The patch below does not apply to the 5.15-stable tree. If someone wants it applied there, or to any other stable or longterm tree, then please email the backport, including the original git commit id to <stable@xxxxxxxxxxxxxxx>. To reproduce the conflict and resubmit, you may use the following commands: git fetch https://git.kernel.org/pub/scm/linux/kernel/git/stable/linux.git/ linux-5.15.y git checkout FETCH_HEAD git cherry-pick -x a337b64f0d5717248a0c894e2618e658e6a9de9f # <resolve conflicts, build, test, etc.> git commit -s git send-email --to '<stable@xxxxxxxxxxxxxxx>' --in-reply-to '2023080739-trinity-overfed-f523@gregkh' --subject-prefix 'PATCH 5.15.y' HEAD^.. Possible dependencies: a337b64f0d57 ("drm/i915: Fix premature release of request's reusable memory") 506006055769 ("drm/i915/active: Fix misuse of non-idle barriers as fence trackers") ad5c99e02047 ("drm/i915: Remove unused bits of i915_vma/active api") f6c466b84cfa ("drm/i915: Add support for moving fence waiting") 544460c33821 ("drm/i915: Multi-BB execbuf") 5851387a422c ("drm/i915/guc: Implement no mid batch preemption for multi-lrc") e5e32171a2cf ("drm/i915/guc: Connect UAPI to GuC multi-lrc interface") d38a9294491d ("drm/i915/guc: Update debugfs for GuC multi-lrc") bc955204919e ("drm/i915/guc: Insert submit fences between requests in parent-child relationship") 6b540bf6f143 ("drm/i915/guc: Implement multi-lrc submission") 99b47aaddfa9 ("drm/i915/guc: Implement parallel context pin / unpin functions") c2aa552ff09d ("drm/i915/guc: Add multi-lrc context registration") 3897df4c0187 ("drm/i915/guc: Introduce context parent-child relationship") 4f3059dc2dbb ("drm/i915: Add logical engine mapping") 1a52faed3131 ("drm/i915/guc: Take GT PM ref when deregistering context") 0ea92ace8b95 ("drm/i915/guc: Move GuC guc_id allocation under submission state sub-struct") 0d8ee5ba8db4 ("drm/i915: Don't back up pinned LMEM context images and rings during suspend") c56ce9565374 ("drm/i915 Implement LMEM backup and restore for suspend / resume") 0d9388635a22 ("drm/i915/ttm: Implement a function to copy the contents of two TTM-based objects") 68c03c0e985e ("drm/i915/debugfs: Do not report currently active engine when describing objects") thanks, greg k-h ------------------ original commit in Linus's tree ------------------ >From a337b64f0d5717248a0c894e2618e658e6a9de9f Mon Sep 17 00:00:00 2001 From: Janusz Krzysztofik <janusz.krzysztofik@xxxxxxxxxxxxxxx> Date: Thu, 20 Jul 2023 11:35:44 +0200 Subject: [PATCH] drm/i915: Fix premature release of request's reusable memory Infinite waits for completion of GPU activity have been observed in CI, mostly inside __i915_active_wait(), triggered by igt@gem_barrier_race or igt@perf@stress-open-close. Root cause analysis, based of ftrace dumps generated with a lot of extra trace_printk() calls added to the code, revealed loops of request dependencies being accidentally built, preventing the requests from being processed, each waiting for completion of another one's activity. After we substitute a new request for a last active one tracked on a timeline, we set up a dependency of our new request to wait on completion of current activity of that previous one. While doing that, we must take care of keeping the old request still in memory until we use its attributes for setting up that await dependency, or we can happen to set up the await dependency on an unrelated request that already reuses the memory previously allocated to the old one, already released. Combined with perf adding consecutive kernel context remote requests to different user context timelines, unresolvable loops of await dependencies can be built, leading do infinite waits. We obtain a pointer to the previous request to wait upon when we substitute it with a pointer to our new request in an active tracker, e.g. in intel_timeline.last_request. In some processing paths we protect that old request from being freed before we use it by getting a reference to it under RCU protection, but in others, e.g. __i915_request_commit() -> __i915_request_add_to_timeline() -> __i915_request_ensure_ordering(), we don't. But anyway, since the requests' memory is SLAB_FAILSAFE_BY_RCU, that RCU protection is not sufficient against reuse of memory. We could protect i915_request's memory from being prematurely reused by calling its release function via call_rcu() and using rcu_read_lock() consequently, as proposed in v1. However, that approach leads to significant (up to 10 times) increase of SLAB utilization by i915_request SLAB cache. Another potential approach is to take a reference to the previous active fence. When updating an active fence