Re: [Qemu-devel] [PATCH] vfio: fix duplicate function call
On 10/17/2016 11:01 PM, Alex Williamson wrote: On Mon, 17 Oct 2016 16:57:08 +0800 Cao jin wrote: Hi, On 10/14/2016 11:50 PM, Alex Williamson wrote: On Fri, 14 Oct 2016 19:16:59 +0800 Cao jin wrote: When vfio device is reset(encounter FLR, or bus reset), if need to do bus reset(vfio_pci_hot_reset_one is called), vfio_pci_pre_reset & vfio_pci_post_reset will be called twice. Signed-off-by: Cao jin --- Also has a little question on vfio_pci_reset. it will be called when encounter bus reset, or FLR. The reset method's priority in this function now is: 1. If has "device specific reset function", then do it 2. If has FLR, then do it. 3. If it can do bus reset(only 1 affected device), then do it 4. If has pm_reset, then do it The question is: why pm reset has low priority than bus reset(if it does can do a bus reset)? why bus reset is not the last choice? In PCI driver of kernel, pls see __pci_dev_reset, we can see, if device support pm reset, it won't do bus reset. The PCI spec doesn't really define what sort of reset is done with a PM reset. My thinking was that if a device advertises an FLR capability then the hardware has made a concerted effort to have a per function reset mechanism available. NoSoftRst- is not terribly common and it's not entirely clear to me that the hardware has made a conscious effort to provide this for the purposes of per function reset mechanism. Therefore I've opt'd to prioritize a bus reset over a PM reset. I still have a question about vfio_pci_reset. I checked commit message in f16f39c3, if I understand right, couldn't we put /* See if we can do our own bus reset */ if (!vfio_pci_hot_reset_one(vdev)) { goto post_reset; } in the 1st priority? Because if there is 1 affected device, then it will do bus reset which is the best reset we can do; if there are more than 1 affected devices, after this patch, vfio_pci_hot_reset_one will do nothing, and then try other reset methods. It's possible, yes, but that disregards that the hardware has gone to the trouble to implement a proper function level reset. As I explained, I de-prioritize PM reset, specifically because I'm not sure if hardware designers are necessarily intending it for the purpose of a device reset. For FLR this is the entire purpose of the interface. We also have a fair bit of experience with the current priority scheme and I would not take it lightly to change without some compelling evidence to prove that a new priority scheme is better than the existing. There do also exist devices which do not behave properly with a secondary bus reset, see drivers/pci/quirks.c:quirk_no_bus_reset() in the kernel tree. It's possible more devices like this exist, but we don't see them because they implement FLR. A bus reset may result in a more complete device reset, but it's also more disruptive to the system. Thanks, Alex I see. Thanks Alex, I think these are valuable info to me, although maybe I still need more time in the future to understand these totally. -- Yours Sincerely, Cao jin
Re: [Qemu-devel] [PATCH] vfio: fix duplicate function call
On Mon, 17 Oct 2016 16:57:08 +0800 Cao jin wrote: > Hi, > > On 10/14/2016 11:50 PM, Alex Williamson wrote: > > On Fri, 14 Oct 2016 19:16:59 +0800 > > Cao jin wrote: > > > >> When vfio device is reset(encounter FLR, or bus reset), if need to do > >> bus reset(vfio_pci_hot_reset_one is called), vfio_pci_pre_reset & > >> vfio_pci_post_reset will be called twice. > >> > >> Signed-off-by: Cao jin > >> --- > >> Also has a little question on vfio_pci_reset. it will be called when > >> encounter > >> bus reset, or FLR. The reset method's priority in this function now is: > >> > >> 1. If has "device specific reset function", then do it > >> 2. If has FLR, then do it. > >> 3. If it can do bus reset(only 1 affected device), then do it > >> 4. If has pm_reset, then do it > >> > >> The question is: why pm reset has low priority than bus reset(if it does > >> can do a bus reset)? why bus reset is not the last choice? In PCI driver > >> of kernel, pls see __pci_dev_reset, we can see, if device support pm reset, > >> it won't do bus reset. > > > > The PCI spec doesn't really define what sort of reset is done with a PM > > reset. My thinking was that if a device advertises an FLR capability > > then the hardware has made a concerted effort to have a per function > > reset mechanism available. NoSoftRst- is not terribly common and it's > > not entirely clear to me that the hardware has made a conscious effort > > to provide this for the purposes of per function reset mechanism. > > Therefore I've opt'd to prioritize a bus reset over a PM reset. > > > > I still have a question about vfio_pci_reset. I checked commit message > in f16f39c3, if I understand right, couldn't we put > > /* See if we can do our own bus reset */ > if (!vfio_pci_hot_reset_one(vdev)) { > goto post_reset; > } > > in the 1st priority? Because if there is 1 affected device, then it will > do bus reset which is the best reset we can do; if there are more than 1 > affected devices, after this patch, vfio_pci_hot_reset_one will do > nothing, and then try other reset methods. It's possible, yes, but that disregards that the hardware has gone to the trouble to implement a proper function level reset. As I explained, I de-prioritize PM reset, specifically because I'm not sure if hardware designers are necessarily intending it for the purpose of a device reset. For FLR this is the entire purpose of the interface. We also have a fair bit of experience with the current priority scheme and I would not take it lightly to change without some compelling evidence to prove that a new priority scheme is better than the existing. There do also exist devices which do not behave properly with a secondary bus reset, see drivers/pci/quirks.c:quirk_no_bus_reset() in the kernel tree. It's possible more devices like this exist, but we don't see them because they implement FLR. A bus reset may result in a more complete device reset, but it's also more disruptive to the system. Thanks, Alex
