From: Tvrtko Ursulin tvrtko.ursulin@intel.com
"Watchdog" aka "restoring hangcheck" aka default request/fence expiry - second post of a somewhat controversial feature, now upgraded to patch status.
I quote the "watchdog" becuase in classical sense watchdog would allow userspace to ping it and so remain alive.
I quote "restoring hangcheck" because this series, contrary to the old hangcheck, is not looking at whether the workload is making any progress from the kernel side either. (Although disclaimer my memory may be leaky - Daniel suspects old hangcheck had some stricter, more indiscriminatory, angles to it. But apart from being prone to both false negatives and false positives I can't remember that myself.)
Short version - ask is to fail any user submissions after a set time period. In this RFC that time is twelve seconds.
Time counts from the moment user submission is "runnable" (implicit and explicit dependencies have been cleared) and keeps counting regardless of the GPU contetion caused by other users of the system.
So semantics are really a bit weak, but again, I understand this is really really wanted by the DRM core even if I am not convinced it is a good idea.
There are some dangers with doing this - text borrowed from a patch in the series:
This can have an effect that workloads which used to work fine will suddenly start failing. Even workloads comprised of short batches but in long dependency chains can be terminated.
And becuase of lack of agreement on usefulness and safety of fence error propagation this partial execution can be invisible to userspace even if it is "listening" to returned fence status.
Another interaction is with hangcheck where care needs to be taken timeout is not set lower or close to three times the heartbeat interval. Otherwise a hang in any application can cause complete termination of all submissions from unrelated clients. Any users modifying the per engine heartbeat intervals therefore need to be aware of this potential denial of service to avoid inadvertently enabling it.
Given all this I am personally not convinced the scheme is a good idea. Intuitively it feels object importers would be better positioned to enforce the time they are willing to wait for something to complete.
v2: * Dropped context param. * Improved commit messages and Kconfig text.
v3: * Log timeouts. * Bump timeout to 20s to see if it helps Tigerlake. * Fix sentinel assert.
v4: * A round of review feedback applied.
Chris Wilson (1): drm/i915: Individual request cancellation
Tvrtko Ursulin (6): drm/i915: Extract active lookup engine to a helper drm/i915: Restrict sentinel requests further drm/i915: Handle async cancellation in sentinel assert drm/i915: Request watchdog infrastructure drm/i915: Fail too long user submissions by default drm/i915: Allow configuring default request expiry via modparam
drivers/gpu/drm/i915/Kconfig.profile | 14 ++ drivers/gpu/drm/i915/gem/i915_gem_context.c | 73 ++++--- .../gpu/drm/i915/gem/i915_gem_context_types.h | 4 + drivers/gpu/drm/i915/gt/intel_context_param.h | 11 +- drivers/gpu/drm/i915/gt/intel_context_types.h | 4 + .../gpu/drm/i915/gt/intel_engine_heartbeat.c | 1 + .../drm/i915/gt/intel_execlists_submission.c | 23 +- .../drm/i915/gt/intel_execlists_submission.h | 2 + drivers/gpu/drm/i915/gt/intel_gt.c | 3 + drivers/gpu/drm/i915/gt/intel_gt.h | 2 + drivers/gpu/drm/i915/gt/intel_gt_requests.c | 28 +++ drivers/gpu/drm/i915/gt/intel_gt_types.h | 7 + drivers/gpu/drm/i915/i915_params.c | 5 + drivers/gpu/drm/i915/i915_params.h | 1 + drivers/gpu/drm/i915/i915_request.c | 129 ++++++++++- drivers/gpu/drm/i915/i915_request.h | 16 +- drivers/gpu/drm/i915/selftests/i915_request.c | 201 ++++++++++++++++++ 17 files changed, 479 insertions(+), 45 deletions(-)
From: Tvrtko Ursulin tvrtko.ursulin@intel.com
Move active engine lookup to exported i915_request_active_engine.
Signed-off-by: Tvrtko Ursulin tvrtko.ursulin@intel.com --- drivers/gpu/drm/i915/gem/i915_gem_context.c | 34 +--------------- drivers/gpu/drm/i915/i915_request.c | 44 +++++++++++++++++++++ drivers/gpu/drm/i915/i915_request.h | 4 ++ 3 files changed, 49 insertions(+), 33 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_context.c b/drivers/gpu/drm/i915/gem/i915_gem_context.c index ca37d93ef5e7..03a2f5f2a11f 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_context.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_context.c @@ -408,38 +408,6 @@ static bool __cancel_engine(struct intel_engine_cs *engine) return intel_engine_pulse(engine) == 0; }
-static bool -__active_engine(struct i915_request *rq, struct intel_engine_cs **active) -{ - struct intel_engine_cs *engine, *locked; - bool ret = false; - - /* - * Serialise with __i915_request_submit() so that it sees - * is-banned?, or we know the request is already inflight. - * - * Note that rq->engine is unstable, and so we double - * check that we have acquired the lock on the final engine. - */ - locked = READ_ONCE(rq->engine); - spin_lock_irq(&locked->sched.lock); - while (unlikely(locked != (engine = READ_ONCE(rq->engine)))) { - spin_unlock(&locked->sched.lock); - locked = engine; - spin_lock(&locked->sched.lock); - } - - if (i915_request_is_active(rq)) { - if (!__i915_request_is_complete(rq)) - *active = locked; - ret = true; - } - - spin_unlock_irq(&locked->sched.lock); - - return ret; -} - static struct intel_engine_cs *active_engine(struct intel_context *ce) { struct intel_engine_cs *engine = NULL; @@ -467,7 +435,7 @@ static struct intel_engine_cs *active_engine(struct intel_context *ce) /* Check with the backend if the request is inflight */ found = true; if (likely(rcu_access_pointer(rq->timeline) == ce->timeline)) - found = __active_engine(rq, &engine); + found = i915_request_active_engine(rq, &engine);
i915_request_put(rq); if (found) diff --git a/drivers/gpu/drm/i915/i915_request.c b/drivers/gpu/drm/i915/i915_request.c index e7b4c4bc41a6..8416b0bc4eb3 100644 --- a/drivers/gpu/drm/i915/i915_request.c +++ b/drivers/gpu/drm/i915/i915_request.c @@ -244,6 +244,50 @@ static void __i915_request_fill(struct i915_request *rq, u8 val) memset(vaddr + head, val, rq->postfix - head); }
+/** + * i915_request_active_engine + * @rq: request to inspect + * @active: pointer in which to return the active engine + * + * Fills the currently active engine to the @active pointer if the request + * is active and still not completed. + * + * Returns true if request was active or false otherwise. + */ +bool +i915_request_active_engine(struct i915_request *rq, + struct intel_engine_cs **active) +{ + struct intel_engine_cs *engine, *locked; + bool ret = false; + + /* + * Serialise with __i915_request_submit() so that it sees + * is-banned?, or we know the request is already inflight. + * + * Note that rq->engine is unstable, and so we double + * check that we have acquired the lock on the final engine. + */ + locked = READ_ONCE(rq->engine); + spin_lock_irq(&locked->sched.lock); + while (unlikely(locked != (engine = READ_ONCE(rq->engine)))) { + spin_unlock(&locked->sched.lock); + locked = engine; + spin_lock(&locked->sched.lock); + } + + if (i915_request_is_active(rq)) { + if (!__i915_request_is_complete(rq)) + *active = locked; + ret = true; + } + + spin_unlock_irq(&locked->sched.lock); + + return ret; +} + + static void remove_from_engine(struct i915_request *rq) { struct intel_engine_cs *engine, *locked; diff --git a/drivers/gpu/drm/i915/i915_request.h b/drivers/gpu/drm/i915/i915_request.h index dd10a6db3d21..f5374bab7e69 100644 --- a/drivers/gpu/drm/i915/i915_request.h +++ b/drivers/gpu/drm/i915/i915_request.h @@ -650,4 +650,8 @@ static inline bool i915_request_use_semaphores(const struct i915_request *rq) return intel_engine_has_semaphores(rq->engine); }
+bool +i915_request_active_engine(struct i915_request *rq, + struct intel_engine_cs **active); + #endif /* I915_REQUEST_H */
On Wed, 24 Mar 2021 at 12:13, Tvrtko Ursulin tvrtko.ursulin@linux.intel.com wrote:
From: Tvrtko Ursulin tvrtko.ursulin@intel.com
Move active engine lookup to exported i915_request_active_engine.
Signed-off-by: Tvrtko Ursulin tvrtko.ursulin@intel.com
Reviewed-by: Matthew Auld matthew.auld@intel.com
From: Chris Wilson chris@chris-wilson.co.uk
Currently, we cancel outstanding requests within a context when the context is closed. We may also want to cancel individual requests using the same graceful preemption mechanism.
v2 (Tvrtko): * Cancel waiters carefully considering no timeline lock and RCU. * Fixed selftests.
v3 (Tvrtko): * Remove error propagation to waiters for now.
