Skip to contentSkip to conversation

nvme: only use power of two io boundaries

PATCH Superseded Keith Busch 4 participants 11 messages Aug 27, 2020 – Aug 28, 2020 NVMe

v2 of 3 Superseded Latest is v3 →

lore
Share & tools
lore

Keith Busch

author maintainer
#0 lore raw

PATCH v2 1/1 nvme: only use power of two io boundaries

The kernel requires a power of two for boundaries because that's the only way it can efficiently split commands that cross them. A controller, however, may report a non-power of two boundary.

The driver had been rounding the controller's value to one the kernel can use, but splitting on the wrong boundary provides no benefit on the device side, and actually incurs additional submission overhead from non-optimal splits. Don't provide any boundary hint if the controller's value can't be used and log a warning when initially observing a disk's unreported io boundary.

---
v1->v2:

Warn for the namespace with unusable io boundary.

 drivers/nvme/host/core.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)
drivers/nvme/host/core.c +6 -2
index 154942fc64eb..9bbbd600621f 100644
@@ -2097,8 +2097,12 @@ static int __nvme_revalidate_disk(struct gendisk *disk, struct nvme_id_ns *id)
 		}
 	}
 
-	if (iob && !blk_queue_is_zoned(ns->queue))
-		blk_queue_chunk_sectors(ns->queue, rounddown_pow_of_two(iob));
+	if (is_power_of_2(iob) && !blk_queue_is_zoned(ns->queue))
+		blk_queue_chunk_sectors(ns->queue, iob);
+	else if (iob && !(disk->flags & GENHD_FL_UP))
+		dev_warn(ctrl->device, "namespace:%u has unused io boundary:%u\n",
+			ns->head->ns_id, iob);
+
 	nvme_update_disk_info(disk, ns, id);
 #ifdef CONFIG_NVME_MULTIPATH
 	if (ns->head->disk) {
Signature · 7 lines
2.24.1


_______________________________________________
Linux-nvme mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-nvme
Reviewed-by on v2 1/1
#1 lore raw

Keith,

Quoted 5 lines
drivers/nvme/host/core.c +0 -0
index 154942fc64eb..9bbbd600621f 100644

drivers/nvme/host/core.c static int __nvme_revalidate_disk(struct gendisk *disk, struct nvme_id_ns *id) · line 2103

Quoted 9 lines
drivers/nvme/host/core.c +4 -2
@@ -2097,8 +2097,12 @@ static int __nvme_revalidate_disk(struct gendisk *disk, struct nvme_id_ns *id)
 		}
 	}
 
-	if (iob && !blk_queue_is_zoned(ns->queue))
-		blk_queue_chunk_sectors(ns->queue, rounddown_pow_of_two(iob));
+	if (is_power_of_2(iob) && !blk_queue_is_zoned(ns->queue))
+		blk_queue_chunk_sectors(ns->queue, iob);
+	else if (iob && !(disk->flags & GENHD_FL_UP))
+		dev_warn(ctrl->device, "namespace:%u has unused io boundary:%u\n",

You used "unusable" in the patch description above. Did you mean that here too?

Maybe something like:

"namespace%u: ignoring unsupported I/O boundary:%u\n"

to make it clear that this is due to a kernel limitation and not the drive reporting garbage.

drivers/nvme/host/core.c static int __nvme_revalidate_disk(struct gendisk *disk, struct nvme_id_ns *id) · line 2108

Quoted 5 lines
drivers/nvme/host/core.c +2 -0
+			ns->head->ns_id, iob);
+
 	nvme_update_disk_info(disk, ns, id);
 #ifdef CONFIG_NVME_MULTIPATH
 	if (ns->head->disk) {

Otherwise OK.

Signature · 6 lines
Martin K. Petersen	Oracle Linux Engineering

_______________________________________________
Linux-nvme mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-nvme
author maintainer
#2 lore raw
Quoted 9 lines · Martin K. Petersen

You used "unusable" in the patch description above. Did you mean that here too?

Maybe something like:

"namespace%u: ignoring unsupported I/O boundary:%u\n"

to make it clear that this is due to a kernel limitation and not the drive reporting garbage.

Thanks, your suggestion definitely improves the clarity. I'll use this for a v3.

