xref: /illumos-gate/usr/src/cmd/mdb/common/kmdb/kaif_start.c (revision 9a016c63ca347047a236dff12f0da83aac8981d1)
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  * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 /*
29  * The main CPU-control loops, used to control masters and slaves.
30  */
31 
32 #include <sys/types.h>
33 
34 #include <kmdb/kaif.h>
35 #include <kmdb/kaif_start.h>
36 #include <kmdb/kmdb_asmutil.h>
37 #include <kmdb/kmdb_dpi_impl.h>
38 #include <kmdb/kmdb_kdi.h>
39 
40 #define	KAIF_SLAVE_CMD_SPIN	0
41 #define	KAIF_SLAVE_CMD_SWITCH	1
42 #define	KAIF_SLAVE_CMD_RESUME	2
43 #define	KAIF_SLAVE_CMD_FLUSH	3
44 #define	KAIF_SLAVE_CMD_REBOOT	4
45 #if defined(__sparc)
46 #define	KAIF_SLAVE_CMD_ACK	5
47 #endif
48 
49 
50 /*
51  * Used to synchronize attempts to set kaif_master_cpuid.  kaif_master_cpuid may
52  * be read without kaif_master_lock, and may be written by the current master
53  * CPU.
54  */
55 int kaif_master_cpuid = KAIF_MASTER_CPUID_UNSET;
56 static uintptr_t kaif_master_lock = 0;
57 
58 /*
59  * Used to ensure that all CPUs leave the debugger together. kaif_loop_lock must
60  * be held to write kaif_looping, but need not be held to read it.
61  */
62 static volatile uint_t kaif_looping;
63 static uintptr_t kaif_loop_lock;
64 
65 static volatile int kaif_slave_cmd;
66 static volatile int kaif_slave_tgt;	/* target cpuid for CMD_SWITCH */
67 
68 static void
69 kaif_lock_enter(uintptr_t *lock)
70 {
71 	while (cas(lock, 0, 1) != 0)
72 		continue;
73 	membar_producer();
74 }
75 
76 static void
77 kaif_lock_exit(uintptr_t *lock)
78 {
79 	*lock = 0;
80 	membar_producer();
81 }
82 
83 static int
84 kaif_master_loop(kaif_cpusave_t *cpusave)
85 {
86 	int notflushed, i;
87 
88 #if defined(__sparc)
89 	kaif_prom_rearm();
90 #endif
91 	kaif_trap_set_debugger();
92 
93 master_loop:
94 	switch (kmdb_dpi_reenter()) {
95 	case KMDB_DPI_CMD_SWITCH_CPU:
96 		/*
97 		 * We assume that the target CPU is a valid slave.  There's no
98 		 * easy way to complain here, so we'll assume that the caller
99 		 * has done the proper checking.
100 		 */
101 		if (kmdb_dpi_switch_target == cpusave->krs_cpu_id)
102 			break;
103 
104 		kaif_slave_tgt = kaif_master_cpuid = kmdb_dpi_switch_target;
105 		cpusave->krs_cpu_state = KAIF_CPU_STATE_SLAVE;
106 		membar_producer();
107 
108 		/*
109 		 * Switch back to the saved trap table before we switch CPUs --
110 		 * we need to make sure that only one CPU is on the debugger's
111 		 * table at a time.
112 		 */
113 		kaif_trap_set_saved(cpusave);
114 
115 		kaif_slave_cmd = KAIF_SLAVE_CMD_SWITCH;
116 
117 		/* The new master is now awake */
118 		return (KAIF_CPU_CMD_SWITCH);
119 
120 	case KMDB_DPI_CMD_RESUME_ALL:
121 	case KMDB_DPI_CMD_RESUME_UNLOAD:
122 		/*
123 		 * Resume everyone, clean up for next entry.
124 		 */
125 		kaif_master_cpuid = KAIF_MASTER_CPUID_UNSET;
126 		membar_producer();
127 		kaif_slave_cmd = KAIF_SLAVE_CMD_RESUME;
128 
129 		if (kmdb_dpi_work_required())
130 			kmdb_dpi_wrintr_fire();
131 
132 		kaif_trap_set_saved(cpusave);
133 
134 		return (KAIF_CPU_CMD_RESUME);
135 
136 	case KMDB_DPI_CMD_RESUME_MASTER:
137 		/*
138 		 * Single-CPU resume, which is performed on the debugger's
139 		 * trap table (so no need to switch back).
