xref: /freebsd/sys/x86/iommu/intel_dmar.h (revision 59f5f100b774de8824fb2fc1a8a11a93bbc2dafd)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2013-2015 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 #ifndef __X86_IOMMU_INTEL_DMAR_H
32 #define	__X86_IOMMU_INTEL_DMAR_H
33 
34 #include <dev/iommu/iommu.h>
35 
36 struct dmar_unit;
37 
38 /*
39  * Locking annotations:
40  * (u) - Protected by iommu unit lock
41  * (d) - Protected by domain lock
42  * (c) - Immutable after initialization
43  */
44 
45 /*
46  * The domain abstraction.  Most non-constant members of the domain
47  * are protected by owning dmar unit lock, not by the domain lock.
48  * Most important, the dmar lock protects the contexts list.
49  *
50  * The domain lock protects the address map for the domain, and list
51  * of unload entries delayed.
52  *
53  * Page tables pages and pages content is protected by the vm object
54  * lock pgtbl_obj, which contains the page tables pages.
55  */
56 struct dmar_domain {
57 	struct iommu_domain iodom;
58 	int domain;			/* (c) DID, written in context entry */
59 	int mgaw;			/* (c) Real max address width */
60 	int agaw;			/* (c) Adjusted guest address width */
61 	int pglvl;			/* (c) The pagelevel */
62 	int awlvl;			/* (c) The pagelevel as the bitmask,
63 					   to set in context entry */
64 	u_int ctx_cnt;			/* (u) Number of contexts owned */
65 	u_int refs;			/* (u) Refs, including ctx */
66 	struct dmar_unit *dmar;		/* (c) */
67 	LIST_ENTRY(dmar_domain) link;	/* (u) Member in the dmar list */
68 	LIST_HEAD(, dmar_ctx) contexts;	/* (u) */
69 	vm_object_t pgtbl_obj;		/* (c) Page table pages */
70 	u_int batch_no;
71 };
72 
73 struct dmar_ctx {
74 	struct iommu_ctx context;
75 	uint64_t last_fault_rec[2];	/* Last fault reported */
76 	LIST_ENTRY(dmar_ctx) link;	/* (u) Member in the domain list */
77 	u_int refs;			/* (u) References from tags */
78 };
79 
80 #define	DMAR_DOMAIN_PGLOCK(dom)		VM_OBJECT_WLOCK((dom)->pgtbl_obj)
81 #define	DMAR_DOMAIN_PGTRYLOCK(dom)	VM_OBJECT_TRYWLOCK((dom)->pgtbl_obj)
82 #define	DMAR_DOMAIN_PGUNLOCK(dom)	VM_OBJECT_WUNLOCK((dom)->pgtbl_obj)
83 #define	DMAR_DOMAIN_ASSERT_PGLOCKED(dom) \
84 	VM_OBJECT_ASSERT_WLOCKED((dom)->pgtbl_obj)
85 
86 #define	DMAR_DOMAIN_LOCK(dom)	mtx_lock(&(dom)->iodom.lock)
87 #define	DMAR_DOMAIN_UNLOCK(dom)	mtx_unlock(&(dom)->iodom.lock)
88 #define	DMAR_DOMAIN_ASSERT_LOCKED(dom) mtx_assert(&(dom)->iodom.lock, MA_OWNED)
89 
90 #define	DMAR2IOMMU(dmar)	(&((dmar)->iommu))
91 #define	IOMMU2DMAR(dmar)	\
92 	__containerof((dmar), struct dmar_unit, iommu)
93 
94 #define	DOM2IODOM(domain)	(&((domain)->iodom))
95 #define	IODOM2DOM(domain)	\
96 	__containerof((domain), struct dmar_domain, iodom)
97 
98 #define	CTX2IOCTX(ctx)		(&((ctx)->context))
99 #define	IOCTX2CTX(ctx)		\
100 	__containerof((ctx), struct dmar_ctx, context)
101 
102 #define	CTX2DOM(ctx)		IODOM2DOM((ctx)->context.domain)
103 #define	CTX2DMAR(ctx)		(CTX2DOM(ctx)->dmar)
