This is a note to let you know that I've just added the patch titled drm/vmwgfx: Replace the hw mutex with a hw spinlock to the 3.18-stable tree which can be found at: http://www.kernel.org/git/?p=linux/kernel/git/stable/stable-queue.git;a=summary The filename of the patch is: drm-vmwgfx-replace-the-hw-mutex-with-a-hw-spinlock.patch and it can be found in the queue-3.18 subdirectory. If you, or anyone else, feels it should not be added to the stable tree, please let <stable@xxxxxxxxxxxxxxx> know about it. >From 496eb6fd2c3fd13f4b914e537598e5c86ce4f52a Mon Sep 17 00:00:00 2001 From: Thomas Hellstrom <thellstrom@xxxxxxxxxx> Date: Wed, 14 Jan 2015 02:33:39 -0800 Subject: drm/vmwgfx: Replace the hw mutex with a hw spinlock From: Thomas Hellstrom <thellstrom@xxxxxxxxxx> commit 496eb6fd2c3fd13f4b914e537598e5c86ce4f52a upstream. Fixes a case where we call vmw_fifo_idle() from within a wait function with task state !TASK_RUNNING, which is illegal. In addition, make the locking fine-grained, so that it is performed once for every read- and write operation. This is of course more costly, but we don't perform much register access in the timing critical paths anyway. Instead we have the extra benefit of being sure that we don't forget the hw lock around register accesses. I think currently the kms code was quite buggy w r t this. This fixes Red Hat Bugzilla Bug 1180796 Signed-off-by: Thomas Hellstrom <thellstrom@xxxxxxxxxx> Reviewed-by: Jakob Bornecrantz <jakob@xxxxxxxxxx> Signed-off-by: Greg Kroah-Hartman <gregkh@xxxxxxxxxxxxxxxxxxx> --- drivers/gpu/drm/vmwgfx/vmwgfx_drv.c | 28 ++++---------------------- drivers/gpu/drm/vmwgfx/vmwgfx_drv.h | 25 +++++++++++++++++++---- drivers/gpu/drm/vmwgfx/vmwgfx_fence.c | 18 +---------------- drivers/gpu/drm/vmwgfx/vmwgfx_fifo.c | 36 ++++++++++++++-------------------- drivers/gpu/drm/vmwgfx/vmwgfx_ioctl.c | 8 +++---- drivers/gpu/drm/vmwgfx/vmwgfx_irq.c | 25 ++++++++--------------- drivers/gpu/drm/vmwgfx/vmwgfx_kms.c | 2 - 7 files changed, 56 insertions(+), 86 deletions(-) --- a/drivers/gpu/drm/vmwgfx/vmwgfx_drv.c +++ b/drivers/gpu/drm/vmwgfx/vmwgfx_drv.c @@ -406,11 +406,9 @@ int vmw_3d_resource_inc(struct vmw_priva if (unlikely(ret != 0)) --dev_priv->num_3d_resources; } else if (unhide_svga) { - mutex_lock(&dev_priv->hw_mutex); vmw_write(dev_priv, SVGA_REG_ENABLE, vmw_read(dev_priv, SVGA_REG_ENABLE) & ~SVGA_REG_ENABLE_HIDE); - mutex_unlock(&dev_priv->hw_mutex); } mutex_unlock(&dev_priv->release_mutex); @@ -433,13 +431,10 @@ void vmw_3d_resource_dec(struct vmw_priv mutex_lock(&dev_priv->release_mutex); if (unlikely(--dev_priv->num_3d_resources == 0)) vmw_release_device(dev_priv); - else if (hide_svga) { - mutex_lock(&dev_priv->hw_mutex); + else if (hide_svga) vmw_write(dev_priv, SVGA_REG_ENABLE, vmw_read(dev_priv, SVGA_REG_ENABLE) | SVGA_REG_ENABLE_HIDE); - mutex_unlock(&dev_priv->hw_mutex); - } n3d = (int32_t) dev_priv->num_3d_resources; mutex_unlock(&dev_priv->release_mutex); @@ -600,12 +595,14 @@ static int vmw_driver_load(struct drm_de dev_priv->dev = dev; dev_priv->vmw_chipset = chipset; dev_priv->last_read_seqno = (uint32_t) -100; - mutex_init(&dev_priv->hw_mutex); mutex_init(&dev_priv->cmdbuf_mutex); mutex_init(&dev_priv->release_mutex); mutex_init(&dev_priv->binding_mutex); rwlock_init(&dev_priv->resource_lock); ttm_lock_init(&dev_priv->reservation_sem); + spin_lock_init(&dev_priv->hw_lock); + spin_lock_init(&dev_priv->waiter_lock); + spin_lock_init(&dev_priv->cap_lock); for (i = vmw_res_context; i < vmw_res_max; ++i) { idr_init(&dev_priv->res_idr[i]); @@ -626,14 +623,11 @@ static int vmw_driver_load(struct drm_de dev_priv->enable_fb = enable_fbdev; - mutex_lock(&dev_priv->hw_mutex); - vmw_write(dev_priv, SVGA_REG_ID, SVGA_ID_2); svga_id = vmw_read(dev_priv, SVGA_REG_ID); if (svga_id != SVGA_ID_2) { ret = -ENOSYS; DRM_ERROR("Unsupported SVGA ID 0x%x\n", svga_id); - mutex_unlock(&dev_priv->hw_mutex); goto out_err0; } @@ -683,10 +677,8 @@ static int vmw_driver_load(struct drm_de dev_priv->prim_bb_mem = dev_priv->vram_size; ret = vmw_dma_masks(dev_priv); - if (unlikely(ret != 0)) { - mutex_unlock(&dev_priv->hw_mutex); + if (unlikely(ret != 0)) goto out_err0; - } /* * Limit back buffer size to VRAM size. Remove this once @@ -695,8 +687,6 @@ static int vmw_driver_load(struct drm_de if (dev_priv->prim_bb_mem > dev_priv->vram_size) dev_priv->prim_bb_mem = dev_priv->vram_size; - mutex_unlock(&dev_priv->hw_mutex); - vmw_print_capabilities(dev_priv->capabilities); if (dev_priv->capabilities & SVGA_CAP_GMR2) { @@ -1161,9 +1151,7 @@ static int vmw_master_set(struct drm_dev if (unlikely(ret != 0)) return ret; vmw_kms_save_vga(dev_priv); - mutex_lock(&dev_priv->hw_mutex); vmw_write(dev_priv, SVGA_REG_TRACES, 0); - mutex_unlock(&dev_priv->hw_mutex); } if (active) { @@ -1197,9 +1185,7 @@ out_no_active_lock: if (!dev_priv->enable_fb) { vmw_kms_restore_vga(dev_priv); vmw_3d_resource_dec(dev_priv, true); - mutex_lock(&dev_priv->hw_mutex); vmw_write(dev_priv, SVGA_REG_TRACES, 1); - mutex_unlock(&dev_priv->hw_mutex); } return ret; } @@ -1234,9 +1220,7 @@ static void vmw_master_drop(struct drm_d DRM_ERROR("Unable to clean VRAM on master drop.\n"); vmw_kms_restore_vga(dev_priv); vmw_3d_resource_dec(dev_priv, true); - mutex_lock(&dev_priv->hw_mutex); vmw_write(dev_priv, SVGA_REG_TRACES, 1); - mutex_unlock(&dev_priv->hw_mutex); } dev_priv->active_master = &dev_priv->fbdev_master; @@ -1368,10 +1352,8 @@ static void vmw_pm_complete(struct devic struct drm_device *dev = pci_get_drvdata(pdev); struct vmw_private *dev_priv = vmw_priv(dev); - mutex_lock(&dev_priv->hw_mutex); vmw_write(dev_priv, SVGA_REG_ID, SVGA_ID_2); (void) vmw_read(dev_priv, SVGA_REG_ID); - mutex_unlock(&dev_priv->hw_mutex); /** * Reclaim 3d reference held by fbdev and potentially --- a/drivers/gpu/drm/vmwgfx/vmwgfx_drv.h +++ b/drivers/gpu/drm/vmwgfx/vmwgfx_drv.h @@ -399,7 +399,8 @@ struct vmw_private { uint32_t memory_size; bool has_gmr; bool