On 2/20/2023 10:35 PM, Xiao Ni wrote:
> On Tue, Feb 21, 2023 at 11:31 AM Jonathan Derrick
> <jonathan.derrick@xxxxxxxxx> wrote:
>>
>>
>>
>> On 2/20/2023 7:38 PM, Xiao Ni wrote:
>>> On Sat, Feb 18, 2023 at 2:36 AM Jonathan Derrick
>>> <jonathan.derrick@xxxxxxxxx> wrote:
>>>>
>>>> From: Jon Derrick <jonathan.derrick@xxxxxxxxx>
>>>>
>>>> If the bitmap space has enough room, size the I/O for the last bitmap
>>>> page write to the optimal I/O size for the storage device. The expanded
>>>> write is checked that it won't overrun the data or metadata.
>>>>
>>>> This change helps increase performance by preventing unnecessary
>>>> device-side read-mod-writes due to non-atomic write unit sizes.
>>>
>>> Hi Jonathan
>>>
>>> Thanks for your patch.
>>>
>>> Could you explain more about the how the optimal io size can affect
>>> the device-side read-mod-writes?
>>
>> The effects of device-side read-mod-writes are a device-specific implementation detail.
>> This is not something expected to be universal as its an abstracted detail.
>>
>> However the NVMe spec allows the storage namespace to provide implementation
>> hints about its underlying media.
>>
>> Both NPWA and NOWS are used in the NVMe driver, where NOWS provides the
>> optimal_io hint:
>>
>> Per NVMe Command Set 1.0b, Figure 97
>> Namespace Preferred Write Alignment (NPWA): This field indicates the recommended
>> write alignment in logical blocks for this namespace
>>
>> Namespace Optimal Write Size (NOWS): This field indicates the size in logical blocks
>> for optimal write performance for this namespace.
>>
>> Per 5.8.2.1 Improved I/O examples (non-normative)
>> When constructing write operations, the host should minimally construct writes
>> that meet the recommendations of NPWG and NPWA, but may achieve optimal write
>> performance by constructing writes that meet the recommendation of NOWS.
>>
>>
>> It then makes sense that an NVMe drive would provide optimal io size hints
>> that matches its underlying media unit size, such that sub-4k writes might
>> invoke a device-side read-mod-write whereas 4k writes become atomic.
>
> Thanks very much for the detailed explanation.
>
> If I want to try it myself, it must use nvme disks that have 4k
> optimal_io_size, right?
> I tried to check in many environments with this command:
> cat /sys/block/nvme4n1/queue/optimal_io_size (it shows 0)
>
> So I need to find a machine that is optimal_io_size is 4096, right?
>
You can force it with something like this:
diff --git a/drivers/nvme/host/core.c b/drivers/nvme/host/core.c
index d307ae4d8a57..9c349ad54352 100644
--- a/drivers/nvme/host/core.c
+++ b/drivers/nvme/host/core.c
@@ -1901,6 +1901,8 @@ static void nvme_update_disk_info(struct gendisk *disk,
/* NOWS = Namespace Optimal Write Size */
io_opt = bs * (1 + le16_to_cpu(id->nows));
}
+ phys_bs = 4096;
+ io_opt = 4096;
blk_queue_logical_block_size(disk->queue, bs);
/*
> Regards
> Xiao
>>
>>>
>>> Regards
>>> Xiao
>>>>
>>>> Example Intel/Solidigm P5520
>>>> Raid10, Chunk-size 64M, bitmap-size 57228 bits
>>>>
>>>> $ mdadm --create /dev/md0 --level=10 --raid-devices=4 /dev/nvme{0,1,2,3}n1 --assume-clean --bitmap=internal --bitmap-chunk=64M
>>>> $ fio --name=test --direct=1 --filename=/dev/md0 --rw=randwrite --bs=4k --runtime=60
>>>>
>>>> Without patch:
>>>> write: IOPS=1676, BW=6708KiB/s (6869kB/s)(393MiB/60001msec); 0 zone resets
>>>>
>>>> With patch:
>>>> write: IOPS=15.7k, BW=61.4MiB/s (64.4MB/s)(3683MiB/60001msec); 0 zone resets
>>>>
>>>> Biosnoop:
>>>> Without patch:
>>>> Time Process PID Device LBA Size Lat
>>>> 1.410377 md0_raid10 6900 nvme0n1 W 16 4096 0.02
>>>> 1.410387 md0_raid10 6900 nvme2n1 W 16 4096 0.02
>>>> 1.410374 md0_raid10 6900 nvme3n1 W 16 4096 0.01
>>>> 1.410381 md0_raid10 6900 nvme1n1 W 16 4096 0.02
>>>> 1.410411 md0_raid10 6900 nvme1n1 W 115346512 4096 0.01
