Hi all,
commit 0227f058aa29 ("net/smc: Unbind r/w buffer size from clcsock
and make them tunable") started to derive the effective buffer size for
SMC connections inconsistently in case a TCP fallback was used and
memory consumption of SMC with the default settings was doubled when
a connection negotiated SMC. That was not what we want.
This series consolidates the resulting effective buffer size that is
used with SMC sockets, which is based on Jan Karcher's effort (see
[1]). For all TCP exchanges (in particular in case of a fall back when
no SMC connection was possible) the values from net.ipv4.tcp_[rw]mem
are used. If SMC succeeds in establishing a SMC connection, the newly
introduced values from net.smc.[rw]mem are used.
net.smc.[rw]mem is initialized to 64kB, respectively. Internal test
have show this to be a good compromise between throughput/latency
and memory consumption. Also net.smc.[rw]mem is now decoupled completely
from any tuning through net.ipv4.tcp_[rw]mem.
If a user chose to tune a socket's receive or send buffer size with
setsockopt, this tuning is now consistently applied to either fall-back
TCP or proper SMC connections over the socket.
Thanks,
Gerd
v2 - v3:
- Rebase to and resolve conflict of second patch with latest net/master.
v1 - v2:
- In second patch, use sock_net() helper as suggested by Tony and demanded
by kernel test robot.
Gerd Bayer (2):
net/smc: Fix setsockopt and sysctl to specify same buffer size again
net/smc: Use correct buffer sizes when switching between TCP and SMC
net/smc/af_smc.c | 77 ++++++++++++++++++++++++++++++--------------
net/smc/smc.h | 2 +-
net/smc/smc_clc.c | 4 +--
net/smc/smc_core.c | 25 +++++++-------
net/smc/smc_sysctl.c | 10 ++++--
5 files changed, 76 insertions(+), 42 deletions(-)
base-commit: 1733d0be68ab1b89358a3b0471ef425fd61de7c5
--
2.41.0
On Fri, 2023-08-04 at 18:30 +0200, Gerd Bayer wrote:
> Hi all,
>
> commit 0227f058aa29 ("net/smc: Unbind r/w buffer size from clcsock
> and make them tunable") started to derive the effective buffer size
> for
> SMC connections inconsistently in case a TCP fallback was used and
> memory consumption of SMC with the default settings was doubled when
> a connection negotiated SMC. That was not what we want.
>
> This series consolidates the resulting effective buffer size that is
> used with SMC sockets, which is based on Jan Karcher's effort (see
> [1]). For all TCP exchanges (in particular in case of a fall back
> when
> no SMC connection was possible) the values from net.ipv4.tcp_[rw]mem
> are used. If SMC succeeds in establishing a SMC connection, the newly
> introduced values from net.smc.[rw]mem are used.
>
> net.smc.[rw]mem is initialized to 64kB, respectively. Internal test
> have show this to be a good compromise between throughput/latency
> and memory consumption. Also net.smc.[rw]mem is now decoupled
> completely
> from any tuning through net.ipv4.tcp_[rw]mem.
>
> If a user chose to tune a socket's receive or send buffer size with
> setsockopt, this tuning is now consistently applied to either fall-
> back
> TCP or proper SMC connections over the socket.
>
> Thanks,
> Gerd
>
> v2 - v3:
> - Rebase to and resolve conflict of second patch with latest
> net/master.
> v1 - v2:
> - In second patch, use sock_net() helper as suggested by Tony and
> demanded
> by kernel test robot.
>
>
> Gerd Bayer (2):
> net/smc: Fix setsockopt and sysctl to specify same buffer size
> again
> net/smc: Use correct buffer sizes when switching between TCP and
> SMC
>
> net/smc/af_smc.c | 77 ++++++++++++++++++++++++++++++------------
> --
> net/smc/smc.h | 2 +-
> net/smc/smc_clc.c | 4 +--
> net/smc/smc_core.c | 25 +++++++-------
> net/smc/smc_sysctl.c | 10 ++++--
> 5 files changed, 76 insertions(+), 42 deletions(-)
>
>
> base-commit: 1733d0be68ab1b89358a3b0471ef425fd61de7c5
Oh boy,
this should have gone as v3 against "net" instead of "net-next".
