Skip to contentSkip to conversation

mm/fadvise: validate offset in generic_fadvise

PATCH Superseded [email protected] 4 participants 11 messages Dec 22, 2025 – Feb 9

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 mm/fadvise: validate offset in generic_fadvise

From: Kevin Lourenco <[email protected]>

When converted to (u64) for page calculations, a negative offset
can produce extremely large page indices. This may lead to issues in certain advice modes (excessive readahead or cache invalidation)

offsets are normally non-negative, but the API does not guarantee this. Since 'len' is already validated, checking 'offset' here is reasonable to prevent potential system instability.

---
 mm/fadvise.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
mm/fadvise.c +1 -1
index 67028e30aa91..b63fe21416ff 100644
@@ -43,7 +43,7 @@ int generic_fadvise(struct file *file, loff_t offset, loff_t len, int advice)
 		return -ESPIPE;
 
 	mapping = file->f_mapping;
-	if (!mapping || len < 0)
+	if (!mapping || len < 0 || offset < 0)
 		return -EINVAL;
 
 	bdi = inode_to_bdi(mapping->host);
Signature · 1 line
2.47.3
maintainer
#1 lore raw
Quoted 9 lines · On Mon, 22 Dec 2025 15:18:17 +0100 [email protected] wrote:

From: Kevin Lourenco <[email protected]>

When converted to (u64) for page calculations, a negative offset
can produce extremely large page indices. This may lead to issues in certain advice modes (excessive readahead or cache invalidation)

offsets are normally non-negative, but the API does not guarantee this. Since 'len' is already validated, checking 'offset' here is reasonable to prevent potential system instability.

Yeah, seems this code really didn't think about negative offsets.

	fadvise(fd, -20, 40, advice);

does make some sense I guess - equivalent to

	fadvise(fd, 0, 20, advice);

but whether this is how the implementation actually treats -20,40 is unclear. Probably not, and what will filemap_flush_range() make of this.

ugh.

I suppose I'll toss this into -next for now, see if it breaks any tests. But more thought is needed.

#2 lore raw

Thank you, Andrew, for taking a look and for queueing in -next.

I'll keep an eye on it and will adjust the patch if needed.

Kevin

Quoted 30 lines · Andrew Morton
Quoted 9 lines · On Mon, 22 Dec 2025 15:18:17 +0100 [email protected] wrote:

From: Kevin Lourenco <[email protected]>

When converted to (u64) for page calculations, a negative offset
can produce extremely large page indices. This may lead to issues in certain advice modes (excessive readahead or cache invalidation)

offsets are normally non-negative, but the API does not guarantee this. Since 'len' is already validated, checking 'offset' here is reasonable to prevent potential system instability.

Yeah, seems this code really didn't think about negative offsets.

        fadvise(fd, -20, 40, advice);

does make some sense I guess - equivalent to

        fadvise(fd, 0, 20, advice);

but whether this is how the implementation actually treats -20,40 is unclear. Probably not, and what will filemap_flush_range() make of this.

ugh.

I suppose I'll toss this into -next for now, see if it breaks any tests. But more thought is needed.

#3 lore raw
In reply to [email protected]
Quoted 8 lines · On 12/22/25 15:18, [email protected] wrote:

From: Kevin Lourenco <[email protected]>

When converted to (u64) for page calculations, a negative offset
can produce extremely large page indices. This may lead to issues in certain advice modes (excessive readahead or cache invalidation)

offsets are normally non-negative, but the API does not guarantee this. Since 'len' is already validated, checking 'offset' here is reasonable to prevent potential system instability.

