Skip to contentSkip to conversation

smb: client: do not let the server grow a write subrequest

PATCH Applied ozgun kultekin 3 participants 3 messages Nov 18 – Nov 19

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 smb: client: do not let the server grow a write subrequest

smb2_writev_callback() takes the byte count from the SMB2 WRITE response, masks it with 0xFFFF if it is larger than the request (for old OS/2 servers), and then copies it into wdata->subreq.len. A server can return a count that is still larger than the request after the mask, for example 0x8002 for a 5-byte write or 0xFFFF for a 4 KiB write, and the subrequest then covers data that was never sent.

netfs uses the subrequest length to decide how far the write collector may advance. With a grown subrequest the collector moves collected_to past the data that was issued: netfs_writeback_unlock_folios() walks folio_queue slots that were never filled, and rolling_buffer_delete_spent() frees folio_queue segments that later in-flight subrequests still use. This is a use-after-free that a malicious or broken server triggers from an ordinary write:

  BUG: KASAN: slab-use-after-free in netfs_limit_iter+0xed4/0x1040
  Read of size 1 at addr ff11000010561511 by task kworker/u8:4/63
  Workqueue: events_unbound netfs_write_collection_worker
   netfs_limit_iter+0xed4/0x1040
   netfs_retry_writes+0x1410/0x1e90
   netfs_write_collection+0x2647/0x3880
   netfs_write_collection_worker+0x1cd/0x3e0
  Allocated by task 141:
   netfs_folioq_alloc+0x39a/0x470
   rolling_buffer_init+0x29/0xc0
   netfs_create_write_req+0x12f/0x840
  Freed by task 42:
   rolling_buffer_delete_spent+0xb1/0x220
   netfs_write_collection+0x1f62/0x3880

Never change the subrequest length from the response. If the count is still larger than the request after the mask, fail the write with -EIO and record it with a new write_overlarge trace point, like the read path does with read_overlarge. A short count still returns -ENOSPC as before.

The folio_queue form of the use-after-free exists since commit cd0277ed0c18 ("netfs: Use new folio_queue data type and iterator instead of xarray iter").

Fixes 3ee1a1fc3981 ("cifs: Cut over to using netfslib")
---
 fs/smb/client/smb2pdu.c | 17 +++++++++++++----
 fs/smb/client/trace.h   |  1 +
 2 files changed, 14 insertions(+), 4 deletions(-)
fs/smb/client/smb2pdu.c +13 -4
index b26bfb4bd..2f05ba613 100644
@@ -5181,14 +5181,23 @@ smb2_writev_callback(struct TCP_Server_Info *server, struct mid_q_entry *mid)
 		if (written > wdata->subreq.len)
 			written &= 0xFFFF;
 
+		/*
+		 * The server must not report more bytes than were sent.  netfs
+		 * uses the subrequest length to work out which folios this
+		 * write has finished with, so it must not be changed here.
+		 */
+		if (written > wdata->subreq.len) {
+			result = smb_EIO2(smb_eio_trace_write_overlarge,
+					  written, wdata->subreq.len);
+			break;
+		}
+
 		cifs_stats_bytes_written(tcon, written);
 
-		if (written < wdata->subreq.len) {
+		if (written < wdata->subreq.len)
 			result = -ENOSPC;
-		} else if (written > 0) {
-			wdata->subreq.len = written;
+		else if (written > 0)
 			__set_bit(NETFS_SREQ_MADE_PROGRESS, &wdata->subreq.flags);
-		}
 		break;
 	case MID_REQUEST_SUBMITTED:
 		trace_netfs_sreq(&wdata->subreq, netfs_sreq_trace_io_req_submitted);
fs/smb/client/trace.h +1 -0
index 5331ce6dc..18760a819 100644
@@ -153,6 +153,7 @@
 	EM(smb_eio_trace_user_iter,			"user_iter") \
 	EM(smb_eio_trace_write_bad_buf_type,		"write_bad_buf_type") \
 	EM(smb_eio_trace_write_mid_state_unknown,	"write_mid_state_unknown") \
+	EM(smb_eio_trace_write_overlarge,		"write_overlarge") \
 	EM(smb_eio_trace_write_rsp_malformed,		"write_rsp_malformed") \
 	E_(smb_eio_trace_write_too_far,			"write_too_far")
 
