Home | History | Annotate | Line # | Download | only in rumpkern
emul.c revision 1.150.14.5
      1  1.150.14.3       tls /*	$NetBSD: emul.c,v 1.150.14.5 2017/12/03 11:39:16 jdolecek Exp $	*/
      2         1.1     pooka 
      3         1.1     pooka /*
      4       1.150     pooka  * Copyright (c) 2007-2011 Antti Kantee.  All Rights Reserved.
      5         1.1     pooka  *
      6         1.1     pooka  * Redistribution and use in source and binary forms, with or without
      7         1.1     pooka  * modification, are permitted provided that the following conditions
      8         1.1     pooka  * are met:
      9         1.1     pooka  * 1. Redistributions of source code must retain the above copyright
     10         1.1     pooka  *    notice, this list of conditions and the following disclaimer.
     11         1.1     pooka  * 2. Redistributions in binary form must reproduce the above copyright
     12         1.1     pooka  *    notice, this list of conditions and the following disclaimer in the
     13         1.1     pooka  *    documentation and/or other materials provided with the distribution.
     14         1.1     pooka  *
     15         1.1     pooka  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS
     16         1.1     pooka  * OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
     17         1.1     pooka  * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
     18         1.1     pooka  * DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
     19         1.1     pooka  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
     20         1.1     pooka  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
     21         1.1     pooka  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
     22         1.1     pooka  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
     23         1.1     pooka  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
     24         1.1     pooka  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
     25         1.1     pooka  * SUCH DAMAGE.
     26         1.1     pooka  */
     27         1.1     pooka 
     28        1.61     pooka #include <sys/cdefs.h>
     29  1.150.14.3       tls __KERNEL_RCSID(0, "$NetBSD: emul.c,v 1.150.14.5 2017/12/03 11:39:16 jdolecek Exp $");
     30         1.9     pooka 
     31         1.1     pooka #include <sys/param.h>
     32  1.150.14.5  jdolecek #include <sys/cprng.h>
     33  1.150.14.4       tls #include <sys/filedesc.h>
     34  1.150.14.5  jdolecek #include <sys/fstrans.h>
     35  1.150.14.5  jdolecek #include <sys/kauth.h>
     36        1.66     pooka #include <sys/module.h>
     37        1.67     pooka #include <sys/reboot.h>
     38  1.150.14.3       tls #include <sys/syscall.h>
     39  1.150.14.5  jdolecek #include <sys/pserialize.h>
     40  1.150.14.5  jdolecek #ifdef LOCKDEBUG
     41       1.148     pooka #include <sys/sleepq.h>
     42  1.150.14.5  jdolecek #endif
     43        1.67     pooka 
     44        1.67     pooka #include <dev/cons.h>
     45         1.1     pooka 
     46  1.150.14.5  jdolecek #include <rump-sys/kern.h>
     47         1.8     pooka 
     48  1.150.14.5  jdolecek #include <rump/rumpuser.h>
     49         1.1     pooka 
     50  1.150.14.4       tls void (*rump_vfs_fini)(void) = (void *)nullop;
     51  1.150.14.4       tls 
     52       1.149     pooka /*
     53       1.149     pooka  * physmem is largely unused (except for nmbcluster calculations),
     54       1.149     pooka  * so pick a default value which suits ZFS.  if an application wants
     55       1.149     pooka  * a very small memory footprint, it can still adjust this before
