In commit
72653ae5303c ("selftests: net: tcp_mmap:
Use huge pages in send path") I made a change to use hugepages
for the buffer used by the client (tx path)
Today, I understood that the cause for poor zerocopy
performance was that after a mmap() for a 512KB memory
zone, kernel uses a single zeropage, mapped 128 times.
This was really the reason for poor tx path performance
in zero copy mode, because this zero page refcount is
under high pressure, especially when TCP ACK packets
are processed on another cpu.
We need either to force a COW on all the memory range,
or use MAP_POPULATE so that a zero page is not abused.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20230120181136.3764521-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
if (buffer == (void *)-1) {
sz = need;
buffer = mmap(NULL, sz, PROT_READ | PROT_WRITE,
- MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
+ MAP_PRIVATE | MAP_ANONYMOUS | MAP_POPULATE,
+ -1, 0);
if (buffer != (void *)-1)
fprintf(stderr, "MAP_HUGETLB attempt failed, look at /sys/kernel/mm/hugepages for optimal performance\n");
}