| Message ID | 20191018115547.19299-1-dplotnikov@virtuozzo.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | virtio: fix IO request length in virtio SCSI/block #PSBM-78839 | expand |
On Fri, Oct 18, 2019 at 02:55:47PM +0300, Denis Plotnikov wrote: > From: "Denis V. Lunev" <den@openvz.org> > > Linux guests submit IO requests no longer than PAGE_SIZE * max_seg > field reported by SCSI controler. Thus typical sequential read with > 1 MB size results in the following pattern of the IO from the guest: > 8,16 1 15754 2.766095122 2071 D R 2095104 + 1008 [dd] > 8,16 1 15755 2.766108785 2071 D R 2096112 + 1008 [dd] > 8,16 1 15756 2.766113486 2071 D R 2097120 + 32 [dd] > 8,16 1 15757 2.767668961 0 C R 2095104 + 1008 [0] > 8,16 1 15758 2.768534315 0 C R 2096112 + 1008 [0] > 8,16 1 15759 2.768539782 0 C R 2097120 + 32 [0] > The IO was generated by > dd if=/dev/sda of=/dev/null bs=1024 iflag=direct > > This effectively means that on rotational disks we will observe 3 IOPS > for each 2 MBs processed. This definitely negatively affects both > guest and host IO performance. > > The cure is relatively simple - we should report lengthy scatter-gather > ability of the SCSI controller. Fortunately the situation here is very > good. VirtIO transport layer can accomodate 1024 items in one request > while we are using only 128. This situation is present since almost > very beginning. 2 items are dedicated for request metadata thus we > should publish VIRTQUEUE_MAX_SIZE - 2 as max_seg. > > The following pattern is observed after the patch: > 8,16 1 9921 2.662721340 2063 D R 2095104 + 1024 [dd] > 8,16 1 9922 2.662737585 2063 D R 2096128 + 1024 [dd] > 8,16 1 9923 2.665188167 0 C R 2095104 + 1024 [0] > 8,16 1 9924 2.665198777 0 C R 2096128 + 1024 [0] > which is much better. > > The dark side of this patch is that we are tweaking guest visible > parameter, though this should be relatively safe as above transport > layer support is present in QEMU/host Linux for a very long time. > The patch adds configurable property for VirtIO SCSI with a new default > and hardcode option for VirtBlock which does not provide good > configurable framework. > > Unfortunately the commit can not be applied as is. For the real cure we > need guest to be fixed to accomodate that queue length, which is done > only in the latest 4.14 kernel. Thus we are going to expose the property > and tweak it on machine type level. > > The problem with the old kernels is that they have > max_segments <= virtqueue_size restriction which cause the guest > crashing in the case of violation. > To fix the case described above in the old kernels we can increase > virtqueue_size to 256 and max_segments to 254. The pitfall here is > that seabios allows the virtqueue_size-s < 128, however, the seabios > patch extending that value to 256 is pending. If I understand correctly you are relying on Indirect Descriptor support in the guest driver in order to exceed the Virtqueue Descriptor Table size. Unfortunately the "max_segments <= virtqueue_size restriction" is required by the VIRTIO 1.1 specification: 2.6.5.3.1 Driver Requirements: Indirect Descriptors A driver MUST NOT create a descriptor chain longer than the Queue Size of the device. So this idea seems to be in violation of the specification? There is a bug in hw/block/virtio-blk.c:virtio_blk_update_config() and hw/scsi/virtio-scsi.c:virtio_scsi_get_config(): virtio_stl_p(vdev, &blkcfg.seg_max, 128 - 2); This number should be the minimum of blk_get_max_iov() and virtio_queue_get_num(), minus 2 for the header and footer. I looked at the Linux SCSI driver code and it seems each HBA has a single max_segments number - it does not vary on a per-device basis. This could be a problem if two host block device with different max_segments are exposed to the guest through the same virtio-scsi controller. Another bug? :( Anyway, if you want ~1024 descriptors you should set Queue Size to 1024. I don't see a spec-compliant way of doing it otherwise. Hopefully I have overlooked something and there is a nice way to solve this. Stefan
On 10/21/19 4:24 PM, Stefan Hajnoczi wrote: > On Fri, Oct 18, 2019 at 02:55:47PM +0300, Denis Plotnikov wrote: >> From: "Denis V. Lunev" <den@openvz.org> >> >> Linux guests submit IO requests no longer than PAGE_SIZE * max_seg >> field reported by SCSI controler. Thus typical sequential read with >> 1 MB size results in the following pattern of the IO from the guest: >> 8,16 1 15754 2.766095122 2071 D R 2095104 + 1008 [dd] >> 8,16 1 15755 2.766108785 2071 D R 2096112 + 1008 [dd] >> 8,16 1 15756 2.766113486 2071 D R 2097120 + 32 [dd] >> 8,16 1 15757 2.767668961 0 C R 2095104 + 1008 [0] >> 8,16 1 15758 2.768534315 0 C R 2096112 + 1008 [0] >> 8,16 1 15759 2.768539782 0 C R 2097120 + 32 [0] >> The IO was generated by >> dd if=/dev/sda of=/dev/null bs=1024 iflag=direct >> >> This effectively means that on rotational disks we will observe 3 IOPS >> for each 2 MBs processed. This definitely negatively affects both >> guest and host IO performance. >> >> The cure is relatively simple - we should report lengthy scatter-gather >> ability of the SCSI controller. Fortunately the situation here is very >> good. VirtIO transport layer can accomodate 1024 items in one request >> while we are using only 128. This situation is present since almost >> very beginning. 2 items are dedicated for request metadata thus we >> should publish VIRTQUEUE_MAX_SIZE - 2 as max_seg. >> >> The following pattern is observed after the patch: >> 8,16 1 9921 2.662721340 2063 D R 2095104 + 1024 [dd] >> 8,16 1 9922 2.662737585 2063 D R 2096128 + 1024 [dd] >> 8,16 1 9923 2.665188167 0 C R 2095104 + 1024 [0] >> 8,16 1 9924 2.665198777 0 C R 2096128 + 1024 [0] >> which is much better. >> >> The dark side of this patch is that we are tweaking guest visible >> parameter, though this should be relatively safe as above transport >> layer support is present in QEMU/host Linux for a very long time. >> The patch adds configurable property for VirtIO SCSI with a new default >> and hardcode option for VirtBlock which does not provide good >> configurable framework. >> >> Unfortunately the commit can not be applied as is. For the real cure we >> need guest to be fixed to accomodate that queue length, which is done >> only in the latest 4.14 kernel. Thus we are going to expose the property >> and tweak it on machine type level. >> >> The problem with the old kernels is that they have >> max_segments <= virtqueue_size restriction which cause the guest >> crashing in the case of violation. >> To fix the case described above in the old kernels we can increase >> virtqueue_size to 256 and max_segments to 254. The pitfall here is >> that seabios allows the virtqueue_size-s < 128, however, the seabios >> patch extending that value to 256 is pending. > If I understand correctly you are relying on Indirect Descriptor support > in the guest driver in order to exceed the Virtqueue Descriptor Table > size. > > Unfortunately the "max_segments <= virtqueue_size restriction" is > required by the VIRTIO 1.1 specification: > > 2.6.5.3.1 Driver Requirements: Indirect Descriptors > > A driver MUST NOT create a descriptor chain longer than the Queue > Size of the device. > > So this idea seems to be in violation of the specification? > > There is a bug in hw/block/virtio-blk.c:virtio_blk_update_config() and > hw/scsi/virtio-scsi.c:virtio_scsi_get_config(): > > virtio_stl_p(vdev, &blkcfg.seg_max, 128 - 2); > > This number should be the minimum of blk_get_max_iov() and > virtio_queue_get_num(), minus 2 for the header and footer. > > I looked at the Linux SCSI driver code and it seems each HBA has a > single max_segments number - it does not vary on a per-device basis. > This could be a problem if two host block device with different > max_segments are exposed to the guest through the same virtio-scsi > controller. Another bug? :( > > Anyway, if you want ~1024 descriptors you should set Queue Size to 1024. > I don't see a spec-compliant way of doing it otherwise. Hopefully I > have overlooked something and there is a nice way to solve this. > > Stefan you are perfectly correct. We need actually 3 changes to improve guest behavior: 1) This patch, which adds property but does not change anything useful 2) The patch to SeaBIOS, which extends maximum allowed queue size. Right now virtque > 128 results in assert (pending in SeaBIOS list). 3) Increase queue size and max_segments inside machine type. We have done that to 256 and 256 -2 respectively. I think that this exact patch with property does not harm and upon the acceptance could start a discussion about default queue length extension. Den
On 21.10.2019 16:24, Stefan Hajnoczi wrote: > On Fri, Oct 18, 2019 at 02:55:47PM +0300, Denis Plotnikov wrote: >> From: "Denis V. Lunev" <den@openvz.org> >> >> Linux guests submit IO requests no longer than PAGE_SIZE * max_seg >> field reported by SCSI controler. Thus typical sequential read with >> 1 MB size results in the following pattern of the IO from the guest: >> 8,16 1 15754 2.766095122 2071 D R 2095104 + 1008 [dd] >> 8,16 1 15755 2.766108785 2071 D R 2096112 + 1008 [dd] >> 8,16 1 15756 2.766113486 2071 D R 2097120 + 32 [dd] >> 8,16 1 15757 2.767668961 0 C R 2095104 + 1008 [0] >> 8,16 1 15758 2.768534315 0 C R 2096112 + 1008 [0] >> 8,16 1 15759 2.768539782 0 C R 2097120 + 32 [0] >> The IO was generated by >> dd if=/dev/sda of=/dev/null bs=1024 iflag=direct >> >> This effectively means that on rotational disks we will observe 3 IOPS >> for each 2 MBs processed. This definitely negatively affects both >> guest and host IO performance. >> >> The cure is relatively simple - we should report lengthy scatter-gather >> ability of the SCSI controller. Fortunately the situation here is very >> good. VirtIO transport layer can accomodate 1024 items in one request >> while we are using only 128. This situation is present since almost >> very beginning. 2 items are dedicated for request metadata thus we >> should publish VIRTQUEUE_MAX_SIZE - 2 as max_seg. >> >> The following pattern is observed after the patch: >> 8,16 1 9921 2.662721340 2063 D R 2095104 + 1024 [dd] >> 8,16 1 9922 2.662737585 2063 D R 2096128 + 1024 [dd] >> 8,16 1 9923 2.665188167 0 C R 2095104 + 1024 [0] >> 8,16 1 9924 2.665198777 0 C R 2096128 + 1024 [0] >> which is much better. >> >> The dark side of this patch is that we are tweaking guest visible >> parameter, though this should be relatively safe as above transport >> layer support is present in QEMU/host Linux for a very long time. >> The patch adds configurable property for VirtIO SCSI with a new default >> and hardcode option for VirtBlock which does not provide good >> configurable framework. >> >> Unfortunately the commit can not be applied as is. For the real cure we >> need guest to be fixed to accomodate that queue length, which is done >> only in the latest 4.14 kernel. Thus we are going to expose the property >> and tweak it on machine type level. >> >> The problem with the old kernels is that they have >> max_segments <= virtqueue_size restriction which cause the guest >> crashing in the case of violation. >> To fix the case described above in the old kernels we can increase >> virtqueue_size to 256 and max_segments to 254. The pitfall here is >> that seabios allows the virtqueue_size-s < 128, however, the seabios >> patch extending that value to 256 is pending. > If I understand correctly you are relying on Indirect Descriptor support > in the guest driver in order to exceed the Virtqueue Descriptor Table > size. > > Unfortunately the "max_segments <= virtqueue_size restriction" is > required by the VIRTIO 1.1 specification: > > 2.6.5.3.1 Driver Requirements: Indirect Descriptors > > A driver MUST NOT create a descriptor chain longer than the Queue > Size of the device. > > So this idea seems to be in violation of the specification? > > There is a bug in hw/block/virtio-blk.c:virtio_blk_update_config() and > hw/scsi/virtio-scsi.c:virtio_scsi_get_config(): > > virtio_stl_p(vdev, &blkcfg.seg_max, 128 - 2); > > This number should be the minimum of blk_get_max_iov() and > virtio_queue_get_num(), minus 2 for the header and footer. Stefan, It seems VitrioSCSI don't have a direct link to blk, apart of VirtIOBlock->blk, and the link to a blk comes with each scsi request. I suspect that idea here is that a single virtioscsi can serve several blk-s. If my assumption is corect, then we can't get blk_get_max_iov() on virtioscsi configuration stage and we shouldn't take into account max_iov and limit max_segments with virtio_queue_get_num()-2 only. Is it so, or is there any other details to take into account? Thanks! Denis > > I looked at the Linux SCSI driver code and it seems each HBA has a > single max_segments number - it does not vary on a per-device basis. > This could be a problem if two host block device with different > max_segments are exposed to the guest through the same virtio-scsi > controller. Another bug? :( > > Anyway, if you want ~1024 descriptors you should set Queue Size to 1024. > I don't see a spec-compliant way of doing it otherwise. Hopefully I > have overlooked something and there is a nice way to solve this. > > Stefan
