| Message ID | 20210926145518.64164-2-mgurtovoy@nvidia.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [1/2] virtio: introduce virtio_dev_to_node helper | expand |
On Sun, Sep 26, 2021 at 05:55:18PM +0300, Max Gurtovoy wrote: > To optimize performance, set the affinity of the block device tagset > according to the virtio device affinity. > > Signed-off-by: Max Gurtovoy <mgurtovoy@nvidia.com> > --- > drivers/block/virtio_blk.c | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c > index 9b3bd083b411..1c68c3e0ebf9 100644 > --- a/drivers/block/virtio_blk.c > +++ b/drivers/block/virtio_blk.c > @@ -774,7 +774,7 @@ static int virtblk_probe(struct virtio_device *vdev) > memset(&vblk->tag_set, 0, sizeof(vblk->tag_set)); > vblk->tag_set.ops = &virtio_mq_ops; > vblk->tag_set.queue_depth = queue_depth; > - vblk->tag_set.numa_node = NUMA_NO_NODE; > + vblk->tag_set.numa_node = virtio_dev_to_node(vdev); > vblk->tag_set.flags = BLK_MQ_F_SHOULD_MERGE; > vblk->tag_set.cmd_size = > sizeof(struct virtblk_req) + I implemented NUMA affinity in the past and could not demonstrate a performance improvement: https://lists.linuxfoundation.org/pipermail/virtualization/2020-June/048248.html The pathological case is when a guest with vNUMA has the virtio-blk-pci device on the "wrong" host NUMA node. Then memory accesses should cross NUMA nodes. Still, it didn't seem to matter. Please share your benchmark results. If you haven't collected data yet you could even combine our patches to see if it helps. Thanks! Stefan
On Sun, Sep 26, 2021 at 05:55:18PM +0300, Max Gurtovoy wrote: > To optimize performance, set the affinity of the block device tagset > according to the virtio device affinity. > > Signed-off-by: Max Gurtovoy <mgurtovoy@nvidia.com> > --- > drivers/block/virtio_blk.c | 2 +- > 1 file changed, 1 insertion(+), 1 deletion(-) > > diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c > index 9b3bd083b411..1c68c3e0ebf9 100644 > --- a/drivers/block/virtio_blk.c > +++ b/drivers/block/virtio_blk.c > @@ -774,7 +774,7 @@ static int virtblk_probe(struct virtio_device *vdev) > memset(&vblk->tag_set, 0, sizeof(vblk->tag_set)); > vblk->tag_set.ops = &virtio_mq_ops; > vblk->tag_set.queue_depth = queue_depth; > - vblk->tag_set.numa_node = NUMA_NO_NODE; > + vblk->tag_set.numa_node = virtio_dev_to_node(vdev); I afraid that by doing it, you will increase chances to see OOM, because in NUMA_NO_NODE, MM will try allocate memory in whole system, while in the latter mode only on specific NUMA which can be depleted. Thanks > vblk->tag_set.flags = BLK_MQ_F_SHOULD_MERGE; > vblk->tag_set.cmd_size = > sizeof(struct virtblk_req) + > -- > 2.18.1 >
On 9/27/2021 2:34 PM, Leon Romanovsky wrote: > On Sun, Sep 26, 2021 at 05:55:18PM +0300, Max Gurtovoy wrote: >> To optimize performance, set the affinity of the block device tagset >> according to the virtio device affinity. >> >> Signed-off-by: Max Gurtovoy <mgurtovoy@nvidia.com> >> --- >> drivers/block/virtio_blk.c | 2 +- >> 1 file changed, 1 insertion(+), 1 deletion(-) >> >> diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c >> index 9b3bd083b411..1c68c3e0ebf9 100644 >> --- a/drivers/block/virtio_blk.c >> +++ b/drivers/block/virtio_blk.c >> @@ -774,7 +774,7 @@ static int virtblk_probe(struct virtio_device *vdev) >> memset(&vblk->tag_set, 0, sizeof(vblk->tag_set)); >> vblk->tag_set.ops = &virtio_mq_ops; >> vblk->tag_set.queue_depth = queue_depth; >> - vblk->tag_set.numa_node = NUMA_NO_NODE; >> + vblk->tag_set.numa_node = virtio_dev_to_node(vdev); > I afraid that by doing it, you will increase chances to see OOM, because > in NUMA_NO_NODE, MM will try allocate memory in whole system, while in > the latter mode only on specific NUMA which can be depleted. This is a common methodology we use in the block layer and in NVMe subsystem and we don't afraid of the OOM issue you raised. This is not new and I guess that the kernel MM will (or should) be handling the fallback you raised. Anyway, if we're doing this in NVMe I don't see a reason to afraid doing it in virtio-blk. Also, I've send a patch that decrease the size of the memory consumption for virtio-blk few weeks ago so I guess we'll be just fine. > > Thanks > >> vblk->tag_set.flags = BLK_MQ_F_SHOULD_MERGE; >> vblk->tag_set.cmd_size = >> sizeof(struct virtblk_req) + >> -- >> 2.18.1 >>
