Skip to contentSkip to conversation

Improve changing the number of hardware queues

PATCH In review Bart Van Assche 2 participants 10 messages Mar 10, 2020 Block layer

lore
Share & tools
lore
#0 lore raw

v4 cover Improve changing the number of hardware queues

Hi Jens,

These patches are what I came up with while analyzing syzbot and blktests complaints related to dynamically changing the number of hardware queues. Please consider these patches for the upstream kernel.

Thanks,

Bart.

Changes compared to v3:
- Added more Reviewed-by tags.

Changes compared to v2:
- Changed patch 2/8 such that .nr_queues is only set if nr_maps == 1.

Changes compared to v1:
- Reworked patch 2/8 as requested by Ming.
- Added three new null_blk patches.

Bart Van Assche (8):
blk-mq: Fix a comment in include/linux/blk-mq.h
blk-mq: Keep set->nr_hw_queues and set->map[].nr_queues in sync
blk-mq: Fix a recently introduced regression in
blk_mq_realloc_hw_ctxs()
null_blk: Suppress an UBSAN complaint triggered when setting
'memory_backed'
null_blk: Fix changing the number of hardware queues
null_blk: Fix the null_add_dev() error path
null_blk: Handle null_add_dev() failures properly
null_blk: Add support for init_hctx() fault injection

 block/blk-mq.c                |   9 ++-
 drivers/block/null_blk_main.c | 124 +++++++++++++++++++++++-----------
 include/linux/blk-mq.h        |   5 +-
 3 files changed, 97 insertions(+), 41 deletions(-)
#1 lore raw

PATCH v4 1/8 blk-mq: Fix a comment in include/linux/blk-mq.h

The 'hctx_list' member of struct blk_mq_hw_ctx is not a list head but instead an entry in q->unused_hctx_list. Fix the comment above this struct member.

Cc André Almeida <[email protected]>
Fixes d386732bc142 ("blk-mq: fill header with kernel-doc")
---
 include/linux/blk-mq.h | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)
include/linux/blk-mq.h +4 -1
index 11cfd6470b1a..31344d5f83e2 100644
@@ -162,7 +162,10 @@ struct blk_mq_hw_ctx {
 	struct dentry		*sched_debugfs_dir;
 #endif
 
-	/** @hctx_list:	List of all hardware queues. */
+	/**
+	 * @hctx_list: if this hctx is not in use, this is an entry in
+	 * q->unused_hctx_list.
+	 */
 	struct list_head	hctx_list;
 
 	/**
#2 lore raw

PATCH v4 2/8 blk-mq: Keep set->nr_hw_queues and set->map[].nr_queues in sync

In reply to Bart Van Assche

blk_mq_map_queues() and multiple .map_queues() implementations expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the number of hardware queues. Hence set .nr_queues before calling these functions. This patch fixes the following kernel warning:

WARNING: CPU: 0 PID: 2501 at include/linux/cpumask.h:137
Call Trace:
 blk_mq_run_hw_queue+0x19d/0x350 block/blk-mq.c:1508
 blk_mq_run_hw_queues+0x112/0x1a0 block/blk-mq.c:1525
 blk_mq_requeue_work+0x502/0x780 block/blk-mq.c:775
 process_one_work+0x9af/0x1740 kernel/workqueue.c:2269
 worker_thread+0x98/0xe40 kernel/workqueue.c:2415
 kthread+0x361/0x430 kernel/kthread.c:255
Cc Christoph Hellwig <[email protected]>
Cc Ming Lei <[email protected]>
Cc Hannes Reinecke <[email protected]>
Cc Johannes Thumshirn <[email protected]>
Reported-by [email protected]
Fixes ed76e329d74a ("blk-mq: abstract out queue map") # v5.0
---
 block/blk-mq.c | 8 ++++++++
 1 file changed, 8 insertions(+)
block/blk-mq.c +8 -0
index 7a16d61e3cd3..f174b6ad94cb 100644
@@ -3027,6 +3027,14 @@ static int blk_mq_alloc_rq_maps(struct blk_mq_tag_set *set)
 
 static int blk_mq_update_queue_map(struct blk_mq_tag_set *set)
 {
+	/*
+	 * blk_mq_map_queues() and multiple .map_queues() implementations
+	 * expect that set->map[HCTX_TYPE_DEFAULT].nr_queues is set to the
+	 * number of hardware queues.
+	 */
+	if (set->nr_maps == 1)
+		set->map[HCTX_TYPE_DEFAULT].nr_queues = set->nr_hw_queues;
+
 	if (set->ops->map_queues && !is_kdump_kernel()) {
 		int i;
 
