On 09/18/2016 06:25 AM, Mintz, Yuval wrote:
Currently, we can have high order page allocations that specify
GFP_ATOMIC when configuring multicast MAC address filters.
For example, we have seen order 2 page allocation failures with
~500 multicast addresses configured.
Convert the allocation for the pending list to be done in PAGE_SIZE
Signed-off-by: Jason Baron <jba...@akamai.com>
While I appreciate the effort, I wonder whether it's worth it:
- The hardware [even in its newer generation] provides an approximate
based classification [I.e., hashed] with 256 bins.
When configuring 500 multicast addresses, one can argue the
difference between multicast-promisc mode and actual configuration
With 256 bins, I think it takes close to: 256*lg(256) or 2,048
multicast addresses to expect to have all bins have at least one hash,
assuming a uniform distribution of the hashes.
Perhaps the easier-to-maintain alternative would simply be to
determine the maximal number of multicast addresses that can be
configured using a single PAGE, and if in need of more than that
simply move into multicast-promisc.
sizeof(struct bnx2x_mcast_list_elem) = 24. So there are 170 per
page on x86. So if we want to fit 2,048 elements, we need 12 pages.
- While GFP_ATOMIC is required in this flow due to the fact it's being
called from sleepless context, I do believe this is mostly a remnant -
it's possible that by slightly changing the locking scheme we can have
the configuration done from sleepless context and simply switch to
Ok if its GFP_KERNEL, I think its still undesirable to do large page
order allocations (unless of course its converted to a single page, but
I'm not sure this makes sense as mentioned).
Regarding the patch itself, only comment I have:
+ elem_group = (struct bnx2x_mcast_elem_group *)
Let's use list_next_entry() instead.
I think it would be easy to add a check to bnx2x_set_rx_mode_inner() to
enforce some maximum number of elements (perhaps 2,048 based on the
above math) for the !CHIP_IS_E1() case on top of what I already posted.