xref: /xnu-8796.121.2/bsd/dev/dtrace/profile_prvd.c (revision c54f35ca767986246321eb901baf8f5ff7923f6a)

/*
 * CDDL HEADER START
 *
 * The contents of this file are subject to the terms of the
 * Common Development and Distribution License (the "License").
 * You may not use this file except in compliance with the License.
 *
 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
 * or http://www.opensolaris.org/os/licensing.
 * See the License for the specific language governing permissions
 * and limitations under the License.
 *
 * When distributing Covered Code, include this CDDL HEADER in each
 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
 * If applicable, add the following below this CDDL HEADER, with the
 * fields enclosed by brackets "[]" replaced with your own identifying
 * information: Portions Copyright [yyyy] [name of copyright owner]
 *
 * CDDL HEADER END
 */
/*
 * Copyright 2007 Sun Microsystems, Inc.  All rights reserved.
 * Use is subject to license terms.
 */

#include <kern/cpu_data.h>
#include <kern/thread.h>
#include <kern/assert.h>
#include <mach/thread_status.h>

#include <sys/param.h>
#include <sys/systm.h>
#include <sys/errno.h>
#include <sys/stat.h>
#include <sys/ioctl.h>
#include <sys/conf.h>
#include <sys/fcntl.h>
#include <miscfs/devfs/devfs.h>

#include <sys/dtrace.h>
#include <sys/dtrace_impl.h>

#include <sys/dtrace_glue.h>

#include <machine/pal_routines.h>

#if defined(__x86_64__)
extern x86_saved_state_t *find_kern_regs(thread_t);
#elif defined(__arm64__)
extern struct arm_saved_state *find_kern_regs(thread_t);
#else
#error Unknown architecture
#endif

extern void profile_init(void);

static dtrace_provider_id_t profile_id;

/*
 * Regardless of platform, the stack frames look like this in the case of the
 * profile provider:
 *
 *	profile_fire
 *	cyclic_expire
 *	cyclic_fire
 *	[ cbe ]
 *	[ interrupt code ]
 *
 * On x86, there are five frames from the generic interrupt code; further, the
 * interrupted instruction appears as its own stack frame, giving us a total of
 * 10.
 *
 * On SPARC, the picture is further complicated because the compiler
 * optimizes away tail-calls -- so the following frames are optimized away:
 *
 *      profile_fire
 *	cyclic_expire
 *
 * This gives three frames.  However, on DEBUG kernels, the cyclic_expire
 * frame cannot be tail-call eliminated, yielding four frames in this case.
 *
 * All of the above constraints lead to the mess below.  Yes, the profile
 * provider should ideally figure this out on-the-fly by hitting one of its own
 * probes and then walking its own stack trace.  This is complicated, however,
 * and the static definition doesn't seem to be overly brittle.  Still, we
 * allow for a manual override in case we get it completely wrong.
 */

#if defined(__x86_64__)
#define PROF_ARTIFICIAL_FRAMES  9
#elif defined(__arm64__)
#define PROF_ARTIFICIAL_FRAMES 8
#else
#error Unknown architecture
#endif

#define PROF_NAMELEN            15

#define PROF_PROFILE            0
#define PROF_TICK               1
#define PROF_PREFIX_PROFILE     "profile-"
#define PROF_PREFIX_TICK        "tick-"

typedef struct profile_probe {
	char            prof_name[PROF_NAMELEN];
	dtrace_id_t     prof_id;
	int             prof_kind;
	hrtime_t        prof_interval;
	cyclic_id_t     prof_cyclic;
} profile_probe_t;

typedef struct profile_probe_percpu {
	hrtime_t        profc_expected;
	hrtime_t        profc_interval;
	profile_probe_t *profc_probe;
} profile_probe_percpu_t;

hrtime_t        profile_interval_min = NANOSEC / 5000;          /* 5000 hz */
int             profile_aframes = 0;                            /* override */

static int profile_rates[] = {
	97, 199, 499, 997, 1999,
	4001, 4999, 0, 0, 0,
	0, 0, 0, 0, 0,
	0, 0, 0, 0, 0
};

static int profile_ticks[] = {
	1, 10, 100, 500, 1000,
	5000, 0, 0, 0, 0,
	0, 0, 0, 0, 0
};

