Skip to contentSkip to conversation

drm/gem-shmem: Track page accessed/dirty status

PATCH Superseded Thomas Zimmermann 3 participants 8 messages Feb 4 – Feb 9

Subsystems (4)

v2 of 4 Superseded Latest is v4 →

lore
Share & tools
lore

Thomas Zimmermann

author maintainer
#0 lore raw

v2 cover drm/gem-shmem: Track page accessed/dirty status

Track page access/dirty status in gem-shmem for better integration with the overall memory management. Gem-shmem has long had two flag bits in struct drm_gem_shmem_object, named pages_mark_accessed_on_put and pages_mark_dirty_on_put, but never used them much; except for some odd cases in drivers. Therefore pages in gem-shmem where never marked correctly. (Other DRM memory managers do some course-grain tracking at least).

Patches 1 and 2 switch from PFN-based mapping to page mapping. The pages are already available; only the mmap handling needs to be adapted.

Patch 3 adds tracking access and dirty status in mmap.

Patch 4 adds tracking access and dirty status in vmap. Becasue there's no fault handling here, we refer to the existing status bits in struct drm_gem_shmem_object. Each page's status will be updated by the page release in drm_gem_put_pages(). The imagination driver requires a small fix to make it work correctly.

Tested with CONFIG_VM=y by running animations on DRM's bochs driver for several hours. This uses gem-shmem's mmap and vmap extensively.

v2:
- fix possible OOB access into page array (Matthew)
- simplify fault-handler error handling (Boris)
- simplify internal interfaces (Matthew)

Thomas Zimmermann (4):
drm/gem-shmem: Return vm_fault_t from drm_gem_shmem_try_map_pmd() drm/gem-shmem: Map pages in mmap fault handler drm/gem-shmem: Track folio accessed/dirty status in mmap drm/gem-shmem: Track folio accessed/dirty status in vmap

 drivers/gpu/drm/drm_gem_shmem_helper.c | 77 ++++++++++++++++++--------
 drivers/gpu/drm/imagination/pvr_gem.c  |  6 +-
 2 files changed, 57 insertions(+), 26 deletions(-)

base-commit: 6e53f6296065672f8a0c7f98b4b6c409dac382b4

Signature · 1 line
2.52.0
author maintainer
#1 lore raw

PATCH v2 1/4 drm/gem-shmem: Return vm_fault_t from drm_gem_shmem_try_map_pmd()

Return the exact VM_FAULT_ mask from drm_gem_shmem_try_map_pmd(). Gives the caller better insight into the result. Return 0 if nothing was done.

If the caller sees VM_FAULT_NOPAGE, drm_gem_shmem_try_map_pmd() added a PMD entry to the page table. As before, return early from the page-fault handler in that case.

---
 drivers/gpu/drm/drm_gem_shmem_helper.c | 14 ++++++--------
 1 file changed, 6 insertions(+), 8 deletions(-)
drivers/gpu/drm/drm_gem_shmem_helper.c +6 -8
index 3871a6d92f77..e7316dc7e921 100644
@@ -550,8 +550,8 @@ int drm_gem_shmem_dumb_create(struct drm_file *file, struct drm_device *dev,
 }
 EXPORT_SYMBOL_GPL(drm_gem_shmem_dumb_create);
 
-static bool drm_gem_shmem_try_map_pmd(struct vm_fault *vmf, unsigned long addr,
-				      struct page *page)
+static vm_fault_t drm_gem_shmem_try_map_pmd(struct vm_fault *vmf, unsigned long addr,
+					    struct page *page)
 {
 #ifdef CONFIG_ARCH_SUPPORTS_PMD_PFNMAP
 	unsigned long pfn = page_to_pfn(page);
@@ -562,12 +562,11 @@ static bool drm_gem_shmem_try_map_pmd(struct vm_fault *vmf, unsigned long addr,
 	    pmd_none(*vmf->pmd) &&
 	    folio_test_pmd_mappable(page_folio(page))) {
 		pfn &= PMD_MASK >> PAGE_SHIFT;
-		if (vmf_insert_pfn_pmd(vmf, pfn, false) == VM_FAULT_NOPAGE)
-			return true;
+		return vmf_insert_pfn_pmd(vmf, pfn, false);
 	}
 #endif
 
