On 2/15/24 2:26 PM, Toke Høiland-Jørgensen wrote:
The BPF_TEST_RUN code in XDP live frame mode creates a new page pool
each time it is called and uses that to allocate the frames used for the
XDP run. This works well if the syscall is used with a high repetitions
number, as it allows for efficient page recycling. However, if used with
a small number of repetitions, the overhead of creating and tearing down
the page pool is significant, and can even lead to system stalls if the
syscall is called in a tight loop.
Now that we have a persistent system page pool instance, it becomes
pretty straight forward to change the test_run code to use it. The only
wrinkle is that we can no longer rely on a custom page init callback
from page_pool itself; instead, we change the test_run code to write a
random cookie value to the beginning of the page as an indicator that
the page has been initialised and can be re-used without copying the
initial data again.
Signed-off-by: Toke Høiland-Jørgensen <toke@xxxxxxxxxx>
[...]
-
/* We create a 'fake' RXQ referencing the original dev, but with an
* xdp_mem_info pointing to our page_pool
*/
xdp_rxq_info_reg(&xdp->rxq, orig_ctx->rxq->dev, 0, 0);
- xdp->rxq.mem.type = MEM_TYPE_PAGE_POOL;
- xdp->rxq.mem.id = pp->xdp_mem_id;
+ xdp->rxq.mem.type = MEM_TYPE_PAGE_POOL; /* mem id is set per-frame below */
xdp->dev = orig_ctx->rxq->dev;
xdp->orig_ctx = orig_ctx;
+ /* We need a random cookie for each run as pages can stick around
+ * between runs in the system page pool
+ */
+ get_random_bytes(&xdp->cookie, sizeof(xdp->cookie));
+
So the assumption is that there is only a tiny chance of collisions with
users outside of xdp test_run. If they do collide however, you'd leak data.
Presumably the 64 bit cookie might suffice.. nit, perhaps makes sense to
explicitly exclude zero cookie?
Acked-by: Daniel Borkmann <daniel@xxxxxxxxxxxxx>
Thanks,
Daniel