tracker, we first lock the new fence, substitute a pointer of the current active fence with the new one, then we lock the substituted fence. With this approach, there is a time window after the substitution and before the lock when the request can be concurrently released by an interrupt handler and its memory reused, then we may happen to lock and return a new, unrelated request. Always get a reference to the current active fence first, before replacing it with a new one. Having it protected from premature release and reuse, lock it and then replace with the new one but only if not yet signalled via a potential concurrent interrupt nor replaced with another one by a potential concurrent thread, otherwise retry, starting from getting a reference to the new current one. Adjust users to not get a reference to the previous active fence themselves and always put the reference got by __i915_active_fence_set() when no longer needed. v3: Fix lockdep splat reports and other issues caused by incorrect use of try_cmpxchg() (use (cmpxchg() != prev) instead) v2: Protect request's memory by getting a reference to it in favor of delegating its release to call_rcu() (Chris) Closes: https://gitlab.freedesktop.org/drm/intel/-/issues/8211 Fixes: df9f85d8582e ("drm/i915: Serialise i915_active_fence_set() with itself") Suggested-by: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx> Signed-off-by: Janusz Krzysztofik <janusz.krzysztofik@xxxxxxxxxxxxxxx> Cc: <stable@xxxxxxxxxxxxxxx> # v5.6+ Reviewed-by: Andi Shyti <andi.shyti@xxxxxxxxxxxxxxx> Signed-off-by: Andi Shyti <andi.shyti@xxxxxxxxxxxxxxx> Link: https://patchwork.freedesktop.org/patch/msgid/20230720093543.832147-2-janusz.krzysztofik@xxxxxxxxxxxxxxx (cherry picked from commit 946e047a3d88d46d15b5c5af0414098e12b243f7) Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin@xxxxxxxxx> diff --git a/drivers/gpu/drm/i915/i915_active.c b/drivers/gpu/drm/i915/i915_active.c index 8ef93889061a..5ec293011d99 100644 --- a/drivers/gpu/drm/i915/i915_active.c +++ b/drivers/gpu/drm/i915/i915_active.c @@ -449,8 +449,11 @@ int i915_active_add_request(struct i915_active *ref, struct i915_request *rq) } } while (unlikely(is_barrier(active))); - if (!__i915_active_fence_set(active, fence)) + fence = __i915_active_fence_set(active, fence); + if (!fence) __i915_active_acquire(ref); + else + dma_fence_put(fence); out: i915_active_release(ref); @@ -469,13 +472,9 @@ __i915_active_set_fence(struct i915_active *ref, return NULL; } - rcu_read_lock(); prev = __i915_active_fence_set(active, fence); - if (prev) - prev = dma_fence_get_rcu(prev); - else + if (!prev) __i915_active_acquire(ref); - rcu_read_unlock(); return prev; } @@ -1019,10 +1018,11 @@ void i915_request_add_active_barriers(struct i915_request *rq) * * Records the new @fence as the last active fence along its timeline in * this active tracker, moving the tracking callbacks from the previous - * fence onto this one. Returns the previous fence (if not already completed), - * which the caller must ensure is executed before the new fence. To ensure - * that the order of fences within the timeline of the i915_active_fence is - * understood, it should be locked by the caller. + * fence onto this one. Gets and returns a reference to the previous fence + * (if not already completed), which the caller must put after making sure + * that it is executed before the new fence. To ensure that the order of + * fences within the timeline of the i915_active_fence is understood, it + * should be locked by the caller. */ struct dma_fence * __i915_active_fence_set(struct i915_active_fence *active, @@ -1031,7 +1031,23 @@ __i915_active_fence_set(struct i915_active_fence *active, struct dma_fence *prev; unsigned long flags; - if (fence == rcu_access_pointer(active->fence)) + /* + * In case of fences embedded in i915_requests, their memory is + * SLAB_FAILSAFE_BY_RCU, then it can be reused right after release + * by new requests. Then, there is a risk of passing back a pointer + * to a new, completely unrelated fence that reuses the same memory + * while tracked under a different active tracker. Combined with i915 + * perf open/close operations that build await dependencies between + * engine kernel context requests and user requests from different + * timelines, this can lead to dependency loops and infinite waits. + * + * As a countermeasure, we try to get a reference to the active->fence + * first, so if we succeed and pass it back to our user then it is not + * released and potentially reused by an unrelated request before the + * user has a chance to set up an await dependency on it. + */ + prev = i915_active_fence_get(active); + if (fence == prev) return fence; GEM_BUG_ON(test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)); @@ -1040,27 +1056,56 @@ __i915_active_fence_set(struct i915_active_fence *active, * Consider that we have two threads arriving (A and B), with * C already resident as the active->fence. * - * A does the xchg first, and so it sees C or NULL depending - * on the timing of the interrupt handler. If it is NULL, the - * previous fence must have been signaled and we know that - * we are first on the timeline. If it is still present, - * we acquire the lock on that fence and serialise with the interrupt - * handler, in the process removing it from any future interrupt - * callback. A will then wait on C before executing (if present). - * - * As B is second, it sees A as the previous fence and so waits for - * it to complete its transition and takes over the occupancy for - * itself -- remembering that it needs to wait on A before executing. + * Both A and B have got a reference to C or NULL, depending on the + * timing of the interrupt handler. Let's assume that if A has got C + * then it has locked C first (before B). * * Note the strong ordering of the timeline also provides consistent * nesting rules for the fence->lock; the inner lock is always the * older lock. */ spin_lock_irqsave(fence->lock, flags); - prev = xchg(__active_fence_slot(active), fence); - if (prev) { - GEM_BUG_ON(prev == fence); + if (prev) spin_lock_nested(prev->lock, SINGLE_DEPTH_NESTING); + + /* + * A does the cmpxchg first, and so it sees C or NULL, as before, or + * something else, depending on the timing of other threads and/or + * interrupt handler. If not the same as before then A unlocks C if + * applicable and retries, starting from an attempt to get a new + * active->fence. Meanwhile, B follows the same path as A. + * Once A succeeds with cmpxch, B fails again, retires, gets A from + * active->fence, locks it as soon as A completes, and possibly + * succeeds with cmpxchg. + */ + while (cmpxchg(__active_fence_slot(active), prev, fence) != prev) { + if (prev) { + spin_unlock(prev->lock); + dma_fence_put(prev); + } + spin_unlock_irqrestore(fence->lock, flags); + + prev = i915_active_fence_get(active); + GEM_BUG_ON(prev == fence); + + spin_lock_irqsave(fence->lock, flags); + if (prev) + spin_lock_nested(prev->lock, SINGLE_DEPTH_NESTING); + } + + /* + * If prev is NULL then the previous fence must have been signaled + * and we know that we are first on the timeline. If it is still + * present then, having the lock on that fence already acquired, we + * serialise with the interrupt handler, in the process of removing it + * from any future interrupt callback. A will then wait on C before + * executing (if present). + * + * As B is second, it sees A as the previous fence and so waits for + * it to complete its transition and takes over the occupancy for + * itself -- remembering that it needs to wait on A before executing. + */ + if (prev) { __list_del_entry(&active->cb.node); spin_unlock(prev->lock); /* serialise with prev->cb_list */ } @@ -1077,11 +1122,7 @@ int i915_active_fence_set(struct i915_active_fence *active, int err = 0; /* Must maintain timeline ordering wrt previous active requests */ - rcu_read_lock(); fence = __i915_active_fence_set(active, &rq->fence); - if (fence) /* but the previous fence may not belong to that timeline! */ - fence = dma_fence_get_rcu(fence); - rcu_read_unlock(); if (fence) { err = i915_request_await_dma_fence(rq, fence); dma_fence_put(fence); diff --git a/drivers/gpu/drm/i915/i915_request.c b/drivers/gpu/drm/i915/i915_request.c index 894068bb37b6..833b73edefdb 100644 --- a/drivers/gpu/drm/i915/i915_request.c +++ b/drivers/gpu/drm/i915/i915_request.c @@ -1661,6 +1661,11 @@ __i915_request_ensure_parallel_ordering(struct i915_request *rq, request_to_parent(rq)->parallel.last_rq = i915_request_get(rq); + /* + * Users have to put a reference potentially got by + * __i915_active_fence_set() to the returned request + * when no longer needed + */ return to_request(__i915_active_fence_set(&timeline->last_request, &rq->fence)); } @@ -1707,6 +1712,10 @@ __i915_request_ensure_ordering(struct i915_request *rq, 0); } + /* + * Users have to put the reference to prev potentially got + * by __i915_active_fence_set() when no longer needed + */ return prev; } @@ -1760,6 +1769,8 @@ __i915_request_add_to_timeline(struct i915_request *rq) prev = __i915_request_ensure_ordering(rq, timeline); else prev = __i915_request_ensure_parallel_ordering(rq, timeline); + if (prev) + i915_request_put(prev); /* * Make sure that no request gazumped us - if it was allocated after
