5 февр. 2019 г., в 12:00, Vladimir Davydov <vdavydov.dev@gmail.com> написал(а):

On Tue, Feb 05, 2019 at 09:28:01AM +0300, Serge Petrenko wrote:

4 февр. 2019 г., в 20:44, Vladimir Davydov <vdavydov.dev@gmail.com> написал(а):

diff --git a/src/fiber.c b/src/fiber.c
index 6f3d0ab78..4dc6b3c5a 100644
--- a/src/fiber.c
+++ b/src/fiber.c
@@ -392,9 +392,17 @@ fiber_join(struct fiber *fiber)
assert(fiber->flags & FIBER_IS_JOINABLE);

if (! fiber_is_dead(fiber)) {
- rlist_add_tail_entry(&fiber->wake, fiber(), state);

do {
+ /*
+  * In case fiber is cancelled during yield
+  * it will be removed from wake queue by a
+  * wakeup following the cancel.
+  * Having multiple entries for the same fiber
+  * doesn't hurt, since wakeup is executed only
+  * once per fiber.
+  */
+ rlist_add_tail_entry(&fiber->wake, fiber(), state);

I don't quite like the idea that cancelling a fiber that is joining
another fiber will have no effect until the other fiber has exited.
Can't we break the loop if fiber_is_cancelled()?

We can do that. But then we have to set FIBER_IS_JOINABLE to false
for the joined fiber so that it executes fiber_recycle().

Why should we? IMO the user should be free to kill a fiber executing
fiber_join. If that happens, the joinable fiber shouldn't be collected
until another fiber joins it successfully. This would be consistent with
pthread_join behavior.

Otherwise it will leak.  Is it ok?