Bug 206919
| Summary: | Very bad performance of sendto function | ||
|---|---|---|---|
| Product: | Networking | Reporter: | Bernat Arlandis (bernat.arlandis) |
| Component: | Other | Assignee: | Stephen Hemminger (stephen) |
| Status: | RESOLVED INSUFFICIENT_DATA | ||
| Severity: | normal | ||
| Priority: | P1 | ||
| Hardware: | All | ||
| OS: | Linux | ||
| Kernel Version: | 5.4.13 | Subsystem: | |
| Regression: | No | Bisected commit-id: | |
| Attachments: | Sysprof capture file | ||
What is your application? What is the protocol? Why not use perf top (modern kernel profiling). netlink is for control operations (ip link show etc). > What is your application? The application is a racing game, rFactor 2, running on Proton (a Wine special version). > What is the protocol? I don't know about the protocol, it's the game's networking code. > Why not use perf top (modern kernel profiling). I didn't know about perf top. I'm relatively new to this kind of tasks. > netlink is for control operations (ip link show etc). What do you mean? Those netlink functions are called inside the sendto function in the kernel. Or maybe I'm missing something. I can investigate more as I gather more information. When dealing with an unknown application, looking at what system calls are being made (with strace) can help. It can be anything; like a Wine doing something to emulate Windows behaviour. Bottom line, with out a specific Linux kernel indication, this is not the right place to get a resolution. Ok. Sorry, I thought the profiling session showed something weird. I'll keep investigating. Thanks for the answer. |
Created attachment 288007 [details] Sysprof capture file Hi, I have a case where the sendto function is slowing down the software I'm trying to use. Sysprof capture file attached. It's probably being called at a high frequency and the __netlink_dump_start function spends a lot of time in the mutex lock, then also in the netlink_dump call. This is causing the threads calling the function to stall. At least this is my interpretation of the profiling session. I don't have any knowledge about the way networking code works. I'd like to know if the flaw is in the user-space software or in the kernel. Can I provide any more info or do some test to help determine what's happening? I'd like to help fix it. My NIC: Intel Corporation I211 Gigabit Network Connection (rev 03) My system: Debian 10 with kernel 5.4.13. Thanks.