diff options
Diffstat (limited to 'drivers/gpu/drm/i915/i915_gem.c')
| -rw-r--r-- | drivers/gpu/drm/i915/i915_gem.c | 752 |
1 files changed, 106 insertions, 646 deletions
diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c index 7399ac7a5629..b359390ba22c 100644 --- a/drivers/gpu/drm/i915/i915_gem.c +++ b/drivers/gpu/drm/i915/i915_gem.c @@ -25,18 +25,8 @@ * */ -#include <drm/drmP.h> #include <drm/drm_vma_manager.h> #include <drm/i915_drm.h> -#include "i915_drv.h" -#include "i915_gem_clflush.h" -#include "i915_vgpu.h" -#include "i915_trace.h" -#include "intel_drv.h" -#include "intel_frontbuffer.h" -#include "intel_mocs.h" -#include "intel_workarounds.h" -#include "i915_gemfs.h" #include <linux/dma-fence-array.h> #include <linux/kthread.h> #include <linux/reservation.h> @@ -47,6 +37,18 @@ #include <linux/pci.h> #include <linux/dma-buf.h> +#include "i915_drv.h" +#include "i915_gem_clflush.h" +#include "i915_gemfs.h" +#include "i915_reset.h" +#include "i915_trace.h" +#include "i915_vgpu.h" + +#include "intel_drv.h" +#include "intel_frontbuffer.h" +#include "intel_mocs.h" +#include "intel_workarounds.h" + static void i915_gem_flush_free_objects(struct drm_i915_private *i915); static bool cpu_write_needs_clflush(struct drm_i915_gem_object *obj) @@ -139,6 +141,8 @@ int i915_mutex_lock_interruptible(struct drm_device *dev) static u32 __i915_gem_park(struct drm_i915_private *i915) { + intel_wakeref_t wakeref; + GEM_TRACE("\n"); lockdep_assert_held(&i915->drm.struct_mutex); @@ -169,14 +173,13 @@ static u32 __i915_gem_park(struct drm_i915_private *i915) i915_pmu_gt_parked(i915); i915_vma_parked(i915); - i915->gt.awake = false; + wakeref = fetch_and_zero(&i915->gt.awake); + GEM_BUG_ON(!wakeref); if (INTEL_GEN(i915) >= 6) gen6_rps_idle(i915); - intel_display_power_put(i915, POWER_DOMAIN_GT_IRQ); - - intel_runtime_pm_put(i915); + intel_display_power_put(i915, POWER_DOMAIN_GT_IRQ, wakeref); return i915->gt.epoch; } @@ -201,12 +204,11 @@ void i915_gem_unpark(struct drm_i915_private *i915) lockdep_assert_held(&i915->drm.struct_mutex); GEM_BUG_ON(!i915->gt.active_requests); + assert_rpm_wakelock_held(i915); if (i915->gt.awake) return; - intel_runtime_pm_get_noresume(i915); - /* * It seems that the DMC likes to transition between the DC states a lot * when there are no connected displays (no active power domains) during @@ -218,9 +220,9 @@ void i915_gem_unpark(struct drm_i915_private *i915) * Work around it by grabbing a GT IRQ power domain whilst there is any * GT activity, preventing any DC state transitions. */ - intel_display_power_get(i915, POWER_DOMAIN_GT_IRQ); + i915->gt.awake = intel_display_power_get(i915, POWER_DOMAIN_GT_IRQ); + GEM_BUG_ON(!i915->gt.awake); - i915->gt.awake = true; if (unlikely(++i915->gt.epoch == 0)) /* keep 0 as invalid */ i915->gt.epoch = 1; @@ -711,8 +713,8 @@ void i915_gem_object_free(struct drm_i915_gem_object *obj) static int i915_gem_create(struct drm_file *file, struct drm_i915_private *dev_priv, - uint64_t size, - uint32_t *handle_p) + u64 size, + u32 *handle_p) { struct drm_i915_gem_object *obj; int ret; @@ -783,6 +785,8 @@ fb_write_origin(struct drm_i915_gem_object *obj, unsigned int domain) void i915_gem_flush_ggtt_writes(struct drm_i915_private *dev_priv) { + intel_wakeref_t wakeref; + /* * No actual flushing is required for the GTT write domain for reads * from the GTT domain. Writes to it "immediately" go to main memory @@ -809,13 +813,13 @@ void i915_gem_flush_ggtt_writes(struct drm_i915_private *dev_priv) i915_gem_chipset_flush(dev_priv); - intel_runtime_pm_get(dev_priv); - spin_lock_irq(&dev_priv->uncore.lock); + with_intel_runtime_pm(dev_priv, wakeref) { + spin_lock_irq(&dev_priv->uncore.lock); - POSTING_READ_FW(RING_HEAD(RENDER_RING_BASE)); + POSTING_READ_FW(RING_HEAD(RENDER_RING_BASE)); - spin_unlock_irq(&dev_priv->uncore.lock); - intel_runtime_pm_put(dev_priv); + spin_unlock_irq(&dev_priv->uncore.lock); + } } static void @@ -859,58 +863,6 @@ flush_write_domain(struct drm_i915_gem_object *obj, unsigned int flush_domains) obj->write_domain = 0; } -static inline int -__copy_to_user_swizzled(char __user *cpu_vaddr, - const char *gpu_vaddr, int gpu_offset, - int length) -{ - int ret, cpu_offset = 0; - - while (length > 