104 #define	DOM2DMAR(domain)	((domain)->dmar)
105 
106 #define	DMAR_INTR_FAULT		0
107 #define	DMAR_INTR_QI		1
108 #define	DMAR_INTR_TOTAL		2
109 
110 struct dmar_unit {
111 	struct iommu_unit iommu;
112 	struct x86_unit_common x86c;
113 	uint16_t segment;
114 	uint64_t base;
115 
116 	/* Resources */
117 	int reg_rid;
118 	struct resource *regs;
119 
120 	/* Hardware registers cache */
121 	uint32_t hw_ver;
122 	uint64_t hw_cap;
123 	uint64_t hw_ecap;
124 	uint32_t hw_gcmd;
125 
126 	/* Data for being a dmar */
127 	LIST_HEAD(, dmar_domain) domains;
128 	struct unrhdr *domids;
129 	vm_object_t ctx_obj;
130 	u_int barrier_flags;
131 
132 	/* Fault handler data */
133 	struct mtx fault_lock;
134 	uint64_t *fault_log;
135 	int fault_log_head;
136 	int fault_log_tail;
137 	int fault_log_size;
138 	struct task fault_task;
139 	struct taskqueue *fault_taskqueue;
140 
141 	/* QI */
142 	int qi_enabled;
143 
144 	/* IR */
145 	int ir_enabled;
146 	vm_paddr_t irt_phys;
147 	dmar_irte_t *irt;
148 	u_int irte_cnt;
149 	vmem_t *irtids;
150 };
151 
152 #define	DMAR_LOCK(dmar)		mtx_lock(&DMAR2IOMMU(dmar)->lock)
153 #define	DMAR_UNLOCK(dmar)	mtx_unlock(&DMAR2IOMMU(dmar)->lock)
154 #define	DMAR_ASSERT_LOCKED(dmar) mtx_assert(&DMAR2IOMMU(dmar)->lock, MA_OWNED)
155 
156 #define	DMAR_FAULT_LOCK(dmar)	mtx_lock_spin(&(dmar)->fault_lock)
157 #define	DMAR_FAULT_UNLOCK(dmar)	mtx_unlock_spin(&(dmar)->fault_lock)
158 #define	DMAR_FAULT_ASSERT_LOCKED(dmar) mtx_assert(&(dmar)->fault_lock, MA_OWNED)
159 
160 #define	DMAR_IS_COHERENT(dmar)	(((dmar)->hw_ecap & DMAR_ECAP_C) != 0)
161 #define	DMAR_HAS_QI(dmar)	(((dmar)->hw_ecap & DMAR_ECAP_QI) != 0)
162 #define	DMAR_X2APIC(dmar) \
163 	(x2apic_mode && ((dmar)->hw_ecap & DMAR_ECAP_EIM) != 0)
164 
165 /* Barrier ids */
166 #define	DMAR_BARRIER_RMRR	0
167 #define	DMAR_BARRIER_USEQ	1
168 
169 SYSCTL_DECL(_hw_iommu_dmar);
170 
171 struct dmar_unit *dmar_find(device_t dev, bool verbose);
172 struct dmar_unit *dmar_find_hpet(device_t dev, uint16_t *rid);
173 struct dmar_unit *dmar_find_ioapic(u_int apic_id, uint16_t *rid);
174 
175 u_int dmar_nd2mask(u_int nd);
176 bool dmar_pglvl_supported(struct dmar_unit *unit, int pglvl);
177 int domain_set_agaw(struct dmar_domain *domain, int mgaw);
178 int dmar_maxaddr2mgaw(struct dmar_unit *unit, iommu_gaddr_t maxaddr,
179     bool allow_less);
180 int domain_is_sp_lvl(struct dmar_domain *domain, int lvl);
181 iommu_gaddr_t domain_page_size(struct dmar_domain *domain, int lvl);
182 int calc_am(struct dmar_unit *unit, iommu_gaddr_t base, iommu_gaddr_t size,
183     iommu_gaddr_t *isizep);
184 int dmar_load_root_entry_ptr(struct dmar_unit *unit);
185 int dmar_inv_ctx_glob(struct dmar_unit *unit);
186 int dmar_inv_iotlb_glob(struct dmar_unit *unit);
187 int dmar_flush_write_bufs(struct dmar_unit *unit);
188 void dmar_flush_pte_to_ram(struct dmar_unit *unit, iommu_pte_t *dst);
