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