On Tue, Oct 22, 2019 at 04:01:57AM +0000, Denis Lunev wrote: > On 10/21/19 4:24 PM, Stefan Hajnoczi wrote: > > On Fri, Oct 18, 2019 at 02:55:47PM +0300, Denis Plotnikov wrote: > >> From: "Denis V. Lunev" <den@openvz.org> > >> > >> Linux guests submit IO requests no longer than PAGE_SIZE * max_seg > >> field reported by SCSI controler. Thus typical sequential read with > >> 1 MB size results in the following pattern of the IO from the guest: > >> 8,16 1 15754 2.766095122 2071 D R 2095104 + 1008 [dd] > >> 8,16 1 15755 2.766108785 2071 D R 2096112 + 1008 [dd] > >> 8,16 1 15756 2.766113486 2071 D R 2097120 + 32 [dd] > >> 8,16 1 15757 2.767668961 0 C R 2095104 + 1008 [0] > >> 8,16 1 15758 2.768534315 0 C R 2096112 + 1008 [0] > >> 8,16 1 15759 2.768539782 0 C R 2097120 + 32 [0] > >> The IO was generated by > >> dd if=/dev/sda of=/dev/null bs=1024 iflag=direct > >> > >> This effectively means that on rotational disks we will observe 3 IOPS > >> for each 2 MBs processed. This definitely negatively affects both > >> guest and host IO performance. > >> > >> The cure is relatively simple - we should report lengthy scatter-gather > >> ability of the SCSI controller. Fortunately the situation here is very > >> good. VirtIO transport layer can accomodate 1024 items in one request > >> while we are using only 128. This situation is present since almost > >> very beginning. 2 items are dedicated for request metadata thus we > >> should publish VIRTQUEUE_MAX_SIZE - 2 as max_seg. > >> > >> The following pattern is observed after the patch: > >> 8,16 1 9921 2.662721340 2063 D R 2095104 + 1024 [dd] > >> 8,16 1 9922 2.662737585 2063 D R 2096128 + 1024 [dd] > >> 8,16 1 9923 2.665188167 0 C R 2095104 + 1024 [0] > >> 8,16 1 9924 2.665198777 0 C R 2096128 + 1024 [0] > >> which is much better. > >> > >> The dark side of this patch is that we are tweaking guest visible > >> parameter, though this should be relatively safe as above transport > >> layer support is present in QEMU/host Linux for a very long time. > >> The patch adds configurable property for VirtIO SCSI with a new default > >> and hardcode option for VirtBlock which does not provide good > >> configurable framework. > >> > >> Unfortunately the commit can not be applied as is. For the real cure we > >> need guest to be fixed to accomodate that queue length, which is done > >> only in the latest 4.14 kernel. Thus we are going to expose the property > >> and tweak it on machine type level. > >> > >> The problem with the old kernels is that they have > >> max_segments <= virtqueue_size restriction which cause the guest > >> crashing in the case of violation. > >> To fix the case described above in the old kernels we can increase > >> virtqueue_size to 256 and max_segments to 254. The pitfall here is > >> that seabios allows the virtqueue_size-s < 128, however, the seabios > >> patch extending that value to 256 is pending. > > If I understand correctly you are relying on Indirect Descriptor support > > in the guest driver in order to exceed the Virtqueue Descriptor Table > > size. > > > > Unfortunately the "max_segments <= virtqueue_size restriction" is > > required by the VIRTIO 1.1 specification: > > > > 2.6.5.3.1 Driver Requirements: Indirect Descriptors > > > > A driver MUST NOT create a descriptor chain longer than the Queue > > Size of the device. > > > > So this idea seems to be in violation of the specification? > > > > There is a bug in hw/block/virtio-blk.c:virtio_blk_update_config() and > > hw/scsi/virtio-scsi.c:virtio_scsi_get_config(): > > > > virtio_stl_p(vdev, &blkcfg.seg_max, 128 - 2); > > > > This number should be the minimum of blk_get_max_iov() and > > virtio_queue_get_num(), minus 2 for the header and footer. > > > > I looked at the Linux SCSI driver code and it seems each HBA has a > > single max_segments number - it does not vary on a per-device basis. > > This could be a problem if two host block device with different > > max_segments are exposed to the guest through the same virtio-scsi > > controller. Another bug? :( > > > > Anyway, if you want ~1024 descriptors you should set Queue Size to 1024. > > I don't see a spec-compliant way of doing it otherwise. Hopefully I > > have overlooked something and there is a nice way to solve this. > > > > Stefan > you are perfectly correct. We need actually 3 changes to improve > guest behavior: > 1) This patch, which adds property but does not change anything > useful This patch is problematic because it causes existing guest drivers to violate the VIRTIO specification (or fail) if the value is set too high. In addition, it does not take into account the virtqueue size so the default value is too low when the user sets -device ...,queue-size=1024. Let's calculate blkcfg.seg_max based on the virtqueue size as mentioned in my previous email instead. There is one caveat with my suggestion: drivers are allowed to access VIRTIO Configuration Space before virtqueue setup has determined the final size. Therefore the value of this field can change after virtqueue setup. Drivers that set a custom virtqueue size would need to read the value after virtqueue setup. (Linux drivers do not modify the virtqueue size so it won't affect them.) Stefan
On 10/23/19 5:17 PM, Stefan Hajnoczi wrote: > On Tue, Oct 22, 2019 at 04:01:57AM +0000, Denis Lunev wrote: >> On 10/21/19 4:24 PM, Stefan Hajnoczi wrote: >>> On Fri, Oct 18, 2019 at 02:55:47PM +0300, Denis Plotnikov wrote: >>>> From: "Denis V. Lunev" <den@openvz.org> >>>> >>>> Linux guests submit IO requests no longer than PAGE_SIZE * max_seg >>>> field reported by SCSI controler. Thus typical sequential read with >>>> 1 MB size results in the following pattern of the IO from the guest: >>>> 8,16 1 15754 2.766095122 2071 D R 2095104 + 1008 [dd] >>>> 8,16 1 15755 2.766108785 2071 D R 2096112 + 1008 [dd] >>>> 8,16 1 15756 2.766113486 2071 D R 2097120 + 32 [dd] >>>> 8,16 1 15757 2.767668961 0 C R 2095104 + 1008 [0] >>>> 8,16 1 15758 2.768534315 0 C R 2096112 + 1008 [0] >>>> 8,16 1 15759 2.768539782 0 C R 2097120 + 32 [0] >>>> The IO was generated by >>>> dd if=/dev/sda of=/dev/null bs=1024 iflag=direct >>>> >>>> This effectively means that on rotational disks we will observe 3 IOPS >>>> for each 2 MBs processed. This definitely negatively affects both >>>> guest and host IO performance. >>>> >>>> The cure is relatively simple - we should report lengthy scatter-gather >>>> ability of the SCSI controller. Fortunately the situation here is very >>>> good. VirtIO transport layer can accomodate 1024 items in one request >>>> while we are using only 128. This situation is present since almost >>>> very beginning. 2 items are dedicated for request metadata thus we >>>> should publish VIRTQUEUE_MAX_SIZE - 2 as max_seg. >>>> >>>> The following pattern is observed after the patch: >>>> 8,16 1 9921 2.662721340 2063 D R 2095104 + 1024 [dd] >>>> 8,16 1 9922 2.662737585 2063 D R 2096128 + 1024 [dd] >>>> 8,16 1 9923 2.665188167 0 C R 2095104 + 1024 [0] >>>> 8,16 1 9924 2.665198777 0 C R 2096128 + 1024 [0] >>>> which is much better. >>>> >>>> The dark side of this patch is that we are tweaking guest visible >>>> parameter, though this should be relatively safe as above transport >>>> layer support is present in QEMU/host Linux for a very long time. >>>> The patch adds configurable property for VirtIO SCSI with a new default >>>> and hardcode option for VirtBlock which does not provide good >>>> configurable framework. >>>> >>>> Unfortunately the commit can not be applied as is. For the real cure we >>>> need guest to be fixed to accomodate that queue length, which is done >>>> only in the latest 4.14 kernel. Thus we are going to expose the property >>>> and tweak it on machine type level. >>>> >>>> The problem with the old kernels is that they have >>>> max_segments <= virtqueue_size restriction which cause the guest >>>> crashing in the case of violation. >>>> To fix the case described above in the old kernels we can increase >>>> virtqueue_size to 256 and max_segments to 254. The pitfall here is >>>> that seabios allows the virtqueue_size-s < 128, however, the seabios >>>> patch extending that value to 256 is pending. >>> If I understand correctly you are relying on Indirect Descriptor support >>> in the guest driver in order to exceed the Virtqueue Descriptor Table >>> size. >>> >>> Unfortunately the "max_segments <= virtqueue_size restriction" is >>> required by the VIRTIO 1.1 specification: >>> >>> 2.6.5.3.1 Driver Requirements: Indirect Descriptors >>> >>> A driver MUST NOT create a descriptor chain longer than the Queue >>> Size of the device. >>> >>> So this idea seems to be in violation of the specification? >>> >>> There is a bug in hw/block/virtio-blk.c:virtio_blk_update_config() and >>> hw/scsi/virtio-scsi.c:virtio_scsi_get_config(): >>> >>> virtio_stl_p(vdev, &blkcfg.seg_max, 128 - 2); >>> >>> This number should be the minimum of blk_get_max_iov() and >>> virtio_queue_get_num(), minus 2 for the header and footer. >>> >>> I looked at the Linux SCSI driver code and it seems each HBA has a >>> single max_segments number - it does not vary on a per-device basis. >>> This could be a problem if two host block device with different >>> max_segments are exposed to the guest through the same virtio-scsi >>> controller. Another bug? :( >>> >>> Anyway, if you want ~1024 descriptors you should set Queue Size to 1024. >>> I don't see a spec-compliant way of doing it otherwise. Hopefully I >>> have overlooked something and there is a nice way to solve this. >>> >>> Stefan >> you are perfectly correct. We need actually 3 changes to improve >> guest behavior: >> 1) This patch, which adds property but does not change anything >> useful > This patch is problematic because it causes existing guest drivers to > violate the VIRTIO specification (or fail) if the value is set too high. > In addition, it does not take into account the virtqueue size so the > default value is too low when the user sets -device ...,queue-size=1024. > > Let's calculate blkcfg.seg_max based on the virtqueue size as mentioned > in my previous email instead. As far as I understand maximum amount of segments could be more than virtqueue size for indirect requests (allowed in VirtIO 1.0). > There is one caveat with my suggestion: drivers are allowed to access > VIRTIO Configuration Space before virtqueue setup has determined the > final size. Therefore the value of this field can change after > virtqueue setup. Drivers that set a custom virtqueue size would need to > read the value after virtqueue setup. (Linux drivers do not modify the > virtqueue size so it won't affect them.) > > Stefan I think that we should do that a little bit different :) We can not change max_segs just if queue size is changed, this should be somehow bound to machine type. Thus I propose to add "automatic" value, i.e. if max_segs is set to 0 the code should set it to queue size -2. This should be default. Alternatively the value from max_segs should be taken. Will this work for you? Please note, currently the specification could also be violated if we will reduce queue size to 64 :) Den
