xref: /illumos-gate/usr/src/uts/i86pc/io/immu_qinv.c (revision 69b3e10436272b0970d58743de375d0dd61046ce)
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  * Portions Copyright 2010 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 /*
28  * Copyright (c) 2009, Intel Corporation.
29  * All rights reserved.
30  */
31 
32 #include <sys/ddi.h>
33 #include <sys/archsystm.h>
34 #include <vm/hat_i86.h>
35 #include <sys/types.h>
36 #include <sys/sysmacros.h>
37 #include <sys/immu.h>
38 
39 /* invalidation queue table entry size */
40 #define	QINV_ENTRY_SIZE		0x10
41 
42 /* max value of Queue Size field of Invalidation Queue Address Register */
43 #define	QINV_MAX_QUEUE_SIZE	0x7
44 
45 /* status data size of invalidation wait descriptor */
46 #define	QINV_SYNC_DATA_SIZE	0x4
47 
48 /* status data value of invalidation wait descriptor */
49 #define	QINV_SYNC_DATA_FENCE	1
50 #define	QINV_SYNC_DATA_UNFENCE	2
51 
52 /* invalidation queue head and tail */
53 #define	QINV_IQA_HEAD(QH)	BITX((QH), 18, 4)
54 #define	QINV_IQA_TAIL_SHIFT	4
55 
56 /* invalidation queue entry structure */
57 typedef struct qinv_inv_dsc {
58 	uint64_t	lo;
59 	uint64_t	hi;
60 } qinv_dsc_t;
61 
62 /*
63  * struct iotlb_cache_node
64  *   the pending data for iotlb flush
65  */
66 typedef struct iotlb_pend_node {
67 	dvcookie_t	*icn_dvcookies;  /* ptr to dvma cookie array */
68 	uint_t		icn_count;  /* valid cookie count */
69 	uint_t		icn_array_size;  /* array size */
70 	list_node_t	node;
71 } qinv_iotlb_pend_node_t;
72 
73 /*
74  * struct iotlb_cache_head
75  *   the pending head for the iotlb flush
76  */
77 typedef struct iotlb_pend_head {
78 	/* the pending node cache list */
79 	kmutex_t	ich_mem_lock;
80 	list_t		ich_mem_list;
81 } qinv_iotlb_pend_head_t;
82 
83 /*
84  * qinv_iotlb_t
85  *   pending data for qiueued invalidation iotlb flush
86  */
87 typedef struct qinv_iotlb {
88 	dvcookie_t	*qinv_iotlb_dvcookies;
89 	uint_t		qinv_iotlb_count;
90 	uint_t		qinv_iotlb_size;
91 	list_node_t	qinv_iotlb_node;
92 } qinv_iotlb_t;
93 
94 /* physical contigous pages for invalidation queue */
95 typedef struct qinv_mem {
96 	kmutex_t	   qinv_mem_lock;
97 	ddi_dma_handle_t   qinv_mem_dma_hdl;
98 	ddi_acc_handle_t   qinv_mem_acc_hdl;
99 	caddr_t		   qinv_mem_vaddr;
100 	paddr_t		   qinv_mem_paddr;
101 	uint_t		   qinv_mem_size;
102 	uint16_t	   qinv_mem_head;
103 	uint16_t	   qinv_mem_tail;
104 } qinv_mem_t;
105 
106 
107 /*
108  * invalidation queue state
109  *   This structure describes the state information of the
110  *   invalidation queue table and related status memeory for
111  *   invalidation wait descriptor
112  *
113  * qinv_table		- invalidation queue table
114  * qinv_sync		- sync status memory for invalidation wait descriptor
115  * qinv_iotlb_pend_node	- pending iotlb node
116  */
117 typedef struct qinv {
118 	qinv_mem_t		qinv_table;
119 	qinv_mem_t		qinv_sync;
120 	qinv_iotlb_pend_head_t qinv_pend_head;
121 	qinv_iotlb_pend_node_t  **qinv_iotlb_pend_node;
122 } qinv_t;
123 
124 
125 /* helper macro for making queue invalidation descriptor */
126 #define	INV_DSC_TYPE(dsc)	((dsc)->lo & 0xF)
127 #define	CC_INV_DSC_HIGH		(0)
