xref: /freebsd/sys/x86/iommu/intel_dmar.h (revision b197d4b893974c9eb4d7b38704c6d5c486235d6f)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
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  * $FreeBSD$
31  */
32 
33 #ifndef __X86_IOMMU_INTEL_DMAR_H
34 #define	__X86_IOMMU_INTEL_DMAR_H
35 
36 #include <dev/iommu/iommu.h>
37 
38 struct dmar_unit;
39 
40 /*
41  * Locking annotations:
42  * (u) - Protected by iommu unit lock
43  * (d) - Protected by domain lock
44  * (c) - Immutable after initialization
45  */
46 
47 /*
48  * The domain abstraction.  Most non-constant members of the domain
49  * are protected by owning dmar unit lock, not by the domain lock.
50  * Most important, the dmar lock protects the contexts list.
51  *
52  * The domain lock protects the address map for the domain, and list
53  * of unload entries delayed.
54  *
55  * Page tables pages and pages content is protected by the vm object
56  * lock pgtbl_obj, which contains the page tables pages.
57  */
58 struct dmar_domain {
59 	struct iommu_domain iodom;
60 	int domain;			/* (c) DID, written in context entry */
61 	int mgaw;			/* (c) Real max address width */
62 	int agaw;			/* (c) Adjusted guest address width */
63 	int pglvl;			/* (c) The pagelevel */
64 	int awlvl;			/* (c) The pagelevel as the bitmask,
65 					   to set in context entry */
66 	u_int ctx_cnt;			/* (u) Number of contexts owned */
67 	u_int refs;			/* (u) Refs, including ctx */
68 	struct dmar_unit *dmar;		/* (c) */
69 	LIST_ENTRY(dmar_domain) link;	/* (u) Member in the dmar list */
70 	LIST_HEAD(, dmar_ctx) contexts;	/* (u) */
71 	vm_object_t pgtbl_obj;		/* (c) Page table pages */
72 	u_int batch_no;
73 };
74 
75 struct dmar_ctx {
76 	struct iommu_ctx context;
77 	uint64_t last_fault_rec[2];	/* Last fault reported */
78 	LIST_ENTRY(dmar_ctx) link;	/* (u) Member in the domain list */
79 	u_int refs;			/* (u) References from tags */
80 };
81 
82 #define	DMAR_DOMAIN_PGLOCK(dom)		VM_OBJECT_WLOCK((dom)->pgtbl_obj)
83 #define	DMAR_DOMAIN_PGTRYLOCK(dom)	VM_OBJECT_TRYWLOCK((dom)->pgtbl_obj)
84 #define	DMAR_DOMAIN_PGUNLOCK(dom)	VM_OBJECT_WUNLOCK((dom)->pgtbl_obj)
85 #define	DMAR_DOMAIN_ASSERT_PGLOCKED(dom) \
86 	VM_OBJECT_ASSERT_WLOCKED((dom)->pgtbl_obj)
87 
88 #define	DMAR_DOMAIN_LOCK(dom)	mtx_lock(&(dom)->iodom.lock)
89 #define	DMAR_DOMAIN_UNLOCK(dom)	mtx_unlock(&(dom)->iodom.lock)
90 #define	DMAR_DOMAIN_ASSERT_LOCKED(dom) mtx_assert(&(dom)->iodom.lock, MA_OWNED)
91 
92 #define	DMAR2IOMMU(dmar)	&((dmar)->iommu)
93 #define	IOMMU2DMAR(dmar)	\
94 	__containerof((dmar), struct dmar_unit, iommu)
95 
96 #define	DOM2IODOM(domain)	&((domain)->iodom)
97 #define	IODOM2DOM(domain)	\
98 	__containerof((domain), struct dmar_domain, iodom)
99 
100 #define	CTX2IOCTX(ctx)		&((ctx)->context)
101 #define	IOCTX2CTX(ctx)		\
102 	__containerof((ctx), struct dmar_ctx, context)
103 
104 #define	CTX2DOM(ctx)		IODOM2DOM((ctx)->context.domain)
105 #define	CTX2DMAR(ctx)		(CTX2DOM(ctx)->dmar)
106 #define	DOM2DMAR(domain)	((domain)->dmar)
107 
108 struct dmar_msi_data {
109 	int irq;
110 	int irq_rid;
111 	struct resource *irq_res;
112 	void *intr_handle;
113 	int (*handler)(void *);
114 	int msi_data_reg;
115 	int msi_addr_reg;