-	return false;
+	return 0;
 }
 
 static vm_fault_t drm_gem_shmem_fault(struct vm_fault *vmf)
@@ -593,10 +592,9 @@ static vm_fault_t drm_gem_shmem_fault(struct vm_fault *vmf)
 		goto out;
 	}
 
-	if (drm_gem_shmem_try_map_pmd(vmf, vmf->address, pages[page_offset])) {
-		ret = VM_FAULT_NOPAGE;
+	ret = drm_gem_shmem_try_map_pmd(vmf, vmf->address, pages[page_offset]);
+	if (ret == VM_FAULT_NOPAGE)
 		goto out;
-	}
 
 	pfn = page_to_pfn(pages[page_offset]);
 	ret = vmf_insert_pfn(vma, vmf->address, pfn);
Signature · 1 line
2.52.0
Reviewed-by on v2 1/4
#2 lore raw

Re: [PATCH v2 1/4] drm/gem-shmem: Return vm_fault_t from drm_gem_shmem_try_map_pmd()

Quoted 9 lines · On Wed, 4 Feb 2026 12:39:29 +0100 Thomas Zimmermann <[email protected]>…

Return the exact VM_FAULT_ mask from drm_gem_shmem_try_map_pmd(). Gives the caller better insight into the result. Return 0 if nothing was done.

If the caller sees VM_FAULT_NOPAGE, drm_gem_shmem_try_map_pmd() added a PMD entry to the page table. As before, return early from the page-fault handler in that case.

Quoted 47 lines
---
 drivers/gpu/drm/drm_gem_shmem_helper.c | 14 ++++++--------
 1 file changed, 6 insertions(+), 8 deletions(-)
drivers/gpu/drm/drm_gem_shmem_helper.c +6 -8
index 3871a6d92f77..e7316dc7e921 100644
@@ -550,8 +550,8 @@ int drm_gem_shmem_dumb_create(struct drm_file *file, struct drm_device *dev,
 }
 EXPORT_SYMBOL_GPL(drm_gem_shmem_dumb_create);
 
-static bool drm_gem_shmem_try_map_pmd(struct vm_fault *vmf, unsigned long addr,
-				      struct page *page)
+static vm_fault_t drm_gem_shmem_try_map_pmd(struct vm_fault *vmf, unsigned long addr,
+					    struct page *page)
 {
 #ifdef CONFIG_ARCH_SUPPORTS_PMD_PFNMAP
 	unsigned long pfn = page_to_pfn(page);
@@ -562,12 +562,11 @@ static bool drm_gem_shmem_try_map_pmd(struct vm_fault *vmf, unsigned long addr,
 	    pmd_none(*vmf->pmd) &&
 	    folio_test_pmd_mappable(page_folio(page))) {
 		pfn &= PMD_MASK >> PAGE_SHIFT;
-		if (vmf_insert_pfn_pmd(vmf, pfn, false) == VM_FAULT_NOPAGE)
-			return true;
+		return vmf_insert_pfn_pmd(vmf, pfn, false);
 	}
 #endif
 
-	return false;
+	return 0;
 }
 
 static vm_fault_t drm_gem_shmem_fault(struct vm_fault *vmf)
@@ -593,10 +592,9 @@ static vm_fault_t drm_gem_shmem_fault(struct vm_fault *vmf)
 		goto out;
 	}
 
