block: Add blk_co_start/end_request() and BDRV_REQ_NO_QUEUE
If a device uses blk_inc/dec_in_flight() in order to build macro operations that involve multiple requests for the block layer and that need to be completed as a unit before the BlockBackend can be considered drained, it sets the stage for a deadlock: When a drain is requested, the inner request at the BlockBackend level will be queued in blk_wait_while_drained() and wait until the drained section ends, but at the same time, drain_begin can only return if the whole macro operation at the device level has completed. Introduce a new interface to allow implementing the logic correctly: Instead of queueing individual requests, blk_co_start_request() calls blk_wait_while_drained() once at the beginning. The individual requests must then set BDRV_REQ_NO_QUEUE to avoid being queued and running into the deadlock; being wrapped in blk_co_start/end_request() makes sure that drain_begin waits for them and they don't sneak in when the BlockBackend is supposed to already be quiescent. Signed-off-by: Kevin Wolf <kwolf@redhat.com> Message-ID: <20260421161132.99878-3-kwolf@redhat.com> Signed-off-by: Kevin Wolf <kwolf@redhat.com>
Kevin Wolf committed
Apr 21, 2026 at 18:11 UTC
34a67637767d3ed1ac813c44effe827bbfba5996
3 files changed
+41
-10
block/block-backend.c
+29
-9
@@ -82,6 +82,7 @@ struct BlockBackend {
82
QemuMutex queued_requests_lock; /* protects queued_requests */
83
CoQueue queued_requests;
84
bool disable_request_queuing; /* atomic */
85
+ int start_request_count; /* atomic */
86
87
VMChangeStateEntry *vmsh;
88
bool force_allow_inactivate;
@@ -1306,10 +1307,16 @@ bool blk_in_drain(BlockBackend *blk)
1307
}
1308
1309
/* To be called between exactly one pair of blk_inc/dec_in_flight() */
1309
-static void coroutine_fn blk_wait_while_drained(BlockBackend *blk)
1310
+static void coroutine_fn blk_wait_while_drained(BlockBackend *blk,
1311
+ BdrvRequestFlags flags)
1312
{
1313
assert(blk->in_flight > 0);
1314
1315
+ if (flags & BDRV_REQ_NO_QUEUE) {
1316
+ assert(qatomic_read(&blk->start_request_count));
1317
+ return;
1318
+ }
1319
+
1320
if (qatomic_read(&blk->quiesce_counter) &&
1321
!qatomic_read(&blk->disable_request_queuing)) {
1322
/*
@@ -1335,7 +1342,7 @@ blk_co_do_preadv_part(BlockBackend *blk, int64_t offset, int64_t bytes,
1342
BlockDriverState *bs;
1343
IO_CODE();
1344
1338
- blk_wait_while_drained(blk);
1345
+ blk_wait_while_drained(blk, flags);
1346
GRAPH_RDLOCK_GUARD();
1347
1348
/* Call blk_bs() only after waiting, the graph may have changed */
@@ -1410,7 +1417,7 @@ blk_co_do_pwritev_part(BlockBackend *blk, int64_t offset, int64_t bytes,
1417
BlockDriverState *bs;
1418
IO_CODE();
1419
1413
- blk_wait_while_drained(blk);
1420
+ blk_wait_while_drained(blk, flags);
1421
GRAPH_RDLOCK_GUARD();
1422
1423
/* Call blk_bs() only after waiting, the graph may have changed */
@@ -1523,6 +1530,19 @@ void blk_dec_in_flight(BlockBackend *blk)
1530
aio_wait_kick();
1531
}
1532
1533
+void coroutine_fn blk_co_start_request(BlockBackend *blk)
1534
+{
1535
+ blk_inc_in_flight(blk);
1536
+ blk_wait_while_drained(blk, 0);
1537
+ qatomic_inc(&blk->start_request_count);
1538
+}
1539
+
1540
+void blk_end_request(BlockBackend *blk)