#3 lore raw

PATCH v4 3/8 blk-mq: Fix a recently introduced regression in blk_mq_realloc_hw_ctxs()

In reply to Bart Van Assche

q->nr_hw_queues must only be updated once it is known that blk_mq_realloc_hw_ctxs() has succeeded. Otherwise it can happen that reallocation fails and that q->nr_hw_queues is larger than the number of allocated hardware queues. This patch fixes the following crash if increasing the number of hardware queues fails:

BUG: KASAN: null-ptr-deref in blk_mq_map_swqueue+0x775/0x810
Write of size 8 at addr 0000000000000118 by task check/977

CPU: 3 PID: 977 Comm: check Not tainted 5.6.0-rc1-dbg+ #8
Hardware name: Bochs Bochs, BIOS Bochs 01/01/2011
Call Trace:
 dump_stack+0xa5/0xe6
 __kasan_report.cold+0x65/0x99
 kasan_report+0x16/0x20
 check_memory_region+0x140/0x1b0
 memset+0x28/0x40
 blk_mq_map_swqueue+0x775/0x810
 blk_mq_update_nr_hw_queues+0x468/0x710
 nullb_device_submit_queues_store+0xf7/0x1a0 [null_blk]
 configfs_write_file+0x1c4/0x250 [configfs]
 __vfs_write+0x4c/0x90
 vfs_write+0x145/0x2c0
 ksys_write+0xd7/0x180
 __x64_sys_write+0x47/0x50
 do_syscall_64+0x6f/0x2f0
 entry_SYSCALL_64_after_hwframe+0x49/0xbe
Cc Christoph Hellwig <[email protected]>
Cc Hannes Reinecke <[email protected]>
Cc Johannes Thumshirn <[email protected]>
Cc Keith Busch <[email protected]>
Fixes ac0d6b926e74 ("block: Reduce the amount of memory required per request queue")
---
 block/blk-mq.c | 1 -
 1 file changed, 1 deletion(-)
block/blk-mq.c +0 -1
index f174b6ad94cb..41f7264e51c9 100644
@@ -2828,7 +2828,6 @@ static void blk_mq_realloc_hw_ctxs(struct blk_mq_tag_set *set,
 			memcpy(new_hctxs, hctxs, q->nr_hw_queues *
 			       sizeof(*hctxs));
 		q->queue_hw_ctx = new_hctxs;
-		q->nr_hw_queues = set->nr_hw_queues;
 		kfree(hctxs);
 		hctxs = new_hctxs;
 	}
#4 lore raw

PATCH v4 4/8 null_blk: Suppress an UBSAN complaint triggered when setting 'memory_backed'

In reply to Bart Van Assche

Although it is not clear to me why UBSAN complains when 'memory_backed' is set, this patch suppresses the UBSAN complaint that is triggered when setting that configfs attribute.

UBSAN: Undefined behaviour in drivers/block/null_blk_main.c:327:1
load of value 16 is not a valid value for type '_Bool'
CPU: 2 PID: 8396 Comm: check Not tainted 5.6.0-rc1-dbg+ #14
Hardware name: Bochs Bochs, BIOS Bochs 01/01/2011
Call Trace:
 dump_stack+0xa5/0xe6
 ubsan_epilogue+0x9/0x26
 __ubsan_handle_load_invalid_value+0x6d/0x76
 nullb_device_memory_backed_store.cold+0x2c/0x38 [null_blk]
 configfs_write_file+0x1c4/0x250 [configfs]
 __vfs_write+0x4c/0x90
 vfs_write+0x145/0x2c0
 ksys_write+0xd7/0x180
 __x64_sys_write+0x47/0x50
 do_syscall_64+0x6f/0x2f0
 entry_SYSCALL_64_after_hwframe+0x49/0xbe
Cc Christoph Hellwig <[email protected]>
Cc Ming Lei <[email protected]>
Cc Hannes Reinecke <[email protected]>
Cc Johannes Thumshirn <[email protected]>
---
 drivers/block/null_blk_main.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/block/null_blk_main.c +1 -1
index 133060431dbd..b8a08b82b882 100644
@@ -276,7 +276,7 @@ nullb_device_##NAME##_store(struct config_item *item, const char *page,	\
 {									\
 	int (*apply_fn)(struct nullb_device *dev, TYPE new_value) = APPLY;\
 	struct nullb_device *dev = to_nullb_device(item);		\
-	TYPE uninitialized_var(new_value);				\
+	TYPE new_value = 0;						\
 	int ret;							\
 									\
 	ret = nullb_device_##TYPE##_attr_store(&new_value, page, count);\
#5 lore raw

