[PATCH 04/20] drm/radeon: convert fence to uint64_t v4

Jerome Glisse j.glisse at gmail.com
Mon May 7 07:39:26 PDT 2012


On Mon, May 7, 2012 at 7:42 AM, Christian König <deathsimple at vodafone.de> wrote:
> From: Jerome Glisse <jglisse at redhat.com>
>
> This convert fence to use uint64_t sequence number intention is
> to use the fact that uin64_t is big enough that we don't need to
> care about wrap around.
>
> Tested with and without writeback using 0xFFFFF000 as initial
> fence sequence and thus allowing to test the wrap around from
> 32bits to 64bits.
>
> v2: Add comment about possible race btw CPU & GPU, add comment
>    stressing that we need 2 dword aligned for R600_WB_EVENT_OFFSET
>    Read fence sequenc in reverse order of GPU write them so we
>    mitigate the race btw CPU and GPU.
>
> v3: Drop the need for ring to emit the 64bits fence, and just have
>    each ring emit the lower 32bits of the fence sequence. We
>    handle the wrap over 32bits in fence_process.
>
> v4: Just a small optimization: Don't reread the last_seq value
>    if loop restarts, since we already know its value anyway.
>    Also start at zero not one for seq value and use pre instead
>    of post increment in emmit, otherwise wait_empty will deadlock.

Why changing that v3 was already good no deadlock. I started at 1
especialy for that, a signaled fence is set to 0 so it always compare
as signaled. Just using preincrement is exactly like starting at one.
I don't see the need for this change but if it makes you happy.

Cheers,
Jerome
>
> Signed-off-by: Jerome Glisse <jglisse at redhat.com>
> Signed-off-by: Christian König <deathsimple at vodafone.de>
> ---
>  drivers/gpu/drm/radeon/radeon.h       |   39 ++++++-----
>  drivers/gpu/drm/radeon/radeon_fence.c |  116 +++++++++++++++++++++++----------
>  drivers/gpu/drm/radeon/radeon_ring.c  |    9 ++-
>  3 files changed, 107 insertions(+), 57 deletions(-)
>
> diff --git a/drivers/gpu/drm/radeon/radeon.h b/drivers/gpu/drm/radeon/radeon.h
> index e99ea81..cdf46bc 100644
> --- a/drivers/gpu/drm/radeon/radeon.h
> +++ b/drivers/gpu/drm/radeon/radeon.h
> @@ -100,28 +100,32 @@ extern int radeon_lockup_timeout;
>  * Copy from radeon_drv.h so we don't have to include both and have conflicting
>  * symbol;
>  */
> -#define RADEON_MAX_USEC_TIMEOUT                100000  /* 100 ms */
> -#define RADEON_FENCE_JIFFIES_TIMEOUT   (HZ / 2)
> +#define RADEON_MAX_USEC_TIMEOUT                        100000  /* 100 ms */
> +#define RADEON_FENCE_JIFFIES_TIMEOUT           (HZ / 2)
>  /* RADEON_IB_POOL_SIZE must be a power of 2 */
> -#define RADEON_IB_POOL_SIZE            16
> -#define RADEON_DEBUGFS_MAX_COMPONENTS  32
> -#define RADEONFB_CONN_LIMIT            4
> -#define RADEON_BIOS_NUM_SCRATCH                8
> +#define RADEON_IB_POOL_SIZE                    16
> +#define RADEON_DEBUGFS_MAX_COMPONENTS          32
> +#define RADEONFB_CONN_LIMIT                    4
> +#define RADEON_BIOS_NUM_SCRATCH                        8
>
>  /* max number of rings */
> -#define RADEON_NUM_RINGS 3
> +#define RADEON_NUM_RINGS                       3
> +
> +/* fence seq are set to this number when signaled */
> +#define RADEON_FENCE_SIGNALED_SEQ              0LL
> +#define RADEON_FENCE_NOTEMITED_SEQ             (~0LL)
>
>  /* internal ring indices */
>  /* r1xx+ has gfx CP ring */
> -#define RADEON_RING_TYPE_GFX_INDEX  0
> +#define RADEON_RING_TYPE_GFX_INDEX             0
>
>  /* cayman has 2 compute CP rings */
> -#define CAYMAN_RING_TYPE_CP1_INDEX 1
> -#define CAYMAN_RING_TYPE_CP2_INDEX 2
> +#define CAYMAN_RING_TYPE_CP1_INDEX             1
> +#define CAYMAN_RING_TYPE_CP2_INDEX             2
>
>  /* hardcode those limit for now */
> -#define RADEON_VA_RESERVED_SIZE                (8 << 20)
> -#define RADEON_IB_VM_MAX_SIZE          (64 << 10)
> +#define RADEON_VA_RESERVED_SIZE                        (8 << 20)
> +#define RADEON_IB_VM_MAX_SIZE                  (64 << 10)
>
>  /*
>  * Errata workarounds.
