xref: /titanic_41/usr/src/uts/common/xen/dtrace/xdt.c (revision a799b1e741b6f59012a469e6b57c40cb8061127b)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 /*
28  * Xen event provider for DTrace
29  *
30  * NOTE: This provider is PRIVATE. It is intended as a short-term solution and
31  * may disappear or be re-implemented at anytime.
32  *
33  * This provider isn't suitable as a general-purpose solution for a number of
34  * reasons. First and foremost, we rely on the Xen tracing mechanism and don't
35  * have any way to gather data other than that collected by the Xen trace
36  * buffers. Further, it does not fit into the DTrace model (see "Interacting
37  * with DTrace" below.)
38  *
39  *
40  * Tracing in Xen
41  * --------------
42  *
43  * Xen implements a tracing facility for generating and collecting execution
44  * event traces from the hypervisor. When tracing is enabled, compiled in
45  * probes record events in contiguous per-CPU trace buffers.
46  *
47  *               +---------+
48  * +------+      |         |
49  * | CPUn |----> | BUFFERn |
50  * +------+      |         |
51  *               +---------+- tbuf.va + (tbuf.size * n)
52  *               :         :
53  *               +---------+
54  * +------+      |         |
55  * | CPU1 |----> | BUFFER1 |
56  * +------+      |         |
57  *               +---------+- tbuf.va + tbuf.size
58  * +------+      |         |
59  * | CPU0 |----> | BUFFER0 |
60  * +------+      |         |
61  *               +---------+- tbuf.va
62  *
63  * Each CPU buffer consists of a metadata header followed by the trace records.
64  * The metadata consists of a producer/consumer pair of pointers into the buffer
65  * that point to the next record to be written and the next record to be read
66  * respectively.
67  *
68  * A trace record can be in one of two forms, depending on if the TSC is
69  * included. The record header indicates whether or not the TSC field is
70  * present.
71  *
72  * 1. Trace record without TSC:
73  * +------------------------------------------------------------+
74  * | HEADER(uint32_t) |            DATA FIELDS                  |
75  * +------------------------------------------------------------+
76  *
77  * 2. Trace record with TSC:
78  * +--------------------------------------------------------------------------+
79  * | HEADER(uint32_t) | TSC(uint64_t) |              DATA FIELDS              |
80  * +--------------------------------------------------------------------------+
81  *
82  * Where,
83  *
84  * HEADER bit field:
85  * +--------------------------------------------------------------------------+
86  * | C |  NDATA  |                        EVENT                               |
87  * +--------------------------------------------------------------------------+
88  *  31  30     28 27                                                         0
89  *
90  * EVENT: Event ID.
91  * NDATA: Number of populated data fields.
92  *     C: TSC included.
93  *
94  * DATA FIELDS:
95  * +--------------------------------------------------------------------------+
96  * | D1(uint32_t) | D2(uint32_t) | D3(uint32_t) |     . . .    | D7(uint32_t) |
97  * +--------------------------------------------------------------------------+
98  *
99  *
100  * Interacting with DTrace
101  * -----------------------
102  *
103  * Every xdt_poll_nsec nano-seconds we poll the trace buffers for data and feed
104  * each entry into dtrace_probe() with the corresponding probe ID for the event.
105  * As a result of this periodic collection implementation probe firings are
106  * asynchronous. This is the only sensible way to implement this form of
107  * provider, but because of its asynchronous nature asking things like
108  * "current CPU" and, more importantly, arbitrary questions about the context
109  * surrounding the probe firing are not meaningful. So, consumers should not
110  * attempt to infer anything beyond what is supplied via the probe arguments.
111  */
112 
113 #include <sys/xpv_user.h>
114 
115 #include <sys/types.h>
116 #include <sys/sysmacros.h>
117 #include <sys/modctl.h>
118 #include <sys/sunddi.h>
119 #include <sys/ddi.h>
120 #include <sys/conf.h>
121 #include <sys/devops.h>
122 #include <sys/stat.h>
123 #include <sys/cmn_err.h>
124 #include <sys/dtrace.h>
125 #include <sys/sdt.h>
126 #include <sys/cyclic.h>
127 #include <vm/seg_kmem.h>
128 #include <vm/hat_i86.h>
129 
130 #include <sys/hypervisor.h>
131 #include <xen/public/trace.h>
132 #include <xen/public/sched.h>
133 
134 #define	XDT_POLL_DEFAULT	100000000	/* default poll interval (ns) */
135 #define	XDT_POLL_MIN		10000000	/* min poll interval (ns) */
136 #define	XDT_TBUF_RETRY		50		/* tbuf disable retry count */
137 
138 /*
139  * The domid must match IDLE_DOMAIN_ID in xen.hg/xen/include/xen/sched.h
140  * in the xVM gate.
