Skip to content

L
linux
Switch branch/tag
  1. 06 Feb, 2023 26 commits
  3. 04 Feb, 2023 14 commits
    • David S. Miller's avatar
      Merge branch 'net-smc-parallelism' · 042b7858
      David S. Miller authored 
      D. Wythe says:

====================
net/smc: optimize the parallelism of SMC-R connections

This patch set attempts to optimize the parallelism of SMC-R connections,
mainly to reduce unnecessary blocking on locks, and to fix exceptions that
occur after thoses optimization.

According to Off-CPU graph, SMC worker's off-CPU as that:

smc_close_passive_work                  (1.09%)
        smcr_buf_unuse                  (1.08%)
                smc_llc_flow_initiate   (1.02%)

smc_listen_work                         (48.17%)
        __mutex_lock.isra.11            (47.96%)

An ideal SMC-R connection process should only block on the IO events
of the network, but it's quite clear that the SMC-R connection now is
queued on the lock most of the time.

The goal of this patchset is to achieve our ideal situation where
network IO events are blocked for the majority of the connection lifetime.

There are three big locks here:

1. smc_client_lgr_pending & smc_server_lgr_pending

2. llc_conf_mutex

3. rmbs_lock & sndbufs_lock

And an implementation issue:

1. confirm/delete rkey msg can't be sent concurrently while
protocol allows indeed.

Unfortunately,The above problems together affect the parallelism of
SMC-R connection. If any of them are not solved. our goal cannot
be achieved.

After this patch set, we can get a quite ideal off-CPU graph as
following:

smc_close_passive_work                                  (41.58%)
        smcr_buf_unuse                                  (41.57%)
                smc_llc_do_delete_rkey                  (41.57%)

smc_listen_work                                         (39.10%)
        smc_clc_wait_msg                                (13.18%)
                tcp_recvmsg_locked                      (13.18)
        smc_listen_find_device                          (25.87%)
                smcr_lgr_reg_rmbs                       (25.87%)
                        smc_llc_do_confirm_rkey         (25.87%)

We can see that most of the waiting times are waiting for network IO
events. This also has a certain performance improvement on our
short-lived conenction wrk/nginx benchmark test:

+--------------+------+------+-------+--------+------+--------+
|conns/qps     |c4    | c8   |  c16  |  c32   | c64  |  c200  |
+--------------+------+------+-------+--------+------+--------+
|SMC-R before  |9.7k  | 10k  |  10k  |  9.9k  | 9.1k |  8.9k  |
+--------------+------+------+-------+--------+------+--------+
|SMC-R now     |13k   | 19k  |  18k  |  16k   | 15k  |  12k   |
+--------------+------+------+-------+--------+------+--------+
|TCP           |15k   | 35k  |  51k  |  80k   | 100k |  162k  |
+--------------+------+------+-------+--------+------+--------+

The reason why the benefit is not obvious after the number of connections
has increased dues to workqueue. If we try to change workqueue to UNBOUND,
we can obtain at least 4-5 times performance improvement, reach up to half
of TCP. However, this is not an elegant solution, the optimization of it
will be much more complicated. But in any case, we will submit relevant
optimization patches as soon as possible.

Please note that the premise here is that the lock related problem
must be solved first, otherwise, no matter how we optimize the workqueue,
there won't be much improvement.

Because there are a lot of related changes to the code, if you have
any questions or suggestions, please let me know.

Thanks
D. Wythe

v1 -> v2:

1. Fix panic in SMC-D scenario
2. Fix lnkc related hashfn calculation exception, caused by operator
priority
3. Only wake up one connection if the lnk is not active
4. Delete obsolete unlock logic in smc_listen_work()
5. PATCH format, do Reverse Christmas tree
6. PATCH format, change all xxx_lnk_xxx function to xxx_link_xxx
7. PATCH format, add correct fix tag for the patches for fixes.
8. PATCH format, fix some spelling error
9. PATCH format, rename slow to do_slow

v2 -> v3:

1. add SMC-D support, remove the concept of link cluster since SMC-D has
no link at all. Replace it by lgr decision maker, who provides suggestions
to SMC-D and SMC-R on whether to create new link group.

