Loading...
1/*
2 * Copyright(c) 2016 - 2017 Intel Corporation. All rights reserved.
3 *
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of version 2 of the GNU General Public License as
6 * published by the Free Software Foundation.
7 *
8 * This program is distributed in the hope that it will be useful, but
9 * WITHOUT ANY WARRANTY; without even the implied warranty of
10 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
11 * General Public License for more details.
12 */
13#include <linux/pagemap.h>
14#include <linux/module.h>
15#include <linux/device.h>
16#include <linux/pfn_t.h>
17#include <linux/cdev.h>
18#include <linux/slab.h>
19#include <linux/dax.h>
20#include <linux/fs.h>
21#include <linux/mm.h>
22#include <linux/mman.h>
23#include "dax-private.h"
24#include "dax.h"
25
26static struct class *dax_class;
27
28/*
29 * Rely on the fact that drvdata is set before the attributes are
30 * registered, and that the attributes are unregistered before drvdata
31 * is cleared to assume that drvdata is always valid.
32 */
33static ssize_t id_show(struct device *dev,
34 struct device_attribute *attr, char *buf)
35{
36 struct dax_region *dax_region = dev_get_drvdata(dev);
37
38 return sprintf(buf, "%d\n", dax_region->id);
39}
40static DEVICE_ATTR_RO(id);
41
42static ssize_t region_size_show(struct device *dev,
43 struct device_attribute *attr, char *buf)
44{
45 struct dax_region *dax_region = dev_get_drvdata(dev);
46
47 return sprintf(buf, "%llu\n", (unsigned long long)
48 resource_size(&dax_region->res));
49}
50static struct device_attribute dev_attr_region_size = __ATTR(size, 0444,
51 region_size_show, NULL);
52
53static ssize_t align_show(struct device *dev,
54 struct device_attribute *attr, char *buf)
55{
56 struct dax_region *dax_region = dev_get_drvdata(dev);
57
58 return sprintf(buf, "%u\n", dax_region->align);
59}
60static DEVICE_ATTR_RO(align);
61
62static struct attribute *dax_region_attributes[] = {
63 &dev_attr_region_size.attr,
64 &dev_attr_align.attr,
65 &dev_attr_id.attr,
66 NULL,
67};
68
69static const struct attribute_group dax_region_attribute_group = {
70 .name = "dax_region",
71 .attrs = dax_region_attributes,
72};
73
74static const struct attribute_group *dax_region_attribute_groups[] = {
75 &dax_region_attribute_group,
76 NULL,
77};
78
79static void dax_region_free(struct kref *kref)
80{
81 struct dax_region *dax_region;
82
83 dax_region = container_of(kref, struct dax_region, kref);
84 kfree(dax_region);
85}
86
87void dax_region_put(struct dax_region *dax_region)
88{
89 kref_put(&dax_region->kref, dax_region_free);
90}
91EXPORT_SYMBOL_GPL(dax_region_put);
92
93static void dax_region_unregister(void *region)
94{
95 struct dax_region *dax_region = region;
96
97 sysfs_remove_groups(&dax_region->dev->kobj,
98 dax_region_attribute_groups);
99 dax_region_put(dax_region);
100}
101
102struct dax_region *alloc_dax_region(struct device *parent, int region_id,
103 struct resource *res, unsigned int align, void *addr,
104 unsigned long pfn_flags)
105{
106 struct dax_region *dax_region;
107
108 /*
109 * The DAX core assumes that it can store its private data in
110 * parent->driver_data. This WARN is a reminder / safeguard for
111 * developers of device-dax drivers.
