Re: [PATCHv4] virtio-spec: virtio network device multiqueue support
On Wed, Sep 19, 2012 at 11:10:10AM +0930, Rusty Russell wrote: Tom Herbert therb...@google.com writes: On Tue, Sep 11, 2012 at 10:49 PM, Rusty Russell ru...@rustcorp.com.auwrote: Perhaps Tom can explain how we avoid out-of-order receive for the accelerated RFS case? It's not clear to me, but we need to be able to do that for virtio-net if it implements accelerated RFS. AFAIK ooo RX is possible with accelerated RFS. We have an algorithm that prevents this for RFS case by deferring a migration to a new queue as long as it's possible that a flow might have outstanding packets on the old queue. I suppose this could be implemented in the device for the HW queues, but I don't think it would be easy to cover all cases where packets were already in transit to the host or other cases where host and device queues are out of sync. Having gone to such great lengths to avoid ooo for RFS, I don't think DaveM would be happy if we allow it for virtio_net. So, how *would* we implement such a thing for a hardware device? What if the device will only change the receive queue if the old receive queue is empty? Cheers, Rusty. I think that would do it in most cases. Or if we want to be more exact we could delay switching a specific flow until no outstanding rx packets for this flow. Not sure it's worth the hassle. -- MST -- To unsubscribe from this list: send the line unsubscribe kvm in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [PATCHv4] virtio-spec: virtio network device multiqueue support
Tom Herbert therb...@google.com writes: On Tue, Sep 11, 2012 at 10:49 PM, Rusty Russell ru...@rustcorp.com.auwrote: Perhaps Tom can explain how we avoid out-of-order receive for the accelerated RFS case? It's not clear to me, but we need to be able to do that for virtio-net if it implements accelerated RFS. AFAIK ooo RX is possible with accelerated RFS. We have an algorithm that prevents this for RFS case by deferring a migration to a new queue as long as it's possible that a flow might have outstanding packets on the old queue. I suppose this could be implemented in the device for the HW queues, but I don't think it would be easy to cover all cases where packets were already in transit to the host or other cases where host and device queues are out of sync. Having gone to such great lengths to avoid ooo for RFS, I don't think DaveM would be happy if we allow it for virtio_net. So, how *would* we implement such a thing for a hardware device? What if the device will only change the receive queue if the old receive queue is empty? Cheers, Rusty. -- To unsubscribe from this list: send the line unsubscribe kvm in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [PATCHv4] virtio-spec: virtio network device multiqueue support
Jason Wang jasow...@redhat.com writes: On 09/10/2012 02:33 PM, Michael S. Tsirkin wrote: A final addition: what you suggest above would be TX follows RX, right? BTW, yes. But it's a weird way to express what the nic is doing. It is in anticipation of something like that, that I made steering programming so generic. I think TX follows RX is more immediately useful for reasons above but we can add both to spec and let drivers and devices decide what they want to support. You mean RX follows TX? ie. accelerated RFS. I agree. Perhaps Tom can explain how we avoid out-of-order receive for the accelerated RFS case? It's not clear to me, but we need to be able to do that for virtio-net if it implements accelerated RFS. AFAIK, ixgbe does rx follows tx. The only differences between ixgbe and virtio-net is that ixgbe driver programs the flow director during packet transmission but we suggest to do it silently in the device for simplicity. Implying the receive queue by xmit will be slightly laggy. Don't know if that's a problem. Cheers, Rusty. -- To unsubscribe from this list: send the line unsubscribe kvm in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [PATCHv4] virtio-spec: virtio network device multiqueue support
On Wed, Sep 12, 2012 at 03:19:11PM +0930, Rusty Russell wrote: Jason Wang jasow...