116 	int msi_uaddr_reg;
117 	void (*enable_intr)(struct dmar_unit *);
118 	void (*disable_intr)(struct dmar_unit *);
119 	const char *name;
120 };
121 
122 #define	DMAR_INTR_FAULT		0
123 #define	DMAR_INTR_QI		1
124 #define	DMAR_INTR_TOTAL		2
125 
126 struct dmar_unit {
127 	struct iommu_unit iommu;
128 	device_t dev;
129 	uint16_t segment;
130 	uint64_t base;
131 
132 	/* Resources */
133 	int reg_rid;
134 	struct resource *regs;
135 
136 	struct dmar_msi_data intrs[DMAR_INTR_TOTAL];
137 
138 	/* Hardware registers cache */
139 	uint32_t hw_ver;
140 	uint64_t hw_cap;
141 	uint64_t hw_ecap;
142 	uint32_t hw_gcmd;
143 
144 	/* Data for being a dmar */
145 	LIST_HEAD(, dmar_domain) domains;
146 	struct unrhdr *domids;
147 	vm_object_t ctx_obj;
148 	u_int barrier_flags;
149 
150 	/* Fault handler data */
151 	struct mtx fault_lock;
152 	uint64_t *fault_log;
153 	int fault_log_head;
154 	int fault_log_tail;
155 	int fault_log_size;
156 	struct task fault_task;
157 	struct taskqueue *fault_taskqueue;
158 
159 	/* QI */
160 	int qi_enabled;
161 	vm_offset_t inv_queue;
162 	vm_size_t inv_queue_size;
163 	uint32_t inv_queue_avail;
164 	uint32_t inv_queue_tail;
165 	volatile uint32_t inv_waitd_seq_hw; /* hw writes there on wait
166 					       descr completion */
167 	uint64_t inv_waitd_seq_hw_phys;
168 	uint32_t inv_waitd_seq; /* next sequence number to use for wait descr */
169 	u_int inv_waitd_gen;	/* seq number generation AKA seq overflows */
170 	u_int inv_seq_waiters;	/* count of waiters for seq */
171 	u_int inv_queue_full;	/* informational counter */
172 
173 	/* IR */
174 	int ir_enabled;
175 	vm_paddr_t irt_phys;
176 	dmar_irte_t *irt;
177 	u_int irte_cnt;
178 	vmem_t *irtids;
179 
180 	/*
181 	 * Delayed freeing of map entries queue processing:
182 	 *
183 	 * tlb_flush_head and tlb_flush_tail are used to implement a FIFO
184 	 * queue that supports concurrent dequeues and enqueues.  However,
185 	 * there can only be a single dequeuer (accessing tlb_flush_head) and
186 	 * a single enqueuer (accessing tlb_flush_tail) at a time.  Since the
187 	 * unit's qi_task is the only dequeuer, it can access tlb_flush_head
188 	 * without any locking.  In contrast, there may be multiple enqueuers,
189 	 * so the enqueuers acquire the iommu unit lock to serialize their
190 	 * accesses to tlb_flush_tail.
191 	 *
192 	 * In this FIFO queue implementation, the key to enabling concurrent
193 	 * dequeues and enqueues is that the dequeuer never needs to access
194 	 * tlb_flush_tail and the enqueuer never needs to access
195 	 * tlb_flush_head.  In particular, tlb_flush_head and tlb_flush_tail
196 	 * are never NULL, so neither a dequeuer nor an enqueuer ever needs to
197 	 * update both.  Instead, tlb_flush_head always points to a "zombie"
198 	 * struct, which previously held the last dequeued item.  Thus, the
199 	 * zombie's next field actually points to the struct holding the first
200 	 * item in the queue.  When an item is dequeued, the current zombie is
201 	 * finally freed, and the struct that held the just dequeued item
202 	 * becomes the new zombie.  When the queue is empty, tlb_flush_tail
203 	 * also points to the zombie.
