On Sat, May 01, 2021 at 08:19:19AM +0200, Anton Lindqvist wrote: > Hi, > In August 2019 I tried to unlock lseek(2) which failed since the vnode > lock could not be acquired without holding the kernel lock back then, > found the hard way. claudio@ recently[1] make it possible to acquire a > vnode lock without holding the kernel lock. I therefore would like to > give this another try. The kernel lock is still required around > VOP_GETATTR() as the underlying file system implementations are not > yet^W MP-safe. > > Comments? OK? > > [1] > https://github.com/openbsd/src/commit/9d6122f62b6ed32d6c956e1d5269114b2f24ea14 > > Index: kern/syscalls.master > =================================================================== > RCS file: /cvs/src/sys/kern/syscalls.master,v > retrieving revision 1.209 > diff -u -p -r1.209 syscalls.master > --- kern/syscalls.master 18 Mar 2021 08:43:38 -0000 1.209 > +++ kern/syscalls.master 29 Apr 2021 19:09:58 -0000 > @@ -349,7 +349,7 @@ > 197 STD { void *sys_mmap(void *addr, size_t len, int prot, \ > int flags, int fd, long pad, off_t pos); } > 198 INDIR { quad_t sys___syscall(quad_t num, ...); } > -199 STD { off_t sys_lseek(int fd, int pad, off_t offset, \ > +199 STD NOLOCK { off_t sys_lseek(int fd, int pad, off_t offset, \ > int whence); } > 200 STD { int sys_truncate(const char *path, int pad, \ > off_t length); } > Index: kern/vfs_vnops.c > =================================================================== > RCS file: /cvs/src/sys/kern/vfs_vnops.c,v > retrieving revision 1.115 > diff -u -p -r1.115 vfs_vnops.c > --- kern/vfs_vnops.c 28 Apr 2021 09:53:53 -0000 1.115 > +++ kern/vfs_vnops.c 29 Apr 2021 19:09:58 -0000 > @@ -657,7 +657,9 @@ vn_seek(struct file *fp, off_t *offset, > newoff = fp->f_offset + *offset; > break; > case SEEK_END: > + KERNEL_LOCK(); > error = VOP_GETATTR(vp, &vattr, cred, p); > + KERNEL_UNLOCK(); > if (error) > goto out; > newoff = *offset + (off_t)vattr.va_size; >
Been running with this for a while. Both make build and a regress run done. Diff make sense. OK claudio@ -- :wq Claudio