v4 (Tvrtko): * Rebase for extracted i915_request_active_engine. (Matt)
Signed-off-by: Chris Wilson chris@chris-wilson.co.uk Signed-off-by: Tvrtko Ursulin tvrtko.ursulin@intel.com Reviewed-by: Matthew Auld matthew.auld@intel.com # v3 --- .../gpu/drm/i915/gt/intel_engine_heartbeat.c | 1 + .../drm/i915/gt/intel_execlists_submission.c | 9 +- drivers/gpu/drm/i915/i915_request.c | 33 ++- drivers/gpu/drm/i915/i915_request.h | 4 +- drivers/gpu/drm/i915/selftests/i915_request.c | 201 ++++++++++++++++++ 5 files changed, 242 insertions(+), 6 deletions(-)
diff --git a/drivers/gpu/drm/i915/gt/intel_engine_heartbeat.c b/drivers/gpu/drm/i915/gt/intel_engine_heartbeat.c index 0b062fad1837..e2fb3ae2aaf3 100644 --- a/drivers/gpu/drm/i915/gt/intel_engine_heartbeat.c +++ b/drivers/gpu/drm/i915/gt/intel_engine_heartbeat.c @@ -314,6 +314,7 @@ int intel_engine_pulse(struct intel_engine_cs *engine) mutex_unlock(&ce->timeline->mutex); }
+ intel_engine_flush_scheduler(engine); intel_engine_pm_put(engine); return err; } diff --git a/drivers/gpu/drm/i915/gt/intel_execlists_submission.c b/drivers/gpu/drm/i915/gt/intel_execlists_submission.c index 85ff5fe861b4..4c2acb5a6c0a 100644 --- a/drivers/gpu/drm/i915/gt/intel_execlists_submission.c +++ b/drivers/gpu/drm/i915/gt/intel_execlists_submission.c @@ -421,6 +421,11 @@ static void reset_active(struct i915_request *rq, ce->lrc.lrca = lrc_update_regs(ce, engine, head); }
+static bool bad_request(const struct i915_request *rq) +{ + return rq->fence.error && i915_request_started(rq); +} + static struct intel_engine_cs * __execlists_schedule_in(struct i915_request *rq) { @@ -433,7 +438,7 @@ __execlists_schedule_in(struct i915_request *rq) !intel_engine_has_heartbeat(engine))) intel_context_set_banned(ce);
- if (unlikely(intel_context_is_banned(ce))) + if (unlikely(intel_context_is_banned(ce) || bad_request(rq))) reset_active(rq, engine);
if (IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)) @@ -1112,7 +1117,7 @@ static unsigned long active_preempt_timeout(struct intel_engine_cs *engine, return 0;
/* Force a fast reset for terminated contexts (ignoring sysfs!) */ - if (unlikely(intel_context_is_banned(rq->context))) + if (unlikely(intel_context_is_banned(rq->context) || bad_request(rq))) return 1;
return READ_ONCE(engine->props.preempt_timeout_ms); diff --git a/drivers/gpu/drm/i915/i915_request.c b/drivers/gpu/drm/i915/i915_request.c index 8416b0bc4eb3..d1a4a3fa7425 100644 --- a/drivers/gpu/drm/i915/i915_request.c +++ b/drivers/gpu/drm/i915/i915_request.c @@ -33,7 +33,10 @@ #include "gem/i915_gem_context.h" #include "gt/intel_breadcrumbs.h" #include "gt/intel_context.h" +#include "gt/intel_engine.h" +#include "gt/intel_engine_heartbeat.h" #include "gt/intel_gpu_commands.h" +#include "gt/intel_reset.h" #include "gt/intel_ring.h" #include "gt/intel_rps.h"
@@ -473,20 +476,22 @@ void __i915_request_skip(struct i915_request *rq) rq->infix = rq->postfix; }
-void i915_request_set_error_once(struct i915_request *rq, int error) +bool i915_request_set_error_once(struct i915_request *rq, int error) { int old;
GEM_BUG_ON(!IS_ERR_VALUE((long)error));
if (i915_request_signaled(rq)) - return; + return false;
old = READ_ONCE(rq->fence.error); do { if (fatal_error(old)) - return; + return false; } while (!try_cmpxchg(&rq->fence.error, &old, error)); + + return true; }
struct i915_request *i915_request_mark_eio(struct i915_request *rq) @@ -653,6 +658,28 @@ void i915_request_unsubmit(struct i915_request *request) spin_unlock_irqrestore(&se->lock, flags); }
+static void __cancel_request(struct i915_request *rq) +{ + struct intel_engine_cs *engine = NULL; + + i915_request_active_engine(rq, &engine); + + if (engine && intel_engine_pulse(engine)) + intel_gt_handle_error(engine->gt, engine->mask, 0, + "request cancellation by %s", + current->comm); +} + +void i915_request_cancel(struct i915_request *rq, int error) +{ + if (!i915_request_set_error_once(rq, error)) + return; + + set_bit(I915_FENCE_FLAG_SENTINEL, &rq->fence.flags); + + __cancel_request(rq); +} + static int __i915_sw_fence_call submit_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state) { diff --git a/drivers/gpu/drm/i915/i915_request.h b/drivers/gpu/drm/i915/i915_request.h index f5374bab7e69..2dea55ea69e1 100644 --- a/drivers/gpu/drm/i915/i915_request.h +++ b/drivers/gpu/drm/i915/i915_request.h @@ -312,7 +312,7 @@ struct i915_request * __must_check i915_request_create(struct intel_context *ce);
void __i915_request_skip(struct i915_request *rq); -void i915_request_set_error_once(struct i915_request *rq, int error); +bool i915_request_set_error_once(struct i915_request *rq, int error); struct i915_request *i915_request_mark_eio(struct i915_request *rq);
struct i915_request *__i915_request_commit(struct i915_request *request); @@ -368,6 +368,8 @@ void i915_request_submit(struct i915_request *request); void __i915_request_unsubmit(struct i915_request *request); void i915_request_unsubmit(struct i915_request *request);
+void i915_request_cancel(struct i915_request *rq, int error); + long i915_request_wait(struct i915_request *rq, unsigned int flags, long timeout) diff --git a/drivers/gpu/drm/i915/selftests/i915_request.c b/drivers/gpu/drm/i915/selftests/i915_request.c index 8035ea7565ed..72f85d9d5e8b 100644 --- a/drivers/gpu/drm/i915/selftests/i915_request.c +++ b/drivers/gpu/drm/i915/selftests/i915_request.c @@ -608,6 +608,206 @@ static int live_nop_request(void *arg) return err; }
+static int __cancel_inactive(struct intel_engine_cs *engine) +{ + struct intel_context *ce; + struct igt_spinner spin; + struct i915_request *rq; + int err = 0; + + if (igt_spinner_init(&spin, engine->gt)) + return -ENOMEM; + + ce = intel_context_create(engine); + if (IS_ERR(ce)) { + err = PTR_ERR(ce); + goto out_spin; + } + + rq = igt_spinner_create_request(&spin, ce, MI_ARB_CHECK); + if (IS_ERR(rq)) { + err = PTR_ERR(rq); + goto out_ce; + } + + pr_debug("%s: Cancelling inactive request\n", engine->name); + i915_request_cancel(rq, -EINTR); + i915_request_get(rq); + i915_request_add(rq); + + if (i915_request_wait(rq, 0, HZ / 5) < 0) { + struct drm_printer p = drm_info_printer(engine->i915->drm.dev); + + pr_err("%s: Failed to cancel inactive request\n", engine->name); + intel_engine_dump(engine, &p, "%s\n", engine->name); + err = -ETIME; + goto out_rq; + } + + if (rq->fence.error != -EINTR) { + pr_err("%s: fence not cancelled (%u)\n", + engine->name, rq->fence.error); + err = -EINVAL; + } + +out_rq: + i915_request_put(rq); +out_ce: + intel_context_put(ce); +out_spin: + igt_spinner_fini(&spin); + if (err) + pr_err("%s: %s error %d\n", __func__, engine->name, err); + return err; +} + +static int __cancel_active(struct intel_engine_cs *engine) +{ + struct intel_context *ce; + struct igt_spinner spin; + struct i915_request *rq; + int err = 0; + + if (igt_spinner_init(&spin, engine->gt)) + return -ENOMEM; + + ce = intel_context_create(engine); + if (IS_ERR(ce)) { + err = PTR_ERR(ce); + goto out_spin; + } + + rq = igt_spinner_create_request(&spin, ce, MI_ARB_CHECK); + if (IS_ERR(rq)) { + err = PTR_ERR(rq); + goto out_ce; + } + + pr_debug("%s: Cancelling active request\n", engine->name); + i915_request_get(rq); + i915_request_add(rq); + if (!igt_wait_for_spinner(&spin, rq)) { + struct drm_printer p = drm_info_printer(engine->i915->drm.dev); + + pr_err("Failed to start spinner on %s\n", engine->name); + intel_engine_dump(engine, &p, "%s\n", engine->name); + err = -ETIME; + goto out_rq; + } + i915_request_cancel(rq, -EINTR); + + if (i915_request_wait(rq, 0, HZ / 5) < 0) { + struct drm_printer p = drm_info_printer(engine->i915->drm.dev); + + pr_err("%s: Failed to cancel active request\n", engine->name); + intel_engine_dump(engine, &p, "%s\n", engine->name); + err = -ETIME; + goto out_rq; + } + + if (rq->fence.error != -EINTR) { + pr_err("%s: fence not cancelled (%u)\n", + engine->name, rq->fence.error); + err = -EINVAL; + } + +out_rq: + i915_request_put(rq); +out_ce: + intel_context_put(ce); +out_spin: + igt_spinner_fini(&spin); + if (err) + pr_err("%s: %s error %d\n", __func__, engine->name, err); + return err; +} + +static int __cancel_completed(struct intel_engine_cs *engine) +{ + struct intel_context *ce; + struct igt_spinner spin; + struct i915_request *rq; + int err = 0; + + if (igt_spinner_init(&spin, engine->gt)) + return -ENOMEM; + + ce = intel_context_create(engine); + if (IS_ERR(ce)) { + err = PTR_ERR(ce); + goto out_spin; + } + + rq = igt_spinner_create_request(&spin, ce, MI_ARB_CHECK); + if (IS_ERR(rq)) { + err = PTR_ERR(rq); + goto out_ce; + } + igt_spinner_end(&spin); + i915_request_get(rq); + i915_request_add(rq); + + if (i915_request_wait(rq, 0, HZ / 5) < 0) { + err = -ETIME; + goto out_rq; + } + + pr_debug("%s: Cancelling completed request\n", engine->name); + i915_request_cancel(rq, -EINTR); + if (rq->fence.error) { + pr_err("%s: fence not cancelled (%u)\n", + engine->name, rq->fence.error); + err = -EINVAL; + } + +out_rq: + i915_request_put(rq); +out_ce: + intel_context_put(ce); +out_spin: + igt_spinner_fini(&spin); + if (err) + pr_err("%s: %s error %d\n", __func__, engine->name, err); + return err; +} + +static int live_cancel_request(void *arg) +{ + struct drm_i915_private *i915 = arg; + struct intel_engine_cs *engine; + + /* + * Check cancellation of requests. We expect to be able to immediately + * cancel active requests, even if they are currently on the GPU. + */ + + for_each_uabi_engine(engine, i915) { + struct igt_live_test t; + int err, err2; + + if (!intel_engine_has_preemption(engine)) + continue; + + err = igt_live_test_begin(&t, i915, __func__, engine->name); + if (err) + return err; + + err = __cancel_inactive(engine); + if (err == 0) + err = __cancel_active(engine); + if (err == 0) + err = __cancel_completed(engine); + + err2 = igt_live_test_end(&t); + if (err) + return err; + if (err2) + return err2; + } + + return 0; +} + static struct i915_vma *empty_batch(struct drm_i915_private *i915) { struct drm_i915_gem_object *obj; @@ -1485,6 +1685,7 @@ int i915_request_live_selftests(struct drm_i915_private *i915) SUBTEST(live_sequential_engines), SUBTEST(live_parallel_engines), SUBTEST(live_empty_request), + SUBTEST(live_cancel_request), SUBTEST(live_breadcrumbs_smoketest), };
On Wed, 24 Mar 2021 at 12:13, Tvrtko Ursulin tvrtko.ursulin@linux.intel.com wrote:
From: Chris Wilson chris@chris-wilson.co.uk
Currently, we cancel outstanding requests within a context when the context is closed. We may also want to cancel individual requests using the same graceful preemption mechanism.
