Skip to contentSkip to conversation

nvme: core: freeze multipath queue early in nvme_update_ns_info()

PATCH In review Martin Wilck 5 participants 7 messages Aug 22, 2024 – Aug 25, 2024 NVMe

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 nvme: core: freeze multipath queue early in nvme_update_ns_info()

For multipath devices, nvme_update_ns_info() needs to freeze both the queue of the path and the queue of the multipath device. For both operations, it waits for one RCU grace period to pass, ~25ms on my test system. By calling blk_freeze_queue_start() for the multipath queue early, we avoid waiting twice; tests using ftrace have shown that the second blk_mq_freeze_queue_wait() call finishes in just a few microseconds. The path queue is unfrozen before calling blk_mq_freeze_queue_wait() on the multipath queue, so that possibly outstanding IO in the multipath queue can be flushed.

I tested this using the "controller rescan under I/O load" test I submitted recently [1].

[1] https://lore.kernel.org/linux-nvme/[email protected]/T/#u

---
 drivers/nvme/host/core.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)
drivers/nvme/host/core.c +6 -2
index 33fa01c599ad..e2454398c660 100644
@@ -2217,6 +2217,9 @@ static int nvme_update_ns_info(struct nvme_ns *ns, struct nvme_ns_info *info)
 	bool unsupported = false;
 	int ret;
 
+	if (nvme_ns_head_multipath(ns->head))
+		blk_freeze_queue_start(ns->head->disk->queue);
+
 	switch (info->ids.csi) {
 	case NVME_CSI_ZNS:
 		if (!IS_ENABLED(CONFIG_BLK_DEV_ZONED)) {
@@ -2254,7 +2257,7 @@ static int nvme_update_ns_info(struct nvme_ns *ns, struct nvme_ns_info *info)
 		struct queue_limits *ns_lim = &ns->disk->queue->limits;
 		struct queue_limits lim;
 
-		blk_mq_freeze_queue(ns->head->disk->queue);
+		blk_mq_freeze_queue_wait(ns->head->disk->queue);
 		/*
 		 * queue_limits mixes values that are the hardware limitations
 		 * for bio splitting with what is the device configuration.
@@ -2288,7 +2291,8 @@ static int nvme_update_ns_info(struct nvme_ns *ns, struct nvme_ns_info *info)
 		nvme_mpath_revalidate_paths(ns);
 
 		blk_mq_unfreeze_queue(ns->head->disk->queue);
-	}
+	} else if (nvme_ns_head_multipath(ns->head))
+		blk_mq_unfreeze_queue(ns->head->disk->queue);
 
 	return ret;
 }
Signature · 1 line
2.35.3
Reviewed-by on v1 1/1
#1 lore raw
Quoted 13 lines · Martin Wilck

For multipath devices, nvme_update_ns_info() needs to freeze both the queue of the path and the queue of the multipath device. For both operations, it waits for one RCU grace period to pass, ~25ms on my test system. By calling blk_freeze_queue_start() for the multipath queue early, we avoid waiting twice; tests using ftrace have shown that the second blk_mq_freeze_queue_wait() call finishes in just a few microseconds. The path queue is unfrozen before calling blk_mq_freeze_queue_wait() on the multipath queue, so that possibly outstanding IO in the multipath queue can be flushed.

I tested this using the "controller rescan under I/O load" test I submitted recently [1].

Quoted 7 lines

[1] https://lore.kernel.org/linux-nvme/[email protected]/T/#u

---
  drivers/nvme/host/core.c | 8 ++++++--
  1 file changed, 6 insertions(+), 2 deletions(-)

Cheers,

Hannes

Signature · 4 lines
Dr. Hannes Reinecke                  Kernel Storage Architect
[email protected]                                +49 911 74053 688
SUSE Software Solutions GmbH, Frankenstr. 146, 90461 Nürnberg
HRB 36809 (AG Nürnberg), GF: I. Totev, A. McDonald, W. Knoblich
Reviewed-by on v1 1/1
#2 lore raw
In reply to Martin Wilck
Quoted 10 lines · Martin Wilck

For multipath devices, nvme_update_ns_info() needs to freeze both the queue of the path and the queue of the multipath device. For both operations, it waits for one RCU grace period to pass, ~25ms on my test system. By calling blk_freeze_queue_start() for the multipath queue early, we avoid waiting twice; tests using ftrace have shown that the second blk_mq_freeze_queue_wait() call finishes in just a few microseconds. The path queue is unfrozen before calling blk_mq_freeze_queue_wait() on the multipath queue, so that possibly outstanding IO in the multipath queue can be flushed.

Quoted 6 lines

I tested this using the "controller rescan under I/O load" test I submitted recently [1].

