xref: /freebsd/sys/compat/linuxkpi/common/src/linux_page.c (revision e3b16f53a6455903a7e814045584fe203d4fff64) 
1 /*-
2  * Copyright (c) 2010 Isilon Systems, Inc.
3  * Copyright (c) 2016 Matthew Macy (mmacy@mattmacy.io)
4  * Copyright (c) 2017 Mellanox Technologies, Ltd.
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice unmodified, this list of conditions, and the following
12  *    disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18  * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20  * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21  * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22  * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26  * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27  */
28 
29 #include <sys/param.h>
30 #include <sys/systm.h>
31 #include <sys/malloc.h>
32 #include <sys/kernel.h>
33 #include <sys/sysctl.h>
34 #include <sys/lock.h>
35 #include <sys/mutex.h>
36 #include <sys/rwlock.h>
37 #include <sys/proc.h>
38 #include <sys/sched.h>
39 #include <sys/memrange.h>
40 
41 #include <machine/bus.h>
42 
43 #include <vm/vm.h>
44 #include <vm/pmap.h>
45 #include <vm/vm_param.h>
46 #include <vm/vm_kern.h>
47 #include <vm/vm_object.h>
48 #include <vm/vm_map.h>
49 #include <vm/vm_page.h>
50 #include <vm/vm_pageout.h>
51 #include <vm/vm_pager.h>
52 #include <vm/vm_radix.h>
53 #include <vm/vm_reserv.h>
54 #include <vm/vm_extern.h>
55 
56 #include <vm/uma.h>
57 #include <vm/uma_int.h>
58 
59 #include <linux/gfp.h>
60 #include <linux/mm.h>
61 #include <linux/preempt.h>
62 #include <linux/fs.h>
63 #include <linux/shmem_fs.h>
64 #include <linux/kernel.h>
65 #include <linux/idr.h>
66 #include <linux/io.h>
67 #include <linux/io-mapping.h>
68 
69 #ifdef __i386__
70 DEFINE_IDR(mtrr_idr);
71 static MALLOC_DEFINE(M_LKMTRR, "idr", "Linux MTRR compat");
72 extern int pat_works;
73 #endif
74 
75 void
76 si_meminfo(struct sysinfo *si)
77 {
78 	si->totalram = physmem;
79 	si->freeram = vm_free_count();
80 	si->totalhigh = 0;
81 	si->freehigh = 0;
82 	si->mem_unit = PAGE_SIZE;
83 }
84 
85 void *
86 linux_page_address(const struct page *page)
87 {
88 
89 	if (page->object != kernel_object) {
90 		return (PMAP_HAS_DMAP ?
91 		    ((void *)(uintptr_t)PHYS_TO_DMAP(page_to_phys(page))) :
92 		    NULL);
93 	}
94 	return ((void *)(uintptr_t)(VM_MIN_KERNEL_ADDRESS +
95 	    IDX_TO_OFF(page->pindex)));
96 }
97 
98 struct page *
99 linux_alloc_pages(gfp_t flags, unsigned int order)
100 {
101 	struct page *page;
102 
103 	if (PMAP_HAS_DMAP) {
104 		unsigned long npages = 1UL << order;
105 		int req = VM_ALLOC_WIRED;
106 
107 		if ((flags & M_ZERO) != 0)
108 			req |= VM_ALLOC_ZERO;
109 		if (order == 0 && (flags & GFP_DMA32) == 0) {
110 			page = vm_page_alloc_noobj(req);
111 			if (page == NULL)
112 				return (NULL);
113 		} else {
114 			vm_paddr_t pmax = (flags & GFP_DMA32) ?