v2 (Tvrtko):
- Cancel waiters carefully considering no timeline lock and RCU.
- Fixed selftests.
v3 (Tvrtko):
- Remove error propagation to waiters for now.
v4 (Tvrtko):
- Rebase for extracted i915_request_active_engine. (Matt)
Signed-off-by: Chris Wilson chris@chris-wilson.co.uk Signed-off-by: Tvrtko Ursulin tvrtko.ursulin@intel.com Reviewed-by: Matthew Auld matthew.auld@intel.com # v3
Reviewed-by: Matthew Auld matthew.auld@intel.com
From: Tvrtko Ursulin tvrtko.ursulin@intel.com
Disallow sentinel requests follow previous sentinels to make request cancellation work better when faced with a chain of requests which have all been marked as in error.
Because in cases where we end up with a stream of cancelled requests we want to turn of request coalescing so they each will get individually skipped by the execlists_schedule_in (which is called per ELSP port, not per request).
Signed-off-by: Tvrtko Ursulin tvrtko.ursulin@intel.com --- drivers/gpu/drm/i915/gt/intel_execlists_submission.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/i915/gt/intel_execlists_submission.c b/drivers/gpu/drm/i915/gt/intel_execlists_submission.c index 4c2acb5a6c0a..4b870eca9693 100644 --- a/drivers/gpu/drm/i915/gt/intel_execlists_submission.c +++ b/drivers/gpu/drm/i915/gt/intel_execlists_submission.c @@ -896,7 +896,7 @@ static bool can_merge_rq(const struct i915_request *prev, if (__i915_request_is_complete(next)) return true;
- if (unlikely((i915_request_flags(prev) ^ i915_request_flags(next)) & + if (unlikely((i915_request_flags(prev) | i915_request_flags(next)) & (BIT(I915_FENCE_FLAG_NOPREEMPT) | BIT(I915_FENCE_FLAG_SENTINEL)))) return false;
On Wed, 24 Mar 2021 at 12:14, Tvrtko Ursulin tvrtko.ursulin@linux.intel.com wrote:
From: Tvrtko Ursulin tvrtko.ursulin@intel.com
Disallow sentinel requests follow previous sentinels to make request cancellation work better when faced with a chain of requests which have all been marked as in error.
Because in cases where we end up with a stream of cancelled requests we want to turn of request coalescing so they each will get individually
turn off
skipped by the execlists_schedule_in (which is called per ELSP port, not per request).
Signed-off-by: Tvrtko Ursulin tvrtko.ursulin@intel.com
Reviewed-by: Matthew Auld matthew.auld@intel.com
On Wed, Mar 24, 2021 at 03:25:55PM +0000, Matthew Auld wrote:
On Wed, 24 Mar 2021 at 12:14, Tvrtko Ursulin tvrtko.ursulin@linux.intel.com wrote:
From: Tvrtko Ursulin tvrtko.ursulin@intel.com
Disallow sentinel requests follow previous sentinels to make request cancellation work better when faced with a chain of requests which have all been marked as in error.
Because in cases where we end up with a stream of cancelled requests we want to turn of request coalescing so they each will get individually
turn off
Fixed while applying. -Daniel
skipped by the execlists_schedule_in (which is called per ELSP port, not per request).
Signed-off-by: Tvrtko Ursulin tvrtko.ursulin@intel.com
Reviewed-by: Matthew Auld matthew.auld@intel.com _______________________________________________ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/intel-gfx
From: Tvrtko Ursulin tvrtko.ursulin@intel.com
With the watchdog cancelling requests asynchronously to preempt-to-busy we need to relax one assert making it apply only to requests not in error.
v2: * Check against the correct request!
v3: * Simplify the check to avoid the question of when to sample the fence error vs sentinel bit.
Signed-off-by: Tvrtko Ursulin tvrtko.ursulin@intel.com --- drivers/gpu/drm/i915/gt/intel_execlists_submission.c | 12 ++++++++---- 1 file changed, 8 insertions(+), 4 deletions(-)
diff --git a/drivers/gpu/drm/i915/gt/intel_execlists_submission.c b/drivers/gpu/drm/i915/gt/intel_execlists_submission.c index 4b870eca9693..9d264d4ffa75 100644 --- a/drivers/gpu/drm/i915/gt/intel_execlists_submission.c +++ b/drivers/gpu/drm/i915/gt/intel_execlists_submission.c @@ -700,9 +700,8 @@ assert_pending_valid(const struct intel_engine_execlists *execlists, { struct intel_engine_cs *engine = container_of(execlists, typeof(*engine), execlists); - struct i915_request * const *port, *rq; + struct i915_request * const *port, *rq, *prev = NULL; struct intel_context *ce = NULL; - bool sentinel = false; u32 ccid = -1;
trace_ports(execlists, msg, execlists->pending); @@ -752,15 +751,20 @@ assert_pending_valid(const struct intel_engine_execlists *execlists, * Sentinels are supposed to be the last request so they flush * the current execution off the HW. Check that they are the only * request in the pending submission. + * + * NB: Due to the async nature of preempt-to-busy and request + * cancellation we need to handle the case where request + * becomes a sentinel in parallel to CSB processing. */ - if (sentinel) { + if (prev && i915_request_has_sentinel(prev) && + !READ_ONCE(prev->fence.error)) { GEM_TRACE_ERR("%s: context:%llx after sentinel in pending[%zd]\n", engine->name, ce->timeline->fence_context, port - execlists->pending); return false; } - sentinel = i915_request_has_sentinel(rq); + prev = rq;
/* * We want virtual requests to only be in the first slot so
On Wed, 24 Mar 2021 at 12:14, Tvrtko Ursulin tvrtko.ursulin@linux.intel.com wrote:
From: Tvrtko Ursulin tvrtko.ursulin@intel.com
With the watchdog cancelling requests asynchronously to preempt-to-busy we need to relax one assert making it apply only to requests not in error.
v2:
- Check against the correct request!
v3:
- Simplify the check to avoid the question of when to sample the fence error vs sentinel bit.
Signed-off-by: Tvrtko Ursulin tvrtko.ursulin@intel.com
Reviewed-by: Matthew Auld matthew.auld@intel.com
From: Tvrtko Ursulin tvrtko.ursulin@intel.com
Prepares the plumbing for setting request/fence expiration time. All code is put in place but is never activated due yet missing ability to actually configure the timer.
Outline of the basic operation:
A timer is started when request is ready for execution. If the request completes (retires) before the timer fires, timer is cancelled and nothing further happens.
If the timer fires request is added to a lockless list and worker queued. Purpose of this is twofold: a) It allows request cancellation from a more friendly context and b) coalesces multiple expirations into a single event of consuming the list.
Worker locklessly consumes the list of expired requests and cancels them all using previous added i915_request_cancel().
Associated timeout value is stored in rq->context.watchdog.timeout_us.
v2: * Log expiration.
v3: * Include more information about user timeline in the log message.
