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
si_meminfo(struct sysinfo * si)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 *
linux_page_address(const struct page * page)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 *
linux_alloc_pages(gfp_t flags,unsigned int order)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
110 if (order == 0 && (flags & GFP_DMA32) == 0) {
111 page = vm_page_alloc_noobj(req);
112 if (page == NULL)
113 return (NULL);
114 } else {
115 vm_paddr_t pmax = (flags & GFP_DMA32) ?
116 BUS_SPACE_MAXADDR_32BIT : BUS_SPACE_MAXADDR;
117
118 if ((flags & __GFP_NORETRY) != 0)
119 req |= VM_ALLOC_NORECLAIM;
120
121 retry:
122 page = vm_page_alloc_noobj_contig(req, npages, 0, pmax,
123 PAGE_SIZE, 0, VM_MEMATTR_DEFAULT);
124 if (page == NULL) {
125 if ((flags & (M_WAITOK | __GFP_NORETRY)) ==
126 M_WAITOK) {
127 int err = vm_page_reclaim_contig(req,
128 npages, 0, pmax, PAGE_SIZE, 0);
129 if (err == ENOMEM)
130 vm_wait(NULL);
131 else if (err != 0)
132 return (NULL);
133 flags &= ~M_WAITOK;
134 goto retry;
135 }
136 return (NULL);
137 }
138 }
139 } else {
140 vm_offset_t vaddr;
141
142 vaddr = linux_alloc_kmem(flags, order);
143 if (vaddr == 0)
144 return (NULL);
145
146 page = virt_to_page((void *)vaddr);
147
148 KASSERT(vaddr == (vm_offset_t)page_address(page),
149 ("Page address mismatch"));
150 }
151
152 return (page);
153 }
154
155 static void
_linux_free_kmem(vm_offset_t addr,unsigned int order)156 _linux_free_kmem(vm_offset_t addr, unsigned int order)
157 {
158 size_t size = ((size_t)PAGE_SIZE) << order;
159
160 kmem_free((void *)addr, size);
161 }
162
163 void
linux_free_pages(struct page * page,unsigned int order)164 linux_free_pages(struct page *page, unsigned int order)
165 {
166 if (PMAP_HAS_DMAP) {
167 unsigned long npages = 1UL << order;
168 unsigned long x;
169
170 for (x = 0; x != npages; x++) {
171 vm_page_t pgo = page + x;
172
173 /*
174 * The "free page" function is used in several
175 * contexts.
176 *
177 * Some pages are allocated by `linux_alloc_pages()`
178 * above, but not all of them are. For instance in the
179 * DRM drivers, some pages come from
180 * `shmem_read_mapping_page_gfp()`.
181 *
182 * That's why we need to check if the page is managed
183 * or not here.
184 */
185 if ((pgo->oflags & VPO_UNMANAGED) == 0) {
186 vm_page_unwire(pgo, PQ_ACTIVE);
187 } else {
188 if (vm_page_unwire_noq(pgo))
189 vm_page_free(pgo);
190 }
191 }
192 } else {
193 vm_offset_t vaddr;
194
195 vaddr = (vm_offset_t)page_address(page);
196
197 _linux_free_kmem(vaddr, order);
198 }
199 }
200
201 void
linux_release_pages(release_pages_arg arg,int nr)202 linux_release_pages(release_pages_arg arg, int nr)
203 {
204 int i;
205
206 CTASSERT(offsetof(struct folio, page) == 0);
207
208 for (i = 0; i < nr; i++)
209 __free_page(arg.pages[i]);
210 }
211
212 vm_offset_t
linux_alloc_kmem(gfp_t flags,unsigned int order)213 linux_alloc_kmem(gfp_t flags, unsigned int order)
214 {
215 size_t size = ((size_t)PAGE_SIZE) << order;
216 void *addr;
217
218 addr = kmem_alloc_contig(size, flags & GFP_NATIVE_MASK, 0,
219 ((flags & GFP_DMA32) == 0) ? -1UL : BUS_SPACE_MAXADDR_32BIT,
220 PAGE_SIZE, 0, VM_MEMATTR_DEFAULT);
221
222 return ((vm_offset_t)addr);
223 }
224
225 void
linux_free_kmem(vm_offset_t addr,unsigned int order)226 linux_free_kmem(vm_offset_t addr, unsigned int order)
227 {
228 KASSERT((addr & ~PAGE_MASK) == 0,
229 ("%s: addr %p is not page aligned", __func__, (void *)addr));
230
231 if (addr >= VM_MIN_KERNEL_ADDRESS && addr < VM_MAX_KERNEL_ADDRESS) {
