xref: /freebsd/sys/x86/iommu/intel_qi.c (revision 076b94438c7d42c1b4661ed1e12e3b12ca69361a)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2013 The FreeBSD Foundation
5  * All rights reserved.
6  *
7  * This software was developed by Konstantin Belousov <kib@FreeBSD.org>
8  * under sponsorship from the FreeBSD Foundation.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29  * SUCH DAMAGE.
30  */
31 
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 #include "opt_acpi.h"
36 
37 #include <sys/param.h>
38 #include <sys/bus.h>
39 #include <sys/kernel.h>
40 #include <sys/malloc.h>
41 #include <sys/memdesc.h>
42 #include <sys/module.h>
43 #include <sys/rman.h>
44 #include <sys/taskqueue.h>
45 #include <sys/time.h>
46 #include <sys/tree.h>
47 #include <sys/vmem.h>
48 #include <machine/bus.h>
49 #include <contrib/dev/acpica/include/acpi.h>
50 #include <contrib/dev/acpica/include/accommon.h>
51 #include <dev/acpica/acpivar.h>
52 #include <vm/vm.h>
53 #include <vm/vm_extern.h>
54 #include <vm/vm_kern.h>
55 #include <vm/vm_page.h>
56 #include <vm/vm_map.h>
57 #include <machine/cpu.h>
58 #include <x86/include/busdma_impl.h>
59 #include <x86/iommu/intel_reg.h>
60 #include <x86/iommu/busdma_dmar.h>
61 #include <x86/iommu/intel_dmar.h>
62 
63 static bool
64 dmar_qi_seq_processed(const struct dmar_unit *unit,
65     const struct dmar_qi_genseq *pseq)
66 {
67 
68 	return (pseq->gen < unit->inv_waitd_gen ||
69 	    (pseq->gen == unit->inv_waitd_gen &&
70 	     pseq->seq <= unit->inv_waitd_seq_hw));
71 }
72 
73 static int
74 dmar_enable_qi(struct dmar_unit *unit)
75 {
76 	int error;
77 
78 	DMAR_ASSERT_LOCKED(unit);
79 	unit->hw_gcmd |= DMAR_GCMD_QIE;
80 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd);
81 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_QIES)
82 	    != 0));
83 	return (error);
84 }
85 
86 static int
87 dmar_disable_qi(struct dmar_unit *unit)
88 {
89 	int error;
90 
91 	DMAR_ASSERT_LOCKED(unit);
92 	unit->hw_gcmd &= ~DMAR_GCMD_QIE;
93 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd);
94 	DMAR_WAIT_UNTIL(((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_QIES)
95 	    == 0));
96 	return (error);
97 }
98 
99 static void
100 dmar_qi_advance_tail(struct dmar_unit *unit)
101 {
102 
103 	DMAR_ASSERT_LOCKED(unit);
104 	dmar_write4(unit, DMAR_IQT_REG, unit->inv_queue_tail);
105 }
106 
107 static void
108 dmar_qi_ensure(struct dmar_unit *unit, int descr_count)
109 {
110 	uint32_t head;
111 	int bytes;
112 
113 	DMAR_ASSERT_LOCKED(unit);
114 	bytes = descr_count << DMAR_IQ_DESCR_SZ_SHIFT;
115 	for (;;) {
116 		if (bytes <= unit->inv_queue_avail)
117 			break;
118 		/* refill */
119 		head = dmar_read4(unit, DMAR_IQH_REG);
120 		head &= DMAR_IQH_MASK;
121 		unit->inv_queue_avail = head - unit->inv_queue_tail -
122 		    DMAR_IQ_DESCR_SZ;
123 		if (head <= unit->inv_queue_tail)
124 			unit->inv_queue_avail += unit->inv_queue_size;
125 		if (bytes <= unit->inv_queue_avail)
126 			break;
127 
128 		/*
129 		 * No space in the queue, do busy wait.  Hardware must
130 		 * make a progress.  But first advance the tail to
131 		 * inform the descriptor streamer about entries we
132 		 * might have already filled, otherwise they could
133 		 * clog the whole queue..
