On Wed, Jan 22, 2020 at 11:07:58AM +0800, Yan Zhao wrote: > On Wed, Jan 22, 2020 at 06:10:38AM +0800, Yan Zhao wrote: > > On Wed, Jan 22, 2020 at 12:51:16AM +0800, Alex Williamson wrote: > > > On Tue, 21 Jan 2020 03:12:07 -0500 > > > Yan Zhao <yan.y.zhao@xxxxxxxxx> wrote: > > > > > > > On Tue, Jan 21, 2020 at 04:01:57AM +0800, Alex Williamson wrote: > > > > > On Sun, 19 Jan 2020 05:06:37 -0500 > > > > > Yan Zhao <yan.y.zhao@xxxxxxxxx> wrote: > > > > > > > > > > > On Thu, Jan 16, 2020 at 11:37:29PM +0800, Alex Williamson wrote: > > > > > > > On Thu, 16 Jan 2020 00:49:41 -0500 > > > > > > > Yan Zhao <yan.y.zhao@xxxxxxxxx> wrote: > > > > > > > > > > > > > > > On Thu, Jan 16, 2020 at 04:06:51AM +0800, Alex Williamson wrote: > > > > > > > > > On Tue, 14 Jan 2020 22:54:55 -0500 > > > > > > > > > Yan Zhao <yan.y.zhao@xxxxxxxxx> wrote: > > > > > > > > > > > > > > > > > > > As a device model, it is better to read/write guest memory using vfio > > > > > > > > > > interface, so that vfio is able to maintain dirty info of device IOVAs. > > > > > > > > > > > > > > > > > > > > Compared to kvm interfaces kvm_read/write_guest(), vfio_dma_rw() has ~600 > > > > > > > > > > cycles more overhead on average. > > > > > > > > > > > > > > > > > > > > ------------------------------------- > > > > > > > > > > | interface | avg cpu cycles | > > > > > > > > > > |-----------------------------------| > > > > > > > > > > | kvm_write_guest | 1554 | > > > > > > > > > > | ----------------------------------| > > > > > > > > > > | kvm_read_guest | 707 | > > > > > > > > > > |-----------------------------------| > > > > > > > > > > | vfio_dma_rw(w) | 2274 | > > > > > > > > > > |-----------------------------------| > > > > > > > > > > | vfio_dma_rw(r) | 1378 | > > > > > > > > > > ------------------------------------- > > > > > > > > > > > > > > > > > > In v1 you had: > > > > > > > > > > > > > > > > > > ------------------------------------- > > > > > > > > > | interface | avg cpu cycles | > > > > > > > > > |-----------------------------------| > > > > > > > > > | kvm_write_guest | 1546 | > > > > > > > > > | ----------------------------------| > > > > > > > > > | kvm_read_guest | 686 | > > > > > > > > > |-----------------------------------| > > > > > > > > > | vfio_iova_rw(w) | 2233 | > > > > > > > > > |-----------------------------------| > > > > > > > > > | vfio_iova_rw(r) | 1262 | > > > > > > > > > ------------------------------------- > > > > > > > > > > > > > > > > > > So the kvm numbers remained within +0.5-3% while the vfio numbers are > > > > > > > > > now +1.8-9.2%. I would have expected the algorithm change to at least > > > > > > > > > not be worse for small accesses and be better for accesses crossing > > > > > > > > > page boundaries. Do you know what happened? > > > > > > > > > > > > > > > > > I only tested the 4 interfaces in GVT's environment, where most of the > > > > > > > > guest memory accesses are less than one page. > > > > > > > > And the different fluctuations should be caused by the locks. > > > > > > > > vfio_dma_rw contends locks with other vfio accesses which are assumed to > > > > > > > > be abundant in the case of GVT. > > > > > > > > > > > > > > Hmm, so maybe it's time to convert vfio_iommu.lock from a mutex to a > > > > > > > rwsem? Thanks, > > > > > > > > > > > > > > > > > > > hi Alex > > > > > > I tested your rwsem patches at (https://lkml.org/lkml/2020/1/16/1869). > > > > > > They works without any runtime error at my side. :) > > > > > > However, I found out that the previous fluctuation may be because I didn't > > > > > > take read/write counts in to account. > > > > > > For example. though the two tests have different avg read/write cycles, > > > > > > their average cycles are almost the same. > > > > > > ______________________________________________________________________ > > > > > > | | avg read | | avg write | | | > > > > > > | | cycles | read cnt | cycles | write cnt | avg cycles | > > > > > > |----------------------------------------------------------------------| > > > > > > | test 1 | 1339 | 29,587,120 | 2258 | 17,098,364 | 1676 | > > > > > > | test 2 | 1340 | 28,454,262 | 2238 | 16,501,788 | 1670 | > > > > > > ---------------------------------------------------------------------- > > > > > > > > > > > > After measuring the exact read/write cnt and cycles of a specific workload, > > > > > > I get below findings: > > > > > > > > > > > > (1) with single VM running glmark2 inside. > > > > > > glmark2: 40M+ read+write cnt, among which 63% is read. > > > > > > among reads, 48% is of PAGE_SIZE, the rest is less than a page. > > > > > > among writes, 100% is less than a page. > > > > > > > > > > > > __________________________________________________ > > > > > > | cycles | read | write | avg | inc | > > > > > > |--------------------------------------------------| > > > > > > | kvm_read/write_page | 694 | 1506 | 993 | / | > > > > > > |--------------------------------------------------| > > > > > > | vfio_dma_rw(mutex) | 1340 | 2248 | 1673 | 680 | > > > > > > |--------------------------------------------------| > > > > > > | vfio_dma_rw(rwsem r) | 1323 | 2198 | 1645 | 653 | > > > > > > --------------------------------------------------- > > > > > > > > > > > > so vfio_dma_rw generally has 650+ more cycles per each read/write. > > > > > > While kvm->srcu is of 160 cycles on average with one vm is running, the > > > > > > cycles spending on locks for vfio_dma_rw spread like this: > > > > > > ___________________________ > > > > > > | cycles | avg | > > > > > > |---------------------------| > > > > > > | iommu->lock | 117 | > > > > > > |---------------------------| > > > > > > | vfio.group_lock | 108 | > > > > > > |---------------------------| > > > > > > | group->unbound_lock | 114 | > > > > > > |---------------------------| > > > > > > | group->device_lock | 115 | > > > > > > |---------------------------| > > > > > > | group->mutex | 113 | > > > > > > --------------------------- > > > > > > > > > > > > I measured the cycles for a mutex without any contention is 104 cycles > > > > > > on average (including time for get_cycles() and measured in the same way > > > > > > as other locks). So the contention of a single lock in a single vm > > > > > > environment is light. probably because there's a vgpu lock hold in GVT already. > > > > > > > > > > > > (2) with two VMs each running glmark2 inside. > > > > > > The contention increases a little. > > > > > > > > > > > > ___________________________________________________ > > > > > > | cycles | read | write | avg | inc | > > > > > > |---------------------------------------------------| > > > > > > | kvm_read/write_page | 1035 | 1832 | 1325 | / | > > > > > > |---------------------------------------------------| > > > > > > | vfio_dma_rw(mutex) | 2104 | 2886 | 2390 | 1065 | > > > > > > |---------------------------------------------------| > > > > > > | vfio_dma_rw(rwsem r) | 1965 | 2778 | 2260 | 935 | > > > > > > --------------------------------------------------- > > > > > > > > > > > > > > > > > > ----------------------------------------------- > > > > > > | avg cycles | one VM | two VMs | > > > > > > |-----------------------------------------------| > > > > > > | iommu lock (mutex) | 117 | 150 | > > > > > > |-----------------------------------|-----------| > > > > > > | iommu lock (rwsem r) | 117 | 156 | > > > > > > |-----------------------------------|-----------| > > > > > > | kvm->srcu | 160 | 213 | > > > > > > ----------------------------------------------- > > > > > > > > > > > > In the kvm case, avg cycles increased 332 cycles, while kvm->srcu only costed > > > > > > 213 cycles. The rest 109 cycles may be spent on atomic operations. > > > > > > But I didn't measure them, as get_cycles() operation itself would influence final > > > > > > cycles by ~20 cycles. > > > > > > > > > > It seems like we need to extend the vfio external user interface so > > > > > that GVT-g can hold the group and container user references across > > > > > multiple calls. For instance if we had a > > > > > vfio_group_get_external_user_from_dev() (based on > > > > > vfio_group_get_external_user()) then i915 could get an opaque > > > > > vfio_group pointer which it could use to call vfio_group_dma_rw() which > > > > > would leave us with only the iommu rw_sem locking. i915 would release > > > > > the reference with vfio_group_put_external_user() when the device is > > > > > released. The same could be done with the pin pages interface to > > > > > streamline that as well. Thoughts? Thanks, > > > > > > > > > hi Alex, > > > > it works! > > > > > > Hurrah! > > > > > > > now the average vfio_dma_rw cycles can reduced to 1198. > > > > one thing I want to propose is that, in sight of dma->task is always user > > > > space process, instead of calling get_task_mm(dma->task), can we just use > > > > "mmget_not_zero(dma->task->mm)"? in this way, the avg cycles can > > > > further reduce to 1051. > > > > > > I'm not an expert there. As noted in the type1 code we hold a > > > reference to the task because it's not advised to hold a long term > > > reference to the mm, so do we know we can look at task->mm without > > > acquiring task_lock()? It's possible this is safe, but it's not > > > abundantly obvious to me. Please research further and provide > > > justification if you think it's correct. Thanks, > > > > > in get_task_mm, > > struct mm_struct *get_task_mm(struct task_struct *task) > > { > > struct mm_struct *mm; > > > > task_lock(task); > > mm = task->mm; > > if (mm) { > > if (task->flags & PF_KTHREAD) > > mm = NULL; > > else > > mmget(mm); > > } > > task_unlock(task); > > return mm; > > } > > task lock is hold only during the call, so the purpose of it is to > > ensure task->flags and task->mm is not changed or gone before mmget(mm) > > or function return. > > so, if we know for sure the task always has no flag PF_THREAD, > > then we only need to ensure mm is not gone before mmget(mm) is done. > > > > static inline void mmget(struct mm_struct *mm) > > { > > atomic_inc(&mm->mm_users); > > } > > > > static inline bool mmget_not_zero(struct mm_struct *mm) > > { > > return atomic_inc_not_zero(&mm->mm_users); > > } > > > > the atomic_inc_not_zero() in mmget_not_zero can ensure mm is not gone > > before its ref count inc. > > > > So, I think the only thing we need to make sure is dma->task is not a > > kernel thread. > > Do you think I can make this assumption? > > > hi Alex > Maybe I can still test PF_KTHREAD without holding task_lock > (task->alloc_lock), as it is only used to protect > "->fs, ->files, ->mm, ->group_info, ->comm, keyring > subscriptions and synchronises with wait4(). Also used in procfs. Also > pins the final release of task.io_context. Also protects ->cpuset and > ->cgroup.subsys[]. And ->vfork_done." > > I checked elsewhere in kernel, e.g. > try_to_wake_up > |->select_task_rq > |->is_per_cpu_kthread > |->if (!(p->flags & PF_KTHREAD)) > task->alloc_lock is not hold there. > > So, I would replace get_task_mm(dma->task) into two steps: > (1) check dma->task->flags & PF_KTHREAD, and (2) mmget_not_zero(mm). > > I'll do more tests and send out new patches after Chinese new year. > > Thanks > Yan > hi Alex after a second thought, I find out that the task->alloc_lock cannot be dropped, as there are chances that dma->task->mm is set to null between checking "dma->task->mm != NULL" and "mmget_not_zero(dma->task->mm)". The chances of this condition can be increased by adding a msleep in-between the two lines of code to mimic possible task schedule. Then it was proved that dma->task->mm could be set to NULL in exit_mm() by killing QEMU. So I have to keep the call to get_task_mm(). :) Thanks Yan