     56       1.149     pooka  * calling rump_init()
     57       1.149     pooka  */
     58       1.149     pooka #define PHYSMEM 512*256
     59  1.150.14.5  jdolecek psize_t physmem = PHYSMEM;
     60       1.149     pooka int nkmempages = PHYSMEM/2; /* from le chapeau */
     61       1.149     pooka #undef PHYSMEM
     62       1.149     pooka 
     63         1.1     pooka struct vnode *rootvp;
     64       1.124     pooka dev_t rootdev = NODEV;
     65       1.149     pooka 
     66        1.19     pooka const int schedppq = 1;
     67        1.53     pooka bool mp_online = false;
     68  1.150.14.5  jdolecek struct timespec boottime;
     69        1.66     pooka int cold = 1;
     70        1.95     pooka int boothowto = AB_SILENT;
     71        1.67     pooka struct tty *constty;
     72         1.1     pooka 
     73       1.135     pooka const struct bdevsw *bdevsw0[255];
     74        1.64     pooka const struct bdevsw **bdevsw = bdevsw0;
     75       1.135     pooka const int sys_cdevsws = 255;
     76       1.135     pooka int max_cdevsws = 255;
     77        1.64     pooka 
     78       1.135     pooka const struct cdevsw *cdevsw0[255];
     79        1.64     pooka const struct cdevsw **cdevsw = cdevsw0;
     80       1.135     pooka const int sys_bdevsws = 255;
     81       1.135     pooka int max_bdevsws = 255;
     82        1.64     pooka 
     83        1.76     pooka int mem_no = 2;
     84        1.76     pooka 
     85  1.150.14.1       tls device_t booted_device;
     86  1.150.14.1       tls device_t booted_wedge;
     87        1.92     pooka int booted_partition;
     88  1.150.14.5  jdolecek const char *booted_method;
     89        1.92     pooka 
     90       1.120     pooka /* XXX: unused */
     91        1.76     pooka kmutex_t tty_lock;
     92       1.120     pooka krwlock_t exec_lock;
     93        1.64     pooka 
     94       1.139     pooka /* sparc doesn't sport constant page size, pretend we have 4k pages */
     95       1.109     pooka #ifdef __sparc__
     96       1.109     pooka int nbpg = 4096;
     97       1.139     pooka int pgofset = 4096-1;
     98       1.139     pooka int pgshift = 12;
     99       1.109     pooka #endif
    100       1.109     pooka 
    101       1.146   tsutsui /* on sun3 VM_MAX_ADDRESS is a const variable */
    102       1.146   tsutsui /* XXX: should be moved into rump.c and initialize for sun3 and sun3x? */
    103       1.141     pooka #ifdef sun3
    104       1.146   tsutsui const vaddr_t kernbase = KERNBASE3;
    105       1.140     pooka #endif
    106       1.140     pooka 
    107       1.117     pooka struct loadavg averunnable = {
    108       1.117     pooka 	{ 0 * FSCALE,
    109       1.117     pooka 	  1 * FSCALE,
    110       1.117     pooka 	  11 * FSCALE, },
    111       1.117     pooka 	FSCALE,
    112       1.117     pooka };
    113       1.117     pooka 
    114  1.150.14.5  jdolecek /*
    115  1.150.14.5  jdolecek  * Include the autogenerated list of auto-loadable syscalls
    116  1.150.14.5  jdolecek  */
    117  1.150.14.5  jdolecek #include <kern/syscalls_autoload.c>
    118  1.150.14.5  jdolecek 
    119       1.126     pooka struct emul emul_netbsd = {
    120       1.126     pooka 	.e_name = "netbsd-rump",
    121       1.126     pooka 	.e_sysent = rump_sysent,
    122  1.150.14.3       tls #ifndef __HAVE_MINIMAL_EMUL
    123  1.150.14.3       tls 	.e_nsysent = SYS_NSYSENT,
    124  1.150.14.3       tls #endif
    125       1.128     pooka 	.e_vm_default_addr = uvm_default_mapaddr,
    126       1.138     pooka #ifdef __HAVE_SYSCALL_INTERN
