xref: /illumos-gate/usr/src/uts/i86pc/io/mp_platform_misc.c (revision b89e420ae1290e425c29db875ec0c0546006eec7)
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 (c) 2010, Oracle and/or its affiliates. All rights reserved.
23  */
24 /*
25  * Copyright (c) 2010, Intel Corporation.
26  * All rights reserved.
27  */
28 
29 /*
30  * PSMI 1.1 extensions are supported only in 2.6 and later versions.
31  * PSMI 1.2 extensions are supported only in 2.7 and later versions.
32  * PSMI 1.3 and 1.4 extensions are supported in Solaris 10.
33  * PSMI 1.5 extensions are supported in Solaris Nevada.
34  * PSMI 1.6 extensions are supported in Solaris Nevada.
35  * PSMI 1.7 extensions are supported in Solaris Nevada.
36  */
37 #define	PSMI_1_7
38 
39 #include <sys/processor.h>
40 #include <sys/time.h>
41 #include <sys/psm.h>
42 #include <sys/smp_impldefs.h>
43 #include <sys/inttypes.h>
44 #include <sys/cram.h>
45 #include <sys/acpi/acpi.h>
46 #include <sys/acpica.h>
47 #include <sys/psm_common.h>
48 #include <sys/apic.h>
49 #include <sys/apic_common.h>
50 #include <sys/pit.h>
51 #include <sys/ddi.h>
52 #include <sys/sunddi.h>
53 #include <sys/ddi_impldefs.h>
54 #include <sys/pci.h>
55 #include <sys/promif.h>
56 #include <sys/x86_archext.h>
57 #include <sys/cpc_impl.h>
58 #include <sys/uadmin.h>
59 #include <sys/panic.h>
60 #include <sys/debug.h>
61 #include <sys/archsystm.h>
62 #include <sys/trap.h>
63 #include <sys/machsystm.h>
64 #include <sys/cpuvar.h>
65 #include <sys/rm_platter.h>
66 #include <sys/privregs.h>
67 #include <sys/cyclic.h>
68 #include <sys/note.h>
69 #include <sys/pci_intr_lib.h>
70 #include <sys/sunndi.h>
71 #include <sys/hpet.h>
72 #include <sys/clock.h>
73 
74 /*
75  * Part of mp_platfrom_common.c that's used only by pcplusmp & xpv_psm
76  * but not apix.
77  * These functions may be moved to xpv_psm later when apix and pcplusmp
78  * are merged together
79  */
80 
81 /*
82  *	Local Function Prototypes
83  */
84 static void apic_mark_vector(uchar_t oldvector, uchar_t newvector);
85 static void apic_xlate_vector_free_timeout_handler(void *arg);
86 static int apic_check_stuck_interrupt(apic_irq_t *irq_ptr, int old_bind_cpu,
87     int new_bind_cpu, int apicindex, int intin_no, int which_irq,
88     struct ioapic_reprogram_data *drep);
89 static int apic_setup_irq_table(dev_info_t *dip, int irqno,
90     struct apic_io_intr *intrp, struct intrspec *ispec, iflag_t *intr_flagp,
91     int type);
92 static void apic_try_deferred_reprogram(int ipl, int vect);
93 static void delete_defer_repro_ent(int which_irq);
94 static void apic_ioapic_wait_pending_clear(int ioapicindex,
95     int intin_no);
96 
97 extern int apic_acpi_translate_pci_irq(dev_info_t *dip, int busid, int devid,
98     int ipin, int *pci_irqp, iflag_t *intr_flagp);
99 extern int apic_handle_pci_pci_bridge(dev_info_t *idip, int child_devno,
100     int child_ipin, struct apic_io_intr **intrp);
101 extern uchar_t acpi_find_ioapic(int irq);
102 extern struct apic_io_intr *apic_find_io_intr_w_busid(int irqno, int busid);
103 extern int apic_find_bus_id(int bustype);
104 extern int apic_find_intin(uchar_t ioapic, uchar_t intin);
105 extern void apic_record_rdt_entry(apic_irq_t *irqptr, int irq);
106 
107 extern	int apic_sci_vect;
108 extern	iflag_t apic_sci_flags;
109 /* ACPI HPET interrupt configuration; -1 if HPET not used */
110 extern	int apic_hpet_vect;
111 extern	iflag_t apic_hpet_flags;
112 extern	int	apic_intr_policy;
113 extern	char *psm_name;
114 
115 /*
116  * number of bits per byte, from <sys/param.h>
117  */
118 #define	UCHAR_MAX	UINT8_MAX
119 
120 /* Max wait time (in repetitions) for flags to clear in an RDT entry. */
121 extern int apic_max_reps_clear_pending;
122 
123 /* The irq # is implicit in the array index: */
124 struct ioapic_reprogram_data apic_reprogram_info[APIC_MAX_VECTOR+1];
125 /*
126  * APIC_MAX_VECTOR + 1 is the maximum # of IRQs as well. ioapic_reprogram_info
127  * is indexed by IRQ number, NOT by vector number.
128  */
129 
130 extern	int	apic_int_busy_mark;
131 extern	int	apic_int_free_mark;
132 extern	int	apic_diff_for_redistribution;
133 extern	int	apic_sample_factor_redistribution;
134 extern	int	apic_redist_cpu_skip;
135 extern	int	apic_num_imbalance;
136 extern	int	apic_num_rebind;
137 
138 /* timeout for xlate_vector, mark_vector */
139 int	apic_revector_timeout = 16 * 10000; /* 160 millisec */
140 
141 extern int	apic_defconf;
142 extern int	apic_irq_translate;
143 
144 extern int	apic_use_acpi_madt_only;	/* 1=ONLY use MADT from ACPI */
145 
146 extern	uchar_t	apic_io_vectbase[MAX_IO_APIC];
147 
148 extern	boolean_t ioapic_mask_workaround[MAX_IO_APIC];
149 
150 /*
151  * First available slot to be used as IRQ index into the apic_irq_table
152  * for those interrupts (like MSI/X) that don't have a physical IRQ.
153  */
154 extern int apic_first_avail_irq;
155 
156 /*
157  * apic_defer_reprogram_lock ensures that only one processor is handling
158  * deferred interrupt programming at *_intr_exit time.
159  */
160 static	lock_t	apic_defer_reprogram_lock;
161 
162 /*
163  * The current number of deferred reprogrammings outstanding
164  */
165 uint_t	apic_reprogram_outstanding = 0;
166 
167 #ifdef DEBUG
168 /*
169  * Counters that keep track of deferred reprogramming stats
170  */
171 uint_t	apic_intr_deferrals = 0;
172 uint_t	apic_intr_deliver_timeouts = 0;
173 uint_t	apic_last_ditch_reprogram_failures = 0;
174 uint_t	apic_deferred_setup_failures = 0;
175 uint_t	apic_defer_repro_total_retries = 0;
176 uint_t	apic_defer_repro_successes = 0;
177 uint_t	apic_deferred_spurious_enters = 0;
178 #endif
179 
180 extern	int	apic_io_max;
181 extern	struct apic_io_intr *apic_io_intrp;
182 
183 uchar_t	apic_vector_to_irq[APIC_MAX_VECTOR+1];
184 
185 extern	uint32_t	eisa_level_intr_mask;
186 	/* At least MSB will be set if EISA bus */
187 
188 extern	int	apic_pci_bus_total;
189 extern	uchar_t	apic_single_pci_busid;
190 
191 /*
192  * Following declarations are for revectoring; used when ISRs at different
193  * IPLs share an irq.
194  */
195 static	lock_t	apic_revector_lock;
196 int	apic_revector_pending = 0;
197 static	uchar_t	*apic_oldvec_to_newvec;
198 static	uchar_t	*apic_newvec_to_oldvec;
199 
200 /* ACPI Interrupt Source Override Structure ptr */
201 extern ACPI_MADT_INTERRUPT_OVERRIDE *acpi_isop;
202 extern int acpi_iso_cnt;
203 
204 /*
205  * Auto-configuration routines
206  */
207 
208 /*
209  * Initialise vector->ipl and ipl->pri arrays. level_intr and irqtable
210  * are also set to NULL. vector->irq is set to a value which cannot map
211  * to a real irq to show that it is free.
212  */
213 void
214 apic_init_common(void)
215 {
216 	int	i, j, indx;
217 	int	*iptr;
218 
219 	/*
220 	 * Initialize apic_ipls from apic_vectortoipl.  This array is
221 	 * used in apic_intr_enter to determine the IPL to use for the
222 	 * corresponding vector.  On some systems, due to hardware errata
223 	 * and interrupt sharing, the IPL may not correspond to the IPL listed
224 	 * in apic_vectortoipl (see apic_addspl and apic_delspl).
225 	 */
226 	for (i = 0; i < (APIC_AVAIL_VECTOR / APIC_VECTOR_PER_IPL); i++) {
227 		indx = i * APIC_VECTOR_PER_IPL;
228 
229 		for (j = 0; j < APIC_VECTOR_PER_IPL; j++, indx++)
230 			apic_ipls[indx] = apic_vectortoipl[i];
231 	}
232 
233 	/* cpu 0 is always up (for now) */
234 	apic_cpus[0].aci_status = APIC_CPU_ONLINE | APIC_CPU_INTR_ENABLE;
235 
236 	iptr = (int *)&apic_irq_table[0];
237 	for (i = 0; i <= APIC_MAX_VECTOR; i++) {
238 		apic_level_intr[i] = 0;
239 		*iptr++ = NULL;
240 		apic_vector_to_irq[i] = APIC_RESV_IRQ;
241 
242 		/* These *must* be initted to B_TRUE! */
243 		apic_reprogram_info[i].done = B_TRUE;
244 		apic_reprogram_info[i].irqp = NULL;
245 		apic_reprogram_info[i].tries = 0;
246 		apic_reprogram_info[i].bindcpu = 0;
247 	}
248 
249 	/*
250 	 * Allocate a dummy irq table entry for the reserved entry.
251 	 * This takes care of the race between removing an irq and
252 	 * clock detecting a CPU in that irq during interrupt load
253 	 * sampling.
254 	 */
255 	apic_irq_table[APIC_RESV_IRQ] =
256 	    kmem_zalloc(sizeof (apic_irq_t), KM_SLEEP);
257 
258 	mutex_init(&airq_mutex, NULL, MUTEX_DEFAULT, NULL);
259 }
260 
261 void
262 ioapic_init_intr(int mask_apic)
263 {
264 	int ioapic_ix;
265 	struct intrspec ispec;
266 	apic_irq_t *irqptr;
267 	int i, j;
268 	ulong_t iflag;
269 
270 	LOCK_INIT_CLEAR(&apic_revector_lock);
271 	LOCK_INIT_CLEAR(&apic_defer_reprogram_lock);
272 
273 	/* mask interrupt vectors */
274 	for (j = 0; j < apic_io_max && mask_apic; j++) {
275 		int intin_max;
276 
277 		ioapic_ix = j;
278 		/* Bits 23-16 define the maximum redirection entries */
279 		intin_max = (ioapic_read(ioapic_ix, APIC_VERS_CMD) >> 16)
280 		    & 0xff;
281 		for (i = 0; i <= intin_max; i++)
282 			ioapic_write(ioapic_ix, APIC_RDT_CMD + 2 * i, AV_MASK);
283 	}
284 
285 	/*
286 	 * Hack alert: deal with ACPI SCI interrupt chicken/egg here
287 	 */
288 	if (apic_sci_vect > 0) {
289 		/*
290 		 * acpica has already done add_avintr(); we just
291 		 * to finish the job by mimicing translate_irq()
292 		 *
293 		 * Fake up an intrspec and setup the tables
294 		 */
295 		ispec.intrspec_vec = apic_sci_vect;
296 		ispec.intrspec_pri = SCI_IPL;
297 
298 		if (apic_setup_irq_table(NULL, apic_sci_vect, NULL,
299 		    &ispec, &apic_sci_flags, DDI_INTR_TYPE_FIXED) < 0) {
300 			cmn_err(CE_WARN, "!apic: SCI setup failed");
301 			return;
302 		}
303 		irqptr = apic_irq_table[apic_sci_vect];
304 
305 		iflag = intr_clear();
306 		lock_set(&apic_ioapic_lock);
307 
308 		/* Program I/O APIC */
309 		(void) apic_setup_io_intr(irqptr, apic_sci_vect, B_FALSE);
310 
311 		lock_clear(&apic_ioapic_lock);
312 		intr_restore(iflag);
313 
314 		irqptr->airq_share++;
315 	}
316 
317 	/*
318 	 * Hack alert: deal with ACPI HPET interrupt chicken/egg here.
