xref: /freebsd/sys/x86/iommu/intel_fault.c (revision a90b9d0159070121c221b966469c3e36d912bf82)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2013 The FreeBSD Foundation
5  *
6  * This software was developed by Konstantin Belousov <kib@FreeBSD.org>
7  * under sponsorship from the FreeBSD Foundation.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  */
30 
31 #include "opt_acpi.h"
32 
33 #include <sys/param.h>
34 #include <sys/bus.h>
35 #include <sys/kernel.h>
36 #include <sys/malloc.h>
37 #include <sys/memdesc.h>
38 #include <sys/module.h>
39 #include <sys/rman.h>
40 #include <sys/taskqueue.h>
41 #include <sys/tree.h>
42 #include <sys/vmem.h>
43 #include <machine/bus.h>
44 #include <contrib/dev/acpica/include/acpi.h>
45 #include <contrib/dev/acpica/include/accommon.h>
46 #include <dev/acpica/acpivar.h>
47 #include <dev/pci/pcireg.h>
48 #include <dev/pci/pcivar.h>
49 #include <vm/vm.h>
50 #include <vm/vm_extern.h>
51 #include <vm/vm_kern.h>
52 #include <vm/vm_page.h>
53 #include <vm/vm_map.h>
54 #include <x86/include/busdma_impl.h>
55 #include <x86/iommu/intel_reg.h>
56 #include <dev/iommu/busdma_iommu.h>
57 #include <x86/iommu/x86_iommu.h>
58 #include <x86/iommu/intel_dmar.h>
59 
60 /*
61  * Fault interrupt handling for DMARs.  If advanced fault logging is
62  * not implemented by hardware, the code emulates it.  Fast interrupt
63  * handler flushes the fault registers into circular buffer at
64  * unit->fault_log, and schedules a task.
65  *
66  * The fast handler is used since faults usually come in bursts, and
67  * number of fault log registers is limited, e.g. down to one for 5400
68  * MCH.  We are trying to reduce the latency for clearing the fault
69  * register file.  The task is usually long-running, since printf() is
70  * slow, but this is not problematic because bursts are rare.
71  *
72  * For the same reason, each translation unit task is executed in its
73  * own thread.
74  *
75  * XXXKIB It seems there is no hardware available which implements
76  * advanced fault logging, so the code to handle AFL is not written.
77  */
78 
79 static int
80 dmar_fault_next(struct dmar_unit *unit, int faultp)
81 {
82 
83 	faultp += 2;
84 	if (faultp == unit->fault_log_size)
85 		faultp = 0;
86 	return (faultp);
87 }
88 
89 static void
90 dmar_fault_intr_clear(struct dmar_unit *unit, uint32_t fsts)
91 {
92 	uint32_t clear;
93 
94 	clear = 0;
95 	if ((fsts & DMAR_FSTS_ITE) != 0) {
96 		printf("DMAR%d: Invalidation timed out\n", unit->iommu.unit);
97 		clear |= DMAR_FSTS_ITE;
98 	}
99 	if ((fsts & DMAR_FSTS_ICE) != 0) {
100 		printf("DMAR%d: Invalidation completion error\n",
101 		    unit->iommu.unit);
102 		clear |= DMAR_FSTS_ICE;
103 	}
104 	if ((fsts & DMAR_FSTS_IQE) != 0) {
105 		printf("DMAR%d: Invalidation queue error\n",
106 		    unit->iommu.unit);
107 		clear |= DMAR_FSTS_IQE;
108 	}
109 	if ((fsts & DMAR_FSTS_APF) != 0) {
110 		printf("DMAR%d: Advanced pending fault\n", unit->iommu.unit);
111 		clear |= DMAR_FSTS_APF;
112 	}
113 	if ((fsts & DMAR_FSTS_AFO) != 0) {