128 #define	CC_INV_DSC_LOW(fm, sid, did, g)	(((uint64_t)(fm) << 48) | \
129 	((uint64_t)(sid) << 32) | \
130 	((uint64_t)(did) << 16) | \
131 	((uint64_t)(g) << 4) | \
132 	1)
133 
134 #define	IOTLB_INV_DSC_HIGH(addr, ih, am) (((uint64_t)(addr)) | \
135 	((uint64_t)(ih) << 6) |	\
136 	((uint64_t)(am)))
137 
138 #define	IOTLB_INV_DSC_LOW(did, dr, dw, g) (((uint64_t)(did) << 16) | \
139 	((uint64_t)(dr) << 7) | \
140 	((uint64_t)(dw) << 6) | \
141 	((uint64_t)(g) << 4) | \
142 	2)
143 
144 #define	DEV_IOTLB_INV_DSC_HIGH(addr, s) (((uint64_t)(addr)) | (s))
145 
146 #define	DEV_IOTLB_INV_DSC_LOW(sid, max_invs_pd) ( \
147 	((uint64_t)(sid) << 32) | \
148 	((uint64_t)(max_invs_pd) << 16) | \
149 	3)
150 
151 #define	IEC_INV_DSC_HIGH (0)
152 #define	IEC_INV_DSC_LOW(idx, im, g) (((uint64_t)(idx) << 32) | \
153 	((uint64_t)(im) << 27) | \
154 	((uint64_t)(g) << 4) | \
155 	4)
156 
157 #define	INV_WAIT_DSC_HIGH(saddr) ((uint64_t)(saddr))
158 
159 #define	INV_WAIT_DSC_LOW(sdata, fn, sw, iflag) (((uint64_t)(sdata) << 32) | \
160 	((uint64_t)(fn) << 6) | \
161 	((uint64_t)(sw) << 5) | \
162 	((uint64_t)(iflag) << 4) | \
163 	5)
164 
165 /*
166  * QS field of Invalidation Queue Address Register
167  * the size of invalidation queue is 1 << (qinv_iqa_qs + 8)
168  */
169 static uint_t qinv_iqa_qs = 6;
170 
171 /*
172  * the invalidate desctiptor type of queued invalidation interface
173  */
174 static char *qinv_dsc_type[] = {
175 	"Reserved",
176 	"Context Cache Invalidate Descriptor",
177 	"IOTLB Invalidate Descriptor",
178 	"Device-IOTLB Invalidate Descriptor",
179 	"Interrupt Entry Cache Invalidate Descriptor",
180 	"Invalidation Wait Descriptor",
181 	"Incorrect queue invalidation type"
182 };
183 
184 #define	QINV_MAX_DSC_TYPE	(sizeof (qinv_dsc_type) / sizeof (char *))
185 
186 /*
187  * the queued invalidation interface functions
188  */
189 static void qinv_submit_inv_dsc(immu_t *immu, qinv_dsc_t *dsc);
190 static void qinv_context_common(immu_t *immu, uint8_t function_mask,
191     uint16_t source_id, uint_t domain_id, ctt_inv_g_t type);
192 static void qinv_iotlb_common(immu_t *immu, uint_t domain_id,
193     uint64_t addr, uint_t am, uint_t hint, tlb_inv_g_t type);
194 static void qinv_iec_common(immu_t *immu, uint_t iidx,
195     uint_t im, uint_t g);
196 static uint_t qinv_alloc_sync_mem_entry(immu_t *immu);
197 static void qinv_wait_async_unfence(immu_t *immu,
198     qinv_iotlb_pend_node_t *node);
199 static void qinv_wait_sync(immu_t *immu);
200 static int qinv_wait_async_finish(immu_t *immu, int *count);
201 /*LINTED*/
202 static void qinv_wait_async_fence(immu_t *immu);
203 /*LINTED*/
204 static void qinv_dev_iotlb_common(immu_t *immu, uint16_t sid,
205     uint64_t addr, uint_t size, uint_t max_invs_pd);
206 
207 
208 /* submit invalidation request descriptor to invalidation queue */
209 static void
210 qinv_submit_inv_dsc(immu_t *immu, qinv_dsc_t *dsc)
211 {
212 	qinv_t *qinv;
213 	qinv_mem_t *qinv_table;
214 	uint_t tail;
215 
216 	qinv = (qinv_t *)immu->immu_qinv;
217 	qinv_table = &(qinv->qinv_table);
218 
219 	mutex_enter(&qinv_table->qinv_mem_lock);
220 	tail = qinv_table->qinv_mem_tail;
221 	qinv_table->qinv_mem_tail++;
222 