319 	 */
320 	if (apic_hpet_vect > 0) {
321 		/*
322 		 * hpet has already done add_avintr(); we just need
323 		 * to finish the job by mimicing translate_irq()
324 		 *
325 		 * Fake up an intrspec and setup the tables
326 		 */
327 		ispec.intrspec_vec = apic_hpet_vect;
328 		ispec.intrspec_pri = CBE_HIGH_PIL;
329 
330 		if (apic_setup_irq_table(NULL, apic_hpet_vect, NULL,
331 		    &ispec, &apic_hpet_flags, DDI_INTR_TYPE_FIXED) < 0) {
332 			cmn_err(CE_WARN, "!apic: HPET setup failed");
333 			return;
334 		}
335 		irqptr = apic_irq_table[apic_hpet_vect];
336 
337 		iflag = intr_clear();
338 		lock_set(&apic_ioapic_lock);
339 
340 		/* Program I/O APIC */
341 		(void) apic_setup_io_intr(irqptr, apic_hpet_vect, B_FALSE);
342 
343 		lock_clear(&apic_ioapic_lock);
344 		intr_restore(iflag);
345 
346 		irqptr->airq_share++;
347 	}
348 }
349 
350 /*
351  * Add mask bits to disable interrupt vector from happening
352  * at or above IPL. In addition, it should remove mask bits
353  * to enable interrupt vectors below the given IPL.
354  *
355  * Both add and delspl are complicated by the fact that different interrupts
356  * may share IRQs. This can happen in two ways.
357  * 1. The same H/W line is shared by more than 1 device
358  * 1a. with interrupts at different IPLs
359  * 1b. with interrupts at same IPL
360  * 2. We ran out of vectors at a given IPL and started sharing vectors.
361  * 1b and 2 should be handled gracefully, except for the fact some ISRs
362  * will get called often when no interrupt is pending for the device.
363  * For 1a, we handle it at the higher IPL.
364  */
365 /*ARGSUSED*/
366 int
367 apic_addspl_common(int irqno, int ipl, int min_ipl, int max_ipl)
368 {
369 	uchar_t vector;
370 	ulong_t iflag;
371 	apic_irq_t *irqptr, *irqheadptr;
372 	int irqindex;
373 
374 	ASSERT(max_ipl <= UCHAR_MAX);
375 	irqindex = IRQINDEX(irqno);
376 
377 	if ((irqindex == -1) || (!apic_irq_table[irqindex]))
378 		return (PSM_FAILURE);
379 
380 	mutex_enter(&airq_mutex);
381 	irqptr = irqheadptr = apic_irq_table[irqindex];
382 
383 	DDI_INTR_IMPLDBG((CE_CONT, "apic_addspl: dip=0x%p type=%d irqno=0x%x "
384 	    "vector=0x%x\n", (void *)irqptr->airq_dip,
385 	    irqptr->airq_mps_intr_index, irqno, irqptr->airq_vector));
386 
387 	while (irqptr) {
388 		if (VIRTIRQ(irqindex, irqptr->airq_share_id) == irqno)
389 			break;
390 		irqptr = irqptr->airq_next;
391 	}
392 	irqptr->airq_share++;
393 
394 	mutex_exit(&airq_mutex);
395 
396 	/* return if it is not hardware interrupt */
397 	if (irqptr->airq_mps_intr_index == RESERVE_INDEX)
398 		return (PSM_SUCCESS);
399 
400 	/* Or if there are more interupts at a higher IPL */
401 	if (ipl != max_ipl)
402 		return (PSM_SUCCESS);
403 
404 	/*
405 	 * if apic_picinit() has not been called yet, just return.
406 	 * At the end of apic_picinit(), we will call setup_io_intr().
407 	 */
408 
409 	if (!apic_picinit_called)
410 		return (PSM_SUCCESS);
411 
412 	/*
413 	 * Upgrade vector if max_ipl is not earlier ipl. If we cannot allocate,
414 	 * return failure.
415 	 */
416 	if (irqptr->airq_ipl != max_ipl &&
417 	    !ioapic_mask_workaround[irqptr->airq_ioapicindex]) {
418 
419 		vector = apic_allocate_vector(max_ipl, irqindex, 1);
420 		if (vector == 0) {
421 			irqptr->airq_share--;
422 			return (PSM_FAILURE);
423 		}
424 		irqptr = irqheadptr;
425 		apic_mark_vector(irqptr->airq_vector, vector);
426 		while (irqptr) {
427 			irqptr->airq_vector = vector;
428 			irqptr->airq_ipl = (uchar_t)max_ipl;
429 			/*
430 			 * reprogram irq being added and every one else
431 			 * who is not in the UNINIT state
432 			 */
433 			if ((VIRTIRQ(irqindex, irqptr->airq_share_id) ==
434 			    irqno) || (irqptr->airq_temp_cpu != IRQ_UNINIT)) {
435 				apic_record_rdt_entry(irqptr, irqindex);
436 
437 				iflag = intr_clear();
438 				lock_set(&apic_ioapic_lock);
439 
440 				(void) apic_setup_io_intr(irqptr, irqindex,
441 				    B_FALSE);
442 
443 				lock_clear(&apic_ioapic_lock);
444 				intr_restore(iflag);
445 			}
446 			irqptr = irqptr->airq_next;
447 		}
448 		return (PSM_SUCCESS);
449 
450 	} else if (irqptr->airq_ipl != max_ipl &&
451 	    ioapic_mask_workaround[irqptr->airq_ioapicindex]) {
452 		/*
453 		 * We cannot upgrade the vector, but we can change
454 		 * the IPL that this vector induces.
455 		 *
456 		 * Note that we subtract APIC_BASE_VECT from the vector
457 		 * here because this array is used in apic_intr_enter
458 		 * (no need to add APIC_BASE_VECT in that hot code
459 		 * path since we can do it in the rarely-executed path
460 		 * here).
461 		 */
462 		apic_ipls[irqptr->airq_vector - APIC_BASE_VECT] =
463 		    (uchar_t)max_ipl;
464 
465 		irqptr = irqheadptr;
466 		while (irqptr) {
467 			irqptr->airq_ipl = (uchar_t)max_ipl;
468 			irqptr = irqptr->airq_next;
469 		}
470 
471 		return (PSM_SUCCESS);
472 	}
473 
474 	ASSERT(irqptr);
475 
476 	iflag = intr_clear();
477 	lock_set(&apic_ioapic_lock);
478 
479 	(void) apic_setup_io_intr(irqptr, irqindex, B_FALSE);
480 
481 	lock_clear(&apic_ioapic_lock);
482 	intr_restore(iflag);
483 
484 	return (PSM_SUCCESS);
485 }
486 
487 /*
488  * Recompute mask bits for the given interrupt vector.
489  * If there is no interrupt servicing routine for this
490  * vector, this function should disable interrupt vector
491  * from happening at all IPLs. If there are still
492  * handlers using the given vector, this function should
493  * disable the given vector from happening below the lowest
494  * IPL of the remaining hadlers.
495  */
496 /*ARGSUSED*/
497 int
498 apic_delspl_common(int irqno, int ipl, int min_ipl, int max_ipl)
499 {
500 	uchar_t vector;
501 	uint32_t bind_cpu;
502 	int intin, irqindex;
503 	int ioapic_ix;
504 	apic_irq_t	*irqptr, *preirqptr, *irqheadptr, *irqp;
505 	ulong_t iflag;
506 
507 	mutex_enter(&airq_mutex);
508 	irqindex = IRQINDEX(irqno);
509 	irqptr = preirqptr = irqheadptr = apic_irq_table[irqindex];
510 
511 	DDI_INTR_IMPLDBG((CE_CONT, "apic_delspl: dip=0x%p type=%d irqno=0x%x "
512 	    "vector=0x%x\n", (void *)irqptr->airq_dip,
513 	    irqptr->airq_mps_intr_index, irqno, irqptr->airq_vector));
514 
515 	while (irqptr) {
516 		if (VIRTIRQ(irqindex, irqptr->airq_share_id) == irqno)
517 			break;
518 		preirqptr = irqptr;
519 		irqptr = irqptr->airq_next;
520 	}
521 	ASSERT(irqptr);
522 
523 	irqptr->airq_share--;
524 
525 	mutex_exit(&airq_mutex);
526 
527 	/*
528 	 * If there are more interrupts at a higher IPL, we don't need
529 	 * to disable anything.
530 	 */
531 	if (ipl < max_ipl)
532 		return (PSM_SUCCESS);
533 
534 	/* return if it is not hardware interrupt */
535 	if (irqptr->airq_mps_intr_index == RESERVE_INDEX)
536 		return (PSM_SUCCESS);
537 
538 	if (!apic_picinit_called) {
539 		/*
540 		 * Clear irq_struct. If two devices shared an intpt
541 		 * line & 1 unloaded before picinit, we are hosed. But, then
542 		 * we hope the machine survive.
543 		 */
544 		irqptr->airq_mps_intr_index = FREE_INDEX;
545 		irqptr->airq_temp_cpu = IRQ_UNINIT;
546 		apic_free_vector(irqptr->airq_vector);
547 		return (PSM_SUCCESS);
548 	}
549 	/*
550 	 * Downgrade vector to new max_ipl if needed. If we cannot allocate,
551 	 * use old IPL. Not very elegant, but it should work.
552 	 */
553 	if ((irqptr->airq_ipl != max_ipl) && (max_ipl != PSM_INVALID_IPL) &&
554 	    !ioapic_mask_workaround[irqptr->airq_ioapicindex]) {
555 		apic_irq_t	*irqp;
556 		if ((vector = apic_allocate_vector(max_ipl, irqno, 1))) {
557 			apic_mark_vector(irqheadptr->airq_vector, vector);
558 			irqp = irqheadptr;
559 			while (irqp) {
560 				irqp->airq_vector = vector;
561 				irqp->airq_ipl = (uchar_t)max_ipl;
562 				if (irqp->airq_temp_cpu != IRQ_UNINIT) {
563 					apic_record_rdt_entry(irqp, irqindex);
564 
565 					iflag = intr_clear();
566 					lock_set(&apic_ioapic_lock);
567 
568 					(void) apic_setup_io_intr(irqp,
569 					    irqindex, B_FALSE);
570 
571 					lock_clear(&apic_ioapic_lock);
572 					intr_restore(iflag);
573 				}
574 				irqp = irqp->airq_next;
575 			}
576 		}
577 
578 	} else if (irqptr->airq_ipl != max_ipl &&
579 	    max_ipl != PSM_INVALID_IPL &&
580 	    ioapic_mask_workaround[irqptr->airq_ioapicindex]) {
581 
582 	/*
583 	 * We cannot downgrade the IPL of the vector below the vector's
584 	 * hardware priority. If we did, it would be possible for a
585 	 * higher-priority hardware vector to interrupt a CPU running at an IPL
586 	 * lower than the hardware priority of the interrupting vector (but
587 	 * higher than the soft IPL of this IRQ). When this happens, we would
588 	 * then try to drop the IPL BELOW what it was (effectively dropping
589 	 * below base_spl) which would be potentially catastrophic.
590 	 *
591 	 * (e.g. Suppose the hardware vector associated with this IRQ is 0x40
592 	 * (hardware IPL of 4).  Further assume that the old IPL of this IRQ
593 	 * was 4, but the new IPL is 1.  If we forced vector 0x40 to result in
594 	 * an IPL of 1, it would be possible for the processor to be executing
595 	 * at IPL 3 and for an interrupt to come in on vector 0x40, interrupting
596 	 * the currently-executing ISR.  When apic_intr_enter consults
597 	 * apic_irqs[], it will return 1, bringing the IPL of the CPU down to 1
598 	 * so even though the processor was running at IPL 4, an IPL 1
599 	 * interrupt will have interrupted it, which must not happen)).
