Messages in this thread | | | Date | Fri, 5 Apr 2024 10:54:00 +0200 | Subject | Re: [PATCH v2 11/15] sched/deadline: Mark DL server as unthrottled before enqueue | From | Daniel Bristot de Oliveira <> |
| |
On 3/13/24 02:24, Joel Fernandes (Google) wrote: > The DL server may not have had its timer started if start_dl_timer() > returns 0 (say the zero-laxity time has already passed). In such cases, > mark the DL task which is about to be enqueued as not throttled and > cancel any previous timers, then do the enqueue. > > This fixes the following crash: > > [ 9.263331] kernel BUG at kernel/sched/deadline.c:1765! > [ 9.282382] Call Trace: > [ 9.282767] <TASK> > [ 9.283086] ? __die_body+0x62/0xb0 > [ 9.283602] ? die+0x9b/0xc0 > [ 9.284036] ? do_trap+0xa3/0x170 > [ 9.284528] ? enqueue_dl_entity+0x45e/0x460 > [ 9.285158] ? enqueue_dl_entity+0x45e/0x460 > [ 9.285791] ? handle_invalid_op+0x65/0x80 > [ 9.286392] ? enqueue_dl_entity+0x45e/0x460 > [ 9.287021] ? exc_invalid_op+0x2f/0x40 > [ 9.287585] ? asm_exc_invalid_op+0x16/0x20 > [ 9.288200] ? find_later_rq+0x120/0x120 > [ 9.288775] ? fair_server_init+0x40/0x40 > [ 9.289364] ? enqueue_dl_entity+0x45e/0x460 > [ 9.289989] ? find_later_rq+0x120/0x120 > [ 9.290564] dl_task_timer+0x1d7/0x2f0 > [ 9.291120] ? find_later_rq+0x120/0x120 > [ 9.291695] __run_hrtimer+0x73/0x1b0 > [ 9.292238] hrtimer_interrupt+0x216/0x2c0 > [ 9.292841] __sysvec_apic_timer_interrupt+0x53/0x140 > [ 9.293581] sysvec_apic_timer_interrupt+0x2d/0x80 > [ 9.294285] asm_sysvec_apic_timer_interrupt+0x16/0x20 > > The crash can easily be reproduced by adding a 100ms delay as follows: > > +int delay_inject_count; > + > static void > enqueue_dl_entity(struct sched_dl_entity *dl_se, int flags) > { > @@ -1827,6 +1830,12 @@ enqueue_dl_entity(struct sched_dl_entity *dl_se, int flags) > setup_new_dl_entity(dl_se); > } > > + // 100ms delay every 20 enqueues. > + if (delay_inject_count++ > 20) { > + mdelay(100); > + delay_inject_count = 0; > + } > + > /* > * If we are still throttled, eg. we got replenished but are a > * zero-laxity task and still got to wait, don't enqueue.
Makes sense, I am adding this in the defer patch v6 as it is a fix for it...
-- Daniel
| |