Skip to contentSkip to conversation

credit: fix race condition in csched_vcpu flags

PATCH In review Igor Pavlikevich 2 participants 3 messages Feb 25, 2013 – Feb 26, 2013

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 credit: fix race condition in csched_vcpu flags

When vcpu stops yielding, it's possible to overwrite CSCHED_FLAG_VCPU_PARKED flag after clearing CSCHED_FLAG_VCPU_YIELD.

332.970218393 -------x d1v1 runstate_continue d1v1 running->running
]332.970248371 -------x d1v1 28005(2:8:5) 2 [ 1 1 ] yielding, flags=0x0002
]332.970251968 -------x d1v1 clearing 0x0002, flags=0
332.970252175 -------x d1v1 runstate_continue d1v1 running->running

]332.970282574 -------x d1v1   28005(2:8:5) 2 [ 1 1 ]                  yielding, flags=0x0002
]332.970283068 x------| d32767v0                                       entering csched_acct(), d1v1 flags=0x0002
]332.970285763 x------| d32767v0   28003(2:8:3) 2 [ 1 1 ]              calling vcpu_pause_nosync() for d1v1
]332.970286647 -------x d1v1                                           clearing 0x0002, flags=0
]332.970286771 x------| d32767v0                                       flags |= CSCHED_FLAG_VCPU_PARKED, flags=0x0003
]332.970286990 -------x d1v1   2800e(2:8:e) 2 [ 1 1cacfca ]
]332.970287122 -------x d1v1   2800f(2:8:f) 3 [ 7fff 1cacfca ffffffff ]
]332.970287242 -------x d1v1   2800a(2:8:a) 4 [ 1 1 7fff 7 ]           exiting schedule(), flags=0
 332.970287417 -------x d1v1 runstate_change d1v1 running->offline
 332.970287629 -------x d?v? runstate_change d32767v7 runnable->running
]332.970288397 x------- d32767v0                                       exiting csched_acct(), flags=0x0003
]332.995349690 x------- d32767v0                                       entering csched_acct(), d1v1 flags=0x0000, flag lost.
]332.995352170 x------- d32767v0   28003(2:8:3) 2 [ 1 1 ]              calling vcpu_pause_nosync() for d1v1 again
Quoted 1 line

From this moment d1v1 has extra +1 in pause_count, so it goes offline right after credit>prv->credits_per_tslice forever.

xen/common/sched_credit.c +2 -0
diff -r 66f563be41d9 -r 4d2fb69d362b xen/common/sched_credit.c
@@ -1155,7 +1155,9 @@ csched_acct(void* dummy)
                 {
                     SCHED_STAT_CRANK(vcpu_park);
                     vcpu_pause_nosync(svc->vcpu);
+                    vcpu_schedule_lock_irq(svc->vcpu);
                     svc->flags |= CSCHED_FLAG_VCPU_PARKED;
+                    vcpu_schedule_unlock_irq(svc->vcpu);
                 }
 
                 /* Lower bound on credits */
#1 lore raw
Quoted 32 lines · Igor Pavlikevich

When vcpu stops yielding, it's possible to overwrite CSCHED_FLAG_VCPU_PARKED flag after clearing CSCHED_FLAG_VCPU_YIELD.

332.970218393 -------x d1v1 runstate_continue d1v1 running->running
]332.970248371 -------x d1v1 28005(2:8:5) 2 [ 1 1 ]
yielding, flags=0x0002
]332.970251968 -------x d1v1
clearing 0x0002, flags=0
332.970252175 -------x d1v1 runstate_continue d1v1 running->running
]332.970282574 -------x d1v1 28005(2:8:5) 2 [ 1 1 ]
yielding, flags=0x0002
]332.970283068 x------| d32767v0
entering csched_acct(), d1v1 flags=0x0002
]332.970285763 x------| d32767v0 28003(2:8:3) 2 [ 1 1 ]
calling vcpu_pause_nosync() for d1v1
]332.970286647 -------x d1v1
clearing 0x0002, flags=0
]332.970286771 x------| d32767v0
flags |= CSCHED_FLAG_VCPU_PARKED, flags=0x0003

