xref: /freebsd/sys/x86/iommu/intel_dmar.h (revision b4af4f93c682e445bf159f0d1ec90b636296c946)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2013-2015 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  * $FreeBSD$
32  */
33 
34 #ifndef __X86_IOMMU_INTEL_DMAR_H
35 #define	__X86_IOMMU_INTEL_DMAR_H
36 
37 #include <dev/iommu/iommu.h>
38 
39 struct dmar_unit;
40 
41 /*
42  * Locking annotations:
43  * (u) - Protected by iommu unit lock
44  * (d) - Protected by domain lock
45  * (c) - Immutable after initialization
46  */
47 
48 /*
49  * The domain abstraction.  Most non-constant members of the domain
50  * are protected by owning dmar unit lock, not by the domain lock.
51  * Most important, the dmar lock protects the contexts list.
52  *
53  * The domain lock protects the address map for the domain, and list
54  * of unload entries delayed.
55  *
56  * Page tables pages and pages content is protected by the vm object
57  * lock pgtbl_obj, which contains the page tables pages.
58  */
59 struct dmar_domain {
60 	struct iommu_domain iodom;
61 	int domain;			/* (c) DID, written in context entry */
62 	int mgaw;			/* (c) Real max address width */
63 	int agaw;			/* (c) Adjusted guest address width */
64 	int pglvl;			/* (c) The pagelevel */
65 	int awlvl;			/* (c) The pagelevel as the bitmask,
66 					   to set in context entry */
67 	u_int ctx_cnt;			/* (u) Number of contexts owned */
68 	u_int refs;			/* (u) Refs, including ctx */
69 	struct dmar_unit *dmar;		/* (c) */
70 	LIST_ENTRY(dmar_domain) link;	/* (u) Member in the dmar list */
71 	LIST_HEAD(, dmar_ctx) contexts;	/* (u) */
72 	vm_object_t pgtbl_obj;		/* (c) Page table pages */
73 	u_int batch_no;
74 };
75 
76 struct dmar_ctx {
77 	struct iommu_ctx context;
78 	uint16_t rid;			/* (c) pci RID */
79 	uint64_t last_fault_rec[2];	/* Last fault reported */
80 	LIST_ENTRY(dmar_ctx) link;	/* (u) Member in the domain list */
81 	u_int refs;			/* (u) References from tags */
82 };
83 
84 #define	DMAR_DOMAIN_PGLOCK(dom)		VM_OBJECT_WLOCK((dom)->pgtbl_obj)
85 #define	DMAR_DOMAIN_PGTRYLOCK(dom)	VM_OBJECT_TRYWLOCK((dom)->pgtbl_obj)
86 #define	DMAR_DOMAIN_PGUNLOCK(dom)	VM_OBJECT_WUNLOCK((dom)->pgtbl_obj)
87 #define	DMAR_DOMAIN_ASSERT_PGLOCKED(dom) \
88 	VM_OBJECT_ASSERT_WLOCKED((dom)->pgtbl_obj)
89 
90 #define	DMAR_DOMAIN_LOCK(dom)	mtx_lock(&(dom)->iodom.lock)
91 #define	DMAR_DOMAIN_UNLOCK(dom)	mtx_unlock(&(dom)->iodom.lock)
92 #define	DMAR_DOMAIN_ASSERT_LOCKED(dom) mtx_assert(&(dom)->iodom.lock, MA_OWNED)
93 
94 struct dmar_msi_data {
95 	int irq;
96 	int irq_rid;
97 	struct resource *irq_res;
98 	void *intr_handle;
99 	int (*handler)(void *);
100 	int msi_data_reg;
101 	int msi_addr_reg;
102 	int msi_uaddr_reg;
103 	void (*enable_intr)(struct dmar_unit *);
104 	void (*disable_intr)(struct dmar_unit *);
105 	const char *name;
106 };
107 
108 #define	DMAR_INTR_FAULT		0
109 #define	DMAR_INTR_QI		1
110 #define	DMAR_INTR_TOTAL		2
111 
112 struct dmar_unit {
113 	struct iommu_unit iommu;
114 	device_t dev;