@redhat.com writes: On 09/10/2012 02:33 PM, Michael S. Tsirkin wrote: A final addition: what you suggest above would be TX follows RX, right? BTW, yes. But it's a weird way to express what the nic is doing. It explains what the system is doing. TX is done by driver, RX by nic. We document both driver and device in the spec so I thought it's fine. any suggestions wellcome. It is in anticipation of something like that, that I made steering programming so generic. I think TX follows RX is more immediately useful for reasons above but we can add both to spec and let drivers and devices decide what they want to support. You mean RX follows TX? ie. accelerated RFS. I agree. Yes that's what I meant. Thanks for the correction. Perhaps Tom can explain how we avoid out-of-order receive for the accelerated RFS case? It's not clear to me, but we need to be able to do that for virtio-net if it implements accelerated RFS. Basically this has tx vq per cpu and relies on scheduler not bouncing threads between cpus too aggressively. Appears to be what ixgbe does. AFAIK, ixgbe does rx follows tx. The only differences between ixgbe and virtio-net is that ixgbe driver programs the flow director during packet transmission but we suggest to do it silently in the device for simplicity. Implying the receive queue by xmit will be slightly laggy. Don't know if that's a problem. Cheers, Rusty. Doesn't seem to be a problem in Jason's testing so far. -- To unsubscribe from this list: send the line unsubscribe kvm in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [PATCHv4] virtio-spec: virtio network device multiqueue support
On Wed, Sep 12, 2012 at 12:57 AM, Michael S. Tsirkin m...@redhat.com wrote: On Wed, Sep 12, 2012 at 03:19:11PM +0930, Rusty Russell wrote: Jason Wang jasow...@redhat.com writes: On 09/10/2012 02:33 PM, Michael S. Tsirkin wrote: A final addition: what you suggest above would be TX follows RX, right? BTW, yes. But it's a weird way to express what the nic is doing. It explains what the system is doing. TX is done by driver, RX by nic. We document both driver and device in the spec so I thought it's fine. any suggestions wellcome. It is in anticipation of something like that, that I made steering programming so generic. I think TX follows RX is more immediately useful for reasons above but we can add both to spec and let drivers and devices decide what they want to support. You mean RX follows TX? ie. accelerated RFS. I agree. RX following TX is logic of flow director I believe. {a}RFS has RX follow CPU where application receive is done on the socket. So in RFS there is no requirement to have a 1-1 correspondence between TX and RX queues, and in fact this allows different number of queues between TX and RX. We found this necessary when using priority HW queues, so that there are more TX queues than RX. Yes that's what I meant. Thanks for the correction. Perhaps Tom can explain how we avoid out-of-order receive for the accelerated RFS case? It's not clear to me, but we need to be able to do that for virtio-net if it implements accelerated RFS. AFAIK ooo RX is still possible with accelerated RFS. We have an algorithm that prevents this for RFS by deferring a migration to a new queue as long as it's possible that a flow might have outstanding packets on the old queue. I suppose this could be implemented in the device for the HW queues, but I don't think it would be easy to cover all cases where packets were already in transit to the host or other cases where host and device queues are out of sync. Basically this has tx vq per cpu and relies on scheduler not bouncing threads between cpus too aggressively. Appears to be what ixgbe does. AFAIK, ixgbe does rx follows tx. The only differences between ixgbe and virtio-net is that ixgbe driver programs the flow director during packet transmission but we suggest to do it silently in the device for simplicity. Implying the receive queue by xmit will be slightly laggy. Don't know if that's a problem. Cheers, Rusty. Doesn't seem to be a problem in Jason's testing so far. -- To unsubscribe from this list: send the line unsubscribe kvm in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [PATCHv4] virtio-spec: virtio network device multiqueue support
On Wed, 2012-09-12 at 07:40 -0700, Tom Herbert wrote: On Wed, Sep 12, 2012 at 12:57 AM, Michael S. Tsirkin m...@redhat.com wrote: On Wed, Sep 12, 2012 at 03:19:11PM +0930, Rusty Russell wrote: [...] Perhaps Tom can explain how we avoid out-of-order receive for the accelerated RFS case? It's not clear to me, but we need to be able to do that for virtio-net if it implements accelerated RFS. AFAIK ooo RX is still possible with accelerated RFS. We have an algorithm that prevents this for RFS by deferring a migration to a new queue as long as it's possible that a flow might have outstanding packets on the old queue. I suppose this could be implemented in the device for the HW queues, but I don't think it would be easy to cover all cases where packets were already in transit to the host or other cases where host and device queues are out of sync. [...] Yes, I couldn't see any way to eliminate the possibility of OOO. The software queue check in RFS should redirect the flow only when it is new or has had an idle period, when I hope only a few packets will be received before we send some kind of response (transport or application layer ACK). So I think that OOO is not that likely in practice, but I don't have the evidence to back that up. If the filter update latency is high enough that a response can overtake the filter update, there may be more of a problem. Ben. -- Ben Hutchings, Staff Engineer, Solarflare Not speaking for my employer; that's the marketing department's job. They asked us to note that Solarflare product names are trademarked. -- To unsubscribe from this list: send the line unsubscribe kvm in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [PATCHv4] virtio-spec: virtio network device multiqueue support
