[PATCH] nvme: configure discard at init time

[PATCH] nvme: configure discard at init time

[Jens Axboe]

Currently nvme reconfigures discard for every disk revalidation.  This
is problematic because any O_WRONLY or O_RDWR open will trigger a
partition scan through udev/systemd, and we will reconfigure discard.
This blows away any user settings, like discard_max_bytes.

Configure discard at init time instead.

Signed-off-by: Jens Axboe <axboe at kernel.dk>

---

I'm open to other suggestions as well, currently it sucks that you'd
have to continually re-configure the discard settings when someone opens
the device for writing.

diff --git a/drivers/nvme/host/core.c b/drivers/nvme/host/core.c
index 9df4f71e58ca..a35aa8050749 100644
--- a/drivers/nvme/host/core.c
+++ b/drivers/nvme/host/core.c
@@ -1347,13 +1347,13 @@ static void nvme_set_chunk_size(struct nvme_ns *ns)
 	blk_queue_chunk_sectors(ns->queue, rounddown_pow_of_two(chunk_size));
 }
 
-static void nvme_config_discard(struct nvme_ctrl *ctrl,
-		unsigned stream_alignment, struct request_queue *queue)
+static void nvme_config_discard(struct nvme_ctrl *ctrl, struct nvme_ns *ns)
 {
+	struct request_queue *queue = ns->queue;
 	u32 size = queue_logical_block_size(queue);
 
-	if (stream_alignment)
-		size *= stream_alignment;
+	if (ctrl->nr_streams && ns->sws && ns->sgs)
+		size *= ns->sws * ns->sgs;
 
 	BUILD_BUG_ON(PAGE_SIZE / sizeof(struct nvme_dsm_range) <
 			NVME_DSM_MAX_RANGES);
@@ -1426,8 +1426,6 @@ static void nvme_update_disk_info(struct gendisk *disk,
 		capacity = 0;
 	set_capacity(disk, capacity);
 
-	if (ns->ctrl->oncs & NVME_CTRL_ONCS_DSM)
-		nvme_config_discard(ns->ctrl, stream_alignment, disk->queue);
 	blk_mq_unfreeze_queue(disk->queue);
 }
 
@@ -3043,6 +3041,9 @@ static void nvme_alloc_ns(struct nvme_ctrl *ctrl, unsigned nsid)
 
 	__nvme_revalidate_disk(disk, id);
 
+	if (ns->ctrl->oncs & NVME_CTRL_ONCS_DSM)
+		nvme_config_discard(ctrl, ns);
+
 	down_write(&ctrl->namespaces_rwsem);
 	list_add_tail(&ns->list, &ctrl->namespaces);
 	up_write(&ctrl->namespaces_rwsem);

-- 
Jens Axboe

[PATCH] nvme: configure discard at init time

[Keith Busch]

On Wed, May 02, 2018@09:52:07AM -0600, Jens Axboe wrote:
> Currently nvme reconfigures discard for every disk revalidation.  This
> is problematic because any O_WRONLY or O_RDWR open will trigger a
> partition scan through udev/systemd, and we will reconfigure discard.
> This blows away any user settings, like discard_max_bytes.
> 
> Configure discard at init time instead.
> 
> Signed-off-by: Jens Axboe <axboe at kernel.dk>
> 
> ---
> 
> I'm open to other suggestions as well, currently it sucks that you'd
> have to continually re-configure the discard settings when someone opens
> the device for writing.

Your suggestion is probably fine. The only problem I can think of is a
_very_ unlikely scenario where a firmware update adds discard support,
then a user would have to reload the module in order to expose the
capability.

How about this?