/*
 * profile_max defines the upper bound on the number of profile probes that
 * can exist (this is to prevent malicious or clumsy users from exhausing
 * system resources by creating a slew of profile probes). At mod load time,
 * this gets its value from PROFILE_MAX_DEFAULT or profile-max-probes if it's
 * present in the profile.conf file.
 */
#define PROFILE_MAX_DEFAULT     1000    /* default max. number of probes */
static uint32_t profile_max;            /* maximum number of profile probes */
static uint32_t profile_total;  /* current number of profile probes */

static void
profile_fire(void *arg)
{
	profile_probe_percpu_t *pcpu = arg;
	profile_probe_t *prof = pcpu->profc_probe;
	hrtime_t late;

	late = dtrace_gethrtime() - pcpu->profc_expected;
	pcpu->profc_expected += pcpu->profc_interval;

#if defined(__x86_64__)
	x86_saved_state_t *kern_regs = find_kern_regs(current_thread());

	if (NULL != kern_regs) {
		/* Kernel was interrupted. */
		dtrace_probe(prof->prof_id, saved_state64(kern_regs)->isf.rip, 0x0, late, 0, 0);
	} else {
		pal_register_cache_state(current_thread(), VALID);
		/* Possibly a user interrupt */
		x86_saved_state_t   *tagged_regs = (x86_saved_state_t *)find_user_regs(current_thread());

		if (NULL == tagged_regs) {
			/* Too bad, so sad, no useful interrupt state. */
			dtrace_probe(prof->prof_id, 0xcafebabe,
			    0x0, late, 0, 0); /* XXX_BOGUS also see profile_usermode() below. */
		} else if (is_saved_state64(tagged_regs)) {
			x86_saved_state64_t *regs = saved_state64(tagged_regs);

			dtrace_probe(prof->prof_id, 0x0, regs->isf.rip, late, 0, 0);
		} else {
			x86_saved_state32_t *regs = saved_state32(tagged_regs);

			dtrace_probe(prof->prof_id, 0x0, regs->eip, late, 0, 0);
		}
	}
#elif defined(__arm64__)
	{
		arm_saved_state_t *arm_kern_regs = (arm_saved_state_t *) find_kern_regs(current_thread());

		// We should only come in here from interrupt context, so we should always have valid kernel regs
		assert(NULL != arm_kern_regs);

		if (saved_state64(arm_kern_regs)->cpsr & 0xF) {
			const uint64_t pc = ml_get_backtrace_pc(arm_kern_regs);

			/* Kernel was interrupted. */
			dtrace_probe(prof->prof_id, pc, 0x0, late, 0, 0);
		} else {
			/* Possibly a user interrupt */
			arm_saved_state_t   *arm_user_regs = (arm_saved_state_t *)find_user_regs(current_thread());

			if (NULL == arm_user_regs) {
				/* Too bad, so sad, no useful interrupt state. */
				dtrace_probe(prof->prof_id, 0xcafebabe, 0x0, late, 0, 0); /* XXX_BOGUS also see profile_usermode() below. */
			} else {
				dtrace_probe(prof->prof_id, 0x0, get_saved_state_pc(arm_user_regs), late, 0, 0);
			}
		}
	}
#else
#error Unknown architecture
#endif
}

static void
profile_tick(void *arg)
{
	profile_probe_t *prof = arg;

