Hi Stefan,
I tested this series with the vDPA block simulator in Linux v6.0.
It worked well, but I had a segfault when blkio_map_mem_region() fails.
In my case, it failed because I forgot to increase the `memlock` limit
with `ulimit -l unlimited` since the simulator still requires locking
all the VM memory.
QEMU failed with this messages:
$ qemu-system-x86_64 ... \
-blockdev
node-name=drive_src1,driver=virtio-blk-vhost-vdpa,path=/dev/vhost-vdpa-0,cache.direct=on
qemu-system-x86_64: -device
virtio-blk-pci,id=src1,bootindex=2,drive=drive_src1,iothread=iothread0: Failed
to add blkio mem region 0x7f60bfe00000 with size 536870912: Bad address (os
error 14)
qemu-system-x86_64: -device
virtio-blk-pci,id=src1,bootindex=2,drive=drive_src1,iothread=iothread0: Failed
to add blkio mem region 0x7f60be400000 with size 16777216: Bad address (os
error 14)
[1] 20803 segmentation fault
IIUC this could be related to a double call to
ram_block_notifier_remove() in the error path of ram_block_added()
(block/block-ram-registrar.c) callback.
Maybe we should call blk_register_buf() only if r->ok is true.
The stack trace is the following:
#0 0x00005641a8b097dd in ram_block_notifier_remove (n=n@entry=0x5641ab8354d8)
at ../hw/core/numa.c:850
#1 0x00005641a89c4701 in ram_block_added
(n=0x5641ab8354d8, host=<optimized out>, size=<optimized out>,
max_size=<optimized out>)
at ../block/block-ram-registrar.c:20
#2 0x00005641a8b083af in ram_block_notify_add_single (rb=0x5641ab3d2810,
opaque=0x5641ab8354d8)
at ../hw/core/numa.c:820
#3 0x00005641a8b92d8d in qemu_ram_foreach_block
(func=func@entry=0x5641a8b08370 <ram_block_notify_add_single>,
opaque=opaque@entry=0x5641ab8354d8)
at ../softmmu/physmem.c:3571
#4 0x00005641a8b097af in ram_block_notifier_add (n=n@entry=0x5641ab8354d8) at
../hw/core/numa.c:844
#5 0x00005641a89c474f in blk_ram_registrar_init (r=r@entry=0x5641ab8354d0,
blk=<optimized out>)
at ../block/block-ram-registrar.c:46
#6 0x00005641a8affe88 in virtio_blk_device_realize (dev=0x5641ab835230,
errp=0x7ffc72657190)
at ../hw/block/virtio-blk.c:1239
#7 0x00005641a8b4b716 in virtio_device_realize (dev=0x5641ab835230,
errp=0x7ffc726571f0)
at ../hw/virtio/virtio.c:3684
#8 0x00005641a8c2049b in device_set_realized (obj=<optimized out>,
value=<optimized out>, errp=0x7ffc72657380)
at ../hw/core/qdev.c:553
#9 0x00005641a8c24738 in property_set_bool
(obj=0x5641ab835230, v=<optimized out>, name=<optimized out>,
opaque=0x5641aa518820, errp=0x7ffc72657380)
at ../qom/object.c:2273
#10 0x00005641a8c27683 in object_property_set
(obj=obj@entry=0x5641ab835230, name=name@entry=0x5641a8e9719a "realized",
v=v@entry=0x5641ab83f5b0, errp=errp@entry=0x7ffc72657380) at
../qom/object.c:1408
#11 0x00005641a8c2a97f in object_property_set_qobject
(obj=obj@entry=0x5641ab835230, name=name@entry=0x5641a8e9719a "realized",
value=value@entry=0x5641ab83f4f0, errp=errp@entry=0x7ffc72657380) at
../qom/qom-qobject.c:28
#12 0x00005641a8c27c84 in object_property_set_bool
(obj=0x5641ab835230, name=0x5641a8e9719a "realized", value=<optimized out>,
errp=0x7ffc72657380)
at ../qom/object.c:1477
#13 0x00005641a8929b74 in pci_qdev_realize (qdev=<optimized out>,
errp=<optimized out>) at ../hw/pci/pci.c:2218
#14 0x00005641a8c2049b in device_set_realized (obj=<optimized out>,
value=<optimized out>, errp=0x7ffc726575b0)
at ../hw/core/qdev.c:553