140 		 */
141 		return (KAIF_CPU_CMD_RESUME_MASTER);
142 
143 	case KMDB_DPI_CMD_FLUSH_CACHES:
144 		kaif_slave_cmd = KAIF_SLAVE_CMD_FLUSH;
145 
146 		/*
147 		 * Wait for the other cpus to finish flushing their caches.
148 		 */
149 		do {
150 			notflushed = 0;
151 			for (i = 0; i < kaif_ncpusave; i++) {
152 				kaif_cpusave_t *save = &kaif_cpusave[i];
153 
154 				if (save->krs_cpu_state ==
155 				    KAIF_CPU_STATE_SLAVE &&
156 				    !save->krs_cpu_flushed) {
157 					notflushed++;
158 					break;
159 				}
160 			}
161 		} while (notflushed > 0);
162 
163 		kaif_slave_cmd = KAIF_SLAVE_CMD_SPIN;
164 		break;
165 
166 #if defined(__i386) || defined(__amd64)
167 	case KMDB_DPI_CMD_REBOOT:
168 		/*
169 		 * Reboot must be initiated by CPU 0.  I could ask why, but I'm
170 		 * afraid that I don't want to know the answer.
171 		 */
172 		if (cpusave->krs_cpu_id == 0)
173 			return (KAIF_CPU_CMD_REBOOT);
174 
175 		kaif_slave_cmd = KAIF_SLAVE_CMD_REBOOT;
176 
177 		/*
178 		 * Spin forever, waiting for CPU 0 (apparently a slave) to
179 		 * reboot the system.
180 		 */
181 		for (;;)
182 			continue;
183 
184 		/*NOTREACHED*/
185 		break;
186 #endif
187 	}
188 
189 	goto master_loop;
190 }
191 
192 static int
193 kaif_slave_loop(kaif_cpusave_t *cpusave)
194 {
195 	int slavecmd, rv;
196 
197 #if defined(__sparc)
198 	/*
199 	 * If the user elects to drop to OBP from the debugger, some OBP
200 	 * implementations will cross-call the slaves.  We have to turn
201 	 * IE back on so we can receive the cross-calls.  If we don't,
202 	 * some OBP implementations will wait forever.
203 	 */
204 	interrupts_on();
205 #endif
206 
207 	/* Wait for duty to call */
208 	for (;;) {
209 		slavecmd = kaif_slave_cmd;
210 
211 		if (slavecmd == KAIF_SLAVE_CMD_SWITCH &&
212 		    kaif_slave_tgt == cpusave->krs_cpu_id) {
213 			kaif_slave_cmd = KAIF_SLAVE_CMD_SPIN;
214 			cpusave->krs_cpu_state = KAIF_CPU_STATE_MASTER;
215 			rv = KAIF_CPU_CMD_SWITCH;
216 			break;
217 
218 		} else if (slavecmd == KAIF_SLAVE_CMD_FLUSH) {
219 			kmdb_kdi_flush_caches();
220 			cpusave->krs_cpu_flushed = 1;
221 			continue;
222 
223 #if defined(__i386) || defined(__amd64)
224 		} else if (slavecmd == KAIF_SLAVE_CMD_REBOOT &&
225 		    cpusave->krs_cpu_id == 0) {
226 			rv = KAIF_CPU_CMD_REBOOT;
227 			break;
228 #endif
229 
230 		} else if (slavecmd == KAIF_SLAVE_CMD_RESUME) {
231 			rv = KAIF_CPU_CMD_RESUME;
232 			break;
233 #if defined(__sparc)
234 		} else if (slavecmd == KAIF_SLAVE_CMD_ACK) {
235 			cpusave->krs_cpu_acked = 1;
236 		} else if (cpusave->krs_cpu_acked &&
237 			slavecmd == KAIF_SLAVE_CMD_SPIN) {
238 			cpusave->krs_cpu_acked = 0;
239 #endif
240 		}
241 	}
242 
243 #if defined(__sparc)
244 	interrupts_off();
245 #endif
246 
247 	return (rv);
248 }
249 
250 static void
251 kaif_select_master(kaif_cpusave_t *cpusave)
252 {
253 	kaif_lock_enter(&kaif_master_lock);
254 
255 	if (kaif_master_cpuid == KAIF_MASTER_CPUID_UNSET) {
256 		/* This is the master. */
257 		kaif_master_cpuid = cpusave->krs_cpu_id;
258 		cpusave->krs_cpu_state = KAIF_CPU_STATE_MASTER;
