9p: dynamically allocate directory entry names
Subsystems (4)
PATCH 1/1 9p: dynamically allocate directory entry names
p9dirent_read() copies names into a fixed 256-byte buffer. A name that is valid on the host filesystem but longer than that buffer makes strscpy() fail and aborts getdents64(), hiding the remaining directory entries.
Keep ownership of the protocol-allocated string in p9_dirent and free it after dir_emit() consumes the name.
--- fs/9p/vfs_dir.c | 6 +++++- include/net/9p/client.h | 2 +- net/9p/protocol.c | 12 ++---------- 3 files changed, 8 insertions(+), 12 deletions(-)
fs/9p/vfs_dir.c +5 -1
@@ -185,8 +185,12 @@ static int v9fs_dir_readdir_dotl(struct file *file, struct dir_context *ctx)if (!dir_emit (ctx, curdirent.d_name,strlen (curdirent.d_name),QID2INO (&curdirent.qid), - curdirent.d_type)) + curdirent.d_type)) { +kfree (curdirent.d_name);return 0 ; + } + +kfree (curdirent.d_name); ctx->pos = curdirent.d_off; rdir->head += err;
include/net/9p/client.h +1 -1
@@ -268,7 +268,7 @@ struct p9_dirent {struct p9_qid qid;u64 d_off;unsigned char d_type; -char d_name[256 ]; +char *d_name; };struct iov_iter ;
net/9p/protocol.c +2 -10
@@ -770,7 +770,7 @@ int p9dirent_read(struct p9_client *clnt, char *buf, int len, {struct p9_fcall fake_pdu;int ret; -char *nameptr; +char *nameptr =NULL ; fake_pdu.size = len; fake_pdu.capacity = len;
@@ -785,15 +785,7 @@ int p9dirent_read(struct p9_client *clnt, char *buf, int len,return ret; } - ret =strscpy (dirent->d_name, nameptr,sizeof (dirent->d_name)); -if (ret <0 ) { -p9_debug (P9_DEBUG_ERROR , -"On the wire dirent name too long: %s\n" , - nameptr); -kfree (nameptr); -return ret; - } -kfree (nameptr); + dirent->d_name = nameptr;return fake_pdu.offset; }
base-commit: 028ef9c96e96197026887c0f092424679298aae8
Signature · 1 line
2.54.0 (Apple Git-157)
Dominique Martinet
maintainer+fsdevel@ & vfs maintainers to Cc TL;DR:
Should the 9p .iterate_shared() dir_emit files with path components > NAME_LEN, as allowed in fs/readdir.c verify_dirent_name() only enforcing length < PATH_MAX, or should we skip any such entry given nothing can interact with them later anyway?
hoobnn wrote on Wed, Aug 26, 2026 at 02:48:19PM +0800:
Quoted 11 lines
p9dirent_read()copies names into a fixed 256-byte buffer. A name that is valid on the host filesystem but longer than that buffer makesstrscpy()fail and abortsgetdents64(), hiding the remaining directory entries.Keep ownership of the protocol-allocated string in
p9_direntand free it afterdir_emit()consumes the name.Fixes 7751bdb3a095 ("9p: readdir implementation for 9p2000.L")Assisted-by Codex:gpt-5Signed-off-by hoobnn <[email protected]>
Sorry, but pseudonyms are not allowed to contribute to the kernel:
please use a real name here.
Okay, I've just tried with a dummy server that adds a few bytes to every filenames, and it's weirder than I expected... The vfs only checks for PATH_MAX length, not NAME_MAX, so e.g. getdents64 will list the long file name but then any operation on that file name will fail with ENAMETOOLONG...
I think it'd make sense to be more clear about this in the commit message.
(The patch subject is also pretty bad, by the way, in that "dynamically allocate ..." sounds like it adds an extra allocation whereas it just reuses the one that exists, so it should be more like "9p: skip intermediate directory entry name copy in p9dirent_read()" and explain a bit more how the current readdir processing works in the commit message e.g. first alloc+copy in p9pdu_vreadf etc)
Anyway, code-wise:
- p9_dirent are very temporary and are meant to get freed immediately, I think it's perfectly fine to skip the strscpy here and pass the pointer directly.
- I'd like a second opinion on whether it's acceptable that we return a file whose lenght is > NAME_MAX to userspace here; we're already calling strlen() again in v9fs_dir_readdir_dotl() so we could just skip the dir_emit() call here if required.