#15 0x00005641a8c24738 in property_set_bool
(obj=0x5641ab82ce70, v=<optimized out>, name=<optimized out>,
opaque=0x5641aa518820, errp=0x7ffc726575b0)
at ../qom/object.c:2273
#16 0x00005641a8c27683 in object_property_set
Thanks,
Stefano
On Tue, Sep 27, 2022 at 9:34 PM Stefan Hajnoczi <stefa...@redhat.com> wrote:
>
> v5:
> - Drop "RFC" since libblkio 1.0 has been released and the library API is
> stable
> - Disable BDRV_REQ_REGISTERED_BUF if we run out of blkio_mem_regions. The
> bounce buffer slow path is taken when there are not enough blkio_mem_regions
> to cover guest RAM. [Hanna & David Hildenbrand]
> - Call ram_block_discard_disable() when mem-region-pinned property is true or
> absent [David Hildenbrand]
> - Use a bounce buffer pool instead of allocating/freeing a buffer for each
> request. This reduces the number of blkio_mem_regions required for bounce
> buffers to 1 and avoids frequent blkio_mem_region_map/unmap() calls.
> - Switch to .bdrv_co_*() instead of .bdrv_aio_*(). Needed for the bounce
> buffer
> pool's CoQueue.
> v4:
> - Patch 1:
> - Add virtio-blk-vhost-user driver [Kevin]
> - Drop .bdrv_parse_filename() and .bdrv_needs_filename for
> virtio-blk-vhost-vdpa [Stefano]
> - Add copyright and license header [Hanna]
> - Drop .bdrv_parse_filename() in favor of --blockdev or json: [Hanna]
> - Clarify that "filename" is always non-NULL for io_uring [Hanna]
> - Check that virtio-blk-vhost-vdpa "path" option is non-NULL [Hanna]
> - Fix virtio-blk-vhost-vdpa cache.direct=off logic [Hanna]
> - Use macros for driver names [Hanna]
> - Assert that the driver name is valid [Hanna]
> - Update "readonly" property name to "read-only" [Hanna]
> - Call blkio_detach_aio_context() in blkio_close() [Hanna]
> - Avoid uint32_t * to int * casts in blkio_refresh_limits() [Hanna]
> - Remove write zeroes and discard from the todo list [Hanna]
> - Use PRIu32 instead of %d for uint32_t [Hanna]
> - Fix error messages with buf-alignment instead of optimal-io-size [Hanna]
> - Call map/unmap APIs since libblkio alloc/free APIs no longer do that
> - Update QAPI schema QEMU version to 7.2
> - Patch 5:
> - Expand the BDRV_REQ_REGISTERED_BUF flag passthrough and drop
> assert(!flags)
> in drivers [Hanna]
> - Patch 7:
> - Fix BLK->BDRV typo [Hanna]
> - Make BlockRAMRegistrar handle failure [Hanna]
> - Patch 8:
> - Replace memory_region_get_fd() approach with qemu_ram_get_fd()
> - Patch 10:
> - Use (void)ret; to discard unused return value [Hanna]
> - libblkio's blkio_unmap_mem_region() API no longer has a return value
> - Check for registered bufs that cross RAMBlocks [Hanna]
> - Patch 11:
> - Handle bdrv_register_buf() errors [Hanna]
> v3:
> - Add virtio-blk-vhost-vdpa for vdpa-blk devices including VDUSE
> - Add discard and write zeroes support
> - Rebase and adopt latest libblkio APIs
> v2:
> - Add BDRV_REQ_REGISTERED_BUF to bs.supported_write_flags [Stefano]
> - Use new blkioq_get_num_completions() API
> - Implement .bdrv_refresh_limits()
>
> This patch series adds a QEMU BlockDriver for libblkio
> (https://gitlab.com/libblkio/libblkio/), a library for high-performance block
> device I/O. This work was presented at KVM Forum 2022 and slides are available
> here:
> https://static.sched.com/hosted_files/kvmforum2022/8c/libblkio-kvm-forum-2022.pdf
>
> The second patch adds the core BlockDriver and most of the libblkio API usage.