112 */
113 if (dev_get_drvdata(parent)) {
114 dev_WARN(parent, "dax core failed to setup private data\n");
115 return NULL;
116 }
117
118 if (!IS_ALIGNED(res->start, align)
119 || !IS_ALIGNED(resource_size(res), align))
120 return NULL;
121
122 dax_region = kzalloc(sizeof(*dax_region), GFP_KERNEL);
123 if (!dax_region)
124 return NULL;
125
126 dev_set_drvdata(parent, dax_region);
127 memcpy(&dax_region->res, res, sizeof(*res));
128 dax_region->pfn_flags = pfn_flags;
129 kref_init(&dax_region->kref);
130 dax_region->id = region_id;
131 ida_init(&dax_region->ida);
132 dax_region->align = align;
133 dax_region->dev = parent;
134 dax_region->base = addr;
135 if (sysfs_create_groups(&parent->kobj, dax_region_attribute_groups)) {
136 kfree(dax_region);
137 return NULL;
138 }
139
140 kref_get(&dax_region->kref);
141 if (devm_add_action_or_reset(parent, dax_region_unregister, dax_region))
142 return NULL;
143 return dax_region;
144}
145EXPORT_SYMBOL_GPL(alloc_dax_region);
146
147static struct dev_dax *to_dev_dax(struct device *dev)
148{
149 return container_of(dev, struct dev_dax, dev);
150}
151
152static ssize_t size_show(struct device *dev,
153 struct device_attribute *attr, char *buf)
154{
155 struct dev_dax *dev_dax = to_dev_dax(dev);
156 unsigned long long size = 0;
157 int i;
158
159 for (i = 0; i < dev_dax->num_resources; i++)
160 size += resource_size(&dev_dax->res[i]);
161
162 return sprintf(buf, "%llu\n", size);
163}
164static DEVICE_ATTR_RO(size);
165
166static struct attribute *dev_dax_attributes[] = {
167 &dev_attr_size.attr,
168 NULL,
169};
170
171static const struct attribute_group dev_dax_attribute_group = {
172 .attrs = dev_dax_attributes,
173};
174
175static const struct attribute_group *dax_attribute_groups[] = {
176 &dev_dax_attribute_group,
177 NULL,
178};
179
180static int check_vma(struct dev_dax *dev_dax, struct vm_area_struct *vma,
181 const char *func)
182{
183 struct dax_region *dax_region = dev_dax->region;
184 struct device *dev = &dev_dax->dev;
185 unsigned long mask;
186
187 if (!dax_alive(dev_dax->dax_dev))
188 return -ENXIO;
189
190 /* prevent private mappings from being established */
191 if ((vma->vm_flags & VM_MAYSHARE) != VM_MAYSHARE) {
192 dev_info(dev, "%s: %s: fail, attempted private mapping\n",
193 current->comm, func);
194 return -EINVAL;
195 }
196
197 mask = dax_region->align - 1;
198 if (vma->vm_start & mask || vma->vm_end & mask) {
199 dev_info(dev, "%s: %s: fail, unaligned vma (%#lx - %#lx, %#lx)\n",
200 current->comm, func, vma->vm_start, vma->vm_end,
201 mask);
202 return -EINVAL;
203 }
204
205 if ((dax_region->pfn_flags & (PFN_DEV|PFN_MAP)) == PFN_DEV
206 && (vma->vm_flags & VM_DONTCOPY) == 0) {
207 dev_info(dev, "%s: %s: fail, dax range requires MADV_DONTFORK\n",
208 current->comm, func);
209 return -EINVAL;
210 }
211
212 if (!vma_is_dax(vma)) {
213 dev_info(dev, "%s: %s: fail, vma is not DAX capable\n",
214 current->comm, func);
215 return -EINVAL;
216 }
217
218 return 0;
219}
220
221/* see "strong" declaration in tools/testing/nvdimm/dax-dev.c */
222__weak phys_addr_t dax_pgoff_to_phys(struct dev_dax *dev_dax, pgoff_t pgoff,
223 unsigned long size)
224{
225 struct resource *res;
226 /* gcc-4.6.3-nolibc for i386 complains that this is uninitialized */
227 phys_addr_t uninitialized_var(phys);
228 int i;
229
