Re: [PATCH 05/24] kvm: x86/mmu: Fix yielding in TDP MMU

From: Sean Christopherson
Date: Wed Jan 20 2021 - 14:33:35 EST

On Tue, Jan 12, 2021, Ben Gardon wrote:
> There are two problems with the way the TDP MMU yields in long running
> functions. 1.) Given certain conditions, the function may not yield
> reliably / frequently enough. 2.) In some functions the TDP iter risks
> not making forward progress if two threads livelock yielding to
> one another.
> Case 1 is possible if for example, a paging structure was very large
> but had few, if any writable entries. wrprot_gfn_range could traverse many
> entries before finding a writable entry and yielding.
> Case 2 is possible if two threads were trying to execute wrprot_gfn_range.
> Each could write protect an entry and then yield. This would reset the
> tdp_iter's walk over the paging structure and the loop would end up
> repeating the same entry over and over, preventing either thread from
> making forward progress.
> Fix these issues by moving the yield to the beginning of the loop,
> before other checks and only yielding if the loop has made forward
> progress since the last yield.

I think it'd be best to split this into two patches, e.g. ensure forward
progress and then yield more agressively. They are two separate bugs, and I
don't think that ensuring forward progress would exacerbate case #1. I'm not
worried about breaking things so much as getting more helpful shortlogs; "Fix
yielding in TDP MMU" doesn't provide any insight into what exactly was broken.
E.g. something like:

KVM: x86/mmu: Ensure forward progress when yielding in TDP MMU iter
KVM: x86/mmu: Yield in TDU MMU iter even if no real work was done

> Fixes: a6a0b05da9f3 ("kvm: x86/mmu: Support dirty logging for the TDP MMU")
> Reviewed-by: Peter Feiner <pfeiner@xxxxxxxxxx>
> Signed-off-by: Ben Gardon <bgardon@xxxxxxxxxx>
> ---
> arch/x86/kvm/mmu/tdp_mmu.c | 83 +++++++++++++++++++++++++++++++-------
> 1 file changed, 69 insertions(+), 14 deletions(-)
> diff --git a/arch/x86/kvm/mmu/tdp_mmu.c b/arch/x86/kvm/mmu/tdp_mmu.c
> index b2784514ca2d..1987da0da66e 100644
> --- a/arch/x86/kvm/mmu/tdp_mmu.c
> +++ b/arch/x86/kvm/mmu/tdp_mmu.c
> @@ -470,9 +470,23 @@ static bool zap_gfn_range(struct kvm *kvm, struct kvm_mmu_page *root,
> gfn_t start, gfn_t end, bool can_yield)
> {
> struct tdp_iter iter;
> + gfn_t last_goal_gfn = start;
> bool flush_needed = false;
> tdp_root_for_each_pte(iter, root, start, end) {
> + /* Ensure forward progress has been made before yielding. */
> + if (can_yield && iter.goal_gfn != last_goal_gfn &&

Make last_goal_gfn a property of the iterator, that way all this logic can be
shoved into tdp_mmu_iter_flush_cond_resched(), and the comments about ensuring
forward progress and effectively invalidating/resetting the iterator (the
comment below) can be a function comment, as opposed to being copied everywhere.
E.g. there can be a big scary warning in the function comment stating that the
caller must restart its loop if the helper yielded.

Tangentially related, the name goal_gfn is quite confusing. "goal" and "end"
are synonyms, but "goal" is often initialized with "start", and it's not used to
terminate the walk. Maybe next_gfn instead? And maybe yielded_gfn, since
last_next_gfn is pretty horrendous.

> + tdp_mmu_iter_flush_cond_resched(kvm, &iter)) {

This isn't quite correct, as tdp_mmu_iter_flush_cond_resched() will do an
expensive remote TLB flush on every yield, even if no flush is needed. The
cleanest solution is likely to drop tdp_mmu_iter_flush_cond_resched() and
instead add a @flush param to tdp_mmu_iter_cond_resched(). If it's tagged
__always_inline, then the callers that unconditionally pass true/false will
optimize out the conditional code.

At that point, I think it would also make sense to fold tdp_iter_refresh_walk()
into tdp_mmu_iter_cond_resched(), because really we shouldn't be mucking with
the guts of the iter except for the yield case.

> + last_goal_gfn = iter.goal_gfn;

Another argument for both renaming goal_gfn and moving last_*_gfn into the iter:
it's not at all obvious that updating the last gfn _after_ tdp_iter_refresh_walk()
is indeed correct.

You can also avoid a local variable by doing max(iter->next_gfn, iter->gfn) when
calling tdp_iter_refresh_walk(). IMO, that's also a bit easier to understand
than an open-coded equivalent.

E.g. putting it all together, with yielded_gfn set by tdp_iter_start():

static __always_inline bool tdp_mmu_iter_cond_resched(struct kvm *kvm,
struct tdp_iter *iter,
bool flush)
/* Ensure forward progress has been made since the last yield. */
if (iter->next_gfn == iter->yielded_gfn)
return false;

if (need_resched() || spin_needbreak(&kvm->mmu_lock)) {
if (flush)

* Restart the walk over the paging structure from the root,
* starting from the highest gfn the iterator had previously
* reached. The entire paging structure, except the root, may
* have been completely torn down and rebuilt while we yielded.
tdp_iter_start(iter, iter->pt_path[iter->root_level - 1],
iter->root_level, iter->min_level,
max(iter->next_gfn, iter->gfn));
return true;

return false;

> + flush_needed = false;
> + /*
> + * Yielding caused the paging structure walk to be
> + * reset so skip to the next iteration to continue the
> + * walk from the root.
> + */
> + continue;
> + }
> +
> if (!is_shadow_present_pte(iter.old_spte))
> continue;