141  */
142 #define	IS_IDLE_DOM(domid)	(domid == 0x7FFFU)
143 
144 /* Macros to extract the domid and cpuid from a HVM trace data field */
145 #define	HVM_DOMID(d)		(d >> 16)
146 #define	HVM_VCPUID(d)		(d & 0xFFFF)
147 
148 #define	XDT_PROBE4(event, cpuid, arg0, arg1, arg2, arg3) {		\
149 	dtrace_id_t id = xdt_probemap[event];				\
150 	if (id)								\
151 		dtrace_probe(id, cpuid, arg0, arg1, arg2, arg3);	\
152 }									\
153 
154 #define	XDT_PROBE3(event, cpuid, arg0, arg1, arg2) \
155 	XDT_PROBE4(event, cpuid, arg0, arg1, arg2, 0)
156 
157 #define	XDT_PROBE2(event, cpuid, arg0, arg1) \
158 	XDT_PROBE4(event, cpuid, arg0, arg1, 0, 0)
159 
160 #define	XDT_PROBE1(event, cpuid, arg0) \
161 	XDT_PROBE4(event, cpuid, arg0, 0, 0, 0)
162 
163 #define	XDT_PROBE0(event, cpuid) \
164 	XDT_PROBE4(event, cpuid, 0, 0, 0, 0)
165 
166 /* Probe classes */
167 #define	XDT_SCHED			0
168 #define	XDT_MEM				1
169 #define	XDT_HVM				2
170 #define	XDT_GEN				3
171 #define	XDT_NCLASSES			4
172 
173 /* Probe events */
174 #define	XDT_EVT_INVALID			(-(int)1)
175 #define	XDT_SCHED_OFF_CPU		0
176 #define	XDT_SCHED_ON_CPU		1
177 #define	XDT_SCHED_IDLE_OFF_CPU		2
178 #define	XDT_SCHED_IDLE_ON_CPU		3
179 #define	XDT_SCHED_BLOCK			4
180 #define	XDT_SCHED_SLEEP			5
181 #define	XDT_SCHED_WAKE			6
182 #define	XDT_SCHED_YIELD			7
183 #define	XDT_SCHED_SHUTDOWN_POWEROFF	8
184 #define	XDT_SCHED_SHUTDOWN_REBOOT	9
185 #define	XDT_SCHED_SHUTDOWN_SUSPEND	10
186 #define	XDT_SCHED_SHUTDOWN_CRASH	11
187 #define	XDT_MEM_PAGE_GRANT_MAP		12
188 #define	XDT_MEM_PAGE_GRANT_UNMAP	13
189 #define	XDT_MEM_PAGE_GRANT_TRANSFER	14
190 #define	XDT_HVM_VMENTRY			15
191 #define	XDT_HVM_VMEXIT			16
192 #define	XDT_TRC_LOST_RECORDS		17
193 #define	XDT_NEVENTS			18
194 
195 typedef struct {
196 	const char	*pr_mod;	/* probe module */
197 	const char	*pr_name;	/* probe name */
198 	int		evt_id;		/* event id */
199 	uint_t		class;		/* probe class */
200 } xdt_probe_t;
201 
202 typedef struct {
203 	uint32_t	trc_mask;	/* trace mask */
204 	uint32_t	cnt;		/* num enabled probes in class */
205 } xdt_classinfo_t;
206 
207 typedef struct {
208 	ulong_t prev_domid;		/* previous dom executed */
209 	ulong_t prev_vcpuid;		/* previous vcpu executed */
210 	ulong_t prev_ctime;		/* time spent on cpu */
211 	ulong_t next_domid;		/* next dom to be scheduled */
212 	ulong_t next_vcpuid;		/* next vcpu to be scheduled */
213 	ulong_t next_wtime;		/* time spent waiting to get on cpu */
214 	ulong_t next_ts;		/* allocated time slice */
215 } xdt_schedinfo_t;
216 
217 static struct {
218 	uint_t cnt;			/* total num of trace buffers */
219 	size_t size;			/* size of each cpu buffer */
220 	mfn_t start_mfn;		/* starting mfn of buffers */
221 	caddr_t va;			/* va buffers are mapped into */
222 
223 	/* per-cpu buffers */
224 	struct t_buf **meta;		/* buffer metadata */
225 	struct t_rec **data;		/* buffer data records */
226 
227 	/* statistics */
228 	uint64_t stat_dropped_recs;	/* records dropped */
229 	uint64_t stat_spurious_cpu;	/* recs with garbage cpuids */
230 	uint64_t stat_spurious_switch;	/* inconsistent vcpu switches */
231 	uint64_t stat_unknown_shutdown;	/* unknown shutdown code */
232 	uint64_t stat_unknown_recs;	/* unknown records */
233 } tbuf;
234 
235 static char *xdt_stats[] = {
236 	"dropped_recs",
237 };
238 
239 /*
240  * Tunable variables
241  *
242  * The following may be tuned by adding a line to /etc/system that
243  * includes both the name of the module ("xdt") and the name of the variable.
244  * For example:
245  *     set xdt:xdt_tbuf_pages = 40
246  */
247 uint_t xdt_tbuf_pages = 20;			/* pages to alloc per-cpu buf */
248 
249 /*
250  * The following may be tuned by adding a line to
251  * /platform/i86xpv/kernel/drv/xdt.conf.
