/** * enum hl_cs_wait_status - cs wait status * @CS_WAIT_STATUS_BUSY: cs was not completed yet * @CS_WAIT_STATUS_COMPLETED: cs completed * @CS_WAIT_STATUS_GONE: cs completed but fence is already gone
*/ enum hl_cs_wait_status {
CS_WAIT_STATUS_BUSY,
CS_WAIT_STATUS_COMPLETED,
CS_WAIT_STATUS_GONE
};
/* * Data used while handling wait/timestamp nodes. * The purpose of this struct is to store the needed data for both operations * in one variable instead of passing large number of arguments to functions.
*/ struct wait_interrupt_data { struct hl_user_interrupt *interrupt; struct hl_mmap_mem_buf *buf; struct hl_mem_mgr *mmg; struct hl_cb *cq_cb;
u64 ts_handle;
u64 ts_offset;
u64 cq_handle;
u64 cq_offset;
u64 target_value;
u64 intr_timeout_us;
};
/* * CS outcome store supports the following operations: * push outcome - store a recent CS outcome in the store * pop outcome - retrieve a SPECIFIC (by seq) CS outcome from the store * It uses 2 lists: used list and free list. * It has a pre-allocated amount of nodes, each node stores * a single CS outcome. * Initially, all the nodes are in the free list. * On push outcome, a node (any) is taken from the free list, its * information is filled in, and the node is moved to the used list. * It is possible, that there are no nodes left in the free list. * In this case, we will lose some information about old outcomes. We * will pop the OLDEST node from the used list, and make it free. * On pop, the node is searched for in the used list (using a search * index). * If found, the node is then removed from the used list, and moved * back to the free list. The outcome data that the node contained is * returned back to the user.
*/
dev_crit(hdev->dev, "SOB release shouldn't be called here, q_idx: %d, sob_id: %d\n",
hw_sob->q_idx, hw_sob->sob_id);
}
void hw_sob_put(struct hl_hw_sob *hw_sob)
{ if (hw_sob)
kref_put(&hw_sob->kref, hl_sob_reset);
}
staticvoid hw_sob_put_err(struct hl_hw_sob *hw_sob)
{ if (hw_sob)
kref_put(&hw_sob->kref, hl_sob_reset_error);
}
void hw_sob_get(struct hl_hw_sob *hw_sob)
{ if (hw_sob)
kref_get(&hw_sob->kref);
}
/** * hl_gen_sob_mask() - Generates a sob mask to be used in a monitor arm packet * @sob_base: sob base id * @sob_mask: sob user mask, each bit represents a sob offset from sob base * @mask: generated mask * * Return: 0 if given parameters are valid
*/ int hl_gen_sob_mask(u16 sob_base, u8 sob_mask, u8 *mask)
{ int i;
if (sob_mask == 0) return -EINVAL;
if (sob_mask == 0x1) {
*mask = ~(1 << (sob_base & 0x7));
} else { /* find msb in order to verify sob range is valid */ for (i = BITS_PER_BYTE - 1 ; i >= 0 ; i--) if (BIT(i) & sob_mask) break;
if (i > (HL_MAX_SOBS_PER_MONITOR - (sob_base & 0x7) - 1)) return -EINVAL;
bool cs_needs_completion(struct hl_cs *cs)
{ /* In case this is a staged CS, only the last CS in sequence should * get a completion, any non staged CS will always get a completion
*/ if (cs->staged_cs && !cs->staged_last) returnfalse;
returntrue;
}
bool cs_needs_timeout(struct hl_cs *cs)
{ /* In case this is a staged CS, only the first CS in sequence should * get a timeout, any non staged CS will always get a timeout
*/ if (cs->staged_cs && !cs->staged_first) returnfalse;
returntrue;
}
staticbool is_cb_patched(struct hl_device *hdev, struct hl_cs_job *job)
{ /* Patched CB is created for external queues jobs */ return (job->queue_type == QUEUE_TYPE_EXT);
}
/* * cs_parser - parse the user command submission * * @hpriv : pointer to the private data of the fd * @job : pointer to the job that holds the command submission info * * The function parses the command submission of the user. It calls the * ASIC specific parser, which returns a list of memory blocks to send * to the device as different command buffers *
*/ staticint cs_parser(struct hl_fpriv *hpriv, struct hl_cs_job *job)
{ struct hl_device *hdev = hpriv->hdev; struct hl_cs_parser parser; int rc;
if (is_cb_patched(hdev, job)) { if (!rc) {
job->patched_cb = parser.patched_cb;
job->job_cb_size = parser.patched_cb_size;
job->contains_dma_pkt = parser.contains_dma_pkt;
atomic_inc(&job->patched_cb->cs_cnt);
}
/* * Whether the parsing worked or not, we don't need the * original CB anymore because it was already parsed and * won't be accessed again for this CS
*/
atomic_dec(&job->user_cb->cs_cnt);
hl_cb_put(job->user_cb);
job->user_cb = NULL;
} elseif (!rc) {
job->job_cb_size = job->user_cb_size;
}
if (is_cb_patched(hdev, job)) {
hl_userptr_delete_list(hdev, &job->userptr_list);
/* * We might arrive here from rollback and patched CB wasn't * created, so we need to check it's not NULL
*/ if (job->patched_cb) {
atomic_dec(&job->patched_cb->cs_cnt);
hl_cb_put(job->patched_cb);
}
}
/* For H/W queue jobs, if a user CB was allocated by driver, * the user CB isn't released in cs_parser() and thus should be * released here. This is also true for INT queues jobs which were * allocated by driver.
*/ if (job->is_kernel_allocated_cb &&
(job->queue_type == QUEUE_TYPE_HW || job->queue_type == QUEUE_TYPE_INT)) {
atomic_dec(&job->user_cb->cs_cnt);
hl_cb_put(job->user_cb);
}
/* * This is the only place where there can be multiple threads * modifying the list at the same time
*/
spin_lock(&cs->job_lock);
list_del(&job->cs_node);
spin_unlock(&cs->job_lock);
hl_debugfs_remove_job(hdev, job);
/* We decrement reference only for a CS that gets completion * because the reference was incremented only for this kind of CS * right before it was scheduled. * * In staged submission, only the last CS marked as 'staged_last' * gets completion, hence its release function will be called from here. * As for all the rest CS's in the staged submission which do not get * completion, their CS reference will be decremented by the * 'staged_last' CS during the CS release flow. * All relevant PQ CI counters will be incremented during the CS release * flow by calling 'hl_hw_queue_update_ci'.
*/ if (cs_needs_completion(cs) &&
(job->queue_type == QUEUE_TYPE_EXT || job->queue_type == QUEUE_TYPE_HW)) {
/* In CS based completions, the timestamp is already available, * so no need to extract it from job
*/ if (hdev->asic_prop.completion_mode == HL_COMPLETION_MODE_JOB)
cs->completion_timestamp = job->timestamp;
cs_put(cs);
}
hl_cs_job_put(job);
}
/* * hl_staged_cs_find_first - locate the first CS in this staged submission * * @hdev: pointer to device structure * @cs_seq: staged submission sequence number * * @note: This function must be called under 'hdev->cs_mirror_lock' * * Find and return a CS pointer with the given sequence
*/ struct hl_cs *hl_staged_cs_find_first(struct hl_device *hdev, u64 cs_seq)
{ struct hl_cs *cs;
/* * staged_cs_get - get CS reference if this CS is a part of a staged CS * * @hdev: pointer to device structure * @cs: current CS * @cs_seq: staged submission sequence number * * Increment CS reference for every CS in this staged submission except for * the CS which get completion.
*/ staticvoid staged_cs_get(struct hl_device *hdev, struct hl_cs *cs)
{ /* Only the last CS in this staged submission will get a completion. * We must increment the reference for all other CS's in this * staged submission. * Once we get a completion we will release the whole staged submission.