v4: * Remove obsolete comment and fix formatting. (Matt)
Signed-off-by: Tvrtko Ursulin tvrtko.ursulin@intel.com Cc: Daniel Vetter daniel.vetter@ffwll.ch Reviewed-by: Matthew Auld matthew.auld@intel.com --- drivers/gpu/drm/i915/gt/intel_context_types.h | 4 ++ .../drm/i915/gt/intel_execlists_submission.h | 2 + drivers/gpu/drm/i915/gt/intel_gt.c | 3 ++ drivers/gpu/drm/i915/gt/intel_gt.h | 2 + drivers/gpu/drm/i915/gt/intel_gt_requests.c | 28 ++++++++++ drivers/gpu/drm/i915/gt/intel_gt_types.h | 7 +++ drivers/gpu/drm/i915/i915_request.c | 52 +++++++++++++++++++ drivers/gpu/drm/i915/i915_request.h | 8 +++ 8 files changed, 106 insertions(+)
diff --git a/drivers/gpu/drm/i915/gt/intel_context_types.h b/drivers/gpu/drm/i915/gt/intel_context_types.h index 0ea18c9e2aca..65a5730a4f5b 100644 --- a/drivers/gpu/drm/i915/gt/intel_context_types.h +++ b/drivers/gpu/drm/i915/gt/intel_context_types.h @@ -99,6 +99,10 @@ struct intel_context { #define CONTEXT_FORCE_SINGLE_SUBMISSION 7 #define CONTEXT_NOPREEMPT 8
+ struct { + u64 timeout_us; + } watchdog; + u32 *lrc_reg_state; union { struct { diff --git a/drivers/gpu/drm/i915/gt/intel_execlists_submission.h b/drivers/gpu/drm/i915/gt/intel_execlists_submission.h index f7bd3fccfee8..4ca9b475e252 100644 --- a/drivers/gpu/drm/i915/gt/intel_execlists_submission.h +++ b/drivers/gpu/drm/i915/gt/intel_execlists_submission.h @@ -6,6 +6,7 @@ #ifndef __INTEL_EXECLISTS_SUBMISSION_H__ #define __INTEL_EXECLISTS_SUBMISSION_H__
+#include <linux/llist.h> #include <linux/types.h>
struct drm_printer; @@ -13,6 +14,7 @@ struct drm_printer; struct i915_request; struct intel_context; struct intel_engine_cs; +struct intel_gt;
enum { INTEL_CONTEXT_SCHEDULE_IN = 0, diff --git a/drivers/gpu/drm/i915/gt/intel_gt.c b/drivers/gpu/drm/i915/gt/intel_gt.c index ca76f93bc03d..8d77dcbad059 100644 --- a/drivers/gpu/drm/i915/gt/intel_gt.c +++ b/drivers/gpu/drm/i915/gt/intel_gt.c @@ -31,6 +31,9 @@ void intel_gt_init_early(struct intel_gt *gt, struct drm_i915_private *i915) INIT_LIST_HEAD(>->closed_vma); spin_lock_init(>->closed_lock);
+ init_llist_head(>->watchdog.list); + INIT_WORK(>->watchdog.work, intel_gt_watchdog_work); + intel_gt_init_buffer_pool(gt); intel_gt_init_reset(gt); intel_gt_init_requests(gt); diff --git a/drivers/gpu/drm/i915/gt/intel_gt.h b/drivers/gpu/drm/i915/gt/intel_gt.h index a17bd8b3195f..7ec395cace69 100644 --- a/drivers/gpu/drm/i915/gt/intel_gt.h +++ b/drivers/gpu/drm/i915/gt/intel_gt.h @@ -78,4 +78,6 @@ static inline bool intel_gt_is_wedged(const struct intel_gt *gt) void intel_gt_info_print(const struct intel_gt_info *info, struct drm_printer *p);
+void intel_gt_watchdog_work(struct work_struct *work); + #endif /* __INTEL_GT_H__ */ diff --git a/drivers/gpu/drm/i915/gt/intel_gt_requests.c b/drivers/gpu/drm/i915/gt/intel_gt_requests.c index 36ec97f79174..fbfd19b2e5f2 100644 --- a/drivers/gpu/drm/i915/gt/intel_gt_requests.c +++ b/drivers/gpu/drm/i915/gt/intel_gt_requests.c @@ -8,6 +8,7 @@ #include "i915_drv.h" /* for_each_engine() */ #include "i915_request.h" #include "intel_engine_heartbeat.h" +#include "intel_execlists_submission.h" #include "intel_gt.h" #include "intel_gt_pm.h" #include "intel_gt_requests.h" @@ -242,4 +243,31 @@ void intel_gt_fini_requests(struct intel_gt *gt) { /* Wait until the work is marked as finished before unloading! */ cancel_delayed_work_sync(>->requests.retire_work); + + flush_work(>->watchdog.work); +} + +void intel_gt_watchdog_work(struct work_struct *work) +{ + struct intel_gt *gt = + container_of(work, typeof(*gt), watchdog.work); + struct i915_request *rq, *rn; + struct llist_node *first; + + first = llist_del_all(>->watchdog.list); + if (!first) + return; + + llist_for_each_entry_safe(rq, rn, first, watchdog.link) { + if (!i915_request_completed(rq)) { + struct dma_fence *f = &rq->fence; + + pr_notice("Fence expiration time out i915-%s:%s:%llx!\n", + f->ops->get_driver_name(f), + f->ops->get_timeline_name(f), + f->seqno); + i915_request_cancel(rq, -EINTR); + } + i915_request_put(rq); + } } diff --git a/drivers/gpu/drm/i915/gt/intel_gt_types.h b/drivers/gpu/drm/i915/gt/intel_gt_types.h index 626af37c7790..d70ebcc6f19f 100644 --- a/drivers/gpu/drm/i915/gt/intel_gt_types.h +++ b/drivers/gpu/drm/i915/gt/intel_gt_types.h @@ -8,10 +8,12 @@
#include <linux/ktime.h> #include <linux/list.h> +#include <linux/llist.h> #include <linux/mutex.h> #include <linux/notifier.h> #include <linux/spinlock.h> #include <linux/types.h> +#include <linux/workqueue.h>
#include "uc/intel_uc.h"
@@ -62,6 +64,11 @@ struct intel_gt { struct delayed_work retire_work; } requests;
+ struct { + struct llist_head list; + struct work_struct work; + } watchdog; + struct intel_wakeref wakeref; atomic_t user_wakeref;
diff --git a/drivers/gpu/drm/i915/i915_request.c b/drivers/gpu/drm/i915/i915_request.c index d1a4a3fa7425..d58052f3410c 100644 --- a/drivers/gpu/drm/i915/i915_request.c +++ b/drivers/gpu/drm/i915/i915_request.c @@ -321,6 +321,53 @@ static void remove_from_engine(struct i915_request *rq) __notify_execute_cb_imm(rq); }
+static void __rq_init_watchdog(struct i915_request *rq) +{ + rq->watchdog.timer.function = NULL; +} + +static enum hrtimer_restart __rq_watchdog_expired(struct hrtimer *hrtimer) +{ + struct i915_request *rq = + container_of(hrtimer, struct i915_request, watchdog.timer); + struct intel_gt *gt = rq->engine->gt; + + if (!i915_request_completed(rq)) { + if (llist_add(&rq->watchdog.link, >->watchdog.list)) + schedule_work(>->watchdog.work); + } else { + i915_request_put(rq); + } + + return HRTIMER_NORESTART; +} + +static void __rq_arm_watchdog(struct i915_request *rq) +{ + struct i915_request_watchdog *wdg = &rq->watchdog; + struct intel_context *ce = rq->context; + + if (!ce->watchdog.timeout_us) + return; + + hrtimer_init(&wdg->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); + wdg->timer.function = __rq_watchdog_expired; + hrtimer_start_range_ns(&wdg->timer, + ns_to_ktime(ce->watchdog.timeout_us * + NSEC_PER_USEC), + NSEC_PER_MSEC, + HRTIMER_MODE_REL); + i915_request_get(rq); +} + +static void __rq_cancel_watchdog(struct i915_request *rq) +{ + struct i915_request_watchdog *wdg = &rq->watchdog; + + if (wdg->timer.function && hrtimer_try_to_cancel(&wdg->timer) > 0) + i915_request_put(rq); +} + bool i915_request_retire(struct i915_request *rq) { if (!__i915_request_is_complete(rq)) @@ -332,6 +379,8 @@ bool i915_request_retire(struct i915_request *rq) trace_i915_request_retire(rq); i915_request_mark_complete(rq);
+ __rq_cancel_watchdog(rq); + /* * We know the GPU must have read the request to have * sent us the seqno + interrupt, so use the position @@ -692,6 +741,8 @@ submit_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state)
if (unlikely(fence->error)) i915_request_set_error_once(request, fence->error); + else + __rq_arm_watchdog(request);
/* * We need to serialize use of the submit_request() callback @@ -879,6 +930,7 @@ __i915_request_create(struct intel_context *ce, gfp_t gfp)
/* No zalloc, everything must be cleared after use */ rq->batch = NULL; + __rq_init_watchdog(rq); GEM_BUG_ON(rq->capture_list); GEM_BUG_ON(!llist_empty(&rq->execute_cb));
diff --git a/drivers/gpu/drm/i915/i915_request.h b/drivers/gpu/drm/i915/i915_request.h index 2dea55ea69e1..5e0946992f1a 100644 --- a/drivers/gpu/drm/i915/i915_request.h +++ b/drivers/gpu/drm/i915/i915_request.h @@ -26,7 +26,9 @@ #define I915_REQUEST_H
#include <linux/dma-fence.h> +#include <linux/hrtimer.h> #include <linux/irq_work.h> +#include <linux/llist.h> #include <linux/lockdep.h>
#include "gem/i915_gem_context_types.h" @@ -289,6 +291,12 @@ struct i915_request { /** timeline->request entry for this request */ struct list_head link;
+ /** Watchdog support fields. */ + struct i915_request_watchdog { + struct llist_node link; + struct hrtimer timer; + } watchdog; + I915_SELFTEST_DECLARE(struct { struct list_head link; unsigned long delay;
On Wed, Mar 24, 2021 at 12:13:33PM +0000, Tvrtko Ursulin wrote:
From: Tvrtko Ursulin tvrtko.ursulin@intel.com
Prepares the plumbing for setting request/fence expiration time. All code is put in place but is never activated due yet missing ability to actually configure the timer.
Outline of the basic operation:
A timer is started when request is ready for execution. If the request completes (retires) before the timer fires, timer is cancelled and nothing further happens.
If the timer fires request is added to a lockless list and worker queued. Purpose of this is twofold: a) It allows request cancellation from a more friendly context and b) coalesces multiple expirations into a single event of consuming the list.
Worker locklessly consumes the list of expired requests and cancels them all using previous added i915_request_cancel().