189 void dmar_flush_ctx_to_ram(struct dmar_unit *unit, dmar_ctx_entry_t *dst);
190 void dmar_flush_root_to_ram(struct dmar_unit *unit, dmar_root_entry_t *dst);
191 int dmar_disable_protected_regions(struct dmar_unit *unit);
192 int dmar_enable_translation(struct dmar_unit *unit);
193 int dmar_disable_translation(struct dmar_unit *unit);
194 int dmar_load_irt_ptr(struct dmar_unit *unit);
195 int dmar_enable_ir(struct dmar_unit *unit);
196 int dmar_disable_ir(struct dmar_unit *unit);
197 bool dmar_barrier_enter(struct dmar_unit *dmar, u_int barrier_id);
198 void dmar_barrier_exit(struct dmar_unit *dmar, u_int barrier_id);
199 uint64_t dmar_get_timeout(void);
200 void dmar_update_timeout(uint64_t newval);
201 
202 int dmar_fault_intr(void *arg);
203 void dmar_enable_fault_intr(struct iommu_unit *unit);
204 void dmar_disable_fault_intr(struct iommu_unit *unit);
205 int dmar_init_fault_log(struct dmar_unit *unit);
206 void dmar_fini_fault_log(struct dmar_unit *unit);
207 
208 int dmar_qi_intr(void *arg);
209 void dmar_enable_qi_intr(struct iommu_unit *unit);
210 void dmar_disable_qi_intr(struct iommu_unit *unit);
211 int dmar_init_qi(struct dmar_unit *unit);
212 void dmar_fini_qi(struct dmar_unit *unit);
213 void dmar_qi_invalidate_locked(struct dmar_domain *domain,
214     struct iommu_map_entry *entry, bool emit_wait);
215 void dmar_qi_invalidate_sync(struct dmar_domain *domain, iommu_gaddr_t start,
216     iommu_gaddr_t size, bool cansleep);
217 void dmar_qi_invalidate_ctx_glob_locked(struct dmar_unit *unit);
218 void dmar_qi_invalidate_iotlb_glob_locked(struct dmar_unit *unit);
219 void dmar_qi_invalidate_iec_glob(struct dmar_unit *unit);
220 void dmar_qi_invalidate_iec(struct dmar_unit *unit, u_int start, u_int cnt);
221 
222 vm_object_t dmar_get_idmap_pgtbl(struct dmar_domain *domain,
223     iommu_gaddr_t maxaddr);
224 void dmar_put_idmap_pgtbl(vm_object_t obj);
225 void dmar_flush_iotlb_sync(struct dmar_domain *domain, iommu_gaddr_t base,
226     iommu_gaddr_t size);
227 int dmar_domain_alloc_pgtbl(struct dmar_domain *domain);
228 void dmar_domain_free_pgtbl(struct dmar_domain *domain);
229 extern const struct iommu_domain_map_ops dmar_domain_map_ops;
230 
231 int dmar_dev_depth(device_t child);
232 void dmar_dev_path(device_t child, int *busno, void *path1, int depth);
233 
234 struct dmar_ctx *dmar_get_ctx_for_dev(struct dmar_unit *dmar, device_t dev,
235     uint16_t rid, bool id_mapped, bool rmrr_init);
236 struct dmar_ctx *dmar_get_ctx_for_devpath(struct dmar_unit *dmar, uint16_t rid,
237     int dev_domain, int dev_busno, const void *dev_path, int dev_path_len,
238     bool id_mapped, bool rmrr_init);
239 int dmar_move_ctx_to_domain(struct dmar_domain *domain, struct dmar_ctx *ctx);
240 void dmar_free_ctx_locked_method(struct iommu_unit *dmar,
241     struct iommu_ctx *ctx);
242 void dmar_free_ctx_method(struct iommu_ctx *ctx);
243 struct dmar_ctx *dmar_find_ctx_locked(struct dmar_unit *dmar, uint16_t rid);
