This gives the mm subsystem the ability to increase fault handling performance by proposing the insertion of a range of pages around the faulty address in a single batch.
v4: - implement map_pages instead of huge_fault v5: - improve patch series progression - use dma_resv_trylock() in map_pages (many thanks to Matthew Wilcox) - validate map_pages range based on end_pgoff instead of start_pgoff Signed-off-by: Loïc Molinari <[email protected]> --- drivers/gpu/drm/drm_gem_shmem_helper.c | 72 ++++++++++++++++++++++---- 1 file changed, 62 insertions(+), 10 deletions(-) diff --git a/drivers/gpu/drm/drm_gem_shmem_helper.c b/drivers/gpu/drm/drm_gem_shmem_helper.c index be89be1c804c..2a9fbc9c3712 100644 --- a/drivers/gpu/drm/drm_gem_shmem_helper.c +++ b/drivers/gpu/drm/drm_gem_shmem_helper.c @@ -567,31 +567,82 @@ int drm_gem_shmem_dumb_create(struct drm_file *file, struct drm_device *dev, } EXPORT_SYMBOL_GPL(drm_gem_shmem_dumb_create); -static vm_fault_t drm_gem_shmem_fault(struct vm_fault *vmf) +static bool drm_gem_shmem_fault_is_valid(struct drm_gem_object *obj, + pgoff_t pgoff) +{ + struct drm_gem_shmem_object *shmem = to_drm_gem_shmem_obj(obj); + + if (drm_WARN_ON_ONCE(obj->dev, !shmem->pages) || + pgoff >= (obj->size >> PAGE_SHIFT) || + shmem->madv < 0) + return false; + + return true; +} + +static vm_fault_t drm_gem_shmem_map_pages(struct vm_fault *vmf, + pgoff_t start_pgoff, + pgoff_t end_pgoff) { struct vm_area_struct *vma = vmf->vma; struct drm_gem_object *obj = vma->vm_private_data; struct drm_gem_shmem_object *shmem = to_drm_gem_shmem_obj(obj); - loff_t num_pages = obj->size >> PAGE_SHIFT; + struct page **pages = shmem->pages; + unsigned long addr, pfn; vm_fault_t ret; - struct page *page; + + start_pgoff -= vma->vm_pgoff; + end_pgoff -= vma->vm_pgoff; + addr = vma->vm_start + (start_pgoff << PAGE_SHIFT); + + /* map_pages is called with the RCU lock for reading (sleep isn't + * allowed) so just fall through to the more heavy-weight fault path. + */ + if (unlikely(!dma_resv_trylock(shmem->base.resv))) + return 0; + + if (unlikely(!drm_gem_shmem_fault_is_valid(obj, end_pgoff))) { + ret = VM_FAULT_SIGBUS; + goto out; + } + + /* Map a range of pages around the faulty address. */ + do { + pfn = page_to_pfn(pages[start_pgoff]); + ret = vmf_insert_pfn(vma, addr, pfn); + addr += PAGE_SIZE; + } while (++start_pgoff <= end_pgoff && ret == VM_FAULT_NOPAGE); + + out: + dma_resv_unlock(shmem->base.resv); + + return ret; +} + +static vm_fault_t drm_gem_shmem_fault(struct vm_fault *vmf) +{ + struct vm_area_struct *vma = vmf->vma; + struct drm_gem_object *obj = vma->vm_private_data; + struct drm_gem_shmem_object *shmem = to_drm_gem_shmem_obj(obj); + struct page **pages = shmem->pages; pgoff_t page_offset; + unsigned long pfn; + vm_fault_t ret; /* Offset to faulty address in the VMA (without the fake offset). */ page_offset = vmf->pgoff - vma->vm_pgoff; dma_resv_lock(shmem->base.resv, NULL); - if (page_offset >= num_pages || - drm_WARN_ON_ONCE(obj->dev, !shmem->pages) || - shmem->madv < 0) { + if (unlikely(!drm_gem_shmem_fault_is_valid(obj, page_offset))) { ret = VM_FAULT_SIGBUS; - } else { - page = shmem->pages[page_offset]; - - ret = vmf_insert_pfn(vma, vmf->address, page_to_pfn(page)); + goto out; } + pfn = page_to_pfn(pages[page_offset]); + ret = vmf_insert_pfn(vma, vmf->address, pfn); + + out: dma_resv_unlock(shmem->base.resv); return ret; @@ -632,6 +683,7 @@ static void drm_gem_shmem_vm_close(struct vm_area_struct *vma) } const struct vm_operations_struct drm_gem_shmem_vm_ops = { + .map_pages = drm_gem_shmem_map_pages, .fault = drm_gem_shmem_fault, .open = drm_gem_shmem_vm_open, .close = drm_gem_shmem_vm_close, -- 2.47.3
