Two new ways to read a file quickly

Posted Mar 6, 2020 19:11 UTC (Fri) by NYKevin (subscriber, #129325)
In reply to: Two new ways to read a file quickly by mm7323
Parent article: Two new ways to read a file quickly

> I think looking at a way to generically batch any system call in an easy to use way might be a better approach, though not easy to get.

We have that. It's called "userspace."

Snark aside, I think any solution in this space needs to be very clear on exactly what set of problems it is solving, and what set of problems are out of scope. Otherwise, I imagine you would inevitably end up with a harder-to-use-but-less-flexible syscall interface.

> Perhaps a getlastfd() 'system call' could also be used to help bridge the problem of open() (or socket() etc...) then read() without needing to use fixed FD numbers or loop beck through userspace.

Would getlastfd() be thread-local, or would it be thread-unsafe? Does the kernel track fds in a thread-local way right now?

Two new ways to read a file quickly

Posted Mar 7, 2020 3:12 UTC (Sat) by josh (subscriber, #17465) [Link]

"userspace" has this increasingly problematic performance bottleneck called "system calls", which io_uring replaces with a much faster mechanism called "shared memory".