230 for (i = 0; i < dev_dax->num_resources; i++) {
231 res = &dev_dax->res[i];
232 phys = pgoff * PAGE_SIZE + res->start;
233 if (phys >= res->start && phys <= res->end)
234 break;
235 pgoff -= PHYS_PFN(resource_size(res));
236 }
237
238 if (i < dev_dax->num_resources) {
239 res = &dev_dax->res[i];
240 if (phys + size - 1 <= res->end)
241 return phys;
242 }
243
244 return -1;
245}
246
247static int __dev_dax_pte_fault(struct dev_dax *dev_dax, struct vm_fault *vmf)
248{
249 struct device *dev = &dev_dax->dev;
250 struct dax_region *dax_region;
251 int rc = VM_FAULT_SIGBUS;
252 phys_addr_t phys;
253 pfn_t pfn;
254 unsigned int fault_size = PAGE_SIZE;
255
256 if (check_vma(dev_dax, vmf->vma, __func__))
257 return VM_FAULT_SIGBUS;
258
259 dax_region = dev_dax->region;
260 if (dax_region->align > PAGE_SIZE) {
261 dev_dbg(dev, "alignment (%#x) > fault size (%#x)\n",
262 dax_region->align, fault_size);
263 return VM_FAULT_SIGBUS;
264 }
265
266 if (fault_size != dax_region->align)
267 return VM_FAULT_SIGBUS;
268
269 phys = dax_pgoff_to_phys(dev_dax, vmf->pgoff, PAGE_SIZE);
270 if (phys == -1) {
271 dev_dbg(dev, "pgoff_to_phys(%#lx) failed\n", vmf->pgoff);
272 return VM_FAULT_SIGBUS;
273 }
274
275 pfn = phys_to_pfn_t(phys, dax_region->pfn_flags);
276
277 rc = vm_insert_mixed(vmf->vma, vmf->address, pfn);
278
279 if (rc == -ENOMEM)
280 return VM_FAULT_OOM;
281 if (rc < 0 && rc != -EBUSY)
282 return VM_FAULT_SIGBUS;
283
284 return VM_FAULT_NOPAGE;
285}
286
287static int __dev_dax_pmd_fault(struct dev_dax *dev_dax, struct vm_fault *vmf)
288{
289 unsigned long pmd_addr = vmf->address & PMD_MASK;
290 struct device *dev = &dev_dax->dev;
291 struct dax_region *dax_region;
292 phys_addr_t phys;
293 pgoff_t pgoff;
294 pfn_t pfn;
295 unsigned int fault_size = PMD_SIZE;
296
297 if (check_vma(dev_dax, vmf->vma, __func__))
298 return VM_FAULT_SIGBUS;
299
300 dax_region = dev_dax->region;
301 if (dax_region->align > PMD_SIZE) {
302 dev_dbg(dev, "alignment (%#x) > fault size (%#x)\n",
303 dax_region->align, fault_size);
304 return VM_FAULT_SIGBUS;
305 }
306
307 /* dax pmd mappings require pfn_t_devmap() */
308 if ((dax_region->pfn_flags & (PFN_DEV|PFN_MAP)) != (PFN_DEV|PFN_MAP)) {
309 dev_dbg(dev, "region lacks devmap flags\n");
310 return VM_FAULT_SIGBUS;
311 }
312
313 if (fault_size < dax_region->align)
314 return VM_FAULT_SIGBUS;
315 else if (fault_size > dax_region->align)
316 return VM_FAULT_FALLBACK;
317
318 /* if we are outside of the VMA */
319 if (pmd_addr < vmf->vma->vm_start ||
320 (pmd_addr + PMD_SIZE) > vmf->vma->vm_end)
321 return VM_FAULT_SIGBUS;
322
323 pgoff = linear_page_index(vmf->vma, pmd_addr);
324 phys = dax_pgoff_to_phys(dev_dax, pgoff, PMD_SIZE);
325 if (phys == -1) {
326 dev_dbg(dev, "pgoff_to_phys(%#lx) failed\n", pgoff);
327 return VM_FAULT_SIGBUS;
328 }
329
330 pfn = phys_to_pfn_t(phys, dax_region->pfn_flags);
331
332 return vmf_insert_pfn_pmd(vmf->vma, vmf->address, vmf->pmd, pfn,
333 vmf->flags & FAULT_FLAG_WRITE);
334}
335
336#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
337static int __dev_dax_pud_fault(struct dev_dax *dev_dax, struct vm_fault *vmf)
338{
339 unsigned long pud_addr = vmf->address & PUD_MASK;
340 struct device *dev = &dev_dax->dev;
341 struct dax_region *dax_region;
342 phys_addr_t phys;
343 pgoff_t pgoff;
344 pfn_t pfn;
345 unsigned int fault_size = PUD_SIZE;
346
