1 // SPDX-License-Identifier: GPL-2.0
2 /* Copyright(c) 2016-2018 Intel Corporation. All rights reserved. */
3 #include <linux/memremap.h>
4 #include <linux/pagemap.h>
5 #include <linux/module.h>
6 #include <linux/device.h>
7 #include <linux/pfn_t.h>
8 #include <linux/cdev.h>
9 #include <linux/slab.h>
10 #include <linux/dax.h>
11 #include <linux/fs.h>
12 #include <linux/mm.h>
13 #include <linux/mman.h>
14 #include "dax-private.h"
15 #include "bus.h"
16
check_vma(struct dev_dax * dev_dax,struct vm_area_struct * vma,const char * func)17 static int check_vma(struct dev_dax *dev_dax, struct vm_area_struct *vma,
18 const char *func)
19 {
20 struct device *dev = &dev_dax->dev;
21 unsigned long mask;
22
23 if (!dax_alive(dev_dax->dax_dev))
24 return -ENXIO;
25
26 /* prevent private mappings from being established */
27 if ((vma->vm_flags & VM_MAYSHARE) != VM_MAYSHARE) {
28 dev_info_ratelimited(dev,
29 "%s: %s: fail, attempted private mapping\n",
30 current->comm, func);
31 return -EINVAL;
32 }
33
34 mask = dev_dax->align - 1;
35 if (vma->vm_start & mask || vma->vm_end & mask) {
36 dev_info_ratelimited(dev,
37 "%s: %s: fail, unaligned vma (%#lx - %#lx, %#lx)\n",
38 current->comm, func, vma->vm_start, vma->vm_end,
39 mask);
40 return -EINVAL;
41 }
42
43 if (!vma_is_dax(vma)) {
44 dev_info_ratelimited(dev,
45 "%s: %s: fail, vma is not DAX capable\n",
46 current->comm, func);
47 return -EINVAL;
48 }
49
50 return 0;
51 }
52
53 /* see "strong" declaration in tools/testing/nvdimm/dax-dev.c */
dax_pgoff_to_phys(struct dev_dax * dev_dax,pgoff_t pgoff,unsigned long size)54 __weak phys_addr_t dax_pgoff_to_phys(struct dev_dax *dev_dax, pgoff_t pgoff,
55 unsigned long size)
56 {
57 int i;
58
59 for (i = 0; i < dev_dax->nr_range; i++) {
60 struct dev_dax_range *dax_range = &dev_dax->ranges[i];
61 struct range *range = &dax_range->range;
62 unsigned long long pgoff_end;
63 phys_addr_t phys;
64
65 pgoff_end = dax_range->pgoff + PHYS_PFN(range_len(range)) - 1;
66 if (pgoff < dax_range->pgoff || pgoff > pgoff_end)
67 continue;
68 phys = PFN_PHYS(pgoff - dax_range->pgoff) + range->start;
69 if (phys + size - 1 <= range->end)
70 return phys;
71 break;
72 }
73 return -1;
74 }
75
dax_set_mapping(struct vm_fault * vmf,pfn_t pfn,unsigned long fault_size)76 static void dax_set_mapping(struct vm_fault *vmf, pfn_t pfn,
77 unsigned long fault_size)
78 {
79 unsigned long i, nr_pages = fault_size / PAGE_SIZE;
80 struct file *filp = vmf->vma->vm_file;
81 struct dev_dax *dev_dax = filp->private_data;
82 pgoff_t pgoff;
83
84 /* mapping is only set on the head */
85 if (dev_dax->pgmap->vmemmap_shift)
86 nr_pages = 1;
87
88 pgoff = linear_page_index(vmf->vma,
89 ALIGN(vmf->address, fault_size));
90
91 for (i = 0; i < nr_pages; i++) {
92 struct page *page = pfn_to_page(pfn_t_to_pfn(pfn) + i);
93
94 page = compound_head(page);
95 if (page->mapping)
96 continue;
97
98 page->mapping = filp->f_mapping;
99 page->index = pgoff + i;
100 }
101 }
102