0) { - int cacheline_end = ALIGN(gpu_offset + 1, 64); - int this_length = min(cacheline_end - gpu_offset, length); - int swizzled_gpu_offset = gpu_offset ^ 64; - - ret = __copy_to_user(cpu_vaddr + cpu_offset, - gpu_vaddr + swizzled_gpu_offset, - this_length); - if (ret) - return ret + length; - - cpu_offset += this_length; - gpu_offset += this_length; - length -= this_length; - } - - return 0; -} - -static inline int -__copy_from_user_swizzled(char *gpu_vaddr, int gpu_offset, - const char __user *cpu_vaddr, - int length) -{ - int ret, cpu_offset = 0; - - while (length > 0) { - int cacheline_end = ALIGN(gpu_offset + 1, 64); - int this_length = min(cacheline_end - gpu_offset, length); - int swizzled_gpu_offset = gpu_offset ^ 64; - - ret = __copy_from_user(gpu_vaddr + swizzled_gpu_offset, - cpu_vaddr + cpu_offset, - this_length); - if (ret) - return ret + length; - - cpu_offset += this_length; - gpu_offset += this_length; - length -= this_length; - } - - return 0; -} - /* * Pins the specified object's pages and synchronizes the object with * GPU accesses. Sets needs_clflush to non-zero if the caller should @@ -1030,72 +982,23 @@ err_unpin: return ret; } -static void -shmem_clflush_swizzled_range(char *addr, unsigned long length, - bool swizzled) -{ - if (unlikely(swizzled)) { - unsigned long start = (unsigned long) addr; - unsigned long end = (unsigned long) addr + length; - - /* For swizzling simply ensure that we always flush both - * channels. Lame, but simple and it works. Swizzled - * pwrite/pread is far from a hotpath - current userspace - * doesn't use it at all. */ - start = round_down(start, 128); - end = round_up(end, 128); - - drm_clflush_virt_range((void *)start, end - start); - } else { - drm_clflush_virt_range(addr, length); - } - -} - -/* Only difference to the fast-path function is that this can handle bit17 - * and uses non-atomic copy and kmap functions. */ static int -shmem_pread_slow(struct page *page, int offset, int length, - char __user *user_data, - bool page_do_bit17_swizzling, bool needs_clflush) +shmem_pread(struct page *page, int offset, int len, char __user *user_data, + bool needs_clflush) { char *vaddr; int ret; vaddr = kmap(page); - if (needs_clflush) - shmem_clflush_swizzled_range(vaddr + offset, length, - page_do_bit17_swizzling); - if (page_do_bit17_swizzling) - ret = __copy_to_user_swizzled(user_data, vaddr, offset, length); - else - ret = __copy_to_user(user_data, vaddr + offset, length); - kunmap(page); - - return ret ? - EFAULT : 0; -} - -static int -shmem_pread(struct page *page, int offset, int length, char __user *user_data, - bool page_do_bit17_swizzling, bool needs_clflush) -{ - int ret; + if (needs_clflush) + drm_clflush_virt_range(vaddr + offset, len); - ret = -ENODEV; - if (!page_do_bit17_swizzling) { - char *vaddr = kmap_atomic(page); + ret = __copy_to_user(user_data, vaddr + offset, len); - if (needs_clflush) - drm_clflush_virt_range(vaddr + offset, length); - ret = __copy_to_user_inatomic(user_data, vaddr + offset, length); - kunmap_atomic(vaddr); - } - if (ret == 0) - return 0; + kunmap(page); - return shmem_pread_slow(page, offset, length, user_data, - page_do_bit17_swizzling, needs_clflush); + return ret ? -EFAULT : 0; } static int @@ -1104,15 +1007,10 @@ i915_gem_shmem_pread(struct drm_i915_gem_object *obj, { char __user *user_data; u64 remain; - unsigned int obj_do_bit17_swizzling; unsigned int needs_clflush; unsigned int idx, offset; int ret; - obj_do_bit17_swizzling = 0; - if (i915_gem_object_needs_bit17_swizzle(obj)) - obj_do_bit17_swizzling = BIT(17); - ret = mutex_lock_interruptible(&obj->base.dev->struct_mutex); if (ret) return ret; @@ -1130,7 +1028,6 @@ i915_gem_shmem_pread(struct drm_i915_gem_object *obj, unsigned int length = min_t(u64, remain, PAGE_SIZE - offset); ret = shmem_pread(page, offset, length, user_data, - page_to_phys(page) & obj_do_bit17_swizzling, needs_clflush); if (ret) break; @@ -1174,6 +1071,7 @@ i915_gem_gtt_pread(struct drm_i915_gem_object *obj, { struct drm_i915_private *i915 = to_i915(obj->base.dev); struct i915_ggtt *ggtt = &i915->ggtt; + intel_wakeref_t wakeref; struct drm_mm_node node; struct i915_vma *vma; void __user *user_data; @@ -1184,7 +1082,7 @@ i915_gem_gtt_pread(struct