xref: /freebsd/sys/x86/iommu/intel_qi.c (revision fba3cde907930eed2adb8a320524bc250338c729)
1 /*-
2  * Copyright (c) 2013 The FreeBSD Foundation
3  * All rights reserved.
4  *
5  * This software was developed by Konstantin Belousov <kib@FreeBSD.org>
6  * under sponsorship from the FreeBSD Foundation.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include <sys/cdefs.h>
31 __FBSDID("$FreeBSD$");
32 
33 #include "opt_acpi.h"
34 
35 #include <sys/param.h>
36 #include <sys/bus.h>
37 #include <sys/kernel.h>
38 #include <sys/malloc.h>
39 #include <sys/memdesc.h>
40 #include <sys/module.h>
41 #include <sys/rman.h>
42 #include <sys/taskqueue.h>
43 #include <sys/tree.h>
44 #include <machine/bus.h>
45 #include <contrib/dev/acpica/include/acpi.h>
46 #include <contrib/dev/acpica/include/accommon.h>
47 #include <dev/acpica/acpivar.h>
48 #include <vm/vm.h>
49 #include <vm/vm_extern.h>
50 #include <vm/vm_kern.h>
51 #include <vm/vm_page.h>
52 #include <vm/vm_map.h>
53 #include <machine/cpu.h>
54 #include <x86/include/busdma_impl.h>
55 #include <x86/iommu/intel_reg.h>
56 #include <x86/iommu/busdma_dmar.h>
57 #include <x86/iommu/intel_dmar.h>
58 
59 static bool
60 dmar_qi_seq_processed(const struct dmar_unit *unit,
61     const struct dmar_qi_genseq *pseq)
62 {
63 
64 	return (pseq->gen < unit->inv_waitd_gen ||
65 	    (pseq->gen == unit->inv_waitd_gen &&
66 	     pseq->seq <= unit->inv_waitd_seq_hw));
67 }
68 
69 static int
70 dmar_enable_qi(struct dmar_unit *unit)
71 {
72 
73 	DMAR_ASSERT_LOCKED(unit);
74 	unit->hw_gcmd |= DMAR_GCMD_QIE;
75 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd);
76 	/* XXXKIB should have a timeout */
77 	while ((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_QIES) == 0)
78 		cpu_spinwait();
79 	return (0);
80 }
81 
82 static int
83 dmar_disable_qi(struct dmar_unit *unit)
84 {
85 
86 	DMAR_ASSERT_LOCKED(unit);
87 	unit->hw_gcmd &= ~DMAR_GCMD_QIE;
88 	dmar_write4(unit, DMAR_GCMD_REG, unit->hw_gcmd);
89 	/* XXXKIB should have a timeout */
90 	while ((dmar_read4(unit, DMAR_GSTS_REG) & DMAR_GSTS_QIES) != 0)
91 		cpu_spinwait();
92 	return (0);
93 }
94 
95 static void
96 dmar_qi_advance_tail(struct dmar_unit *unit)
97 {
98 
99 	DMAR_ASSERT_LOCKED(unit);
100 	dmar_write4(unit, DMAR_IQT_REG, unit->inv_queue_tail);
101 }
102 
103 static void
104 dmar_qi_ensure(struct dmar_unit *unit, int descr_count)
105 {
106 	uint32_t head;
107 	int bytes;
108 
109 	DMAR_ASSERT_LOCKED(unit);
110 	bytes = descr_count << DMAR_IQ_DESCR_SZ_SHIFT;
111 	for (;;) {
112 		if (bytes <= unit->inv_queue_avail)
113 			break;
114 		/* refill */
115 		head = dmar_read4(unit, DMAR_IQH_REG);
116 		head &= DMAR_IQH_MASK;
117 		unit->inv_queue_avail = head - unit->inv_queue_tail -
118 		    DMAR_IQ_DESCR_SZ;
119 		if (head <= unit->inv_queue_tail)
120 			unit->inv_queue_avail += unit->inv_queue_size;
121 		if (bytes <= unit->inv_queue_avail)
122 			break;
123 
124 		/*
125 		 * No space in the queue, do busy wait.  Hardware must
126 		 * make a progress.  But first advance the tail to
127 		 * inform the descriptor streamer about entries we
128 		 * might have already filled, otherwise they could
129 		 * clog the whole queue..