__dev_dax_pte_fault(struct dev_dax * dev_dax,struct vm_fault * vmf)103 static vm_fault_t __dev_dax_pte_fault(struct dev_dax *dev_dax,
104 struct vm_fault *vmf)
105 {
106 struct device *dev = &dev_dax->dev;
107 phys_addr_t phys;
108 pfn_t pfn;
109 unsigned int fault_size = PAGE_SIZE;
110
111 if (check_vma(dev_dax, vmf->vma, __func__))
112 return VM_FAULT_SIGBUS;
113
114 if (dev_dax->align > PAGE_SIZE) {
115 dev_dbg(dev, "alignment (%#x) > fault size (%#x)\n",
116 dev_dax->align, fault_size);
117 return VM_FAULT_SIGBUS;
118 }
119
120 if (fault_size != dev_dax->align)
121 return VM_FAULT_SIGBUS;
122
123 phys = dax_pgoff_to_phys(dev_dax, vmf->pgoff, PAGE_SIZE);
124 if (phys == -1) {
125 dev_dbg(dev, "pgoff_to_phys(%#lx) failed\n", vmf->pgoff);
126 return VM_FAULT_SIGBUS;
127 }
128
129 pfn = phys_to_pfn_t(phys, PFN_DEV|PFN_MAP);
130
131 dax_set_mapping(vmf, pfn, fault_size);
132
133 return vmf_insert_mixed(vmf->vma, vmf->address, pfn);
134 }
135
__dev_dax_pmd_fault(struct dev_dax * dev_dax,struct vm_fault * vmf)136 static vm_fault_t __dev_dax_pmd_fault(struct dev_dax *dev_dax,
137 struct vm_fault *vmf)
138 {
139 unsigned long pmd_addr = vmf->address & PMD_MASK;
140 struct device *dev = &dev_dax->dev;
141 phys_addr_t phys;
142 pgoff_t pgoff;
143 pfn_t pfn;
144 unsigned int fault_size = PMD_SIZE;
145
146 if (check_vma(dev_dax, vmf->vma, __func__))
147 return VM_FAULT_SIGBUS;
148
149 if (dev_dax->align > PMD_SIZE) {
150 dev_dbg(dev, "alignment (%#x) > fault size (%#x)\n",
151 dev_dax->align, fault_size);
152 return VM_FAULT_SIGBUS;
153 }
154
155 if (fault_size < dev_dax->align)
156 return VM_FAULT_SIGBUS;
157 else if (fault_size > dev_dax->align)
158 return VM_FAULT_FALLBACK;
159
160 /* if we are outside of the VMA */
161 if (pmd_addr < vmf->vma->vm_start ||
162 (pmd_addr + PMD_SIZE) > vmf->vma->vm_end)
163 return VM_FAULT_SIGBUS;
164
165 pgoff = linear_page_index(vmf->vma, pmd_addr);
166 phys = dax_pgoff_to_phys(dev_dax, pgoff, PMD_SIZE);
167 if (phys == -1) {
168 dev_dbg(dev, "pgoff_to_phys(%#lx) failed\n", pgoff);
169 return VM_FAULT_SIGBUS;
170 }
171
172 pfn = phys_to_pfn_t(phys, PFN_DEV|PFN_MAP);
173
174 dax_set_mapping(vmf, pfn, fault_size);
175
176 return vmf_insert_pfn_pmd(vmf, pfn, vmf->flags & FAULT_FLAG_WRITE);
177 }
178
179 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
__dev_dax_pud_fault(struct dev_dax * dev_dax,struct vm_fault * vmf)180 static vm_fault_t __dev_dax_pud_fault(struct dev_dax *dev_dax,
181 struct vm_fault *vmf)
182 {
183 unsigned long pud_addr = vmf->address & PUD_MASK;
184 struct device *dev = &dev_dax->dev;
185 phys_addr_t phys;
186 pgoff_t pgoff;
187 pfn_t pfn;
188 unsigned int fault_size = PUD_SIZE;
189
190
191 if (check_vma(dev_dax, vmf->vma, __func__))
192 return VM_FAULT_SIGBUS;
193
194 if (dev_dax->align > PUD_SIZE) {
195 dev_dbg(dev, "alignment (%#x) > fault size (%#x)\n",
196 dev_dax->align, fault_size);
197 return VM_FAULT_SIGBUS;
198 }
199
200 if (fault_size < dev_dax->align)
201 return VM_FAULT_SIGBUS;
202 else if (fault_size > dev_dax->align)
