Besides deferring the call to housekeeping_update(), commit 6df415aa46ec ("cgroup/cpuset: Defer housekeeping_update() calls from CPU hotplug to workqueue") also defers the rebuild_sched_domains() call to the workqueue. So a new offline CPU may still be in a sched domain or new online CPU not showing up in the sched domains for a short transition period. That could be a problem in some corner cases and can be the cause of a reported test failure[1]. Fix it by calling rebuild_sched_domains_cpuslocked() directly in hotplug as before. If isolated partition invalidation or recreation is being done, the housekeeping_update() call to update the housekeeping cpumasks will still be deferred to a workqueue. In commit 3bfe47967191 ("cgroup/cpuset: Move housekeeping_update()/rebuild_sched_domains() together"), housekeeping_update() is called before rebuild_sched_domains() because it needs to access the HK_TYPE_DOMAIN housekeeping cpumask. That is now changed to use the static HK_TYPE_DOMAIN_BOOT cpumask as HK_TYPE_DOMAIN cpumask is now changeable at run time. As a result, we can move the rebuild_sched_domains() call before housekeeping_update() with the slight advantage that it will be done in the same cpus_read_lock critical section without the possibility of interference by a concurrent cpu hot add/remove operation. As it doesn't make sense to acquire cpuset_mutex/cpuset_top_mutex after calling housekeeping_update() and immediately release them again, move the cpuset_full_unlock() operation inside update_hk_sched_domains() and rename it to cpuset_update_sd_hk_unlock() to signify that it will release the full set of locks. [1] https://lore.kernel.org/lkml/1a89aceb-48db-4edd-a730-b445e41221fe@nvidia.com
Unified · Split
| @@ -879,7 +879,7 @@ generate_doms: · 878 unmodified lines | ||
| 879 | 879 | /* |
| 880 | 880 | * Cgroup v2 doesn't support domain attributes, just set all of them |
| 881 | 881 | * to SD_ATTR_INIT. Also non-isolating partition root CPUs are a |
| 882 | − * subset of HK_TYPE_DOMAIN housekeeping CPUs. | |
| 882 | + * subset of HK_TYPE_DOMAIN_BOOT housekeeping CPUs. | |
| 883 | 883 | */ |
| 884 | 884 | for (i = 0; i < ndoms; i++) { |
| 885 | 885 | /* |
| @@ -888,7 +888,7 @@ generate_doms: · 2 unmodified lines | ||
| 888 | 888 | */ |
| 889 | 889 | if (!csa || csa[i] == &top_cpuset) |
| 890 | 890 | cpumask_and(doms[i], top_cpuset.effective_cpus, |
| 891 | − housekeeping_cpumask(HK_TYPE_DOMAIN)); | |
| 891 | + housekeeping_cpumask(HK_TYPE_DOMAIN_BOOT)); | |
| 892 | 892 | else |
| 893 | 893 | cpumask_copy(doms[i], csa[i]->effective_cpus); |
| 894 | 894 | if (dattr) |
| @@ -1329,17 +1329,22 @@ static bool prstate_housekeeping_conflict(int prstate, struct cpumask *new_cpus) · 434 unmodified lines | ||
| 1329 | 1329 | } |
| 1330 | 1330 | |
| 1331 | 1331 | /* |
| 1332 | − * update_hk_sched_domains - Update HK cpumasks & rebuild sched domains | |
| 1332 | + * cpuset_update_sd_hk_unlock - Rebuild sched domains, update HK & unlock | |
| 1333 | 1333 | * |
| 1334 | − * Update housekeeping cpumasks and rebuild sched domains if necessary. | |
| 1335 | − * This should be called at the end of cpuset or hotplug actions. | |
| 1334 | + * Update housekeeping cpumasks and rebuild sched domains if necessary and | |
| 1335 | + * then do a cpuset_full_unlock(). | |
| 1336 | + * This should be called at the end of cpuset operation. | |
| 1336 | 1337 | */ |
| 1337 | −static void update_hk_sched_domains(void) | |
| 1338 | +static void cpuset_update_sd_hk_unlock(void) | |
| 1339 | + __releases(&cpuset_mutex) | |
| 1340 | + __releases(&cpuset_top_mutex) | |
| 1338 | 1341 | { |
| 1342 | + /* force_sd_rebuild will be cleared in rebuild_sched_domains_locked() */ | |
| 1343 | + if (force_sd_rebuild) | |
| 1344 | + rebuild_sched_domains_locked(); | |
| 1345 | + | |
| 1339 | 1346 | if (update_housekeeping) { |
| 1340 | − /* Updating HK cpumasks implies rebuild sched domains */ | |
| 1341 | 1347 | update_housekeeping = false; |
| 1342 | − force_sd_rebuild = true; | |
| 1343 | 1348 | cpumask_copy(isolated_hk_cpus, isolated_cpus); |
| 1344 | 1349 | |
| 1345 | 1350 | /* |
| @@ -1350,22 +1355,19 @@ static void update_hk_sched_domains(void) · 4 unmodified lines | ||
| 1350 | 1355 | mutex_unlock(&cpuset_mutex); |