232 _linux_free_kmem(addr, order);
233 } else {
234 vm_page_t page;
235
236 page = PHYS_TO_VM_PAGE(DMAP_TO_PHYS(addr));
237 linux_free_pages(page, order);
238 }
239 }
240
241 static int
linux_get_user_pages_internal(vm_map_t map,unsigned long start,int nr_pages,int write,struct page ** pages)242 linux_get_user_pages_internal(vm_map_t map, unsigned long start, int nr_pages,
243 int write, struct page **pages)
244 {
245 vm_prot_t prot;
246 size_t len;
247 int count;
248
249 prot = write ? (VM_PROT_READ | VM_PROT_WRITE) : VM_PROT_READ;
250 len = ptoa((vm_offset_t)nr_pages);
251 count = vm_fault_quick_hold_pages(map, start, len, prot, pages, nr_pages);
252 return (count == -1 ? -EFAULT : nr_pages);
253 }
254
255 int
__get_user_pages_fast(unsigned long start,int nr_pages,int write,struct page ** pages)256 __get_user_pages_fast(unsigned long start, int nr_pages, int write,
257 struct page **pages)
258 {
259 vm_map_t map;
260 vm_page_t *mp;
261 vm_offset_t va;
262 vm_offset_t end;
263 vm_prot_t prot;
264 int count;
265
266 if (nr_pages == 0 || in_interrupt())
267 return (0);
268
269 MPASS(pages != NULL);
270 map = &curthread->td_proc->p_vmspace->vm_map;
271 end = start + ptoa((vm_offset_t)nr_pages);
272 if (!vm_map_range_valid(map, start, end))
273 return (-EINVAL);
274 prot = write ? (VM_PROT_READ | VM_PROT_WRITE) : VM_PROT_READ;
275 for (count = 0, mp = pages, va = start; va < end;
276 mp++, va += PAGE_SIZE, count++) {
277 *mp = pmap_extract_and_hold(map->pmap, va, prot);
278 if (*mp == NULL)
279 break;
280
281 if ((prot & VM_PROT_WRITE) != 0 &&
282 (*mp)->dirty != VM_PAGE_BITS_ALL) {
283 /*
284 * Explicitly dirty the physical page. Otherwise, the
285 * caller's changes may go unnoticed because they are
286 * performed through an unmanaged mapping or by a DMA
287 * operation.
288 *
289 * The object lock is not held here.
290 * See vm_page_clear_dirty_mask().
291 */
292 vm_page_dirty(*mp);
293 }
294 }
295 return (count);
296 }
297
298 long
get_user_pages_remote(struct task_struct * task,struct mm_struct * mm,unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages,struct vm_area_struct ** vmas)299 get_user_pages_remote(struct task_struct *task, struct mm_struct *mm,
300 unsigned long start, unsigned long nr_pages, unsigned int gup_flags,
301 struct page **pages, struct vm_area_struct **vmas)
302 {
303 vm_map_t map;
304
305 map = &task->task_thread->td_proc->p_vmspace->vm_map;
306 return (linux_get_user_pages_internal(map, start, nr_pages,
307 !!(gup_flags & FOLL_WRITE), pages));
308 }
309
310 long
lkpi_get_user_pages(unsigned long start,unsigned long nr_pages,unsigned int gup_flags,struct page ** pages)311 lkpi_get_user_pages(unsigned long start, unsigned long nr_pages,
312 unsigned int gup_flags, struct page **pages)
313 {
314 vm_map_t map;
315
316 map = &curthread->td_proc->p_vmspace->vm_map;
317 return (linux_get_user_pages_internal(map, start, nr_pages,
318 !!(gup_flags & FOLL_WRITE), pages));
319 }
320
321 int
is_vmalloc_addr(const void * addr)322 is_vmalloc_addr(const void *addr)
323 {
324 return (vtoslab((vm_offset_t)addr & ~UMA_SLAB_MASK) != NULL);
325 }
326
327 vm_fault_t
lkpi_vmf_insert_pfn_prot_locked(struct vm_area_struct * vma,unsigned long addr,unsigned long pfn,pgprot_t prot)328 lkpi_vmf_insert_pfn_prot_locked(struct vm_area_struct *vma, unsigned long addr,
329 unsigned long pfn, pgprot_t prot)
330 {
331 struct pctrie_iter pages;
332 vm_object_t vm_obj = vma->vm_obj;
333 vm_object_t tmp_obj;
334 vm_page_t page;
335 vm_pindex_t pindex;
336
337 VM_OBJECT_ASSERT_WLOCKED(vm_obj);
