Re: [Regression] [iommu/iova] iova_rbtree_lock contended seriously causing performance bottleneck (bisect done; commit found)

From: Ethan Zhao
Date: Tue Jan 02 2024 - 23:57:45 EST



On 1/3/2024 12:55 PM, Ethan Zhao wrote:
part of the mail got lost after sent
past here

Issue:
When network throughput is large (>1GB/s), the performance of
application nginx hits bottleneck, most of cpu cycles are cost
by spin lock, it takes a lot of cycles to aquire iova_rbtree_lock
for  __alloc_and_insert_iova_range().
flame graph


queued_spin_lock_slowpath-------------\
do_raw_spin_lock \
_raw_spin_lock_irqsave \
__alloc_and_insert_iova_range \
alloc_iova_fast \
iommu_dma_alloc_iova \
__iommu_dma_map \
iommu_dma_map_page \
ice_tx_map.isra.0 \
ice_xmit_frame_ring \
dev_hard_start_xmit \
sch_direct_xmit \
__dev_xmit_skb \
__dev_queue_xmit \
ip_finish_output2 \
__ip_queue_xmit \
__tcp_transmit_skb \
tcp_write_xmit \
__tcp_push_pending_frames \
tcp_sendmsg_locked \
tcp_sendmsg \
tcp_sendmsg \
sock_write_iter \
do_iter_readv_writev \
do_iter_write \
vfs_writev \
do_writev \
do_syscall_64 \
entry_SYSCALL_64_after_hwframe \
__GI___writev \
[nginx] \
ngx_linux_sendfile_chain \
ngx_http_write_filter \
ngx_output_chain \
ngx_http_send_response \
ngx_http_script_return_code \
[nginx] \
ngx_http_core_run_phases \
ngx_http_process_request \

Setup:
1. configure server nginx with following nginx.conf (appended to the tail)

2. request server with WRK
./wrk -t 64 -c 1024 -d 40 --latency http://$server_ip:10802/1KB.json

Debugging summary:
Bisect identified "commit 371d7955e310 iommu/iova: Improve restart logic"
as cause.

nginx.conf
user nginx;
worker_processes 16;
error_log /var/log/nginx/error.log crit;
pid /var/run/nginx.pid;
events {
worker_connections 4000;
use epoll;
multi_accept on;
}
http {
include /etc/nginx/mime.types;
default_type application/octet-stream;
log_format main '$remote_addr - $remote_user [$time_local] "$request" '
'$status $body_bytes_sent "$http_referer" '
'"$http_user_agent" "$http_x_forwarded_for"';
access_log off;
sendfile on;
tcp_nopush on;
tcp_nodelay on;
keepalive_timeout 65;
keepalive_requests 20480;
gzip_min_length 10240;
gzip_comp_level 1;
gzip_vary on;
gzip_disable msie6;
gzip_proxied expired no-cache no-store private auth;
gzip_types
# text/html is always compressed by HttpGzipModule
text/css
text/javascript
text/xml
text/plain
text/x-component
application/javascript
application/x-javascript
application/json
application/xml
application/rss+xml
application/atom+xml
font/truetype
font/opentype
application/vnd.ms-fontobject
image/svg+xml;

reset_timedout_connection on;
client_body_timeout 10;
send_timeout 2;
include /etc/nginx/conf.d/*.conf;
server {
listen 10802;
server_name localhost;
location / {
root /usr/share/nginx/html;
index index.html index.htm;
}
error_page 500 502 503 504 /50x.html;
location = /50x.html {
#root /usr/share/nginx/html;
root /ramdisk;
}
location /1KB.json {
return 202 '{"status":"success","result":"\
Hello from NGINX, 2KB test\
Nanchang, which was the capital of Yuzhang Prefecture during the HanDynasty, \
now falls under the jurisdiction of Hongzhou. It straddles the borderof the \
influence of the Ye and Zhen constellations , and is adjacent to theHeng \
and the Lu mountains . The three rivers enfold it like the frontpart \
of a garment and the five lakes encircle it like a girdle. Itcontrols \
nature’s jewels. The radiance of its legendary sword shootsdirectly upward \
between the constellations Niu and Dou. Its talented peopleare outstanding,\
and the spirit of intelligence pervades the place. This wasthe place where Xu \
Ru spent the night on his visit to Chen Fan (10). The mightyHongzhou spreads \
meteors chasing one another.\
"}';
}
}
}


Thanks,

Ethan