xref: /freebsd/sys/arm64/iommu/iommu_pmap.c (revision f7c32ed617858bcd22f8d1b03199099d50125721)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2020-2021 Ruslan Bukin <br@bsdpad.com>
5  * Copyright (c) 2014-2021 Andrew Turner
6  * Copyright (c) 2014-2016 The FreeBSD Foundation
7  * All rights reserved.
8  *
9  * This work was supported by Innovate UK project 105694, "Digital Security
10  * by Design (DSbD) Technology Platform Prototype".
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  */
33 
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
36 
37 /*
38  *	Manages physical address maps for ARM SMMUv3 and ARM Mali GPU.
39  */
40 
41 #include "opt_vm.h"
42 
43 #include <sys/param.h>
44 #include <sys/ktr.h>
45 #include <sys/mutex.h>
46 #include <sys/rwlock.h>
47 
48 #include <vm/vm.h>
49 #include <vm/vm_param.h>
50 #include <vm/vm_page.h>
51 #include <vm/vm_map.h>
52 #include <vm/vm_object.h>
53 #include <vm/vm_pageout.h>
54 #include <vm/vm_radix.h>
55 
56 #include <machine/machdep.h>
57 
58 #include <arm64/iommu/iommu_pmap.h>
59 #include <arm64/iommu/iommu_pte.h>
60 
61 #define	IOMMU_PAGE_SIZE		4096
62 
63 #define	NL0PG		(IOMMU_PAGE_SIZE/(sizeof (pd_entry_t)))
64 #define	NL1PG		(IOMMU_PAGE_SIZE/(sizeof (pd_entry_t)))
65 #define	NL2PG		(IOMMU_PAGE_SIZE/(sizeof (pd_entry_t)))
66 #define	NL3PG		(IOMMU_PAGE_SIZE/(sizeof (pt_entry_t)))
67 
68 #define	NUL0E		IOMMU_L0_ENTRIES
69 #define	NUL1E		(NUL0E * NL1PG)
70 #define	NUL2E		(NUL1E * NL2PG)
71 
72 #define	iommu_l0_pindex(v)	(NUL2E + NUL1E + ((v) >> IOMMU_L0_SHIFT))
73 #define	iommu_l1_pindex(v)	(NUL2E + ((v) >> IOMMU_L1_SHIFT))
74 #define	iommu_l2_pindex(v)	((v) >> IOMMU_L2_SHIFT)
75 
76 /* This code assumes all L1 DMAP entries will be used */
77 CTASSERT((DMAP_MIN_ADDRESS  & ~IOMMU_L0_OFFSET) == DMAP_MIN_ADDRESS);
78 CTASSERT((DMAP_MAX_ADDRESS  & ~IOMMU_L0_OFFSET) == DMAP_MAX_ADDRESS);
79 
80 static vm_page_t _pmap_alloc_l3(pmap_t pmap, vm_pindex_t ptepindex);
81 static void _pmap_unwire_l3(pmap_t pmap, vm_offset_t va, vm_page_t m,
82     struct spglist *free);
83 
84 /*
85  * These load the old table data and store the new value.
86  * They need to be atomic as the System MMU may write to the table at
87  * the same time as the CPU.
88  */
89 #define	pmap_load(table)		(*table)
90 #define	pmap_clear(table)		atomic_store_64(table, 0)
91 #define	pmap_store(table, entry)	atomic_store_64(table, entry)
92 
93 /********************/
94 /* Inline functions */
95 /********************/
96 
97 static __inline pd_entry_t *
98 pmap_l0(pmap_t pmap, vm_offset_t va)
99 {
100 
101 	return (&pmap->pm_l0[iommu_l0_index(va)]);
102 }
103 
104 static __inline pd_entry_t *
105 pmap_l0_to_l1(pd_entry_t *l0, vm_offset_t va)
106 {
107 	pd_entry_t *l1;
108 
109 	l1 = (pd_entry_t *)PHYS_TO_DMAP(pmap_load(l0) & ~ATTR_MASK);
110 	return (&l1[iommu_l1_index(va)]);
111 }
112 
113 static __inline pd_entry_t *
114 pmap_l1(pmap_t pmap, vm_offset_t va)
115 {
116 	pd_entry_t *l0;
117 
118 	l0 = pmap_l0(pmap, va);
119 	if ((pmap_load(l0) & ATTR_DESCR_MASK) != IOMMU_L0_TABLE)
120 		return (NULL);
121 
122 	return (pmap_l0_to_l1(l0, va));
123 }
124 
125 static __inline pd_entry_t *
126 pmap_l1_to_l2(pd_entry_t *l1p, vm_offset_t va)
127 {
128 	pd_entry_t l1, *l2p;
129 
130 	l1 = pmap_load(l1p);
131 
132 	/*
133 	 * The valid bit may be clear if pmap_update_entry() is concurrently
134 	 * modifying the entry, so for KVA only the entry type may be checked.
