On Tue, 18 Apr 2017 01:02:12 +0530
Kirti Wankhede <kwankh...@nvidia.com> wrote:
> On 4/18/2017 12:49 AM, Alex Williamson wrote:
> > On Tue, 18 Apr 2017 00:35:06 +0530
> > Kirti Wankhede <kwankh...@nvidia.com> wrote:
> >
> >> On 4/17/2017 8:02 PM, Alex Williamson wrote:
> >>> On Mon, 17 Apr 2017 14:47:54 +0800
> >>> Peter Xu <pet...@redhat.com> wrote:
> >>>
> >>>> On Sun, Apr 16, 2017 at 07:42:27PM -0600, Alex Williamson wrote:
> >>>>
> >>>> [...]
> >>>>
> >>>>> -static void vfio_lock_acct(struct task_struct *task, long npage)
> >>>>> +static int vfio_lock_acct(struct task_struct *task, long npage, bool
> >>>>> lock_cap)
> >>>>> {
> >>>>> - struct vwork *vwork;
> >>>>> struct mm_struct *mm;
> >>>>> bool is_current;
> >>>>> + int ret;
> >>>>>
> >>>>> if (!npage)
> >>>>> - return;
> >>>>> + return 0;
> >>>>>
> >>>>> is_current = (task->mm == current->mm);
> >>>>>
> >>>>> mm = is_current ? task->mm : get_task_mm(task);
> >>>>> if (!mm)
> >>>>> - return; /* process exited */
> >>>>> + return -ESRCH; /* process exited */
> >>>>>
> >>>>> - if (down_write_trylock(&mm->mmap_sem)) {
> >>>>> - mm->locked_vm += npage;
> >>>>> - up_write(&mm->mmap_sem);
> >>>>> - if (!is_current)
> >>>>> - mmput(mm);
> >>>>> - return;
> >>>>> - }
> >>>>> + ret = down_write_killable(&mm->mmap_sem);
> >>>>> + if (!ret) {
> >>>>> + if (npage < 0 || lock_cap) {
> >>>>
> >>>> Nit: maybe we can avoid passing in lock_cap in all the callers of
> >>>> vfio_lock_acct() and fetch it via has_capability() only if npage < 0?
> >>>> IMHO that'll keep the vfio_lock_acct() interface cleaner, and we won't
> >>>> need to pass in "false" any time when doing unpins.
> >>>
> >>> Unfortunately vfio_pin_pages_remote() needs to know about lock_cap
> >>> since it tests whether the user is exceeding their locked memory
> >>> limit. The other callers could certainly get away with
> >>> vfio_lock_acct() testing the capability itself but that would add a
> >>> redundant call for the most common user. I'm not a big fan of passing
> >>> a lock_cap bool either, but it seemed the best fix for now. The
> >>> cleanest alternative I can up with is this (untested):
> >>>
> >>
> >> In my opinion, passing 'bool lock_cap' looks much clean and simple.
> >>
> >> Reviewed-by: Kirti Wankhede <kwankh...@nvidia.com>
> >
> > Well shoot, I was just starting to warm up to the bool*. I like that
> > we're not presuming the polarity for the callers we expect to be
> > removing pages and I generally just dislike passing fixed bool
> > parameters to change the function behavior. I've cleaned it up a bit
> > further and was starting to do some testing on this which I'd propose
> > for v5. Does it change your opinion?
>
> If passing fixed bool parameter is the concern then I would lean towards
> Peter's suggestion. vfio_pin_pages_remote() will check lock capability
> outside vfio_lock_acct() and again in vfio_lock_acct(). At other places,
> it will be takes care within vfio_lock_acct()
Sorry, I don't see that as a viable option. Testing for CAP_IPC_LOCK in
both vfio_pin_pages_remote() and vfio_lock_acct() results in over a
10% performance hit on the mapping path with a custom micro-benchmark.
In fact, it suggests we should probably pass that from even higher in
the call stack. Thanks,
Alex
> >
> > commit cd61c5f507d614ac14b75b0a548c8738deff88ea
> > Author: Alex Williamson <alex.william...@redhat.com>
> > Date: Thu Apr 13 14:10:15 2017 -0600
> >
> > vfio/type1: Remove locked page accounting workqueue
> >
> > If the mmap_sem is contented then the vfio type1 IOMMU backend will
> > defer locked page accounting updates to a workqueue task. This has a
> > few problems and depending on which side the user tries to play, they
> > might be over-penalized for unmaps that haven't yet been accounted or
> > race the workqueue to enter more mappings than they're allowed. The
> > original intent of this workqueue mechanism seems to be focused on
> > reducing latency through the ioctl, but we cannot do so at the cost
> > of correctness. Remove this workqueue mechanism and update the
> > callers to allow for failure. We can also now recheck the limit under
> > write lock to make sure we don't exceed it.
