smb: client: fix request buffer leak in smb2_new_read_req()
PATCH 1/1 smb: client: fix request buffer leak in smb2_new_read_req()
smb2_new_read_req() allocates the request buffer with smb2_plain_req_init() but only publishes it to the caller with *buf = req at the very end of the function. Two error returns sit in between:
rc = smb2_plain_req_init(SMB2_READ, io_parms->tcon, server, (void **) &req, total_len); if (rc) return rc; if (server == NULL) return -ECONNABORTED; [...] rdata->mr = smbd_register_mr(server->smbd_conn, &rdata->subreq.io_iter, true, need_invalidate); if (!rdata->mr) return -EAGAIN;
On either of them the buffer is neither released nor handed back, so it is leaked. The caller cannot clean up after it: smb2_async_readv() does 'goto out' on a non-zero return, which skips the cifs_small_buf_release(buf) at async_readv_out, and buf has not been assigned at that point in any case.
The write path has never had this problem. smb2_async_writev() registers the memory region inline and jumps to its release label instead of returning:
wdata->mr = smbd_register_mr(...);
if (!wdata->mr) {
rc = -EAGAIN;
goto async_writev_out;
}Commit b7972092199f ("cifs: smbd: Retry on memory registration failure") changed both sides from -ENOBUFS to -EAGAIN in a single patch, which puts the two shapes next to each other.
Only the -EAGAIN return is reachable in practice, because smb2_plain_req_init() calls smb2_reconnect() first and that already fails with -EIO when server is NULL, before anything is allocated. Both returns are given the same treatment here rather than leaving one of them correct only by accident.
Because -EAGAIN is a replayable error, the failure also reaches the retry block at the end of smb2_async_readv(), which marks the subrequest NETFS_SREQ_NEED_RETRY, so a failing registration can be retried rather than ending the I/O, and every attempt that reaches it leaks another buffer. smb2_should_replay() short-circuits on tcon->retry, so on a hard mount the attempt count is not bounded by the retrans setting.
Only the asynchronous read path is affected. The synchronous SMB2_read() caller passes rdata == NULL and the memory registration block is guarded on rdata.
The memory registration failure path was pointed out by the Sashiko AI reviewer while it was reviewing an unrelated patch to smb2_async_readv().
---
Tested by fault injection, not on real hardware. I have no SMB Direct setup, so a throwaway debug patch forced the memory registration failure branch on both the read and the write side, with a countdown module parameter, and the resulting error paths were measured. The mount was ordinary SMB2 over TCP; nothing was simulated beyond making the registration return NULL.
Detector: small_buf_alloc_count, the counter behind "SMB Small Req/Resp Buffer" in /proc/fs/cifs/Stats, incremented in cifs_small_buf_get() and decremented in cifs_small_buf_release(). Second detector: kmemleak.
Same kernel, same test, without and with the patch:
unpatched patched
clean read, no injection +0 +0 5 reads, one forced read MR failure each +5 0 5 writes, one forced write MR failure each +0 +0 buffers still allocated after umount 5 0 kmemleak objects under smb2_new_read_req 1 0
Both runs consumed all five injected failures on each side and returned the same errors to userspace, so the only difference is the leak. The write column is the control: the same forced failure on the path that already has the release does not leak.
kmemleak on the unpatched run, the 448 byte object's contents starting \xfeSMB:
unreferenced object 0xffffa241013cca80 (size 448):
comm "dd"
backtrace:
cifs_small_buf_get+0x15/0x30
__smb2_plain_req_init+0x33/0x230
smb2_new_read_req.constprop.0+0x93/0x2e0
smb2_async_readv+0xfd/0x3c0
cifs_issue_read+0x87/0x170kmemleak found one of the five, which is the usual false negative when the address is still lying around in stale stack memory. The counter found all five.
Two limits worth stating. The five failures came from five separate reads rather than five retries of one read: in this configuration the read was unbuffered and netfs returned EAGAIN to userspace instead of reissuing, while the write side did retry and completed successfully. And forcing the branch says nothing about how often a real memory registration fails, so the reachability argument in the changelog is from reading the code, not from measurement.
Build: x86_64, gcc 15.2.1, W=1, clean in both CONFIG_CIFS_SMB_DIRECT=y, which is what compiles the changed hunk, and CONFIG_CIFS_SMB_DIRECT=n, which checks that the new label is still reached. checkpatch --strict reports 0 errors, 0 warnings, 0 checks. Applies to cifs-2.6/for-next at fa724e235cfd and on top of the earlier patch in this thread.