135 	 */
136 	KASSERT(va >= VM_MAX_USER_ADDRESS || (l1 & ATTR_DESCR_VALID) != 0,
137 	    ("%s: L1 entry %#lx for %#lx is invalid", __func__, l1, va));
138 	KASSERT((l1 & ATTR_DESCR_TYPE_MASK) == ATTR_DESCR_TYPE_TABLE,
139 	    ("%s: L1 entry %#lx for %#lx is a leaf", __func__, l1, va));
140 	l2p = (pd_entry_t *)PHYS_TO_DMAP(l1 & ~ATTR_MASK);
141 	return (&l2p[iommu_l2_index(va)]);
142 }
143 
144 static __inline pd_entry_t *
145 pmap_l2(pmap_t pmap, vm_offset_t va)
146 {
147 	pd_entry_t *l1;
148 
149 	l1 = pmap_l1(pmap, va);
150 	if ((pmap_load(l1) & ATTR_DESCR_MASK) != IOMMU_L1_TABLE)
151 		return (NULL);
152 
153 	return (pmap_l1_to_l2(l1, va));
154 }
155 
156 static __inline pt_entry_t *
157 pmap_l2_to_l3(pd_entry_t *l2p, vm_offset_t va)
158 {
159 	pd_entry_t l2;
160 	pt_entry_t *l3p;
161 
162 	l2 = pmap_load(l2p);
163 
164 	/*
165 	 * The valid bit may be clear if pmap_update_entry() is concurrently
166 	 * modifying the entry, so for KVA only the entry type may be checked.
167 	 */
168 	KASSERT(va >= VM_MAX_USER_ADDRESS || (l2 & ATTR_DESCR_VALID) != 0,
169 	    ("%s: L2 entry %#lx for %#lx is invalid", __func__, l2, va));
170 	KASSERT((l2 & ATTR_DESCR_TYPE_MASK) == ATTR_DESCR_TYPE_TABLE,
171 	    ("%s: L2 entry %#lx for %#lx is a leaf", __func__, l2, va));
172 	l3p = (pt_entry_t *)PHYS_TO_DMAP(l2 & ~ATTR_MASK);
173 	return (&l3p[iommu_l3_index(va)]);
174 }
175 
176 /*
177  * Returns the lowest valid pde for a given virtual address.
178  * The next level may or may not point to a valid page or block.
179  */
180 static __inline pd_entry_t *
181 pmap_pde(pmap_t pmap, vm_offset_t va, int *level)
182 {
183 	pd_entry_t *l0, *l1, *l2, desc;
184 
185 	l0 = pmap_l0(pmap, va);
186 	desc = pmap_load(l0) & ATTR_DESCR_MASK;
187 	if (desc != IOMMU_L0_TABLE) {
188 		*level = -1;
189 		return (NULL);
190 	}
191 
192 	l1 = pmap_l0_to_l1(l0, va);
193 	desc = pmap_load(l1) & ATTR_DESCR_MASK;
194 	if (desc != IOMMU_L1_TABLE) {
195 		*level = 0;
196 		return (l0);
197 	}
198 
199 	l2 = pmap_l1_to_l2(l1, va);
200 	desc = pmap_load(l2) & ATTR_DESCR_MASK;
201 	if (desc != IOMMU_L2_TABLE) {
202 		*level = 1;
203 		return (l1);
204 	}
205 
206 	*level = 2;
207 	return (l2);
208 }
209 
210 /*
211  * Returns the lowest valid pte block or table entry for a given virtual
212  * address. If there are no valid entries return NULL and set the level to
213  * the first invalid level.