Resending ASAP.
Sorry for the noise,
Gerd
On Fri, 04 Aug 2023 18:55:26 +0200 Gerd Bayer wrote:
> this should have gone as v3 against "net" instead of "net-next".
> Resending ASAP.
>
> Sorry for the noise,
Less apologizing more reading of the rules, please.
https://www.kernel.org/doc/html/next/process/maintainer-netdev.html
Commit 0227f058aa29 ("net/smc: Unbind r/w buffer size from clcsock
and make them tunable") introduced the net.smc.rmem and net.smc.wmem
sysctls to specify the size of buffers to be used for SMC type
connections. This created a regression for users that specified the
buffer size via setsockopt() as the effective buffer size was now
doubled.
Re-introduce the division by 2 in the SMC buffer create code and level
this out by duplicating the net.smc.[rw]mem values used for initializing
sk_rcvbuf/sk_sndbuf at socket creation time. This gives users of both
methods (setsockopt or sysctl) the effective buffer size that they
expect.
Initialize net.smc.[rw]mem from its own constant of 64kB, respectively.
Internal performance tests show that this value is a good compromise
between throughput/latency and memory consumption. Also, this decouples
it from any tuning that was done to net.ipv4.tcp_[rw]mem[1] before the
module for SMC protocol was loaded. Check that no more than INT_MAX / 2
is assigned to net.smc.[rw]mem, in order to avoid any overflow condition
when that is doubled for use in sk_sndbuf or sk_rcvbuf.
While at it, drop the confusing sk_buf_size variable from
__smc_buf_create and name "compressed" buffer size variables more
consistently.
Background:
Before the commit mentioned above, SMC's buffer allocator in
__smc_buf_create() always used half of the sockets' sk_rcvbuf/sk_sndbuf
value as initial value to search for appropriate buffers. If the search
resorted to using a bigger buffer when all buffers of the specified
size were busy, the duplicate of the used effective buffer size is
stored back to sk_rcvbuf/sk_sndbuf.
When available, buffers of exactly the size that a user had specified as
input to setsockopt() were used, despite setsockopt()'s documentation in
"man 7 socket" talking of a mandatory duplication:
[...]
SO_SNDBUF
Sets or gets the maximum socket send buffer in bytes.
The kernel doubles this value (to allow space for book‐
keeping overhead) when it is set using setsockopt(2),
and this doubled value is returned by getsockopt(2).
The default value is set by the
/proc/sys/net/core/wmem_default file and the maximum
allowed value is set by the /proc/sys/net/core/wmem_max
file. The minimum (doubled) value for this option is
2048.
[...]
Fixes: 0227f058aa29 ("net/smc: Unbind r/w buffer size from clcsock and make them tunable")
Co-developed-by: Jan Karcher <[email protected]>
Signed-off-by: Jan Karcher <[email protected]>
Reviewed-by: Wenjia Zhang <[email protected]>
Reviewed-by: Tony Lu <[email protected]>
Signed-off-by: Gerd Bayer <[email protected]>
---
net/smc/af_smc.c | 4 ++--