347
348 if (check_vma(dev_dax, vmf->vma, __func__))
349 return VM_FAULT_SIGBUS;
350
351 dax_region = dev_dax->region;
352 if (dax_region->align > PUD_SIZE) {
353 dev_dbg(dev, "alignment (%#x) > fault size (%#x)\n",
354 dax_region->align, fault_size);
355 return VM_FAULT_SIGBUS;
356 }
357
358 /* dax pud mappings require pfn_t_devmap() */
359 if ((dax_region->pfn_flags & (PFN_DEV|PFN_MAP)) != (PFN_DEV|PFN_MAP)) {
360 dev_dbg(dev, "region lacks devmap flags\n");
361 return VM_FAULT_SIGBUS;
362 }
363
364 if (fault_size < dax_region->align)
365 return VM_FAULT_SIGBUS;
366 else if (fault_size > dax_region->align)
367 return VM_FAULT_FALLBACK;
368
369 /* if we are outside of the VMA */
370 if (pud_addr < vmf->vma->vm_start ||
371 (pud_addr + PUD_SIZE) > vmf->vma->vm_end)
372 return VM_FAULT_SIGBUS;
373
374 pgoff = linear_page_index(vmf->vma, pud_addr);
375 phys = dax_pgoff_to_phys(dev_dax, pgoff, PUD_SIZE);
376 if (phys == -1) {
377 dev_dbg(dev, "pgoff_to_phys(%#lx) failed\n", pgoff);
378 return VM_FAULT_SIGBUS;
379 }
380
381 pfn = phys_to_pfn_t(phys, dax_region->pfn_flags);
382
383 return vmf_insert_pfn_pud(vmf->vma, vmf->address, vmf->pud, pfn,
384 vmf->flags & FAULT_FLAG_WRITE);
385}
386#else
387static int __dev_dax_pud_fault(struct dev_dax *dev_dax, struct vm_fault *vmf)
388{
389 return VM_FAULT_FALLBACK;
390}
391#endif /* !CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
392
393static int dev_dax_huge_fault(struct vm_fault *vmf,
394 enum page_entry_size pe_size)
395{
396 int rc, id;
397 struct file *filp = vmf->vma->vm_file;
398 struct dev_dax *dev_dax = filp->private_data;
399
400 dev_dbg(&dev_dax->dev, "%s: %s (%#lx - %#lx) size = %d\n", current->comm,
401 (vmf->flags & FAULT_FLAG_WRITE) ? "write" : "read",
402 vmf->vma->vm_start, vmf->vma->vm_end, pe_size);
403
404 id = dax_read_lock();
405 switch (pe_size) {
406 case PE_SIZE_PTE:
407 rc = __dev_dax_pte_fault(dev_dax, vmf);
408 break;
409 case PE_SIZE_PMD:
410 rc = __dev_dax_pmd_fault(dev_dax, vmf);
411 break;
412 case PE_SIZE_PUD:
413 rc = __dev_dax_pud_fault(dev_dax, vmf);
414 break;
415 default:
416 rc = VM_FAULT_SIGBUS;
417 }
418 dax_read_unlock(id);
419
420 return rc;
421}
422
423static int dev_dax_fault(struct vm_fault *vmf)
424{
425 return dev_dax_huge_fault(vmf, PE_SIZE_PTE);
426}
427
428static int dev_dax_split(struct vm_area_struct *vma, unsigned long addr)
429{
430 struct file *filp = vma->vm_file;
431 struct dev_dax *dev_dax = filp->private_data;
432 struct dax_region *dax_region = dev_dax->region;
433
434 if (!IS_ALIGNED(addr, dax_region->align))
435 return -EINVAL;
436 return 0;
437}
438
439static unsigned long dev_dax_pagesize(struct vm_area_struct *vma)
440{
441 struct file *filp = vma->vm_file;
442 struct dev_dax *dev_dax = filp->private_data;
443 struct dax_region *dax_region = dev_dax->region;
444
445 return dax_region->align;
446}
447
448static const struct vm_operations_struct dax_vm_ops = {
449 .fault = dev_dax_fault,
450 .huge_fault = dev_dax_huge_fault,
451 .split = dev_dax_split,
452 .pagesize = dev_dax_pagesize,
453};
454
455static int dax_mmap(struct file *filp, struct vm_area_struct *vma)
456{
457 struct dev_dax *dev_dax = filp->private_data;
458 int rc, id;
459
460 dev_dbg(&dev_dax->dev, "trace\n");
461
462 /*
463 * We lock to check dax_dev liveness and will re-check at
464 * fault time.
