If the device used by SMC-D supports merging local sndbuf to peer DMB,
then create sndbuf descriptor and attach it to peer DMB once peer
token is obtained, and detach and free the sndbuf descriptor when the
connection is freed.
Signed-off-by: Wen Gu <guwen@linux.alibaba.com>
Reviewed-by: Wenjia Zhang <wenjia@linux.ibm.com>
Reviewed-and-tested-by: Jan Karcher <jaka@linux.ibm.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
According to latest release of SMCv2.1[1], the term 'virtual ISM' has
been changed to 'Emulated-ISM' to avoid the ambiguity of the word
'virtual' in different contexts. So the names or comments in the code
need be modified accordingly.
[1] https://www.ibm.com/support/pages/node/7112343
Signed-off-by: Wen Gu <guwen@linux.alibaba.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Reviewed-by: Wenjia Zhang <wenjia@linux.ibm.com>
Link: https://lore.kernel.org/r/20240205033317.127269-1-guwen@linux.alibaba.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The logic to determine if SMC-D link group matches is incorrect. The
correct logic should be that it only returns true when the GID is the
same, and the SMC-D device is the same and the extended GID is the same
(in the case of virtual ISM).
It can be fixed by adding brackets around the conditional (or ternary)
operator expression. But for better readability and maintainability, it
has been changed to an if-else statement.
Reported-by: Matthew Rosato <mjrosato@linux.ibm.com>
Closes: https://lore.kernel.org/r/13579588-eb9d-4626-a063-c0b77ed80f11@linux.ibm.com
Fixes: b40584d145 ("net/smc: compatible with 128-bits extended GID of virtual ISM device")
Link: https://lore.kernel.org/r/13579588-eb9d-4626-a063-c0b77ed80f11@linux.ibm.com
Signed-off-by: Wen Gu <guwen@linux.alibaba.com>
Reviewed-by: Alexandra Winter <wintera@linux.ibm.com>
Link: https://lore.kernel.org/r/20240125123916.77928-1-guwen@linux.alibaba.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Virtual ISM devices introduced in SMCv2.1 requires a 128 bit extended
GID vs. the existing ISM 64bit GID. So the 2nd 64 bit of extended GID
should be included in SMC-D linkgroup netlink attribute as well.
Signed-off-by: Wen Gu <guwen@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
According to virtual ISM support feature defined by SMCv2.1, GIDs of
virtual ISM device are UUIDs defined by RFC4122, which are 128-bits
long. So some adaptation work is required. And note that the GIDs of
existing platform firmware ISM devices still remain 64-bits long.
Signed-off-by: Wen Gu <guwen@linux.alibaba.com>
Reviewed-by: Alexandra Winter <wintera@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Add SMC_NLA_LGR_R_V2_MAX_CONNS and SMC_NLA_LGR_R_V2_MAX_LINKS
to SMCR v2 linkgroup netlink attribute SMC_NLA_LGR_R_V2 for
linkgroup's detail info showing.
Signed-off-by: Guangguan Wang <guangguan.wang@linux.alibaba.com>
Reviewed-by: Jan Karcher <jaka@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Support max links per lgr negotiation in clc handshake for SMCR v2.1,
which is one of smc v2.1 features. Server makes decision for the final
value of max links based on the client preferred max links and
self-preferred max links. Here use the minimum value of the client
preferred max links and server preferred max links.
Client Server
Proposal(max links(client preferred))
-------------------------------------->
Accept(max links(accepted value))
accepted value=min(client preferred, server preferred)
<-------------------------------------
Confirm(max links(accepted value))
------------------------------------->
Signed-off-by: Guangguan Wang <guangguan.wang@linux.alibaba.com>
Reviewed-by: Tony Lu <tonylu@linux.alibaba.com>
Reviewed-by: Jan Karcher <jaka@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Support max connections per lgr negotiation for SMCR v2.1,
which is one of smc v2.1 features. Server makes decision for
the final value of max conns based on the client preferred
max conns and self-preferred max conns. Here use the minimum
value of client preferred max conns and server preferred max
conns.