drm_i915_gem_object *obj, if (ret) return ret; - intel_runtime_pm_get(i915); + wakeref = intel_runtime_pm_get(i915); vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0, PIN_MAPPABLE | PIN_NONFAULT | @@ -1257,7 +1155,7 @@ out_unpin: i915_vma_unpin(vma); } out_unlock: - intel_runtime_pm_put(i915); + intel_runtime_pm_put(i915, wakeref); mutex_unlock(&i915->drm.struct_mutex); return ret; @@ -1358,6 +1256,7 @@ i915_gem_gtt_pwrite_fast(struct drm_i915_gem_object *obj, { struct drm_i915_private *i915 = to_i915(obj->base.dev); struct i915_ggtt *ggtt = &i915->ggtt; + intel_wakeref_t wakeref; struct drm_mm_node node; struct i915_vma *vma; u64 remain, offset; @@ -1376,13 +1275,14 @@ i915_gem_gtt_pwrite_fast(struct drm_i915_gem_object *obj, * This easily dwarfs any performance advantage from * using the cache bypass of indirect GGTT access. */ - if (!intel_runtime_pm_get_if_in_use(i915)) { + wakeref = intel_runtime_pm_get_if_in_use(i915); + if (!wakeref) { ret = -EFAULT; goto out_unlock; } } else { /* No backing pages, no fallback, we must force GGTT access */ - intel_runtime_pm_get(i915); + wakeref = intel_runtime_pm_get(i915); } vma = i915_gem_object_ggtt_pin(obj, NULL, 0, 0, @@ -1464,39 +1364,12 @@ out_unpin: i915_vma_unpin(vma); } out_rpm: - intel_runtime_pm_put(i915); + intel_runtime_pm_put(i915, wakeref); out_unlock: mutex_unlock(&i915->drm.struct_mutex); return ret; } -static int -shmem_pwrite_slow(struct page *page, int offset, int length, - char __user *user_data, - bool page_do_bit17_swizzling, - bool needs_clflush_before, - bool needs_clflush_after) -{ - char *vaddr; - int ret; - - vaddr = kmap(page); - if (unlikely(needs_clflush_before || page_do_bit17_swizzling)) - shmem_clflush_swizzled_range(vaddr + offset, length, - page_do_bit17_swizzling); - if (page_do_bit17_swizzling) - ret = __copy_from_user_swizzled(vaddr, offset, user_data, - length); - else - ret = __copy_from_user(vaddr + offset, user_data, length); - if (needs_clflush_after) - shmem_clflush_swizzled_range(vaddr + offset, length, - page_do_bit17_swizzling); - kunmap(page); - - return ret ? -EFAULT : 0; -} - /* Per-page copy function for the shmem pwrite fastpath. * Flushes invalid cachelines before writing to the target if * needs_clflush_before is set and flushes out any written cachelines after @@ -1504,31 +1377,24 @@ shmem_pwrite_slow(struct page *page, int offset, int length, */ static int shmem_pwrite(struct page *page, int offset, int len, char __user *user_data, - bool page_do_bit17_swizzling, bool needs_clflush_before, bool needs_clflush_after) { + char *vaddr; int ret; - ret = -ENODEV; - if (!page_do_bit17_swizzling) { - char *vaddr = kmap_atomic(page); + vaddr = kmap(page); - if (needs_clflush_before) - drm_clflush_virt_range(vaddr + offset, len); - ret = __copy_from_user_inatomic(vaddr + offset, user_data, len); - if (needs_clflush_after) - drm_clflush_virt_range(vaddr + offset, len); + if (needs_clflush_before) + drm_clflush_virt_range(vaddr + offset, len); - kunmap_atomic(vaddr); - } - if (ret == 0) - return ret; + ret = __copy_from_user(vaddr + offset, user_data, len); + if (!ret && needs_clflush_after) + drm_clflush_virt_range(vaddr + offset, len); + + kunmap(page); - return shmem_pwrite_slow(page, offset, len, user_data, - page_do_bit17_swizzling, - needs_clflush_before, - needs_clflush_after); + return ret ? -EFAULT : 0; } static int @@ -1538,7 +1404,6 @@ i915_gem_shmem_pwrite(struct drm_i915_gem_object *obj, struct drm_i915_private *i915 = to_i915(obj->base.dev); void __user *user_data; u64 remain; - unsigned int obj_do_bit17_swizzling; unsigned int partial_cacheline_write; unsigned int needs_clflush; unsigned int offset, idx; @@ -1553,10 +1418,6 @@ i915_gem_shmem_pwrite(struct drm_i915_gem_object *obj, if (ret) return ret; - obj_do_bit17_swizzling = 0; - if (i915_gem_object_needs_bit17_swizzle(obj)) - obj_do_bit17_swizzling = BIT(17); - /* If we don't overwrite a cacheline completely we need to be * careful to have up-to-date data by first clflushing. Don't * overcomplicate things and flush the entire patch. @@ -1573,7 +1434,6 @@ i915_gem_shmem_pwrite(struct drm_i915_gem_object *obj, unsigned int length = min_t(u64, remain, PAGE_SIZE - offset); ret = shmem_pwrite(page, offset, length, user_data, - page_to_phys(page) & obj_do_bit17_swizzling, (offset | length) & partial_cacheline_write, needs_clflush & CLFLUSH_AFTER); if (ret) @@ -1713,8 +1573,8 @@ i915_gem_set_domain_ioctl(struct drm_device *dev, void *data, { struct drm_i915_gem_set_domain *args = data; struct drm_i915_gem_object *obj; - uint32_t read_domains = args->read_domains; - uint32_t write_domain = args->write_domain; + u32 read_domains = args->read_domains; + u32 write_domain = args->write_domain; int err; /* Only handle setting domains to types used by the CPU. */ @@ -1896,7 +1756,7 @@ i915_gem_mmap_ioctl(struct drm_device *dev, void *data, if (IS_ERR((void *)addr)) return addr; - args->addr_ptr = (uint64_t) addr; + args->addr_ptr = (u64)addr; return 0; } @@ -2009,6 +1869,7 @@ vm_fault_t i915_gem_fault(struct vm_fault *vmf) struct drm_i915_private *dev_priv = to_i915(dev); struct i915_ggtt *ggtt = &dev_priv->ggtt; bool write = area->vm_flags & VM_WRITE; + intel_wakeref_t wakeref; struct i915_vma *vma; pgoff_t page_offset; int ret; @@ -2038,7 +1899,7 @@ vm_fault_t i915_gem_fault(struct vm_fault *vmf) if (ret) goto err; - intel_runtime_pm_get(dev_priv); + wakeref = intel_runtime_pm_get(dev_priv); ret = i915_mutex_lock_interruptible(dev); if (ret) @@ -2116,7 +1977,7 @@ err_unpin: err_unlock: mutex_unlock(&dev->struct_mutex); err_rpm: - intel_runtime_pm_put(dev_priv); + intel_runtime_pm_put(dev_priv, wakeref); i915_gem_object_unpin_pages(obj); err: switch (ret) { @@ -2189,6 +2050,7 @@ void i915_gem_release_mmap(struct drm_i915_gem_object *obj) { struct drm_i915_private *i915 = to_i915(obj->base.dev); + intel_wakeref_t wakeref; /* Serialisation between user GTT access and our code depends upon * revoking the CPU's PTE whilst the mutex is held. The next user @@ -2199,7 +2061,7 @@ i915_gem_release_mmap(struct drm_i915_gem_object *obj) * wakeref. */ lockdep_assert_held(&i915->drm.struct_mutex); - intel_runtime_pm_get(i915); + wakeref = intel_runtime_pm_get(i915); if (!obj->userfault_count) goto out; @@ -2216,7 +2078,7 @@ i915_gem_release_mmap(struct drm_i915_gem_object *obj) wmb(); out: - intel_runtime_pm_put(i915); + intel_runtime_pm_put(i915, wakeref); } void i915_gem_runtime_suspend(struct drm_i915_private *dev_priv) @@ -2296,8 +2158,8 @@ static void i915_gem_object_free_mmap_offset(struct drm_i915_gem_object *obj) int i915_gem_mmap_gtt(struct drm_file *file, struct drm_device *dev, - uint32_t handle, - uint64_t *offset) + u32 handle, + u64 *offset) { struct drm_i915_gem_object *obj; int ret; @@ -2444,8 +2306,8 @@ __i915_gem_object_unset_pages(struct drm_i915_gem_object *obj) struct sg_table *pages; pages = fetch_and_zero(&obj->mm.pages); - if (!pages) - return NULL; + if (IS_ERR_OR_NULL(pages)) + return pages; spin_lock(&i915->mm.obj_lock); list_del(&obj->mm.link); @@ -2469,22 +2331,23 @@ __i915_gem_object_unset_pages(struct drm_i915_gem_object *obj) return pages; } -void __i915_gem_object_put_pages(struct drm_i915_gem_object *obj, - enum i915_mm_subclass subclass) +int __i915_gem_object_put_pages(struct drm_i915_gem_object *obj, + enum i915_mm_subclass subclass) { struct sg_table *pages; + int ret; if (i915_gem_object_has_pinned_pages(obj)) - return; + return -EBUSY; GEM_BUG_ON(obj->bind_count); - if (!i915_gem_object_has_pages(obj)) - return; /* May be called by shrinker from within get_pages() (on another bo) */ mutex_lock_nested(&obj->mm.lock, subclass); - if (unlikely(atomic_read(&obj->mm.pages_pin_count))) + if (unlikely(atomic_read(&obj->mm.pages_pin_count))) { + ret = -EBUSY; goto unlock; + } /* * ->put_pages might need to allocate memory for the bit17 swizzle @@ -2492,11 +2355,24 @@ void __i915_gem_object_put_pages(struct drm_i915_gem_object *obj, * lists early. */ pages = __i915_gem_object_unset_pages(obj); + + /* + * XXX Temporary hijinx to avoid updating all backends to handle + * NULL pages. In the future, when we have more asynchronous + * get_pages backends we should be better able to handle the + * cancellation of the async task in a more uniform manner. + */ + if (!pages && !i915_gem_object_needs_async_cancel(obj)) + pages = ERR_PTR(-EINVAL); + if (!IS_ERR(pages)) obj->ops->put_pages(obj, pages); + ret = 0; unlock: mutex_unlock(&obj->mm.lock); + + return