    127       1.138     pooka 	.e_syscall_intern = syscall_intern,
    128       1.138     pooka #endif
    129  1.150.14.5  jdolecek 	.e_sc_autoload = netbsd_syscalls_autoload,
    130       1.126     pooka };
    131       1.126     pooka 
    132  1.150.14.4       tls cprng_strong_t *kern_cprng;
    133  1.150.14.4       tls 
    134  1.150.14.4       tls /* not used, but need the symbols for pointer comparisons */
    135  1.150.14.4       tls syncobj_t mutex_syncobj, rw_syncobj;
    136  1.150.14.4       tls 
    137        1.20     pooka int
    138        1.20     pooka kpause(const char *wmesg, bool intr, int timeo, kmutex_t *mtx)
    139        1.20     pooka {
    140        1.20     pooka 	extern int hz;
    141  1.150.14.4       tls 	int rv __diagused;
    142        1.78     pooka 	uint64_t sec, nsec;
    143       1.127     pooka 
    144        1.20     pooka 	if (mtx)
    145        1.20     pooka 		mutex_exit(mtx);
    146        1.40     pooka 
    147        1.78     pooka 	sec = timeo / hz;
    148        1.78     pooka 	nsec = (timeo % hz) * (1000000000 / hz);
    149  1.150.14.3       tls 	rv = rumpuser_clock_sleep(RUMPUSER_CLOCK_RELWALL, sec, nsec);
    150  1.150.14.3       tls 	KASSERT(rv == 0);
    151  1.150.14.3       tls 
    152        1.20     pooka 	if (mtx)
    153        1.20     pooka 		mutex_enter(mtx);
    154        1.20     pooka 
    155        1.20     pooka 	return 0;
    156        1.20     pooka }
    157        1.27     pooka 
    158        1.34      yamt vaddr_t
    159  1.150.14.3       tls calc_cache_size(vsize_t vasz, int pct, int va_pct)
    160        1.34      yamt {
    161        1.34      yamt 	paddr_t t;
    162        1.34      yamt 
    163        1.34      yamt 	t = (paddr_t)physmem * pct / 100 * PAGE_SIZE;
    164        1.34      yamt 	if ((vaddr_t)t != t) {
    165        1.34      yamt 		panic("%s: needs tweak", __func__);
    166        1.34      yamt 	}
    167        1.34      yamt 	return t;
    168        1.34      yamt }
    169        1.35    martin 
    170  1.150.14.5  jdolecek #define	RETURN_ADDRESS	(uintptr_t)__builtin_return_address(0)
    171  1.150.14.5  jdolecek 
    172        1.41     pooka void
    173        1.41     pooka assert_sleepable(void)
    174        1.41     pooka {
    175  1.150.14.5  jdolecek 	const char *reason = NULL;
    176        1.41     pooka 
    177        1.41     pooka 	/* always sleepable, although we should improve this */
    178  1.150.14.5  jdolecek 
    179  1.150.14.5  jdolecek 	if (!pserialize_not_in_read_section()) {
    180  1.150.14.5  jdolecek 		reason = "pserialize";
    181  1.150.14.5  jdolecek 	}
    182  1.150.14.5  jdolecek 
    183  1.150.14.5  jdolecek 	if (reason) {
    184  1.150.14.5  jdolecek 		panic("%s: %s caller=%p", __func__, reason,
    185  1.150.14.5  jdolecek 		    (void *)RETURN_ADDRESS);
    186  1.150.14.5  jdolecek 	}
    187        1.41     pooka }
    188        1.43     pooka 
    189        1.66     pooka void
    190        1.80    cegger module_init_md(void)
    191        1.66     pooka {
    192        1.66     pooka 
    193        1.66     pooka 	/*
    194        1.66     pooka 	 * Nothing for now.  However, we should load the librump
    195        1.66     pooka 	 * symbol table.
    196        1.66     pooka 	 */
    197        1.66     pooka }
    198        1.67     pooka 
    199  1.150.14.5  jdolecek /*
    200  1.150.14.5  jdolecek  * Try to emulate all the MD definitions of DELAY() / delay().
    201  1.150.14.5  jdolecek  * Would be nice to fix the #defines in MD headers, but this quicker.