Client Server
Proposal(max conns(client preferred))
------------------------------------>
Accept(max conns(accepted value))
accepted value=min(client preferred, server preferred)
<-----------------------------------
Confirm(max conns(accepted value))
----------------------------------->
Signed-off-by: Guangguan Wang <guangguan.wang@linux.alibaba.com>
Reviewed-by: Tony Lu <tonylu@linux.alibaba.com>
Reviewed-by: Jan Karcher <jaka@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Commit 0227f058aa ("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: 0227f058aa ("net/smc: Unbind r/w buffer size from clcsock and make them tunable")
Co-developed-by: Jan Karcher <jaka@linux.ibm.com>
Signed-off-by: Jan Karcher <jaka@linux.ibm.com>
Reviewed-by: Wenjia Zhang <wenjia@linux.ibm.com>
Reviewed-by: Tony Lu <tonylu@linux.alibaba.com>
Signed-off-by: Gerd Bayer <gbayer@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We found a crash when using SMCRv2 with 2 Mellanox ConnectX-4. It
can be reproduced by:
- smc_run nginx
- smc_run wrk -t 32 -c 500 -d 30 http://<ip>:<port>
BUG: kernel NULL pointer dereference, address: 0000000000000014
#PF: supervisor read access in kernel mode
#PF: error_code(0x0000) - not-present page
PGD 8000000108713067 P4D 8000000108713067 PUD 151127067 PMD 0
Oops: 0000 [#1] PREEMPT SMP PTI
CPU: 4 PID: 2441 Comm: kworker/4:249 Kdump: loaded Tainted: G W E 6.4.0-rc1+ #42
Workqueue: smc_hs_wq smc_listen_work [smc]
RIP: 0010:smc_clc_send_confirm_accept+0x284/0x580 [smc]
RSP: 0018:ffffb8294b2d7c78 EFLAGS: 00010a06
RAX: ffff8f1873238880 RBX: ffffb8294b2d7dc8 RCX: 0000000000000000
RDX: 00000000000000b4 RSI: 0000000000000001 RDI: 0000000000b40c00
RBP: ffffb8294b2d7db8 R08: ffff8f1815c5860c R09: 0000000000000000
R10: 0000000000000400 R11: 0000000000000000 R12: ffff8f1846f56180
R13: ffff8f1815c5860c R14: 0000000000000001 R15: 0000000000000001
FS: 0000000000000000(0000) GS:ffff8f1aefd00000(0000) knlGS:0000000000000000
CS: 0010 DS: 0000 ES: 0000 CR0: 0000000080050033
CR2: 0000000000000014 CR3: 00000001027a0001 CR4: 00000000003706e0
DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
Call Trace:
<TASK>
? mlx5_ib_map_mr_sg+0xa1/0xd0 [mlx5_ib]
? smcr_buf_map_link+0x24b/0x290 [smc]
? __smc_buf_create+0x4ee/0x9b0 [smc]
smc_clc_send_accept+0x4c/0xb0 [smc]
smc_listen_work+0x346/0x650 [smc]
? __schedule+0x279/0x820
process_one_work+0x1e5/0x3f0
worker_thread+0x4d/0x2f0
? __pfx_worker_thread+0x10/0x10
kthread+0xe5/0x120
? __pfx_kthread+0x10/0x10
ret_from_fork+0x2c/0x50
</TASK>
During the CLC handshake, server sequentially tries available SMCRv2
and SMCRv1 devices in smc_listen_work().
If an SMCRv2 device is found. SMCv2 based link group and link will be
assigned to the connection. Then assumed that some buffer assignment
errors happen later in the CLC handshake, such as RMB registration
failure, server will give up SMCRv2 and try SMCRv1 device instead. But
the resources assigned to the connection won't be reset.
When server tries SMCRv1 device, the connection creation process will
be executed again. Since conn->lnk has been assigned when trying SMCRv2,
it will not be set to the correct SMCRv1 link in
smcr_lgr_conn_assign_link(). So in such situation, conn->lgr points to
correct SMCRv1 link group but conn->lnk points to the SMCRv2 link
mistakenly.
Then in smc_clc_send_confirm_accept(), conn->rmb_desc->mr[link->link_idx]
will be accessed. Since the link->link_idx is not correct, the related
MR may not have been initialized, so crash happens.
| Try SMCRv2 device first
| |-> conn->lgr: assign existed SMCRv2 link group;
| |-> conn->link: assign existed SMCRv2 link (link_idx may be 1 in SMC_LGR_SYMMETRIC);
| |-> sndbuf & RMB creation fails, quit;
|
| Try SMCRv1 device then
| |-> conn->lgr: create SMCRv1 link group and assign;
| |-> conn->link: keep SMCRv2 link mistakenly;
| |-> sndbuf & RMB creation succeed, only RMB->mr[link_idx = 0]
| initialized.
|
| Then smc_clc_send_confirm_accept() accesses
| conn->rmb_desc->mr[conn->link->link_idx, which is 1], then crash.
v
This patch tries to fix this by cleaning conn->lnk before assigning
link. In addition, it is better to reset the connection and clean the
resources assigned if trying SMCRv2 failed in buffer creation or
registration.
Fixes: e49300a6bf ("net/smc: add listen processing for SMC-Rv2")
Link: https://lore.kernel.org/r/20220523055056.2078994-1-liuyacan@corp.netease.com/
Signed-off-by: Wen Gu <guwen@linux.alibaba.com>
Reviewed-by: Tony Lu <tonylu@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
There is a certain probability that following
exceptions will occur in the wrk benchmark test:
Running 10s test @ http://11.213.45.6:80
8 threads and 64 connections
Thread Stats Avg Stdev Max +/- Stdev
Latency 3.72ms 13.94ms 245.33ms 94.17%
Req/Sec 1.96k 713.67 5.41k 75.16%
155262 requests in 10.10s, 23.10MB read
Non-2xx or 3xx responses: 3
We will find that the error is HTTP 400 error, which is a serious
exception in our test, which means the application data was
corrupted.