Hi,

we tend to break lines as 72 chars in the patch description. I assume Andrew will fix that up or already did it :)

Quoted 11 lines
---
  mm/fadvise.c | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
mm/fadvise.c +0 -0
index 67028e30aa91..b63fe21416ff 100644

mm/fadvise.c int generic_fadvise(struct file *file, loff_t offset, loff_t len, int advice) · line 49

Quoted 8 lines
mm/fadvise.c +1 -1
@@ -43,7 +43,7 @@ int generic_fadvise(struct file *file, loff_t offset, loff_t len, int advice)
  		return -ESPIPE;
 
  	mapping = file->f_mapping;
-	if (!mapping || len < 0)
+	if (!mapping || len < 0 || offset < 0)
  		return -EINVAL;
 
  	bdi = inode_to_bdi(mapping->host);

The man page of fadvise64()/posix_madvise() is a bit unclear. It doesn't really specify what's supposed to happen on negative size or offset.

Staring at test cases in LTP, we seem to have:

* Check the value that posix_fadvise returns for wrong ADVISE value
* Check the value that posix_fadvise returns for wrong file descriptor
* Check the value that posix_fadvise returns for wrong ADVISE value
* Check the value that posix_fadvise returns for pipe descriptor

And we primarily only seem to test what's documented in the man page to fail.

Which raises the questions:
(1) Could we accidentally break some users out there?
(2) Should we update the man page to document what is supposed to happen with negative size or offset.

Signature · 3 lines
Cheers

David
#4 lore raw

Hi David,

Thank you for your thoughts and recommendations. Well noted, I will now start limiting the commit message to those specifications :)

For point (1):

I see POSIX definition like this: "The byte position in the file where the next I/O operation begins."
(https://pubs.opengroup.org/onlinepubs/9699919799/basedefs/V1_chap03.html?utm_source=chatgpt.com),
so what is almost certain is that a negative offset makes no sense / there is no hidden mechanism or anything. I don't think we can accidentally break users, my intuition is that nobody ever intentionally passes a negative offset (probably the same for len).

We could rely exclusively on the user to ensure the offset is positive
and follow POSIX recommendation. But since errors can occur, I think
it can be hard to debug where a simple test, the same for len, could
have made their experience easier. I think this is also the path
FreeBSD chose: https://github.com/freebsd/freebsd-src/blob/main/sys/kern/vfs_syscalls.c#L4918

For point (2):

Since I think nobody intentionally passes negative values and since that makes no sense semantically speaking, I think we can consider it unnecessary to update the man page right now and only think about it in the future if a negative offset gains a new meaning, for example.

Wdyt?

Quoted 58 lines · David Hildenbrand (Red Hat)
Quoted 8 lines · On 12/22/25 15:18, [email protected] wrote:

From: Kevin Lourenco <[email protected]>

When converted to (u64) for page calculations, a negative offset
can produce extremely large page indices. This may lead to issues in certain advice modes (excessive readahead or cache invalidation)

offsets are normally non-negative, but the API does not guarantee this. Since 'len' is already validated, checking 'offset' here is reasonable to prevent potential system instability.

Hi,

we tend to break lines as 72 chars in the patch description. I assume Andrew will fix that up or already did it :)

Quoted 11 lines
---
  mm/fadvise.c | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
mm/fadvise.c +0 -0
index 67028e30aa91..b63fe21416ff 100644

mm/fadvise.c int generic_fadvise(struct file *file, loff_t offset, loff_t len, int advice) · line 49

Quoted 8 lines
mm/fadvise.c +1 -1
@@ -43,7 +43,7 @@ int generic_fadvise(struct file *file, loff_t offset, loff_t len, int advice)
              return -ESPIPE;
 
      mapping = file->f_mapping;
-     if (!mapping || len < 0)
+     if (!mapping || len < 0 || offset < 0)
              return -EINVAL;
 
      bdi = inode_to_bdi(mapping->host);

The man page of fadvise64()/posix_madvise() is a bit unclear. It doesn't really specify what's supposed to happen on negative size or offset.

Staring at test cases in LTP, we seem to have:

* Check the value that posix_fadvise returns for wrong ADVISE value
* Check the value that posix_fadvise returns for wrong file descriptor
* Check the value that posix_fadvise returns for wrong ADVISE value
* Check the value that posix_fadvise returns for pipe descriptor

And we primarily only seem to test what's documented in the man page to fail.