#if defined(__x86_64__)
	x86_saved_state_t *kern_regs = find_kern_regs(current_thread());

	if (NULL != kern_regs) {
		/* Kernel was interrupted. */
		dtrace_probe(prof->prof_id, saved_state64(kern_regs)->isf.rip, 0x0, 0, 0, 0);
	} else {
		pal_register_cache_state(current_thread(), VALID);
		/* Possibly a user interrupt */
		x86_saved_state_t   *tagged_regs = (x86_saved_state_t *)find_user_regs(current_thread());

		if (NULL == tagged_regs) {
			/* Too bad, so sad, no useful interrupt state. */
			dtrace_probe(prof->prof_id, 0xcafebabe,
			    0x0, 0, 0, 0); /* XXX_BOGUS also see profile_usermode() below. */
		} else if (is_saved_state64(tagged_regs)) {
			x86_saved_state64_t *regs = saved_state64(tagged_regs);

			dtrace_probe(prof->prof_id, 0x0, regs->isf.rip, 0, 0, 0);
		} else {
			x86_saved_state32_t *regs = saved_state32(tagged_regs);

			dtrace_probe(prof->prof_id, 0x0, regs->eip, 0, 0, 0);
		}
	}
#elif defined(__arm64__)
	{
		arm_saved_state_t *arm_kern_regs = (arm_saved_state_t *) find_kern_regs(current_thread());

		if (NULL != arm_kern_regs) {
			const uint64_t pc = ml_get_backtrace_pc(arm_kern_regs);

			/* Kernel was interrupted. */
			dtrace_probe(prof->prof_id, pc, 0x0, 0, 0, 0);
		} else {
			/* Possibly a user interrupt */
			arm_saved_state_t   *arm_user_regs = (arm_saved_state_t *)find_user_regs(current_thread());

			if (NULL == arm_user_regs) {
				/* Too bad, so sad, no useful interrupt state. */
				dtrace_probe(prof->prof_id, 0xcafebabe, 0x0, 0, 0, 0); /* XXX_BOGUS also see profile_usermode() below. */
			} else {
				dtrace_probe(prof->prof_id, 0x0, get_saved_state_pc(arm_user_regs), 0, 0, 0);
			}
		}
	}

#else
#error Unknown architecture
#endif
}

static void
profile_create(hrtime_t interval, const char *name, int kind)
{
	profile_probe_t *prof;

	if (interval < profile_interval_min) {
		return;
	}

	if (dtrace_probe_lookup(profile_id, NULL, NULL, name) != 0) {
		return;
	}

	os_atomic_inc(&profile_total, relaxed);
	if (profile_total > profile_max) {
		os_atomic_dec(&profile_total, relaxed);
		return;
	}

	if (PROF_TICK == kind) {
		prof = kmem_zalloc(sizeof(profile_probe_t), KM_SLEEP);
	} else {
		prof = kmem_zalloc(sizeof(profile_probe_t) + NCPU * sizeof(profile_probe_percpu_t), KM_SLEEP);
	}

	(void) strlcpy(prof->prof_name, name, sizeof(prof->prof_name));
	prof->prof_interval = interval;
	prof->prof_cyclic = CYCLIC_NONE;
	prof->prof_kind = kind;
	prof->prof_id = dtrace_probe_create(profile_id,
	    NULL, NULL, name,
	    profile_aframes ? profile_aframes : PROF_ARTIFICIAL_FRAMES, prof);
}

/*ARGSUSED*/
static void
profile_provide(void *arg, const dtrace_probedesc_t *desc)
{
#pragma unused(arg) /* __APPLE__ */
	int i, j, rate, kind;
	hrtime_t val = 0, mult = 1, len;
	const char *name, *suffix = NULL;

	const struct {
		const char *prefix;
		int kind;
	} types[] = {
		{ PROF_PREFIX_PROFILE, PROF_PROFILE },
		{ PROF_PREFIX_TICK, PROF_TICK },
		{ NULL, 0 }
	};

