Christoph,
On 2/21/18 11:48, Damien Le Moal wrote:
> Christoph,
>
> On 2/21/18 11:39, Christoph Hellwig wrote:
>> On Tue, Feb 20, 2018 at 03:06:10PM +0900, Damien Le Moal wrote:
>>> For a zoned block device mount, an array of zone types for the device is
>>> allocated and initialized in order to determine if a section is stored
>>> on a sequential zone (zone reset needed) or a conventional zone (no zone
>>> reset needed and regular discard applies). Considering this usage, the
>>> zone types stored in memory can be replaced with a bitmap to indicate
>>> equivalent information, that is, if a zone is sequential or not. This
>>> reduces the memory usage for the device mount by roughly 8 (on a 14TB
>>> disk with zones of 256 MB, the zone type array consumes 13x4KB pages
>>> while the bitmap uses only 2x4KB pages.
>>>
>>> This patch changes the f2fs_dev_info structure blkz_type field to the
>>> bitmap blkz_seq. Access to this bitmap is done using the function
>>> f2fs_blkz_is_seq(), which is a rewrite of the function get_blkz_type().
>>
>> Is there any way we could just provide a block layer helper to
>> figure this out so that the file system code could be simplified even
>> more?
>
> I thought about that before sending the patch...
>
> For a physical drive, the block device queue already has the bitmap
> indicating sequential zones, and a helper could be used in that case.
> But that is not true if the FS block device is a logical device provided
> by something like dm. E.g. if the FS mounts a zoned block device that is
> a part of a real disk split by dm-linear, then there is no sequential
> zone bitmap available, and the FS has to discover that by itself.
>
> Currently, the sequential zone bitmap is initialized in the scsi driver
> only. We could move that initialization to the block device layer at
> some point when the bdev is created and requests can be sent to the
> underlying dev, but before the bdev is exposed. Any suggestion of an
> appropriate place to do that ? I do not see any obvious path that works
> in all cases (real disks and dm).
Answering to myself :)
I was thinking of optimizing by reusing the request queue sequentila
zone bitmap. But you may have been thinking of something more simple like:
struct blk_zone_info {
unsigned int nr_zones;
unsigned long *seq_zones:
};
struct blk_zone_info *blk_get_zone_info(struct block_device *bdev);
void blk_put_zone_info(struct block_device *bdev);
bool blk_zone_is_seq(struct block_device *bdev, sector_t sect);
Which is in essence what f2fs currently code.
That is easy to write. To avoid a lot of report zones, the blk_zone_info
structure can be added to the block device struct for caching (the zone
types and number of zones never change, even with hybrid SMR drives).
Would this be better ?
Best regards.
--
Damien Le Moal,
Western Digital