134 		 */
135 		dmar_qi_advance_tail(unit);
136 		unit->inv_queue_full++;
137 		cpu_spinwait();
138 	}
139 	unit->inv_queue_avail -= bytes;
140 }
141 
142 static void
143 dmar_qi_emit(struct dmar_unit *unit, uint64_t data1, uint64_t data2)
144 {
145 
146 	DMAR_ASSERT_LOCKED(unit);
147 	*(volatile uint64_t *)(unit->inv_queue + unit->inv_queue_tail) = data1;
148 	unit->inv_queue_tail += DMAR_IQ_DESCR_SZ / 2;
149 	KASSERT(unit->inv_queue_tail <= unit->inv_queue_size,
150 	    ("tail overflow 0x%x 0x%jx", unit->inv_queue_tail,
151 	    (uintmax_t)unit->inv_queue_size));
152 	unit->inv_queue_tail &= unit->inv_queue_size - 1;
153 	*(volatile uint64_t *)(unit->inv_queue + unit->inv_queue_tail) = data2;
154 	unit->inv_queue_tail += DMAR_IQ_DESCR_SZ / 2;
155 	KASSERT(unit->inv_queue_tail <= unit->inv_queue_size,
156 	    ("tail overflow 0x%x 0x%jx", unit->inv_queue_tail,
157 	    (uintmax_t)unit->inv_queue_size));
158 	unit->inv_queue_tail &= unit->inv_queue_size - 1;
159 }
160 
161 static void
162 dmar_qi_emit_wait_descr(struct dmar_unit *unit, uint32_t seq, bool intr,
163     bool memw, bool fence)
164 {
165 
166 	DMAR_ASSERT_LOCKED(unit);
167 	dmar_qi_emit(unit, DMAR_IQ_DESCR_WAIT_ID |
168 	    (intr ? DMAR_IQ_DESCR_WAIT_IF : 0) |
169 	    (memw ? DMAR_IQ_DESCR_WAIT_SW : 0) |
170 	    (fence ? DMAR_IQ_DESCR_WAIT_FN : 0) |
171 	    (memw ? DMAR_IQ_DESCR_WAIT_SD(seq) : 0),
172 	    memw ? unit->inv_waitd_seq_hw_phys : 0);
173 }
174 
175 static void
176 dmar_qi_emit_wait_seq(struct dmar_unit *unit, struct dmar_qi_genseq *pseq,
177     bool emit_wait)
178 {
179 	struct dmar_qi_genseq gsec;
180 	uint32_t seq;
181 
182 	KASSERT(pseq != NULL, ("wait descriptor with no place for seq"));
183 	DMAR_ASSERT_LOCKED(unit);
184 	if (unit->inv_waitd_seq == 0xffffffff) {
185 		gsec.gen = unit->inv_waitd_gen;
186 		gsec.seq = unit->inv_waitd_seq;
187 		dmar_qi_ensure(unit, 1);
188 		dmar_qi_emit_wait_descr(unit, gsec.seq, false, true, false);
189 		dmar_qi_advance_tail(unit);
190 		while (!dmar_qi_seq_processed(unit, &gsec))
191 			cpu_spinwait();
192 		unit->inv_waitd_gen++;
193 		unit->inv_waitd_seq = 1;
194 	}
195 	seq = unit->inv_waitd_seq++;
196 	pseq->gen = unit->inv_waitd_gen;
197 	pseq->seq = seq;
198 	if (emit_wait) {
199 		dmar_qi_ensure(unit, 1);
200 		dmar_qi_emit_wait_descr(unit, seq, true, true, false);
201 	}
202 }
203 
204 static void
205 dmar_qi_wait_for_seq(struct dmar_unit *unit, const struct dmar_qi_genseq *gseq,
206     bool nowait)
207 {
208 
209 	DMAR_ASSERT_LOCKED(unit);
210 	unit->inv_seq_waiters++;
211 	while (!dmar_qi_seq_processed(unit, gseq)) {
212 		if (cold || nowait) {
213 			cpu_spinwait();
214 		} else {
215 			msleep(&unit->inv_seq_waiters, &unit->lock, 0,
216 			    "dmarse", hz);
217 		}
218 	}
219 	unit->inv_seq_waiters--;
220 }
221 
222 void
223 dmar_qi_invalidate_locked(struct dmar_domain *domain, dmar_gaddr_t base,
