Skip to contentSkip to conversation

netfs: Fix missing xas_retry() calls in xarray iteration

PATCH Superseded David Howells 2 participants 3 messages Nov 3, 2022 – Nov 4, 2022

Subsystems (2)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore

David Howells

author maintainer
#0 lore raw

PATCH 1/1 netfs: Fix missing xas_retry() calls in xarray iteration

netfslib has a number of places in which it performs iteration of an xarray whilst being under the RCU read lock. It *should* call xas_retry() as the first thing inside of the loop and do "continue" if it returns true in case the xarray walker passed out a special value indicating that the walk needs to be redone from the root[*].

Fix this by adding the missing retry checks.

[*] I wonder if this should be done inside xas_find(), xas_next_node() and
suchlike, but I'm told that's not an simple change to effect.

This can cause an oops like that below. Note the faulting address - this is an internal value (|0x2) returned from xarray.

BUG: kernel NULL pointer dereference, address: 0000000000000402
...
RIP: 0010:netfs_rreq_unlock+0xef/0x380 [netfs]
...
Call Trace:
 netfs_rreq_assess+0xa6/0x240 [netfs]
 netfs_readpage+0x173/0x3b0 [netfs]
 ? init_wait_var_entry+0x50/0x50
 filemap_read_page+0x33/0xf0
 filemap_get_pages+0x2f2/0x3f0
 filemap_read+0xaa/0x320
 ? do_filp_open+0xb2/0x150
 ? rmqueue+0x3be/0xe10
 ceph_read_iter+0x1fe/0x680 [ceph]
 ? new_sync_read+0x115/0x1a0
 new_sync_read+0x115/0x1a0
 vfs_read+0xf3/0x180
 ksys_read+0x5f/0xe0
 do_syscall_64+0x38/0x90
 entry_SYSCALL_64_after_hwframe+0x44/0xae
Fixes 3d3c95046742 ("netfs: Provide readahead and readpage netfs helpers")
cc Matthew Wilcox <[email protected]>

---

 fs/netfs/buffered_read.c |    9 +++++++--
 fs/netfs/io.c            |    3 +++
 2 files changed, 10 insertions(+), 2 deletions(-)
fs/netfs/buffered_read.c +7 -2
index 0ce535852151..baf668fb4315 100644
@@ -46,10 +46,15 @@ void netfs_rreq_unlock_folios(struct netfs_io_request *rreq)
 
 	rcu_read_lock();
 	xas_for_each(&xas, folio, last_page) {
-		unsigned int pgpos = (folio_index(folio) - start_page) * PAGE_SIZE;
-		unsigned int pgend = pgpos + folio_size(folio);
+		unsigned int pgpos, pgend;
 		bool pg_failed = false;
 
+		if (xas_retry(&xas, folio))
+			continue;
+
+		pgpos = (folio_index(folio) - start_page) * PAGE_SIZE;
+		pgend = pgpos + folio_size(folio);
+
 		for (;;) {
 			if (!subreq) {
 				pg_failed = true;
fs/netfs/io.c +3 -0
index 428925899282..e374767d1b68 100644
@@ -121,6 +121,9 @@ static void netfs_rreq_unmark_after_write(struct netfs_io_request *rreq,
 		XA_STATE(xas, &rreq->mapping->i_pages, subreq->start / PAGE_SIZE);
 
 		xas_for_each(&xas, folio, (subreq->start + subreq->len - 1) / PAGE_SIZE) {
+			if (xas_retry(&xas, folio))
+				continue;
+
 			/* We might have multiple writes from the same huge
 			 * folio, but we mustn't unlock a folio more than once.
 			 */
#1 lore raw

fs/netfs/buffered_read.c void netfs_rreq_unlock_folios(struct netfs_io_request *rreq) · line 49

Quoted 8 lines · David Howells

+++ b/fs/netfs/buffered_read.c

fs/netfs/buffered_read.c +1 -2
@@ -46,10 +46,15 @@ void netfs_rreq_unlock_folios(struct netfs_io_request *rreq)
 
 	rcu_read_lock();
 	xas_for_each(&xas, folio, last_page) {
-		unsigned int pgpos = (folio_index(folio) - start_page) * PAGE_SIZE;
-		unsigned int pgend = pgpos + folio_size(folio);
+		unsigned int pgpos, pgend;

"unsigned int" assumes that the number of bytes isn't going to exceed 32 bits. I tend to err on the side of safety here and use size_t.

fs/netfs/buffered_read.c void netfs_rreq_unlock_folios(struct netfs_io_request *rreq) · line 56

Quoted 7 lines
fs/netfs/buffered_read.c +5 -0
 		bool pg_failed = false;
 
+		if (xas_retry(&xas, folio))
+			continue;
+
+		pgpos = (folio_index(folio) - start_page) * PAGE_SIZE;
+		pgend = pgpos + folio_size(folio);

What happens if start_page is somewhere inside folio? Seems to me that pgend ends up overhanging into the next folio?

author maintainer
#2 lore raw
Quoted 2 lines · Matthew Wilcox

"unsigned int" assumes that the number of bytes isn't going to exceed 32 bits. I tend to err on the side of safety here and use size_t.

Not unreasonable.

Quoted 5 lines

fs/netfs/buffered_read.c void netfs_rreq_unlock_folios(struct netfs_io_request *rreq) · line 56

Quoted 2 lines
fs/netfs/buffered_read.c +2 -0
+		pgpos = (folio_index(folio) - start_page) * PAGE_SIZE;
+		pgend = pgpos + folio_size(folio);

What happens if start_page is somewhere inside folio? Seems to me that pgend ends up overhanging into the next folio?

Yeah, I think my maths is dodgy. I should probably use folio_pos() and/or offset_in_folio().

David