115 			    BUS_SPACE_MAXADDR_32BIT : BUS_SPACE_MAXADDR;
116 		retry:
117 			page = vm_page_alloc_noobj_contig(req, npages, 0, pmax,
118 			    PAGE_SIZE, 0, VM_MEMATTR_DEFAULT);
119 			if (page == NULL) {
120 				if ((flags & (M_WAITOK | __GFP_NORETRY)) ==
121 				    M_WAITOK) {
122 					int err = vm_page_reclaim_contig(req,
123 					    npages, 0, pmax, PAGE_SIZE, 0);
124 					if (err == ENOMEM)
125 						vm_wait(NULL);
126 					else if (err != 0)
127 						return (NULL);
128 					flags &= ~M_WAITOK;
129 					goto retry;
130 				}
131 				return (NULL);
132 			}
133 		}
134 	} else {
135 		vm_offset_t vaddr;
136 
137 		vaddr = linux_alloc_kmem(flags, order);
138 		if (vaddr == 0)
139 			return (NULL);
140 
141 		page = virt_to_page((void *)vaddr);
142 
143 		KASSERT(vaddr == (vm_offset_t)page_address(page),
144 		    ("Page address mismatch"));
145 	}
146 
147 	return (page);
148 }
149 
150 static void
151 _linux_free_kmem(vm_offset_t addr, unsigned int order)
152 {
153 	size_t size = ((size_t)PAGE_SIZE) << order;
154 
155 	kmem_free((void *)addr, size);
156 }
157 
158 void
159 linux_free_pages(struct page *page, unsigned int order)
160 {
161 	if (PMAP_HAS_DMAP) {
162 		unsigned long npages = 1UL << order;
163 		unsigned long x;
164 
165 		for (x = 0; x != npages; x++) {
166 			vm_page_t pgo = page + x;
167 
168 			/*
169 			 * The "free page" function is used in several
170 			 * contexts.
171 			 *
172 			 * Some pages are allocated by `linux_alloc_pages()`
173 			 * above, but not all of them are. For instance in the
174 			 * DRM drivers, some pages come from
175 			 * `shmem_read_mapping_page_gfp()`.
176 			 *
177 			 * That's why we need to check if the page is managed
178 			 * or not here.
179 			 */
180 			if ((pgo->oflags & VPO_UNMANAGED) == 0) {
181 				vm_page_unwire(pgo, PQ_ACTIVE);
182 			} else {
183 				if (vm_page_unwire_noq(pgo))
184 					vm_page_free(pgo);
185 			}
186 		}
187 	} else {
188 		vm_offset_t vaddr;
189 
190 		vaddr = (vm_offset_t)page_address(page);
191 
192 		_linux_free_kmem(vaddr, order);
193 	}
194 }
195 
196 void
197 linux_release_pages(release_pages_arg arg, int nr)
198 {
199 	int i;
200 
201 	CTASSERT(offsetof(struct folio, page) == 0);
202 
203 	for (i = 0; i < nr; i++)
204 		__free_page(arg.pages[i]);
205 }
206 
207 vm_offset_t
208 linux_alloc_kmem(gfp_t flags, unsigned int order)
209 {
210 	size_t size = ((size_t)PAGE_SIZE) << order;
211 	void *addr;
212 
213 	addr = kmem_alloc_contig(size, flags & GFP_NATIVE_MASK, 0,
214 	    ((flags & GFP_DMA32) == 0) ? -1UL : BUS_SPACE_MAXADDR_32BIT,
215 	    PAGE_SIZE, 0, VM_MEMATTR_DEFAULT);
216 
217 	return ((vm_offset_t)addr);
218 }
219 
220 void
221 linux_free_kmem(vm_offset_t addr, unsigned int order)
222 {
223 	KASSERT((addr & ~PAGE_MASK) == 0,
224 	    ("%s: addr %p is not page aligned", __func__, (void *)addr));
225 
226 	if (addr >= VM_MIN_KERNEL_ADDRESS && addr < VM_MAX_KERNEL_ADDRESS) {
227 		_linux_free_kmem(addr, order);
228 	} else {
229 		vm_page_t page;
230 
231 		page = PHYS_TO_VM_PAGE(DMAP_TO_PHYS(addr));
232 		linux_free_pages(page, order);
233 	}
234 }
235 
236 static int