244 struct iommu_ctx *dmar_get_ctx(struct iommu_unit *iommu, device_t dev,
245     uint16_t rid, bool id_mapped, bool rmrr_init);
246 void dmar_domain_unload_entry(struct iommu_map_entry *entry, bool free,
247     bool cansleep);
248 void dmar_domain_unload(struct iommu_domain *iodom,
249     struct iommu_map_entries_tailq *entries, bool cansleep);
250 
251 void dmar_dev_parse_rmrr(struct dmar_domain *domain, int dev_domain,
252     int dev_busno, const void *dev_path, int dev_path_len,
253     struct iommu_map_entries_tailq *rmrr_entries);
254 int dmar_instantiate_rmrr_ctxs(struct iommu_unit *dmar);
255 
256 void dmar_quirks_post_ident(struct dmar_unit *dmar);
257 void dmar_quirks_pre_use(struct iommu_unit *dmar);
258 
259 int dmar_init_irt(struct dmar_unit *unit);
260 void dmar_fini_irt(struct dmar_unit *unit);
261 int dmar_alloc_msi_intr(device_t src, u_int *cookies, u_int count);
262 int dmar_map_msi_intr(device_t src, u_int cpu, u_int vector, u_int cookie,
263     uint64_t *addr, uint32_t *data);
264 int dmar_unmap_msi_intr(device_t src, u_int cookie);
265 int dmar_map_ioapic_intr(u_int ioapic_id, u_int cpu, u_int vector, bool edge,
266     bool activehi, int irq, u_int *cookie, uint32_t *hi, uint32_t *lo);
267 int dmar_unmap_ioapic_intr(u_int ioapic_id, u_int *cookie);
268 
269 extern int haw;
270 extern int dmar_rmrr_enable;
271 
272 static inline uint32_t
273 dmar_read4(const struct dmar_unit *unit, int reg)
274 {
275 
276 	return (bus_read_4(unit->regs, reg));
277 }
278 
279 static inline uint64_t
280 dmar_read8(const struct dmar_unit *unit, int reg)
281 {
282 #ifdef __i386__
283 	uint32_t high, low;
284 
285 	low = bus_read_4(unit->regs, reg);
286 	high = bus_read_4(unit->regs, reg + 4);
287 	return (low | ((uint64_t)high << 32));
288 #else
289 	return (bus_read_8(unit->regs, reg));
290 #endif
291 }
292 
293 static inline void
294 dmar_write4(const struct dmar_unit *unit, int reg, uint32_t val)
295 {
296 
297 	KASSERT(reg != DMAR_GCMD_REG || (val & DMAR_GCMD_TE) ==
298 	    (unit->hw_gcmd & DMAR_GCMD_TE),
299 	    ("dmar%d clearing TE 0x%08x 0x%08x", unit->iommu.unit,
300 	    unit->hw_gcmd, val));
301 	bus_write_4(unit->regs, reg, val);
302 }
303 
304 static inline void
305 dmar_write8(const struct dmar_unit *unit, int reg, uint64_t val)
306 {
307 
308 	KASSERT(reg != DMAR_GCMD_REG, ("8byte GCMD write"));
309 #ifdef __i386__
310 	uint32_t high, low;
311 
312 	low = val;
313 	high = val >> 32;
314 	bus_write_4(unit->regs, reg, low);
315 	bus_write_4(unit->regs, reg + 4, high);
316 #else
317 	bus_write_8(unit->regs, reg, val);
318 #endif
319 }
320 
321 /*
322  * dmar_pte_store and dmar_pte_clear ensure that on i386, 32bit writes
323  * are issued in the correct order.  For store, the lower word,
324  * containing the P or R and W bits, is set only after the high word
325  * is written.  For clear, the P bit is cleared first, then the high
326  * word is cleared.