223 	if (qinv_table->qinv_mem_tail == qinv_table->qinv_mem_size)
224 		qinv_table->qinv_mem_tail = 0;
225 
226 	while (qinv_table->qinv_mem_head == qinv_table->qinv_mem_tail) {
227 		/*
228 		 * inv queue table exhausted, wait hardware to fetch
229 		 * next descriptor
230 		 */
231 		qinv_table->qinv_mem_head = QINV_IQA_HEAD(
232 		    immu_regs_get64(immu, IMMU_REG_INVAL_QH));
233 	}
234 
235 	bcopy(dsc, qinv_table->qinv_mem_vaddr + tail * QINV_ENTRY_SIZE,
236 	    QINV_ENTRY_SIZE);
237 
238 	immu_regs_put64(immu, IMMU_REG_INVAL_QT,
239 	    qinv_table->qinv_mem_tail << QINV_IQA_TAIL_SHIFT);
240 
241 	mutex_exit(&qinv_table->qinv_mem_lock);
242 }
243 
244 /* queued invalidation interface -- invalidate context cache */
245 static void
246 qinv_context_common(immu_t *immu, uint8_t function_mask,
247     uint16_t source_id, uint_t domain_id, ctt_inv_g_t type)
248 {
249 	qinv_dsc_t dsc;
250 
251 	dsc.lo = CC_INV_DSC_LOW(function_mask, source_id, domain_id, type);
252 	dsc.hi = CC_INV_DSC_HIGH;
253 
254 	qinv_submit_inv_dsc(immu, &dsc);
255 }
256 
257 /* queued invalidation interface -- invalidate iotlb */
258 static void
259 qinv_iotlb_common(immu_t *immu, uint_t domain_id,
260     uint64_t addr, uint_t am, uint_t hint, tlb_inv_g_t type)
261 {
262 	qinv_dsc_t dsc;
263 	uint8_t dr = 0;
264 	uint8_t dw = 0;
265 
266 	if (IMMU_CAP_GET_DRD(immu->immu_regs_cap))
267 		dr = 1;
268 	if (IMMU_CAP_GET_DWD(immu->immu_regs_cap))
269 		dw = 1;
270 
271 	switch (type) {
272 	case TLB_INV_G_PAGE:
273 		if (!IMMU_CAP_GET_PSI(immu->immu_regs_cap) ||
274 		    am > IMMU_CAP_GET_MAMV(immu->immu_regs_cap) ||
275 		    addr & IMMU_PAGEOFFSET) {
276 			type = TLB_INV_G_DOMAIN;
277 			goto qinv_ignore_psi;
278 		}
279 		dsc.lo = IOTLB_INV_DSC_LOW(domain_id, dr, dw, type);
280 		dsc.hi = IOTLB_INV_DSC_HIGH(addr, hint, am);
281 		break;
282 
283 	qinv_ignore_psi:
284 	case TLB_INV_G_DOMAIN:
285 		dsc.lo = IOTLB_INV_DSC_LOW(domain_id, dr, dw, type);
286 		dsc.hi = 0;
287 		break;
288 
289 	case TLB_INV_G_GLOBAL:
290 		dsc.lo = IOTLB_INV_DSC_LOW(0, dr, dw, type);
291 		dsc.hi = 0;
292 		break;
293 	default:
294 		ddi_err(DER_WARN, NULL, "incorrect iotlb flush type");
295 		return;
296 	}
297 
298 	qinv_submit_inv_dsc(immu, &dsc);
299 }
300 
301 /* queued invalidation interface -- invalidate dev_iotlb */
302 static void
303 qinv_dev_iotlb_common(immu_t *immu, uint16_t sid,
304     uint64_t addr, uint_t size, uint_t max_invs_pd)
305 {
306 	qinv_dsc_t dsc;
307 
308 	dsc.lo = DEV_IOTLB_INV_DSC_LOW(sid, max_invs_pd);
309 	dsc.hi = DEV_IOTLB_INV_DSC_HIGH(addr, size);
310 
311 	qinv_submit_inv_dsc(immu, &dsc);
312 }
313 
314 /* queued invalidation interface -- invalidate interrupt entry cache */
315 static void
316 qinv_iec_common(immu_t *immu, uint_t iidx, uint_t im, uint_t g)
317 {
318 	qinv_dsc_t dsc;
319 
320 	dsc.lo = IEC_INV_DSC_LOW(iidx, im, g);
321 	dsc.hi = IEC_INV_DSC_HIGH;
322 
323 	qinv_submit_inv_dsc(immu, &dsc);
324 }
325 
326 /*
327  * alloc free entry from sync status table
328  */
329 static uint_t
330 qinv_alloc_sync_mem_entry(immu_t *immu)
331 {
332 	qinv_mem_t *sync_mem;
333 	uint_t tail;
334 	qinv_t *qinv;
335 
336 	qinv = (qinv_t *)immu->immu_qinv;