On 9/27/2021 11:09 AM, Stefan Hajnoczi wrote: > On Sun, Sep 26, 2021 at 05:55:18PM +0300, Max Gurtovoy wrote: >> To optimize performance, set the affinity of the block device tagset >> according to the virtio device affinity. >> >> Signed-off-by: Max Gurtovoy <mgurtovoy@nvidia.com> >> --- >> drivers/block/virtio_blk.c | 2 +- >> 1 file changed, 1 insertion(+), 1 deletion(-) >> >> diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c >> index 9b3bd083b411..1c68c3e0ebf9 100644 >> --- a/drivers/block/virtio_blk.c >> +++ b/drivers/block/virtio_blk.c >> @@ -774,7 +774,7 @@ static int virtblk_probe(struct virtio_device *vdev) >> memset(&vblk->tag_set, 0, sizeof(vblk->tag_set)); >> vblk->tag_set.ops = &virtio_mq_ops; >> vblk->tag_set.queue_depth = queue_depth; >> - vblk->tag_set.numa_node = NUMA_NO_NODE; >> + vblk->tag_set.numa_node = virtio_dev_to_node(vdev); >> vblk->tag_set.flags = BLK_MQ_F_SHOULD_MERGE; >> vblk->tag_set.cmd_size = >> sizeof(struct virtblk_req) + > I implemented NUMA affinity in the past and could not demonstrate a > performance improvement: > https://lists.linuxfoundation.org/pipermail/virtualization/2020-June/048248.html > > The pathological case is when a guest with vNUMA has the virtio-blk-pci > device on the "wrong" host NUMA node. Then memory accesses should cross > NUMA nodes. Still, it didn't seem to matter. I think the reason you didn't see any improvement is since you didn't use the right device for the node query. See my patch 1/2. I can try integrating these patches in my series and fix it. BTW, we might not see a big improvement because of other bottlenecks but this is known perf optimization we use often in block storage drivers. > > Please share your benchmark results. If you haven't collected data yet > you could even combine our patches to see if it helps. Thanks! > > Stefan
On Mon, Sep 27, 2021 at 08:25:09PM +0300, Max Gurtovoy wrote: > > On 9/27/2021 2:34 PM, Leon Romanovsky wrote: > > On Sun, Sep 26, 2021 at 05:55:18PM +0300, Max Gurtovoy wrote: > > > To optimize performance, set the affinity of the block device tagset > > > according to the virtio device affinity. > > > > > > Signed-off-by: Max Gurtovoy <mgurtovoy@nvidia.com> > > > --- > > > drivers/block/virtio_blk.c | 2 +- > > > 1 file changed, 1 insertion(+), 1 deletion(-) > > > > > > diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c > > > index 9b3bd083b411..1c68c3e0ebf9 100644 > > > --- a/drivers/block/virtio_blk.c > > > +++ b/drivers/block/virtio_blk.c > > > @@ -774,7 +774,7 @@ static int virtblk_probe(struct virtio_device *vdev) > > > memset(&vblk->tag_set, 0, sizeof(vblk->tag_set)); > > > vblk->tag_set.ops = &virtio_mq_ops; > > > vblk->tag_set.queue_depth = queue_depth; > > > - vblk->tag_set.numa_node = NUMA_NO_NODE; > > > + vblk->tag_set.numa_node = virtio_dev_to_node(vdev); > > I afraid that by doing it, you will increase chances to see OOM, because > > in NUMA_NO_NODE, MM will try allocate memory in whole system, while in > > the latter mode only on specific NUMA which can be depleted. > > This is a common methodology we use in the block layer and in NVMe subsystem > and we don't afraid of the OOM issue you raised. There are many reasons for that, but we are talking about virtio here and not about NVMe. > > This is not new and I guess that the kernel MM will (or should) be handling > the fallback you raised. I afraid that it is not. Can you point me to the place where such fallback is implemented? > > Anyway, if we're doing this in NVMe I don't see a reason to afraid doing it > in virtio-blk. Still, it is nice to have some empirical data to support this copy/paste. There are too many myths related to optimizations, so finally it will be good to get some supportive data. Thanks
On Mon, Sep 27, 2021 at 08:39:30PM +0300, Max Gurtovoy wrote: > > On 9/27/2021 11:09 AM, Stefan Hajnoczi wrote: > > On Sun, Sep 26, 2021 at 05:55:18PM +0300, Max Gurtovoy wrote: > > > To optimize performance, set the affinity of the block device tagset > > > according to the virtio device affinity. > > > > > > Signed-off-by: Max Gurtovoy <mgurtovoy@nvidia.com> > > > --- > > > drivers/block/virtio_blk.c | 2 +- > > > 1 file changed, 1 insertion(+), 1 deletion(-) > > > > > > diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c > > > index 9b3bd083b411..1c68c3e0ebf9 100644 > > > --- a/drivers/block/virtio_blk.c > > > +++ b/drivers/block/virtio_blk.c > > > @@ -774,7 +774,7 @@ static int virtblk_probe(struct virtio_device *vdev) > > > memset(&vblk->tag_set, 0, sizeof(vblk->tag_set)); > > > vblk->tag_set.ops = &virtio_mq_ops; > > > vblk->tag_set.queue_depth = queue_depth; > > > - vblk->tag_set.numa_node = NUMA_NO_NODE; > > > + vblk->tag_set.numa_node = virtio_dev_to_node(vdev); > > > vblk->tag_set.flags = BLK_MQ_F_SHOULD_MERGE; > > > vblk->tag_set.cmd_size = > > > sizeof(struct virtblk_req) + > > I implemented NUMA affinity in the past and could not demonstrate a > > performance improvement: > > https://lists.linuxfoundation.org/pipermail/virtualization/2020-June/048248.html > > > > The pathological case is when a guest with vNUMA has the virtio-blk-pci > > device on the "wrong" host NUMA node. Then memory accesses should cross > > NUMA nodes. Still, it didn't seem to matter. > > I think the reason you didn't see any improvement is since you didn't use > the right device for the node query. See my patch 1/2. That doesn't seem to be the case. Please see drivers/base/core.c:device_add(): /* use parent numa_node */ if (parent && (dev_to_node(dev) == NUMA_NO_NODE)) set_dev_node(dev, dev_to_node(parent)); IMO it's cleaner to use dev_to_node(&vdev->dev) than to directly access the parent. Have I missed something? > > I can try integrating these patches in my series and fix it. > > BTW, we might not see a big improvement because of other bottlenecks but > this is known perf optimization we use often in block storage drivers. Let's see benchmark results. Otherwise this is just dead code that adds complexity. Stefan