PATCH v4 5/8 null_blk: Fix changing the number of hardware queues

In reply to Bart Van Assche

Instead of initializing null_blk hardware queues explicitly after the request queue has been created, provide .init_hctx() and .exit_hctx() callback functions. The latter functions are not only called during request queue allocation but also when the number of hardware queues changes. Allocate nr_cpu_ids queues during initialization to support increasing the number of hardware queues above the initial hardware queue count.

This change fixes increasing the number of hardware queues above the initial number of hardware queues and also keeps nullb->nr_queues in sync with the number of hardware queues.

Cc Christoph Hellwig <[email protected]>
Cc Ming Lei <[email protected]>
Cc Hannes Reinecke <[email protected]>
Cc Johannes Thumshirn <[email protected]>
Fixes 45919fbfe1c4 ("null_blk: Enable modifying 'submit_queues' after an instance has been configured")
---
 drivers/block/null_blk_main.c | 103 ++++++++++++++++++++++------------
 1 file changed, 66 insertions(+), 37 deletions(-)
drivers/block/null_blk_main.c +66 -37
index b8a08b82b882..83e31b81db3c 100644
@@ -302,6 +302,12 @@ static int nullb_apply_submit_queues(struct nullb_device *dev,
 	if (!nullb)
 		return 0;
 
+	/*
+	 * Make sure that null_init_hctx() does not access nullb->queues[] past
+	 * the end of that array.
+	 */
+	if (submit_queues > nr_cpu_ids)
+		return -EINVAL;
 	set = nullb->tag_set;
 	blk_mq_update_nr_hw_queues(set, submit_queues);
 	return set->nr_hw_queues == submit_queues ? 0 : -ENOMEM;
@@ -1408,12 +1414,6 @@ static blk_status_t null_queue_rq(struct blk_mq_hw_ctx *hctx,
 	return null_handle_cmd(cmd, sector, nr_sectors, req_op(bd->rq));
 }
 
-static const struct blk_mq_ops null_mq_ops = {
-	.queue_rq       = null_queue_rq,
-	.complete	= null_complete_rq,
-	.timeout	= null_timeout_rq,
-};
-
 static void cleanup_queue(struct nullb_queue *nq)
 {
 	kfree(nq->tag_map);
@@ -1430,6 +1430,43 @@ static void cleanup_queues(struct nullb *nullb)
 	kfree(nullb->queues);
 }
 
+static void null_exit_hctx(struct blk_mq_hw_ctx *hctx, unsigned int hctx_idx)
+{
+	struct nullb_queue *nq = hctx->driver_data;
+	struct nullb *nullb = nq->dev->nullb;
+
+	nullb->nr_queues--;
+}
+
+static void null_init_queue(struct nullb *nullb, struct nullb_queue *nq)
+{
+	init_waitqueue_head(&nq->wait);
+	nq->queue_depth = nullb->queue_depth;
+	nq->dev = nullb->dev;
+}
+
+static int null_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data,
+			  unsigned int hctx_idx)
+{
+	struct nullb *nullb = hctx->queue->queuedata;
+	struct nullb_queue *nq;
+
+	nq = &nullb->queues[hctx_idx];
+	hctx->driver_data = nq;
+	null_init_queue(nullb, nq);
+	nullb->nr_queues++;
+
+	return 0;
+}
+
+static const struct blk_mq_ops null_mq_ops = {
+	.queue_rq       = null_queue_rq,
+	.complete	= null_complete_rq,
+	.timeout	= null_timeout_rq,
+	.init_hctx	= null_init_hctx,
+	.exit_hctx	= null_exit_hctx,
+};
+
 static void null_del_dev(struct nullb *nullb)
 {
 	struct nullb_device *dev = nullb->dev;
@@ -1473,33 +1510,6 @@ static const struct block_device_operations null_ops = {
 	.report_zones	= null_report_zones,
 };
 