252  * For example:
253  *     xdt_poll_nsec = 200000000;
254  */
255 static hrtime_t xdt_poll_nsec;			/* trace buffer poll interval */
256 
257 /*
258  * Internal variables
259  */
260 static dev_info_t *xdt_devi;
261 static dtrace_provider_id_t xdt_id;
262 static uint_t xdt_ncpus;			/* total number of phys CPUs */
263 static uint32_t cur_trace_mask;			/* current trace mask */
264 static xdt_schedinfo_t *xdt_cpu_schedinfo;	/* per-cpu sched info */
265 dtrace_id_t xdt_probemap[XDT_NEVENTS];		/* map of enabled probes */
266 dtrace_id_t xdt_prid[XDT_NEVENTS];		/* IDs of registered events */
267 static cyclic_id_t xdt_cyclic = CYCLIC_NONE;
268 static kstat_t *xdt_kstats;
269 static xdt_classinfo_t xdt_classinfo[XDT_NCLASSES];
270 
271 static xdt_probe_t xdt_probe[] = {
272 	/* Sched probes */
273 	{ "sched", "off-cpu", XDT_SCHED_OFF_CPU, XDT_SCHED },
274 	{ "sched", "on-cpu", XDT_SCHED_ON_CPU, XDT_SCHED },
275 	{ "sched", "idle-off-cpu", XDT_SCHED_IDLE_OFF_CPU, XDT_SCHED },
276 	{ "sched", "idle-on-cpu", XDT_SCHED_IDLE_ON_CPU, XDT_SCHED },
277 	{ "sched", "block", XDT_SCHED_BLOCK, XDT_SCHED },
278 	{ "sched", "sleep", XDT_SCHED_SLEEP, XDT_SCHED },
279 	{ "sched", "wake", XDT_SCHED_WAKE, XDT_SCHED },
280 	{ "sched", "yield", XDT_SCHED_YIELD, XDT_SCHED },
281 	{ "sched", "shutdown-poweroff", XDT_SCHED_SHUTDOWN_POWEROFF,
282 		XDT_SCHED },
283 	{ "sched", "shutdown-reboot", XDT_SCHED_SHUTDOWN_REBOOT, XDT_SCHED },
284 	{ "sched", "shutdown-suspend", XDT_SCHED_SHUTDOWN_SUSPEND, XDT_SCHED },
285 	{ "sched", "shutdown-crash", XDT_SCHED_SHUTDOWN_CRASH, XDT_SCHED },
286 
287 	/* Memory probes */
288 	{ "mem", "page-grant-map", XDT_MEM_PAGE_GRANT_MAP, XDT_MEM },
289 	{ "mem", "page-grant-unmap", XDT_MEM_PAGE_GRANT_UNMAP, XDT_MEM },
290 	{ "mem", "page-grant-transfer", XDT_MEM_PAGE_GRANT_TRANSFER, XDT_MEM },
291 
292 	/* HVM probes */
293 	{ "hvm", "vmentry", XDT_HVM_VMENTRY, XDT_HVM },
294 	{ "hvm", "vmexit", XDT_HVM_VMEXIT, XDT_HVM },
295 
296 	/* Trace buffer related probes */
297 	{ "trace", "records-lost", XDT_TRC_LOST_RECORDS, XDT_GEN },
298 
299 	{ NULL }
300 };
301 
302 static inline uint32_t
303 xdt_nr_active_probes()
304 {
305 	int i;
306 	uint32_t tot = 0;
307 
308 	for (i = 0; i < XDT_NCLASSES; i++)
309 		tot += xdt_classinfo[i].cnt;
310 
311 	return (tot);
312 }
313 
314 static void
315 xdt_init_trace_masks(void)
316 {
317 	xdt_classinfo[XDT_SCHED].trc_mask = TRC_SCHED;
318 	xdt_classinfo[XDT_MEM].trc_mask = TRC_MEM;
319 	xdt_classinfo[XDT_HVM].trc_mask = TRC_HVM;
320 	xdt_classinfo[XDT_GEN].trc_mask = TRC_GEN;
321 }
322 
323 static int
324 xdt_kstat_update(kstat_t *ksp, int flag)
325 {
326 	kstat_named_t *knp;
327 
328 	if (flag != KSTAT_READ)
329 		return (EACCES);
330 
331 	knp = ksp->ks_data;
332 
333 	/*
334 	 * Assignment order should match that of the names in
335 	 * xdt_stats.
336 	 */
337 	(knp++)->value.ui64 = tbuf.stat_dropped_recs;
338 
339 	return (0);
340 }
341 
342 static void
343 xdt_kstat_init(void)
344 {
345 	int nstats = sizeof (xdt_stats) / sizeof (xdt_stats[0]);
346 	char **cp = xdt_stats;
347 	kstat_named_t *knp;
348 
349 	if ((xdt_kstats = kstat_create("xdt", 0, "trace_statistics", "misc",
350 	    KSTAT_TYPE_NAMED, nstats, 0)) == NULL)
351 		return;
352 
353 	xdt_kstats->ks_update = xdt_kstat_update;
354 
355 	knp = xdt_kstats->ks_data;
356 	while (nstats > 0) {
357 		kstat_named_init(knp, *cp, KSTAT_DATA_UINT64);
358 		knp++;
359 		cp++;
360 		nstats--;
361 	}
362 
363 	kstat_install(xdt_kstats);
364 }
365 
366 static int
367 xdt_sysctl_tbuf(xen_sysctl_tbuf_op_t *tbuf_op)
368 {
369 	xen_sysctl_t op;
370 	int xerr;
371 
372 	op.cmd = XEN_SYSCTL_tbuf_op;
373 	op.interface_version = XEN_SYSCTL_INTERFACE_VERSION;
374 	op.u.tbuf_op = *tbuf_op;
375 
376 	if ((xerr = HYPERVISOR_sysctl(&op)) != 0)
377 		return (xen_xlate_errcode(xerr));
378 
379 	*tbuf_op = op.u.tbuf_op;
380 	return (0);
381 }
382 
383 static int
384 xdt_map_trace_buffers(mfn_t mfn, caddr_t va, size_t len)
385 {
386 	x86pte_t pte;
387 	caddr_t const sva = va;
388 	caddr_t const eva = va + len;
389 	int xerr;
390 
391 	ASSERT(mfn != MFN_INVALID);
392 	ASSERT(va != NULL);
393 	ASSERT(IS_PAGEALIGNED(len));
394 
395 	for (; va < eva; va += MMU_PAGESIZE) {
396 		/*
397 		 * Ask the HAT to load a throwaway mapping to page zero, then
398 		 * overwrite it with the hypervisor mapping. It gets removed
399 		 * later via hat_unload().