---
diff --git a/drivers/nvme/host/core.c b/drivers/nvme/host/core.c
index a3771c5729f5..18191547e4bd 100644
--- a/drivers/nvme/host/core.c
+++ b/drivers/nvme/host/core.c
@@ -1353,6 +1353,10 @@ static void nvme_config_discard(struct nvme_ctrl *ctrl,
 {
 	u32 size = queue_logical_block_size(queue);
 
+	/* don't reset discard queue limits */
+	if (blk_queue_flag_test_and_set(QUEUE_FLAG_DISCARD, queue))
+		return;
+
 	if (stream_alignment)
 		size *= stream_alignment;
 
@@ -1364,7 +1368,6 @@ static void nvme_config_discard(struct nvme_ctrl *ctrl,
 
 	blk_queue_max_discard_sectors(queue, UINT_MAX);
 	blk_queue_max_discard_segments(queue, NVME_DSM_MAX_RANGES);
-	blk_queue_flag_set(QUEUE_FLAG_DISCARD, queue);
 
 	if (ctrl->quirks & NVME_QUIRK_DEALLOCATE_ZEROES)
 		blk_queue_max_write_zeroes_sectors(queue, UINT_MAX);
--

[PATCH] nvme: configure discard at init time

[Jens Axboe]

On 5/2/18 10:32 AM, Keith Busch wrote:
> On Wed, May 02, 2018@09:52:07AM -0600, Jens Axboe wrote:
>> Currently nvme reconfigures discard for every disk revalidation.  This
>> is problematic because any O_WRONLY or O_RDWR open will trigger a
>> partition scan through udev/systemd, and we will reconfigure discard.
>> This blows away any user settings, like discard_max_bytes.
>>
>> Configure discard at init time instead.
>>
>> Signed-off-by: Jens Axboe <axboe at kernel.dk>
>>
>> ---
>>
>> I'm open to other suggestions as well, currently it sucks that you'd
>> have to continually re-configure the discard settings when someone opens
>> the device for writing.
> 
> Your suggestion is probably fine. The only problem I can think of is a
> _very_ unlikely scenario where a firmware update adds discard support,
> then a user would have to reload the module in order to expose the
> capability.
> 
> How about this?

If we add discard through a firmware upgrade, we should also handle
the case where we lose discard through a firmware upgrade/downgrade.
Not that any of them are likely to ever happen, but...

How about this? Also handles the case where streams values are updated.


diff --git a/drivers/nvme/host/core.c b/drivers/nvme/host/core.c
index 9df4f71e58ca..b0c1f1ce8226 100644
--- a/drivers/nvme/host/core.c
+++ b/drivers/nvme/host/core.c
@@ -1347,13 +1347,19 @@ static void nvme_set_chunk_size(struct nvme_ns *ns)
 	blk_queue_chunk_sectors(ns->queue, rounddown_pow_of_two(chunk_size));
 }
 
-static void nvme_config_discard(struct nvme_ctrl *ctrl,
-		unsigned stream_alignment, struct request_queue *queue)
+static void nvme_config_discard(struct nvme_ns *ns)
 {
+	struct nvme_ctrl *ctrl = ns->ctrl;
+	struct request_queue *queue = ns->queue;
 	u32 size = queue_logical_block_size(queue);
 
-	if (stream_alignment)
-		size *= stream_alignment;
+	if (!(ctrl->oncs & NVME_CTRL_ONCS_DSM)) {
+		blk_queue_flag_clear(QUEUE_FLAG_DISCARD, queue);
+		return;
+	}
+
+	if (ctrl->nr_streams && ns->sws && ns->sgs)
+		size *= ns->sws * ns->sgs;
 
 	BUILD_BUG_ON(PAGE_SIZE / sizeof(struct nvme_dsm_range) <
 			NVME_DSM_MAX_RANGES);
@@ -1361,6 +1367,10 @@ static void nvme_config_discard(struct nvme_ctrl *ctrl,
 	queue->limits.discard_alignment = 0;
 	queue->limits.discard_granularity = size;
 