204 	 */
205 	struct iommu_map_entry *tlb_flush_head;
206 	struct iommu_map_entry *tlb_flush_tail;
207 	struct task qi_task;
208 	struct taskqueue *qi_taskqueue;
209 };
210 
211 #define	DMAR_LOCK(dmar)		mtx_lock(&(dmar)->iommu.lock)
212 #define	DMAR_UNLOCK(dmar)	mtx_unlock(&(dmar)->iommu.lock)
213 #define	DMAR_ASSERT_LOCKED(dmar) mtx_assert(&(dmar)->iommu.lock, MA_OWNED)
214 
215 #define	DMAR_FAULT_LOCK(dmar)	mtx_lock_spin(&(dmar)->fault_lock)
216 #define	DMAR_FAULT_UNLOCK(dmar)	mtx_unlock_spin(&(dmar)->fault_lock)
217 #define	DMAR_FAULT_ASSERT_LOCKED(dmar) mtx_assert(&(dmar)->fault_lock, MA_OWNED)
218 
219 #define	DMAR_IS_COHERENT(dmar)	(((dmar)->hw_ecap & DMAR_ECAP_C) != 0)
220 #define	DMAR_HAS_QI(dmar)	(((dmar)->hw_ecap & DMAR_ECAP_QI) != 0)
221 #define	DMAR_X2APIC(dmar) \
222 	(x2apic_mode && ((dmar)->hw_ecap & DMAR_ECAP_EIM) != 0)
223 
224 /* Barrier ids */
225 #define	DMAR_BARRIER_RMRR	0
226 #define	DMAR_BARRIER_USEQ	1
227 
228 struct dmar_unit *dmar_find(device_t dev, bool verbose);
229 struct dmar_unit *dmar_find_hpet(device_t dev, uint16_t *rid);
230 struct dmar_unit *dmar_find_ioapic(u_int apic_id, uint16_t *rid);
231 
232 u_int dmar_nd2mask(u_int nd);
233 bool dmar_pglvl_supported(struct dmar_unit *unit, int pglvl);
234 int domain_set_agaw(struct dmar_domain *domain, int mgaw);
235 int dmar_maxaddr2mgaw(struct dmar_unit *unit, iommu_gaddr_t maxaddr,
236     bool allow_less);
237 vm_pindex_t pglvl_max_pages(int pglvl);
238 int domain_is_sp_lvl(struct dmar_domain *domain, int lvl);
239 iommu_gaddr_t pglvl_page_size(int total_pglvl, int lvl);
240 iommu_gaddr_t domain_page_size(struct dmar_domain *domain, int lvl);
241 int calc_am(struct dmar_unit *unit, iommu_gaddr_t base, iommu_gaddr_t size,
242     iommu_gaddr_t *isizep);
243 struct vm_page *dmar_pgalloc(vm_object_t obj, vm_pindex_t idx, int flags);
244 void dmar_pgfree(vm_object_t obj, vm_pindex_t idx, int flags);
245 void *dmar_map_pgtbl(vm_object_t obj, vm_pindex_t idx, int flags,
246     struct sf_buf **sf);
247 void dmar_unmap_pgtbl(struct sf_buf *sf);
248 int dmar_load_root_entry_ptr(struct dmar_unit *unit);
249 int dmar_inv_ctx_glob(struct dmar_unit *unit);
250 int dmar_inv_iotlb_glob(struct dmar_unit *unit);
251 int dmar_flush_write_bufs(struct dmar_unit *unit);
252 void dmar_flush_pte_to_ram(struct dmar_unit *unit, dmar_pte_t *dst);
253 void dmar_flush_ctx_to_ram(struct dmar_unit *unit, dmar_ctx_entry_t *dst);
254 void dmar_flush_root_to_ram(struct dmar_unit *unit, dmar_root_entry_t *dst);
255 int dmar_disable_protected_regions(struct dmar_unit *unit);
256 int dmar_enable_translation(struct dmar_unit *unit);
257 int dmar_disable_translation(struct dmar_unit *unit);
258 int dmar_load_irt_ptr(struct dmar_unit *unit);
259 int dmar_enable_ir(struct dmar_unit *unit);
260 int dmar_disable_ir(struct dmar_unit *unit);
261 bool dmar_barrier_enter(struct dmar_unit *dmar, u_int barrier_id);
262 void dmar_barrier_exit(struct dmar_unit *dmar, u_int barrier_id);
263 uint64_t dmar_get_timeout(void);
264 void dmar_update_timeout(uint64_t newval);
265 
266 int dmar_fault_intr(void *arg);
267 void dmar_enable_fault_intr(struct dmar_unit *unit);
268 void dmar_disable_fault_intr(struct dmar_unit *unit);
269 int dmar_init_fault_log(struct dmar_unit *unit);
270 void dmar_fini_fault_log(struct dmar_unit *unit);
271 
272 int dmar_qi_intr(void *arg);
273 void dmar_enable_qi_intr(struct dmar_unit *unit);
274 void dmar_disable_qi_intr(struct dmar_unit *unit);
275 int dmar_init_qi(struct dmar_unit *unit);
276 void dmar_fini_qi(struct dmar_unit *unit);
277 void dmar_qi_invalidate_locked(struct dmar_domain *domain,
278     struct iommu_map_entry *entry, bool emit_wait);
279 void dmar_qi_invalidate_sync(struct dmar_domain *domain, iommu_gaddr_t start,