600 	 *
601 	 * Effectively, this means that the hardware priority corresponding to
602 	 * the IRQ's IPL (in apic_ipls[]) cannot be lower than the vector's
603 	 * hardware priority.
604 	 *
605 	 * (In the above example, then, after removal of the IPL 4 device's
606 	 * interrupt handler, the new IPL will continue to be 4 because the
607 	 * hardware priority that IPL 1 implies is lower than the hardware
608 	 * priority of the vector used.)
609 	 */
610 		/* apic_ipls is indexed by vector, starting at APIC_BASE_VECT */
611 		const int apic_ipls_index = irqptr->airq_vector -
612 		    APIC_BASE_VECT;
613 		const int vect_inherent_hwpri = irqptr->airq_vector >>
614 		    APIC_IPL_SHIFT;
615 
616 		/*
617 		 * If there are still devices using this IRQ, determine the
618 		 * new ipl to use.
619 		 */
620 		if (irqptr->airq_share) {
621 			int vect_desired_hwpri, hwpri;
622 
623 			ASSERT(max_ipl < MAXIPL);
624 			vect_desired_hwpri = apic_ipltopri[max_ipl] >>
625 			    APIC_IPL_SHIFT;
626 
627 			/*
628 			 * If the desired IPL's hardware priority is lower
629 			 * than that of the vector, use the hardware priority
630 			 * of the vector to determine the new IPL.
631 			 */
632 			hwpri = (vect_desired_hwpri < vect_inherent_hwpri) ?
633 			    vect_inherent_hwpri : vect_desired_hwpri;
634 
635 			/*
636 			 * Now, to get the right index for apic_vectortoipl,
637 			 * we need to subtract APIC_BASE_VECT from the
638 			 * hardware-vector-equivalent (in hwpri).  Since hwpri
639 			 * is already shifted, we shift APIC_BASE_VECT before
640 			 * doing the subtraction.
641 			 */
642 			hwpri -= (APIC_BASE_VECT >> APIC_IPL_SHIFT);
643 
644 			ASSERT(hwpri >= 0);
645 			ASSERT(hwpri < MAXIPL);
646 			max_ipl = apic_vectortoipl[hwpri];
647 			apic_ipls[apic_ipls_index] = max_ipl;
648 
649 			irqp = irqheadptr;
650 			while (irqp) {
651 				irqp->airq_ipl = (uchar_t)max_ipl;
652 				irqp = irqp->airq_next;
653 			}
654 		} else {
655 			/*
656 			 * No more devices on this IRQ, so reset this vector's
657 			 * element in apic_ipls to the original IPL for this
658 			 * vector
659 			 */
660 			apic_ipls[apic_ipls_index] =
661 			    apic_vectortoipl[vect_inherent_hwpri];
662 		}
663 	}
664 
665 	/*
666 	 * If there are still active interrupts, we are done.
667 	 */
668 	if (irqptr->airq_share)
669 		return (PSM_SUCCESS);
670 
671 	iflag = intr_clear();
672 	lock_set(&apic_ioapic_lock);
673 
674 	if (irqptr->airq_mps_intr_index == MSI_INDEX) {
675 		/*
676 		 * Disable the MSI vector
677 		 * Make sure we only disable on the last
678 		 * of the multi-MSI support
679 		 */
680 		if (i_ddi_intr_get_current_nenables(irqptr->airq_dip) == 1) {
681 			apic_pci_msi_disable_mode(irqptr->airq_dip,
682 			    DDI_INTR_TYPE_MSI);
683 		}
684 	} else if (irqptr->airq_mps_intr_index == MSIX_INDEX) {
685 		/*
686 		 * Disable the MSI-X vector
687 		 * needs to clear its mask and addr/data for each MSI-X
688 		 */
689 		apic_pci_msi_unconfigure(irqptr->airq_dip, DDI_INTR_TYPE_MSIX,
690 		    irqptr->airq_origirq);
691 		/*
692 		 * Make sure we only disable on the last MSI-X
693 		 */
694 		if (i_ddi_intr_get_current_nenables(irqptr->airq_dip) == 1) {
695 			apic_pci_msi_disable_mode(irqptr->airq_dip,
696 			    DDI_INTR_TYPE_MSIX);
697 		}
698 	} else {
699 		/*
700 		 * The assumption here is that this is safe, even for
701 		 * systems with IOAPICs that suffer from the hardware
702 		 * erratum because all devices have been quiesced before
703 		 * they unregister their interrupt handlers.  If that
704 		 * assumption turns out to be false, this mask operation
705 		 * can induce the same erratum result we're trying to
706 		 * avoid.
707 		 */
708 		ioapic_ix = irqptr->airq_ioapicindex;
709 		intin = irqptr->airq_intin_no;
710 		ioapic_write(ioapic_ix, APIC_RDT_CMD + 2 * intin, AV_MASK);
711 	}
712 
713 	apic_vt_ops->apic_intrmap_free_entry(&irqptr->airq_intrmap_private);
714 
715 	/*
716 	 * This irq entry is the only one in the chain.
717 	 */
718 	if (irqheadptr->airq_next == NULL) {
719 		ASSERT(irqheadptr == irqptr);
720 		bind_cpu = irqptr->airq_temp_cpu;
721 		if (((uint32_t)bind_cpu != IRQ_UNBOUND) &&
722 		    ((uint32_t)bind_cpu != IRQ_UNINIT)) {
723 			ASSERT(apic_cpu_in_range(bind_cpu));
724 			if (bind_cpu & IRQ_USER_BOUND) {
725 				/* If hardbound, temp_cpu == cpu */
726 				bind_cpu &= ~IRQ_USER_BOUND;
727 				apic_cpus[bind_cpu].aci_bound--;
728 			} else
729 				apic_cpus[bind_cpu].aci_temp_bound--;
730 		}
731 		irqptr->airq_temp_cpu = IRQ_UNINIT;
732 		irqptr->airq_mps_intr_index = FREE_INDEX;
733 		lock_clear(&apic_ioapic_lock);
734 		intr_restore(iflag);
735 		apic_free_vector(irqptr->airq_vector);
736 		return (PSM_SUCCESS);
737 	}
738 
739 	/*
740 	 * If we get here, we are sharing the vector and there are more than
741 	 * one active irq entries in the chain.
742 	 */
743 	lock_clear(&apic_ioapic_lock);
744 	intr_restore(iflag);
745 
746 	mutex_enter(&airq_mutex);
747 	/* Remove the irq entry from the chain */
748 	if (irqptr == irqheadptr) { /* The irq entry is at the head */
749 		apic_irq_table[irqindex] = irqptr->airq_next;
750 	} else {
751 		preirqptr->airq_next = irqptr->airq_next;
752 	}
753 	/* Free the irq entry */
754 	kmem_free(irqptr, sizeof (apic_irq_t));
755 	mutex_exit(&airq_mutex);
756 
757 	return (PSM_SUCCESS);
758 }
759 
760 /*
761  * apic_introp_xlate() replaces apic_translate_irq() and is
762  * called only from apic_intr_ops().  With the new ADII framework,
763  * the priority can no longer be retrieved through i_ddi_get_intrspec().
764  * It has to be passed in from the caller.
765  *
766  * Return value:
767  *      Success: irqno for the given device
768  *      Failure: -1
769  */
770 int
771 apic_introp_xlate(dev_info_t *dip, struct intrspec *ispec, int type)
772 {
773 	char dev_type[16];
774 	int dev_len, pci_irq, newirq, bustype, devid, busid, i;
775 	int irqno = ispec->intrspec_vec;
776 	ddi_acc_handle_t cfg_handle;
777 	uchar_t ipin;
778 	struct apic_io_intr *intrp;
779 	iflag_t intr_flag;
780 	ACPI_SUBTABLE_HEADER	*hp;
781 	ACPI_MADT_INTERRUPT_OVERRIDE *isop;
782 	apic_irq_t *airqp;
783 	int parent_is_pci_or_pciex = 0;
784 	int child_is_pciex = 0;
785 
786 	DDI_INTR_IMPLDBG((CE_CONT, "apic_introp_xlate: dip=0x%p name=%s "
787 	    "type=%d irqno=0x%x\n", (void *)dip, ddi_get_name(dip), type,
788 	    irqno));
789 
790 	dev_len = sizeof (dev_type);
791 	if (ddi_getlongprop_buf(DDI_DEV_T_ANY, ddi_get_parent(dip),
792 	    DDI_PROP_DONTPASS, "device_type", (caddr_t)dev_type,
793 	    &dev_len) == DDI_PROP_SUCCESS) {
794 		if ((strcmp(dev_type, "pci") == 0) ||
795 		    (strcmp(dev_type, "pciex") == 0))
796 			parent_is_pci_or_pciex = 1;
797 	}
798 
799 	if (ddi_getlongprop_buf(DDI_DEV_T_ANY, dip,
800 	    DDI_PROP_DONTPASS, "compatible", (caddr_t)dev_type,
801 	    &dev_len) == DDI_PROP_SUCCESS) {
802 		if (strstr(dev_type, "pciex"))
803 			child_is_pciex = 1;
804 	}
805 
806 	if (DDI_INTR_IS_MSI_OR_MSIX(type)) {
807 		if ((airqp = apic_find_irq(dip, ispec, type)) != NULL) {
808 			airqp->airq_iflag.bustype =
809 			    child_is_pciex ? BUS_PCIE : BUS_PCI;
810 			return (apic_vector_to_irq[airqp->airq_vector]);
811 		}
812 		return (apic_setup_irq_table(dip, irqno, NULL, ispec,
813 		    NULL, type));
814 	}
815 
816 	bustype = 0;
817 
818 	/* check if we have already translated this irq */
819 	mutex_enter(&airq_mutex);
820 	newirq = apic_min_device_irq;
821 	for (; newirq <= apic_max_device_irq; newirq++) {
822 		airqp = apic_irq_table[newirq];
823 		while (airqp) {
824 			if ((airqp->airq_dip == dip) &&
825 			    (airqp->airq_origirq == irqno) &&
826 			    (airqp->airq_mps_intr_index != FREE_INDEX)) {
827 
828 				mutex_exit(&airq_mutex);
829 				return (VIRTIRQ(newirq, airqp->airq_share_id));
830 			}
831 			airqp = airqp->airq_next;
832 		}
833 	}
834 	mutex_exit(&airq_mutex);
835 
836 	if (apic_defconf)
837 		goto defconf;
838 
839 	if ((dip == NULL) || (!apic_irq_translate && !apic_enable_acpi))
840 		goto nonpci;
841 
842 	if (parent_is_pci_or_pciex) {
843 		/* pci device */
844 		if (acpica_get_bdf(dip, &busid, &devid, NULL) != 0)
845 			goto nonpci;
846 		if (busid == 0 && apic_pci_bus_total == 1)
847 			busid = (int)apic_single_pci_busid;
848 
849 		if (pci_config_setup(dip, &cfg_handle) != DDI_SUCCESS)
850 			return (-1);
851 		ipin = pci_config_get8(cfg_handle, PCI_CONF_IPIN) - PCI_INTA;
852 		pci_config_teardown(&cfg_handle);
853 		if (apic_enable_acpi && !apic_use_acpi_madt_only) {
854 			if (apic_acpi_translate_pci_irq(dip, busid, devid,
855 			    ipin, &pci_irq, &intr_flag) != ACPI_PSM_SUCCESS)
856 				return (-1);
857 
858 			intr_flag.bustype = child_is_pciex ? BUS_PCIE : BUS_PCI;
859 			return (apic_setup_irq_table(dip, pci_irq, NULL, ispec,