> Three libblkio drivers are included:
> - io_uring
> - virtio-blk-vhost-user
> - virtio-blk-vhost-vdpa
>
> The remainder of the patch series reworks the existing QEMU
> bdrv_register_buf()
> API so virtio-blk emulation efficiently map guest RAM for libblkio - some
> libblkio drivers require that I/O buffer memory is pre-registered (think VFIO,
> vhost, etc).
>
> Vladimir requested performance results that show the effect of the
> BDRV_REQ_REGISTERED_BUF flag. I ran the patches against qemu-storage-daemon's
> vhost-user-blk export with iodepth=1 bs=512 to see the per-request overhead
> due
> to bounce buffer allocation/mapping:
>
> Name IOPS Error
> bounce-buf 4373.81 ± 0.01%
> registered-buf 13062.80 ± 0.67%
>
> The BDRV_REQ_REGISTERED_BUF optimization version is about 3x faster.
>
> See the BlockDriver struct in block/blkio.c for a list of APIs that still need
> to be implemented. The core functionality is covered.
>
> Regarding the design: each libblkio driver is a separately named BlockDriver.
> That means there is an "io_uring" BlockDriver and not a generic "libblkio"
> BlockDriver. This way QAPI and open parameters are type-safe and mandatory
> parameters can be checked by QEMU.
>
> Stefan Hajnoczi (12):
> coroutine: add flag to re-queue at front of CoQueue
> blkio: add libblkio block driver
> numa: call ->ram_block_removed() in ram_block_notifer_remove()
> block: pass size to bdrv_unregister_buf()
> block: use BdrvRequestFlags type for supported flag fields
> block: add BDRV_REQ_REGISTERED_BUF request flag
> block: return errors from bdrv_register_buf()
> block: add BlockRAMRegistrar
> exec/cpu-common: add qemu_ram_get_fd()
> stubs: add qemu_ram_block_from_host() and qemu_ram_get_fd()
> blkio: implement BDRV_REQ_REGISTERED_BUF optimization
> virtio-blk: use BDRV_REQ_REGISTERED_BUF optimization hint
>
> MAINTAINERS | 7 +
> meson_options.txt | 2 +
> qapi/block-core.json | 53 +-
> meson.build | 9 +
> include/block/block-common.h | 9 +
> include/block/block-global-state.h | 10 +-
> include/block/block_int-common.h | 15 +-
> include/exec/cpu-common.h | 1 +
> include/hw/virtio/virtio-blk.h | 2 +
> include/qemu/coroutine.h | 15 +-
> include/sysemu/block-backend-global-state.h | 4 +-
> include/sysemu/block-ram-registrar.h | 37 +
> block.c | 14 +
> block/blkio.c | 1017 +++++++++++++++++++
> block/blkverify.c | 4 +-
> block/block-backend.c | 8 +-
> block/block-ram-registrar.c | 54 +
> block/crypto.c | 4 +-
> block/file-posix.c | 1 -
> block/gluster.c | 1 -
> block/io.c | 101 +-
> block/mirror.c | 2 +
> block/nbd.c | 1 -
> block/nvme.c | 20 +-
> block/parallels.c | 1 -
> block/qcow.c | 2 -
> block/qed.c | 1 -
> block/raw-format.c | 2 +
> block/replication.c | 1 -
> block/ssh.c | 1 -
> block/vhdx.c | 1 -
> hw/block/virtio-blk.c | 39 +-
> hw/core/numa.c | 17 +
> qemu-img.c | 6 +-
> softmmu/physmem.c | 5 +
> stubs/physmem.c | 13 +
> tests/qtest/modules-test.c | 3 +
> util/qemu-coroutine-lock.c | 9 +-
> util/vfio-helpers.c | 5 +-
> block/meson.build | 2 +
> scripts/meson-buildoptions.sh | 3 +
> stubs/meson.build | 1 +
> 42 files changed, 1412 insertions(+), 91 deletions(-)
> create mode 100644 include/sysemu/block-ram-registrar.h
> create mode 100644 block/blkio.c
> create mode 100644 block/block-ram-registrar.c
> create mode 100644 stubs/physmem.c
>
> --
> 2.37.3
>