Consider the following scenarios:
CPU0 CPU1
buf_desc->used = 0;
cmpxchg(buf_desc->used, 0, 1)
deal_with(buf_desc)
memset(buf_desc->cpu_addr,0);
This will cause the data received by a victim connection to be cleared,
thus triggering an HTTP 400 error in the server.
This patch exchange the order between clear used and memset, add
barrier to ensure memory consistency.
Fixes: 1c5526968e ("net/smc: Clear memory when release and reuse buffer")
Signed-off-by: D. Wythe <alibuda@linux.alibaba.com>
Reviewed-by: Wenjia Zhang <wenjia@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
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: D. Wythe <alibuda@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
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: D. Wythe <alibuda@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
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: D. Wythe <alibuda@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The struct device for ISM devices was part of struct smcd_dev. Move to
struct ism_dev, provide a new API call in struct smcd_ops, and convert
existing SMCD code accordingly.
Furthermore, remove struct smcd_dev from struct ism_dev.
This is the final part of a bigger overhaul of the interfaces between SMC
and ISM.
Signed-off-by: Stefan Raspl <raspl@linux.ibm.com>
Signed-off-by: Jan Karcher <jaka@linux.ibm.com>
Signed-off-by: Wenjia Zhang <wenjia@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We separate the code implementing the struct smcd_ops API in the ISM
device driver from the functions that may be used by other exploiters of
ISM devices.
Note: We start out small, and don't offer the whole breadth of the ISM
device for public use, as many functions are specific to or likely only
ever used in the context of SMC-D.
This is the third part of a bigger overhaul of the interfaces between SMC
and ISM.
Signed-off-by: Stefan Raspl <raspl@linux.ibm.com>
Signed-off-by: Jan Karcher <jaka@linux.ibm.com>
Signed-off-by: Wenjia Zhang <wenjia@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Register the smc module with the new ism device driver API.
This is the second part of a bigger overhaul of the interfaces between SMC
and ISM.
Signed-off-by: Stefan Raspl <raspl@linux.ibm.com>
Signed-off-by: Jan Karcher <jaka@linux.ibm.com>
Signed-off-by: Wenjia Zhang <wenjia@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If smc_wr_alloc_lgr_mem() fails then return an error code. Don't return
success.
Fixes: 8799e310fb ("net/smc: add v2 support to the work request layer")
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Reviewed-by: Wenjia Zhang <wenjia@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, SMC uses smc->sk.sk_{rcv|snd}buf to create buffers for
send buffer and RMB. And the values of buffer size are from tcp_{w|r}mem
in clcsock.
The buffer size from TCP socket doesn't fit SMC well. Generally, buffers
are usually larger than TCP for SMC-R/-D to get higher performance, for
they are different underlay devices and paths.
So this patch unbinds buffer size from TCP, and introduces two sysctl
knobs to tune them independently. Also, these knobs are per net
namespace and work for containers.
Signed-off-by: Tony Lu <tonylu@linux.alibaba.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
There might be a potential race between SMC-R buffer map and
link group termination.
smc_smcr_terminate_all() | smc_connect_rdma()
--------------------------------------------------------------
| smc_conn_create()
for links in smcibdev |
schedule links down |
| smc_buf_create()
| \- smcr_buf_map_usable_links()
| \- no usable links found,
| (rmb->mr = NULL)
|
| smc_clc_send_confirm()
| \- access conn->rmb_desc->mr[]->rkey
| (panic)
During reboot and IB device module remove, all links will be set
down and no usable links remain in link groups. In such situation
smcr_buf_map_usable_links() should return an error and stop the
CLC flow accessing to uninitialized mr.
Fixes: b9247544c1 ("net/smc: convert static link ID instances to support multiple links")
Signed-off-by: Wen Gu <guwen@linux.alibaba.com>
Link: https://lore.kernel.org/r/1663656189-32090-1-git-send-email-guwen@linux.alibaba.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Extend SMC-R link group netlink attribute SMC_GEN_LGR_SMCR.
Introduce SMC_NLA_LGR_R_BUF_TYPE to show the buffer type of
SMC-R link group.
Signed-off-by: Wen Gu <guwen@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
On long-running enterprise production servers, high-order contiguous
memory pages are usually very rare and in most cases we can only get
fragmented pages.
When replacing TCP with SMC-R in such production scenarios, attempting
to allocate high-order physically contiguous sndbufs and RMBs may result
in frequent memory compaction, which will cause unexpected hung issue
and further stability risks.