Associated timeout value is stored in rq->context.watchdog.timeout_us.
v2:
- Log expiration.
v3:
- Include more information about user timeline in the log message.
v4:
- Remove obsolete comment and fix formatting. (Matt)
Signed-off-by: Tvrtko Ursulin tvrtko.ursulin@intel.com Cc: Daniel Vetter daniel.vetter@ffwll.ch Reviewed-by: Matthew Auld matthew.auld@intel.com
drivers/gpu/drm/i915/gt/intel_context_types.h | 4 ++ .../drm/i915/gt/intel_execlists_submission.h | 2 + drivers/gpu/drm/i915/gt/intel_gt.c | 3 ++ drivers/gpu/drm/i915/gt/intel_gt.h | 2 + drivers/gpu/drm/i915/gt/intel_gt_requests.c | 28 ++++++++++ drivers/gpu/drm/i915/gt/intel_gt_types.h | 7 +++ drivers/gpu/drm/i915/i915_request.c | 52 +++++++++++++++++++ drivers/gpu/drm/i915/i915_request.h | 8 +++ 8 files changed, 106 insertions(+)
diff --git a/drivers/gpu/drm/i915/gt/intel_context_types.h b/drivers/gpu/drm/i915/gt/intel_context_types.h index 0ea18c9e2aca..65a5730a4f5b 100644 --- a/drivers/gpu/drm/i915/gt/intel_context_types.h +++ b/drivers/gpu/drm/i915/gt/intel_context_types.h @@ -99,6 +99,10 @@ struct intel_context { #define CONTEXT_FORCE_SINGLE_SUBMISSION 7 #define CONTEXT_NOPREEMPT 8
- struct {
u64 timeout_us;- } watchdog;
- u32 *lrc_reg_state; union { struct {
diff --git a/drivers/gpu/drm/i915/gt/intel_execlists_submission.h b/drivers/gpu/drm/i915/gt/intel_execlists_submission.h index f7bd3fccfee8..4ca9b475e252 100644 --- a/drivers/gpu/drm/i915/gt/intel_execlists_submission.h +++ b/drivers/gpu/drm/i915/gt/intel_execlists_submission.h @@ -6,6 +6,7 @@ #ifndef __INTEL_EXECLISTS_SUBMISSION_H__ #define __INTEL_EXECLISTS_SUBMISSION_H__
+#include <linux/llist.h> #include <linux/types.h>
struct drm_printer; @@ -13,6 +14,7 @@ struct drm_printer; struct i915_request; struct intel_context; struct intel_engine_cs; +struct intel_gt;
enum { INTEL_CONTEXT_SCHEDULE_IN = 0, diff --git a/drivers/gpu/drm/i915/gt/intel_gt.c b/drivers/gpu/drm/i915/gt/intel_gt.c index ca76f93bc03d..8d77dcbad059 100644 --- a/drivers/gpu/drm/i915/gt/intel_gt.c +++ b/drivers/gpu/drm/i915/gt/intel_gt.c @@ -31,6 +31,9 @@ void intel_gt_init_early(struct intel_gt *gt, struct drm_i915_private *i915) INIT_LIST_HEAD(>->closed_vma); spin_lock_init(>->closed_lock);
- init_llist_head(>->watchdog.list);
- INIT_WORK(>->watchdog.work, intel_gt_watchdog_work);
- intel_gt_init_buffer_pool(gt); intel_gt_init_reset(gt); intel_gt_init_requests(gt);
diff --git a/drivers/gpu/drm/i915/gt/intel_gt.h b/drivers/gpu/drm/i915/gt/intel_gt.h index a17bd8b3195f..7ec395cace69 100644 --- a/drivers/gpu/drm/i915/gt/intel_gt.h +++ b/drivers/gpu/drm/i915/gt/intel_gt.h @@ -78,4 +78,6 @@ static inline bool intel_gt_is_wedged(const struct intel_gt *gt) void intel_gt_info_print(const struct intel_gt_info *info, struct drm_printer *p);
+void intel_gt_watchdog_work(struct work_struct *work);
#endif /* __INTEL_GT_H__ */ diff --git a/drivers/gpu/drm/i915/gt/intel_gt_requests.c b/drivers/gpu/drm/i915/gt/intel_gt_requests.c index 36ec97f79174..fbfd19b2e5f2 100644 --- a/drivers/gpu/drm/i915/gt/intel_gt_requests.c +++ b/drivers/gpu/drm/i915/gt/intel_gt_requests.c @@ -8,6 +8,7 @@ #include "i915_drv.h" /* for_each_engine() */ #include "i915_request.h" #include "intel_engine_heartbeat.h" +#include "intel_execlists_submission.h" #include "intel_gt.h" #include "intel_gt_pm.h" #include "intel_gt_requests.h" @@ -242,4 +243,31 @@ void intel_gt_fini_requests(struct intel_gt *gt) { /* Wait until the work is marked as finished before unloading! */ cancel_delayed_work_sync(>->requests.retire_work);
- flush_work(>->watchdog.work);
+}
+void intel_gt_watchdog_work(struct work_struct *work) +{
- struct intel_gt *gt =
container_of(work, typeof(*gt), watchdog.work);- struct i915_request *rq, *rn;
- struct llist_node *first;
- first = llist_del_all(>->watchdog.list);
- if (!first)
return;- llist_for_each_entry_safe(rq, rn, first, watchdog.link) {
if (!i915_request_completed(rq)) {struct dma_fence *f = &rq->fence;pr_notice("Fence expiration time out i915-%s:%s:%llx!\n",f->ops->get_driver_name(f),f->ops->get_timeline_name(f),f->seqno);i915_request_cancel(rq, -EINTR);}i915_request_put(rq);- }
} diff --git a/drivers/gpu/drm/i915/gt/intel_gt_types.h b/drivers/gpu/drm/i915/gt/intel_gt_types.h index 626af37c7790..d70ebcc6f19f 100644 --- a/drivers/gpu/drm/i915/gt/intel_gt_types.h +++ b/drivers/gpu/drm/i915/gt/intel_gt_types.h @@ -8,10 +8,12 @@
#include <linux/ktime.h> #include <linux/list.h> +#include <linux/llist.h> #include <linux/mutex.h> #include <linux/notifier.h> #include <linux/spinlock.h> #include <linux/types.h> +#include <linux/workqueue.h>
#include "uc/intel_uc.h"
@@ -62,6 +64,11 @@ struct intel_gt { struct delayed_work retire_work; } requests;
- struct {
struct llist_head list;struct work_struct work;- } watchdog;
- struct intel_wakeref wakeref; atomic_t user_wakeref;
diff --git a/drivers/gpu/drm/i915/i915_request.c b/drivers/gpu/drm/i915/i915_request.c index d1a4a3fa7425..d58052f3410c 100644 --- a/drivers/gpu/drm/i915/i915_request.c +++ b/drivers/gpu/drm/i915/i915_request.c @@ -321,6 +321,53 @@ static void remove_from_engine(struct i915_request *rq) __notify_execute_cb_imm(rq); }
+static void __rq_init_watchdog(struct i915_request *rq) +{
- rq->watchdog.timer.function = NULL;
+}
+static enum hrtimer_restart __rq_watchdog_expired(struct hrtimer *hrtimer) +{
- struct i915_request *rq =
container_of(hrtimer, struct i915_request, watchdog.timer);- struct intel_gt *gt = rq->engine->gt;
- if (!i915_request_completed(rq)) {
if (llist_add(&rq->watchdog.link, >->watchdog.list))schedule_work(>->watchdog.work);- } else {
i915_request_put(rq);- }
- return HRTIMER_NORESTART;
+}
+static void __rq_arm_watchdog(struct i915_request *rq) +{
- struct i915_request_watchdog *wdg = &rq->watchdog;
- struct intel_context *ce = rq->context;
- if (!ce->watchdog.timeout_us)
return;- hrtimer_init(&wdg->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
- wdg->timer.function = __rq_watchdog_expired;
- hrtimer_start_range_ns(&wdg->timer,
ns_to_ktime(ce->watchdog.timeout_us *NSEC_PER_USEC),NSEC_PER_MSEC,HRTIMER_MODE_REL);- i915_request_get(rq);
Shouldn't we grab the new reference before we arm the timer? Either way since fairly academic I went ahead and applied, but if you agree pls do a follow up patch. -Daniel
+}
+static void __rq_cancel_watchdog(struct i915_request *rq) +{
- struct i915_request_watchdog *wdg = &rq->watchdog;
- if (wdg->timer.function && hrtimer_try_to_cancel(&wdg->timer) > 0)
i915_request_put(rq);+}
bool i915_request_retire(struct i915_request *rq) { if (!__i915_request_is_complete(rq)) @@ -332,6 +379,8 @@ bool i915_request_retire(struct i915_request *rq) trace_i915_request_retire(rq); i915_request_mark_complete(rq);
- __rq_cancel_watchdog(rq);
- /*
- We know the GPU must have read the request to have
- sent us the seqno + interrupt, so use the position
@@ -692,6 +741,8 @@ submit_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state)
if (unlikely(fence->error)) i915_request_set_error_once(request, fence->error);
else__rq_arm_watchdog(request);/*
- We need to serialize use of the submit_request() callback
@@ -879,6 +930,7 @@ __i915_request_create(struct intel_context *ce, gfp_t gfp)
/* No zalloc, everything must be cleared after use */ rq->batch = NULL;
- __rq_init_watchdog(rq); GEM_BUG_ON(rq->capture_list); GEM_BUG_ON(!llist_empty(&rq->execute_cb));
diff --git a/drivers/gpu/drm/i915/i915_request.h b/drivers/gpu/drm/i915/i915_request.h index 2dea55ea69e1..5e0946992f1a 100644 --- a/drivers/gpu/drm/i915/i915_request.h +++ b/drivers/gpu/drm/i915/i915_request.h @@ -26,7 +26,9 @@ #define I915_REQUEST_H
#include <linux/dma-fence.h> +#include <linux/hrtimer.h> #include <linux/irq_work.h> +#include <linux/llist.h> #include <linux/lockdep.h>
#include "gem/i915_gem_context_types.h" @@ -289,6 +291,12 @@ struct i915_request { /** timeline->request entry for this request */ struct list_head link;
- /** Watchdog support fields. */
- struct i915_request_watchdog {
struct llist_node link;struct hrtimer timer;- } watchdog;
- I915_SELFTEST_DECLARE(struct { struct list_head link; unsigned long delay;
-- 2.27.0
On 26/03/2021 00:00, Daniel Vetter wrote:
On Wed, Mar 24, 2021 at 12:13:33PM +0000, Tvrtko Ursulin wrote:
[snip]
+static enum hrtimer_restart __rq_watchdog_expired(struct hrtimer *hrtimer) +{
- struct i915_request *rq =
container_of(hrtimer, struct i915_request, watchdog.timer);- struct intel_gt *gt = rq->engine->gt;
- if (!i915_request_completed(rq)) {
if (llist_add(&rq->watchdog.link, >->watchdog.list))schedule_work(>->watchdog.work);- } else {
i915_request_put(rq);- }
- return HRTIMER_NORESTART;
+}
+static void __rq_arm_watchdog(struct i915_request *rq) +{
- struct i915_request_watchdog *wdg = &rq->watchdog;
- struct intel_context *ce = rq->context;
- if (!ce->watchdog.timeout_us)
return;- hrtimer_init(&wdg->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
- wdg->timer.function = __rq_watchdog_expired;
- hrtimer_start_range_ns(&wdg->timer,
ns_to_ktime(ce->watchdog.timeout_us *NSEC_PER_USEC),NSEC_PER_MSEC,HRTIMER_MODE_REL);- i915_request_get(rq);
Shouldn't we grab the new reference before we arm the timer? Either way since fairly academic I went ahead and applied, but if you agree pls do a follow up patch.