400 		 */
401 		hat_devload(kas.a_hat, va, MMU_PAGESIZE, (pfn_t)0,
402 		    PROT_READ | HAT_UNORDERED_OK,
403 		    HAT_LOAD_NOCONSIST | HAT_LOAD);
404 
405 		pte = mmu_ptob((x86pte_t)mfn) | PT_VALID | PT_USER
406 		    | PT_FOREIGN | PT_WRITABLE;
407 
408 		xerr = HYPERVISOR_update_va_mapping_otherdomain((ulong_t)va,
409 		    pte, UVMF_INVLPG | UVMF_LOCAL, DOMID_XEN);
410 
411 		if (xerr != 0) {
412 			/* unmap pages loaded so far */
413 			size_t ulen = (uintptr_t)(va + MMU_PAGESIZE) -
414 			    (uintptr_t)sva;
415 			hat_unload(kas.a_hat, sva, ulen, HAT_UNLOAD_UNMAP);
416 			return (xen_xlate_errcode(xerr));
417 		}
418 
419 		mfn++;
420 	}
421 
422 	return (0);
423 }
424 
425 static int
426 xdt_attach_trace_buffers(void)
427 {
428 	xen_sysctl_tbuf_op_t tbuf_op;
429 	size_t len;
430 	int err;
431 	uint_t i;
432 
433 	/*
434 	 * Xen does not support trace buffer re-sizing. If the buffers
435 	 * have already been allocated we just use them as is.
436 	 */
437 	tbuf_op.cmd  = XEN_SYSCTL_TBUFOP_get_info;
438 	if ((err = xdt_sysctl_tbuf(&tbuf_op)) != 0)
439 		return (err);
440 
441 	if (tbuf_op.size == 0) {
442 		/* set trace buffer size */
443 		tbuf_op.cmd  = XEN_SYSCTL_TBUFOP_set_size;
444 		tbuf_op.size = xdt_tbuf_pages;
445 		(void) xdt_sysctl_tbuf(&tbuf_op);
446 
447 		/* get trace buffer info */
448 		tbuf_op.cmd  = XEN_SYSCTL_TBUFOP_get_info;
449 		if ((err = xdt_sysctl_tbuf(&tbuf_op)) != 0)
450 			return (err);
451 
452 		if (tbuf_op.size == 0) {
453 			cmn_err(CE_NOTE, "Couldn't allocate trace buffers.");
454 			return (ENOBUFS);
455 		}
456 	}
457 
458 	tbuf.size = tbuf_op.size;
459 	tbuf.start_mfn = (mfn_t)tbuf_op.buffer_mfn;
460 	tbuf.cnt = xdt_ncpus;
461 
462 	ASSERT(tbuf.start_mfn != MFN_INVALID);
463 	ASSERT(tbuf.cnt > 0);
464 
465 	len = tbuf.size * tbuf.cnt;
466 	tbuf.va = vmem_alloc(heap_arena, len, VM_SLEEP);
467 
468 	if ((err = xdt_map_trace_buffers(tbuf.start_mfn, tbuf.va, len)) != 0) {
469 		vmem_free(heap_arena, tbuf.va, len);
470 		tbuf.va = NULL;
471 		return (err);
472 	}
473 
474 	tbuf.meta = (struct t_buf **)kmem_alloc(tbuf.cnt * sizeof (*tbuf.meta),
475 	    KM_SLEEP);
476 	tbuf.data = (struct t_rec **)kmem_alloc(tbuf.cnt * sizeof (*tbuf.data),
477 	    KM_SLEEP);
478 
479 	for (i = 0; i < tbuf.cnt; i++) {
480 		void *cpu_buf = (void *)(tbuf.va + (tbuf.size * i));
481 		tbuf.meta[i] = cpu_buf;
482 		tbuf.data[i] = (struct t_rec *)((uintptr_t)cpu_buf +
483 		    sizeof (struct t_buf));
484 
485 		/* throw away stale trace records */
486 		tbuf.meta[i]->cons = tbuf.meta[i]->prod;
487 	}
488 
489 	return (0);
490 }
491 
492 static void
493 xdt_detach_trace_buffers(void)
494 {
495 	size_t len = tbuf.size * tbuf.cnt;
496 
497 	ASSERT(tbuf.va != NULL);
498 
499 	hat_unload(kas.a_hat, tbuf.va, len,
500 	    HAT_UNLOAD_UNMAP | HAT_UNLOAD_UNLOCK);
501 	vmem_free(heap_arena, tbuf.va, len);
502 	kmem_free(tbuf.meta, tbuf.cnt * sizeof (*tbuf.meta));
503 	kmem_free(tbuf.data, tbuf.cnt * sizeof (*tbuf.data));
504 }
505 
506 static inline size_t
507 xdt_process_rec(uint_t cpuid, struct t_rec *rec)
508 {
509 	xdt_schedinfo_t *sp = &xdt_cpu_schedinfo[cpuid];
510 	int eid;
511 	uint32_t *data;
512 	size_t rec_size;
513 
514 	ASSERT(rec != NULL);
515 	ASSERT(xdt_ncpus == xpv_nr_phys_cpus());
516 
517 	if (cpuid >= xdt_ncpus) {
518 		tbuf.stat_spurious_cpu++;
519 		goto done;
520 	}
521 
522 	data = rec->cycles_included ? rec->u.cycles.extra_u32 :
523 	    rec->u.nocycles.extra_u32;
524 
525 	switch (rec->event) {
526 	/*
527 	 * Sched probes
528 	 */
529 	case TRC_SCHED_SWITCH_INFPREV:
530 		/*
531 		 * Info on vCPU being de-scheduled
532 		 *
533 		 * data[0] = prev domid
534 		 * data[1] = time spent on pcpu
535 		 */
536 		sp->prev_domid = data[0];
537 		sp->prev_ctime = data[1];
538 		break;
539 
540 	case TRC_SCHED_SWITCH_INFNEXT:
541 		/*
542 		 * Info on next vCPU to be scheduled
543 		 *
544 		 * data[0] = next domid
545 		 * data[1] = time spent waiting to get on cpu
546 		 * data[2] = time slice
547 		 */
548 		sp->next_domid = data[0];
549 		sp->next_wtime = data[1];
550 		sp->next_ts = data[2];
551 		break;
552 
553 	case TRC_SCHED_SWITCH:
554 		/*
555 		 * vCPU switch
556 		 *
557 		 * data[0] = prev domid
558 		 * data[1] = prev vcpuid
559 		 * data[2] = next domid
560 		 * data[3] = next vcpuid
561 		 */
562 		if (data[0] != sp->prev_domid &&
563 		    data[2] != sp->next_domid) {
564 			/* prev and next info don't match doms being sched'd */
565 			tbuf.stat_spurious_switch++;
566 			goto done;
567 		}
568 
569 		sp->prev_vcpuid = data[1];
570 		sp->next_vcpuid = data[3];