Michael S. Tsirkin m...@redhat.com writes: In other words RPS is a hack to speed up networking on cheapo hardware, this is one of the reasons it is off by default. Good hardware has multiple receive queues. We can implement a good one so we do not need RPS. Also not all guest OS-es support RPS. Does this clarify? Ok, thanks. BTW, I found a better description by Tom Herbert, BTW: https://code.google.com/p/kernel/wiki/NetScalingGuide Now, I find the description of VIRTIO_NET_CTRL_STEERING_RX_FOLLOWS_TX confusing: 1) AFAICT it turns on multiqueue rx, with no semantics attached. I have no idea why it's called what it is. Why? 2) We've said we can remove steering methods, but we haven't actually defined any, as we've left it completely open. If I were a driver author, it leaves me completely baffled on how to implement the spec :( What are we actually planning to implement at the moment? For best performance, packets from a single connection should utilize the paired transmit and receive queues from the same virtqueue pair; for example both transmitqN and receiveqN. This rule makes it possible to optimize processing on the device side, but this is not a hard requirement: devices should function correctly even when this rule is not followed. Why is this true? I don't actually see why the queues are in pairs at all; are tx and rx not completely independent? So why does it matter? When the steering rule is modified, some packets can still be outstanding in one or more of the virtqueues. Device is not required to wait for these packets to be consumed before delivering packets using the new streering rule. Drivers modifying the steering rule at a high rate (e.g. adaptively in response to changes in the workload) are recommended to complete processing of the receive queue(s) utilized by the original steering before processing any packets delivered by the modified steering rule. How can this be done? This isn't actually possible without taking the queue down, since more packets are incoming. Cheers, Rusty. -- To unsubscribe from this list: send the line unsubscribe kvm in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [PATCHv4] virtio-spec: virtio network device multiqueue support
On Mon, Sep 10, 2012 at 11:42:25AM +0930, Rusty Russell wrote: OK, I read the spec (pasted below for easy of reading), but I'm still confused over how this will work. I thought normal net drivers have the hardware provide an rxhash for each packet, and we map that to CPU to queue the packet on[1]. We hope that the receiving process migrates to that CPU, so xmit queue matches. This ony works sometimes. For example it's common to pin netperf to a cpu to get consistent performance. Proper hardware must obey what applications want it to do, not the other way around. For virtio this would mean a new per-packet rxhash value, right? Why are we doing something different? What am I missing? Thanks, Rusty. [1] Everything I Know About Networking I Learned From LWN: https://lwn.net/Articles/362339/ I think you missed this: Some network interfaces can help with the distribution of incoming packets; they have multiple receive queues and multiple interrupt lines. Others, though, are equipped with a single queue, meaning that the driver for that hardware must deal with all incoming packets in a single, serialized stream. Parallelizing such a stream requires some intelligence on the part of the host operating system. In other words RPS is a hack to speed up networking on cheapo hardware, this is one of the reasons it is off by default. Good hardware has multiple receive queues. We can implement a good one so we do not need RPS. Also not all guest OS-es support RPS. Does this clarify? --- Transmit Packet Steering When VIRTIO_NET_F_MULTIQUEUE feature bit is negotiated, guest can use any of multiple configured transmit queues to transmit a given packet. To avoid packet reordering by device (which generally leads to performance degradation) driver should attempt to utilize the same transmit virtqueue for all packets of a given transmit flow. For bi-directional protocols (in practice, TCP), a given network connection can utilize both transmit and receive queues. For best performance, packets from a single connection should utilize the paired transmit and receive queues from the same virtqueue