    202  1.150.14.5  jdolecek  *
    203  1.150.14.5  jdolecek  * XXX: we'd need a rumpuser_clock_sleep_nowrap() here.  Since we
    204  1.150.14.5  jdolecek  * don't have it in the current hypercall revision, busyloop.
    205  1.150.14.5  jdolecek  * Note that rather than calibrate a loop delay and work with that,
    206  1.150.14.5  jdolecek  * get call gettime (which does not block) in a loop to make sure
    207  1.150.14.5  jdolecek  * we didn't get virtual ghosttime.  That might be slightly inaccurate
    208  1.150.14.5  jdolecek  * for very small delays ...
    209  1.150.14.5  jdolecek  *
    210  1.150.14.5  jdolecek  * The other option would be to run a thread in the hypervisor which
    211  1.150.14.5  jdolecek  * sleeps for us and we can wait for it using rumpuser_cv_wait_nowrap()
    212  1.150.14.5  jdolecek  * Probably too fussy.  Better just wait for hypercall rev 18 ;)
    213  1.150.14.5  jdolecek  */
    214        1.69     pooka static void
    215        1.69     pooka rump_delay(unsigned int us)
    216        1.67     pooka {
    217  1.150.14.5  jdolecek 	struct timespec target, tmp;
    218  1.150.14.5  jdolecek 	uint64_t sec, sec_ini, sec_now;
    219  1.150.14.5  jdolecek 	long nsec, nsec_ini, nsec_now;
    220  1.150.14.5  jdolecek 	int loops;
    221  1.150.14.5  jdolecek 
    222  1.150.14.5  jdolecek 	rumpuser_clock_gettime(RUMPUSER_CLOCK_ABSMONO, &sec_ini, &nsec_ini);
    223  1.150.14.5  jdolecek 
    224  1.150.14.5  jdolecek #ifdef __mac68k__
    225  1.150.14.5  jdolecek 	sec = us / 1000;
    226  1.150.14.5  jdolecek 	nsec = (us % 1000) * 1000000;
    227  1.150.14.5  jdolecek #else
    228        1.78     pooka 	sec = us / 1000000;
    229        1.78     pooka 	nsec = (us % 1000000) * 1000;
    230  1.150.14.5  jdolecek #endif
    231  1.150.14.5  jdolecek 
    232  1.150.14.5  jdolecek 	target.tv_sec = sec_ini;
    233  1.150.14.5  jdolecek 	tmp.tv_sec = sec;
    234  1.150.14.5  jdolecek 	target.tv_nsec = nsec_ini;
    235  1.150.14.5  jdolecek 	tmp.tv_nsec = nsec;
    236  1.150.14.5  jdolecek 	timespecadd(&target, &tmp, &target);
    237        1.67     pooka 
    238        1.78     pooka 	if (__predict_false(sec != 0))
    239        1.67     pooka 		printf("WARNING: over 1s delay\n");
    240        1.67     pooka 
    241  1.150.14.5  jdolecek 	for (loops = 0; loops < 1000*1000*100; loops++) {
    242  1.150.14.5  jdolecek 		struct timespec cur;
    243  1.150.14.5  jdolecek 
    244  1.150.14.5  jdolecek 		rumpuser_clock_gettime(RUMPUSER_CLOCK_ABSMONO,
    245  1.150.14.5  jdolecek 		    &sec_now, &nsec_now);
    246  1.150.14.5  jdolecek 		cur.tv_sec = sec_now;
    247  1.150.14.5  jdolecek 		cur.tv_nsec = nsec_now;
    248  1.150.14.5  jdolecek 		if (timespeccmp(&cur, &target, >=)) {
    249  1.150.14.5  jdolecek 			return;
    250  1.150.14.5  jdolecek 		}
    251  1.150.14.5  jdolecek 	}
    252  1.150.14.5  jdolecek 	printf("WARNING: DELAY ESCAPED\n");
    253        1.67     pooka }
    254        1.69     pooka void (*delay_func)(unsigned int) = rump_delay;
    255  1.150.14.5  jdolecek __strong_alias(delay,rump_delay);
    256  1.150.14.5  jdolecek __strong_alias(_delay,rump_delay);
    257  1.150.14.5  jdolecek 
    258  1.150.14.5  jdolecek /* Weak aliases for fstrans to be used unless librumpvfs is present. */
    259  1.150.14.5  jdolecek 
    260  1.150.14.5  jdolecek void rump_fstrans_start(struct mount *);
    261  1.150.14.5  jdolecek void
    262  1.150.14.5  jdolecek rump_fstrans_start(struct mount *mp)