214  */
215 static __inline pt_entry_t *
216 pmap_pte(pmap_t pmap, vm_offset_t va, int *level)
217 {
218 	pd_entry_t *l1, *l2, desc;
219 	pt_entry_t *l3;
220 
221 	l1 = pmap_l1(pmap, va);
222 	if (l1 == NULL) {
223 		*level = 0;
224 		return (NULL);
225 	}
226 	desc = pmap_load(l1) & ATTR_DESCR_MASK;
227 	if (desc == IOMMU_L1_BLOCK) {
228 		*level = 1;
229 		return (l1);
230 	}
231 
232 	if (desc != IOMMU_L1_TABLE) {
233 		*level = 1;
234 		return (NULL);
235 	}
236 
237 	l2 = pmap_l1_to_l2(l1, va);
238 	desc = pmap_load(l2) & ATTR_DESCR_MASK;
239 	if (desc == IOMMU_L2_BLOCK) {
240 		*level = 2;
241 		return (l2);
242 	}
243 
244 	if (desc != IOMMU_L2_TABLE) {
245 		*level = 2;
246 		return (NULL);
247 	}
248 
249 	*level = 3;
250 	l3 = pmap_l2_to_l3(l2, va);
251 	if ((pmap_load(l3) & ATTR_DESCR_MASK) != IOMMU_L3_PAGE)
252 		return (NULL);
253 
254 	return (l3);
255 }
256 
257 static __inline int
258 pmap_l3_valid(pt_entry_t l3)
259 {
260 
261 	return ((l3 & ATTR_DESCR_MASK) == IOMMU_L3_PAGE);
262 }
263 
264 CTASSERT(IOMMU_L1_BLOCK == IOMMU_L2_BLOCK);
265 
266 static __inline void
267 pmap_resident_count_inc(pmap_t pmap, int count)
268 {
269 
270 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
271 	pmap->pm_stats.resident_count += count;
272 }
273 
274 static __inline void
275 pmap_resident_count_dec(pmap_t pmap, int count)
276 {
277 
278 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
279 	KASSERT(pmap->pm_stats.resident_count >= count,
280 	    ("pmap %p resident count underflow %ld %d", pmap,
281 	    pmap->pm_stats.resident_count, count));
282 	pmap->pm_stats.resident_count -= count;
283 }
284 
285 /***************************************************
286  * Page table page management routines.....
287  ***************************************************/
288 /*
289  * Schedule the specified unused page table page to be freed.  Specifically,
290  * add the page to the specified list of pages that will be released to the
291  * physical memory manager after the TLB has been updated.
292  */
293 static __inline void
294 pmap_add_delayed_free_list(vm_page_t m, struct spglist *free,
295     boolean_t set_PG_ZERO)
296 {
297 
298 	if (set_PG_ZERO)
299 		m->flags |= PG_ZERO;
300 	else
301 		m->flags &= ~PG_ZERO;
302 	SLIST_INSERT_HEAD(free, m, plinks.s.ss);
303 }
304 
305 /***************************************************
306  * Low level mapping routines.....
307  ***************************************************/
308 
309 /*
310  * Decrements a page table page's reference count, which is used to record the
311  * number of valid page table entries within the page.  If the reference count
312  * drops to zero, then the page table page is unmapped.  Returns TRUE if the
313  * page table page was unmapped and FALSE otherwise.
314  */
315 static inline boolean_t
316 pmap_unwire_l3(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free)
317 {
318 
319 	--m->ref_count;
320 	if (m->ref_count == 0) {
321 		_pmap_unwire_l3(pmap, va, m, free);
322 		return (TRUE);
323 	} else
324 		return (FALSE);
325 }
326 
327 static void
328 _pmap_unwire_l3(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free)
329 {
330 
331 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
332 	/*
333 	 * unmap the page table page
334 	 */
335 	if (m->pindex >= (NUL2E + NUL1E)) {
336 		/* l1 page */
337 		pd_entry_t *l0;
338 
339 		l0 = pmap_l0(pmap, va);
340 		pmap_clear(l0);
341 	} else if (m->pindex >= NUL2E) {
342 		/* l2 page */
343 		pd_entry_t *l1;
344 
345 		l1 = pmap_l1(pmap, va);