has_mob; - struct mutex hw_mutex; + spinlock_t hw_lock; + spinlock_t cap_lock; /* * VGA registers. @@ -449,8 +450,9 @@ struct vmw_private { atomic_t marker_seq; wait_queue_head_t fence_queue; wait_queue_head_t fifo_queue; - int fence_queue_waiters; /* Protected by hw_mutex */ - int goal_queue_waiters; /* Protected by hw_mutex */ + spinlock_t waiter_lock; + int fence_queue_waiters; /* Protected by waiter_lock */ + int goal_queue_waiters; /* Protected by waiter_lock */ atomic_t fifo_queue_waiters; uint32_t last_read_seqno; spinlock_t irq_lock; @@ -553,20 +555,35 @@ static inline struct vmw_master *vmw_mas return (struct vmw_master *) master->driver_priv; } +/* + * The locking here is fine-grained, so that it is performed once + * for every read- and write operation. This is of course costly, but we + * don't perform much register access in the timing critical paths anyway. + * Instead we have the extra benefit of being sure that we don't forget + * the hw lock around register accesses. + */ static inline void vmw_write(struct vmw_private *dev_priv, unsigned int offset, uint32_t value) { + unsigned long irq_flags; + + spin_lock_irqsave(&dev_priv->hw_lock, irq_flags); outl(offset, dev_priv->io_start + VMWGFX_INDEX_PORT); outl(value, dev_priv->io_start + VMWGFX_VALUE_PORT); + spin_unlock_irqrestore(&dev_priv->hw_lock, irq_flags); } static inline uint32_t vmw_read(struct vmw_private *dev_priv, unsigned int offset) { - uint32_t val; + unsigned long irq_flags; + u32 val; + spin_lock_irqsave(&dev_priv->hw_lock, irq_flags); outl(offset, dev_priv->io_start + VMWGFX_INDEX_PORT); val = inl(dev_priv->io_start + VMWGFX_VALUE_PORT); + spin_unlock_irqrestore(&dev_priv->hw_lock, irq_flags); + return val; } --- a/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c +++ b/drivers/gpu/drm/vmwgfx/vmwgfx_fence.c @@ -35,7 +35,7 @@ struct vmw_fence_manager { struct vmw_private *dev_priv; spinlock_t lock; struct list_head fence_list; - struct work_struct work, ping_work; + struct work_struct work; u32 user_fence_size; u32 fence_size; u32 event_fence_action_size; @@ -134,14 +134,6 @@ static const char *vmw_fence_get_timelin return "svga"; } -static void vmw_fence_ping_func(struct work_struct *work) -{ - struct vmw_fence_manager *fman = - container_of(work, struct vmw_fence_manager, ping_work); - - vmw_fifo_ping_host(fman->dev_priv, SVGA_SYNC_GENERIC); -} - static bool vmw_fence_enable_signaling(struct fence *f) { struct vmw_fence_obj *fence = @@ -155,11 +147,7 @@ static bool vmw_fence_enable_signaling(s if (seqno - fence->base.seqno < VMW_FENCE_WRAP) return false; - if (mutex_trylock(&dev_priv->hw_mutex)) { - vmw_fifo_ping_host_locked(dev_priv, SVGA_SYNC_GENERIC); - mutex_unlock(&dev_priv->hw_mutex); - } else - schedule_work(&fman->ping_work); + vmw_fifo_ping_host(dev_priv, SVGA_SYNC_GENERIC); return true; } @@ -305,7 +293,6 @@ struct vmw_fence_manager *vmw_fence_mana