As found by tb@ and visa@, `f_mtx' need to block interrupts as long as it can be taken w/ and w/o the KERNEL_LOCK(). Otherwise a deadlock is possible if an interrupt tries to grab the KERNEL_LOCK().
I'm not switching to a rwlock because code paths are short, I don't want to introduce new sleeping points and in the long run we should be using SRPs or atomic operations for reference counts. ok? Index: kern/kern_descrip.c =================================================================== RCS file: /cvs/src/sys/kern/kern_descrip.c,v retrieving revision 1.158 diff -u -p -r1.158 kern_descrip.c --- kern/kern_descrip.c 8 May 2018 09:03:58 -0000 1.158 +++ kern/kern_descrip.c 28 May 2018 09:23:31 -0000 @@ -957,7 +957,11 @@ restart: */ numfiles++; fp = pool_get(&file_pool, PR_WAITOK|PR_ZERO); - mtx_init(&fp->f_mtx, IPL_NONE); + /* + * We need to block interrupts as long as `f_mtx' is being taken + * with and without the KERNEL_LOCK(). + */ + mtx_init(&fp->f_mtx, IPL_VM); fp->f_iflags = FIF_LARVAL; if ((fq = p->p_fd->fd_ofiles[0]) != NULL) { LIST_INSERT_AFTER(fq, fp, f_list);