*/ if (!cs->staged_last)
cs_get(cs);
}
/* * staged_cs_put - put a CS in case it is part of staged submission * * @hdev: pointer to device structure * @cs: CS to put * * This function decrements a CS reference (for a non completion CS)
*/ staticvoid staged_cs_put(struct hl_device *hdev, struct hl_cs *cs)
{ /* We release all CS's in a staged submission except the last * CS which we have never incremented its reference.
*/ if (!cs_needs_completion(cs))
cs_put(cs);
}
/* We need to handle tdr only once for the complete staged submission. * Hence, we choose the CS that reaches this function first which is * the CS marked as 'staged_last'. * In case single staged cs was submitted which has both first and last * indications, then "cs_find_first" below will return NULL, since we * removed the cs node from the list before getting here, * in such cases just continue with the cs to cancel it's TDR work.
*/ if (cs->staged_cs && cs->staged_last) {
first_cs = hl_staged_cs_find_first(hdev, cs->staged_sequence); if (first_cs)
cs = first_cs;
}
spin_unlock(&hdev->cs_mirror_lock);
/* Don't cancel TDR in case this CS was timedout because we might be * running from the TDR context
*/ if (cs->timedout || hdev->timeout_jiffies == MAX_SCHEDULE_TIMEOUT) return;
if (cs->tdr_active)
cancel_delayed_work_sync(&cs->work_tdr);
spin_lock(&hdev->cs_mirror_lock);
/* queue TDR for next CS */
list_for_each_entry(iter, &hdev->cs_mirror_list, mirror_node) if (cs_needs_timeout(iter)) {
next = iter; break;
}
/* * force_complete_multi_cs - complete all contexts that wait on multi-CS * * @hdev: pointer to habanalabs device structure
*/ staticvoid force_complete_multi_cs(struct hl_device *hdev)
{ int i;
for (i = 0; i < MULTI_CS_MAX_USER_CTX; i++) { struct multi_cs_completion *mcs_compl;
mcs_compl = &hdev->multi_cs_completion[i];
spin_lock(&mcs_compl->lock);
if (!mcs_compl->used) {
spin_unlock(&mcs_compl->lock); continue;
}
/* when calling force complete no context should be waiting on * multi-cS. * We are calling the function as a protection for such case * to free any pending context and print error message
*/
dev_err(hdev->dev, "multi-CS completion context %d still waiting when calling force completion\n",
i);
complete_all(&mcs_compl->completion);
spin_unlock(&mcs_compl->lock);
}
}
/* * complete_multi_cs - complete all waiting entities on multi-CS * * @hdev: pointer to habanalabs device structure * @cs: CS structure * The function signals a waiting entity that has an overlapping stream masters * with the completed CS. * For example: * - a completed CS worked on stream master QID 4, multi CS completion * is actively waiting on stream master QIDs 3, 5. don't send signal as no * common stream master QID * - a completed CS worked on stream master QID 4, multi CS completion * is actively waiting on stream master QIDs 3, 4. send signal as stream * master QID 4 is common
*/ staticvoid complete_multi_cs(struct hl_device *hdev, struct hl_cs *cs)
{ struct hl_fence *fence = cs->fence; int i;
/* in case of multi CS check for completion only for the first CS */ if (cs->staged_cs && !cs->staged_first) return;
for (i = 0; i < MULTI_CS_MAX_USER_CTX; i++) { struct multi_cs_completion *mcs_compl;
mcs_compl = &hdev->multi_cs_completion[i]; if (!mcs_compl->used) continue;
spin_lock(&mcs_compl->lock);
/* * complete if: * 1. still waiting for completion * 2. the completed CS has at least one overlapping stream * master with the stream masters in the completion
*/ if (mcs_compl->used &&
(fence->stream_master_qid_map &
mcs_compl->stream_master_qid_map)) { /* extract the timestamp only of first completed CS */ if (!mcs_compl->timestamp)
mcs_compl->timestamp = ktime_to_ns(fence->timestamp);
complete_all(&mcs_compl->completion);
/* * Setting mcs_handling_done inside the lock ensures * at least one fence have mcs_handling_done set to * true before wait for mcs finish. This ensures at * least one CS will be set as completed when polling * mcs fences.
*/
fence->mcs_handling_done = true;
}
spin_unlock(&mcs_compl->lock);
} /* In case CS completed without mcs completion initialized */
fence->mcs_handling_done = true;
}
staticinlinevoid cs_release_sob_reset_handler(struct hl_device *hdev, struct hl_cs *cs, struct hl_cs_compl *hl_cs_cmpl)
{ /* Skip this handler if the cs wasn't submitted, to avoid putting * the hw_sob twice, since this case already handled at this point, * also skip if the hw_sob pointer wasn't set.
*/ if (!hl_cs_cmpl->hw_sob || !cs->submitted) return;
spin_lock(&hl_cs_cmpl->lock);
/* * we get refcount upon reservation of signals or signal/wait cs for the * hw_sob object, and need to put it when the first staged cs * (which contains the encaps signals) or cs signal/wait is completed.
*/ if ((hl_cs_cmpl->type == CS_TYPE_SIGNAL) ||
(hl_cs_cmpl->type == CS_TYPE_WAIT) ||
(hl_cs_cmpl->type == CS_TYPE_COLLECTIVE_WAIT) ||
(!!hl_cs_cmpl->encaps_signals)) {
dev_dbg(hdev->dev, "CS 0x%llx type %d finished, sob_id: %d, sob_val: %u\n",
hl_cs_cmpl->cs_seq,
hl_cs_cmpl->type,
hl_cs_cmpl->hw_sob->sob_id,
hl_cs_cmpl->sob_val);
hw_sob_put(hl_cs_cmpl->hw_sob);
if (hl_cs_cmpl->type == CS_TYPE_COLLECTIVE_WAIT)
hdev->asic_funcs->reset_sob_group(hdev,
hl_cs_cmpl->sob_group);
}
/* * Although if we reached here it means that all external jobs have * finished, because each one of them took refcnt to CS, we still * need to go over the internal jobs and complete them. Otherwise, we * will have leaked memory and what's worse, the CS object (and * potentially the CTX object) could be released, while the JOB * still holds a pointer to them (but no reference).
*/
list_for_each_entry_safe(job, tmp, &cs->job_list, cs_node)
hl_complete_job(hdev, job);
if (!cs->submitted) { /* * In case the wait for signal CS was submitted, the fence put * occurs in init_signal_wait_cs() or collective_wait_init_cs() * right before hanging on the PQ.
*/ if (cs->type == CS_TYPE_WAIT ||
cs->type == CS_TYPE_COLLECTIVE_WAIT)
hl_fence_put(cs->signal_fence);
goto out;
}
/* Need to update CI for all queue jobs that does not get completion */
hl_hw_queue_update_ci(cs);
/* remove CS from CS mirror list */
spin_lock(&hdev->cs_mirror_lock);
list_del_init(&cs->mirror_node);
spin_unlock(&hdev->cs_mirror_lock);
cs_handle_tdr(hdev, cs);
if (cs->staged_cs) { /* the completion CS decrements reference for the entire * staged submission
*/ if (cs->staged_last) { struct hl_cs *staged_cs, *tmp_cs;
/* A staged CS will be a member in the list only after it * was submitted. We used 'cs_mirror_lock' when inserting * it to list so we will use it again when removing it
*/ if (cs->submitted) {
spin_lock(&hdev->cs_mirror_lock);
list_del(&cs->staged_cs_node);
spin_unlock(&hdev->cs_mirror_lock);
}
/* decrement refcount to handle when first staged cs * with encaps signals is completed.