130 		 */
131 		dmar_qi_advance_tail(unit);
132 		unit->inv_queue_full++;
133 		cpu_spinwait();
134 	}
135 	unit->inv_queue_avail -= bytes;
136 }
137 
138 static void
139 dmar_qi_emit(struct dmar_unit *unit, uint64_t data1, uint64_t data2)
140 {
141 
142 	DMAR_ASSERT_LOCKED(unit);
143 	*(volatile uint64_t *)(unit->inv_queue + unit->inv_queue_tail) = data1;
144 	unit->inv_queue_tail += DMAR_IQ_DESCR_SZ / 2;
145 	KASSERT(unit->inv_queue_tail <= unit->inv_queue_size,
146 	    ("tail overflow 0x%x 0x%jx", unit->inv_queue_tail,
147 	    (uintmax_t)unit->inv_queue_size));
148 	unit->inv_queue_tail &= unit->inv_queue_size - 1;
149 	*(volatile uint64_t *)(unit->inv_queue + unit->inv_queue_tail) = data2;
150 	unit->inv_queue_tail += DMAR_IQ_DESCR_SZ / 2;
151 	KASSERT(unit->inv_queue_tail <= unit->inv_queue_size,
152 	    ("tail overflow 0x%x 0x%jx", unit->inv_queue_tail,
153 	    (uintmax_t)unit->inv_queue_size));
154 	unit->inv_queue_tail &= unit->inv_queue_size - 1;
155 }
156 
157 static void
158 dmar_qi_emit_wait_descr(struct dmar_unit *unit, uint32_t seq, bool intr,
159     bool memw, bool fence)
160 {
161 
162 	DMAR_ASSERT_LOCKED(unit);
163 	dmar_qi_emit(unit, DMAR_IQ_DESCR_WAIT_ID |
164 	    (intr ? DMAR_IQ_DESCR_WAIT_IF : 0) |
165 	    (memw ? DMAR_IQ_DESCR_WAIT_SW : 0) |
166 	    (fence ? DMAR_IQ_DESCR_WAIT_FN : 0) |
167 	    (memw ? DMAR_IQ_DESCR_WAIT_SD(seq) : 0),
168 	    memw ? unit->inv_waitd_seq_hw_phys : 0);
169 }
170 
171 static void
172 dmar_qi_emit_wait_seq(struct dmar_unit *unit, struct dmar_qi_genseq *pseq)
173 {
174 	struct dmar_qi_genseq gsec;
175 	uint32_t seq;
176 
177 	KASSERT(pseq != NULL, ("wait descriptor with no place for seq"));
178 	DMAR_ASSERT_LOCKED(unit);
179 	if (unit->inv_waitd_seq == 0xffffffff) {
180 		gsec.gen = unit->inv_waitd_gen;
181 		gsec.seq = unit->inv_waitd_seq;
182 		dmar_qi_ensure(unit, 1);
183 		dmar_qi_emit_wait_descr(unit, gsec.seq, false, true, false);
184 		dmar_qi_advance_tail(unit);
185 		while (!dmar_qi_seq_processed(unit, &gsec))
186 			cpu_spinwait();
187 		unit->inv_waitd_gen++;
188 		unit->inv_waitd_seq = 1;
189 	}
190 	seq = unit->inv_waitd_seq++;
191 	pseq->gen = unit->inv_waitd_gen;
192 	pseq->seq = seq;
193 	dmar_qi_emit_wait_descr(unit, seq, true, true, false);
194 }
195 
196 static void
197 dmar_qi_wait_for_seq(struct dmar_unit *unit, const struct dmar_qi_genseq *gseq)
198 {
199 
200 	DMAR_ASSERT_LOCKED(unit);
201 	unit->inv_seq_waiters++;
202 	while (!dmar_qi_seq_processed(unit, gseq)) {
203 		if (cold) {
204 			cpu_spinwait();
205 		} else {
206 			msleep(&unit->inv_seq_waiters, &unit->lock, 0,
207 			    "dmarse", hz);
208 		}
209 	}
210 	unit->inv_seq_waiters--;
211 }
212 
213 void
214 dmar_qi_invalidate_locked(struct dmar_ctx *ctx, dmar_gaddr_t base,
215     dmar_gaddr_t size, struct dmar_qi_genseq *pseq)
216 {
217 	struct dmar_unit *unit;
218 	dmar_gaddr_t isize;
219 	int am;
220 
221 	unit = ctx->dmar;