On Fri, Oct 18, 2019 at 02:55:47PM +0300, Denis Plotnikov wrote: > From: "Denis V. Lunev" <den@openvz.org> > > Linux guests submit IO requests no longer than PAGE_SIZE * max_seg > field reported by SCSI controler. Thus typical sequential read with > 1 MB size results in the following pattern of the IO from the guest: > 8,16 1 15754 2.766095122 2071 D R 2095104 + 1008 [dd] > 8,16 1 15755 2.766108785 2071 D R 2096112 + 1008 [dd] > 8,16 1 15756 2.766113486 2071 D R 2097120 + 32 [dd] > 8,16 1 15757 2.767668961 0 C R 2095104 + 1008 [0] > 8,16 1 15758 2.768534315 0 C R 2096112 + 1008 [0] > 8,16 1 15759 2.768539782 0 C R 2097120 + 32 [0] > The IO was generated by > dd if=/dev/sda of=/dev/null bs=1024 iflag=direct > > This effectively means that on rotational disks we will observe 3 IOPS > for each 2 MBs processed. This definitely negatively affects both > guest and host IO performance. > > The cure is relatively simple - we should report lengthy scatter-gather > ability of the SCSI controller. Fortunately the situation here is very > good. VirtIO transport layer can accomodate 1024 items in one request > while we are using only 128. This situation is present since almost > very beginning. 2 items are dedicated for request metadata thus we > should publish VIRTQUEUE_MAX_SIZE - 2 as max_seg. > > The following pattern is observed after the patch: > 8,16 1 9921 2.662721340 2063 D R 2095104 + 1024 [dd] > 8,16 1 9922 2.662737585 2063 D R 2096128 + 1024 [dd] > 8,16 1 9923 2.665188167 0 C R 2095104 + 1024 [0] > 8,16 1 9924 2.665198777 0 C R 2096128 + 1024 [0] > which is much better. > > The dark side of this patch is that we are tweaking guest visible > parameter, though this should be relatively safe as above transport > layer support is present in QEMU/host Linux for a very long time. > The patch adds configurable property for VirtIO SCSI with a new default > and hardcode option for VirtBlock which does not provide good > configurable framework. > > Unfortunately the commit can not be applied as is. For the real cure we > need guest to be fixed to accomodate that queue length, which is done > only in the latest 4.14 kernel. Thus we are going to expose the property > and tweak it on machine type level. > > The problem with the old kernels is that they have > max_segments <= virtqueue_size restriction which cause the guest > crashing in the case of violation. This isn't just in the guests: virtio spec also seems to imply this, or at least be vague on this point. So I think it'll need a feature bit. Doing that in a safe way will also allow being compatible with old guests. The only downside is it's a bit more work as we need to spec this out and add guest support. > To fix the case described above in the old kernels we can increase > virtqueue_size to 256 and max_segments to 254. The pitfall here is > that seabios allows the virtqueue_size-s < 128, however, the seabios > patch extending that value to 256 is pending. And the fix here is just to limit large vq size to virtio 1.0. In that mode it's fine I think: /* check if the queue is available */ if (vp->use_modern) { num = vp_read(&vp->common, virtio_pci_common_cfg, queue_size); if (num > MAX_QUEUE_NUM) { vp_write(&vp->common, virtio_pci_common_cfg, queue_size, MAX_QUEUE_NUM); num = vp_read(&vp->common, virtio_pci_common_cfg, queue_size); } } else { num = vp_read(&vp->legacy, virtio_pci_legacy, queue_num); } > CC: "Michael S. Tsirkin" <mst@redhat.com> > CC: Stefan Hajnoczi <stefanha@redhat.com> > CC: Kevin Wolf <kwolf@redhat.com> > CC: Max Reitz <mreitz@redhat.com> > CC: Gerd Hoffmann <kraxel@redhat.com> > Signed-off-by: Denis V. Lunev <den@openvz.org> > Signed-off-by: Denis Plotnikov <dplotnikov@virtuozzo.com> > --- > hw/block/virtio-blk.c | 3 ++- > hw/scsi/vhost-scsi.c | 2 ++ > hw/scsi/virtio-scsi.c | 4 +++- > include/hw/virtio/virtio-blk.h | 1 + > include/hw/virtio/virtio-scsi.h | 1 + > 5 files changed, 9 insertions(+), 2 deletions(-) > > diff --git a/hw/block/virtio-blk.c b/hw/block/virtio-blk.c > index 06e57a4d39..b2eaeeaf67 100644 > --- a/hw/block/virtio-blk.c > +++ b/hw/block/virtio-blk.c > @@ -903,7 +903,7 @@ static void virtio_blk_update_config(VirtIODevice *vdev, uint8_t *config) > blk_get_geometry(s->blk, &capacity); > memset(&blkcfg, 0, sizeof(blkcfg)); > virtio_stq_p(vdev, &blkcfg.capacity, capacity); > - virtio_stl_p(vdev, &blkcfg.seg_max, 128 - 2); > + virtio_stl_p(vdev, &blkcfg.seg_max, s->conf.max_segments); > virtio_stw_p(vdev, &blkcfg.geometry.cylinders, conf->cyls); > virtio_stl_p(vdev, &blkcfg.blk_size, blk_size); > virtio_stw_p(vdev, &blkcfg.min_io_size, conf->min_io_size / blk_size); > @@ -1240,6 +1240,7 @@ static Property virtio_blk_properties[] = { > conf.max_discard_sectors, BDRV_REQUEST_MAX_SECTORS), > DEFINE_PROP_UINT32("max-write-zeroes-sectors", VirtIOBlock, > conf.max_write_zeroes_sectors, BDRV_REQUEST_MAX_SECTORS), > + DEFINE_PROP_UINT32("max_segments", VirtIOBlock, conf.max_segments, 126), > DEFINE_PROP_END_OF_LIST(), > }; > I'd worry that it's too easy to create a broken config with this parameter. > diff --git a/hw/scsi/vhost-scsi.c b/hw/scsi/vhost-scsi.c > index 61e2e57da9..fa3b377807 100644 > --- a/hw/scsi/vhost-scsi.c > +++ b/hw/scsi/vhost-scsi.c > @@ -242,6 +242,8 @@ static Property vhost_scsi_properties[] = { > DEFINE_PROP_BIT64("t10_pi", VHostSCSICommon, host_features, > VIRTIO_SCSI_F_T10_PI, > false), > + DEFINE_PROP_UINT32("max_segments", VirtIOSCSICommon, conf.max_segments, > + 126), > DEFINE_PROP_END_OF_LIST(), > }; > > diff --git a/hw/scsi/virtio-scsi.c b/hw/scsi/virtio-scsi.c > index 839f120256..8b070ddeed 100644 > --- a/hw/scsi/virtio-scsi.c > +++ b/hw/scsi/virtio-scsi.c > @@ -650,7 +650,7 @@ static void virtio_scsi_get_config(VirtIODevice *vdev, > VirtIOSCSICommon *s = VIRTIO_SCSI_COMMON(vdev); > > virtio_stl_p(vdev, &scsiconf->num_queues, s->conf.num_queues); > - virtio_stl_p(vdev, &scsiconf->seg_max, 128 - 2); > + virtio_stl_p(vdev, &scsiconf->seg_max, s->conf.max_segments); > virtio_stl_p(vdev, &scsiconf->max_sectors, s->conf.max_sectors); > virtio_stl_p(vdev, &scsiconf->cmd_per_lun, s->conf.cmd_per_lun); > virtio_stl_p(vdev, &scsiconf->event_info_size, sizeof(VirtIOSCSIEvent)); > @@ -948,6 +948,8 @@ static Property virtio_scsi_properties[] = { > VIRTIO_SCSI_F_CHANGE, true), > DEFINE_PROP_LINK("iothread", VirtIOSCSI, parent_obj.conf.iothread, > TYPE_IOTHREAD, IOThread *), > + DEFINE_PROP_UINT32("max_segments", VirtIOSCSI, parent_obj.conf.max_segments, > + 126), > DEFINE_PROP_END_OF_LIST(), > }; > > diff --git a/include/hw/virtio/virtio-blk.h b/include/hw/virtio/virtio-blk.h > index cddcfbebe9..22da23a4a3 100644 > --- a/include/hw/virtio/virtio-blk.h > +++ b/include/hw/virtio/virtio-blk.h > @@ -40,6 +40,7 @@ struct VirtIOBlkConf > uint16_t queue_size; > uint32_t max_discard_sectors; > uint32_t max_write_zeroes_sectors; > + uint32_t max_segments; > }; > > struct VirtIOBlockDataPlane; > diff --git a/include/hw/virtio/virtio-scsi.h b/include/hw/virtio/virtio-scsi.h > index 4c0bcdb788..1e5805eec4 100644 > --- a/include/hw/virtio/virtio-scsi.h > +++ b/include/hw/virtio/virtio-scsi.h > @@ -49,6 +49,7 @@ struct VirtIOSCSIConf { > uint32_t num_queues; > uint32_t virtqueue_size; > uint32_t max_sectors; > + uint32_t max_segments; > uint32_t cmd_per_lun; > #ifdef CONFIG_VHOST_SCSI > char *vhostfd; > -- > 2.17.0
On Mon, Oct 21, 2019 at 02:24:55PM +0100, Stefan Hajnoczi wrote: > On Fri, Oct 18, 2019 at 02:55:47PM +0300, Denis Plotnikov wrote: > > From: "Denis V. Lunev" <den@openvz.org> > > > > Linux guests submit IO requests no longer than PAGE_SIZE * max_seg > > field reported by SCSI controler. Thus typical sequential read with > > 1 MB size results in the following pattern of the IO from the guest: > > 8,16 1 15754 2.766095122 2071 D R 2095104 + 1008 [dd] > > 8,16 1 15755 2.766108785 2071 D R 2096112 + 1008 [dd] > > 8,16 1 15756 2.766113486 2071 D R 2097120 + 32 [dd] > > 8,16 1 15757 2.767668961 0 C R 2095104 + 1008 [0] > > 8,16 1 15758 2.768534315 0 C R 2096112 + 1008 [0] > > 8,16 1 15759 2.768539782 0 C R 2097120 + 32 [0] > > The IO was generated by > > dd if=/dev/sda of=/dev/null bs=1024 iflag=direct > > > > This effectively means that on rotational disks we will observe 3 IOPS > > for each 2 MBs processed. This definitely negatively affects both > > guest and host IO performance. > > > > The cure is relatively simple - we should report lengthy scatter-gather > > ability of the SCSI controller. Fortunately the situation here is very > > good. VirtIO transport layer can accomodate 1024 items in one request > > while we are using only 128. This situation is present since almost > > very beginning. 2 items are dedicated for request metadata thus we > > should publish VIRTQUEUE_MAX_SIZE - 2 as max_seg. > > > > The following pattern is observed after the patch: > > 8,16 1 9921 2.662721340 2063 D R 2095104 + 1024 [dd] > > 8,16 1 9922 2.662737585 2063 D R 2096128 + 1024 [dd] > > 8,16 1 9923 2.665188167 0 C R 2095104 + 1024 [0] > > 8,16 1 9924 2.665198777 0 C R 2096128 + 1024 [0] > > which is much better. > > > > The dark side of this patch is that we are tweaking guest visible > > parameter, though this should be relatively safe as above transport > > layer support is present in QEMU/host Linux for a very long time. > > The patch adds configurable property for VirtIO SCSI with a new default > > and hardcode option for VirtBlock which does not provide good > > configurable framework. > > > > Unfortunately the commit can not be applied as is. For the real cure we > > need guest to be fixed to accomodate that queue length, which is done > > only in the latest 4.14 kernel. Thus we are going to expose the property > > and tweak it on machine type level. > > > > The problem with the old kernels is that they have > > max_segments <= virtqueue_size restriction which cause the guest > > crashing in the case of violation. > > To fix the case described above in the old kernels we can increase > > virtqueue_size to 256 and max_segments to 254. The pitfall here is > > that seabios allows the virtqueue_size-s < 128, however, the seabios > > patch extending that value to 256 is pending. > > If I understand correctly you are relying on Indirect Descriptor support > in the guest driver in order to exceed the Virtqueue Descriptor Table > size. > > Unfortunately the "max_segments <= virtqueue_size restriction" is > required by the VIRTIO 1.1 specification: > > 2.6.5.3.1 Driver Requirements: Indirect Descriptors > > A driver MUST NOT create a descriptor chain longer than the Queue > Size of the device. > > So this idea seems to be in violation of the specification? > > There is a bug in hw/block/virtio-blk.c:virtio_blk_update_config() and > hw/scsi/virtio-scsi.c:virtio_scsi_get_config(): > > virtio_stl_p(vdev, &blkcfg.seg_max, 128 - 2); > > This number should be the minimum of blk_get_max_iov() and > virtio_queue_get_num(), minus 2 for the header and footer. > > I looked at the Linux SCSI driver code and it seems each HBA has a > single max_segments number - it does not vary on a per-device basis. > This could be a problem if two host block device with different > max_segments are exposed to the guest through the same virtio-scsi > controller. Another bug? :( > > Anyway, if you want ~1024 descriptors you should set Queue Size to 1024. > I don't see a spec-compliant way of doing it otherwise. Hopefully I > have overlooked something and there is a nice way to solve this. > > Stefan We can extend the spec of course. And we can also have different vq sizes between legacy and modern interfaces.