203 return VM_FAULT_FALLBACK;
204
205 /* if we are outside of the VMA */
206 if (pud_addr < vmf->vma->vm_start ||
207 (pud_addr + PUD_SIZE) > vmf->vma->vm_end)
208 return VM_FAULT_SIGBUS;
209
210 pgoff = linear_page_index(vmf->vma, pud_addr);
211 phys = dax_pgoff_to_phys(dev_dax, pgoff, PUD_SIZE);
212 if (phys == -1) {
213 dev_dbg(dev, "pgoff_to_phys(%#lx) failed\n", pgoff);
214 return VM_FAULT_SIGBUS;
215 }
216
217 pfn = phys_to_pfn_t(phys, PFN_DEV|PFN_MAP);
218
219 dax_set_mapping(vmf, pfn, fault_size);
220
221 return vmf_insert_pfn_pud(vmf, pfn, vmf->flags & FAULT_FLAG_WRITE);
222 }
223 #else
__dev_dax_pud_fault(struct dev_dax * dev_dax,struct vm_fault * vmf)224 static vm_fault_t __dev_dax_pud_fault(struct dev_dax *dev_dax,
225 struct vm_fault *vmf)
226 {
227 return VM_FAULT_FALLBACK;
228 }
229 #endif /* !CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
230
dev_dax_huge_fault(struct vm_fault * vmf,unsigned int order)231 static vm_fault_t dev_dax_huge_fault(struct vm_fault *vmf, unsigned int order)
232 {
233 struct file *filp = vmf->vma->vm_file;
234 vm_fault_t rc = VM_FAULT_SIGBUS;
235 int id;
236 struct dev_dax *dev_dax = filp->private_data;
237
238 dev_dbg(&dev_dax->dev, "%s: %s (%#lx - %#lx) order:%d\n", current->comm,
239 (vmf->flags & FAULT_FLAG_WRITE) ? "write" : "read",
240 vmf->vma->vm_start, vmf->vma->vm_end, order);
241
242 id = dax_read_lock();
243 if (order == 0)
244 rc = __dev_dax_pte_fault(dev_dax, vmf);
245 else if (order == PMD_ORDER)
246 rc = __dev_dax_pmd_fault(dev_dax, vmf);
247 else if (order == PUD_ORDER)
248 rc = __dev_dax_pud_fault(dev_dax, vmf);
249 else
250 rc = VM_FAULT_SIGBUS;
251
252 dax_read_unlock(id);
253
254 return rc;
255 }
256
dev_dax_fault(struct vm_fault * vmf)257 static vm_fault_t dev_dax_fault(struct vm_fault *vmf)
258 {
259 return dev_dax_huge_fault(vmf, 0);
260 }
261
dev_dax_may_split(struct vm_area_struct * vma,unsigned long addr)262 static int dev_dax_may_split(struct vm_area_struct *vma, unsigned long addr)
263 {
264 struct file *filp = vma->vm_file;
265 struct dev_dax *dev_dax = filp->private_data;
266
267 if (!IS_ALIGNED(addr, dev_dax->align))
268 return -EINVAL;
269 return 0;
270 }
271
dev_dax_pagesize(struct vm_area_struct * vma)272 static unsigned long dev_dax_pagesize(struct vm_area_struct *vma)
273 {
274 struct file *filp = vma->vm_file;
275 struct dev_dax *dev_dax = filp->private_data;
276
277 return dev_dax->align;
278 }
279
280 static const struct vm_operations_struct dax_vm_ops = {
281 .fault = dev_dax_fault,
282 .huge_fault = dev_dax_huge_fault,
283 .may_split = dev_dax_may_split,
284 .pagesize = dev_dax_pagesize,
285 };
286
dax_mmap(struct file * filp,struct vm_area_struct * vma)287 static int dax_mmap(struct file *filp, struct vm_area_struct *vma)
288 {
289 struct dev_dax *dev_dax = filp->private_data;
290 int rc, id;
291
292 dev_dbg(&dev_dax->dev, "trace\n");
293
294 /*
295 * We lock to check dax_dev liveness and will re-check at
296 * fault time.