-static void null_init_queue(struct nullb *nullb, struct nullb_queue *nq)
-{
-	BUG_ON(!nullb);
-	BUG_ON(!nq);
-
-	init_waitqueue_head(&nq->wait);
-	nq->queue_depth = nullb->queue_depth;
-	nq->dev = nullb->dev;
-}
-
-static void null_init_queues(struct nullb *nullb)
-{
-	struct request_queue *q = nullb->q;
-	struct blk_mq_hw_ctx *hctx;
-	struct nullb_queue *nq;
-	int i;
-
-	queue_for_each_hw_ctx(q, hctx, i) {
-		if (!hctx->nr_ctx || !hctx->tags)
-			continue;
-		nq = &nullb->queues[i];
-		hctx->driver_data = nq;
-		null_init_queue(nullb, nq);
-		nullb->nr_queues++;
-	}
-}
-
 static int setup_commands(struct nullb_queue *nq)
 {
 	struct nullb_cmd *cmd;
@@ -1526,8 +1536,7 @@ static int setup_commands(struct nullb_queue *nq)
 
 static int setup_queues(struct nullb *nullb)
 {
-	nullb->queues = kcalloc(nullb->dev->submit_queues,
-				sizeof(struct nullb_queue),
+	nullb->queues = kcalloc(nr_cpu_ids, sizeof(struct nullb_queue),
 				GFP_KERNEL);
 	if (!nullb->queues)
 		return -ENOMEM;
@@ -1673,6 +1682,27 @@ static bool null_setup_fault(void)
 	return true;
 }
 
+/*
+ * This function is identical to blk_mq_init_queue() except that it sets
+ * queuedata before .init_hctx is called.
+ */
+static struct request_queue *nullb_alloc_queue(struct nullb *nullb)
+{
+	struct request_queue *uninit_q, *q;
+	struct blk_mq_tag_set *set = nullb->tag_set;
+
+	uninit_q = blk_alloc_queue_node(GFP_KERNEL, set->numa_node);
+	if (!uninit_q)
+		return ERR_PTR(-ENOMEM);
+
+	uninit_q->queuedata = nullb;
+	q = blk_mq_init_allocated_queue(set, uninit_q, false);
+	if (IS_ERR(q))
+		blk_cleanup_queue(uninit_q);
+
+	return q;
+}
+
 static int null_add_dev(struct nullb_device *dev)
 {
 	struct nullb *nullb;
@@ -1712,12 +1742,11 @@ static int null_add_dev(struct nullb_device *dev)
 			goto out_cleanup_queues;
 
 		nullb->tag_set->timeout = 5 * HZ;
-		nullb->q = blk_mq_init_queue(nullb->tag_set);
+		nullb->q = nullb_alloc_queue(nullb);
 		if (IS_ERR(nullb->q)) {
 			rv = -ENOMEM;
 			goto out_cleanup_tags;
 		}
-		null_init_queues(nullb);
 	} else if (dev->queue_mode == NULL_Q_BIO) {
 		nullb->q = blk_alloc_queue_node(GFP_KERNEL, dev->home_node);
 		if (!nullb->q) {
#6 lore raw

PATCH v4 6/8 null_blk: Fix the null_add_dev() error path

In reply to Bart Van Assche

If null_add_dev() fails, clear dev->nullb.

