Home | History | Annotate | Line # | Download | only in rumpkern
emul.c revision 1.176
      1  1.176     pooka /*	$NetBSD: emul.c,v 1.176 2016/01/18 23:21:28 pooka 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  1.176     pooka __KERNEL_RCSID(0, "$NetBSD: emul.c,v 1.176 2016/01/18 23:21:28 pooka Exp $");
     30    1.9     pooka 
     31    1.1     pooka #include <sys/param.h>
     32  1.176     pooka #include <sys/cprng.h>
     33  1.176     pooka #include <sys/filedesc.h>
     34    1.1     pooka #include <sys/kauth.h>
     35   1.66     pooka #include <sys/module.h>
     36   1.67     pooka #include <sys/reboot.h>
     37  1.154     pooka #include <sys/syscall.h>
     38   1.67     pooka 
     39   1.67     pooka #include <dev/cons.h>
     40    1.1     pooka 
     41   1.44     pooka #include <rump/rumpuser.h>
     42   1.44     pooka 
     43   1.10     pooka #include "rump_private.h"
     44    1.1     pooka 
     45  1.165     pooka void (*rump_vfs_fini)(void) = (void *)nullop;
     46  1.165     pooka 
     47  1.149     pooka /*
     48  1.149     pooka  * physmem is largely unused (except for nmbcluster calculations),
     49  1.149     pooka  * so pick a default value which suits ZFS.  if an application wants
     50  1.149     pooka  * a very small memory footprint, it can still adjust this before
     51  1.149     pooka  * calling rump_init()
     52  1.149     pooka  */
     53  1.149     pooka #define PHYSMEM 512*256
     54  1.149     pooka int physmem = PHYSMEM;
     55  1.149     pooka int nkmempages = PHYSMEM/2; /* from le chapeau */
     56  1.149     pooka #undef PHYSMEM
     57  1.149     pooka 
     58  1.162     pooka struct lwp lwp0 = {
     59  1.162     pooka 	.l_lid = 1,
     60  1.162     pooka 	.l_proc = &proc0,
     61  1.162     pooka 	.l_fd = &filedesc0,
     62  1.162     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  1.175     pooka 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  1.151       chs device_t booted_device;
     86  1.151       chs device_t booted_wedge;
     87   1.92     pooka int booted_partition;
     88   1.92     pooka 
     89  1.120     pooka /* XXX: unused */
     90   1.76     pooka kmutex_t tty_lock;
     91  1.120     pooka krwlock_t exec_lock;
     92   1.64     pooka 
     93  1.125     pooka struct lwplist alllwp = LIST_HEAD_INITIALIZER(alllwp);
     94  1.125     pooka 
     95  1.139     pooka /* sparc doesn't sport constant page size, pretend we have 4k pages */
     96  1.109     pooka #ifdef __sparc__
     97  1.109     pooka int nbpg = 4096;
     98  1.139     pooka int pgofset = 4096-1;
     99  1.139     pooka int pgshift = 12;
    100  1.109     pooka #endif
    101  1.109     pooka 
    102  1.146   tsutsui /* on sun3 VM_MAX_ADDRESS is a const variable */
    103  1.146   tsutsui /* XXX: should be moved into rump.c and initialize for sun3 and sun3x? */
    104  1.141     pooka #ifdef sun3
    105  1.146   tsutsui const vaddr_t kernbase = KERNBASE3;
    106  1.140     pooka #endif
    107  1.140     pooka 
    108  1.117     pooka struct loadavg averunnable = {
    109  1.117     pooka 	{ 0 * FSCALE,
    110  1.117     pooka 	  1 * FSCALE,
    111  1.117     pooka 	  11 * FSCALE, },
    112  1.117     pooka 	FSCALE,
    113  1.117     pooka };
    114  1.117     pooka 
    115  1.174  pgoyette /*
    116  1.174  pgoyette  * Include the autogenerated list of auto-loadable syscalls
    117  1.174  pgoyette  */
    118  1.174  pgoyette #include <kern/syscalls_autoload.c>
    119  1.174  pgoyette 
    120  1.126     pooka struct emul emul_netbsd = {
    121  1.126     pooka 	.e_name = "netbsd-rump",
    122  1.126     pooka 	.e_sysent = rump_sysent,
    123  1.154     pooka #ifndef __HAVE_MINIMAL_EMUL
