Check we have enough space in the completion queue
136 * Add -1 to counter (decrement) unless counter was already 0
137 * In that case, CQ is full so we can't submit a new CB because
138 * we won't get ack on its completion
139 * atomic_add_unless will return 0 if counter was already 0
141 if (atomic_add_negative(num_of_entries * -1, free_slots)) {
142 dev_dbg(hdev->dev, "No space for %d on CQ %d\n",
143 num_of_entries, q->hw_queue_id);
144 atomic_add(num_of_entries, free_slots);
153 * int_queue_sanity_checks - perform some sanity checks on internal queue
155 * @hdev : pointer to hl_device structure
156 * @q : pointer to hl_hw_queue structure
157 * @num_of_entries : how many entries to check for space
159 * H/W queues spinlock should be taken before calling this function
161 * Perform the following:
162 * - Make sure we have enough space in the h/w queue
165 static int int_queue_sanity_checks(struct hl_device *hdev,
166 struct hl_hw_queue *q,
171 if (num_of_entries > q->int_queue_len) {
173 "Cannot populate queue %u with %u jobs\n",
174 q->hw_queue_id, num_of_entries);
178 /* Check we have enough space in the queue */
179 free_slots_cnt = queue_free_slots(q, q->int_queue_len);
181 if (free_slots_cnt < num_of_entries) {
182 dev_dbg(hdev->dev, "Queue %d doesn't have room for %d CBs\n",
183 q->hw_queue_id, num_of_entries);
191 * hw_queue_sanity_checks() - Make sure we have enough space in the h/w queue
192 * @hdev: Pointer to hl_device structure.
193 * @q: Pointer to hl_hw_queue structure.
194 * @num_of_entries: How many entries to check for space.
196 * Notice: We do not reserve queue entries so this function mustn't be called
197 * more than once per CS for the same queue
200 static int hw_queue_sanity_checks(struct hl_device *hdev, struct hl_hw_queue *q,
205 /* Check we have enough space in the queue */
206 free_slots_cnt = queue_free_slots(q, HL_QUEUE_LENGTH);
208 if (free_slots_cnt < num_of_entries) {
209 dev_dbg(hdev->dev, "Queue %d doesn't have room for %d CBs\n",
210 q->hw_queue_id, num_of_entries);
218 * hl_hw_queue_send_cb_no_cmpl - send a single CB (not a JOB) without completion
220 * @hdev: pointer to hl_device structure
221 * @hw_queue_id: Queue's type
222 * @cb_size: size of CB
223 * @cb_ptr: pointer to CB location
225 * This function sends a single CB, that must NOT generate a completion entry.
226 * Sending CPU messages can be done instead via 'hl_hw_queue_submit_bd()'
228 int hl_hw_queue_send_cb_no_cmpl(struct hl_device *hdev, u32 hw_queue_id,
229 u32 cb_size, u64 cb_ptr)
231 struct hl_hw_queue *q = &hdev->kernel_queues[hw_queue_id];
234 hdev->asic_funcs->hw_queues_lock(hdev);
236 if (hdev->disabled) {
242 * hl_hw_queue_send_cb_no_cmpl() is called for queues of a H/W queue
243 * type only on init phase, when the queues are empty and being tested,
244 * so there is no need for sanity checks.
