UCX osc: properly release exclusive lock to avoid lockup #6933
Conversation
| int ret = OMPI_SUCCESS; | ||
|
|
||
| ret = opal_common_ucx_wpmem_fetch(module->state_mem, | ||
| UCP_ATOMIC_FETCH_OP_SWAP, TARGET_LOCK_UNLOCKED, |
There was a problem hiding this comment.
can you please elaborate a bit about the hang? If we set TARGET_LOCK_UNLOCKED (which is zero), start_shared should succeed, because result_value is supposed to be less than TARGET_LOCK_EXCLUSIVE
There was a problem hiding this comment.
Assume that process 0 holds an exclusive lock and process 1 tries to acquire a shared lock. Process 1 will retrieve the lock value (which is TARGET_LOCK_EXCLUSIVE) and add 1 it, making it TARGET_LOCK_EXCLUSIVE + 1. In the meantime, however, process 0 releases the lock by resetting it to TARGET_LOCK_UNLOCKED (which is 0). Since Process 1 has seen the lock as being taken exclusively, it will subtract 1 again, leading to a value of -1 in the lock variable (the lock is thus out of sync).
The next time Process 1 tries to take a shared lock, it will get the value -1 in a variable of type uint64_t, which is definitely >=TARGET_LOCK_EXCLUSIVE.
I realized, though, the my patch used -((int64_t)TARGET_LOCK_UNLOCKED) instead of -((int64_t)TARGET_LOCK_EXCLUSIVE). That should be fixed now. Sorry if that led to confusion.
Signed-off-by: Joseph Schuchart <schuchart@hlrs.de>
devreal
force-pushed
the
osc-ucx-excl-lock
branch
from
08cb638 to
a5cc380
Compare
|
I notice this was cherry-picked to the 3.1.x branch. Do we want this for the 4.0 branches? Edit: I see it was merged to 4.0.x a while ago. |
|
4.0.x: #6934 |
Fixes #6931
Signed-off-by: Joseph Schuchart schuchart@hlrs.de