So this patch is aimed to allow SMC-R link group to use virtually
contiguous sndbufs and RMBs to avoid potential issues mentioned above.
Whether to use physically or virtually contiguous buffers can be set
by sysctl smcr_buf_type.
Note that using virtually contiguous buffers will bring an acceptable
performance regression, which can be mainly divided into two parts:
1) regression in data path, which is brought by additional address
translation of sndbuf by RNIC in Tx. But in general, translating
address through MTT is fast.
Taking 256KB sndbuf and RMB as an example, the comparisons in qperf
latency and bandwidth test with physically and virtually contiguous
buffers are as follows:
- client:
smc_run taskset -c <cpu> qperf <server> -oo msg_size:1:64K:*2\
-t 5 -vu tcp_{bw|lat}
- server:
smc_run taskset -c <cpu> qperf
[latency]
msgsize tcp smcr smcr-use-virt-buf
1 11.17 us 7.56 us 7.51 us (-0.67%)
2 10.65 us 7.74 us 7.56 us (-2.31%)
4 11.11 us 7.52 us 7.59 us ( 0.84%)
8 10.83 us 7.55 us 7.51 us (-0.48%)
16 11.21 us 7.46 us 7.51 us ( 0.71%)
32 10.65 us 7.53 us 7.58 us ( 0.61%)
64 10.95 us 7.74 us 7.80 us ( 0.76%)
128 11.14 us 7.83 us 7.87 us ( 0.47%)
256 10.97 us 7.94 us 7.92 us (-0.28%)
512 11.23 us 7.94 us 8.20 us ( 3.25%)
1024 11.60 us 8.12 us 8.20 us ( 0.96%)
2048 14.04 us 8.30 us 8.51 us ( 2.49%)
4096 16.88 us 9.13 us 9.07 us (-0.64%)
8192 22.50 us 10.56 us 11.22 us ( 6.26%)
16384 28.99 us 12.88 us 13.83 us ( 7.37%)
32768 40.13 us 16.76 us 16.95 us ( 1.16%)
65536 68.70 us 24.68 us 24.85 us ( 0.68%)
[bandwidth]
msgsize tcp smcr smcr-use-virt-buf
1 1.65 MB/s 1.59 MB/s 1.53 MB/s (-3.88%)
2 3.32 MB/s 3.17 MB/s 3.08 MB/s (-2.67%)
4 6.66 MB/s 6.33 MB/s 6.09 MB/s (-3.85%)
8 13.67 MB/s 13.45 MB/s 11.97 MB/s (-10.99%)
16 25.36 MB/s 27.15 MB/s 24.16 MB/s (-11.01%)
32 48.22 MB/s 54.24 MB/s 49.41 MB/s (-8.89%)
64 106.79 MB/s 107.32 MB/s 99.05 MB/s (-7.71%)
128 210.21 MB/s 202.46 MB/s 201.02 MB/s (-0.71%)
256 400.81 MB/s 416.81 MB/s 393.52 MB/s (-5.59%)
512 746.49 MB/s 834.12 MB/s 809.99 MB/s (-2.89%)
1024 1292.33 MB/s 1641.96 MB/s 1571.82 MB/s (-4.27%)
2048 2007.64 MB/s 2760.44 MB/s 2717.68 MB/s (-1.55%)
4096 2665.17 MB/s 4157.44 MB/s 4070.76 MB/s (-2.09%)
8192 3159.72 MB/s 4361.57 MB/s 4270.65 MB/s (-2.08%)
16384 4186.70 MB/s 4574.13 MB/s 4501.17 MB/s (-1.60%)
32768 4093.21 MB/s 4487.42 MB/s 4322.43 MB/s (-3.68%)
65536 4057.14 MB/s 4735.61 MB/s 4555.17 MB/s (-3.81%)
2) regression in buffer initialization and destruction path, which is
brought by additional MR operations of sndbufs. But thanks to link
group buffer reuse mechanism, the impact of this kind of regression
decreases as times of buffer reuse increases.
Taking 256KB sndbuf and RMB as an example, latency of some key SMC-R
buffer-related function obtained by bpftrace are as follows:
Function Phys-bufs Virt-bufs
smcr_new_buf_create() 67154 ns 79164 ns
smc_ib_buf_map_sg() 525 ns 928 ns
smc_ib_get_memory_region() 162294 ns 161191 ns
smc_wr_reg_send() 9957 ns 9635 ns
smc_ib_put_memory_region() 203548 ns 198374 ns
smc_ib_buf_unmap_sg() 508 ns 1158 ns
------------
Test environment notes:
1. Above tests run on 2 VMs within the same Host.
2. The NIC is ConnectX-4Lx, using SRIOV and passing through 2 VFs to
the each VM respectively.
3. VMs' vCPUs are binded to different physical CPUs, and the binded
physical CPUs are isolated by `isolcpus=xxx` cmdline.