pair; for example both transmitqN and receiveqN. This rule makes it possible to optimize processing on the device side, but this is not a hard requirement: devices should function correctly even when this rule is not followed. Driver selects an active steering rule using VIRTIO_NET_CTRL_STEERING command (this controls both which virtqueue is selected for a given packet for receive and notifies the device which virtqueues are about to be used for transmit). This command accepts a single out argument in the following format: #define VIRTIO_NET_CTRL_STEERING 4 The field rule specifies the function used to select transmit virtqueue for a given packet; the field param makes it possible to pass an extra parameter if appropriate. When rule is set to VIRTIO_NET_CTRL_STEERING_SINGLE (this is the default) all packets are steered to the default virtqueue transmitq (1); param is unused; this is the default. With any other rule, When rule is set to VIRTIO_NET_CTRL_STEERING_RX_FOLLOWS_TX packets are steered by driver to the first N=(param+1) multiqueue virtqueues transmitq1...transmitqN; the default transmitq is unused. Driver must have configured all these (param+1) virtqueues beforehand. Supported steering rules can be added and removed in the future. Driver should check that the request to change the steering rule was successful by checking ack values of the command. As selecting a specific steering is an optimization feature, drivers should avoid hard failure and fall back on using a supported steering rule if this command fails. The default steering rule is VIRTIO_NET_CTRL_STEERING_SINGLE. It will not be removed. When the steering rule is modified, some packets can still be outstanding in one or more of the transmit virtqueues. Since drivers might choose to modify the current steering rule at a high rate (e.g. adaptively in response to changes in the workload) to avoid reordering packets, device is recommended to complete processing of the transmit queue(s) utilized by the original steering before processing any packets delivered by the modified steering rule. For debugging, the current steering rule can also be read from the configuration space. Receive Packet Steering When VIRTIO_NET_F_MULTIQUEUE feature bit is negotiated, device can use any of multiple configured receive queues to pass a given packet to driver. Driver controls which virtqueue is selected in practice by configuring packet steering rule using VIRTIO_NET_CTRL_STEERING command, as described above[sub:Transmit-Packet-Steering]. The field rule specifies the function used to select receive virtqueue for a given packet; the field param makes it possible to pass an extra parameter
Re: [PATCHv4] virtio-spec: virtio network device multiqueue support
On Mon, Sep 10, 2012 at 09:16:29AM +0300, Michael S. Tsirkin wrote: On Mon, Sep 10, 2012 at 11:42:25AM +0930, Rusty Russell wrote: OK, I read the spec (pasted below for easy of reading), but I'm still confused over how this will work. I thought normal net drivers have the hardware provide an rxhash for each packet, and we map that to CPU to queue the packet on[1]. We hope that the receiving process migrates to that CPU, so xmit queue matches. This ony works sometimes. For example it's common to pin netperf to a cpu to get consistent performance. Proper hardware must obey what applications want it to do, not the other way around. For virtio this would mean a new per-packet rxhash value, right? Why are we doing something different? What am I missing? Thanks, Rusty. [1] Everything I Know About Networking I Learned From LWN: https://lwn.net/Articles/362339/ I think you missed this: Some network interfaces can help with the distribution of incoming packets; they have multiple receive queues and multiple interrupt lines. Others, though, are equipped with a single queue, meaning that the driver for that hardware must deal with all incoming packets in a single, serialized stream. Parallelizing such a stream requires some intelligence on the part of the host operating system. In other words RPS is a hack to speed up networking on cheapo hardware, this is one of the reasons it is off by default. Good hardware has multiple receive queues. We can implement a good one so we do not need RPS. Also not all guest OS-es support RPS. Does this clarify? I would like to add that on many processors, sending IPCs between guest CPUs requires exits on sending *and* receiving path, making it very expensive. --- Transmit Packet Steering When VIRTIO_NET_F_MULTIQUEUE feature bit is negotiated, guest can use any of multiple configured transmit queues to transmit a given packet. To avoid packet reordering by device (which generally leads to performance degradation) driver should attempt to utilize the same transmit virtqueue for all packets of a given transmit