Re: svn commit: r357641 - head/sys/kern
On 2/6/20, Michael Tuexen wrote:
>
>
>> On 6. Feb 2020, at 21:51, Jeff Roberson wrote:
>>
>> Author: jeff
>> Date: Thu Feb 6 20:51:46 2020
>> New Revision: 357641
>> URL: https://svnweb.freebsd.org/changeset/base/357641
>>
>> Log:
>> Fix a race in smr_advance() that could result in unnecessary poll calls.
>>
>> This was relatively harmless but surprising to see in counters. The
>> race occurred when rd_seq was read after the goal was updated and we
>> incorrectly calculated the delta between them.
>>
>> Reviewed by:rlibby
>> Differential Revision: https://reviews.freebsd.org/D23464
>>
>> Modified:
>> head/sys/kern/subr_smr.c
>>
>> Modified: head/sys/kern/subr_smr.c
>> ==
>> --- head/sys/kern/subr_smr.c Thu Feb 6 20:47:50 2020(r357640)
>> +++ head/sys/kern/subr_smr.c Thu Feb 6 20:51:46 2020(r357641)
>> @@ -160,7 +160,7 @@ static uma_zone_t smr_zone;
>> #define SMR_SEQ_INCR(UINT_MAX / 1)
>> #define SMR_SEQ_INIT(UINT_MAX - 10)
>> /* Force extra polls to test the integer overflow detection. */
>> -#define SMR_SEQ_MAX_DELTA (1000)
>> +#define SMR_SEQ_MAX_DELTA (SMR_SEQ_INCR * 32)
>> #define SMR_SEQ_MAX_ADVANCE SMR_SEQ_MAX_DELTA / 2
>> #endif
>>
>> @@ -188,7 +188,7 @@ smr_seq_t
>> smr_advance(smr_t smr)
>> {
>> smr_shared_t s;
>> -smr_seq_t goal;
>> +smr_seq_t goal, s_rd_seq;
>>
>> /*
>> * It is illegal to enter while in an smr section.
>> @@ -203,12 +203,18 @@ smr_advance(smr_t smr)
>> atomic_thread_fence_rel();
>>
>> /*
>> + * Load the current read seq before incrementing the goal so
>> + * we are guaranteed it is always < goal.
>> + */
>> +s = zpcpu_get(smr)->c_shared;
>> +s_rd_seq = atomic_load_acq_int(&s->s_rd_seq);
>> +
>> +/*
>> * Increment the shared write sequence by 2. Since it is
>> * initialized to 1 this means the only valid values are
>> * odd and an observed value of 0 in a particular CPU means
>> * it is not currently in a read section.
>> */
>> -s = zpcpu_get(smr)->c_shared;
>> goal = atomic_fetchadd_int(&s->s_wr_seq, SMR_SEQ_INCR) + SMR_SEQ_INCR;
>> counter_u64_add(advance, 1);
>>
>> @@ -217,7 +223,7 @@ smr_advance(smr_t smr)
>> * far ahead of the read sequence number. This keeps the
>> * wrap detecting arithmetic working in pathological cases.
>> */
>> -if (goal - atomic_load_int(&s->s_rd_seq) >= SMR_SEQ_MAX_DELTA) {
>> +if (SMR_SEQ_DELTA(goal, s_rd_seq) >= SMR_SEQ_MAX_DELTA) {
> SMR_SEQ_DELTA is not defined, therefore compilation fails.
https://reviews.freebsd.org/D23464#516866
https://reviews.freebsd.org/D23464#516881
Jung-uk Kim
> Best regards
> Michael
>> counter_u64_add(advance_wait, 1);
>> smr_wait(smr, goal - SMR_SEQ_MAX_ADVANCE);
>> }
signature.asc
Description: OpenPGP digital signature
Re: svn commit: r357641 - head/sys/kern
> On 6. Feb 2020, at 21:51, Jeff Roberson wrote:
>
> Author: jeff
> Date: Thu Feb 6 20:51:46 2020
> New Revision: 357641
> URL: https://svnweb.freebsd.org/changeset/base/357641
>
> Log:
> Fix a race in smr_advance() that could result in unnecessary poll calls.
>
> This was relatively harmless but surprising to see in counters. The
> race occurred when rd_seq was read after the goal was updated and we
> incorrectly calculated the delta between them.
>
> Reviewed by: rlibby
> Differential Revision: https://reviews.freebsd.org/D23464
>
> Modified:
> head/sys/kern/subr_smr.c
>
> Modified: head/sys/kern/subr_smr.c
> ==
> --- head/sys/kern/subr_smr.c Thu Feb 6 20:47:50 2020(r357640)
> +++ head/sys/kern/subr_smr.c Thu Feb 6 20:51:46 2020(r357641)
> @@ -160,7 +160,7 @@ static uma_zone_t smr_zone;
> #define SMR_SEQ_INCR(UINT_MAX / 1)
> #define SMR_SEQ_INIT(UINT_MAX - 10)
> /* Force extra polls to test the integer overflow detection. */
> -#define SMR_SEQ_MAX_DELTA (1000)
> +#define SMR_SEQ_MAX_DELTA (SMR_SEQ_INCR * 32)
> #define SMR_SEQ_MAX_ADVANCE SMR_SEQ_MAX_DELTA / 2
> #endif
>
> @@ -188,7 +188,7 @@ smr_seq_t
> smr_advance(smr_t smr)
> {
> smr_shared_t s;
> - smr_seq_t goal;
> + smr_seq_t goal, s_rd_seq;
>
> /*
>* It is illegal to enter while in an smr section.