346 		pmap_clear(l1);
347 	} else {
348 		/* l3 page */
349 		pd_entry_t *l2;
350 
351 		l2 = pmap_l2(pmap, va);
352 		pmap_clear(l2);
353 	}
354 	pmap_resident_count_dec(pmap, 1);
355 	if (m->pindex < NUL2E) {
356 		/* We just released an l3, unhold the matching l2 */
357 		pd_entry_t *l1, tl1;
358 		vm_page_t l2pg;
359 
360 		l1 = pmap_l1(pmap, va);
361 		tl1 = pmap_load(l1);
362 		l2pg = PHYS_TO_VM_PAGE(tl1 & ~ATTR_MASK);
363 		pmap_unwire_l3(pmap, va, l2pg, free);
364 	} else if (m->pindex < (NUL2E + NUL1E)) {
365 		/* We just released an l2, unhold the matching l1 */
366 		pd_entry_t *l0, tl0;
367 		vm_page_t l1pg;
368 
369 		l0 = pmap_l0(pmap, va);
370 		tl0 = pmap_load(l0);
371 		l1pg = PHYS_TO_VM_PAGE(tl0 & ~ATTR_MASK);
372 		pmap_unwire_l3(pmap, va, l1pg, free);
373 	}
374 
375 	/*
376 	 * Put page on a list so that it is released after
377 	 * *ALL* TLB shootdown is done
378 	 */
379 	pmap_add_delayed_free_list(m, free, TRUE);
380 }
381 
382 static int
383 iommu_pmap_pinit_levels(pmap_t pmap, int levels)
384 {
385 	vm_page_t m;
386 
387 	/*
388 	 * allocate the l0 page
389 	 */
390 	m = vm_page_alloc_noobj(VM_ALLOC_WAITOK | VM_ALLOC_WIRED |
391 	    VM_ALLOC_ZERO);
392 	pmap->pm_l0_paddr = VM_PAGE_TO_PHYS(m);
393 	pmap->pm_l0 = (pd_entry_t *)PHYS_TO_DMAP(pmap->pm_l0_paddr);
394 
395 	vm_radix_init(&pmap->pm_root);
396 	bzero(&pmap->pm_stats, sizeof(pmap->pm_stats));
397 
398 	MPASS(levels == 3 || levels == 4);
399 	pmap->pm_levels = levels;
400 
401 	/*
402 	 * Allocate the level 1 entry to use as the root. This will increase
403 	 * the refcount on the level 1 page so it won't be removed until
404 	 * pmap_release() is called.
405 	 */
406 	if (pmap->pm_levels == 3) {
407 		PMAP_LOCK(pmap);
408 		m = _pmap_alloc_l3(pmap, NUL2E + NUL1E);
409 		PMAP_UNLOCK(pmap);
410 	}
411 	pmap->pm_ttbr = VM_PAGE_TO_PHYS(m);
412 
413 	return (1);
414 }
415 
416 int
417 iommu_pmap_pinit(pmap_t pmap)
418 {
419 
420 	return (iommu_pmap_pinit_levels(pmap, 4));
421 }
422 
423 /*
424  * This routine is called if the desired page table page does not exist.
425  *
426  * If page table page allocation fails, this routine may sleep before
427  * returning NULL.  It sleeps only if a lock pointer was given.
428  *
429  * Note: If a page allocation fails at page table level two or three,
430  * one or two pages may be held during the wait, only to be released
431  * afterwards.  This conservative approach is easily argued to avoid
432  * race conditions.
433  */
434 static vm_page_t
435 _pmap_alloc_l3(pmap_t pmap, vm_pindex_t ptepindex)
436 {
437 	vm_page_t m, l1pg, l2pg;
438 
439 	PMAP_LOCK_ASSERT(pmap, MA_OWNED);
440 
441 	/*
442 	 * Allocate a page table page.
443 	 */
444 	if ((m = vm_page_alloc_noobj(VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) {
445 		/*
446 		 * Indicate the need to retry.  While waiting, the page table
447 		 * page may have been allocated.
448 		 */
449 		return (NULL);
450 	}
451 	m->pindex = ptepindex;
452 
453 	/*
454 	 * Because of AArch64's weak memory consistency model, we must have a
455 	 * barrier here to ensure that the stores for zeroing "m", whether by
456 	 * pmap_zero_page() or an earlier function, are visible before adding
457 	 * "m" to the page table.  Otherwise, a page table walk by another
458 	 * processor's MMU could see the mapping to "m" and a stale, non-zero
459 	 * PTE within "m".
460 	 */
461 	dmb(ishst);
462 
463 	/*
464 	 * Map the pagetable page into the process address space, if
465 	 * it isn't already there.
