[PATCH v2] fuse: back uncached readdir buffers with pages

Matthew R. Ochs posted 1 patch 1 month, 2 weeks ago
There is a newer version of this series
fs/fuse/readdir.c | 67 ++++++++++++++++++++++++++++++++++++++++-------
1 file changed, 57 insertions(+), 10 deletions(-)
[PATCH v2] fuse: back uncached readdir buffers with pages
Posted by Matthew R. Ochs 1 month, 2 weeks ago
Commit dabb90391028 ("fuse: increase readdir buffer size") changed
fuse_readdir_uncached() to size its temporary buffer from ctx->count.
That is useful for overlayfs and other in-kernel callers that use
INT_MAX to indicate an unlimited directory read.

The buffer is capped by fc->max_pages converted to bytes with PAGE_SIZE.
However, fc->max_pages is a page-count limit, while fc->max_write is the
negotiated byte-sized payload limit. Using only fc->max_pages can produce
a READDIR request larger than the server is prepared to handle, especially
when the server and client use different page sizes.

The larger buffer is also currently supplied as a kvec output argument.
For virtiofs, kvec arguments are copied through req->argbuf, which is
allocated with kmalloc(..., GFP_ATOMIC). A large readdir buffer can
therefore require a multi-megabyte contiguous atomic allocation and fail
with -ENOMEM.

This was observed with a 64K-page guest on a 4K-page host, using an
overlayfs mount whose lower directory is on virtiofs. Reading a merged
directory through overlayfs failed with:

  ls: reading directory '<path>': Cannot allocate memory

Avoid the oversized request and the large bounce-buffer allocation by
capping the requested byte size by both fc->max_pages and fc->max_write,
then backing the uncached readdir output with pages and setting out_pages.
The virtiofs transport can then pass the pages as scatter-gather entries
instead of copying the output through argbuf.

Map the pages with vm_map_ram() only while parsing the returned dirents,
so the existing parser can continue to operate on a linear kernel mapping.

Fixes: dabb90391028 ("fuse: increase readdir buffer size")
Cc: stable@vger.kernel.org
Signed-off-by: Matthew R. Ochs <mochs@nvidia.com>
---
v2:
- Reworked uncached readdir to use output pages and out_pages, per Miklos.
- Cap the requested byte size by both fc->max_pages and fc->max_write.
- Map pages with vm_map_ram() only while parsing returned dirents.
- Verified with --overlay-rwdir across 4K/64K host and guest page sizes.
- Link to v1: https://lore.kernel.org/all/20260428021304.2338592-1-mochs@nvidia.com/

 fs/fuse/readdir.c | 67 ++++++++++++++++++++++++++++++++++++++++-------
 1 file changed, 57 insertions(+), 10 deletions(-)

diff --git a/fs/fuse/readdir.c b/fs/fuse/readdir.c
index db5ae8ec1030..27162084a683 100644
--- a/fs/fuse/readdir.c
+++ b/fs/fuse/readdir.c
@@ -12,6 +12,7 @@
 #include <linux/posix_acl.h>
 #include <linux/pagemap.h>
 #include <linux/highmem.h>
+#include <linux/vmalloc.h>
 
 static bool fuse_use_readdirplus(struct inode *dir, struct dir_context *ctx)
 {
@@ -343,17 +344,45 @@ static int fuse_readdir_uncached(struct file *file, struct dir_context *ctx)
 	struct fuse_mount *fm = get_fuse_mount(inode);
 	struct fuse_conn *fc = fm->fc;
 	struct fuse_io_args ia = {};
-	struct fuse_args *args = &ia.ap.args;
+	struct fuse_args_pages *ap = &ia.ap;
+	struct fuse_args *args = &ap->args;
+	struct page **pages;
 	void *buf;
-	size_t bufsize = clamp((unsigned int) ctx->count, PAGE_SIZE, fc->max_pages << PAGE_SHIFT);
+	size_t max_bufsize = min_t(size_t, (size_t)fc->max_pages << PAGE_SHIFT,
+				   fc->max_write);
+	size_t count = ctx->count > 0 ? ctx->count : PAGE_SIZE;
+	size_t bufsize = min_t(size_t, max_t(size_t, count, PAGE_SIZE),
+			       max_bufsize);
+	unsigned int nr_pages = DIV_ROUND_UP(bufsize, PAGE_SIZE);
 	u64 attr_version = 0, evict_ctr = 0;
 	bool locked;
+	unsigned int nr_alloc = 0;
+	unsigned int i;
 