net/smc/smc.h | 2 +-
net/smc/smc_clc.c | 4 ++--
net/smc/smc_core.c | 25 ++++++++++++-------------
net/smc/smc_sysctl.c | 10 ++++++++--
5 files changed, 25 insertions(+), 20 deletions(-)
diff --git a/net/smc/af_smc.c b/net/smc/af_smc.c
index 0c013d2b5d8f..5b878e523abf 100644
--- a/net/smc/af_smc.c
+++ b/net/smc/af_smc.c
@@ -378,8 +378,8 @@ static struct sock *smc_sock_alloc(struct net *net, struct socket *sock,
sk->sk_state = SMC_INIT;
sk->sk_destruct = smc_destruct;
sk->sk_protocol = protocol;
- WRITE_ONCE(sk->sk_sndbuf, READ_ONCE(net->smc.sysctl_wmem));
- WRITE_ONCE(sk->sk_rcvbuf, READ_ONCE(net->smc.sysctl_rmem));
+ WRITE_ONCE(sk->sk_sndbuf, 2 * READ_ONCE(net->smc.sysctl_wmem));
+ WRITE_ONCE(sk->sk_rcvbuf, 2 * READ_ONCE(net->smc.sysctl_rmem));
smc = smc_sk(sk);
INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
INIT_WORK(&smc->connect_work, smc_connect_work);
diff --git a/net/smc/smc.h b/net/smc/smc.h
index 2eeea4cdc718..1f2b912c43d1 100644
--- a/net/smc/smc.h
+++ b/net/smc/smc.h
@@ -161,7 +161,7 @@ struct smc_connection {
struct smc_buf_desc *sndbuf_desc; /* send buffer descriptor */
struct smc_buf_desc *rmb_desc; /* RMBE descriptor */
- int rmbe_size_short;/* compressed notation */
+ int rmbe_size_comp; /* compressed notation */
int rmbe_update_limit;
/* lower limit for consumer
* cursor update
diff --git a/net/smc/smc_clc.c b/net/smc/smc_clc.c
index b9b8b07aa702..c90d9e5dda54 100644
--- a/net/smc/smc_clc.c
+++ b/net/smc/smc_clc.c
@@ -1007,7 +1007,7 @@ static int smc_clc_send_confirm_accept(struct smc_sock *smc,
clc->d0.gid =
conn->lgr->smcd->ops->get_local_gid(conn->lgr->smcd);
clc->d0.token = conn->rmb_desc->token;
- clc->d0.dmbe_size = conn->rmbe_size_short;
+ clc->d0.dmbe_size = conn->rmbe_size_comp;
clc->d0.dmbe_idx = 0;
memcpy(&clc->d0.linkid, conn->lgr->id, SMC_LGR_ID_SIZE);
if (version == SMC_V1) {
@@ -1050,7 +1050,7 @@ static int smc_clc_send_confirm_accept(struct smc_sock *smc,
clc->r0.qp_mtu = min(link->path_mtu, link->peer_mtu);
break;
}
- clc->r0.rmbe_size = conn->rmbe_size_short;
+ clc->r0.rmbe_size = conn->rmbe_size_comp;
clc->r0.rmb_dma_addr = conn->rmb_desc->is_vm ?
cpu_to_be64((uintptr_t)conn->rmb_desc->cpu_addr) :
cpu_to_be64((u64)sg_dma_address
diff --git a/net/smc/smc_core.c b/net/smc/smc_core.c
index 3f465faf2b68..6b78075404d7 100644
--- a/net/smc/smc_core.c
+++ b/net/smc/smc_core.c
@@ -2309,31 +2309,30 @@ static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
struct smc_connection *conn = &smc->conn;
struct smc_link_group *lgr = conn->lgr;
struct list_head *buf_list;
- int bufsize, bufsize_short;
+ int bufsize, bufsize_comp;
struct rw_semaphore *lock; /* lock buffer list */
bool is_dgraded = false;
- int sk_buf_size;
if (is_rmb)
/* use socket recv buffer size (w/o overhead) as start value */
- sk_buf_size = smc->sk.sk_rcvbuf;
+ bufsize = smc->sk.sk_rcvbuf / 2;
else
/* use socket send buffer size (w/o overhead) as start value */
