Skip to contentSkip to conversation

ramfs: nommu: fix error rollback and reader races in ramfs_nommu_expand_for_mapping()

PATCH In review Daniel Palmer 2 participants 4 messages Jul 13 – Jul 14

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 ramfs: nommu: fix error rollback and reader races in ramfs_nommu_expand_for_mapping()

ramfs_nommu_expand_for_mapping() sets the new i_size before it has allocated or inserted any of the contiguous backing pages.

If alloc_pages() or add_to_page_cache_lru() fails, the inode is left with an inflated i_size and possibly a partial run of pages. As ramfs_nommu_setattr() treats a truncate to the current i_size as a no-op, the expansion cannot be retried and shared mmap() of the file fails with -ENOSYS.

Setting i_size early also races with lockless readers: buffered reads and splice do not take i_rwsem, so once the new size is visible a concurrent read can instantiate a zero-filled folio, making the expansion's add_to_page_cache_lru() fail with -EEXIST.

Fix this by taking mapping->invalidate_lock around the insertion, evicting any stray folios first, and only publishing i_size once every page is in place.

On failure, after freeing the pages that were allocated but not inserted truncate the mapping back to empty so already inserted pages are also disposed of.

Fixes 642fb4d1f1dd ("[PATCH] NOMMU: Provide shared-writable mmap support on ramfs")
Assisted-by Claude:claude-5-fable # expanded my fix to address the reader race etc.

---

sashiko found these issues on a previous patch to fix creating memfd's. I quickly fixed up the i_size part and then fed it to fable for review before sending and it said the reader thing needed to be fixed too and expanded the commit message to reflect that.

I am not sure about the locking part at all so hopefully someone can chime in on that. Maybe sashiko will say the things it reported and fable fixed aren't a problem after all.. :).

I tested the happy path on my mc68000 virt machine and manually stepped through the code to make sure its actually being executed and tested on real hardware.

 fs/ramfs/file-nommu.c | 21 +++++++++++++++++----
 1 file changed, 17 insertions(+), 4 deletions(-)
fs/ramfs/file-nommu.c +17 -4
index fb471bf88ab7..6819f4790065 100644
@@ -80,8 +80,6 @@ int ramfs_nommu_expand_for_mapping(struct inode *inode, size_t newsize)
 	if (ret)
 		return ret;
 
-	i_size_write(inode, newsize);
-
 	/* allocate enough contiguous pages to be able to satisfy the
 	 * request */
 	pages = alloc_pages(gfp, order);
@@ -99,9 +97,15 @@ int ramfs_nommu_expand_for_mapping(struct inode *inode, size_t newsize)
 		__free_page(pages + loop);
 
 	/* clear the memory we allocated */
-	newsize = PAGE_SIZE * npages;
 	data = page_address(pages);
-	memset(data, 0, newsize);
+	memset(data, 0, PAGE_SIZE * npages);
+
+	/* block the read and splice paths from instantiating pagecache
+	 * folios whilst we build the contiguous mapping, and evict any
+	 * folio they may already have instantiated (the read path only
+	 * checks i_size after folio lookup/creation) */
+	filemap_invalidate_lock(inode->i_mapping);
+	truncate_inode_pages(inode->i_mapping, 0);
 
 	/* attach all the pages to the inode's address space */
 	for (loop = 0; loop < npages; loop++) {
@@ -120,11 +124,20 @@ int ramfs_nommu_expand_for_mapping(struct inode *inode, size_t newsize)
 		put_page(page);
 	}
 
+	/* only publish the new size once all of the backing pages are in
+	 * place, so that readers never observe a size without the pages to
+	 * back it and a failure leaves the inode unchanged at size 0 */
+	i_size_write(inode, newsize);
+	filemap_invalidate_unlock(inode->i_mapping);
 	return 0;
 
 add_error:
+	/* free the pages we hadn't inserted yet... */
 	while (loop < npages)
 		__free_page(pages + loop++);
+	/* ...and evict the ones we had; they belong to the page cache now */
+	truncate_inode_pages(inode->i_mapping, 0);
+	filemap_invalidate_unlock(inode->i_mapping);
 	return ret;
 }
 
Signature · 1 line
2.53.0
#1 lore raw

Hello Daniel,

thanks for cc-ing me for the patch.

Quoted 23 lines · Daniel Palmer

ramfs_nommu_expand_for_mapping() sets the new i_size before it has allocated or inserted any of the contiguous backing pages.

If alloc_pages() or add_to_page_cache_lru() fails, the inode is left with an inflated i_size and possibly a partial run of pages. As ramfs_nommu_setattr() treats a truncate to the current i_size as a no-op, the expansion cannot be retried and shared mmap() of the file fails with -ENOSYS.

Setting i_size early also races with lockless readers: buffered reads and splice do not take i_rwsem, so once the new size is visible a concurrent read can instantiate a zero-filled folio, making the expansion's add_to_page_cache_lru() fail with -EEXIST.

Fix this by taking mapping->invalidate_lock around the insertion, evicting any stray folios first, and only publishing i_size once every page is in place.

On failure, after freeing the pages that were allocated but not inserted truncate the mapping back to empty so already inserted pages are also disposed of.