Absolutely true.. my bad.
Regards,
Tvrtko
From: Tvrtko Ursulin tvrtko.ursulin@intel.com
A new Kconfig option CONFIG_DRM_I915_REQUEST_TIMEOUT is added, defaulting to 20s, and this timeout is applied to all users contexts using the previously added watchdog facility.
Result of this is that any user submission will simply fail after this timeout, either causing a reset (for non-preemptable), or incomplete results.
This can have an effect that workloads which used to work fine will suddenly start failing. Even workloads comprised of short batches but in long dependency chains can be terminated.
And because of lack of agreement on usefulness and safety of fence error propagation this partial execution can be invisible to userspace even if it is "listening" to returned fence status.
Another interaction is with hangcheck where care needs to be taken timeout is not set lower or close to three times the heartbeat interval. Otherwise a hang in any application can cause complete termination of all submissions from unrelated clients. Any users modifying the per engine heartbeat intervals therefore need to be aware of this potential denial of service to avoid inadvertently enabling it.
Given all this I am personally not convinced the scheme is a good idea. Intuitively it feels object importers would be better positioned to enforce the time they are willing to wait for something to complete.
v2: * Improved commit message and Kconfig text. * Pull in some helper code from patch which got dropped.
v3: * Bump timeout to 20s to see if it helps Tigerlake.
Signed-off-by: Tvrtko Ursulin tvrtko.ursulin@intel.com Cc: Daniel Vetter daniel.vetter@ffwll.ch Acked-by: Matthew Auld matthew.auld@intel.com --- drivers/gpu/drm/i915/Kconfig.profile | 14 +++++++ drivers/gpu/drm/i915/gem/i915_gem_context.c | 38 +++++++++++++++++++ .../gpu/drm/i915/gem/i915_gem_context_types.h | 4 ++ drivers/gpu/drm/i915/gt/intel_context_param.h | 11 +++++- 4 files changed, 66 insertions(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/i915/Kconfig.profile b/drivers/gpu/drm/i915/Kconfig.profile index 35bbe2b80596..39328567c200 100644 --- a/drivers/gpu/drm/i915/Kconfig.profile +++ b/drivers/gpu/drm/i915/Kconfig.profile @@ -1,3 +1,17 @@ +config DRM_I915_REQUEST_TIMEOUT + int "Default timeout for requests (ms)" + default 20000 # milliseconds + help + Configures the default timeout after which any user submissions will + be forcefully terminated. + + Beware setting this value lower, or close to heartbeat interval + rounded to whole seconds times three, in order to avoid allowing + misbehaving applications causing total rendering failure in unrelated + clients. + + May be 0 to disable the timeout. + config DRM_I915_FENCE_TIMEOUT int "Timeout for unsignaled foreign fences (ms, jiffy granularity)" default 10000 # milliseconds diff --git a/drivers/gpu/drm/i915/gem/i915_gem_context.c b/drivers/gpu/drm/i915/gem/i915_gem_context.c index 03a2f5f2a11f..33ff1a6a7724 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_context.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_context.c @@ -233,6 +233,8 @@ static void intel_context_set_gem(struct intel_context *ce, if (ctx->sched.priority >= I915_PRIORITY_NORMAL && intel_engine_has_timeslices(ce->engine)) __set_bit(CONTEXT_USE_SEMAPHORES, &ce->flags); + + intel_context_set_watchdog_us(ce, ctx->watchdog.timeout_us); }
static void __free_engines(struct i915_gem_engines *e, unsigned int count) @@ -820,6 +822,40 @@ static void __assign_timeline(struct i915_gem_context *ctx, context_apply_all(ctx, __apply_timeline, timeline); }
+static int __apply_watchdog(struct intel_context *ce, void *timeout_us) +{ + return intel_context_set_watchdog_us(ce, (uintptr_t)timeout_us); +} + +static int +__set_watchdog(struct i915_gem_context *ctx, unsigned long timeout_us) +{ + int ret; + + ret = context_apply_all(ctx, __apply_watchdog, + (void *)(uintptr_t)timeout_us); + if (!ret) + ctx->watchdog.timeout_us = timeout_us; + + return ret; +} + +static void __set_default_fence_expiry(struct i915_gem_context *ctx) +{ + struct drm_i915_private *i915 = ctx->i915; + int ret; + + if (!IS_ACTIVE(CONFIG_DRM_I915_REQUEST_TIMEOUT)) + return; + + /* Default expiry for user fences. */ + ret = __set_watchdog(ctx, CONFIG_DRM_I915_REQUEST_TIMEOUT * 1000); + if (ret) + drm_notice(&i915->drm, + "Failed to configure default fence expiry! (%d)", + ret); +} + static struct i915_gem_context * i915_gem_create_context(struct drm_i915_private *i915, unsigned int flags) { @@ -864,6 +900,8 @@ i915_gem_create_context(struct drm_i915_private *i915, unsigned int flags) intel_timeline_put(timeline); }
+ __set_default_fence_expiry(ctx); + trace_i915_context_create(ctx);
return ctx; diff --git a/drivers/gpu/drm/i915/gem/i915_gem_context_types.h b/drivers/gpu/drm/i915/gem/i915_gem_context_types.h index d5bc75508048..f17da7e26c43 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_context_types.h +++ b/drivers/gpu/drm/i915/gem/i915_gem_context_types.h @@ -150,6 +150,10 @@ struct i915_gem_context { */ atomic_t active_count;
+ struct { + u64 timeout_us; + } watchdog; + /** * @hang_timestamp: The last time(s) this context caused a GPU hang */ diff --git a/drivers/gpu/drm/i915/gt/intel_context_param.h b/drivers/gpu/drm/i915/gt/intel_context_param.h index f053d8633fe2..3ecacc675f41 100644 --- a/drivers/gpu/drm/i915/gt/intel_context_param.h +++ b/drivers/gpu/drm/i915/gt/intel_context_param.h @@ -6,9 +6,18 @@ #ifndef INTEL_CONTEXT_PARAM_H #define INTEL_CONTEXT_PARAM_H
-struct intel_context; +#include <linux/types.h> + +#include "intel_context.h"
int intel_context_set_ring_size(struct intel_context *ce, long sz); long intel_context_get_ring_size(struct intel_context *ce);
+static inline int +intel_context_set_watchdog_us(struct intel_context *ce, u64 timeout_us) +{ + ce->watchdog.timeout_us = timeout_us; + return 0; +} + #endif /* INTEL_CONTEXT_PARAM_H */
From: Tvrtko Ursulin tvrtko.ursulin@intel.com
Module parameter is added (request_timeout_ms) to allow configuring the default request/fence expiry.
Default value is inherited from CONFIG_DRM_I915_REQUEST_TIMEOUT.
Signed-off-by: Tvrtko Ursulin tvrtko.ursulin@intel.com Cc: Daniel Vetter daniel.vetter@ffwll.ch Acked-by: Daniel Vetter daniel.vetter@ffwll.ch --- drivers/gpu/drm/i915/gem/i915_gem_context.c | 5 +++-- drivers/gpu/drm/i915/i915_params.c | 5 +++++ drivers/gpu/drm/i915/i915_params.h | 1 + 3 files changed, 9 insertions(+), 2 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_context.c b/drivers/gpu/drm/i915/gem/i915_gem_context.c index 33ff1a6a7724..0e8f0476e01f 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_context.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_context.c @@ -845,11 +845,12 @@ static void __set_default_fence_expiry(struct i915_gem_context *ctx) struct drm_i915_private *i915 = ctx->i915; int ret;
- if (!IS_ACTIVE(CONFIG_DRM_I915_REQUEST_TIMEOUT)) + if (!IS_ACTIVE(CONFIG_DRM_I915_REQUEST_TIMEOUT) || + !i915->params.request_timeout_ms) return;
/* Default expiry for user fences. */ - ret = __set_watchdog(ctx, CONFIG_DRM_I915_REQUEST_TIMEOUT * 1000); + ret = __set_watchdog(ctx, i915->params.request_timeout_ms * 1000); if (ret) drm_notice(&i915->drm, "Failed to configure default fence expiry! (%d)", diff --git a/drivers/gpu/drm/i915/i915_params.c b/drivers/gpu/drm/i915/i915_params.c index 6939634e56ed..0320878d96b0 100644 --- a/drivers/gpu/drm/i915/i915_params.c +++ b/drivers/gpu/drm/i915/i915_params.c @@ -197,6 +197,11 @@ i915_param_named_unsafe(fake_lmem_start, ulong, 0400, "Fake LMEM start offset (default: 0)"); #endif
+#if CONFIG_DRM_I915_REQUEST_TIMEOUT +i915_param_named_unsafe(request_timeout_ms, uint, 0600, + "Default request/fence/batch buffer expiration timeout."); +#endif + static __always_inline void _print_param(struct drm_printer *p, const char *name, const char *type, diff --git a/drivers/gpu/drm/i915/i915_params.h b/drivers/gpu/drm/i915/i915_params.h index 48f47e44e848..34ebb0662547 100644 --- a/drivers/gpu/drm/i915/i915_params.h +++ b/drivers/gpu/drm/i915/i915_params.h @@ -72,6 +72,7 @@ struct drm_printer; param(int, enable_dpcd_backlight, -1, 0600) \ param(char *, force_probe, CONFIG_DRM_I915_FORCE_PROBE, 0400) \ param(unsigned long, fake_lmem_start, 0, 0400) \ + param(unsigned int, request_timeout_ms, CONFIG_DRM_I915_REQUEST_TIMEOUT, 0600) \ /* leave bools at the end to not create holes */ \ param(bool, enable_hangcheck, true, 0600) \ param(bool, load_detect_test, false, 0600) \
On Wed, Mar 24, 2021 at 12:13:35PM +0000, Tvrtko Ursulin wrote:
From: Tvrtko Ursulin tvrtko.ursulin@intel.com
Module parameter is added (request_timeout_ms) to allow configuring the default request/fence expiry.