On 9/27/2021 9:23 PM, Leon Romanovsky wrote: > On Mon, Sep 27, 2021 at 08:25:09PM +0300, Max Gurtovoy wrote: >> On 9/27/2021 2:34 PM, Leon Romanovsky wrote: >>> On Sun, Sep 26, 2021 at 05:55:18PM +0300, Max Gurtovoy wrote: >>>> To optimize performance, set the affinity of the block device tagset >>>> according to the virtio device affinity. >>>> >>>> Signed-off-by: Max Gurtovoy <mgurtovoy@nvidia.com> >>>> --- >>>> drivers/block/virtio_blk.c | 2 +- >>>> 1 file changed, 1 insertion(+), 1 deletion(-) >>>> >>>> diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c >>>> index 9b3bd083b411..1c68c3e0ebf9 100644 >>>> --- a/drivers/block/virtio_blk.c >>>> +++ b/drivers/block/virtio_blk.c >>>> @@ -774,7 +774,7 @@ static int virtblk_probe(struct virtio_device *vdev) >>>> memset(&vblk->tag_set, 0, sizeof(vblk->tag_set)); >>>> vblk->tag_set.ops = &virtio_mq_ops; >>>> vblk->tag_set.queue_depth = queue_depth; >>>> - vblk->tag_set.numa_node = NUMA_NO_NODE; >>>> + vblk->tag_set.numa_node = virtio_dev_to_node(vdev); >>> I afraid that by doing it, you will increase chances to see OOM, because >>> in NUMA_NO_NODE, MM will try allocate memory in whole system, while in >>> the latter mode only on specific NUMA which can be depleted. >> This is a common methodology we use in the block layer and in NVMe subsystem >> and we don't afraid of the OOM issue you raised. > There are many reasons for that, but we are talking about virtio here > and not about NVMe. Ok. what reasons ? > >> This is not new and I guess that the kernel MM will (or should) be handling >> the fallback you raised. > I afraid that it is not. Can you point me to the place where such > fallback is implemented? > >> Anyway, if we're doing this in NVMe I don't see a reason to afraid doing it >> in virtio-blk. > Still, it is nice to have some empirical data to support this copy/paste. I'm not sure what you meant when you say that, but this was just an example as everyone else saw. No copy/paste. Anyhow, taking good ideas from other Linux subsystems is common and I don't see a problem doing so. I'll let the storage experts to comment on that and the maintainers to decide whether they would like to take this optimization among other optimizations I've sent to this subsystem. I'll send a v2 in few days. Thanks. > > There are too many myths related to optimizations, so finally it will be > good to get some supportive data. > > Thanks
On Tue, Sep 28, 2021 at 06:59:15PM +0300, Max Gurtovoy wrote: > > On 9/27/2021 9:23 PM, Leon Romanovsky wrote: > > On Mon, Sep 27, 2021 at 08:25:09PM +0300, Max Gurtovoy wrote: > > > On 9/27/2021 2:34 PM, Leon Romanovsky wrote: > > > > On Sun, Sep 26, 2021 at 05:55:18PM +0300, Max Gurtovoy wrote: > > > > > To optimize performance, set the affinity of the block device tagset > > > > > according to the virtio device affinity. > > > > > > > > > > Signed-off-by: Max Gurtovoy <mgurtovoy@nvidia.com> > > > > > --- > > > > > drivers/block/virtio_blk.c | 2 +- > > > > > 1 file changed, 1 insertion(+), 1 deletion(-) > > > > > > > > > > diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c > > > > > index 9b3bd083b411..1c68c3e0ebf9 100644 > > > > > --- a/drivers/block/virtio_blk.c > > > > > +++ b/drivers/block/virtio_blk.c > > > > > @@ -774,7 +774,7 @@ static int virtblk_probe(struct virtio_device *vdev) > > > > > memset(&vblk->tag_set, 0, sizeof(vblk->tag_set)); > > > > > vblk->tag_set.ops = &virtio_mq_ops; > > > > > vblk->tag_set.queue_depth = queue_depth; > > > > > - vblk->tag_set.numa_node = NUMA_NO_NODE; > > > > > + vblk->tag_set.numa_node = virtio_dev_to_node(vdev); > > > > I afraid that by doing it, you will increase chances to see OOM, because > > > > in NUMA_NO_NODE, MM will try allocate memory in whole system, while in > > > > the latter mode only on specific NUMA which can be depleted. > > > This is a common methodology we use in the block layer and in NVMe subsystem > > > and we don't afraid of the OOM issue you raised. > > There are many reasons for that, but we are talking about virtio here > > and not about NVMe. > > Ok. what reasons ? For example, NVMe are physical devices that rely on DMA operations, PCI connectivity e.t.c to operate. Such systems indeed can benefit from NUMA locality hints. At the end, these devices are physically connected to that NUMA node. In our case, virtio-blk is a software interface that doesn't have all these limitations. On the contrary, the virtio-blk can be created on one CPU and moved later to be close to the QEMU which can run on another NUMA node. Also this patch increases chances to get OOM by factor of NUMA nodes. Before your patch, the virtio_blk can allocate from X memory, after your patch it will be X/NUMB_NUMA_NODES. In addition, it has all chances to even hurt performance. So yes, post v2, but as Stefan and I asked, please provide supportive performance results, because what was done for another subsystem doesn't mean that it will be applicable here. Thanks
