9pfs: fix deep path truncation in V9fsPath
PATCH 1/1 9pfs: fix deep path truncation in V9fsPath
V9fsPath.size tracks the length of backend path data. Storing it in a uint16_t truncates local backend paths longer than 65535 bytes, so later path copies can end up much smaller than the string data they are supposed to describe.
A guest can reach this with normal 9p filesystem operations by creating and walking a sufficiently deep directory tree on the local backend. On an ASan build, calling readdir() in that deep directory aborts the host process with:
ERROR: AddressSanitizer: heap-buffer-overflow
#0 __interceptor_strrchr
#1 g_path_get_dirname
#2 local_lstat
#3 v9fs_co_lstat
#4 v9fs_getattrFix this by storing V9fsPath lengths in size_t.
Resolves: https://gitlab.com/qemu-project/qemu/-/work_items/3358
---
Runtime reproducer:
confirmed on current master (11.0.50) with an x86_64 ASan build and a local 9p backend
guest actions:
- mount the 9p share
- create a 260-level directory tree with 255-byte names
- walk back to the deepest directory
- call readdir()
host abort:
ERROR: AddressSanitizer: heap-buffer-overflow
#0 __interceptor_strrchr
#1 g_path_get_dirname
#2 local_lstat
#3 v9fs_co_lstat
#4 v9fs_getattrfsdev/file-op-9p.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
fsdev/file-op-9p.h +1 -1
@@ -112,7 +112,7 @@ struct FsContext { };struct V9fsPath { -uint16_t size; +size_t size;char *data; };P9ARRAY_DECLARE_TYPE (V9fsPath);
Signature · 1 line
2.34.1
Please ignore my [PATCH] 9pfs: fix deep path truncation in V9fsPath
Hi,
Please ignore my [PATCH] 9pfs: fix deep path truncation in V9fsPath.
Sorry. This issue has already been submitted, and the existing GitLab report already includes a concrete fix proposal:
https://gitlab.com/qemu-project/qemu/-/work_items/3358
Please ignore my patch, and sorry for the noise.
Thanks,
Jia Jia
Quoted 17 lines · On Tuesday, 28 April 2026 09:46:14 CEST Jia Jia wrote:
V9fsPath.size tracks the length of backend path data. Storing it in a
uint16_ttruncates local backend paths longer than 65535 bytes, so later path copies can end up much smaller than the string data they are supposed to describe.A guest can reach this with normal 9p filesystem operations by creating and walking a sufficiently deep directory tree on the local backend. On an ASan build, calling
readdir()in that deep directory aborts the host process with:ERROR: AddressSanitizer: heap-buffer-overflow
#0 __interceptor_strrchr #1 g_path_get_dirname #2 local_lstat #3 v9fs_co_lstat #4 v9fs_getattr
Quoted 6 lines
Fix this by storing V9fsPath lengths in
size_t.Resolves: https://gitlab.com/qemu-project/qemu/-/work_items/3358
Signed-off-by Jia Jia <[email protected]>---
Hi Jia, thanks for looking at this issue!
Quoted 26 lines
Runtime reproducer:
confirmed on current master (11.0.50) with anx86_64ASan build and a local 9p backendguest actions:
- mount the 9p share
- create a 260-level directory tree with 255-byte names
- walk back to the deepest directory
- callreaddir()host abort:
ERROR: AddressSanitizer: heap-buffer-overflow #0 __interceptor_strrchr #1 g_path_get_dirname #2 local_lstat #3 v9fs_co_lstat #4 v9fs_getattrfsdev/file-op-9p.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)fsdev/file-op-9p.h +0
-0
fsdev/file-op-9p.h struct FsContext { · line 118
Quoted 8 lines
fsdev/file-op-9p.h +1
-1@@ -112,7 +112,7 @@ struct FsContext { };struct V9fsPath { -uint16_t size; +size_t size;char *data; };P9ARRAY_DECLARE_TYPE (V9fsPath);
I fear it is not that simple. Just changing this data type would only move the problem and furthermore turn a small OOB into a giant OOB: g_vasprintf() (called by v9fs_path_sprintf()) has as return type gint -> int, and v9fs_path_sprintf() returns -1 on error, and this change would then implicitly cast -1 to a giant unsigned value.
AFAICS the only way for truly fixing this is by getting rid of dragging full paths around with FIDs in general, which would be a massive change though.
On short term I only see a possible mitigation: adding error handling to v9fs_path_sprintf(), and for FIDs where v9fs_fix_path() fails, making the fids inaccessible for good (i.e. immediately closing those FIDs).
/Christian
Hi Christian,
I agree. Just widening V9fsPath.size is not a real fix; given that g_vasprintf() returns a signed value and v9fs_path_sprintf() can return -1, my change could indeed make the failure case worse by turning that into a huge unsigned size.
I also realized afterwards that this issue had already been reported and that my patch was essentially repeating the same incomplete direction. That is why I sent the follow-up asking people to ignore it:
https://lore.kernel.org/all/[email protected]/
If I understood your suggestion correctly, the short-term mitigation would be roughly:
- make v9fs_path_sprintf() fail on both g_vasprintf() failure and overlong
results;
- propagate that error through callers such as local_name_to_path();
- make v9fs_fix_path() return an error as well;
- and if a live fid path cannot be rebuilt during rename/fixup, clunk or otherwise invalidate that fid immediately instead of leaving it reachable with invalid path state.
So this is not just a local check in v9fs_path_sprintf(), but also error propagation plus fid invalidation on fixup failure.
Sorry for the noise, and thanks for the clarification.
Thanks,
Jia Jia
On Thursday, 30 April 2026 14:52:41 CEST Jia Jia wrote:
[...]
Quoted 10 lines
If I understood your suggestion correctly, the short-term mitigation would be roughly:
- make
v9fs_path_sprintf()fail on bothg_vasprintf()failure and overlong
results;
- propagate that error through callers such aslocal_name_to_path();
- makev9fs_fix_path()return an error as well;
- and if a live fid path cannot be rebuilt during rename/fixup, clunk or otherwise invalidate that fid immediately instead of leaving it
reachable with invalid path state.
Yes, that's roughly what I had in mind as short-term mitigation. Additionally I would suggest calling error_report_once() somewhere, as it would be an unusual incident that should be logged.
Quoted 2 lines
So this is not just a local check in
v9fs_path_sprintf(), but also error propagation plus fid invalidation on fixup failure.
Right, unfortunately I don't see an easier way to address this. E.g. one might think to just add a max. path length check when creating a new file/dir, which would fix the reported vector, but then it could still be triggered by moving a dir into a subdir, and the moved dir could have a deep directory structure, so one simple path length check would not be sufficient.
/Christian