860 			    &intr_flag, type));
861 		} else {
862 			pci_irq = ((devid & 0x1f) << 2) | (ipin & 0x3);
863 			if ((intrp = apic_find_io_intr_w_busid(pci_irq, busid))
864 			    == NULL) {
865 				if ((pci_irq = apic_handle_pci_pci_bridge(dip,
866 				    devid, ipin, &intrp)) == -1)
867 					return (-1);
868 			}
869 			return (apic_setup_irq_table(dip, pci_irq, intrp, ispec,
870 			    NULL, type));
871 		}
872 	} else if (strcmp(dev_type, "isa") == 0)
873 		bustype = BUS_ISA;
874 	else if (strcmp(dev_type, "eisa") == 0)
875 		bustype = BUS_EISA;
876 
877 nonpci:
878 	if (apic_enable_acpi && !apic_use_acpi_madt_only) {
879 		/* search iso entries first */
880 		if (acpi_iso_cnt != 0) {
881 			hp = (ACPI_SUBTABLE_HEADER *)acpi_isop;
882 			i = 0;
883 			while (i < acpi_iso_cnt) {
884 				if (hp->Type ==
885 				    ACPI_MADT_TYPE_INTERRUPT_OVERRIDE) {
886 					isop =
887 					    (ACPI_MADT_INTERRUPT_OVERRIDE *) hp;
888 					if (isop->Bus == 0 &&
889 					    isop->SourceIrq == irqno) {
890 						newirq = isop->GlobalIrq;
891 						intr_flag.intr_po =
892 						    isop->IntiFlags &
893 						    ACPI_MADT_POLARITY_MASK;
894 						intr_flag.intr_el =
895 						    (isop->IntiFlags &
896 						    ACPI_MADT_TRIGGER_MASK)
897 						    >> 2;
898 						intr_flag.bustype = BUS_ISA;
899 
900 						return (apic_setup_irq_table(
901 						    dip, newirq, NULL, ispec,
902 						    &intr_flag, type));
903 
904 					}
905 					i++;
906 				}
907 				hp = (ACPI_SUBTABLE_HEADER *)(((char *)hp) +
908 				    hp->Length);
909 			}
910 		}
911 		intr_flag.intr_po = INTR_PO_ACTIVE_HIGH;
912 		intr_flag.intr_el = INTR_EL_EDGE;
913 		intr_flag.bustype = BUS_ISA;
914 		return (apic_setup_irq_table(dip, irqno, NULL, ispec,
915 		    &intr_flag, type));
916 	} else {
917 		if (bustype == 0)	/* not initialized */
918 			bustype = eisa_level_intr_mask ? BUS_EISA : BUS_ISA;
919 		for (i = 0; i < 2; i++) {
920 			if (((busid = apic_find_bus_id(bustype)) != -1) &&
921 			    ((intrp = apic_find_io_intr_w_busid(irqno, busid))
922 			    != NULL)) {
923 				if ((newirq = apic_setup_irq_table(dip, irqno,
924 				    intrp, ispec, NULL, type)) != -1) {
925 					return (newirq);
926 				}
927 				goto defconf;
928 			}
929 			bustype = (bustype == BUS_EISA) ? BUS_ISA : BUS_EISA;
930 		}
931 	}
932 
933 /* MPS default configuration */
934 defconf:
935 	newirq = apic_setup_irq_table(dip, irqno, NULL, ispec, NULL, type);
936 	if (newirq == -1)
937 		return (-1);
938 	ASSERT(IRQINDEX(newirq) == irqno);
939 	ASSERT(apic_irq_table[irqno]);
940 	return (newirq);
941 }
942 
943 /*
944  * Attempt to share vector with someone else
945  */
946 static int
947 apic_share_vector(int irqno, iflag_t *intr_flagp, short intr_index, int ipl,
948 	uchar_t ioapicindex, uchar_t ipin, apic_irq_t **irqptrp)
949 {
950 #ifdef DEBUG
951 	apic_irq_t *tmpirqp = NULL;
952 #endif /* DEBUG */
953 	apic_irq_t *irqptr, dummyirq;
954 	int	newirq, chosen_irq = -1, share = 127;
955 	int	lowest, highest, i;
956 	uchar_t	share_id;
957 
958 	DDI_INTR_IMPLDBG((CE_CONT, "apic_share_vector: irqno=0x%x "
959 	    "intr_index=0x%x ipl=0x%x\n", irqno, intr_index, ipl));
960 
961 	highest = apic_ipltopri[ipl] + APIC_VECTOR_MASK;
962 	lowest = apic_ipltopri[ipl-1] + APIC_VECTOR_PER_IPL;
963 
964 	if (highest < lowest) /* Both ipl and ipl-1 map to same pri */
965 		lowest -= APIC_VECTOR_PER_IPL;
966 	dummyirq.airq_mps_intr_index = intr_index;
967 	dummyirq.airq_ioapicindex = ioapicindex;
968 	dummyirq.airq_intin_no = ipin;
969 	if (intr_flagp)
970 		dummyirq.airq_iflag = *intr_flagp;
971 	apic_record_rdt_entry(&dummyirq, irqno);
972 	for (i = lowest; i <= highest; i++) {
973 		newirq = apic_vector_to_irq[i];
974 		if (newirq == APIC_RESV_IRQ)
975 			continue;
976 		irqptr = apic_irq_table[newirq];
977 
978 		if ((dummyirq.airq_rdt_entry & 0xFF00) !=
979 		    (irqptr->airq_rdt_entry & 0xFF00))
980 			/* not compatible */
981 			continue;
982 
983 		if (irqptr->airq_share < share) {
984 			share = irqptr->airq_share;
985 			chosen_irq = newirq;
986 		}
987 	}
988 	if (chosen_irq != -1) {
989 		/*
990 		 * Assign a share id which is free or which is larger
991 		 * than the largest one.
992 		 */
993 		share_id = 1;
994 		mutex_enter(&airq_mutex);
995 		irqptr = apic_irq_table[chosen_irq];
996 		while (irqptr) {
997 			if (irqptr->airq_mps_intr_index == FREE_INDEX) {
998 				share_id = irqptr->airq_share_id;
999 				break;
1000 			}
1001 			if (share_id <= irqptr->airq_share_id)
1002 				share_id = irqptr->airq_share_id + 1;
1003 #ifdef DEBUG
1004 			tmpirqp = irqptr;
1005 #endif /* DEBUG */
1006 			irqptr = irqptr->airq_next;
1007 		}
1008 		if (!irqptr) {
1009 			irqptr = kmem_zalloc(sizeof (apic_irq_t), KM_SLEEP);
1010 			irqptr->airq_temp_cpu = IRQ_UNINIT;
1011 			irqptr->airq_next =
1012 			    apic_irq_table[chosen_irq]->airq_next;
1013 			apic_irq_table[chosen_irq]->airq_next = irqptr;
1014 #ifdef	DEBUG
1015 			tmpirqp = apic_irq_table[chosen_irq];
1016 #endif /* DEBUG */
1017 		}
1018 		irqptr->airq_mps_intr_index = intr_index;
1019 		irqptr->airq_ioapicindex = ioapicindex;
1020 		irqptr->airq_intin_no = ipin;
1021 		if (intr_flagp)
1022 			irqptr->airq_iflag = *intr_flagp;
1023 		irqptr->airq_vector = apic_irq_table[chosen_irq]->airq_vector;
1024 		irqptr->airq_share_id = share_id;
1025 		apic_record_rdt_entry(irqptr, irqno);
1026 		*irqptrp = irqptr;
1027 #ifdef	DEBUG
1028 		/* shuffle the pointers to test apic_delspl path */
1029 		if (tmpirqp) {
1030 			tmpirqp->airq_next = irqptr->airq_next;
1031 			irqptr->airq_next = apic_irq_table[chosen_irq];
1032 			apic_irq_table[chosen_irq] = irqptr;
1033 		}
1034 #endif /* DEBUG */
1035 		mutex_exit(&airq_mutex);
1036 		return (VIRTIRQ(chosen_irq, share_id));
1037 	}
1038 	return (-1);
1039 }
1040 
1041 /*
1042  * Allocate/Initialize the apic_irq_table[] entry for given irqno. If the entry
1043  * is used already, we will try to allocate a new irqno.
1044  *
1045  * Return value:
1046  *	Success: irqno
1047  *	Failure: -1
1048  */
1049 static int
1050 apic_setup_irq_table(dev_info_t *dip, int irqno, struct apic_io_intr *intrp,
1051     struct intrspec *ispec, iflag_t *intr_flagp, int type)
1052 {
1053 	int origirq = ispec->intrspec_vec;
1054 	uchar_t ipl = ispec->intrspec_pri;
1055 	int	newirq, intr_index;
1056 	uchar_t	ipin, ioapic, ioapicindex, vector;
1057 	apic_irq_t *irqptr;
1058 	major_t	major;
1059 	dev_info_t	*sdip;
1060 
1061 	DDI_INTR_IMPLDBG((CE_CONT, "apic_setup_irq_table: dip=0x%p type=%d "
1062 	    "irqno=0x%x origirq=0x%x\n", (void *)dip, type, irqno, origirq));
1063 
1064 	ASSERT(ispec != NULL);
1065 
1066 	major =  (dip != NULL) ? ddi_driver_major(dip) : 0;
1067 
1068 	if (DDI_INTR_IS_MSI_OR_MSIX(type)) {
1069 		/* MSI/X doesn't need to setup ioapic stuffs */
1070 		ioapicindex = 0xff;
1071 		ioapic = 0xff;
1072 		ipin = (uchar_t)0xff;
1073 		intr_index = (type == DDI_INTR_TYPE_MSI) ? MSI_INDEX :
1074 		    MSIX_INDEX;
1075 		mutex_enter(&airq_mutex);
1076 		if ((irqno = apic_allocate_irq(apic_first_avail_irq)) == -1) {
1077 			mutex_exit(&airq_mutex);
1078 			/* need an irq for MSI/X to index into autovect[] */
1079 			cmn_err(CE_WARN, "No interrupt irq: %s instance %d",
1080 			    ddi_get_name(dip), ddi_get_instance(dip));
1081 			return (-1);
1082 		}
1083 		mutex_exit(&airq_mutex);
1084 
1085 	} else if (intrp != NULL) {
1086 		intr_index = (int)(intrp - apic_io_intrp);
1087 		ioapic = intrp->intr_destid;
1088 		ipin = intrp->intr_destintin;
1089 		/* Find ioapicindex. If destid was ALL, we will exit with 0. */
1090 		for (ioapicindex = apic_io_max - 1; ioapicindex; ioapicindex--)
1091 			if (apic_io_id[ioapicindex] == ioapic)
1092 				break;
1093 		ASSERT((ioapic == apic_io_id[ioapicindex]) ||
1094 		    (ioapic == INTR_ALL_APIC));
1095 
1096 		/* check whether this intin# has been used by another irqno */
1097 		if ((newirq = apic_find_intin(ioapicindex, ipin)) != -1) {
1098 			return (newirq);
1099 		}
1100 
1101 	} else if (intr_flagp != NULL) {
1102 		/* ACPI case */
1103 		intr_index = ACPI_INDEX;
1104 		ioapicindex = acpi_find_ioapic(irqno);
1105 		ASSERT(ioapicindex != 0xFF);
1106 		ioapic = apic_io_id[ioapicindex];
1107 		ipin = irqno - apic_io_vectbase[ioapicindex];
1108 		if (apic_irq_table[irqno] &&
1109 		    apic_irq_table[irqno]->airq_mps_intr_index == ACPI_INDEX) {
1110 			ASSERT(apic_irq_table[irqno]->airq_intin_no == ipin &&
1111 			    apic_irq_table[irqno]->airq_ioapicindex ==
1112 			    ioapicindex);
1113 			return (irqno);
1114 		}
1115 
1116 	} else {
1117 		/* default configuration */
1118 		ioapicindex = 0;
1119 		ioapic = apic_io_id[ioapicindex];
1120 		ipin = (uchar_t)irqno;
1121 		intr_index = DEFAULT_INDEX;
1122 	}
1123 
1124 	if (ispec == NULL) {
1125 		APIC_VERBOSE_IOAPIC((CE_WARN, "No intrspec for irqno = %x\n",