-	if (drm_gem_shmem_try_map_pmd(vmf, vmf->address, pages[page_offset])) {
-		ret = VM_FAULT_NOPAGE;
+	ret = drm_gem_shmem_try_map_pmd(vmf, vmf->address, pages[page_offset]);
+	if (ret == VM_FAULT_NOPAGE)
 		goto out;
-	}
 
 	pfn = page_to_pfn(pages[page_offset]);
 	ret = vmf_insert_pfn(vma, vmf->address, pfn);
author maintainer
#3 lore raw

PATCH v2 2/4 drm/gem-shmem: Map pages in mmap fault handler

In reply to Thomas Zimmermann

Gem-shmem operates on pages instead of I/O memory ranges, so use them for mmap. This will allow for tracking page dirty/accessed flags. If hugepage support is available, insert the page's folio if possible. Otherwise fall back to mapping individual pages.

As the PFN is no longer required for hugepage mappings, simplify the related code and make it depend on CONFIG_TRANSPARENT_HUGEPAGE. Prepare for tracking folio status.

v2:
- do not look up the page before testing page-array bounds (Matthew)
- simplify error handling in drm_gem_shmem_fault() (Boris)
- adapt to changes in drm_gem_shmem_try_mmap_pmd()

---
 drivers/gpu/drm/drm_gem_shmem_helper.c | 49 +++++++++++++++-----------
 1 file changed, 29 insertions(+), 20 deletions(-)
drivers/gpu/drm/drm_gem_shmem_helper.c +29 -20
index e7316dc7e921..24553dec070d 100644
@@ -553,16 +553,17 @@ EXPORT_SYMBOL_GPL(drm_gem_shmem_dumb_create);
 static vm_fault_t drm_gem_shmem_try_map_pmd(struct vm_fault *vmf, unsigned long addr,
 					    struct page *page)
 {
-#ifdef CONFIG_ARCH_SUPPORTS_PMD_PFNMAP
-	unsigned long pfn = page_to_pfn(page);
-	unsigned long paddr = pfn << PAGE_SHIFT;
+#ifdef CONFIG_TRANSPARENT_HUGEPAGE
+	phys_addr_t paddr = page_to_phys(page);
 	bool aligned = (addr & ~PMD_MASK) == (paddr & ~PMD_MASK);
 
-	if (aligned &&
-	    pmd_none(*vmf->pmd) &&
-	    folio_test_pmd_mappable(page_folio(page))) {
-		pfn &= PMD_MASK >> PAGE_SHIFT;
-		return vmf_insert_pfn_pmd(vmf, pfn, false);
+	if (aligned && pmd_none(*vmf->pmd)) {
+		struct folio *folio = page_folio(page);
+
+		if (folio_test_pmd_mappable(folio)) {
+			/* Read-only mapping; split upon write fault */
+			return vmf_insert_folio_pmd(vmf, folio, false);
+		}
 	}
 #endif
 
@@ -575,13 +576,10 @@ static vm_fault_t drm_gem_shmem_fault(struct vm_fault *vmf)
 	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;
-	vm_fault_t ret;
 	struct page **pages = shmem->pages;
-	pgoff_t page_offset;
-	unsigned long pfn;
-
-	/* Offset to faulty address in the VMA. */
-	page_offset = vmf->pgoff - vma->vm_pgoff;
+	pgoff_t page_offset = vmf->pgoff - vma->vm_pgoff; /* page offset within VMA */
+	struct page *page;
+	vm_fault_t ret;
 
 	dma_resv_lock(shmem->base.resv, NULL);
 
@@ -592,14 +590,25 @@ static vm_fault_t drm_gem_shmem_fault(struct vm_fault *vmf)
 		goto out;
 	}
 
-	ret = drm_gem_shmem_try_map_pmd(vmf, vmf->address, pages[page_offset]);
-	if (ret == VM_FAULT_NOPAGE)
+	page = pages[page_offset];
+	if (!page) {
+		ret = VM_FAULT_SIGBUS;
 		goto out;
+	}
+
+	ret = drm_gem_shmem_try_map_pmd(vmf, vmf->address, page);
+	if (ret != VM_FAULT_NOPAGE) {
+		struct folio *folio = page_folio(page);
+
+		get_page(page);
 