INIT_LIST_HEAD(&fman->fence_list); INIT_LIST_HEAD(&fman->cleanup_list); INIT_WORK(&fman->work, &vmw_fence_work_func); - INIT_WORK(&fman->ping_work, &vmw_fence_ping_func); fman->fifo_down = true; fman->user_fence_size = ttm_round_pot(sizeof(struct vmw_user_fence)); fman->fence_size = ttm_round_pot(sizeof(struct vmw_fence_obj)); @@ -323,7 +310,6 @@ void vmw_fence_manager_takedown(struct v bool lists_empty; (void) cancel_work_sync(&fman->work); - (void) cancel_work_sync(&fman->ping_work); spin_lock_irqsave(&fman->lock, irq_flags); lists_empty = list_empty(&fman->fence_list) && --- a/drivers/gpu/drm/vmwgfx/vmwgfx_fifo.c +++ b/drivers/gpu/drm/vmwgfx/vmwgfx_fifo.c @@ -44,10 +44,10 @@ bool vmw_fifo_have_3d(struct vmw_private if (!dev_priv->has_mob) return false; - mutex_lock(&dev_priv->hw_mutex); + spin_lock(&dev_priv->cap_lock); vmw_write(dev_priv, SVGA_REG_DEV_CAP, SVGA3D_DEVCAP_3D); result = vmw_read(dev_priv, SVGA_REG_DEV_CAP); - mutex_unlock(&dev_priv->hw_mutex); + spin_unlock(&dev_priv->cap_lock); return (result != 0); } @@ -120,7 +120,6 @@ int vmw_fifo_init(struct vmw_private *de DRM_INFO("height %d\n", vmw_read(dev_priv, SVGA_REG_HEIGHT)); DRM_INFO("bpp %d\n", vmw_read(dev_priv, SVGA_REG_BITS_PER_PIXEL)); - mutex_lock(&dev_priv->hw_mutex); dev_priv->enable_state = vmw_read(dev_priv, SVGA_REG_ENABLE); dev_priv->config_done_state = vmw_read(dev_priv, SVGA_REG_CONFIG_DONE); dev_priv->traces_state = vmw_read(dev_priv, SVGA_REG_TRACES); @@ -143,7 +142,6 @@ int vmw_fifo_init(struct vmw_private *de mb(); vmw_write(dev_priv, SVGA_REG_CONFIG_DONE, 1); - mutex_unlock(&dev_priv->hw_mutex); max = ioread32(fifo_mem + SVGA_FIFO_MAX); min = ioread32(fifo_mem + SVGA_FIFO_MIN); @@ -160,31 +158,28 @@ int vmw_fifo_init(struct vmw_private *de return vmw_fifo_send_fence(dev_priv, &dummy); } -void vmw_fifo_ping_host_locked(struct vmw_private *dev_priv, uint32_t reason) +void vmw_fifo_ping_host(struct vmw_private *dev_priv, uint32_t reason) { __le32 __iomem *fifo_mem = dev_priv->mmio_virt; + static DEFINE_SPINLOCK(ping_lock); + unsigned long irq_flags; + /* + * The ping_lock is needed because we don't have an atomic + * test-and-set of the SVGA_FIFO_BUSY register. + */ + spin_lock_irqsave(&ping_lock, irq_flags); if (unlikely(ioread32(fifo_mem + SVGA_FIFO_BUSY) == 0)) { iowrite32(1, fifo_mem + SVGA_FIFO_BUSY); vmw_write(dev_priv, SVGA_REG_SYNC, reason); } -} - -void vmw_fifo_ping_host(struct vmw_private *dev_priv, uint32_t reason) -{ - mutex_lock(&dev_priv->hw_mutex); - - vmw_fifo_ping_host_locked(dev_priv, reason); - - mutex_unlock(&dev_priv->hw_mutex); + spin_unlock_irqrestore(&ping_lock, irq_flags); } void vmw_fifo_release(struct vmw_private *dev_priv, struct vmw_fifo_state *fifo) { __le32 __iomem *fifo_mem = dev_priv->mmio_virt; - mutex_lock(&dev_priv->hw_mutex); - vmw_write(dev_priv, SVGA_REG_SYNC, SVGA_SYNC_GENERIC); while (vmw_read(dev_priv, SVGA_REG_BUSY) != 0) ; @@ -198,7 +193,6 @@ void vmw_fifo_release(struct vmw_private