224     dmar_gaddr_t size, struct dmar_qi_genseq *pseq, bool emit_wait)
225 {
226 	struct dmar_unit *unit;
227 	dmar_gaddr_t isize;
228 	int am;
229 
230 	unit = domain->dmar;
231 	DMAR_ASSERT_LOCKED(unit);
232 	for (; size > 0; base += isize, size -= isize) {
233 		am = calc_am(unit, base, size, &isize);
234 		dmar_qi_ensure(unit, 1);
235 		dmar_qi_emit(unit, DMAR_IQ_DESCR_IOTLB_INV |
236 		    DMAR_IQ_DESCR_IOTLB_PAGE | DMAR_IQ_DESCR_IOTLB_DW |
237 		    DMAR_IQ_DESCR_IOTLB_DR |
238 		    DMAR_IQ_DESCR_IOTLB_DID(domain->domain),
239 		    base | am);
240 	}
241 	dmar_qi_emit_wait_seq(unit, pseq, emit_wait);
242 	dmar_qi_advance_tail(unit);
243 }
244 
245 void
246 dmar_qi_invalidate_ctx_glob_locked(struct dmar_unit *unit)
247 {
248 	struct dmar_qi_genseq gseq;
249 
250 	DMAR_ASSERT_LOCKED(unit);
251 	dmar_qi_ensure(unit, 2);
252 	dmar_qi_emit(unit, DMAR_IQ_DESCR_CTX_INV | DMAR_IQ_DESCR_CTX_GLOB, 0);
253 	dmar_qi_emit_wait_seq(unit, &gseq, true);
254 	dmar_qi_advance_tail(unit);
255 	dmar_qi_wait_for_seq(unit, &gseq, false);
256 }
257 
258 void
259 dmar_qi_invalidate_iotlb_glob_locked(struct dmar_unit *unit)
260 {
261 	struct dmar_qi_genseq gseq;
262 
263 	DMAR_ASSERT_LOCKED(unit);
264 	dmar_qi_ensure(unit, 2);
265 	dmar_qi_emit(unit, DMAR_IQ_DESCR_IOTLB_INV | DMAR_IQ_DESCR_IOTLB_GLOB |
266 	    DMAR_IQ_DESCR_IOTLB_DW | DMAR_IQ_DESCR_IOTLB_DR, 0);
267 	dmar_qi_emit_wait_seq(unit, &gseq, true);
268 	dmar_qi_advance_tail(unit);
269 	dmar_qi_wait_for_seq(unit, &gseq, false);
270 }
271 
272 void
273 dmar_qi_invalidate_iec_glob(struct dmar_unit *unit)
274 {
275 	struct dmar_qi_genseq gseq;
276 
277 	DMAR_ASSERT_LOCKED(unit);
278 	dmar_qi_ensure(unit, 2);
279 	dmar_qi_emit(unit, DMAR_IQ_DESCR_IEC_INV, 0);
280 	dmar_qi_emit_wait_seq(unit, &gseq, true);
281 	dmar_qi_advance_tail(unit);
282 	dmar_qi_wait_for_seq(unit, &gseq, false);
283 }
284 
285 void
286 dmar_qi_invalidate_iec(struct dmar_unit *unit, u_int start, u_int cnt)
287 {
288 	struct dmar_qi_genseq gseq;
289 	u_int c, l;
290 
291 	DMAR_ASSERT_LOCKED(unit);
292 	KASSERT(start < unit->irte_cnt && start < start + cnt &&
293 	    start + cnt <= unit->irte_cnt,
294 	    ("inv iec overflow %d %d %d", unit->irte_cnt, start, cnt));
295 	for (; cnt > 0; cnt -= c, start += c) {
296 		l = ffs(start | cnt) - 1;
297 		c = 1 << l;
298 		dmar_qi_ensure(unit, 1);
299 		dmar_qi_emit(unit, DMAR_IQ_DESCR_IEC_INV |
300 		    DMAR_IQ_DESCR_IEC_IDX | DMAR_IQ_DESCR_IEC_IIDX(start) |
301 		    DMAR_IQ_DESCR_IEC_IM(l), 0);
302 	}
303 	dmar_qi_ensure(unit, 1);
304 	dmar_qi_emit_wait_seq(unit, &gseq, true);
305 	dmar_qi_advance_tail(unit);
306 
307 	/*
308 	 * The caller of the function, in particular,
309 	 * dmar_ir_program_irte(), may be called from the context
310 	 * where the sleeping is forbidden (in fact, the
311 	 * intr_table_lock mutex may be held, locked from
312 	 * intr_shuffle_irqs()).  Wait for the invalidation completion
313 	 * using the busy wait.