-	buf = kvmalloc(bufsize, GFP_KERNEL);
-	if (!buf)
+	pages = kvcalloc(nr_pages, sizeof(*pages), GFP_KERNEL);
+	if (!pages)
 		return -ENOMEM;
 
-	args->out_args[0].value = buf;
+	while (nr_alloc < nr_pages) {
+		unsigned int last = nr_alloc;
+
+		nr_alloc = alloc_pages_bulk(GFP_KERNEL, nr_pages, pages);
+		if (nr_alloc == last)
+			goto nomem;
+	}
+
+	ap->folios = fuse_folios_alloc(nr_pages, GFP_KERNEL, &ap->descs);
+	if (!ap->folios)
+		goto nomem;
+
+	for (i = 0; i < nr_pages; i++) {
+		ap->folios[i] = page_folio(pages[i]);
+		ap->descs[i].length = min_t(size_t,
+					    bufsize - (size_t)i * PAGE_SIZE,
+					    PAGE_SIZE);
+	}
+	ap->num_folios = nr_pages;
+	args->out_pages = true;
 
 	plus = fuse_use_readdirplus(inode, ctx);
 	if (plus) {
@@ -372,17 +401,35 @@ static int fuse_readdir_uncached(struct file *file, struct dir_context *ctx)
 
 			if (ff->open_flags & FOPEN_CACHE_DIR)
 				fuse_readdir_cache_end(file, ctx->pos);
-		} else if (plus) {
-			res = parse_dirplusfile(buf, res, file, ctx, attr_version,
-						evict_ctr);
 		} else {
-			res = parse_dirfile(buf, res, file, ctx);
+			buf = vm_map_ram(pages, nr_pages, -1);
+			if (!buf) {
+				res = -ENOMEM;
+			} else {
+				if (plus)
+					res = parse_dirplusfile(buf, res, file, ctx,
+								attr_version,
+								evict_ctr);
+				else
+					res = parse_dirfile(buf, res, file, ctx);
+
+				vm_unmap_ram(buf, nr_pages);
+			}
 		}
 	}
 
-	kvfree(buf);
 	fuse_invalidate_atime(inode);
+
+out:
+	kfree(ap->folios);
+	for (i = 0; i < nr_alloc; i++)
+		__free_page(pages[i]);
+	kvfree(pages);
 	return res;
+
+nomem:
+	res = -ENOMEM;
+	goto out;
 }
 
 enum fuse_parse_result {
-- 
2.50.1
Re: [PATCH v2] fuse: back uncached readdir buffers with pages
Posted by Bernd Schubert 1 month, 2 weeks ago

On 4/29/26 01:29, Matthew R. Ochs wrote:
> [You don't often get email from mochs@nvidia.com. Learn why this is important at https://aka.ms/LearnAboutSenderIdentification ]
> 
> Commit dabb90391028 ("fuse: increase readdir buffer size") changed
> fuse_readdir_uncached() to size its temporary buffer from ctx->count.
> That is useful for overlayfs and other in-kernel callers that use
> INT_MAX to indicate an unlimited directory read.
> 
> The buffer is capped by fc->max_pages converted to bytes with PAGE_SIZE.
> However, fc->max_pages is a page-count limit, while fc->max_write is the
> negotiated byte-sized payload limit. Using only fc->max_pages can produce
> a READDIR request larger than the server is prepared to handle, especially
> when the server and client use different page sizes.
> 
> The larger buffer is also currently supplied as a kvec output argument.
> For virtiofs, kvec arguments are copied through req->argbuf, which is
> allocated with kmalloc(..., GFP_ATOMIC). A large readdir buffer can
> therefore require a multi-megabyte contiguous atomic allocation and fail
> with -ENOMEM.
> 
> This was observed with a 64K-page guest on a 4K-page host, using an
> overlayfs mount whose lower directory is on virtiofs. Reading a merged
> directory through overlayfs failed with:
> 
>   ls: reading directory '<path>': Cannot allocate memory
> 
> Avoid the oversized request and the large bounce-buffer allocation by
> capping the requested byte size by both fc->max_pages and fc->max_write,
> then backing the uncached readdir output with pages and setting out_pages.
> The virtiofs transport can then pass the pages as scatter-gather entries
> instead of copying the output through argbuf.
> 
> Map the pages with vm_map_ram() only while parsing the returned dirents,
> so the existing parser can continue to operate on a linear kernel mapping.
> 
> Fixes: dabb90391028 ("fuse: increase readdir buffer size")
> Cc: stable@vger.kernel.org
> Signed-off-by: Matthew R. Ochs <mochs@nvidia.com>