115 	uint16_t segment;
116 	uint64_t base;
117 
118 	/* Resources */
119 	int reg_rid;
120 	struct resource *regs;
121 
122 	struct dmar_msi_data intrs[DMAR_INTR_TOTAL];
123 
124 	/* Hardware registers cache */
125 	uint32_t hw_ver;
126 	uint64_t hw_cap;
127 	uint64_t hw_ecap;
128 	uint32_t hw_gcmd;
129 
130 	/* Data for being a dmar */
131 	LIST_HEAD(, dmar_domain) domains;
132 	struct unrhdr *domids;
133 	vm_object_t ctx_obj;
134 	u_int barrier_flags;
135 
136 	/* Fault handler data */
137 	struct mtx fault_lock;
138 	uint64_t *fault_log;
139 	int fault_log_head;
140 	int fault_log_tail;
141 	int fault_log_size;
142 	struct task fault_task;
143 	struct taskqueue *fault_taskqueue;
144 
145 	/* QI */
146 	int qi_enabled;
147 	vm_offset_t inv_queue;
148 	vm_size_t inv_queue_size;
149 	uint32_t inv_queue_avail;
150 	uint32_t inv_queue_tail;
151 	volatile uint32_t inv_waitd_seq_hw; /* hw writes there on wait
152 					       descr completion */
153 	uint64_t inv_waitd_seq_hw_phys;
154 	uint32_t inv_waitd_seq; /* next sequence number to use for wait descr */
155 	u_int inv_waitd_gen;	/* seq number generation AKA seq overflows */
156 	u_int inv_seq_waiters;	/* count of waiters for seq */
157 	u_int inv_queue_full;	/* informational counter */
158 
159 	/* IR */
160 	int ir_enabled;
161 	vm_paddr_t irt_phys;
162 	dmar_irte_t *irt;
163 	u_int irte_cnt;
164 	vmem_t *irtids;
165 
166 	/* Delayed freeing of map entries queue processing */
167 	struct iommu_map_entries_tailq tlb_flush_entries;
168 	struct task qi_task;
169 	struct taskqueue *qi_taskqueue;
170 };
171 
172 #define	DMAR_LOCK(dmar)		mtx_lock(&(dmar)->iommu.lock)
173 #define	DMAR_UNLOCK(dmar)	mtx_unlock(&(dmar)->iommu.lock)
174 #define	DMAR_ASSERT_LOCKED(dmar) mtx_assert(&(dmar)->iommu.lock, MA_OWNED)
175 
176 #define	DMAR_FAULT_LOCK(dmar)	mtx_lock_spin(&(dmar)->fault_lock)
177 #define	DMAR_FAULT_UNLOCK(dmar)	mtx_unlock_spin(&(dmar)->fault_lock)
178 #define	DMAR_FAULT_ASSERT_LOCKED(dmar) mtx_assert(&(dmar)->fault_lock, MA_OWNED)
179 
180 #define	DMAR_IS_COHERENT(dmar)	(((dmar)->hw_ecap & DMAR_ECAP_C) != 0)
181 #define	DMAR_HAS_QI(dmar)	(((dmar)->hw_ecap & DMAR_ECAP_QI) != 0)
182 #define	DMAR_X2APIC(dmar) \
183 	(x2apic_mode && ((dmar)->hw_ecap & DMAR_ECAP_EIM) != 0)
184 
185 /* Barrier ids */
186 #define	DMAR_BARRIER_RMRR	0
187 #define	DMAR_BARRIER_USEQ	1
188 
189 struct dmar_unit *dmar_find(device_t dev, bool verbose);
190 struct dmar_unit *dmar_find_hpet(device_t dev, uint16_t *rid);
191 struct dmar_unit *dmar_find_ioapic(u_int apic_id, uint16_t *rid);
192 
193 u_int dmar_nd2mask(u_int nd);
194 bool dmar_pglvl_supported(struct dmar_unit *unit, int pglvl);
195 int domain_set_agaw(struct dmar_domain *domain, int mgaw);
196 int dmar_maxaddr2mgaw(struct dmar_unit *unit, iommu_gaddr_t maxaddr,
197     bool allow_less);
198 vm_pindex_t pglvl_max_pages(int pglvl);
199 int domain_is_sp_lvl(struct dmar_domain *domain, int lvl);
200 iommu_gaddr_t pglvl_page_size(int total_pglvl, int lvl);