259 		kaif_slave_cmd = KAIF_SLAVE_CMD_SPIN;
260 
261 		membar_producer();
262 
263 		kmdb_kdi_stop_other_cpus(cpusave->krs_cpu_id,
264 		    kaif_slave_entry);
265 
266 	} else {
267 		/* The master was already chosen - go be a slave */
268 		cpusave->krs_cpu_state = KAIF_CPU_STATE_SLAVE;
269 		membar_producer();
270 	}
271 
272 	kaif_lock_exit(&kaif_master_lock);
273 }
274 
275 int
276 kaif_main_loop(kaif_cpusave_t *cpusave)
277 {
278 	int cmd;
279 
280 	if (kaif_master_cpuid == KAIF_MASTER_CPUID_UNSET) {
281 		if (!kmdb_dpi_resume_requested &&
282 		    kmdb_kdi_get_unload_request()) {
283 			/*
284 			 * Special case: Unload requested before first debugger
285 			 * entry.  Don't stop the world, as there's nothing to
286 			 * clean up that can't be handled by the running kernel.
287 			 */
288 			cpusave->krs_cpu_state = KAIF_CPU_STATE_NONE;
289 			return (KAIF_CPU_CMD_RESUME);
290 		}
291 
292 		kaif_select_master(cpusave);
293 
294 #ifdef __sparc
295 		if (kaif_master_cpuid == cpusave->krs_cpu_id) {
296 			/*
297 			 * Everyone has arrived, so we can disarm the post-PROM
298 			 * entry point.
299 			 */
300 			*kaif_promexitarmp = 0;
301 			membar_producer();
302 		}
303 #endif
304 	} else if (kaif_master_cpuid == cpusave->krs_cpu_id) {
305 		cpusave->krs_cpu_state = KAIF_CPU_STATE_MASTER;
306 	} else {
307 		cpusave->krs_cpu_state = KAIF_CPU_STATE_SLAVE;
308 	}
309 
310 	cpusave->krs_cpu_flushed = 0;
311 
312 	kaif_lock_enter(&kaif_loop_lock);
313 	kaif_looping++;
314 	kaif_lock_exit(&kaif_loop_lock);
315 
316 	/*
317 	 * We know who the master and slaves are, so now they can go off
318 	 * to their respective loops.
319 	 */
320 	do {
321 		if (kaif_master_cpuid == cpusave->krs_cpu_id)
322 			cmd = kaif_master_loop(cpusave);
323 		else
324 			cmd = kaif_slave_loop(cpusave);
325 	} while (cmd == KAIF_CPU_CMD_SWITCH);
326 
327 	kaif_lock_enter(&kaif_loop_lock);
328 	kaif_looping--;
329 	kaif_lock_exit(&kaif_loop_lock);
330 
331 	cpusave->krs_cpu_state = KAIF_CPU_STATE_NONE;
332 
333 	if (cmd == KAIF_CPU_CMD_RESUME) {
334 		/*
335 		 * By this point, the master has directed the slaves to resume,
336 		 * and everyone is making their way to this point.  We're going
337 		 * to block here until all CPUs leave the master and slave
338 		 * loops.  When all have arrived, we'll turn them all loose.
339 		 * This barrier is required for two reasons:
340 		 *
341 		 * 1. There exists a race condition whereby a CPU could reenter
342 		 *    the debugger while another CPU is still in the slave loop
343 		 *    from this debugger entry.  This usually happens when the
344 		 *    current master releases the slaves, and makes it back to
345 		 *    the world before the slaves notice the release.  The
346 		 *    former master then triggers a debugger entry, and attempts
347 		 *    to stop the slaves for this entry before they've even
348 		 *    resumed from the last one.  When the slaves arrive here,
349 		 *    they'll have re-disabled interrupts, and will thus ignore
350 		 *    cross-calls until they finish resuming.
