On Tue, Jun 08, 2021 at 11:49:04PM +0800, Wang Shanker wrote:
> > 2021年06月07日 21:07，Ming Lei <ming.lei@xxxxxxxxxx> 写道：
> >
> > On Sun, Jun 06, 2021 at 04:54:09AM +0800, Wang Shanker wrote:
> >> Hi, all
> >>
> >> I'm writing to report my recent findings about the handling of discard
> >> operations. As indicated by a few tests, discard operation cannot be
> >> correctly merged, which leads to poor performance of RAID456 on discard
> >> requests. I'm not quite familiar with block subsystem, so please correct
> >> me if there are any mistakes in the following analysis.
> >>
> >> In blk_discard_mergable(), we can see the handling of merging discard
> >> operations goes through different processes, decided by whether we have
> >> more than one queue_max_discard_segments. If the device requires the
> >> sectors should be contiguous in one discard operation, the merging process
> >> will be the same as that for normal read/write operations. Otherwise,
> >> bio_attempt_discard_merge will try to merge as many bios as the device
> >> allows, ignoring the contiguity. Sadly, for both cases, there are problems.
> >>
> >> For devices requiring contiguous sector ranges(such as scsi disks),
> >> bio_attempt_front_merge() or bio_attempt_back_merge() will be handling
> >> the merging process, and both control flows will arrive at
> >> ll_new_hw_segment(), where the following statement:
> >>
> >> req->nr_phys_segments + nr_phys_segs > blk_rq_get_max_segments(req)
> >>
> >> can never be true, since blk_rq_get_max_segments(req) will always be 1.
> >> As a result, no discard operations shall be merged.
> >
> > OK, that looks a bug, and the following change may fix the issue:
> >
> > diff --git a/block/blk-merge.c b/block/blk-merge.c
> > index 4d97fb6dd226..65210e9a8efa 100644
> > --- a/block/blk-merge.c
> > +++ b/block/blk-merge.c
> > @@ -559,10 +559,14 @@ static inline unsigned int blk_rq_get_max_segments(struct request *rq)
> > static inline int ll_new_hw_segment(struct request *req, struct bio *bio,
> > unsigned int nr_phys_segs)
> > {
> > - if (req->nr_phys_segments + nr_phys_segs > blk_rq_get_max_segments(req))
> > + if (blk_integrity_merge_bio(req->q, req, bio) == false)
> > goto no_merge;
> >
> > - if (blk_integrity_merge_bio(req->q, req, bio) == false)
> > + /* discard request merge won't add new segment */
> > + if (req_op(req) == REQ_OP_DISCARD)
> > + return 1;
> > +
> > + if (req->nr_phys_segments + nr_phys_segs > blk_rq_get_max_segments(req))
> > goto no_merge;
> >
> > /*
> >
>
> Many thanks for this patch. I wonder whether it will be applied soon or it will go
> through a formal [PATCH] process.
>
> >>
> >> For devices supporting multiple segments of sector ranges,
> >> bio_attempt_discard_merge() will take over the process. Indeed it will merge
> >> some bios. But how many bios can be merged into one request? In the
> >> implementation, the maximum number of bios is limited mainly by
> >> queue_max_discard_segments (also by blk_rq_get_max_sectors, but it's not where
> >> the problem is). However, it is not the case, since bio_attempt_discard_merge
> >> is not aware of the contiguity of bios. Suppose there are 20 contiguous bios.
> >> They should be considered as only one segment instead 20 of them.
> >
> > Right, so far ELEVATOR_DISCARD_MERGE doesn't merge bios actually, but it
> > can be supported without much difficulty.
> >
> >>
> >> You may wonder the importance of merging discard operations. In the
> >> implementation of RAID456, bios are committed in 4k trunks (they call
> >> them as stripes in the code and the size is determined by DEFAULT_STRIPE_SIZE).
> >> The proper merging of the bios is of vital importance for a reasonable
> >> operating performance of RAID456 devices. In fact, I met this problem
> >> when attempting to create a raid5 volume on a bunch of Nvme SSDs enabling trim
> >> support. Someone also reported similar issues in the linux-raid list
> >> (https://www.spinics.net/lists/raid/msg62108.html). In that post, the author
> >> reported that ``lots of small 4k discard requests that get merged into larger
> >> 512k chunks submitted to devices". This can be explained by my above discovery
> >> because nvme allows 128 segments at the maximum in a dsm instruction.
> >>
> >> The above two scenarios can be reproduced utilizing latest QEMU, with emulated
> >> scsi drives (for the first scenario) or nvme drives (for the second scenario)
> >> and enabling the trace of scsi_disk_emulate_command_UNMAP or pci_nvme_dsm_deallocate.