314 	 *
315 	 * The impact on the interrupt input setup code is small, the
316 	 * expected overhead is comparable with the chipset register
317 	 * read.  It is more harmful for the parallel DMA operations,
318 	 * since we own the dmar unit lock until whole invalidation
319 	 * queue is processed, which includes requests possibly issued
320 	 * before our request.
321 	 */
322 	dmar_qi_wait_for_seq(unit, &gseq, true);
323 }
324 
325 int
326 dmar_qi_intr(void *arg)
327 {
328 	struct dmar_unit *unit;
329 
330 	unit = arg;
331 	KASSERT(unit->qi_enabled, ("dmar%d: QI is not enabled", unit->unit));
332 	taskqueue_enqueue(unit->qi_taskqueue, &unit->qi_task);
333 	return (FILTER_HANDLED);
334 }
335 
336 static void
337 dmar_qi_task(void *arg, int pending __unused)
338 {
339 	struct dmar_unit *unit;
340 	struct dmar_map_entry *entry;
341 	uint32_t ics;
342 
343 	unit = arg;
344 
345 	DMAR_LOCK(unit);
346 	for (;;) {
347 		entry = TAILQ_FIRST(&unit->tlb_flush_entries);
348 		if (entry == NULL)
349 			break;
350 		if (!dmar_qi_seq_processed(unit, &entry->gseq))
351 			break;
352 		TAILQ_REMOVE(&unit->tlb_flush_entries, entry, dmamap_link);
353 		DMAR_UNLOCK(unit);
354 		dmar_domain_free_entry(entry, (entry->flags &
355 		    DMAR_MAP_ENTRY_QI_NF) == 0);
356 		DMAR_LOCK(unit);
357 	}
358 	ics = dmar_read4(unit, DMAR_ICS_REG);
359 	if ((ics & DMAR_ICS_IWC) != 0) {
360 		ics = DMAR_ICS_IWC;
361 		dmar_write4(unit, DMAR_ICS_REG, ics);
362 	}
363 	if (unit->inv_seq_waiters > 0)
364 		wakeup(&unit->inv_seq_waiters);
365 	DMAR_UNLOCK(unit);
366 }
367 
368 int
369 dmar_init_qi(struct dmar_unit *unit)
370 {
371 	uint64_t iqa;
372 	uint32_t ics;
373 	int qi_sz;
374 
375 	if (!DMAR_HAS_QI(unit) || (unit->hw_cap & DMAR_CAP_CM) != 0)
376 		return (0);
377 	unit->qi_enabled = 1;
378 	TUNABLE_INT_FETCH("hw.dmar.qi", &unit->qi_enabled);
379 	if (!unit->qi_enabled)
380 		return (0);
381 
382 	TAILQ_INIT(&unit->tlb_flush_entries);
383 	TASK_INIT(&unit->qi_task, 0, dmar_qi_task, unit);
384 	unit->qi_taskqueue = taskqueue_create_fast("dmarqf", M_WAITOK,
385 	    taskqueue_thread_enqueue, &unit->qi_taskqueue);
386 	taskqueue_start_threads(&unit->qi_taskqueue, 1, PI_AV,
387 	    "dmar%d qi taskq", unit->unit);
388 
389 	unit->inv_waitd_gen = 0;
390 	unit->inv_waitd_seq = 1;
391 
392 	qi_sz = DMAR_IQA_QS_DEF;
393 	TUNABLE_INT_FETCH("hw.dmar.qi_size", &qi_sz);
394 	if (qi_sz > DMAR_IQA_QS_MAX)
395 		qi_sz = DMAR_IQA_QS_MAX;