4. NICs' queue number are set to 1.
Signed-off-by: Wen Gu <guwen@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This patch introduces a new SMC-R specific element buf_type
in struct smc_link_group, for recording the value of sysctl
smcr_buf_type when link group is created.
New created link group will create and reuse buffers of the
type specified by buf_type.
Signed-off-by: Wen Gu <guwen@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Some CPU, such as Xeon, can guarantee DMA cache coherency.
So it is no need to use dma sync APIs to flush cache on such CPUs.
In order to avoid calling dma sync APIs on the IO path, use the
dma_need_sync to check whether smc_buf_desc needs dma sync when
creating smc_buf_desc.
Signed-off-by: Guangguan Wang <guangguan.wang@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
smc_ib_sync_sg_for_cpu/device are the ops used for dma memory cache
consistency. Smc sndbufs are dma buffers, where CPU writes data to
it and PCIE device reads data from it. So for sndbufs,
smc_ib_sync_sg_for_device is needed and smc_ib_sync_sg_for_cpu is
redundant as PCIE device will not write the buffers. Smc rmbs
are dma buffers, where PCIE device write data to it and CPU read
data from it. So for rmbs, smc_ib_sync_sg_for_cpu is needed and
smc_ib_sync_sg_for_device is redundant as CPU will not write the buffers.
Signed-off-by: Guangguan Wang <guangguan.wang@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The problem of SMC_CLC_DECL_ERR_REGRMB on the server is very clear.
Based on the fact that whether a new SMC connection can be accepted or
not depends on not only the limit of conn nums, but also the available
entries of rtoken. Since the rtoken release is trigger by peer, while
the conn nums is decrease by local, tons of thing can happen in this
time difference.
This only thing that needs to be mentioned is that now all connection
creations are completely protected by smc_server_lgr_pending lock, it's
enough to check only the available entries in rtokens_used_mask.
Fixes: cd6851f303 ("smc: remote memory buffers (RMBs)")
Signed-off-by: D. Wythe <alibuda@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The main reason for this unexpected SMC_CLC_DECL_ERR_REGRMB in client
dues to following execution sequence:
Server Conn A: Server Conn B: Client Conn B:
smc_lgr_unregister_conn
smc_lgr_register_conn
smc_clc_send_accept ->
smc_rtoken_add
smcr_buf_unuse
-> Client Conn A:
smc_rtoken_delete
smc_lgr_unregister_conn() makes current link available to assigned to new
incoming connection, while smcr_buf_unuse() has not executed yet, which
means that smc_rtoken_add may fail because of insufficient rtoken_entry,
reversing their execution order will avoid this problem.
Fixes: 3e034725c0 ("net/smc: common functions for RMBs and send buffers")
Signed-off-by: D. Wythe <alibuda@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
rmbe_update_limit is used to limit announcing receive
window updating too frequently. RFC7609 request a minimal
increase in the window size of 10% of the receive buffer
space. But current implementation used:
min_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2)
and SOCK_MIN_SNDBUF / 2 == 2304 Bytes, which is almost
always less then 10% of the receive buffer space.
This causes the receiver always sending CDC message to
update its consumer cursor when it consumes more then 2K
of data. And as a result, we may encounter something like
"TCP silly window syndrome" when sending 2.5~8K message.
This patch fixes this using max(rmbe_size / 10, SOCK_MIN_SNDBUF / 2).
With this patch and SMC autocorking enabled, qperf 2K/4K/8K
tcp_bw test shows 45%/75%/40% increase in throughput respectively.
Signed-off-by: Dust Li <dust.li@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
A hung_task is observed when removing SMC-R devices. Suppose that
a link group has two active links(lnk_A, lnk_B) associated with two
different SMC-R devices(dev_A, dev_B). When dev_A is removed, the
link group will be removed from smc_lgr_list and added into
lgr_linkdown_list. lnk_A will be cleared and smcibdev(A)->lnk_cnt
will reach to zero. However, when dev_B is removed then, the link
group can't be found in smc_lgr_list and lnk_B won't be cleared,
making smcibdev->lnk_cnt never reaches zero, which causes a hung_task.
This patch fixes this issue by restoring the implementation of
smc_smcr_terminate_all() to what it was before commit 349d43127d
("net/smc: fix kernel panic caused by race of smc_sock"). The original
implementation also satisfies the intention that make sure QP destroy
earlier than CQ destroy because we will always wait for smcibdev->lnk_cnt
reaches zero, which guarantees QP has been destroyed.
Fixes: 349d43127d ("net/smc: fix kernel panic caused by race of smc_sock")
Signed-off-by: Wen Gu <guwen@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We encountered some crashes caused by the race between SMC-R
link access and link clear that triggered by abnormal link
group termination, such as port error.