ret; } bool i915_sg_trim(struct sg_table *orig_st) @@ -3000,61 +2876,6 @@ i915_gem_object_pwrite_gtt(struct drm_i915_gem_object *obj, return 0; } -static void i915_gem_client_mark_guilty(struct drm_i915_file_private *file_priv, - const struct i915_gem_context *ctx) -{ - unsigned int score; - unsigned long prev_hang; - - if (i915_gem_context_is_banned(ctx)) - score = I915_CLIENT_SCORE_CONTEXT_BAN; - else - score = 0; - - prev_hang = xchg(&file_priv->hang_timestamp, jiffies); - if (time_before(jiffies, prev_hang + I915_CLIENT_FAST_HANG_JIFFIES)) - score += I915_CLIENT_SCORE_HANG_FAST; - - if (score) { - atomic_add(score, &file_priv->ban_score); - - DRM_DEBUG_DRIVER("client %s: gained %u ban score, now %u\n", - ctx->name, score, - atomic_read(&file_priv->ban_score)); - } -} - -static void i915_gem_context_mark_guilty(struct i915_gem_context *ctx) -{ - unsigned int score; - bool banned, bannable; - - atomic_inc(&ctx->guilty_count); - - bannable = i915_gem_context_is_bannable(ctx); - score = atomic_add_return(CONTEXT_SCORE_GUILTY, &ctx->ban_score); - banned = score >= CONTEXT_SCORE_BAN_THRESHOLD; - - /* Cool contexts don't accumulate client ban score */ - if (!bannable) - return; - - if (banned) { - DRM_DEBUG_DRIVER("context %s: guilty %d, score %u, banned\n", - ctx->name, atomic_read(&ctx->guilty_count), - score); - i915_gem_context_set_banned(ctx); - } - - if (!IS_ERR_OR_NULL(ctx->file_priv)) - i915_gem_client_mark_guilty(ctx->file_priv, ctx); -} - -static void i915_gem_context_mark_innocent(struct i915_gem_context *ctx) -{ - atomic_inc(&ctx->active_count); -} - struct i915_request * i915_gem_find_active_request(struct intel_engine_cs *engine) { @@ -3085,366 +2906,6 @@ i915_gem_find_active_request(struct intel_engine_cs *engine) return active; } -/* - * Ensure irq handler finishes, and not run again. - * Also return the active request so that we only search for it once. - */ -struct i915_request * -i915_gem_reset_prepare_engine(struct intel_engine_cs *engine) -{ - struct i915_request *request; - - /* - * During the reset sequence, we must prevent the engine from - * entering RC6. As the context state is undefined until we restart - * the engine, if it does enter RC6 during the reset, the state - * written to the powercontext is undefined and so we may lose - * GPU state upon resume, i.e. fail to restart after a reset. - */ - intel_uncore_forcewake_get(engine->i915, FORCEWAKE_ALL); - - request = engine->reset.prepare(engine); - if (request && request->fence.error == -EIO) - request = ERR_PTR(-EIO); /* Previous reset failed! */ - - return request; -} - -int i915_gem_reset_prepare(struct drm_i915_private *dev_priv) -{ - struct intel_engine_cs *engine; - struct i915_request *request; - enum intel_engine_id id; - int err = 0; - - for_each_engine(engine, dev_priv, id) { - request = i915_gem_reset_prepare_engine(engine); - if (IS_ERR(request)) { - err = PTR_ERR(request); - continue; - } - - engine->hangcheck.active_request = request; - } - - i915_gem_revoke_fences(dev_priv); - intel_uc_sanitize(dev_priv); - - return err; -} - -static void engine_skip_context(struct i915_request *request) -{ - struct intel_engine_cs *engine = request->engine; - struct i915_gem_context *hung_ctx = request->gem_context; - struct i915_timeline *timeline = request->timeline; - unsigned long flags; - - GEM_BUG_ON(timeline == &engine->timeline); - - spin_lock_irqsave(&engine->timeline.lock, flags); - spin_lock(&timeline->lock); - - list_for_each_entry_continue(request, &engine->timeline.requests, link) - if (request->gem_context == hung_ctx) - i915_request_skip(request, -EIO); - - list_for_each_entry(request, &timeline->requests, link) - i915_request_skip(request, -EIO); - - spin_unlock(&timeline->lock); - spin_unlock_irqrestore(&engine->timeline.lock, flags); -} - -/* Returns the request if it was guilty of the hang */ -static struct i915_request * -i915_gem_reset_request(struct intel_engine_cs *engine, - struct i915_request *request, - bool stalled) -{ - /* The guilty request will get skipped on a hung engine. - * - * Users of client default contexts do not rely on logical - * state preserved between batches so it is safe to execute - * queued requests following the