*/ if (hl_cs_cmpl->encaps_signals)
kref_put(&hl_cs_cmpl->encaps_sig_hdl->refcount,
hl_encaps_release_handle_and_put_ctx);
}
/* We need to mark an error for not submitted because in that case * the hl fence release flow is different. Mainly, we don't need * to handle hw_sob for signal/wait
*/ if (cs->timedout)
cs->fence->error = -ETIMEDOUT; elseif (cs->aborted)
cs->fence->error = -EIO; elseif (!cs->submitted)
cs->fence->error = -EBUSY;
if (unlikely(cs->skip_reset_on_timeout)) {
dev_err(hdev->dev, "Command submission %llu completed after %llu (s)\n",
cs->sequence,
div_u64(jiffies - cs->submission_time_jiffies, HZ));
}
switch (cs->type) { case CS_TYPE_SIGNAL:
dev_err(hdev->dev, "Signal command submission %llu has not finished in %u seconds!\n",
cs->sequence, timeout_sec); break;
case CS_TYPE_WAIT:
dev_err(hdev->dev, "Wait command submission %llu has not finished in %u seconds!\n",
cs->sequence, timeout_sec); break;
case CS_TYPE_COLLECTIVE_WAIT:
dev_err(hdev->dev, "Collective Wait command submission %llu has not finished in %u seconds!\n",
cs->sequence, timeout_sec); break;
default:
dev_err(hdev->dev, "Command submission %llu has not finished in %u seconds!\n",
cs->sequence, timeout_sec); break;
}
rc = hl_state_dump(hdev); if (rc)
dev_err(hdev->dev, "Error during system state dump %d\n", rc);
cs_cmpl->cs_seq = ctx->cs_sequence;
other = ctx->cs_pending[cs_cmpl->cs_seq &
(hdev->asic_prop.max_pending_cs - 1)];
if (other && !completion_done(&other->completion)) { /* If the following statement is true, it means we have reached * a point in which only part of the staged submission was * submitted and we don't have enough room in the 'cs_pending' * array for the rest of the submission. * This causes a deadlock because this CS will never be * completed as it depends on future CS's for completion.
*/ if (other->cs_sequence == user_sequence)
dev_crit_ratelimited(hdev->dev, "Staged CS %llu deadlock due to lack of resources",
user_sequence);
dev_dbg_ratelimited(hdev->dev, "Rejecting CS because of too many in-flights CS\n");
atomic64_inc(&ctx->cs_counters.max_cs_in_flight_drop_cnt);
atomic64_inc(&cntr->max_cs_in_flight_drop_cnt);
rc = -EAGAIN; goto free_fence;
}
/* * release_reserved_encaps_signals() - release reserved encapsulated signals. * @hdev: pointer to habanalabs device structure * * Release reserved encapsulated signals which weren't un-reserved, or for which a CS with * encapsulated signals wasn't submitted and thus weren't released as part of CS roll-back. * For these signals need also to put the refcount of the H/W SOB which was taken at the * reservation.
*/ staticvoid release_reserved_encaps_signals(struct hl_device *hdev)
{ struct hl_ctx *ctx = hl_get_compute_ctx(hdev); struct hl_cs_encaps_sig_handle *handle; struct hl_encaps_signals_mgr *mgr;
u32 id;
if (!ctx) return;
mgr = &ctx->sig_mgr;
idr_for_each_entry(&mgr->handles, handle, id) if (handle->cs_seq == ULLONG_MAX)
kref_put(&handle->refcount, hl_encaps_release_handle_and_put_sob_ctx);
if (!skip_wq_flush) {
flush_workqueue(hdev->ts_free_obj_wq);
/* flush all completions before iterating over the CS mirror list in * order to avoid a race with the release functions
*/ for (i = 0 ; i < hdev->asic_prop.completion_queues_count ; i++)
flush_workqueue(hdev->cq_wq[i]);
flush_workqueue(hdev->cs_cmplt_wq);
}
/* Make sure we don't have leftovers in the CS mirror list */
list_for_each_entry_safe(cs, tmp, &hdev->cs_mirror_list, mirror_node) {
cs_get(cs);
cs->aborted = true;
dev_warn_ratelimited(hdev->dev, "Killing CS %d.%llu\n",
cs->ctx->asid, cs->sequence);
cs_rollback(hdev, cs);
cs_put(cs);
}
/* We iterate through the user interrupt requests and waking up all * user threads waiting for interrupt completion. We iterate the * list under a lock, this is why all user threads, once awake, * will wait on the same lock and will release the waiting object upon * unlock.
*/
for (i = 0 ; i < prop->user_interrupt_count ; i++) {
interrupt = &hdev->user_interrupt[i];
wake_pending_user_interrupt_threads(interrupt);
}
/* This must be checked here to prevent out-of-bounds access to * hw_queues_props array
*/ if (chunk->queue_index >= asic->max_queues) {
dev_err(hdev->dev, "Queue index %d is invalid\n",
chunk->queue_index); return -EINVAL;
}
if (hw_queue_prop->type == QUEUE_TYPE_NA) {
dev_err(hdev->dev, "Queue index %d is not applicable\n",
chunk->queue_index); return -EINVAL;
}
if (hw_queue_prop->binned) {
dev_err(hdev->dev, "Queue index %d is binned out\n",
chunk->queue_index); return -EINVAL;
}
if (hw_queue_prop->driver_only) {
dev_err(hdev->dev, "Queue index %d is restricted for the kernel driver\n",
chunk->queue_index); return -EINVAL;
}
/* When hw queue type isn't QUEUE_TYPE_HW, * USER_ALLOC_CB flag shall be referred as "don't care".
*/ if (hw_queue_prop->type == QUEUE_TYPE_HW) { if (chunk->cs_chunk_flags & HL_CS_CHUNK_FLAGS_USER_ALLOC_CB) { if (!(hw_queue_prop->cb_alloc_flags & CB_ALLOC_USER)) {
dev_err(hdev->dev, "Queue index %d doesn't support user CB\n",
chunk->queue_index); return -EINVAL;
}
*is_kernel_allocated_cb = false;
} else { if (!(hw_queue_prop->cb_alloc_flags &
CB_ALLOC_KERNEL)) {
dev_err(hdev->dev, "Queue index %d doesn't support kernel CB\n",
chunk->queue_index); return -EINVAL;
}
if (num_chunks > HL_MAX_JOBS_PER_CS) {
atomic64_inc(&ctx->cs_counters.validation_drop_cnt);
atomic64_inc(&hdev->aggregated_cs_counters.validation_drop_cnt);
dev_err(hdev->dev, "Number of chunks can NOT be larger than %d\n",
HL_MAX_JOBS_PER_CS); return -EINVAL;
}
*cs_chunk_array = kmalloc_array(num_chunks, sizeof(**cs_chunk_array),
GFP_ATOMIC); if (!*cs_chunk_array)
*cs_chunk_array = kmalloc_array(num_chunks, sizeof(**cs_chunk_array), GFP_KERNEL); if (!*cs_chunk_array) {
atomic64_inc(&ctx->cs_counters.out_of_mem_drop_cnt);
atomic64_inc(&hdev->aggregated_cs_counters.out_of_mem_drop_cnt); return -ENOMEM;
}
size_to_copy = num_chunks * sizeof(struct hl_cs_chunk); if (copy_from_user(*cs_chunk_array, chunks, size_to_copy)) {
atomic64_inc(&ctx->cs_counters.validation_drop_cnt);
atomic64_inc(&hdev->aggregated_cs_counters.validation_drop_cnt);
dev_err(hdev->dev, "Failed to copy cs chunk array from user\n");
kfree(*cs_chunk_array); return -EFAULT;
}
if (cs->staged_first) { /* Staged CS sequence is the first CS sequence */
INIT_LIST_HEAD(&cs->staged_cs_node);
cs->staged_sequence = cs->sequence;
if (cs->encaps_signals)
cs->encaps_sig_hdl_id = encaps_signal_handle;
} else { /* User sequence will be validated in 'hl_hw_queue_schedule_cs' * under the cs_mirror_lock
*/
cs->staged_sequence = sequence;
}
/* Increment CS reference if needed */
staged_cs_get(hdev, cs);
cs->staged_cs = true;
return 0;
}
static u32 get_stream_master_qid_mask(struct hl_device *hdev, u32 qid)
{ int i;
for (i = 0; i < hdev->stream_master_qid_arr_size; i++) if (qid == hdev->stream_master_qid_arr[i]) return BIT(i);
/* If this is a staged submission we must return the staged sequence * rather than the internal CS sequence
*/ if (cs->staged_cs)
*cs_seq = cs->staged_sequence;
/* Validate ALL the CS chunks before submitting the CS */ for (i = 0 ; i < num_chunks ; i++) { struct hl_cs_chunk *chunk = &cs_chunk_array[i]; enum hl_queue_type queue_type; bool is_kernel_allocated_cb;
/* * store which stream are being used for external/HW * queues of this CS