]332.970286990 -------x d1v1   2800e(2:8:e) 2 [ 1 1cacfca ]
]332.970287122 -------x d1v1   2800f(2:8:f) 3 [ 7fff 1cacfca ffffffff ]
]332.970287242 -------x d1v1   2800a(2:8:a) 4 [ 1 1 7fff 7 ]

exiting schedule(), flags=0
332.970287417 -------x d1v1 runstate_change d1v1 running->offline 332.970287629 -------x d?v? runstate_change d32767v7 runnable->running
]332.970288397 x------- d32767v0
exiting csched_acct(), flags=0x0003
]332.995349690 x------- d32767v0
entering csched_acct(), d1v1 flags=0x0000, flag lost.
]332.995352170 x------- d32767v0 28003(2:8:3) 2 [ 1 1 ]
calling vcpu_pause_nosync() for d1v1 again

Quoted 3 lines

From this moment d1v1 has extra +1 in pause_count, so it goes offline right after credit>prv->credits_per_tslice forever.

Good catch -- I guess we don't really test the scheduler "cap" functionality enough to catch this one.

Unfortunately I think the real problem is that the ->flags field should be uniformly using the atomic bit operations.

Can you try the attached patch and see if it works for you?

-George

commit c72abbcf71f12ae5e9c5dda4a85bce23dbf4868f

Author: George Dunlap <[email protected]>
Date:   Tue Feb 26 12:44:48 2013 +0000

credit1: Use atomic bit operations for the flags structure

The flags structure is not protected by locks (or more precisely, it is protected using an inconsistent set of locks); we therefore need to make sure that all accesses are atomic-safe. This is particulary important in the case of the PARKED flag, which if clobbered while changing the YIELD bit will leave a vcpu wedged in an offline state.

Using the atomic bitops also requires us to change the size of the "flags" element.

xen/common/sched_credit.c +10 -13
index df2d076..ecdbd76 100644
@@ -46,8 +46,8 @@
 /*
  * Flags
  */
-#define CSCHED_FLAG_VCPU_PARKED    0x0001  /* VCPU over capped credits */
-#define CSCHED_FLAG_VCPU_YIELD     0x0002  /* VCPU yielding */
+#define CSCHED_FLAG_VCPU_PARKED    0x0  /* VCPU over capped credits */
+#define CSCHED_FLAG_VCPU_YIELD     0x1  /* VCPU yielding */
 
 
 /*
@@ -137,7 +137,7 @@ struct csched_vcpu {
     struct vcpu *vcpu;
     atomic_t credit;
     s_time_t start_time;   /* When we were scheduled (used for credit) */
-    uint16_t flags;
+    unsigned flags;
     int16_t pri;
 #ifdef CSCHED_STATS
     struct {
@@ -220,7 +220,7 @@ __runq_insert(unsigned int cpu, struct csched_vcpu *svc)
     /* If the vcpu yielded, try to put it behind one lower-priority
      * runnable vcpu if we can.  The next runq_sort will bring it forward
      * within 30ms if the queue too long. */
-    if ( svc->flags & CSCHED_FLAG_VCPU_YIELD
+    if ( test_bit(CSCHED_FLAG_VCPU_YIELD, &svc->flags)
          && __runq_elem(iter)->pri > CSCHED_PRI_IDLE )
     {
         iter=iter->next;
@@ -811,7 +811,7 @@ csched_vcpu_wake(const struct scheduler *ops, struct vcpu *vc)
      * those.
      */
     if ( svc->pri == CSCHED_PRI_TS_UNDER &&
-         !(svc->flags & CSCHED_FLAG_VCPU_PARKED) )
+         !test_bit(CSCHED_FLAG_VCPU_PARKED, &svc->flags) )
     {
         svc->pri = CSCHED_PRI_TS_BOOST;
     }
@@ -824,10 +824,10 @@ csched_vcpu_wake(const struct scheduler *ops, struct vcpu *vc)
 static void
 csched_vcpu_yield(const struct scheduler *ops, struct vcpu *vc)
 {
-    struct csched_vcpu * const sv = CSCHED_VCPU(vc);
+    struct csched_vcpu * const svc = CSCHED_VCPU(vc);
 