Josef and Joanne had spent quite some time to allow to use large folios
- maybe we should make use of it? Attached is a totally untested patch
and that ignores all of Miklos' comments for now. Also cannot be that
easily back ported

Thanks,
Bernd


> ---
> v2:
> - Reworked uncached readdir to use output pages and out_pages, per Miklos.
> - Cap the requested byte size by both fc->max_pages and fc->max_write.
> - Map pages with vm_map_ram() only while parsing returned dirents.
> - Verified with --overlay-rwdir across 4K/64K host and guest page sizes.
> - Link to v1: https://lore.kernel.org/all/20260428021304.2338592-1-mochs@nvidia.com/
> 
>  fs/fuse/readdir.c | 67 ++++++++++++++++++++++++++++++++++++++++-------
>  1 file changed, 57 insertions(+), 10 deletions(-)
> 
> diff --git a/fs/fuse/readdir.c b/fs/fuse/readdir.c
> index db5ae8ec1030..27162084a683 100644
> --- a/fs/fuse/readdir.c
> +++ b/fs/fuse/readdir.c
> @@ -12,6 +12,7 @@
>  #include <linux/posix_acl.h>
>  #include <linux/pagemap.h>
>  #include <linux/highmem.h>
> +#include <linux/vmalloc.h>
> 
>  static bool fuse_use_readdirplus(struct inode *dir, struct dir_context *ctx)
>  {
> @@ -343,17 +344,45 @@ static int fuse_readdir_uncached(struct file *file, struct dir_context *ctx)
>         struct fuse_mount *fm = get_fuse_mount(inode);
>         struct fuse_conn *fc = fm->fc;
>         struct fuse_io_args ia = {};
> -       struct fuse_args *args = &ia.ap.args;
> +       struct fuse_args_pages *ap = &ia.ap;
> +       struct fuse_args *args = &ap->args;
> +       struct page **pages;
>         void *buf;
> -       size_t bufsize = clamp((unsigned int) ctx->count, PAGE_SIZE, fc->max_pages << PAGE_SHIFT);
> +       size_t max_bufsize = min_t(size_t, (size_t)fc->max_pages << PAGE_SHIFT,
> +                                  fc->max_write);
> +       size_t count = ctx->count > 0 ? ctx->count : PAGE_SIZE;
> +       size_t bufsize = min_t(size_t, max_t(size_t, count, PAGE_SIZE),
> +                              max_bufsize);
> +       unsigned int nr_pages = DIV_ROUND_UP(bufsize, PAGE_SIZE);
>         u64 attr_version = 0, evict_ctr = 0;
>         bool locked;
> +       unsigned int nr_alloc = 0;
> +       unsigned int i;
> 
> -       buf = kvmalloc(bufsize, GFP_KERNEL);
> -       if (!buf)
> +       pages = kvcalloc(nr_pages, sizeof(*pages), GFP_KERNEL);
> +       if (!pages)
>                 return -ENOMEM;
> 
> -       args->out_args[0].value = buf;
> +       while (nr_alloc < nr_pages) {
> +               unsigned int last = nr_alloc;
> +
> +               nr_alloc = alloc_pages_bulk(GFP_KERNEL, nr_pages, pages);
> +               if (nr_alloc == last)
> +                       goto nomem;
> +       }
> +
> +       ap->folios = fuse_folios_alloc(nr_pages, GFP_KERNEL, &ap->descs);
> +       if (!ap->folios)
> +               goto nomem;
> +
> +       for (i = 0; i < nr_pages; i++) {
> +               ap->folios[i] = page_folio(pages[i]);
> +               ap->descs[i].length = min_t(size_t,
> +                                           bufsize - (size_t)i * PAGE_SIZE,
> +                                           PAGE_SIZE);
> +       }
> +       ap->num_folios = nr_pages;
> +       args->out_pages = true;
> 
>         plus = fuse_use_readdirplus(inode, ctx);
>         if (plus) {
> @@ -372,17 +401,35 @@ static int fuse_readdir_uncached(struct file *file, struct dir_context *ctx)
> 
>                         if (ff->open_flags & FOPEN_CACHE_DIR)
>                                 fuse_readdir_cache_end(file, ctx->pos);
> -               } else if (plus) {
> -                       res = parse_dirplusfile(buf, res, file, ctx, attr_version,
> -                                               evict_ctr);
>                 } else {
> -                       res = parse_dirfile(buf, res, file, ctx);
> +                       buf = vm_map_ram(pages, nr_pages, -1);
> +                       if (!buf) {
> +                               res = -ENOMEM;
> +                       } else {
> +                               if (plus)
> +                                       res = parse_dirplusfile(buf, res, file, ctx,
> +                                                               attr_version,
> +                                                               evict_ctr);
> +                               else
> +                                       res = parse_dirfile(buf, res, file, ctx);
> +
> +                               vm_unmap_ram(buf, nr_pages);
> +                       }
>                 }
>         }
> 
> -       kvfree(buf);
>         fuse_invalidate_atime(inode);
> +
> +out:
> +       kfree(ap->folios);
> +       for (i = 0; i < nr_alloc; i++)
> +               __free_page(pages[i]);
> +       kvfree(pages);
>         return res;
> +
> +nomem:
> +       res = -ENOMEM;
> +       goto out;
>  }
> 
>  enum fuse_parse_result {
> --
> 2.50.1
> 
Re: [PATCH v2] fuse: back uncached readdir buffers with pages
Posted by Miklos Szeredi 1 month, 2 weeks ago
On Wed, 29 Apr 2026 at 11:29, Bernd Schubert <bschubert@ddn.com> wrote:

