X-Git-Url: http://ftp.safe.ca/?a=blobdiff_plain;f=drivers%2Fgpu%2Fdrm%2Fi915%2Fi915_gem.c;h=5b46623d62d4d811432f406e9429f64588b7b667;hb=f2b115e69d46344ae7afcaad5823496d2a0d8650;hp=e5d2bdf2cc9b664b91d085c2db1d3358848df4ea;hpb=90f959bcb386da2c71613dcefc6a285e054a539e;p=safe%2Fjmp%2Flinux-2.6 diff --git a/drivers/gpu/drm/i915/i915_gem.c b/drivers/gpu/drm/i915/i915_gem.c index e5d2bdf..5b46623 100644 --- a/drivers/gpu/drm/i915/i915_gem.c +++ b/drivers/gpu/drm/i915/i915_gem.c @@ -29,6 +29,8 @@ #include "drm.h" #include "i915_drm.h" #include "i915_drv.h" +#include "i915_trace.h" +#include "intel_drv.h" #include #include @@ -43,18 +45,19 @@ static int i915_gem_object_set_cpu_read_domain_range(struct drm_gem_object *obj, uint64_t offset, uint64_t size); static void i915_gem_object_set_to_full_cpu_read_domain(struct drm_gem_object *obj); -static int i915_gem_object_get_pages(struct drm_gem_object *obj); -static void i915_gem_object_put_pages(struct drm_gem_object *obj); static int i915_gem_object_wait_rendering(struct drm_gem_object *obj); static int i915_gem_object_bind_to_gtt(struct drm_gem_object *obj, unsigned alignment); -static int i915_gem_object_get_fence_reg(struct drm_gem_object *obj, bool write); static void i915_gem_clear_fence_reg(struct drm_gem_object *obj); -static int i915_gem_evict_something(struct drm_device *dev); +static int i915_gem_evict_something(struct drm_device *dev, int min_size); +static int i915_gem_evict_from_inactive_list(struct drm_device *dev); static int i915_gem_phys_pwrite(struct drm_device *dev, struct drm_gem_object *obj, struct drm_i915_gem_pwrite *args, struct drm_file *file_priv); +static LIST_HEAD(shrink_list); +static DEFINE_SPINLOCK(shrink_list_lock); + int i915_gem_do_init(struct drm_device *dev, unsigned long start, unsigned long end) { @@ -114,7 +117,8 @@ i915_gem_create_ioctl(struct drm_device *dev, void *data, { struct drm_i915_gem_create *args = data; struct drm_gem_object *obj; - int handle, ret; + int ret; + u32 handle; args->size = roundup(args->size, PAGE_SIZE); @@ -143,15 +147,27 @@ fast_shmem_read(struct page **pages, int length) { char __iomem *vaddr; - int ret; + int unwritten; vaddr = kmap_atomic(pages[page_base >> PAGE_SHIFT], KM_USER0); if (vaddr == NULL) return -ENOMEM; - ret = __copy_to_user_inatomic(data, vaddr + page_offset, length); + unwritten = __copy_to_user_inatomic(data, vaddr + page_offset, length); kunmap_atomic(vaddr, KM_USER0); - return ret; + if (unwritten) + return -EFAULT; + + return 0; +} + +static int i915_gem_object_needs_bit17_swizzle(struct drm_gem_object *obj) +{ + drm_i915_private_t *dev_priv = obj->dev->dev_private; + struct drm_i915_gem_object *obj_priv = obj->driver_private; + + return dev_priv->mm.bit_6_swizzle_x == I915_BIT_6_SWIZZLE_9_10_17 && + obj_priv->tiling_mode != I915_TILING_NONE; } static inline int @@ -181,6 +197,64 @@ slow_shmem_copy(struct page *dst_page, return 0; } +static inline int +slow_shmem_bit17_copy(struct page *gpu_page, + int gpu_offset, + struct page *cpu_page, + int cpu_offset, + int length, + int is_read) +{ + char *gpu_vaddr, *cpu_vaddr; + + /* Use the unswizzled path if this page isn't affected. */ + if ((page_to_phys(gpu_page) & (1 << 17)) == 0) { + if (is_read) + return slow_shmem_copy(cpu_page, cpu_offset, + gpu_page, gpu_offset, length); + else + return slow_shmem_copy(gpu_page, gpu_offset, + cpu_page, cpu_offset, length); + } + + gpu_vaddr = kmap_atomic(gpu_page, KM_USER0); + if (gpu_vaddr == NULL) + return -ENOMEM; + + cpu_vaddr = kmap_atomic(cpu_page, KM_USER1); + if (cpu_vaddr == NULL) { + kunmap_atomic(gpu_vaddr, KM_USER0); + return -ENOMEM; + } + + /* Copy the data, XORing A6 with A17 (1). The user already knows he's + * XORing with the other bits (A9 for Y, A9 and A10 for X) + */ + 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; + + if (is_read) { + memcpy(cpu_vaddr + cpu_offset, + gpu_vaddr + swizzled_gpu_offset, + this_length); + } else { + memcpy(gpu_vaddr + swizzled_gpu_offset, + cpu_vaddr + cpu_offset, + this_length); + } + cpu_offset += this_length; + gpu_offset += this_length; + length -= this_length; + } + + kunmap_atomic(cpu_vaddr, KM_USER1); + kunmap_atomic(gpu_vaddr, KM_USER0); + + return 0; +} + /** * This is the fast shmem pread path, which attempts to copy_from_user directly * from the backing pages of the object to the user's address space. On a @@ -247,6 +321,45 @@ fail_unlock: return ret; } +static inline gfp_t +i915_gem_object_get_page_gfp_mask (struct drm_gem_object *obj) +{ + return mapping_gfp_mask(obj->filp->f_path.dentry->d_inode->i_mapping); +} + +static inline void +i915_gem_object_set_page_gfp_mask (struct drm_gem_object *obj, gfp_t gfp) +{ + mapping_set_gfp_mask(obj->filp->f_path.dentry->d_inode->i_mapping, gfp); +} + +static int +i915_gem_object_get_pages_or_evict(struct drm_gem_object *obj) +{ + int ret; + + ret = i915_gem_object_get_pages(obj); + + /* If we've insufficient memory to map in the pages, attempt + * to make some space by throwing out some old buffers. + */ + if (ret == -ENOMEM) { + struct drm_device *dev = obj->dev; + gfp_t gfp; + + ret = i915_gem_evict_something(dev, obj->size); + if (ret) + return ret; + + gfp = i915_gem_object_get_page_gfp_mask(obj); + i915_gem_object_set_page_gfp_mask(obj, gfp & ~__GFP_NORETRY); + ret = i915_gem_object_get_pages(obj); + i915_gem_object_set_page_gfp_mask (obj, gfp); + } + + return ret; +} + /** * This is the fallback shmem pread path, which allocates temporary storage * in kernel space to copy_to_user into outside of the struct_mutex, so we @@ -269,6 +382,7 @@ i915_gem_shmem_pread_slow(struct drm_device *dev, struct drm_gem_object *obj, int page_length; int ret; uint64_t data_ptr = args->data_ptr; + int do_bit17_swizzling; remain = args->size; @@ -280,23 +394,25 @@ i915_gem_shmem_pread_slow(struct drm_device *dev, struct drm_gem_object *obj, last_data_page = (data_ptr + args->size - 1) / PAGE_SIZE; num_pages = last_data_page - first_data_page + 1; - user_pages = kcalloc(num_pages, sizeof(struct page *), GFP_KERNEL); + user_pages = drm_calloc_large(num_pages, sizeof(struct page *)); if (user_pages == NULL) return -ENOMEM; down_read(&mm->mmap_sem); pinned_pages = get_user_pages(current, mm, (uintptr_t)args->data_ptr, - num_pages, 0, 0, user_pages, NULL); + num_pages, 1, 0, user_pages, NULL); up_read(&mm->mmap_sem); if (pinned_pages < num_pages) { ret = -EFAULT; goto fail_put_user_pages; } + do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj); + mutex_lock(&dev->struct_mutex); - ret = i915_gem_object_get_pages(obj); - if (ret != 0) + ret = i915_gem_object_get_pages_or_evict(obj); + if (ret) goto fail_unlock; ret = i915_gem_object_set_cpu_read_domain_range(obj, args->offset, @@ -327,11 +443,20 @@ i915_gem_shmem_pread_slow(struct drm_device *dev, struct drm_gem_object *obj, if ((data_page_offset + page_length) > PAGE_SIZE) page_length = PAGE_SIZE - data_page_offset; - ret = slow_shmem_copy(user_pages[data_page_index], - data_page_offset, - obj_priv->pages[shmem_page_index], - shmem_page_offset, - page_length); + if (do_bit17_swizzling) { + ret = slow_shmem_bit17_copy(obj_priv->pages[shmem_page_index], + shmem_page_offset, + user_pages[data_page_index], + data_page_offset, + page_length, + 1); + } else { + ret = slow_shmem_copy(user_pages[data_page_index], + data_page_offset, + obj_priv->pages[shmem_page_index], + shmem_page_offset, + page_length); + } if (ret) goto fail_put_pages; @@ -349,7 +474,7 @@ fail_put_user_pages: SetPageDirty(user_pages[i]); page_cache_release(user_pages[i]); } - kfree(user_pages); + drm_free_large(user_pages); return ret; } @@ -383,9 +508,14 @@ i915_gem_pread_ioctl(struct drm_device *dev, void *data, return -EINVAL; } - ret = i915_gem_shmem_pread_fast(dev, obj, args, file_priv); - if (ret != 0) + if (i915_gem_object_needs_bit17_swizzle(obj)) { ret = i915_gem_shmem_pread_slow(dev, obj, args, file_priv); + } else { + ret = i915_gem_shmem_pread_fast(dev, obj, args, file_priv); + if (ret != 0) + ret = i915_gem_shmem_pread_slow(dev, obj, args, + file_priv); + } drm_gem_object_unreference(obj); @@ -446,13 +576,16 @@ fast_shmem_write(struct page **pages, int length) { char __iomem *vaddr; + unsigned long unwritten; vaddr = kmap_atomic(pages[page_base >> PAGE_SHIFT], KM_USER0); if (vaddr == NULL) return -ENOMEM; - __copy_from_user_inatomic(vaddr + page_offset, data, length); + unwritten = __copy_from_user_inatomic(vaddr + page_offset, data, length); kunmap_atomic(vaddr, KM_USER0); + if (unwritten) + return -EFAULT; return 0; } @@ -561,7 +694,7 @@ i915_gem_gtt_pwrite_slow(struct drm_device *dev, struct drm_gem_object *obj, last_data_page = (data_ptr + args->size - 1) / PAGE_SIZE; num_pages = last_data_page - first_data_page + 1; - user_pages = kcalloc(num_pages, sizeof(struct page *), GFP_KERNEL); + user_pages = drm_calloc_large(num_pages, sizeof(struct page *)); if (user_pages == NULL) return -ENOMEM; @@ -631,7 +764,7 @@ out_unlock: out_unpin_pages: for (i = 0; i < pinned_pages; i++) page_cache_release(user_pages[i]); - kfree(user_pages); + drm_free_large(user_pages); return ret; } @@ -724,6 +857,7 @@ i915_gem_shmem_pwrite_slow(struct drm_device *dev, struct drm_gem_object *obj, int page_length; int ret; uint64_t data_ptr = args->data_ptr; + int do_bit17_swizzling; remain = args->size; @@ -735,7 +869,7 @@ i915_gem_shmem_pwrite_slow(struct drm_device *dev, struct drm_gem_object *obj, last_data_page = (data_ptr + args->size - 1) / PAGE_SIZE; num_pages = last_data_page - first_data_page + 1; - user_pages = kcalloc(num_pages, sizeof(struct page *), GFP_KERNEL); + user_pages = drm_calloc_large(num_pages, sizeof(struct page *)); if (user_pages == NULL) return -ENOMEM; @@ -748,10 +882,12 @@ i915_gem_shmem_pwrite_slow(struct drm_device *dev, struct drm_gem_object *obj, goto fail_put_user_pages; } + do_bit17_swizzling = i915_gem_object_needs_bit17_swizzle(obj); + mutex_lock(&dev->struct_mutex); - ret = i915_gem_object_get_pages(obj); - if (ret != 0) + ret = i915_gem_object_get_pages_or_evict(obj); + if (ret) goto fail_unlock; ret = i915_gem_object_set_to_cpu_domain(obj, 1); @@ -782,11 +918,20 @@ i915_gem_shmem_pwrite_slow(struct drm_device *dev, struct drm_gem_object *obj, if ((data_page_offset + page_length) > PAGE_SIZE) page_length = PAGE_SIZE - data_page_offset; - ret = slow_shmem_copy(obj_priv->pages[shmem_page_index], - shmem_page_offset, - user_pages[data_page_index], - data_page_offset, - page_length); + if (do_bit17_swizzling) { + ret = slow_shmem_bit17_copy(obj_priv->pages[shmem_page_index], + shmem_page_offset, + user_pages[data_page_index], + data_page_offset, + page_length, + 0); + } else { + ret = slow_shmem_copy(obj_priv->pages[shmem_page_index], + shmem_page_offset, + user_pages[data_page_index], + data_page_offset, + page_length); + } if (ret) goto fail_put_pages; @@ -802,7 +947,7 @@ fail_unlock: fail_put_user_pages: for (i = 0; i < pinned_pages; i++) page_cache_release(user_pages[i]); - kfree(user_pages); + drm_free_large(user_pages); return ret; } @@ -851,6 +996,8 @@ i915_gem_pwrite_ioctl(struct drm_device *dev, void *data, ret = i915_gem_gtt_pwrite_slow(dev, obj, args, file_priv); } + } else if (i915_gem_object_needs_bit17_swizzle(obj)) { + ret = i915_gem_shmem_pwrite_slow(dev, obj, args, file_priv); } else { ret = i915_gem_shmem_pwrite_fast(dev, obj, args, file_priv); if (ret == -EFAULT) { @@ -877,8 +1024,10 @@ int i915_gem_set_domain_ioctl(struct drm_device *dev, void *data, struct drm_file *file_priv) { + struct drm_i915_private *dev_priv = dev->dev_private; struct drm_i915_gem_set_domain *args = data; struct drm_gem_object *obj; + struct drm_i915_gem_object *obj_priv; uint32_t read_domains = args->read_domains; uint32_t write_domain = args->write_domain; int ret; @@ -887,10 +1036,10 @@ i915_gem_set_domain_ioctl(struct drm_device *dev, void *data, return -ENODEV; /* Only handle setting domains to types used by the CPU. */ - if (write_domain & ~(I915_GEM_DOMAIN_CPU | I915_GEM_DOMAIN_GTT)) + if (write_domain & I915_GEM_GPU_DOMAINS) return -EINVAL; - if (read_domains & ~(I915_GEM_DOMAIN_CPU | I915_GEM_DOMAIN_GTT)) + if (read_domains & I915_GEM_GPU_DOMAINS) return -EINVAL; /* Having something in the write domain implies it's in the read @@ -902,15 +1051,27 @@ i915_gem_set_domain_ioctl(struct drm_device *dev, void *data, obj = drm_gem_object_lookup(dev, file_priv, args->handle); if (obj == NULL) return -EBADF; + obj_priv = obj->driver_private; mutex_lock(&dev->struct_mutex); + + intel_mark_busy(dev, obj); + #if WATCH_BUF - DRM_INFO("set_domain_ioctl %p(%d), %08x %08x\n", + DRM_INFO("set_domain_ioctl %p(%zd), %08x %08x\n", obj, obj->size, read_domains, write_domain); #endif if (read_domains & I915_GEM_DOMAIN_GTT) { ret = i915_gem_object_set_to_gtt_domain(obj, write_domain != 0); + /* Update the LRU on the fence for the CPU access that's + * about to occur. + */ + if (obj_priv->fence_reg != I915_FENCE_REG_NONE) { + list_move_tail(&obj_priv->fence_list, + &dev_priv->mm.fence_list); + } + /* Silently promote "you're not bound, there was nothing to do" * to success, since the client was just asking us to * make sure everything was done. @@ -949,7 +1110,7 @@ i915_gem_sw_finish_ioctl(struct drm_device *dev, void *data, } #if WATCH_BUF - DRM_INFO("%s: sw_finish %d (%p %d)\n", + DRM_INFO("%s: sw_finish %d (%p %zd)\n", __func__, args->handle, obj, obj->size); #endif obj_priv = obj->driver_private; @@ -1038,22 +1199,22 @@ int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf) /* Now bind it into the GTT if needed */ mutex_lock(&dev->struct_mutex); if (!obj_priv->gtt_space) { - ret = i915_gem_object_bind_to_gtt(obj, obj_priv->gtt_alignment); - if (ret) { - mutex_unlock(&dev->struct_mutex); - return VM_FAULT_SIGBUS; - } - list_add(&obj_priv->list, &dev_priv->mm.inactive_list); + ret = i915_gem_object_bind_to_gtt(obj, 0); + if (ret) + goto unlock; + + list_add_tail(&obj_priv->list, &dev_priv->mm.inactive_list); + + ret = i915_gem_object_set_to_gtt_domain(obj, write); + if (ret) + goto unlock; } /* Need a new fence register? */ - if (obj_priv->fence_reg == I915_FENCE_REG_NONE && - obj_priv->tiling_mode != I915_TILING_NONE) { - ret = i915_gem_object_get_fence_reg(obj, write); - if (ret) { - mutex_unlock(&dev->struct_mutex); - return VM_FAULT_SIGBUS; - } + if (obj_priv->tiling_mode != I915_TILING_NONE) { + ret = i915_gem_object_get_fence_reg(obj); + if (ret) + goto unlock; } pfn = ((dev->agp->base + obj_priv->gtt_offset) >> PAGE_SHIFT) + @@ -1061,17 +1222,18 @@ int i915_gem_fault(struct vm_area_struct *vma, struct vm_fault *vmf) /* Finally, remap it using the new GTT offset */ ret = vm_insert_pfn(vma, (unsigned long)vmf->virtual_address, pfn); - +unlock: mutex_unlock(&dev->struct_mutex); switch (ret) { + case 0: + case -ERESTARTSYS: + return VM_FAULT_NOPAGE; case -ENOMEM: case -EAGAIN: return VM_FAULT_OOM; - case -EFAULT: - return VM_FAULT_SIGBUS; default: - return VM_FAULT_NOPAGE; + return VM_FAULT_SIGBUS; } } @@ -1098,8 +1260,7 @@ i915_gem_create_mmap_offset(struct drm_gem_object *obj) /* Set the object up for mmap'ing */ list = &obj->map_list; - list->map = drm_calloc(1, sizeof(struct drm_map_list), - DRM_MEM_DRIVER); + list->map = kzalloc(sizeof(struct drm_map_list), GFP_KERNEL); if (!list->map) return -ENOMEM; @@ -1139,11 +1300,36 @@ i915_gem_create_mmap_offset(struct drm_gem_object *obj) out_free_mm: drm_mm_put_block(list->file_offset_node); out_free_list: - drm_free(list->map, sizeof(struct drm_map_list), DRM_MEM_DRIVER); + kfree(list->map); return ret; } +/** + * i915_gem_release_mmap - remove physical page mappings + * @obj: obj in question + * + * Preserve the reservation of the mmaping with the DRM core code, but + * relinquish ownership of the pages back to the system. + * + * It is vital that we remove the page mapping if we have mapped a tiled + * object through the GTT and then lose the fence register due to + * resource pressure. Similarly if the object has been moved out of the + * aperture, than pages mapped into userspace must be revoked. Removing the + * mapping will then trigger a page fault on the next user access, allowing + * fixup by i915_gem_fault(). + */ +void +i915_gem_release_mmap(struct drm_gem_object *obj) +{ + struct drm_device *dev = obj->dev; + struct drm_i915_gem_object *obj_priv = obj->driver_private; + + if (dev->dev_mapping) + unmap_mapping_range(dev->dev_mapping, + obj_priv->mmap_offset, obj->size, 1); +} + static void i915_gem_free_mmap_offset(struct drm_gem_object *obj) { @@ -1161,7 +1347,7 @@ i915_gem_free_mmap_offset(struct drm_gem_object *obj) } if (list->map) { - drm_free(list->map, sizeof(struct drm_map), DRM_MEM_DRIVER); + kfree(list->map); list->map = NULL; } @@ -1240,6 +1426,14 @@ i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data, obj_priv = obj->driver_private; + if (obj_priv->madv != I915_MADV_WILLNEED) { + DRM_ERROR("Attempting to mmap a purgeable buffer\n"); + drm_gem_object_unreference(obj); + mutex_unlock(&dev->struct_mutex); + return -EINVAL; + } + + if (!obj_priv->mmap_offset) { ret = i915_gem_create_mmap_offset(obj); if (ret) { @@ -1251,28 +1445,18 @@ i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data, args->offset = obj_priv->mmap_offset; - obj_priv->gtt_alignment = i915_gem_get_gtt_alignment(obj); - - /* Make sure the alignment is correct for fence regs etc */ - if (obj_priv->agp_mem && - (obj_priv->gtt_offset & (obj_priv->gtt_alignment - 1))) { - drm_gem_object_unreference(obj); - mutex_unlock(&dev->struct_mutex); - return -EINVAL; - } - /* * Pull it into the GTT so that we have a page list (makes the * initial fault faster and any subsequent flushing possible). */ if (!obj_priv->agp_mem) { - ret = i915_gem_object_bind_to_gtt(obj, obj_priv->gtt_alignment); + ret = i915_gem_object_bind_to_gtt(obj, 0); if (ret) { drm_gem_object_unreference(obj); mutex_unlock(&dev->struct_mutex); return ret; } - list_add(&obj_priv->list, &dev_priv->mm.inactive_list); + list_add_tail(&obj_priv->list, &dev_priv->mm.inactive_list); } drm_gem_object_unreference(obj); @@ -1281,7 +1465,7 @@ i915_gem_mmap_gtt_ioctl(struct drm_device *dev, void *data, return 0; } -static void +void i915_gem_object_put_pages(struct drm_gem_object *obj) { struct drm_i915_gem_object *obj_priv = obj->driver_private; @@ -1289,22 +1473,32 @@ i915_gem_object_put_pages(struct drm_gem_object *obj) int i; BUG_ON(obj_priv->pages_refcount == 0); + BUG_ON(obj_priv->madv == __I915_MADV_PURGED); if (--obj_priv->pages_refcount != 0) return; - for (i = 0; i < page_count; i++) - if (obj_priv->pages[i] != NULL) { - if (obj_priv->dirty) - set_page_dirty(obj_priv->pages[i]); + if (obj_priv->tiling_mode != I915_TILING_NONE) + i915_gem_object_save_bit_17_swizzle(obj); + + if (obj_priv->madv == I915_MADV_DONTNEED) + obj_priv->dirty = 0; + + for (i = 0; i < page_count; i++) { + if (obj_priv->pages[i] == NULL) + break; + + if (obj_priv->dirty) + set_page_dirty(obj_priv->pages[i]); + + if (obj_priv->madv == I915_MADV_WILLNEED) mark_page_accessed(obj_priv->pages[i]); - page_cache_release(obj_priv->pages[i]); - } + + page_cache_release(obj_priv->pages[i]); + } obj_priv->dirty = 0; - drm_free(obj_priv->pages, - page_count * sizeof(struct page *), - DRM_MEM_DRIVER); + drm_free_large(obj_priv->pages); obj_priv->pages = NULL; } @@ -1321,8 +1515,10 @@ i915_gem_object_move_to_active(struct drm_gem_object *obj, uint32_t seqno) obj_priv->active = 1; } /* Move from whatever list we were on to the tail of execution. */ + spin_lock(&dev_priv->mm.active_list_lock); list_move_tail(&obj_priv->list, &dev_priv->mm.active_list); + spin_unlock(&dev_priv->mm.active_list_lock); obj_priv->last_rendering_seqno = seqno; } @@ -1338,6 +1534,26 @@ i915_gem_object_move_to_flushing(struct drm_gem_object *obj) obj_priv->last_rendering_seqno = 0; } +/* Immediately discard the backing storage */ +static void +i915_gem_object_truncate(struct drm_gem_object *obj) +{ + struct drm_i915_gem_object *obj_priv = obj->driver_private; + struct inode *inode; + + inode = obj->filp->f_path.dentry->d_inode; + if (inode->i_op->truncate) + inode->i_op->truncate (inode); + + obj_priv->madv = __I915_MADV_PURGED; +} + +static inline int +i915_gem_object_is_purgeable(struct drm_i915_gem_object *obj_priv) +{ + return obj_priv->madv == I915_MADV_DONTNEED; +} + static void i915_gem_object_move_to_inactive(struct drm_gem_object *obj) { @@ -1367,16 +1583,21 @@ i915_gem_object_move_to_inactive(struct drm_gem_object *obj) * * Returned sequence numbers are nonzero on success. */ -static uint32_t -i915_add_request(struct drm_device *dev, uint32_t flush_domains) +uint32_t +i915_add_request(struct drm_device *dev, struct drm_file *file_priv, + uint32_t flush_domains) { drm_i915_private_t *dev_priv = dev->dev_private; + struct drm_i915_file_private *i915_file_priv = NULL; struct drm_i915_gem_request *request; uint32_t seqno; int was_empty; RING_LOCALS; - request = drm_calloc(1, sizeof(*request), DRM_MEM_DRIVER); + if (file_priv != NULL) + i915_file_priv = file_priv->driver_priv; + + request = kzalloc(sizeof(*request), GFP_KERNEL); if (request == NULL) return 0; @@ -1396,12 +1617,18 @@ i915_add_request(struct drm_device *dev, uint32_t flush_domains) OUT_RING(MI_USER_INTERRUPT); ADVANCE_LP_RING(); - DRM_DEBUG("%d\n", seqno); + DRM_DEBUG_DRIVER("%d\n", seqno); request->seqno = seqno; request->emitted_jiffies = jiffies; was_empty = list_empty(&dev_priv->mm.request_list); list_add_tail(&request->list, &dev_priv->mm.request_list); + if (i915_file_priv) { + list_add_tail(&request->client_list, + &i915_file_priv->mm.request_list); + } else { + INIT_LIST_HEAD(&request->client_list); + } /* Associate any objects on the flushing list matching the write * domain we're flushing with our flush. @@ -1415,15 +1642,24 @@ i915_add_request(struct drm_device *dev, uint32_t flush_domains) if ((obj->write_domain & flush_domains) == obj->write_domain) { + uint32_t old_write_domain = obj->write_domain; + obj->write_domain = 0; i915_gem_object_move_to_active(obj, seqno); + + trace_i915_gem_object_change_domain(obj, + obj->read_domains, + old_write_domain); } } } - if (was_empty && !dev_priv->mm.suspended) - schedule_delayed_work(&dev_priv->mm.retire_work, HZ); + if (!dev_priv->mm.suspended) { + mod_timer(&dev_priv->hangcheck_timer, jiffies + DRM_I915_HANGCHECK_PERIOD); + if (was_empty) + queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, HZ); + } return seqno; } @@ -1461,9 +1697,12 @@ i915_gem_retire_request(struct drm_device *dev, { drm_i915_private_t *dev_priv = dev->dev_private; + trace_i915_gem_request_retire(dev, request->seqno); + /* Move any buffers on the active list that are no longer referenced * by the ringbuffer to the flushing/inactive lists as appropriate. */ + spin_lock(&dev_priv->mm.active_list_lock); while (!list_empty(&dev_priv->mm.active_list)) { struct drm_gem_object *obj; struct drm_i915_gem_object *obj_priv; @@ -1478,7 +1717,7 @@ i915_gem_retire_request(struct drm_device *dev, * this seqno. */ if (obj_priv->last_rendering_seqno != request->seqno) - return; + goto out; #if WATCH_LRU DRM_INFO("%s: retire %d moves to inactive list %p\n", @@ -1487,15 +1726,28 @@ i915_gem_retire_request(struct drm_device *dev, if (obj->write_domain != 0) i915_gem_object_move_to_flushing(obj); - else + else { + /* Take a reference on the object so it won't be + * freed while the spinlock is held. The list + * protection for this spinlock is safe when breaking + * the lock like this since the next thing we do + * is just get the head of the list again. + */ + drm_gem_object_reference(obj); i915_gem_object_move_to_inactive(obj); + spin_unlock(&dev_priv->mm.active_list_lock); + drm_gem_object_unreference(obj); + spin_lock(&dev_priv->mm.active_list_lock); + } } +out: + spin_unlock(&dev_priv->mm.active_list_lock); } /** * Returns true if seq1 is later than seq2. */ -static int +bool i915_seqno_passed(uint32_t seq1, uint32_t seq2) { return (int32_t)(seq1 - seq2) >= 0; @@ -1518,7 +1770,7 @@ i915_gem_retire_requests(struct drm_device *dev) drm_i915_private_t *dev_priv = dev->dev_private; uint32_t seqno; - if (!dev_priv->hw_status_page) + if (!dev_priv->hw_status_page || list_empty(&dev_priv->mm.request_list)) return; seqno = i915_get_gem_seqno(dev); @@ -1533,14 +1785,21 @@ i915_gem_retire_requests(struct drm_device *dev) retiring_seqno = request->seqno; if (i915_seqno_passed(seqno, retiring_seqno) || - dev_priv->mm.wedged) { + atomic_read(&dev_priv->mm.wedged)) { i915_gem_retire_request(dev, request); list_del(&request->list); - drm_free(request, sizeof(*request), DRM_MEM_DRIVER); + list_del(&request->client_list); + kfree(request); } else break; } + + if (unlikely (dev_priv->trace_irq_seqno && + i915_seqno_passed(dev_priv->trace_irq_seqno, seqno))) { + i915_user_irq_put(dev); + dev_priv->trace_irq_seqno = 0; + } } void @@ -1557,33 +1816,53 @@ i915_gem_retire_work_handler(struct work_struct *work) i915_gem_retire_requests(dev); if (!dev_priv->mm.suspended && !list_empty(&dev_priv->mm.request_list)) - schedule_delayed_work(&dev_priv->mm.retire_work, HZ); + queue_delayed_work(dev_priv->wq, &dev_priv->mm.retire_work, HZ); mutex_unlock(&dev->struct_mutex); } -/** - * Waits for a sequence number to be signaled, and cleans up the - * request and object lists appropriately for that event. - */ -static int -i915_wait_request(struct drm_device *dev, uint32_t seqno) +int +i915_do_wait_request(struct drm_device *dev, uint32_t seqno, int interruptible) { drm_i915_private_t *dev_priv = dev->dev_private; + u32 ier; int ret = 0; BUG_ON(seqno == 0); + if (atomic_read(&dev_priv->mm.wedged)) + return -EIO; + if (!i915_seqno_passed(i915_get_gem_seqno(dev), seqno)) { + if (IS_IRONLAKE(dev)) + ier = I915_READ(DEIER) | I915_READ(GTIER); + else + ier = I915_READ(IER); + if (!ier) { + DRM_ERROR("something (likely vbetool) disabled " + "interrupts, re-enabling\n"); + i915_driver_irq_preinstall(dev); + i915_driver_irq_postinstall(dev); + } + + trace_i915_gem_request_wait_begin(dev, seqno); + dev_priv->mm.waiting_gem_seqno = seqno; i915_user_irq_get(dev); - ret = wait_event_interruptible(dev_priv->irq_queue, - i915_seqno_passed(i915_get_gem_seqno(dev), - seqno) || - dev_priv->mm.wedged); + if (interruptible) + ret = wait_event_interruptible(dev_priv->irq_queue, + i915_seqno_passed(i915_get_gem_seqno(dev), seqno) || + atomic_read(&dev_priv->mm.wedged)); + else + wait_event(dev_priv->irq_queue, + i915_seqno_passed(i915_get_gem_seqno(dev), seqno) || + atomic_read(&dev_priv->mm.wedged)); + i915_user_irq_put(dev); dev_priv->mm.waiting_gem_seqno = 0; + + trace_i915_gem_request_wait_end(dev, seqno); } - if (dev_priv->mm.wedged) + if (atomic_read(&dev_priv->mm.wedged)) ret = -EIO; if (ret && ret != -ERESTARTSYS) @@ -1601,6 +1880,16 @@ i915_wait_request(struct drm_device *dev, uint32_t seqno) return ret; } +/** + * Waits for a sequence number to be signaled, and cleans up the + * request and object lists appropriately for that event. + */ +static int +i915_wait_request(struct drm_device *dev, uint32_t seqno) +{ + return i915_do_wait_request(dev, seqno, 1); +} + static void i915_gem_flush(struct drm_device *dev, uint32_t invalidate_domains, @@ -1614,12 +1903,13 @@ i915_gem_flush(struct drm_device *dev, DRM_INFO("%s: invalidate %08x flush %08x\n", __func__, invalidate_domains, flush_domains); #endif + trace_i915_gem_request_flush(dev, dev_priv->mm.next_gem_seqno, + invalidate_domains, flush_domains); if (flush_domains & I915_GEM_DOMAIN_CPU) drm_agp_chipset_flush(dev); - if ((invalidate_domains | flush_domains) & ~(I915_GEM_DOMAIN_CPU | - I915_GEM_DOMAIN_GTT)) { + if ((invalidate_domains | flush_domains) & I915_GEM_GPU_DOMAINS) { /* * read/write caches: * @@ -1668,7 +1958,7 @@ i915_gem_flush(struct drm_device *dev, #endif BEGIN_LP_RING(2); OUT_RING(cmd); - OUT_RING(0); /* noop */ + OUT_RING(MI_NOOP); ADVANCE_LP_RING(); } } @@ -1713,7 +2003,6 @@ i915_gem_object_unbind(struct drm_gem_object *obj) { struct drm_device *dev = obj->dev; struct drm_i915_gem_object *obj_priv = obj->driver_private; - loff_t offset; int ret = 0; #if WATCH_BUF @@ -1728,6 +2017,12 @@ i915_gem_object_unbind(struct drm_gem_object *obj) return -EINVAL; } + /* blow away mappings if mapped through GTT */ + i915_gem_release_mmap(obj); + + if (obj_priv->fence_reg != I915_FENCE_REG_NONE) + i915_gem_clear_fence_reg(obj); + /* Move the object to the CPU domain to ensure that * any possible CPU writes while it's not in the GTT * are flushed when we go to remap it. This will @@ -1741,23 +2036,16 @@ i915_gem_object_unbind(struct drm_gem_object *obj) return ret; } + BUG_ON(obj_priv->active); + if (obj_priv->agp_mem != NULL) { drm_unbind_agp(obj_priv->agp_mem); drm_free_agp(obj_priv->agp_mem, obj->size / PAGE_SIZE); obj_priv->agp_mem = NULL; } - BUG_ON(obj_priv->active); - - /* blow away mappings if mapped through GTT */ - offset = ((loff_t) obj->map_list.hash.key) << PAGE_SHIFT; - if (dev->dev_mapping) - unmap_mapping_range(dev->dev_mapping, offset, obj->size, 1); - - if (obj_priv->fence_reg != I915_FENCE_REG_NONE) - i915_gem_clear_fence_reg(obj); - i915_gem_object_put_pages(obj); + BUG_ON(obj_priv->pages_refcount); if (obj_priv->gtt_space) { atomic_dec(&dev->gtt_count); @@ -1771,40 +2059,113 @@ i915_gem_object_unbind(struct drm_gem_object *obj) if (!list_empty(&obj_priv->list)) list_del_init(&obj_priv->list); + if (i915_gem_object_is_purgeable(obj_priv)) + i915_gem_object_truncate(obj); + + trace_i915_gem_object_unbind(obj); + + return 0; +} + +static struct drm_gem_object * +i915_gem_find_inactive_object(struct drm_device *dev, int min_size) +{ + drm_i915_private_t *dev_priv = dev->dev_private; + struct drm_i915_gem_object *obj_priv; + struct drm_gem_object *best = NULL; + struct drm_gem_object *first = NULL; + + /* Try to find the smallest clean object */ + list_for_each_entry(obj_priv, &dev_priv->mm.inactive_list, list) { + struct drm_gem_object *obj = obj_priv->obj; + if (obj->size >= min_size) { + if ((!obj_priv->dirty || + i915_gem_object_is_purgeable(obj_priv)) && + (!best || obj->size < best->size)) { + best = obj; + if (best->size == min_size) + return best; + } + if (!first) + first = obj; + } + } + + return best ? best : first; +} + +static int +i915_gem_evict_everything(struct drm_device *dev) +{ + drm_i915_private_t *dev_priv = dev->dev_private; + uint32_t seqno; + int ret; + bool lists_empty; + + spin_lock(&dev_priv->mm.active_list_lock); + lists_empty = (list_empty(&dev_priv->mm.inactive_list) && + list_empty(&dev_priv->mm.flushing_list) && + list_empty(&dev_priv->mm.active_list)); + spin_unlock(&dev_priv->mm.active_list_lock); + + if (lists_empty) + return -ENOSPC; + + /* Flush everything (on to the inactive lists) and evict */ + i915_gem_flush(dev, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS); + seqno = i915_add_request(dev, NULL, I915_GEM_GPU_DOMAINS); + if (seqno == 0) + return -ENOMEM; + + ret = i915_wait_request(dev, seqno); + if (ret) + return ret; + + ret = i915_gem_evict_from_inactive_list(dev); + if (ret) + return ret; + + spin_lock(&dev_priv->mm.active_list_lock); + lists_empty = (list_empty(&dev_priv->mm.inactive_list) && + list_empty(&dev_priv->mm.flushing_list) && + list_empty(&dev_priv->mm.active_list)); + spin_unlock(&dev_priv->mm.active_list_lock); + BUG_ON(!lists_empty); + return 0; } static int -i915_gem_evict_something(struct drm_device *dev) +i915_gem_evict_something(struct