p9_client_write doesn't check for negative "written" from server
If the kernel's 9p client sends a write request, and the server returns a successful Rwrite reply, but with a negative "written" value, then this line towards the end of p9_client_write():
iov_iter_revert(from, count - written - iov_iter_count(from));
can pass an unroll argument to iov_iter_revert() that's larger than the number of bytes in the iov. This causes iov_iter_folioq_revert() to try to back up beyond the beginning of the iov data, and this code
if (slot == 0) {
folioq = folioq->prev;
slot = folioq_nr_slots(folioq);tries to dereference a NULL folioq, since folioq->prev is NULL.
I've attached a demo.
# cc 9p2a.c # ./a.out ... BUG: kernel NULL pointer dereference, address: 000000000000011e #PF: supervisor read access in kernel mode #PF: error_code(0x0000) - not-present page PGD 0 P4D 0 Oops: Oops: 0000 [#1] SMP DEBUG_PAGEALLOC PTI CPU: 1 UID: 0 PID: 1613 Comm: sh Not tainted 6.13.0-rc3-00017-gf44d154d6e3d #12 Hardware name: FreeBSD BHYVE/BHYVE, BIOS 14.0 10/17/2021 RIP: 0010:iov_iter_revert+0xc9/0x150 Code: cc cc cc 48 29 f1 48 89 4f 08 c3 cc cc cc cc 48 c7 47 08 00 00 00 00 0f b6 57 20 bf 00 10 00 00 eb 1f 83 ea 01 89 d1 49 89 f9 <41> 0f b6 8c 08 00 01 00 00 49 d3 e1 4c 89 c9 49 39 f1 73 15 4c 29 RSP: 0018:ffffc900020dfae8 EFLAGS: 00010246 Call Trace: <TASK> ? __die+0x1e/0x60 ? page_fault_oops+0x157/0x450 ? check_object+0xc2/0x430 ? exc_page_fault+0x66/0x140 ? asm_exc_page_fault+0x26/0x30 ? entry_SYSCALL_64_after_hwframe+0x76/0x7e ? iov_iter_revert+0xc9/0x150 p9_client_write+0x115/0x250 v9fs_issue_write+0x38/0x70 netfs_advance_write+0xc6/0x2a0 netfs_write_folio+0x2e7/0x9a0 netfs_writepages+0xd7/0x3e0 ? filemap_dirty_folio+0x5b/0x70 do_writepages+0xc6/0x240 __filemap_fdatawrite_range+0x9a/0xb0 v9fs_dir_release+0x153/0x180 ? task_work_add+0x71/0xe0 __fput+0xe8/0x290 task_work_run+0x57/0xa0 syscall_exit_to_user_mode+0x105/0x110 do_syscall_64+0x4b/0xd0 entry_SYSCALL_64_after_hwframe+0x76/0x7e RIP: 0033:0x7fdfc55fe96b
Robert Morris
[email protected]
Dominique Martinet
maintainerHi,
[email protected] wrote on Tue, Dec 17, 2024 at 10:17:11AM -0500:
Quoted 8 lines
If the kernel's 9p client sends a write request, and the server returns a successful Rwrite reply, but with a negative "written" value, then this line towards the end of
p9_client_write():iov_iter_revert(from, count - written - iov_iter_count(from));can pass an unroll argument to
iov_iter_revert()that's larger than the number of bytes in the iov
Thanks for the report, I'm surprised this didn't come up sooner through syzbot...
That is correct, written (and simlarily received in p9_client_read_once) should be made unsigned.
They're actually compared with 'rsize' which is an int but is set from fid->iounit which is a u32, so just making these three variables u32 will sort the issue.
The patch is trivial so would you like to send it?
If you don't want to (or there is no reply in a while) then I'll send a patch this weekend crediting you as reported-by.
Signature · 1 line
Dominique