337 	sync_mem = &qinv->qinv_sync;
338 
339 sync_mem_exhausted:
340 	mutex_enter(&sync_mem->qinv_mem_lock);
341 	tail = sync_mem->qinv_mem_tail;
342 	sync_mem->qinv_mem_tail++;
343 	if (sync_mem->qinv_mem_tail == sync_mem->qinv_mem_size)
344 		sync_mem->qinv_mem_tail = 0;
345 
346 	if (sync_mem->qinv_mem_head == sync_mem->qinv_mem_tail) {
347 		/* should never happen */
348 		ddi_err(DER_WARN, NULL, "sync mem exhausted");
349 		sync_mem->qinv_mem_tail = tail;
350 		mutex_exit(&sync_mem->qinv_mem_lock);
351 		delay(IMMU_ALLOC_RESOURCE_DELAY);
352 		goto sync_mem_exhausted;
353 	}
354 	mutex_exit(&sync_mem->qinv_mem_lock);
355 
356 	return (tail);
357 }
358 
359 /*
360  * queued invalidation interface -- invalidation wait descriptor
361  *   fence flag not set, need status data to indicate the invalidation
362  *   wait descriptor completion
363  */
364 static void
365 qinv_wait_async_unfence(immu_t *immu, qinv_iotlb_pend_node_t *node)
366 {
367 	qinv_dsc_t dsc;
368 	qinv_mem_t *sync_mem;
369 	uint64_t saddr;
370 	uint_t tail;
371 	qinv_t *qinv;
372 
373 	qinv = (qinv_t *)immu->immu_qinv;
374 	sync_mem = &qinv->qinv_sync;
375 	tail = qinv_alloc_sync_mem_entry(immu);
376 
377 	/* plant an iotlb pending node */
378 	qinv->qinv_iotlb_pend_node[tail] = node;
379 
380 	saddr = sync_mem->qinv_mem_paddr + tail * QINV_SYNC_DATA_SIZE;
381 
382 	/*
383 	 * sdata = QINV_SYNC_DATA_UNFENCE, fence = 0, sw = 1, if = 0
384 	 * indicate the invalidation wait descriptor completion by
385 	 * performing a coherent DWORD write to the status address,
386 	 * not by generating an invalidation completion event
387 	 */
388 	dsc.lo = INV_WAIT_DSC_LOW(QINV_SYNC_DATA_UNFENCE, 0, 1, 0);
389 	dsc.hi = INV_WAIT_DSC_HIGH(saddr);
390 
391 	qinv_submit_inv_dsc(immu, &dsc);
392 }
393 
394 /*
395  * queued invalidation interface -- invalidation wait descriptor
396  *   fence flag set, indicate descriptors following the invalidation
397  *   wait descriptor must be processed by hardware only after the
398  *   invalidation wait descriptor completes.
399  */
400 static void
401 qinv_wait_async_fence(immu_t *immu)
402 {
403 	qinv_dsc_t dsc;
404 
405 	/* sw = 0, fence = 1, iflag = 0 */
406 	dsc.lo = INV_WAIT_DSC_LOW(0, 1, 0, 0);
407 	dsc.hi = 0;
408 	qinv_submit_inv_dsc(immu, &dsc);
409 }
410 
411 /*
412  * queued invalidation interface -- invalidation wait descriptor
413  *   wait until the invalidation request finished
414  */
415 static void
416 qinv_wait_sync(immu_t *immu)
417 {
418 	qinv_dsc_t dsc;
419 	qinv_mem_t *sync_mem;
420 	uint64_t saddr;
421 	uint_t tail;
422 	qinv_t *qinv;
423 	volatile uint32_t *status;
424 
425 	qinv = (qinv_t *)immu->immu_qinv;
426 	sync_mem = &qinv->qinv_sync;
427 	tail = qinv_alloc_sync_mem_entry(immu);
428 	saddr = sync_mem->qinv_mem_paddr + tail * QINV_SYNC_DATA_SIZE;
429 	status = (uint32_t *)(sync_mem->qinv_mem_vaddr + tail *
430 	    QINV_SYNC_DATA_SIZE);
431 
432 	/*
433 	 * sdata = QINV_SYNC_DATA_FENCE, fence = 1, sw = 1, if = 0
434 	 * indicate the invalidation wait descriptor completion by
435 	 * performing a coherent DWORD write to the status address,
436 	 * not by generating an invalidation completion event