drm_device *dev, int min_size) { drm_i915_private_t *dev_priv = dev->dev_private; struct drm_gem_object *obj; - struct drm_i915_gem_object *obj_priv; - int ret = 0; + int ret; for (;;) { + i915_gem_retire_requests(dev); + /* If there's an inactive buffer available now, grab it * and be done. */ - if (!list_empty(&dev_priv->mm.inactive_list)) { - obj_priv = list_first_entry(&dev_priv->mm.inactive_list, - struct drm_i915_gem_object, - list); - obj = obj_priv->obj; - BUG_ON(obj_priv->pin_count != 0); + obj = i915_gem_find_inactive_object(dev, min_size); + if (obj) { + struct drm_i915_gem_object *obj_priv; + #if WATCH_LRU DRM_INFO("%s: evicting %p\n", __func__, obj); #endif + obj_priv = obj->driver_private; + BUG_ON(obj_priv->pin_count != 0); BUG_ON(obj_priv->active); /* Wait on the rendering and unbind the buffer. */ - ret = i915_gem_object_unbind(obj); - break; + return i915_gem_object_unbind(obj); } /* If we didn't get anything, but the ring is still processing - * things, wait for one of those things to finish and hopefully - * leave us a buffer to evict. + * things, wait for the next to finish and hopefully leave us + * a buffer to evict. */ if (!list_empty(&dev_priv->mm.request_list)) { struct drm_i915_gem_request *request; @@ -1815,16 +2176,9 @@ i915_gem_evict_something(struct drm_device *dev) ret = i915_wait_request(dev, request->seqno); if (ret) - break; + return ret; - /* if waiting caused an object to become inactive, - * then loop around and wait for it. Otherwise, we - * assume that waiting freed and unbound something, - * so there should now be some space in the GTT - */ - if (!list_empty(&dev_priv->mm.inactive_list)) - continue; - break; + continue; } /* If we didn't have anything on the request list but there @@ -1833,49 +2187,47 @@ i915_gem_evict_something(struct drm_device *dev) * will get moved to inactive. */ if (!list_empty(&dev_priv->mm.flushing_list)) { - obj_priv = list_first_entry(&dev_priv->mm.flushing_list, - struct drm_i915_gem_object, - list); - obj = obj_priv->obj; + struct drm_i915_gem_object *obj_priv; - i915_gem_flush(dev, - obj->write_domain, - obj->write_domain); - i915_add_request(dev, obj->write_domain); + /* Find an object that we can immediately reuse */ + list_for_each_entry(obj_priv, &dev_priv->mm.flushing_list, list) { + obj = obj_priv->obj; + if (obj->size >= min_size) + break; - obj = NULL; - continue; - } + obj = NULL; + } - DRM_ERROR("inactive empty %d request empty %d " - "flushing empty %d\n", - list_empty(&dev_priv->mm.inactive_list), - list_empty(&dev_priv->mm.request_list), - list_empty(&dev_priv->mm.flushing_list)); - /* If we didn't do any of the above, there's nothing to be done - * and we just can't fit it in. - */ - return -ENOMEM; - } - return ret; -} + if (obj != NULL) { + uint32_t seqno; -static int -i915_gem_evict_everything(struct drm_device *dev) -{ - int ret; + i915_gem_flush(dev, + obj->write_domain, + obj->write_domain); + seqno = i915_add_request(dev, NULL, obj->write_domain); + if (seqno == 0) + return -ENOMEM; - for (;;) { - ret = i915_gem_evict_something(dev); - if (ret != 0) - break; + ret = i915_wait_request(dev, seqno); + if (ret) + return ret; + + continue; + } + } + + /* If we didn't do any of the above, there's no single buffer + * large enough to swap out for the new one, so just evict + * everything and start again. (This should be rare.) + */ + if (!list_empty (&dev_priv->mm.inactive_list)) + return i915_gem_evict_from_inactive_list(dev); + else + return i915_gem_evict_everything(dev); } - if (ret == -ENOMEM) - return 0; - return ret; } -static int +int i915_gem_object_get_pages(struct drm_gem_object *obj) { struct drm_i915_gem_object *obj_priv = obj->driver_private; @@ -1893,10 +2245,8 @@ i915_gem_object_get_pages(struct drm_gem_object *obj) */ page_count = obj->size / PAGE_SIZE; BUG_ON(obj_priv->pages != NULL); - obj_priv->pages = drm_calloc(page_count, sizeof(struct page *), - DRM_MEM_DRIVER); + obj_priv->pages = drm_calloc_large(page_count, sizeof(struct page *)); if (obj_priv->pages == NULL) { - DRM_ERROR("Faled to allocate page list\n"); obj_priv->pages_refcount--; return -ENOMEM; } @@ -1907,12 +2257,15 @@ i915_gem_object_get_pages(struct drm_gem_object *obj) page = read_mapping_page(mapping, i, NULL); if (IS_ERR(page)) { ret = PTR_ERR(page); - DRM_ERROR("read_mapping_page failed: %d\n", ret); i915_gem_object_put_pages(obj); return ret; } obj_priv->pages[i] = page; } + + if (obj_priv->tiling_mode != I915_TILING_NONE) + i915_gem_object_do_bit_17_swizzle(obj); + return 0; } @@ -1987,31 +2340,34 @@ static void i830_write_fence_reg(struct drm_i915_fence_reg *reg) int regnum = obj_priv->fence_reg; uint32_t val; uint32_t pitch_val; + uint32_t fence_size_bits; - if ((obj_priv->gtt_offset & ~I915_FENCE_START_MASK) || + if ((obj_priv->gtt_offset & ~I830_FENCE_START_MASK) || (obj_priv->gtt_offset & (obj->size - 1))) { - WARN(1, "%s: object 0x%08x not 1M or size aligned\n", + WARN(1, "%s: object 0x%08x not 512K or size aligned\n", __func__, obj_priv->gtt_offset); return; } - pitch_val = (obj_priv->stride / 128) - 1; + pitch_val = obj_priv->stride / 128; + pitch_val = ffs(pitch_val) - 1; + WARN_ON(pitch_val > I830_FENCE_MAX_PITCH_VAL); val = obj_priv->gtt_offset; if (obj_priv->tiling_mode == I915_TILING_Y) val |= 1 << I830_FENCE_TILING_Y_SHIFT; - val |= I830_FENCE_SIZE_BITS(obj->size); + fence_size_bits = I830_FENCE_SIZE_BITS(obj->size); + WARN_ON(fence_size_bits & ~0x00000f00); + val |= fence_size_bits; val |= pitch_val << I830_FENCE_PITCH_SHIFT; val |= I830_FENCE_REG_VALID; I915_WRITE(FENCE_REG_830_0 + (regnum * 4), val); - } /** * i915_gem_object_get_fence_reg - set up a fence reg for an object * @obj: object to map through a fence reg - * @write: object is about to be written * * When mapping objects through the GTT, userspace wants to be able to write * to them without having to worry about swizzling if the object is tiled. @@ -2022,8 +2378,8 @@ static void i830_write_fence_reg(struct drm_i915_fence_reg *reg) * It then sets up the reg based on the object's properties: address, pitch * and tiling format. */ -static int -i915_gem_object_get_fence_reg(struct drm_gem_object *obj, bool write) +int +i915_gem_object_get_fence_reg(struct drm_gem_object *obj) { struct drm_device *dev = obj->dev; struct drm_i915_private *dev_priv = dev->dev_private; @@ -2032,6 +2388,12 @@ i915_gem_object_get_fence_reg(struct drm_gem_object *obj, bool write) struct drm_i915_gem_object *old_obj_priv = NULL; int i, ret, avail; + /* Just update our place in the LRU if our fence is getting used. */ + if (obj_priv->fence_reg != I915_FENCE_REG_NONE) { + list_move_tail(&obj_priv->fence_list, &dev_priv->mm.fence_list); + return 0; + } + switch (obj_priv->tiling_mode) { case I915_TILING_NONE: WARN(1, "allocating a fence for non-tiled object?\n"); @@ -2053,7 +2415,6 @@ i915_gem_object_get_fence_reg(struct drm_gem_object *obj, bool write) } /* First try to find a free reg */ -try_again: avail = 0; for (i = dev_priv->fence_reg_start; i < dev_priv->num_fence_regs; i++) { reg = &dev_priv->fence_regs[i]; @@ -2067,70 +2428,62 @@ try_again: /* None available, try to steal one or wait for a user to finish */ if (i == dev_priv->num_fence_regs) { - uint32_t seqno = dev_priv->mm.next_gem_seqno; - loff_t offset; + struct drm_gem_object *old_obj = NULL; if (avail == 0) - return -ENOMEM; - - for (i = dev_priv->fence_reg_start; - i < dev_priv->num_fence_regs; i++) { - uint32_t this_seqno; + return -ENOSPC; - reg = &dev_priv->fence_regs[i]; - old_obj_priv = reg->obj->driver_private; + list_for_each_entry(old_obj_priv, &dev_priv->mm.fence_list, + fence_list) { + old_obj = old_obj_priv->obj; if (old_obj_priv->pin_count) continue; + /* Take a reference, as otherwise the wait_rendering + * below may cause the object to get freed out from + * under us. + */ + drm_gem_object_reference(old_obj); + /* i915 uses fences for GPU access to tiled buffers */ if (IS_I965G(dev) || !old_obj_priv->active) break; - /* find the seqno of the first available fence */ - this_seqno = old_obj_priv->last_rendering_seqno; - if (this_seqno != 0 && - reg->obj->write_domain == 0 && - i915_seqno_passed(seqno, this_seqno)) - seqno = this_seqno; - } - - /* - * Now things get ugly... we have to wait for one of the - * objects to finish before trying again. - */ - if (i == dev_priv->num_fence_regs) { - if (seqno == dev_priv->mm.next_gem_seqno) { - i915_gem_flush(dev, - I915_GEM_GPU_DOMAINS, - I915_GEM_GPU_DOMAINS); - seqno = i915_add_request(dev, - I915_GEM_GPU_DOMAINS); - if (seqno == 0) - return -ENOMEM; + /* This brings the object to the head of the LRU if it + * had been written to. The only way this should + * result in us waiting longer than the expected + * optimal amount of time is if there was a + * fence-using buffer later that was read-only. + */ + i915_gem_object_flush_gpu_write_domain(old_obj); + ret = i915_gem_object_wait_rendering(old_obj); + if (ret != 0) { + drm_gem_object_unreference(old_obj); + return ret; } - ret = i915_wait_request(dev, seqno); - if (ret) - return ret; - goto try_again; + break; } - BUG_ON(old_obj_priv->active || - (reg->obj->write_domain & I915_GEM_GPU_DOMAINS)); - /* * Zap this virtual mapping so we can set up a fence again * for this object next time we need it. */ - offset = ((loff_t) reg->obj->map_list.hash.key) << PAGE_SHIFT; - if (dev->dev_mapping) - unmap_mapping_range(dev->dev_mapping, offset, - reg->obj->size, 1); + i915_gem_release_mmap(old_obj); + + i = old_obj_priv->fence_reg; + reg = &dev_priv->fence_regs[i]; + old_obj_priv->fence_reg = I915_FENCE_REG_NONE; + list_del_init(&old_obj_priv->fence_list); + + drm_gem_object_unreference(old_obj); } obj_priv->fence_reg = i; + list_add_tail(&obj_priv->fence_list, &dev_priv->mm.fence_list); + reg->obj = obj; if (IS_I965G(dev)) @@ -2140,6 +2493,8 @@ try_again: else i830_write_fence_reg(reg); + trace_i915_gem_object_get_fence(obj, i, obj_priv->tiling_mode); + return 0; } @@ -2173,6 +2528,43 @@ i915_gem_clear_fence_reg(struct drm_gem_object *obj) dev_priv->fence_regs[obj_priv->fence_reg].obj = NULL; obj_priv->fence_reg = I915_FENCE_REG_NONE; + list_del_init(&obj_priv->fence_list); +} + +/** + * i915_gem_object_put_fence_reg - waits on outstanding fenced access + * to the buffer to finish, and then resets the fence register. + * @obj: tiled object holding a fence register. + * + * Zeroes out the fence register itself and clears out the associated + * data