	const struct {
		const char *name;
		hrtime_t mult;
	} suffixes[] = {
		{ "ns", NANOSEC / NANOSEC },
		{ "nsec", NANOSEC / NANOSEC },
		{ "us", NANOSEC / MICROSEC },
		{ "usec", NANOSEC / MICROSEC },
		{ "ms", NANOSEC / MILLISEC },
		{ "msec", NANOSEC / MILLISEC },
		{ "s", NANOSEC / SEC },
		{ "sec", NANOSEC / SEC },
		{ "m", NANOSEC * (hrtime_t)60 },
		{ "min", NANOSEC * (hrtime_t)60 },
		{ "h", NANOSEC * (hrtime_t)(60 * 60) },
		{ "hour", NANOSEC * (hrtime_t)(60 * 60) },
		{ "d", NANOSEC * (hrtime_t)(24 * 60 * 60) },
		{ "day", NANOSEC * (hrtime_t)(24 * 60 * 60) },
		{ "hz", 0 },
		{ NULL, 0 }
	};

	if (desc == NULL) {
		char n[PROF_NAMELEN];

		/*
		 * If no description was provided, provide all of our probes.
		 */
		for (i = 0; i < (int)(sizeof(profile_rates) / sizeof(int)); i++) {
			if ((rate = profile_rates[i]) == 0) {
				continue;
			}

			(void) snprintf(n, PROF_NAMELEN, "%s%d",
			    PROF_PREFIX_PROFILE, rate);
			profile_create(NANOSEC / rate, n, PROF_PROFILE);
		}

		for (i = 0; i < (int)(sizeof(profile_ticks) / sizeof(int)); i++) {
			if ((rate = profile_ticks[i]) == 0) {
				continue;
			}

			(void) snprintf(n, PROF_NAMELEN, "%s%d",
			    PROF_PREFIX_TICK, rate);
			profile_create(NANOSEC / rate, n, PROF_TICK);
		}

		return;
	}

	name = desc->dtpd_name;

	for (i = 0; types[i].prefix != NULL; i++) {
		len = strlen(types[i].prefix);

		if (strncmp(name, types[i].prefix, len) != 0) {
			continue;
		}
		break;
	}

	if (types[i].prefix == NULL) {
		return;
	}

	kind = types[i].kind;
	j = strlen(name) - len;

	/*
	 * We need to start before any time suffix.
	 */
	for (j = strlen(name); j >= len; j--) {
		if (name[j] >= '0' && name[j] <= '9') {
			break;
		}
		suffix = &name[j];
	}

	if (!suffix) {
		suffix = &name[strlen(name)];
	}

	/*
	 * Now determine the numerical value present in the probe name.
	 */
	for (; j >= len; j--) {
		if (name[j] < '0' || name[j] > '9') {
			return;
		}

		val += (name[j] - '0') * mult;
		mult *= (hrtime_t)10;
	}

	if (val == 0) {
		return;
	}

	/*
	 * Look-up the suffix to determine the multiplier.
	 */
	for (i = 0, mult = 0; suffixes[i].name != NULL; i++) {
		/* APPLE NOTE: Darwin employs size bounded string operations */
		if (strncasecmp(suffixes[i].name, suffix, strlen(suffixes[i].name) + 1) == 0) {
			mult = suffixes[i].mult;
			break;
		}
	}

	if (suffixes[i].name == NULL && *suffix != '\0') {
		return;
	}

	if (mult == 0) {
		/*
		 * The default is frequency-per-second.
		 */
		val = NANOSEC / val;
	} else {
		val *= mult;
	}

	profile_create(val, name, kind);
}

/*ARGSUSED*/
static void
profile_destroy(void *arg, dtrace_id_t id, void *parg)
{
#pragma unused(arg,id) /* __APPLE__ */
	profile_probe_t *prof = parg;

	ASSERT(prof->prof_cyclic == CYCLIC_NONE);

	if (prof->prof_kind == PROF_TICK) {
		kmem_free(prof, sizeof(profile_probe_t));
	} else {
		kmem_free(prof, sizeof(profile_probe_t) + NCPU * sizeof(profile_probe_percpu_t));
	}