*/ if (hdev->supports_wait_for_multi_cs)
stream_master_qid_map |=
get_stream_master_qid_mask(hdev,
chunk->queue_index);
}
if (queue_type == QUEUE_TYPE_HW)
using_hw_queues = true;
job = hl_cs_allocate_job(hdev, queue_type,
is_kernel_allocated_cb); if (!job) {
atomic64_inc(&ctx->cs_counters.out_of_mem_drop_cnt);
atomic64_inc(&cntr->out_of_mem_drop_cnt);
dev_err(hdev->dev, "Failed to allocate a new job\n");
rc = -ENOMEM; if (is_kernel_allocated_cb) goto release_cb;
/* * Increment CS reference. When CS reference is 0, CS is * done and can be signaled to user and free all its resources * Only increment for JOB on external or H/W queues, because * only for those JOBs we get completion
*/ if (cs_needs_completion(cs) &&
(job->queue_type == QUEUE_TYPE_EXT ||
job->queue_type == QUEUE_TYPE_HW))
cs_get(cs);
hl_debugfs_add_job(hdev, job);
rc = cs_parser(hpriv, job); if (rc) {
atomic64_inc(&ctx->cs_counters.parsing_drop_cnt);
atomic64_inc(&cntr->parsing_drop_cnt);
dev_err(hdev->dev, "Failed to parse JOB %d.%llu.%d, err %d, rejecting the CS\n",
cs->ctx->asid, cs->sequence, job->id, rc); goto free_cs_object;
}
}
/* We allow a CS with any queue type combination as long as it does * not get a completion
*/ if (int_queues_only && cs_needs_completion(cs)) {
atomic64_inc(&ctx->cs_counters.validation_drop_cnt);
atomic64_inc(&cntr->validation_drop_cnt);
dev_err(hdev->dev, "Reject CS %d.%llu since it contains only internal queues jobs and needs completion\n",
cs->ctx->asid, cs->sequence);
rc = -EINVAL; goto free_cs_object;
}
if (using_hw_queues)
INIT_WORK(&cs->finish_work, cs_completion);
/* * store the (external/HW queues) streams used by the CS in the * fence object for multi-CS completion
*/ if (hdev->supports_wait_for_multi_cs)
cs->fence->stream_master_qid_map = stream_master_qid_map;
rc = hl_hw_queue_schedule_cs(cs); if (rc) { if (rc != -EAGAIN)
dev_err(hdev->dev, "Failed to submit CS %d.%llu to H/W queues, error %d\n",
cs->ctx->asid, cs->sequence, rc); goto free_cs_object;
}
release_cb:
atomic_dec(&cb->cs_cnt);
hl_cb_put(cb);
free_cs_object:
cs_rollback(hdev, cs);
*cs_seq = ULLONG_MAX; /* The path below is both for good and erroneous exits */
put_cs: /* We finished with the CS in this function, so put the ref */
cs_put(cs);
free_cs_chunk_array:
kfree(cs_chunk_array);
out: return rc;
}
if (hdev->supports_ctx_switch)
do_ctx_switch = atomic_cmpxchg(&ctx->thread_ctx_switch_token, 1, 0);
if (do_ctx_switch || (args->in.cs_flags & HL_CS_FLAGS_FORCE_RESTORE)) {
mutex_lock(&hpriv->restore_phase_mutex);
if (do_ctx_switch) {
rc = hdev->asic_funcs->context_switch(hdev, ctx->asid); if (rc) {
dev_err_ratelimited(hdev->dev, "Failed to switch to context %d, rejecting CS! %d\n",
ctx->asid, rc); /* * If we timedout, or if the device is not IDLE * while we want to do context-switch (-EBUSY), * we need to soft-reset because QMAN is * probably stuck. However, we can't call to * reset here directly because of deadlock, so * need to do it at the very end of this * function
*/ if ((rc == -ETIMEDOUT) || (rc == -EBUSY))
need_soft_reset = true;
mutex_unlock(&hpriv->restore_phase_mutex); goto out;
}
}
/* * hl_cs_signal_sob_wraparound_handler: handle SOB value wrapaound case. * if the SOB value reaches the max value move to the other SOB reserved * to the queue. * @hdev: pointer to device structure * @q_idx: stream queue index * @hw_sob: the H/W SOB used in this signal CS. * @count: signals count * @encaps_sig: tells whether it's reservation for encaps signals or not. * * Note that this function must be called while hw_queues_lock is taken.
*/ int hl_cs_signal_sob_wraparound_handler(struct hl_device *hdev, u32 q_idx, struct hl_hw_sob **hw_sob, u32 count, bool encaps_sig)
/* check for wraparound */ if (prop->next_sob_val + count >= HL_MAX_SOB_VAL) { /* * Decrement as we reached the max value. * The release function won't be called here as we've * just incremented the refcount right before calling this * function.
*/
hw_sob_put_err(sob);
/* * check the other sob value, if it still in use then fail * otherwise make the switch
*/
other_sob_offset = (prop->curr_sob_offset + 1) % HL_RSVD_SOBS;
other_sob = &prop->hw_sob[other_sob_offset];
/* * next_sob_val always points to the next available signal * in the sob, so in encaps signals it will be the next one * after reserving the required amount.
*/ if (encaps_sig)
prop->next_sob_val = count + 1; else
prop->next_sob_val = count;
/* only two SOBs are currently in use */
prop->curr_sob_offset = other_sob_offset;
*hw_sob = other_sob;
/* * check if other_sob needs reset, then do it before using it * for the reservation or the next signal cs. * we do it here, and for both encaps and regular signal cs * cases in order to avoid possible races of two kref_put * of the sob which can occur at the same time if we move the * sob reset(kref_put) to cs_do_release function. * in addition, if we have combination of cs signal and * encaps, and at the point we need to reset the sob there was * no more reservations and only signal cs keep coming, * in such case we need signal_cs to put the refcount and * reset the sob.
*/ if (other_sob->need_reset)
hw_sob_put(other_sob);
if (encaps_sig) { /* set reset indication for the sob */
sob->need_reset = true;
hw_sob_get(other_sob);
}
dev_dbg(hdev->dev, "switched to SOB %d, q_idx: %d\n",
prop->curr_sob_offset, q_idx);
} else {
prop->next_sob_val += count;
}
if (encaps_signals) {
*signal_seq = chunk->encaps_signal_seq; return 0;
}
signal_seq_arr_len = chunk->num_signal_seq_arr;
/* currently only one signal seq is supported */ if (signal_seq_arr_len != 1) {
atomic64_inc(&ctx->cs_counters.validation_drop_cnt);
atomic64_inc(&hdev->aggregated_cs_counters.validation_drop_cnt);
dev_err(hdev->dev, "Wait for signal CS supports only one signal CS seq\n"); return -EINVAL;
}
signal_seq_arr = kmalloc_array(signal_seq_arr_len, sizeof(*signal_seq_arr),
GFP_ATOMIC); if (!signal_seq_arr)
signal_seq_arr = kmalloc_array(signal_seq_arr_len, sizeof(*signal_seq_arr),
GFP_KERNEL); if (!signal_seq_arr) {
atomic64_inc(&ctx->cs_counters.out_of_mem_drop_cnt);
atomic64_inc(&hdev->aggregated_cs_counters.out_of_mem_drop_cnt); return -ENOMEM;
}
size_to_copy = signal_seq_arr_len * sizeof(*signal_seq_arr); if (copy_from_user(signal_seq_arr,
u64_to_user_ptr(chunk->signal_seq_arr),
size_to_copy)) {
atomic64_inc(&ctx->cs_counters.validation_drop_cnt);
atomic64_inc(&hdev->aggregated_cs_counters.validation_drop_cnt);
dev_err(hdev->dev, "Failed to copy signal seq array from user\n");
rc = -EFAULT; goto out;
}
/* currently it is guaranteed to have only one signal seq */
*signal_seq = signal_seq_arr[0];
job = hl_cs_allocate_job(hdev, q_type, true); if (!job) {
atomic64_inc(&ctx->cs_counters.out_of_mem_drop_cnt);
atomic64_inc(&cntr->out_of_mem_drop_cnt);
dev_err(hdev->dev, "Failed to allocate a new job\n"); return -ENOMEM;
}
if ((cs->type == CS_TYPE_WAIT || cs->type == CS_TYPE_COLLECTIVE_WAIT)
&& cs->encaps_signals)
job->encaps_sig_wait_offset = encaps_signal_offset; /* * No need in parsing, user CB is the patched CB. * We call hl_cb_destroy() out of two reasons - we don't need the CB in * the CB idr anymore and to decrement its refcount as it was * incremented inside hl_cb_kernel_create().