flow. For bi-directional protocols (in practice, TCP), a given network connection can utilize both transmit and receive queues. For best performance, packets from a single connection should utilize the paired transmit and receive queues from the same virtqueue pair; for example both transmitqN and receiveqN. This rule makes it possible to optimize processing on the device side, but this is not a hard requirement: devices should function correctly even when this rule is not followed. Driver selects an active steering rule using VIRTIO_NET_CTRL_STEERING command (this controls both which virtqueue is selected for a given packet for receive and notifies the device which virtqueues are about to be used for transmit). This command accepts a single out argument in the following format: #define VIRTIO_NET_CTRL_STEERING 4 The field rule specifies the function used to select transmit virtqueue for a given packet; the field param makes it possible to pass an extra parameter if appropriate. When rule is set to VIRTIO_NET_CTRL_STEERING_SINGLE (this is the default) all packets are steered to the default virtqueue transmitq (1); param is unused; this is the default. With any other rule, When rule is set to VIRTIO_NET_CTRL_STEERING_RX_FOLLOWS_TX packets are steered by driver to the first N=(param+1) multiqueue virtqueues transmitq1...transmitqN; the default transmitq is unused. Driver must have configured all these (param+1) virtqueues beforehand. Supported steering rules can be added and removed in the future. Driver should check that the request to change the steering rule was successful by checking ack values of the command. As selecting a specific steering is an optimization feature, drivers should avoid hard failure and fall back on using a supported steering rule if this command fails. The default steering rule is VIRTIO_NET_CTRL_STEERING_SINGLE. It will not be removed. When the steering rule is modified, some packets can still be outstanding in one or more of the transmit virtqueues. Since drivers might choose to modify the current steering rule at a high rate (e.g. adaptively in response to changes in the workload) to avoid reordering packets, device is recommended to complete processing of the transmit queue(s) utilized by the original steering before processing any packets delivered by the modified steering rule. For debugging, the current steering rule can also be read from the configuration space. Receive Packet Steering When VIRTIO_NET_F_MULTIQUEUE feature bit is negotiated, device can use any of multiple configured receive queues to pass a given packet to driver. Driver controls which virtqueue is
Re: [PATCHv4] virtio-spec: virtio network device multiqueue support
On Mon, Sep 10, 2012 at 09:27:38AM +0300, Michael S. Tsirkin wrote: On Mon, Sep 10, 2012 at 09:16:29AM +0300, Michael S. Tsirkin wrote: On Mon, Sep 10, 2012 at 11:42:25AM +0930, Rusty Russell wrote: OK, I read the spec (pasted below for easy of reading), but I'm still confused over how this will work. I thought normal net drivers have the hardware provide an rxhash for each packet, and we map that to CPU to queue the packet on[1]. We hope that the receiving process migrates to that CPU, so xmit queue matches. This ony works sometimes. For example it's common to pin netperf to a cpu to get consistent performance. Proper hardware must obey what applications want it to do, not the other way around. For virtio this would mean a new per-packet rxhash value, right? Why are we doing something different? What am I missing? Thanks, Rusty. [1] Everything I Know About Networking I Learned From LWN: https://lwn.net/Articles/362339/ I think you missed this: Some network interfaces can help with the distribution of incoming packets; they have multiple receive queues and multiple interrupt lines. Others, though, are equipped with a single queue, meaning that the driver for that hardware must deal with all incoming packets in a single, serialized stream. Parallelizing such a stream requires some intelligence on the part of the host operating system. In other words RPS is a hack to speed up networking on cheapo hardware, this is one of the reasons it is off by default. Good hardware has multiple receive queues. We can implement a good one so we do not need RPS. Also not all guest OS-es support RPS. Does this clarify? I would like to add that on many processors, sending IPCs between guest CPUs requires exits on sending *and* receiving path, making it very expensive. A final addition: what you suggest above would be TX follows RX, right? It is in anticipation of something like that, that I made steering programming so generic. I think TX follows RX is more immediately useful for reasons above but we can add both to spec and let drivers and devices decide what they want to support. Pls let me