_______________________________________________
Linux-nvme mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-nvme

#3 lore raw
In reply to Keith Busch

drivers/nvme/host/core.c static int __nvme_revalidate_disk(struct gendisk *disk, struct nvme_id_ns *id) · line 2105

Quoted 6 lines · Keith Busch
drivers/nvme/host/core.c +6 -0
+	if (is_power_of_2(iob) && !blk_queue_is_zoned(ns->queue))
+		blk_queue_chunk_sectors(ns->queue, iob);
+	else if (iob && !(disk->flags & GENHD_FL_UP))
+		dev_warn(ctrl->device, "namespace:%u has unused io boundary:%u\n",
+			ns->head->ns_id, iob);
+

Similar to the zoned stuff I just find this super hard to read.

I wonder if we need to untangle this a little. Something like

static inline bool nvme_first_scan(struct nvme_ns *)

{
	/* add a comment here */
	return !(ns->disk->flags & GENHD_FL_UP)
}

static void nvme_set_chunk_sectors(struct nvme_ns *ns, struct nvme_id_ns *id)

{
	u32 chunk_sectors = nvme_lba_to_sect(ns, le16_to_cpu(id->noiob));

	if ((ctrl->quirks & NVME_QUIRK_STRIPE_SIZE) &&
	    is_power_of_2(ctrl->max_hw_sectors))
	    	chunk_sectors = ctrl->max_hw_sectors;

	if (!chunk_sectors)
		return;

	if (!is_power_of_2(chunk_sectors)) {
		if (nvme_first_scan(ns)) {
			dev_warn(ctrl->device,
				"namespace%u: ignoring unaligned I/O boundary: %u\n",
				ns->head->ns_id, chunk_sectors);
		}
		return;
	}

	if (blk_queue_is_zoned(ns->queue)) {
		if (nvme_first_scan(ns)) {
			dev_info(ctrl->device,
				"namespace%u: ignoring I/O boundary for zoned namespace\n",
				ns->head->ns_id);
		}
		return;
	}
			
	blk_queue_chunk_sectors(ns->queue, chunk_sectors);
}
Quoted 5 lines
drivers/nvme/host/core.c +0 -0
 	nvme_update_disk_info(disk, ns, id);
 #ifdef CONFIG_NVME_MULTIPATH
 	if (ns->head->disk) {
Signature · 1 line
2.24.1

---end quoted text---

_______________________________________________
Linux-nvme mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-nvme

author maintainer
#4 lore raw
Quoted 11 lines · Christoph Hellwig

drivers/nvme/host/core.c static int __nvme_revalidate_disk(struct gendisk *disk, struct nvme_id_ns *id) · line 2105

Quoted 6 lines · Keith Busch
drivers/nvme/host/core.c +6 -0
+	if (is_power_of_2(iob) && !blk_queue_is_zoned(ns->queue))
+		blk_queue_chunk_sectors(ns->queue, iob);
+	else if (iob && !(disk->flags & GENHD_FL_UP))
+		dev_warn(ctrl->device, "namespace:%u has unused io boundary:%u\n",
+			ns->head->ns_id, iob);
+

Similar to the zoned stuff I just find this super hard to read.

I wonder if we need to untangle this a little. Something like

That's fine. Do you want to see a prep patch splitting chunk settings first then power-of-2 check second, or okay as a single patch?

_______________________________________________
Linux-nvme mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-nvme

#5 lore raw
Quoted 11 lines · Keith Busch
Quoted 10 lines · Christoph Hellwig

drivers/nvme/host/core.c static int __nvme_revalidate_disk(struct gendisk *disk, struct nvme_id_ns *id) · line 2105

Quoted 6 lines · Keith Busch
drivers/nvme/host/core.c +6 -0
+	if (is_power_of_2(iob) && !blk_queue_is_zoned(ns->queue))
+		blk_queue_chunk_sectors(ns->queue, iob);
+	else if (iob && !(disk->flags & GENHD_FL_UP))
+		dev_warn(ctrl->device, "namespace:%u has unused io boundary:%u\n",
+			ns->head->ns_id, iob);
+

Similar to the zoned stuff I just find this super hard to read.

Quoted 4 lines
Quoted 1 line

I wonder if we need to untangle this a little. Something like

That's fine. Do you want to see a prep patch splitting chunk settings first then power-of-2 check second, or okay as a single patch?

I don't really care, whatever suits best.

_______________________________________________
Linux-nvme mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-nvme

#6 lore raw
In reply to Keith Busch
Quoted 3 lines · Keith Busch

The kernel requires a power of two for boundaries because that's the only way it can efficiently split commands that cross them. A controller, however, may report a non-power of two boundary.

I can understand that using a power of 2 limit for splitting can be more efficient CPU-wise, but I do not think the kernel "requires" such a limit alignment for splitting BIOs. BLK_SAFE_MAX_SECTORS is 255 after all. What I am missing here ?