hang. Non default contexts - * rely on preserved state, so skipping a batch loses the - * evolution of the state and it needs to be considered corrupted. - * Executing more queued batches on top of corrupted state is - * risky. But we take the risk by trying to advance through - * the queued requests in order to make the client behaviour - * more predictable around resets, by not throwing away random - * amount of batches it has prepared for execution. Sophisticated - * clients can use gem_reset_stats_ioctl and dma fence status - * (exported via sync_file info ioctl on explicit fences) to observe - * when it loses the context state and should rebuild accordingly. - * - * The context ban, and ultimately the client ban, mechanism are safety - * valves if client submission ends up resulting in nothing more than - * subsequent hangs. - */ - - if (i915_request_completed(request)) { - GEM_TRACE("%s pardoned global=%d (fence %llx:%lld), current %d\n", - engine->name, request->global_seqno, - request->fence.context, request->fence.seqno, - intel_engine_get_seqno(engine)); - stalled = false; - } - - if (stalled) { - i915_gem_context_mark_guilty(request->gem_context); - i915_request_skip(request, -EIO); - - /* If this context is now banned, skip all pending requests. */ - if (i915_gem_context_is_banned(request->gem_context)) - engine_skip_context(request); - } else { - /* - * Since this is not the hung engine, it may have advanced - * since the hang declaration. Double check by refinding - * the active request at the time of the reset. - */ - request = i915_gem_find_active_request(engine); - if (request) { - unsigned long flags; - - i915_gem_context_mark_innocent(request->gem_context); - dma_fence_set_error(&request->fence, -EAGAIN); - - /* Rewind the engine to replay the incomplete rq */ - spin_lock_irqsave(&engine->timeline.lock, flags); - request = list_prev_entry(request, link); - if (&request->link == &engine->timeline.requests) - request = NULL; - spin_unlock_irqrestore(&engine->timeline.lock, flags); - } - } - - return request; -} - -void i915_gem_reset_engine(struct intel_engine_cs *engine, - struct i915_request *request, - bool stalled) -{ - /* - * Make sure this write is visible before we re-enable the interrupt - * handlers on another CPU, as tasklet_enable() resolves to just - * a compiler barrier which is insufficient for our purpose here. - */ - smp_store_mb(engine->irq_posted, 0); - - if (request) - request = i915_gem_reset_request(engine, request, stalled); - - /* Setup the CS to resume from the breadcrumb of the hung request */ - engine->reset.reset(engine, request); -} - -void i915_gem_reset(struct drm_i915_private *dev_priv, - unsigned int stalled_mask) -{ - struct intel_engine_cs *engine; - enum intel_engine_id id; - - lockdep_assert_held(&dev_priv->drm.struct_mutex); - - i915_retire_requests(dev_priv); - - for_each_engine(engine, dev_priv, id) { - struct intel_context *ce; - - i915_gem_reset_engine(engine, - engine->hangcheck.active_request, - stalled_mask & ENGINE_MASK(id)); - ce = fetch_and_zero(&engine->last_retired_context); - if (ce) - intel_context_unpin(ce); - - /* - * Ostensibily, we always want a context loaded for powersaving, - * so if the engine is idle after the reset, send a request - * to load our scratch kernel_context. - * - * More mysteriously, if we leave the engine idle after a reset, - * the next userspace batch may hang, with what appears to be - * an incoherent read by the CS (presumably stale TLB). An - * empty request appears sufficient to paper over the glitch. - */ - if (intel_engine_is_idle(engine)) { - struct i915_request *rq; - - rq = i915_request_alloc(engine, - dev_priv->kernel_context); - if (!IS_ERR(rq)) - i915_request_add(rq); - } - } - - i915_gem_restore_fences(dev_priv); -} - -void i915_gem_reset_finish_engine(struct intel_engine_cs *engine) -{ - engine->reset.finish(engine); - - intel_uncore_forcewake_put(engine->i915, FORCEWAKE_ALL); -} - -void i915_gem_reset_finish(struct drm_i915_private *dev_priv) -{ - struct intel_engine_cs *engine; - enum intel_engine_id id; - - lockdep_assert_held(&dev_priv->drm.struct_mutex); - - for_each_engine(engine, dev_priv, id) { - engine->hangcheck.active_request = NULL; - i915_gem_reset_finish_engine(engine); - } -} - -static void nop_submit_request(struct i915_request *request) -{ - unsigned long flags; - - GEM_TRACE("%s fence %llx:%lld -> -EIO\n", - request->engine->name, - request->fence.context, request->fence.seqno); - dma_fence_set_error(&request->fence, -EIO); - - spin_lock_irqsave(&request->engine->timeline.lock, flags); - __i915_request_submit(request); - intel_engine_init_global_seqno(request->engine, request->global_seqno); - spin_unlock_irqrestore(&request->engine->timeline.lock, flags); -} - -void i915_gem_set_wedged(struct drm_i915_private *i915) -{ - struct intel_engine_cs *engine; - enum intel_engine_id id; - - GEM_TRACE("start\n"); - - if (GEM_SHOW_DEBUG()) { - struct drm_printer p = drm_debug_printer(__func__); - - for_each_engine(engine, i915, id) - intel_engine_dump(engine, &p, "%s\n", engine->name); - } - - if (test_and_set_bit(I915_WEDGED, &i915->gpu_error.flags)) - goto out; - - /* - * First, stop submission to hw, but do not yet complete requests by - * rolling the global seqno forward (since this would complete requests - * for which we haven't set the fence error to EIO yet). - */ - for_each_engine(engine, i915, id) - i915_gem_reset_prepare_engine(engine); - - /* Even if the GPU reset fails, it should still stop the engines */ - if (INTEL_GEN(i915) >= 5) - intel_gpu_reset(i915, ALL_ENGINES); - - for_each_engine(engine, i915, id) { - engine->submit_request = nop_submit_request; - engine->schedule = NULL; - } - i915->caps.scheduler = 0; - - /* - * Make sure no request can slip through without getting completed by - * either this call here to intel_engine_init_global_seqno, or the one - * in nop_submit_request. - */ - synchronize_rcu(); - - /* Mark all executing requests as skipped */ - for_each_engine(engine, i915, id) - engine->cancel_requests(engine); - - for_each_engine(engine, i915, id) { - i915_gem_reset_finish_engine(engine); - intel_engine_wakeup(engine); - } - -out: - GEM_TRACE("end\n"); - - wake_up_all(&i915->gpu_error.reset_queue); -} - -bool i915_gem_unset_wedged(struct drm_i915_private *i915) -{ - struct i915_timeline *tl; - - lockdep_assert_held(&i915->drm.struct_mutex); - if (!test_bit(I915_WEDGED, &i915->gpu_error.flags)) - return true; - - GEM_TRACE("start\n"); - - /* - * Before unwedging, make sure that all pending operations - * are flushed and errored out - we may have requests waiting upon - * third party fences. We marked all inflight requests as EIO, and - * every execbuf since returned EIO, for consistency we want all - * the currently pending requests to also be marked as EIO, which - * is done inside our nop_submit_request - and so we must wait. - * - * No more can be submitted until we reset the wedged bit. - */ - list_for_each_entry(tl, &i915->gt.timelines, link) { - struct i915_request *rq; - - rq = i915_gem_active_peek(&tl->last_request, - &i915->drm.struct_mutex); - if (!rq) - continue; - - /* - * We can't use our normal waiter as we want to - * avoid recursively trying to handle the current - * reset. The basic dma_fence_default_wait() installs - * a callback for dma_fence_signal(), which is - * triggered by our nop handler (indirectly, the - * callback enables the signaler thread which is - * woken by the nop_submit_request() advancing the seqno - * and when the seqno passes the fence, the signaler - * then signals the fence waking us up). - */ - if (dma_fence_default_wait(&rq->fence, true, - MAX_SCHEDULE_TIMEOUT) < 0) - return false; - } - i915_retire_requests(i915); - GEM_BUG_ON(i915->gt.active_requests); - - if (!intel_gpu_reset(i915, ALL_ENGINES)) - intel_engines_sanitize(i915); - - /* - * Undo nop_submit_request. We prevent all new i915 requests from - * being queued (by disallowing execbuf whilst wedged) so having - * waited for all active requests above, we know the system is idle - * and do not have to worry about a thread being inside - * engine->submit_request() as we swap over. So unlike installing - * the nop_submit_request on reset, we can do this from normal - * context and do not require stop_machine(). - */ - intel_engines_reset_default_submission(i915); - i915_gem_contexts_lost(i915); - - GEM_TRACE("end\n"); - - smp_mb__before_atomic(); /* complete takeover