297 */
298 id = dax_read_lock();
299 rc = check_vma(dev_dax, vma, __func__);
300 dax_read_unlock(id);
301 if (rc)
302 return rc;
303
304 vma->vm_ops = &dax_vm_ops;
305 vm_flags_set(vma, VM_HUGEPAGE);
306 return 0;
307 }
308
309 /* return an unmapped area aligned to the dax region specified alignment */
dax_get_unmapped_area(struct file * filp,unsigned long addr,unsigned long len,unsigned long pgoff,unsigned long flags)310 static unsigned long dax_get_unmapped_area(struct file *filp,
311 unsigned long addr, unsigned long len, unsigned long pgoff,
312 unsigned long flags)
313 {
314 unsigned long off, off_end, off_align, len_align, addr_align, align;
315 struct dev_dax *dev_dax = filp ? filp->private_data : NULL;
316
317 if (!dev_dax || addr)
318 goto out;
319
320 align = dev_dax->align;
321 off = pgoff << PAGE_SHIFT;
322 off_end = off + len;
323 off_align = round_up(off, align);
324
325 if ((off_end <= off_align) || ((off_end - off_align) < align))
326 goto out;
327
328 len_align = len + align;
329 if ((off + len_align) < off)
330 goto out;
331
332 addr_align = current->mm->get_unmapped_area(filp, addr, len_align,
333 pgoff, flags);
334 if (!IS_ERR_VALUE(addr_align)) {
335 addr_align += (off - addr_align) & (align - 1);
336 return addr_align;
337 }
338 out:
339 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
340 }
341
342 static const struct address_space_operations dev_dax_aops = {
343 .dirty_folio = noop_dirty_folio,
344 };
345
dax_open(struct inode * inode,struct file * filp)346 static int dax_open(struct inode *inode, struct file *filp)
347 {
348 struct dax_device *dax_dev = inode_dax(inode);
349 struct inode *__dax_inode = dax_inode(dax_dev);
350 struct dev_dax *dev_dax = dax_get_private(dax_dev);
351
352 dev_dbg(&dev_dax->dev, "trace\n");
353 inode->i_mapping = __dax_inode->i_mapping;
354 inode->i_mapping->host = __dax_inode;
355 inode->i_mapping->a_ops = &dev_dax_aops;
356 filp->f_mapping = inode->i_mapping;
357 filp->f_wb_err = filemap_sample_wb_err(filp->f_mapping);
358 filp->f_sb_err = file_sample_sb_err(filp);
359 filp->private_data = dev_dax;
360 inode->i_flags = S_DAX;
361
362 return 0;
363 }
364
dax_release(struct inode * inode,struct file * filp)365 static int dax_release(struct inode *inode, struct file *filp)
366 {
367 struct dev_dax *dev_dax = filp->private_data;
368
369 dev_dbg(&dev_dax->dev, "trace\n");
370 return 0;
371 }
372
373 static const struct file_operations dax_fops = {
374 .llseek = noop_llseek,
375 .owner = THIS_MODULE,
376 .open = dax_open,
377 .release = dax_release,
378 .get_unmapped_area = dax_get_unmapped_area,
379 .mmap = dax_mmap,
380 .mmap_supported_flags = MAP_SYNC,
381 };
382
dev_dax_cdev_del(void * cdev)383 static void dev_dax_cdev_del(void *cdev)