    263  1.150.14.5  jdolecek {
    264  1.150.14.5  jdolecek 
    265  1.150.14.5  jdolecek }
    266  1.150.14.5  jdolecek __weak_alias(fstrans_start,rump_fstrans_start);
    267  1.150.14.5  jdolecek 
    268  1.150.14.5  jdolecek int rump_fstrans_start_nowait(struct mount *);
    269  1.150.14.5  jdolecek int
    270  1.150.14.5  jdolecek rump_fstrans_start_nowait(struct mount *mp)
    271  1.150.14.5  jdolecek {
    272  1.150.14.5  jdolecek 
    273  1.150.14.5  jdolecek 	return 0;
    274  1.150.14.5  jdolecek }
    275  1.150.14.5  jdolecek __weak_alias(fstrans_start_nowait,rump_fstrans_start_nowait);
    276  1.150.14.5  jdolecek 
    277  1.150.14.5  jdolecek void rump_fstrans_done(struct mount *);
    278  1.150.14.5  jdolecek void
    279  1.150.14.5  jdolecek rump_fstrans_done(struct mount *mp)
    280  1.150.14.5  jdolecek {
    281  1.150.14.5  jdolecek 
    282  1.150.14.5  jdolecek }
    283  1.150.14.5  jdolecek __weak_alias(fstrans_done,rump_fstrans_done);
    284        1.67     pooka 
    285       1.142     pooka /*
    286       1.142     pooka  * Provide weak aliases for tty routines used by printf.
    287       1.142     pooka  * They will be used unless the rumpkern_tty component is present.
    288       1.142     pooka  */
    289       1.142     pooka 
    290       1.142     pooka int rump_ttycheckoutq(struct tty *, int);
    291       1.105     pooka int
    292       1.142     pooka rump_ttycheckoutq(struct tty *tp, int wait)
    293        1.67     pooka {
    294        1.67     pooka 
    295        1.67     pooka 	return 1;
    296        1.67     pooka }
    297       1.142     pooka __weak_alias(ttycheckoutq,rump_ttycheckoutq);
    298        1.67     pooka 
    299       1.142     pooka int rump_tputchar(int, int, struct tty *);
    300       1.142     pooka int
    301       1.142     pooka rump_tputchar(int c, int flags, struct tty *tp)
    302       1.142     pooka {
    303       1.142     pooka 
    304       1.142     pooka 	cnputc(c);
    305       1.142     pooka 	return 0;
    306       1.142     pooka }
    307       1.142     pooka __weak_alias(tputchar,rump_tputchar);
    308       1.142     pooka 
    309        1.67     pooka void
    310        1.67     pooka cnputc(int c)
    311        1.67     pooka {
    312        1.67     pooka 
    313  1.150.14.3       tls 	rumpuser_putchar(c);
    314        1.67     pooka }
    315        1.67     pooka 
    316        1.67     pooka void
    317        1.80    cegger cnflush(void)
    318        1.67     pooka {
    319        1.67     pooka 
    320        1.67     pooka 	/* done */
    321        1.67     pooka }
    322        1.67     pooka 
    323       1.138     pooka void
    324  1.150.14.4       tls resettodr(void)
    325       1.138     pooka {
    326       1.138     pooka 
    327  1.150.14.4       tls 	/* setting clocks is not in the jurisdiction of rump kernels */
    328       1.138     pooka }
    329       1.144     pooka 
    330  1.150.14.4       tls #ifdef __HAVE_SYSCALL_INTERN
    331  1.150.14.4       tls void
    332  1.150.14.4       tls syscall_intern(struct proc *p)
    333       1.147     pooka {
    334       1.147     pooka 
    335  1.150.14.4       tls 	p->p_emuldata = NULL;
    336       1.147     pooka }
    337  1.150.14.4       tls #endif
    338       1.147     pooka 
    339  1.150.14.4       tls #ifdef LOCKDEBUG
    340       1.147     pooka void
    341  1.150.14.4       tls turnstile_print(volatile void *obj, void (*pr)(const char *, ...))