> >
> > vfio_pin_pages_remote() also now necessarily includes an unwind path
> > which we can jump to directly if the consecutive page pinning finds
> > that we're exceeding the user's memory limits. This avoids the
> > current lazy approach which does accounting and mapping up to the
> > fault, only to return an error on the next iteration to unwind the
> > entire vfio_dma.
> >
> > Cc: sta...@vger.kernel.org
> > Signed-off-by: Alex Williamson <alex.william...@redhat.com>
> >
> > diff --git a/drivers/vfio/vfio_iommu_type1.c
> > b/drivers/vfio/vfio_iommu_type1.c
> > index 32d2633092a3..a8a079ba9477 100644
> > --- a/drivers/vfio/vfio_iommu_type1.c
> > +++ b/drivers/vfio/vfio_iommu_type1.c
> > @@ -246,69 +246,46 @@ static int vfio_iova_put_vfio_pfn(struct vfio_dma
> > *dma, struct vfio_pfn *vpfn)
> > return ret;
> > }
> >
> > -struct vwork {
> > - struct mm_struct *mm;
> > - long npage;
> > - struct work_struct work;
> > -};
> > -
> > -/* delayed decrement/increment for locked_vm */
> > -static void vfio_lock_acct_bg(struct work_struct *work)
> > -{
> > - struct vwork *vwork = container_of(work, struct vwork, work);
> > - struct mm_struct *mm;
> > -
> > - mm = vwork->mm;
> > - down_write(&mm->mmap_sem);
> > - mm->locked_vm += vwork->npage;
> > - up_write(&mm->mmap_sem);
> > - mmput(mm);
> > - kfree(vwork);
> > -}
> > -
> > -static void vfio_lock_acct(struct task_struct *task, long npage)
> > +static int vfio_lock_acct(struct task_struct *task, long npage, bool
> > *lock_cap)
> > {
> > - struct vwork *vwork;
> > struct mm_struct *mm;
> > bool is_current;
> > + int ret;
> >
> > if (!npage)
> > - return;
> > + return 0;
> >
> > is_current = (task->mm == current->mm);
> >
> > mm = is_current ? task->mm : get_task_mm(task);
> > if (!mm)
> > - return; /* process exited */
> > + return -ESRCH; /* process exited */
> >
> > - if (down_write_trylock(&mm->mmap_sem)) {
> > - mm->locked_vm += npage;
> > - up_write(&mm->mmap_sem);
> > - if (!is_current)
> > - mmput(mm);
> > - return;
> > - }
> > + ret = down_write_killable(&mm->mmap_sem);
> > + if (!ret) {
> > + if (npage > 0) {
> > + if (lock_cap ? !*lock_cap :
> > + !has_capability(task, CAP_IPC_LOCK)) {
> > + unsigned long limit;
> > +
> > + limit = task_rlimit(task,
> > + RLIMIT_MEMLOCK) >> PAGE_SHIFT;
> > +
> > + if (mm->locked_vm + npage > limit)
> > + ret = -ENOMEM;
> > + }
> > + }
> > +
> > + if (!ret)
> > + mm->locked_vm += npage;
> >
> > - if (is_current) {
> > - mm = get_task_mm(task);
> > - if (!mm)
> > - return;
> > + up_write(&mm->mmap_sem);
> > }
> >
> > - /*
> > - * Couldn't get mmap_sem lock, so must setup to update
> > - * mm->locked_vm later. If locked_vm were atomic, we
> > - * wouldn't need this silliness
> > - */
> > - vwork = kmalloc(sizeof(struct vwork), GFP_KERNEL);
> > - if (WARN_ON(!vwork)) {
> > + if (!is_current)
> > mmput(mm);
> > - return;
> > - }
> > - INIT_WORK(&vwork->work, vfio_lock_acct_bg);
> > - vwork->mm = mm;
> > - vwork->npage = npage;
> > - schedule_work(&vwork->work);
> > +
> > + return ret;
> > }
> >
> > /*
> > @@ -405,7 +382,7 @@ static int vaddr_get_pfn(struct mm_struct *mm, unsigned
> > long vaddr,
> > static long vfio_pin_pages_remote(struct vfio_dma *dma, unsigned long
