Messages in this thread | | | Date | Mon, 11 May 2009 23:04:33 +0200 | From | Ingo Molnar <> | Subject | Re: [PATCH 1/5] Split wait_noreap_copyout() |
| |
* Vitaly Mayatskikh <v.mayatskih@gmail.com> wrote:
> At Mon, 11 May 2009 14:17:08 +0200, Ingo Molnar wrote: > > > > if (put_user(signal, &infop->si_signo) || > > > put_user(0, &infop->si_errno) || > > > put_user((short)why, &infop->si_code) || > > > put_user(pid, &infop->si_pid) || > > > put_user(uid, &infop->si_uid) || > > > put_user(status, &infop->si_status)) > > > return -EFAULT; > > > > For best assembly code this should generally be written as a series > > of: > > > > __uaccess_err |= __put_user(x, ptr); > > __uaccess_err |= __put_user(y, ptr); > > __uaccess_err |= __put_user(z, ptr); > > > > As this generates non-dependent, compressed, branch-less code. > > Yeah, my first intention was to eliminate a lot of branches in one > place. It's terrible for CPU pipeline, I bet. > > > See the (new) put_user_try / put_user_ex() / put_user_catch() > > abstraction in arch/x86/include/asm/uaccess.h, and how all the > > x86 signal code makes use of that to optimize such patterns of > > per field user copies. > > So, there's catch block to handle GPF and the code inside of `try' > block is still branch-less, right? I was thinking of minimized > version of struct siginfo (up to si_uid) and copying it with > single copy_to_user(), but the idea with try/catch is definitely > much better.
It creates really nice assembly code. Hiroshi-san experimented with it a lot until he found this form.
Regarding potentially generalizing that facility into generic code, it relies on the exception code filling in current_thread_info()->uaccess_err with -EFAULT. So it needs architecture level support. It also kind of relies on current_thread_info()->uaccess_err being super-optimal - which it is on x86. (the assembler can optimize it)
But a compatible wrapper could be added, for architectures that dont support, or that dont need support.
Ingo
| |