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/cdefs.h>
30 #include <sys/param.h>
31 #include <sys/systm.h>
32 #include <sys/malloc.h>
33 #include <sys/kernel.h>
34 #include <sys/sysctl.h>
35 #include <sys/lock.h>
36 #include <sys/mutex.h>
37 #include <sys/rwlock.h>
38 #include <sys/proc.h>
39 #include <sys/sched.h>
40 #include <sys/memrange.h>
41 
42 #include <machine/bus.h>
43 
44 #include <vm/vm.h>
45 #include <vm/pmap.h>
46 #include <vm/vm_param.h>
47 #include <vm/vm_kern.h>
48 #include <vm/vm_object.h>
49 #include <vm/vm_map.h>
50 #include <vm/vm_page.h>
51 #include <vm/vm_pageout.h>
52 #include <vm/vm_pager.h>
53 #include <vm/vm_radix.h>
54 #include <vm/vm_reserv.h>
55 #include <vm/vm_extern.h>
56 
57 #include <vm/uma.h>
58 #include <vm/uma_int.h>
59 
60 #include <linux/gfp.h>
61 #include <linux/mm.h>
62 #include <linux/preempt.h>
63 #include <linux/fs.h>
64 #include <linux/shmem_fs.h>
65 #include <linux/kernel.h>
66 #include <linux/idr.h>
67 #include <linux/io.h>
68 #include <linux/io-mapping.h>
69 
70 #ifdef __i386__
71 DEFINE_IDR(mtrr_idr);
72 static MALLOC_DEFINE(M_LKMTRR, "idr", "Linux MTRR compat");
73 extern int pat_works;
74 #endif
75 
76 void
77 si_meminfo(struct sysinfo *si)
78 {
79 	si->totalram = physmem;
80 	si->freeram = vm_free_count();
81 	si->totalhigh = 0;
82 	si->freehigh = 0;
83 	si->mem_unit = PAGE_SIZE;
84 }
85 
86 void *
87 linux_page_address(struct page *page)
88 {
89 
90 	if (page->object != kernel_object) {
91 		return (PMAP_HAS_DMAP ?
92 		    ((void *)(uintptr_t)PHYS_TO_DMAP(page_to_phys(page))) :
93 		    NULL);
94 	}
95 	return ((void *)(uintptr_t)(VM_MIN_KERNEL_ADDRESS +
96 	    IDX_TO_OFF(page->pindex)));
97 }
98 
99 struct page *
100 linux_alloc_pages(gfp_t flags, unsigned int order)
101 {
102 	struct page *page;
103 
104 	if (PMAP_HAS_DMAP) {
105 		unsigned long npages = 1UL << order;
106 		int req = VM_ALLOC_WIRED;
107 
108 		if ((flags & M_ZERO) != 0)
109 			req |= VM_ALLOC_ZERO;
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 		retry:
118 			page = vm_page_alloc_noobj_contig(req, npages, 0, pmax,
119 			    PAGE_SIZE, 0, VM_MEMATTR_DEFAULT);
120 			if (page == NULL) {
121 				if (flags & M_WAITOK) {
122 					if (!vm_page_reclaim_contig(req,
123 					    npages, 0, pmax, PAGE_SIZE, 0)) {
124 						vm_wait(NULL);
125 					}
126 					flags &= ~M_WAITOK;
127 					goto retry;
128 				}
129 				return (NULL);
130 			}
131 		}
132 	} else {
133 		vm_offset_t vaddr;
134 
135 		vaddr = linux_alloc_kmem(flags, order);
136 		if (vaddr == 0)
137 			return (NULL);
138 
139 		page = virt_to_page((void *)vaddr);
140 
141 		KASSERT(vaddr == (vm_offset_t)page_address(page),
142 		    ("Page address mismatch"));
143 	}
144 
145 	return (page);
146 }
147 
148 void
149 linux_free_pages(struct page *page, unsigned int order)
150 {
151 	if (PMAP_HAS_DMAP) {
152 		unsigned long npages = 1UL << order;
153 		unsigned long x;
154 
155 		for (x = 0; x != npages; x++) {
156 			vm_page_t pgo = page + x;
157 
158 			if (vm_page_unwire_noq(pgo))
159 				vm_page_free(pgo);
160 		}
161 	} else {
162 		vm_offset_t vaddr;
163 
164 		vaddr = (vm_offset_t)page_address(page);
165 
166 		linux_free_kmem(vaddr, order);
167 	}
168 }
169 
170 vm_offset_t
171 linux_alloc_kmem(gfp_t flags, unsigned int order)
172 {
173 	size_t size = ((size_t)PAGE_SIZE) << order;
174 	void *addr;