Thanks,
Signature · 1 line
Dominique Martinet | Asmadeus
Jan Kara
reviewerQuoted 5 lines · Dominique Martinet
+fsdevel@ & vfs maintainers to Cc TL;DR:
Should the 9p .iterate_shared()dir_emitfiles with path components >NAME_LEN, as allowed infs/readdir.cverify_dirent_name()only enforcing length <PATH_MAX, or should we skip any such entry given nothing can interact with them later anyway?
So my take on this is that I'd refuse any entries larger than NAME_MAX (255). As you say it is very likely something in userspace or other parts of the kernel (fs drivers, ...) will break with such names so I don't think there's a good reason to provoke these bugs. For example POSIX states readdir(2) should be returning component of at most NAME_MAX length and although we don't take POSIX very seriously in the kernel in this case I don't think there's a good enough reason to deviate from it.
Honza
Quoted 45 lines
hoobnn wrote on Wed, Aug 26, 2026 at 02:48:19PM +0800:
Quoted 11 lines
p9dirent_read()copies names into a fixed 256-byte buffer. A name that is valid on the host filesystem but longer than that buffer makesstrscpy()fail and abortsgetdents64(), hiding the remaining directory entries.Keep ownership of the protocol-allocated string in
p9_direntand free it afterdir_emit()consumes the name.Fixes 7751bdb3a095 ("9p: readdir implementation for 9p2000.L")Assisted-by Codex:gpt-5Signed-off-by hoobnn <[email protected]>Sorry, but pseudonyms are not allowed to contribute to the kernel:
please use a real name here.Okay, I've just tried with a dummy server that adds a few bytes to every filenames, and it's weirder than I expected... The vfs only checks for
PATH_MAXlength, notNAME_MAX, so e.g. getdents64 will list the long file name but then any operation on that file name will fail with ENAMETOOLONG...
I think it'd make sense to be more clear about this in the commit message.
(The patch subject is also pretty bad, by the way, in that "dynamically allocate ..." sounds like it adds an extra allocation whereas it just reuses the one that exists, so it should be more like "9p: skip intermediate directory entry name copy inp9dirent_read()" and explain a bit more how the current readdir processing works in the commit message e.g. first alloc+copy inp9pdu_vreadfetc)Anyway, code-wise:
-p9_direntare very temporary and are meant to get freed immediately, I think it's perfectly fine to skip the strscpy here and pass the pointer directly.
- I'd like a second opinion on whether it's acceptable that we return a file whose lenght is >NAME_MAXto userspace here; we're already callingstrlen()again inv9fs_dir_readdir_dotl()so we could just skip thedir_emit()call here if required.Thanks,
Signature · 1 line
Dominique Martinet | Asmadeus
Signature · 2 lines
Jan Kara <[email protected]> SUSE Labs, CR
Dominique Martinet
maintainerJan Kara wrote on Wed, Sep 16, 2026 at 11:47:00AM +0200:
Quoted 7 lines · Dominique Martinet
Quoted 5 lines
+fsdevel@ & vfs maintainers to Cc TL;DR:
Should the 9p .iterate_shared()dir_emitfiles with path components >NAME_LEN, as allowed infs/readdir.cverify_dirent_name()only enforcing length <PATH_MAX, or should we skip any such entry given nothing can interact with them later anyway?
Quoted 7 lines
So my take on this is that I'd refuse any entries larger than
NAME_MAX(255). As you say it is very likely something in userspace or other parts of the kernel (fs drivers, ...) will break with such names so I don't think there's a good reason to provoke these bugs. For example POSIX states readdir(2) should be returning component of at mostNAME_MAXlength and although we don't take POSIX very seriously in the kernel in this case I don't think there's a good enough reason to deviate from it.
Thanks for taking the time to confirm this! Let's go with that direction then.
@hoobnn: as said in my previous mail I can't take your patch unless you give me a name, but the idea behind the patch is trivial so if I don't hear back from you in say ~2 weeks I'll send a patch under my name that reimplements what you did, plus check the strlen() result that's already done in v9fs_dir_readdir_dotl() so entries > NAME_MAX are skipped over.
If you reply then there's no hurry on my end and you can take your time, please send a v2 of this as two patches, one skipping the extra copy into p9_dirent as done here and one for the limit
Signature · 1 line
Dominique