- sk_buf_size = smc->sk.sk_sndbuf;
+ bufsize = smc->sk.sk_sndbuf / 2;
- for (bufsize_short = smc_compress_bufsize(sk_buf_size, is_smcd, is_rmb);
- bufsize_short >= 0; bufsize_short--) {
+ for (bufsize_comp = smc_compress_bufsize(bufsize, is_smcd, is_rmb);
+ bufsize_comp >= 0; bufsize_comp--) {
if (is_rmb) {
lock = &lgr->rmbs_lock;
- buf_list = &lgr->rmbs[bufsize_short];
+ buf_list = &lgr->rmbs[bufsize_comp];
} else {
lock = &lgr->sndbufs_lock;
- buf_list = &lgr->sndbufs[bufsize_short];
+ buf_list = &lgr->sndbufs[bufsize_comp];
}
- bufsize = smc_uncompress_bufsize(bufsize_short);
+ bufsize = smc_uncompress_bufsize(bufsize_comp);
/* check for reusable slot in the link group */
- buf_desc = smc_buf_get_slot(bufsize_short, lock, buf_list);
+ buf_desc = smc_buf_get_slot(bufsize_comp, lock, buf_list);
if (buf_desc) {
buf_desc->is_dma_need_sync = 0;
SMC_STAT_RMB_SIZE(smc, is_smcd, is_rmb, bufsize);
@@ -2377,8 +2376,8 @@ static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
if (is_rmb) {
conn->rmb_desc = buf_desc;
- conn->rmbe_size_short = bufsize_short;
- smc->sk.sk_rcvbuf = bufsize;
+ conn->rmbe_size_comp = bufsize_comp;
+ smc->sk.sk_rcvbuf = bufsize * 2;
atomic_set(&conn->bytes_to_rcv, 0);
conn->rmbe_update_limit =
smc_rmb_wnd_update_limit(buf_desc->len);
@@ -2386,7 +2385,7 @@ static int __smc_buf_create(struct smc_sock *smc, bool is_smcd, bool is_rmb)
smc_ism_set_conn(conn); /* map RMB/smcd_dev to conn */
} else {
conn->sndbuf_desc = buf_desc;
- smc->sk.sk_sndbuf = bufsize;
+ smc->sk.sk_sndbuf = bufsize * 2;
atomic_set(&conn->sndbuf_space, bufsize);
}
return 0;
diff --git a/net/smc/smc_sysctl.c b/net/smc/smc_sysctl.c
index b6f79fabb9d3..0b2a957ca5f5 100644
--- a/net/smc/smc_sysctl.c
+++ b/net/smc/smc_sysctl.c
@@ -21,6 +21,10 @@
static int min_sndbuf = SMC_BUF_MIN_SIZE;
static int min_rcvbuf = SMC_BUF_MIN_SIZE;
+static int max_sndbuf = INT_MAX / 2;
+static int max_rcvbuf = INT_MAX / 2;
+static const int net_smc_wmem_init = (64 * 1024);
+static const int net_smc_rmem_init = (64 * 1024);
static struct ctl_table smc_table[] = {
{
@@ -53,6 +57,7 @@ static struct ctl_table smc_table[] = {
.mode = 0644,
.proc_handler = proc_dointvec_minmax,
.extra1 = &min_sndbuf,
+ .extra2 = &max_sndbuf,
},
{
.procname = "rmem",
@@ -61,6 +66,7 @@ static struct ctl_table smc_table[] = {
.mode = 0644,
.proc_handler = proc_dointvec_minmax,
.extra1 = &min_rcvbuf,
+ .extra2 = &max_rcvbuf,
},
{ }
};
@@ -88,8 +94,8 @@ int __net_init smc_sysctl_net_init(struct net *net)
net->smc.sysctl_autocorking_size = SMC_AUTOCORKING_DEFAULT_SIZE;
net->smc.sysctl_smcr_buf_type = SMCR_PHYS_CONT_BUFS;
net->smc.sysctl_smcr_testlink_time = SMC_LLC_TESTLINK_DEFAULT_TIME;
- WRITE_ONCE(net->smc.sysctl_wmem, READ_ONCE(net->ipv4.sysctl_tcp_wmem[1]));
- WRITE_ONCE(net->smc.sysctl_rmem, READ_ONCE(net->ipv4.sysctl_tcp_rmem[1]));
+ WRITE_ONCE(net->smc.sysctl_wmem, net_smc_wmem_init);
+ WRITE_ONCE(net->smc.sysctl_rmem, net_smc_rmem_init);
return 0;
--
2.41.0