> @@ -254,8 +258,9 @@ struct radeon_fence_driver {
>        uint32_t                        scratch_reg;
>        uint64_t                        gpu_addr;
>        volatile uint32_t               *cpu_addr;
> -       atomic_t                        seq;
> -       uint32_t                        last_seq;
> +       /* seq is protected by ring emission lock */
> +       uint64_t                        seq;
> +       atomic64_t                      last_seq;
>        unsigned long                   last_activity;
>        wait_queue_head_t               queue;
>        struct list_head                emitted;
> @@ -268,11 +273,9 @@ struct radeon_fence {
>        struct kref                     kref;
>        struct list_head                list;
>        /* protected by radeon_fence.lock */
> -       uint32_t                        seq;
> -       bool                            emitted;
> -       bool                            signaled;
> +       uint64_t                        seq;
>        /* RB, DMA, etc. */
> -       int                             ring;
> +       unsigned                        ring;
>        struct radeon_semaphore         *semaphore;
>  };
>
> diff --git a/drivers/gpu/drm/radeon/radeon_fence.c b/drivers/gpu/drm/radeon/radeon_fence.c
> index 5bb78bf..feb2bbc 100644
> --- a/drivers/gpu/drm/radeon/radeon_fence.c
> +++ b/drivers/gpu/drm/radeon/radeon_fence.c
> @@ -66,14 +66,14 @@ int radeon_fence_emit(struct radeon_device *rdev, struct radeon_fence *fence)
>        unsigned long irq_flags;
>
>        write_lock_irqsave(&rdev->fence_lock, irq_flags);
> -       if (fence->emitted) {
> +       if (fence->seq && fence->seq < RADEON_FENCE_NOTEMITED_SEQ) {
>                write_unlock_irqrestore(&rdev->fence_lock, irq_flags);
>                return 0;
>        }
> -       fence->seq = atomic_add_return(1, &rdev->fence_drv[fence->ring].seq);
> +       /* we are protected by the ring emission mutex */
> +       fence->seq = ++rdev->fence_drv[fence->ring].seq;
>        radeon_fence_ring_emit(rdev, fence->ring, fence);
>        trace_radeon_fence_emit(rdev->ddev, fence->seq);
> -       fence->emitted = true;
>        /* are we the first fence on a previusly idle ring? */
>        if (list_empty(&rdev->fence_drv[fence->ring].emitted)) {
>                rdev->fence_drv[fence->ring].last_activity = jiffies;
> @@ -87,14 +87,60 @@ static bool radeon_fence_poll_locked(struct radeon_device *rdev, int ring)
>  {
>        struct radeon_fence *fence;
>        struct list_head *i, *n;
> -       uint32_t seq;
> +       uint64_t seq, last_seq;
> +       unsigned count_loop = 0;
>        bool wake = false;
>
> -       seq = radeon_fence_read(rdev, ring);
> -       if (seq == rdev->fence_drv[ring].last_seq)
> -               return false;
> +       /* Note there is a scenario here for an infinite loop but it's
> +        * very unlikely to happen. For it to happen, the current polling
> +        * process need to be interrupted by another process and another
> +        * process needs to update the last_seq btw the atomic read and
> +        * xchg of the current process.