201 iommu_gaddr_t domain_page_size(struct dmar_domain *domain, int lvl);
202 int calc_am(struct dmar_unit *unit, iommu_gaddr_t base, iommu_gaddr_t size,
203     iommu_gaddr_t *isizep);
204 struct vm_page *dmar_pgalloc(vm_object_t obj, vm_pindex_t idx, int flags);
205 void dmar_pgfree(vm_object_t obj, vm_pindex_t idx, int flags);
206 void *dmar_map_pgtbl(vm_object_t obj, vm_pindex_t idx, int flags,
207     struct sf_buf **sf);
208 void dmar_unmap_pgtbl(struct sf_buf *sf);
209 int dmar_load_root_entry_ptr(struct dmar_unit *unit);
210 int dmar_inv_ctx_glob(struct dmar_unit *unit);
211 int dmar_inv_iotlb_glob(struct dmar_unit *unit);
212 int dmar_flush_write_bufs(struct dmar_unit *unit);
213 void dmar_flush_pte_to_ram(struct dmar_unit *unit, dmar_pte_t *dst);
214 void dmar_flush_ctx_to_ram(struct dmar_unit *unit, dmar_ctx_entry_t *dst);
215 void dmar_flush_root_to_ram(struct dmar_unit *unit, dmar_root_entry_t *dst);
216 int dmar_enable_translation(struct dmar_unit *unit);
217 int dmar_disable_translation(struct dmar_unit *unit);
218 int dmar_load_irt_ptr(struct dmar_unit *unit);
219 int dmar_enable_ir(struct dmar_unit *unit);
220 int dmar_disable_ir(struct dmar_unit *unit);
221 bool dmar_barrier_enter(struct dmar_unit *dmar, u_int barrier_id);
222 void dmar_barrier_exit(struct dmar_unit *dmar, u_int barrier_id);
223 uint64_t dmar_get_timeout(void);
224 void dmar_update_timeout(uint64_t newval);
225 
226 int dmar_fault_intr(void *arg);
227 void dmar_enable_fault_intr(struct dmar_unit *unit);
228 void dmar_disable_fault_intr(struct dmar_unit *unit);
229 int dmar_init_fault_log(struct dmar_unit *unit);
230 void dmar_fini_fault_log(struct dmar_unit *unit);
231 
232 int dmar_qi_intr(void *arg);
233 void dmar_enable_qi_intr(struct dmar_unit *unit);
234 void dmar_disable_qi_intr(struct dmar_unit *unit);
235 int dmar_init_qi(struct dmar_unit *unit);
236 void dmar_fini_qi(struct dmar_unit *unit);
237 void dmar_qi_invalidate_locked(struct dmar_domain *domain, iommu_gaddr_t start,
238     iommu_gaddr_t size, struct iommu_qi_genseq *psec, bool emit_wait);
239 void dmar_qi_invalidate_ctx_glob_locked(struct dmar_unit *unit);
240 void dmar_qi_invalidate_iotlb_glob_locked(struct dmar_unit *unit);
241 void dmar_qi_invalidate_iec_glob(struct dmar_unit *unit);
242 void dmar_qi_invalidate_iec(struct dmar_unit *unit, u_int start, u_int cnt);
243 
244 vm_object_t domain_get_idmap_pgtbl(struct dmar_domain *domain,
245     iommu_gaddr_t maxaddr);
246 void put_idmap_pgtbl(vm_object_t obj);
247 void domain_flush_iotlb_sync(struct dmar_domain *domain, iommu_gaddr_t base,
248     iommu_gaddr_t size);
249 int domain_alloc_pgtbl(struct dmar_domain *domain);
250 void domain_free_pgtbl(struct dmar_domain *domain);
251 void domain_pgtbl_init(struct dmar_domain *domain);
252 
253 int dmar_dev_depth(device_t child);
254 void dmar_dev_path(device_t child, int *busno, void *path1, int depth);
255 
256 struct dmar_ctx *dmar_get_ctx_for_dev(struct dmar_unit *dmar, device_t dev,