Here is an example of this kind of crashes:
BUG: kernel NULL pointer dereference, address: 0000000000000000
Workqueue: smc_hs_wq smc_listen_work [smc]
RIP: 0010:smc_llc_flow_initiate+0x44/0x190 [smc]
Call Trace:
<TASK>
? __smc_buf_create+0x75a/0x950 [smc]
smcr_lgr_reg_rmbs+0x2a/0xbf [smc]
smc_listen_work+0xf72/0x1230 [smc]
? process_one_work+0x25c/0x600
process_one_work+0x25c/0x600
worker_thread+0x4f/0x3a0
? process_one_work+0x600/0x600
kthread+0x15d/0x1a0
? set_kthread_struct+0x40/0x40
ret_from_fork+0x1f/0x30
</TASK>
smc_listen_work() __smc_lgr_terminate()
---------------------------------------------------------------
| smc_lgr_free()
| |- smcr_link_clear()
| |- memset(lnk, 0)
smc_listen_rdma_reg() |
|- smcr_lgr_reg_rmbs() |
|- smc_llc_flow_initiate() |
|- access lnk->lgr (panic) |
These crashes are similarly caused by clearing SMC-R link
resources when some functions is still accessing to them.
This patch tries to fix the issue by introducing reference
count of SMC-R links and ensuring that the sensitive resources
of links won't be cleared until reference count reaches zero.
The operation to the SMC-R link reference count can be concluded
as follows:
object [hold or initialized as 1] [put]
--------------------------------------------------------------------
links smcr_link_init() smcr_link_clear()
connections smc_conn_create() smc_conn_free()
Through this way, the clear of SMC-R links is later than the
free of all the smc connections above it, thus avoiding the
unsafe reference to SMC-R links.
Signed-off-by: Wen Gu <guwen@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
It is no longer suitable to identify whether a smc connection
is registered in a link group through checking if conn->lgr
is NULL, because conn->lgr won't be reset even the connection
is unregistered from a link group.
So this patch introduces a new helper smc_conn_lgr_valid() and
replaces all the check of conn->lgr in original implementation
with the new helper to judge if conn->lgr is valid to use.
Signed-off-by: Wen Gu <guwen@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We encountered some crashes caused by the race between the access
and the termination of link groups.
Here are some of panic stacks we met:
1) Race between smc_clc_wait_msg() and __smc_lgr_terminate()
BUG: kernel NULL pointer dereference, address: 00000000000002f0
Workqueue: smc_hs_wq smc_listen_work [smc]
RIP: 0010:smc_clc_wait_msg+0x3eb/0x5c0 [smc]
Call Trace:
<TASK>
? smc_clc_send_accept+0x45/0xa0 [smc]
? smc_clc_send_accept+0x45/0xa0 [smc]
smc_listen_work+0x783/0x1220 [smc]
? finish_task_switch+0xc4/0x2e0
? process_one_work+0x1ad/0x3c0
process_one_work+0x1ad/0x3c0
worker_thread+0x4c/0x390
? rescuer_thread+0x320/0x320
kthread+0x149/0x190
? set_kthread_struct+0x40/0x40
ret_from_fork+0x1f/0x30
</TASK>
smc_listen_work() abnormal case like port error
---------------------------------------------------------------
| __smc_lgr_terminate()
| |- smc_conn_kill()
| |- smc_lgr_unregister_conn()
| |- set conn->lgr = NULL
smc_clc_wait_msg() |
|- access conn->lgr (panic) |
2) Race between smc_setsockopt() and __smc_lgr_terminate()
BUG: kernel NULL pointer dereference, address: 00000000000002e8
RIP: 0010:smc_setsockopt+0x17a/0x280 [smc]
Call Trace:
<TASK>
__sys_setsockopt+0xfc/0x190
__x64_sys_setsockopt+0x20/0x30
do_syscall_64+0x34/0x90
entry_SYSCALL_64_after_hwframe+0x44/0xae
</TASK>
smc_setsockopt() abnormal case like port error
--------------------------------------------------------------
| __smc_lgr_terminate()
| |- smc_conn_kill()
| |- smc_lgr_unregister_conn()
| |- set conn->lgr = NULL
mod_delayed_work() |
|- access conn->lgr (panic) |
There are some other panic places and they are caused by the
similar reason as described above, which is accessing link
group after termination, thus getting a NULL pointer or invalid
resource.
Currently, there seems to be no synchronization between the
link group access and a sudden termination of it. This patch
tries to fix this by introducing reference count of link group
and not freeing link group until reference count is zero.
Link group might be referred to by links or smc connections. So
the operation to the link group reference count can be concluded
as follows:
object [hold or initialized as 1] [put]
-------------------------------------------------------------------
link group smc_lgr_create() smc_lgr_free()
connections smc_conn_create() smc_conn_free()
links smcr_link_init() smcr_link_clear()
Througth this way, we extend the life cycle of link group and
ensure it is longer than the life cycle of connections and links
above it, so that avoid invalid access to link group after its
termination.