+	/* If discard is already enabled, don't reset queue limits */
+	if (blk_queue_flag_test_and_set(QUEUE_FLAG_DISCARD, queue))
+		return;
+
 	blk_queue_max_discard_sectors(queue, UINT_MAX);
 	blk_queue_max_discard_segments(queue, NVME_DSM_MAX_RANGES);
 	blk_queue_flag_set(QUEUE_FLAG_DISCARD, queue);
@@ -1407,10 +1417,6 @@ static void nvme_update_disk_info(struct gendisk *disk,
 {
 	sector_t capacity = le64_to_cpup(&id->nsze) << (ns->lba_shift - 9);
 	unsigned short bs = 1 << ns->lba_shift;
-	unsigned stream_alignment = 0;
-
-	if (ns->ctrl->nr_streams && ns->sws && ns->sgs)
-		stream_alignment = ns->sws * ns->sgs;
 
 	blk_mq_freeze_queue(disk->queue);
 	blk_integrity_unregister(disk);
@@ -1427,7 +1433,7 @@ static void nvme_update_disk_info(struct gendisk *disk,
 	set_capacity(disk, capacity);
 
 	if (ns->ctrl->oncs & NVME_CTRL_ONCS_DSM)
-		nvme_config_discard(ns->ctrl, stream_alignment, disk->queue);
+		nvme_config_discard(ns);
 	blk_mq_unfreeze_queue(disk->queue);
 }
 

-- 
Jens Axboe

[PATCH] nvme: configure discard at init time

[Keith Busch]

On Wed, May 02, 2018@10:41:03AM -0600, Jens Axboe wrote:
> If we add discard through a firmware upgrade, we should also handle
> the case where we lose discard through a firmware upgrade/downgrade.
> Not that any of them are likely to ever happen, but...
> 
> How about this? Also handles the case where streams values are updated.

Good call. You also have it updating granularity when a format changes
the logical block size, so that's also a good thing.

One minor issue below:
 
> -static void nvme_config_discard(struct nvme_ctrl *ctrl,
> -		unsigned stream_alignment, struct request_queue *queue)
> +static void nvme_config_discard(struct nvme_ns *ns)
>  {
> +	struct nvme_ctrl *ctrl = ns->ctrl;
> +	struct request_queue *queue = ns->queue;
>  	u32 size = queue_logical_block_size(queue);
>  
> -	if (stream_alignment)
> -		size *= stream_alignment;
> +	if (!(ctrl->oncs & NVME_CTRL_ONCS_DSM)) {
> +		blk_queue_flag_clear(QUEUE_FLAG_DISCARD, queue);
> +		return;
> +	}
> +
> +	if (ctrl->nr_streams && ns->sws && ns->sgs)
> +		size *= ns->sws * ns->sgs;
>  
>  	BUILD_BUG_ON(PAGE_SIZE / sizeof(struct nvme_dsm_range) <
>  			NVME_DSM_MAX_RANGES);
> @@ -1361,6 +1367,10 @@ static void nvme_config_discard(struct nvme_ctrl *ctrl,
>  	queue->limits.discard_alignment = 0;
>  	queue->limits.discard_granularity = size;
>  
> +	/* If discard is already enabled, don't reset queue limits */
> +	if (blk_queue_flag_test_and_set(QUEUE_FLAG_DISCARD, queue))
> +		return;
> +
>  	blk_queue_max_discard_sectors(queue, UINT_MAX);
>  	blk_queue_max_discard_segments(queue, NVME_DSM_MAX_RANGES);
>  	blk_queue_flag_set(QUEUE_FLAG_DISCARD, queue);
> @@ -1407,10 +1417,6 @@ static void nvme_update_disk_info(struct gendisk *disk,
>  {
>  	sector_t capacity = le64_to_cpup(&id->nsze) << (ns->lba_shift - 9);
>  	unsigned short bs = 1 << ns->lba_shift;
> -	unsigned stream_alignment = 0;
> -
> -	if (ns->ctrl->nr_streams && ns->sws && ns->sgs)
> -		stream_alignment = ns->sws * ns->sgs;
>  
>  	blk_mq_freeze_queue(disk->queue);
>  	blk_integrity_unregister(disk);
> @@ -1427,7 +1433,7 @@ static void nvme_update_disk_info(struct gendisk *disk,
>  	set_capacity(disk, capacity);
>  
>  	if (ns->ctrl->oncs & NVME_CTRL_ONCS_DSM)
> -		nvme_config_discard(ns->ctrl, stream_alignment, disk->queue);
> +		nvme_config_discard(ns);