466 	 */
467 
468 	if (ptepindex >= (NUL2E + NUL1E)) {
469 		pd_entry_t *l0;
470 		vm_pindex_t l0index;
471 
472 		l0index = ptepindex - (NUL2E + NUL1E);
473 		l0 = &pmap->pm_l0[l0index];
474 		pmap_store(l0, VM_PAGE_TO_PHYS(m) | IOMMU_L0_TABLE);
475 	} else if (ptepindex >= NUL2E) {
476 		vm_pindex_t l0index, l1index;
477 		pd_entry_t *l0, *l1;
478 		pd_entry_t tl0;
479 
480 		l1index = ptepindex - NUL2E;
481 		l0index = l1index >> IOMMU_L0_ENTRIES_SHIFT;
482 
483 		l0 = &pmap->pm_l0[l0index];
484 		tl0 = pmap_load(l0);
485 		if (tl0 == 0) {
486 			/* recurse for allocating page dir */
487 			if (_pmap_alloc_l3(pmap, NUL2E + NUL1E + l0index)
488 			    == NULL) {
489 				vm_page_unwire_noq(m);
490 				vm_page_free_zero(m);
491 				return (NULL);
492 			}
493 		} else {
494 			l1pg = PHYS_TO_VM_PAGE(tl0 & ~ATTR_MASK);
495 			l1pg->ref_count++;
496 		}
497 
498 		l1 = (pd_entry_t *)PHYS_TO_DMAP(pmap_load(l0) & ~ATTR_MASK);
499 		l1 = &l1[ptepindex & Ln_ADDR_MASK];
500 		pmap_store(l1, VM_PAGE_TO_PHYS(m) | IOMMU_L1_TABLE);
501 	} else {
502 		vm_pindex_t l0index, l1index;
503 		pd_entry_t *l0, *l1, *l2;
504 		pd_entry_t tl0, tl1;
505 
506 		l1index = ptepindex >> Ln_ENTRIES_SHIFT;
507 		l0index = l1index >> IOMMU_L0_ENTRIES_SHIFT;
508 
509 		l0 = &pmap->pm_l0[l0index];
510 		tl0 = pmap_load(l0);
511 		if (tl0 == 0) {
512 			/* recurse for allocating page dir */
513 			if (_pmap_alloc_l3(pmap, NUL2E + l1index) == NULL) {
514 				vm_page_unwire_noq(m);
515 				vm_page_free_zero(m);
516 				return (NULL);
517 			}
518 			tl0 = pmap_load(l0);
519 			l1 = (pd_entry_t *)PHYS_TO_DMAP(tl0 & ~ATTR_MASK);
520 			l1 = &l1[l1index & Ln_ADDR_MASK];
521 		} else {
522 			l1 = (pd_entry_t *)PHYS_TO_DMAP(tl0 & ~ATTR_MASK);
523 			l1 = &l1[l1index & Ln_ADDR_MASK];
524 			tl1 = pmap_load(l1);
525 			if (tl1 == 0) {
526 				/* recurse for allocating page dir */
527 				if (_pmap_alloc_l3(pmap, NUL2E + l1index)
528 				    == NULL) {
529 					vm_page_unwire_noq(m);
530 					vm_page_free_zero(m);
531 					return (NULL);
532 				}
533 			} else {
534 				l2pg = PHYS_TO_VM_PAGE(tl1 & ~ATTR_MASK);
535 				l2pg->ref_count++;
536 			}
537 		}
538 
539 		l2 = (pd_entry_t *)PHYS_TO_DMAP(pmap_load(l1) & ~ATTR_MASK);
540 		l2 = &l2[ptepindex & Ln_ADDR_MASK];
541 		pmap_store(l2, VM_PAGE_TO_PHYS(m) | IOMMU_L2_TABLE);
542 	}
543 
544 	pmap_resident_count_inc(pmap, 1);
545 
546 	return (m);
547 }
548 
549 /***************************************************
550  * Pmap allocation/deallocation routines.
551  ***************************************************/
552 
553 /*
554  * Release any resources held by the given physical map.
555  * Called when a pmap initialized by pmap_pinit is being released.
556  * Should only be called if the map contains no valid mappings.
