Two performance improvements when recording wine #1777
Conversation
src/preload/preload.c
Outdated
| if (set && (how == SIG_BLOCK || how == SIG_SETMASK)) { | ||
| local_memcpy(&modified_set, set, sizeof(kernel_sigset_t)); | ||
| // SIGSTKFLT (PerfCounters::TIME_SLICE_SIGNAL) and | ||
| // SIGPWR(SYSCALLBUF_DESCHED_SIGNAL) are used by |
|
Looks good! |
|
This causes a number of test failures... |
|
Oops, sorry. I will take a look. |
|
|
|
Ah yes, of course. Should have realized that (well, should have remembered that I realized that ;)). This gets our understanding of which signals are blocked out of whack, which as we learned before is a bad idea. I had thought about using an extra thread-local field to hold a copy of the current sigmask that rr can reload when it needs to. Thoughts? |
|
This looks like it should be easy but it's not :-( |
|
ah right. I thought about that and then forgot about it. |
|
The way we handle this for |
|
that could work, but I don't think it needs to be that fancy, because we don't really need to know the full history. |
|
Storing a signal mask in |
|
Maybe stick it in |
|
Unless I'm missing something, we can't put this in |
|
It can go in |
|
Ok, try this. Passed tests for me, but I'm running them again just to be sure ;). Also GitHub appears to be confused by me having put a newer commit in front of the older ones, but in the actual git history they should be in proper order. |
In preparation of being able to buffer system calls that affect the signal mask, keep a copy of blocked_sigs in syscallbuf_hdr, save it whenever modified due to an event not visible to preload, and reload it whenever queried whether a particular signal is blocked.
…state A priori, there's nothing wrong with keeping the existing TLS descriptor if the VM is forked, since the two threads won't clobber each other. This fixes a performance problem in single threaded processes that fork (where thread locals get initialized only because of the preload library), since the forked process would never set a new TLS descriptor.
|
Found one small thing, but now should be good to go. |
|
Merged. Thanks!!! |
|
My pleasure as always ;). |
2 participants
First add syscall buffering for rt_sigprocmask, second if a clone forks the address space, keep the state of thread local initialization since there won't be an explicit re-initialization (and without this the syscallbuf is effectively disabled).