465 */
466 id = dax_read_lock();
467 rc = check_vma(dev_dax, vma, __func__);
468 dax_read_unlock(id);
469 if (rc)
470 return rc;
471
472 vma->vm_ops = &dax_vm_ops;
473 vma->vm_flags |= VM_MIXEDMAP | VM_HUGEPAGE;
474 return 0;
475}
476
477/* return an unmapped area aligned to the dax region specified alignment */
478static unsigned long dax_get_unmapped_area(struct file *filp,
479 unsigned long addr, unsigned long len, unsigned long pgoff,
480 unsigned long flags)
481{
482 unsigned long off, off_end, off_align, len_align, addr_align, align;
483 struct dev_dax *dev_dax = filp ? filp->private_data : NULL;
484 struct dax_region *dax_region;
485
486 if (!dev_dax || addr)
487 goto out;
488
489 dax_region = dev_dax->region;
490 align = dax_region->align;
491 off = pgoff << PAGE_SHIFT;
492 off_end = off + len;
493 off_align = round_up(off, align);
494
495 if ((off_end <= off_align) || ((off_end - off_align) < align))
496 goto out;
497
498 len_align = len + align;
499 if ((off + len_align) < off)
500 goto out;
501
502 addr_align = current->mm->get_unmapped_area(filp, addr, len_align,
503 pgoff, flags);
504 if (!IS_ERR_VALUE(addr_align)) {
505 addr_align += (off - addr_align) & (align - 1);
506 return addr_align;
507 }
508 out:
509 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
510}
511
512static int dax_open(struct inode *inode, struct file *filp)
513{
514 struct dax_device *dax_dev = inode_dax(inode);
515 struct inode *__dax_inode = dax_inode(dax_dev);
516 struct dev_dax *dev_dax = dax_get_private(dax_dev);
517
518 dev_dbg(&dev_dax->dev, "trace\n");
519 inode->i_mapping = __dax_inode->i_mapping;
520 inode->i_mapping->host = __dax_inode;
521 filp->f_mapping = inode->i_mapping;
522 filp->f_wb_err = filemap_sample_wb_err(filp->f_mapping);
523 filp->private_data = dev_dax;
524 inode->i_flags = S_DAX;
525
526 return 0;
527}
528
529static int dax_release(struct inode *inode, struct file *filp)
530{
531 struct dev_dax *dev_dax = filp->private_data;
532
533 dev_dbg(&dev_dax->dev, "trace\n");
534 return 0;
535}
536
537static const struct file_operations dax_fops = {
538 .llseek = noop_llseek,
539 .owner = THIS_MODULE,
540 .open = dax_open,
541 .release = dax_release,
542 .get_unmapped_area = dax_get_unmapped_area,
543 .mmap = dax_mmap,
544 .mmap_supported_flags = MAP_SYNC,
545};
546
547static void dev_dax_release(struct device *dev)
548{
549 struct dev_dax *dev_dax = to_dev_dax(dev);
550 struct dax_region *dax_region = dev_dax->region;
551 struct dax_device *dax_dev = dev_dax->dax_dev;
552
553 if (dev_dax->id >= 0)
554 ida_simple_remove(&dax_region->ida, dev_dax->id);
555 dax_region_put(dax_region);
556 put_dax(dax_dev);
557 kfree(dev_dax);
558}
559
560static void kill_dev_dax(struct dev_dax *dev_dax)
561{
562 struct dax_device *dax_dev = dev_dax->dax_dev;