175 
176 	if ((flags & GFP_DMA32) == 0) {
177 		addr = kmem_malloc(size, flags & GFP_NATIVE_MASK);
178 	} else {
179 		addr = kmem_alloc_contig(size, flags & GFP_NATIVE_MASK, 0,
180 		    BUS_SPACE_MAXADDR_32BIT, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT);
181 	}
182 	return ((vm_offset_t)addr);
183 }
184 
185 void
186 linux_free_kmem(vm_offset_t addr, unsigned int order)
187 {
188 	size_t size = ((size_t)PAGE_SIZE) << order;
189 
190 	kmem_free((void *)addr, size);
191 }
192 
193 static int
194 linux_get_user_pages_internal(vm_map_t map, unsigned long start, int nr_pages,
195     int write, struct page **pages)
196 {
197 	vm_prot_t prot;
198 	size_t len;
199 	int count;
200 
201 	prot = write ? (VM_PROT_READ | VM_PROT_WRITE) : VM_PROT_READ;
202 	len = ptoa((vm_offset_t)nr_pages);
203 	count = vm_fault_quick_hold_pages(map, start, len, prot, pages, nr_pages);
204 	return (count == -1 ? -EFAULT : nr_pages);
205 }
206 
207 int
208 __get_user_pages_fast(unsigned long start, int nr_pages, int write,
209     struct page **pages)
210 {
211 	vm_map_t map;
212 	vm_page_t *mp;
213 	vm_offset_t va;
214 	vm_offset_t end;
215 	vm_prot_t prot;
216 	int count;
217 
218 	if (nr_pages == 0 || in_interrupt())
219 		return (0);
220 
221 	MPASS(pages != NULL);
222 	map = &curthread->td_proc->p_vmspace->vm_map;
223 	end = start + ptoa((vm_offset_t)nr_pages);
224 	if (!vm_map_range_valid(map, start, end))
225 		return (-EINVAL);
226 	prot = write ? (VM_PROT_READ | VM_PROT_WRITE) : VM_PROT_READ;
227 	for (count = 0, mp = pages, va = start; va < end;
228 	    mp++, va += PAGE_SIZE, count++) {
229 		*mp = pmap_extract_and_hold(map->pmap, va, prot);
230 		if (*mp == NULL)
231 			break;
232 
233 		if ((prot & VM_PROT_WRITE) != 0 &&
234 		    (*mp)->dirty != VM_PAGE_BITS_ALL) {
235 			/*
236 			 * Explicitly dirty the physical page.  Otherwise, the
237 			 * caller's changes may go unnoticed because they are
238 			 * performed through an unmanaged mapping or by a DMA
239 			 * operation.
240 			 *
241 			 * The object lock is not held here.
242 			 * See vm_page_clear_dirty_mask().
243 			 */
244 			vm_page_dirty(*mp);
245 		}
246 	}
247 	return (count);
248 }
249 
250 long
251 get_user_pages_remote(struct task_struct *task, struct mm_struct *mm,
252     unsigned long start, unsigned long nr_pages, unsigned int gup_flags,
253     struct page **pages, struct vm_area_struct **vmas)
254 {
255 	vm_map_t map;
256 
257 	map = &task->task_thread->td_proc->p_vmspace->vm_map;
258 	return (linux_get_user_pages_internal(map, start, nr_pages,
259 	    !!(gup_flags & FOLL_WRITE), pages));
260 }
261 
262 long
263 get_user_pages(unsigned long start, unsigned long nr_pages,
264     unsigned int gup_flags, struct page **pages, struct vm_area_struct **vmas)
265 {
266 	vm_map_t map;
267 
268 	map = &curthread->td_proc->p_vmspace->vm_map;
269 	return (linux_get_user_pages_internal(map, start, nr_pages,
270 	    !!(gup_flags & FOLL_WRITE), pages));
271 }
272 
273 int
274 is_vmalloc_addr(const void *addr)
275 {
276 	return (vtoslab((vm_offset_t)addr & ~UMA_SLAB_MASK) != NULL);
277 }
278 
279 vm_fault_t
280 lkpi_vmf_insert_pfn_prot_locked(struct vm_area_struct *vma, unsigned long addr,
281     unsigned long pfn, pgprot_t prot)
282 {
283 	vm_object_t vm_obj = vma->vm_obj;
284 	vm_object_t tmp_obj;
285 	vm_page_t page;
