Lockdep noticed the following 3-way lockup scenario: sys_perf_event_open() perf_event_alloc() perf_try_init_event() #0 ctx = perf_event_ctx_lock_nested(1) perf_swevent_init() swevent_hlist_get() #1 mutex_lock(&pmus_lock) perf_event_init_cpu() #1 mutex_lock(&pmus_lock) #2 mutex_lock(&ctx->mutex) sys_perf_event_open() mutex_lock_double() #2 mutex_lock() #0 mutex_lock_nested() And while we need that perf_event_ctx_lock_nested() for HW PMUs such that they can iterate the sibling list, trying to match it to the available counters, the software PMUs need do no such thing. Exclude them. In particular the swevent triggers the above invertion, while the tpevent PMU triggers a more elaborate one through their event_mutex.
Unified · Split
| @@ -9199,7 +9199,13 @@ static int perf_try_init_event(struct pmu *pmu, struct perf_event *event) · 9198 unmodified lines | ||
| 9199 | 9199 | if (!try_module_get(pmu->module)) |
| 9200 | 9200 | return -ENODEV; |
| 9201 | 9201 | |
| 9202 | − if (event->group_leader != event) { | |
| 9202 | + /* | |
| 9203 | + * A number of pmu->event_init() methods iterate the sibling_list to, | |
| 9204 | + * for example, validate if the group fits on the PMU. Therefore, | |
| 9205 | + * if this is a sibling event, acquire the ctx->mutex to protect | |
| 9206 | + * the sibling_list. | |
| 9207 | + */ | |
| 9208 | + if (event->group_leader != event && pmu->task_ctx_nr != perf_sw_context) { | |
| 9203 | 9209 | /* |
| 9204 | 9210 | * This ctx->mutex can nest when we're called through |
| 9205 | 9211 | * inheritance. See the perf_event_ctx_lock_nested() comment. |