    124  1.154     pooka 	.e_nsysent = SYS_NSYSENT,
    125  1.154     pooka #endif
    126  1.128     pooka 	.e_vm_default_addr = uvm_default_mapaddr,
    127  1.138     pooka #ifdef __HAVE_SYSCALL_INTERN
    128  1.138     pooka 	.e_syscall_intern = syscall_intern,
    129  1.138     pooka #endif
    130  1.174  pgoyette 	.e_sc_autoload = netbsd_syscalls_autoload,
    131  1.126     pooka };
    132  1.126     pooka 
    133  1.138     pooka u_int nprocs = 1;
    134   1.20     pooka 
    135  1.161     pooka cprng_strong_t *kern_cprng;
    136  1.161     pooka 
    137  1.164     pooka /* not used, but need the symbols for pointer comparisons */
    138  1.164     pooka syncobj_t mutex_syncobj, rw_syncobj;
    139  1.164     pooka 
    140   1.20     pooka int
    141   1.20     pooka kpause(const char *wmesg, bool intr, int timeo, kmutex_t *mtx)
    142   1.20     pooka {
    143   1.20     pooka 	extern int hz;
    144  1.166    justin 	int rv __diagused;
    145   1.78     pooka 	uint64_t sec, nsec;
    146  1.127     pooka 
    147   1.20     pooka 	if (mtx)
    148   1.20     pooka 		mutex_exit(mtx);
    149   1.40     pooka 
    150   1.78     pooka 	sec = timeo / hz;
    151   1.78     pooka 	nsec = (timeo % hz) * (1000000000 / hz);
    152  1.158     pooka 	rv = rumpuser_clock_sleep(RUMPUSER_CLOCK_RELWALL, sec, nsec);
    153  1.156     pooka 	KASSERT(rv == 0);
    154  1.156     pooka 
    155   1.20     pooka 	if (mtx)
    156   1.20     pooka 		mutex_enter(mtx);
    157   1.20     pooka 
    158   1.20     pooka 	return 0;
    159   1.20     pooka }
    160   1.27     pooka 
    161   1.27     pooka void
    162   1.32       bjs lwp_unsleep(lwp_t *l, bool cleanup)
    163   1.32       bjs {
    164   1.32       bjs 
    165   1.32       bjs 	KASSERT(mutex_owned(l->l_mutex));
    166   1.32       bjs 
    167  1.104     rmind 	(*l->l_syncobj->sobj_unsleep)(l, cleanup);
    168   1.32       bjs }
    169   1.34      yamt 
    170  1.145     pooka void
    171  1.145     pooka lwp_update_creds(struct lwp *l)
    172  1.145     pooka {
    173  1.145     pooka 	struct proc *p;
    174  1.145     pooka 	kauth_cred_t oldcred;
    175  1.145     pooka 
    176  1.145     pooka 	p = l->l_proc;
    177  1.145     pooka 	oldcred = l->l_cred;
    178  1.145     pooka 	l->l_prflag &= ~LPR_CRMOD;
    179  1.145     pooka 
    180  1.145     pooka 	mutex_enter(p->p_lock);
    181  1.145     pooka 	kauth_cred_hold(p->p_cred);
    182  1.145     pooka 	l->l_cred = p->p_cred;
    183  1.145     pooka 	mutex_exit(p->p_lock);
    184  1.145     pooka 
    185  1.145     pooka 	if (oldcred != NULL)
    186  1.145     pooka 		kauth_cred_free(oldcred);
    187  1.145     pooka }
    188  1.145     pooka 
    189   1.34      yamt vaddr_t
    190  1.155      para calc_cache_size(vsize_t vasz, int pct, int va_pct)
    191   1.34      yamt {
    192   1.34      yamt 	paddr_t t;
    193   1.34      yamt 
    194   1.34      yamt 	t = (paddr_t)physmem * pct / 100 * PAGE_SIZE;
    195   1.34      yamt 	if ((vaddr_t)t != t) {
    196   1.34      yamt 		panic("%s: needs tweak", __func__);
    197   1.34      yamt 	}
    198   1.34      yamt 	return t;
    199   1.34      yamt }
    200   1.35    martin 
    201   1.41     pooka void
    202   1.41     pooka assert_sleepable(void)
    203   1.41     pooka {
    204   1.41     pooka 
    205   1.41     pooka 	/* always sleepable, although we should improve this */
    206   1.41     pooka }
    207   1.43     pooka 
    208   1.66     pooka void
    209   1.80    cegger module_init_md(void)
    210   1.66     pooka {
    211   1.66     pooka 
    212   1.66     pooka 	/*
    213   1.66     pooka 	 * Nothing for now.  However, we should load the librump
    214   1.66     pooka 	 * symbol table.
