There is a deadlock between the irq and fctx locks, the irq handling takes irq then fctx lock the fence signalling takes fctx then irq lock This splits the fence signalling path so the code that hits the irq lock is done in a separate work queue. This seems to fix crashes/hangs when using nouveau gsp with i915 primary GPU. Signed-off-by: Dave Airlie <airlied@xxxxxxxxxx> --- drivers/gpu/drm/nouveau/nouveau_fence.c | 28 ++++++++++++++++++++----- drivers/gpu/drm/nouveau/nouveau_fence.h | 5 ++++- 2 files changed, 27 insertions(+), 6 deletions(-) diff --git a/drivers/gpu/drm/nouveau/nouveau_fence.c b/drivers/gpu/drm/nouveau/nouveau_fence.c index ca762ea55413..5057d976fa57 100644 --- a/drivers/gpu/drm/nouveau/nouveau_fence.c +++ b/drivers/gpu/drm/nouveau/nouveau_fence.c @@ -62,7 +62,7 @@ nouveau_fence_signal(struct nouveau_fence *fence) if (test_bit(DMA_FENCE_FLAG_USER_BITS, &fence->base.flags)) { struct nouveau_fence_chan *fctx = nouveau_fctx(fence); - if (!--fctx->notify_ref) + if (atomic_dec_and_test(&fctx->notify_ref)) drop = 1; } @@ -103,6 +103,7 @@ nouveau_fence_context_kill(struct nouveau_fence_chan *fctx, int error) void nouveau_fence_context_del(struct nouveau_fence_chan *fctx) { + cancel_work_sync(&fctx->allow_block_work); nouveau_fence_context_kill(fctx, 0); nvif_event_dtor(&fctx->event); fctx->dead = 1; @@ -167,6 +168,18 @@ nouveau_fence_wait_uevent_handler(struct nvif_event *event, void *repv, u32 repc return ret; } +static void +nouveau_fence_work_allow_block(struct work_struct *work) +{ + struct nouveau_fence_chan *fctx = container_of(work, struct nouveau_fence_chan, + allow_block_work); + + if (atomic_read(&fctx->notify_ref) == 0) + nvif_event_block(&fctx->event); + else + nvif_event_allow(&fctx->event); +} + void nouveau_fence_context_new(struct nouveau_channel *chan, struct nouveau_fence_chan *fctx) { @@ -178,6 +191,7 @@ nouveau_fence_context_new(struct nouveau_channel *chan, struct nouveau_fence_cha } args; int ret; + INIT_WORK(&fctx->allow_block_work, nouveau_fence_work_allow_block); INIT_LIST_HEAD(&fctx->flip); INIT_LIST_HEAD(&fctx->pending); spin_lock_init(&fctx->lock); @@ -521,15 +535,19 @@ static bool nouveau_fence_enable_signaling(struct dma_fence *f) struct nouveau_fence *fence = from_fence(f); struct nouveau_fence_chan *fctx = nouveau_fctx(fence); bool ret; + bool do_work; - if (!fctx->notify_ref++) - nvif_event_allow(&fctx->event); + if (atomic_inc_return(&fctx->notify_ref) == 0) + do_work = true; ret = nouveau_fence_no_signaling(f); if (ret) set_bit(DMA_FENCE_FLAG_USER_BITS, &fence->base.flags); - else if (!--fctx->notify_ref) - nvif_event_block(&fctx->event); + else if (atomic_dec_and_test(&fctx->notify_ref)) + do_work = true; + + if (do_work) + schedule_work(&fctx->allow_block_work); return ret; } diff --git a/drivers/gpu/drm/nouveau/nouveau_fence.h b/drivers/gpu/drm/nouveau/nouveau_fence.h index 64d33ae7f356..28f5cf013b89 100644 --- a/drivers/gpu/drm/nouveau/nouveau_fence.h +++ b/drivers/gpu/drm/nouveau/nouveau_fence.h @@ -3,6 +3,7 @@ #define __NOUVEAU_FENCE_H__ #include <linux/dma-fence.h> +#include <linux/workqueue.h> #include <nvif/event.h> struct nouveau_drm; @@ -45,7 +46,9 @@ struct nouveau_fence_chan { char name[32]; struct nvif_event event; - int notify_ref, dead, killed; + struct work_struct allow_block_work; + atomic_t notify_ref; + int dead, killed; }; struct nouveau_fence_priv { -- 2.43.0