257     uint16_t rid, bool id_mapped, bool rmrr_init);
258 struct dmar_ctx *dmar_get_ctx_for_devpath(struct dmar_unit *dmar, uint16_t rid,
259     int dev_domain, int dev_busno, const void *dev_path, int dev_path_len,
260     bool id_mapped, bool rmrr_init);
261 int dmar_move_ctx_to_domain(struct dmar_domain *domain, struct dmar_ctx *ctx);
262 void dmar_free_ctx_locked(struct dmar_unit *dmar, struct dmar_ctx *ctx);
263 void dmar_free_ctx(struct dmar_ctx *ctx);
264 struct dmar_ctx *dmar_find_ctx_locked(struct dmar_unit *dmar, uint16_t rid);
265 void dmar_domain_unload_entry(struct iommu_map_entry *entry, bool free);
266 void dmar_domain_unload(struct dmar_domain *domain,
267     struct iommu_map_entries_tailq *entries, bool cansleep);
268 void dmar_domain_free_entry(struct iommu_map_entry *entry, bool free);
269 
270 void dmar_dev_parse_rmrr(struct dmar_domain *domain, int dev_domain,
271     int dev_busno, const void *dev_path, int dev_path_len,
272     struct iommu_map_entries_tailq *rmrr_entries);
273 int dmar_instantiate_rmrr_ctxs(struct iommu_unit *dmar);
274 
275 void dmar_quirks_post_ident(struct dmar_unit *dmar);
276 void dmar_quirks_pre_use(struct iommu_unit *dmar);
277 
278 int dmar_init_irt(struct dmar_unit *unit);
279 void dmar_fini_irt(struct dmar_unit *unit);
280 
281 extern iommu_haddr_t dmar_high;
282 extern int haw;
283 extern int dmar_tbl_pagecnt;
284 extern int dmar_batch_coalesce;
285 
286 static inline uint32_t
287 dmar_read4(const struct dmar_unit *unit, int reg)
288 {
289 
290 	return (bus_read_4(unit->regs, reg));
291 }
292 
293 static inline uint64_t
294 dmar_read8(const struct dmar_unit *unit, int reg)
295 {
296 #ifdef __i386__
297 	uint32_t high, low;
298 
299 	low = bus_read_4(unit->regs, reg);
300 	high = bus_read_4(unit->regs, reg + 4);
301 	return (low | ((uint64_t)high << 32));
302 #else
303 	return (bus_read_8(unit->regs, reg));
304 #endif
305 }
306 
307 static inline void
308 dmar_write4(const struct dmar_unit *unit, int reg, uint32_t val)
309 {
310 
311 	KASSERT(reg != DMAR_GCMD_REG || (val & DMAR_GCMD_TE) ==
312 	    (unit->hw_gcmd & DMAR_GCMD_TE),
313 	    ("dmar%d clearing TE 0x%08x 0x%08x", unit->iommu.unit,
314 	    unit->hw_gcmd, val));
315 	bus_write_4(unit->regs, reg, val);
316 }
317 
318 static inline void
319 dmar_write8(const struct dmar_unit *unit, int reg, uint64_t val)
320 {
321 
322 	KASSERT(reg != DMAR_GCMD_REG, ("8byte GCMD write"));
323 #ifdef __i386__
324 	uint32_t high, low;
325 
326 	low = val;
327 	high = val >> 32;
328 	bus_write_4(unit->regs, reg, low);
329 	bus_write_4(unit->regs, reg + 4, high);
330 #else
331 	bus_write_8(unit->regs, reg, val);
332 #endif
333 }
334 
335 /*
336  * dmar_pte_store and dmar_pte_clear ensure that on i386, 32bit writes
337  * are issued in the correct order.  For store, the lower word,
338  * containing the P or R and W bits, is set only after the high word
339  * is written.  For clear, the P bit is cleared first, then the high
340  * word is cleared.