396 	unit->inv_queue_size = (1ULL << qi_sz) * PAGE_SIZE;
397 	/* Reserve one descriptor to prevent wraparound. */
398 	unit->inv_queue_avail = unit->inv_queue_size - DMAR_IQ_DESCR_SZ;
399 
400 	/* The invalidation queue reads by DMARs are always coherent. */
401 	unit->inv_queue = kmem_alloc_contig(unit->inv_queue_size, M_WAITOK |
402 	    M_ZERO, 0, dmar_high, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT);
403 	unit->inv_waitd_seq_hw_phys = pmap_kextract(
404 	    (vm_offset_t)&unit->inv_waitd_seq_hw);
405 
406 	DMAR_LOCK(unit);
407 	dmar_write8(unit, DMAR_IQT_REG, 0);
408 	iqa = pmap_kextract(unit->inv_queue);
409 	iqa |= qi_sz;
410 	dmar_write8(unit, DMAR_IQA_REG, iqa);
411 	dmar_enable_qi(unit);
412 	ics = dmar_read4(unit, DMAR_ICS_REG);
413 	if ((ics & DMAR_ICS_IWC) != 0) {
414 		ics = DMAR_ICS_IWC;
415 		dmar_write4(unit, DMAR_ICS_REG, ics);
416 	}
417 	dmar_enable_qi_intr(unit);
418 	DMAR_UNLOCK(unit);
419 
420 	return (0);
421 }
422 
423 void
424 dmar_fini_qi(struct dmar_unit *unit)
425 {
426 	struct dmar_qi_genseq gseq;
427 
428 	if (unit->qi_enabled)
429 		return;
430 	taskqueue_drain(unit->qi_taskqueue, &unit->qi_task);
431 	taskqueue_free(unit->qi_taskqueue);
432 	unit->qi_taskqueue = NULL;
433 
434 	DMAR_LOCK(unit);
435 	/* quisce */
436 	dmar_qi_ensure(unit, 1);
437 	dmar_qi_emit_wait_seq(unit, &gseq, true);
438 	dmar_qi_advance_tail(unit);
439 	dmar_qi_wait_for_seq(unit, &gseq, false);
440 	/* only after the quisce, disable queue */
441 	dmar_disable_qi_intr(unit);
442 	dmar_disable_qi(unit);
443 	KASSERT(unit->inv_seq_waiters == 0,
444 	    ("dmar%d: waiters on disabled queue", unit->unit));
445 	DMAR_UNLOCK(unit);
446 
447 	kmem_free(unit->inv_queue, unit->inv_queue_size);
448 	unit->inv_queue = 0;
449 	unit->inv_queue_size = 0;
450 	unit->qi_enabled = 0;
451 }
452 
453 void
454 dmar_enable_qi_intr(struct dmar_unit *unit)
455 {
456 	uint32_t iectl;
457 
458 	DMAR_ASSERT_LOCKED(unit);
459 	KASSERT(DMAR_HAS_QI(unit), ("dmar%d: QI is not supported", unit->unit));
460 	iectl = dmar_read4(unit, DMAR_IECTL_REG);
461 	iectl &= ~DMAR_IECTL_IM;
462 	dmar_write4(unit, DMAR_IECTL_REG, iectl);
463 }
464 
465 void
466 dmar_disable_qi_intr(struct dmar_unit *unit)
467 {
468 	uint32_t iectl;
469 
470 	DMAR_ASSERT_LOCKED(unit);
471 	KASSERT(DMAR_HAS_QI(unit), ("dmar%d: QI is not supported", unit->unit));
472 	iectl = dmar_read4(unit, DMAR_IECTL_REG);
473 	dmar_write4(unit, DMAR_IECTL_REG, iectl | DMAR_IECTL_IM);
474 }
475