246 if (q->queue_type != QUEUE_TYPE_HW) {
247 rc = ext_queue_sanity_checks(hdev, q, 1, false);
252 hl_hw_queue_submit_bd(hdev, q, 0, cb_size, cb_ptr);
255 hdev->asic_funcs->hw_queues_unlock(hdev);
261 * ext_queue_schedule_job - submit a JOB to an external queue
263 * @job: pointer to the job that needs to be submitted to the queue
265 * This function must be called when the scheduler mutex is taken
268 static void ext_queue_schedule_job(struct hl_cs_job *job)
270 struct hl_device *hdev = job->cs->ctx->hdev;
271 struct hl_hw_queue *q = &hdev->kernel_queues[job->hw_queue_id];
272 struct hl_cq_entry cq_pkt;
281 * Update the JOB ID inside the BD CTL so the device would know what
282 * to write in the completion queue
284 ctl = ((q->pi << BD_CTL_SHADOW_INDEX_SHIFT) & BD_CTL_SHADOW_INDEX_MASK);
286 cb = job->patched_cb;
287 len = job->job_cb_size;
288 ptr = cb->bus_address;
290 /* Skip completion flow in case this is a non completion CS */
291 if (!cs_needs_completion(job->cs))
294 cq_pkt.data = cpu_to_le32(
295 ((q->pi << CQ_ENTRY_SHADOW_INDEX_SHIFT)
296 & CQ_ENTRY_SHADOW_INDEX_MASK) |
297 FIELD_PREP(CQ_ENTRY_SHADOW_INDEX_VALID_MASK, 1) |
298 FIELD_PREP(CQ_ENTRY_READY_MASK, 1));
301 * No need to protect pi_offset because scheduling to the
302 * H/W queues is done under the scheduler mutex
304 * No need to check if CQ is full because it was already
305 * checked in ext_queue_sanity_checks
307 cq = &hdev->completion_queue[q->cq_id];
308 cq_addr = cq->bus_address + cq->pi * sizeof(struct hl_cq_entry);
310 hdev->asic_funcs->add_end_of_cb_packets(hdev, cb->kernel_address, len,
312 le32_to_cpu(cq_pkt.data),
314 job->contains_dma_pkt);
316 q->shadow_queue[hl_pi_2_offset(q->pi)] = job;
318 cq->pi = hl_cq_inc_ptr(cq->pi);
321 hl_hw_queue_submit_bd(hdev, q, ctl, len, ptr);
325 * int_queue_schedule_job - submit a JOB to an internal queue
327 * @job: pointer to the job that needs to be submitted to the queue
329 * This function must be called when the scheduler mutex is taken
332 static void int_queue_schedule_job(struct hl_cs_job *job)
334 struct hl_device *hdev = job->cs->ctx->hdev;
335 struct hl_hw_queue *q = &hdev->kernel_queues[job->hw_queue_id];
340 bd.len = cpu_to_le32(job->job_cb_size);
342 if (job->is_kernel_allocated_cb)
343 /* bus_address is actually a mmu mapped address
344 * allocated from an internal pool
346 bd.ptr = cpu_to_le64(job->user_cb->bus_address);
348 bd.ptr = cpu_to_le64((u64) (uintptr_t) job->user_cb);
350 pi = q->kernel_address + (q->pi & (q->int_queue_len - 1)) * sizeof(bd);
353 q->pi &= ((q->int_queue_len << 1) - 1);
355 hdev->asic_funcs->pqe_write(hdev, pi, &bd);
357 hdev->asic_funcs->ring_doorbell(hdev, q->hw_queue_id, q->pi);
361 * hw_queue_schedule_job - submit a JOB to a H/W queue
363 * @job: pointer to the job that needs to be submitted to the queue
365 * This function must be called when the scheduler mutex is taken
368 static void hw_queue_schedule_job(struct hl_cs_job *job)
370 struct hl_device *hdev = job->cs->ctx->hdev;
371 struct hl_hw_queue *q = &hdev->kernel_queues[job->hw_queue_id];
373 u32 offset, ctl, len;
376 * Upon PQE completion, COMP_DATA is used as the write data to the
377 * completion queue (QMAN HBW message), and COMP_OFFSET is used as the
378 * write address offset in the SM block (QMAN LBW message).
379 * The write address offset is calculated as "COMP_OFFSET << 2".
381 offset = job->cs->sequence & (hdev->asic_prop.max_pending_cs - 1);
382 ctl = ((offset << BD_CTL_COMP_OFFSET_SHIFT) & BD_CTL_COMP_OFFSET_MASK) |
383 ((q->pi << BD_CTL_COMP_DATA_SHIFT) & BD_CTL_COMP_DATA_MASK);
385 len = job->job_cb_size;
388 * A patched CB is created only if a user CB was allocated by driver and
389 * MMU is disabled. If MMU is enabled, the user CB should be used
390 * instead. If the user CB wasn't allocated by driver, assume that it
394 ptr = job->patched_cb->bus_address;