Since nvme_config_discard now handles disabling the queue limit, we need
to call this unconditionally regardless of the ONCS_DSM

[PATCH] nvme: configure discard at init time

[Jens Axboe]

On 5/2/18 10:54 AM, Keith Busch wrote:
> On Wed, May 02, 2018@10:41:03AM -0600, Jens Axboe wrote:
>> If we add discard through a firmware upgrade, we should also handle
>> the case where we lose discard through a firmware upgrade/downgrade.
>> Not that any of them are likely to ever happen, but...
>>
>> How about this? Also handles the case where streams values are updated.
> 
> Good call. You also have it updating granularity when a format changes
> the logical block size, so that's also a good thing.
> 
> One minor issue below:
>  
>> -static void nvme_config_discard(struct nvme_ctrl *ctrl,
>> -		unsigned stream_alignment, struct request_queue *queue)
>> +static void nvme_config_discard(struct nvme_ns *ns)
>>  {
>> +	struct nvme_ctrl *ctrl = ns->ctrl;
>> +	struct request_queue *queue = ns->queue;
>>  	u32 size = queue_logical_block_size(queue);
>>  
>> -	if (stream_alignment)
>> -		size *= stream_alignment;
>> +	if (!(ctrl->oncs & NVME_CTRL_ONCS_DSM)) {
>> +		blk_queue_flag_clear(QUEUE_FLAG_DISCARD, queue);
>> +		return;
>> +	}
>> +
>> +	if (ctrl->nr_streams && ns->sws && ns->sgs)
>> +		size *= ns->sws * ns->sgs;
>>  
>>  	BUILD_BUG_ON(PAGE_SIZE / sizeof(struct nvme_dsm_range) <
>>  			NVME_DSM_MAX_RANGES);
>> @@ -1361,6 +1367,10 @@ static void nvme_config_discard(struct nvme_ctrl *ctrl,
>>  	queue->limits.discard_alignment = 0;
>>  	queue->limits.discard_granularity = size;
>>  
>> +	/* If discard is already enabled, don't reset queue limits */
>> +	if (blk_queue_flag_test_and_set(QUEUE_FLAG_DISCARD, queue))
>> +		return;
>> +
>>  	blk_queue_max_discard_sectors(queue, UINT_MAX);
>>  	blk_queue_max_discard_segments(queue, NVME_DSM_MAX_RANGES);
>>  	blk_queue_flag_set(QUEUE_FLAG_DISCARD, queue);
>> @@ -1407,10 +1417,6 @@ static void nvme_update_disk_info(struct gendisk *disk,
>>  {
>>  	sector_t capacity = le64_to_cpup(&id->nsze) << (ns->lba_shift - 9);
>>  	unsigned short bs = 1 << ns->lba_shift;
>> -	unsigned stream_alignment = 0;
>> -
>> -	if (ns->ctrl->nr_streams && ns->sws && ns->sgs)
>> -		stream_alignment = ns->sws * ns->sgs;
>>  
>>  	blk_mq_freeze_queue(disk->queue);
>>  	blk_integrity_unregister(disk);
>> @@ -1427,7 +1433,7 @@ static void nvme_update_disk_info(struct gendisk *disk,
>>  	set_capacity(disk, capacity);
>>  
>>  	if (ns->ctrl->oncs & NVME_CTRL_ONCS_DSM)
>> -		nvme_config_discard(ns->ctrl, stream_alignment, disk->queue);
>> +		nvme_config_discard(ns);
> 
> Since nvme_config_discard now handles disabling the queue limit, we need
> to call this unconditionally regardless of the ONCS_DSM

Oh yeah, I forgot making that change. I'll send out a proper version.

-- 
Jens Axboe