286 	vm_pindex_t pindex;
287 
288 	VM_OBJECT_ASSERT_WLOCKED(vm_obj);
289 	pindex = OFF_TO_IDX(addr - vma->vm_start);
290 	if (vma->vm_pfn_count == 0)
291 		vma->vm_pfn_first = pindex;
292 	MPASS(pindex <= OFF_TO_IDX(vma->vm_end));
293 
294 retry:
295 	page = vm_page_grab(vm_obj, pindex, VM_ALLOC_NOCREAT);
296 	if (page == NULL) {
297 		page = PHYS_TO_VM_PAGE(IDX_TO_OFF(pfn));
298 		if (!vm_page_busy_acquire(page, VM_ALLOC_WAITFAIL))
299 			goto retry;
300 		if (page->object != NULL) {
301 			tmp_obj = page->object;
302 			vm_page_xunbusy(page);
303 			VM_OBJECT_WUNLOCK(vm_obj);
304 			VM_OBJECT_WLOCK(tmp_obj);
305 			if (page->object == tmp_obj &&
306 			    vm_page_busy_acquire(page, VM_ALLOC_WAITFAIL)) {
307 				KASSERT(page->object == tmp_obj,
308 				    ("page has changed identity"));
309 				KASSERT((page->oflags & VPO_UNMANAGED) == 0,
310 				    ("page does not belong to shmem"));
311 				vm_pager_page_unswapped(page);
312 				if (pmap_page_is_mapped(page)) {
313 					vm_page_xunbusy(page);
314 					VM_OBJECT_WUNLOCK(tmp_obj);
315 					printf("%s: page rename failed: page "
316 					    "is mapped\n", __func__);
317 					VM_OBJECT_WLOCK(vm_obj);
318 					return (VM_FAULT_NOPAGE);
319 				}
320 				vm_page_remove(page);
321 			}
322 			VM_OBJECT_WUNLOCK(tmp_obj);
323 			VM_OBJECT_WLOCK(vm_obj);
324 			goto retry;
325 		}
326 		if (vm_page_insert(page, vm_obj, pindex)) {
327 			vm_page_xunbusy(page);
328 			return (VM_FAULT_OOM);
329 		}
330 		vm_page_valid(page);
331 	}
332 	pmap_page_set_memattr(page, pgprot2cachemode(prot));
333 	vma->vm_pfn_count++;
334 
335 	return (VM_FAULT_NOPAGE);
336 }
337 
338 int
339 lkpi_remap_pfn_range(struct vm_area_struct *vma, unsigned long start_addr,
340     unsigned long start_pfn, unsigned long size, pgprot_t prot)
341 {
342 	vm_object_t vm_obj;
343 	unsigned long addr, pfn;
344 	int err = 0;
345 
346 	vm_obj = vma->vm_obj;
347 
348 	VM_OBJECT_WLOCK(vm_obj);
349 	for (addr = start_addr, pfn = start_pfn;
350 	    addr < start_addr + size;
351 	    addr += PAGE_SIZE) {
352 		vm_fault_t ret;
353 retry:
354 		ret = lkpi_vmf_insert_pfn_prot_locked(vma, addr, pfn, prot);
355 
356 		if ((ret & VM_FAULT_OOM) != 0) {
357 			VM_OBJECT_WUNLOCK(vm_obj);
358 			vm_wait(NULL);
359 			VM_OBJECT_WLOCK(vm_obj);
360 			goto retry;
361 		}
362 
363 		if ((ret & VM_FAULT_ERROR) != 0) {
364 			err = -EFAULT;
365 			break;
366 		}
367 
368 		pfn++;
369 	}
370 	VM_OBJECT_WUNLOCK(vm_obj);
371 
372 	if (unlikely(err)) {
373 		zap_vma_ptes(vma, start_addr,
374 		    (pfn - start_pfn) << PAGE_SHIFT);
375 		return (err);
376 	}
377 
378 	return (0);
379 }
380 
381 int
382 lkpi_io_mapping_map_user(struct io_mapping *iomap,
383     struct vm_area_struct *vma, unsigned long addr,
384     unsigned long pfn, unsigned long size)
385 {
386 	pgprot_t prot;
387 	int ret;
388 
389 	prot = cachemode2protval(iomap->attr);
390 	ret = lkpi_remap_pfn_range(vma, addr, pfn, size, prot);
391 
392 	return (ret);
393 }
394 
395 /*
396  * Although FreeBSD version of unmap_mapping_range has semantics and types of
397  * parameters compatible with Linux version, the values passed in are different
398  * @obj should match to vm_private_data field of vm_area_struct returned by
399  *      mmap file operation handler, see linux_file_mmap_single() sources
400  * @holelen should match to size of area to be munmapped.