> +        *
> +        * More over for this to go in infinite loop there need to be
> +        * continuously new fence signaled ie radeon_fence_read needs
> +        * to return a different value each time for both the currently
> +        * polling process and the other process that xchg the last_seq
> +        * btw atomic read and xchg of the current process. And the
> +        * value the other process set as last seq must be higher than
> +        * the seq value we just read. Which means that current process
> +        * need to be interrupted after radeon_fence_read and before
> +        * atomic xchg.
> +        *
> +        * To be even more safe we count the number of time we loop and
> +        * we bail after 10 loop just accepting the fact that we might
> +        * have temporarly set the last_seq not to the true real last
> +        * seq but to an older one.
> +        */
> +       last_seq = atomic64_read(&rdev->fence_drv[ring].last_seq);
> +       do {
> +               seq = radeon_fence_read(rdev, ring);
> +               seq |= last_seq & 0xffffffff00000000LL;
> +               if (seq < last_seq) {
> +                       seq += 0x100000000LL;
> +               }
>
> -       rdev->fence_drv[ring].last_seq = seq;
> +               if (!wake && seq == last_seq) {
> +                       return false;
> +               }
> +               /* If we loop over we don't want to return without
> +                * checking if a fence is signaled as it means that the
> +                * seq we just read is different from the previous on.
> +                */
> +               wake = true;
> +               if ((count_loop++) > 10) {
> +                       /* We looped over too many time leave with the
> +                        * fact that we might have set an older fence
> +                        * seq then the current real last seq as signaled
> +                        * by the hw.
> +                        */
> +                       break;
> +               }
> +               last_seq = seq;
> +       } while (atomic64_xchg(&rdev->fence_drv[ring].last_seq, seq) > seq);
> +
> +       /* reset wake to false */
> +       wake = false;
>        rdev->fence_drv[ring].last_activity = jiffies;
>
>        n = NULL;
> @@ -112,7 +158,7 @@ static bool radeon_fence_poll_locked(struct radeon_device *rdev, int ring)
>                        n = i->prev;
>                        list_move_tail(i, &rdev->fence_drv[ring].signaled);
>                        fence = list_entry(i, struct radeon_fence, list);
> -                       fence->signaled = true;
> +                       fence->seq = RADEON_FENCE_SIGNALED_SEQ;
>                        i = n;
>                } while (i != &rdev->fence_drv[ring].emitted);
>                wake = true;
> @@ -128,7 +174,7 @@ static void radeon_fence_destroy(struct kref *kref)
>        fence = container_of(kref, struct radeon_fence, kref);
>        write_lock_irqsave(&fence->rdev->fence_lock, irq_flags);
>        list_del(&fence->list);
> -       fence->emitted = false;
> +       fence->seq = RADEON_FENCE_NOTEMITED_SEQ;
>        write_unlock_irqrestore(&fence->rdev->fence_lock, irq_flags);
>        if (fence->semaphore)
>                radeon_semaphore_free(fence->rdev, fence->semaphore);
> @@ -145,9 +191,7 @@ int radeon_fence_create(struct radeon_device *rdev,
>        }
>        kref_init(&((*fence)->kref));
>        (*fence)->rdev = rdev;
> -       (*fence)->emitted = false;
> -       (*fence)->signaled = false;
> -       (*fence)->seq = 0;
> +       (*fence)->seq = RADEON_FENCE_NOTEMITED_SEQ;
>        (*fence)->ring = ring;
>        (*fence)->semaphore = NULL;
>        INIT_LIST_HEAD(&(*fence)->list);
> @@ -163,18 +207,18 @@ bool radeon_fence_signaled(struct radeon_fence *fence)
>                return true;
>
>        write_lock_irqsave(&fence->rdev->fence_lock, irq_flags);
> -       signaled = fence->signaled;
> +       signaled = (fence->seq == RADEON_FENCE_SIGNALED_SEQ);