571 
572 		XDT_PROBE3(IS_IDLE_DOM(sp->prev_domid)?
573 		    XDT_SCHED_IDLE_OFF_CPU:XDT_SCHED_OFF_CPU,
574 		    cpuid, sp->prev_domid, sp->prev_vcpuid, sp->prev_ctime);
575 
576 		XDT_PROBE4(IS_IDLE_DOM(sp->next_domid)?
577 		    XDT_SCHED_IDLE_ON_CPU:XDT_SCHED_ON_CPU,
578 		    cpuid, sp->next_domid, sp->next_vcpuid, sp->next_wtime,
579 		    sp->next_ts);
580 		break;
581 
582 	case TRC_SCHED_BLOCK:
583 		/*
584 		 * vCPU blocked
585 		 *
586 		 * data[0] = domid
587 		 * data[1] = vcpuid
588 		 */
589 		XDT_PROBE2(XDT_SCHED_BLOCK, cpuid, data[0], data[1]);
590 		break;
591 
592 	case TRC_SCHED_SLEEP:
593 		/*
594 		 * Put vCPU to sleep
595 		 *
596 		 * data[0] = domid
597 		 * data[1] = vcpuid
598 		 */
599 		XDT_PROBE2(XDT_SCHED_SLEEP, cpuid, data[0], data[1]);
600 		break;
601 
602 	case TRC_SCHED_WAKE:
603 		/*
604 		 * Wake up vCPU
605 		 *
606 		 * data[0] = domid
607 		 * data[1] = vcpuid
608 		 */
609 		XDT_PROBE2(XDT_SCHED_WAKE, cpuid, data[0], data[1]);
610 		break;
611 
612 	case TRC_SCHED_YIELD:
613 		/*
614 		 * vCPU yielded
615 		 *
616 		 * data[0] = domid
617 		 * data[1] = vcpuid
618 		 */
619 		XDT_PROBE2(XDT_SCHED_YIELD, cpuid, data[0], data[1]);
620 		break;
621 
622 	case TRC_SCHED_SHUTDOWN:
623 		/*
624 		 * Guest shutting down
625 		 *
626 		 * data[0] = domid
627 		 * data[1] = initiating vcpu
628 		 * data[2] = shutdown code
629 		 */
630 		switch (data[2]) {
631 		case SHUTDOWN_poweroff:
632 			eid = XDT_SCHED_SHUTDOWN_POWEROFF;
633 			break;
634 		case SHUTDOWN_reboot:
635 			eid = XDT_SCHED_SHUTDOWN_REBOOT;
636 			break;
637 		case SHUTDOWN_suspend:
638 			eid = XDT_SCHED_SHUTDOWN_SUSPEND;
639 			break;
640 		case SHUTDOWN_crash:
641 			eid = XDT_SCHED_SHUTDOWN_CRASH;
642 			break;
643 		default:
644 			tbuf.stat_unknown_shutdown++;
645 			goto done;
646 		}
647 
648 		XDT_PROBE1(eid, cpuid, data[0]);
649 		break;
650 
651 	/*
652 	 * Mem probes
653 	 */
654 	case TRC_MEM_PAGE_GRANT_MAP:
655 		/*
656 		 * Guest mapped page grant
657 		 *
658 		 * data[0] = domid
659 		 */
660 		XDT_PROBE1(XDT_MEM_PAGE_GRANT_MAP, cpuid, data[0]);
661 		break;
662 
663 	case TRC_MEM_PAGE_GRANT_UNMAP:
664 		/*
665 		 * Guest unmapped page grant
666 		 *
667 		 * data[0] = domid
668 		 */
669 		XDT_PROBE1(XDT_MEM_PAGE_GRANT_UNMAP, cpuid, data[0]);
670 		break;
671 
672 	case TRC_MEM_PAGE_GRANT_TRANSFER:
673 		/*
674 		 * Page grant is being transferred
675 		 *
676 		 * data[0] = target domid
677 		 */
678 		XDT_PROBE1(XDT_MEM_PAGE_GRANT_TRANSFER, cpuid, data[0]);
679 		break;
680 
681 	/*
682 	 * HVM probes
683 	 */
684 	case TRC_HVM_VMENTRY:
685 		/*
686 		 * Return to guest via vmx_launch/vmrun
687 		 *
688 		 * data[0] = (domid<<16 + vcpuid)
689 		 */
690 		XDT_PROBE2(XDT_HVM_VMENTRY, cpuid, HVM_DOMID(data[0]),
691 		    HVM_VCPUID(data[0]));
692 		break;
693 
694 	case TRC_HVM_VMEXIT64:
695 		/*
696 		 * Entry into VMEXIT handler
697 		 *
698 		 * data[0] = (domid<<16 + vcpuid)
699 		 * data[1] = cpu vendor specific exit code
700 		 * data[2] = guest rip(0:31)
701 		 * data[3] = guest rip(32:64)
702 		 */
703 		XDT_PROBE4(XDT_HVM_VMEXIT, cpuid, HVM_DOMID(data[0]),
704 		    HVM_VCPUID(data[0]), data[1],
705 		    ((uint64_t)data[3]<<32) | data[2]);
706 		break;
707 
708 	case TRC_LOST_RECORDS:
709 		XDT_PROBE0(XDT_TRC_LOST_RECORDS, cpuid);
710 		tbuf.stat_dropped_recs++;
711 		break;
712 
713 	default:
714 		tbuf.stat_unknown_recs++;
715 		break;
716 	}
717 
718 done:
719 	rec_size = 4 + (rec->cycles_included ? 8 : 0) + (rec->extra_u32 * 4);
720 	return (rec_size);
721 }
722 
723 /*ARGSUSED*/
724 static void
725 xdt_tbuf_scan(void *arg)
726 {