Not marked for stable. The leak is real but I cannot say how often the path is taken in the field. If you think an unbounded leak under persistent memory registration failure warrants a backport, please add the tag.
Per Documentation/process/generated-content.rst: this patch, the analysis in its changelog and the test harness were produced with the assistance of Claude (claude-opus-5). The memory registration failure path was surfaced by the Sashiko AI reviewer on an unrelated patch. The claims were checked against the tree rather than taken from the tools, and the numbers above are from runs I can reproduce.
fs/smb/client/smb2pdu.c | 16 ++++++++++++---- 1 file changed, 12 insertions(+), 4 deletions(-)
fs/smb/client/smb2pdu.c +12 -4
@@ -4564,8 +4564,10 @@ smb2_new_read_req(void **buf, unsigned int *total_len,if (rc)return rc; -if (server ==NULL ) -return -ECONNABORTED ; +if (!server) { + rc = -ECONNABORTED ; + goto free_req; + } shdr = &req->hdr; shdr->Id.SyncId.ProcessId =cpu_to_le32 (io_parms->pid);
@@ -4596,8 +4598,10 @@ smb2_new_read_req(void **buf, unsigned int *total_len, rdata->mr =smbd_register_mr (server->smbd_conn, &rdata->subreq.io_iter, true, need_invalidate); -if (!rdata->mr) -return -EAGAIN ; +if (!rdata->mr) { + rc = -EAGAIN ; + goto free_req; + } req->Channel =SMB2_CHANNEL_RDMA_V1_INVALIDATE ;if (need_invalidate)
@@ -4638,6 +4642,10 @@ smb2_new_read_req(void **buf, unsigned int *total_len, *buf = req;return rc; + +free_req: +cifs_small_buf_release (req); +return rc; } static void
Signature · 1 line
2.54.0
Namjae Jeon
maintainerQuoted 62 lines · Christopher Lusk
smb2_new_read_req()allocates the request buffer withsmb2_plain_req_init()but only publishes it to the caller with *buf = req at the very end of the function. Two error returns sit in between:rc = smb2_plain_req_init(SMB2_READ, io_parms->tcon, server, (void **) &req, total_len); if (rc) return rc; if (server == NULL) return -ECONNABORTED; [...] rdata->mr = smbd_register_mr(server->smbd_conn, &rdata->subreq.io_iter, true, need_invalidate); if (!rdata->mr) return -EAGAIN;On either of them the buffer is neither released nor handed back, so it is leaked. The caller cannot clean up after it:
smb2_async_readv()does 'goto out' on a non-zero return, which skips thecifs_small_buf_release(buf)atasync_readv_out, and buf has not been assigned at that point in any case.The write path has never had this problem.
smb2_async_writev()registers the memory region inline and jumps to its release label instead of returning:wdata->mr = smbd_register_mr(...); if (!wdata->mr) { rc = -EAGAIN; goto async_writev_out; }Commit
b7972092199f("cifs: smbd: Retry on memory registration failure") changed both sides from-ENOBUFSto-EAGAINin a single patch, which puts the two shapes next to each other.Only the
-EAGAINreturn is reachable in practice, becausesmb2_plain_req_init()callssmb2_reconnect()first and that already fails with-EIOwhen server is NULL, before anything is allocated. Both returns are given the same treatment here rather than leaving one of them correct only by accident.Because
-EAGAINis a replayable error, the failure also reaches the retry block at the end ofsmb2_async_readv(), which marks the subrequestNETFS_SREQ_NEED_RETRY, so a failing registration can be retried rather than ending the I/O, and every attempt that reaches it leaks another buffer.smb2_should_replay()short-circuits ontcon->retry, so on a hard mount the attempt count is not bounded by the retrans setting.Only the asynchronous read path is affected. The synchronous
SMB2_read()caller passes rdata == NULL and the memory registration block is guarded on rdata.The memory registration failure path was pointed out by the Sashiko AI reviewer while it was reviewing an unrelated patch to
smb2_async_readv().
Quoted 5 lines
Fixes bd3dcc6a22a9 ("CIFS: SMBD: Upper layer performs SMB read via RDMA write through memory registration")Assisted-by Claude:claude-opus-5Signed-off-by Christopher Lusk <[email protected]>
I will apply it to #for-next.
Thanks!