On 9/28/2021 7:27 PM, Leon Romanovsky wrote: > On Tue, Sep 28, 2021 at 06:59:15PM +0300, Max Gurtovoy wrote: >> On 9/27/2021 9:23 PM, Leon Romanovsky wrote: >>> On Mon, Sep 27, 2021 at 08:25:09PM +0300, Max Gurtovoy wrote: >>>> On 9/27/2021 2:34 PM, Leon Romanovsky wrote: >>>>> On Sun, Sep 26, 2021 at 05:55:18PM +0300, Max Gurtovoy wrote: >>>>>> To optimize performance, set the affinity of the block device tagset >>>>>> according to the virtio device affinity. >>>>>> >>>>>> Signed-off-by: Max Gurtovoy <mgurtovoy@nvidia.com> >>>>>> --- >>>>>> drivers/block/virtio_blk.c | 2 +- >>>>>> 1 file changed, 1 insertion(+), 1 deletion(-) >>>>>> >>>>>> diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c >>>>>> index 9b3bd083b411..1c68c3e0ebf9 100644 >>>>>> --- a/drivers/block/virtio_blk.c >>>>>> +++ b/drivers/block/virtio_blk.c >>>>>> @@ -774,7 +774,7 @@ static int virtblk_probe(struct virtio_device *vdev) >>>>>> memset(&vblk->tag_set, 0, sizeof(vblk->tag_set)); >>>>>> vblk->tag_set.ops = &virtio_mq_ops; >>>>>> vblk->tag_set.queue_depth = queue_depth; >>>>>> - vblk->tag_set.numa_node = NUMA_NO_NODE; >>>>>> + vblk->tag_set.numa_node = virtio_dev_to_node(vdev); >>>>> I afraid that by doing it, you will increase chances to see OOM, because >>>>> in NUMA_NO_NODE, MM will try allocate memory in whole system, while in >>>>> the latter mode only on specific NUMA which can be depleted. >>>> This is a common methodology we use in the block layer and in NVMe subsystem >>>> and we don't afraid of the OOM issue you raised. >>> There are many reasons for that, but we are talking about virtio here >>> and not about NVMe. >> Ok. what reasons ? > For example, NVMe are physical devices that rely on DMA operations, > PCI connectivity e.t.c to operate. Such systems indeed can benefit from > NUMA locality hints. At the end, these devices are physically connected > to that NUMA node. FYI Virtio devices are also physical devices that have PCI interface and rely on DMA operations. from virtio spec: "Virtio devices use normal bus mechanisms of interrupts and DMA which should be familiar to any device driver author". Also we develop virtio HW at NVIDIA for blk and net devices with our SNAP technology. These devices are connected via PCI bus to the host. We also support SRIOV. Same it true also for paravirt devices that are emulated by QEMU but still the guest sees them as PCI devices. > > In our case, virtio-blk is a software interface that doesn't have all > these limitations. On the contrary, the virtio-blk can be created on one > CPU and moved later to be close to the QEMU which can run on another NUMA > node. Not at all. virtio is HW interface. I don't understand what are you saying here ? > > Also this patch increases chances to get OOM by factor of NUMA nodes. This is common practice in Linux for storage drivers. Why does it bothers you at all ? I already decreased the memory footprint for virtio blk devices. > Before your patch, the virtio_blk can allocate from X memory, after your > patch it will be X/NUMB_NUMA_NODES. So go ahead and change all the block layer if it bothers you so much. Also please change the NVMe subsystem when you do it. And lets see what the community will say. > In addition, it has all chances to even hurt performance. > > So yes, post v2, but as Stefan and I asked, please provide supportive > performance results, because what was done for another subsystem doesn't > mean that it will be applicable here. I will measure the perf but even if we wont see an improvement since it might not be the bottleneck, this changes should be merged since this is the way the block layer is optimized. This is a micro optimization that commonly used also in other subsystem. And non of your above reasons (PCI, SW device, DMA) is true. Virtio blk device is in 99% a PCI device (paravirt or real HW) exactly like any other PCI device you are familiar with. It's connected physically to some slot, it has a BAR, MMIO, configuration space, etc.. Thanks. > > Thanks
On Wed, Sep 29, 2021 at 02:28:08AM +0300, Max Gurtovoy wrote: > > On 9/28/2021 7:27 PM, Leon Romanovsky wrote: > > On Tue, Sep 28, 2021 at 06:59:15PM +0300, Max Gurtovoy wrote: > > > On 9/27/2021 9:23 PM, Leon Romanovsky wrote: > > > > On Mon, Sep 27, 2021 at 08:25:09PM +0300, Max Gurtovoy wrote: > > > > > On 9/27/2021 2:34 PM, Leon Romanovsky wrote: > > > > > > On Sun, Sep 26, 2021 at 05:55:18PM +0300, Max Gurtovoy wrote: > > > > > > > To optimize performance, set the affinity of the block device tagset > > > > > > > according to the virtio device affinity. > > > > > > > > > > > > > > Signed-off-by: Max Gurtovoy <mgurtovoy@nvidia.com> > > > > > > > --- > > > > > > > drivers/block/virtio_blk.c | 2 +- > > > > > > > 1 file changed, 1 insertion(+), 1 deletion(-) > > > > > > > > > > > > > > diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c > > > > > > > index 9b3bd083b411..1c68c3e0ebf9 100644 > > > > > > > --- a/drivers/block/virtio_blk.c > > > > > > > +++ b/drivers/block/virtio_blk.c > > > > > > > @@ -774,7 +774,7 @@ static int virtblk_probe(struct virtio_device *vdev) > > > > > > > memset(&vblk->tag_set, 0, sizeof(vblk->tag_set)); > > > > > > > vblk->tag_set.ops = &virtio_mq_ops; > > > > > > > vblk->tag_set.queue_depth = queue_depth; > > > > > > > - vblk->tag_set.numa_node = NUMA_NO_NODE; > > > > > > > + vblk->tag_set.numa_node = virtio_dev_to_node(vdev); > > > > > > I afraid that by doing it, you will increase chances to see OOM, because > > > > > > in NUMA_NO_NODE, MM will try allocate memory in whole system, while in > > > > > > the latter mode only on specific NUMA which can be depleted. > > > > > This is a common methodology we use in the block layer and in NVMe subsystem > > > > > and we don't afraid of the OOM issue you raised. > > > > There are many reasons for that, but we are talking about virtio here > > > > and not about NVMe. > > > Ok. what reasons ? > > For example, NVMe are physical devices that rely on DMA operations, > > PCI connectivity e.t.c to operate. Such systems indeed can benefit from > > NUMA locality hints. At the end, these devices are physically connected > > to that NUMA node. > > FYI Virtio devices are also physical devices that have PCI interface and > rely on DMA operations. > > from virtio spec: "Virtio devices use normal bus mechanisms of interrupts > and DMA which should be familiar > to any device driver author". Yes, this is how bus in Linux is implemented, there is nothing new here. > > Also we develop virtio HW at NVIDIA for blk and net devices with our SNAP > technology. > > These devices are connected via PCI bus to the host. How all these related to general virtio-blk implementation? > > We also support SRIOV. > > Same it true also for paravirt devices that are emulated by QEMU but still > the guest sees them as PCI devices. Yes, the key word here - "emulated". > > > > > In our case, virtio-blk is a software interface that doesn't have all > > these limitations. On the contrary, the virtio-blk can be created on one > > CPU and moved later to be close to the QEMU which can run on another NUMA > > node. > > Not at all. virtio is HW interface. Virtio are para-virtualized devices that are represented as HW interfaces in the guest OS. They are not needed to be real devices in the hypervisor, which is my (and probably most of the world) use case. My QEMU command line contains something like that: "-drive file=IMAGE.img,if=virtio" > > I don't understand what are you saying here ? > > > > > Also this patch increases chances to get OOM by factor of NUMA nodes. > > This is common practice in Linux for storage drivers. Why does it bothers > you at all ? Do I need a reason to ask for a clarification for publicly posted patch in open mailing list? I use virtio and care about it. > > I already decreased the memory footprint for virtio blk devices. As I wrote before, you decreased by several KB, but by this patch you limited available memory in magnitudes. > > > > Before your patch, the virtio_blk can allocate from X memory, after your > > patch it will be X/NUMB_NUMA_NODES. > > So go ahead and change all the block layer if it bothers you so much. > > Also please change the NVMe subsystem when you do it. I suggest less radical approach - don't take patches without proven benefit. We are in 2021, let's rely on NUMA node policy. > > And lets see what the community will say. Stephen asked you for performance data too. I'm not alone here. > > > In addition, it has all chances to even hurt performance. > > > > So yes, post v2, but as Stefan and I asked, please provide supportive > > performance results, because what was done for another subsystem doesn't > > mean that it will be applicable here. > > I will measure the perf but even if we wont see an improvement since it > might not be the bottleneck, this changes should be merged since this is the > way the block layer is optimized. This is not acceptance criteria to merge patches. > > This is a micro optimization that commonly used also in other subsystem. And > non of your above reasons (PCI, SW device, DMA) is true. Every subsystem is different, in some it makes sense, in others it doesn't. We (RDMA) had very long discussion (together with perf data) and heavily tailored test to measure influence of per-node allocations and guess what? We didn't see any performance advantage. https://lore.kernel.org/linux-rdma/c34a864803f9bbd33d3f856a6ba2dd595ab708a7.1620729033.git.leonro@nvidia.com/ > > Virtio blk device is in 99% a PCI device (paravirt or real HW) exactly like > any other PCI device you are familiar with. > > It's connected physically to some slot, it has a BAR, MMIO, configuration > space, etc.. In general case, it is far from being true. > > Thanks. > > > > > Thanks