This patch fixes the following KASAN complaint:

BUG: KASAN: use-after-free in nullb_device_submit_queues_store+0xcf/0x160 [null_blk]
Read of size 8 at addr ffff88803280fc30 by task check/8409

Call Trace:
 dump_stack+0xa5/0xe6
 print_address_description.constprop.0+0x26/0x260
 __kasan_report.cold+0x7b/0x99
 kasan_report+0x16/0x20
 __asan_load8+0x58/0x90
 nullb_device_submit_queues_store+0xcf/0x160 [null_blk]
 configfs_write_file+0x1c4/0x250 [configfs]
 __vfs_write+0x4c/0x90
 vfs_write+0x145/0x2c0
 ksys_write+0xd7/0x180
 __x64_sys_write+0x47/0x50
 do_syscall_64+0x6f/0x2f0
 entry_SYSCALL_64_after_hwframe+0x49/0xbe
RIP: 0033:0x7ff370926317
Code: 64 89 02 48 c7 c0 ff ff ff ff eb bb 0f 1f 80 00 00 00 00 f3 0f 1e fa 64 8b 04 25 18 00 00 00 85 c0 75 10 b8 01 00 00 00 0f 05 <48> 3d 00 f0 ff ff 77 51 c3 48 83 ec 28 48 89 54 24 18 48 89 74 24
RSP: 002b:00007fff2dd2da48 EFLAGS: 00000246 ORIG_RAX: 0000000000000001
RAX: ffffffffffffffda RBX: 0000000000000002 RCX: 00007ff370926317
RDX: 0000000000000002 RSI: 0000559437ef23f0 RDI: 0000000000000001
RBP: 0000559437ef23f0 R08: 000000000000000a R09: 0000000000000001
R10: 0000559436703471 R11: 0000000000000246 R12: 0000000000000002
R13: 00007ff370a006a0 R14: 00007ff370a014a0 R15: 00007ff370a008a0

Allocated by task 8409:
 save_stack+0x23/0x90
 __kasan_kmalloc.constprop.0+0xcf/0xe0
 kasan_kmalloc+0xd/0x10
 kmem_cache_alloc_node_trace+0x129/0x4c0
 null_add_dev+0x24a/0xe90 [null_blk]
 nullb_device_power_store+0x1b6/0x270 [null_blk]
 configfs_write_file+0x1c4/0x250 [configfs]
 __vfs_write+0x4c/0x90
 vfs_write+0x145/0x2c0
 ksys_write+0xd7/0x180
 __x64_sys_write+0x47/0x50
 do_syscall_64+0x6f/0x2f0
 entry_SYSCALL_64_after_hwframe+0x49/0xbe

Freed by task 8409:
 save_stack+0x23/0x90
 __kasan_slab_free+0x112/0x160
 kasan_slab_free+0x12/0x20
 kfree+0xdf/0x250
 null_add_dev+0xaf3/0xe90 [null_blk]
 nullb_device_power_store+0x1b6/0x270 [null_blk]
 configfs_write_file+0x1c4/0x250 [configfs]
 __vfs_write+0x4c/0x90
 vfs_write+0x145/0x2c0
 ksys_write+0xd7/0x180
 __x64_sys_write+0x47/0x50
 do_syscall_64+0x6f/0x2f0
 entry_SYSCALL_64_after_hwframe+0x49/0xbe
Cc Christoph Hellwig <[email protected]>
Cc Ming Lei <[email protected]>
Cc Hannes Reinecke <[email protected]>
Cc Johannes Thumshirn <[email protected]>
Fixes 2984c8684f96 ("nullb: factor disk parameters")
---
 drivers/block/null_blk_main.c | 1 +
 1 file changed, 1 insertion(+)
drivers/block/null_blk_main.c +1 -0
index 83e31b81db3c..80d3dcb8722c 100644
@@ -1817,6 +1817,7 @@ static int null_add_dev(struct nullb_device *dev)
 	cleanup_queues(nullb);
 out_free_nullb:
 	kfree(nullb);
+	dev->nullb = NULL;
 out:
 	return rv;
 }
#7 lore raw

PATCH v4 7/8 null_blk: Handle null_add_dev() failures properly

In reply to Bart Van Assche

If null_add_dev() fails then null_del_dev() is called with a NULL argument. Make null_del_dev() handle this scenario correctly. This patch fixes the following KASAN complaint:

null-ptr-deref in null_del_dev+0x28/0x280 [null_blk]