[1] https://lore.kernel.org/linux-nvme/[email protected]/T/#u

#3 lore raw
In reply to Martin Wilck
Quoted 25 lines · Martin Wilck

For multipath devices, nvme_update_ns_info() needs to freeze both the queue of the path and the queue of the multipath device. For both operations, it waits for one RCU grace period to pass, ~25ms on my test system. By calling blk_freeze_queue_start() for the multipath queue early, we avoid waiting twice; tests using ftrace have shown that the second blk_mq_freeze_queue_wait() call finishes in just a few microseconds. The path queue is unfrozen before calling blk_mq_freeze_queue_wait() on the multipath queue, so that possibly outstanding IO in the multipath queue can be flushed.

I tested this using the "controller rescan under I/O load" test I submitted recently [1].

[1] https://lore.kernel.org/linux-nvme/[email protected]/T/#u

---
 drivers/nvme/host/core.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)
drivers/nvme/host/core.c +0 -0
index 33fa01c599ad..e2454398c660 100644

drivers/nvme/host/core.c static int nvme_update_ns_info(struct nvme_ns *ns, struct nvme_ns_info *info) · line 2222

Quoted 6 lines
drivers/nvme/host/core.c +3 -0
@@ -2217,6 +2217,9 @@ static int nvme_update_ns_info(struct nvme_ns *ns, struct nvme_ns_info *info)
 	bool unsupported = false;
 	int ret;
 
+	if (nvme_ns_head_multipath(ns->head))
+		blk_freeze_queue_start(ns->head->disk->queue);
+

From someone reading this code, it looks quite similar to nvme_mpath_start_freeze().

Perhaps create a new helper, with proper kdoc, and possibly also add kdoc to nvme_mpath_start_freeze(), so that a user can easily tell (from the kdoc) when to use which function.

Kind regards,
Niklas

#4 lore raw
Quoted 22 lines · Niklas Cassel
Quoted 21 lines · Martin Wilck

For multipath devices, nvme_update_ns_info() needs to freeze both the queue of the path and the queue of the multipath device. For both operations, it waits for one RCU grace period to pass, ~25ms on my test system. By calling blk_freeze_queue_start() for the multipath queue early, we avoid waiting twice; tests using ftrace have shown that the second blk_mq_freeze_queue_wait() call finishes in just a few microseconds. The path queue is unfrozen before calling blk_mq_freeze_queue_wait() on the multipath queue, so that possibly outstanding IO in the multipath queue can be flushed.

I tested this using the "controller rescan under I/O load" test I submitted recently [1].

[1]
https://lore.kernel.org/linux-nvme/[email protected]/T/#u

---
 drivers/nvme/host/core.c | 8 ++++++--

 1 file changed, 6 insertions(+), 2 deletions(-)
Quoted 15 lines

drivers/nvme/host/core.c static int nvme_update_ns_info(struct nvme_ns *ns, struct nvme_ns_info *info) · line 2222

Quoted 12 lines
drivers/nvme/host/core.c +0 -0
index 33fa01c599ad..e2454398c660 100644
@@ -2217,6 +2217,9 @@ static int nvme_update_ns_info(struct nvme_ns
 *ns, struct nvme_ns_info *info)
  	bool unsupported = false;
  	int ret;

drivers/nvme/host/core.c +3 -0
+	if (nvme_ns_head_multipath(ns->head))
+		blk_freeze_queue_start(ns->head->disk->queue);
+

From someone reading this code, it looks quite similar to nvme_mpath_start_freeze().

That function takes a struct nvme_subsystem as argument, and walks over all namespaces in that subsystem, whereas here we're just acting on a single namespace.

Quoted 5 lines

Perhaps create a new helper, with proper kdoc, and possibly also add kdoc to
nvme_mpath_start_freeze(), so that a user can easily tell (from the kdoc)
when to use which function.

What's the benefit of introducing such a trivial helper, used only in a single place of the code?
Thanks for the suggestion, but I'd like to see other maintainers' opinions about this.

Thanks,
Martin

#5 lore raw
Quoted 43 lines · Martin Wilck
Quoted 22 lines · Niklas Cassel
Quoted 21 lines · Martin Wilck

For multipath devices, nvme_update_ns_info() needs to freeze both the queue of the path and the queue of the multipath device. For both operations, it waits for one RCU grace period to pass, ~25ms on my test system. By calling blk_freeze_queue_start() for the multipath queue early, we avoid waiting twice; tests using ftrace have shown that the second blk_mq_freeze_queue_wait() call finishes in just a few microseconds. The path queue is unfrozen before calling blk_mq_freeze_queue_wait() on the multipath queue, so that possibly outstanding IO in the multipath queue can be flushed.

I tested this using the "controller rescan under I/O load" test I submitted recently [1].

[1]
https://lore.kernel.org/linux-nvme/[email protected]/T/#u

---
 drivers/nvme/host/core.c | 8 ++++++--

 1 file changed, 6 insertions(+), 2 deletions(-)
Quoted 15 lines

drivers/nvme/host/core.c static int nvme_update_ns_info(struct nvme_ns *ns, struct nvme_ns_info *info) · line 2222

Quoted 12 lines
drivers/nvme/host/core.c +0 -0
index 33fa01c599ad..e2454398c660 100644
@@ -2217,6 +2217,9 @@ static int nvme_update_ns_info(struct nvme_ns
 *ns, struct nvme_ns_info *info)
  	bool unsupported = false;
  	int ret;

drivers/nvme/host/core.c +3 -0
+	if (nvme_ns_head_multipath(ns->head))
+		blk_freeze_queue_start(ns->head->disk->queue);
+

From someone reading this code, it looks quite similar to nvme_mpath_start_freeze().