1541
+{
1542
+ qatomic_dec(&blk->start_request_count);
1543
+ blk_dec_in_flight(blk);
1544
+}
1545
+
1546
static void error_callback_bh(void *opaque)
1547
{
1548
struct BlockBackendAIOCB *acb = opaque;
@@ -1741,7 +1761,7 @@ blk_co_do_ioctl(BlockBackend *blk, unsigned long int req, void *buf)
1761
{
1762
IO_CODE();
1763
1744
- blk_wait_while_drained(blk);
1764
+ blk_wait_while_drained(blk, 0);
1765
GRAPH_RDLOCK_GUARD();
1766
1767
if (!blk_co_is_available(blk)) {
@@ -1788,7 +1808,7 @@ blk_co_do_pdiscard(BlockBackend *blk, int64_t offset, int64_t bytes)
1808
int ret;
1809
IO_CODE();
1810
1791
- blk_wait_while_drained(blk);
1811
+ blk_wait_while_drained(blk, 0);
1812
GRAPH_RDLOCK_GUARD();
1813
1814
ret = blk_check_byte_request(blk, offset, bytes);
@@ -1834,7 +1854,7 @@ int coroutine_fn blk_co_pdiscard(BlockBackend *blk, int64_t offset,
1854
static int coroutine_fn blk_co_do_flush(BlockBackend *blk)
1855
{
1856
IO_CODE();
1837
- blk_wait_while_drained(blk);
1857
+ blk_wait_while_drained(blk, 0);
1858
GRAPH_RDLOCK_GUARD();
1859
1860
if (!blk_co_is_available(blk)) {
@@ -2009,7 +2029,7 @@ int coroutine_fn blk_co_zone_report(BlockBackend *blk, int64_t offset,
2029
IO_CODE();
2030
2031
blk_inc_in_flight(blk); /* increase before waiting */
2012
- blk_wait_while_drained(blk);
2032
+ blk_wait_while_drained(blk, 0);
2033
GRAPH_RDLOCK_GUARD();
2034
if (!blk_is_available(blk)) {
2035
blk_dec_in_flight(blk);
@@ -2034,7 +2054,7 @@ int coroutine_fn blk_co_zone_mgmt(BlockBackend *blk, BlockZoneOp op,
2054
IO_CODE();
2055
2056
blk_inc_in_flight(blk);
2037
- blk_wait_while_drained(blk);
2057
+ blk_wait_while_drained(blk, 0);
2058
GRAPH_RDLOCK_GUARD();
2059
2060
ret = blk_check_byte_request(blk, offset, len);
@@ -2058,7 +2078,7 @@ int coroutine_fn blk_co_zone_append(BlockBackend *blk, int64_t *offset,
2078
IO_CODE();
2079
2080
blk_inc_in_flight(blk);
2061
- blk_wait_while_drained(blk);
2081
+ blk_wait_while_drained(blk, flags);
2082
GRAPH_RDLOCK_GUARD();
2083
if (!blk_is_available(blk)) {
2084
blk_dec_in_flight(blk);
include/block/block-common.h
+10
-1
@@ -215,8 +215,17 @@ typedef enum {
215
*/
216
BDRV_REQ_NO_WAIT = 0x400,
217
218
+ /*
219
+ * Used between blk_co_start_request() and blk_end_request() to avoid
220
+ * that the request waits in a drained BlockBackend until the drained
221
+ * section ends. Waiting would cause a deadlock because drain waits for
222
+ * blk_end_request() to be called, but the request never completes
223
+ * because it waits for the drain to end.
224
+ */
225
+ BDRV_REQ_NO_QUEUE = 0x800,
226
+
227
/* Mask of valid flags */
219
- BDRV_REQ_MASK = 0x7ff,
228
+ BDRV_REQ_MASK = 0xfff,
229
} BdrvRequestFlags;
230
231
#define BDRV_O_NO_SHARE 0x0001 /* don't share permissions */
include/system/block-backend-io.h
+2
@@ -71,6 +71,8 @@ BlockAIOCB *blk_aio_ioctl(BlockBackend *blk, unsigned long int req, void *buf,
71
72
void blk_inc_in_flight(BlockBackend *blk);
73
void blk_dec_in_flight(BlockBackend *blk);
74
+void coroutine_fn blk_co_start_request(BlockBackend *blk);
75
+void blk_end_request(BlockBackend *blk);
76
77
bool coroutine_fn GRAPH_RDLOCK blk_co_is_inserted(BlockBackend *blk);
78
bool co_wrapper_mixed_bdrv_rdlock blk_is_inserted(BlockBackend *blk);