Default value is inherited from CONFIG_DRM_I915_REQUEST_TIMEOUT.
Signed-off-by: Tvrtko Ursulin tvrtko.ursulin@intel.com Cc: Daniel Vetter daniel.vetter@ffwll.ch Acked-by: Daniel Vetter daniel.vetter@ffwll.ch
Entire series merged, thanks for the patches. -Daniel
drivers/gpu/drm/i915/gem/i915_gem_context.c | 5 +++-- drivers/gpu/drm/i915/i915_params.c | 5 +++++ drivers/gpu/drm/i915/i915_params.h | 1 + 3 files changed, 9 insertions(+), 2 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_context.c b/drivers/gpu/drm/i915/gem/i915_gem_context.c index 33ff1a6a7724..0e8f0476e01f 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_context.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_context.c @@ -845,11 +845,12 @@ static void __set_default_fence_expiry(struct i915_gem_context *ctx) struct drm_i915_private *i915 = ctx->i915; int ret;
- if (!IS_ACTIVE(CONFIG_DRM_I915_REQUEST_TIMEOUT))
if (!IS_ACTIVE(CONFIG_DRM_I915_REQUEST_TIMEOUT) ||
!i915->params.request_timeout_ms)return;
/* Default expiry for user fences. */
- ret = __set_watchdog(ctx, CONFIG_DRM_I915_REQUEST_TIMEOUT * 1000);
- ret = __set_watchdog(ctx, i915->params.request_timeout_ms * 1000); if (ret) drm_notice(&i915->drm, "Failed to configure default fence expiry! (%d)",
diff --git a/drivers/gpu/drm/i915/i915_params.c b/drivers/gpu/drm/i915/i915_params.c index 6939634e56ed..0320878d96b0 100644 --- a/drivers/gpu/drm/i915/i915_params.c +++ b/drivers/gpu/drm/i915/i915_params.c @@ -197,6 +197,11 @@ i915_param_named_unsafe(fake_lmem_start, ulong, 0400, "Fake LMEM start offset (default: 0)"); #endif
+#if CONFIG_DRM_I915_REQUEST_TIMEOUT +i915_param_named_unsafe(request_timeout_ms, uint, 0600,
"Default request/fence/batch buffer expiration timeout.");+#endif
static __always_inline void _print_param(struct drm_printer *p, const char *name, const char *type, diff --git a/drivers/gpu/drm/i915/i915_params.h b/drivers/gpu/drm/i915/i915_params.h index 48f47e44e848..34ebb0662547 100644 --- a/drivers/gpu/drm/i915/i915_params.h +++ b/drivers/gpu/drm/i915/i915_params.h @@ -72,6 +72,7 @@ struct drm_printer; param(int, enable_dpcd_backlight, -1, 0600) \ param(char *, force_probe, CONFIG_DRM_I915_FORCE_PROBE, 0400) \ param(unsigned long, fake_lmem_start, 0, 0400) \
- param(unsigned int, request_timeout_ms, CONFIG_DRM_I915_REQUEST_TIMEOUT, 0600) \ /* leave bools at the end to not create holes */ \ param(bool, enable_hangcheck, true, 0600) \ param(bool, load_detect_test, false, 0600) \
-- 2.27.0
On Wed, Mar 24, 2021 at 12:13:28PM +0000, Tvrtko Ursulin wrote:
From: Tvrtko Ursulin tvrtko.ursulin@intel.com
"Watchdog" aka "restoring hangcheck" aka default request/fence expiry - second post of a somewhat controversial feature, now upgraded to patch status.
I quote the "watchdog" becuase in classical sense watchdog would allow userspace to ping it and so remain alive.
I quote "restoring hangcheck" because this series, contrary to the old hangcheck, is not looking at whether the workload is making any progress from the kernel side either. (Although disclaimer my memory may be leaky - Daniel suspects old hangcheck had some stricter, more indiscriminatory, angles to it. But apart from being prone to both false negatives and false positives I can't remember that myself.)
Short version - ask is to fail any user submissions after a set time period. In this RFC that time is twelve seconds.
Time counts from the moment user submission is "runnable" (implicit and explicit dependencies have been cleared) and keeps counting regardless of the GPU contetion caused by other users of the system.
So semantics are really a bit weak, but again, I understand this is really really wanted by the DRM core even if I am not convinced it is a good idea.
There are some dangers with doing this - text borrowed from a patch in the series:
This can have an effect that workloads which used to work fine will suddenly start failing. Even workloads comprised of short batches but in long dependency chains can be terminated.
And becuase of lack of agreement on usefulness and safety of fence error propagation this partial execution can be invisible to userspace even if it is "listening" to returned fence status.
Another interaction is with hangcheck where care needs to be taken timeout is not set lower or close to three times the heartbeat interval. Otherwise a hang in any application can cause complete termination of all submissions from unrelated clients. Any users modifying the per engine heartbeat intervals therefore need to be aware of this potential denial of service to avoid inadvertently enabling it.
Given all this I am personally not convinced the scheme is a good idea. Intuitively it feels object importers would be better positioned to enforce the time they are willing to wait for something to complete.
v2:
- Dropped context param.
- Improved commit messages and Kconfig text.
v3:
- Log timeouts.
- Bump timeout to 20s to see if it helps Tigerlake.
I think 20s is a bit much, and seems like problem is still there in igt. I think we need look at that and figure out what to do with it. And then go back down with the timeout somewhat again since 20s is quite a long time. Irrespective of all the additional gaps/opens around watchdog timeout. -Daniel
- Fix sentinel assert.
v4:
- A round of review feedback applied.
Chris Wilson (1): drm/i915: Individual request cancellation
Tvrtko Ursulin (6): drm/i915: Extract active lookup engine to a helper drm/i915: Restrict sentinel requests further drm/i915: Handle async cancellation in sentinel assert drm/i915: Request watchdog infrastructure drm/i915: Fail too long user submissions by default drm/i915: Allow configuring default request expiry via modparam
drivers/gpu/drm/i915/Kconfig.profile | 14 ++ drivers/gpu/drm/i915/gem/i915_gem_context.c | 73 ++++--- .../gpu/drm/i915/gem/i915_gem_context_types.h | 4 + drivers/gpu/drm/i915/gt/intel_context_param.h | 11 +- drivers/gpu/drm/i915/gt/intel_context_types.h | 4 + .../gpu/drm/i915/gt/intel_engine_heartbeat.c | 1 + .../drm/i915/gt/intel_execlists_submission.c | 23 +- .../drm/i915/gt/intel_execlists_submission.h | 2 + drivers/gpu/drm/i915/gt/intel_gt.c | 3 + drivers/gpu/drm/i915/gt/intel_gt.h | 2 + drivers/gpu/drm/i915/gt/intel_gt_requests.c | 28 +++ drivers/gpu/drm/i915/gt/intel_gt_types.h | 7 + drivers/gpu/drm/i915/i915_params.c | 5 + drivers/gpu/drm/i915/i915_params.h | 1 + drivers/gpu/drm/i915/i915_request.c | 129 ++++++++++- drivers/gpu/drm/i915/i915_request.h | 16 +- drivers/gpu/drm/i915/selftests/i915_request.c | 201 ++++++++++++++++++ 17 files changed, 479 insertions(+), 45 deletions(-)
-- 2.27.0
Intel-gfx mailing list Intel-gfx@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/intel-gfx
On 26/03/2021 09:10, Daniel Vetter wrote:
On Wed, Mar 24, 2021 at 12:13:28PM +0000, Tvrtko Ursulin wrote:
From: Tvrtko Ursulin tvrtko.ursulin@intel.com
"Watchdog" aka "restoring hangcheck" aka default request/fence expiry - second post of a somewhat controversial feature, now upgraded to patch status.
I quote the "watchdog" becuase in classical sense watchdog would allow userspace to ping it and so remain alive.
I quote "restoring hangcheck" because this series, contrary to the old hangcheck, is not looking at whether the workload is making any progress from the kernel side either. (Although disclaimer my memory may be leaky - Daniel suspects old hangcheck had some stricter, more indiscriminatory, angles to it. But apart from being prone to both false negatives and false positives I can't remember that myself.)
Short version - ask is to fail any user submissions after a set time period. In this RFC that time is twelve seconds.
Time counts from the moment user submission is "runnable" (implicit and explicit dependencies have been cleared) and keeps counting regardless of the GPU contetion caused by other users of the system.
So semantics are really a bit weak, but again, I understand this is really really wanted by the DRM core even if I am not convinced it is a good idea.
There are some dangers with doing this - text borrowed from a patch in the series:
This can have an effect that workloads which used to work fine will suddenly start failing. Even workloads comprised of short batches but in long dependency chains can be terminated.