1126 		    irqno));
1127 	} else if ((vector = apic_allocate_vector(ipl, irqno, 0)) == 0) {
1128 		if ((newirq = apic_share_vector(irqno, intr_flagp, intr_index,
1129 		    ipl, ioapicindex, ipin, &irqptr)) != -1) {
1130 			irqptr->airq_ipl = ipl;
1131 			irqptr->airq_origirq = (uchar_t)origirq;
1132 			irqptr->airq_dip = dip;
1133 			irqptr->airq_major = major;
1134 			sdip = apic_irq_table[IRQINDEX(newirq)]->airq_dip;
1135 			/* This is OK to do really */
1136 			if (sdip == NULL) {
1137 				cmn_err(CE_WARN, "Sharing vectors: %s"
1138 				    " instance %d and SCI",
1139 				    ddi_get_name(dip), ddi_get_instance(dip));
1140 			} else {
1141 				cmn_err(CE_WARN, "Sharing vectors: %s"
1142 				    " instance %d and %s instance %d",
1143 				    ddi_get_name(sdip), ddi_get_instance(sdip),
1144 				    ddi_get_name(dip), ddi_get_instance(dip));
1145 			}
1146 			return (newirq);
1147 		}
1148 		/* try high priority allocation now  that share has failed */
1149 		if ((vector = apic_allocate_vector(ipl, irqno, 1)) == 0) {
1150 			cmn_err(CE_WARN, "No interrupt vector: %s instance %d",
1151 			    ddi_get_name(dip), ddi_get_instance(dip));
1152 			return (-1);
1153 		}
1154 	}
1155 
1156 	mutex_enter(&airq_mutex);
1157 	if (apic_irq_table[irqno] == NULL) {
1158 		irqptr = kmem_zalloc(sizeof (apic_irq_t), KM_SLEEP);
1159 		irqptr->airq_temp_cpu = IRQ_UNINIT;
1160 		apic_irq_table[irqno] = irqptr;
1161 	} else {
1162 		irqptr = apic_irq_table[irqno];
1163 		if (irqptr->airq_mps_intr_index != FREE_INDEX) {
1164 			/*
1165 			 * The slot is used by another irqno, so allocate
1166 			 * a free irqno for this interrupt
1167 			 */
1168 			newirq = apic_allocate_irq(apic_first_avail_irq);
1169 			if (newirq == -1) {
1170 				mutex_exit(&airq_mutex);
1171 				return (-1);
1172 			}
1173 			irqno = newirq;
1174 			irqptr = apic_irq_table[irqno];
1175 			if (irqptr == NULL) {
1176 				irqptr = kmem_zalloc(sizeof (apic_irq_t),
1177 				    KM_SLEEP);
1178 				irqptr->airq_temp_cpu = IRQ_UNINIT;
1179 				apic_irq_table[irqno] = irqptr;
1180 			}
1181 			vector = apic_modify_vector(vector, newirq);
1182 		}
1183 	}
1184 	apic_max_device_irq = max(irqno, apic_max_device_irq);
1185 	apic_min_device_irq = min(irqno, apic_min_device_irq);
1186 	mutex_exit(&airq_mutex);
1187 	irqptr->airq_ioapicindex = ioapicindex;
1188 	irqptr->airq_intin_no = ipin;
1189 	irqptr->airq_ipl = ipl;
1190 	irqptr->airq_vector = vector;
1191 	irqptr->airq_origirq = (uchar_t)origirq;
1192 	irqptr->airq_share_id = 0;
1193 	irqptr->airq_mps_intr_index = (short)intr_index;
1194 	irqptr->airq_dip = dip;
1195 	irqptr->airq_major = major;
1196 	irqptr->airq_cpu = apic_bind_intr(dip, irqno, ioapic, ipin);
1197 	if (intr_flagp)
1198 		irqptr->airq_iflag = *intr_flagp;
1199 
1200 	if (!DDI_INTR_IS_MSI_OR_MSIX(type)) {
1201 		/* setup I/O APIC entry for non-MSI/X interrupts */
1202 		apic_record_rdt_entry(irqptr, irqno);
1203 	}
1204 	return (irqno);
1205 }
1206 
1207 /*
1208  * return the cpu to which this intr should be bound.
1209  * Check properties or any other mechanism to see if user wants it
1210  * bound to a specific CPU. If so, return the cpu id with high bit set.
1211  * If not, use the policy to choose a cpu and return the id.
1212  */
1213 uint32_t
1214 apic_bind_intr(dev_info_t *dip, int irq, uchar_t ioapicid, uchar_t intin)
1215 {
1216 	int	instance, instno, prop_len, bind_cpu, count;
1217 	uint_t	i, rc;
1218 	uint32_t cpu;
1219 	major_t	major;
1220 	char	*name, *drv_name, *prop_val, *cptr;
1221 	char	prop_name[32];
1222 	ulong_t iflag;
1223 
1224 
1225 	if (apic_intr_policy == INTR_LOWEST_PRIORITY)
1226 		return (IRQ_UNBOUND);
1227 
1228 	if (apic_nproc == 1)
1229 		return (0);
1230 
1231 	drv_name = NULL;
1232 	rc = DDI_PROP_NOT_FOUND;
1233 	major = (major_t)-1;
1234 	if (dip != NULL) {
1235 		name = ddi_get_name(dip);
1236 		major = ddi_name_to_major(name);
1237 		drv_name = ddi_major_to_name(major);
1238 		instance = ddi_get_instance(dip);
1239 		if (apic_intr_policy == INTR_ROUND_ROBIN_WITH_AFFINITY) {
1240 			i = apic_min_device_irq;
1241 			for (; i <= apic_max_device_irq; i++) {
1242 
1243 				if ((i == irq) || (apic_irq_table[i] == NULL) ||
1244 				    (apic_irq_table[i]->airq_mps_intr_index
1245 				    == FREE_INDEX))
1246 					continue;
1247 
1248 				if ((apic_irq_table[i]->airq_major == major) &&
1249 				    (!(apic_irq_table[i]->airq_cpu &
1250 				    IRQ_USER_BOUND))) {
1251 
1252 					cpu = apic_irq_table[i]->airq_cpu;
1253 
1254 					cmn_err(CE_CONT,
1255 					    "!%s: %s (%s) instance #%d "
1256 					    "irq 0x%x vector 0x%x ioapic 0x%x "
1257 					    "intin 0x%x is bound to cpu %d\n",
1258 					    psm_name,
1259 					    name, drv_name, instance, irq,
1260 					    apic_irq_table[irq]->airq_vector,
1261 					    ioapicid, intin, cpu);
1262 					return (cpu);
1263 				}
1264 			}
1265 		}
1266 		/*
1267 		 * search for "drvname"_intpt_bind_cpus property first, the
1268 		 * syntax of the property should be "a[,b,c,...]" where
1269 		 * instance 0 binds to cpu a, instance 1 binds to cpu b,
1270 		 * instance 3 binds to cpu c...
1271 		 * ddi_getlongprop() will search /option first, then /
1272 		 * if "drvname"_intpt_bind_cpus doesn't exist, then find
1273 		 * intpt_bind_cpus property.  The syntax is the same, and
1274 		 * it applies to all the devices if its "drvname" specific
1275 		 * property doesn't exist
1276 		 */
1277 		(void) strcpy(prop_name, drv_name);
1278 		(void) strcat(prop_name, "_intpt_bind_cpus");
1279 		rc = ddi_getlongprop(DDI_DEV_T_ANY, dip, 0, prop_name,
1280 		    (caddr_t)&prop_val, &prop_len);
1281 		if (rc != DDI_PROP_SUCCESS) {
1282 			rc = ddi_getlongprop(DDI_DEV_T_ANY, dip, 0,
1283 			    "intpt_bind_cpus", (caddr_t)&prop_val, &prop_len);
1284 		}
1285 	}
1286 	if (rc == DDI_PROP_SUCCESS) {
1287 		for (i = count = 0; i < (prop_len - 1); i++)
1288 			if (prop_val[i] == ',')
1289 				count++;
1290 		if (prop_val[i-1] != ',')
1291 			count++;
1292 		/*
1293 		 * if somehow the binding instances defined in the
1294 		 * property are not enough for this instno., then
1295 		 * reuse the pattern for the next instance until
1296 		 * it reaches the requested instno
1297 		 */
1298 		instno = instance % count;
1299 		i = 0;
1300 		cptr = prop_val;
1301 		while (i < instno)
1302 			if (*cptr++ == ',')
1303 				i++;
1304 		bind_cpu = stoi(&cptr);
1305 		kmem_free(prop_val, prop_len);
1306 		/* if specific CPU is bogus, then default to next cpu */
1307 		if (!apic_cpu_in_range(bind_cpu)) {
1308 			cmn_err(CE_WARN, "%s: %s=%s: CPU %d not present",
1309 			    psm_name, prop_name, prop_val, bind_cpu);
1310 			rc = DDI_PROP_NOT_FOUND;
1311 		} else {
1312 			/* indicate that we are bound at user request */
1313 			bind_cpu |= IRQ_USER_BOUND;
1314 		}
1315 		/*
1316 		 * no need to check apic_cpus[].aci_status, if specific CPU is
1317 		 * not up, then post_cpu_start will handle it.
1318 		 */
1319 	}
1320 	if (rc != DDI_PROP_SUCCESS) {
1321 		iflag = intr_clear();
1322 		lock_set(&apic_ioapic_lock);
1323 		bind_cpu = apic_get_next_bind_cpu();
1324 		lock_clear(&apic_ioapic_lock);
1325 		intr_restore(iflag);
1326 	}
1327 
1328 	if (drv_name != NULL)
1329 		cmn_err(CE_CONT, "!%s: %s (%s) instance %d irq 0x%x "
1330 		    "vector 0x%x ioapic 0x%x intin 0x%x is bound to cpu %d\n",
1331 		    psm_name, name, drv_name, instance, irq,
1332 		    apic_irq_table[irq]->airq_vector, ioapicid, intin,
1333 		    bind_cpu & ~IRQ_USER_BOUND);
1334 	else
1335 		cmn_err(CE_CONT, "!%s: irq 0x%x "
1336 		    "vector 0x%x ioapic 0x%x intin 0x%x is bound to cpu %d\n",
1337 		    psm_name, irq, apic_irq_table[irq]->airq_vector, ioapicid,
1338 		    intin, bind_cpu & ~IRQ_USER_BOUND);
1339 
1340 	return ((uint32_t)bind_cpu);
1341 }
1342 
1343 /*
1344  * Mark vector as being in the process of being deleted. Interrupts
1345  * may still come in on some CPU. The moment an interrupt comes with
1346  * the new vector, we know we can free the old one. Called only from
1347  * addspl and delspl with interrupts disabled. Because an interrupt
1348  * can be shared, but no interrupt from either device may come in,
1349  * we also use a timeout mechanism, which we arbitrarily set to
1350  * apic_revector_timeout microseconds.
1351  */
1352 static void
1353 apic_mark_vector(uchar_t oldvector, uchar_t newvector)
1354 {
1355 	ulong_t iflag;
1356 
1357 	iflag = intr_clear();
1358 	lock_set(&apic_revector_lock);
1359 	if (!apic_oldvec_to_newvec) {
1360 		apic_oldvec_to_newvec =
1361 		    kmem_zalloc(sizeof (newvector) * APIC_MAX_VECTOR * 2,
1362 		    KM_NOSLEEP);
1363 
1364 		if (!apic_oldvec_to_newvec) {
1365 			/*
1366 			 * This failure is not catastrophic.
1367 			 * But, the oldvec will never be freed.