280     iommu_gaddr_t size, bool cansleep);
281 void dmar_qi_invalidate_ctx_glob_locked(struct dmar_unit *unit);
282 void dmar_qi_invalidate_iotlb_glob_locked(struct dmar_unit *unit);
283 void dmar_qi_invalidate_iec_glob(struct dmar_unit *unit);
284 void dmar_qi_invalidate_iec(struct dmar_unit *unit, u_int start, u_int cnt);
285 
286 vm_object_t domain_get_idmap_pgtbl(struct dmar_domain *domain,
287     iommu_gaddr_t maxaddr);
288 void put_idmap_pgtbl(vm_object_t obj);
289 void domain_flush_iotlb_sync(struct dmar_domain *domain, iommu_gaddr_t base,
290     iommu_gaddr_t size);
291 int domain_alloc_pgtbl(struct dmar_domain *domain);
292 void domain_free_pgtbl(struct dmar_domain *domain);
293 extern const struct iommu_domain_map_ops dmar_domain_map_ops;
294 
295 int dmar_dev_depth(device_t child);
296 void dmar_dev_path(device_t child, int *busno, void *path1, int depth);
297 
298 struct dmar_ctx *dmar_get_ctx_for_dev(struct dmar_unit *dmar, device_t dev,
299     uint16_t rid, bool id_mapped, bool rmrr_init);
300 struct dmar_ctx *dmar_get_ctx_for_devpath(struct dmar_unit *dmar, uint16_t rid,
301     int dev_domain, int dev_busno, const void *dev_path, int dev_path_len,
302     bool id_mapped, bool rmrr_init);
303 int dmar_move_ctx_to_domain(struct dmar_domain *domain, struct dmar_ctx *ctx);
304 void dmar_free_ctx_locked(struct dmar_unit *dmar, struct dmar_ctx *ctx);
305 void dmar_free_ctx(struct dmar_ctx *ctx);
306 struct dmar_ctx *dmar_find_ctx_locked(struct dmar_unit *dmar, uint16_t rid);
307 void dmar_domain_free_entry(struct iommu_map_entry *entry, bool free);
308 
309 void dmar_dev_parse_rmrr(struct dmar_domain *domain, int dev_domain,
310     int dev_busno, const void *dev_path, int dev_path_len,
311     struct iommu_map_entries_tailq *rmrr_entries);
312 int dmar_instantiate_rmrr_ctxs(struct iommu_unit *dmar);
313 
314 void dmar_quirks_post_ident(struct dmar_unit *dmar);
315 void dmar_quirks_pre_use(struct iommu_unit *dmar);
316 
317 int dmar_init_irt(struct dmar_unit *unit);
318 void dmar_fini_irt(struct dmar_unit *unit);
319 
320 extern iommu_haddr_t dmar_high;
321 extern int haw;
322 extern int dmar_tbl_pagecnt;
323 extern int dmar_batch_coalesce;
324 
325 static inline uint32_t
326 dmar_read4(const struct dmar_unit *unit, int reg)
327 {
328 
329 	return (bus_read_4(unit->regs, reg));
330 }
331 
332 static inline uint64_t
333 dmar_read8(const struct dmar_unit *unit, int reg)
334 {
335 #ifdef __i386__
336 	uint32_t high, low;
337 
338 	low = bus_read_4(unit->regs, reg);
339 	high = bus_read_4(unit->regs, reg + 4);
340 	return (low | ((uint64_t)high << 32));
341 #else
342 	return (bus_read_8(unit->regs, reg));
343 #endif
344 }
345 
346 static inline void
347 dmar_write4(const struct dmar_unit *unit, int reg, uint32_t val)
348 {
349 
350 	KASSERT(reg != DMAR_GCMD_REG || (val & DMAR_GCMD_TE) ==
351 	    (unit->hw_gcmd & DMAR_GCMD_TE),
352 	    ("dmar%d clearing TE 0x%08x 0x%08x", unit->iommu.unit,
353 	    unit->hw_gcmd, val));
354 	bus_write_4(unit->regs, reg, val);
355 }
356 
357 static inline void
358 dmar_write8(const struct dmar_unit *unit, int reg, uint64_t val)
359 {
360 
361 	KASSERT(reg != DMAR_GCMD_REG, ("8byte GCMD write"));
362 #ifdef __i386__
363 	uint32_t high, low;
364 
365 	low = val;
366 	high = val >> 32;
367 	bus_write_4(unit->regs, reg, low);
368 	bus_write_4(unit->regs, reg + 4, high);
369 #else
370 	bus_write_8(unit->regs, reg, val);
371 #endif
372 }
373 
374 /*
375  * dmar_pte_store and dmar_pte_clear ensure that on i386, 32bit writes
376  * are issued in the correct order.  For store, the lower word,
377  * containing the P or R and W bits, is set only after the high word
378  * is written.  For clear, the P bit is cleared first, then the high
379  * word is cleared.