structures in dev_priv and obj_priv. + */ +int +i915_gem_object_put_fence_reg(struct drm_gem_object *obj) +{ + struct drm_device *dev = obj->dev; + struct drm_i915_gem_object *obj_priv = obj->driver_private; + + if (obj_priv->fence_reg == I915_FENCE_REG_NONE) + return 0; + + /* On the i915, GPU access to tiled buffers is via a fence, + * therefore we must wait for any outstanding access to complete + * before clearing the fence. + */ + if (!IS_I965G(dev)) { + int ret; + + i915_gem_object_flush_gpu_write_domain(obj); + i915_gem_object_flush_gtt_write_domain(obj); + ret = i915_gem_object_wait_rendering(obj); + if (ret != 0) + return ret; + } + + i915_gem_clear_fence_reg (obj); + + return 0; } /** @@ -2185,13 +2577,20 @@ i915_gem_object_bind_to_gtt(struct drm_gem_object *obj, unsigned alignment) drm_i915_private_t *dev_priv = dev->dev_private; struct drm_i915_gem_object *obj_priv = obj->driver_private; struct drm_mm_node *free_space; - int page_count, ret; + bool retry_alloc = false; + int ret; if (dev_priv->mm.suspended) return -EBUSY; + + if (obj_priv->madv != I915_MADV_WILLNEED) { + DRM_ERROR("Attempting to bind a purgeable object\n"); + return -EINVAL; + } + if (alignment == 0) alignment = i915_gem_get_gtt_alignment(obj); - if (alignment & (PAGE_SIZE - 1)) { + if (alignment & (i915_gem_get_gtt_alignment(obj) - 1)) { DRM_ERROR("Invalid object alignment requested %u\n", alignment); return -EINVAL; } @@ -2214,47 +2613,67 @@ i915_gem_object_bind_to_gtt(struct drm_gem_object *obj, unsigned alignment) #if WATCH_LRU DRM_INFO("%s: GTT full, evicting something\n", __func__); #endif - if (list_empty(&dev_priv->mm.inactive_list) && - list_empty(&dev_priv->mm.flushing_list) && - list_empty(&dev_priv->mm.active_list)) { - DRM_ERROR("GTT full, but LRU list empty\n"); - return -ENOMEM; - } - - ret = i915_gem_evict_something(dev); - if (ret != 0) { - if (ret != -ERESTARTSYS) - DRM_ERROR("Failed to evict a buffer %d\n", ret); + ret = i915_gem_evict_something(dev, obj->size); + if (ret) return ret; - } + goto search_free; } #if WATCH_BUF - DRM_INFO("Binding object of size %d at 0x%08x\n", + DRM_INFO("Binding object of size %zd at 0x%08x\n", obj->size, obj_priv->gtt_offset); #endif + if (retry_alloc) { + i915_gem_object_set_page_gfp_mask (obj, + i915_gem_object_get_page_gfp_mask (obj) & ~__GFP_NORETRY); + } ret = i915_gem_object_get_pages(obj); + if (retry_alloc) { + i915_gem_object_set_page_gfp_mask (obj, + i915_gem_object_get_page_gfp_mask (obj) | __GFP_NORETRY); + } if (ret) { drm_mm_put_block(obj_priv->gtt_space); obj_priv->gtt_space = NULL; + + if (ret == -ENOMEM) { + /* first try to clear up some space from the GTT */ + ret = i915_gem_evict_something(dev, obj->size); + if (ret) { + /* now try to shrink everyone else */ + if (! retry_alloc) { + retry_alloc = true; + goto search_free; + } + + return ret; + } + + goto search_free; + } + return ret; } - page_count = obj->size / PAGE_SIZE; /* Create an AGP memory structure pointing at our pages, and bind it * into the GTT. */ obj_priv->agp_mem = drm_agp_bind_pages(dev, obj_priv->pages, - page_count, + obj->size >> PAGE_SHIFT, obj_priv->gtt_offset, obj_priv->agp_type); if (obj_priv->agp_mem == NULL) { i915_gem_object_put_pages(obj); drm_mm_put_block(obj_priv->gtt_space); obj_priv->gtt_space = NULL; - return -ENOMEM; + + ret = i915_gem_evict_something(dev, obj->size); + if (ret) + return ret; + + goto search_free; } atomic_inc(&dev->gtt_count); atomic_add(obj->size, &dev->gtt_memory); @@ -2263,8 +2682,10 @@ i915_gem_object_bind_to_gtt(struct drm_gem_object *obj, unsigned alignment) * wasn't in the GTT, there shouldn't be any way it could have been in * a GPU cache */ - BUG_ON(obj->read_domains & ~(I915_GEM_DOMAIN_CPU|I915_GEM_DOMAIN_GTT)); - BUG_ON(obj->write_domain & ~(I915_GEM_DOMAIN_CPU|I915_GEM_DOMAIN_GTT)); + BUG_ON(obj->read_domains & I915_GEM_GPU_DOMAINS); + BUG_ON(obj->write_domain & I915_GEM_GPU_DOMAINS); + + trace_i915_gem_object_bind(obj, obj_priv->gtt_offset); return 0; } @@ -2281,6 +2702,8 @@ i915_gem_clflush_object(struct drm_gem_object *obj) if (obj_priv->pages == NULL) return; + trace_i915_gem_object_clflush(obj); + drm_clflush_pages(obj_priv->pages, obj->size / PAGE_SIZE); } @@ -2290,21 +2713,29 @@ i915_gem_object_flush_gpu_write_domain(struct drm_gem_object *obj) { struct drm_device *dev = obj->dev; uint32_t seqno; + uint32_t old_write_domain; if ((obj->write_domain & I915_GEM_GPU_DOMAINS) == 0) return; /* Queue the GPU write cache flushing we need. */ + old_write_domain = obj->write_domain; i915_gem_flush(dev, 0, obj->write_domain); - seqno = i915_add_request(dev, obj->write_domain); + seqno = i915_add_request(dev, NULL, obj->write_domain); obj->write_domain = 0; i915_gem_object_move_to_active(obj, seqno); + + trace_i915_gem_object_change_domain(obj, + obj->read_domains, + old_write_domain); } /** Flushes the GTT write domain for the object if it's dirty. */ static void i915_gem_object_flush_gtt_write_domain(struct drm_gem_object *obj) { + uint32_t old_write_domain; + if (obj->write_domain != I915_GEM_DOMAIN_GTT) return; @@ -2312,7 +2743,12 @@ i915_gem_object_flush_gtt_write_domain(struct drm_gem_object *obj) * to it immediately go to main memory as far as we know, so there's * no chipset flush. It also doesn't land in render cache. */ + old_write_domain = obj->write_domain; obj->write_domain = 0; + + trace_i915_gem_object_change_domain(obj, + obj->read_domains, + old_write_domain); } /** Flushes the CPU write domain for the object if it's dirty. */ @@ -2320,13 +2756,35 @@ static void i915_gem_object_flush_cpu_write_domain(struct drm_gem_object *obj) { struct drm_device *dev = obj->dev; + uint32_t old_write_domain; if (obj->write_domain != I915_GEM_DOMAIN_CPU) return; i915_gem_clflush_object(obj); drm_agp_chipset_flush(dev); + old_write_domain = obj->write_domain; obj->write_domain = 0; + + trace_i915_gem_object_change_domain(obj, + obj->read_domains, + old_write_domain); +} + +void +i915_gem_object_flush_write_domain(struct drm_gem_object *obj) +{ + switch (obj->write_domain) { + case I915_GEM_DOMAIN_GTT: + i915_gem_object_flush_gtt_write_domain(obj); + break; + case I915_GEM_DOMAIN_CPU: + i915_gem_object_flush_cpu_write_domain(obj); + break; + default: + i915_gem_object_flush_gpu_write_domain(obj); + break; + } } /** @@ -2339,6 +2797,7 @@ int i915_gem_object_set_to_gtt_domain(struct drm_gem_object *obj, int write) { struct drm_i915_gem_object *obj_priv = obj->driver_private; + uint32_t old_write_domain, old_read_domains; int ret; /* Not valid to be called on unbound objects. */ @@ -2351,6 +2810,9 @@ i915_gem_object_set_to_gtt_domain(struct drm_gem_object *obj, int write) if (ret != 0) return ret; + old_write_domain = obj->write_domain; + old_read_domains = obj->read_domains; + /* If we're writing through the GTT domain, then CPU and GPU caches * will need to be invalidated at next use. */ @@ -2369,6 +2831,10 @@ i915_gem_object_set_to_gtt_domain(struct drm_gem_object *obj, int write) obj_priv->dirty = 1; } + trace_i915_gem_object_change_domain(obj, + old_read_domains, + old_write_domain); + return 0; } @@ -2381,6 +2847,7 @@ i915_gem_object_set_to_gtt_domain(struct drm_gem_object *obj, int write) static int i915_gem_object_set_to_cpu_domain(struct drm_gem_object *obj, int write) { + uint32_t old_write_domain, old_read_domains; int ret; i915_gem_object_flush_gpu_write_domain(obj); @@ -2396,6 +2863,9 @@ i915_gem_object_set_to_cpu_domain(struct drm_gem_object *obj, int write) */ i915_gem_object_set_to_full_cpu_read_domain(obj); + old_write_domain = obj->write_domain; + old_read_domains = obj->read_domains; + /* Flush the CPU cache if it's still invalid. */ if ((obj->read_domains & I915_GEM_DOMAIN_CPU) == 0) { i915_gem_clflush_object(obj); @@ -2416,6 +2886,10 @@ i915_gem_object_set_to_cpu_domain(struct drm_gem_object *obj, int write) obj->write_domain = I915_GEM_DOMAIN_CPU; } + trace_i915_gem_object_change_domain(obj, + old_read_domains, + old_write_domain); + return 0; } @@ -2537,10 +3011,13 @@ i915_gem_object_set_to_gpu_domain(struct drm_gem_object *obj) struct drm_i915_gem_object *obj_priv = obj->driver_private; uint32_t invalidate_domains = 0; uint32_t flush_domains = 0; + uint32_t old_read_domains; BUG_ON(obj->pending_read_domains & I915_GEM_DOMAIN_CPU); BUG_ON(obj->pending_write_domain == I915_GEM_DOMAIN_CPU); + intel_mark_busy(dev, obj); + #if WATCH_BUF DRM_INFO("%s: object %p read %08x -> %08x write %08x -> %08x\n", __func__, obj, @@ -2581,6 +3058,8 @@ i915_gem_object_set_to_gpu_domain(struct drm_gem_object *obj) i915_gem_clflush_object(obj); } + old_read_domains = obj->read_domains; + /* The actual obj->write_domain will be updated with * pending_write_domain after we emit the accumulated flush for all * of our domain changes in execbuffers (which clears objects' @@ -2599,6 +3078,10 @@ i915_gem_object_set_to_gpu_domain(struct drm_gem_object *obj) obj->read_domains, obj->write_domain, dev->invalidate_domains, dev->flush_domains); #endif + + trace_i915_gem_object_change_domain(obj, + old_read_domains, + obj->write_domain); } /** @@ -2630,8 +3113,7 @@ i915_gem_object_set_to_full_cpu_read_domain(struct drm_gem_object *obj) /* Free the page_cpu_valid mappings which are now stale, whether * or not we've got I915_GEM_DOMAIN_CPU. */ - drm_free(obj_priv->page_cpu_valid, obj->size / PAGE_SIZE, - DRM_MEM_DRIVER); + kfree(obj_priv->page_cpu_valid); obj_priv->page_cpu_valid = NULL; } @@ -2652,6 +3134,7 @@ i915_gem_object_set_cpu_read_domain_range(struct drm_gem_object *obj, uint64_t offset, uint64_t size) { struct drm_i915_gem_object *obj_priv = obj->driver_private; + uint32_t old_read_domains; int i, ret; if (offset == 0 && size == obj->size) @@ -2673,8 +3156,8 @@ i915_gem_object_set_cpu_read_domain_range(struct drm_gem_object *obj, * newly adding I915_GEM_DOMAIN_CPU */ if (obj_priv->page_cpu_valid == NULL) { - obj_priv->page_cpu_valid = drm_calloc(1, obj->size / PAGE_SIZE, - DRM_MEM_DRIVER); + obj_priv->page_cpu_valid = kzalloc(obj->size / PAGE_SIZE, + GFP_KERNEL); if (obj_priv->page_cpu_valid == NULL) return -ENOMEM; } else if ((obj->read_domains & I915_GEM_DOMAIN_CPU) == 0) @@ -2698,8 +3181,13 @@ i915_gem_object_set_cpu_read_domain_range(struct drm_gem_object *obj, */ BUG_ON((obj->write_domain & ~I915_GEM_DOMAIN_CPU) != 0); + old_read_domains = obj->read_domains; obj->read_domains |= I915_GEM_DOMAIN_CPU; + trace_i915_gem_object_change_domain(obj, + old_read_domains, + obj->write_domain); + return 0; } @@ -2743,6 +3231,21 @@ i915_gem_object_pin_and_relocate(struct drm_gem_object *obj, } target_obj_priv = target_obj->driver_private; +#if WATCH_RELOC + DRM_INFO("%s: obj %p offset %08x target %d " + "read %08x write %08x gtt %08x " + "presumed %08x delta %08x\n", + __func__, + obj, + (int) reloc->offset, + (int) reloc->target_handle, + (int) reloc->read_domains, + (int) reloc->write_domain, + (int) target_obj_priv->gtt_offset, + (int) reloc->presumed_offset, + reloc->delta); +#endif + /* The target buffer should have appeared before us in the * exec_object list, so it should have a GTT space bound by now. */ @@ -2754,25 +3257,7 @@ i915_gem_object_pin_and_relocate(struct drm_gem_object *obj, return -EINVAL; } - if (reloc->offset > obj->size - 4) { - DRM_ERROR("Relocation beyond object bounds: " - "obj %p target %d offset %d size %d.