*/
job->patched_cb = job->user_cb;
job->job_cb_size = job->user_cb_size;
hl_cb_destroy(&hdev->kernel_mem_mgr, cb->buf->handle);
/* increment refcount as for external queues we get completion */
cs_get(cs);
if (!hw_queue_prop->supports_sync_stream) {
dev_err(hdev->dev, "Queue index %d does not support sync stream operations\n",
q_idx);
rc = -EINVAL; goto out;
}
/* * Increment the SOB value by count by user request * to reserve those signals * check if the signals amount to reserve is not exceeding the max sob * value, if yes then switch sob.
*/
rc = hl_cs_signal_sob_wraparound_handler(hdev, q_idx, &hw_sob, count, true); if (rc) {
dev_err(hdev->dev, "Failed to switch SOB\n");
hdev->asic_funcs->hw_queues_unlock(hdev);
rc = -EINVAL; goto remove_idr;
} /* set the hw_sob to the handle after calling the sob wraparound handler * since sob could have changed.
*/
handle->hw_sob = hw_sob;
/* store the current sob value for unreserve validity check, and * signal offset support
*/
handle->pre_sob_val = prop->next_sob_val - handle->count;
/* Check if sob_val got out of sync due to other * signal submission requests which were handled * between the reserve-unreserve calls or SOB switch * upon reaching SOB max value.
*/ if (encaps_sig_hdl->pre_sob_val + encaps_sig_hdl->count
!= prop->next_sob_val ||
sob_addr != encaps_sig_hdl->hw_sob->sob_addr) {
dev_err(hdev->dev, "Cannot unreserve signals, SOB val ran out of sync, expected: %u, actual val: %u\n",
encaps_sig_hdl->pre_sob_val,
(prop->next_sob_val - encaps_sig_hdl->count));
if (!hw_queue_prop->supports_sync_stream) {
atomic64_inc(&ctx->cs_counters.validation_drop_cnt);
atomic64_inc(&cntr->validation_drop_cnt);
dev_err(hdev->dev, "Queue index %d does not support sync stream operations\n",
q_idx);
rc = -EINVAL; goto free_cs_chunk_array;
}
if (cs_type == CS_TYPE_COLLECTIVE_WAIT) { if (!(hw_queue_prop->collective_mode == HL_COLLECTIVE_MASTER)) {
atomic64_inc(&ctx->cs_counters.validation_drop_cnt);
atomic64_inc(&cntr->validation_drop_cnt);
dev_err(hdev->dev, "Queue index %d is invalid\n", q_idx);
rc = -EINVAL; goto free_cs_chunk_array;
}
if (!hdev->nic_ports_mask) {
atomic64_inc(&ctx->cs_counters.validation_drop_cnt);
atomic64_inc(&cntr->validation_drop_cnt);
dev_err(hdev->dev, "Collective operations not supported when NIC ports are disabled");
rc = -EINVAL; goto free_cs_chunk_array;
}
if (is_wait_cs) {
rc = cs_ioctl_extract_signal_seq(hdev, chunk, &signal_seq,
ctx, cs_encaps_signals); if (rc) goto free_cs_chunk_array;
if (cs_encaps_signals) { /* check if cs sequence has encapsulated * signals handle
*/ struct idr *idp;
u32 id;
spin_lock(&ctx->sig_mgr.lock);
idp = &ctx->sig_mgr.handles;
idr_for_each_entry(idp, encaps_sig_hdl, id) { if (encaps_sig_hdl->cs_seq == signal_seq) { /* get refcount to protect removing this handle from idr, * needed when multiple wait cs are used with offset * to wait on reserved encaps signals. * Since kref_put of this handle is executed outside the * current lock, it is possible that the handle refcount * is 0 but it yet to be removed from the list. In this * case need to consider the handle as not valid.
*/ if (kref_get_unless_zero(&encaps_sig_hdl->refcount))
handle_found = true; break;
}
}
spin_unlock(&ctx->sig_mgr.lock);
if (!handle_found) { /* treat as signal CS already finished */
dev_dbg(hdev->dev, "Cannot find encapsulated signals handle for seq 0x%llx\n",
signal_seq);
rc = 0; goto free_cs_chunk_array;
}
/* validate also the signal offset value */ if (chunk->encaps_signal_offset >
encaps_sig_hdl->count) {
dev_err(hdev->dev, "offset(%u) value exceed max reserved signals count(%u)!\n",
chunk->encaps_signal_offset,
encaps_sig_hdl->count);
rc = -EINVAL; goto free_cs_chunk_array;
}
}
sig_fence = hl_ctx_get_fence(ctx, signal_seq); if (IS_ERR(sig_fence)) {
atomic64_inc(&ctx->cs_counters.validation_drop_cnt);
atomic64_inc(&cntr->validation_drop_cnt);
dev_err(hdev->dev, "Failed to get signal CS with seq 0x%llx\n",
signal_seq);
rc = PTR_ERR(sig_fence); goto free_cs_chunk_array;
}
if (!sig_fence) { /* signal CS already finished */
rc = 0; goto free_cs_chunk_array;
}
if (sig_waitcs_cmpl->type != CS_TYPE_SIGNAL &&
!staged_cs_with_encaps_signals) {
atomic64_inc(&ctx->cs_counters.validation_drop_cnt);
atomic64_inc(&cntr->validation_drop_cnt);
dev_err(hdev->dev, "CS seq 0x%llx is not of a signal/encaps-signal CS\n",
signal_seq);
hl_fence_put(sig_fence);
rc = -EINVAL; goto free_cs_chunk_array;
}
if (completion_done(&sig_fence->completion)) { /* signal CS already finished */
hl_fence_put(sig_fence);
rc = 0; goto free_cs_chunk_array;
}
}
rc = allocate_cs(hdev, ctx, cs_type, ULLONG_MAX, &cs, flags, timeout); if (rc) { if (is_wait_cs)
hl_fence_put(sig_fence);
goto free_cs_chunk_array;
}
/* * Save the signal CS fence for later initialization right before * hanging the wait CS on the queue. * for encaps signals case, we save the cs sequence and handle pointer * for later initialization.
*/ if (is_wait_cs) {
cs->signal_fence = sig_fence; /* store the handle pointer, so we don't have to * look for it again, later on the flow * when we need to set SOB info in hw_queue.
*/ if (cs->encaps_signals)
cs->encaps_sig_hdl = encaps_sig_hdl;
}
if (q_type == QUEUE_TYPE_HW)
INIT_WORK(&cs->finish_work, cs_completion);
rc = hl_hw_queue_schedule_cs(cs); if (rc) { /* In case wait cs failed here, it means the signal cs * already completed. we want to free all it's related objects * but we don't want to fail the ioctl.