563 struct inode *inode = dax_inode(dax_dev);
564
565 kill_dax(dax_dev);
566 unmap_mapping_range(inode->i_mapping, 0, 0, 1);
567}
568
569static void unregister_dev_dax(void *dev)
570{
571 struct dev_dax *dev_dax = to_dev_dax(dev);
572 struct dax_device *dax_dev = dev_dax->dax_dev;
573 struct inode *inode = dax_inode(dax_dev);
574 struct cdev *cdev = inode->i_cdev;
575
576 dev_dbg(dev, "trace\n");
577
578 kill_dev_dax(dev_dax);
579 cdev_device_del(cdev, dev);
580 put_device(dev);
581}
582
583struct dev_dax *devm_create_dev_dax(struct dax_region *dax_region,
584 int id, struct resource *res, int count)
585{
586 struct device *parent = dax_region->dev;
587 struct dax_device *dax_dev;
588 struct dev_dax *dev_dax;
589 struct inode *inode;
590 struct device *dev;
591 struct cdev *cdev;
592 int rc, i;
593
594 if (!count)
595 return ERR_PTR(-EINVAL);
596
597 dev_dax = kzalloc(sizeof(*dev_dax) + sizeof(*res) * count, GFP_KERNEL);
598 if (!dev_dax)
599 return ERR_PTR(-ENOMEM);
600
601 for (i = 0; i < count; i++) {
602 if (!IS_ALIGNED(res[i].start, dax_region->align)
603 || !IS_ALIGNED(resource_size(&res[i]),
604 dax_region->align)) {
605 rc = -EINVAL;
606 break;
607 }
608 dev_dax->res[i].start = res[i].start;
609 dev_dax->res[i].end = res[i].end;
610 }
611
612 if (i < count)
613 goto err_id;
614
615 if (id < 0) {
616 id = ida_simple_get(&dax_region->ida, 0, 0, GFP_KERNEL);
617 dev_dax->id = id;
618 if (id < 0) {
619 rc = id;
620 goto err_id;
621 }
622 } else {
623 /* region provider owns @id lifetime */
624 dev_dax->id = -1;
625 }
626
627 /*
628 * No 'host' or dax_operations since there is no access to this
629 * device outside of mmap of the resulting character device.
630 */
631 dax_dev = alloc_dax(dev_dax, NULL, NULL);
632 if (!dax_dev) {
633 rc = -ENOMEM;
634 goto err_dax;
635 }
636
637 /* from here on we're committed to teardown via dax_dev_release() */
638 dev = &dev_dax->dev;
639 device_initialize(dev);
640
641 inode = dax_inode(dax_dev);
642 cdev = inode->i_cdev;
643 cdev_init(cdev, &dax_fops);
644 cdev->owner = parent->driver->owner;
645
646 dev_dax->num_resources = count;
647 dev_dax->dax_dev = dax_dev;
648 dev_dax->region = dax_region;
649 kref_get(&dax_region->kref);
650
651 dev->devt = inode->i_rdev;
652 dev->class = dax_class;
653 dev->parent = parent;
654 dev->groups = dax_attribute_groups;
655 dev->release = dev_dax_release;
656 dev_set_name(dev, "dax%d.%d", dax_region->id, id);
657
658 rc = cdev_device_add(cdev, dev);
659 if (rc) {
660 kill_dev_dax(dev_dax);
661 put_device(dev);
662 return ERR_PTR(rc);
663 }
664
665 rc = devm_add_action_or_reset(dax_region->dev, unregister_dev_dax, dev);
666 if (rc)
667 return ERR_PTR(rc);
668
669 return dev_dax;
670
671 err_dax:
672 if (dev_dax->id >= 0)