On 10/24/19 12:28 AM, Michael S. Tsirkin wrote: > On Fri, Oct 18, 2019 at 02:55:47PM +0300, Denis Plotnikov wrote: >> From: "Denis V. Lunev" <den@openvz.org> >> >> Linux guests submit IO requests no longer than PAGE_SIZE * max_seg >> field reported by SCSI controler. Thus typical sequential read with >> 1 MB size results in the following pattern of the IO from the guest: >> 8,16 1 15754 2.766095122 2071 D R 2095104 + 1008 [dd] >> 8,16 1 15755 2.766108785 2071 D R 2096112 + 1008 [dd] >> 8,16 1 15756 2.766113486 2071 D R 2097120 + 32 [dd] >> 8,16 1 15757 2.767668961 0 C R 2095104 + 1008 [0] >> 8,16 1 15758 2.768534315 0 C R 2096112 + 1008 [0] >> 8,16 1 15759 2.768539782 0 C R 2097120 + 32 [0] >> The IO was generated by >> dd if=/dev/sda of=/dev/null bs=1024 iflag=direct >> >> This effectively means that on rotational disks we will observe 3 IOPS >> for each 2 MBs processed. This definitely negatively affects both >> guest and host IO performance. >> >> The cure is relatively simple - we should report lengthy scatter-gather >> ability of the SCSI controller. Fortunately the situation here is very >> good. VirtIO transport layer can accomodate 1024 items in one request >> while we are using only 128. This situation is present since almost >> very beginning. 2 items are dedicated for request metadata thus we >> should publish VIRTQUEUE_MAX_SIZE - 2 as max_seg. >> >> The following pattern is observed after the patch: >> 8,16 1 9921 2.662721340 2063 D R 2095104 + 1024 [dd] >> 8,16 1 9922 2.662737585 2063 D R 2096128 + 1024 [dd] >> 8,16 1 9923 2.665188167 0 C R 2095104 + 1024 [0] >> 8,16 1 9924 2.665198777 0 C R 2096128 + 1024 [0] >> which is much better. >> >> The dark side of this patch is that we are tweaking guest visible >> parameter, though this should be relatively safe as above transport >> layer support is present in QEMU/host Linux for a very long time. >> The patch adds configurable property for VirtIO SCSI with a new default >> and hardcode option for VirtBlock which does not provide good >> configurable framework. >> >> Unfortunately the commit can not be applied as is. For the real cure we >> need guest to be fixed to accomodate that queue length, which is done >> only in the latest 4.14 kernel. Thus we are going to expose the property >> and tweak it on machine type level. >> >> The problem with the old kernels is that they have >> max_segments <= virtqueue_size restriction which cause the guest >> crashing in the case of violation. > This isn't just in the guests: virtio spec also seems to imply this, > or at least be vague on this point. > > So I think it'll need a feature bit. > Doing that in a safe way will also allow being compatible with old guests. > > The only downside is it's a bit more work as we need to > spec this out and add guest support. > >> To fix the case described above in the old kernels we can increase >> virtqueue_size to 256 and max_segments to 254. The pitfall here is >> that seabios allows the virtqueue_size-s < 128, however, the seabios >> patch extending that value to 256 is pending. > > And the fix here is just to limit large vq size to virtio 1.0. > In that mode it's fine I think: > > > /* check if the queue is available */ > if (vp->use_modern) { > num = vp_read(&vp->common, virtio_pci_common_cfg, queue_size); > if (num > MAX_QUEUE_NUM) { > vp_write(&vp->common, virtio_pci_common_cfg, queue_size, > MAX_QUEUE_NUM); > num = vp_read(&vp->common, virtio_pci_common_cfg, queue_size); > } > } else { > num = vp_read(&vp->legacy, virtio_pci_legacy, queue_num); > } you mean to put the code like this into virtio_pci_realize() inside QEMU? If no, can you pls clarify which component should be touched. Den
On Thu, Oct 24, 2019 at 11:34:34AM +0000, Denis Lunev wrote: > On 10/24/19 12:28 AM, Michael S. Tsirkin wrote: > > On Fri, Oct 18, 2019 at 02:55:47PM +0300, Denis Plotnikov wrote: > >> From: "Denis V. Lunev" <den@openvz.org> > >> > >> Linux guests submit IO requests no longer than PAGE_SIZE * max_seg > >> field reported by SCSI controler. Thus typical sequential read with > >> 1 MB size results in the following pattern of the IO from the guest: > >> 8,16 1 15754 2.766095122 2071 D R 2095104 + 1008 [dd] > >> 8,16 1 15755 2.766108785 2071 D R 2096112 + 1008 [dd] > >> 8,16 1 15756 2.766113486 2071 D R 2097120 + 32 [dd] > >> 8,16 1 15757 2.767668961 0 C R 2095104 + 1008 [0] > >> 8,16 1 15758 2.768534315 0 C R 2096112 + 1008 [0] > >> 8,16 1 15759 2.768539782 0 C R 2097120 + 32 [0] > >> The IO was generated by > >> dd if=/dev/sda of=/dev/null bs=1024 iflag=direct > >> > >> This effectively means that on rotational disks we will observe 3 IOPS > >> for each 2 MBs processed. This definitely negatively affects both > >> guest and host IO performance. > >> > >> The cure is relatively simple - we should report lengthy scatter-gather > >> ability of the SCSI controller. Fortunately the situation here is very > >> good. VirtIO transport layer can accomodate 1024 items in one request > >> while we are using only 128. This situation is present since almost > >> very beginning. 2 items are dedicated for request metadata thus we > >> should publish VIRTQUEUE_MAX_SIZE - 2 as max_seg. > >> > >> The following pattern is observed after the patch: > >> 8,16 1 9921 2.662721340 2063 D R 2095104 + 1024 [dd] > >> 8,16 1 9922 2.662737585 2063 D R 2096128 + 1024 [dd] > >> 8,16 1 9923 2.665188167 0 C R 2095104 + 1024 [0] > >> 8,16 1 9924 2.665198777 0 C R 2096128 + 1024 [0] > >> which is much better. > >> > >> The dark side of this patch is that we are tweaking guest visible > >> parameter, though this should be relatively safe as above transport > >> layer support is present in QEMU/host Linux for a very long time. > >> The patch adds configurable property for VirtIO SCSI with a new default > >> and hardcode option for VirtBlock which does not provide good > >> configurable framework. > >> > >> Unfortunately the commit can not be applied as is. For the real cure we > >> need guest to be fixed to accomodate that queue length, which is done > >> only in the latest 4.14 kernel. Thus we are going to expose the property > >> and tweak it on machine type level. > >> > >> The problem with the old kernels is that they have > >> max_segments <= virtqueue_size restriction which cause the guest > >> crashing in the case of violation. > > This isn't just in the guests: virtio spec also seems to imply this, > > or at least be vague on this point. > > > > So I think it'll need a feature bit. > > Doing that in a safe way will also allow being compatible with old guests. > > > > The only downside is it's a bit more work as we need to > > spec this out and add guest support. > > > >> To fix the case described above in the old kernels we can increase > >> virtqueue_size to 256 and max_segments to 254. The pitfall here is > >> that seabios allows the virtqueue_size-s < 128, however, the seabios > >> patch extending that value to 256 is pending. > > > > And the fix here is just to limit large vq size to virtio 1.0. > > In that mode it's fine I think: > > > > > > /* check if the queue is available */ > > if (vp->use_modern) { > > num = vp_read(&vp->common, virtio_pci_common_cfg, queue_size); > > if (num > MAX_QUEUE_NUM) { > > vp_write(&vp->common, virtio_pci_common_cfg, queue_size, > > MAX_QUEUE_NUM); > > num = vp_read(&vp->common, virtio_pci_common_cfg, queue_size); > > } > > } else { > > num = vp_read(&vp->legacy, virtio_pci_legacy, queue_num); > > } > > you mean to put the code like this into virtio_pci_realize() inside QEMU? > > If no, can you pls clarify which component should be touched. > > Den I mean: - add an API to change the default queue size - add a validate features callback, in there check and for modern flag set in features increase the queue size maybe all this is too much work, we could block this for transitional devices, but your patch does not do it, you need to check that legacy is enabled not that modern is not disabled.
On Wed, Oct 23, 2019 at 05:28:17PM -0400, Michael S. Tsirkin wrote: > On Fri, Oct 18, 2019 at 02:55:47PM +0300, Denis Plotnikov wrote: > > From: "Denis V. Lunev" <den@openvz.org> > > > > Linux guests submit IO requests no longer than PAGE_SIZE * max_seg > > field reported by SCSI controler. Thus typical sequential read with > > 1 MB size results in the following pattern of the IO from the guest: > > 8,16 1 15754 2.766095122 2071 D R 2095104 + 1008 [dd] > > 8,16 1 15755 2.766108785 2071 D R 2096112 + 1008 [dd] > > 8,16 1 15756 2.766113486 2071 D R 2097120 + 32 [dd] > > 8,16 1 15757 2.767668961 0 C R 2095104 + 1008 [0] > > 8,16 1 15758 2.768534315 0 C R 2096112 + 1008 [0] > > 8,16 1 15759 2.768539782 0 C R 2097120 + 32 [0] > > The IO was generated by > > dd if=/dev/sda of=/dev/null bs=1024 iflag=direct > > > > This effectively means that on rotational disks we will observe 3 IOPS > > for each 2 MBs processed. This definitely negatively affects both > > guest and host IO performance. > > > > The cure is relatively simple - we should report lengthy scatter-gather > > ability of the SCSI controller. Fortunately the situation here is very > > good. VirtIO transport layer can accomodate 1024 items in one request > > while we are using only 128. This situation is present since almost > > very beginning. 2 items are dedicated for request metadata thus we > > should publish VIRTQUEUE_MAX_SIZE - 2 as max_seg. > > > > The following pattern is observed after the patch: > > 8,16 1 9921 2.662721340 2063 D R 2095104 + 1024 [dd] > > 8,16 1 9922 2.662737585 2063 D R 2096128 + 1024 [dd] > > 8,16 1 9923 2.665188167 0 C R 2095104 + 1024 [0] > > 8,16 1 9924 2.665198777 0 C R 2096128 + 1024 [0] > > which is much better. > > > > The dark side of this patch is that we are tweaking guest visible > > parameter, though this should be relatively safe as above transport > > layer support is present in QEMU/host Linux for a very long time. > > The patch adds configurable property for VirtIO SCSI with a new default > > and hardcode option for VirtBlock which does not provide good > > configurable framework. > > > > Unfortunately the commit can not be applied as is. For the real cure we > > need guest to be fixed to accomodate that queue length, which is done > > only in the latest 4.14 kernel. Thus we are going to expose the property > > and tweak it on machine type level. > > > > The problem with the old kernels is that they have > > max_segments <= virtqueue_size restriction which cause the guest > > crashing in the case of violation. > > This isn't just in the guests: virtio spec also seems to imply this, > or at least be vague on this point. > > So I think it'll need a feature bit. > Doing that in a safe way will also allow being compatible with old guests. The spec is quite explicit about this: 2.6.5 The Virtqueue Descriptor Table The number of descriptors in the table is defined by the queue size for this virtqueue: this is the maximum possible descriptor chain length. and: 2.6.5.3.1 Driver Requirements: Indirect Descriptors A driver MUST NOT create a descriptor chain longer than the Queue Size of the device. If some drivers or devices allow longer descriptor chains today that's an implementation quirk but a new feature bit is definitely required to officially allow longer descriptor chains. Stefan
On 06.11.2019 15:03, Michael S. Tsirkin wrote: > On Thu, Oct 24, 2019 at 11:34:34AM +0000, Denis Lunev wrote: >> On 10/24/19 12:28 AM, Michael S. Tsirkin wrote: >>> On Fri, Oct 18, 2019 at 02:55:47PM +0300, Denis Plotnikov wrote: >>>> From: "Denis V. Lunev" <den@openvz.org> >>>> >>>> Linux guests submit IO requests no longer than PAGE_SIZE * max_seg >>>> field reported by SCSI controler. Thus typical sequential read with >>>> 1 MB size results in the following pattern of the IO from the guest: >>>> 8,16 1 15754 2.766095122 2071 D R 2095104 + 1008 [dd] >>>> 8,16 1 15755 2.766108785 2071 D R 2096112 + 1008 [dd] >>>> 8,16 1 15756 2.766113486 2071 D R 2097120 + 32 [dd] >>>> 8,16 1 15757 2.767668961 0 C R 2095104 + 1008 [0] >>>> 8,16 1 15758 2.768534315 0 C R 2096112 + 1008 [0] >>>> 8,16 1 15759 2.768539782 0 C R 2097120 + 32 [0] >>>> The IO was generated by >>>> dd if=/dev/sda of=/dev/null bs=1024 iflag=direct >>>> >>>> This effectively means that on rotational disks we will observe 3 IOPS >>>> for each 2 MBs processed. This definitely negatively affects both >>>> guest and host IO performance. >>>> >>>> The cure is relatively simple - we should report lengthy scatter-gather >>>> ability of the SCSI controller. Fortunately the situation here is very >>>> good. VirtIO transport layer can accomodate 1024 items in one request >>>> while we are using only 128. This situation is present since almost >>>> very beginning. 