On 2024/5/10 16:20, Vlastimil Babka wrote:
On 5/10/24 9:59 AM, wuqiang.matt wrote:
On 2024/5/7 21:55, Vlastimil Babka wrote:
>>
+ } while (!try_cmpxchg_acquire(&slot->tail, &tail, tail + 1));
+
+ /* now the tail position is reserved for the given obj */
+ WRITE_ONCE(slot->entries[tail & slot->mask], obj);
+ /* update sequence to make this obj available for pop() */
+ smp_store_release(&slot->last, tail + 1);
+
+ return 0;
+}
/**
* objpool_push() - reclaim the object and return back to objpool
@@ -134,7 +219,19 @@ void *objpool_pop(struct objpool_head *pool);
* return: 0 or error code (it fails only when user tries to push
* the same object multiple times or wrong "objects" into objpool)
*/
-int objpool_push(void *obj, struct objpool_head *pool);
+static inline int objpool_push(void *obj, struct objpool_head *pool)
+{
+ unsigned long flags;
+ int rc;
+
+ /* disable local irq to avoid preemption & interruption */
+ raw_local_irq_save(flags);
+ rc = __objpool_try_add_slot(obj, pool, raw_smp_processor_id());
And IIUC, we could in theory objpool_pop() on one cpu, then later another
cpu might do objpool_push() and cause the latter cpu's pool to go over
capacity? Is there some implicit requirements of objpool users to take care
of having matched cpu for pop and push? Are the current objpool users
obeying this requirement? (I can see the selftests do, not sure about the
actual users).
Or am I missing something? Thanks.
The objects are all pre-allocated along with creation of the new objpool
and the total number of objects never exceeds the capacity on local node.
Aha, I see, the capacity of entries is enough to hold objects from all nodes
in the most unfortunate case they all end up freed from a single cpu.
So objpool_push() would always find an available slot from the ring-array
for the given object to insert back. objpool_pop() would try looping all
the percpu slots until an object is found or whole objpool is empty.
So it's correct, but seems rather wasteful to have the whole capacity for
entries replicated on every cpu? It does make objpool_push() simple and
fast, but as you say, objpool_pop() still has to search potentially all
non-local percpu slots, with disabled irqs, which is far from ideal.
Yes, it's a trade-off between performance and memory usage, with a slight
increase of memory consumption for a significant improvement of performance.
The reason of disabling local irqs is objpool uses a 32bit sequence number
as the state description of each element. It could likely overflow and go
back with the same value for extreme cases. 64bit value could eliminate the
collision but seems too heavy.
And the "abort if the slot was already full" comment for
objpool_try_add_slot() seems still misleading? Maybe that was your initial
idea but changed later?
Right, the comments are just left unchanged during iterations. The original
implementation kept each percpu ring-array very compact and objpool_push will
try looping all cpu nodes to return the given object to objpool.
Actually my new update would remove objpool_try_add_slot and integrate it's
functionality into objpool_push. I'll submit the new patch when I finish the
verification.
Currently kretprobe is the only actual usecase of objpool.
I'm testing an updated objpool in our HIDS project for critical pathes,
which is widely deployed on servers inside my company. The new version
eliminates the raw_local_irq_save and raw_local_irq_restore pair of
objpool_push and gains up to 5% of performance boost.
Mind Ccing me and linux-mm once you are posting that?
Sure, I'll make sure to let you know.
Thanks,
Vlastimil
Regards,
Matt Wu
