On 2/21/24 00:46, Nilay Shroff wrote: > > On 2/21/24 02:11, Keith Busch wrote: >> From: Keith Busch <kbusch@xxxxxxxxxx> >> >> If a user runs something like `blkdiscard -z /dev/sda`, and the device >> does not have an efficient write zero offload, the kernel will dispatch >> long chains of bio's using the ZERO_PAGE for the entire capacity of the >> device. If the capacity is very large, this process could take a long >> time in an uninterruptable state, which the user may want to abort. >> Check between batches for the user's request to kill the process so they >> don't need to wait potentially many hours. >> >> Signed-off-by: Keith Busch <kbusch@xxxxxxxxxx> >> --- >> block/blk-lib.c | 2 ++ >> 1 file changed, 2 insertions(+) >> >> diff --git a/block/blk-lib.c b/block/blk-lib.c >> index e59c3069e8351..d5c334aa98e0d 100644 >> --- a/block/blk-lib.c >> +++ b/block/blk-lib.c >> @@ -190,6 +190,8 @@ static int __blkdev_issue_zero_pages(struct block_device *bdev, >> break; >> } >> cond_resched(); >> + if (fatal_signal_pending(current)) >> + break; >> } >> >> *biop = bio; > I have an NVMe device which supports write offload. The total size of this disk is ~1.5 TB. > When I tried zero out the whole NVMe drive it took more than 10 minutes. Please see below: > > # cat /sys/block/nvme0n1/queue/write_zeroes_max_bytes > 8388608 > > # nvme id-ns /dev/nvme0n1 -H > NVME Identify Namespace 1: > nsze : 0x1749a956 > ncap : 0x1749a956 > nuse : 0x1749a956 > <snip> > nvmcap : 1600321314816 > <snip> > LBA Format 0 : Metadata Size: 0 bytes - Data Size: 4096 bytes - Relative Performance: 0 Best (in use) > <snip> > > # time blkdiscard -z /dev/nvme0n1 > > real 10m27.514s > user 0m0.000s > sys 0m0.369s > > So shouldn't we need to add the same code (allowing user to kill the process) under > __blkdev_issue_write_zeroes()? I think even though a drive supports "write zero offload", if > drive has a very large capacity then it would take up a lot of time to zero out the complete drive. > Yes the time required may not be in hours in this case but it could be in tens of minutes depending > on the drive capacity. > > Thanks, > --Nilay > > this is long standing problem with discard and write-zeroes, if we are going this route then we might as well make this change for rest of the callers in blk-lib.c .. -ck