Signed-off-by: Wen Gu <guwen@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
SMC connections might fail to be registered in a link group due to
unable to find a usable link during its creation. As a result,
smc_conn_create() will return a failure and most resources related
to the connection won't be applied or initialized, such as
conn->abort_work or conn->lnk.
If smc_conn_free() is invoked later, it will try to access the
uninitialized resources related to the connection, thus causing
a warning or crash.
This patch tries to fix this by resetting conn->lgr to NULL if an
abnormal exit occurs in smc_lgr_register_conn(), thus avoiding the
access to uninitialized resources in smc_conn_free().
Meanwhile, the new created link group should be terminated if smc
connections can't be registered in it. So smc_lgr_cleanup_early() is
modified to take care of link group only and invoked to terminate
unusable link group by smc_conn_create(). The call to smc_conn_free()
is moved out from smc_lgr_cleanup_early() to smc_conn_abort().
Fixes: 56bc3b2094 ("net/smc: assign link to a new connection")
Suggested-by: Karsten Graul <kgraul@linux.ibm.com>
Signed-off-by: Wen Gu <guwen@linux.alibaba.com>
Acked-by: Karsten Graul <kgraul@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This adds net namespace ID to the kernel log, net_cookie is unique in
the whole system. It is useful in container environment.
Signed-off-by: Tony Lu <tonylu@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This adds net namespace ID to diag of linkgroup, helps us to distinguish
different namespaces, and net_cookie is unique in the whole system.
Signed-off-by: Tony Lu <tonylu@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, rdma device supports exclusive net namespace isolation,
however linkgroup doesn't know and support ibdev net namespace.
Applications in the containers don't want to share the nics if we
enabled rdma exclusive mode. Every net namespaces should have their own
linkgroups.
This patch introduce a new field net for linkgroup, which is standing
for the ibdev net namespace in the linkgroup. The net in linkgroup is
initialized with the net namespace of link's ibdev. It compares the net
of linkgroup and sock or ibdev before choose it, if no matched, create
new one in current net namespace. If rdma net namespace exclusive mode
is not enabled, it behaves as before.
Signed-off-by: Tony Lu <tonylu@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
drivers/net/ethernet/mellanox/mlx5/core/en_tc.c
commit 077cdda764 ("net/mlx5e: TC, Fix memory leak with rules with internal port")
commit 31108d142f ("net/mlx5: Fix some error handling paths in 'mlx5e_tc_add_fdb_flow()'")
commit 4390c6edc0 ("net/mlx5: Fix some error handling paths in 'mlx5e_tc_add_fdb_flow()'")
https://lore.kernel.org/all/20211229065352.30178-1-saeed@kernel.org/
net/smc/smc_wr.c
commit 49dc9013e3 ("net/smc: Use the bitmap API when applicable")
commit 349d43127d ("net/smc: fix kernel panic caused by race of smc_sock")
bitmap_zero()/memset() is removed by the fix
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
A crash occurs when smc_cdc_tx_handler() tries to access smc_sock
but smc_release() has already freed it.
[ 4570.695099] BUG: unable to handle page fault for address: 000000002eae9e88
[ 4570.696048] #PF: supervisor write access in kernel mode
[ 4570.696728] #PF: error_code(0x0002) - not-present page
[ 4570.697401] PGD 0 P4D 0
[ 4570.697716] Oops: 0002 [#1] PREEMPT SMP NOPTI
[ 4570.698228] CPU: 0 PID: 0 Comm: swapper/0 Not tainted 5.16.0-rc4+ #111
[ 4570.699013] Hardware name: Alibaba Cloud Alibaba Cloud ECS, BIOS 8c24b4c 04/0
[ 4570.699933] RIP: 0010:_raw_spin_lock+0x1a/0x30
<...>
[ 4570.711446] Call Trace:
[ 4570.711746] <IRQ>
[ 4570.711992] smc_cdc_tx_handler+0x41/0xc0
[ 4570.712470] smc_wr_tx_tasklet_fn+0x213/0x560
[ 4570.712981] ? smc_cdc_tx_dismisser+0x10/0x10
[ 4570.713489] tasklet_action_common.isra.17+0x66/0x140
[ 4570.714083] __do_softirq+0x123/0x2f4
[ 4570.714521] irq_exit_rcu+0xc4/0xf0
[ 4570.714934] common_interrupt+0xba/0xe0
Though smc_cdc_tx_handler() checked the existence of smc connection,
smc_release() may have already dismissed and released the smc socket
before smc_cdc_tx_handler() further visits it.
smc_cdc_tx_handler() |smc_release()
if (!conn) |
|
|smc_cdc_tx_dismiss_slots()
| smc_cdc_tx_dismisser()
|
|sock_put(&smc->sk) <- last sock_put,
| smc_sock freed
bh_lock_sock(&smc->sk) (panic) |
To make sure we won't receive any CDC messages after we free the
smc_sock, add a refcount on the smc_connection for inflight CDC
message(posted to the QP but haven't received related CQE), and
don't release the smc_connection until all the inflight CDC messages
haven been done, for both success or failed ones.