Re: [Qemu-devel] [PATCH] vfio: fix duplicate function call
Hi, On 10/14/2016 11:50 PM, Alex Williamson wrote: On Fri, 14 Oct 2016 19:16:59 +0800 Cao jin wrote: When vfio device is reset(encounter FLR, or bus reset), if need to do bus reset(vfio_pci_hot_reset_one is called), vfio_pci_pre_reset & vfio_pci_post_reset will be called twice. Signed-off-by: Cao jin --- Also has a little question on vfio_pci_reset. it will be called when encounter bus reset, or FLR. The reset method's priority in this function now is: 1. If has "device specific reset function", then do it 2. If has FLR, then do it. 3. If it can do bus reset(only 1 affected device), then do it 4. If has pm_reset, then do it The question is: why pm reset has low priority than bus reset(if it does can do a bus reset)? why bus reset is not the last choice? In PCI driver of kernel, pls see __pci_dev_reset, we can see, if device support pm reset, it won't do bus reset. The PCI spec doesn't really define what sort of reset is done with a PM reset. My thinking was that if a device advertises an FLR capability then the hardware has made a concerted effort to have a per function reset mechanism available. NoSoftRst- is not terribly common and it's not entirely clear to me that the hardware has made a conscious effort to provide this for the purposes of per function reset mechanism. Therefore I've opt'd to prioritize a bus reset over a PM reset. I still have a question about vfio_pci_reset. I checked commit message in f16f39c3, if I understand right, couldn't we put /* See if we can do our own bus reset */ if (!vfio_pci_hot_reset_one(vdev)) { goto post_reset; } in the 1st priority? Because if there is 1 affected device, then it will do bus reset which is the best reset we can do; if there are more than 1 affected devices, after this patch, vfio_pci_hot_reset_one will do nothing, and then try other reset methods. -- Yours Sincerely, Cao jin
Re: [Qemu-devel] [PATCH] vfio: fix duplicate function call
On 10/14/2016 11:50 PM, Alex Williamson wrote: On Fri, 14 Oct 2016 19:16:59 +0800 Cao jin wrote: When vfio device is reset(encounter FLR, or bus reset), if need to do bus reset(vfio_pci_hot_reset_one is called), vfio_pci_pre_reset & vfio_pci_post_reset will be called twice. Signed-off-by: Cao jin --- Also has a little question on vfio_pci_reset. it will be called when encounter bus reset, or FLR. The reset method's priority in this function now is: 1. If has "device specific reset function", then do it 2. If has FLR, then do it. 3. If it can do bus reset(only 1 affected device), then do it 4. If has pm_reset, then do it The question is: why pm reset has low priority than bus reset(if it does can do a bus reset)? why bus reset is not the last choice? In PCI driver of kernel, pls see __pci_dev_reset, we can see, if device support pm reset, it won't do bus reset. The PCI spec doesn't really define what sort of reset is done with a PM reset. My thinking was that if a device advertises an FLR capability then the hardware has made a concerted effort to have a per function reset mechanism available. NoSoftRst- is not terribly common and it's not entirely clear to me that the hardware has made a conscious effort to provide this for the purposes of per function reset mechanism. Therefore I've opt'd to prioritize a bus reset over a PM reset. I see, thanks for your infomation. -- Yours Sincerely, Cao jin
Re: [Qemu-devel] [PATCH] vfio: fix duplicate function call
On Fri, 14 Oct 2016 19:16:59 +0800 Cao jin wrote: > When vfio device is reset(encounter FLR, or bus reset), if need to do > bus reset(vfio_pci_hot_reset_one is called), vfio_pci_pre_reset & > vfio_pci_post_reset will be called twice. > > Signed-off-by: Cao jin > --- > Also has a little question on vfio_pci_reset. it will be called when encounter > bus reset, or FLR. The reset method's priority in this function now is: > > 1. If has "device specific reset function", then do it > 2. If has FLR, then do it. > 3. If it can do bus reset(only 1 affected device), then do it > 4. If has pm_reset, then do it > > The question is: why pm reset has low priority than bus reset(if it does > can do a bus reset)? why bus reset is not the last choice? In PCI driver > of kernel, pls see __pci_dev_reset, we can see, if device support pm reset, > it won't do bus reset. The PCI spec doesn't really define what sort of reset is done with a PM reset. My thinking was that if a device advertises an FLR capability then the hardware has made a concerted effort to have a per function reset mechanism available. NoSoftRst- is not terribly common and it's not entirely clear to me that the hardware has made a conscious effort to provide this for the purposes of per function reset mechanism. Therefore I've opt'd to prioritize a bus reset over a PM reset. > hw/vfio/pci.c | 8 ++-- > 1 file changed, 6 insertions(+), 2 deletions(-) > > diff --git a/hw/vfio/pci.c b/hw/vfio/pci.c > index cce3024..ca4d1c1 100644 > --- a/hw/vfio/pci.c > +++ b/hw/vfio/pci.c > @@ -1930,7 +1930,9 @@ static int vfio_pci_hot_reset(VFIOPCIDevice *vdev, bool > single) > > trace_vfio_pci_hot_reset(vdev->vbasedev.name, single ? "one" : "multi"); > > -vfio_pci_pre_reset(vdev); > +if (!single) { > +vfio_pci_pre_reset(vdev); > +} > vdev->vbasedev.needs_reset = false; > > info = g_malloc0(sizeof(*info)); > @@ -2088,7 +2090,9 @@ out: > } > } > out_single: > -vfio_pci_post_reset(vdev); > +if (!single) { > +vfio_pci_post_reset(vdev); > +} > g_free(info); > > return ret; Looks ok to me, I'll queue it. Thanks, Alex