727 	uint_t cpuid;
728 	size_t tbuf_data_size;
729 	struct t_rec *rec;
730 	uintptr_t data;
731 	uint32_t prod, cons;
732 	uint32_t offset, end_offset;
733 
734 	tbuf_data_size = tbuf.size - sizeof (struct t_buf);
735 
736 	/* scan all cpu buffers for new records */
737 	for (cpuid = 0; cpuid < tbuf.cnt; cpuid++) {
738 		cons = tbuf.meta[cpuid]->cons;
739 		prod = tbuf.meta[cpuid]->prod;
740 		membar_consumer(); /* read prod /then/ data */
741 
742 		/* see <xen/public/trace.h> */
743 		ASSERT(cons < 2 * tbuf_data_size);
744 		ASSERT(prod < 2 * tbuf_data_size);
745 
746 		if (prod == cons)
747 			continue;
748 
749 		offset = cons % tbuf_data_size;
750 		end_offset = prod % tbuf_data_size;
751 
752 		if (offset >= end_offset) {
753 			/* read up to the end of the buffer */
754 			while (offset != tbuf_data_size) {
755 				data = (uintptr_t)tbuf.data[cpuid] + offset;
756 				rec = (struct t_rec *)data;
757 				ASSERT((caddr_t)rec < tbuf.va + (tbuf.size *
758 				    (cpuid + 1)));
759 				offset += xdt_process_rec(cpuid, rec);
760 			}
761 			offset = 0; /* wrap around */
762 		}
763 
764 		while (offset != end_offset) {
765 			data = (uintptr_t)tbuf.data[cpuid] + offset;
766 			rec = (struct t_rec *)data;
767 			ASSERT((caddr_t)rec < tbuf.va + (tbuf.size *
768 			    (cpuid + 1)));
769 			offset += xdt_process_rec(cpuid, rec);
770 		}
771 
772 		membar_exit(); /* read data /then/ update cons */
773 		tbuf.meta[cpuid]->cons = prod;
774 	}
775 }
776 
777 static void
778 xdt_cyclic_enable(void)
779 {
780 	cyc_handler_t hdlr;
781 	cyc_time_t when;
782 
783 	ASSERT(MUTEX_HELD(&cpu_lock));
784 
785 	hdlr.cyh_func = xdt_tbuf_scan;
786 	hdlr.cyh_arg = NULL;
787 	hdlr.cyh_level = CY_LOW_LEVEL;
788 
789 	when.cyt_interval = xdt_poll_nsec;
790 	when.cyt_when = dtrace_gethrtime() + when.cyt_interval;
791 
792 	xdt_cyclic = cyclic_add(&hdlr, &when);
793 }
794 
795 static void
796 xdt_probe_create(xdt_probe_t *p)
797 {
798 	ASSERT(p != NULL && p->pr_mod != NULL);
799 
800 	if (dtrace_probe_lookup(xdt_id, p->pr_mod, NULL, p->pr_name) != 0)
801 		return;
802 
803 	xdt_prid[p->evt_id] = dtrace_probe_create(xdt_id, p->pr_mod, NULL,
804 	    p->pr_name, dtrace_mach_aframes(), p);
805 }
806 
807 /*ARGSUSED*/
808 static void
809 xdt_provide(void *arg, const dtrace_probedesc_t *desc)
810 {
811 	const char *mod, *name;
812 	int i;
813 
814 	if (desc == NULL) {
815 		for (i = 0; xdt_probe[i].pr_mod != NULL; i++) {
816 			xdt_probe_create(&xdt_probe[i]);
817 		}
818 	} else {
819 		mod = desc->dtpd_mod;
820 		name = desc->dtpd_name;
821 		for (i = 0; xdt_probe[i].pr_mod != NULL; i++) {
822 			int l1 = strlen(xdt_probe[i].pr_name);
823 			int l2 = strlen(xdt_probe[i].pr_mod);
824 			if (strncmp(name, xdt_probe[i].pr_name, l1) == 0 &&
825 			    strncmp(mod, xdt_probe[i].pr_mod, l2) == 0)
826 				break;
827 		}
828 
829 		if (xdt_probe[i].pr_mod == NULL)
830 			return;
831 		xdt_probe_create(&xdt_probe[i]);
832 	}
833 
834 }
835 
836 /*ARGSUSED*/
837 static void
838 xdt_destroy(void *arg, dtrace_id_t id, void *parg)
839 {
840 	xdt_probe_t *p = parg;
841 	xdt_prid[p->evt_id] = 0;
842 }
843 
844 static void
845 xdt_set_trace_mask(uint32_t mask)
846 {
847 	xen_sysctl_tbuf_op_t tbuf_op;
848 
849 	tbuf_op.cmd  = XEN_SYSCTL_TBUFOP_set_evt_mask;
850 	tbuf_op.evt_mask = mask;
851 	(void) xdt_sysctl_tbuf(&tbuf_op);
852 }
853 
854 /*ARGSUSED*/
855 static int
856 xdt_enable(void *arg, dtrace_id_t id, void *parg)
857 {
858 	xdt_probe_t *p = parg;
859 	xen_sysctl_tbuf_op_t tbuf_op;
860 
861 	ASSERT(MUTEX_HELD(&cpu_lock));
862 	ASSERT(xdt_prid[p->evt_id] != 0);
863 
864 	xdt_probemap[p->evt_id] = xdt_prid[p->evt_id];