*/ if (is_wait_cs)
rc = 0; elseif (rc != -EAGAIN)
dev_err(hdev->dev, "Failed to submit CS %d.%llu to H/W queues, error %d\n",
ctx->asid, cs->sequence, rc); goto free_cs_object;
}
rc = HL_CS_STATUS_SUCCESS; if (is_wait_cs)
wait_cs_submitted = true; goto put_cs;
free_cs_object:
cs_rollback(hdev, cs);
*cs_seq = ULLONG_MAX; /* The path below is both for good and erroneous exits */
put_cs: /* We finished with the CS in this function, so put the ref */
cs_put(cs);
free_cs_chunk_array: if (!wait_cs_submitted && cs_encaps_signals && handle_found && is_wait_cs)
kref_put(&encaps_sig_hdl->refcount, hl_encaps_release_handle_and_put_ctx);
kfree(cs_chunk_array);
out: return rc;
}
/* In case this is a staged CS, user should supply the CS sequence */ if ((flags & HL_CS_FLAGS_STAGED_SUBMISSION) &&
!(flags & HL_CS_FLAGS_STAGED_SUBMISSION_FIRST))
cs_seq = args->in.seq;
if (IS_ERR(fence)) {
rc = PTR_ERR(fence); if (rc == -EINVAL)
dev_notice_ratelimited(hdev->dev, "Can't wait on CS %llu because current CS is at seq %llu\n",
seq, ctx->cs_sequence); return rc;
}
if (!fence) { if (!hl_pop_cs_outcome(&ctx->outcome_store, seq, ×tamp_kt, &error)) {
dev_dbg(hdev->dev, "Can't wait on seq %llu because current CS is at seq %llu (Fence is gone)\n",
seq, ctx->cs_sequence);
*status = CS_WAIT_STATUS_GONE; return 0;
}
/* * hl_cs_poll_fences - iterate CS fences to check for CS completion * * @mcs_data: multi-CS internal data * @mcs_compl: multi-CS completion structure * * @return 0 on success, otherwise non 0 error code * * The function iterates on all CS sequence in the list and set bit in * completion_bitmap for each completed CS. * While iterating, the function sets the stream map of each fence in the fence * array in the completion QID stream map to be used by CSs to perform * completion to the multi-CS context. * This function shall be called after taking context ref
*/ staticint hl_cs_poll_fences(struct multi_cs_data *mcs_data, struct multi_cs_completion *mcs_compl)
{ struct hl_fence **fence_ptr = mcs_data->fence_arr; struct hl_device *hdev = mcs_data->ctx->hdev; int i, rc, arr_len = mcs_data->arr_len;
u64 *seq_arr = mcs_data->seq_arr;
ktime_t max_ktime, first_cs_time; enum hl_cs_wait_status status;
/* get all fences under the same lock */
rc = hl_ctx_get_fences(mcs_data->ctx, seq_arr, fence_ptr, arr_len); if (rc) return rc;
/* * re-initialize the completion here to handle 2 possible cases: * 1. CS will complete the multi-CS prior clearing the completion. in which * case the fence iteration is guaranteed to catch the CS completion. * 2. the completion will occur after re-init of the completion. * in which case we will wake up immediately in wait_for_completion.
*/
reinit_completion(&mcs_compl->completion);
/* * set to maximum time to verify timestamp is valid: if at the end * this value is maintained- no timestamp was updated
*/
max_ktime = ktime_set(KTIME_SEC_MAX, 0);
first_cs_time = max_ktime;
for (i = 0; i < arr_len; i++, fence_ptr++) { struct hl_fence *fence = *fence_ptr;
/* * In order to prevent case where we wait until timeout even though a CS associated * with the multi-CS actually completed we do things in the below order: * 1. for each fence set it's QID map in the multi-CS completion QID map. This way * any CS can, potentially, complete the multi CS for the specific QID (note * that once completion is initialized, calling complete* and then wait on the * completion will cause it to return at once) * 2. only after allowing multi-CS completion for the specific QID we check whether * the specific CS already completed (and thus the wait for completion part will * be skipped). if the CS not completed it is guaranteed that completing CS will * wake up the completion.
*/ if (fence)
mcs_compl->stream_master_qid_map |= fence->stream_master_qid_map;
/* * function won't sleep as it is called with timeout 0 (i.e. * poll the fence)
*/
rc = hl_wait_for_fence(mcs_data->ctx, seq_arr[i], fence, &status, 0, NULL); if (rc) {
dev_err(hdev->dev, "wait_for_fence error :%d for CS seq %llu\n",
rc, seq_arr[i]); break;
}
switch (status) { case CS_WAIT_STATUS_BUSY: /* CS did not finished, QID to wait on already stored */ break; case CS_WAIT_STATUS_COMPLETED: /* * Using mcs_handling_done to avoid possibility of mcs_data * returns to user indicating CS completed before it finished * all of its mcs handling, to avoid race the next time the * user waits for mcs. * note: when reaching this case fence is definitely not NULL * but NULL check was added to overcome static analysis
*/ if (fence && !fence->mcs_handling_done) { /* * in case multi CS is completed but MCS handling not done * we "complete" the multi CS to prevent it from waiting * until time-out and the "multi-CS handling done" will have * another chance at the next iteration
*/
complete_all(&mcs_compl->completion); break;
}
mcs_data->completion_bitmap |= BIT(i); /* * For all completed CSs we take the earliest timestamp. * For this we have to validate that the timestamp is * earliest of all timestamps so far.
*/ if (fence && mcs_data->update_ts &&
(ktime_compare(fence->timestamp, first_cs_time) < 0))
first_cs_time = fence->timestamp; break; case CS_WAIT_STATUS_GONE:
mcs_data->update_ts = false;
mcs_data->gone_cs = true; /* * It is possible to get an old sequence numbers from user * which related to already completed CSs and their fences * already gone. In this case, CS set as completed but * no need to consider its QID for mcs completion.
*/
mcs_data->completion_bitmap |= BIT(i); break; default:
dev_err(hdev->dev, "Invalid fence status\n");
rc = -EINVAL; break;
}
}
hl_fences_put(mcs_data->fence_arr, arr_len);
if (mcs_data->update_ts &&
(ktime_compare(first_cs_time, max_ktime) != 0))
mcs_data->timestamp = ktime_to_ns(first_cs_time);
/* * If the value in nanoseconds is larger than 64 bit, use the largest * 64 bit value.
*/ if (usecs >= ((u64)(U64_MAX / NSEC_PER_USEC))) return nsecs_to_jiffies(U64_MAX);
return nsecs_to_jiffies(usecs * NSEC_PER_USEC);
}
/* * hl_wait_multi_cs_completion_init - init completion structure * * @hdev: pointer to habanalabs device structure * @stream_master_bitmap: stream master QIDs map, set bit indicates stream * master QID to wait on * * @return valid completion struct pointer on success, otherwise error pointer * * up to MULTI_CS_MAX_USER_CTX calls can be done concurrently to the driver. * the function gets the first available completion (by marking it "used") * and initialize its values.