And becuase of lack of agreement on usefulness and safety of fence error propagation this partial execution can be invisible to userspace even if it is "listening" to returned fence status.
Another interaction is with hangcheck where care needs to be taken timeout is not set lower or close to three times the heartbeat interval. Otherwise a hang in any application can cause complete termination of all submissions from unrelated clients. Any users modifying the per engine heartbeat intervals therefore need to be aware of this potential denial of service to avoid inadvertently enabling it.
Given all this I am personally not convinced the scheme is a good idea. Intuitively it feels object importers would be better positioned to enforce the time they are willing to wait for something to complete.
v2:
- Dropped context param.
- Improved commit messages and Kconfig text.
v3:
- Log timeouts.
- Bump timeout to 20s to see if it helps Tigerlake.
I think 20s is a bit much, and seems like problem is still there in igt. I think we need look at that and figure out what to do with it. And then go back down with the timeout somewhat again since 20s is quite a long time. Irrespective of all the additional gaps/opens around watchdog timeout.
1)
The relationship with the hearbeat is the first issue. There we have 3x heartbeat period (each rounded to full second) before sending a high-prio pulse which can cause a preempt timeout and hence a reset/kicking out of a non-compliant request.
Defaults for those values mean default expiry shouldn't be lower than 3x rounded hearbeat interval + preempt timeout, currently ~9.75s. In practice even 12s which I tried initially was too aggressive due slacks on some platforms.
2)
20s seems to work apart that it shows the general regression unconditional default expiry adds. Either some existing IGTs which create long runnable chains, or the far-fence test which explicitly demonstrates this. AFAIK, and apart from the can_merge_rq yet unexplained oops, this is the only class of IGT failures which can appear.
So you could tweak it lower, if you also decide to make real hang detection stricter. But doing that also worsens the regression with loaded systems.
I only can have a large shrug/dontknow here since I wish we went more towards my suggestion of emulating setrlimit(RLIMIT_CPU). Meaning at least going with GPU time instead of elapsed time and possibly even leaving the policy of setting it to sysadmins. That would fit much better with our hangcheck, but, doesn't fit the drm core mandate.. hence I really don't know.
Regards,
Tvrtko
-Daniel
- Fix sentinel assert.
v4:
- A round of review feedback applied.
Chris Wilson (1): drm/i915: Individual request cancellation
Tvrtko Ursulin (6): drm/i915: Extract active lookup engine to a helper drm/i915: Restrict sentinel requests further drm/i915: Handle async cancellation in sentinel assert drm/i915: Request watchdog infrastructure drm/i915: Fail too long user submissions by default drm/i915: Allow configuring default request expiry via modparam
drivers/gpu/drm/i915/Kconfig.profile | 14 ++ drivers/gpu/drm/i915/gem/i915_gem_context.c | 73 ++++--- .../gpu/drm/i915/gem/i915_gem_context_types.h | 4 + drivers/gpu/drm/i915/gt/intel_context_param.h | 11 +- drivers/gpu/drm/i915/gt/intel_context_types.h | 4 + .../gpu/drm/i915/gt/intel_engine_heartbeat.c | 1 + .../drm/i915/gt/intel_execlists_submission.c | 23 +- .../drm/i915/gt/intel_execlists_submission.h | 2 + drivers/gpu/drm/i915/gt/intel_gt.c | 3 + drivers/gpu/drm/i915/gt/intel_gt.h | 2 + drivers/gpu/drm/i915/gt/intel_gt_requests.c | 28 +++ drivers/gpu/drm/i915/gt/intel_gt_types.h | 7 + drivers/gpu/drm/i915/i915_params.c | 5 + drivers/gpu/drm/i915/i915_params.h | 1 + drivers/gpu/drm/i915/i915_request.c | 129 ++++++++++- drivers/gpu/drm/i915/i915_request.h | 16 +- drivers/gpu/drm/i915/selftests/i915_request.c | 201 ++++++++++++++++++ 17 files changed, 479 insertions(+), 45 deletions(-)
-- 2.27.0
Intel-gfx mailing list Intel-gfx@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/intel-gfx
On Fri, Mar 26, 2021 at 10:31:10AM +0000, Tvrtko Ursulin wrote:
On 26/03/2021 09:10, Daniel Vetter wrote:
On Wed, Mar 24, 2021 at 12:13:28PM +0000, Tvrtko Ursulin wrote:
From: Tvrtko Ursulin tvrtko.ursulin@intel.com
"Watchdog" aka "restoring hangcheck" aka default request/fence expiry - second post of a somewhat controversial feature, now upgraded to patch status.
I quote the "watchdog" becuase in classical sense watchdog would allow userspace to ping it and so remain alive.
I quote "restoring hangcheck" because this series, contrary to the old hangcheck, is not looking at whether the workload is making any progress from the kernel side either. (Although disclaimer my memory may be leaky - Daniel suspects old hangcheck had some stricter, more indiscriminatory, angles to it. But apart from being prone to both false negatives and false positives I can't remember that myself.)
Short version - ask is to fail any user submissions after a set time period. In this RFC that time is twelve seconds.
Time counts from the moment user submission is "runnable" (implicit and explicit dependencies have been cleared) and keeps counting regardless of the GPU contetion caused by other users of the system.
So semantics are really a bit weak, but again, I understand this is really really wanted by the DRM core even if I am not convinced it is a good idea.
There are some dangers with doing this - text borrowed from a patch in the series:
This can have an effect that workloads which used to work fine will suddenly start failing. Even workloads comprised of short batches but in long dependency chains can be terminated.
And becuase of lack of agreement on usefulness and safety of fence error propagation this partial execution can be invisible to userspace even if it is "listening" to returned fence status.
Another interaction is with hangcheck where care needs to be taken timeout is not set lower or close to three times the heartbeat interval. Otherwise a hang in any application can cause complete termination of all submissions from unrelated clients. Any users modifying the per engine heartbeat intervals therefore need to be aware of this potential denial of service to avoid inadvertently enabling it.
Given all this I am personally not convinced the scheme is a good idea. Intuitively it feels object importers would be better positioned to enforce the time they are willing to wait for something to complete.
v2:
- Dropped context param.
- Improved commit messages and Kconfig text.
v3:
- Log timeouts.
- Bump timeout to 20s to see if it helps Tigerlake.
I think 20s is a bit much, and seems like problem is still there in igt. I think we need look at that and figure out what to do with it. And then go back down with the timeout somewhat again since 20s is quite a long time. Irrespective of all the additional gaps/opens around watchdog timeout.
The relationship with the hearbeat is the first issue. There we have 3x heartbeat period (each rounded to full second) before sending a high-prio pulse which can cause a preempt timeout and hence a reset/kicking out of a non-compliant request.
Defaults for those values mean default expiry shouldn't be lower than 3x rounded hearbeat interval + preempt timeout, currently ~9.75s. In practice even 12s which I tried initially was too aggressive due slacks on some platforms.
Hm, would be good to put that as a comment next to the module param, or something like that. Maybe even a sanity check to make sure these two values are consistent (i.e. if watchdog is less than 3.5x the heartbeat, we complain in dmesg).
20s seems to work apart that it shows the general regression unconditional default expiry adds. Either some existing IGTs which create long runnable chains, or the far-fence test which explicitly demonstrates this. AFAIK, and apart from the can_merge_rq yet unexplained oops, this is the only class of IGT failures which can appear.
So you could tweak it lower, if you also decide to make real hang detection stricter. But doing that also worsens the regression with loaded systems.
I only can have a large shrug/dontknow here since I wish we went more towards my suggestion of emulating setrlimit(RLIMIT_CPU). Meaning at least going with GPU time instead of elapsed time and possibly even leaving the policy of setting it to sysadmins. That would fit much better with our hangcheck, but, doesn't fit the drm core mandate.. hence I really don't know.
The bikeshed will come back when we wire up drm/scheduler as the frontend for guc scheduler backend. I guess we can tackle it then. -Daniel
Regards,
Tvrtko
-Daniel
- Fix sentinel assert.
v4:
- A round of review feedback applied.
Chris Wilson (1): drm/i915: Individual request cancellation
Tvrtko Ursulin (6): drm/i915: Extract active lookup engine to a helper drm/i915: Restrict sentinel requests further drm/i915: Handle async cancellation in sentinel assert drm/i915: Request watchdog infrastructure drm/i915: Fail too long user submissions by default drm/i915: Allow configuring default request expiry via modparam
drivers/gpu/drm/i915/Kconfig.profile | 14 ++ drivers/gpu/drm/i915/gem/i915_gem_context.c | 73 ++++--- .../gpu/drm/i915/gem/i915_gem_context_types.h | 4 + drivers/gpu/drm/i915/gt/intel_context_param.h | 11 +- drivers/gpu/drm/i915/gt/intel_context_types.h | 4 + .../gpu/drm/i915/gt/intel_engine_heartbeat.c | 1 + .../drm/i915/gt/intel_execlists_submission.c | 23 +- .../drm/i915/gt/intel_execlists_submission.h | 2 + drivers/gpu/drm/i915/gt/intel_gt.c | 3 + drivers/gpu/drm/i915/gt/intel_gt.h | 2 + drivers/gpu/drm/i915/gt/intel_gt_requests.c | 28 +++ drivers/gpu/drm/i915/gt/intel_gt_types.h | 7 + drivers/gpu/drm/i915/i915_params.c | 5 + drivers/gpu/drm/i915/i915_params.h | 1 + drivers/gpu/drm/i915/i915_request.c | 129 ++++++++++- drivers/gpu/drm/i915/i915_request.h | 16 +- drivers/gpu/drm/i915/selftests/i915_request.c | 201 ++++++++++++++++++ 17 files changed, 479 insertions(+), 45 deletions(-)
-- 2.27.0
Intel-gfx mailing list Intel-gfx@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/intel-gfx
dri-devel@lists.freedesktop.org
-
Daniel Vetter -
Matthew Auld -
Tvrtko Ursulin