437 	 */
438 	dsc.lo = INV_WAIT_DSC_LOW(QINV_SYNC_DATA_FENCE, 1, 1, 0);
439 	dsc.hi = INV_WAIT_DSC_HIGH(saddr);
440 
441 	qinv_submit_inv_dsc(immu, &dsc);
442 
443 	while ((*status) != QINV_SYNC_DATA_FENCE)
444 		iommu_cpu_nop();
445 	*status = QINV_SYNC_DATA_UNFENCE;
446 }
447 
448 /* get already completed invalidation wait requests */
449 static int
450 qinv_wait_async_finish(immu_t *immu, int *cnt)
451 {
452 	qinv_mem_t *sync_mem;
453 	int index;
454 	qinv_t *qinv;
455 	volatile uint32_t *value;
456 
457 	ASSERT((*cnt) == 0);
458 
459 	qinv = (qinv_t *)immu->immu_qinv;
460 	sync_mem = &qinv->qinv_sync;
461 
462 	mutex_enter(&sync_mem->qinv_mem_lock);
463 	index = sync_mem->qinv_mem_head;
464 	value = (uint32_t *)(sync_mem->qinv_mem_vaddr + index
465 	    * QINV_SYNC_DATA_SIZE);
466 	while (*value == QINV_SYNC_DATA_UNFENCE) {
467 		*value = 0;
468 		(*cnt)++;
469 		sync_mem->qinv_mem_head++;
470 		if (sync_mem->qinv_mem_head == sync_mem->qinv_mem_size) {
471 			sync_mem->qinv_mem_head = 0;
472 			value = (uint32_t *)(sync_mem->qinv_mem_vaddr);
473 		} else
474 			value = (uint32_t *)((char *)value +
475 			    QINV_SYNC_DATA_SIZE);
476 	}
477 
478 	mutex_exit(&sync_mem->qinv_mem_lock);
479 	if ((*cnt) > 0)
480 		return (index);
481 	else
482 		return (-1);
483 }
484 
485 /*
486  * call ddi_dma_mem_alloc to allocate physical contigous
487  * pages for invalidation queue table
488  */
489 static int
490 qinv_setup(immu_t *immu)
491 {
492 	qinv_t *qinv;
493 	size_t size;
494 
495 	ddi_dma_attr_t qinv_dma_attr = {
496 		DMA_ATTR_V0,
497 		0U,
498 		0xffffffffU,
499 		0xffffffffU,
500 		MMU_PAGESIZE, /* page aligned */
501 		0x1,
502 		0x1,
503 		0xffffffffU,
504 		0xffffffffU,
505 		1,
506 		4,
507 		0
508 	};
509 
510 	ddi_device_acc_attr_t qinv_acc_attr = {
511 		DDI_DEVICE_ATTR_V0,
512 		DDI_NEVERSWAP_ACC,
513 		DDI_STRICTORDER_ACC
514 	};
515 
516 	mutex_init(&(immu->immu_qinv_lock), NULL, MUTEX_DRIVER, NULL);
517 
518 
519 	mutex_enter(&(immu->immu_qinv_lock));
520 
521 	immu->immu_qinv = NULL;
522 	if (!IMMU_ECAP_GET_QI(immu->immu_regs_excap) ||
523 	    immu_qinv_enable == B_FALSE) {
524 		mutex_exit(&(immu->immu_qinv_lock));
525 		return (DDI_SUCCESS);
526 	}
527 
528 	if (qinv_iqa_qs > QINV_MAX_QUEUE_SIZE)
529 		qinv_iqa_qs = QINV_MAX_QUEUE_SIZE;
530 
531 	qinv = kmem_zalloc(sizeof (qinv_t), KM_SLEEP);
532 
533 	if (ddi_dma_alloc_handle(root_devinfo,
534 	    &qinv_dma_attr, DDI_DMA_SLEEP, NULL,
535 	    &(qinv->qinv_table.qinv_mem_dma_hdl)) != DDI_SUCCESS) {
536 		ddi_err(DER_WARN, root_devinfo,
537 		    "alloc invalidation queue table handler failed");
538 		goto queue_table_handle_failed;
539 	}
540 
541 	if (ddi_dma_alloc_handle(root_devinfo,
542 	    &qinv_dma_attr, DDI_DMA_SLEEP, NULL,
543 	    &(qinv->qinv_sync.qinv_mem_dma_hdl)) != DDI_SUCCESS) {
544 		ddi_err(DER_WARN, root_devinfo,
545 		    "alloc invalidation queue sync mem handler failed");
546 		goto sync_table_handle_failed;
547 	}
548 
549 	qinv->qinv_table.qinv_mem_size = (1 << (qinv_iqa_qs + 8));
550 	size = qinv->qinv_table.qinv_mem_size * QINV_ENTRY_SIZE;
551 
552 	/* alloc physical contiguous pages for invalidation queue */
