On 21.01.25 9:15, Pavel Tikhomirov wrote:
On 1/21/25 01:34, Alexander Atanasov wrote:
On 20.01.25 12:08, Pavel Tikhomirov wrote:
On 12/6/24 05:56, Alexander Atanasov wrote:
+static inline int ploop_runners_add_work(struct ploop *ploop,
struct pio *pio)
+{
+ int i;
+ struct ploop_worker *wrkr;
+
+ if (++ploop->last_used_runner >= ploop->nkt_runners)
+ ploop->last_used_runner = 0;
+ wrkr = ploop->kt_runners[ploop->last_used_runner];
This can potentially lead to out of bound read, as we don't have
locks around wrapping last_used_runner increment to zero (i.e. it's
not atomic), here we can see ploop->last_used_runner >= nkt_runners.
Reworked - runners are circular list - READ_ONCE/WRITE_ONCE used to
switch to ->next /see updated patches/
In updated patches:
+static inline int ploop_runners_add_work(struct ploop *ploop, struct
pio *pio)
+{
+ struct ploop_worker *wrkr;
+
+ wrkr = READ_ONCE(ploop->last_used_runner)->next; // A
+ WRITE_ONCE(ploop->last_used_runner, wrkr); // B
No out of bound now. But note that as we have two operations thus it's
not atomic, and thus this list iteration can go "backwards" sometimes:
A1 -> A2 -> B2 -> A3 -> B3 -> ... AN -> BN -> B1
Yes, we may send two requests to same runner or skip but this does not
matter - we do not require strict scheduling here.
Note: I see (in final version, with all patches applied) multiple
stacks which probably can run concurrently:
+-< ploop_runners_add_work
+-< ploop_index_wb_submit
| +-< ploop_grow_relocate_cluster
| | +-< ploop_process_resize_cmd
| | | +-< ploop_resize
| | | | +-< ploop_message
| +-< ploop_grow_update_header
| | +-< ploop_process_resize_cmd
| | | +-< ploop_resize
| | | | +-< ploop_message
| +-< ploop_submit_metadata_writeback
| | +-< do_ploop_run_work
| | | +-< do_ploop_work
| | | +-< ploop_worker
+-< process_ploop_fsync_work
| +-< do_ploop_run_work
| | +-< do_ploop_work
| | +-< ploop_worker
+-< ploop_runners_add_work_list
| +-< do_ploop_run_work
| | +-< do_ploop_work
| | +-< ploop_worker
+
+ atomic_inc(&ploop->kt_worker->inflight_pios);
+ llist_add((struct llist_node *)(&pio->list), &wrkr->work_llist);
+ wake_up_process(wrkr->task);
+
+ return 0;
+}
--
Regards,
Alexander Atanasov
_______________________________________________
Devel mailing list
Devel@openvz.org
https://lists.openvz.org/mailman/listinfo/devel