Which raises the questions:
(1) Could we accidentally break some users out there?
(2) Should we update the man page to document what is supposed to happen with negative size or offset.

--
Cheers

David

#5 lore raw
Quoted 1 line · Kevin Lourenco

Hi David,

Hi!

Quoted 11 lines

Thank you for your thoughts and recommendations. Well noted, I will now start limiting the commit message to those specifications :)

For point (1):

I see POSIX definition like this: "The byte position in the file where the next I/O operation begins."
(https://pubs.opengroup.org/onlinepubs/9699919799/basedefs/V1_chap03.html?utm_source=chatgpt.com),
so what is almost certain is that a negative offset makes no sense / there is no hidden mechanism or anything.

All I find in that regard is

"... starting at offset and continuing for len bytes. The specified range need not currently exist in the file. If len is zero, all data from offset to the largest possible value of the file offset for that file shall be specified. "

So, something with a negative offset sure does not exist in the file.

Which makes me wonder whether we are here in a scenario where we would simply ignore everything "negative" (clamp offset to 0), simply return 0 (because it's not documented to create an error), or whether we are in undefined behavior land and can simply return EINVAL.

Quoted 3 lines

I don't think we can
accidentally break users, my intuition is that nobody ever intentionally passes a negative offset (probably the same for len).

Some programs do stupid things, you would be surprised :)

Quoted 6 lines

We could rely exclusively on the user to ensure the offset is positive
and follow POSIX recommendation. But since errors can occur, I think
it can be hard to debug where a simple test, the same for len, could
have made their experience easier. I think this is also the path
FreeBSD chose: https://github.com/freebsd/freebsd-src/blob/main/sys/kern/vfs_syscalls.c#L4918

That's very good thing to note!

Interestingly, they also check

offset > OFF_MAX - len

So checking that offset+len will not overflow. Do we want something similar? (check_add_overflow() etc. )

Quoted 5 lines

For point (2):

Since I think nobody intentionally passes negative values and since that makes no sense semantically speaking, I think we can consider it

Well, then why document negative length? ;) We should be as clear as possible in the documentation.

If POSIX would have been clearer we wouldn't have this discussion :)

Quoted 4 lines

unnecessary to update the man page right now and only think about it in the future if a negative offset gains a new meaning, for example.

Wdyt?

Likely we should really update the man page to reflect reality. "Starting with Linux v7.0, posix_fadvise() will fail with ...".

Given that FreeBSD rejects negative offsets I guess we are good.

Signature · 3 lines
Cheers

David
maintainer
#6 lore raw
Quoted 11 lines · On Tue, 6 Jan 2026 20:46:14 +0100 "David Hildenbrand (Red Hat)"…
Quoted 2 lines

Since I think nobody intentionally passes negative values and since that makes no sense semantically speaking, I think we can consider it

Well, then why document negative length? ;) We should be as clear as possible in the documentation.

If POSIX would have been clearer we wouldn't have this discussion :)

Quoted 3 lines

unnecessary to update the man page right now and only think about it in the future if a negative offset gains a new meaning, for example.

Quoted 6 lines
Quoted 1 line

Wdyt?

Likely we should really update the man page to reflect reality. "Starting with Linux v7.0, posix_fadvise() will fail with ...".

Given that FreeBSD rejects negative offsets I guess we are good.

I'm still not sure what to do with this.

Kevin, would you have time to update the changelog with the additional things we've discussed: linux manpages, POSIX, freebsd etc? Then please resend, cc'ing linux-fsdevel, Jan Kara and anyone else we can think of ;)

Thanks.

#7 lore raw

H i Andrew !

Sorry for the late reply, I’m a bit short on time right now, but I’ll try to send an updated changelog this week.

Quoted 29 lines · Andrew Morton
Quoted 11 lines · On Tue, 6 Jan 2026 20:46:14 +0100 "David Hildenbrand (Red Hat)"…
Quoted 2 lines

Since I think nobody intentionally passes negative values and since that makes no sense semantically speaking, I think we can consider it

Well, then why document negative length? ;) We should be as clear as possible in the documentation.