vmw_write(dev_priv, SVGA_REG_TRACES, dev_priv->traces_state); - mutex_unlock(&dev_priv->hw_mutex); vmw_marker_queue_takedown(&fifo->marker_queue); if (likely(fifo->static_buffer != NULL)) { @@ -271,7 +265,7 @@ static int vmw_fifo_wait(struct vmw_priv return vmw_fifo_wait_noirq(dev_priv, bytes, interruptible, timeout); - mutex_lock(&dev_priv->hw_mutex); + spin_lock(&dev_priv->waiter_lock); if (atomic_add_return(1, &dev_priv->fifo_queue_waiters) > 0) { spin_lock_irqsave(&dev_priv->irq_lock, irq_flags); outl(SVGA_IRQFLAG_FIFO_PROGRESS, @@ -280,7 +274,7 @@ static int vmw_fifo_wait(struct vmw_priv vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask); spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags); } - mutex_unlock(&dev_priv->hw_mutex); + spin_unlock(&dev_priv->waiter_lock); if (interruptible) ret = wait_event_interruptible_timeout @@ -296,14 +290,14 @@ static int vmw_fifo_wait(struct vmw_priv else if (likely(ret > 0)) ret = 0; - mutex_lock(&dev_priv->hw_mutex); + spin_lock(&dev_priv->waiter_lock); if (atomic_dec_and_test(&dev_priv->fifo_queue_waiters)) { spin_lock_irqsave(&dev_priv->irq_lock, irq_flags); dev_priv->irq_mask &= ~SVGA_IRQFLAG_FIFO_PROGRESS; vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask); spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags); } - mutex_unlock(&dev_priv->hw_mutex); + spin_unlock(&dev_priv->waiter_lock); return ret; } --- a/drivers/gpu/drm/vmwgfx/vmwgfx_ioctl.c +++ b/drivers/gpu/drm/vmwgfx/vmwgfx_ioctl.c @@ -135,13 +135,13 @@ static int vmw_fill_compat_cap(struct vm (pair_offset + max_size * sizeof(SVGA3dCapPair)) / sizeof(u32); compat_cap->header.type = SVGA3DCAPS_RECORD_DEVCAPS; - mutex_lock(&dev_priv->hw_mutex); + spin_lock(&dev_priv->cap_lock); for (i = 0; i < max_size; ++i) { vmw_write(dev_priv, SVGA_REG_DEV_CAP, i); compat_cap->pairs[i][0] = i; compat_cap->pairs[i][1] = vmw_read(dev_priv, SVGA_REG_DEV_CAP); } - mutex_unlock(&dev_priv->hw_mutex); + spin_unlock(&dev_priv->cap_lock); return 0; } @@ -191,12 +191,12 @@ int vmw_get_cap_3d_ioctl(struct drm_devi if (num > SVGA3D_DEVCAP_MAX) num = SVGA3D_DEVCAP_MAX; - mutex_lock(&dev_priv->hw_mutex); + spin_lock(&dev_priv->cap_lock); for (i = 0; i < num; ++i) { vmw_write(dev_priv, SVGA_REG_DEV_CAP, i); *bounce32++ = vmw_read(dev_priv, SVGA_REG_DEV_CAP); } - mutex_unlock(&dev_priv->hw_mutex); + spin_unlock(&dev_priv->cap_lock); } else if (gb_objects) { ret = vmw_fill_compat_cap(dev_priv, bounce, size); if (unlikely(ret != 0)) --- a/drivers/gpu/drm/vmwgfx/vmwgfx_irq.c +++ b/drivers/gpu/drm/vmwgfx/vmwgfx_irq.c @@ -62,13 +62,8 @@ irqreturn_t vmw_irq_handler(int irq, voi static bool vmw_fifo_idle(struct vmw_private *dev_priv, uint32_t seqno) { - uint32_t busy; - mutex_lock(&dev_priv->hw_mutex); - busy = vmw_read(dev_priv, SVGA_REG_BUSY); - mutex_unlock(&dev_priv->hw_mutex); - - return (busy == 0); + return (vmw_read(dev_priv, SVGA_REG_BUSY) == 0); } void vmw_update_seqno(struct vmw_private *dev_priv, @@ -184,7 +179,7 @@ int vmw_fallback_wait(struct vmw_private void vmw_seqno_waiter_add(struct