2 items are dedicated for request metadata thus we >>>> should publish VIRTQUEUE_MAX_SIZE - 2 as max_seg. >>>> >>>> The following pattern is observed after the patch: >>>> 8,16 1 9921 2.662721340 2063 D R 2095104 + 1024 [dd] >>>> 8,16 1 9922 2.662737585 2063 D R 2096128 + 1024 [dd] >>>> 8,16 1 9923 2.665188167 0 C R 2095104 + 1024 [0] >>>> 8,16 1 9924 2.665198777 0 C R 2096128 + 1024 [0] >>>> which is much better. >>>> >>>> The dark side of this patch is that we are tweaking guest visible >>>> parameter, though this should be relatively safe as above transport >>>> layer support is present in QEMU/host Linux for a very long time. >>>> The patch adds configurable property for VirtIO SCSI with a new default >>>> and hardcode option for VirtBlock which does not provide good >>>> configurable framework. >>>> >>>> Unfortunately the commit can not be applied as is. For the real cure we >>>> need guest to be fixed to accomodate that queue length, which is done >>>> only in the latest 4.14 kernel. Thus we are going to expose the property >>>> and tweak it on machine type level. >>>> >>>> The problem with the old kernels is that they have >>>> max_segments <= virtqueue_size restriction which cause the guest >>>> crashing in the case of violation. >>> This isn't just in the guests: virtio spec also seems to imply this, >>> or at least be vague on this point. >>> >>> So I think it'll need a feature bit. >>> Doing that in a safe way will also allow being compatible with old guests. >>> >>> The only downside is it's a bit more work as we need to >>> spec this out and add guest support. >>> >>>> To fix the case described above in the old kernels we can increase >>>> virtqueue_size to 256 and max_segments to 254. The pitfall here is >>>> that seabios allows the virtqueue_size-s < 128, however, the seabios >>>> patch extending that value to 256 is pending. >>> And the fix here is just to limit large vq size to virtio 1.0. >>> In that mode it's fine I think: >>> >>> >>> /* check if the queue is available */ >>> if (vp->use_modern) { >>> num = vp_read(&vp->common, virtio_pci_common_cfg, queue_size); >>> if (num > MAX_QUEUE_NUM) { >>> vp_write(&vp->common, virtio_pci_common_cfg, queue_size, >>> MAX_QUEUE_NUM); >>> num = vp_read(&vp->common, virtio_pci_common_cfg, queue_size); >>> } >>> } else { >>> num = vp_read(&vp->legacy, virtio_pci_legacy, queue_num); >>> } >> you mean to put the code like this into virtio_pci_realize() inside QEMU? >> >> If no, can you pls clarify which component should be touched. >> >> Den > I mean: > - add an API to change the default queue size > - add a validate features callback, in there check and for modern > flag set in features increase the queue size > > maybe all this is too much work, we could block this > for transitional devices, but your patch does not do it, > you need to check that legacy is enabled not that modern > is not disabled. To develop the idea of how to adjust queue size further I'd like to summarize what we have: 1. Variatly of gusts without(?) queue size limitations which can support queue sizes up to MAX(1024) 2. seabios setups with two possible max queue size limitations: 128 and 256 (recently commited) 3. non-sebios setups with unknown max queue-size limitations Taking into account that queue size may be limited in bios(efi), to safely support gueue sizes > 128 we need to distinguish those how can support greater_than_128 from those who can't. seabios potentially can't do it, so, as far as I understood, the idea is to start with queue size=128 and then increase the queue size when the guest driver is engaged. To achieve that, we need to 1. understand, which driver is currently working with a virtio device: seabios, guest, other. Things here are quite complex, since we can't modify any guest, seabios or other drivers to explicitly tell that to device 2. be able to increase queue size dynamically (re-create queues?). At the time, this functionality is absent, at least in qemu virtio-scsi. Is it possible by design? 3. choose a place for queue size extending (re-creation). VirtioDeviceClass->reset? I actually don't know how to do it reliably, so would really appreciate sone help or advice. You've mentioned that old seabios won't use the modern interface, so would it be ok, if we * define DEFAULT_QUEUE_SIZE = 128 * leave queues creation as is at VirtioDeviceClass->realize() with queue_size = conf.queue_size * on VirtioDeviceClass->reset() we check if the device accessed through "legacy" interface if so, then (in pseudocode) if (current_queue_size > DEFAULT_QUEUE_SIZE) { for (queue in all_queues) { reduce_queue_size(queue, DEFAULT_QUEUE_SIZE) // recreate_queue() ? } } else if (conf.queue_size > current_queue_size) { for (queue in all_queues) { increase_queue_size(queue, conf.queue_size) } } Might this approach work? Does it what you meant? Denis > > >
On Wed, Nov 13, 2019 at 12:38:48PM +0000, Denis Plotnikov wrote: > > > On 06.11.2019 15:03, Michael S. Tsirkin wrote: > > On Thu, Oct 24, 2019 at 11:34:34AM +0000, Denis Lunev wrote: > >> On 10/24/19 12:28 AM, Michael S. Tsirkin wrote: > >>> On Fri, Oct 18, 2019 at 02:55:47PM +0300, Denis Plotnikov wrote: > >>>> From: "Denis V. Lunev" <den@openvz.org> > >>>> > >>>> Linux guests submit IO requests no longer than PAGE_SIZE * max_seg > >>>> field reported by SCSI controler. Thus typical sequential read with > >>>> 1 MB size results in the following pattern of the IO from the guest: > >>>> 8,16 1 15754 2.766095122 2071 D R 2095104 + 1008 [dd] > >>>> 8,16 1 15755 2.766108785 2071 D R 2096112 + 1008 [dd] > >>>> 8,16 1 15756 2.766113486 2071 D R 2097120 + 32 [dd] > >>>> 8,16 1 15757 2.767668961 0 C R 2095104 + 1008 [0] > >>>> 8,16 1 15758 2.768534315 0 C R 2096112 + 1008 [0] > >>>> 8,16 1 15759 2.768539782 0 C R 2097120 + 32 [0] > >>>> The IO was generated by > >>>> dd if=/dev/sda of=/dev/null bs=1024 iflag=direct > >>>> > >>>> This effectively means that on rotational disks we will observe 3 IOPS > >>>> for each 2 MBs processed. This definitely negatively affects both > >>>> guest and host IO performance. > >>>> > >>>> The cure is relatively simple - we should report lengthy scatter-gather > >>>> ability of the SCSI controller. Fortunately the situation here is very > >>>> good. VirtIO transport layer can accomodate 1024 items in one request > >>>> while we are using only 128. This situation is present since almost > >>>> very beginning. 2 items are dedicated for request metadata thus we > >>>> should publish VIRTQUEUE_MAX_SIZE - 2 as max_seg. > >>>> > >>>> The following pattern is observed after the patch: > >>>> 8,16 1 9921 2.662721340 2063 D R 2095104 + 1024 [dd] > >>>> 8,16 1 9922 2.662737585 2063 D R 2096128 + 1024 [dd] > >>>> 8,16 1 9923 2.665188167 0 C R 2095104 + 1024 [0] > >>>> 8,16 1 9924 2.665198777 0 C R 2096128 + 1024 [0] > >>>> which is much better. > >>>> > >>>> The dark side of this patch is that we are tweaking guest visible > >>>> parameter, though this should be relatively safe as above transport > >>>> layer support is present in QEMU/host Linux for a very long time. > >>>> The patch adds configurable property for VirtIO SCSI with a new default > >>>> and hardcode option for VirtBlock which does not provide good > >>>> configurable framework. > >>>> > >>>> Unfortunately the commit can not be applied as is. For the real cure we > >>>> need guest to be fixed to accomodate that queue length, which is done > >>>> only in the latest 4.14 kernel. Thus we are going to expose the property > >>>> and tweak it on machine type level. > >>>> > >>>> The problem with the old kernels is that they have > >>>> max_segments <= virtqueue_size restriction which cause the guest > >>>> crashing in the case of violation. > >>> This isn't just in the guests: virtio spec also seems to imply this, > >>> or at least be vague on this point. > >>> > >>> So I think it'll need a feature bit. > >>> Doing that in a safe way will also allow being compatible with old guests. > >>> > >>> The only downside is it's a bit more work as we need to > >>> spec this out and add guest support. > >>> > >>>> To fix the case described above in the old kernels we can increase > >>>> virtqueue_size to 256 and max_segments to 254. The pitfall here is > >>>> that seabios allows the virtqueue_size-s < 128, however, the seabios > >>>> patch extending that value to 256 is pending. > >>> And the fix here is just to limit large vq size to virtio 1.0. > >>> In that mode it's fine I think: > >>> > >>> > >>> /* check if the queue is available */ > >>> if (vp->use_modern) { > >>> num = vp_read(&vp->common, virtio_pci_common_cfg, queue_size); > >>> if (num > MAX_QUEUE_NUM) { > >>> vp_write(&vp->common, virtio_pci_common_cfg, queue_size, > >>> MAX_QUEUE_NUM); > >>> num = vp_read(&vp->common, virtio_pci_common_cfg, queue_size); > >>> } > >>> } else { > >>> num = vp_read(&vp->legacy, virtio_pci_legacy, queue_num); > >>> } > >> you mean to put the code like this into virtio_pci_realize() inside QEMU? > >> > >> If no, can you pls clarify which component should be touched. > >> > >> Den > > I mean: > > - add an API to change the default queue size > > - add a validate features callback, in there check and for modern > > flag set in features increase the queue size > > > > maybe all this is too much work, we could block this > > for transitional devices, but your patch does not do it, > > you need to check that legacy is enabled not that modern > > is not disabled. > To develop the idea of how to adjust queue size further I'd like to > summarize what we have: > > 1. Variatly of gusts without(?) queue size limitations which can support > queue sizes up to MAX(1024) > > 2. seabios setups with two possible max queue size limitations: 128 and > 256 (recently commited) > > 3. non-sebios setups with unknown max queue-size limitations > > Taking into account that queue size may be limited in bios(efi), to > safely support gueue sizes > 128 we need to distinguish those how can > support greater_than_128 from those who can't. > seabios potentially can't do it, so, as far as I understood, the idea is > to start with queue size=128 and then increase the queue size when the > guest driver is engaged. > > To achieve that, we need to > > 1. understand, which driver is currently working with a virtio device: > seabios, guest, other. Things > here are quite complex, since we can't modify any guest, seabios or > other drivers to explicitly tell > that to device Anyone negotiating VIRTIO_1 > 2. be able to increase queue size dynamically (re-create queues?). At > the time, this functionality > is absent, at least in qemu virtio-scsi. > Is it possible by design? Why not, it's just an array. This is what I meant when I said we need an API to resize a queue. > 3. choose a place for queue size extending (re-creation). > VirtioDeviceClass->reset? Definitely not reset, that gets you back to original state. > I actually don't know how to do it reliably, so would really appreciate > sone help or advice. validate features sounds like a good place. this is why I wrote "add a validate features callback". > > You've mentioned that old seabios won't use the modern interface, so > would it be ok, if we > > * define DEFAULT_QUEUE_SIZE = 128 > * leave queues creation as is at VirtioDeviceClass->realize() > with queue_size = conf.queue_size > * on VirtioDeviceClass->reset() we check if the device accessed > through "legacy" interface > if so, then (in pseudocode) > if (current_queue_size > DEFAULT_QUEUE_SIZE) { > for (queue in all_queues) { > reduce_queue_size(queue, DEFAULT_QUEUE_SIZE) // > recreate_queue() ? > } > } > else > if (conf.queue_size > current_queue_size) { > for (queue in all_queues) { > increase_queue_size(queue, conf.queue_size) > } > } > > Might this approach work? Does it what you meant? > > Denis I don't think you can do anything useful in reset. We need to check features after they have been negotiated. So we'd start with a small queue min(DEFAULT_QUEUE_SIZE, current_queue_size)? and if VIRTIO_1 is set increase the size. This is very compatible but it is certainly ugly as we are second-guessing the user. Simpler idea: add a new property that is simply unsupported with legacy. E.g. "modern-queue-size" ? If someone sets it, legacy must be disabled otherwise we fail. Way less compatible but hey. > > > > > > >
