"Maciej W. Rozycki" <firstname.lastname@example.org> writes:
> Here is the second version of the sys__test_and_set() syscall suite. A
> glibc patch is included this time as well.
> There are two small changes to the sys__test_and_set() implementation:
> 1. verify_area() is now called for the ll/sc version as well. Otherwise
> one could pass a KSEG address and gain unauthorized access.
> 2. The fuction now returns immediately without performing a write access
> if the value stored in the memory wouldn't change. This is to avoid the
> need of a potentially costful sc operation; for consistency, this is also
> done for the non-ll/sc version.
> The glibc patch should be fairly obvious. There is no inline version of
> the _test_and_set() function for MIPS I anymore -- while previously it
> saved an extra stack frame just to call sysmips(), this would be pointless
> now (well, not quite as long as we fallback to sysmips(), actually, but
> that is a temporary compatibility bit that will soon get removed, I hope).
> Note that while sys__test_and_set() never returns an error, there might
> one happen if someone tries to execute the syscall running a kernel that
> does not support it. Hence we fall back to sysmips().
> The official entry point is _test_and_set(). There is also the
> ___test_and_set() entry point defined, mostly for completeness for MIPS
> II+ systems, to be sure all syscalls actually have their wrappers
> exported. Not to be used under normal circumstances, though.
> Andreas, what do you think: Should we fall back to sysmips() as in the
> following patch (at a considerable performance hit -- without the fallback
> the entire ___test_and_set() wrapper is seven instructions long) or just
> require a specific minimum kernel version bail out at the compile time if
> no __NR__test_and_set is defined? Granted, pthreads don't run for
> MIPS/Linux for a long time, so it's possible the user base is not that
> large such an abrupt switch would be impossible. Especially as sysmips()
> seems to be continuously in flux for the last few months. I assume the
> switch to the new syscall would be mandatory for glibc 2.3 in any case.
Do it the following way:
- Define in sysdeps/unix/sysv/linux/kernel-features a new macro, e.g.
__ASSUME_TEST_AND_SET with the appropriate guards
- Do *both* implementations like this:
fast code without fallback
slow code that first tries kernel call and then falls back to sysmips
This way you get the fast one if you configure glibc with
--enable-kernel=2.4.6 if we assume that 2.4.6 is the first kernel with
Check other places in glibc for details how this can be done.
> I'm open to constructive feedback. An open question is whether returning
> the result in v1 is clean. I believe it is -- I haven't been convinced
> that storing the result in a memory location passed as the third argument
> is cleaner. Certainly it's not faster and v1 is still dedicated to be a
> result register. It's used by sys_pipe() this way, for example.
SuSE Labs email@example.com