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