395 else if (job->is_kernel_allocated_cb)
396 ptr = job->user_cb->bus_address;
398 ptr = (u64) (uintptr_t) job->user_cb;
400 hl_hw_queue_submit_bd(hdev, q, ctl, len, ptr);
403 static int init_signal_cs(struct hl_device *hdev,
404 struct hl_cs_job *job, struct hl_cs_compl *cs_cmpl)
406 struct hl_sync_stream_properties *prop;
407 struct hl_hw_sob *hw_sob;
411 q_idx = job->hw_queue_id;
412 prop = &hdev->kernel_queues[q_idx].sync_stream_prop;
413 hw_sob = &prop->hw_sob[prop->curr_sob_offset];
415 cs_cmpl->hw_sob = hw_sob;
416 cs_cmpl->sob_val = prop->next_sob_val;
419 "generate signal CB, sob_id: %d, sob val: 0x%x, q_idx: %d\n",
420 cs_cmpl->hw_sob->sob_id, cs_cmpl->sob_val, q_idx);
422 /* we set an EB since we must make sure all oeprations are done
423 * when sending the signal
425 hdev->asic_funcs->gen_signal_cb(hdev, job->patched_cb,
426 cs_cmpl->hw_sob->sob_id, 0, true);
428 rc = hl_cs_signal_sob_wraparound_handler(hdev, q_idx, &hw_sob, 1);
433 static void init_wait_cs(struct hl_device *hdev, struct hl_cs *cs,
434 struct hl_cs_job *job, struct hl_cs_compl *cs_cmpl)
436 struct hl_cs_compl *signal_cs_cmpl;
437 struct hl_sync_stream_properties *prop;
438 struct hl_gen_wait_properties wait_prop;
441 q_idx = job->hw_queue_id;
442 prop = &hdev->kernel_queues[q_idx].sync_stream_prop;
444 signal_cs_cmpl = container_of(cs->signal_fence,
448 /* copy the SOB id and value of the signal CS */
449 cs_cmpl->hw_sob = signal_cs_cmpl->hw_sob;
450 cs_cmpl->sob_val = signal_cs_cmpl->sob_val;
453 "generate wait CB, sob_id: %d, sob_val: 0x%x, mon_id: %d, q_idx: %d\n",
454 cs_cmpl->hw_sob->sob_id, cs_cmpl->sob_val,
455 prop->base_mon_id, q_idx);
457 wait_prop.data = (void *) job->patched_cb;
458 wait_prop.sob_base = cs_cmpl->hw_sob->sob_id;
459 wait_prop.sob_mask = 0x1;
460 wait_prop.sob_val = cs_cmpl->sob_val;
461 wait_prop.mon_id = prop->base_mon_id;
462 wait_prop.q_idx = q_idx;
464 hdev->asic_funcs->gen_wait_cb(hdev, &wait_prop);
466 kref_get(&cs_cmpl->hw_sob->kref);
468 * Must put the signal fence after the SOB refcnt increment so
469 * the SOB refcnt won't turn 0 and reset the SOB before the
470 * wait CS was submitted.
473 hl_fence_put(cs->signal_fence);
474 cs->signal_fence = NULL;
478 * init_signal_wait_cs - initialize a signal/wait CS
479 * @cs: pointer to the signal/wait CS
481 * H/W queues spinlock should be taken before calling this function
483 static int init_signal_wait_cs(struct hl_cs *cs)
485 struct hl_ctx *ctx = cs->ctx;
486 struct hl_device *hdev = ctx->hdev;
487 struct hl_cs_job *job;
488 struct hl_cs_compl *cs_cmpl =
489 container_of(cs->fence, struct hl_cs_compl, base_fence);
492 /* There is only one job in a signal/wait CS */
493 job = list_first_entry(&cs->job_list, struct hl_cs_job,
496 if (cs->type & CS_TYPE_SIGNAL)
497 rc = init_signal_cs(hdev, job, cs_cmpl);
498 else if (cs->type & CS_TYPE_WAIT)
499 init_wait_cs(hdev, cs, job, cs_cmpl);
505 * hl_hw_queue_schedule_cs - schedule a command submission
506 * @cs: pointer to the CS
508 int hl_hw_queue_schedule_cs(struct hl_cs *cs)
510 enum hl_device_status status;
511 struct hl_cs_counters_atomic *cntr;
512 struct hl_ctx *ctx = cs->ctx;
513 struct hl_device *hdev = ctx->hdev;
514 struct hl_cs_job *job, *tmp;
515 struct hl_hw_queue *q;
516 int rc = 0, i, cq_cnt;
520 cntr = &hdev->aggregated_cs_counters;
522 hdev->asic_funcs->hw_queues_lock(hdev);
524 if (!hl_device_operational(hdev, &status)) {
525 atomic64_inc(&cntr->device_in_reset_drop_cnt);
526 atomic64_inc(&ctx->cs_counters.device_in_reset_drop_cnt);
528 "device is %s, CS rejected!\n", hdev->status[status]);
533 max_queues = hdev->asic_prop.max_queues;
535 q = &hdev->kernel_queues[0];
536 for (i = 0, cq_cnt = 0 ; i < max_queues ; i++, q++) {
537 if (cs->jobs_in_queue_cnt[i]) {
538 switch (q->queue_type) {
540 rc = ext_queue_sanity_checks(hdev, q,
541 cs->jobs_in_queue_cnt[i],
542 cs_needs_completion(cs) ?