237 linux_get_user_pages_internal(vm_map_t map, unsigned long start, int nr_pages,
238     int write, struct page **pages)
239 {
240 	vm_prot_t prot;
241 	size_t len;
242 	int count;
243 
244 	prot = write ? (VM_PROT_READ | VM_PROT_WRITE) : VM_PROT_READ;
245 	len = ptoa((vm_offset_t)nr_pages);
246 	count = vm_fault_quick_hold_pages(map, start, len, prot, pages, nr_pages);
247 	return (count == -1 ? -EFAULT : nr_pages);
248 }
249 
250 int
251 __get_user_pages_fast(unsigned long start, int nr_pages, int write,
252     struct page **pages)
253 {
254 	vm_map_t map;
255 	vm_page_t *mp;
256 	vm_offset_t va;
257 	vm_offset_t end;
258 	vm_prot_t prot;
259 	int count;
260 
261 	if (nr_pages == 0 || in_interrupt())
262 		return (0);
263 
264 	MPASS(pages != NULL);
265 	map = &curthread->td_proc->p_vmspace->vm_map;
266 	end = start + ptoa((vm_offset_t)nr_pages);
267 	if (!vm_map_range_valid(map, start, end))
268 		return (-EINVAL);
269 	prot = write ? (VM_PROT_READ | VM_PROT_WRITE) : VM_PROT_READ;
270 	for (count = 0, mp = pages, va = start; va < end;
271 	    mp++, va += PAGE_SIZE, count++) {
272 		*mp = pmap_extract_and_hold(map->pmap, va, prot);
273 		if (*mp == NULL)
274 			break;
275 
276 		if ((prot & VM_PROT_WRITE) != 0 &&
277 		    (*mp)->dirty != VM_PAGE_BITS_ALL) {
278 			/*
279 			 * Explicitly dirty the physical page.  Otherwise, the
280 			 * caller's changes may go unnoticed because they are
281 			 * performed through an unmanaged mapping or by a DMA
282 			 * operation.
283 			 *
284 			 * The object lock is not held here.
285 			 * See vm_page_clear_dirty_mask().
286 			 */
287 			vm_page_dirty(*mp);
288 		}
289 	}
290 	return (count);
291 }
292 
293 long
294 get_user_pages_remote(struct task_struct *task, struct mm_struct *mm,
295     unsigned long start, unsigned long nr_pages, unsigned int gup_flags,
296     struct page **pages, struct vm_area_struct **vmas)
297 {
298 	vm_map_t map;
299 
300 	map = &task->task_thread->td_proc->p_vmspace->vm_map;
301 	return (linux_get_user_pages_internal(map, start, nr_pages,
302 	    !!(gup_flags & FOLL_WRITE), pages));
303 }
304 
305 long
306 lkpi_get_user_pages(unsigned long start, unsigned long nr_pages,
307     unsigned int gup_flags, struct page **pages)
308 {
309 	vm_map_t map;
310 
311 	map = &curthread->td_proc->p_vmspace->vm_map;
312 	return (linux_get_user_pages_internal(map, start, nr_pages,
313 	    !!(gup_flags & FOLL_WRITE), pages));
314 }
315 
316 int
317 is_vmalloc_addr(const void *addr)
318 {
319 	return (vtoslab((vm_offset_t)addr & ~UMA_SLAB_MASK) != NULL);
320 }
321 
322 vm_fault_t
323 lkpi_vmf_insert_pfn_prot_locked(struct vm_area_struct *vma, unsigned long addr,
324     unsigned long pfn, pgprot_t prot)
325 {
326 	struct pctrie_iter pages;
327 	vm_object_t vm_obj = vma->vm_obj;
328 	vm_object_t tmp_obj;
329 	vm_page_t page;
330 	vm_pindex_t pindex;
331 
332 	VM_OBJECT_ASSERT_WLOCKED(vm_obj);
333 	vm_page_iter_init(&pages, vm_obj);
334 	pindex = OFF_TO_IDX(addr - vma->vm_start);
335 	if (vma->vm_pfn_count == 0)
336 		vma->vm_pfn_first = pindex;
337 	MPASS(pindex <= OFF_TO_IDX(vma->vm_end));
338 
339 retry:
340 	page = vm_page_grab_iter(vm_obj, pindex, VM_ALLOC_NOCREAT, &pages);