     /* Let the scheduler know that this vcpu is trying to yield */
-    sv->flags |= CSCHED_FLAG_VCPU_YIELD;
+    set_bit(CSCHED_FLAG_VCPU_YIELD, &svc->flags);
 }
 
 static int
@@ -1151,11 +1151,10 @@ csched_acct(void* dummy)
                 /* Park running VCPUs of capped-out domains */
                 if ( sdom->cap != 0U &&
                      credit < -credit_cap &&
-                     !(svc->flags & CSCHED_FLAG_VCPU_PARKED) )
+                     !test_and_set_bit(CSCHED_FLAG_VCPU_PARKED, &svc->flags) )
                 {
                     SCHED_STAT_CRANK(vcpu_park);
                     vcpu_pause_nosync(svc->vcpu);
-                    svc->flags |= CSCHED_FLAG_VCPU_PARKED;
                 }
 
                 /* Lower bound on credits */
@@ -1171,7 +1170,7 @@ csched_acct(void* dummy)
                 svc->pri = CSCHED_PRI_TS_UNDER;
 
                 /* Unpark any capped domains whose credits go positive */
-                if ( svc->flags & CSCHED_FLAG_VCPU_PARKED)
+                if ( test_and_clear_bit(CSCHED_FLAG_VCPU_PARKED, &svc->flags) )
                 {
                     /*
                      * It's important to unset the flag AFTER the unpause()
@@ -1180,7 +1179,6 @@ csched_acct(void* dummy)
                      */
                     SCHED_STAT_CRANK(vcpu_unpark);
                     vcpu_unpause(svc->vcpu);
-                    svc->flags &= ~CSCHED_FLAG_VCPU_PARKED;
                 }
 
                 /* Upper bound on credits means VCPU stops earning */
@@ -1442,8 +1440,7 @@ csched_schedule(
     /*
      * Clear YIELD flag before scheduling out
      */
-    if ( scurr->flags & CSCHED_FLAG_VCPU_YIELD )
-        scurr->flags &= ~(CSCHED_FLAG_VCPU_YIELD);
+    clear_bit(CSCHED_FLAG_VCPU_YIELD, &scurr->flags);
 
     /*
      * SMP Load balance:

_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xen.org/xen-devel

#2 lore raw
Quoted 38 lines · George Dunlap
Quoted 32 lines · Igor Pavlikevich

When vcpu stops yielding, it's possible to overwrite CSCHED_FLAG_VCPU_PARKED flag after clearing CSCHED_FLAG_VCPU_YIELD.

332.970218393 -------x d1v1 runstate_continue d1v1 running->running
]332.970248371 -------x d1v1 28005(2:8:5) 2 [ 1 1 ]
yielding, flags=0x0002
]332.970251968 -------x d1v1
clearing 0x0002, flags=0
332.970252175 -------x d1v1 runstate_continue d1v1 running->running
]332.970282574 -------x d1v1 28005(2:8:5) 2 [ 1 1 ]
yielding, flags=0x0002
]332.970283068 x------| d32767v0
entering csched_acct(), d1v1 flags=0x0002
]332.970285763 x------| d32767v0 28003(2:8:3) 2 [ 1 1 ]
calling vcpu_pause_nosync() for d1v1
]332.970286647 -------x d1v1
clearing 0x0002, flags=0
]332.970286771 x------| d32767v0
flags |= CSCHED_FLAG_VCPU_PARKED, flags=0x0003

]332.970286990 -------x d1v1   2800e(2:8:e) 2 [ 1 1cacfca ]
]332.970287122 -------x d1v1   2800f(2:8:f) 3 [ 7fff 1cacfca ffffffff ]
]332.970287242 -------x d1v1   2800a(2:8:a) 4 [ 1 1 7fff 7 ]

exiting schedule(), flags=0
332.970287417 -------x d1v1 runstate_change d1v1 running->offline 332.970287629 -------x d?v? runstate_change d32767v7 runnable->running
]332.970288397 x------- d32767v0
exiting csched_acct(), flags=0x0003
]332.995349690 x------- d32767v0
entering csched_acct(), d1v1 flags=0x0000, flag lost.
]332.995352170 x------- d32767v0 28003(2:8:3) 2 [ 1 1 ]
calling vcpu_pause_nosync() for d1v1 again

Quoted 3 lines

From this moment d1v1 has extra +1 in pause_count, so it goes offline right after credit>prv->credits_per_tslice forever.

Quoted 9 lines

Good catch -- I guess we don't really test the scheduler "cap" functionality enough to catch this one.

Unfortunately I think the real problem is that the ->flags field should be uniformly using the atomic bit operations.

Can you try the attached patch and see if it works for you?

Thanks for help, your patch works well.

_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xen.org/xen-devel