546 rc = int_queue_sanity_checks(hdev, q,
547 cs->jobs_in_queue_cnt[i]);
550 rc = hw_queue_sanity_checks(hdev, q,
551 cs->jobs_in_queue_cnt[i]);
554 dev_err(hdev->dev, "Queue type %d is invalid\n",
562 &ctx->cs_counters.queue_full_drop_cnt);
563 atomic64_inc(&cntr->queue_full_drop_cnt);
567 if (q->queue_type == QUEUE_TYPE_EXT)
572 if ((cs->type == CS_TYPE_SIGNAL) || (cs->type == CS_TYPE_WAIT)) {
573 rc = init_signal_wait_cs(cs);
575 dev_err(hdev->dev, "Failed to submit signal cs\n");
578 } else if (cs->type == CS_TYPE_COLLECTIVE_WAIT)
579 hdev->asic_funcs->collective_wait_init_cs(cs);
582 spin_lock(&hdev->cs_mirror_lock);
584 /* Verify staged CS exists and add to the staged list */
585 if (cs->staged_cs && !cs->staged_first) {
586 struct hl_cs *staged_cs;
588 staged_cs = hl_staged_cs_find_first(hdev, cs->staged_sequence);
591 "Cannot find staged submission sequence %llu",
592 cs->staged_sequence);
594 goto unlock_cs_mirror;
597 if (is_staged_cs_last_exists(hdev, staged_cs)) {
599 "Staged submission sequence %llu already submitted",
600 cs->staged_sequence);
602 goto unlock_cs_mirror;
605 list_add_tail(&cs->staged_cs_node, &staged_cs->staged_cs_node);
607 /* update stream map of the first CS */
608 if (hdev->supports_wait_for_multi_cs)
609 staged_cs->fence->stream_map |= cs->fence->stream_map;
612 list_add_tail(&cs->mirror_node, &hdev->cs_mirror_list);
614 /* Queue TDR if the CS is the first entry and if timeout is wanted */
615 first_entry = list_first_entry(&hdev->cs_mirror_list,
616 struct hl_cs, mirror_node) == cs;
617 if ((hdev->timeout_jiffies != MAX_SCHEDULE_TIMEOUT) &&
618 first_entry && cs_needs_timeout(cs)) {
619 cs->tdr_active = true;
620 schedule_delayed_work(&cs->work_tdr, cs->timeout_jiffies);
624 spin_unlock(&hdev->cs_mirror_lock);
626 list_for_each_entry_safe(job, tmp, &cs->job_list, cs_node)
627 switch (job->queue_type) {
629 ext_queue_schedule_job(job);
632 int_queue_schedule_job(job);
635 hw_queue_schedule_job(job);
641 cs->submitted = true;
646 spin_unlock(&hdev->cs_mirror_lock);
648 q = &hdev->kernel_queues[0];
649 for (i = 0 ; (i < max_queues) && (cq_cnt > 0) ; i++, q++) {
650 if ((q->queue_type == QUEUE_TYPE_EXT) &&
651 (cs->jobs_in_queue_cnt[i])) {
652 atomic_t *free_slots =
653 &hdev->completion_queue[i].free_slots_cnt;
654 atomic_add(cs->jobs_in_queue_cnt[i], free_slots);
660 hdev->asic_funcs->hw_queues_unlock(hdev);
666 * hl_hw_queue_inc_ci_kernel - increment ci for kernel's queue
668 * @hdev: pointer to hl_device structure
669 * @hw_queue_id: which queue to increment its ci
671 void hl_hw_queue_inc_ci_kernel(struct hl_device *hdev, u32 hw_queue_id)
673 struct hl_hw_queue *q = &hdev->kernel_queues[hw_queue_id];
678 static int ext_and_cpu_queue_init(struct hl_device *hdev, struct hl_hw_queue *q,
685 p = hdev->asic_funcs->cpu_accessible_dma_pool_alloc(hdev,
686 HL_QUEUE_SIZE_IN_BYTES,
689 p = hdev->asic_funcs->asic_dma_alloc_coherent(hdev,
690 HL_QUEUE_SIZE_IN_BYTES,
692 GFP_KERNEL | __GFP_ZERO);
696 q->kernel_address = p;
698 q->shadow_queue = kmalloc_array(HL_QUEUE_LENGTH,
699 sizeof(*q->shadow_queue),
701 if (!q->shadow_queue) {
703 "Failed to allocate shadow queue for H/W queue %d\n",
709 /* Make sure read/write pointers are initialized to start of queue */
710 atomic_set(&q->ci, 0);