    215   1.66     pooka 	 */
    216   1.66     pooka }
    217   1.67     pooka 
    218  1.167     pooka /*
    219  1.167     pooka  * Try to emulate all the MD definitions of DELAY() / delay().
    220  1.167     pooka  * Would be nice to fix the #defines in MD headers, but this quicker.
    221  1.168     pooka  *
    222  1.168     pooka  * XXX: we'd need a rumpuser_clock_sleep_nowrap() here.  Since we
    223  1.168     pooka  * don't have it in the current hypercall revision, busyloop.
    224  1.168     pooka  * Note that rather than calibrate a loop delay and work with that,
    225  1.168     pooka  * get call gettime (which does not block) in a loop to make sure
    226  1.168     pooka  * we didn't get virtual ghosttime.  That might be slightly inaccurate
    227  1.168     pooka  * for very small delays ...
    228  1.168     pooka  *
    229  1.168     pooka  * The other option would be to run a thread in the hypervisor which
    230  1.168     pooka  * sleeps for us and we can wait for it using rumpuser_cv_wait_nowrap()
    231  1.168     pooka  * Probably too fussy.  Better just wait for hypercall rev 18 ;)
    232  1.167     pooka  */
    233   1.69     pooka static void
    234   1.69     pooka rump_delay(unsigned int us)
    235   1.67     pooka {
    236  1.168     pooka 	struct timespec target, tmp;
    237  1.168     pooka 	uint64_t sec, sec_ini, sec_now;
    238  1.168     pooka 	long nsec, nsec_ini, nsec_now;
    239  1.168     pooka 	int loops;
    240  1.168     pooka 
    241  1.168     pooka 	rumpuser_clock_gettime(RUMPUSER_CLOCK_ABSMONO, &sec_ini, &nsec_ini);
    242   1.67     pooka 
    243  1.167     pooka #ifdef __mac68k__
    244  1.167     pooka 	sec = us / 1000;
    245  1.167     pooka 	nsec = (us % 1000) * 1000000;
    246  1.167     pooka #else
    247   1.78     pooka 	sec = us / 1000000;
    248   1.78     pooka 	nsec = (us % 1000000) * 1000;
    249  1.167     pooka #endif
    250   1.67     pooka 
    251  1.168     pooka 	target.tv_sec = sec_ini;
    252  1.168     pooka 	tmp.tv_sec = sec;
    253  1.168     pooka 	target.tv_nsec = nsec_ini;
    254  1.168     pooka 	tmp.tv_nsec = nsec;
    255  1.168     pooka 	timespecadd(&target, &tmp, &target);
    256  1.168     pooka 
    257   1.78     pooka 	if (__predict_false(sec != 0))
    258   1.67     pooka 		printf("WARNING: over 1s delay\n");
    259   1.67     pooka 
    260  1.168     pooka 	for (loops = 0; loops < 1000*1000*100; loops++) {
    261  1.168     pooka 		struct timespec cur;
    262  1.168     pooka 
    263  1.168     pooka 		rumpuser_clock_gettime(RUMPUSER_CLOCK_ABSMONO,
    264  1.168     pooka 		    &sec_now, &nsec_now);
    265  1.168     pooka 		cur.tv_sec = sec_now;
    266  1.168     pooka 		cur.tv_nsec = nsec_now;
    267  1.168     pooka 		if (timespeccmp(&cur, &target, >=)) {
    268  1.168     pooka 			return;
    269  1.168     pooka 		}
    270  1.168     pooka 	}
    271  1.168     pooka 	printf("WARNING: DELAY ESCAPED\n");
    272   1.67     pooka }
    273   1.69     pooka void (*delay_func)(unsigned int) = rump_delay;
    274  1.167     pooka __strong_alias(delay,rump_delay);
    275  1.167     pooka __strong_alias(_delay,rump_delay);
    276   1.67     pooka 
    277  1.142     pooka /*
    278  1.142     pooka  * Provide weak aliases for tty routines used by printf.
    279  1.142     pooka  * They will be used unless the rumpkern_tty component is present.