    342       1.147     pooka {
    343       1.147     pooka 
    344       1.147     pooka 	/* nada */
    345       1.147     pooka }
    346  1.150.14.4       tls #endif
    347       1.148     pooka 
    348       1.148     pooka void
    349  1.150.14.4       tls cpu_reboot(int howto, char *bootstr)
    350       1.148     pooka {
    351  1.150.14.4       tls 	int ruhow = 0;
    352  1.150.14.4       tls 	void *finiarg;
    353       1.148     pooka 
    354  1.150.14.4       tls 	printf("rump kernel halting...\n");
    355  1.150.14.4       tls 
    356  1.150.14.4       tls 	if (!RUMP_LOCALPROC_P(curproc))
    357  1.150.14.5  jdolecek 		finiarg = RUMP_SPVM2CTL(curproc->p_vmspace);
    358  1.150.14.4       tls 	else
    359  1.150.14.4       tls 		finiarg = NULL;
    360  1.150.14.4       tls 
    361  1.150.14.4       tls 	/* dump means we really take the dive here */
    362  1.150.14.4       tls 	if ((howto & RB_DUMP) || panicstr) {
    363  1.150.14.4       tls 		ruhow = RUMPUSER_PANIC;
    364  1.150.14.4       tls 		goto out;
    365  1.150.14.4       tls 	}
    366  1.150.14.4       tls 
    367  1.150.14.4       tls 	/* try to sync */
    368  1.150.14.4       tls 	if (!((howto & RB_NOSYNC) || panicstr)) {
    369  1.150.14.4       tls 		rump_vfs_fini();
    370  1.150.14.4       tls 	}
    371  1.150.14.4       tls 
    372  1.150.14.4       tls 	doshutdownhooks();
    373  1.150.14.4       tls 
    374  1.150.14.4       tls 	/* your wish is my command */
    375  1.150.14.4       tls 	if (howto & RB_HALT) {
    376  1.150.14.5  jdolecek 		printf("rump kernel halted (with RB_HALT, not exiting)\n");
    377  1.150.14.5  jdolecek 		rump_sysproxy_fini(finiarg);
    378  1.150.14.4       tls 		for (;;) {
    379  1.150.14.4       tls 			rumpuser_clock_sleep(RUMPUSER_CLOCK_RELWALL, 10, 0);
    380  1.150.14.4       tls 		}
    381  1.150.14.4       tls 	}
    382  1.150.14.4       tls 
    383  1.150.14.4       tls 	/* this function is __dead, we must exit */
    384  1.150.14.4       tls  out:
    385  1.150.14.5  jdolecek 	rump_sysproxy_fini(finiarg);
    386  1.150.14.4       tls 	rumpuser_exit(ruhow);
    387       1.148     pooka }
    388  1.150.14.5  jdolecek 
    389  1.150.14.5  jdolecek const char *
    390  1.150.14.5  jdolecek cpu_getmodel(void)
    391  1.150.14.5  jdolecek {
    392  1.150.14.5  jdolecek 
    393  1.150.14.5  jdolecek 	return "rumpcore (virtual)";
    394  1.150.14.5  jdolecek }
    395