That function takes a struct nvme_subsystem as argument, and walks over all namespaces in that subsystem, whereas here we're just acting on a single namespace.

Quoted 8 lines
Quoted 5 lines

Perhaps create a new helper, with proper kdoc, and possibly also add kdoc to
nvme_mpath_start_freeze(), so that a user can easily tell (from the kdoc)
when to use which function.

What's the benefit of introducing such a trivial helper, used only in a single place of the code?

To make nvme_update_ns_info() easier to read, and to keep the existing code style of calling nvme_mpath_*() functions unconditionally.

Sure, with my suggestion you would need a mpath_freeze() and mpath_unfreeze() helper.

But if you add a helper that simply does:
if (nvme_ns_head_multipath(ns->head))

	blk_freeze_queue_start(ns->head->disk->queue);

nvme_update_ns_info() could call the helper unconditionally, just like how e.g. nvme_passthru_start() calls nvme_mpath_start_freeze() unconditionally for both multipath and non-multipath NS heads:
https://github.com/torvalds/linux/blob/v6.11-rc4/drivers/nvme/host/core.c#L1205

(And how core.c calls many other nvme_mpath_*() functions unconditionally.)

Quoted 2 lines

Thanks for the suggestion, but I'd like to see other maintainers' opinions about this.

Of course, a suggestion is just a suggestion :)

Kind regards,
Niklas

maintainer
#6 lore raw
In reply to Martin Wilck

The patch title should explain what it is fixing, i.e.  the reason of its existence. Perhaps something like:
nvme: shorten multipath ns update duration

Quoted 25 lines · Martin Wilck

For multipath devices, nvme_update_ns_info() needs to freeze both the queue of the path and the queue of the multipath device. For both operations, it waits for one RCU grace period to pass, ~25ms on my test system. By calling blk_freeze_queue_start() for the multipath queue early, we avoid waiting twice; tests using ftrace have shown that the second blk_mq_freeze_queue_wait() call finishes in just a few microseconds. The path queue is unfrozen before calling blk_mq_freeze_queue_wait() on the multipath queue, so that possibly outstanding IO in the multipath queue can be flushed.

I tested this using the "controller rescan under I/O load" test I submitted recently [1].

[1] https://lore.kernel.org/linux-nvme/[email protected]/T/#u

---
  drivers/nvme/host/core.c | 8 ++++++--
  1 file changed, 6 insertions(+), 2 deletions(-)
drivers/nvme/host/core.c +0 -0
index 33fa01c599ad..e2454398c660 100644

drivers/nvme/host/core.c static int nvme_update_ns_info(struct nvme_ns *ns, struct nvme_ns_info *info) · line 2221

Quoted 5 lines
drivers/nvme/host/core.c +2 -0
@@ -2217,6 +2217,9 @@ static int nvme_update_ns_info(struct nvme_ns *ns, struct nvme_ns_info *info)
  	bool unsupported = false;
  	int ret;
 
+	if (nvme_ns_head_multipath(ns->head))
+		blk_freeze_queue_start(ns->head->disk->queue);

Lets add a small comment here.

Quoted 14 lines
drivers/nvme/host/core.c +2 -1
+
  	switch (info->ids.csi) {
  	case NVME_CSI_ZNS:
  		if (!IS_ENABLED(CONFIG_BLK_DEV_ZONED)) {
@@ -2254,7 +2257,7 @@ static int nvme_update_ns_info(struct nvme_ns *ns, struct nvme_ns_info *info)
  		struct queue_limits *ns_lim = &ns->disk->queue->limits;
  		struct queue_limits lim;
 
-		blk_mq_freeze_queue(ns->head->disk->queue);
+		blk_mq_freeze_queue_wait(ns->head->disk->queue);
  		/*
  		 * queue_limits mixes values that are the hardware limitations
  		 * for bio splitting with what is the device configuration.

drivers/nvme/host/core.c static int nvme_update_ns_info(struct nvme_ns *ns, struct nvme_ns_info *info) · line 2295

Quoted 6 lines
drivers/nvme/host/core.c +2 -1
@@ -2288,7 +2291,8 @@ static int nvme_update_ns_info(struct nvme_ns *ns, struct nvme_ns_info *info)
  		nvme_mpath_revalidate_paths(ns);
 
  		blk_mq_unfreeze_queue(ns->head->disk->queue);
-	}
+	} else if (nvme_ns_head_multipath(ns->head))
+		blk_mq_unfreeze_queue(ns->head->disk->queue);

Right now if ret!=0 you are unfreezing without waiting, you need to wait for the freeze to complete before unfreezing.

You should restructure the code to make it so that the freeze_start, freeze_wait, unfreeze are paired regardless of the code flows.