673 ida_simple_remove(&dax_region->ida, dev_dax->id);
674 err_id:
675 kfree(dev_dax);
676
677 return ERR_PTR(rc);
678}
679EXPORT_SYMBOL_GPL(devm_create_dev_dax);
680
681static int __init dax_init(void)
682{
683 dax_class = class_create(THIS_MODULE, "dax");
684 return PTR_ERR_OR_ZERO(dax_class);
685}
686
687static void __exit dax_exit(void)
688{
689 class_destroy(dax_class);
690}
691
692MODULE_AUTHOR("Intel Corporation");
693MODULE_LICENSE("GPL v2");
694subsys_initcall(dax_init);
695module_exit(dax_exit);
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
17static 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 */
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
76static vm_fault_t __dev_dax_pte_fault(struct dev_dax *dev_dax,
77 struct vm_fault *vmf, pfn_t *pfn)
78{
79 struct device *dev = &dev_dax->dev;
80 phys_addr_t phys;
81 unsigned int fault_size = PAGE_SIZE;
82
83 if (check_vma(dev_dax, vmf->vma, __func__))
84 return VM_FAULT_SIGBUS;
85
86 if (dev_dax->align > PAGE_SIZE) {
87 dev_dbg(dev, "alignment (%#x) > fault size (%#x)\n",
88 dev_dax->align, fault_size);
89 return VM_FAULT_SIGBUS;
90 }
91
92 if (fault_size != dev_dax->align)
93 return VM_FAULT_SIGBUS;
94
95 phys = dax_pgoff_to_phys(dev_dax, vmf->pgoff, PAGE_SIZE);
96 if (phys == -1) {
97 dev_dbg(dev, "pgoff_to_phys(%#lx) failed\n", vmf->pgoff);
98 return VM_FAULT_SIGBUS;
99 }
100
101 *pfn = phys_to_pfn_t(phys, PFN_DEV|PFN_MAP);
102
103 return vmf_insert_mixed(vmf->vma, vmf->address, *pfn);
104}
105
106static vm_fault_t __dev_dax_pmd_fault(struct dev_dax *dev_dax,
107 struct vm_fault *vmf, pfn_t *pfn)
108{
109 unsigned long pmd_addr = vmf->address & PMD_MASK;
110 struct device *dev = &dev_dax->dev;
111 phys_addr_t phys;
112 pgoff_t pgoff;
113 unsigned int fault_size = PMD_SIZE;
114
115 if (check_vma(dev_dax, vmf->vma, __func__))
116 return VM_FAULT_SIGBUS;
117
118 if (dev_dax->align > PMD_SIZE) {
119 dev_dbg(dev, "alignment (%#x) > fault size (%#x)\n",
120 dev_dax->align, fault_size);
121 return VM_FAULT_SIGBUS;
122 }
123
124 if (fault_size < dev_dax->align)
125 return VM_FAULT_SIGBUS;
126 else if (fault_size > dev_dax->align)
127 return VM_FAULT_FALLBACK;
128
129 /* if we are outside of the VMA */
130 if (pmd_addr < vmf->vma->vm_start ||
131 (pmd_addr + PMD_SIZE) > vmf->vma->vm_end)
132 return VM_FAULT_SIGBUS;
133
134 pgoff = linear_page_index(vmf->vma, pmd_addr);
135 phys = dax_pgoff_to_phys(dev_dax, pgoff, PMD_SIZE);
136 if (phys == -1) {
137 dev_dbg(dev, "pgoff_to_phys(%#lx) failed\n", pgoff);
138 return VM_FAULT_SIGBUS;
139 }
140
141 *pfn = phys_to_pfn_t(phys, PFN_DEV|PFN_MAP);
142
143 return vmf_insert_pfn_pmd(vmf, *pfn, vmf->flags & FAULT_FLAG_WRITE);
144}
145
146#ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
147static vm_fault_t __dev_dax_pud_fault(struct dev_dax *dev_dax,