557  */
558 void
559 iommu_pmap_release(pmap_t pmap)
560 {
561 	boolean_t rv;
562 	struct spglist free;
563 	vm_page_t m;
564 
565 	if (pmap->pm_levels != 4) {
566 		KASSERT(pmap->pm_stats.resident_count == 1,
567 		    ("pmap_release: pmap resident count %ld != 0",
568 		    pmap->pm_stats.resident_count));
569 		KASSERT((pmap->pm_l0[0] & ATTR_DESCR_VALID) == ATTR_DESCR_VALID,
570 		    ("pmap_release: Invalid l0 entry: %lx", pmap->pm_l0[0]));
571 
572 		SLIST_INIT(&free);
573 		m = PHYS_TO_VM_PAGE(pmap->pm_ttbr);
574 		PMAP_LOCK(pmap);
575 		rv = pmap_unwire_l3(pmap, 0, m, &free);
576 		PMAP_UNLOCK(pmap);
577 		MPASS(rv == TRUE);
578 		vm_page_free_pages_toq(&free, true);
579 	}
580 
581 	KASSERT(pmap->pm_stats.resident_count == 0,
582 	    ("pmap_release: pmap resident count %ld != 0",
583 	    pmap->pm_stats.resident_count));
584 	KASSERT(vm_radix_is_empty(&pmap->pm_root),
585 	    ("pmap_release: pmap has reserved page table page(s)"));
586 
587 	m = PHYS_TO_VM_PAGE(pmap->pm_l0_paddr);
588 	vm_page_unwire_noq(m);
589 	vm_page_free_zero(m);
590 }
591 
592 /***************************************************
593  * page management routines.
594  ***************************************************/
595 
596 /*
597  * Add a single Mali GPU entry. This function does not sleep.
598  */
599 int
600 pmap_gpu_enter(pmap_t pmap, vm_offset_t va, vm_paddr_t pa,
601     vm_prot_t prot, u_int flags)
602 {
603 	pd_entry_t *pde;
604 	pt_entry_t new_l3, orig_l3;
605 	pt_entry_t *l3;
606 	vm_page_t mpte;
607 	pd_entry_t *l1p;
608 	pd_entry_t *l2p;
609 	int lvl;
610 	int rv;
611 
612 	KASSERT(pmap != kernel_pmap, ("kernel pmap used for GPU"));
613 	KASSERT(va < VM_MAXUSER_ADDRESS, ("wrong address space"));
614 	KASSERT((va & PAGE_MASK) == 0, ("va is misaligned"));
615 	KASSERT((pa & PAGE_MASK) == 0, ("pa is misaligned"));
616 
617 	new_l3 = (pt_entry_t)(pa | ATTR_SH(ATTR_SH_IS) | IOMMU_L3_BLOCK);
618 
619 	if ((prot & VM_PROT_WRITE) != 0)
620 		new_l3 |= ATTR_S2_S2AP(ATTR_S2_S2AP_WRITE);
621 	if ((prot & VM_PROT_READ) != 0)
622 		new_l3 |= ATTR_S2_S2AP(ATTR_S2_S2AP_READ);
623 	if ((prot & VM_PROT_EXECUTE) == 0)
624 		new_l3 |= ATTR_S2_XN(ATTR_S2_XN_ALL);
625 
626 	CTR2(KTR_PMAP, "pmap_gpu_enter: %.16lx -> %.16lx", va, pa);
627 
628 	PMAP_LOCK(pmap);
629 
630 	/*
631 	 * In the case that a page table page is not
632 	 * resident, we are creating it here.
633 	 */
634 retry:
635 	pde = pmap_pde(pmap, va, &lvl);
636 	if (pde != NULL && lvl == 2) {
637 		l3 = pmap_l2_to_l3(pde, va);
638 	} else {
639 		mpte = _pmap_alloc_l3(pmap, iommu_l2_pindex(va));
640 		if (mpte == NULL) {
641 			CTR0(KTR_PMAP, "pmap_enter: mpte == NULL");
642 			rv = KERN_RESOURCE_SHORTAGE;
643 			goto out;
644 		}
645 
646 		/*
647 		 * Ensure newly created l1, l2 are visible to GPU.
648 		 * l0 is already visible by similar call in panfrost driver.
649 		 * The cache entry for l3 handled below.