2. Fix the corruption problem described by PATCH 'fix application
data exception' on SMC-D.

v3 -> v4:

1. Fix panic caused by uninitialization map.

v4 -> v5:

1. Make SMC-D buf creation be serial to avoid Potential error
2. Add a flag to synchronize the success of the first contact
with the ready of the link group, including SMC-D and SMC-R.
3. Fixed possible reference count leak in smc_llc_flow_start().
4. reorder the patch, make bugfix PATCH be ahead.

v5 -> v6:

1. Separate the bugfix patches to make it independent.
2. Merge patch 'fix SMC_CLC_DECL_ERR_REGRMB without smc_server_lgr_pending'
with patch 'remove locks smc_client_lgr_pending and smc_server_lgr_pending'
3. Format code styles, including alignment and reverse christmas tree
style.
4. Fix a possible memory leak in smc_llc_rmt_delete_rkey()
and smc_llc_rmt_conf_rkey().

v6 -> v7:

1. Discard patch attempting to remove global locks
2. Discard patch attempting make confirm/delete rkey process concurrently
====================
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
      042b7858
    • D. Wythe's avatar
      net/smc: replace mutex rmbs_lock and sndbufs_lock with rw_semaphore · aff7bfed
      D. Wythe authored 
      It's clear that rmbs_lock and sndbufs_lock are aims to protect the
rmbs list or the sndbufs list.

During connection establieshment, smc_buf_get_slot() will always
be invoked, and it only performs read semantics in rmbs list and
sndbufs list.

Based on the above considerations, we replace mutex with rw_semaphore.
Only smc_buf_get_slot() use down_read() to allow smc_buf_get_slot()
run concurrently, other part use down_write() to keep exclusive
semantics.
Signed-off-by: default avatarD. Wythe <alibuda@linux.alibaba.com>
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
      aff7bfed
    • D. Wythe's avatar
      net/smc: reduce unnecessary blocking in smcr_lgr_reg_rmbs() · 4da68744
      D. Wythe authored 
      Unlike smc_buf_create() and smcr_buf_unuse(), smcr_lgr_reg_rmbs() is
exclusive when assigned rmb_desc was not registered, although it can be
executed in parallel when assigned rmb_desc was registered already
and only performs read semtamics on it. Hence, we can not simply replace
it with read semaphore.

The idea here is that if the assigned rmb_desc was registered already,
use read semaphore to protect the critical section, once the assigned
rmb_desc was not registered, keep using keep write semaphore still
to keep its exclusivity.

Thanks to the reusable features of rmb_desc, which allows us to execute
in parallel in most cases.
Signed-off-by: default avatarD. Wythe <alibuda@linux.alibaba.com>
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
      4da68744
    • D. Wythe's avatar
      net/smc: use read semaphores to reduce unnecessary blocking in smc_buf_create() & smcr_buf_unuse() · f6421014
      D. Wythe authored 
      Following is part of Off-CPU graph during frequent SMC-R short-lived
processing:

process_one_work				(51.19%)
smc_close_passive_work			(28.36%)
	smcr_buf_unuse				(28.34%)
	rwsem_down_write_slowpath		(28.22%)

smc_listen_work				(22.83%)
	smc_clc_wait_msg			(1.84%)
	smc_buf_create				(20.45%)
		smcr_buf_map_usable_links
		rwsem_down_write_slowpath	(20.43%)
	smcr_lgr_reg_rmbs			(0.53%)
		rwsem_down_write_slowpath	(0.43%)
		smc_llc_do_confirm_rkey		(0.08%)

We can clearly see that during the connection establishment time,
waiting time of connections is not on IO, but on llc_conf_mutex.

