Now that mpi has unlocked uvm's fault handler, we can unlock the mmap syscall to handle MAP_ANON without the big lock.
sys_mmap() still protects the !MAP_ANON case, i.e. file based mappings, with the KERNEL_LOCK() itself, which is why unlocking the syscall will only change locking behaviour for anonymous mappings. A previous to unlock file based mappings was reverted, see the following from https://marc.info/?l=openbsd-tech&m=160155434212888&w=2 : commit 38802bc07455f2a4f8cdde272850a5eab5dfa6c8 from: mpi <m...@openbsd.org> date: Wed Oct 7 12:26:20 2020 UTC Do not release the KERNEL_LOCK() when mmap(2)ing files. Previous attempt to unlock amap & anon exposed a race in vnode reference counting. So be conservative with the code paths that we're not fully moving out of the KERNEL_LOCK() to allow us to concentrate on one area at a time. ... So here's a first small step. I've been running with this for months on a few amd64, arm64 and sparc64 boxes without problems; they've been daily drivers and/or have been building releases and ports. Feedback? Objections? OK? Index: sys/kern/syscalls.master =================================================================== RCS file: /cvs/src/sys/kern/syscalls.master,v retrieving revision 1.221 diff -u -p -r1.221 syscalls.master --- sys/kern/syscalls.master 23 Dec 2021 18:50:31 -0000 1.221 +++ sys/kern/syscalls.master 31 Dec 2021 09:14:00 -0000 @@ -126,7 +126,7 @@ struct sigaction *osa); } 47 STD NOLOCK { gid_t sys_getgid(void); } 48 STD NOLOCK { int sys_sigprocmask(int how, sigset_t mask); } -49 STD { void *sys_mmap(void *addr, size_t len, int prot, \ +49 STD NOLOCK { void *sys_mmap(void *addr, size_t len, int prot, \ int flags, int fd, off_t pos); } 50 STD { int sys_setlogin(const char *namebuf); } #ifdef ACCOUNTING