> > vaddr,
> > long npage, unsigned long *pfn_base)
> > {
> > - unsigned long limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
> > + unsigned long pfn = 0, limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
> > bool lock_cap = capable(CAP_IPC_LOCK);
> > long ret, pinned = 0, lock_acct = 0;
> > bool rsvd;
> > @@ -442,8 +419,6 @@ static long vfio_pin_pages_remote(struct vfio_dma *dma,
> > unsigned long vaddr,
> > /* Lock all the consecutive pages from pfn_base */
> > for (vaddr += PAGE_SIZE, iova += PAGE_SIZE; pinned < npage;
> > pinned++, vaddr += PAGE_SIZE, iova += PAGE_SIZE) {
> > - unsigned long pfn = 0;
> > -
> > ret = vaddr_get_pfn(current->mm, vaddr, dma->prot, &pfn);
> > if (ret)
> > break;
> > @@ -460,14 +435,25 @@ static long vfio_pin_pages_remote(struct vfio_dma
> > *dma, unsigned long vaddr,
> > put_pfn(pfn, dma->prot);
> > pr_warn("%s: RLIMIT_MEMLOCK (%ld) exceeded\n",
> > __func__, limit << PAGE_SHIFT);
> > - break;
> > + ret = -ENOMEM;
> > + goto unpin_out;
> > }
> > lock_acct++;
> > }
> > }
> >
> > out:
> > - vfio_lock_acct(current, lock_acct);
> > + ret = vfio_lock_acct(current, lock_acct, &lock_cap);
> > +
> > +unpin_out:
> > + if (ret) {
> > + if (!rsvd) {
> > + for (pfn = *pfn_base ; pinned ; pfn++, pinned--)
> > + put_pfn(pfn, dma->prot);
> > + }
> > +
> > + return ret;
> > + }
> >
> > return pinned;
> > }
> > @@ -488,7 +474,7 @@ static long vfio_unpin_pages_remote(struct vfio_dma
> > *dma, dma_addr_t iova,
> > }
> >
> > if (do_accounting)
> > - vfio_lock_acct(dma->task, locked - unlocked);
> > + vfio_lock_acct(dma->task, locked - unlocked, NULL);
> >
> > return unlocked;
> > }
> > @@ -522,8 +508,14 @@ static int vfio_pin_page_external(struct vfio_dma
> > *dma, unsigned long vaddr,
> > goto pin_page_exit;
> > }
> >
> > - if (!rsvd && do_accounting)
> > - vfio_lock_acct(dma->task, 1);
> > + if (!rsvd && do_accounting) {
> > + ret = vfio_lock_acct(dma->task, 1, &lock_cap);
> > + if (ret) {
> > + put_pfn(*pfn_base, dma->prot);
> > + goto pin_page_exit;
> > + }
> > + }
> > +
> > ret = 1;
> >
> > pin_page_exit:
> > @@ -543,7 +535,7 @@ static int vfio_unpin_page_external(struct vfio_dma
> > *dma, dma_addr_t iova,
> > unlocked = vfio_iova_put_vfio_pfn(dma, vpfn);
> >
> > if (do_accounting)
> > - vfio_lock_acct(dma->task, -unlocked);
> > + vfio_lock_acct(dma->task, -unlocked, NULL);
> >
> > return unlocked;
> > }
> > @@ -740,7 +732,7 @@ static long vfio_unmap_unpin(struct vfio_iommu *iommu,
> > struct vfio_dma *dma,
> >
> > dma->iommu_mapped = false;
> > if (do_accounting) {
> > - vfio_lock_acct(dma->task, -unlocked);
> > + vfio_lock_acct(dma->task, -unlocked, NULL);
> > return 0;
> > }
> > return unlocked;
> > @@ -1382,7 +1374,7 @@ static void vfio_iommu_unmap_unpin_reaccount(struct
> > vfio_iommu *iommu)
> > if (!is_invalid_reserved_pfn(vpfn->pfn))
> > locked++;
> > }
> > - vfio_lock_acct(dma->task, locked - unlocked);
> > + vfio_lock_acct(dma->task, locked - unlocked, NULL);
> > }
> > }
> >
> >
> > Patch 2/2 would clearly change the &lock_cap in
> > vfio_pin_page_external() to a NULL, so only _remote passes a pointer
> > there. Thanks,
> >
> > Alex
> >