327  *
328  * dmar_pte_update updates the pte.  For amd64, the update is atomic.
329  * For i386, it first disables the entry by clearing the word
330  * containing the P bit, and then defer to dmar_pte_store.  The locked
331  * cmpxchg8b is probably available on any machine having DMAR support,
332  * but interrupt translation table may be mapped uncached.
333  */
334 static inline void
335 dmar_pte_store1(volatile uint64_t *dst, uint64_t val)
336 {
337 #ifdef __i386__
338 	volatile uint32_t *p;
339 	uint32_t hi, lo;
340 
341 	hi = val >> 32;
342 	lo = val;
343 	p = (volatile uint32_t *)dst;
344 	*(p + 1) = hi;
345 	*p = lo;
346 #else
347 	*dst = val;
348 #endif
349 }
350 
351 static inline void
352 dmar_pte_store(volatile uint64_t *dst, uint64_t val)
353 {
354 
355 	KASSERT(*dst == 0, ("used pte %p oldval %jx newval %jx",
356 	    dst, (uintmax_t)*dst, (uintmax_t)val));
357 	dmar_pte_store1(dst, val);
358 }
359 
360 static inline void
361 dmar_pte_update(volatile uint64_t *dst, uint64_t val)
362 {
363 
364 #ifdef __i386__
365 	volatile uint32_t *p;
366 
367 	p = (volatile uint32_t *)dst;
368 	*p = 0;
369 #endif
370 	dmar_pte_store1(dst, val);
371 }
372 
373 static inline void
374 dmar_pte_clear(volatile uint64_t *dst)
375 {
376 #ifdef __i386__
377 	volatile uint32_t *p;
378 
379 	p = (volatile uint32_t *)dst;
380 	*p = 0;
381 	*(p + 1) = 0;
382 #else
383 	*dst = 0;
384 #endif
385 }
386 
387 extern struct timespec dmar_hw_timeout;
388 
389 #define	DMAR_WAIT_UNTIL(cond)					\
390 {								\
391 	struct timespec last, curr;				\
392 	bool forever;						\
393 								\
394 	if (dmar_hw_timeout.tv_sec == 0 &&			\
395 	    dmar_hw_timeout.tv_nsec == 0) {			\
396 		forever = true;					\
397 	} else {						\
398 		forever = false;				\
399 		nanouptime(&curr);				\
400 		timespecadd(&curr, &dmar_hw_timeout, &last);	\
401 	}							\
402 	for (;;) {						\
403 		if (cond) {					\
404 			error = 0;				\
405 			break;					\
406 		}						\
407 		nanouptime(&curr);				\
408 		if (!forever && timespeccmp(&last, &curr, <)) {	\
409 			error = ETIMEDOUT;			\
410 			break;					\
411 		}						\
412 		cpu_spinwait();					\
413 	}							\
414 }
415 
416 #ifdef INVARIANTS
417 #define	TD_PREP_PINNED_ASSERT						\
418 	int old_td_pinned;						\
419 	old_td_pinned = curthread->td_pinned
420 #define	TD_PINNED_ASSERT						\
421 	KASSERT(curthread->td_pinned == old_td_pinned,			\
422 	    ("pin count leak: %d %d %s:%d", curthread->td_pinned,	\
423 	    old_td_pinned, __FILE__, __LINE__))
424 #else
425 #define	TD_PREP_PINNED_ASSERT
426 #define	TD_PINNED_ASSERT
427 #endif
428 
429 #endif
430