222 	DMAR_ASSERT_LOCKED(unit);
223 	for (; size > 0; base += isize, size -= isize) {
224 		am = calc_am(unit, base, size, &isize);
225 		dmar_qi_ensure(unit, 1);
226 		dmar_qi_emit(unit, DMAR_IQ_DESCR_IOTLB_INV |
227 		    DMAR_IQ_DESCR_IOTLB_PAGE | DMAR_IQ_DESCR_IOTLB_DW |
228 		    DMAR_IQ_DESCR_IOTLB_DR |
229 		    DMAR_IQ_DESCR_IOTLB_DID(ctx->domain),
230 		    base | am);
231 	}
232 	if (pseq != NULL) {
233 		dmar_qi_ensure(unit, 1);
234 		dmar_qi_emit_wait_seq(unit, pseq);
235 	}
236 	dmar_qi_advance_tail(unit);
237 }
238 
239 void
240 dmar_qi_invalidate_ctx_glob_locked(struct dmar_unit *unit)
241 {
242 	struct dmar_qi_genseq gseq;
243 
244 	DMAR_ASSERT_LOCKED(unit);
245 	dmar_qi_ensure(unit, 2);
246 	dmar_qi_emit(unit, DMAR_IQ_DESCR_CTX_INV | DMAR_IQ_DESCR_CTX_GLOB, 0);
247 	dmar_qi_emit_wait_seq(unit, &gseq);
248 	dmar_qi_advance_tail(unit);
249 	dmar_qi_wait_for_seq(unit, &gseq);
250 }
251 
252 void
253 dmar_qi_invalidate_iotlb_glob_locked(struct dmar_unit *unit)
254 {
255 	struct dmar_qi_genseq gseq;
256 
257 	DMAR_ASSERT_LOCKED(unit);
258 	dmar_qi_ensure(unit, 2);
259 	dmar_qi_emit(unit, DMAR_IQ_DESCR_IOTLB_INV | DMAR_IQ_DESCR_IOTLB_GLOB |
260 	    DMAR_IQ_DESCR_IOTLB_DW | DMAR_IQ_DESCR_IOTLB_DR, 0);
261 	dmar_qi_emit_wait_seq(unit, &gseq);
262 	dmar_qi_advance_tail(unit);
263 	dmar_qi_wait_for_seq(unit, &gseq);
264 }
265 
266 int
267 dmar_qi_intr(void *arg)
268 {
269 	struct dmar_unit *unit;
270 
271 	unit = arg;
272 	KASSERT(unit->qi_enabled, ("dmar%d: QI is not enabled", unit->unit));
273 	taskqueue_enqueue_fast(unit->qi_taskqueue, &unit->qi_task);
274 	return (FILTER_HANDLED);
275 }
276 
277 static void
278 dmar_qi_task(void *arg, int pending __unused)
279 {
280 	struct dmar_unit *unit;
281 	struct dmar_map_entry *entry;
282 	uint32_t ics;
283 
284 	unit = arg;
285 
286 	DMAR_LOCK(unit);
287 	for (;;) {
288 		entry = TAILQ_FIRST(&unit->tlb_flush_entries);
289 		if (entry == NULL)
290 			break;
291 		if ((entry->gseq.gen == 0 && entry->gseq.seq == 0) ||
292 		    !dmar_qi_seq_processed(unit, &entry->gseq))
293 			break;
294 		TAILQ_REMOVE(&unit->tlb_flush_entries, entry, dmamap_link);
295 		DMAR_UNLOCK(unit);
296 		dmar_ctx_free_entry(entry, (entry->flags &
297 		    DMAR_MAP_ENTRY_QI_NF) == 0);
298 		DMAR_LOCK(unit);
299 	}
300 	ics = dmar_read4(unit, DMAR_ICS_REG);
301 	if ((ics & DMAR_ICS_IWC) != 0) {
302 		ics = DMAR_ICS_IWC;
303 		dmar_write4(unit, DMAR_ICS_REG, ics);
304 	}
305 	if (unit->inv_seq_waiters > 0)
306 		wakeup(&unit->inv_seq_waiters);
307 	DMAR_UNLOCK(unit);
308 }
309 
310 int
311 dmar_init_qi(struct dmar_unit *unit)
312 {
313 	uint64_t iqa;
314 	uint32_t ics;
315 	int qi_sz;
316 
317 	if (!DMAR_HAS_QI(unit) || (unit->hw_cap & DMAR_CAP_CM) != 0)
318 		return (0);
319 	unit->qi_enabled = 1;
320 	TUNABLE_INT_FETCH("hw.dmar.qi", &unit->qi_enabled);
321 	if (!unit->qi_enabled)
322 		return (0);
323 
324 	TAILQ_INIT(&unit->tlb_flush_entries);