know. --- Transmit Packet Steering When VIRTIO_NET_F_MULTIQUEUE feature bit is negotiated, guest can use any of multiple configured transmit queues to transmit a given packet. To avoid packet reordering by device (which generally leads to performance degradation) driver should attempt to utilize the same transmit virtqueue for all packets of a given transmit flow. For bi-directional protocols (in practice, TCP), a given network connection can utilize both transmit and receive queues. For best performance, packets from a single connection should utilize the paired transmit and receive queues from the same virtqueue pair; for example both transmitqN and receiveqN. This rule makes it possible to optimize processing on the device side, but this is not a hard requirement: devices should function correctly even when this rule is not followed. Driver selects an active steering rule using VIRTIO_NET_CTRL_STEERING command (this controls both which virtqueue is selected for a given packet for receive and notifies the device which virtqueues are about to be used for transmit). This command accepts a single out argument in the following format: #define VIRTIO_NET_CTRL_STEERING 4 The field rule specifies the function used to select transmit virtqueue for a given packet; the field param makes it possible to pass an extra parameter if appropriate. When rule is set to VIRTIO_NET_CTRL_STEERING_SINGLE (this is the default) all packets are steered to the default virtqueue transmitq (1); param is unused; this is the default. With any other rule, When rule is set to VIRTIO_NET_CTRL_STEERING_RX_FOLLOWS_TX packets are steered by driver to the first N=(param+1) multiqueue virtqueues transmitq1...transmitqN; the default transmitq is unused. Driver must have configured all these (param+1) virtqueues beforehand. Supported steering rules can be added and removed in the future. Driver should check that the request to change the steering rule was successful by checking ack values of the command. As selecting a specific steering is an optimization feature, drivers should avoid hard failure and fall back on using a supported steering rule if this command fails. The default steering rule is VIRTIO_NET_CTRL_STEERING_SINGLE. It will not be removed. When the steering rule is modified, some packets can still be outstanding in one or more of the transmit virtqueues. Since drivers might choose to modify the current steering rule at a high rate (e.g. adaptively in response to changes in the workload) to avoid reordering packets, device
Re: [PATCHv4] virtio-spec: virtio network device multiqueue support
On 09/10/2012 02:33 PM, Michael S. Tsirkin wrote: On Mon, Sep 10, 2012 at 09:27:38AM +0300, Michael S. Tsirkin wrote: On Mon, Sep 10, 2012 at 09:16:29AM +0300, Michael S. Tsirkin wrote: On Mon, Sep 10, 2012 at 11:42:25AM +0930, Rusty Russell wrote: OK, I read the spec (pasted below for easy of reading), but I'm still confused over how this will work. I thought normal net drivers have the hardware provide an rxhash for each packet, and we map that to CPU to queue the packet on[1]. We hope that the receiving process migrates to that CPU, so xmit queue matches. This ony works sometimes. For example it's common to pin netperf to a cpu to get consistent performance. Proper hardware must obey what applications want it to do, not the other way around. For virtio this would mean a new per-packet rxhash value, right? Why are we doing something different? What am I missing? Thanks, Rusty. [1] Everything I Know About Networking I Learned From LWN: https://lwn.net/Articles/362339/ I think you missed this: Some network interfaces can help with the distribution of incoming packets; they have multiple receive queues and multiple interrupt lines. Others, though, are equipped with a single queue, meaning that the driver for that hardware must deal with all incoming packets in a single, serialized stream. Parallelizing such a stream requires some intelligence on the part of the host operating system. In other words RPS is a hack to speed up networking on cheapo hardware, this is one of the reasons it is off by default. Good hardware has multiple receive queues. We can implement a good one so we do not need RPS. Also not all guest OS-es support RPS. Does this clarify? I would like to add that on many processors, sending IPCs between guest CPUs requires exits on sending *and* receiving path, making it very expensive. A final addition: what you suggest above would be TX follows RX, right? It is in anticipation of something like that, that I made steering programming so generic. I think TX follows RX is more immediately useful for reasons above but we can add both to spec and let drivers and devices decide what they want to support. Pls let me know. AFAIK, ixgbe does rx follows tx. The only