650 		 */
651 
652 		l1p = pmap_l1(pmap, va);
653 		l2p = pmap_l2(pmap, va);
654 		cpu_dcache_wb_range((vm_offset_t)l1p, sizeof(pd_entry_t));
655 		cpu_dcache_wb_range((vm_offset_t)l2p, sizeof(pd_entry_t));
656 
657 		goto retry;
658 	}
659 
660 	orig_l3 = pmap_load(l3);
661 	KASSERT(!pmap_l3_valid(orig_l3), ("l3 is valid"));
662 
663 	/* New mapping */
664 	pmap_store(l3, new_l3);
665 
666 	cpu_dcache_wb_range((vm_offset_t)l3, sizeof(pt_entry_t));
667 
668 	pmap_resident_count_inc(pmap, 1);
669 	dsb(ishst);
670 
671 	rv = KERN_SUCCESS;
672 out:
673 	PMAP_UNLOCK(pmap);
674 
675 	return (rv);
676 }
677 
678 /*
679  * Remove a single Mali GPU entry.
680  */
681 int
682 pmap_gpu_remove(pmap_t pmap, vm_offset_t va)
683 {
684 	pd_entry_t *pde;
685 	pt_entry_t *pte;
686 	int lvl;
687 	int rc;
688 
689 	KASSERT((va & PAGE_MASK) == 0, ("va is misaligned"));
690 	KASSERT(pmap != kernel_pmap, ("kernel pmap used for GPU"));
691 
692 	PMAP_LOCK(pmap);
693 
694 	pde = pmap_pde(pmap, va, &lvl);
695 	if (pde == NULL || lvl != 2) {
696 		rc = KERN_FAILURE;
697 		goto out;
698 	}
699 
700 	pte = pmap_l2_to_l3(pde, va);
701 
702 	pmap_resident_count_dec(pmap, 1);
703 	pmap_clear(pte);
704 	cpu_dcache_wb_range((vm_offset_t)pte, sizeof(pt_entry_t));
705 	rc = KERN_SUCCESS;
706 
707 out:
708 	PMAP_UNLOCK(pmap);
709 
710 	return (rc);
711 }
712 
713 /*
714  * Add a single SMMU entry. This function does not sleep.
715  */
716 int
717 pmap_smmu_enter(pmap_t pmap, vm_offset_t va, vm_paddr_t pa,
718     vm_prot_t prot, u_int flags)
719 {
720 	pd_entry_t *pde;
721 	pt_entry_t new_l3, orig_l3;
722 	pt_entry_t *l3;
723 	vm_page_t mpte;
724 	int lvl;
725 	int rv;
726 
727 	KASSERT(va < VM_MAXUSER_ADDRESS, ("wrong address space"));
728 
729 	va = trunc_page(va);
730 	new_l3 = (pt_entry_t)(pa | ATTR_DEFAULT |
731 	    ATTR_S1_IDX(VM_MEMATTR_DEVICE) | IOMMU_L3_PAGE);
732 	if ((prot & VM_PROT_WRITE) == 0)
733 		new_l3 |= ATTR_S1_AP(ATTR_S1_AP_RO);
734 	new_l3 |= ATTR_S1_XN; /* Execute never. */
735 	new_l3 |= ATTR_S1_AP(ATTR_S1_AP_USER);
736 	new_l3 |= ATTR_S1_nG; /* Non global. */
737 
738 	CTR2(KTR_PMAP, "pmap_senter: %.16lx -> %.16lx", va, pa);
739 
740 	PMAP_LOCK(pmap);
741 
742 	/*
743 	 * In the case that a page table page is not
744 	 * resident, we are creating it here.
745 	 */
746 retry:
747 	pde = pmap_pde(pmap, va, &lvl);
748 	if (pde != NULL && lvl == 2) {
749 		l3 = pmap_l2_to_l3(pde, va);
750 	} else {
751 		mpte = _pmap_alloc_l3(pmap, iommu_l2_pindex(va));
752 		if (mpte == NULL) {
753 			CTR0(KTR_PMAP, "pmap_enter: mpte == NULL");
754 			rv = KERN_RESOURCE_SHORTAGE;
755 			goto out;
756 		}
757 		goto retry;
758 	}
759 
760 	orig_l3 = pmap_load(l3);
761 	KASSERT(!pmap_l3_valid(orig_l3), ("l3 is valid"));
762 
763 	/* New mapping */
764 	pmap_store(l3, new_l3);
765 	pmap_resident_count_inc(pmap, 1);
766 	dsb(ishst);
767 
768 	rv = KERN_SUCCESS;
769 out:
770 	PMAP_UNLOCK(pmap);
771 
772 	return (rv);
773 }
774 
775 /*
776  * Remove a single SMMU entry.