341 	if (page == NULL) {
342 		page = PHYS_TO_VM_PAGE(IDX_TO_OFF(pfn));
343 		if (!vm_page_busy_acquire(page, VM_ALLOC_WAITFAIL)) {
344 			pctrie_iter_reset(&pages);
345 			goto retry;
346 		}
347 		if (page->object != NULL) {
348 			tmp_obj = page->object;
349 			vm_page_xunbusy(page);
350 			VM_OBJECT_WUNLOCK(vm_obj);
351 			VM_OBJECT_WLOCK(tmp_obj);
352 			if (page->object == tmp_obj &&
353 			    vm_page_busy_acquire(page, VM_ALLOC_WAITFAIL)) {
354 				KASSERT(page->object == tmp_obj,
355 				    ("page has changed identity"));
356 				KASSERT((page->oflags & VPO_UNMANAGED) == 0,
357 				    ("page does not belong to shmem"));
358 				vm_pager_page_unswapped(page);
359 				if (pmap_page_is_mapped(page)) {
360 					vm_page_xunbusy(page);
361 					VM_OBJECT_WUNLOCK(tmp_obj);
362 					printf("%s: page rename failed: page "
363 					    "is mapped\n", __func__);
364 					VM_OBJECT_WLOCK(vm_obj);
365 					return (VM_FAULT_NOPAGE);
366 				}
367 				vm_page_remove(page);
368 			}
369 			VM_OBJECT_WUNLOCK(tmp_obj);
370 			pctrie_iter_reset(&pages);
371 			VM_OBJECT_WLOCK(vm_obj);
372 			goto retry;
373 		}
374 		if (vm_page_iter_insert(page, vm_obj, pindex, &pages) != 0) {
375 			vm_page_xunbusy(page);
376 			return (VM_FAULT_OOM);
377 		}
378 		vm_page_valid(page);
379 	}
380 	pmap_page_set_memattr(page, pgprot2cachemode(prot));
381 	vma->vm_pfn_count++;
382 
383 	return (VM_FAULT_NOPAGE);
384 }
385 
386 int
387 lkpi_remap_pfn_range(struct vm_area_struct *vma, unsigned long start_addr,
388     unsigned long start_pfn, unsigned long size, pgprot_t prot)
389 {
390 	vm_object_t vm_obj;
391 	unsigned long addr, pfn;
392 	int err = 0;
393 
394 	vm_obj = vma->vm_obj;
395 
396 	VM_OBJECT_WLOCK(vm_obj);
397 	for (addr = start_addr, pfn = start_pfn;
398 	    addr < start_addr + size;
399 	    addr += PAGE_SIZE) {
400 		vm_fault_t ret;
401 retry:
402 		ret = lkpi_vmf_insert_pfn_prot_locked(vma, addr, pfn, prot);
403 
404 		if ((ret & VM_FAULT_OOM) != 0) {
405 			VM_OBJECT_WUNLOCK(vm_obj);
406 			vm_wait(NULL);
407 			VM_OBJECT_WLOCK(vm_obj);
408 			goto retry;
409 		}
410 
411 		if ((ret & VM_FAULT_ERROR) != 0) {
412 			err = -EFAULT;
413 			break;
414 		}
415 
416 		pfn++;
417 	}
418 	VM_OBJECT_WUNLOCK(vm_obj);
419 
420 	if (unlikely(err)) {
421 		zap_vma_ptes(vma, start_addr,
422 		    (pfn - start_pfn) << PAGE_SHIFT);
423 		return (err);
424 	}
425 
426 	return (0);
427 }
428 
429 int
430 lkpi_io_mapping_map_user(struct io_mapping *iomap,
431     struct vm_area_struct *vma, unsigned long addr,
432     unsigned long pfn, unsigned long size)
433 {
434 	pgprot_t prot;
435 	int ret;
436 
437 	prot = cachemode2protval(iomap->attr);
438 	ret = lkpi_remap_pfn_range(vma, addr, pfn, size, prot);
439 
440 	return (ret);
441 }
442 
443 /*
444  * Although FreeBSD version of unmap_mapping_range has semantics and types of
445  * parameters compatible with Linux version, the values passed in are different
446  * @obj should match to vm_private_data field of vm_area_struct returned by
447  *      mmap file operation handler, see linux_file_mmap_single() sources
448  * @holelen should match to size of area to be munmapped.