On 13.11.2019 16:18, Michael S. Tsirkin wrote: > On Wed, Nov 13, 2019 at 12:38:48PM +0000, Denis Plotnikov wrote: >> >> On 06.11.2019 15:03, Michael S. Tsirkin wrote: >>> On Thu, Oct 24, 2019 at 11:34:34AM +0000, Denis Lunev wrote: >>>> On 10/24/19 12:28 AM, Michael S. Tsirkin wrote: >>>>> On Fri, Oct 18, 2019 at 02:55:47PM +0300, Denis Plotnikov wrote: >>>>>> From: "Denis V. Lunev" <den@openvz.org> >>>>>> >>>>>> Linux guests submit IO requests no longer than PAGE_SIZE * max_seg >>>>>> field reported by SCSI controler. Thus typical sequential read with >>>>>> 1 MB size results in the following pattern of the IO from the guest: >>>>>> 8,16 1 15754 2.766095122 2071 D R 2095104 + 1008 [dd] >>>>>> 8,16 1 15755 2.766108785 2071 D R 2096112 + 1008 [dd] >>>>>> 8,16 1 15756 2.766113486 2071 D R 2097120 + 32 [dd] >>>>>> 8,16 1 15757 2.767668961 0 C R 2095104 + 1008 [0] >>>>>> 8,16 1 15758 2.768534315 0 C R 2096112 + 1008 [0] >>>>>> 8,16 1 15759 2.768539782 0 C R 2097120 + 32 [0] >>>>>> The IO was generated by >>>>>> dd if=/dev/sda of=/dev/null bs=1024 iflag=direct >>>>>> >>>>>> This effectively means that on rotational disks we will observe 3 IOPS >>>>>> for each 2 MBs processed. This definitely negatively affects both >>>>>> guest and host IO performance. >>>>>> >>>>>> The cure is relatively simple - we should report lengthy scatter-gather >>>>>> ability of the SCSI controller. Fortunately the situation here is very >>>>>> good. VirtIO transport layer can accomodate 1024 items in one request >>>>>> while we are using only 128. This situation is present since almost >>>>>> very beginning. 2 items are dedicated for request metadata thus we >>>>>> should publish VIRTQUEUE_MAX_SIZE - 2 as max_seg. >>>>>> >>>>>> The following pattern is observed after the patch: >>>>>> 8,16 1 9921 2.662721340 2063 D R 2095104 + 1024 [dd] >>>>>> 8,16 1 9922 2.662737585 2063 D R 2096128 + 1024 [dd] >>>>>> 8,16 1 9923 2.665188167 0 C R 2095104 + 1024 [0] >>>>>> 8,16 1 9924 2.665198777 0 C R 2096128 + 1024 [0] >>>>>> which is much better. >>>>>> >>>>>> The dark side of this patch is that we are tweaking guest visible >>>>>> parameter, though this should be relatively safe as above transport >>>>>> layer support is present in QEMU/host Linux for a very long time. >>>>>> The patch adds configurable property for VirtIO SCSI with a new default >>>>>> and hardcode option for VirtBlock which does not provide good >>>>>> configurable framework. >>>>>> >>>>>> Unfortunately the commit can not be applied as is. For the real cure we >>>>>> need guest to be fixed to accomodate that queue length, which is done >>>>>> only in the latest 4.14 kernel. Thus we are going to expose the property >>>>>> and tweak it on machine type level. >>>>>> >>>>>> The problem with the old kernels is that they have >>>>>> max_segments <= virtqueue_size restriction which cause the guest >>>>>> crashing in the case of violation. >>>>> This isn't just in the guests: virtio spec also seems to imply this, >>>>> or at least be vague on this point. >>>>> >>>>> So I think it'll need a feature bit. >>>>> Doing that in a safe way will also allow being compatible with old guests. >>>>> >>>>> The only downside is it's a bit more work as we need to >>>>> spec this out and add guest support. >>>>> >>>>>> To fix the case described above in the old kernels we can increase >>>>>> virtqueue_size to 256 and max_segments to 254. The pitfall here is >>>>>> that seabios allows the virtqueue_size-s < 128, however, the seabios >>>>>> patch extending that value to 256 is pending. >>>>> And the fix here is just to limit large vq size to virtio 1.0. >>>>> In that mode it's fine I think: >>>>> >>>>> >>>>> /* check if the queue is available */ >>>>> if (vp->use_modern) { >>>>> num = vp_read(&vp->common, virtio_pci_common_cfg, queue_size); >>>>> if (num > MAX_QUEUE_NUM) { >>>>> vp_write(&vp->common, virtio_pci_common_cfg, queue_size, >>>>> MAX_QUEUE_NUM); >>>>> num = vp_read(&vp->common, virtio_pci_common_cfg, queue_size); >>>>> } >>>>> } else { >>>>> num = vp_read(&vp->legacy, virtio_pci_legacy, queue_num); >>>>> } >>>> you mean to put the code like this into virtio_pci_realize() inside QEMU? >>>> >>>> If no, can you pls clarify which component should be touched. >>>> >>>> Den >>> I mean: >>> - add an API to change the default queue size >>> - add a validate features callback, in there check and for modern >>> flag set in features increase the queue size >>> >>> maybe all this is too much work, we could block this >>> for transitional devices, but your patch does not do it, >>> you need to check that legacy is enabled not that modern >>> is not disabled. >> To develop the idea of how to adjust queue size further I'd like to >> summarize what we have: >> >> 1. Variatly of gusts without(?) queue size limitations which can support >> queue sizes up to MAX(1024) >> >> 2. seabios setups with two possible max queue size limitations: 128 and >> 256 (recently commited) >> >> 3. non-sebios setups with unknown max queue-size limitations >> >> Taking into account that queue size may be limited in bios(efi), to >> safely support gueue sizes > 128 we need to distinguish those how can >> support greater_than_128 from those who can't. >> seabios potentially can't do it, so, as far as I understood, the idea is >> to start with queue size=128 and then increase the queue size when the >> guest driver is engaged. >> >> To achieve that, we need to >> >> 1. understand, which driver is currently working with a virtio device: >> seabios, guest, other. Things >> here are quite complex, since we can't modify any guest, seabios or >> other drivers to explicitly tell >> that to device > Anyone negotiating VIRTIO_1 > >> 2. be able to increase queue size dynamically (re-create queues?). At >> the time, this functionality >> is absent, at least in qemu virtio-scsi. >> Is it possible by design? > Why not, it's just an array. > This is what I meant when I said we need an API to resize a queue. > >> 3. choose a place for queue size extending (re-creation). >> VirtioDeviceClass->reset? > Definitely not reset, that gets you back to original state. > >> I actually don't know how to do it reliably, so would really appreciate >> sone help or advice. > validate features sounds like a good place. > this is why I wrote "add a validate features callback". > >> You've mentioned that old seabios won't use the modern interface, so >> would it be ok, if we >> >> * define DEFAULT_QUEUE_SIZE = 128 >> * leave queues creation as is at VirtioDeviceClass->realize() >> with queue_size = conf.queue_size >> * on VirtioDeviceClass->reset() we check if the device accessed >> through "legacy" interface >> if so, then (in pseudocode) >> if (current_queue_size > DEFAULT_QUEUE_SIZE) { >> for (queue in all_queues) { >> reduce_queue_size(queue, DEFAULT_QUEUE_SIZE) // >> recreate_queue() ? >> } >> } >> else >> if (conf.queue_size > current_queue_size) { >> for (queue in all_queues) { >> increase_queue_size(queue, conf.queue_size) >> } >> } >> >> Might this approach work? Does it what you meant? >> >> Denis > > I don't think you can do anything useful in reset. We need to check > features after they have been negotiated. So we'd start with a small > queue min(DEFAULT_QUEUE_SIZE, current_queue_size)? > and if VIRTIO_1 is set increase the size. > > This is very compatible but it is certainly ugly as we are > second-guessing the user. > > > Simpler idea: add a new property that is simply > unsupported with legacy. E.g. "modern-queue-size" ? > If someone sets it, legacy must be disabled otherwise we fail. > > Way less compatible but hey. If I got the idea correctly, in that case the old seabios won't start. Hence we won't achieve what we want: increase the queue size in the guests which use the old seabios. May be the ugly way is worth to be implemented since it would allow to add some performance to the existing guests? Denis > > >>> >>>
On 06.11.2019 15:03, Michael S. Tsirkin wrote: > On Thu, Oct 24, 2019 at 11:34:34AM +0000, Denis Lunev wrote: >> On 10/24/19 12:28 AM, Michael S. Tsirkin wrote: >>> On Fri, Oct 18, 2019 at 02:55:47PM +0300, Denis Plotnikov wrote: >>>> From: "Denis V. Lunev" <den@openvz.org> >>>> >>>> Linux guests submit IO requests no longer than PAGE_SIZE * max_seg >>>> field reported by SCSI controler. Thus typical sequential read with >>>> 1 MB size results in the following pattern of the IO from the guest: >>>> 8,16 1 15754 2.766095122 2071 D R 2095104 + 1008 [dd] >>>> 8,16 1 15755 2.766108785 2071 D R 2096112 + 1008 [dd] >>>> 8,16 1 15756 2.766113486 2071 D R 2097120 + 32 [dd] >>>> 8,16 1 15757 2.767668961 0 C R 2095104 + 1008 [0] >>>> 8,16 1 15758 2.768534315 0 C R 2096112 + 1008 [0] >>>> 8,16 1 15759 2.768539782 0 C R 2097120 + 32 [0] >>>> The IO was generated by >>>> dd if=/dev/sda of=/dev/null bs=1024 iflag=direct >>>> >>>> This effectively means that on rotational disks we will observe 3 IOPS >>>> for each 2 MBs processed. This definitely negatively affects both >>>> guest and host IO performance. >>>> >>>> The cure is relatively simple - we should report lengthy scatter-gather >>>> ability of the SCSI controller. Fortunately the situation here is very >>>> good. VirtIO transport layer can accomodate 1024 items in one request >>>> while we are using only 128. This situation is present since almost >>>> very beginning. 2 items are dedicated for request metadata thus we >>>> should publish VIRTQUEUE_MAX_SIZE - 2 as max_seg. >>>> >>>> The following pattern is observed after the patch: >>>> 8,16 1 9921 2.662721340 2063 D R 2095104 + 1024 [dd] >>>> 8,16 1 9922 2.662737585 2063 D R 2096128 + 1024 [dd] >>>> 8,16 1 9923 2.665188167 0 C R 2095104 + 1024 [0] >>>> 8,16 1 9924 2.665198777 0 C R 2096128 + 1024 [0] >>>> which is much better. >>>> >>>> The dark side of this patch is that we are tweaking guest visible >>>> parameter, though this should be relatively safe as above transport >>>> layer support is present in QEMU/host Linux for a very long time. >>>> The patch adds configurable property for VirtIO SCSI with a new default >>>> and hardcode option for VirtBlock which does not provide good >>>> configurable framework. >>>> >>>> Unfortunately the commit can not be applied as is. For the real cure we >>>> need guest to be fixed to accomodate that queue length, which is done >>>> only in the latest 4.14 kernel. Thus we are going to expose the property >>>> and tweak it on machine type level. >>>> >>>> The problem with the old kernels is that they have >>>> max_segments <= virtqueue_size restriction which cause the guest >>>> crashing in the case of violation. >>> This isn't just in the guests: virtio spec also seems to imply this, >>> or at least be vague on this point. >>> >>> So I think it'll need a feature bit. >>> Doing that in a safe way will also allow being compatible with old guests. >>> >>> The only downside is it's a bit more work as we need to >>> spec this out and add guest support. >>> >>>> To fix the case described above in the old kernels we can increase >>>> virtqueue_size to 256 and max_segments to 254. The pitfall here is >>>> that seabios allows the virtqueue_size-s < 128, however, the seabios >>>> patch extending that value to 256 is pending. >>> And the fix here is just to limit large vq size to virtio 1.0. >>> In that mode it's fine I think: >>> >>> >>> /* check if the queue is available */ >>> if (vp->use_modern) { >>> num = vp_read(&vp->common, virtio_pci_common_cfg, queue_size); >>> if (num > MAX_QUEUE_NUM) { >>> vp_write(&vp->common, virtio_pci_common_cfg, queue_size, >>> MAX_QUEUE_NUM); >>> num = vp_read(&vp->common, virtio_pci_common_cfg, queue_size); >>> } >>> } else { >>> num = vp_read(&vp->legacy, virtio_pci_legacy, queue_num); >>> } The same seabios snippet, but more detailed: vp_find_vq() { ... /* check if the queue is available */ if (vp->use_modern) { num = vp_read(&vp->common, virtio_pci_common_cfg, queue_size); if (num > MAX_QUEUE_NUM) { vp_write(&vp->common, virtio_pci_common_cfg, queue_size, MAX_QUEUE_NUM); num = vp_read(&vp->common, virtio_pci_common_cfg, queue_size); } } else { num = vp_read(&vp->legacy, virtio_pci_legacy, queue_num); } if (!num) { dprintf(1, "ERROR: queue size is 0\n"); goto fail; } if (num > MAX_QUEUE_NUM) { dprintf(1, "ERROR: queue size %d > %d\n", num, MAX_QUEUE_NUM); goto fail; } ... } It turned out that the problem is here, but not because of the seabios code. The virtqueue size is written and then incorrect value is re-read. Thanks to Roman Kagan (rkagan@virtuozzo.com) for investigating the root cause of the problem. As the code states, for the modern devices, seabios reads the queue size and if it's greater than seabios can support, reduce the queue size to the max seabios supported value. This doesn't work. The reason is that the size is read from the virtio device, virtio_pci_common_read() { ... case VIRTIO_PCI_COMMON_Q_SIZE: val = virtio_queue_get_num(vdev, vdev->queue_sel); break; ... } but is written to the proxy virtio_pci_common_write() { ... case VIRTIO_PCI_COMMON_Q_SIZE: proxy->vqs[vdev->queue_sel].num = val; break; ... }. The final stage of the size setting is propagated it from the proxy to the device on virtqueue enabling: virtio_cpi_common_write() { ... case VIRTIO_PCI_COMMON_Q_ENABLE: virtio_queue_set_num(vdev, vdev->queue_sel, proxy->vqs[vdev->queue_sel].num); virtio_queue_set_rings(vdev, vdev->queue_sel, ((uint64_t)proxy->vqs[vdev->queue_sel].desc[1]) << 32 | proxy->vqs[vdev->queue_sel].desc[0], ((uint64_t)proxy->vqs[vdev->queue_sel].avail[1]) << 32 | proxy->vqs[vdev->queue_sel].avail[0], ((uint64_t)proxy->vqs[vdev->queue_sel].used[1]) << 32 | proxy->vqs[vdev->queue_sel].used[0]); proxy->vqs[vdev->queue_sel].enabled = 1; break; ... }. So we have the following workflow: suppose the device has virtqueue size = 256 and seabios MAX_QUEUE_NUM = 128. In that case the seabios works like: 1. if vp_modern read the size (256) 2. 256 > 128 3. write virtqueue size = 128 4. re-read virtqueue size = 256 !!! 5. fail because of the check if (num > MAX_QUEUE_NUM) { dprintf(1, "ERROR: queue size %d > %d\n", num, MAX_QUEUE_NUM); goto fail; } To fix the issue, we need to read and write the virtqueue size from the same place. Should we do with the proxy? Is there any reason to read from the device and write to the proxy? Furthermore, the size setting has a few flaws: 1. The size being set should be a power of 2 2. The size being set should be less or equal to the virtqueue size (and be greater that 2?) Denis >> you mean to put the code like this into virtio_pci_realize() inside QEMU? >> >> If no, can you pls clarify which component should be touched. >> >> Den > I mean: > - add an API to change the default queue size > - add a validate features callback, in there check and for modern > flag set in features increase the queue size > > maybe all this is too much work, we could block this > for transitional devices, but your patch does not do it, > you need to check that legacy is enabled not that modern > is not disabled. > > >