| 1351 | 1356 | cpus_read_unlock(); |
| 1352 | 1357 | WARN_ON_ONCE(housekeeping_update(isolated_hk_cpus)); |
| 1353 | − cpus_read_lock(); | |
| 1354 | − mutex_lock(&cpuset_mutex); | |
| 1358 | + mutex_unlock(&cpuset_top_mutex); | |
| 1359 | + } else { | |
| 1360 | + cpuset_full_unlock(); | |
| 1355 | 1361 | } |
| 1356 | − /* force_sd_rebuild will be cleared in rebuild_sched_domains_locked() */ | |
| 1357 | − if (force_sd_rebuild) | |
| 1358 | − rebuild_sched_domains_locked(); | |
| 1359 | 1362 | } |
| 1360 | 1363 | |
| 1361 | 1364 | /* |
| 1362 | − * Work function to invoke update_hk_sched_domains() | |
| 1365 | + * Work function to invoke cpuset_update_sd_hk_unlock() | |
| 1363 | 1366 | */ |
| 1364 | 1367 | static void hk_sd_workfn(struct work_struct *work) |
| 1365 | 1368 | { |
| 1366 | 1369 | cpuset_full_lock(); |
| 1367 | − update_hk_sched_domains(); | |
| 1368 | − cpuset_full_unlock(); | |
| 1370 | + cpuset_update_sd_hk_unlock(); | |
| 1369 | 1371 | } |
| 1370 | 1372 | |
| 1371 | 1373 | /** |
| @@ -3230,8 +3232,7 @@ ssize_t cpuset_write_resmask(struct kernfs_open_file *of, · 1858 unmodified lines | ||
| 3230 | 3232 | |
| 3231 | 3233 | free_cpuset(trialcs); |
| 3232 | 3234 | out_unlock: |
| 3233 | − update_hk_sched_domains(); | |
| 3234 | − cpuset_full_unlock(); | |
| 3235 | + cpuset_update_sd_hk_unlock(); | |
| 3235 | 3236 | if (of_cft(of)->private == FILE_MEMLIST) |
| 3236 | 3237 | schedule_flush_migrate_mm(); |
| 3237 | 3238 | return retval ?: nbytes; |
| @@ -3338,8 +3339,7 @@ static ssize_t cpuset_partition_write(struct kernfs_open_file *of, char *buf, · 100 unmodified lines | ||
| 3338 | 3339 | cpuset_full_lock(); |
| 3339 | 3340 | if (is_cpuset_online(cs)) |
| 3340 | 3341 | retval = update_prstate(cs, val); |
| 3341 | − update_hk_sched_domains(); | |
| 3342 | − cpuset_full_unlock(); | |
| 3342 | + cpuset_update_sd_hk_unlock(); | |
| 3343 | 3343 | return retval ?: nbytes; |
| 3344 | 3344 | } |
| 3345 | 3345 | |
| @@ -3513,8 +3513,7 @@ static void cpuset_css_killed(struct cgroup_subsys_state *css) · 167 unmodified lines | ||
| 3513 | 3513 | /* Reset valid partition back to member */ |
| 3514 | 3514 | if (is_partition_valid(cs)) |
| 3515 | 3515 | update_prstate(cs, PRS_MEMBER); |
| 3516 | − update_hk_sched_domains(); | |
| 3517 | − cpuset_full_unlock(); | |
| 3516 | + cpuset_update_sd_hk_unlock(); | |
| 3518 | 3517 | } |
| 3519 | 3518 | |
| 3520 | 3519 | static void cpuset_css_free(struct cgroup_subsys_state *css) |
| @@ -3923,11 +3922,13 @@ static void cpuset_handle_hotplug(void) · 402 unmodified lines | ||
| 3923 | 3922 | rcu_read_unlock(); |
| 3924 | 3923 | } |
| 3925 | 3924 | |
| 3926 | − | |
| 3927 | 3925 | /* |
| 3928 | − * Queue a work to call housekeeping_update() & rebuild_sched_domains() | |
| 3929 | − * There will be a slight delay before the HK_TYPE_DOMAIN housekeeping | |
| 3930 | − * cpumask can correctly reflect what is in isolated_cpus. | |
| 3926 | + * rebuild_sched_domains() will always be called directly if needed | |
| 3927 | + * to make sure that newly added or removed CPU will be reflected in | |
| 3928 | + * the sched domains. However, if isolated partition invalidation | |
| 3929 | + * or recreation is being done (update_housekeeping set), a work item | |
| 3930 | + * will be queued to call housekeeping_update() to update the | |
| 3931 | + * corresponding housekeeping cpumasks after some slight delay. | |
| 3931 | 3932 | * |
| 3932 | 3933 | * We rely on WORK_STRUCT_PENDING_BIT to not requeue a work item that |
| 3933 | 3934 | * is still pending. Before the pending bit is cleared, the work data |
| @@ -3936,8 +3937,10 @@ static void cpuset_handle_hotplug(void) · 2 unmodified lines | ||
| 3936 | 3937 | * previously queued work. Since hk_sd_workfn() doesn't use the work |
| 3937 | 3938 | * item at all, this is not a problem. |
| 3938 | 3939 | */ |
| 3939 | − if (update_housekeeping || force_sd_rebuild) | |
| 3940 | − queue_work(system_unbound_wq, &hk_sd_work); | |
| 3940 | + if (force_sd_rebuild) | |
| 3941 | + rebuild_sched_domains_cpuslocked(); | |
| 3942 | + if (update_housekeeping) | |
| 3943 | + queue_work(system_dfl_wq, &hk_sd_work); | |
| 3941 | 3944 | |
| 3942 | 3945 | free_tmpmasks(ptmp); |
| 3943 | 3946 | } |