> Josef and Joanne had spent quite some time to allow to use large folios
> - maybe we should make use of it? Attached is a totally untested patch
> and that ignores all of Miklos' comments for now. Also cannot be that
> easily back ported

I'd be happier if the VM infrastructure for folio arrays was available
first, then used in fuse.  Not the other way round.

Thanks,
Miklos
Re: [PATCH v2] fuse: back uncached readdir buffers with pages
Posted by Bernd Schubert 1 month, 2 weeks ago

On 4/29/26 12:38, Miklos Szeredi wrote:
> On Wed, 29 Apr 2026 at 11:29, Bernd Schubert <bschubert@ddn.com> wrote:
> 
>> Josef and Joanne had spent quite some time to allow to use large folios
>> - maybe we should make use of it? Attached is a totally untested patch
>> and that ignores all of Miklos' comments for now. Also cannot be that
>> easily back ported
> 
> I'd be happier if the VM infrastructure for folio arrays was available
> first, then used in fuse.  Not the other way round.

It is only the the missing vm_map part for folios? I had found a patch

https://lists.freedesktop.org/archives/dri-devel/2025-March/497993.html

and added the comment therefore. Maybe we can bring it up with Matthew
or someone else from mm next week.

A bit a pity if there is generic support for large folios and fuse
internals for random reasons then still use single pages.

Thanks,
Bernd
Re: [PATCH v2] fuse: back uncached readdir buffers with pages
Posted by Miklos Szeredi 1 month, 2 weeks ago
On Wed, 29 Apr 2026 at 01:30, Matthew R. Ochs <mochs@nvidia.com> wrote:

> The larger buffer is also currently supplied as a kvec output argument.
> For virtiofs, kvec arguments are copied through req->argbuf, which is
> allocated with kmalloc(..., GFP_ATOMIC). A large readdir buffer can
> therefore require a multi-megabyte contiguous atomic allocation and fail
> with -ENOMEM.