On 9/29/2021 9:50 AM, Leon Romanovsky wrote: > On Wed, Sep 29, 2021 at 02:28:08AM +0300, Max Gurtovoy wrote: >> On 9/28/2021 7:27 PM, Leon Romanovsky wrote: >>> On Tue, Sep 28, 2021 at 06:59:15PM +0300, Max Gurtovoy wrote: >>>> On 9/27/2021 9:23 PM, Leon Romanovsky wrote: >>>>> On Mon, Sep 27, 2021 at 08:25:09PM +0300, Max Gurtovoy wrote: >>>>>> On 9/27/2021 2:34 PM, Leon Romanovsky wrote: >>>>>>> On Sun, Sep 26, 2021 at 05:55:18PM +0300, Max Gurtovoy wrote: >>>>>>>> To optimize performance, set the affinity of the block device tagset >>>>>>>> according to the virtio device affinity. >>>>>>>> >>>>>>>> Signed-off-by: Max Gurtovoy <mgurtovoy@nvidia.com> >>>>>>>> --- >>>>>>>> drivers/block/virtio_blk.c | 2 +- >>>>>>>> 1 file changed, 1 insertion(+), 1 deletion(-) >>>>>>>> >>>>>>>> diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c >>>>>>>> index 9b3bd083b411..1c68c3e0ebf9 100644 >>>>>>>> --- a/drivers/block/virtio_blk.c >>>>>>>> +++ b/drivers/block/virtio_blk.c >>>>>>>> @@ -774,7 +774,7 @@ static int virtblk_probe(struct virtio_device *vdev) >>>>>>>> memset(&vblk->tag_set, 0, sizeof(vblk->tag_set)); >>>>>>>> vblk->tag_set.ops = &virtio_mq_ops; >>>>>>>> vblk->tag_set.queue_depth = queue_depth; >>>>>>>> - vblk->tag_set.numa_node = NUMA_NO_NODE; >>>>>>>> + vblk->tag_set.numa_node = virtio_dev_to_node(vdev); >>>>>>> I afraid that by doing it, you will increase chances to see OOM, because >>>>>>> in NUMA_NO_NODE, MM will try allocate memory in whole system, while in >>>>>>> the latter mode only on specific NUMA which can be depleted. >>>>>> This is a common methodology we use in the block layer and in NVMe subsystem >>>>>> and we don't afraid of the OOM issue you raised. >>>>> There are many reasons for that, but we are talking about virtio here >>>>> and not about NVMe. >>>> Ok. what reasons ? >>> For example, NVMe are physical devices that rely on DMA operations, >>> PCI connectivity e.t.c to operate. Such systems indeed can benefit from >>> NUMA locality hints. At the end, these devices are physically connected >>> to that NUMA node. >> FYI Virtio devices are also physical devices that have PCI interface and >> rely on DMA operations. >> >> from virtio spec: "Virtio devices use normal bus mechanisms of interrupts >> and DMA which should be familiar >> to any device driver author". > Yes, this is how bus in Linux is implemented, there is nothing new here. So why you said that virtio is not a PCI device with DMA capabilities ? > >> Also we develop virtio HW at NVIDIA for blk and net devices with our SNAP >> technology. >> >> These devices are connected via PCI bus to the host. > How all these related to general virtio-blk implementation? They use the same driver. We develop HW virtio devices for bare metal cloud and also for virtualized cloud that use the SRIOV feature of the PF (real PF). > >> We also support SRIOV. >> >> Same it true also for paravirt devices that are emulated by QEMU but still >> the guest sees them as PCI devices. > Yes, the key word here - "emulated". It doesn't matter. The guest kernel doesn't know if it's a paravirt device or real NVIDIA HW virtio SNAP device. And FYI, a guest can also have 2 NUMA nodes and can benefit from this patch. > >>> In our case, virtio-blk is a software interface that doesn't have all >>> these limitations. On the contrary, the virtio-blk can be created on one >>> CPU and moved later to be close to the QEMU which can run on another NUMA >>> node. >> Not at all. virtio is HW interface. > Virtio are para-virtualized devices that are represented as HW interfaces > in the guest OS. They are not needed to be real devices in the hypervisor, > which is my (and probably most of the world) use case. Again, the kernel doesn't care or know if its a paravirt device or not. And it shouldn't care. This patch is for kernel driver and not QEMU. > > My QEMU command line contains something like that: "-drive file=IMAGE.img,if=virtio" This is one option. For NVIDIA HW device, you pass a virtio device exactly how you pass a mlx5 device - using vfio + vfio_pci. > >> I don't understand what are you saying here ? >> >>> Also this patch increases chances to get OOM by factor of NUMA nodes. >> This is common practice in Linux for storage drivers. Why does it bothers >> you at all ? > Do I need a reason to ask for a clarification for publicly posted patch > in open mailing list? > > I use virtio and care about it. I meant, why don't you want to change the entire block layer and NVMe subsystem ? Why only this patch bothers you ? > >> I already decreased the memory footprint for virtio blk devices. > As I wrote before, you decreased by several KB, but by this patch you > limited available memory in magnitudes. > >> >>> Before your patch, the virtio_blk can allocate from X memory, after your >>> patch it will be X/NUMB_NUMA_NODES. >> So go ahead and change all the block layer if it bothers you so much. >> >> Also please change the NVMe subsystem when you do it. > I suggest less radical approach - don't take patches without proven > benefit. > > We are in 2021, let's rely on NUMA node policy. I'm trying to add NUMA policy here. Exactly. > >> And lets see what the community will say. > Stephen asked you for performance data too. I'm not alone here. I