	ASSERT(profile_total >= 1);
	os_atomic_dec(&profile_total, relaxed);
}

/*ARGSUSED*/
static void
profile_online(void *arg, dtrace_cpu_t *cpu, cyc_handler_t *hdlr, cyc_time_t *when)
{
#pragma unused(cpu) /* __APPLE__ */
	profile_probe_t *prof = arg;
	profile_probe_percpu_t *pcpu;

	pcpu = ((profile_probe_percpu_t *)(&(prof[1]))) + cpu_number();
	pcpu->profc_probe = prof;

	hdlr->cyh_func = profile_fire;
	hdlr->cyh_arg = pcpu;
	hdlr->cyh_level = CY_HIGH_LEVEL;

	when->cyt_interval = prof->prof_interval;
	when->cyt_when = dtrace_gethrtime() + when->cyt_interval;

	pcpu->profc_expected = when->cyt_when;
	pcpu->profc_interval = when->cyt_interval;
}

/*ARGSUSED*/
static void
profile_offline(void *arg, dtrace_cpu_t *cpu, void *oarg)
{
	profile_probe_percpu_t *pcpu = oarg;

	ASSERT(pcpu->profc_probe == arg);
#pragma unused(pcpu,arg,cpu) /* __APPLE__ */
}

/*ARGSUSED*/
static int
profile_enable(void *arg, dtrace_id_t id, void *parg)
{
#pragma unused(arg,id) /* __APPLE__ */
	profile_probe_t *prof = parg;
	cyc_omni_handler_t omni;
	cyc_handler_t hdlr;
	cyc_time_t when;

	ASSERT(prof->prof_interval != 0);
	ASSERT(MUTEX_HELD(&cpu_lock));

	if (prof->prof_kind == PROF_TICK) {
		hdlr.cyh_func = profile_tick;
		hdlr.cyh_arg = prof;
		hdlr.cyh_level = CY_HIGH_LEVEL;

		when.cyt_interval = prof->prof_interval;
#if !defined(__APPLE__)
		when.cyt_when = dtrace_gethrtime() + when.cyt_interval;
#else
		when.cyt_when = 0;
#endif /* __APPLE__ */
	} else {
		ASSERT(prof->prof_kind == PROF_PROFILE);
		omni.cyo_online = profile_online;
		omni.cyo_offline = profile_offline;
		omni.cyo_arg = prof;
	}

	if (prof->prof_kind == PROF_TICK) {
		prof->prof_cyclic = cyclic_timer_add(&hdlr, &when);
	} else {
		prof->prof_cyclic = (cyclic_id_t)cyclic_add_omni(&omni); /* cast puns cyclic_id_list_t with cyclic_id_t */
	}

	return 0;
}

/*ARGSUSED*/
static void
profile_disable(void *arg, dtrace_id_t id, void *parg)
{
	profile_probe_t *prof = parg;

	ASSERT(prof->prof_cyclic != CYCLIC_NONE);
	ASSERT(MUTEX_HELD(&cpu_lock));

#pragma unused(arg,id)
	if (prof->prof_kind == PROF_TICK) {
		cyclic_timer_remove(prof->prof_cyclic);
	} else {
		cyclic_remove_omni((cyclic_id_list_t)prof->prof_cyclic); /* cast puns cyclic_id_list_t with cyclic_id_t */
	}
	prof->prof_cyclic = CYCLIC_NONE;
}

static uint64_t
profile_getarg(void *arg, dtrace_id_t id, void *parg, int argno, int aframes)
{
#pragma unused(arg, id, parg, argno, aframes)
	/*
	 * All the required arguments for the profile probe are passed directly
	 * to dtrace_probe, and we do not go through dtrace_getarg which doesn't
	 * know how to hop to the kernel stack from the interrupt stack like
	 * dtrace_getpcstack
	 */
	return 0;
}

static void
profile_getargdesc(void *arg, dtrace_id_t id, void *parg, dtrace_argdesc_t *desc)
{
#pragma unused(arg, id)
	profile_probe_t *prof = parg;
	const char *argdesc = NULL;
	switch (desc->dtargd_ndx) {
	case 0:
		argdesc = "void*";
		break;
	case 1:
		argdesc = "user_addr_t";
		break;
	case 2:
		if (prof->prof_kind == PROF_PROFILE) {
			argdesc = "hrtime_t";
		}
		break;
	}
	if (argdesc) {
		strlcpy(desc->dtargd_native, argdesc, DTRACE_ARGTYPELEN);
	} else {
		desc->dtargd_ndx = DTRACE_ARGNONE;
	}
}