Ping! On 25.11.2019 12:16, Denis Plotnikov wrote: > > > On 06.11.2019 15:03, Michael S. Tsirkin wrote: >> On Thu, Oct 24, 2019 at 11:34:34AM +0000, Denis Lunev wrote: >>> On 10/24/19 12:28 AM, Michael S. Tsirkin wrote: >>>> On Fri, Oct 18, 2019 at 02:55:47PM +0300, Denis Plotnikov wrote: >>>>> From: "Denis V. Lunev" <den@openvz.org> >>>>> >>>>> Linux guests submit IO requests no longer than PAGE_SIZE * max_seg >>>>> field reported by SCSI controler. Thus typical sequential read with >>>>> 1 MB size results in the following pattern of the IO from the guest: >>>>> 8,16 1 15754 2.766095122 2071 D R 2095104 + 1008 [dd] >>>>> 8,16 1 15755 2.766108785 2071 D R 2096112 + 1008 [dd] >>>>> 8,16 1 15756 2.766113486 2071 D R 2097120 + 32 [dd] >>>>> 8,16 1 15757 2.767668961 0 C R 2095104 + 1008 [0] >>>>> 8,16 1 15758 2.768534315 0 C R 2096112 + 1008 [0] >>>>> 8,16 1 15759 2.768539782 0 C R 2097120 + 32 [0] >>>>> The IO was generated by >>>>> dd if=/dev/sda of=/dev/null bs=1024 iflag=direct >>>>> >>>>> This effectively means that on rotational disks we will observe 3 >>>>> IOPS >>>>> for each 2 MBs processed. This definitely negatively affects both >>>>> guest and host IO performance. >>>>> >>>>> The cure is relatively simple - we should report lengthy >>>>> scatter-gather >>>>> ability of the SCSI controller. Fortunately the situation here is >>>>> very >>>>> good. VirtIO transport layer can accomodate 1024 items in one request >>>>> while we are using only 128. This situation is present since almost >>>>> very beginning. 2 items are dedicated for request metadata thus we >>>>> should publish VIRTQUEUE_MAX_SIZE - 2 as max_seg. >>>>> >>>>> The following pattern is observed after the patch: >>>>> 8,16 1 9921 2.662721340 2063 D R 2095104 + 1024 [dd] >>>>> 8,16 1 9922 2.662737585 2063 D R 2096128 + 1024 [dd] >>>>> 8,16 1 9923 2.665188167 0 C R 2095104 + 1024 [0] >>>>> 8,16 1 9924 2.665198777 0 C R 2096128 + 1024 [0] >>>>> which is much better. >>>>> >>>>> The dark side of this patch is that we are tweaking guest visible >>>>> parameter, though this should be relatively safe as above transport >>>>> layer support is present in QEMU/host Linux for a very long time. >>>>> The patch adds configurable property for VirtIO SCSI with a new >>>>> default >>>>> and hardcode option for VirtBlock which does not provide good >>>>> configurable framework. >>>>> >>>>> Unfortunately the commit can not be applied as is. For the real >>>>> cure we >>>>> need guest to be fixed to accomodate that queue length, which is done >>>>> only in the latest 4.14 kernel. Thus we are going to expose the >>>>> property >>>>> and tweak it on machine type level. >>>>> >>>>> The problem with the old kernels is that they have >>>>> max_segments <= virtqueue_size restriction which cause the guest >>>>> crashing in the case of violation. >>>> This isn't just in the guests: virtio spec also seems to imply this, >>>> or at least be vague on this point. >>>> >>>> So I think it'll need a feature bit. >>>> Doing that in a safe way will also allow being compatible with old >>>> guests. >>>> >>>> The only downside is it's a bit more work as we need to >>>> spec this out and add guest support. >>>> >>>>> To fix the case described above in the old kernels we can increase >>>>> virtqueue_size to 256 and max_segments to 254. The pitfall here is >>>>> that seabios allows the virtqueue_size-s < 128, however, the seabios >>>>> patch extending that value to 256 is pending. >>>> And the fix here is just to limit large vq size to virtio 1.0. >>>> In that mode it's fine I think: >>>> >>>> >>>> /* check if the queue is available */ >>>> if (vp->use_modern) { >>>> num = vp_read(&vp->common, virtio_pci_common_cfg, queue_size); >>>> if (num > MAX_QUEUE_NUM) { >>>> vp_write(&vp->common, virtio_pci_common_cfg, queue_size, >>>> MAX_QUEUE_NUM); >>>> num = vp_read(&vp->common, virtio_pci_common_cfg, >>>> queue_size); >>>> } >>>> } else { >>>> num = vp_read(&vp->legacy, virtio_pci_legacy, queue_num); >>>> } > The same seabios snippet, but more detailed: > > vp_find_vq() > { > ... > /* check if the queue is available */ > if (vp->use_modern) { > num = vp_read(&vp->common, virtio_pci_common_cfg, queue_size); > if (num > MAX_QUEUE_NUM) { > vp_write(&vp->common, virtio_pci_common_cfg, queue_size, > MAX_QUEUE_NUM); > num = vp_read(&vp->common, virtio_pci_common_cfg, queue_size); > } > } else { > num = vp_read(&vp->legacy, virtio_pci_legacy, queue_num); > } > if (!num) { > dprintf(1, "ERROR: queue size is 0\n"); > goto fail; > } > if (num > MAX_QUEUE_NUM) { > dprintf(1, "ERROR: queue size %d > %d\n", num, MAX_QUEUE_NUM); > goto fail; > } > ... > } > > It turned out that the problem is here, but not because of the seabios > code. > The virtqueue size is written and then incorrect value is re-read. > Thanks to Roman Kagan (rkagan@virtuozzo.com) for investigating the > root cause of the problem. > > As the code states, for the modern devices, seabios reads the queue > size and if it's > greater than seabios can support, reduce the queue size to the max > seabios supported value. > > This doesn't work. > > The reason is that the size is read from the virtio device, > > virtio_pci_common_read() > { > ... > case VIRTIO_PCI_COMMON_Q_SIZE: > val = virtio_queue_get_num(vdev, vdev->queue_sel); > break; > ... > } > > but is written to the proxy > > virtio_pci_common_write() > { > ... > case VIRTIO_PCI_COMMON_Q_SIZE: > proxy->vqs[vdev->queue_sel].num = val; > break; > ... > }. > > The final stage of the size setting is propagated it from the proxy to > the device on virtqueue enabling: > > virtio_cpi_common_write() > { > ... > case VIRTIO_PCI_COMMON_Q_ENABLE: > virtio_queue_set_num(vdev, vdev->queue_sel, > proxy->vqs[vdev->queue_sel].num); > virtio_queue_set_rings(vdev, vdev->queue_sel, > ((uint64_t)proxy->vqs[vdev->queue_sel].desc[1]) << 32 | > proxy->vqs[vdev->queue_sel].desc[0], > ((uint64_t)proxy->vqs[vdev->queue_sel].avail[1]) << 32 | > proxy->vqs[vdev->queue_sel].avail[0], > ((uint64_t)proxy->vqs[vdev->queue_sel].used[1]) << 32 | > proxy->vqs[vdev->queue_sel].used[0]); > proxy->vqs[vdev->queue_sel].enabled = 1; > break; > ... > }. > > So we have the following workflow: > suppose the device has virtqueue size = 256 and seabios MAX_QUEUE_NUM > = 128. > In that case the seabios works like: > > 1. if vp_modern read the size (256) > 2. 256 > 128 > 3. write virtqueue size = 128 > 4. re-read virtqueue size = 256 !!! > 5. fail because of the check > if (num > MAX_QUEUE_NUM) { > dprintf(1, "ERROR: queue size %d > %d\n", num, MAX_QUEUE_NUM); > goto fail; > } > > To fix the issue, we need to read and write the virtqueue size from > the same place. > Should we do with the proxy? > Is there any reason to read from the device and write to the proxy? > > Furthermore, the size setting has a few flaws: > > 1. The size being set should be a power of 2 > 2. The size being set should be less or equal to the virtqueue size > (and be greater that 2?) > > Denis >>> you mean to put the code like this into virtio_pci_realize() inside >>> QEMU? >>> >>> If no, can you pls clarify which component should be touched. >>> >>> Den >> I mean: >> - add an API to change the default queue size >> - add a validate features callback, in there check and for modern >> flag set in features increase the queue size >> >> maybe all this is too much work, we could block this >> for transitional devices, but your patch does not do it, >> you need to check that legacy is enabled not that modern >> is not disabled. >> >> >> >
On 05.12.2019 10:59, Denis Plotnikov wrote: > Ping! > > On 25.11.2019 12:16, Denis Plotnikov wrote: >> >> >> On 06.11.2019 15:03, Michael S. Tsirkin wrote: >>> On Thu, Oct 24, 2019 at 11:34:34AM +0000, Denis Lunev wrote: >>>> On 10/24/19 12:28 AM, Michael S. Tsirkin wrote: >>>>> On Fri, Oct 18, 2019 at 02:55:47PM +0300, Denis Plotnikov wrote: >>>>>> From: "Denis V. Lunev" <den@openvz.org> >>>>>> >>>>>> Linux guests submit IO requests no longer than PAGE_SIZE * max_seg >>>>>> field reported by SCSI controler. Thus typical sequential read with >>>>>> 1 MB size results in the following pattern of the IO from the guest: >>>>>> 8,16 1 15754 2.766095122 2071 D R 2095104 + 1008 >>>>>> [dd] >>>>>> 8,16 1 15755 2.766108785 2071 D R 2096112 + 1008 >>>>>> [dd] >>>>>> 8,16 1 15756 2.766113486 2071 D R 2097120 + 32 [dd] >>>>>> 8,16 1 15757 2.767668961 0 C R 2095104 + 1008 [0] >>>>>> 8,16 1 15758 2.768534315 0 C R 2096112 + 1008 [0] >>>>>> 8,16 1 15759 2.768539782 0 C R 2097120 + 32 [0] >>>>>> The IO was generated by >>>>>> dd if=/dev/sda of=/dev/null bs=1024 iflag=direct >>>>>> >>>>>> This effectively means that on rotational disks we will observe 3 >>>>>> IOPS >>>>>> for each 2 MBs processed. This definitely negatively affects both >>>>>> guest and host IO performance. >>>>>> >>>>>> The cure is relatively simple - we should report lengthy >>>>>> scatter-gather >>>>>> ability of the SCSI controller. Fortunately the situation here is >>>>>> very >>>>>> good. VirtIO transport layer can accomodate 1024 items in one >>>>>> request >>>>>> while we are using only 128. This situation is present since almost >>>>>> very beginning. 2 items are dedicated for request metadata thus we >>>>>> should publish VIRTQUEUE_MAX_SIZE - 2 as max_seg. >>>>>> >>>>>> The following pattern is observed after the patch: >>>>>> 8,16 1 9921 2.662721340 2063 D R 2095104 + 1024 >>>>>> [dd] >>>>>> 8,16 1 9922 2.662737585 2063 D R 2096128 + 1024 >>>>>> [dd] >>>>>> 8,16 1 9923 2.665188167 0 C R 2095104 + 1024 [0] >>>>>> 8,16 1 9924 2.665198777 0 C R 2096128 + 1024 [0] >>>>>> which is much better. >>>>>> >>>>>> The dark side of this patch is that we are tweaking guest visible >>>>>> parameter, though this should be relatively safe as above transport >>>>>> layer support is present in QEMU/host Linux for a very long time. >>>>>> The patch adds configurable property for VirtIO SCSI with a new >>>>>> default >>>>>> and hardcode option for VirtBlock which does not provide good >>>>>> configurable framework. >>>>>> >>>>>> Unfortunately the commit can not be applied as is. For the real >>>>>> cure we >>>>>> need guest to be fixed to accomodate that queue length, which is >>>>>> done >>>>>> only in the latest 4.14 kernel. Thus we are going to expose the >>>>>> property >>>>>> and tweak it on machine type level. >>>>>> >>>>>> The problem with the old kernels is that they have >>>>>> max_segments <= virtqueue_size restriction which cause the guest >>>>>> crashing in the case of violation. >>>>> This isn't just in the guests: virtio spec also seems to imply this, >>>>> or at least be vague on this point. >>>>> >>>>> So I think it'll need a feature bit. >>>>> Doing that in a safe way will also allow being compatible with old >>>>> guests. >>>>> >>>>> The only downside is it's a bit more work as we need to >>>>> spec this out and add guest support. >>>>> >>>>>> To fix the case described above in the old kernels we can increase >>>>>> virtqueue_size to 256 and max_segments to 254. The pitfall here is >>>>>> that seabios allows the virtqueue_size-s < 128, however, the seabios >>>>>> patch extending that value to 256 is pending. >>>>> And the fix here is just to limit large vq size to virtio 1.0. >>>>> In that mode it's fine I think: >>>>> >>>>> >>>>> /* check if the queue is available */ >>>>> if (vp->use_modern) { >>>>> num = vp_read(&vp->common, virtio_pci_common_cfg, >>>>> queue_size); >>>>> if (num > MAX_QUEUE_NUM) { >>>>> vp_write(&vp->common, virtio_pci_common_cfg, queue_size, >>>>> MAX_QUEUE_NUM); >>>>> num = vp_read(&vp->common, virtio_pci_common_cfg, >>>>> queue_size); >>>>> } >>>>> } else { >>>>> num = vp_read(&vp->legacy, virtio_pci_legacy, queue_num); >>>>> } >> The same seabios snippet, but more detailed: >> >> vp_find_vq() >> { >> ... >> /* check if the queue is available */ >> if (vp->use_modern) { >> num = vp_read(&vp->common, virtio_pci_common_cfg, queue_size); >> if (num > MAX_QUEUE_NUM) { >> vp_write(&vp->common, virtio_pci_common_cfg, queue_size, >> MAX_QUEUE_NUM); >> num = vp_read(&vp->common, virtio_pci_common_cfg, >> queue_size); >> } >> } else { >> num = vp_read(&vp->legacy, virtio_pci_legacy, queue_num); >> } >> if (!num) { >> dprintf(1, "ERROR: queue size is 0\n"); >> goto fail; >> } >> if (num > MAX_QUEUE_NUM) { >> dprintf(1, "ERROR: queue size %d > %d\n", num, MAX_QUEUE_NUM); >> goto fail; >> } >> ... >> } >> >> It turned out that the problem is here, but not because of the >> seabios code. >> The virtqueue size is written and then incorrect value is re-read. >> Thanks to Roman Kagan (rkagan@virtuozzo.com) for investigating the >> root cause of the problem. >> >> As the code states, for the modern devices, seabios reads the queue >> size and if it's >> greater than seabios can support, reduce the queue size to the max >> seabios supported value. >> >> This doesn't work. >> >> The reason is that the size is read from the virtio device, >> >> virtio_pci_common_read() >> { >> ... >> case VIRTIO_PCI_COMMON_Q_SIZE: >> val = virtio_queue_get_num(vdev, vdev->queue_sel); >> break; >> ... >> } >> >> but is written to the proxy >> >> virtio_pci_common_write() >> { >> ... >> case VIRTIO_PCI_COMMON_Q_SIZE: >> proxy->vqs[vdev->queue_sel].num = val; >> break; >> ... >> }. >> >> The final stage of the size setting is propagated it from the proxy >> to the device on virtqueue enabling: >> >> virtio_cpi_common_write() >> { >> ... >> case VIRTIO_PCI_COMMON_Q_ENABLE: >> virtio_queue_set_num(vdev, vdev->queue_sel, >> proxy->vqs[vdev->queue_sel].num); >> virtio_queue_set_rings(vdev, vdev->queue_sel, >> ((uint64_t)proxy->vqs[vdev->queue_sel].desc[1]) << 32 | >> proxy->vqs[vdev->queue_sel].desc[0], >> ((uint64_t)proxy->vqs[vdev->queue_sel].avail[1]) << 32 | >> proxy->vqs[vdev->queue_sel].avail[0], >> ((uint64_t)proxy->vqs[vdev->queue_sel].used[1]) << 32 | >> proxy->vqs[vdev->queue_sel].used[0]); >> proxy->vqs[vdev->queue_sel].enabled = 1; >> break; >> ... >> }. >> >> So we have the following workflow: >> suppose the device has virtqueue size = 256 and seabios MAX_QUEUE_NUM >> = 128. >> In that case the seabios works like: >> >> 1. if vp_modern read the size (256) >> 2. 256 > 128 >> 3. write virtqueue size = 128 >> 4. re-read virtqueue size = 256 !!! >> 5. fail because of the check >> if (num > MAX_QUEUE_NUM) { >> dprintf(1, "ERROR: queue size %d > %d\n", num, MAX_QUEUE_NUM); >> goto fail; >> } >> >> To fix the issue, we need to read and write the virtqueue size from >> the same place. >> Should we do with the proxy? >> Is there any reason to read from the device and write to the proxy? >> >> Furthermore, the size setting has a few flaws: >> >> 1. The size being set should be a power of 2 >> 2. The size being set should be less or equal to the virtqueue size >> (and be greater that 2?) >> >> Denis >>>> you mean to put the code like this into virtio_pci_realize() inside >>>> QEMU? >>>> >>>> If no, can you pls clarify which component should be touched. >>>> >>>> Den >>> I mean: >>> - add an API to change the default queue size >>> - add a validate features callback, in there check and for modern >>> flag set in features increase the queue size >>> >>> maybe all this is too much work, we could block this >>> for transitional devices, but your patch does not do it, >>> you need to check that legacy is enabled not that modern >>> is not disabled. >>> >>> >>> >> >