380  *
381  * dmar_pte_update updates the pte.  For amd64, the update is atomic.
382  * For i386, it first disables the entry by clearing the word
383  * containing the P bit, and then defer to dmar_pte_store.  The locked
384  * cmpxchg8b is probably available on any machine having DMAR support,
385  * but interrupt translation table may be mapped uncached.
386  */
387 static inline void
388 dmar_pte_store1(volatile uint64_t *dst, uint64_t val)
389 {
390 #ifdef __i386__
391 	volatile uint32_t *p;
392 	uint32_t hi, lo;
393 
394 	hi = val >> 32;
395 	lo = val;
396 	p = (volatile uint32_t *)dst;
397 	*(p + 1) = hi;
398 	*p = lo;
399 #else
400 	*dst = val;
401 #endif
402 }
403 
404 static inline void
405 dmar_pte_store(volatile uint64_t *dst, uint64_t val)
406 {
407 
408 	KASSERT(*dst == 0, ("used pte %p oldval %jx newval %jx",
409 	    dst, (uintmax_t)*dst, (uintmax_t)val));
410 	dmar_pte_store1(dst, val);
411 }
412 
413 static inline void
414 dmar_pte_update(volatile uint64_t *dst, uint64_t val)
415 {
416 
417 #ifdef __i386__
418 	volatile uint32_t *p;
419 
420 	p = (volatile uint32_t *)dst;
421 	*p = 0;
422 #endif
423 	dmar_pte_store1(dst, val);
424 }
425 
426 static inline void
427 dmar_pte_clear(volatile uint64_t *dst)
428 {
429 #ifdef __i386__
430 	volatile uint32_t *p;
431 
432 	p = (volatile uint32_t *)dst;
433 	*p = 0;
434 	*(p + 1) = 0;
435 #else
436 	*dst = 0;
437 #endif
438 }
439 
440 extern struct timespec dmar_hw_timeout;
441 
442 #define	DMAR_WAIT_UNTIL(cond)					\
443 {								\
444 	struct timespec last, curr;				\
445 	bool forever;						\
446 								\
447 	if (dmar_hw_timeout.tv_sec == 0 &&			\
448 	    dmar_hw_timeout.tv_nsec == 0) {			\
449 		forever = true;					\
450 	} else {						\
451 		forever = false;				\
452 		nanouptime(&curr);				\
453 		timespecadd(&curr, &dmar_hw_timeout, &last);	\
454 	}							\
455 	for (;;) {						\
456 		if (cond) {					\
457 			error = 0;				\
458 			break;					\
459 		}						\
460 		nanouptime(&curr);				\
461 		if (!forever && timespeccmp(&last, &curr, <)) {	\
462 			error = ETIMEDOUT;			\
463 			break;					\
464 		}						\
465 		cpu_spinwait();					\
466 	}							\
467 }
468 
469 #ifdef INVARIANTS
470 #define	TD_PREP_PINNED_ASSERT						\
471 	int old_td_pinned;						\
472 	old_td_pinned = curthread->td_pinned
473 #define	TD_PINNED_ASSERT						\
474 	KASSERT(curthread->td_pinned == old_td_pinned,			\
475 	    ("pin count leak: %d %d %s:%d", curthread->td_pinned,	\
476 	    old_td_pinned, __FILE__, __LINE__))
477 #else
478 #define	TD_PREP_PINNED_ASSERT
479 #define	TD_PINNED_ASSERT
480 #endif
481 
482 #endif
483