before enabling execbuf */ - clear_bit(I915_WEDGED, &i915->gpu_error.flags); - - return true; -} - static void i915_gem_retire_work_handler(struct work_struct *work) { @@ -4856,8 +4317,9 @@ static void __i915_gem_free_objects(struct drm_i915_private *i915, struct llist_node *freed) { struct drm_i915_gem_object *obj, *on; + intel_wakeref_t wakeref; - intel_runtime_pm_get(i915); + wakeref = intel_runtime_pm_get(i915); llist_for_each_entry_safe(obj, on, freed, freed) { struct i915_vma *vma, *vn; @@ -4918,7 +4380,7 @@ static void __i915_gem_free_objects(struct drm_i915_private *i915, if (on) cond_resched(); } - intel_runtime_pm_put(i915); + intel_runtime_pm_put(i915, wakeref); } static void i915_gem_flush_free_objects(struct drm_i915_private *i915) @@ -5027,13 +4489,13 @@ void __i915_gem_object_release_unless_active(struct drm_i915_gem_object *obj) void i915_gem_sanitize(struct drm_i915_private *i915) { - int err; + intel_wakeref_t wakeref; GEM_TRACE("\n"); mutex_lock(&i915->drm.struct_mutex); - intel_runtime_pm_get(i915); + wakeref = intel_runtime_pm_get(i915); intel_uncore_forcewake_get(i915, FORCEWAKE_ALL); /* @@ -5053,14 +4515,10 @@ void i915_gem_sanitize(struct drm_i915_private *i915) * it may impact the display and we are uncertain about the stability * of the reset, so this could be applied to even earlier gen. */ - err = -ENODEV; - if (INTEL_GEN(i915) >= 5 && intel_has_gpu_reset(i915)) - err = WARN_ON(intel_gpu_reset(i915, ALL_ENGINES)); - if (!err) - intel_engines_sanitize(i915); + intel_engines_sanitize(i915, false); intel_uncore_forcewake_put(i915, FORCEWAKE_ALL); - intel_runtime_pm_put(i915); + intel_runtime_pm_put(i915, wakeref); i915_gem_contexts_lost(i915); mutex_unlock(&i915->drm.struct_mutex); @@ -5068,11 +4526,12 @@ void i915_gem_sanitize(struct drm_i915_private *i915) int i915_gem_suspend(struct drm_i915_private *i915) { + intel_wakeref_t wakeref; int ret; GEM_TRACE("\n"); - intel_runtime_pm_get(i915); + wakeref = intel_runtime_pm_get(i915); intel_suspend_gt_powersave(i915); mutex_lock(&i915->drm.struct_mutex); @@ -5124,12 +4583,12 @@ int i915_gem_suspend(struct drm_i915_private *i915) if (WARN_ON(!intel_engines_are_idle(i915))) i915_gem_set_wedged(i915); /* no hope, discard everything */ - intel_runtime_pm_put(i915); + intel_runtime_pm_put(i915, wakeref); return 0; err_unlock: mutex_unlock(&i915->drm.struct_mutex); - intel_runtime_pm_put(i915); + intel_runtime_pm_put(i915, wakeref); return ret; } @@ -5223,15 +4682,15 @@ void i915_gem_init_swizzling(struct drm_i915_private *dev_priv) I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) | DISP_TILE_SURFACE_SWIZZLING); - if (IS_GEN5(dev_priv)) + if (IS_GEN(dev_priv, 5)) return; I915_WRITE(TILECTL, I915_READ(TILECTL) | TILECTL_SWZCTL); - if (IS_GEN6(dev_priv)) + if (IS_GEN(dev_priv, 6)) I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_SNB)); - else if (IS_GEN7(dev_priv)) + else if (IS_GEN(dev_priv, 7)) I915_WRITE(ARB_MODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_IVB)); - else if (IS_GEN8(dev_priv)) + else if (IS_GEN(dev_priv, 8)) I915_WRITE(GAMTARBMODE, _MASKED_BIT_ENABLE(ARB_MODE_SWIZZLE_BDW)); else BUG(); @@ -5253,10 +4712,10 @@ static void init_unused_rings(struct drm_i915_private *dev_priv) init_unused_ring(dev_priv, SRB1_BASE); init_unused_ring(dev_priv, SRB2_BASE); init_unused_ring(dev_priv, SRB3_BASE); - } else if (IS_GEN2(dev_priv)) { + } else if (IS_GEN(dev_priv, 2)) { init_unused_ring(dev_priv, SRB0_BASE); init_unused_ring(dev_priv, SRB1_BASE); - } else if (IS_GEN3(dev_priv)) { + } else if (IS_GEN(dev_priv, 3)) { init_unused_ring(dev_priv, PRB1_BASE); init_unused_ring(dev_priv, PRB2_BASE); } @@ -5580,7 +5039,7 @@ int i915_gem_init(struct drm_i915_private *dev_priv) } ret = i915_gem_init_scratch(dev_priv, - IS_GEN2(dev_priv) ? SZ_256K : PAGE_SIZE); + IS_GEN(dev_priv, 2) ? SZ_256K : PAGE_SIZE); if (ret) { GEM_BUG_ON(ret == -EIO); goto err_ggtt; @@ -5840,6 +5299,7 @@ int i915_gem_init_early(struct drm_i915_private *dev_priv) i915_gem_idle_work_handler); init_waitqueue_head(&dev_priv->gpu_error.wait_queue); init_waitqueue_head(&dev_priv->gpu_error.reset_queue); + mutex_init(&dev_priv->gpu_error.wedge_mutex); atomic_set(&dev_priv->mm.bsd_engine_dispatch_index, 0); |