384 {
385 cdev_del(cdev);
386 }
387
dev_dax_kill(void * dev_dax)388 static void dev_dax_kill(void *dev_dax)
389 {
390 kill_dev_dax(dev_dax);
391 }
392
dev_dax_probe(struct dev_dax * dev_dax)393 static int dev_dax_probe(struct dev_dax *dev_dax)
394 {
395 struct dax_device *dax_dev = dev_dax->dax_dev;
396 struct device *dev = &dev_dax->dev;
397 struct dev_pagemap *pgmap;
398 struct inode *inode;
399 struct cdev *cdev;
400 void *addr;
401 int rc, i;
402
403 if (static_dev_dax(dev_dax)) {
404 if (dev_dax->nr_range > 1) {
405 dev_warn(dev,
406 "static pgmap / multi-range device conflict\n");
407 return -EINVAL;
408 }
409
410 pgmap = dev_dax->pgmap;
411 } else {
412 if (dev_dax->pgmap) {
413 dev_warn(dev,
414 "dynamic-dax with pre-populated page map\n");
415 return -EINVAL;
416 }
417
418 pgmap = devm_kzalloc(dev,
419 struct_size(pgmap, ranges, dev_dax->nr_range - 1),
420 GFP_KERNEL);
421 if (!pgmap)
422 return -ENOMEM;
423
424 pgmap->nr_range = dev_dax->nr_range;
425 dev_dax->pgmap = pgmap;
426
427 for (i = 0; i < dev_dax->nr_range; i++) {
428 struct range *range = &dev_dax->ranges[i].range;
429 pgmap->ranges[i] = *range;
430 }
431 }
432
433 for (i = 0; i < dev_dax->nr_range; i++) {
434 struct range *range = &dev_dax->ranges[i].range;
435
436 if (!devm_request_mem_region(dev, range->start,
437 range_len(range), dev_name(dev))) {
438 dev_warn(dev, "mapping%d: %#llx-%#llx could not reserve range\n",
439 i, range->start, range->end);
440 return -EBUSY;
441 }
442 }
443
444 pgmap->type = MEMORY_DEVICE_GENERIC;
445 if (dev_dax->align > PAGE_SIZE)
446 pgmap->vmemmap_shift =
447 order_base_2(dev_dax->align >> PAGE_SHIFT);
448 addr = devm_memremap_pages(dev, pgmap);
449 if (IS_ERR(addr))
450 return PTR_ERR(addr);
451
452 inode = dax_inode(dax_dev);
453 cdev = inode->i_cdev;
454 cdev_init(cdev, &dax_fops);
455 cdev->owner = dev->driver->owner;
456 cdev_set_parent(cdev, &dev->kobj);
457 rc = cdev_add(cdev, dev->devt, 1);
458 if (rc)
459 return rc;
460
461 rc = devm_add_action_or_reset(dev, dev_dax_cdev_del, cdev);
462 if (rc)
463 return rc;
464
465 run_dax(dax_dev);
466 return devm_add_action_or_reset(dev, dev_dax_kill, dev_dax);
467 }
468
469 static struct dax_device_driver device_dax_driver = {
470 .probe = dev_dax_probe,
471 .type = DAXDRV_DEVICE_TYPE,
472 };
473
dax_init(void)474 static int __init dax_init(void)
475 {
476 return dax_driver_register(&device_dax_driver);
477 }
478
dax_exit(void)479 static void __exit dax_exit(void)
480 {
481 dax_driver_unregister(&device_dax_driver);
482 }
483
484 MODULE_AUTHOR("Intel Corporation");
485 MODULE_LICENSE("GPL v2");
486 module_init(dax_init);
487 module_exit(dax_exit);
488 MODULE_ALIAS_DAX_DEVICE(0);
489