865 	xdt_classinfo[p->class].cnt++;
866 
867 	if (xdt_classinfo[p->class].cnt == 1) {
868 		/* set the trace mask for this class */
869 		cur_trace_mask |= xdt_classinfo[p->class].trc_mask;
870 		xdt_set_trace_mask(cur_trace_mask);
871 	}
872 
873 	if (xdt_cyclic == CYCLIC_NONE) {
874 		tbuf_op.cmd = XEN_SYSCTL_TBUFOP_enable;
875 		if (xdt_sysctl_tbuf(&tbuf_op) != 0) {
876 			cmn_err(CE_NOTE, "Couldn't enable hypervisor tracing.");
877 			return (-1);
878 		}
879 
880 		xdt_cyclic_enable();
881 	}
882 	return (0);
883 }
884 
885 /*ARGSUSED*/
886 static void
887 xdt_disable(void *arg, dtrace_id_t id, void *parg)
888 {
889 	xdt_probe_t *p = parg;
890 	xen_sysctl_tbuf_op_t tbuf_op;
891 	int i, err;
892 
893 	ASSERT(MUTEX_HELD(&cpu_lock));
894 	ASSERT(xdt_probemap[p->evt_id] != 0);
895 	ASSERT(xdt_probemap[p->evt_id] == xdt_prid[p->evt_id]);
896 	ASSERT(xdt_classinfo[p->class].cnt > 0);
897 
898 	/*
899 	 * We could be here in the slight window between the cyclic firing and
900 	 * a call to dtrace_probe() occurring. We need to be careful if we tear
901 	 * down any shared state.
902 	 */
903 
904 	xdt_probemap[p->evt_id] = 0;
905 	xdt_classinfo[p->class].cnt--;
906 
907 	if (xdt_nr_active_probes() == 0) {
908 		cur_trace_mask = 0;
909 
910 		if (xdt_cyclic == CYCLIC_NONE)
911 			return;
912 
913 		/*
914 		 * We will try to disable the trace buffers. If we fail for some
915 		 * reason we will try again, up to a count of XDT_TBUF_RETRY.
916 		 * If we still aren't successful we try to set the trace mask
917 		 * to 0 in order to prevent trace records from being written.
918 		 */
919 		tbuf_op.cmd = XEN_SYSCTL_TBUFOP_disable;
920 		i = 0;
921 		do {
922 			err = xdt_sysctl_tbuf(&tbuf_op);
923 		} while ((err != 0) && (++i < XDT_TBUF_RETRY));
924 
925 		if (err != 0) {
926 			cmn_err(CE_NOTE,
927 			    "Couldn't disable hypervisor tracing.");
928 			xdt_set_trace_mask(0);
929 		} else {
930 			cyclic_remove(xdt_cyclic);
931 			xdt_cyclic = CYCLIC_NONE;
932 			/*
933 			 * We don't bother making the hypercall to set
934 			 * the trace mask, since it will be reset when
935 			 * tracing is re-enabled.
936 			 */
937 		}
938 	} else if (xdt_classinfo[p->class].cnt == 0) {
939 		cur_trace_mask ^= xdt_classinfo[p->class].trc_mask;
940 		/* other probes are enabled, so add the sub-class mask back */
941 		cur_trace_mask |= 0xF000;
942 		xdt_set_trace_mask(cur_trace_mask);
943 	}
944 }
945 
946 static dtrace_pattr_t xdt_attr = {
947 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_PLATFORM },
948 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_PLATFORM },
949 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
950 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_PLATFORM },
951 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_PLATFORM },
952 };
953 
954 static dtrace_pops_t xdt_pops = {
955 	xdt_provide,		/* dtps_provide() */
956 	NULL,			/* dtps_provide_module() */
957 	xdt_enable,		/* dtps_enable() */
958 	xdt_disable,		/* dtps_disable() */
959 	NULL,			/* dtps_suspend() */
960 	NULL,			/* dtps_resume() */
961 	NULL,			/* dtps_getargdesc() */
962 	NULL,			/* dtps_getargval() */
963 	NULL,			/* dtps_usermode() */
964 	xdt_destroy		/* dtps_destroy() */
965 };
966 
967 static int
968 xdt_attach(dev_info_t *devi, ddi_attach_cmd_t cmd)
969 {
970 	int val;
971 
972 	if (!DOMAIN_IS_INITDOMAIN(xen_info))
973 		return (DDI_FAILURE);
974 
975 	switch (cmd) {
976 	case DDI_ATTACH:
977 		break;
978 
979 	case DDI_RESUME:
980 		/*
981 		 * We might support proper suspend/resume in the future, so,
982 		 * return DDI_FAILURE for now.