401  */
402 void
403 lkpi_unmap_mapping_range(void *obj, loff_t const holebegin __unused,
404     loff_t const holelen, int even_cows __unused)
405 {
406 	vm_object_t devobj;
407 	vm_page_t page;
408 	int i, page_count;
409 
410 	devobj = cdev_pager_lookup(obj);
411 	if (devobj != NULL) {
412 		page_count = OFF_TO_IDX(holelen);
413 
414 		VM_OBJECT_WLOCK(devobj);
415 retry:
416 		for (i = 0; i < page_count; i++) {
417 			page = vm_page_lookup(devobj, i);
418 			if (page == NULL)
419 				continue;
420 			if (!vm_page_busy_acquire(page, VM_ALLOC_WAITFAIL))
421 				goto retry;
422 			cdev_pager_free_page(devobj, page);
423 		}
424 		VM_OBJECT_WUNLOCK(devobj);
425 		vm_object_deallocate(devobj);
426 	}
427 }
428 
429 int
430 lkpi_arch_phys_wc_add(unsigned long base, unsigned long size)
431 {
432 #ifdef __i386__
433 	struct mem_range_desc *mrdesc;
434 	int error, id, act;
435 
436 	/* If PAT is available, do nothing */
437 	if (pat_works)
438 		return (0);
439 
440 	mrdesc = malloc(sizeof(*mrdesc), M_LKMTRR, M_WAITOK);
441 	mrdesc->mr_base = base;
442 	mrdesc->mr_len = size;
443 	mrdesc->mr_flags = MDF_WRITECOMBINE;
444 	strlcpy(mrdesc->mr_owner, "drm", sizeof(mrdesc->mr_owner));
445 	act = MEMRANGE_SET_UPDATE;
446 	error = mem_range_attr_set(mrdesc, &act);
447 	if (error == 0) {
448 		error = idr_get_new(&mtrr_idr, mrdesc, &id);
449 		MPASS(idr_find(&mtrr_idr, id) == mrdesc);
450 		if (error != 0) {
451 			act = MEMRANGE_SET_REMOVE;
452 			mem_range_attr_set(mrdesc, &act);
453 		}
454 	}
455 	if (error != 0) {
456 		free(mrdesc, M_LKMTRR);
457 		pr_warn(
458 		    "Failed to add WC MTRR for [%p-%p]: %d; "
459 		    "performance may suffer\n",
460 		    (void *)base, (void *)(base + size - 1), error);
461 	} else
462 		pr_warn("Successfully added WC MTRR for [%p-%p]\n",
463 		    (void *)base, (void *)(base + size - 1));
464 
465 	return (error != 0 ? -error : id + __MTRR_ID_BASE);
466 #else
467 	return (0);
468 #endif
469 }
470 
471 void
472 lkpi_arch_phys_wc_del(int reg)
473 {
474 #ifdef __i386__
475 	struct mem_range_desc *mrdesc;
476 	int act;
477 
478 	/* Check if arch_phys_wc_add() failed. */
479 	if (reg < __MTRR_ID_BASE)
480 		return;
481 
482 	mrdesc = idr_find(&mtrr_idr, reg - __MTRR_ID_BASE);
483 	MPASS(mrdesc != NULL);
484 	idr_remove(&mtrr_idr, reg - __MTRR_ID_BASE);
485 	act = MEMRANGE_SET_REMOVE;
486 	mem_range_attr_set(mrdesc, &act);
487 	free(mrdesc, M_LKMTRR);
488 #endif
489 }
490 
491 /*
492  * This is a highly simplified version of the Linux page_frag_cache.
493  * We only support up-to 1 single page as fragment size and we will
494  * always return a full page.  This may be wasteful on small objects
495  * but the only known consumer (mt76) is either asking for a half-page
496  * or a full page.  If this was to become a problem we can implement
497  * a more elaborate version.
498  */
499 void *
500 linuxkpi_page_frag_alloc(struct page_frag_cache *pfc,
501     size_t fragsz, gfp_t gfp)
502 {
503 	vm_page_t pages;
504 
505 	if (fragsz == 0)
506 		return (NULL);
507 
508 	KASSERT(fragsz <= PAGE_SIZE, ("%s: fragsz %zu > PAGE_SIZE not yet "
509 	    "supported", __func__, fragsz));
510 
511 	pages = alloc_pages(gfp, flsl(howmany(fragsz, PAGE_SIZE) - 1));
512 	if (pages == NULL)
513 		return (NULL);
514 	pfc->va = linux_page_address(pages);
515 
516 	/* Passed in as "count" to __page_frag_cache_drain(). Unused by us. */
517 	pfc->pagecnt_bias = 0;
518 
519 	return (pfc->va);
520 }
521 
522 void
523 linuxkpi_page_frag_free(void *addr)
524 {
525 	vm_page_t page;
526 
527 	page = virt_to_page(addr);
528 	linux_free_pages(page, 0);
529 }
530 
531 void
532 linuxkpi__page_frag_cache_drain(struct page *page, size_t count __unused)
533 {
534 
535 	linux_free_pages(page, 0);
536 }
537