Signature · 1 line
2.39.5
maintainer Reviewed-by on v1 1/1
#1 lore raw
Quoted 37 lines · ozgun kultekin

smb2_writev_callback() takes the byte count from the SMB2 WRITE response, masks it with 0xFFFF if it is larger than the request (for old OS/2 servers), and then copies it into wdata->subreq.len. A server can return a count that is still larger than the request after the mask, for example 0x8002 for a 5-byte write or 0xFFFF for a 4 KiB write, and the subrequest then covers data that was never sent.

netfs uses the subrequest length to decide how far the write collector may advance. With a grown subrequest the collector moves collected_to past the data that was issued: netfs_writeback_unlock_folios() walks folio_queue slots that were never filled, and rolling_buffer_delete_spent() frees folio_queue segments that later in-flight subrequests still use. This is a use-after-free that a malicious or broken server triggers from an ordinary write:

  BUG: KASAN: slab-use-after-free in netfs_limit_iter+0xed4/0x1040
  Read of size 1 at addr ff11000010561511 by task kworker/u8:4/63
  Workqueue: events_unbound netfs_write_collection_worker
   netfs_limit_iter+0xed4/0x1040
   netfs_retry_writes+0x1410/0x1e90
   netfs_write_collection+0x2647/0x3880
   netfs_write_collection_worker+0x1cd/0x3e0
  Allocated by task 141:
   netfs_folioq_alloc+0x39a/0x470
   rolling_buffer_init+0x29/0xc0
   netfs_create_write_req+0x12f/0x840
  Freed by task 42:
   rolling_buffer_delete_spent+0xb1/0x220
   netfs_write_collection+0x1f62/0x3880

Never change the subrequest length from the response. If the count is still larger than the request after the mask, fail the write with -EIO and record it with a new write_overlarge trace point, like the read path does with read_overlarge. A short count still returns -ENOSPC as before.

Quoted 7 lines

The folio_queue form of the use-after-free exists since commit cd0277ed0c18 ("netfs: Use new folio_queue data type and iterator instead of xarray iter").

Fixes 3ee1a1fc3981 ("cifs: Cut over to using netfslib")

Looks good.

Thanks!

maintainer Applied
#2 lore raw
In reply to ozgun kultekin
Quoted 30 lines · ozgun kultekin

smb2_writev_callback() takes the byte count from the SMB2 WRITE response, masks it with 0xFFFF if it is larger than the request (for old OS/2 servers), and then copies it into wdata->subreq.len. A server can return a count that is still larger than the request after the mask, for example 0x8002 for a 5-byte write or 0xFFFF for a 4 KiB write, and the subrequest then covers data that was never sent.

netfs uses the subrequest length to decide how far the write collector may advance. With a grown subrequest the collector moves collected_to past the data that was issued: netfs_writeback_unlock_folios() walks folio_queue slots that were never filled, and rolling_buffer_delete_spent() frees folio_queue segments that later in-flight subrequests still use. This is a use-after-free that a malicious or broken server triggers from an ordinary write:

  BUG: KASAN: slab-use-after-free in netfs_limit_iter+0xed4/0x1040
  Read of size 1 at addr ff11000010561511 by task kworker/u8:4/63
  Workqueue: events_unbound netfs_write_collection_worker
   netfs_limit_iter+0xed4/0x1040
   netfs_retry_writes+0x1410/0x1e90
   netfs_write_collection+0x2647/0x3880
   netfs_write_collection_worker+0x1cd/0x3e0
  Allocated by task 141:
   netfs_folioq_alloc+0x39a/0x470
   rolling_buffer_init+0x29/0xc0
   netfs_create_write_req+0x12f/0x840
  Freed by task 42:
   rolling_buffer_delete_spent+0xb1/0x220
   netfs_write_collection+0x1f62/0x3880
Quoted 6 lines

Never change the subrequest length from the response. If the count is still larger than the request after the mask, fail the write with -EIO and record it with a new write_overlarge trace point, like the read path does with read_overlarge. A short count still returns -ENOSPC as before.
...

Applied.