>        /* if we are shuting down report all fence as signaled */
>        if (fence->rdev->shutdown) {
>                signaled = true;
>        }
> -       if (!fence->emitted) {
> +       if (fence->seq == RADEON_FENCE_NOTEMITED_SEQ) {
>                WARN(1, "Querying an unemitted fence : %p !\n", fence);
>                signaled = true;
>        }
>        if (!signaled) {
>                radeon_fence_poll_locked(fence->rdev, fence->ring);
> -               signaled = fence->signaled;
> +               signaled = (fence->seq == RADEON_FENCE_SIGNALED_SEQ);
>        }
>        write_unlock_irqrestore(&fence->rdev->fence_lock, irq_flags);
>        return signaled;
> @@ -183,8 +227,8 @@ bool radeon_fence_signaled(struct radeon_fence *fence)
>  int radeon_fence_wait(struct radeon_fence *fence, bool intr)
>  {
>        struct radeon_device *rdev;
> -       unsigned long irq_flags, timeout;
> -       u32 seq;
> +       unsigned long irq_flags, timeout, last_activity;
> +       uint64_t seq;
>        int i, r;
>        bool signaled;
>
> @@ -207,7 +251,9 @@ int radeon_fence_wait(struct radeon_fence *fence, bool intr)
>                        timeout = 1;
>                }
>                /* save current sequence value used to check for GPU lockups */
> -               seq = rdev->fence_drv[fence->ring].last_seq;
> +               seq = atomic64_read(&rdev->fence_drv[fence->ring].last_seq);
> +               /* Save current last activity valuee, used to check for GPU lockups */
> +               last_activity = rdev->fence_drv[fence->ring].last_activity;
>                read_unlock_irqrestore(&rdev->fence_lock, irq_flags);
>
>                trace_radeon_fence_wait_begin(rdev->ddev, seq);
> @@ -235,24 +281,23 @@ int radeon_fence_wait(struct radeon_fence *fence, bool intr)
>                        }
>
>                        write_lock_irqsave(&rdev->fence_lock, irq_flags);
> -                       /* check if sequence value has changed since last_activity */
> -                       if (seq != rdev->fence_drv[fence->ring].last_seq) {
> +                       /* test if somebody else has already decided that this is a lockup */
> +                       if (last_activity != rdev->fence_drv[fence->ring].last_activity) {
>                                write_unlock_irqrestore(&rdev->fence_lock, irq_flags);
>                                continue;
>                        }
>
> -                       /* change sequence value on all rings, so nobody else things there is a lockup */
> -                       for (i = 0; i < RADEON_NUM_RINGS; ++i)
> -                               rdev->fence_drv[i].last_seq -= 0x10000;
> -
> -                       rdev->fence_drv[fence->ring].last_activity = jiffies;
>                        write_unlock_irqrestore(&rdev->fence_lock, irq_flags);
>
>                        if (radeon_ring_is_lockup(rdev, fence->ring, &rdev->ring[fence->ring])) {
> -
>                                /* good news we believe it's a lockup */
> -                               printk(KERN_WARNING "GPU lockup (waiting for 0x%08X last fence id 0x%08X)\n",
> -                                    fence->seq, seq);
> +                               dev_warn(rdev->dev, "GPU lockup (waiting for 0x%016llx last fence id 0x%016llx)\n",
> +                                        fence->seq, seq);
> +
> +                               /* change last activity so nobody else think there is a lockup */
> +                               for (i = 0; i < RADEON_NUM_RINGS; ++i) {
> +                                       rdev->fence_drv[i].last_activity = jiffies;
> +                               }
>
>                                /* mark the ring as not ready any more */
>                                rdev->ring[fence->ring].ready = false;
> @@ -387,9 +432,9 @@ int radeon_fence_driver_start_ring(struct radeon_device *rdev, int ring)
>        }
>        rdev->fence_drv[ring].cpu_addr = &rdev->wb.wb[index/4];