> >> The detailed process reproducing is as follows:
> >>
> >> 1. create a rootfs (e.g. using debootstrap) under ./rootfs/ ;
> >> 2. obtain a kernel image vmlinuz and generate a initramfs image initrd.img ;
> >> 3. create 3 empty sparse disk images:
> >> # truncate -s 1T disk1 disk2 disk3
> >> 4. using the following qemu command to start the guest vm (here 9p is used
> >> as the rootfs because we don't want the io operations on the rootfs influence
> >> the debugging of the block layer of the guest vm)
> >> # qemu-system-x86_64 \
> >> -cpu kvm64 -machine pc,accel=kvm -smp cpus=2,cores=2,sockets=1 -m 2G \
> >> -chardev stdio,mux=on,id=char0,signal=off \
> >> -fsdev local,path=./rootfs,security_model=passthrough,id=rootfs \
> >> -device virtio-9p,fsdev=rootfs,mount_tag=rootfs \
> >> -monitor chardev:char0 \
> >> -device isa-serial,baudbase=1500000,chardev=char0,index=0,id=ttyS0 \
> >> -nographic \
> >> -kernel vmlinuz -initrd initrd.img \
> >> -append 'root=rootfs rw rootfstype=9p rootflags=trans=virtio,msize=524288 console=ttyS0,1500000 nokaslr' \
> >> -blockdev driver=raw,node-name=nvme1,file.driver=file,file.filename=disk1 \
> >> -blockdev driver=raw,node-name=nvme2,file.driver=file,file.filename=disk2 \
> >> -blockdev driver=raw,node-name=nvme3,file.driver=file,file.filename=disk3 \
> >> -trace pci_nvme_dsm_deallocate,file=nvmetrace.log \
> >> -device nvme,drive=nvme1,logical_block_size=4096,discard_granularity=2097152,physical_block_size=4096,serial=NVME1 \
> >> -device nvme,drive=nvme2,logical_block_size=4096,discard_granularity=2097152,physical_block_size=4096,serial=NVME2 \
> >> -device nvme,drive=nvme3,logical_block_size=4096,discard_granularity=2097152,physical_block_size=4096,serial=NVME3
> >> 5. enable trim support of the raid456 module:
> >> # modprobe raid456
> >> # echo Y > /sys/module/raid456/parameters/devices_handle_discard_safely
> >> 6. using mdaam to create a raid5 device in the guest vm:
> >> # mdadm --create --level=5 --raid-devices=3 /dev/md/test /dev/nvme*n1
> >> 7. and issue a discard request on the dm device: (limit the size of
> >> discard request because discarding all the 2T data is too slow)
> >> # blkdiscard -o 0 -l 1M -p 1M --verbose /dev/md/test
> >> 8. in nvmetrace.log, there are many pci_nvme_dsm_deallocate events of 4k
> >> length (nlb 1).
> >
> > 4kb should be the discard segment length, instead of discard request
> > length, which should be 512k in the above test.
>
> Actually, what are received by the nvme controller are discard requests
> with 128 segments of 4k, instead of one segment of 512k.
Right, I am just wondering if this way makes a difference wrt. single
range/segment discard request from device viewpoint, but anyway it is
better to send less segment.
>
> >
> >>
> >> Similarly, the problem with scsi devices can be emulated using the following
> >> options for qemu:
> >>
> >> -device virtio-scsi,id=scsi \
> >> -device scsi-hd,drive=nvme1,bus=scsi.0,logical_block_size=4096,discard_granularity=2097152,physical_block_size=4096,serial=NVME1 \
> >> -device scsi-hd,drive=nvme2,bus=scsi.0,logical_block_size=4096,discard_granularity=2097152,physical_block_size=4096,serial=NVME2 \
> >> -device scsi-hd,drive=nvme3,bus=scsi.0,logical_block_size=4096,discard_granularity=2097152,physical_block_size=4096,serial=NVME3 \
> >> -trace scsi_disk_emulate_command_UNMAP,file=scsitrace.log
> >>
> >>
> >> Despite the discovery, I cannot come up with a proper fix of this issue due
> >> to my lack of familiarity of the block subsystem. I expect your kind feedback
> >> on this. Thanks in advance.
> >
> > In the above setting and raid456 test, I observe that rq->nr_phys_segments can
> > reach 128, but queue_max_discard_segments() reports 256. So discard
> > request size can be 512KB, which is the max size when you run 1MB discard on
> > raid456. However, if the discard length on raid456 is increased, the
> > current way will become inefficient.
>
> Exactly.
>
> I suggest that bio's can be merged and be calculated as one segment if they are
> contiguous and contain no data.
Fine.
>
> And I also discovered later that, even normal long write requests, e.g.
> a 10m write, will be split into 4k bio's. The maximum number of bio's which can
> be merged into one request is limited by queue_max_segments, regardless
> of whether those bio's are contiguous. In my test environment, for scsi devices,
> queue_max_segments can be 254, which means about 1m size of requests. For nvme
> devices(e.g. Intel DC P4610), queue_max_segments is only 33 since their mdts is 5,
> which results in only 132k of requests.
Here what matters is queue_max_discard_segments().
>
> So, I would also suggest that raid456 should be improved to issue bigger bio's to
> underlying drives.
Right, that should be root solution.
Cc Xiao, I remembered that he worked on this area.
Thanks,
Ming
