KVM Archive on lore.kernel.org
 help / color / Atom feed
From: Peter Xu <peterx@redhat.com>
To: kvm@vger.kernel.org, linux-kernel@vger.kernel.org
Cc: Paolo Bonzini <pbonzini@redhat.com>,
	peterx@redhat.com, Sean Christopherson <seanjc@google.com>,
	Andrew Jones <drjones@redhat.com>
Subject: [PATCH v4 1/2] KVM: selftests: Sync data verify of dirty logging with guest sync
Date: Tue, 20 Apr 2021 11:39:28 -0400
Message-ID: <20210420153929.482810-2-peterx@redhat.com> (raw)
In-Reply-To: <20210420153929.482810-1-peterx@redhat.com>

This fixes a bug that can trigger with e.g. "taskset -c 0 ./dirty_log_test" or
when the testing host is very busy.

A similar previous attempt is done [1] but that is not enough, the reason is
stated in the reply [2].

As a summary (partly quotting from [2]):

The problem is I think one guest memory write operation (of this specific test)
contains a few micro-steps when page is during kvm dirty tracking (here I'm
only considering write-protect rather than pml but pml should be similar at
least when the log buffer is full):

  (1) Guest read 'iteration' number into register, prepare to write, page fault
  (2) Set dirty bit in either dirty bitmap or dirty ring
  (3) Return to guest, data written

When we verify the data, we assumed that all these steps are "atomic", say,
when (1) happened for this page, we assume (2) & (3) must have happened.  We
had some trick to workaround "un-atomicity" of above three steps, as previous
version of this patch wanted to fix atomicity of step (2)+(3) by explicitly
letting the main thread wait for at least one vmenter of vcpu thread, which
should work.  However what I overlooked is probably that we still have race
when (1) and (2) can be interrupted.

One example calltrace when it could happen that we read an old interation, got
interrupted before even setting the dirty bit and flushing data:


It means iteration number cached in vcpu register can be very old when dirty
bit set and data flushed.

So far I don't see an easy way to guarantee all steps 1-3 atomicity but to sync
at the GUEST_SYNC() point of guest code when we do verification of the dirty
bits as what this patch does.

[1] https://lore.kernel.org/lkml/20210413213641.23742-1-peterx@redhat.com/
[2] https://lore.kernel.org/lkml/20210417140956.GV4440@xz-x1/

Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Sean Christopherson <seanjc@google.com>
Cc: Andrew Jones <drjones@redhat.com>
Signed-off-by: Peter Xu <peterx@redhat.com>
 tools/testing/selftests/kvm/dirty_log_test.c | 62 ++++++++++++++++----
 1 file changed, 51 insertions(+), 11 deletions(-)