717 hdev->asic_funcs->cpu_accessible_dma_pool_free(hdev,
718 HL_QUEUE_SIZE_IN_BYTES,
721 hdev->asic_funcs->asic_dma_free_coherent(hdev,
722 HL_QUEUE_SIZE_IN_BYTES,
729 static int int_queue_init(struct hl_device *hdev, struct hl_hw_queue *q)
733 p = hdev->asic_funcs->get_int_queue_base(hdev, q->hw_queue_id,
734 &q->bus_address, &q->int_queue_len);
737 "Failed to get base address for internal queue %d\n",
742 q->kernel_address = p;
744 atomic_set(&q->ci, 0);
749 static int cpu_queue_init(struct hl_device *hdev, struct hl_hw_queue *q)
751 return ext_and_cpu_queue_init(hdev, q, true);
754 static int ext_queue_init(struct hl_device *hdev, struct hl_hw_queue *q)
756 return ext_and_cpu_queue_init(hdev, q, false);
759 static int hw_queue_init(struct hl_device *hdev, struct hl_hw_queue *q)
763 p = hdev->asic_funcs->asic_dma_alloc_coherent(hdev,
764 HL_QUEUE_SIZE_IN_BYTES,
766 GFP_KERNEL | __GFP_ZERO);
770 q->kernel_address = p;
772 /* Make sure read/write pointers are initialized to start of queue */
773 atomic_set(&q->ci, 0);
779 static void sync_stream_queue_init(struct hl_device *hdev, u32 q_idx)
781 struct hl_sync_stream_properties *sync_stream_prop;
782 struct asic_fixed_properties *prop = &hdev->asic_prop;
783 struct hl_hw_sob *hw_sob;
784 int sob, reserved_mon_idx, queue_idx;
786 sync_stream_prop = &hdev->kernel_queues[q_idx].sync_stream_prop;
788 /* We use 'collective_mon_idx' as a running index in order to reserve
789 * monitors for collective master/slave queues.
790 * collective master queue gets 2 reserved monitors
791 * collective slave queue gets 1 reserved monitor
793 if (hdev->kernel_queues[q_idx].collective_mode ==
794 HL_COLLECTIVE_MASTER) {
795 reserved_mon_idx = hdev->collective_mon_idx;
797 /* reserve the first monitor for collective master queue */
798 sync_stream_prop->collective_mstr_mon_id[0] =
799 prop->collective_first_mon + reserved_mon_idx;
801 /* reserve the second monitor for collective master queue */
802 sync_stream_prop->collective_mstr_mon_id[1] =
803 prop->collective_first_mon + reserved_mon_idx + 1;
805 hdev->collective_mon_idx += HL_COLLECTIVE_RSVD_MSTR_MONS;
806 } else if (hdev->kernel_queues[q_idx].collective_mode ==
807 HL_COLLECTIVE_SLAVE) {
808 reserved_mon_idx = hdev->collective_mon_idx++;
810 /* reserve a monitor for collective slave queue */
811 sync_stream_prop->collective_slave_mon_id =
812 prop->collective_first_mon + reserved_mon_idx;
815 if (!hdev->kernel_queues[q_idx].supports_sync_stream)
818 queue_idx = hdev->sync_stream_queue_idx++;
820 sync_stream_prop->base_sob_id = prop->sync_stream_first_sob +
821 (queue_idx * HL_RSVD_SOBS);
822 sync_stream_prop->base_mon_id = prop->sync_stream_first_mon +
823 (queue_idx * HL_RSVD_MONS);
824 sync_stream_prop->next_sob_val = 1;
825 sync_stream_prop->curr_sob_offset = 0;
827 for (sob = 0 ; sob < HL_RSVD_SOBS ; sob++) {
828 hw_sob = &sync_stream_prop->hw_sob[sob];
830 hw_sob->sob_id = sync_stream_prop->base_sob_id + sob;
831 hw_sob->q_idx = q_idx;
832 kref_init(&hw_sob->kref);
836 static void sync_stream_queue_reset(struct hl_device *hdev, u32 q_idx)
838 struct hl_sync_stream_properties *prop =
839 &hdev->kernel_queues[q_idx].sync_stream_prop;
842 * In case we got here due to a stuck CS, the refcnt might be bigger
843 * than 1 and therefore we reset it.