>>>> 1.410418 md0_raid10 6900 nvme0n1 W 115346512 4096 0.02
>>>> 1.410915 md0_raid10 6900 nvme2n1 W 24 3584 0.43
>>>> 1.410935 md0_raid10 6900 nvme3n1 W 24 3584 0.45
>>>> 1.411124 md0_raid10 6900 nvme1n1 W 24 3584 0.64
>>>> 1.411147 md0_raid10 6900 nvme0n1 W 24 3584 0.66
>>>> 1.411176 md0_raid10 6900 nvme3n1 W 2019022184 4096 0.01
>>>> 1.411189 md0_raid10 6900 nvme2n1 W 2019022184 4096 0.02
>>>>
>>>> With patch:
>>>> Time Process PID Device LBA Size Lat
>>>> 5.747193 md0_raid10 727 nvme0n1 W 16 4096 0.01
>>>> 5.747192 md0_raid10 727 nvme1n1 W 16 4096 0.02
>>>> 5.747195 md0_raid10 727 nvme3n1 W 16 4096 0.01
>>>> 5.747202 md0_raid10 727 nvme2n1 W 16 4096 0.02
>>>> 5.747229 md0_raid10 727 nvme3n1 W 1196223704 4096 0.02
>>>> 5.747224 md0_raid10 727 nvme0n1 W 1196223704 4096 0.01
>>>> 5.747279 md0_raid10 727 nvme0n1 W 24 4096 0.01
>>>> 5.747279 md0_raid10 727 nvme1n1 W 24 4096 0.02
>>>> 5.747284 md0_raid10 727 nvme3n1 W 24 4096 0.02
>>>> 5.747291 md0_raid10 727 nvme2n1 W 24 4096 0.02
>>>> 5.747314 md0_raid10 727 nvme2n1 W 2234636712 4096 0.01
>>>> 5.747317 md0_raid10 727 nvme1n1 W 2234636712 4096 0.02
>>>>
>>>> Signed-off-by: Jon Derrick <jonathan.derrick@xxxxxxxxx>
>>>> ---
>>>> v1->v2: Add more information to commit message
>>>>
>>>> drivers/md/md-bitmap.c | 27 ++++++++++++++++++---------
>>>> 1 file changed, 18 insertions(+), 9 deletions(-)
>>>>
>>>> diff --git a/drivers/md/md-bitmap.c b/drivers/md/md-bitmap.c
>>>> index e7cc6ba1b657..569297ea9b99 100644
>>>> --- a/drivers/md/md-bitmap.c
>>>> +++ b/drivers/md/md-bitmap.c
>>>> @@ -220,6 +220,7 @@ static int write_sb_page(struct bitmap *bitmap, struct page *page, int wait)
>>>> rdev = NULL;
>>>> while ((rdev = next_active_rdev(rdev, mddev)) != NULL) {
>>>> int size = PAGE_SIZE;
>>>> + int optimal_size = PAGE_SIZE;
>>>> loff_t offset = mddev->bitmap_info.offset;
>>>>
>>>> bdev = (rdev->meta_bdev) ? rdev->meta_bdev : rdev->bdev;
>>>> @@ -228,9 +229,14 @@ static int write_sb_page(struct bitmap *bitmap, struct page *page, int wait)
>>>> int last_page_size = store->bytes & (PAGE_SIZE-1);
>>>> if (last_page_size == 0)
>>>> last_page_size = PAGE_SIZE;
>>>> - size = roundup(last_page_size,
>>>> - bdev_logical_block_size(bdev));
>>>> + size = roundup(last_page_size, bdev_logical_block_size(bdev));
>>>> + if (bdev_io_opt(bdev) > bdev_logical_block_size(bdev))
>>>> + optimal_size = roundup(last_page_size, bdev_io_opt(bdev));
>>>> + else
>>>> + optimal_size = size;
>>>> }
>>>> +
>>>> +
>>>> /* Just make sure we aren't corrupting data or
>>>> * metadata
>>>> */
>>>> @@ -246,9 +252,11 @@ static int write_sb_page(struct bitmap *bitmap, struct page *page, int wait)
>>>> goto bad_alignment;
>>>> } else if (offset < 0) {
>>>> /* DATA BITMAP METADATA */
>>>> - if (offset
>>>> - + (long)(page->index * (PAGE_SIZE/512))
>>>> - + size/512 > 0)
>>>> + loff_t off = offset + (long)(page->index * (PAGE_SIZE/512));
>>>> + if (size != optimal_size &&
>>>> + off + optimal_size/512 <= 0)
>>>> + size = optimal_size;
>>>> + else if (off + size/512 > 0)
>>>> /* bitmap runs in to metadata */
>>>> goto bad_alignment;
>>>> if (rdev->data_offset + mddev->dev_sectors
>>>> @@ -257,10 +265,11 @@ static int write_sb_page(struct bitmap *bitmap, struct page *page, int wait)
>>>> goto bad_alignment;
>>>> } else if (rdev->sb_start < rdev->data_offset) {
>>>> /* METADATA BITMAP DATA */
>>>> - if (rdev->sb_start
>>>> - + offset
>>>> - + page->index*(PAGE_SIZE/512) + size/512
>>>> - > rdev->data_offset)
>>>> + loff_t off = rdev->sb_start + offset + page->index*(PAGE_SIZE/512);
>>>> + if (size != optimal_size &&
>>>> + off + optimal_size/512 <= rdev->data_offset)
>>>> + size = optimal_size;
>>>> + else if (off + size/512 > rdev->data_offset)
>>>> /* bitmap runs in to data */
>>>> goto bad_alignment;
>>>> } else {
>>>> --
>>>> 2.27.0
>>>>
>>>
>>
>