If POSIX would have been clearer we wouldn't have this discussion :)

Quoted 3 lines

unnecessary to update the man page right now and only think about it in the future if a negative offset gains a new meaning, for example.

Quoted 6 lines
Quoted 1 line

Wdyt?

Likely we should really update the man page to reflect reality. "Starting with Linux v7.0, posix_fadvise() will fail with ...".

Given that FreeBSD rejects negative offsets I guess we are good.

I'm still not sure what to do with this.

Kevin, would you have time to update the changelog with the additional things we've discussed: linux manpages, POSIX, freebsd etc? Then please resend, cc'ing linux-fsdevel, Jan Kara and anyone else we can think of ;)

Thanks.

#8 lore raw

PATCH v2 1/1 mm/fadvise: validate offset in generic_fadvise

From: Kevin Lourenco <[email protected]>

When converted to (u64) for page calculations, a negative offset can produce extremely large page indices. This may lead to issues in certain advice modes (excessive readahead or cache invalidation).

Reject negative offsets with -EINVAL for consistent argument validation and to avoid silent misbehavior.

POSIX and the man page do not clearly define behavior for negative offset/len. FreeBSD rejects negative offsets as well, so failing with -EINVAL is consistent with existing practice. The man page can be updated separately to document the Linux behavior.

---
 mm/fadvise.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
mm/fadvise.c +1 -1
index 67028e30aa91..b63fe21416ff 100644
@@ -43,7 +43,7 @@ int generic_fadvise(struct file *file, loff_t offset, loff_t len, int advice)
 		return -ESPIPE;
 
 	mapping = file->f_mapping;
-	if (!mapping || len < 0)
+	if (!mapping || len < 0 || offset < 0)
 		return -EINVAL;
 
 	bdi = inode_to_bdi(mapping->host);
Signature · 1 line
2.52.0
maintainer Acked-by on v2 1/1
#9 lore raw
Quoted 9 lines · On 2/8/26 14:57, [email protected] wrote:

From: Kevin Lourenco <[email protected]>

When converted to (u64) for page calculations, a negative offset can produce extremely large page indices. This may lead to issues in certain advice modes (excessive readahead or cache invalidation).

Reject negative offsets with -EINVAL for consistent argument validation and to avoid silent misbehavior.

Quoted 7 lines

POSIX and the man page do not clearly define behavior for negative offset/len. FreeBSD rejects negative offsets as well, so failing with -EINVAL is consistent with existing practice. The man page can be updated separately to document the Linux behavior.

---

Signature · 3 lines
Cheers,

David
reviewer Reviewed-by on v2 1/1
#10 lore raw
In reply to [email protected]
Quoted 9 lines · On Sun 08-02-26 14:57:38, [email protected] wrote:

From: Kevin Lourenco <[email protected]>

When converted to (u64) for page calculations, a negative offset can produce extremely large page indices. This may lead to issues in certain advice modes (excessive readahead or cache invalidation).

Reject negative offsets with -EINVAL for consistent argument validation and to avoid silent misbehavior.

Quoted 6 lines

POSIX and the man page do not clearly define behavior for negative offset/len. FreeBSD rejects negative offsets as well, so failing with -EINVAL is consistent with existing practice. The man page can be updated separately to document the Linux behavior.

Indeed. That looks like an oversight. Feel free to add:

Honza

Quoted 20 lines
---
 mm/fadvise.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
mm/fadvise.c +1 -1
index 67028e30aa91..b63fe21416ff 100644
@@ -43,7 +43,7 @@ int generic_fadvise(struct file *file, loff_t offset, loff_t len, int advice)
 		return -ESPIPE;
 
 	mapping = file->f_mapping;
-	if (!mapping || len < 0)
+	if (!mapping || len < 0 || offset < 0)
 		return -EINVAL;
 
 	bdi = inode_to_bdi(mapping->host);
Signature · 1 line
2.52.0
Signature · 2 lines
Jan Kara <[email protected]>
SUSE Labs, CR