net/rds: Optimize rds_ib_laddr_check

rds_ib_laddr_check() creates a CM_ID and attempts to bind the address
in question to it. This in order to qualify the allegedly local
address as a usable IB/RoCE address.

In the field, ExaWatcher runs rds-ping to all ports in the fabric from
all local ports. This using all active ToS'es. In a full rack system,
we have 14 cell servers and eight db servers. Typically, 6 ToS'es are
used. This implies 528 rds-ping invocations per ExaWatcher's "RDSinfo"
interval.

Adding to this, each rds-ping invocation creates eight sockets and
binds the local address to them:

socket(AF_RDS, SOCK_SEQPACKET, 0)       = 3
bind(3, {sa_family=AF_INET, sin_port=htons(0),
	sin_addr=inet_addr("192.168.36.2")}, 16) = 0
socket(AF_RDS, SOCK_SEQPACKET, 0)       = 4
bind(4, {sa_family=AF_INET, sin_port=htons(0),
	sin_addr=inet_addr("192.168.36.2")}, 16) = 0
socket(AF_RDS, SOCK_SEQPACKET, 0)       = 5
bind(5, {sa_family=AF_INET, sin_port=htons(0),
	sin_addr=inet_addr("192.168.36.2")}, 16) = 0
socket(AF_RDS, SOCK_SEQPACKET, 0)       = 6
bind(6, {sa_family=AF_INET, sin_port=htons(0),
	sin_addr=inet_addr("192.168.36.2")}, 16) = 0
socket(AF_RDS, SOCK_SEQPACKET, 0)       = 7
bind(7, {sa_family=AF_INET, sin_port=htons(0),
	sin_addr=inet_addr("192.168.36.2")}, 16) = 0
socket(AF_RDS, SOCK_SEQPACKET, 0)       = 8
bind(8, {sa_family=AF_INET, sin_port=htons(0),
	sin_addr=inet_addr("192.168.36.2")}, 16) = 0
socket(AF_RDS, SOCK_SEQPACKET, 0)       = 9
bind(9, {sa_family=AF_INET, sin_port=htons(0),
	sin_addr=inet_addr("192.168.36.2")}, 16) = 0
socket(AF_RDS, SOCK_SEQPACKET, 0)       = 10
bind(10, {sa_family=AF_INET, sin_port=htons(0),
	sin_addr=inet_addr("192.168.36.2")}, 16) = 0

So, at every interval ExaWatcher executes rds-ping's, 4224 CM_IDs are
allocated, considering this full-rack system. After the a CM_ID has
been allocated, rdma_bind_addr() is called, with the port number being
zero. This implies that the CMA will attempt to search for an un-used
ephemeral port. Simplified, the algorithm is to start at a random
position in the available port space, and then if needed, iterate
until an un-used port is found.

The book-keeping of used ports uses the idr system, which again uses
slab to allocate new struct idr_layer's. The size is 2092 bytes and
slab tries to reduce the wasted space. Hence, it chooses an order:3
allocation, for which 15 idr_layer structs will fit and only 1388
bytes are wasted per the 32KiB order:3 chunk.

Although this order:3 allocation seems like a good space/speed
trade-off, it does not resonate well with how it used by the CMA. The
combination of the randomized starting point in the port space (which
has close to zero spatial locality) and the close proximity in time of
the 4224 invocations of the rds-ping's, creates a memory hog for
order:3 allocations.

These costly allocations may need reclaims and/or compaction. At
worst, they may fail and produce a stack trace such as (from uek4):

[<ffffffff811a72d5>] __inc_zone_page_state+0x35/0x40
[<ffffffff811c2e97>] page_add_file_rmap+0x57/0x60
[<ffffffffa37ca1df>] remove_migration_pte+0x3f/0x3c0 [ksplice_6cn872bt_vmlinux_new]
[<ffffffff811c3de8>] rmap_walk+0xd8/0x340
[<ffffffff811e8860>] remove_migration_ptes+0x40/0x50
[<ffffffff811ea83c>] migrate_pages+0x3ec/0x890
[<ffffffff811afa0d>] compact_zone+0x32d/0x9a0
[<ffffffff811b00ed>] compact_zone_order+0x6d/0x90
[<ffffffff811b03b2>] try_to_compact_pages+0x102/0x270
[<ffffffff81190e56>] __alloc_pages_direct_compact+0x46/0x100
[<ffffffff8119165b>] __alloc_pages_nodemask+0x74b/0xaa0
[<ffffffff811d8411>] alloc_pages_current+0x91/0x110
[<ffffffff811e3b0b>] new_slab+0x38b/0x480
[<ffffffffa41323c7>] __slab_alloc+0x3b7/0x4a0 [ksplice_s0dk66a8_vmlinux_new]
[<ffffffff811e42ab>] kmem_cache_alloc+0x1fb/0x250
[<ffffffff8131fdd6>] idr_layer_alloc+0x36/0x90
[<ffffffff8132029c>] idr_get_empty_slot+0x28c/0x3d0
[<ffffffff813204ad>] idr_alloc+0x4d/0xf0
[<ffffffffa051727d>] cma_alloc_port+0x4d/0xa0 [rdma_cm]
[<ffffffffa0517cbe>] rdma_bind_addr+0x2ae/0x5b0 [rdma_cm]
[<ffffffffa09d8083>] rds_ib_laddr_check+0x83/0x2c0 [ksplice_6l2xst5i_rds_rdma_new]
[<ffffffffa05f892b>] rds_trans_get_preferred+0x5b/0xa0 [rds]
[<ffffffffa05f09f2>] rds_bind+0x212/0x280 [rds]
[<ffffffff815b4016>] SYSC_bind+0xe6/0x120
[<ffffffff815b4d3e>] SyS_bind+0xe/0x10
[<ffffffff816b031a>] system_call_fastpath+0x18/0xd4

To avoid these excessive calls to rdma_bind_addr(), we optimize
rds_ib_laddr_check() by simply checking if the address in question has
been used before. The rds_rdma module keeps track of addresses
associated with IB devices, and the function rds_ib_get_device() is
used to determine if the address already has been qualified as a valid
local address. If not found, we call the legacy rds_ib_laddr_check(),
now renamed to rds_ib_laddr_check_cm().

Signed-off-by: Håkon Bugge <haakon.bugge@oracle.com>
Signed-off-by: Somasundaram Krishnasamy <somasundaram.krishnasamy@oracle.com>
Signed-off-by: Gerd Rausch <gerd.rausch@oracle.com>
Signed-off-by: Allison Henderson <achender@kernel.org>
Link: https://patch.msgid.link/20260408080420.540032-2-achender@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

0 files changed, 0 insertions, 0 deletions