553 	if (ddi_dma_mem_alloc(qinv->qinv_table.qinv_mem_dma_hdl,
554 	    size,
555 	    &qinv_acc_attr,
556 	    DDI_DMA_CONSISTENT | IOMEM_DATA_UNCACHED,
557 	    DDI_DMA_SLEEP,
558 	    NULL,
559 	    &(qinv->qinv_table.qinv_mem_vaddr),
560 	    &size,
561 	    &(qinv->qinv_table.qinv_mem_acc_hdl)) != DDI_SUCCESS) {
562 		ddi_err(DER_WARN, root_devinfo,
563 		    "alloc invalidation queue table failed");
564 		goto queue_table_mem_failed;
565 	}
566 
567 	ASSERT(!((uintptr_t)qinv->qinv_table.qinv_mem_vaddr & MMU_PAGEOFFSET));
568 	bzero(qinv->qinv_table.qinv_mem_vaddr, size);
569 
570 	/* get the base physical address of invalidation request queue */
571 	qinv->qinv_table.qinv_mem_paddr = pfn_to_pa(
572 	    hat_getpfnum(kas.a_hat, qinv->qinv_table.qinv_mem_vaddr));
573 
574 	qinv->qinv_table.qinv_mem_head = qinv->qinv_table.qinv_mem_tail = 0;
575 
576 	qinv->qinv_sync.qinv_mem_size = qinv->qinv_table.qinv_mem_size;
577 	size = qinv->qinv_sync.qinv_mem_size * QINV_SYNC_DATA_SIZE;
578 
579 	/* alloc status memory for invalidation wait descriptor */
580 	if (ddi_dma_mem_alloc(qinv->qinv_sync.qinv_mem_dma_hdl,
581 	    size,
582 	    &qinv_acc_attr,
583 	    DDI_DMA_CONSISTENT | IOMEM_DATA_UNCACHED,
584 	    DDI_DMA_SLEEP,
585 	    NULL,
586 	    &(qinv->qinv_sync.qinv_mem_vaddr),
587 	    &size,
588 	    &(qinv->qinv_sync.qinv_mem_acc_hdl)) != DDI_SUCCESS) {
589 		ddi_err(DER_WARN, root_devinfo,
590 		    "alloc invalidation queue sync mem failed");
591 		goto sync_table_mem_failed;
592 	}
593 
594 	ASSERT(!((uintptr_t)qinv->qinv_sync.qinv_mem_vaddr & MMU_PAGEOFFSET));
595 	bzero(qinv->qinv_sync.qinv_mem_vaddr, size);
596 	qinv->qinv_sync.qinv_mem_paddr = pfn_to_pa(
597 	    hat_getpfnum(kas.a_hat, qinv->qinv_sync.qinv_mem_vaddr));
598 
599 	qinv->qinv_sync.qinv_mem_head = qinv->qinv_sync.qinv_mem_tail = 0;
600 
601 	mutex_init(&(qinv->qinv_table.qinv_mem_lock), NULL, MUTEX_DRIVER, NULL);
602 	mutex_init(&(qinv->qinv_sync.qinv_mem_lock), NULL, MUTEX_DRIVER, NULL);
603 
604 	/*
605 	 * init iotlb pend node for submitting invalidation iotlb
606 	 * queue request
607 	 */
608 	qinv->qinv_iotlb_pend_node = (qinv_iotlb_pend_node_t **)
609 	    kmem_zalloc(qinv->qinv_sync.qinv_mem_size
610 	    * sizeof (qinv_iotlb_pend_node_t *), KM_SLEEP);
611 
612 	/* set invalidation queue structure */
613 	immu->immu_qinv = qinv;
614 
615 	mutex_exit(&(immu->immu_qinv_lock));
616 
617 	return (DDI_SUCCESS);
618 
619 sync_table_mem_failed:
620 	ddi_dma_mem_free(&(qinv->qinv_table.qinv_mem_acc_hdl));
621 
622 queue_table_mem_failed:
623 	ddi_dma_free_handle(&(qinv->qinv_sync.qinv_mem_dma_hdl));
624 
625 sync_table_handle_failed:
626 	ddi_dma_free_handle(&(qinv->qinv_table.qinv_mem_dma_hdl));
627 
628 queue_table_handle_failed:
629 	kmem_free(qinv, sizeof (qinv_t));
630 
631 	mutex_exit(&(immu->immu_qinv_lock));
632 
633 	return (DDI_FAILURE);
634 }
635 
636 /*
637  * ###########################################################################
638  *
639  * Functions exported by immu_qinv.c
640  *
641  * ###########################################################################
642  */
643 
644 /*
645  * initialize invalidation request queue structure.