*/ staticstruct multi_cs_completion *hl_wait_multi_cs_completion_init(struct hl_device *hdev)
{ struct multi_cs_completion *mcs_compl; int i;
/* find free multi_cs completion structure */ for (i = 0; i < MULTI_CS_MAX_USER_CTX; i++) {
mcs_compl = &hdev->multi_cs_completion[i];
spin_lock(&mcs_compl->lock); if (!mcs_compl->used) {
mcs_compl->used = 1;
mcs_compl->timestamp = 0; /* * init QID map to 0 to avoid completion by CSs. the actual QID map * to multi-CS CSs will be set incrementally at a later stage
*/
mcs_compl->stream_master_qid_map = 0;
spin_unlock(&mcs_compl->lock); break;
}
spin_unlock(&mcs_compl->lock);
}
if (i == MULTI_CS_MAX_USER_CTX) {
dev_err(hdev->dev, "no available multi-CS completion structure\n"); return ERR_PTR(-ENOMEM);
} return mcs_compl;
}
/* * hl_wait_multi_cs_completion_fini - return completion structure and set as * unused * * @mcs_compl: pointer to the completion structure
*/ staticvoid hl_wait_multi_cs_completion_fini( struct multi_cs_completion *mcs_compl)
{ /* * free completion structure, do it under lock to be in-sync with the * thread that signals completion
*/
spin_lock(&mcs_compl->lock);
mcs_compl->used = 0;
spin_unlock(&mcs_compl->lock);
}
/* * hl_wait_multi_cs_completion - wait for first CS to complete * * @mcs_data: multi-CS internal data * * @return 0 on success, otherwise non 0 error code
*/ staticint hl_wait_multi_cs_completion(struct multi_cs_data *mcs_data, struct multi_cs_completion *mcs_compl)
{ long completion_rc;
if (completion_rc == -ERESTARTSYS) return completion_rc;
mcs_data->wait_status = completion_rc;
return 0;
}
/* * hl_multi_cs_completion_init - init array of multi-CS completion structures * * @hdev: pointer to habanalabs device structure
*/ void hl_multi_cs_completion_init(struct hl_device *hdev)
{ struct multi_cs_completion *mcs_cmpl; int i;
for (i = 0; i < MULTI_CS_MAX_USER_CTX; i++) {
mcs_cmpl = &hdev->multi_cs_completion[i];
mcs_cmpl->used = 0;
spin_lock_init(&mcs_cmpl->lock);
init_completion(&mcs_cmpl->completion);
}
}
/* * hl_multi_cs_wait_ioctl - implementation of the multi-CS wait ioctl * * @hpriv: pointer to the private data of the fd * @data: pointer to multi-CS wait ioctl in/out args *
*/ staticint hl_multi_cs_wait_ioctl(struct hl_fpriv *hpriv, void *data)
{ struct multi_cs_completion *mcs_compl; struct hl_device *hdev = hpriv->hdev; struct multi_cs_data mcs_data = {}; union hl_wait_cs_args *args = data; struct hl_ctx *ctx = hpriv->ctx; struct hl_fence **fence_arr; void __user *seq_arr;
u32 size_to_copy;
u64 *cs_seq_arr;
u8 seq_arr_len; int rc, i;
for (i = 0 ; i < sizeof(args->in.pad) ; i++) if (args->in.pad[i]) {
dev_dbg(hdev->dev, "Padding bytes must be 0\n"); return -EINVAL;
}
if (!hdev->supports_wait_for_multi_cs) {
dev_err(hdev->dev, "Wait for multi CS is not supported\n"); return -EPERM;
}
seq_arr_len = args->in.seq_arr_len;
if (seq_arr_len > HL_WAIT_MULTI_CS_LIST_MAX_LEN) {
dev_err(hdev->dev, "Can wait only up to %d CSs, input sequence is of length %u\n",
HL_WAIT_MULTI_CS_LIST_MAX_LEN, seq_arr_len); return -EINVAL;
}
/* allocate memory for sequence array */
cs_seq_arr =
kmalloc_array(seq_arr_len, sizeof(*cs_seq_arr), GFP_KERNEL); if (!cs_seq_arr) return -ENOMEM;
/* copy CS sequence array from user */
seq_arr = (void __user *) (uintptr_t) args->in.seq;
size_to_copy = seq_arr_len * sizeof(*cs_seq_arr); if (copy_from_user(cs_seq_arr, seq_arr, size_to_copy)) {
dev_err(hdev->dev, "Failed to copy multi-cs sequence array from user\n");
rc = -EFAULT; goto free_seq_arr;
}
/* allocate array for the fences */
fence_arr = kmalloc_array(seq_arr_len, sizeof(struct hl_fence *), GFP_KERNEL); if (!fence_arr) {
rc = -ENOMEM; goto free_seq_arr;
}
/* initialize the multi-CS internal data */
mcs_data.ctx = ctx;
mcs_data.seq_arr = cs_seq_arr;
mcs_data.fence_arr = fence_arr;
mcs_data.arr_len = seq_arr_len;
hl_ctx_get(ctx);
/* wait (with timeout) for the first CS to be completed */
mcs_data.timeout_jiffies = hl_usecs64_to_jiffies(args->in.timeout_us);
mcs_compl = hl_wait_multi_cs_completion_init(hdev); if (IS_ERR(mcs_compl)) {
rc = PTR_ERR(mcs_compl); goto put_ctx;
}
/* poll all CS fences, extract timestamp */
mcs_data.update_ts = true;
rc = hl_cs_poll_fences(&mcs_data, mcs_compl); /* * skip wait for CS completion when one of the below is true: * - an error on the poll function * - one or more CS in the list completed * - the user called ioctl with timeout 0
*/ if (rc || mcs_data.completion_bitmap || !args->in.timeout_us) goto completion_fini;
while (true) {
rc = hl_wait_multi_cs_completion(&mcs_data, mcs_compl); if (rc || (mcs_data.wait_status == 0)) break;
/* * poll fences once again to update the CS map. * no timestamp should be updated this time.
*/
mcs_data.update_ts = false;
rc = hl_cs_poll_fences(&mcs_data, mcs_compl);
if (rc || mcs_data.completion_bitmap) break;
/* * if hl_wait_multi_cs_completion returned before timeout (i.e. * it got a completion) it either got completed by CS in the multi CS list * (in which case the indication will be non empty completion_bitmap) or it * got completed by CS submitted to one of the shared stream master but * not in the multi CS list (in which case we should wait again but modify * the timeout and set timestamp as zero to let a CS related to the current * multi-CS set a new, relevant, timestamp)
*/
mcs_data.timeout_jiffies = mcs_data.wait_status;
mcs_compl->timestamp = 0;
}
if (mcs_data.completion_bitmap) {
args->out.status = HL_WAIT_CS_STATUS_COMPLETED;
args->out.cs_completion_map = mcs_data.completion_bitmap;
/* if timestamp not 0- it's valid */ if (mcs_data.timestamp) {
args->out.timestamp_nsec = mcs_data.timestamp;
args->out.flags |= HL_WAIT_CS_STATUS_FLAG_TIMESTAMP_VLD;
}
/* update if some CS was gone */ if (!mcs_data.timestamp)
args->out.flags |= HL_WAIT_CS_STATUS_FLAG_GONE;
} else {
args->out.status = HL_WAIT_CS_STATUS_BUSY;
}
if (rc == -ERESTARTSYS) {
dev_err_ratelimited(hdev->dev, "user process got signal while waiting for CS handle %llu\n",
seq); return -EINTR;
}
memset(args, 0, sizeof(*args));
if (rc) { if (rc == -ETIMEDOUT) {
dev_err_ratelimited(hdev->dev, "CS %llu has timed-out while user process is waiting for it\n",
seq);
args->out.status = HL_WAIT_CS_STATUS_TIMEDOUT;
} elseif (rc == -EIO) {
dev_err_ratelimited(hdev->dev, "CS %llu has been aborted while user process is waiting for it\n",
seq);
args->out.status = HL_WAIT_CS_STATUS_ABORTED;
} return rc;
}
if (timestamp) {
args->out.flags |= HL_WAIT_CS_STATUS_FLAG_TIMESTAMP_VLD;
args->out.timestamp_nsec = timestamp;
}
switch (status) { case CS_WAIT_STATUS_GONE:
args->out.flags |= HL_WAIT_CS_STATUS_FLAG_GONE;
fallthrough; case CS_WAIT_STATUS_COMPLETED:
args->out.status = HL_WAIT_CS_STATUS_COMPLETED; break; case CS_WAIT_STATUS_BUSY: default:
args->out.status = HL_WAIT_CS_STATUS_BUSY; break;
}
/* Validate ts_offset not exceeding last max */ if (*req_event_record >= ts_cb_last) {
dev_err(dev, "Ts offset(%llu) exceeds max CB offset(0x%llx)\n",
ts_offset, (u64)(uintptr_t)ts_cb_last); return -EINVAL;
}
if (need_lock)
spin_unlock_irqrestore(&interrupt->ts_list_lock, flags);
/* Put refcounts that were taken when we registered the event */ if (ts_rec_found) {
hl_mmap_mem_buf_put(record->ts_reg_info.buf);
hl_cb_put(record->ts_reg_info.cq_cb);
}
}
rc = validate_and_get_ts_record(data->buf->mmg->dev, ts_buff, data->ts_offset,
&req_offset_record); if (rc) return rc;
/* In case the node already registered, need to unregister first then re-use */ if (req_offset_record->ts_reg_info.in_use) { /* * Since interrupt here can be different than the one the node currently registered * on, and we don't want to lock two lists while we're doing unregister, so * unlock the new interrupt wait list here and acquire the lock again after you done
*/ if (data->interrupt->interrupt_id !=
req_offset_record->ts_reg_info.interrupt->interrupt_id) {
if (need_lock)
spin_lock_irqsave(&data->interrupt->ts_list_lock, *flags);
}
/* Fill up the new registration node info and add it to the list */
req_offset_record->ts_reg_info.in_use = true;
req_offset_record->ts_reg_info.buf = data->buf;
req_offset_record->ts_reg_info.timestamp_kernel_addr =
(u64 *) ts_buff->user_buff_address + data->ts_offset;
req_offset_record->ts_reg_info.interrupt = data->interrupt;
set_record_cq_info(req_offset_record, data->cq_cb, data->cq_offset,
data->target_value);
/* get ts buffer record */
rc = ts_get_and_handle_kernel_record(hdev, ctx, data, &flags, &pend); if (rc) {
spin_unlock_irqrestore(&data->interrupt->ts_list_lock, flags); goto put_ts_buff;
}
/* We check for completion value as interrupt could have been received * before we add the timestamp node to the ts list.