Read of size 8 at addr 0000000000000000 by task find/1062

Call Trace:
 dump_stack+0xa5/0xe6
 __kasan_report.cold+0x65/0x99
 kasan_report+0x16/0x20
 __asan_load8+0x58/0x90
 null_del_dev+0x28/0x280 [null_blk]
 nullb_group_drop_item+0x7e/0xa0 [null_blk]
 client_drop_item+0x53/0x80 [configfs]
 configfs_rmdir+0x395/0x4e0 [configfs]
 vfs_rmdir+0xb6/0x220
 do_rmdir+0x238/0x2c0
 __x64_sys_unlinkat+0x75/0x90
 do_syscall_64+0x6f/0x2f0
 entry_SYSCALL_64_after_hwframe+0x49/0xbe
Cc Christoph Hellwig <[email protected]>
Cc Ming Lei <[email protected]>
Cc Hannes Reinecke <[email protected]>
Cc Johannes Thumshirn <[email protected]>
---
 drivers/block/null_blk_main.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)
drivers/block/null_blk_main.c +6 -1
index 80d3dcb8722c..d21c82c8863f 100644
@@ -1469,7 +1469,12 @@ static const struct blk_mq_ops null_mq_ops = {
 
 static void null_del_dev(struct nullb *nullb)
 {
-	struct nullb_device *dev = nullb->dev;
+	struct nullb_device *dev;
+
+	if (!nullb)
+		return;
+
+	dev = nullb->dev;
 
 	ida_simple_remove(&nullb_indexes, nullb->index);
 
#8 lore raw

PATCH v4 8/8 null_blk: Add support for init_hctx() fault injection

In reply to Bart Van Assche

This makes it possible to test the error path in blk_mq_realloc_hw_ctxs() and also several error paths in null_blk.

Cc Christoph Hellwig <[email protected]>
Cc Ming Lei <[email protected]>
Cc Hannes Reinecke <[email protected]>
Cc Johannes Thumshirn <[email protected]>
---
 drivers/block/null_blk_main.c | 11 +++++++++++
 1 file changed, 11 insertions(+)
drivers/block/null_blk_main.c +11 -0
index d21c82c8863f..89bb16a99007 100644
@@ -23,6 +23,7 @@
 #ifdef CONFIG_BLK_DEV_NULL_BLK_FAULT_INJECTION
 static DECLARE_FAULT_ATTR(null_timeout_attr);
 static DECLARE_FAULT_ATTR(null_requeue_attr);
+static DECLARE_FAULT_ATTR(null_init_hctx_attr);
 #endif
 
 static inline u64 mb_per_tick(int mbps)
@@ -101,6 +102,9 @@ module_param_string(timeout, g_timeout_str, sizeof(g_timeout_str), 0444);
 
 static char g_requeue_str[80];
 module_param_string(requeue, g_requeue_str, sizeof(g_requeue_str), 0444);
+
+static char g_init_hctx_str[80];
+module_param_string(init_hctx, g_init_hctx_str, sizeof(g_init_hctx_str), 0444);
 #endif
 
 static int g_queue_mode = NULL_Q_MQ;
@@ -1451,6 +1455,11 @@ static int null_init_hctx(struct blk_mq_hw_ctx *hctx, void *driver_data,
 	struct nullb *nullb = hctx->queue->queuedata;
 	struct nullb_queue *nq;
 
+#ifdef CONFIG_BLK_DEV_NULL_BLK_FAULT_INJECTION
+	if (g_init_hctx_str[0] && should_fail(&null_init_hctx_attr, 1))
+		return -EFAULT;
+#endif
+
 	nq = &nullb->queues[hctx_idx];
 	hctx->driver_data = nq;
 	null_init_queue(nullb, nq);
@@ -1683,6 +1692,8 @@ static bool null_setup_fault(void)
 		return false;
 	if (!__null_setup_fault(&null_requeue_attr, g_requeue_str))
 		return false;
+	if (!__null_setup_fault(&null_init_hctx_attr, g_init_hctx_str))
+		return false;
 #endif
 	return true;
 }
maintainer
#9 lore raw
In reply to Bart Van Assche
Quoted 5 lines · Bart Van Assche

Hi Jens,

These patches are what I came up with while analyzing syzbot and blktests complaints related to dynamically changing the number of hardware queues. Please consider these patches for the upstream kernel.

Applied for 5.7, thanks Bart.

Signature · 1 line
Jens Axboe