\n", - obj, reloc->target_handle, - (int) reloc->offset, (int) obj->size); - drm_gem_object_unreference(target_obj); - i915_gem_object_unpin(obj); - return -EINVAL; - } - if (reloc->offset & 3) { - DRM_ERROR("Relocation not 4-byte aligned: " - "obj %p target %d offset %d.\n", - obj, reloc->target_handle, - (int) reloc->offset); - drm_gem_object_unreference(target_obj); - i915_gem_object_unpin(obj); - return -EINVAL; - } - + /* Validate that the target is in a valid r/w GPU domain */ if (reloc->write_domain & I915_GEM_DOMAIN_CPU || reloc->read_domains & I915_GEM_DOMAIN_CPU) { DRM_ERROR("reloc with read/write CPU domains: " @@ -2786,7 +3271,6 @@ i915_gem_object_pin_and_relocate(struct drm_gem_object *obj, i915_gem_object_unpin(obj); return -EINVAL; } - if (reloc->write_domain && target_obj->pending_write_domain && reloc->write_domain != target_obj->pending_write_domain) { DRM_ERROR("Write domain conflict: " @@ -2801,21 +3285,6 @@ i915_gem_object_pin_and_relocate(struct drm_gem_object *obj, return -EINVAL; } -#if WATCH_RELOC - DRM_INFO("%s: obj %p offset %08x target %d " - "read %08x write %08x gtt %08x " - "presumed %08x delta %08x\n", - __func__, - obj, - (int) reloc->offset, - (int) reloc->target_handle, - (int) reloc->read_domains, - (int) reloc->write_domain, - (int) target_obj_priv->gtt_offset, - (int) reloc->presumed_offset, - reloc->delta); -#endif - target_obj->pending_read_domains |= reloc->read_domains; target_obj->pending_write_domain |= reloc->write_domain; @@ -2827,6 +3296,37 @@ i915_gem_object_pin_and_relocate(struct drm_gem_object *obj, continue; } + /* Check that the relocation address is valid... */ + if (reloc->offset > obj->size - 4) { + DRM_ERROR("Relocation beyond object bounds: " + "obj %p target %d offset %d size %d.\n", + obj, reloc->target_handle, + (int) reloc->offset, (int) obj->size); + drm_gem_object_unreference(target_obj); + i915_gem_object_unpin(obj); + return -EINVAL; + } + if (reloc->offset & 3) { + DRM_ERROR("Relocation not 4-byte aligned: " + "obj %p target %d offset %d.\n", + obj, reloc->target_handle, + (int) reloc->offset); + drm_gem_object_unreference(target_obj); + i915_gem_object_unpin(obj); + return -EINVAL; + } + + /* and points to somewhere within the target object. */ + if (reloc->delta >= target_obj->size) { + DRM_ERROR("Relocation beyond target object bounds: " + "obj %p target %d delta %d size %d.\n", + obj, reloc->target_handle, + (int) reloc->delta, (int) target_obj->size); + drm_gem_object_unreference(target_obj); + i915_gem_object_unpin(obj); + return -EINVAL; + } + ret = i915_gem_object_set_to_gtt_domain(obj, 1); if (ret != 0) { drm_gem_object_unreference(target_obj); @@ -2879,19 +3379,13 @@ i915_dispatch_gem_execbuffer(struct drm_device *dev, drm_i915_private_t *dev_priv = dev->dev_private; int nbox = exec->num_cliprects; int i = 0, count; - uint32_t exec_start, exec_len; + uint32_t exec_start, exec_len; RING_LOCALS; exec_start = (uint32_t) exec_offset + exec->batch_start_offset; exec_len = (uint32_t) exec->batch_len; - if ((exec_start | exec_len) & 0x7) { - DRM_ERROR("alignment\n"); - return -EINVAL; - } - - if (!exec_start) - return -EINVAL; + trace_i915_gem_request_submit(dev, dev_priv->mm.next_gem_seqno + 1); count = nbox ? nbox : 1; @@ -2933,6 +3427,10 @@ i915_dispatch_gem_execbuffer(struct drm_device *dev, /* Throttle our rendering by waiting until the ring has completed our requests * emitted over 20 msec ago. * + * Note that if we were to use the current jiffies each time around the loop, + * we wouldn't escape the function with any frames outstanding if the time to + * render a frame was over 20ms. + * * This should get us reasonable parallelism between CPU and GPU but also * relatively low latency when blocking on a particular request to finish. */ @@ -2941,15 +3439,25 @@ i915_gem_ring_throttle(struct drm_device *dev, struct drm_file *file_priv) { struct drm_i915_file_private *i915_file_priv = file_priv->driver_priv; int ret = 0; - uint32_t seqno; + unsigned long recent_enough = jiffies - msecs_to_jiffies(20); mutex_lock(&dev->struct_mutex); - seqno = i915_file_priv->mm.last_gem_throttle_seqno; - i915_file_priv->mm.last_gem_throttle_seqno = - i915_file_priv->mm.last_gem_seqno; - if (seqno) - ret = i915_wait_request(dev, seqno); + while (!list_empty(&i915_file_priv->mm.request_list)) { + struct drm_i915_gem_request *request; + + request = list_first_entry(&i915_file_priv->mm.request_list, + struct drm_i915_gem_request, + client_list); + + if (time_after_eq(request->emitted_jiffies, recent_enough)) + break; + + ret = i915_wait_request(dev, request->seqno); + if (ret != 0) + break; + } mutex_unlock(&dev->struct_mutex); + return ret; } @@ -2968,7 +3476,7 @@ i915_gem_get_relocs_from_user(struct drm_i915_gem_exec_object *exec_list, reloc_count += exec_list[i].relocation_count; } - *relocs = drm_calloc(reloc_count, sizeof(**relocs), DRM_MEM_DRIVER); + *relocs = drm_calloc_large(reloc_count, sizeof(**relocs)); if (*relocs == NULL) return -ENOMEM; @@ -2982,16 +3490,15 @@ i915_gem_get_relocs_from_user(struct drm_i915_gem_exec_object *exec_list, exec_list[i].relocation_count * sizeof(**relocs)); if (ret != 0) { - drm_free(*relocs, reloc_count * sizeof(**relocs), - DRM_MEM_DRIVER); + drm_free_large(*relocs); *relocs = NULL; - return ret; + return -EFAULT; } reloc_index += exec_list[i].relocation_count; } - return ret; + return 0; } static int @@ -3000,24 +3507,82 @@ i915_gem_put_relocs_to_user(struct drm_i915_gem_exec_object *exec_list, struct drm_i915_gem_relocation_entry *relocs) { uint32_t reloc_count = 0, i; - int ret; + int ret = 0; for (i = 0; i < buffer_count; i++) { struct drm_i915_gem_relocation_entry __user *user_relocs; + int unwritten; user_relocs = (void __user *)(uintptr_t)exec_list[i].relocs_ptr; - if (ret == 0) { - ret = copy_to_user(user_relocs, - &relocs[reloc_count], - exec_list[i].relocation_count * - sizeof(*relocs)); + unwritten = copy_to_user(user_relocs, + &relocs[reloc_count], + exec_list[i].relocation_count * + sizeof(*relocs)); + + if (unwritten) { + ret = -EFAULT; + goto err; } reloc_count += exec_list[i].relocation_count; } - drm_free(relocs, reloc_count * sizeof(*relocs), DRM_MEM_DRIVER); +err: + drm_free_large(relocs); + + return ret; +} + +static int +i915_gem_check_execbuffer (struct drm_i915_gem_execbuffer *exec, + uint64_t exec_offset) +{ + uint32_t exec_start, exec_len; + + exec_start = (uint32_t) exec_offset + exec->batch_start_offset; + exec_len = (uint32_t) exec->batch_len; + + if ((exec_start | exec_len) & 0x7) + return -EINVAL; + + if (!exec_start) + return -EINVAL; + + return 0; +} + +static int +i915_gem_wait_for_pending_flip(struct drm_device *dev, + struct drm_gem_object **object_list, + int count) +{ + drm_i915_private_t *dev_priv = dev->dev_private; + struct drm_i915_gem_object *obj_priv; + DEFINE_WAIT(wait); + int i, ret = 0; + + for (;;) { + prepare_to_wait(&dev_priv->pending_flip_queue, + &wait, TASK_INTERRUPTIBLE); + for (i = 0; i < count; i++) { + obj_priv = object_list[i]->driver_private; + if (atomic_read(&obj_priv->pending_flip) > 0) + break; + } + if (i == count) + break; + + if (!signal_pending(current)) { + mutex_unlock(&dev->struct_mutex); + schedule(); + mutex_lock(&dev->struct_mutex); + continue; + } + ret = -ERESTARTSYS; + break; + } + finish_wait(&dev_priv->pending_flip_queue, &wait); return ret; } @@ -3027,7 +3592,6 @@ i915_gem_execbuffer(struct drm_device *dev, void *data, struct drm_file *file_priv) { drm_i915_private_t *dev_priv = dev->dev_private; - struct drm_i915_file_private *i915_file_priv = file_priv->driver_priv; struct drm_i915_gem_execbuffer *args = data; struct drm_i915_gem_exec_object *exec_list = NULL; struct drm_gem_object **object_list = NULL; @@ -3038,7 +3602,7 @@ i915_gem_execbuffer(struct drm_device *dev, void *data, int ret, ret2, i, pinned = 0; uint64_t exec_offset; uint32_t seqno, flush_domains, reloc_index; - int pin_tries; + int pin_tries, flips; #if WATCH_EXEC DRM_INFO("buffers_ptr %d buffer_count %d len %08x\n", @@ -3050,10 +3614,8 @@ i915_gem_execbuffer(struct drm_device *dev, void *data, return -EINVAL; } /* Copy in the exec list from userland */ - exec_list = drm_calloc(sizeof(*exec_list), args->buffer_count, - DRM_MEM_DRIVER); - object_list = drm_calloc(sizeof(*object_list), args->buffer_count, - DRM_MEM_DRIVER); + exec_list = drm_malloc_ab(sizeof(*exec_list), args->buffer_count); + object_list = drm_malloc_ab(sizeof(*object_list), args->buffer_count); if (exec_list == NULL || object_list == NULL) { DRM_ERROR("Failed to allocate exec or object list " "for %d buffers\n", @@ -3072,8 +3634,8 @@ i915_gem_execbuffer(struct drm_device *dev, void *data, } if (args->num_cliprects != 0) { - cliprects = drm_calloc(args->num_cliprects, sizeof(*cliprects), - DRM_MEM_DRIVER); + cliprects = kcalloc(args->num_cliprects, sizeof(*cliprects), + GFP_KERNEL); if (cliprects == NULL) goto pre_mutex_err; @@ -3097,21 +3659,20 @@ i915_gem_execbuffer(struct drm_device *dev, void *data, i915_verify_inactive(dev, __FILE__, __LINE__); - if (dev_priv->mm.wedged) { - DRM_ERROR("Execbuf while wedged\n"); + if (atomic_read(&dev_priv->mm.wedged)) { mutex_unlock(&dev->struct_mutex); ret = -EIO; goto pre_mutex_err; } if (dev_priv->mm.suspended) { - DRM_ERROR("Execbuf while VT-switched.