338 vm_page_iter_init(&pages, vm_obj);
339 pindex = OFF_TO_IDX(addr - vma->vm_start);
340 if (vma->vm_pfn_count == 0)
341 vma->vm_pfn_first = pindex;
342 MPASS(pindex <= OFF_TO_IDX(vma->vm_end));
343
344 retry:
345 page = vm_page_grab_iter(vm_obj, pindex, VM_ALLOC_NOCREAT, &pages);
346 if (page == NULL) {
347 page = PHYS_TO_VM_PAGE(IDX_TO_OFF(pfn));
348 if (page == NULL) {
349 pctrie_iter_reset(&pages);
350 return (VM_FAULT_SIGBUS);
351 }
352 if (!vm_page_busy_acquire(page, VM_ALLOC_WAITFAIL)) {
353 pctrie_iter_reset(&pages);
354 goto retry;
355 }
356 if (page->object != NULL) {
357 tmp_obj = page->object;
358 vm_page_xunbusy(page);
359 VM_OBJECT_WUNLOCK(vm_obj);
360 VM_OBJECT_WLOCK(tmp_obj);
361 if (page->object == tmp_obj &&
362 vm_page_busy_acquire(page, VM_ALLOC_WAITFAIL)) {
363 KASSERT(page->object == tmp_obj,
364 ("page has changed identity"));
365 KASSERT((page->oflags & VPO_UNMANAGED) == 0,
366 ("page does not belong to shmem"));
367 vm_pager_page_unswapped(page);
368 if (pmap_page_is_mapped(page)) {
369 vm_page_xunbusy(page);
370 VM_OBJECT_WUNLOCK(tmp_obj);
371 printf("%s: page rename failed: page "
372 "is mapped\n", __func__);
373 VM_OBJECT_WLOCK(vm_obj);
374 return (VM_FAULT_NOPAGE);
375 }
376 vm_page_remove(page);
377 }
378 VM_OBJECT_WUNLOCK(tmp_obj);
379 pctrie_iter_reset(&pages);
380 VM_OBJECT_WLOCK(vm_obj);
381 goto retry;
382 }
383 if (vm_page_iter_insert(page, vm_obj, pindex, &pages) != 0) {
384 vm_page_xunbusy(page);
385 return (VM_FAULT_OOM);
386 }
387 vm_page_valid(page);
388 }
389 pmap_page_set_memattr(page, pgprot2cachemode(prot));
390 vma->vm_pfn_count++;
391
392 return (VM_FAULT_NOPAGE);
393 }
394
395 int
lkpi_remap_pfn_range(struct vm_area_struct * vma,unsigned long start_addr,unsigned long start_pfn,unsigned long size,pgprot_t prot)396 lkpi_remap_pfn_range(struct vm_area_struct *vma, unsigned long start_addr,
397 unsigned long start_pfn, unsigned long size, pgprot_t prot)
398 {
399 vm_object_t vm_obj;
400 unsigned long addr, pfn;
401 int err = 0;
402
403 vm_obj = vma->vm_obj;
404
405 VM_OBJECT_WLOCK(vm_obj);
406 for (addr = start_addr, pfn = start_pfn;
407 addr < start_addr + size;
408 addr += PAGE_SIZE) {
409 vm_fault_t ret;
410 retry:
411 ret = lkpi_vmf_insert_pfn_prot_locked(vma, addr, pfn, prot);
412
413 if ((ret & VM_FAULT_OOM) != 0) {
414 VM_OBJECT_WUNLOCK(vm_obj);
415 vm_wait(NULL);
416 VM_OBJECT_WLOCK(vm_obj);
417 goto retry;
418 }
419
420 if ((ret & VM_FAULT_ERROR) != 0) {
421 err = -EFAULT;
422 break;
423 }
424
425 pfn++;
426 }
427 VM_OBJECT_WUNLOCK(vm_obj);
428
429 if (unlikely(err)) {
430 zap_vma_ptes(vma, start_addr,
431 (pfn - start_pfn) << PAGE_SHIFT);
432 return (err);
433 }
434
435 return (0);
436 }
437
438 int
lkpi_io_mapping_map_user(struct io_mapping * iomap,struct vm_area_struct * vma,unsigned long addr,unsigned long pfn,unsigned long size)439 lkpi_io_mapping_map_user(struct io_mapping *iomap,
440 struct vm_area_struct *vma, unsigned long addr,
441 unsigned long pfn, unsigned long size)
442 {
443 pgprot_t prot;
444 int ret;
445
446 prot = cachemode2protval(iomap->attr);
447 ret = lkpi_remap_pfn_range(vma, addr, pfn, size, prot);
448
449 return (ret);
450 }
451
452 /*
453 * Although FreeBSD version of unmap_mapping_range has semantics and types of
454 * parameters compatible with Linux version, the values passed in are different
455 * @obj should match to vm_private_data field of vm_area_struct returned by
456 * mmap file operation handler, see linux_file_mmap_single() sources
457 * @holelen should match to size of area to be munmapped.