>        rdev->fence_drv[ring].gpu_addr = rdev->wb.gpu_addr + index;
> -       radeon_fence_write(rdev, atomic_read(&rdev->fence_drv[ring].seq), ring);
> +       radeon_fence_write(rdev, rdev->fence_drv[ring].seq, ring);
>        rdev->fence_drv[ring].initialized = true;
> -       DRM_INFO("fence driver on ring %d use gpu addr 0x%08Lx and cpu addr 0x%p\n",
> +       DRM_INFO("fence driver on ring %d use gpu addr 0x%016llx and cpu addr 0x%p\n",
>                 ring, rdev->fence_drv[ring].gpu_addr, rdev->fence_drv[ring].cpu_addr);
>        write_unlock_irqrestore(&rdev->fence_lock, irq_flags);
>        return 0;
> @@ -400,7 +445,8 @@ static void radeon_fence_driver_init_ring(struct radeon_device *rdev, int ring)
>        rdev->fence_drv[ring].scratch_reg = -1;
>        rdev->fence_drv[ring].cpu_addr = NULL;
>        rdev->fence_drv[ring].gpu_addr = 0;
> -       atomic_set(&rdev->fence_drv[ring].seq, 0);
> +       rdev->fence_drv[ring].seq = 0;
> +       atomic64_set(&rdev->fence_drv[ring].last_seq, 0);
>        INIT_LIST_HEAD(&rdev->fence_drv[ring].emitted);
>        INIT_LIST_HEAD(&rdev->fence_drv[ring].signaled);
>        init_waitqueue_head(&rdev->fence_drv[ring].queue);
> @@ -458,12 +504,12 @@ static int radeon_debugfs_fence_info(struct seq_file *m, void *data)
>                        continue;
>
>                seq_printf(m, "--- ring %d ---\n", i);
> -               seq_printf(m, "Last signaled fence 0x%08X\n",
> -                          radeon_fence_read(rdev, i));
> +               seq_printf(m, "Last signaled fence 0x%016lx\n",
> +                          atomic64_read(&rdev->fence_drv[i].last_seq));
>                if (!list_empty(&rdev->fence_drv[i].emitted)) {
>                        fence = list_entry(rdev->fence_drv[i].emitted.prev,
>                                           struct radeon_fence, list);
> -                       seq_printf(m, "Last emitted fence %p with 0x%08X\n",
> +                       seq_printf(m, "Last emitted fence %p with 0x%016llx\n",
>                                   fence,  fence->seq);
>                }
>        }
> diff --git a/drivers/gpu/drm/radeon/radeon_ring.c b/drivers/gpu/drm/radeon/radeon_ring.c
> index a4d60ae..4ae222b 100644
> --- a/drivers/gpu/drm/radeon/radeon_ring.c
> +++ b/drivers/gpu/drm/radeon/radeon_ring.c
> @@ -82,7 +82,7 @@ bool radeon_ib_try_free(struct radeon_device *rdev, struct radeon_ib *ib)
>        bool done = false;
>
>        /* only free ib which have been emited */
> -       if (ib->fence && ib->fence->emitted) {
> +       if (ib->fence && ib->fence->seq < RADEON_FENCE_NOTEMITED_SEQ) {
>                if (radeon_fence_signaled(ib->fence)) {
>                        radeon_fence_unref(&ib->fence);
>                        radeon_sa_bo_free(rdev, &ib->sa_bo);
> @@ -149,8 +149,9 @@ retry:
>        /* this should be rare event, ie all ib scheduled none signaled yet.
>         */
>        for (i = 0; i < RADEON_IB_POOL_SIZE; i++) {
> -               if (rdev->ib_pool.ibs[idx].fence && rdev->ib_pool.ibs[idx].fence->emitted) {
> -                       r = radeon_fence_wait(rdev->ib_pool.ibs[idx].fence, false);
> +               struct radeon_fence *fence = rdev->ib_pool.ibs[idx].fence;
> +               if (fence && fence->seq < RADEON_FENCE_NOTEMITED_SEQ) {
> +                       r = radeon_fence_wait(fence, false);
>                        if (!r) {
>                                goto retry;
>                        }
> @@ -173,7 +174,7 @@ void radeon_ib_free(struct radeon_device *rdev, struct radeon_ib **ib)
>                return;
>        }
>        radeon_mutex_lock(&rdev->ib_pool.mutex);
> -       if (tmp->fence && !tmp->fence->emitted) {
> +       if (tmp->fence && tmp->fence->seq == RADEON_FENCE_NOTEMITED_SEQ) {
>                radeon_sa_bo_free(rdev, &tmp->sa_bo);
>                radeon_fence_unref(&tmp->fence);
>        }
> --
> 1.7.5.4
>


More information about the dri-devel mailing list