diff --git a/tools/testing/selftests/kvm/dirty_log_test.c b/tools/testing/selftests/kvm/dirty_log_test.c
index bb2752d78fe3..fc87e2f11d3d 100644
--- a/tools/testing/selftests/kvm/dirty_log_test.c
+++ b/tools/testing/selftests/kvm/dirty_log_test.c
@@ -17,6 +17,7 @@
 #include <linux/bitmap.h>
 #include <linux/bitops.h>
 #include <asm/barrier.h>
+#include <linux/atomic.h>
 #include "kvm_util.h"
 #include "test_util.h"
@@ -137,12 +138,20 @@ static uint64_t host_clear_count;
 static uint64_t host_track_next_count;
 /* Whether dirty ring reset is requested, or finished */
-static sem_t dirty_ring_vcpu_stop;
-static sem_t dirty_ring_vcpu_cont;
+static sem_t sem_vcpu_stop;
+static sem_t sem_vcpu_cont;
+ * This is only set by main thread, and only cleared by vcpu thread.  It is
+ * used to request vcpu thread to stop at the next GUEST_SYNC, since GUEST_SYNC
+ * is the only place that we'll guarantee both "dirty bit" and "dirty data"
+ * will match.  E.g., SIG_IPI won't guarantee that (e.g., when vcpu interrupted
+ * after setting dirty bit but before data flushed).
+ */
+static atomic_t vcpu_sync_stop_requested;
  * This is updated by the vcpu thread to tell the host whether it's a
  * ring-full event.  It should only be read until a sem_wait() of
- * dirty_ring_vcpu_stop and before vcpu continues to run.
+ * sem_vcpu_stop and before vcpu continues to run.
 static bool dirty_ring_vcpu_ring_full;
@@ -234,6 +243,17 @@ static void clear_log_collect_dirty_pages(struct kvm_vm *vm, int slot,
 	kvm_vm_clear_dirty_log(vm, slot, bitmap, 0, num_pages);
+/* Should only be called after a GUEST_SYNC */
+static void vcpu_handle_sync_stop(void)
+	if (atomic_read(&vcpu_sync_stop_requested)) {
+		/* It means main thread is sleeping waiting */
+		atomic_set(&vcpu_sync_stop_requested, false);
+		sem_post(&sem_vcpu_stop);
+		sem_wait_until(&sem_vcpu_cont);
+	}
 static void default_after_vcpu_run(struct kvm_vm *vm, int ret, int err)
 	struct kvm_run *run = vcpu_state(vm, VCPU_ID);
@@ -244,6 +264,8 @@ static void default_after_vcpu_run(struct kvm_vm *vm, int ret, int err)
 		    "Invalid guest sync status: exit_reason=%s\n",
+	vcpu_handle_sync_stop();
 static bool dirty_ring_supported(void)
@@ -301,13 +323,13 @@ static void dirty_ring_wait_vcpu(void)
 	/* This makes sure that hardware PML cache flushed */
-	sem_wait_until(&dirty_ring_vcpu_stop);
+	sem_wait_until(&sem_vcpu_stop);
 static void dirty_ring_continue_vcpu(void)
 	pr_info("Notifying vcpu to continue\n");
-	sem_post(&dirty_ring_vcpu_cont);
+	sem_post(&sem_vcpu_cont);
 static void dirty_ring_collect_dirty_pages(struct kvm_vm *vm, int slot,
@@ -355,17 +377,17 @@ static void dirty_ring_after_vcpu_run(struct kvm_vm *vm, int ret, int err)
 	/* A ucall-sync or ring-full event is allowed */
 	if (get_ucall(vm, VCPU_ID, NULL) == UCALL_SYNC) {
 		/* We should allow this to continue */
-		;
+		vcpu_handle_sync_stop();
 	} else if (run->exit_reason == KVM_EXIT_DIRTY_RING_FULL ||
 		   (ret == -1 && err == EINTR)) {
 		/* Update the flag first before pause */
 			   run->exit_reason == KVM_EXIT_DIRTY_RING_FULL);
-		sem_post(&dirty_ring_vcpu_stop);
+		sem_post(&sem_vcpu_stop);
 		pr_info("vcpu stops because %s...\n",
 			dirty_ring_vcpu_ring_full ?
 			"dirty ring is full" : "vcpu is kicked out");
-		sem_wait_until(&dirty_ring_vcpu_cont);
+		sem_wait_until(&sem_vcpu_cont);
 		pr_info("vcpu continues now.\n");
 	} else {
 		TEST_ASSERT(false, "Invalid guest sync status: "
@@ -377,7 +399,7 @@ static void dirty_ring_after_vcpu_run(struct kvm_vm *vm, int ret, int err)
 static void dirty_ring_before_vcpu_join(void)
 	/* Kick another round of vcpu just to make sure it will quit */
-	sem_post(&dirty_ring_vcpu_cont);
+	sem_post(&sem_vcpu_cont);
 struct log_mode {
@@ -768,7 +790,25 @@ static void run_test(enum vm_guest_mode mode, void *arg)
 		usleep(p->interval * 1000);
 		log_mode_collect_dirty_pages(vm, TEST_MEM_SLOT_INDEX,
 					     bmap, host_num_pages);
+		/*
+		 * See vcpu_sync_stop_requested definition for details on why
+		 * we need to stop vcpu when verify data.
+		 */
+		atomic_set(&vcpu_sync_stop_requested, true);
+		sem_wait_until(&sem_vcpu_stop);
+		/*
+		 * NOTE: for dirty ring, it's possible that we didn't stop at
+		 * GUEST_SYNC but instead we stopped because ring is full;
+		 * that's okay too because ring full means we're only missing
+		 * the flush of the last page, and since we handle dirty ring
+		 * last page specifically, so verify will still pass.
+		 */
+		assert(host_log_mode == LOG_MODE_DIRTY_RING ||
+		       atomic_read(&vcpu_sync_stop_requested) == false);
 		vm_dirty_log_verify(mode, bmap);
+		sem_post(&sem_vcpu_cont);
 		sync_global_to_guest(vm, iteration);
@@ -819,8 +859,8 @@ int main(int argc, char *argv[])
 	int opt, i;
-	sem_init(&dirty_ring_vcpu_stop, 0, 0);
-	sem_init(&dirty_ring_vcpu_cont, 0, 0);
+	sem_init(&sem_vcpu_stop, 0, 0);
+	sem_init(&sem_vcpu_cont, 0, 0);

  reply index

Thread overview: 3+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2021-04-20 15:39 [PATCH v4 0/2] KVM: selftests: fix races in dirty log test Peter Xu
2021-04-20 15:39 ` Peter Xu [this message]
2021-04-20 15:39 ` [PATCH v4 2/2] KVM: selftests: Wait for vcpu thread before signal setup Peter Xu

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=20210420153929.482810-2-peterx@redhat.com \
    --to=peterx@redhat.com \
    --cc=drjones@redhat.com \
    --cc=kvm@vger.kernel.org \
    --cc=linux-kernel@vger.kernel.org \
    --cc=pbonzini@redhat.com \
    --cc=seanjc@google.com \


* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

KVM Archive on lore.kernel.org

Archives are clonable:
	git clone --mirror https://lore.kernel.org/kvm/0 kvm/git/0.git

	# If you have public-inbox 1.1+ installed, you may
	# initialize and index your mirror using the following commands:
	public-inbox-init -V2 kvm kvm/ https://lore.kernel.org/kvm \
	public-inbox-index kvm

Example config snippet for mirrors

Newsgroup available over NNTP:

AGPL code for this site: git clone https://public-inbox.org/public-inbox.git