\n"); mutex_unlock(&dev->struct_mutex); ret = -EBUSY; goto pre_mutex_err; } /* Look up object handles */ + flips = 0; for (i = 0; i < args->buffer_count; i++) { object_list[i] = drm_gem_object_lookup(dev, file_priv, exec_list[i].handle); @@ -3130,6 +3691,14 @@ i915_gem_execbuffer(struct drm_device *dev, void *data, goto err; } obj_priv->in_execbuffer = true; + flips += atomic_read(&obj_priv->pending_flip); + } + + if (flips > 0) { + ret = i915_gem_wait_for_pending_flip(dev, object_list, + args->buffer_count); + if (ret) + goto err; } /* Pin and relocate */ @@ -3154,9 +3723,24 @@ i915_gem_execbuffer(struct drm_device *dev, void *data, break; /* error other than GTT full, or we've already tried again */ - if (ret != -ENOMEM || pin_tries >= 1) { - if (ret != -ERESTARTSYS) - DRM_ERROR("Failed to pin buffers %d\n", ret); + if (ret != -ENOSPC || pin_tries >= 1) { + if (ret != -ERESTARTSYS) { + unsigned long long total_size = 0; + for (i = 0; i < args->buffer_count; i++) + total_size += object_list[i]->size; + DRM_ERROR("Failed to pin buffer %d of %d, total %llu bytes: %d\n", + pinned+1, args->buffer_count, + total_size, ret); + DRM_ERROR("%d objects [%d pinned], " + "%d object bytes [%d pinned], " + "%d/%d gtt bytes\n", + atomic_read(&dev->object_count), + atomic_read(&dev->pin_count), + atomic_read(&dev->object_memory), + atomic_read(&dev->pin_memory), + atomic_read(&dev->gtt_memory), + dev->gtt_total); + } goto err; } @@ -3167,14 +3751,26 @@ i915_gem_execbuffer(struct drm_device *dev, void *data, /* evict everyone we can from the aperture */ ret = i915_gem_evict_everything(dev); - if (ret) + if (ret && ret != -ENOSPC) goto err; } /* Set the pending read domains for the batch buffer to COMMAND */ batch_obj = object_list[args->buffer_count-1]; - batch_obj->pending_read_domains = I915_GEM_DOMAIN_COMMAND; - batch_obj->pending_write_domain = 0; + if (batch_obj->pending_write_domain) { + DRM_ERROR("Attempting to use self-modifying batch buffer\n"); + ret = -EINVAL; + goto err; + } + batch_obj->pending_read_domains |= I915_GEM_DOMAIN_COMMAND; + + /* Sanity check the batch buffer, prior to moving objects */ + exec_offset = exec_list[args->buffer_count - 1].offset; + ret = i915_gem_check_execbuffer (args, exec_offset); + if (ret != 0) { + DRM_ERROR("execbuf with invalid offset/length\n"); + goto err; + } i915_verify_inactive(dev, __FILE__, __LINE__); @@ -3205,13 +3801,18 @@ i915_gem_execbuffer(struct drm_device *dev, void *data, dev->invalidate_domains, dev->flush_domains); if (dev->flush_domains) - (void)i915_add_request(dev, dev->flush_domains); + (void)i915_add_request(dev, file_priv, + dev->flush_domains); } for (i = 0; i < args->buffer_count; i++) { struct drm_gem_object *obj = object_list[i]; + uint32_t old_write_domain = obj->write_domain; obj->write_domain = obj->pending_write_domain; + trace_i915_gem_object_change_domain(obj, + obj->read_domains, + old_write_domain); } i915_verify_inactive(dev, __FILE__, __LINE__); @@ -3223,10 +3824,8 @@ i915_gem_execbuffer(struct drm_device *dev, void *data, } #endif - exec_offset = exec_list[args->buffer_count - 1].offset; - #if WATCH_EXEC - i915_gem_dump_object(object_list[args->buffer_count - 1], + i915_gem_dump_object(batch_obj, args->batch_len, __func__, ~0); @@ -3254,9 +3853,8 @@ i915_gem_execbuffer(struct drm_device *dev, void *data, * *some* interrupts representing completion of buffers that we can * wait on when trying to clear up gtt space). */ - seqno = i915_add_request(dev, flush_domains); + seqno = i915_add_request(dev, file_priv, flush_domains); BUG_ON(seqno == 0); - i915_file_priv->mm.last_gem_seqno = seqno; for (i = 0; i < args->buffer_count; i++) { struct drm_gem_object *obj = object_list[i]; @@ -3291,10 +3889,12 @@ err: (uintptr_t) args->buffers_ptr, exec_list, sizeof(*exec_list) * args->buffer_count); - if (ret) + if (ret) { + ret = -EFAULT; DRM_ERROR("failed to copy %d exec entries " "back to user (%d)\n", args->buffer_count, ret); + } } /* Copy the updated relocations out regardless of current error @@ -3312,12 +3912,9 @@ err: } pre_mutex_err: - drm_free(object_list, sizeof(*object_list) * args->buffer_count, - DRM_MEM_DRIVER); - drm_free(exec_list, sizeof(*exec_list) * args->buffer_count, - DRM_MEM_DRIVER); - drm_free(cliprects, sizeof(*cliprects) * args->num_cliprects, - DRM_MEM_DRIVER); + drm_free_large(object_list); + drm_free_large(exec_list); + kfree(cliprects); return ret; } @@ -3332,20 +3929,15 @@ i915_gem_object_pin(struct drm_gem_object *obj, uint32_t alignment) i915_verify_inactive(dev, __FILE__, __LINE__); if (obj_priv->gtt_space == NULL) { ret = i915_gem_object_bind_to_gtt(obj, alignment); - if (ret != 0) { - if (ret != -EBUSY && ret != -ERESTARTSYS) - DRM_ERROR("Failure to bind: %d\n", ret); + if (ret) return ret; - } } /* * Pre-965 chips need a fence register set up in order to * properly handle tiled surfaces. */ - if (!IS_I965G(dev) && - obj_priv->fence_reg == I915_FENCE_REG_NONE && - obj_priv->tiling_mode != I915_TILING_NONE) { - ret = i915_gem_object_get_fence_reg(obj, true); + if (!IS_I965G(dev) && obj_priv->tiling_mode != I915_TILING_NONE) { + ret = i915_gem_object_get_fence_reg(obj); if (ret != 0) { if (ret != -EBUSY && ret != -ERESTARTSYS) DRM_ERROR("Failure to install fence: %d\n", @@ -3362,8 +3954,7 @@ i915_gem_object_pin(struct drm_gem_object *obj, uint32_t alignment) atomic_inc(&dev->pin_count); atomic_add(obj->size, &dev->pin_memory); if (!obj_priv->active && - (obj->write_domain & ~(I915_GEM_DOMAIN_CPU | - I915_GEM_DOMAIN_GTT)) == 0 && + (obj->write_domain & I915_GEM_GPU_DOMAINS) == 0 && !list_empty(&obj_priv->list)) list_del_init(&obj_priv->list); } @@ -3390,8 +3981,7 @@ i915_gem_object_unpin(struct drm_gem_object *obj) */ if (obj_priv->pin_count == 0) { if (!obj_priv->active && - (obj->write_domain & ~(I915_GEM_DOMAIN_CPU | - I915_GEM_DOMAIN_GTT)) == 0) + (obj->write_domain & I915_GEM_GPU_DOMAINS) == 0) list_move_tail(&obj_priv->list, &dev_priv->mm.inactive_list); atomic_dec(&dev->pin_count); @@ -3420,6 +4010,13 @@ i915_gem_pin_ioctl(struct drm_device *dev, void *data, } obj_priv = obj->driver_private; + if (obj_priv->madv != I915_MADV_WILLNEED) { + DRM_ERROR("Attempting to pin a purgeable buffer\n"); + drm_gem_object_unreference(obj); + mutex_unlock(&dev->struct_mutex); + return -EINVAL; + } + if (obj_priv->pin_filp != NULL && obj_priv->pin_filp != file_priv) { DRM_ERROR("Already pinned in i915_gem_pin_ioctl(): %d\n", args->handle); @@ -3495,15 +4092,14 @@ i915_gem_busy_ioctl(struct drm_device *dev, void *data, struct drm_gem_object *obj; struct drm_i915_gem_object *obj_priv; - mutex_lock(&dev->struct_mutex); obj = drm_gem_object_lookup(dev, file_priv, args->handle); if (obj == NULL) { DRM_ERROR("Bad handle in i915_gem_busy_ioctl(): %d\n", args->handle); - mutex_unlock(&dev->struct_mutex); return -EBADF; } + mutex_lock(&dev->struct_mutex); /* Update the active list for the hardware's current position. * Otherwise this only updates on a delayed timer or when irqs are * actually unmasked, and our working set ends up being larger than @@ -3533,11 +4129,61 @@ i915_gem_throttle_ioctl(struct drm_device *dev, void *data, return i915_gem_ring_throttle(dev, file_priv); } +int +i915_gem_madvise_ioctl(struct drm_device *dev, void *data, + struct drm_file *file_priv) +{ + struct drm_i915_gem_madvise *args = data; + struct drm_gem_object *obj; + struct drm_i915_gem_object *obj_priv; + + switch (args->madv) { + case I915_MADV_DONTNEED: + case I915_MADV_WILLNEED: + break; + default: + return -EINVAL; + } + + obj = drm_gem_object_lookup(dev, file_priv, args->handle); + if (obj == NULL) { + DRM_ERROR("Bad handle in i915_gem_madvise_ioctl(): %d\n", + args->handle); + return -EBADF; + } + + mutex_lock(&dev->struct_mutex); + obj_priv = obj->driver_private; + + if (obj_priv->pin_count) { + drm_gem_object_unreference(obj); + mutex_unlock(&dev->struct_mutex); + + DRM_ERROR("Attempted i915_gem_madvise_ioctl() on a pinned object\n"); + return -EINVAL; + } + + if (obj_priv->madv != __I915_MADV_PURGED) + obj_priv->madv = args->madv; + + /* if the object is no longer bound, discard its backing storage */ + if (i915_gem_object_is_purgeable(obj_priv) && + obj_priv->gtt_space == NULL) + i915_gem_object_truncate(obj); + + args->retained = obj_priv->madv != __I915_MADV_PURGED; + + drm_gem_object_unreference(obj); + mutex_unlock(&dev->struct_mutex); + + return 0; +} + int i915_gem_init_object(struct drm_gem_object *obj) { struct drm_i915_gem_object *obj_priv; - obj_priv = drm_calloc(1, sizeof(*obj_priv), DRM_MEM_DRIVER); + obj_priv = kzalloc(sizeof(*obj_priv), GFP_KERNEL); if (obj_priv == NULL) return -ENOMEM; @@ -3556,6 +4202,10 @@ int i915_gem_init_object(struct drm_gem_object *obj) obj_priv->obj = obj; obj_priv->fence_reg = I915_FENCE_REG_NONE; INIT_LIST_HEAD(&obj_priv->list); + INIT_LIST_HEAD(&obj_priv->fence_list); + obj_priv->madv = I915_MADV_WILLNEED; + + trace_i915_gem_object_create(obj); return 0; } @@ -3565,6 +4215,8 @@ void i915_gem_free_object(struct drm_gem_object *obj) struct drm_device *dev = obj->dev; struct drm_i915_gem_object *obj_priv = obj->driver_private; + trace_i915_gem_object_destroy(obj); + while (obj_priv->pin_count > 0) i915_gem_object_unpin(obj); @@ -3573,42 +4225,35 @@ void i915_gem_free_object(struct drm_gem_object *obj) i915_gem_object_unbind(obj); - i915_gem_free_mmap_offset(obj); + if (obj_priv->mmap_offset) + i915_gem_free_mmap_offset(obj); - drm_free(obj_priv->page_cpu_valid, 1, DRM_MEM_DRIVER); - drm_free(obj->driver_private, 1, DRM_MEM_DRIVER); + kfree(obj_priv->page_cpu_valid); + kfree(obj_priv->bit_17); + kfree(obj->driver_private); } -/** Unbinds all objects that are on the given buffer list. */ +/** Unbinds all inactive objects. */ static int -i915_gem_evict_from_list(struct drm_device *dev, struct list_head *head) +i915_gem_evict_from_inactive_list(struct drm_device *dev) { - struct drm_gem_object *obj; - struct drm_i915_gem_object *obj_priv; - int ret; + drm_i915_private_t *dev_priv = dev->dev_private; - while (!list_empty(head)) { - obj_priv = list_first_entry(head, - struct drm_i915_gem_object, - list); - obj = obj_priv->obj; + while (!list_empty(&dev_priv->mm.inactive_list)) { + struct drm_gem_object *obj; + int ret; - if (obj_priv->pin_count != 0) { - DRM_ERROR("Pinned object in unbind list\n"); - mutex_unlock(&dev->struct_mutex); - return -EINVAL; - } + obj = list_first_entry(&dev_priv->mm.inactive_list, + struct drm_i915_gem_object, + list)->obj; ret = i915_gem_object_unbind(obj); if (ret != 0) { - DRM_ERROR("Error unbinding object in LeaveVT: %d\n", - ret); - mutex_unlock(&dev->struct_mutex); + DRM_ERROR("Error unbinding object: %d\n", ret); return ret; } } - return 0; } @@ -3630,6 +4275,7 @@ i915_gem_idle(struct