What is more important, the core critical area (smcr_buf_unuse() &
smc_buf_create()) only perfroms read semantics on links, we can
easily replace it with read semaphore.
Signed-off-by: default avatarD. Wythe <alibuda@linux.alibaba.com>
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
      f6421014
    • D. Wythe's avatar
      net/smc: llc_conf_mutex refactor, replace it with rw_semaphore · b5dd4d69
      D. Wythe authored 
      llc_conf_mutex was used to protect links and link related configurations
in the same link group, for example, add or delete links. However,
in most cases, the protected critical area has only read semantics and
with no write semantics at all, such as obtaining a usable link or an
available rmb_desc.

This patch do simply code refactoring, replace mutex with rw_semaphore,
replace mutex_lock with down_write and replace mutex_unlock with
up_write.

Theoretically, this replacement is equivalent, but after this patch,
we can distinguish lock granularity according to different semantics
of critical areas.
Signed-off-by: default avatarD. Wythe <alibuda@linux.alibaba.com>
Signed-off-by: default avatarDavid S. Miller <davem@davemloft.net>
      b5dd4d69
    • Jakub Kicinski's avatar
      Merge branch 'updates-to-enetc-txq-management' · 88c940cc
      Jakub Kicinski authored 
      Vladimir Oltean says:

====================
Updates to ENETC TXQ management

The set ensures that the number of TXQs given by enetc to the network
stack (mqprio or TX hashing) + the number of TXQs given to XDP never
exceeds the number of available TXQs.

These are the first 4 patches of series "[v5,net-next,00/17] ENETC
mqprio/taprio cleanup" from here:
https://patchwork.kernel.org/project/netdevbpf/cover/20230202003621.2679603-1-vladimir.oltean@nxp.com/

There is no change in this version compared to there. I split them off
because this contains a fix for net-next and it would be good if it
could go in quickly. I also did it to reduce the patch count of that
other series, if I need to respin it again.
====================

Link: https://lore.kernel.org/r/20230203001116.3814809-1-vladimir.oltean@nxp.comSigned-off-by: default avatarJakub Kicinski <kuba@kernel.org>
      88c940cc
    • Vladimir Oltean's avatar
      net: enetc: ensure we always have a minimum number of TXQs for stack · 800db2d1
      Vladimir Oltean authored 
      Currently it can happen that an mqprio qdisc is installed with num_tc 8,
and this will reserve 8 (out of 8) TXQs for the network stack. Then we
can attach an XDP program, and this will crop 2 TXQs, leaving just 6 for
mqprio. That's not what the user requested, and we should fail it.

On the other hand, if mqprio isn't requested, we still give the 8 TXQs
to the network stack (with hashing among a single traffic class), but
then, cropping 2 TXQs for XDP is fine, because the user didn't
explicitly ask for any number of TXQs, so no expectations are violated.

Simply put, the logic that mqprio should impose a minimum number of TXQs
for the network never existed. Let's say (more or less arbitrarily) that
without mqprio, the driver expects a minimum number of TXQs equal to the
number of CPUs (on NXP LS1028A, that is either 1, or 2). And with mqprio,
mqprio gives the minimum required number of TXQs.
Signed-off-by: default avatarVladimir Oltean <vladimir.oltean@nxp.com>
Reviewed-by: default avatarSimon Horman <simon.horman@corigine.com>
Signed-off-by: default avatarJakub Kicinski <kuba@kernel.org>
      800db2d1
    • Vladimir Oltean's avatar
      net: enetc: recalculate num_real_tx_queues when XDP program attaches · 4ea1dd74
      Vladimir Oltean authored 
      Since the blamed net-next commit, enetc_setup_xdp_prog() no longer goes
through enetc_open(), and therefore, the function which was supposed to
detect whether a BPF program exists (in order to crop some TX queues
from network stack usage), enetc_num_stack_tx_queues(), no longer gets
called.

We can move the netif_set_real_num_rx_queues() call to enetc_alloc_msix()
(probe time), since it is a runtime invariant. We can do the same thing
with netif_set_real_num_tx_queues(), and let enetc_reconfigure_xdp_cb()
explicitly recalculate and change the number of stack TX queues.