    280  1.142     pooka  */
    281  1.142     pooka 
    282  1.142     pooka int rump_ttycheckoutq(struct tty *, int);
    283  1.105     pooka int
    284  1.142     pooka rump_ttycheckoutq(struct tty *tp, int wait)
    285   1.67     pooka {
    286   1.67     pooka 
    287   1.67     pooka 	return 1;
    288   1.67     pooka }
    289  1.142     pooka __weak_alias(ttycheckoutq,rump_ttycheckoutq);
    290   1.67     pooka 
    291  1.142     pooka int rump_tputchar(int, int, struct tty *);
    292  1.142     pooka int
    293  1.142     pooka rump_tputchar(int c, int flags, struct tty *tp)
    294  1.142     pooka {
    295  1.142     pooka 
    296  1.142     pooka 	cnputc(c);
    297  1.142     pooka 	return 0;
    298  1.142     pooka }
    299  1.142     pooka __weak_alias(tputchar,rump_tputchar);
    300  1.142     pooka 
    301   1.67     pooka void
    302   1.67     pooka cnputc(int c)
    303   1.67     pooka {
    304   1.67     pooka 
    305  1.157     pooka 	rumpuser_putchar(c);
    306   1.67     pooka }
    307   1.67     pooka 
    308   1.67     pooka void
    309   1.80    cegger cnflush(void)
    310   1.67     pooka {
    311   1.67     pooka 
    312   1.67     pooka 	/* done */
    313   1.67     pooka }
    314   1.67     pooka 
    315  1.163     pooka void
    316  1.163     pooka resettodr(void)
    317  1.163     pooka {
    318  1.163     pooka 
    319  1.163     pooka 	/* setting clocks is not in the jurisdiction of rump kernels */
    320  1.163     pooka }
    321  1.163     pooka 
    322  1.138     pooka #ifdef __HAVE_SYSCALL_INTERN
    323  1.138     pooka void
    324  1.138     pooka syscall_intern(struct proc *p)
    325  1.138     pooka {
    326  1.138     pooka 
    327  1.160     pooka 	p->p_emuldata = NULL;
    328  1.138     pooka }
    329  1.138     pooka #endif
    330  1.144     pooka 
    331  1.148     pooka #ifdef LOCKDEBUG
    332  1.148     pooka void
    333  1.148     pooka turnstile_print(volatile void *obj, void (*pr)(const char *, ...))
    334  1.148     pooka {
    335  1.148     pooka 
    336  1.148     pooka 	/* nada */
    337  1.148     pooka }
    338  1.148     pooka #endif
    339  1.165     pooka 
    340  1.165     pooka void
    341  1.165     pooka cpu_reboot(int howto, char *bootstr)
    342  1.165     pooka {
    343  1.165     pooka 	int ruhow = 0;
    344  1.165     pooka 	void *finiarg;
    345  1.165     pooka 
    346  1.165     pooka 	printf("rump kernel halting...\n");
    347  1.165     pooka 
    348  1.165     pooka 	if (!RUMP_LOCALPROC_P(curproc))
    349  1.170     pooka 		finiarg = RUMP_SPVM2CTL(curproc->p_vmspace);
    350  1.165     pooka 	else
    351  1.165     pooka 		finiarg = NULL;
    352  1.165     pooka 
    353  1.165     pooka 	/* dump means we really take the dive here */
    354  1.165     pooka 	if ((howto & RB_DUMP) || panicstr) {
    355  1.165     pooka 		ruhow = RUMPUSER_PANIC;
    356  1.165     pooka 		goto out;
    357  1.165     pooka 	}
    358  1.165     pooka 
    359  1.165     pooka 	/* try to sync */
    360  1.165     pooka 	if (!((howto & RB_NOSYNC) || panicstr)) {
    361  1.165     pooka 		rump_vfs_fini();
    362  1.165     pooka 	}
    363  1.165     pooka 
    364  1.165     pooka 	doshutdownhooks();
    365  1.165     pooka 
    366  1.165     pooka 	/* your wish is my command */
    367  1.165     pooka 	if (howto & RB_HALT) {
    368  1.172     pooka 		printf("rump kernel halted (with RB_HALT, not exiting)\n");
    369  1.169     pooka 		rump_sysproxy_fini(finiarg);
    370  1.165     pooka 		for (;;) {
    371  1.165     pooka 			rumpuser_clock_sleep(RUMPUSER_CLOCK_RELWALL, 10, 0);
    372  1.165     pooka 		}
    373  1.165     pooka 	}
    374  1.165     pooka 
    375  1.165     pooka 	/* this function is __dead, we must exit */
    376  1.165     pooka  out:
    377  1.169     pooka 	rump_sysproxy_fini(finiarg);
    378  1.165     pooka 	rumpuser_exit(ruhow);
    379  1.165     pooka }
    380  1.173     pooka 
    381  1.173     pooka const char *
    382  1.173     pooka cpu_getmodel(void)
    383  1.173     pooka {
    384  1.173     pooka 
    385  1.173     pooka 	return "rumpcore (virtual)";
    386  1.173     pooka }
    387