drm_device *dev) * We need to replace this with a semaphore, or something. */ dev_priv->mm.suspended = 1; + del_timer(&dev_priv->hangcheck_timer); /* Cancel the retire work handler, wait for it to finish if running */ @@ -3641,9 +4287,8 @@ i915_gem_idle(struct drm_device *dev) /* Flush the GPU along with all non-CPU write domains */ - i915_gem_flush(dev, ~(I915_GEM_DOMAIN_CPU|I915_GEM_DOMAIN_GTT), - ~(I915_GEM_DOMAIN_CPU|I915_GEM_DOMAIN_GTT)); - seqno = i915_add_request(dev, ~I915_GEM_DOMAIN_CPU); + i915_gem_flush(dev, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS); + seqno = i915_add_request(dev, NULL, I915_GEM_GPU_DOMAINS); if (seqno == 0) { mutex_unlock(&dev->struct_mutex); @@ -3660,7 +4305,7 @@ i915_gem_idle(struct drm_device *dev) if (last_seqno == cur_seqno) { if (stuck++ > 100) { DRM_ERROR("hardware wedged\n"); - dev_priv->mm.wedged = 1; + atomic_set(&dev_priv->mm.wedged, 1); DRM_WAKEUP(&dev_priv->irq_queue); break; } @@ -3672,7 +4317,8 @@ i915_gem_idle(struct drm_device *dev) i915_gem_retire_requests(dev); - if (!dev_priv->mm.wedged) { + spin_lock(&dev_priv->mm.active_list_lock); + if (!atomic_read(&dev_priv->mm.wedged)) { /* Active and flushing should now be empty as we've * waited for a sequence higher than any pending execbuffer */ @@ -3690,28 +4336,41 @@ i915_gem_idle(struct drm_device *dev) * the GPU domains and just stuff them onto inactive. */ while (!list_empty(&dev_priv->mm.active_list)) { - struct drm_i915_gem_object *obj_priv; - - obj_priv = list_first_entry(&dev_priv->mm.active_list, - struct drm_i915_gem_object, - list); - obj_priv->obj->write_domain &= ~I915_GEM_GPU_DOMAINS; - i915_gem_object_move_to_inactive(obj_priv->obj); + struct drm_gem_object *obj; + uint32_t old_write_domain; + + obj = list_first_entry(&dev_priv->mm.active_list, + struct drm_i915_gem_object, + list)->obj; + old_write_domain = obj->write_domain; + obj->write_domain &= ~I915_GEM_GPU_DOMAINS; + i915_gem_object_move_to_inactive(obj); + + trace_i915_gem_object_change_domain(obj, + obj->read_domains, + old_write_domain); } + spin_unlock(&dev_priv->mm.active_list_lock); while (!list_empty(&dev_priv->mm.flushing_list)) { - struct drm_i915_gem_object *obj_priv; - - obj_priv = list_first_entry(&dev_priv->mm.flushing_list, - struct drm_i915_gem_object, - list); - obj_priv->obj->write_domain &= ~I915_GEM_GPU_DOMAINS; - i915_gem_object_move_to_inactive(obj_priv->obj); + struct drm_gem_object *obj; + uint32_t old_write_domain; + + obj = list_first_entry(&dev_priv->mm.flushing_list, + struct drm_i915_gem_object, + list)->obj; + old_write_domain = obj->write_domain; + obj->write_domain &= ~I915_GEM_GPU_DOMAINS; + i915_gem_object_move_to_inactive(obj); + + trace_i915_gem_object_change_domain(obj, + obj->read_domains, + old_write_domain); } /* Move all inactive buffers out of the GTT. */ - ret = i915_gem_evict_from_list(dev, &dev_priv->mm.inactive_list); + ret = i915_gem_evict_from_inactive_list(dev); WARN_ON(!list_empty(&dev_priv->mm.inactive_list)); if (ret) { mutex_unlock(&dev->struct_mutex); @@ -3766,7 +4425,7 @@ i915_gem_init_hws(struct drm_device *dev) memset(dev_priv->hw_status_page, 0, PAGE_SIZE); I915_WRITE(HWS_PGA, dev_priv->status_gfx_addr); I915_READ(HWS_PGA); /* posting read */ - DRM_DEBUG("hws offset: 0x%08x\n", dev_priv->status_gfx_addr); + DRM_DEBUG_DRIVER("hws offset: 0x%08x\n", dev_priv->status_gfx_addr); return 0; } @@ -3827,7 +4486,6 @@ i915_gem_init_ringbuffer(struct drm_device *dev) /* Set up the kernel mapping for the ring. */ ring->Size = obj->size; - ring->tail_mask = obj->size - 1; ring->map.offset = dev->agp->base + obj_priv->gtt_offset; ring->map.size = obj->size; @@ -3934,19 +4592,24 @@ i915_gem_entervt_ioctl(struct drm_device *dev, void *data, if (drm_core_check_feature(dev, DRIVER_MODESET)) return 0; - if (dev_priv->mm.wedged) { + if (atomic_read(&dev_priv->mm.wedged)) { DRM_ERROR("Reenabling wedged hardware, good luck\n"); - dev_priv->mm.wedged = 0; + atomic_set(&dev_priv->mm.wedged, 0); } mutex_lock(&dev->struct_mutex); dev_priv->mm.suspended = 0; ret = i915_gem_init_ringbuffer(dev); - if (ret != 0) + if (ret != 0) { + mutex_unlock(&dev->struct_mutex); return ret; + } + spin_lock(&dev_priv->mm.active_list_lock); BUG_ON(!list_empty(&dev_priv->mm.active_list)); + spin_unlock(&dev_priv->mm.active_list_lock); + BUG_ON(!list_empty(&dev_priv->mm.flushing_list)); BUG_ON(!list_empty(&dev_priv->mm.inactive_list)); BUG_ON(!list_empty(&dev_priv->mm.request_list)); @@ -3961,15 +4624,11 @@ int i915_gem_leavevt_ioctl(struct drm_device *dev, void *data, struct drm_file *file_priv) { - int ret; - if (drm_core_check_feature(dev, DRIVER_MODESET)) return 0; - ret = i915_gem_idle(dev); drm_irq_uninstall(dev); - - return ret; + return i915_gem_idle(dev); } void @@ -3988,16 +4647,23 @@ i915_gem_lastclose(struct drm_device *dev) void i915_gem_load(struct drm_device *dev) { + int i; drm_i915_private_t *dev_priv = dev->dev_private; + spin_lock_init(&dev_priv->mm.active_list_lock); INIT_LIST_HEAD(&dev_priv->mm.active_list); INIT_LIST_HEAD(&dev_priv->mm.flushing_list); INIT_LIST_HEAD(&dev_priv->mm.inactive_list); INIT_LIST_HEAD(&dev_priv->mm.request_list); + INIT_LIST_HEAD(&dev_priv->mm.fence_list); INIT_DELAYED_WORK(&dev_priv->mm.retire_work, i915_gem_retire_work_handler); dev_priv->mm.next_gem_seqno = 1; + spin_lock(&shrink_list_lock); + list_add(&dev_priv->mm.shrink_list, &shrink_list); + spin_unlock(&shrink_list_lock); + /* Old X drivers will take 0-2 for front, back, depth buffers */ dev_priv->fence_reg_start = 3; @@ -4006,7 +4672,19 @@ i915_gem_load(struct drm_device *dev) else dev_priv->num_fence_regs = 8; + /* Initialize fence registers to zero */ + if (IS_I965G(dev)) { + for (i = 0; i < 16; i++) + I915_WRITE64(FENCE_REG_965_0 + (i * 8), 0); + } else { + for (i = 0; i < 8; i++) + I915_WRITE(FENCE_REG_830_0 + (i * 4), 0); + if (IS_I945G(dev) || IS_I945GM(dev) || IS_G33(dev)) + for (i = 0; i < 8; i++) + I915_WRITE(FENCE_REG_945_8 + (i * 4), 0); + } i915_gem_detect_bit_6_swizzle(dev); + init_waitqueue_head(&dev_priv->pending_flip_queue); } /* @@ -4023,7 +4701,7 @@ int i915_gem_init_phys_object(struct drm_device *dev, if (dev_priv->mm.phys_objs[id - 1] || !size) return 0; - phys_obj = drm_calloc(1, sizeof(struct drm_i915_gem_phys_object), DRM_MEM_DRIVER); + phys_obj = kzalloc(sizeof(struct drm_i915_gem_phys_object), GFP_KERNEL); if (!phys_obj) return -ENOMEM; @@ -4042,7 +4720,7 @@ int i915_gem_init_phys_object(struct drm_device *dev, return 0; kfree_obj: - drm_free(phys_obj, sizeof(struct drm_i915_gem_phys_object), DRM_MEM_DRIVER); + kfree(phys_obj); return ret; } @@ -4102,6 +4780,8 @@ void i915_gem_detach_phys_object(struct drm_device *dev, } drm_clflush_pages(obj_priv->pages, page_count); drm_agp_chipset_flush(dev); + + i915_gem_object_put_pages(obj); out: obj_priv->phys_obj->cur_obj = NULL; obj_priv->phys_obj = NULL; @@ -4159,6 +4839,8 @@ i915_gem_attach_phys_object(struct drm_device *dev, kunmap_atomic(src, KM_USER0); } + i915_gem_object_put_pages(obj); + return 0; out: return ret; @@ -4177,7 +4859,7 @@ i915_gem_phys_pwrite(struct drm_device *dev, struct drm_gem_object *obj, user_data = (char __user *) (uintptr_t) args->data_ptr; obj_addr = obj_priv->phys_obj->handle->vaddr + args->offset; - DRM_DEBUG("obj_addr %p, %lld\n", obj_addr, args->size); + DRM_DEBUG_DRIVER("obj_addr %p, %lld\n", obj_addr, args->size); ret = copy_from_user(obj_addr, user_data, args->size); if (ret) return -EFAULT; @@ -4185,3 +4867,130 @@ i915_gem_phys_pwrite(struct drm_device *dev, struct drm_gem_object *obj, drm_agp_chipset_flush(dev); return 0; } + +void i915_gem_release(struct drm_device * dev, struct drm_file *file_priv) +{ + struct drm_i915_file_private *i915_file_priv = file_priv->driver_priv; + + /* Clean up our request list when the client is going away, so that + * later retire_requests won't dereference our soon-to-be-gone + * file_priv. + */ + mutex_lock(&dev->struct_mutex); + while (!list_empty(&i915_file_priv->mm.request_list)) + list_del_init(i915_file_priv->mm.request_list.next); + mutex_unlock(&dev->struct_mutex); +} + +static int +i915_gem_shrink(int nr_to_scan, gfp_t gfp_mask) +{ + drm_i915_private_t *dev_priv, *next_dev; + struct drm_i915_gem_object *obj_priv, *next_obj; + int cnt = 0; + int would_deadlock = 1; + + /* "fast-path" to count number of available objects */ + if (nr_to_scan == 0) { + spin_lock(&shrink_list_lock); + list_for_each_entry(dev_priv, &shrink_list, mm.shrink_list) { + struct drm_device *dev = dev_priv->dev; + + if (mutex_trylock(&dev->struct_mutex)) { + list_for_each_entry(obj_priv, + &dev_priv->mm.inactive_list, + list) + cnt++; + mutex_unlock(&dev->struct_mutex); + } + } + spin_unlock(&shrink_list_lock); + + return (cnt / 100) * sysctl_vfs_cache_pressure; + } + + spin_lock(&shrink_list_lock); + + /* first scan for clean buffers */ + list_for_each_entry_safe(dev_priv, next_dev, + &shrink_list, mm.shrink_list) { + struct drm_device *dev = dev_priv->dev; + + if (! mutex_trylock(&dev->struct_mutex)) + continue; + + spin_unlock(&shrink_list_lock); + + i915_gem_retire_requests(dev); + + list_for_each_entry_safe(obj_priv, next_obj, + &dev_priv->mm.inactive_list, + list) { + if (i915_gem_object_is_purgeable(obj_priv)) { + i915_gem_object_unbind(obj_priv->obj); + if (--nr_to_scan <= 0) + break; + } + } + + spin_lock(&shrink_list_lock); + mutex_unlock(&dev->struct_mutex); + + would_deadlock = 0; + + if (nr_to_scan <= 0) + break; + } + + /* second pass, evict/count anything still on the inactive list */ + list_for_each_entry_safe(dev_priv, next_dev, + &shrink_list, mm.shrink_list) { + struct drm_device *dev = dev_priv->dev; + + if (! mutex_trylock(&dev->struct_mutex)) + continue; + + spin_unlock(&shrink_list_lock); + + list_for_each_entry_safe(obj_priv, next_obj, + &dev_priv->mm.inactive_list, + list) { + if (nr_to_scan > 0) { + i915_gem_object_unbind(obj_priv->obj); + nr_to_scan--; + } else + cnt++; + } + + spin_lock(&shrink_list_lock); + mutex_unlock(&dev->struct_mutex); + + would_deadlock = 0; + } + + spin_unlock(&shrink_list_lock); + + if (would_deadlock) + return -1; + else if (cnt > 0) + return (cnt / 100) * sysctl_vfs_cache_pressure; + else + return 0; +} + +static struct shrinker shrinker = { + .shrink = i915_gem_shrink, + .seeks = DEFAULT_SEEKS, +}; + +__init void +i915_gem_shrinker_init(void) +{ + register_shrinker(&shrinker); +} + +__exit void +i915_gem_shrinker_exit(void) +{ + unregister_shrinker(&shrinker); +}