458 */
459 void
lkpi_unmap_mapping_range(void * obj,loff_t const holebegin __unused,loff_t const holelen __unused,int even_cows __unused)460 lkpi_unmap_mapping_range(void *obj, loff_t const holebegin __unused,
461 loff_t const holelen __unused, int even_cows __unused)
462 {
463 vm_object_t devobj;
464
465 devobj = cdev_pager_lookup(obj);
466 if (devobj != NULL) {
467 cdev_mgtdev_pager_free_pages(devobj);
468 vm_object_deallocate(devobj);
469 }
470 }
471
472 int
lkpi_arch_phys_wc_add(unsigned long base,unsigned long size)473 lkpi_arch_phys_wc_add(unsigned long base, unsigned long size)
474 {
475 #ifdef __i386__
476 struct mem_range_desc *mrdesc;
477 int error, id, act;
478
479 /* If PAT is available, do nothing */
480 if (pat_works)
481 return (0);
482
483 mrdesc = malloc(sizeof(*mrdesc), M_LKMTRR, M_WAITOK);
484 mrdesc->mr_base = base;
485 mrdesc->mr_len = size;
486 mrdesc->mr_flags = MDF_WRITECOMBINE;
487 strlcpy(mrdesc->mr_owner, "drm", sizeof(mrdesc->mr_owner));
488 act = MEMRANGE_SET_UPDATE;
489 error = mem_range_attr_set(mrdesc, &act);
490 if (error == 0) {
491 error = idr_get_new(&mtrr_idr, mrdesc, &id);
492 MPASS(idr_find(&mtrr_idr, id) == mrdesc);
493 if (error != 0) {
494 act = MEMRANGE_SET_REMOVE;
495 mem_range_attr_set(mrdesc, &act);
496 }
497 }
498 if (error != 0) {
499 free(mrdesc, M_LKMTRR);
500 pr_warn(
501 "Failed to add WC MTRR for [%p-%p]: %d; "
502 "performance may suffer\n",
503 (void *)base, (void *)(base + size - 1), error);
504 } else
505 pr_warn("Successfully added WC MTRR for [%p-%p]\n",
506 (void *)base, (void *)(base + size - 1));
507
508 return (error != 0 ? -error : id + __MTRR_ID_BASE);
509 #else
510 return (0);
511 #endif
512 }
513
514 void
lkpi_arch_phys_wc_del(int reg)515 lkpi_arch_phys_wc_del(int reg)
516 {
517 #ifdef __i386__
518 struct mem_range_desc *mrdesc;
519 int act;
520
521 /* Check if arch_phys_wc_add() failed. */
522 if (reg < __MTRR_ID_BASE)
523 return;
524
525 mrdesc = idr_find(&mtrr_idr, reg - __MTRR_ID_BASE);
526 MPASS(mrdesc != NULL);
527 idr_remove(&mtrr_idr, reg - __MTRR_ID_BASE);
528 act = MEMRANGE_SET_REMOVE;
529 mem_range_attr_set(mrdesc, &act);
530 free(mrdesc, M_LKMTRR);
531 #endif
532 }
533
534 /*
535 * This is a highly simplified version of the Linux page_frag_cache.
536 * We only support up-to 1 single page as fragment size and we will
537 * always return a full page. This may be wasteful on small objects
538 * but the only known consumer (mt76) is either asking for a half-page
539 * or a full page. If this was to become a problem we can implement
540 * a more elaborate version.
541 */
542 void *
linuxkpi_page_frag_alloc(struct page_frag_cache * pfc,size_t fragsz,gfp_t gfp)543 linuxkpi_page_frag_alloc(struct page_frag_cache *pfc,
544 size_t fragsz, gfp_t gfp)
545 {
546 vm_page_t pages;
547
548 if (fragsz == 0)
549 return (NULL);
550
551 KASSERT(fragsz <= PAGE_SIZE, ("%s: fragsz %zu > PAGE_SIZE not yet "
552 "supported", __func__, fragsz));
553
554 pages = alloc_pages(gfp, flsl(howmany(fragsz, PAGE_SIZE) - 1));
555 if (pages == NULL)
556 return (NULL);
557 pfc->va = linux_page_address(pages);
558
559 /* Passed in as "count" to __page_frag_cache_drain(). Unused by us. */
560 pfc->pagecnt_bias = 0;
561
562 return (pfc->va);
563 }
564
565 void
linuxkpi_page_frag_free(void * addr)566 linuxkpi_page_frag_free(void *addr)
567 {
568 vm_page_t page;
569
570 page = virt_to_page(addr);
571 linux_free_pages(page, 0);
572 }
573
574 void
linuxkpi__page_frag_cache_drain(struct page * page,size_t count __unused)575 linuxkpi__page_frag_cache_drain(struct page *page, size_t count __unused)
576 {
577
578 linux_free_pages(page, 0);
579 }
580