Using refcount on CDC messages brings another problem: when the link
is going to be destroyed, smcr_link_clear() will reset the QP, which
then remove all the pending CQEs related to the QP in the CQ. To make
sure all the CQEs will always come back so the refcount on the
smc_connection can always reach 0, smc_ib_modify_qp_reset() was replaced
by smc_ib_modify_qp_error().
And remove the timeout in smc_wr_tx_wait_no_pending_sends() since we
need to wait for all pending WQEs done, or we may encounter use-after-
free when handling CQEs.
For IB device removal routine, we need to wait for all the QPs on that
device been destroyed before we can destroy CQs on the device, or
the refcount on smc_connection won't reach 0 and smc_sock cannot be
released.
Fixes: 5f08318f61 ("smc: connection data control (CDC)")
Reported-by: Wen Gu <guwen@linux.alibaba.com>
Signed-off-by: Dust Li <dust.li@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
We found smc_llc_send_link_delete_all() sometimes wait
for 2s timeout when testing with RDMA link up/down.
It is possible when a smc_link is in ACTIVATING state,
the underlaying QP is still in RESET or RTR state, which
cannot send any messages out.
smc_llc_send_link_delete_all() use smc_link_usable() to
checks whether the link is usable, if the QP is still in
RESET or RTR state, but the smc_link is in ACTIVATING, this
LLC message will always fail without any CQE entering the
CQ, and we will always wait 2s before timeout.
Since we cannot send any messages through the QP before
the QP enter RTS. I add a wrapper smc_link_sendable()
which checks the state of QP along with the link state.
And replace smc_link_usable() with smc_link_sendable()
in all LLC & CDC message sending routine.
Fixes: 5f08318f61 ("smc: connection data control (CDC)")
Signed-off-by: Dust Li <dust.li@linux.alibaba.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, buffers are cleared when smc connections are created and
buffers are reused. This slows down the speed of establishing new
connections. In most cases, the applications want to establish
connections as quickly as possible.
This patch moves memset() from connection creation path to release and
buffer unuse path, this trades off between speed of establishing and
release.
Test environments:
- CPU Intel Xeon Platinum 8 core, mem 32 GiB, nic Mellanox CX4
- socket sndbuf / rcvbuf: 16384 / 131072 bytes
- w/o first round, 5 rounds, avg, 100 conns batch per round
- smc_buf_create() use bpftrace kprobe, introduces extra latency
Latency benchmarks for smc_buf_create():
w/o patch : 19040.0 ns
w/ patch : 1932.6 ns
ratio : 10.2% (-89.8%)
Latency benchmarks for socket create and connect:
w/o patch : 143.3 us
w/ patch : 102.2 us
ratio : 71.3% (-28.7%)
The latency of establishing connections is reduced by 28.7%.
Signed-off-by: Tony Lu <tonylu@linux.alibaba.com>
Reviewed-by: Wen Gu <guwen@linux.alibaba.com>
Signed-off-by: Karsten Graul <kgraul@linux.ibm.com>
Link: https://lore.kernel.org/r/20211203113331.2818873-1-kgraul@linux.ibm.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Coverity reports a possible NULL dereferencing problem:
in smc_vlan_by_tcpsk():
6. returned_null: netdev_lower_get_next returns NULL (checked 29 out of 30 times).
7. var_assigned: Assigning: ndev = NULL return value from netdev_lower_get_next.
1623 ndev = (struct net_device *)netdev_lower_get_next(ndev, &lower);
CID 1468509 (#1 of 1): Dereference null return value (NULL_RETURNS)
8. dereference: Dereferencing a pointer that might be NULL ndev when calling is_vlan_dev.
1624 if (is_vlan_dev(ndev)) {
Remove the manual implementation and use netdev_walk_all_lower_dev() to
iterate over the lower devices. While on it remove an obsolete function
parameter comment.
Fixes: cb9d43f677 ("net/smc: determine vlan_id of stacked net_device")
Suggested-by: Julian Wiedmann <jwi@linux.ibm.com>
Signed-off-by: Karsten Graul <kgraul@linux.ibm.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The link_id is supposed to be unique, but smcr_next_link_id() doesn't
skip the used link_id as expected. So the patch fixes this.
Fixes: 026c381fb4 ("net/smc: introduce link_idx for link group array")
Signed-off-by: Wen Gu <guwen@linux.alibaba.com>
Reviewed-by: Tony Lu <tonylu@linux.alibaba.com>
Acked-by: Karsten Graul <kgraul@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Wen Gu