351 		 *
352 		 * 2. At the time of this writing, there exists a SPARC bug that
353 		 *    causes an apparently unsolicited interrupt vector trap
354 		 *    from OBP to one of the slaves.  This wouldn't normally be
355 		 *    a problem but for the fact that the cross-called CPU
356 		 *    encounters some sort of failure while in OBP.  OBP
357 		 *    recovers by executing the debugger-hook word, which sends
358 		 *    the slave back into the debugger, triggering a debugger
359 		 *    fault.  This problem seems to only happen during resume,
360 		 *    the result being that all CPUs save for the cross-called
361 		 *    one make it back into the world, while the cross-called
362 		 *    one is stuck at the debugger fault prompt.  Leave the
363 		 *    world in that state too long, and you'll get a mondo
364 		 *    timeout panic.  If we hold everyone here, we can give the
365 		 *    the user a chance to trigger a panic for further analysis.
366 		 *    To trigger the bug, "pool_unlock:b :c" and "while : ; do
367 		 *    psrset -p ; done".
368 		 *
369 		 * When the second item is fixed, the barrier can move into
370 		 * kaif_select_master(), immediately prior to the setting of
371 		 * kaif_master_cpuid.
372 		 */
373 		while (kaif_looping != 0)
374 			continue;
375 	}
376 
377 	return (cmd);
378 }
379 
380 
381 #if defined(__sparc)
382 
383 static int slave_loop_barrier_failures = 0;	/* for debug */
384 
385 /*
386  * There exist a race condition observed by some
387  * platforms where the kmdb master cpu exits to OBP via
388  * prom_enter_mon (e.g. "$q" command) and then later re-enter
389  * kmdb (typing "go") while the slaves are still proceeding
390  * from the OBP idle-loop back to the kmdb slave loop. The
391  * problem arises when the master cpu now back in kmdb proceed
392  * to re-enter OBP (e.g. doing a prom_read() from the kmdb main
393  * loop) while the slaves are still trying to get out of (the
394  * previous trip in) OBP into the safety of the kmdb slave loop.
395  * This routine forces the slaves to explicitly acknowledge
396  * that they are back in the slave loop. The master cpu can
397  * call this routine to ensure that all slave cpus are back
398  * in the slave loop before proceeding.
399  */
400 void
401 kaif_slave_loop_barrier(void)
402 {
403 	extern void kdi_usecwait(clock_t);
404 	int i;
405 	int not_acked;
406 	int timeout_count = 0;
407 
408 	kaif_slave_cmd = KAIF_SLAVE_CMD_ACK;
409 
410 	/*
411 	 * Wait for slave cpus to explicitly acknowledge
412 	 * that they are spinning in the slave loop.
413 	 */
414 	do {
415 		not_acked = 0;
416 		for (i = 0; i < kaif_ncpusave; i++) {
417 			kaif_cpusave_t *save = &kaif_cpusave[i];
418 
419 			if (save->krs_cpu_state ==
420 			    KAIF_CPU_STATE_SLAVE &&
421 			    !save->krs_cpu_acked) {
422 				not_acked++;
423 				break;
424 			}
425 		}
426 
427 		if (not_acked == 0)
428 			break;
429 
430 		/*
431 		 * Play it safe and do a timeout delay.
432 		 * We will do at most kaif_ncpusave delays before
433 		 * bailing out of this barrier.
434 		 */
435 		kdi_usecwait(200);
436 
437 	} while (++timeout_count < kaif_ncpusave);
438 
439 	if (not_acked > 0)
440 		/*
441 		 * we cannot establish a barrier with all
442 		 * the slave cpus coming back from OBP
443 		 * Record this fact for future debugging
444 		 */
445 		slave_loop_barrier_failures++;
446 
447 	kaif_slave_cmd = KAIF_SLAVE_CMD_SPIN;
448 }
449 #endif
450