Shouldn't this be max_read?  Here "read" and "write" refer to
direction of I/O on the filesystem, not on the fuse device (see
fuse/file.c)

> @@ -343,17 +344,45 @@ static int fuse_readdir_uncached(struct file *file, struct dir_context *ctx)
>         struct fuse_mount *fm = get_fuse_mount(inode);
>         struct fuse_conn *fc = fm->fc;
>         struct fuse_io_args ia = {};
> -       struct fuse_args *args = &ia.ap.args;
> +       struct fuse_args_pages *ap = &ia.ap;
> +       struct fuse_args *args = &ap->args;
> +       struct page **pages;
>         void *buf;
> -       size_t bufsize = clamp((unsigned int) ctx->count, PAGE_SIZE, fc->max_pages << PAGE_SHIFT);
> +       size_t max_bufsize = min_t(size_t, (size_t)fc->max_pages << PAGE_SHIFT,

No need to cast if using the min_t variant.

> +                                  fc->max_write);
> +       size_t count = ctx->count > 0 ? ctx->count : PAGE_SIZE;

This is open coding the max_t(size_t, count, PAGE_SIZE) in the next
line.  Just delete.

> +       size_t bufsize = min_t(size_t, max_t(size_t, count, PAGE_SIZE),
> +                              max_bufsize);

What's wrong with the clamp() construct used originally?

> +       unsigned int nr_pages = DIV_ROUND_UP(bufsize, PAGE_SIZE);
>         u64 attr_version = 0, evict_ctr = 0;
>         bool locked;
> +       unsigned int nr_alloc = 0;
> +       unsigned int i;
>
> -       buf = kvmalloc(bufsize, GFP_KERNEL);
> -       if (!buf)
> +       pages = kvcalloc(nr_pages, sizeof(*pages), GFP_KERNEL);

 struct page **pages __free(kvfree) = kvcalloc(nr_pages,
sizeof(*pages), GFP_KERNEL);


> +       if (!pages)
>                 return -ENOMEM;
>
> -       args->out_args[0].value = buf;
> +       while (nr_alloc < nr_pages) {
> +               unsigned int last = nr_alloc;
> +
> +               nr_alloc = alloc_pages_bulk(GFP_KERNEL, nr_pages, pages);
> +               if (nr_alloc == last)
> +                       goto nomem;
> +       }

I'd try this without the loop for less complexity.  Falling back to a
shorter read shouldn't be a problem, as long as this doesn't happen
very often.

> +
> +       ap->folios = fuse_folios_alloc(nr_pages, GFP_KERNEL, &ap->descs);
> +       if (!ap->folios)
> +               goto nomem;
> +
> +       for (i = 0; i < nr_pages; i++) {
> +               ap->folios[i] = page_folio(pages[i]);
> +               ap->descs[i].length = min_t(size_t,
> +                                           bufsize - (size_t)i * PAGE_SIZE,
> +                                           PAGE_SIZE);
> +       }
> +       ap->num_folios = nr_pages;
> +       args->out_pages = true;
>
>         plus = fuse_use_readdirplus(inode, ctx);
>         if (plus) {
> @@ -372,17 +401,35 @@ static int fuse_readdir_uncached(struct file *file, struct dir_context *ctx)
>
>                         if (ff->open_flags & FOPEN_CACHE_DIR)
>                                 fuse_readdir_cache_end(file, ctx->pos);
> -               } else if (plus) {
> -                       res = parse_dirplusfile(buf, res, file, ctx, attr_version,
> -                                               evict_ctr);
>                 } else {
> -                       res = parse_dirfile(buf, res, file, ctx);
> +                       buf = vm_map_ram(pages, nr_pages, -1);
> +                       if (!buf) {
> +                               res = -ENOMEM;
> +                       } else {
> +                               if (plus)
> +                                       res = parse_dirplusfile(buf, res, file, ctx,
> +                                                               attr_version,
> +                                                               evict_ctr);
> +                               else
> +                                       res = parse_dirfile(buf, res, file, ctx);
> +
> +                               vm_unmap_ram(buf, nr_pages);
> +                       }
>                 }
>         }
>
> -       kvfree(buf);
>         fuse_invalidate_atime(inode);
> +
> +out:
> +       kfree(ap->folios);
> +       for (i = 0; i < nr_alloc; i++)
> +               __free_page(pages[i]);

release_pages()

> +       kvfree(pages);
>         return res;
> +
> +nomem:
> +       res = -ENOMEM;
> +       goto out;

Usual pattern is to just do res = -ENOMEM before each goto out (or
just the first if nothing else modifies res), so no double jump unless
absolutely necessary.

Thanks,
Miklos