1368 			 */
1369 			apic_error |= APIC_ERR_MARK_VECTOR_FAIL;
1370 			lock_clear(&apic_revector_lock);
1371 			intr_restore(iflag);
1372 			return;
1373 		}
1374 		apic_newvec_to_oldvec = &apic_oldvec_to_newvec[APIC_MAX_VECTOR];
1375 	}
1376 
1377 	/* See if we already did this for drivers which do double addintrs */
1378 	if (apic_oldvec_to_newvec[oldvector] != newvector) {
1379 		apic_oldvec_to_newvec[oldvector] = newvector;
1380 		apic_newvec_to_oldvec[newvector] = oldvector;
1381 		apic_revector_pending++;
1382 	}
1383 	lock_clear(&apic_revector_lock);
1384 	intr_restore(iflag);
1385 	(void) timeout(apic_xlate_vector_free_timeout_handler,
1386 	    (void *)(uintptr_t)oldvector, drv_usectohz(apic_revector_timeout));
1387 }
1388 
1389 /*
1390  * xlate_vector is called from intr_enter if revector_pending is set.
1391  * It will xlate it if needed and mark the old vector as free.
1392  */
1393 uchar_t
1394 apic_xlate_vector(uchar_t vector)
1395 {
1396 	uchar_t	newvector, oldvector = 0;
1397 
1398 	lock_set(&apic_revector_lock);
1399 	/* Do we really need to do this ? */
1400 	if (!apic_revector_pending) {
1401 		lock_clear(&apic_revector_lock);
1402 		return (vector);
1403 	}
1404 	if ((newvector = apic_oldvec_to_newvec[vector]) != 0)
1405 		oldvector = vector;
1406 	else {
1407 		/*
1408 		 * The incoming vector is new . See if a stale entry is
1409 		 * remaining
1410 		 */
1411 		if ((oldvector = apic_newvec_to_oldvec[vector]) != 0)
1412 			newvector = vector;
1413 	}
1414 
1415 	if (oldvector) {
1416 		apic_revector_pending--;
1417 		apic_oldvec_to_newvec[oldvector] = 0;
1418 		apic_newvec_to_oldvec[newvector] = 0;
1419 		apic_free_vector(oldvector);
1420 		lock_clear(&apic_revector_lock);
1421 		/* There could have been more than one reprogramming! */
1422 		return (apic_xlate_vector(newvector));
1423 	}
1424 	lock_clear(&apic_revector_lock);
1425 	return (vector);
1426 }
1427 
1428 void
1429 apic_xlate_vector_free_timeout_handler(void *arg)
1430 {
1431 	ulong_t iflag;
1432 	uchar_t oldvector, newvector;
1433 
1434 	oldvector = (uchar_t)(uintptr_t)arg;
1435 	iflag = intr_clear();
1436 	lock_set(&apic_revector_lock);
1437 	if ((newvector = apic_oldvec_to_newvec[oldvector]) != 0) {
1438 		apic_free_vector(oldvector);
1439 		apic_oldvec_to_newvec[oldvector] = 0;
1440 		apic_newvec_to_oldvec[newvector] = 0;
1441 		apic_revector_pending--;
1442 	}
1443 
1444 	lock_clear(&apic_revector_lock);
1445 	intr_restore(iflag);
1446 }
1447 
1448 /*
1449  * Bind interrupt corresponding to irq_ptr to bind_cpu.
1450  * Must be called with interrupts disabled and apic_ioapic_lock held
1451  */
1452 int
1453 apic_rebind(apic_irq_t *irq_ptr, int bind_cpu,
1454     struct ioapic_reprogram_data *drep)
1455 {
1456 	int			ioapicindex, intin_no;
1457 	uint32_t		airq_temp_cpu;
1458 	apic_cpus_info_t	*cpu_infop;
1459 	uint32_t		rdt_entry;
1460 	int			which_irq;
1461 	ioapic_rdt_t		irdt;
1462 
1463 	which_irq = apic_vector_to_irq[irq_ptr->airq_vector];
1464 
1465 	intin_no = irq_ptr->airq_intin_no;
1466 	ioapicindex = irq_ptr->airq_ioapicindex;
1467 	airq_temp_cpu = irq_ptr->airq_temp_cpu;
1468 	if (airq_temp_cpu != IRQ_UNINIT && airq_temp_cpu != IRQ_UNBOUND) {
1469 		if (airq_temp_cpu & IRQ_USER_BOUND)
1470 			/* Mask off high bit so it can be used as array index */
1471 			airq_temp_cpu &= ~IRQ_USER_BOUND;
1472 
1473 		ASSERT(apic_cpu_in_range(airq_temp_cpu));
1474 	}
1475 
1476 	/*
1477 	 * Can't bind to a CPU that's not accepting interrupts:
1478 	 */
1479 	cpu_infop = &apic_cpus[bind_cpu & ~IRQ_USER_BOUND];
1480 	if (!(cpu_infop->aci_status & APIC_CPU_INTR_ENABLE))
1481 		return (1);
1482 
1483 	/*
1484 	 * If we are about to change the interrupt vector for this interrupt,
1485 	 * and this interrupt is level-triggered, attached to an IOAPIC,
1486 	 * has been delivered to a CPU and that CPU has not handled it
1487 	 * yet, we cannot reprogram the IOAPIC now.
1488 	 */
1489 	if (!APIC_IS_MSI_OR_MSIX_INDEX(irq_ptr->airq_mps_intr_index)) {
1490 
1491 		rdt_entry = READ_IOAPIC_RDT_ENTRY_LOW_DWORD(ioapicindex,
1492 		    intin_no);
1493 
1494 		if ((irq_ptr->airq_vector != RDT_VECTOR(rdt_entry)) &&
1495 		    apic_check_stuck_interrupt(irq_ptr, airq_temp_cpu,
1496 		    bind_cpu, ioapicindex, intin_no, which_irq, drep) != 0) {
1497 
1498 			return (0);
1499 		}
1500 
1501 		/*
1502 		 * NOTE: We do not unmask the RDT here, as an interrupt MAY
1503 		 * still come in before we have a chance to reprogram it below.
1504 		 * The reprogramming below will simultaneously change and
1505 		 * unmask the RDT entry.
1506 		 */
1507 
1508 		if ((uint32_t)bind_cpu == IRQ_UNBOUND) {
1509 			irdt.ir_lo =  AV_LDEST | AV_LOPRI |
1510 			    irq_ptr->airq_rdt_entry;
1511 
1512 			irdt.ir_hi = AV_TOALL >> APIC_ID_BIT_OFFSET;
1513 
1514 			apic_vt_ops->apic_intrmap_alloc_entry(
1515 			    &irq_ptr->airq_intrmap_private, NULL,
1516 			    DDI_INTR_TYPE_FIXED, 1, ioapicindex);
1517 			apic_vt_ops->apic_intrmap_map_entry(
1518 			    irq_ptr->airq_intrmap_private, (void *)&irdt,
1519 			    DDI_INTR_TYPE_FIXED, 1);
1520 			apic_vt_ops->apic_intrmap_record_rdt(
1521 			    irq_ptr->airq_intrmap_private, &irdt);
1522 
1523 			/* Write the RDT entry -- no specific CPU binding */
1524 			WRITE_IOAPIC_RDT_ENTRY_HIGH_DWORD(ioapicindex, intin_no,
1525 			    irdt.ir_hi | AV_TOALL);
1526 
1527 			if (airq_temp_cpu != IRQ_UNINIT && airq_temp_cpu !=
1528 			    IRQ_UNBOUND)
1529 				apic_cpus[airq_temp_cpu].aci_temp_bound--;
1530 
1531 			/*
1532 			 * Write the vector, trigger, and polarity portion of
1533 			 * the RDT
1534 			 */
1535 			WRITE_IOAPIC_RDT_ENTRY_LOW_DWORD(ioapicindex, intin_no,
1536 			    irdt.ir_lo);
1537 
1538 			irq_ptr->airq_temp_cpu = IRQ_UNBOUND;
1539 			return (0);
1540 		}
1541 	}
1542 
1543 	if (bind_cpu & IRQ_USER_BOUND) {
1544 		cpu_infop->aci_bound++;
1545 	} else {
1546 		cpu_infop->aci_temp_bound++;
1547 	}
1548 	ASSERT(apic_cpu_in_range(bind_cpu));
1549 
1550 	if ((airq_temp_cpu != IRQ_UNBOUND) && (airq_temp_cpu != IRQ_UNINIT)) {
1551 		apic_cpus[airq_temp_cpu].aci_temp_bound--;
1552 	}
1553 	if (!APIC_IS_MSI_OR_MSIX_INDEX(irq_ptr->airq_mps_intr_index)) {
1554 
1555 		irdt.ir_lo = AV_PDEST | AV_FIXED | irq_ptr->airq_rdt_entry;
1556 		irdt.ir_hi = cpu_infop->aci_local_id;
1557 
1558 		apic_vt_ops->apic_intrmap_alloc_entry(
1559 		    &irq_ptr->airq_intrmap_private, NULL, DDI_INTR_TYPE_FIXED,
1560 		    1, ioapicindex);
1561 		apic_vt_ops->apic_intrmap_map_entry(
1562 		    irq_ptr->airq_intrmap_private,
1563 		    (void *)&irdt, DDI_INTR_TYPE_FIXED, 1);
1564 		apic_vt_ops->apic_intrmap_record_rdt(
1565 		    irq_ptr->airq_intrmap_private, &irdt);
1566 
1567 		/* Write the RDT entry -- bind to a specific CPU: */
1568 		WRITE_IOAPIC_RDT_ENTRY_HIGH_DWORD(ioapicindex, intin_no,
1569 		    irdt.ir_hi);
1570 
1571 		/* Write the vector, trigger, and polarity portion of the RDT */
1572 		WRITE_IOAPIC_RDT_ENTRY_LOW_DWORD(ioapicindex, intin_no,
1573 		    irdt.ir_lo);
1574 
1575 	} else {
1576 		int type = (irq_ptr->airq_mps_intr_index == MSI_INDEX) ?
1577 		    DDI_INTR_TYPE_MSI : DDI_INTR_TYPE_MSIX;
1578 		if (type == DDI_INTR_TYPE_MSI) {
1579 			if (irq_ptr->airq_ioapicindex ==
1580 			    irq_ptr->airq_origirq) {
1581 				/* first one */
1582 				DDI_INTR_IMPLDBG((CE_CONT, "apic_rebind: call "
1583 				    "apic_pci_msi_enable_vector\n"));
1584 				apic_pci_msi_enable_vector(irq_ptr,
1585 				    type, which_irq, irq_ptr->airq_vector,
1586 				    irq_ptr->airq_intin_no,
1587 				    cpu_infop->aci_local_id);
1588 			}
1589 			if ((irq_ptr->airq_ioapicindex +
1590 			    irq_ptr->airq_intin_no - 1) ==
1591 			    irq_ptr->airq_origirq) { /* last one */
1592 				DDI_INTR_IMPLDBG((CE_CONT, "apic_rebind: call "
1593 				    "apic_pci_msi_enable_mode\n"));
1594 				apic_pci_msi_enable_mode(irq_ptr->airq_dip,
1595 				    type, which_irq);
1596 			}
1597 		} else { /* MSI-X */
1598 			apic_pci_msi_enable_vector(irq_ptr, type,
1599 			    irq_ptr->airq_origirq, irq_ptr->airq_vector, 1,
1600 			    cpu_infop->aci_local_id);
1601 			apic_pci_msi_enable_mode(irq_ptr->airq_dip, type,
1602 			    irq_ptr->airq_origirq);
1603 		}
1604 	}
1605 	irq_ptr->airq_temp_cpu = (uint32_t)bind_cpu;
1606 	apic_redist_cpu_skip &= ~(1 << (bind_cpu & ~IRQ_USER_BOUND));
1607 	return (0);
1608 }
1609 
1610 static void
1611 apic_last_ditch_clear_remote_irr(int ioapic_ix, int intin_no)
1612 {
1613 	if ((READ_IOAPIC_RDT_ENTRY_LOW_DWORD(ioapic_ix, intin_no)
1614 	    & AV_REMOTE_IRR) != 0) {
1615 		/*
1616 		 * Trying to clear the bit through normal
1617 		 * channels has failed.  So as a last-ditch
1618 		 * effort, try to set the trigger mode to
1619 		 * edge, then to level.  This has been
1620 		 * observed to work on many systems.