646  */
647 void
648 immu_qinv_setup(list_t *listp)
649 {
650 	immu_t *immu;
651 
652 	if (immu_qinv_enable == B_FALSE) {
653 		return;
654 	}
655 
656 	immu = list_head(listp);
657 	for (; immu; immu = list_next(listp, immu)) {
658 		if (qinv_setup(immu) == DDI_SUCCESS) {
659 			immu->immu_qinv_setup = B_TRUE;
660 		}
661 	}
662 }
663 
664 void
665 immu_qinv_startup(immu_t *immu)
666 {
667 	qinv_t *qinv;
668 	uint64_t qinv_reg_value;
669 
670 	if (immu->immu_qinv_setup == B_FALSE) {
671 		return;
672 	}
673 
674 	qinv = (qinv_t *)immu->immu_qinv;
675 	qinv_reg_value = qinv->qinv_table.qinv_mem_paddr | qinv_iqa_qs;
676 	immu_regs_qinv_enable(immu, qinv_reg_value);
677 	immu->immu_qinv_running = B_TRUE;
678 }
679 
680 /*
681  * queued invalidation interface
682  *   function based context cache invalidation
683  */
684 void
685 immu_qinv_context_fsi(immu_t *immu, uint8_t function_mask,
686     uint16_t source_id, uint_t domain_id)
687 {
688 	qinv_context_common(immu, function_mask, source_id,
689 	    domain_id, CTT_INV_G_DEVICE);
690 	qinv_wait_sync(immu);
691 }
692 
693 /*
694  * queued invalidation interface
695  *   domain based context cache invalidation
696  */
697 void
698 immu_qinv_context_dsi(immu_t *immu, uint_t domain_id)
699 {
700 	qinv_context_common(immu, 0, 0, domain_id, CTT_INV_G_DOMAIN);
701 	qinv_wait_sync(immu);
702 }
703 
704 /*
705  * queued invalidation interface
706  *   invalidation global context cache
707  */
708 void
709 immu_qinv_context_gbl(immu_t *immu)
710 {
711 	qinv_context_common(immu, 0, 0, 0, CTT_INV_G_GLOBAL);
712 	qinv_wait_sync(immu);
713 }
714 
715 /*
716  * queued invalidation interface
717  *   paged based iotlb invalidation
718  */
719 void
720 immu_inv_iotlb_psi(immu_t *immu, uint_t domain_id,
721 	uint64_t dvma, uint_t count, uint_t hint)
722 {
723 	uint_t am = 0;
724 	uint_t max_am;
725 
726 	max_am = IMMU_CAP_GET_MAMV(immu->immu_regs_cap);
727 
728 	/* choose page specified invalidation */
729 	if (IMMU_CAP_GET_PSI(immu->immu_regs_cap)) {
730 		while (am <= max_am) {
731 			if ((ADDR_AM_OFFSET(IMMU_BTOP(dvma), am) + count)
732 			    <= ADDR_AM_MAX(am)) {
733 				qinv_iotlb_common(immu, domain_id,
734 				    dvma, am, hint, TLB_INV_G_PAGE);
735 				break;
736 			}
737 			am++;
738 		}
739 		if (am > max_am) {
740 			qinv_iotlb_common(immu, domain_id,
741 			    dvma, 0, hint, TLB_INV_G_DOMAIN);
742 		}
743 
744 	/* choose domain invalidation */
745 	} else {
746 		qinv_iotlb_common(immu, domain_id, dvma,
747 		    0, hint, TLB_INV_G_DOMAIN);
748 	}
749 }
750 
751 /*
752  * queued invalidation interface
753  *   domain based iotlb invalidation
754  */
755 void
756 immu_qinv_iotlb_dsi(immu_t *immu, uint_t domain_id)
757 {
758 	qinv_iotlb_common(immu, domain_id, 0, 0, 0, TLB_INV_G_DOMAIN);
759 	qinv_wait_sync(immu);
760 }
761 
762 /*
763  * queued invalidation interface
764  *    global iotlb invalidation
765  */
766 void
767 immu_qinv_iotlb_gbl(immu_t *immu)
768 {
769 	qinv_iotlb_common(immu, 0, 0, 0, 0, TLB_INV_G_GLOBAL);
770 	qinv_wait_sync(immu);
771 }
772 
773 
774 
775 /*
776  * the plant wait operation for queued invalidation interface
777  */
778 void
779 immu_qinv_plant(immu_t *immu, dvcookie_t *dvcookies,
780 	uint_t count, uint_t array_size)
781 {
782 	qinv_t *qinv;
783 	qinv_iotlb_pend_node_t *node = NULL;
784 	qinv_iotlb_pend_head_t *head;
785 
786 	qinv = (qinv_t *)immu->immu_qinv;
787 
788 	head = &(qinv->qinv_pend_head);
789 	mutex_enter(&(head->ich_mem_lock));
