]> git.ipfire.org Git - thirdparty/kernel/stable.git/commitdiff
nfs: fix UAF in direct writes
authorJosef Bacik <josef@toxicpanda.com>
Fri, 1 Mar 2024 16:49:57 +0000 (11:49 -0500)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Wed, 3 Apr 2024 13:32:17 +0000 (15:32 +0200)
[ Upstream commit 17f46b803d4f23c66cacce81db35fef3adb8f2af ]

In production we have been hitting the following warning consistently

------------[ cut here ]------------
refcount_t: underflow; use-after-free.
WARNING: CPU: 17 PID: 1800359 at lib/refcount.c:28 refcount_warn_saturate+0x9c/0xe0
Workqueue: nfsiod nfs_direct_write_schedule_work [nfs]
RIP: 0010:refcount_warn_saturate+0x9c/0xe0
PKRU: 55555554
Call Trace:
 <TASK>
 ? __warn+0x9f/0x130
 ? refcount_warn_saturate+0x9c/0xe0
 ? report_bug+0xcc/0x150
 ? handle_bug+0x3d/0x70
 ? exc_invalid_op+0x16/0x40
 ? asm_exc_invalid_op+0x16/0x20
 ? refcount_warn_saturate+0x9c/0xe0
 nfs_direct_write_schedule_work+0x237/0x250 [nfs]
 process_one_work+0x12f/0x4a0
 worker_thread+0x14e/0x3b0
 ? ZSTD_getCParams_internal+0x220/0x220
 kthread+0xdc/0x120
 ? __btf_name_valid+0xa0/0xa0
 ret_from_fork+0x1f/0x30

This is because we're completing the nfs_direct_request twice in a row.

The source of this is when we have our commit requests to submit, we
process them and send them off, and then in the completion path for the
commit requests we have

if (nfs_commit_end(cinfo.mds))
nfs_direct_write_complete(dreq);

However since we're submitting asynchronous requests we sometimes have
one that completes before we submit the next one, so we end up calling
complete on the nfs_direct_request twice.

The only other place we use nfs_generic_commit_list() is in
__nfs_commit_inode, which wraps this call in a

nfs_commit_begin();
nfs_commit_end();

Which is a common pattern for this style of completion handling, one
that is also repeated in the direct code with get_dreq()/put_dreq()
calls around where we process events as well as in the completion paths.

Fix this by using the same pattern for the commit requests.

Before with my 200 node rocksdb stress running this warning would pop
every 10ish minutes.  With my patch the stress test has been running for
several hours without popping.

Signed-off-by: Josef Bacik <josef@toxicpanda.com>
Cc: stable@vger.kernel.org
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>
fs/nfs/direct.c
fs/nfs/write.c
include/linux/nfs_fs.h

index c03926a1cc73bc42af75cfa9e0fba8dc5b134258..d8cd4f69e33a02c27e95e9a2e82f401006de5e8e 100644 (file)
@@ -667,10 +667,17 @@ static void nfs_direct_commit_schedule(struct nfs_direct_req *dreq)
        LIST_HEAD(mds_list);
 
        nfs_init_cinfo_from_dreq(&cinfo, dreq);
+       nfs_commit_begin(cinfo.mds);
        nfs_scan_commit(dreq->inode, &mds_list, &cinfo);
        res = nfs_generic_commit_list(dreq->inode, &mds_list, 0, &cinfo);
-       if (res < 0) /* res == -ENOMEM */
-               nfs_direct_write_reschedule(dreq);
+       if (res < 0) { /* res == -ENOMEM */
+               spin_lock(&dreq->lock);
+               if (dreq->flags == 0)
+                       dreq->flags = NFS_ODIRECT_RESCHED_WRITES;
+               spin_unlock(&dreq->lock);
+       }
+       if (nfs_commit_end(cinfo.mds))
+               nfs_direct_write_complete(dreq);
 }
 
 static void nfs_direct_write_clear_reqs(struct nfs_direct_req *dreq)
index bb79d3a886ae83d15395371ec735b7d0e6075bae..5d9dc6c053255f4c38100f3e96346a67223cd6c7 100644 (file)
@@ -1650,7 +1650,7 @@ static int wait_on_commit(struct nfs_mds_commit_info *cinfo)
                                       !atomic_read(&cinfo->rpcs_out));
 }
 
-static void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
+void nfs_commit_begin(struct nfs_mds_commit_info *cinfo)
 {
        atomic_inc(&cinfo->rpcs_out);
 }
index f5ce7b1011461385968723826f4052f3f492a4af..d59116ac82099d0b5985a7f2b57a2e69e9e860dc 100644 (file)
@@ -611,6 +611,7 @@ int nfs_wb_folio_cancel(struct inode *inode, struct folio *folio);
 extern int  nfs_commit_inode(struct inode *, int);
 extern struct nfs_commit_data *nfs_commitdata_alloc(void);
 extern void nfs_commit_free(struct nfs_commit_data *data);
+void nfs_commit_begin(struct nfs_mds_commit_info *cinfo);
 bool nfs_commit_end(struct nfs_mds_commit_info *cinfo);
 
 static inline bool nfs_have_writebacks(const struct inode *inode)