On 2019/3/8 下午10:58, Jerome Glisse wrote:
On Fri, Mar 08, 2019 at 04:50:36PM +0800, Jason Wang wrote:
On 2019/3/8 上午3:16, Andrea Arcangeli wrote:
On Thu, Mar 07, 2019 at 12:56:45PM -0500, Michael S. Tsirkin wrote:
On Thu, Mar 07, 2019 at 10:47:22AM -0500, Michael S. Tsirkin wrote:
On Wed, Mar 06, 2019 at 02:18:12AM -0500, Jason Wang wrote:
+static const struct mmu_notifier_ops vhost_mmu_notifier_ops = {
+ .invalidate_range = vhost_invalidate_range,
+};
+
void vhost_dev_init(struct vhost_dev *dev,
struct vhost_virtqueue **vqs, int nvqs, int iov_limit)
{
I also wonder here: when page is write protected then
it does not look like .invalidate_range is invoked.
E.g. mm/ksm.c calls
mmu_notifier_invalidate_range_start and
mmu_notifier_invalidate_range_end but not mmu_notifier_invalidate_range.
Similarly, rmap in page_mkclean_one will not call
mmu_notifier_invalidate_range.
If I'm right vhost won't get notified when page is write-protected since you
didn't install start/end notifiers. Note that end notifier can be called
with page locked, so it's not as straight-forward as just adding a call.
Writing into a write-protected page isn't a good idea.
Note that documentation says:
it is fine to delay the mmu_notifier_invalidate_range
call to mmu_notifier_invalidate_range_end() outside the page table lock.
implying it's called just later.
OK I missed the fact that _end actually calls
mmu_notifier_invalidate_range internally. So that part is fine but the
fact that you are trying to take page lock under VQ mutex and take same
mutex within notifier probably means it's broken for ksm and rmap at
least since these call invalidate with lock taken.
Yes this lock inversion needs more thoughts.
And generally, Andrea told me offline one can not take mutex under
the notifier callback. I CC'd Andrea for why.
Yes, the problem then is the ->invalidate_page is called then under PT
lock so it cannot take mutex, you also cannot take the page_lock, it
can at most take a spinlock or trylock_page.
So it must switch back to the _start/_end methods unless you rewrite
the locking.
The difference with _start/_end, is that ->invalidate_range avoids the
_start callback basically, but to avoid the _start callback safely, it
has to be called in between the ptep_clear_flush and the set_pte_at
whenever the pfn changes like during a COW. So it cannot be coalesced
in a single TLB flush that invalidates all sptes in a range like we
prefer for performance reasons for example in KVM. It also cannot
sleep.
In short ->invalidate_range must be really fast (it shouldn't require
to send IPI to all other CPUs like KVM may require during an
invalidate_range_start) and it must not sleep, in order to prefer it
to _start/_end.
I.e. the invalidate of the secondary MMU that walks the linux
pagetables in hardware (in vhost case with GUP in software) has to
happen while the linux pagetable is zero, otherwise a concurrent
hardware pagetable lookup could re-instantiate a mapping to the old
page in between the set_pte_at and the invalidate_range_end (which
internally calls ->invalidate_range). Jerome documented it nicely in
Documentation/vm/mmu_notifier.rst .
Right, I've actually gone through this several times but some details were
missed by me obviously.
Now you don't really walk the pagetable in hardware in vhost, but if
you use gup_fast after usemm() it's similar.
For vhost the invalidate would be really fast, there are no IPI to
deliver at all, the problem is just the mutex.
Yes. A possible solution is to introduce a valid flag for VA. Vhost may only
try to access kernel VA when it was valid. Invalidate_range_start() will
clear this under the protection of the vq mutex when it can block. Then
invalidate_range_end() then can clear this flag. An issue is blockable is
always false for range_end().
Note that there can be multiple asynchronous concurrent invalidate_range
callbacks. So a flag does not work but a counter of number of active
invalidation would. See how KSM is doing it for instance in kvm_main.c
The pattern for this kind of thing is:
my_invalidate_range_start(start,end) {
...
if (mystruct_overlap(mystruct, start, end)) {
mystruct_lock();
mystruct->invalidate_count++;
...
mystruct_unlock();
}
}
my_invalidate_range_end(start,end) {
...
if (mystruct_overlap(mystruct, start, end)) {
mystruct_lock();
mystruct->invalidate_count--;
...
mystruct_unlock();
}
}
my_access_va(mystruct) {
again:
wait_on(!mystruct->invalidate_count)
mystruct_lock();
if (mystruct->invalidate_count) {
mystruct_unlock();
goto again;
}
GUP();
...
mystruct_unlock();
}
Cheers,
Jérôme
Yes, this should work.
Thanks