said I'll have a V2. I also would like to hear the opinion of the block maintainers like Jens and Christoph regarding numa affinity for block drivers. >>> In addition, it has all chances to even hurt performance. >>> >>> So yes, post v2, but as Stefan and I asked, please provide supportive >>> performance results, because what was done for another subsystem doesn't >>> mean that it will be applicable here. >> I will measure the perf but even if we wont see an improvement since it >> might not be the bottleneck, this changes should be merged since this is the >> way the block layer is optimized. > This is not acceptance criteria to merge patches. > >> This is a micro optimization that commonly used also in other subsystem. And >> non of your above reasons (PCI, SW device, DMA) is true. > Every subsystem is different, in some it makes sense, in others it doesn't. But you were wrong saying that virtio device is not PCI HW device that uses DMA. Do you understand the solution now ? > > We (RDMA) had very long discussion (together with perf data) and heavily tailored > test to measure influence of per-node allocations and guess what? We didn't see > any performance advantage. > > https://lore.kernel.org/linux-rdma/c34a864803f9bbd33d3f856a6ba2dd595ab708a7.1620729033.git.leonro@nvidia.com/ So go ahead and change all the kernel or the block layer. As you said, for RDMA subsystem it might not be a good idea. I don't want to discuss RDMA considerations in this thread. Lets talk storage and virtio. > >> Virtio blk device is in 99% a PCI device (paravirt or real HW) exactly like >> any other PCI device you are familiar with. >> >> It's connected physically to some slot, it has a BAR, MMIO, configuration >> space, etc.. > In general case, it is far from being true. it's exactly true. But let give MST and Stephan to comment. > >> Thanks. >> >>> Thanks
On 9/28/2021 9:47 AM, Stefan Hajnoczi wrote: > On Mon, Sep 27, 2021 at 08:39:30PM +0300, Max Gurtovoy wrote: >> On 9/27/2021 11:09 AM, Stefan Hajnoczi wrote: >>> On Sun, Sep 26, 2021 at 05:55:18PM +0300, Max Gurtovoy wrote: >>>> To optimize performance, set the affinity of the block device tagset >>>> according to the virtio device affinity. >>>> >>>> Signed-off-by: Max Gurtovoy <mgurtovoy@nvidia.com> >>>> --- >>>> drivers/block/virtio_blk.c | 2 +- >>>> 1 file changed, 1 insertion(+), 1 deletion(-) >>>> >>>> diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c >>>> index 9b3bd083b411..1c68c3e0ebf9 100644 >>>> --- a/drivers/block/virtio_blk.c >>>> +++ b/drivers/block/virtio_blk.c >>>> @@ -774,7 +774,7 @@ static int virtblk_probe(struct virtio_device *vdev) >>>> memset(&vblk->tag_set, 0, sizeof(vblk->tag_set)); >>>> vblk->tag_set.ops = &virtio_mq_ops; >>>> vblk->tag_set.queue_depth = queue_depth; >>>> - vblk->tag_set.numa_node = NUMA_NO_NODE; >>>> + vblk->tag_set.numa_node = virtio_dev_to_node(vdev); >>>> vblk->tag_set.flags = BLK_MQ_F_SHOULD_MERGE; >>>> vblk->tag_set.cmd_size = >>>> sizeof(struct virtblk_req) + >>> I implemented NUMA affinity in the past and could not demonstrate a >>> performance improvement: >>> https://lists.linuxfoundation.org/pipermail/virtualization/2020-June/048248.html >>> >>> The pathological case is when a guest with vNUMA has the virtio-blk-pci >>> device on the "wrong" host NUMA node. Then memory accesses should cross >>> NUMA nodes. Still, it didn't seem to matter. >> I think the reason you didn't see any improvement is since you didn't use >> the right device for the node query. See my patch 1/2. > That doesn't seem to be the case. Please see > drivers/base/core.c:device_add(): > > /* use parent numa_node */ > if (parent && (dev_to_node(dev) == NUMA_NO_NODE)) > set_dev_node(dev, dev_to_node(parent)); > > IMO it's cleaner to use dev_to_node(&vdev->dev) than to directly access > the parent. > > Have I missed something? but dev_to_node(dev) is 0 IMO. who set it to NUMA_NO_NODE ? > >> I can try integrating these patches in my series and fix it. >> >> BTW, we might not see a big improvement because of other bottlenecks but >> this is known perf optimization we use often in block storage drivers. > Let's see benchmark results. Otherwise this is just dead code that adds > complexity. > > Stefan
On Wed, Sep 29, 2021 at 06:07:52PM +0300, Max Gurtovoy wrote: > > On 9/28/2021 9:47 AM, Stefan Hajnoczi wrote: > > On Mon, Sep 27, 2021 at 08:39:30PM +0300, Max Gurtovoy wrote: > > > On 9/27/2021 11:09 AM, Stefan Hajnoczi wrote: > > > > On Sun, Sep 26, 2021 at 05:55:18PM +0300, Max Gurtovoy wrote: > > > > > To optimize performance, set the affinity of the block device tagset > > > > > according to the virtio device affinity. > > > > > > > > > > Signed-off-by: Max Gurtovoy <mgurtovoy@nvidia.com> > > > > > --- > > > > > drivers/block/virtio_blk.c | 2 +- > > > > > 1 file changed, 1 insertion(+), 1 deletion(-) > > > > > > > > > > diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c > > > > > index 9b3bd083b411..1c68c3e0ebf9 100644 > > > > > --- a/drivers/block/virtio_blk.c > > > > > +++ b/drivers/block/virtio_blk.c > > > > > @@ -774,7 +774,7 @@ static int virtblk_probe(struct virtio_device *vdev) > > > > > memset(&vblk->tag_set, 0, sizeof(vblk->tag_set)); > > > > > vblk->tag_set.ops = &virtio_mq_ops; > > > > > vblk->tag_set.queue_depth = queue_depth; > > > > > - vblk->tag_set.numa_node = NUMA_NO_NODE; > > > > > + vblk->tag_set.numa_node = virtio_dev_to_node(vdev); > > > > > vblk->tag_set.flags = BLK_MQ_F_SHOULD_MERGE; > > > > > vblk->tag_set.cmd_size = > > > > > sizeof(struct virtblk_req) + > > > > I implemented NUMA affinity in the past and could not demonstrate a > > > > performance improvement: > > > > https://lists.linuxfoundation.org/pipermail/virtualization/2020-June/048248.html > > > > > > > > The pathological case is when a guest with vNUMA has the virtio-blk-pci > > > > device on the "wrong" host NUMA node. Then memory accesses should cross > > > > NUMA nodes. Still, it didn't seem to matter. > > > I think the reason you didn't see any improvement is since you didn't use > > > the right device for the node query. See my patch 1/2. > > That doesn't seem to be the case. Please see > > drivers/base/core.c:device_add(): > > > > /* use parent numa_node */ > > if (parent && (dev_to_node(dev) == NUMA_NO_NODE)) > > set_dev_node(dev, dev_to_node(parent)); > > > > IMO it's cleaner to use dev_to_node(&vdev->dev) than to directly access > > the parent. > > > > Have I missed something? > > but dev_to_node(dev) is 0 IMO. > > who set it to NUMA_NO_NODE ? drivers/virtio/virtio.c:register_virtio_device(): device_initialize(&dev->dev); drivers/base/core.c:device_initialize(): set_dev_node(dev, -1); Stefan
On 9/30/2021 4:16 PM, Stefan Hajnoczi wrote: > On Wed, Sep 29, 2021 at 06:07:52PM +0300, Max Gurtovoy wrote: >> On 9/28/2021 9:47 AM, Stefan Hajnoczi wrote: >>> On Mon, Sep 27, 2021 at 08:39:30PM +0300, Max Gurtovoy wrote: >>>> On 9/27/2021 11:09 AM, Stefan Hajnoczi wrote: >>>>> On Sun, Sep 26, 2021 at 05:55:18PM +0300, Max Gurtovoy wrote: >>>>>> To optimize performance, set the affinity of the block device tagset >>>>>> according to the virtio device affinity. >>>>>> >>>>>> Signed-off-by: Max Gurtovoy <mgurtovoy@nvidia.com> >>>>>> --- >>>>>> drivers/block/virtio_blk.c | 2 +- >>>>>> 1 file changed, 1 insertion(+), 1 deletion(-) >>>>>> >>>>>> diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c >>>>>> index 9b3bd083b411..1c68c3e0ebf9 100644 >>>>>> --- a/drivers/block/virtio_blk.c >>>>>> +++ b/drivers/block/virtio_blk.c >>>>>> @@ -774,7 +774,7 @@ static int virtblk_probe(struct virtio_device *vdev) >>>>>> memset(&vblk->tag_set, 0, sizeof(vblk->tag_set)); >>>>>> vblk->tag_set.ops = &virtio_mq_ops; >>>>>> vblk->tag_set.queue_depth = queue_depth; >>>>>> - vblk->tag_set.numa_node = NUMA_NO_NODE; >>>>>> + vblk->tag_set.numa_node = virtio_dev_to_node(vdev); >>>>>> vblk->tag_set.flags = BLK_MQ_F_SHOULD_MERGE; >>>>>> vblk->tag_set.cmd_size = >>>>>> sizeof(struct virtblk_req) + >>>>> I implemented NUMA affinity in the past and could not demonstrate a >>>>> performance improvement: >>>>> https://lists.linuxfoundation.org/pipermail/virtualization/2020-June/048248.html >>>>> >>>>> The pathological case is when a guest with vNUMA has the virtio-blk-pci >>>>> device on the "wrong" host NUMA node. Then memory accesses should cross >>>>> NUMA nodes. Still, it didn't seem to matter. >>>> I think the reason you didn't see any improvement is since you didn't use >>>> the right device for the node query. See my patch 1/2. >>> That doesn't seem to be the case. Please see >>> drivers/base/core.c:device_add(): >>> >>> /* use parent numa_node */ >>> if (parent && (dev_to_node(dev) == NUMA_NO_NODE)) >>> set_dev_node(dev, dev_to_node(parent)); >>> >>> IMO it's cleaner to use dev_to_node(&vdev->dev) than to directly access >>> the parent. >>> >>> Have I missed something? >> but dev_to_node(dev) is 0 IMO. >> >> who set it to NUMA_NO_NODE ? > drivers/virtio/virtio.c:register_virtio_device(): > > device_initialize(&dev->dev); > > drivers/base/core.c:device_initialize(): > > set_dev_node(dev, -1); Ohh I was searching NUMA_NO_NODE. I guess the initial commit from Christoph 15 years ago was before adding this macro. I'll send a patch to fix it. I hope I'll have a system to check your patches next week. > > Stefan
diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c index 9b3bd083b411..1c68c3e0ebf9 100644 --- a/drivers/block/virtio_blk.c +++ b/drivers/block/virtio_blk.c @@ -774,7 +774,7 @@ static int virtblk_probe(struct virtio_device *vdev) memset(&vblk->tag_set, 0, sizeof(vblk->tag_set)); vblk->tag_set.ops = &virtio_mq_ops; vblk->tag_set.queue_depth = queue_depth; - vblk->tag_set.numa_node = NUMA_NO_NODE; + vblk->tag_set.numa_node = virtio_dev_to_node(vdev); vblk->tag_set.flags = BLK_MQ_F_SHOULD_MERGE; vblk->tag_set.cmd_size = sizeof(struct virtblk_req) +
To optimize performance, set the affinity of the block device tagset according to the virtio device affinity. Signed-off-by: Max Gurtovoy <mgurtovoy@nvidia.com> --- drivers/block/virtio_blk.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)