1621 		 */
1622 		WRITE_IOAPIC_RDT_ENTRY_LOW_DWORD(ioapic_ix,
1623 		    intin_no,
1624 		    READ_IOAPIC_RDT_ENTRY_LOW_DWORD(ioapic_ix,
1625 		    intin_no) & ~AV_LEVEL);
1626 
1627 		WRITE_IOAPIC_RDT_ENTRY_LOW_DWORD(ioapic_ix,
1628 		    intin_no,
1629 		    READ_IOAPIC_RDT_ENTRY_LOW_DWORD(ioapic_ix,
1630 		    intin_no) | AV_LEVEL);
1631 
1632 		/*
1633 		 * If the bit's STILL set, this interrupt may
1634 		 * be hosed.
1635 		 */
1636 		if ((READ_IOAPIC_RDT_ENTRY_LOW_DWORD(ioapic_ix,
1637 		    intin_no) & AV_REMOTE_IRR) != 0) {
1638 
1639 			prom_printf("%s: Remote IRR still "
1640 			    "not clear for IOAPIC %d intin %d.\n"
1641 			    "\tInterrupts to this pin may cease "
1642 			    "functioning.\n", psm_name, ioapic_ix,
1643 			    intin_no);
1644 #ifdef DEBUG
1645 			apic_last_ditch_reprogram_failures++;
1646 #endif
1647 		}
1648 	}
1649 }
1650 
1651 /*
1652  * This function is protected by apic_ioapic_lock coupled with the
1653  * fact that interrupts are disabled.
1654  */
1655 static void
1656 delete_defer_repro_ent(int which_irq)
1657 {
1658 	ASSERT(which_irq >= 0);
1659 	ASSERT(which_irq <= 255);
1660 	ASSERT(LOCK_HELD(&apic_ioapic_lock));
1661 
1662 	if (apic_reprogram_info[which_irq].done)
1663 		return;
1664 
1665 	apic_reprogram_info[which_irq].done = B_TRUE;
1666 
1667 #ifdef DEBUG
1668 	apic_defer_repro_total_retries +=
1669 	    apic_reprogram_info[which_irq].tries;
1670 
1671 	apic_defer_repro_successes++;
1672 #endif
1673 
1674 	if (--apic_reprogram_outstanding == 0) {
1675 
1676 		setlvlx = psm_intr_exit_fn();
1677 	}
1678 }
1679 
1680 
1681 /*
1682  * Interrupts must be disabled during this function to prevent
1683  * self-deadlock.  Interrupts are disabled because this function
1684  * is called from apic_check_stuck_interrupt(), which is called
1685  * from apic_rebind(), which requires its caller to disable interrupts.
1686  */
1687 static void
1688 add_defer_repro_ent(apic_irq_t *irq_ptr, int which_irq, int new_bind_cpu)
1689 {
1690 	ASSERT(which_irq >= 0);
1691 	ASSERT(which_irq <= 255);
1692 	ASSERT(!interrupts_enabled());
1693 
1694 	/*
1695 	 * On the off-chance that there's already a deferred
1696 	 * reprogramming on this irq, check, and if so, just update the
1697 	 * CPU and irq pointer to which the interrupt is targeted, then return.
1698 	 */
1699 	if (!apic_reprogram_info[which_irq].done) {
1700 		apic_reprogram_info[which_irq].bindcpu = new_bind_cpu;
1701 		apic_reprogram_info[which_irq].irqp = irq_ptr;
1702 		return;
1703 	}
1704 
1705 	apic_reprogram_info[which_irq].irqp = irq_ptr;
1706 	apic_reprogram_info[which_irq].bindcpu = new_bind_cpu;
1707 	apic_reprogram_info[which_irq].tries = 0;
1708 	/*
1709 	 * This must be the last thing set, since we're not
1710 	 * grabbing any locks, apic_try_deferred_reprogram() will
1711 	 * make its decision about using this entry iff done
1712 	 * is false.
1713 	 */
1714 	apic_reprogram_info[which_irq].done = B_FALSE;
1715 
1716 	/*
1717 	 * If there were previously no deferred reprogrammings, change
1718 	 * setlvlx to call apic_try_deferred_reprogram()
1719 	 */
1720 	if (++apic_reprogram_outstanding == 1) {
1721 
1722 		setlvlx = apic_try_deferred_reprogram;
1723 	}
1724 }
1725 
1726 static void
1727 apic_try_deferred_reprogram(int prev_ipl, int irq)
1728 {
1729 	int reproirq;
1730 	ulong_t iflag;
1731 	struct ioapic_reprogram_data *drep;
1732 
1733 	(*psm_intr_exit_fn())(prev_ipl, irq);
1734 
1735 	if (!lock_try(&apic_defer_reprogram_lock)) {
1736 		return;
1737 	}
1738 
1739 	/*
1740 	 * Acquire the apic_ioapic_lock so that any other operations that
1741 	 * may affect the apic_reprogram_info state are serialized.
1742 	 * It's still possible for the last deferred reprogramming to clear
1743 	 * between the time we entered this function and the time we get to
1744 	 * the for loop below.  In that case, *setlvlx will have been set
1745 	 * back to *_intr_exit and drep will be NULL. (There's no way to
1746 	 * stop that from happening -- we would need to grab a lock before
1747 	 * calling *setlvlx, which is neither realistic nor prudent).
1748 	 */
1749 	iflag = intr_clear();
1750 	lock_set(&apic_ioapic_lock);
1751 
1752 	/*
1753 	 * For each deferred RDT entry, try to reprogram it now.  Note that
1754 	 * there is no lock acquisition to read apic_reprogram_info because
1755 	 * '.done' is set only after the other fields in the structure are set.
1756 	 */
1757 
1758 	drep = NULL;
1759 	for (reproirq = 0; reproirq <= APIC_MAX_VECTOR; reproirq++) {
1760 		if (apic_reprogram_info[reproirq].done == B_FALSE) {
1761 			drep = &apic_reprogram_info[reproirq];
1762 			break;
1763 		}
1764 	}
1765 
1766 	/*
1767 	 * Either we found a deferred action to perform, or
1768 	 * we entered this function spuriously, after *setlvlx
1769 	 * was restored to point to *_intr_exit.  Any other
1770 	 * permutation is invalid.
1771 	 */
1772 	ASSERT(drep != NULL || *setlvlx == psm_intr_exit_fn());
1773 
1774 	/*
1775 	 * Though we can't really do anything about errors
1776 	 * at this point, keep track of them for reporting.
1777 	 * Note that it is very possible for apic_setup_io_intr
1778 	 * to re-register this very timeout if the Remote IRR bit
1779 	 * has not yet cleared.
1780 	 */
1781 
1782 #ifdef DEBUG
1783 	if (drep != NULL) {
1784 		if (apic_setup_io_intr(drep, reproirq, B_TRUE) != 0) {
1785 			apic_deferred_setup_failures++;
1786 		}
1787 	} else {
1788 		apic_deferred_spurious_enters++;
1789 	}
1790 #else
1791 	if (drep != NULL)
1792 		(void) apic_setup_io_intr(drep, reproirq, B_TRUE);
1793 #endif
1794 
1795 	lock_clear(&apic_ioapic_lock);
1796 	intr_restore(iflag);
1797 
1798 	lock_clear(&apic_defer_reprogram_lock);
1799 }
1800 
1801 static void
1802 apic_ioapic_wait_pending_clear(int ioapic_ix, int intin_no)
1803 {
1804 	int waited;
1805 
1806 	/*
1807 	 * Wait for the delivery pending bit to clear.
1808 	 */
1809 	if ((READ_IOAPIC_RDT_ENTRY_LOW_DWORD(ioapic_ix, intin_no) &
1810 	    (AV_LEVEL|AV_PENDING)) == (AV_LEVEL|AV_PENDING)) {
1811 
1812 		/*
1813 		 * If we're still waiting on the delivery of this interrupt,
1814 		 * continue to wait here until it is delivered (this should be
1815 		 * a very small amount of time, but include a timeout just in
1816 		 * case).
1817 		 */
1818 		for (waited = 0; waited < apic_max_reps_clear_pending;
1819 		    waited++) {
1820 			if ((READ_IOAPIC_RDT_ENTRY_LOW_DWORD(ioapic_ix,
1821 			    intin_no) & AV_PENDING) == 0) {
1822 				break;
1823 			}
1824 		}
1825 	}
1826 }
1827 
1828 
1829 /*
1830  * Checks to see if the IOAPIC interrupt entry specified has its Remote IRR
1831  * bit set.  Calls functions that modify the function that setlvlx points to,
1832  * so that the reprogramming can be retried very shortly.
1833  *
1834  * This function will mask the RDT entry if the interrupt is level-triggered.
1835  * (The caller is responsible for unmasking the RDT entry.)
1836  *
1837  * Returns non-zero if the caller should defer IOAPIC reprogramming.
1838  */
1839 static int
1840 apic_check_stuck_interrupt(apic_irq_t *irq_ptr, int old_bind_cpu,
1841     int new_bind_cpu, int ioapic_ix, int intin_no, int which_irq,
1842     struct ioapic_reprogram_data *drep)
1843 {
1844 	int32_t			rdt_entry;
1845 	int			waited;
1846 	int			reps = 0;
1847 
1848 	/*
1849 	 * Wait for the delivery pending bit to clear.
1850 	 */
1851 	do {
1852 		++reps;
1853 
1854 		apic_ioapic_wait_pending_clear(ioapic_ix, intin_no);
1855 
1856 		/*
1857 		 * Mask the RDT entry, but only if it's a level-triggered
1858 		 * interrupt
1859 		 */
1860 		rdt_entry = READ_IOAPIC_RDT_ENTRY_LOW_DWORD(ioapic_ix,
1861 		    intin_no);
1862 		if ((rdt_entry & (AV_LEVEL|AV_MASK)) == AV_LEVEL) {
1863 
1864 			/* Mask it */
1865 			WRITE_IOAPIC_RDT_ENTRY_LOW_DWORD(ioapic_ix, intin_no,
1866 			    AV_MASK | rdt_entry);
1867 		}
1868 
1869 		if ((rdt_entry & AV_LEVEL) == AV_LEVEL) {
1870 			/*
1871 			 * If there was a race and an interrupt was injected
1872 			 * just before we masked, check for that case here.
1873 			 * Then, unmask the RDT entry and try again.  If we're
1874 			 * on our last try, don't unmask (because we want the
1875 			 * RDT entry to remain masked for the rest of the
1876 			 * function).
1877 			 */
1878 			rdt_entry = READ_IOAPIC_RDT_ENTRY_LOW_DWORD(ioapic_ix,
1879 			    intin_no);
1880 			if ((rdt_entry & AV_PENDING) &&
1881 			    (reps < apic_max_reps_clear_pending)) {
1882 				/* Unmask it */
1883 				WRITE_IOAPIC_RDT_ENTRY_LOW_DWORD(ioapic_ix,
1884 				    intin_no, rdt_entry & ~AV_MASK);
1885 			}
1886 		}
1887 
1888 	} while ((rdt_entry & AV_PENDING) &&
1889 	    (reps < apic_max_reps_clear_pending));
1890 
1891 #ifdef DEBUG
1892 		if (rdt_entry & AV_PENDING)
1893 			apic_intr_deliver_timeouts++;
1894 #endif
1895 
1896 	/*
1897 	 * If the remote IRR bit is set, then the interrupt has been sent
1898 	 * to a CPU for processing.  We have no choice but to wait for
1899 	 * that CPU to process the interrupt, at which point the remote IRR
1900 	 * bit will be cleared.
1901 	 */
1902 	if ((READ_IOAPIC_RDT_ENTRY_LOW_DWORD(ioapic_ix, intin_no) &
1903 	    (AV_LEVEL|AV_REMOTE_IRR)) == (AV_LEVEL|AV_REMOTE_IRR)) {
1904 
1905 		/*
1906 		 * If the CPU that this RDT is bound to is NOT the current
1907 		 * CPU, wait until that CPU handles the interrupt and ACKs
1908 		 * it.  If this interrupt is not bound to any CPU (that is,
1909 		 * if it's bound to the logical destination of "anyone"), it
1910 		 * may have been delivered to the current CPU so handle that
1911 		 * case by deferring the reprogramming (below).