Quoted 22 lines

The driver had been rounding the controller's value to one the kernel can use, but splitting on the wrong boundary provides no benefit on the device side, and actually incurs additional submission overhead from non-optimal splits. Don't provide any boundary hint if the controller's value can't be used and log a warning when initially observing a disk's unreported io boundary.

---
v1->v2:

Warn for the namespace with unusable io boundary.

 drivers/nvme/host/core.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)
drivers/nvme/host/core.c +0 -0
index 154942fc64eb..9bbbd600621f 100644

drivers/nvme/host/core.c static int __nvme_revalidate_disk(struct gendisk *disk, struct nvme_id_ns *id) · line 2104

Quoted 10 lines
drivers/nvme/host/core.c +5 -2
@@ -2097,8 +2097,12 @@ static int __nvme_revalidate_disk(struct gendisk *disk, struct nvme_id_ns *id)
 		}
 	}
 
-	if (iob && !blk_queue_is_zoned(ns->queue))
-		blk_queue_chunk_sectors(ns->queue, rounddown_pow_of_two(iob));
+	if (is_power_of_2(iob) && !blk_queue_is_zoned(ns->queue))
+		blk_queue_chunk_sectors(ns->queue, iob);
+	else if (iob && !(disk->flags & GENHD_FL_UP))
+		dev_warn(ctrl->device, "namespace:%u has unused io boundary:%u\n",
+			ns->head->ns_id, iob);

Isn't this going to generate a warning all the time for a ZNS drive ?

Quoted 5 lines
drivers/nvme/host/core.c +1 -0
+
 	nvme_update_disk_info(disk, ns, id);
 #ifdef CONFIG_NVME_MULTIPATH
 	if (ns->head->disk) {
Signature · 7 lines
Damien Le Moal
Western Digital Research

_______________________________________________
Linux-nvme mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-nvme
author maintainer
#7 lore raw
Quoted 9 lines · Damien Le Moal
Quoted 3 lines · Keith Busch

The kernel requires a power of two for boundaries because that's the only way it can efficiently split commands that cross them. A controller, however, may report a non-power of two boundary.

I can understand that using a power of 2 limit for splitting can be more efficient CPU-wise, but I do not think the kernel "requires" such a limit alignment for splitting BIOs. BLK_SAFE_MAX_SECTORS is 255 after all. What I am missing here ?

Ah, those two limits are a bit different. The max sectors just checks length, but chunk_sectors needs to consider length and offset, so checking if an IO crosses the boundary requires division. For CPU efficiency, power of 2 alignment is enforced through a BUG_ON() in blk_queue_chunk_sectors().

Quoted 9 lines

drivers/nvme/host/core.c static int __nvme_revalidate_disk(struct gendisk *disk, struct nvme_id_ns *id) · line 2104

Quoted 7 lines
drivers/nvme/host/core.c +5 -2
-	if (iob && !blk_queue_is_zoned(ns->queue))
-		blk_queue_chunk_sectors(ns->queue, rounddown_pow_of_two(iob));
+	if (is_power_of_2(iob) && !blk_queue_is_zoned(ns->queue))
+		blk_queue_chunk_sectors(ns->queue, iob);
+	else if (iob && !(disk->flags & GENHD_FL_UP))
+		dev_warn(ctrl->device, "namespace:%u has unused io boundary:%u\n",
+			ns->head->ns_id, iob);

Isn't this going to generate a warning all the time for a ZNS drive ?

Yes, if a ZNS drive reports NOIOB, this will warn on it's initial discovery.

As long as zone sizes reuse the queue's chunk_sectors, the zone size needs to take precedence since it's more than just a hint compared to NOIOB. If people complain about the warning, that may indicate we ought to find a way to get around overloading the chunk_sectors queue limit, but I suspect ZNS drives won't report an iob value.

_______________________________________________
Linux-nvme mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-nvme

#8 lore raw
Quoted 11 lines · Keith Busch
Quoted 9 lines · Damien Le Moal
Quoted 3 lines · Keith Busch

The kernel requires a power of two for boundaries because that's the only way it can efficiently split commands that cross them. A controller, however, may report a non-power of two boundary.

I can understand that using a power of 2 limit for splitting can be more efficient CPU-wise, but I do not think the kernel "requires" such a limit alignment for splitting BIOs. BLK_SAFE_MAX_SECTORS is 255 after all. What I am missing here ?

Quoted 5 lines

Ah, those two limits are a bit different. The max sectors just checks length, but chunk_sectors needs to consider length and offset, so checking if an IO crosses the boundary requires division. For CPU efficiency, power of 2 alignment is enforced through a BUG_ON() in blk_queue_chunk_sectors().