845 kref_init(&prop->hw_sob[prop->curr_sob_offset].kref);
846 prop->curr_sob_offset = 0;
847 prop->next_sob_val = 1;
851 * queue_init - main initialization function for H/W queue object
853 * @hdev: pointer to hl_device device structure
854 * @q: pointer to hl_hw_queue queue structure
855 * @hw_queue_id: The id of the H/W queue
857 * Allocate dma-able memory for the queue and initialize fields
858 * Returns 0 on success
860 static int queue_init(struct hl_device *hdev, struct hl_hw_queue *q,
865 q->hw_queue_id = hw_queue_id;
867 switch (q->queue_type) {
869 rc = ext_queue_init(hdev, q);
872 rc = int_queue_init(hdev, q);
875 rc = cpu_queue_init(hdev, q);
878 rc = hw_queue_init(hdev, q);
884 dev_crit(hdev->dev, "wrong queue type %d during init\n",
890 sync_stream_queue_init(hdev, q->hw_queue_id);
901 * hw_queue_fini - destroy queue
903 * @hdev: pointer to hl_device device structure
904 * @q: pointer to hl_hw_queue queue structure
906 * Free the queue memory
908 static void queue_fini(struct hl_device *hdev, struct hl_hw_queue *q)
914 * If we arrived here, there are no jobs waiting on this queue
915 * so we can safely remove it.
916 * This is because this function can only called when:
917 * 1. Either a context is deleted, which only can occur if all its
919 * 2. A context wasn't able to be created due to failure or timeout,
920 * which means there are no jobs on the queue yet
922 * The only exception are the queues of the kernel context, but
923 * if they are being destroyed, it means that the entire module is
924 * being removed. If the module is removed, it means there is no open
925 * user context. It also means that if a job was submitted by
926 * the kernel driver (e.g. context creation), the job itself was
927 * released by the kernel driver when a timeout occurred on its
928 * Completion. Thus, we don't need to release it again.
931 if (q->queue_type == QUEUE_TYPE_INT)
934 kfree(q->shadow_queue);
936 if (q->queue_type == QUEUE_TYPE_CPU)
937 hdev->asic_funcs->cpu_accessible_dma_pool_free(hdev,
938 HL_QUEUE_SIZE_IN_BYTES,
941 hdev->asic_funcs->asic_dma_free_coherent(hdev,
942 HL_QUEUE_SIZE_IN_BYTES,
947 int hl_hw_queues_create(struct hl_device *hdev)
949 struct asic_fixed_properties *asic = &hdev->asic_prop;
950 struct hl_hw_queue *q;
951 int i, rc, q_ready_cnt;
953 hdev->kernel_queues = kcalloc(asic->max_queues,
954 sizeof(*hdev->kernel_queues), GFP_KERNEL);
956 if (!hdev->kernel_queues) {
957 dev_err(hdev->dev, "Not enough memory for H/W queues\n");
961 /* Initialize the H/W queues */
962 for (i = 0, q_ready_cnt = 0, q = hdev->kernel_queues;
963 i < asic->max_queues ; i++, q_ready_cnt++, q++) {
965 q->queue_type = asic->hw_queues_props[i].type;
966 q->supports_sync_stream =
967 asic->hw_queues_props[i].supports_sync_stream;
968 q->collective_mode = asic->hw_queues_props[i].collective_mode;
969 rc = queue_init(hdev, q, i);
972 "failed to initialize queue %d\n", i);
980 for (i = 0, q = hdev->kernel_queues ; i < q_ready_cnt ; i++, q++)
983 kfree(hdev->kernel_queues);
988 void hl_hw_queues_destroy(struct hl_device *hdev)
990 struct hl_hw_queue *q;
991 u32 max_queues = hdev->asic_prop.max_queues;
994 for (i = 0, q = hdev->kernel_queues ; i < max_queues ; i++, q++)
997 kfree(hdev->kernel_queues);
1000 void hl_hw_queue_reset(struct hl_device *hdev, bool hard_reset)
1002 struct hl_hw_queue *q;
1003 u32 max_queues = hdev->asic_prop.max_queues;
1006 for (i = 0, q = hdev->kernel_queues ; i < max_queues ; i++, q++) {
1008 ((!hard_reset) && (q->queue_type == QUEUE_TYPE_CPU)))
1011 atomic_set(&q->ci, 0);
1013 if (q->supports_sync_stream)
1014 sync_stream_queue_reset(hdev, q->hw_queue_id);