eal: fix memory initialization deadlock
Checks
Commit Message
The issue arose due to the change in the DPDK read-write lock
implementation. That change added a new flag, RTE_RWLOCK_WAIT, designed
to prevent new read locks while a write lock is in the queue. However,
this change has led to a scenario where a recursive read lock, where a
lock is acquired twice by the same execution thread, can initiate a
sequence of events resulting in a deadlock:
Process 1 takes the first read lock.
Process 2 attempts to take a write lock, triggering RTE_RWLOCK_WAIT due
to the presence of a read lock. This makes process 2 enter a wait loop
until the read lock is released.
Process 1 tries to take a second read lock. However, since a write lock
is waiting (due to RTE_RWLOCK_WAIT), it also enters a wait loop until
the write lock is acquired and then released.
Both processes end up in a blocked state, unable to proceed, resulting
in a deadlock scenario.
Following these changes, the RW-lock no longer supports
recursion, implying that a single thread shouldn't obtain a read lock if
it already possesses one. The problem arises during initialization: the
rte_eal_init() function acquires the memory_hotplug_lock, and later on,
the sequence of calls rte_eal_memory_init() -> eal_memalloc_init() ->
rte_memseg_list_walk() acquires it again without releasing it. This
scenario introduces the risk of a potential deadlock when concurrent
write locks are applied to the same memory_hotplug_lock. To address this
we resolved the issue by replacing rte_memseg_list_walk() with
rte_memseg_list_walk_thread_unsafe().
Bugzilla ID: 1277
Fixes: 832cecc03d77 ("rwlock: prevent readers from starving writers")
Cc: stable@dpdk.org
Signed-off-by: Artemy Kovalyov <artemyko@nvidia.com>
---
lib/eal/include/generic/rte_rwlock.h | 4 ++++
lib/eal/linux/eal_memalloc.c | 7 +++++--
2 files changed, 9 insertions(+), 2 deletions(-)
Comments
2023-08-30 13:33 (UTC+0300), Artemy Kovalyov:
> Following these changes, the RW-lock no longer supports
> recursion, implying that a single thread shouldn't obtain a read lock if
> it already possesses one. The problem arises during initialization: the
> rte_eal_init() function acquires the memory_hotplug_lock, and later on,
> the sequence of calls rte_eal_memory_init() -> eal_memalloc_init() ->
> rte_memseg_list_walk() acquires it again without releasing it. This
> scenario introduces the risk of a potential deadlock when concurrent
> write locks are applied to the same memory_hotplug_lock. To address this
> we resolved the issue by replacing rte_memseg_list_walk() with
> rte_memseg_list_walk_thread_unsafe().
There is another call to rte_memseg_list_walk() during initialization:
from eal_dynmem_hugepage_init(), please address it too.
@@ -80,6 +80,10 @@ rte_rwlock_init(rte_rwlock_t *rwl)
/**
* Take a read lock. Loop until the lock is held.
*
+ * @note The RW lock isn't recursive, so calling this function on the same
+ * lock twice without releasing it could potentially result in a deadlock
+ * scenario when a write lock is involved.
+ *
* @param rwl
* A pointer to a rwlock structure.
*/
@@ -1740,7 +1740,10 @@ eal_memalloc_init(void)
eal_get_internal_configuration();
if (rte_eal_process_type() == RTE_PROC_SECONDARY)
- if (rte_memseg_list_walk(secondary_msl_create_walk, NULL) < 0)
+ /* memory_hotplug_lock is taken in rte_eal_init(), so it's
+ * safe to call thread-unsafe version.
+ */
+ if (rte_memseg_list_walk_thread_unsafe(secondary_msl_create_walk, NULL) < 0)
return -1;
if (rte_eal_process_type() == RTE_PROC_PRIMARY &&
internal_conf->in_memory) {
@@ -1778,7 +1781,7 @@ eal_memalloc_init(void)
}
/* initialize all of the fd lists */
- if (rte_memseg_list_walk(fd_list_create_walk, NULL))
+ if (rte_memseg_list_walk_thread_unsafe(fd_list_create_walk, NULL))
return -1;
return 0;
}