On Mon, Nov 25, 2019 at 09:16:10AM +0000, Denis Plotnikov wrote: > > > On 06.11.2019 15:03, Michael S. Tsirkin wrote: > > On Thu, Oct 24, 2019 at 11:34:34AM +0000, Denis Lunev wrote: > >> On 10/24/19 12:28 AM, Michael S. Tsirkin wrote: > >>> On Fri, Oct 18, 2019 at 02:55:47PM +0300, Denis Plotnikov wrote: > >>>> From: "Denis V. Lunev" <den@openvz.org> > >>>> > >>>> Linux guests submit IO requests no longer than PAGE_SIZE * max_seg > >>>> field reported by SCSI controler. Thus typical sequential read with > >>>> 1 MB size results in the following pattern of the IO from the guest: > >>>> 8,16 1 15754 2.766095122 2071 D R 2095104 + 1008 [dd] > >>>> 8,16 1 15755 2.766108785 2071 D R 2096112 + 1008 [dd] > >>>> 8,16 1 15756 2.766113486 2071 D R 2097120 + 32 [dd] > >>>> 8,16 1 15757 2.767668961 0 C R 2095104 + 1008 [0] > >>>> 8,16 1 15758 2.768534315 0 C R 2096112 + 1008 [0] > >>>> 8,16 1 15759 2.768539782 0 C R 2097120 + 32 [0] > >>>> The IO was generated by > >>>> dd if=/dev/sda of=/dev/null bs=1024 iflag=direct > >>>> > >>>> This effectively means that on rotational disks we will observe 3 IOPS > >>>> for each 2 MBs processed. This definitely negatively affects both > >>>> guest and host IO performance. > >>>> > >>>> The cure is relatively simple - we should report lengthy scatter-gather > >>>> ability of the SCSI controller. Fortunately the situation here is very > >>>> good. VirtIO transport layer can accomodate 1024 items in one request > >>>> while we are using only 128. This situation is present since almost > >>>> very beginning. 2 items are dedicated for request metadata thus we > >>>> should publish VIRTQUEUE_MAX_SIZE - 2 as max_seg. > >>>> > >>>> The following pattern is observed after the patch: > >>>> 8,16 1 9921 2.662721340 2063 D R 2095104 + 1024 [dd] > >>>> 8,16 1 9922 2.662737585 2063 D R 2096128 + 1024 [dd] > >>>> 8,16 1 9923 2.665188167 0 C R 2095104 + 1024 [0] > >>>> 8,16 1 9924 2.665198777 0 C R 2096128 + 1024 [0] > >>>> which is much better. > >>>> > >>>> The dark side of this patch is that we are tweaking guest visible > >>>> parameter, though this should be relatively safe as above transport > >>>> layer support is present in QEMU/host Linux for a very long time. > >>>> The patch adds configurable property for VirtIO SCSI with a new default > >>>> and hardcode option for VirtBlock which does not provide good > >>>> configurable framework. > >>>> > >>>> Unfortunately the commit can not be applied as is. For the real cure we > >>>> need guest to be fixed to accomodate that queue length, which is done > >>>> only in the latest 4.14 kernel. Thus we are going to expose the property > >>>> and tweak it on machine type level. > >>>> > >>>> The problem with the old kernels is that they have > >>>> max_segments <= virtqueue_size restriction which cause the guest > >>>> crashing in the case of violation. > >>> This isn't just in the guests: virtio spec also seems to imply this, > >>> or at least be vague on this point. > >>> > >>> So I think it'll need a feature bit. > >>> Doing that in a safe way will also allow being compatible with old guests. > >>> > >>> The only downside is it's a bit more work as we need to > >>> spec this out and add guest support. > >>> > >>>> To fix the case described above in the old kernels we can increase > >>>> virtqueue_size to 256 and max_segments to 254. The pitfall here is > >>>> that seabios allows the virtqueue_size-s < 128, however, the seabios > >>>> patch extending that value to 256 is pending. > >>> And the fix here is just to limit large vq size to virtio 1.0. > >>> In that mode it's fine I think: > >>> > >>> > >>> /* check if the queue is available */ > >>> if (vp->use_modern) { > >>> num = vp_read(&vp->common, virtio_pci_common_cfg, queue_size); > >>> if (num > MAX_QUEUE_NUM) { > >>> vp_write(&vp->common, virtio_pci_common_cfg, queue_size, > >>> MAX_QUEUE_NUM); > >>> num = vp_read(&vp->common, virtio_pci_common_cfg, queue_size); > >>> } > >>> } else { > >>> num = vp_read(&vp->legacy, virtio_pci_legacy, queue_num); > >>> } > The same seabios snippet, but more detailed: > > vp_find_vq() > { > ... > /* check if the queue is available */ > if (vp->use_modern) { > num = vp_read(&vp->common, virtio_pci_common_cfg, queue_size); > if (num > MAX_QUEUE_NUM) { > vp_write(&vp->common, virtio_pci_common_cfg, queue_size, > MAX_QUEUE_NUM); > num = vp_read(&vp->common, virtio_pci_common_cfg, queue_size); So how about we drop this last line in bios? Will fix things for existing hypervisors. spec does not say guests need to re-read it. > } > } else { > num = vp_read(&vp->legacy, virtio_pci_legacy, queue_num); > } > if (!num) { > dprintf(1, "ERROR: queue size is 0\n"); > goto fail; > } > if (num > MAX_QUEUE_NUM) { > dprintf(1, "ERROR: queue size %d > %d\n", num, MAX_QUEUE_NUM); > goto fail; > } > ... > } > > It turned out that the problem is here, but not because of the seabios code. > The virtqueue size is written and then incorrect value is re-read. > Thanks to Roman Kagan (rkagan@virtuozzo.com) for investigating the root > cause of the problem. > > As the code states, for the modern devices, seabios reads the queue size > and if it's > greater than seabios can support, reduce the queue size to the max > seabios supported value. > > This doesn't work. > > The reason is that the size is read from the virtio device, > > virtio_pci_common_read() > { > ... > case VIRTIO_PCI_COMMON_Q_SIZE: > val = virtio_queue_get_num(vdev, vdev->queue_sel); > break; > ... > } > > but is written to the proxy > > virtio_pci_common_write() > { > ... > case VIRTIO_PCI_COMMON_Q_SIZE: > proxy->vqs[vdev->queue_sel].num = val; > break; > ... > }. Yea that's a bug. Here's a hacky way to fix it. But I think really we should just get rid of the two copies down the road. diff --git a/hw/virtio/virtio-pci.c b/hw/virtio/virtio-pci.c index c6b47a9c73..e5c759e19e 100644 --- a/hw/virtio/virtio-pci.c +++ b/hw/virtio/virtio-pci.c @@ -1256,6 +1256,8 @@ static void virtio_pci_common_write(void *opaque, hwaddr addr, break; case VIRTIO_PCI_COMMON_Q_SIZE: proxy->vqs[vdev->queue_sel].num = val; + virtio_queue_set_num(vdev, vdev->queue_sel, + proxy->vqs[vdev->queue_sel].num); break; case VIRTIO_PCI_COMMON_Q_MSIX: msix_vector_unuse(&proxy->pci_dev, > The final stage of the size setting is propagated it from the proxy to > the device on virtqueue enabling: > > virtio_cpi_common_write() > { > ... > case VIRTIO_PCI_COMMON_Q_ENABLE: > virtio_queue_set_num(vdev, vdev->queue_sel, > proxy->vqs[vdev->queue_sel].num); > virtio_queue_set_rings(vdev, vdev->queue_sel, > ((uint64_t)proxy->vqs[vdev->queue_sel].desc[1]) << 32 | > proxy->vqs[vdev->queue_sel].desc[0], > ((uint64_t)proxy->vqs[vdev->queue_sel].avail[1]) << 32 | > proxy->vqs[vdev->queue_sel].avail[0], > ((uint64_t)proxy->vqs[vdev->queue_sel].used[1]) << 32 | > proxy->vqs[vdev->queue_sel].used[0]); > proxy->vqs[vdev->queue_sel].enabled = 1; > break; > ... > }. > > So we have the following workflow: > suppose the device has virtqueue size = 256 and seabios MAX_QUEUE_NUM = 128. > In that case the seabios works like: > > 1. if vp_modern read the size (256) > 2. 256 > 128 > 3. write virtqueue size = 128 > 4. re-read virtqueue size = 256 !!! bios probably should not re-read size, it's a waste of cpu cycles anyway. > 5. fail because of the check > if (num > MAX_QUEUE_NUM) { > dprintf(1, "ERROR: queue size %d > %d\n", num, MAX_QUEUE_NUM); > goto fail; > } > > To fix the issue, we need to read and write the virtqueue size from the > same place. > Should we do with the proxy? > Is there any reason to read from the device and write to the proxy? > > Furthermore, the size setting has a few flaws: > > 1. The size being set should be a power of 2 > 2. The size being set should be less or equal to the virtqueue size (and > be greater that 2?) I think 1 is checked in virtio_queue_set_num. I guess we should check 2 as well? > > Denis > >> you mean to put the code like this into virtio_pci_realize() inside QEMU? > >> > >> If no, can you pls clarify which component should be touched. > >> > >> Den > > I mean: > > - add an API to change the default queue size > > - add a validate features callback, in there check and for modern > > flag set in features increase the queue size > > > > maybe all this is too much work, we could block this > > for transitional devices, but your patch does not do it, > > you need to check that legacy is enabled not that modern > > is not disabled. > > > > > > >
diff --git a/hw/block/virtio-blk.c b/hw/block/virtio-blk.c index 06e57a4d39..b2eaeeaf67 100644 --- a/hw/block/virtio-blk.c +++ b/hw/block/virtio-blk.c @@ -903,7 +903,7 @@ static void virtio_blk_update_config(VirtIODevice *vdev, uint8_t *config) blk_get_geometry(s->blk, &capacity); memset(&blkcfg, 0, sizeof(blkcfg)); virtio_stq_p(vdev, &blkcfg.capacity, capacity); - virtio_stl_p(vdev, &blkcfg.seg_max, 128 - 2); + virtio_stl_p(vdev, &blkcfg.seg_max, s->conf.max_segments); virtio_stw_p(vdev, &blkcfg.geometry.cylinders, conf->cyls); virtio_stl_p(vdev, &blkcfg.blk_size, blk_size); virtio_stw_p(vdev, &blkcfg.min_io_size, conf->min_io_size / blk_size); @@ -1240,6 +1240,7 @@ static Property virtio_blk_properties[] = { conf.max_discard_sectors, BDRV_REQUEST_MAX_SECTORS), DEFINE_PROP_UINT32("max-write-zeroes-sectors", VirtIOBlock, conf.max_write_zeroes_sectors, BDRV_REQUEST_MAX_SECTORS), + DEFINE_PROP_UINT32("max_segments", VirtIOBlock, conf.max_segments, 126), DEFINE_PROP_END_OF_LIST(), }; diff --git a/hw/scsi/vhost-scsi.c b/hw/scsi/vhost-scsi.c index 61e2e57da9..fa3b377807 100644 --- a/hw/scsi/vhost-scsi.c +++ b/hw/scsi/vhost-scsi.c @@ -242,6 +242,8 @@ static Property vhost_scsi_properties[] = { DEFINE_PROP_BIT64("t10_pi", VHostSCSICommon, host_features, VIRTIO_SCSI_F_T10_PI, false), + DEFINE_PROP_UINT32("max_segments", VirtIOSCSICommon, conf.max_segments, + 126), DEFINE_PROP_END_OF_LIST(), }; diff --git a/hw/scsi/virtio-scsi.c b/hw/scsi/virtio-scsi.c index 839f120256..8b070ddeed 100644 --- a/hw/scsi/virtio-scsi.c +++ b/hw/scsi/virtio-scsi.c @@ -650,7 +650,7 @@ static void virtio_scsi_get_config(VirtIODevice *vdev, VirtIOSCSICommon *s = VIRTIO_SCSI_COMMON(vdev); virtio_stl_p(vdev, &scsiconf->num_queues, s->conf.num_queues); - virtio_stl_p(vdev, &scsiconf->seg_max, 128 - 2); + virtio_stl_p(vdev, &scsiconf->seg_max, s->conf.max_segments); virtio_stl_p(vdev, &scsiconf->max_sectors, s->conf.max_sectors); virtio_stl_p(vdev, &scsiconf->cmd_per_lun, s->conf.cmd_per_lun); virtio_stl_p(vdev, &scsiconf->event_info_size, sizeof(VirtIOSCSIEvent)); @@ -948,6 +948,8 @@ static Property virtio_scsi_properties[] = { VIRTIO_SCSI_F_CHANGE, true), DEFINE_PROP_LINK("iothread", VirtIOSCSI, parent_obj.conf.iothread, TYPE_IOTHREAD, IOThread *), + DEFINE_PROP_UINT32("max_segments", VirtIOSCSI, parent_obj.conf.max_segments, + 126), DEFINE_PROP_END_OF_LIST(), }; diff --git a/include/hw/virtio/virtio-blk.h b/include/hw/virtio/virtio-blk.h index cddcfbebe9..22da23a4a3 100644 --- a/include/hw/virtio/virtio-blk.h +++ b/include/hw/virtio/virtio-blk.h @@ -40,6 +40,7 @@ struct VirtIOBlkConf uint16_t queue_size; uint32_t max_discard_sectors; uint32_t max_write_zeroes_sectors; + uint32_t max_segments; }; struct VirtIOBlockDataPlane; diff --git a/include/hw/virtio/virtio-scsi.h b/include/hw/virtio/virtio-scsi.h index 4c0bcdb788..1e5805eec4 100644 --- a/include/hw/virtio/virtio-scsi.h +++ b/include/hw/virtio/virtio-scsi.h @@ -49,6 +49,7 @@ struct VirtIOSCSIConf { uint32_t num_queues; uint32_t virtqueue_size; uint32_t max_sectors; + uint32_t max_segments; uint32_t cmd_per_lun; #ifdef CONFIG_VHOST_SCSI char *vhostfd;