-	pfn = page_to_pfn(pages[page_offset]);
-	ret = vmf_insert_pfn(vma, vmf->address, pfn);
+		folio_lock(folio);
+
+		vmf->page = page;
+		ret = VM_FAULT_LOCKED;
+	}
 
- out:
+out:
 	dma_resv_unlock(shmem->base.resv);
 
 	return ret;
@@ -689,7 +698,7 @@ int drm_gem_shmem_mmap(struct drm_gem_shmem_object *shmem, struct vm_area_struct
 	if (ret)
 		return ret;
 
-	vm_flags_set(vma, VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP);
+	vm_flags_mod(vma, VM_DONTEXPAND | VM_DONTDUMP, VM_PFNMAP);
 	vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
 	if (shmem->map_wc)
 		vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
Signature · 1 line
2.52.0
#4 lore raw

Re: [PATCH v2 2/4] drm/gem-shmem: Map pages in mmap fault handler

drivers/gpu/drm/drm_gem_shmem_helper.c static vm_fault_t drm_gem_shmem_fault(struct vm_fault *vmf) · line 603

Quoted 5 lines · Thomas Zimmermann
drivers/gpu/drm/drm_gem_shmem_helper.c +5 -0
+	ret = drm_gem_shmem_try_map_pmd(vmf, vmf->address, page);
+	if (ret != VM_FAULT_NOPAGE) {
+		struct folio *folio = page_folio(page);
+
+		get_page(page);

folio_get(folio);

Quoted 14 lines
drivers/gpu/drm/drm_gem_shmem_helper.c +6 -3
-	pfn = page_to_pfn(pages[page_offset]);
-	ret = vmf_insert_pfn(vma, vmf->address, pfn);
+		folio_lock(folio);
+
+		vmf->page = page;
+		ret = VM_FAULT_LOCKED;
+	}
 
- out:
+out:
 	dma_resv_unlock(shmem->base.resv);
 
 	return ret;

drivers/gpu/drm/drm_gem_shmem_helper.c int drm_gem_shmem_mmap(struct drm_gem_shmem_object *shmem, struct vm_area_struct · line 701

Quoted 5 lines
drivers/gpu/drm/drm_gem_shmem_helper.c +1 -1
@@ -689,7 +698,7 @@ int drm_gem_shmem_mmap(struct drm_gem_shmem_object *shmem, struct vm_area_struct
 	if (ret)
 		return ret;
 
-	vm_flags_set(vma, VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP);
+	vm_flags_mod(vma, VM_DONTEXPAND | VM_DONTDUMP, VM_PFNMAP);

Do you need to explicitly clear VM_PFNMAP here? I'm not familiar with the DRM stack; maybe that's set for you higher in the stack.

author maintainer
#5 lore raw

Re: [PATCH v2 2/4] drm/gem-shmem: Map pages in mmap fault handler

Hi,

I came across commit 8b93d1d7dbd5 ("drm/shmem-helper: Switch to vmf_insert_pfn") from 2021, which makes it very clear the PFNMAP is strongly preferred over pages. I totally forgot about that change. The next iteration of this series will therefore not contain this patch.

Best regards
Thomas

Am 04.02.26 um 17:03 schrieb Matthew Wilcox:

Quoted 30 lines · Thomas Zimmermann

drivers/gpu/drm/drm_gem_shmem_helper.c static vm_fault_t drm_gem_shmem_fault(struct vm_fault *vmf) · line 603

Quoted 5 lines
drivers/gpu/drm/drm_gem_shmem_helper.c +5 -0
+	ret = drm_gem_shmem_try_map_pmd(vmf, vmf->address, page);
+	if (ret != VM_FAULT_NOPAGE) {
+		struct folio *folio = page_folio(page);
+
+		get_page(page);

folio_get(folio);