983 		 */
984 		return (DDI_FAILURE);
985 
986 	default:
987 		return (DDI_FAILURE);
988 	}
989 
990 	xdt_ncpus = xpv_nr_phys_cpus();
991 	ASSERT(xdt_ncpus > 0);
992 
993 	if (ddi_create_minor_node(devi, "xdt", S_IFCHR, 0, DDI_PSEUDO, 0) ==
994 	    DDI_FAILURE || xdt_attach_trace_buffers() != 0 ||
995 	    dtrace_register("xdt", &xdt_attr, DTRACE_PRIV_KERNEL, NULL,
996 	    &xdt_pops, NULL, &xdt_id) != 0) {
997 		if (tbuf.va != NULL)
998 			xdt_detach_trace_buffers();
999 		ddi_remove_minor_node(devi, NULL);
1000 		return (DDI_FAILURE);
1001 	}
1002 
1003 	val = ddi_getprop(DDI_DEV_T_ANY, devi, DDI_PROP_DONTPASS,
1004 	    "xdt_poll_nsec", XDT_POLL_DEFAULT);
1005 	xdt_poll_nsec = MAX(val, XDT_POLL_MIN);
1006 
1007 	xdt_cpu_schedinfo = (xdt_schedinfo_t *)kmem_alloc(xdt_ncpus *
1008 	    sizeof (xdt_schedinfo_t), KM_SLEEP);
1009 	xdt_init_trace_masks();
1010 	xdt_kstat_init();
1011 
1012 	xdt_devi = devi;
1013 	ddi_report_dev(devi);
1014 	return (DDI_SUCCESS);
1015 }
1016 
1017 static int
1018 xdt_detach(dev_info_t *devi, ddi_detach_cmd_t cmd)
1019 {
1020 	switch (cmd) {
1021 	case DDI_DETACH:
1022 		break;
1023 
1024 	case DDI_SUSPEND:
1025 		/*
1026 		 * We might support proper suspend/resume in the future. So
1027 		 * return DDI_FAILURE for now.
1028 		 */
1029 		return (DDI_FAILURE);
1030 
1031 	default:
1032 		return (DDI_FAILURE);
1033 	}
1034 
1035 	if (dtrace_unregister(xdt_id) != 0)
1036 		return (DDI_FAILURE);
1037 
1038 	xdt_detach_trace_buffers();
1039 	kmem_free(xdt_cpu_schedinfo, xdt_ncpus * sizeof (xdt_schedinfo_t));
1040 	if (xdt_cyclic != CYCLIC_NONE)
1041 		cyclic_remove(xdt_cyclic);
1042 	if (xdt_kstats != NULL)
1043 		kstat_delete(xdt_kstats);
1044 	xdt_devi = (void *)0;
1045 	ddi_remove_minor_node(devi, NULL);
1046 
1047 	return (DDI_SUCCESS);
1048 }
1049 
1050 /*ARGSUSED*/
1051 static int
1052 xdt_info(dev_info_t *devi, ddi_info_cmd_t infocmd, void *arg, void **result)
1053 {
1054 	int error;
1055 
1056 	switch (infocmd) {
1057 	case DDI_INFO_DEVT2DEVINFO:
1058 		*result = xdt_devi;
1059 		error = DDI_SUCCESS;
1060 		break;
1061 	case DDI_INFO_DEVT2INSTANCE:
1062 		*result = (void *)0;
1063 		error = DDI_SUCCESS;
1064 		break;
1065 	default:
1066 		error = DDI_FAILURE;
1067 	}
1068 	return (error);
1069 }
1070 
1071 static struct cb_ops xdt_cb_ops = {
1072 	nulldev,		/* open(9E) */
1073 	nodev,			/* close(9E) */
1074 	nodev,			/* strategy(9E) */
1075 	nodev,			/* print(9E) */
1076 	nodev,			/* dump(9E) */
1077 	nodev,			/* read(9E) */
1078 	nodev,			/* write(9E) */
1079 	nodev,			/* ioctl(9E) */
1080 	nodev,			/* devmap(9E) */
1081 	nodev,			/* mmap(9E) */
1082 	nodev,			/* segmap(9E) */
1083 	nochpoll,		/* chpoll(9E) */
1084 	ddi_prop_op,		/* prop_op(9E) */
1085 	NULL,			/* streamtab(9S) */
1086 	D_MP | D_64BIT | D_NEW	/* cb_flag */
1087 };
1088 
1089 static struct dev_ops xdt_ops = {
1090 	DEVO_REV,		/* devo_rev */
1091 	0,			/* devo_refcnt */
1092 	xdt_info,		/* getinfo(9E) */
1093 	nulldev,		/* identify(9E) */
1094 	nulldev,		/* probe(9E) */
1095 	xdt_attach,		/* attach(9E) */
1096 	xdt_detach,		/* detach(9E) */
1097 	nulldev,		/* devo_reset */
1098 	&xdt_cb_ops,		/* devo_cb_ops */
1099 	NULL,			/* devo_bus_ops */
1100 	NULL,			/* power(9E) */
1101 	ddi_quiesce_not_needed,	/* devo_quiesce */
1102 };
1103 
1104 
1105 static struct modldrv modldrv = {
1106 	&mod_driverops,
1107 	"Hypervisor event tracing",
1108 	&xdt_ops
1109 };
1110 
1111 static struct modlinkage modlinkage = {
1112 	MODREV_1,
1113 	&modldrv,
1114 	NULL
1115 };
1116 
1117 int
1118 _init(void)
1119 {
1120 	return (mod_install(&modlinkage));
1121 }
1122 
1123 int
1124 _fini(void)
1125 {
1126 	return (mod_remove(&modlinkage));
1127 }
1128 
1129 int
1130 _info(struct modinfo *modinfop)
1131 {
1132 	return (mod_info(&modlinkage, modinfop));
1133 }
1134