> @@ -203,12 +203,18 @@ smr_advance(smr_t smr)
> atomic_thread_fence_rel();
>
> /*
> + * Load the current read seq before incrementing the goal so
> + * we are guaranteed it is always < goal.
> + */
> + s = zpcpu_get(smr)->c_shared;
> + s_rd_seq = atomic_load_acq_int(&s->s_rd_seq);
> +
> + /*
>* Increment the shared write sequence by 2. Since it is
>* initialized to 1 this means the only valid values are
>* odd and an observed value of 0 in a particular CPU means
>* it is not currently in a read section.
>*/
> - s = zpcpu_get(smr)->c_shared;
> goal = atomic_fetchadd_int(&s->s_wr_seq, SMR_SEQ_INCR) + SMR_SEQ_INCR;
> counter_u64_add(advance, 1);
>
> @@ -217,7 +223,7 @@ smr_advance(smr_t smr)
>* far ahead of the read sequence number. This keeps the
>* wrap detecting arithmetic working in pathological cases.
>*/
> - if (goal - atomic_load_int(&s->s_rd_seq) >= SMR_SEQ_MAX_DELTA) {
> + if (SMR_SEQ_DELTA(goal, s_rd_seq) >= SMR_SEQ_MAX_DELTA) {
SMR_SEQ_DELTA is not defined, therefore compilation fails.
Best regards
Michael
> counter_u64_add(advance_wait, 1);
> smr_wait(smr, goal - SMR_SEQ_MAX_ADVANCE);
> }
___
svn-src-all@freebsd.org mailing list
https://lists.freebsd.org/mailman/listinfo/svn-src-all
To unsubscribe, send any mail to "svn-src-all-unsubscr...@freebsd.org"
svn commit: r357641 - head/sys/kern
Author: jeff
Date: Thu Feb 6 20:51:46 2020
New Revision: 357641
URL: https://svnweb.freebsd.org/changeset/base/357641
Log:
Fix a race in smr_advance() that could result in unnecessary poll calls.
This was relatively harmless but surprising to see in counters. The
race occurred when rd_seq was read after the goal was updated and we
incorrectly calculated the delta between them.
Reviewed by: rlibby
Differential Revision:https://reviews.freebsd.org/D23464
Modified:
head/sys/kern/subr_smr.c
Modified: head/sys/kern/subr_smr.c
==
--- head/sys/kern/subr_smr.cThu Feb 6 20:47:50 2020(r357640)
+++ head/sys/kern/subr_smr.cThu Feb 6 20:51:46 2020(r357641)
@@ -160,7 +160,7 @@ static uma_zone_t smr_zone;
#defineSMR_SEQ_INCR(UINT_MAX / 1)
#defineSMR_SEQ_INIT(UINT_MAX - 10)
/* Force extra polls to test the integer overflow detection. */
-#defineSMR_SEQ_MAX_DELTA (1000)
+#defineSMR_SEQ_MAX_DELTA (SMR_SEQ_INCR * 32)
#defineSMR_SEQ_MAX_ADVANCE SMR_SEQ_MAX_DELTA / 2
#endif
@@ -188,7 +188,7 @@ smr_seq_t
smr_advance(smr_t smr)
{
smr_shared_t s;
- smr_seq_t goal;
+ smr_seq_t goal, s_rd_seq;
/*
* It is illegal to enter while in an smr section.
@@ -203,12 +203,18 @@ smr_advance(smr_t smr)
atomic_thread_fence_rel();
/*
+* Load the current read seq before incrementing the goal so
+* we are guaranteed it is always < goal.
+*/
+ s = zpcpu_get(smr)->c_shared;
+ s_rd_seq = atomic_load_acq_int(&s->s_rd_seq);
+
+ /*
* Increment the shared write sequence by 2. Since it is
* initialized to 1 this means the only valid values are
* odd and an observed value of 0 in a particular CPU means
* it is not currently in a read section.
*/
- s = zpcpu_get(smr)->c_shared;
goal = atomic_fetchadd_int(&s->s_wr_seq, SMR_SEQ_INCR) + SMR_SEQ_INCR;
counter_u64_add(advance, 1);
@@ -217,7 +223,7 @@ smr_advance(smr_t smr)
* far ahead of the read sequence number. This keeps the
* wrap detecting arithmetic working in pathological cases.
*/
- if (goal - atomic_load_int(&s->s_rd_seq) >= SMR_SEQ_MAX_DELTA) {
+ if (SMR_SEQ_DELTA(goal, s_rd_seq) >= SMR_SEQ_MAX_DELTA) {
counter_u64_add(advance_wait, 1);
smr_wait(smr, goal - SMR_SEQ_MAX_ADVANCE);
}
___
svn-src-all@freebsd.org mailing list
https://lists.freebsd.org/mailman/listinfo/svn-src-all
To unsubscribe, send any mail to "svn-src-all-unsubscr...@freebsd.org"
