Messages in this thread | | | Date | Wed, 27 Nov 2013 18:01:13 -0800 | Subject | Re: [PATCH] Add a text_poke syscall v2 | From | Linus Torvalds <> |
| |
On Wed, Nov 27, 2013 at 3:28 PM, H. Peter Anvin <hpa@zytor.com> wrote: > > The timeout bit was an acknowledgment that some kind of batching > interface is necessary.
That's just moronic. People would make up totally random timeouts, so from an interface standpoint it's just horrid, horrid.
Giving user space random knobs that you don't understand yourself, and the monkeys in user space are guaranteed to mis-use is just entirely the wrong thing to do.
Much better to then just making the interface itself be about batching, which isn't as hard as you make it out to be. Make it an array of those addr/replace/len things. And we have that "restart_block" for system calls, and we'd limit batching to some random smallish number ("128 instructions, just because"), while still being easily interruptible in between those blocks. That limits you to two IPI's per 128 instructions replaced - and at that point even *that* is just an internal kernel random tuning thing, not some insane user interface.
But is such batching really even worth it? If' it's not *that* much more effort, maybe it's worth it, but do we have known users that really would have thousands and thousands of cases all at once?
Linus
| |