| Message ID | 20211112153216.630452-1-matthew.auld@intel.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [1/6] drm/i915: move the pre_pin earlier | expand |
On Fri, 2021-11-12 at 15:32 +0000, Matthew Auld wrote: > In intel_context_do_pin_ww, when calling into the pre_pin hook(which > is > passed the ww context) it could in theory return -EDEADLK(which is > very > likely with debug kernels), once we start adding more ww locking in > there, > like in the next patch. If so then we need to be mindful of having to > restart the do_pin at this point. > > If this is the kernel_context, or some other early in-kernel context > where we have yet to setup the default_state, then we always inhibit > the > context restore, and instead rely on the delayed active_release to > set > the CONTEXT_VALID_BIT for us(if we even care), which should indicate > that we have context switched away, and that our newly saved context > state should now be valid. However, since we currently grab the > active > reference before the potential ww dance, we can end up setting the > CONTEXT_VALID_BIT much too early, if we need to backoff, and then > upon > re-trying the do_pin, we could potentially cause the hardware to > incorrectly load some garbage context state when later context > switching > to that context, but at the very least this will trigger the > GEM_BUG_ON() in __engine_unpark. For now let's just move any ww dance > stuff prior to arming the active reference. > > For normal user contexts this shouldn't be a concern, since we should > already have the default_state ready when initialising the lrc state, > and so there should be no concern with active_release somehow > prematurely setting the CONTEXT_VALID_BIT. > > Signed-off-by: Matthew Auld <matthew.auld@intel.com> > Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com> > Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> > --- > drivers/gpu/drm/i915/gt/intel_context.c | 8 ++++---- > 1 file changed, 4 insertions(+), 4 deletions(-) > > diff --git a/drivers/gpu/drm/i915/gt/intel_context.c > b/drivers/gpu/drm/i915/gt/intel_context.c > index 5634d14052bc..ad44860faaf3 100644 > --- a/drivers/gpu/drm/i915/gt/intel_context.c > +++ b/drivers/gpu/drm/i915/gt/intel_context.c > @@ -228,17 +228,17 @@ int __intel_context_do_pin_ww(struct > intel_context *ce, > if (err) > return err; > > - err = i915_active_acquire(&ce->active); > + err = ce->ops->pre_pin(ce, ww, &vaddr); > if (err) > goto err_ctx_unpin; > > - err = ce->ops->pre_pin(ce, ww, &vaddr); > + err = i915_active_acquire(&ce->active); > if (err) > - goto err_release; > + goto err_post_unpin; Hmm, If i915_active_acquire() fails, wouldn't we end up calling i915_active_release() here? Thanks, Thomas
diff --git a/drivers/gpu/drm/i915/gt/intel_context.c b/drivers/gpu/drm/i915/gt/intel_context.c index 5634d14052bc..ad44860faaf3 100644 --- a/drivers/gpu/drm/i915/gt/intel_context.c +++ b/drivers/gpu/drm/i915/gt/intel_context.c @@ -228,17 +228,17 @@ int __intel_context_do_pin_ww(struct intel_context *ce, if (err) return err; - err = i915_active_acquire(&ce->active); + err = ce->ops->pre_pin(ce, ww, &vaddr); if (err) goto err_ctx_unpin; - err = ce->ops->pre_pin(ce, ww, &vaddr); + err = i915_active_acquire(&ce->active); if (err) - goto err_release; + goto err_post_unpin; err = mutex_lock_interruptible(&ce->pin_mutex); if (err) - goto err_post_unpin; + goto err_release; intel_engine_pm_might_get(ce->engine);
In intel_context_do_pin_ww, when calling into the pre_pin hook(which is passed the ww context) it could in theory return -EDEADLK(which is very likely with debug kernels), once we start adding more ww locking in there, like in the next patch. If so then we need to be mindful of having to restart the do_pin at this point. If this is the kernel_context, or some other early in-kernel context where we have yet to setup the default_state, then we always inhibit the context restore, and instead rely on the delayed active_release to set the CONTEXT_VALID_BIT for us(if we even care), which should indicate that we have context switched away, and that our newly saved context state should now be valid. However, since we currently grab the active reference before the potential ww dance, we can end up setting the CONTEXT_VALID_BIT much too early, if we need to backoff, and then upon re-trying the do_pin, we could potentially cause the hardware to incorrectly load some garbage context state when later context switching to that context, but at the very least this will trigger the GEM_BUG_ON() in __engine_unpark. For now let's just move any ww dance stuff prior to arming the active reference. For normal user contexts this shouldn't be a concern, since we should already have the default_state ready when initialising the lrc state, and so there should be no concern with active_release somehow prematurely setting the CONTEXT_VALID_BIT. Signed-off-by: Matthew Auld <matthew.auld@intel.com> Cc: Thomas Hellström <thomas.hellstrom@linux.intel.com> Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com> --- drivers/gpu/drm/i915/gt/intel_context.c | 8 ++++---- 1 file changed, 4 insertions(+), 4 deletions(-)