449  */
450 void
451 lkpi_unmap_mapping_range(void *obj, loff_t const holebegin __unused,
452     loff_t const holelen __unused, int even_cows __unused)
453 {
454 	vm_object_t devobj;
455 
456 	devobj = cdev_pager_lookup(obj);
457 	if (devobj != NULL) {
458 		cdev_mgtdev_pager_free_pages(devobj);
459 		vm_object_deallocate(devobj);
460 	}
461 }
462 
463 int
464 lkpi_arch_phys_wc_add(unsigned long base, unsigned long size)
465 {
466 #ifdef __i386__
467 	struct mem_range_desc *mrdesc;
468 	int error, id, act;
469 
470 	/* If PAT is available, do nothing */
471 	if (pat_works)
472 		return (0);
473 
474 	mrdesc = malloc(sizeof(*mrdesc), M_LKMTRR, M_WAITOK);
475 	mrdesc->mr_base = base;
476 	mrdesc->mr_len = size;
477 	mrdesc->mr_flags = MDF_WRITECOMBINE;
478 	strlcpy(mrdesc->mr_owner, "drm", sizeof(mrdesc->mr_owner));
479 	act = MEMRANGE_SET_UPDATE;
480 	error = mem_range_attr_set(mrdesc, &act);
481 	if (error == 0) {
482 		error = idr_get_new(&mtrr_idr, mrdesc, &id);
483 		MPASS(idr_find(&mtrr_idr, id) == mrdesc);
484 		if (error != 0) {
485 			act = MEMRANGE_SET_REMOVE;
486 			mem_range_attr_set(mrdesc, &act);
487 		}
488 	}
489 	if (error != 0) {
490 		free(mrdesc, M_LKMTRR);
491 		pr_warn(
492 		    "Failed to add WC MTRR for [%p-%p]: %d; "
493 		    "performance may suffer\n",
494 		    (void *)base, (void *)(base + size - 1), error);
495 	} else
496 		pr_warn("Successfully added WC MTRR for [%p-%p]\n",
497 		    (void *)base, (void *)(base + size - 1));
498 
499 	return (error != 0 ? -error : id + __MTRR_ID_BASE);
500 #else
501 	return (0);
502 #endif
503 }
504 
505 void
506 lkpi_arch_phys_wc_del(int reg)
507 {
508 #ifdef __i386__
509 	struct mem_range_desc *mrdesc;
510 	int act;
511 
512 	/* Check if arch_phys_wc_add() failed. */
513 	if (reg < __MTRR_ID_BASE)
514 		return;
515 
516 	mrdesc = idr_find(&mtrr_idr, reg - __MTRR_ID_BASE);
517 	MPASS(mrdesc != NULL);
518 	idr_remove(&mtrr_idr, reg - __MTRR_ID_BASE);
519 	act = MEMRANGE_SET_REMOVE;
520 	mem_range_attr_set(mrdesc, &act);
521 	free(mrdesc, M_LKMTRR);
522 #endif
523 }
524 
525 /*
526  * This is a highly simplified version of the Linux page_frag_cache.
527  * We only support up-to 1 single page as fragment size and we will
528  * always return a full page.  This may be wasteful on small objects
529  * but the only known consumer (mt76) is either asking for a half-page
530  * or a full page.  If this was to become a problem we can implement
531  * a more elaborate version.
532  */
533 void *
534 linuxkpi_page_frag_alloc(struct page_frag_cache *pfc,
535     size_t fragsz, gfp_t gfp)
536 {
537 	vm_page_t pages;
538 
539 	if (fragsz == 0)
540 		return (NULL);
541 
542 	KASSERT(fragsz <= PAGE_SIZE, ("%s: fragsz %zu > PAGE_SIZE not yet "
543 	    "supported", __func__, fragsz));
544 
545 	pages = alloc_pages(gfp, flsl(howmany(fragsz, PAGE_SIZE) - 1));
546 	if (pages == NULL)
547 		return (NULL);
548 	pfc->va = linux_page_address(pages);
549 
550 	/* Passed in as "count" to __page_frag_cache_drain(). Unused by us. */
551 	pfc->pagecnt_bias = 0;
552 
553 	return (pfc->va);
554 }
555 
556 void
557 linuxkpi_page_frag_free(void *addr)
558 {
559 	vm_page_t page;
560 
561 	page = virt_to_page(addr);
562 	linux_free_pages(page, 0);
563 }
564 
565 void
566 linuxkpi__page_frag_cache_drain(struct page *page, size_t count __unused)
567 {
568 
569 	linux_free_pages(page, 0);
570 }
571