*/ if (*pend->cq_kernel_addr >= data->target_value) {
spin_unlock_irqrestore(&data->interrupt->ts_list_lock, flags);
/* We check for completion value as interrupt could have been received * before we add the wait node to the wait list.
*/ if (*pend->cq_kernel_addr >= data->target_value || (!data->intr_timeout_us)) {
spin_unlock_irqrestore(&data->interrupt->wait_list_lock, flags);
/* Add pending user interrupt to relevant list for the interrupt * handler to monitor. * Note that we cannot have sorted list by target value, * in order to shorten the list pass loop, since * same list could have nodes for different cq counter handle.
*/
list_add_tail(&pend->list_node, &data->interrupt->wait_list_head);
spin_unlock_irqrestore(&data->interrupt->wait_list_lock, flags);
/* Wait for interrupt handler to signal completion */
completion_rc = wait_for_completion_interruptible_timeout(&pend->fence.completion,
timeout); if (completion_rc > 0) { if (pend->fence.error == -EIO) {
dev_err_ratelimited(hdev->dev, "interrupt based wait ioctl aborted(error:%d) due to a reset cycle initiated\n",
pend->fence.error);
rc = -EIO;
*status = HL_WAIT_CS_STATUS_ABORTED;
} else {
*status = HL_WAIT_CS_STATUS_COMPLETED;
}
} else { if (completion_rc == -ERESTARTSYS) {
dev_err_ratelimited(hdev->dev, "user process got signal while waiting for interrupt ID %d\n",
data->interrupt->interrupt_id);
rc = -EINTR;
*status = HL_WAIT_CS_STATUS_ABORTED;
} else { /* The wait has timed-out. We don't know anything beyond that * because the workload was not submitted through the driver. * Therefore, from driver's perspective, the workload is still * executing.
*/
rc = 0;
*status = HL_WAIT_CS_STATUS_BUSY;
}
}
/* * We keep removing the node from list here, and not at the irq handler * for completion timeout case. and if it's a registration * for ts record, the node will be deleted in the irq handler after * we reach the target value.
*/
spin_lock_irqsave(&data->interrupt->wait_list_lock, flags);
list_del(&pend->list_node);
spin_unlock_irqrestore(&data->interrupt->wait_list_lock, flags);
/* Add pending user interrupt to relevant list for the interrupt * handler to monitor
*/
spin_lock_irqsave(&interrupt->wait_list_lock, flags);
list_add_tail(&pend->list_node, &interrupt->wait_list_head);
spin_unlock_irqrestore(&interrupt->wait_list_lock, flags);
/* We check for completion value as interrupt could have been received * before we added the node to the wait list
*/ if (copy_from_user(&completion_value, u64_to_user_ptr(user_address), 8)) {
dev_err(hdev->dev, "Failed to copy completion value from user\n");
rc = -EFAULT; goto remove_pending_user_interrupt;
}
if (completion_value >= target_value) {
*status = HL_WAIT_CS_STATUS_COMPLETED; /* There was no interrupt, we assume the completion is now. */
pend->fence.timestamp = ktime_get();
} else {
*status = HL_WAIT_CS_STATUS_BUSY;
}
if (!timeout_us || (*status == HL_WAIT_CS_STATUS_COMPLETED)) goto remove_pending_user_interrupt;
wait_again: /* Wait for interrupt handler to signal completion */
completion_rc = wait_for_completion_interruptible_timeout(&pend->fence.completion,
timeout);
/* If timeout did not expire we need to perform the comparison. * If comparison fails, keep waiting until timeout expires
*/ if (completion_rc > 0) {
spin_lock_irqsave(&interrupt->wait_list_lock, flags); /* reinit_completion must be called before we check for user * completion value, otherwise, if interrupt is received after * the comparison and before the next wait_for_completion, * we will reach timeout and fail
*/
reinit_completion(&pend->fence.completion);
spin_unlock_irqrestore(&interrupt->wait_list_lock, flags);
if (copy_from_user(&completion_value, u64_to_user_ptr(user_address), 8)) {
dev_err(hdev->dev, "Failed to copy completion value from user\n");
rc = -EFAULT;
goto remove_pending_user_interrupt;
}
if (completion_value >= target_value) {
*status = HL_WAIT_CS_STATUS_COMPLETED;
} elseif (pend->fence.error) {
dev_err_ratelimited(hdev->dev, "interrupt based wait ioctl aborted(error:%d) due to a reset cycle initiated\n",
pend->fence.error); /* set the command completion status as ABORTED */
*status = HL_WAIT_CS_STATUS_ABORTED;
} else {
timeout = completion_rc; goto wait_again;
}
} elseif (completion_rc == -ERESTARTSYS) {
dev_err_ratelimited(hdev->dev, "user process got signal while waiting for interrupt ID %d\n",
interrupt->interrupt_id);
rc = -EINTR;
} else { /* The wait has timed-out. We don't know anything beyond that * because the workload wasn't submitted through the driver. * Therefore, from driver's perspective, the workload is still * executing.
*/
rc = 0;
*status = HL_WAIT_CS_STATUS_BUSY;
}
/* Check if the requested core is enabled */ if (!(prop->decoder_enabled_mask & BIT(interrupt_id))) {
dev_err(hdev->dev, "interrupt on a disabled core(%u) not allowed",
interrupt_id); return -EINVAL;
}
if (args->in.flags & HL_WAIT_CS_FLAGS_REGISTER_INTERRUPT) { /* * Allow only one registration at a time. this is needed in order to prevent * issues while handling the flow of re-use of the same offset. * Since the registration flow is protected only by the interrupt lock, * re-use flow might request to move ts node to another interrupt list, * and in such case we're not protected.
*/
mutex_lock(&hpriv->ctx->ts_reg_lock);
/* If the device is not operational, or if an error has happened and user should release the * device, there is no point in waiting for any command submission or user interrupt.
*/ if (!hl_device_operational(hpriv->hdev, NULL) || hdev->reset_info.watchdog_active) return -EBUSY;
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.