Quoted 14 lines
drivers/gpu/drm/drm_gem_shmem_helper.c +6 -3
-	pfn = page_to_pfn(pages[page_offset]);
-	ret = vmf_insert_pfn(vma, vmf->address, pfn);
+		folio_lock(folio);
+
+		vmf->page = page;
+		ret = VM_FAULT_LOCKED;
+	}
 
- out:
+out:
  	dma_resv_unlock(shmem->base.resv);
 
  	return ret;

drivers/gpu/drm/drm_gem_shmem_helper.c int drm_gem_shmem_mmap(struct drm_gem_shmem_object *shmem, struct vm_area_struct · line 701

Quoted 5 lines
drivers/gpu/drm/drm_gem_shmem_helper.c +1 -1
@@ -689,7 +698,7 @@ int drm_gem_shmem_mmap(struct drm_gem_shmem_object *shmem, struct vm_area_struct
  	if (ret)
  		return ret;
 
-	vm_flags_set(vma, VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP);
+	vm_flags_mod(vma, VM_DONTEXPAND | VM_DONTDUMP, VM_PFNMAP);

Do you need to explicitly clear VM_PFNMAP here? I'm not familiar with the DRM stack; maybe that's set for you higher in the stack.

Signature · 6 lines
--
Thomas Zimmermann
Graphics Driver Developer
SUSE Software Solutions Germany GmbH
Frankenstr. 146, 90461 Nürnberg, Germany, www.suse.com
GF: Jochen Jaser, Andrew McDonald, Werner Knoblich, (HRB 36809, AG Nürnberg)
author maintainer
#6 lore raw

PATCH v2 3/4 drm/gem-shmem: Track folio accessed/dirty status in mmap

In reply to Thomas Zimmermann

Invoke folio_mark_accessed() in mmap page faults to add the folio to the memory manager's LRU list. Userspace invokes mmap to get the memory for software rendering. Compositors do the same when creating the final on-screen image, so keeping the pages in LRU makes sense. Avoids paging out graphics buffers when under memory pressure.

In page_mkwrite, further invoke the folio_mark_dirty() to add the folio for writeback, should the underlying file be paged out from system memory. This rarely happens in practice, yet it would corrupt the buffer content.

This has little effect on a system's hardware-accelerated rendering, which only mmaps for an initial setup of textures, meshes, shaders, etc.

v2:
- adapt to changes in drm_gem_shmem_try_mmap_pmd()

---
 drivers/gpu/drm/drm_gem_shmem_helper.c | 20 +++++++++++++++++++-
 1 file changed, 19 insertions(+), 1 deletion(-)
drivers/gpu/drm/drm_gem_shmem_helper.c +19 -1
index 24553dec070d..c2ee30967c41 100644
@@ -597,12 +597,17 @@ static vm_fault_t drm_gem_shmem_fault(struct vm_fault *vmf)
 	}
 
 	ret = drm_gem_shmem_try_map_pmd(vmf, vmf->address, page);
-	if (ret != VM_FAULT_NOPAGE) {
+	if (ret == VM_FAULT_NOPAGE) {
+		struct folio *folio = page_folio(page);
+
+		folio_mark_accessed(folio);
+	} else {
 		struct folio *folio = page_folio(page);
 
 		get_page(page);
 
 		folio_lock(folio);
+		folio_mark_accessed(folio);
 
 		vmf->page = page;
 		ret = VM_FAULT_LOCKED;
@@ -648,10 +653,23 @@ static void drm_gem_shmem_vm_close(struct vm_area_struct *vma)
 	drm_gem_vm_close(vma);
 }
 