341  *
342  * dmar_pte_update updates the pte.  For amd64, the update is atomic.
343  * For i386, it first disables the entry by clearing the word
344  * containing the P bit, and then defer to dmar_pte_store.  The locked
345  * cmpxchg8b is probably available on any machine having DMAR support,
346  * but interrupt translation table may be mapped uncached.
347  */
348 static inline void
349 dmar_pte_store1(volatile uint64_t *dst, uint64_t val)
350 {
351 #ifdef __i386__
352 	volatile uint32_t *p;
353 	uint32_t hi, lo;
354 
355 	hi = val >> 32;
356 	lo = val;
357 	p = (volatile uint32_t *)dst;
358 	*(p + 1) = hi;
359 	*p = lo;
360 #else
361 	*dst = val;
362 #endif
363 }
364 
365 static inline void
366 dmar_pte_store(volatile uint64_t *dst, uint64_t val)
367 {
368 
369 	KASSERT(*dst == 0, ("used pte %p oldval %jx newval %jx",
370 	    dst, (uintmax_t)*dst, (uintmax_t)val));
371 	dmar_pte_store1(dst, val);
372 }
373 
374 static inline void
375 dmar_pte_update(volatile uint64_t *dst, uint64_t val)
376 {
377 
378 #ifdef __i386__
379 	volatile uint32_t *p;
380 
381 	p = (volatile uint32_t *)dst;
382 	*p = 0;
383 #endif
384 	dmar_pte_store1(dst, val);
385 }
386 
387 static inline void
388 dmar_pte_clear(volatile uint64_t *dst)
389 {
390 #ifdef __i386__
391 	volatile uint32_t *p;
392 
393 	p = (volatile uint32_t *)dst;
394 	*p = 0;
395 	*(p + 1) = 0;
396 #else
397 	*dst = 0;
398 #endif
399 }
400 
401 extern struct timespec dmar_hw_timeout;
402 
403 #define	DMAR_WAIT_UNTIL(cond)					\
404 {								\
405 	struct timespec last, curr;				\
406 	bool forever;						\
407 								\
408 	if (dmar_hw_timeout.tv_sec == 0 &&			\
409 	    dmar_hw_timeout.tv_nsec == 0) {			\
410 		forever = true;					\
411 	} else {						\
412 		forever = false;				\
413 		nanouptime(&curr);				\
414 		timespecadd(&curr, &dmar_hw_timeout, &last);	\
415 	}							\
416 	for (;;) {						\
417 		if (cond) {					\
418 			error = 0;				\
419 			break;					\
420 		}						\
421 		nanouptime(&curr);				\
422 		if (!forever && timespeccmp(&last, &curr, <)) {	\
423 			error = ETIMEDOUT;			\
424 			break;					\
425 		}						\
426 		cpu_spinwait();					\
427 	}							\
428 }
429 
430 #ifdef INVARIANTS
431 #define	TD_PREP_PINNED_ASSERT						\
432 	int old_td_pinned;						\
433 	old_td_pinned = curthread->td_pinned
434 #define	TD_PINNED_ASSERT						\
435 	KASSERT(curthread->td_pinned == old_td_pinned,			\
436 	    ("pin count leak: %d %d %s:%d", curthread->td_pinned,	\
437 	    old_td_pinned, __FILE__, __LINE__))
438 #else
439 #define	TD_PREP_PINNED_ASSERT
440 #define	TD_PINNED_ASSERT
441 #endif
442 
443 #endif
444