On Mon, Dec 21, 2020 at 02:55:12PM -0800, Nadav Amit wrote:
> wouldn’t mmap_write_downgrade() be executed before mprotect_fixup() (so
I assume you mean "in" mprotect_fixup, after change_protection.
If you would downgrade the mmap_lock to read there, then it'd severely
slowdown the non contention case, if there's more than vma that needs
change_protection.
You'd need to throw away the prev->vm_next info and you'd need to do a
new find_vma after droping the mmap_lock for reading and re-taking the
mmap_lock for writing at every iteration of the loop.
To do less harm to the non-contention case you could perhaps walk
vma->vm_next and check if it's outside the mprotect range and only
downgrade in such case. So let's assume we intend to optimize with
mmap_write_downgrade only the last vma.
The problem is once you had to take mmap_lock for writing, you already
stalled for I/O and waited all concurrent page faults and blocked them
as well for the vma allocations in split_vma, so that extra boost in
SMP scalability you get is lost in the noise there at best.
And the risk is that at worst that extra locked op of
mmap_write_downgrade() will hurt SMP scalability because it would
increase the locked ops of mprotect on the hottest false-shared
cacheline by 50% and that may outweight the benefit from unblocking
the page faults half a usec sooner on large systems.
But the ultimate reason why mprotect cannot do mmap_write_downgrade()
while userfaultfd_writeprotect can do mmap_read_lock and avoid the
mmap_write_lock altogether, is that mprotect leaves no mark in the
pte/hugepmd that allows to detect when the TLB is stale in order to
redirect the page fault in a dead end (handle_userfault() or
do_numa_page) until after the TLB has been flushed as it happens in
the the 4 cases below:
/*
* STALE_TLB_WARNING: while the uffd_wp bit is set, the TLB
* can be stale. We cannot allow do_wp_page to proceed or
* it'll wrongly assume that nobody can still be writing to
* the page if !pte_write.
*/
if (userfaultfd_pte_wp(vma, *vmf->pte)) {
/*
* STALE_TLB_WARNING: while the uffd_wp bit is set,
* the TLB can be stale. We cannot allow wp_huge_pmd()
* to proceed or it'll wrongly assume that nobody can
* still be writing to the page if !pmd_write.
*/
if (userfaultfd_huge_pmd_wp(vmf->vma, orig_pmd))
/*
* STALE_TLB_WARNING: if the pte is NUMA protnone the TLB can
* be stale.
*/
if (pte_protnone(vmf->orig_pte) && vma_is_accessible(vmf->vma))
/*
* STALE_TLB_WARNING: if the pmd is NUMA
* protnone the TLB can be stale.
*/
if (pmd_protnone(orig_pmd) && vma_is_accessible(vma))
Thanks,
Andrea