790 	node = list_head(&(head->ich_mem_list));
791 	if (node) {
792 		list_remove(&(head->ich_mem_list), node);
793 	}
794 	mutex_exit(&(head->ich_mem_lock));
795 
796 	/* no cache, alloc one */
797 	if (node == NULL) {
798 		node = kmem_zalloc(sizeof (qinv_iotlb_pend_node_t), KM_SLEEP);
799 	}
800 	node->icn_dvcookies = dvcookies;
801 	node->icn_count = count;
802 	node->icn_array_size = array_size;
803 
804 	/* plant an invalidation wait descriptor, not wait its completion */
805 	qinv_wait_async_unfence(immu, node);
806 }
807 
808 /*
809  * the reap wait operation for queued invalidation interface
810  */
811 void
812 immu_qinv_reap(immu_t *immu)
813 {
814 	int index, cnt = 0;
815 	qinv_iotlb_pend_node_t *node;
816 	qinv_iotlb_pend_head_t *head;
817 	qinv_t *qinv;
818 
819 	qinv = (qinv_t *)immu->immu_qinv;
820 	head = &(qinv->qinv_pend_head);
821 
822 	index = qinv_wait_async_finish(immu, &cnt);
823 
824 	while (cnt--) {
825 		node = qinv->qinv_iotlb_pend_node[index];
826 		if (node == NULL)
827 			continue;
828 		mutex_enter(&(head->ich_mem_lock));
829 		list_insert_head(&(head->ich_mem_list), node);
830 		mutex_exit(&(head->ich_mem_lock));
831 		qinv->qinv_iotlb_pend_node[index] = NULL;
832 		index++;
833 		if (index == qinv->qinv_sync.qinv_mem_size)
834 			index = 0;
835 	}
836 }
837 
838 
839 /* queued invalidation interface -- global invalidate interrupt entry cache */
840 void
841 immu_qinv_intr_global(immu_t *immu)
842 {
843 	qinv_iec_common(immu, 0, 0, IEC_INV_GLOBAL);
844 	qinv_wait_sync(immu);
845 }
846 
847 /* queued invalidation interface -- invalidate single interrupt entry cache */
848 void
849 immu_qinv_intr_one_cache(immu_t *immu, uint_t iidx)
850 {
851 	qinv_iec_common(immu, iidx, 0, IEC_INV_INDEX);
852 	qinv_wait_sync(immu);
853 }
854 
855 /* queued invalidation interface -- invalidate interrupt entry caches */
856 void
857 immu_qinv_intr_caches(immu_t *immu, uint_t iidx, uint_t cnt)
858 {
859 	uint_t	i, mask = 0;
860 
861 	ASSERT(cnt != 0);
862 
863 	/* requested interrupt count is not a power of 2 */
864 	if (!ISP2(cnt)) {
865 		for (i = 0; i < cnt; i++) {
866 			qinv_iec_common(immu, iidx + cnt, 0, IEC_INV_INDEX);
867 		}
868 		qinv_wait_sync(immu);
869 		return;
870 	}
871 
872 	while ((2 << mask) < cnt) {
873 		mask++;
874 	}
875 
876 	if (mask > IMMU_ECAP_GET_MHMV(immu->immu_regs_excap)) {
877 		for (i = 0; i < cnt; i++) {
878 			qinv_iec_common(immu, iidx + cnt, 0, IEC_INV_INDEX);
879 		}
880 		qinv_wait_sync(immu);
881 		return;
882 	}
883 
884 	qinv_iec_common(immu, iidx, mask, IEC_INV_INDEX);
885 
886 	qinv_wait_sync(immu);
887 }
888 
889 void
890 immu_qinv_report_fault(immu_t *immu)
891 {
892 	uint16_t head;
893 	qinv_dsc_t *dsc;
894 	qinv_t *qinv;
895 
896 	/* access qinv data */
897 	mutex_enter(&(immu->immu_qinv_lock));
898 
899 	qinv = (qinv_t *)(immu->immu_qinv);
900 
901 	head = QINV_IQA_HEAD(
902 	    immu_regs_get64(immu, IMMU_REG_INVAL_QH));
903 
904 	dsc = (qinv_dsc_t *)(qinv->qinv_table.qinv_mem_vaddr
905 	    + (head * QINV_ENTRY_SIZE));
906 
907 	/* report the error */
908 	ddi_err(DER_WARN, immu->immu_dip,
909 	    "generated a fault when fetching a descriptor from the"
910 	    "\tinvalidation queue, or detects that the fetched"
911 	    "\tdescriptor is invalid. The head register is "
912 	    "0x%" PRIx64
913 	    "\tthe type is %s",
914 	    head,
915 	    qinv_dsc_type[MIN(INV_DSC_TYPE(dsc), QINV_MAX_DSC_TYPE)]);
916 
917 	mutex_exit(&(immu->immu_qinv_lock));
918 }
919