On Wed, Apr 06, 2022 at 10:52:38AM +0800, Liao Chang wrote: > Kernel check for pending softirqs periodically, they are performed in a > few points of kernel code, such as irq_exit() and __local_bh_enable_ip(), > softirqs that have been activated by a given CPU must be executed on the > same CPU, this characteristic of softirq is always a potentially > "dangerous" operation, because one CPU might be end up very busy while > the other are most idle. > > Above concern is proven in a networking user case: recenlty, we > engineer find out the time used for connection re-establishment on > kernel v5.10 is 300 times larger than v4.19, meanwhile, softirq > monopolize almost 99% of CPU. This problem stem from that the connection > between Sender and Receiver node get lost, the NIC driver on Sender node > will keep raising NET_TX softirq before connection recovery. The system > log show that most of softirq is performed from __local_bh_enable_ip(), > since __local_bh_enable_ip is used widley in kernel code, it is very > easy to run out most of CPU, and the user-mode application can't obtain > enough CPU cycles to establish connection as soon as possible. Shouldn't you fix that bug instead? This seems like papering over the bad effects of a bug and would make it harder to find bugs like this in the future. Essentially, it's the same as a screaming hardware interrupt, except that it's a software interrupt, so we can fix the bug instead of working around broken hardware.