/*
 * APPLE NOTE:  profile_usermode call not supported.
 */
static int
profile_usermode(void *arg, dtrace_id_t id, void *parg)
{
#pragma unused(arg,id,parg)
	return 1; /* XXX_BOGUS */
}

static dtrace_pattr_t profile_attr = {
	{ DTRACE_STABILITY_EVOLVING, DTRACE_STABILITY_EVOLVING, DTRACE_CLASS_COMMON },
	{ DTRACE_STABILITY_UNSTABLE, DTRACE_STABILITY_UNSTABLE, DTRACE_CLASS_UNKNOWN },
	{ DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
	{ DTRACE_STABILITY_EVOLVING, DTRACE_STABILITY_EVOLVING, DTRACE_CLASS_COMMON },
	{ DTRACE_STABILITY_EVOLVING, DTRACE_STABILITY_EVOLVING, DTRACE_CLASS_COMMON },
};

static dtrace_pops_t profile_pops = {
	.dtps_provide =         profile_provide,
	.dtps_provide_module =  NULL,
	.dtps_enable =          profile_enable,
	.dtps_disable =         profile_disable,
	.dtps_suspend =         NULL,
	.dtps_resume =          NULL,
	.dtps_getargdesc =      profile_getargdesc,
	.dtps_getargval =       profile_getarg,
	.dtps_usermode =        profile_usermode,
	.dtps_destroy =         profile_destroy
};

static int
profile_attach(dev_info_t *devi)
{
	if (ddi_create_minor_node(devi, "profile", S_IFCHR, 0,
	    DDI_PSEUDO, 0) == DDI_FAILURE ||
	    dtrace_register("profile", &profile_attr,
	    DTRACE_PRIV_KERNEL | DTRACE_PRIV_USER, NULL,
	    &profile_pops, NULL, &profile_id) != 0) {
		ddi_remove_minor_node(devi, NULL);
		return DDI_FAILURE;
	}

	profile_max = PROFILE_MAX_DEFAULT;

	return DDI_SUCCESS;
}

/*
 * APPLE NOTE:  profile_detach not implemented
 */
#if !defined(__APPLE__)
static int
profile_detach(dev_info_t *devi, ddi_detach_cmd_t cmd)
{
	switch (cmd) {
	case DDI_DETACH:
		break;
	case DDI_SUSPEND:
		return DDI_SUCCESS;
	default:
		return DDI_FAILURE;
	}

	if (dtrace_unregister(profile_id) != 0) {
		return DDI_FAILURE;
	}

	ddi_remove_minor_node(devi, NULL);
	return DDI_SUCCESS;
}
#endif /* __APPLE__ */

d_open_t _profile_open;

int
_profile_open(dev_t dev, int flags, int devtype, struct proc *p)
{
#pragma unused(dev,flags,devtype,p)
	return 0;
}

#define PROFILE_MAJOR  -24 /* let the kernel pick the device number */

static const struct cdevsw profile_cdevsw =
{
	.d_open = _profile_open,
	.d_close = eno_opcl,
	.d_read = eno_rdwrt,
	.d_write = eno_rdwrt,
	.d_ioctl = eno_ioctl,
	.d_stop = eno_stop,
	.d_reset = eno_reset,
	.d_select = eno_select,
	.d_mmap = eno_mmap,
	.d_strategy = eno_strat,
	.d_reserved_1 = eno_getc,
	.d_reserved_2 = eno_putc,
};

void
profile_init( void )
{
	int majdevno = cdevsw_add(PROFILE_MAJOR, &profile_cdevsw);

	if (majdevno < 0) {
		printf("profile_init: failed to allocate a major number!\n");
		return;
	}

	profile_attach((dev_info_t*)(uintptr_t)majdevno);
}
#undef PROFILE_MAJOR