vmw_private *dev_priv) { - mutex_lock(&dev_priv->hw_mutex); + spin_lock(&dev_priv->waiter_lock); if (dev_priv->fence_queue_waiters++ == 0) { unsigned long irq_flags; @@ -195,12 +190,12 @@ void vmw_seqno_waiter_add(struct vmw_pri vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask); spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags); } - mutex_unlock(&dev_priv->hw_mutex); + spin_unlock(&dev_priv->waiter_lock); } void vmw_seqno_waiter_remove(struct vmw_private *dev_priv) { - mutex_lock(&dev_priv->hw_mutex); + spin_lock(&dev_priv->waiter_lock); if (--dev_priv->fence_queue_waiters == 0) { unsigned long irq_flags; @@ -209,13 +204,13 @@ void vmw_seqno_waiter_remove(struct vmw_ vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask); spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags); } - mutex_unlock(&dev_priv->hw_mutex); + spin_unlock(&dev_priv->waiter_lock); } void vmw_goal_waiter_add(struct vmw_private *dev_priv) { - mutex_lock(&dev_priv->hw_mutex); + spin_lock(&dev_priv->waiter_lock); if (dev_priv->goal_queue_waiters++ == 0) { unsigned long irq_flags; @@ -226,12 +221,12 @@ void vmw_goal_waiter_add(struct vmw_priv vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask); spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags); } - mutex_unlock(&dev_priv->hw_mutex); + spin_unlock(&dev_priv->waiter_lock); } void vmw_goal_waiter_remove(struct vmw_private *dev_priv) { - mutex_lock(&dev_priv->hw_mutex); + spin_lock(&dev_priv->waiter_lock); if (--dev_priv->goal_queue_waiters == 0) { unsigned long irq_flags; @@ -240,7 +235,7 @@ void vmw_goal_waiter_remove(struct vmw_p vmw_write(dev_priv, SVGA_REG_IRQMASK, dev_priv->irq_mask); spin_unlock_irqrestore(&dev_priv->irq_lock, irq_flags); } - mutex_unlock(&dev_priv->hw_mutex); + spin_unlock(&dev_priv->waiter_lock); } int vmw_wait_seqno(struct vmw_private *dev_priv, @@ -315,9 +310,7 @@ void vmw_irq_uninstall(struct drm_device if (!(dev_priv->capabilities & SVGA_CAP_IRQMASK)) return; - mutex_lock(&dev_priv->hw_mutex); vmw_write(dev_priv, SVGA_REG_IRQMASK, 0); - mutex_unlock(&dev_priv->hw_mutex); status = inl(dev_priv->io_start + VMWGFX_IRQSTATUS_PORT); outl(status, dev_priv->io_start + VMWGFX_IRQSTATUS_PORT); --- a/drivers/gpu/drm/vmwgfx/vmwgfx_kms.c +++ b/drivers/gpu/drm/vmwgfx/vmwgfx_kms.c @@ -1828,9 +1828,7 @@ vmw_du_connector_detect(struct drm_conne struct vmw_private *dev_priv = vmw_priv(dev); struct vmw_display_unit *du = vmw_connector_to_du(connector); - mutex_lock(&dev_priv->hw_mutex); num_displays = vmw_read(dev_priv, SVGA_REG_NUM_DISPLAYS); - mutex_unlock(&dev_priv->hw_mutex); return ((vmw_connector_to_du(connector)->unit < num_displays && du->pref_active) ? Patches currently in stable-queue which might be from thellstrom@xxxxxxxxxx are queue-3.18/drm-vmwgfx-replace-the-hw-mutex-with-a-hw-spinlock.patch -- To unsubscribe from this list: send the line "unsubscribe stable" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html