1912 		 */
1913 		if ((old_bind_cpu != IRQ_UNBOUND) &&
1914 		    (old_bind_cpu != IRQ_UNINIT) &&
1915 		    (old_bind_cpu != psm_get_cpu_id())) {
1916 			for (waited = 0; waited < apic_max_reps_clear_pending;
1917 			    waited++) {
1918 				if ((READ_IOAPIC_RDT_ENTRY_LOW_DWORD(ioapic_ix,
1919 				    intin_no) & AV_REMOTE_IRR) == 0) {
1920 
1921 					delete_defer_repro_ent(which_irq);
1922 
1923 					/* Remote IRR has cleared! */
1924 					return (0);
1925 				}
1926 			}
1927 		}
1928 
1929 		/*
1930 		 * If we waited and the Remote IRR bit is still not cleared,
1931 		 * AND if we've invoked the timeout APIC_REPROGRAM_MAX_TIMEOUTS
1932 		 * times for this interrupt, try the last-ditch workaround:
1933 		 */
1934 		if (drep && drep->tries >= APIC_REPROGRAM_MAX_TRIES) {
1935 
1936 			apic_last_ditch_clear_remote_irr(ioapic_ix, intin_no);
1937 
1938 			/* Mark this one as reprogrammed: */
1939 			delete_defer_repro_ent(which_irq);
1940 
1941 			return (0);
1942 		} else {
1943 #ifdef DEBUG
1944 			apic_intr_deferrals++;
1945 #endif
1946 
1947 			/*
1948 			 * If waiting for the Remote IRR bit (above) didn't
1949 			 * allow it to clear, defer the reprogramming.
1950 			 * Add a new deferred-programming entry if the
1951 			 * caller passed a NULL one (and update the existing one
1952 			 * in case anything changed).
1953 			 */
1954 			add_defer_repro_ent(irq_ptr, which_irq, new_bind_cpu);
1955 			if (drep)
1956 				drep->tries++;
1957 
1958 			/* Inform caller to defer IOAPIC programming: */
1959 			return (1);
1960 		}
1961 
1962 	}
1963 
1964 	/* Remote IRR is clear */
1965 	delete_defer_repro_ent(which_irq);
1966 
1967 	return (0);
1968 }
1969 
1970 /*
1971  * Called to migrate all interrupts at an irq to another cpu.
1972  * Must be called with interrupts disabled and apic_ioapic_lock held
1973  */
1974 int
1975 apic_rebind_all(apic_irq_t *irq_ptr, int bind_cpu)
1976 {
1977 	apic_irq_t	*irqptr = irq_ptr;
1978 	int		retval = 0;
1979 
1980 	while (irqptr) {
1981 		if (irqptr->airq_temp_cpu != IRQ_UNINIT)
1982 			retval |= apic_rebind(irqptr, bind_cpu, NULL);
1983 		irqptr = irqptr->airq_next;
1984 	}
1985 
1986 	return (retval);
1987 }
1988 
1989 /*
1990  * apic_intr_redistribute does all the messy computations for identifying
1991  * which interrupt to move to which CPU. Currently we do just one interrupt
1992  * at a time. This reduces the time we spent doing all this within clock
1993  * interrupt. When it is done in idle, we could do more than 1.
1994  * First we find the most busy and the most free CPU (time in ISR only)
1995  * skipping those CPUs that has been identified as being ineligible (cpu_skip)
1996  * Then we look for IRQs which are closest to the difference between the
1997  * most busy CPU and the average ISR load. We try to find one whose load
1998  * is less than difference.If none exists, then we chose one larger than the
1999  * difference, provided it does not make the most idle CPU worse than the
2000  * most busy one. In the end, we clear all the busy fields for CPUs. For
2001  * IRQs, they are cleared as they are scanned.
2002  */
2003 void
2004 apic_intr_redistribute(void)
2005 {
2006 	int busiest_cpu, most_free_cpu;
2007 	int cpu_free, cpu_busy, max_busy, min_busy;
2008 	int min_free, diff;
2009 	int average_busy, cpus_online;
2010 	int i, busy;
2011 	ulong_t iflag;
2012 	apic_cpus_info_t *cpu_infop;
2013 	apic_irq_t *min_busy_irq = NULL;
2014 	apic_irq_t *max_busy_irq = NULL;
2015 
2016 	busiest_cpu = most_free_cpu = -1;
2017 	cpu_free = cpu_busy = max_busy = average_busy = 0;
2018 	min_free = apic_sample_factor_redistribution;
2019 	cpus_online = 0;
2020 	/*
2021 	 * Below we will check for CPU_INTR_ENABLE, bound, temp_bound, temp_cpu
2022 	 * without ioapic_lock. That is OK as we are just doing statistical
2023 	 * sampling anyway and any inaccuracy now will get corrected next time
2024 	 * The call to rebind which actually changes things will make sure
2025 	 * we are consistent.
2026 	 */
2027 	for (i = 0; i < apic_nproc; i++) {
2028 		if (apic_cpu_in_range(i) &&
2029 		    !(apic_redist_cpu_skip & (1 << i)) &&
2030 		    (apic_cpus[i].aci_status & APIC_CPU_INTR_ENABLE)) {
2031 
2032 			cpu_infop = &apic_cpus[i];
2033 			/*
2034 			 * If no unbound interrupts or only 1 total on this
2035 			 * CPU, skip
2036 			 */
2037 			if (!cpu_infop->aci_temp_bound ||
2038 			    (cpu_infop->aci_bound + cpu_infop->aci_temp_bound)
2039 			    == 1) {
2040 				apic_redist_cpu_skip |= 1 << i;
2041 				continue;
2042 			}
2043 
2044 			busy = cpu_infop->aci_busy;
2045 			average_busy += busy;
2046 			cpus_online++;
2047 			if (max_busy < busy) {
2048 				max_busy = busy;
2049 				busiest_cpu = i;
2050 			}
2051 			if (min_free > busy) {
2052 				min_free = busy;
2053 				most_free_cpu = i;
2054 			}
2055 			if (busy > apic_int_busy_mark) {
2056 				cpu_busy |= 1 << i;
2057 			} else {
2058 				if (busy < apic_int_free_mark)
2059 					cpu_free |= 1 << i;
2060 			}
2061 		}
2062 	}
2063 	if ((cpu_busy && cpu_free) ||
2064 	    (max_busy >= (min_free + apic_diff_for_redistribution))) {
2065 
2066 		apic_num_imbalance++;
2067 #ifdef	DEBUG
2068 		if (apic_verbose & APIC_VERBOSE_IOAPIC_FLAG) {
2069 			prom_printf(
2070 			    "redistribute busy=%x free=%x max=%x min=%x",
2071 			    cpu_busy, cpu_free, max_busy, min_free);
2072 		}
2073 #endif /* DEBUG */
2074 
2075 
2076 		average_busy /= cpus_online;
2077 
2078 		diff = max_busy - average_busy;
2079 		min_busy = max_busy; /* start with the max possible value */
2080 		max_busy = 0;
2081 		min_busy_irq = max_busy_irq = NULL;
2082 		i = apic_min_device_irq;
2083 		for (; i <= apic_max_device_irq; i++) {
2084 			apic_irq_t *irq_ptr;
2085 			/* Change to linked list per CPU ? */
2086 			if ((irq_ptr = apic_irq_table[i]) == NULL)
2087 				continue;
2088 			/* Check for irq_busy & decide which one to move */
2089 			/* Also zero them for next round */
2090 			if ((irq_ptr->airq_temp_cpu == busiest_cpu) &&
2091 			    irq_ptr->airq_busy) {
2092 				if (irq_ptr->airq_busy < diff) {
2093 					/*
2094 					 * Check for least busy CPU,
2095 					 * best fit or what ?
2096 					 */
2097 					if (max_busy < irq_ptr->airq_busy) {
2098 						/*
2099 						 * Most busy within the
2100 						 * required differential
2101 						 */
2102 						max_busy = irq_ptr->airq_busy;
2103 						max_busy_irq = irq_ptr;
2104 					}
2105 				} else {
2106 					if (min_busy > irq_ptr->airq_busy) {
2107 						/*
2108 						 * least busy, but more than
2109 						 * the reqd diff
2110 						 */
2111 						if (min_busy <
2112 						    (diff + average_busy -
2113 						    min_free)) {
2114 							/*
2115 							 * Making sure new cpu
2116 							 * will not end up
2117 							 * worse
2118 							 */
2119 							min_busy =
2120 							    irq_ptr->airq_busy;
2121 
2122 							min_busy_irq = irq_ptr;
2123 						}
2124 					}
2125 				}
2126 			}
2127 			irq_ptr->airq_busy = 0;
2128 		}
2129 
2130 		if (max_busy_irq != NULL) {
2131 #ifdef	DEBUG
2132 			if (apic_verbose & APIC_VERBOSE_IOAPIC_FLAG) {
2133 				prom_printf("rebinding %x to %x",
2134 				    max_busy_irq->airq_vector, most_free_cpu);
2135 			}
2136 #endif /* DEBUG */
2137 			iflag = intr_clear();
2138 			if (lock_try(&apic_ioapic_lock)) {
2139 				if (apic_rebind_all(max_busy_irq,
2140 				    most_free_cpu) == 0) {
2141 					/* Make change permenant */
2142 					max_busy_irq->airq_cpu =
2143 					    (uint32_t)most_free_cpu;
2144 				}
2145 				lock_clear(&apic_ioapic_lock);
2146 			}
2147 			intr_restore(iflag);
2148 
2149 		} else if (min_busy_irq != NULL) {
2150 #ifdef	DEBUG
2151 			if (apic_verbose & APIC_VERBOSE_IOAPIC_FLAG) {
2152 				prom_printf("rebinding %x to %x",
2153 				    min_busy_irq->airq_vector, most_free_cpu);
2154 			}
2155 #endif /* DEBUG */
2156 
2157 			iflag = intr_clear();
2158 			if (lock_try(&apic_ioapic_lock)) {
2159 				if (apic_rebind_all(min_busy_irq,
2160 				    most_free_cpu) == 0) {
2161 					/* Make change permenant */
2162 					min_busy_irq->airq_cpu =
2163 					    (uint32_t)most_free_cpu;
2164 				}
2165 				lock_clear(&apic_ioapic_lock);
2166 			}
2167 			intr_restore(iflag);
2168 
2169 		} else {
2170 			if (cpu_busy != (1 << busiest_cpu)) {
2171 				apic_redist_cpu_skip |= 1 << busiest_cpu;
2172 				/*
2173 				 * We leave cpu_skip set so that next time we
2174 				 * can choose another cpu
2175 				 */
2176 			}
2177 		}
2178 		apic_num_rebind++;
2179 	} else {
2180 		/*
2181 		 * found nothing. Could be that we skipped over valid CPUs
2182 		 * or we have balanced everything. If we had a variable
2183 		 * ticks_for_redistribution, it could be increased here.
2184 		 * apic_int_busy, int_free etc would also need to be
2185 		 * changed.
2186 		 */
2187 		if (apic_redist_cpu_skip)
2188 			apic_redist_cpu_skip = 0;
2189 	}
2190 	for (i = 0; i < apic_nproc; i++) {
2191 		if (apic_cpu_in_range(i)) {
2192 			apic_cpus[i].aci_busy = 0;
2193 		}
2194 	}
2195 }
2196 
2197 void
2198 apic_cleanup_busy(void)
2199 {
2200 	int i;
2201 	apic_irq_t *irq_ptr;
2202 
2203 	for (i = 0; i < apic_nproc; i++) {
2204 		if (apic_cpu_in_range(i)) {
2205 			apic_cpus[i].aci_busy = 0;
2206 		}
2207 	}
2208 
2209 	for (i = apic_min_device_irq; i <= apic_max_device_irq; i++) {
2210 		if ((irq_ptr = apic_irq_table[i]) != NULL)
2211 			irq_ptr->airq_busy = 0;
2212 	}
2213 }
2214