OK. Got it.

Quoted 11 lines
Quoted 10 lines

drivers/nvme/host/core.c static int __nvme_revalidate_disk(struct gendisk *disk, struct nvme_id_ns *id) · line 2104

Quoted 8 lines
drivers/nvme/host/core.c +5 -2
-	if (iob && !blk_queue_is_zoned(ns->queue))
-		blk_queue_chunk_sectors(ns->queue, rounddown_pow_of_two(iob));
+	if (is_power_of_2(iob) && !blk_queue_is_zoned(ns->queue))
+		blk_queue_chunk_sectors(ns->queue, iob);
+	else if (iob && !(disk->flags & GENHD_FL_UP))
+		dev_warn(ctrl->device, "namespace:%u has unused io boundary:%u\n",
+			ns->head->ns_id, iob);

Isn't this going to generate a warning all the time for a ZNS drive ?

Quoted 8 lines

Yes, if a ZNS drive reports NOIOB, this will warn on it's initial discovery.

As long as zone sizes reuse the queue's chunk_sectors, the zone size needs to take precedence since it's more than just a hint compared to NOIOB. If people complain about the warning, that may indicate we ought to find a way to get around overloading the chunk_sectors queue limit, but I suspect ZNS drives won't report an iob value.

Understood. But for ZNS, since the zone size would take precedence over iob for chunk sectors, wouldn't it make sense to have the entire code block under a if (!blk_queue_is_zoned(ns->queue)) ? Something like this:

	if (!blk_queue_is_zoned(ns->queue)) {
		if (is_power_of_2(iob))
			blk_queue_chunk_sectors(ns->queue, iob);
		else if (iob && !(disk->flags & GENHD_FL_UP))
			dev_warn(ctrl->device,
				"namespace:%u has unused io boundary:%u\n",
				ns->head->ns_id, iob);
	}

Thanks !

Signature · 7 lines
Damien Le Moal
Western Digital Research

_______________________________________________
Linux-nvme mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-nvme
author maintainer
#9 lore raw
Quoted 11 lines · Damien Le Moal
Quoted 5 lines · Keith Busch

As long as zone sizes reuse the queue's chunk_sectors, the zone size needs to take precedence since it's more than just a hint compared to NOIOB. If people complain about the warning, that may indicate we ought to find a way to get around overloading the chunk_sectors queue limit, but I suspect ZNS drives won't report an iob value.

Understood. But for ZNS, since the zone size would take precedence over iob for chunk sectors, wouldn't it make sense to have the entire code block under a if (!blk_queue_is_zoned(ns->queue)) ? Something like this:

Quoted 8 lines
	if (!blk_queue_is_zoned(ns->queue)) {
		if (is_power_of_2(iob))
			blk_queue_chunk_sectors(ns->queue, iob);
		else if (iob && !(disk->flags & GENHD_FL_UP))
			dev_warn(ctrl->device,
				"namespace:%u has unused io boundary:%u\n",
				ns->head->ns_id, iob);
	}

Oh, I see. I think we should have the warning for ZNS too, though, just to get an indication that devices really want this kind of consideration.

_______________________________________________
Linux-nvme mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-nvme

#10 lore raw
Quoted 21 lines · Keith Busch
Quoted 11 lines · Damien Le Moal
Quoted 5 lines · Keith Busch

As long as zone sizes reuse the queue's chunk_sectors, the zone size needs to take precedence since it's more than just a hint compared to NOIOB. If people complain about the warning, that may indicate we ought to find a way to get around overloading the chunk_sectors queue limit, but I suspect ZNS drives won't report an iob value.

Understood. But for ZNS, since the zone size would take precedence over iob for chunk sectors, wouldn't it make sense to have the entire code block under a if (!blk_queue_is_zoned(ns->queue)) ? Something like this:

Quoted 8 lines
	if (!blk_queue_is_zoned(ns->queue)) {
		if (is_power_of_2(iob))
			blk_queue_chunk_sectors(ns->queue, iob);
		else if (iob && !(disk->flags & GENHD_FL_UP))
			dev_warn(ctrl->device,
				"namespace:%u has unused io boundary:%u\n",
				ns->head->ns_id, iob);
	}
Quoted 3 lines

Oh, I see. I think we should have the warning for ZNS too, though, just to get an indication that devices really want this kind of consideration.

I see. The use of chunk_sectors for both the zone size and iob renders iob impossible to satisfy. So indeed it makes sense to get the warning. Thanks !

Signature · 7 lines
Damien Le Moal
Western Digital Research

_______________________________________________
Linux-nvme mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-nvme