differences between ixgbe and virtio-net is that ixgbe driver programs the flow director during packet transmission but we suggest to do it silently in the device for simplicity. Even with this, more co-operation is still needed for the driver ( e.g ixgbe try to use per-cpu queue by setting affinity hint and using cpuid to choose the txq which could be reused in virtio-net driver). --- Transmit Packet Steering When VIRTIO_NET_F_MULTIQUEUE feature bit is negotiated, guest can use any of multiple configured transmit queues to transmit a given packet. To avoid packet reordering by device (which generally leads to performance degradation) driver should attempt to utilize the same transmit virtqueue for all packets of a given transmit flow. For bi-directional protocols (in practice, TCP), a given network connection can utilize both transmit and receive queues. For best performance, packets from a single connection should utilize the paired transmit and receive queues from the same virtqueue pair; for example both transmitqN and receiveqN. This rule makes it possible to optimize processing on the device side, but this is not a hard requirement: devices should function correctly even when this rule is not followed. Driver selects an active steering rule using VIRTIO_NET_CTRL_STEERING command (this controls both which virtqueue is selected for a given packet for receive and notifies the device which virtqueues are about to be used for transmit). This command accepts a single out argument in the following format: #define VIRTIO_NET_CTRL_STEERING 4 The field rule specifies the function used to select transmit virtqueue for a given packet; the field param makes it possible to pass an extra parameter if appropriate. When rule is set to VIRTIO_NET_CTRL_STEERING_SINGLE (this is the default) all packets are steered to the default virtqueue transmitq (1); param is unused; this is the default. With any other rule, When rule is set to VIRTIO_NET_CTRL_STEERING_RX_FOLLOWS_TX packets are steered by driver to the first N=(param+1) multiqueue virtqueues transmitq1...transmitqN; the default transmitq is unused. Driver must have configured all these (param+1) virtqueues beforehand. Supported steering rules can be added and removed in the future. Driver should check that the request to change the steering rule was successful by checking ack values of the command. As selecting a specific steering is an optimization feature, drivers should avoid hard failure and fall back on using a supported steering rule if this command fails. The default steering rule is VIRTIO_NET_CTRL_STEERING_SINGLE. It will not be removed. When the steering rule is
Re: [PATCHv4] virtio-spec: virtio network device multiqueue support
On 09/09/2012 07:12 PM, Rusty Russell wrote: OK, I read the spec (pasted below for easy of reading), but I'm still confused over how this will work. I thought normal net drivers have the hardware provide an rxhash for each packet, and we map that to CPU to queue the packet on[1]. We hope that the receiving process migrates to that CPU, so xmit queue matches. For virtio this would mean a new per-packet rxhash value, right? Why are we doing something different? What am I missing? Thanks, Rusty. [1] Everything I Know About Networking I Learned From LWN: https://lwn.net/Articles/362339/ In my taxonomy at least, multi-queue predates RPS and RFS and is simply where the NIC via some means, perhaps a headers hash, separates incoming frames to different queues. RPS can be thought of as doing something similar inside the host. That could be used to get a spread from an otherwise dumb NIC (certainly that is what one of its predecessors - Inbound Packet Scheduling - used it for in HP-UX 10.20), or it could be used to augment the multi-queue support of a not-so-dump NIC - say if said NIC had a limit of queues that was rather lower than the number of cores/threads in the host. Indeed some driver/NIC combinations provide a hash value to the host for the host to use as it sees fit. However, there is still the matter of a single thread of an application servicing multiple connections, each of which would hash to different locations. RFS (Receive Flow Steering) then goes one step further, and looks-up where the flow endpoint was last accessed and steers the traffic there. The idea being that a thread of execution servicing multiple flows will have the traffic of those flows sent to the same place. It then allows the scheduler to decide where things should be run rather than the networking code. rick jones -- To unsubscribe from this list: send the line unsubscribe kvm in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
Re: [PATCHv4] virtio-spec: virtio network device multiqueue support