777  */
778 int
779 pmap_smmu_remove(pmap_t pmap, vm_offset_t va)
780 {
781 	pt_entry_t *pte;
782 	int lvl;
783 	int rc;
784 
785 	PMAP_LOCK(pmap);
786 
787 	pte = pmap_pte(pmap, va, &lvl);
788 	KASSERT(lvl == 3,
789 	    ("Invalid SMMU pagetable level: %d != 3", lvl));
790 
791 	if (pte != NULL) {
792 		pmap_resident_count_dec(pmap, 1);
793 		pmap_clear(pte);
794 		rc = KERN_SUCCESS;
795 	} else
796 		rc = KERN_FAILURE;
797 
798 	PMAP_UNLOCK(pmap);
799 
800 	return (rc);
801 }
802 
803 /*
804  * Remove all the allocated L1, L2 pages from SMMU pmap.
805  * All the L3 entires must be cleared in advance, otherwise
806  * this function panics.
807  */
808 void
809 iommu_pmap_remove_pages(pmap_t pmap)
810 {
811 	pd_entry_t l0e, *l1, l1e, *l2, l2e;
812 	pt_entry_t *l3, l3e;
813 	vm_page_t m, m0, m1;
814 	vm_offset_t sva;
815 	vm_paddr_t pa;
816 	vm_paddr_t pa0;
817 	vm_paddr_t pa1;
818 	int i, j, k, l;
819 
820 	PMAP_LOCK(pmap);
821 
822 	for (sva = VM_MINUSER_ADDRESS, i = iommu_l0_index(sva);
823 	    (i < Ln_ENTRIES && sva < VM_MAXUSER_ADDRESS); i++) {
824 		l0e = pmap->pm_l0[i];
825 		if ((l0e & ATTR_DESCR_VALID) == 0) {
826 			sva += IOMMU_L0_SIZE;
827 			continue;
828 		}
829 		pa0 = l0e & ~ATTR_MASK;
830 		m0 = PHYS_TO_VM_PAGE(pa0);
831 		l1 = (pd_entry_t *)PHYS_TO_DMAP(pa0);
832 
833 		for (j = iommu_l1_index(sva); j < Ln_ENTRIES; j++) {
834 			l1e = l1[j];
835 			if ((l1e & ATTR_DESCR_VALID) == 0) {
836 				sva += IOMMU_L1_SIZE;
837 				continue;
838 			}
839 			if ((l1e & ATTR_DESCR_MASK) == IOMMU_L1_BLOCK) {
840 				sva += IOMMU_L1_SIZE;
841 				continue;
842 			}
843 			pa1 = l1e & ~ATTR_MASK;
844 			m1 = PHYS_TO_VM_PAGE(pa1);
845 			l2 = (pd_entry_t *)PHYS_TO_DMAP(pa1);
846 
847 			for (k = iommu_l2_index(sva); k < Ln_ENTRIES; k++) {
848 				l2e = l2[k];
849 				if ((l2e & ATTR_DESCR_VALID) == 0) {
850 					sva += IOMMU_L2_SIZE;
851 					continue;
852 				}
853 				pa = l2e & ~ATTR_MASK;
854 				m = PHYS_TO_VM_PAGE(pa);
855 				l3 = (pt_entry_t *)PHYS_TO_DMAP(pa);
856 
857 				for (l = iommu_l3_index(sva); l < Ln_ENTRIES;
858 				    l++, sva += IOMMU_L3_SIZE) {
859 					l3e = l3[l];
860 					if ((l3e & ATTR_DESCR_VALID) == 0)
861 						continue;
862 					panic("%s: l3e found for va %jx\n",
863 					    __func__, sva);
864 				}
865 
866 				vm_page_unwire_noq(m1);
867 				vm_page_unwire_noq(m);
868 				pmap_resident_count_dec(pmap, 1);
869 				vm_page_free(m);
870 				pmap_clear(&l2[k]);
871 			}
872 
873 			vm_page_unwire_noq(m0);
874 			pmap_resident_count_dec(pmap, 1);
875 			vm_page_free(m1);
876 			pmap_clear(&l1[j]);
877 		}
878 
879 		pmap_resident_count_dec(pmap, 1);
880 		vm_page_free(m0);
881 		pmap_clear(&pmap->pm_l0[i]);
882 	}
883 
884 	KASSERT(pmap->pm_stats.resident_count == 0,
885 	    ("Invalid resident count %jd", pmap->pm_stats.resident_count));
886 
887 	PMAP_UNLOCK(pmap);
888 }
889