Fixes: c33bfaf9 ("net: enetc: set up XDP program under enetc_reconfigure()")
Signed-off-by: default avatarVladimir Oltean <vladimir.oltean@nxp.com>
Reviewed-by: default avatarSimon Horman <simon.horman@corigine.com>
Signed-off-by: default avatarJakub Kicinski <kuba@kernel.org>
      4ea1dd74
    • Vladimir Oltean's avatar
      net: enetc: allow the enetc_reconfigure() callback to fail · 46a0ecf9
      Vladimir Oltean authored 
      enetc_reconfigure() was modified in commit c33bfaf9 ("net: enetc:
set up XDP program under enetc_reconfigure()") to take an optional
callback that runs while the netdev is down, but this callback currently
cannot fail.

Code up the error handling so that the interface is restarted with the
old resources if the callback fails.
Signed-off-by: default avatarVladimir Oltean <vladimir.oltean@nxp.com>
Reviewed-by: default avatarSimon Horman <simon.horman@corigine.com>
Signed-off-by: default avatarJakub Kicinski <kuba@kernel.org>
      46a0ecf9
    • Vladimir Oltean's avatar
      net: enetc: simplify enetc_num_stack_tx_queues() · 1c81a9b3
      Vladimir Oltean authored 
      We keep a pointer to the xdp_prog in the private netdev structure as
well; what's replicated per RX ring is done so just for more convenient
access from the NAPI poll procedure.

Simplify enetc_num_stack_tx_queues() by looking at priv->xdp_prog rather
than iterating through the information replicated per RX ring.
Signed-off-by: default avatarVladimir Oltean <vladimir.oltean@nxp.com>
Reviewed-by: default avatarSimon Horman <simon.horman@corigine.com>
Signed-off-by: default avatarJakub Kicinski <kuba@kernel.org>
      1c81a9b3
    • Jakub Kicinski's avatar
      Merge branch 'raw-add-drop-reasons-and-use-another-hash-function' · 8788260e
      Jakub Kicinski authored 
      Eric Dumazet says:

====================
raw: add drop reasons and use another hash function

Two first patches add drop reasons to raw input processing.

Last patch spreads RAW sockets in the shared hash tables
to avoid long hash buckets in some cases.
====================

Link: https://lore.kernel.org/r/20230202094100.3083177-1-edumazet@google.comSigned-off-by: default avatarJakub Kicinski <kuba@kernel.org>
      8788260e
    • Eric Dumazet's avatar
      raw: use net_hash_mix() in hash function · 6579f5ba
      Eric Dumazet authored 
      Some applications seem to rely on RAW sockets.

If they use private netns, we can avoid piling all RAW
sockets bound to a given protocol into a single bucket.

Also place (struct raw_hashinfo).lock into its own
cache line to limit false sharing.

Alternative would be to have per-netns hashtables,
but this seems too expensive for most netns
where RAW sockets are not used.
Signed-off-by: default avatarEric Dumazet <edumazet@google.com>
Signed-off-by: default avatarJakub Kicinski <kuba@kernel.org>
      6579f5ba
    • Eric Dumazet's avatar
      ipv4: raw: add drop reasons · 42186e6c
      Eric Dumazet authored 
      Use existing helpers and drop reason codes for RAW input path.
Signed-off-by: default avatarEric Dumazet <edumazet@google.com>
Signed-off-by: default avatarJakub Kicinski <kuba@kernel.org>
      42186e6c
    • Eric Dumazet's avatar
      ipv6: raw: add drop reasons · 8d8ebd77
      Eric Dumazet authored 
      Use existing helpers and drop reason codes for RAW input path.
Signed-off-by: default avatarEric Dumazet <edumazet@google.com>
Signed-off-by: default avatarJakub Kicinski <kuba@kernel.org>
      8d8ebd77
GitLab Nexedi Edition | About GitLab | About Nexedi | 沪ICP备2021021310号-2 | 沪ICP备2021021310号-7