114 		printf("DMAR%d: Advanced fault overflow\n", unit->iommu.unit);
115 		clear |= DMAR_FSTS_AFO;
116 	}
117 	if (clear != 0)
118 		dmar_write4(unit, DMAR_FSTS_REG, clear);
119 }
120 
121 int
122 dmar_fault_intr(void *arg)
123 {
124 	struct dmar_unit *unit;
125 	uint64_t fault_rec[2];
126 	uint32_t fsts;
127 	int fri, frir, faultp;
128 	bool enqueue;
129 
130 	unit = arg;
131 	enqueue = false;
132 	fsts = dmar_read4(unit, DMAR_FSTS_REG);
133 	dmar_fault_intr_clear(unit, fsts);
134 
135 	if ((fsts & DMAR_FSTS_PPF) == 0)
136 		goto done;
137 
138 	fri = DMAR_FSTS_FRI(fsts);
139 	for (;;) {
140 		frir = (DMAR_CAP_FRO(unit->hw_cap) + fri) * 16;
141 		fault_rec[1] = dmar_read8(unit, frir + 8);
142 		if ((fault_rec[1] & DMAR_FRCD2_F) == 0)
143 			break;
144 		fault_rec[0] = dmar_read8(unit, frir);
145 		dmar_write4(unit, frir + 12, DMAR_FRCD2_F32);
146 		DMAR_FAULT_LOCK(unit);
147 		faultp = unit->fault_log_head;
148 		if (dmar_fault_next(unit, faultp) == unit->fault_log_tail) {
149 			/* XXXKIB log overflow */
150 		} else {
151 			unit->fault_log[faultp] = fault_rec[0];
152 			unit->fault_log[faultp + 1] = fault_rec[1];
153 			unit->fault_log_head = dmar_fault_next(unit, faultp);
154 			enqueue = true;
155 		}
156 		DMAR_FAULT_UNLOCK(unit);
157 		fri += 1;
158 		if (fri >= DMAR_CAP_NFR(unit->hw_cap))
159 			fri = 0;
160 	}
161 
162 done:
163 	/*
164 	 * On SandyBridge, due to errata BJ124, IvyBridge errata
165 	 * BV100, and Haswell errata HSD40, "Spurious Intel VT-d
166 	 * Interrupts May Occur When the PFO Bit is Set".  Handle the
167 	 * cases by clearing overflow bit even if no fault is
168 	 * reported.
169 	 *
170 	 * On IvyBridge, errata BV30 states that clearing clear
171 	 * DMAR_FRCD2_F bit in the fault register causes spurious
172 	 * interrupt.  Do nothing.
173 	 *
174 	 */
175 	if ((fsts & DMAR_FSTS_PFO) != 0) {
176 		printf("DMAR%d: Fault Overflow\n", unit->iommu.unit);
177 		dmar_write4(unit, DMAR_FSTS_REG, DMAR_FSTS_PFO);
178 	}
179 
180 	if (enqueue) {
181 		taskqueue_enqueue(unit->fault_taskqueue,
182 		    &unit->fault_task);
183 	}
184 	return (FILTER_HANDLED);
185 }
186 
187 static void
188 dmar_fault_task(void *arg, int pending __unused)
189 {
190 	struct dmar_unit *unit;
191 	struct dmar_ctx *ctx;
192 	uint64_t fault_rec[2];
193 	int sid, bus, slot, func, faultp;
194 
195 	unit = arg;
196 	DMAR_FAULT_LOCK(unit);
197 	for (;;) {
198 		faultp = unit->fault_log_tail;
199 		if (faultp == unit->fault_log_head)
200 			break;
201 
202 		fault_rec[0] = unit->fault_log[faultp];
203 		fault_rec[1] = unit->fault_log[faultp + 1];
204 		unit->fault_log_tail = dmar_fault_next(unit, faultp);
205 		DMAR_FAULT_UNLOCK(unit);
206 
207 		sid = DMAR_FRCD2_SID(fault_rec[1]);
208 		printf("DMAR%d: ", unit->iommu.unit);
209 		DMAR_LOCK(unit);
210 		ctx = dmar_find_ctx_locked(unit, sid);
211 		if (ctx == NULL) {
212 			printf("<unknown dev>:");
213 
214 			/*
215 			 * Note that the slot and function will not be correct
216 			 * if ARI is in use, but without a ctx entry we have
217 			 * no way of knowing whether ARI is in use or not.