+static vm_fault_t drm_gem_shmem_page_mkwrite(struct vm_fault *vmf)
+{
+	struct folio *folio = page_folio(vmf->page);
+
+	file_update_time(vmf->vma->vm_file);
+
+	folio_lock(folio);
+	folio_mark_dirty(folio);
+
+	return VM_FAULT_LOCKED;
+}
+
 const struct vm_operations_struct drm_gem_shmem_vm_ops = {
 	.fault = drm_gem_shmem_fault,
 	.open = drm_gem_shmem_vm_open,
 	.close = drm_gem_shmem_vm_close,
+	.page_mkwrite = drm_gem_shmem_page_mkwrite,
 };
 EXPORT_SYMBOL_GPL(drm_gem_shmem_vm_ops);
 
Signature · 1 line
2.52.0
author maintainer
#7 lore raw

PATCH v2 4/4 drm/gem-shmem: Track folio accessed/dirty status in vmap

In reply to Thomas Zimmermann

On successful vmap, set the page_mark_accessed_on_put and _dirty_on_put flags in the gem-shmem object. Signals that the contained pages require LRU and dirty tracking when they are being released back to SHMEM. Clear these flags on put, so that buffer remains quiet until the next call to vmap. There's no means of handling dirty status in vmap as there's no write-only mapping available.

Both flags, _accessed_on_put and _dirty_on_put, have always been part of the gem-shmem object, but never used much. So most drivers did not track the page status correctly.

Only the v3d and imagination drivers make limited use of _dirty_on_put. In the case of imagination, move the flag setting from init to cleanup. This ensures writeback of modified pages but does not interfere with the internal vmap/vunmap calls. V3d already implements this behaviour.

---
 drivers/gpu/drm/drm_gem_shmem_helper.c | 4 ++++
 drivers/gpu/drm/imagination/pvr_gem.c  | 6 ++++--
 2 files changed, 8 insertions(+), 2 deletions(-)
drivers/gpu/drm/drm_gem_shmem_helper.c +4 -0
index c2ee30967c41..a0b0912c1d42 100644
@@ -265,6 +265,8 @@ void drm_gem_shmem_put_pages_locked(struct drm_gem_shmem_object *shmem)
 				  shmem->pages_mark_dirty_on_put,
 				  shmem->pages_mark_accessed_on_put);
 		shmem->pages = NULL;
+		shmem->pages_mark_accessed_on_put = false;
+		shmem->pages_mark_dirty_on_put = false;
 	}
 }
 EXPORT_SYMBOL_GPL(drm_gem_shmem_put_pages_locked);
@@ -397,6 +399,8 @@ int drm_gem_shmem_vmap_locked(struct drm_gem_shmem_object *shmem,
 		} else {
 			iosys_map_set_vaddr(map, shmem->vaddr);
 			refcount_set(&shmem->vmap_use_count, 1);
+			shmem->pages_mark_accessed_on_put = true;
+			shmem->pages_mark_dirty_on_put = true;
 		}
 	}
 
drivers/gpu/drm/imagination/pvr_gem.c +4 -2
index c07c9a915190..307b02c916d4 100644
@@ -25,7 +25,10 @@
 
 static void pvr_gem_object_free(struct drm_gem_object *obj)
 {
-	drm_gem_shmem_object_free(obj);
+	struct drm_gem_shmem_object *shmem_obj = to_drm_gem_shmem_obj(obj);
+
+	shmem_obj->pages_mark_dirty_on_put = true;
+	drm_gem_shmem_free(shmem_obj);
 }
 
 static struct dma_buf *pvr_gem_export(struct drm_gem_object *obj, int flags)
@@ -363,7 +366,6 @@ pvr_gem_object_create(struct pvr_device *pvr_dev, size_t size, u64 flags)
 	if (IS_ERR(shmem_obj))
 		return ERR_CAST(shmem_obj);
 
-	shmem_obj->pages_mark_dirty_on_put = true;
 	shmem_obj->map_wc = !(flags & PVR_BO_CPU_CACHED);
 	pvr_obj = shmem_gem_to_pvr_gem(shmem_obj);
 	pvr_obj->flags = flags;
Signature · 1 line
2.52.0