325 	TASK_INIT(&unit->qi_task, 0, dmar_qi_task, unit);
326 	unit->qi_taskqueue = taskqueue_create_fast("dmar", M_WAITOK,
327 	    taskqueue_thread_enqueue, &unit->qi_taskqueue);
328 	taskqueue_start_threads(&unit->qi_taskqueue, 1, PI_AV,
329 	    "dmar%d qi taskq", unit->unit);
330 
331 	unit->inv_waitd_gen = 0;
332 	unit->inv_waitd_seq = 1;
333 
334 	qi_sz = DMAR_IQA_QS_DEF;
335 	TUNABLE_INT_FETCH("hw.dmar.qi_size", &qi_sz);
336 	if (qi_sz > DMAR_IQA_QS_MAX)
337 		qi_sz = DMAR_IQA_QS_MAX;
338 	unit->inv_queue_size = (1ULL << qi_sz) * PAGE_SIZE;
339 	/* Reserve one descriptor to prevent wraparound. */
340 	unit->inv_queue_avail = unit->inv_queue_size - DMAR_IQ_DESCR_SZ;
341 
342 	/* The invalidation queue reads by DMARs are always coherent. */
343 	unit->inv_queue = kmem_alloc_contig(kernel_arena, unit->inv_queue_size,
344 	    M_WAITOK | M_ZERO, 0, dmar_high, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT);
345 	unit->inv_waitd_seq_hw_phys = pmap_kextract(
346 	    (vm_offset_t)&unit->inv_waitd_seq_hw);
347 
348 	DMAR_LOCK(unit);
349 	dmar_write8(unit, DMAR_IQT_REG, 0);
350 	iqa = pmap_kextract(unit->inv_queue);
351 	iqa |= qi_sz;
352 	dmar_write8(unit, DMAR_IQA_REG, iqa);
353 	dmar_enable_qi(unit);
354 	ics = dmar_read4(unit, DMAR_ICS_REG);
355 	if ((ics & DMAR_ICS_IWC) != 0) {
356 		ics = DMAR_ICS_IWC;
357 		dmar_write4(unit, DMAR_ICS_REG, ics);
358 	}
359 	DMAR_UNLOCK(unit);
360 
361 	return (0);
362 }
363 
364 void
365 dmar_fini_qi(struct dmar_unit *unit)
366 {
367 	struct dmar_qi_genseq gseq;
368 
369 	if (unit->qi_enabled)
370 		return;
371 	taskqueue_drain(unit->qi_taskqueue, &unit->qi_task);
372 	taskqueue_free(unit->qi_taskqueue);
373 	unit->qi_taskqueue = NULL;
374 
375 	DMAR_LOCK(unit);
376 	/* quisce */
377 	dmar_qi_ensure(unit, 1);
378 	dmar_qi_emit_wait_seq(unit, &gseq);
379 	dmar_qi_advance_tail(unit);
380 	dmar_qi_wait_for_seq(unit, &gseq);
381 	/* only after the quisce, disable queue */
382 	dmar_disable_qi(unit);
383 	KASSERT(unit->inv_seq_waiters == 0,
384 	    ("dmar%d: waiters on disabled queue", unit->unit));
385 	DMAR_UNLOCK(unit);
386 
387 	kmem_free(kernel_arena, unit->inv_queue, unit->inv_queue_size);
388 	unit->inv_queue = 0;
389 	unit->inv_queue_size = 0;
390 	unit->qi_enabled = 0;
391 }
392 
393 void
394 dmar_enable_qi_intr(struct dmar_unit *unit)
395 {
396 	uint32_t iectl;
397 
398 	DMAR_ASSERT_LOCKED(unit);
399 	KASSERT(DMAR_HAS_QI(unit), ("dmar%d: QI is not supported", unit->unit));
400 	iectl = dmar_read4(unit, DMAR_IECTL_REG);
401 	iectl &= ~DMAR_IECTL_IM;
402 	dmar_write4(unit, DMAR_IECTL_REG, iectl);
403 }
404 
405 void
406 dmar_disable_qi_intr(struct dmar_unit *unit)
407 {
408 	uint32_t iectl;
409 
410 	DMAR_ASSERT_LOCKED(unit);
411 	KASSERT(DMAR_HAS_QI(unit), ("dmar%d: QI is not supported", unit->unit));
412 	iectl = dmar_read4(unit, DMAR_IECTL_REG);
413 	dmar_write4(unit, DMAR_IECTL_REG, iectl | DMAR_IECTL_IM);
414 }
415