218 			 */
219 			bus = PCI_RID2BUS(sid);
220 			slot = PCI_RID2SLOT(sid);
221 			func = PCI_RID2FUNC(sid);
222 		} else {
223 			ctx->context.flags |= IOMMU_CTX_FAULTED;
224 			ctx->last_fault_rec[0] = fault_rec[0];
225 			ctx->last_fault_rec[1] = fault_rec[1];
226 			device_print_prettyname(ctx->context.tag->owner);
227 			bus = pci_get_bus(ctx->context.tag->owner);
228 			slot = pci_get_slot(ctx->context.tag->owner);
229 			func = pci_get_function(ctx->context.tag->owner);
230 		}
231 		DMAR_UNLOCK(unit);
232 		printf(
233 		    "pci%d:%d:%d sid %x fault acc %x adt 0x%x reason 0x%x "
234 		    "addr %jx\n",
235 		    bus, slot, func, sid, DMAR_FRCD2_T(fault_rec[1]),
236 		    DMAR_FRCD2_AT(fault_rec[1]), DMAR_FRCD2_FR(fault_rec[1]),
237 		    (uintmax_t)fault_rec[0]);
238 		DMAR_FAULT_LOCK(unit);
239 	}
240 	DMAR_FAULT_UNLOCK(unit);
241 }
242 
243 static void
244 dmar_clear_faults(struct dmar_unit *unit)
245 {
246 	uint32_t frec, frir, fsts;
247 	int i;
248 
249 	for (i = 0; i < DMAR_CAP_NFR(unit->hw_cap); i++) {
250 		frir = (DMAR_CAP_FRO(unit->hw_cap) + i) * 16;
251 		frec = dmar_read4(unit, frir + 12);
252 		if ((frec & DMAR_FRCD2_F32) == 0)
253 			continue;
254 		dmar_write4(unit, frir + 12, DMAR_FRCD2_F32);
255 	}
256 	fsts = dmar_read4(unit, DMAR_FSTS_REG);
257 	dmar_write4(unit, DMAR_FSTS_REG, fsts);
258 }
259 
260 int
261 dmar_init_fault_log(struct dmar_unit *unit)
262 {
263 
264 	mtx_init(&unit->fault_lock, "dmarflt", NULL, MTX_SPIN);
265 	unit->fault_log_size = 256; /* 128 fault log entries */
266 	TUNABLE_INT_FETCH("hw.dmar.fault_log_size", &unit->fault_log_size);
267 	if (unit->fault_log_size % 2 != 0)
268 		panic("hw.dmar_fault_log_size must be even");
269 	unit->fault_log = malloc(sizeof(uint64_t) * unit->fault_log_size,
270 	    M_DEVBUF, M_WAITOK | M_ZERO);
271 
272 	TASK_INIT(&unit->fault_task, 0, dmar_fault_task, unit);
273 	unit->fault_taskqueue = taskqueue_create_fast("dmarff", M_WAITOK,
274 	    taskqueue_thread_enqueue, &unit->fault_taskqueue);
275 	taskqueue_start_threads(&unit->fault_taskqueue, 1, PI_AV,
276 	    "dmar%d fault taskq", unit->iommu.unit);
277 
278 	DMAR_LOCK(unit);
279 	dmar_disable_fault_intr(unit);
280 	dmar_clear_faults(unit);
281 	dmar_enable_fault_intr(unit);
282 	DMAR_UNLOCK(unit);
283 
284 	return (0);
285 }
286 
287 void
288 dmar_fini_fault_log(struct dmar_unit *unit)
289 {
290 
291 	if (unit->fault_taskqueue == NULL)
292 		return;
293 
294 	DMAR_LOCK(unit);
295 	dmar_disable_fault_intr(unit);
296 	DMAR_UNLOCK(unit);
297 
298 	taskqueue_drain(unit->fault_taskqueue, &unit->fault_task);
299 	taskqueue_free(unit->fault_taskqueue);
300 	unit->fault_taskqueue = NULL;
301 	mtx_destroy(&unit->fault_lock);
302 
303 	free(unit->fault_log, M_DEVBUF);
304 	unit->fault_log = NULL;
305 	unit->fault_log_head = unit->fault_log_tail = 0;
306 }
307 
308 void
309 dmar_enable_fault_intr(struct dmar_unit *unit)
310 {
311 	uint32_t fectl;
312 
313 	DMAR_ASSERT_LOCKED(unit);
314 	fectl = dmar_read4(unit, DMAR_FECTL_REG);
315 	fectl &= ~DMAR_FECTL_IM;
316 	dmar_write4(unit, DMAR_FECTL_REG, fectl);
317 }
318 
319 void
320 dmar_disable_fault_intr(struct dmar_unit *unit)
321 {
322 	uint32_t fectl;
323 
324 	DMAR_ASSERT_LOCKED(unit);
325 	fectl = dmar_read4(unit, DMAR_FECTL_REG);
326 	dmar_write4(unit, DMAR_FECTL_REG, fectl | DMAR_FECTL_IM);
327 }
328