OK, I read the spec (pasted below for easy of reading), but I'm still confused over how this will work. I thought normal net drivers have the hardware provide an rxhash for each packet, and we map that to CPU to queue the packet on[1]. We hope that the receiving process migrates to that CPU, so xmit queue matches. For virtio this would mean a new per-packet rxhash value, right? Why are we doing something different? What am I missing? Thanks, Rusty. [1] Everything I Know About Networking I Learned From LWN: https://lwn.net/Articles/362339/ --- Transmit Packet Steering When VIRTIO_NET_F_MULTIQUEUE feature bit is negotiated, guest can use any of multiple configured transmit queues to transmit a given packet. To avoid packet reordering by device (which generally leads to performance degradation) driver should attempt to utilize the same transmit virtqueue for all packets of a given transmit flow. For bi-directional protocols (in practice, TCP), a given network connection can utilize both transmit and receive queues. For best performance, packets from a single connection should utilize the paired transmit and receive queues from the same virtqueue pair; for example both transmitqN and receiveqN. This rule makes it possible to optimize processing on the device side, but this is not a hard requirement: devices should function correctly even when this rule is not followed. Driver selects an active steering rule using VIRTIO_NET_CTRL_STEERING command (this controls both which virtqueue is selected for a given packet for receive and notifies the device which virtqueues are about to be used for transmit). This command accepts a single out argument in the following format: #define VIRTIO_NET_CTRL_STEERING 4 The field rule specifies the function used to select transmit virtqueue for a given packet; the field param makes it possible to pass an extra parameter if appropriate. When rule is set to VIRTIO_NET_CTRL_STEERING_SINGLE (this is the default) all packets are steered to the default virtqueue transmitq (1); param is unused; this is the default. With any other rule, When rule is set to VIRTIO_NET_CTRL_STEERING_RX_FOLLOWS_TX packets are steered by driver to the first N=(param+1) multiqueue virtqueues transmitq1...transmitqN; the default transmitq is unused. Driver must have configured all these (param+1) virtqueues beforehand. Supported steering rules can be added and removed in the future. Driver should check that the request to change the steering rule was successful by checking ack values of the command. As selecting a specific steering is an optimization feature, drivers should avoid hard failure and fall back on using a supported steering rule if this command fails. The default steering rule is VIRTIO_NET_CTRL_STEERING_SINGLE. It will not be removed. When the steering rule is modified, some packets can still be outstanding in one or more of the transmit virtqueues. Since drivers might choose to modify the current steering rule at a high rate (e.g. adaptively in response to changes in the workload) to avoid reordering packets, device is recommended to complete processing of the transmit queue(s) utilized by the original steering before processing any packets delivered by the modified steering rule. For debugging, the current steering rule can also be read from the configuration space. Receive Packet Steering When VIRTIO_NET_F_MULTIQUEUE feature bit is negotiated, device can use any of multiple configured receive queues to pass a given packet to driver. Driver controls which virtqueue is selected in practice by configuring packet steering rule using VIRTIO_NET_CTRL_STEERING command, as described above[sub:Transmit-Packet-Steering]. The field rule specifies the function used to select receive virtqueue for a given packet; the field param makes it possible to pass an extra parameter if appropriate. When rule is set to VIRTIO_NET_CTRL_STEERING_SINGLE all packets are steered to the default virtqueue receiveq (0); param is unused; this is the default. When rule is set to VIRTIO_NET_CTRL_STEERING_RX_FOLLOWS_TX packets are steered by host to the first N=(param+1) multiqueue virtqueues receiveq1...receiveqN; the default receiveq is unused. Driver must have configured all these (param+1) virtqueues beforehand. For best performance for bi-directional flows (such as TCP) device should detect the flow to virtqueue pair mapping on transmit and select the receive virtqueue from the same virtqueue pair. For uni-directional flows, or when this mapping information is missing, a device-specific steering function is used. Supported steering rules can be added and removed in the future. Driver should probe for supported rules by checking ack values of the command. When the steering rule is modified, some packets can still be outstanding in one or more of the virtqueues. Device is not required to wait for these packets to be consumed before