Quoted 6 lines
Fixes 642fb4d1f1dd ("[PATCH] NOMMU: Provide shared-writable mmap support on ramfs")
Assisted-by Claude:claude-5-fable # expanded my fix to address the reader race etc.

---

I have tested this patch in addition to your previous patch ([email protected]) over LTP test with my local nommu-revert patchset.

I run the following test (which I have never looked at thus doesn't have much knowledge about the test itself).

https://github.com/linux-test-project/ltp/blob/master/testcases/kernel/syscalls/memfd_create/memfd_create01.c

without first patch, it indeed returns EFBIG on memfe_createfd() syscall. applying both patches with this test gives me still failure with the following messages:

```
/__w/linux/linux/lib/tst_test.c:2060: TINFO: LTP version: 20260529
/__w/linux/linux/lib/tst_test.c:2063: TINFO: Tested kernel: 7.1.0-rc1-gfd4e5701bb02-dirty #1 Mon Jul 13 18:17:11 UTC 2026 um(nommu)/x86_64
/__w/linux/linux/lib/tst_kconfig.c:90: TINFO: Parsing kernel config &#x27;/proc/config.gz&#x27;
/__w/linux/linux/lib/tst_test.c:1880: TINFO: Overall timeout per run is 0h 05m 24s
/__w/linux/linux/testcases/kernel/syscalls/memfd_create/memfd_create01.c:241: TINFO: Basic tests + set/get seals
/__w/linux/linux/testcases/kernel/syscalls/memfd_create/memfd_create01.c:243: TPASS: memfd_create(ltp_memfd_create01, 2) succeeded
/__w/linux/linux/testcases/kernel/syscalls/memfd_create/memfd_create01.c:243: TPASS: ftruncate(3, 8192) succeeded
/__w/linux/linux/testcases/kernel/syscalls/memfd_create/memfd_create_common.c:212: TBROK: fcntl(3,F_GET_SEALS,...) failed: EINVAL (22)

Summary:
passed   2
failed   0
broken   1
skipped  0
warnings 0
```

which fails at the location below (in memfd_create_common.c#L207)

	int ret = SAFE_FCNTL((fd), F_GET_SEALS);

so fcntl(fd, F_GET_SEALS) always fails with nommu due to lack of support of CONFIG_SHMEM and CONFIG_HUGETLBFS.

``` mm/memfd.c (in kernel)
static unsigned int *memfd_file_seals_ptr(struct file *file)
{
	if (shmem_file(file))
		return &SHMEM_I(file_inode(file))->seals;

#ifdef CONFIG_HUGETLBFS
	if (is_file_hugepages(file))
		return &HUGETLBFS_I(file_inode(file))->seals;
#endif

	return NULL;
}
```

so I was wondering if it is still interesting to support the syscall (memfd_createfd) on nommu even the fcntl always fails.

My test is used on riscv-nommu (buildroot) and UML + !MMU, which gives me the same results.

-- Hajime

#2 lore raw

Hi Hajime,

Quoted 2 lines · Hajime Tazaki

so I was wondering if it is still interesting to support the syscall (memfd_createfd) on nommu even the fcntl always fails.

I think these fcntls will also fail on MMU=y where the full tmpfs is not enabled for size reasons. I'm not sure if it's actually possible to detect if tmpfs is really tmpfs or not. (I have this exact issue with a test for nolibc).

memfd_createfd is still useful because it basically allows you to create anonymous temporary files that you can pass to another process and mmap() in both.
I use this for doing DNS resolution in my nolibc system: A process that needs DNS resolution creates an memfd for the result and then forks the resolver which mmap()s the memfd, fills in the result and then exits. This avoids having the big resolver code in everything that needs it since I don't have any shared libraries :).

BTW: Could you share a link to your LTP fork? I will add it to my m68k environment.

Thanks,

Daniel

#3 lore raw
Quoted 8 lines · Daniel Palmer
Quoted 3 lines

so I was wondering if it is still interesting to support the syscall (memfd_createfd) on nommu even the fcntl always fails.

I think these fcntls will also fail on MMU=y where the full tmpfs is not enabled for size reasons. I'm not sure if it's actually possible to detect if tmpfs is really tmpfs or not. (I have this exact issue with a test for nolibc).

thanks, I'll look into detail.

Quoted 8 lines

memfd_createfd is still useful because it basically allows you to create anonymous temporary files that you can pass to another process and mmap() in both.
I use this for doing DNS resolution in my nolibc system: A process that needs DNS resolution creates an memfd for the result and then forks the resolver which mmap()s the memfd, fills in the result and then exits. This avoids having the big resolver code in everything that needs it since I don't have any shared libraries :).

I understand, this might be a good test case with the actual use.

Quoted 2 lines

BTW: Could you share a link to your LTP fork? I will add it to my m68k environment.

forgive me in advance :)
links are only the current working tree; it might be renamed/removed in future.

here is a fork/branch of ltp:
https://github.com/thehajime/ltp/tree/fix-nommu

it is triggered from my UML branch with github actions (below),

https://github.com/thehajime/linux/blob/zpoline-nommu-v6.10/.github/workflows/ci.yml

using this dirty yaml file to build and execute part of testcases.

https://github.com/thehajime/linux/blob/zpoline-nommu-v6.10/.github/workflows/ci-docker-build.yml

-- Hajime