148 struct vm_fault *vmf, pfn_t *pfn)
149{
150 unsigned long pud_addr = vmf->address & PUD_MASK;
151 struct device *dev = &dev_dax->dev;
152 phys_addr_t phys;
153 pgoff_t pgoff;
154 unsigned int fault_size = PUD_SIZE;
155
156
157 if (check_vma(dev_dax, vmf->vma, __func__))
158 return VM_FAULT_SIGBUS;
159
160 if (dev_dax->align > PUD_SIZE) {
161 dev_dbg(dev, "alignment (%#x) > fault size (%#x)\n",
162 dev_dax->align, fault_size);
163 return VM_FAULT_SIGBUS;
164 }
165
166 if (fault_size < dev_dax->align)
167 return VM_FAULT_SIGBUS;
168 else if (fault_size > dev_dax->align)
169 return VM_FAULT_FALLBACK;
170
171 /* if we are outside of the VMA */
172 if (pud_addr < vmf->vma->vm_start ||
173 (pud_addr + PUD_SIZE) > vmf->vma->vm_end)
174 return VM_FAULT_SIGBUS;
175
176 pgoff = linear_page_index(vmf->vma, pud_addr);
177 phys = dax_pgoff_to_phys(dev_dax, pgoff, PUD_SIZE);
178 if (phys == -1) {
179 dev_dbg(dev, "pgoff_to_phys(%#lx) failed\n", pgoff);
180 return VM_FAULT_SIGBUS;
181 }
182
183 *pfn = phys_to_pfn_t(phys, PFN_DEV|PFN_MAP);
184
185 return vmf_insert_pfn_pud(vmf, *pfn, vmf->flags & FAULT_FLAG_WRITE);
186}
187#else
188static vm_fault_t __dev_dax_pud_fault(struct dev_dax *dev_dax,
189 struct vm_fault *vmf, pfn_t *pfn)
190{
191 return VM_FAULT_FALLBACK;
192}
193#endif /* !CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
194
195static vm_fault_t dev_dax_huge_fault(struct vm_fault *vmf,
196 enum page_entry_size pe_size)
197{
198 struct file *filp = vmf->vma->vm_file;
199 unsigned long fault_size;
200 vm_fault_t rc = VM_FAULT_SIGBUS;
201 int id;
202 pfn_t pfn;
203 struct dev_dax *dev_dax = filp->private_data;
204
205 dev_dbg(&dev_dax->dev, "%s: %s (%#lx - %#lx) size = %d\n", current->comm,
206 (vmf->flags & FAULT_FLAG_WRITE) ? "write" : "read",
207 vmf->vma->vm_start, vmf->vma->vm_end, pe_size);
208
209 id = dax_read_lock();
210 switch (pe_size) {
211 case PE_SIZE_PTE:
212 fault_size = PAGE_SIZE;
213 rc = __dev_dax_pte_fault(dev_dax, vmf, &pfn);
214 break;
215 case PE_SIZE_PMD:
216 fault_size = PMD_SIZE;
217 rc = __dev_dax_pmd_fault(dev_dax, vmf, &pfn);
218 break;
219 case PE_SIZE_PUD:
220 fault_size = PUD_SIZE;
221 rc = __dev_dax_pud_fault(dev_dax, vmf, &pfn);
222 break;
223 default:
224 rc = VM_FAULT_SIGBUS;
225 }
226
227 if (rc == VM_FAULT_NOPAGE) {
228 unsigned long i;
229 pgoff_t pgoff;
230
231 /*
232 * In the device-dax case the only possibility for a
233 * VM_FAULT_NOPAGE result is when device-dax capacity is
234 * mapped. No need to consider the zero page, or racing
235 * conflicting mappings.
236 */
237 pgoff = linear_page_index(vmf->vma, vmf->address
238 & ~(fault_size - 1));
239 for (i = 0; i < fault_size / PAGE_SIZE; i++) {
240 struct page *page;
241
242 page = pfn_to_page(pfn_t_to_pfn(pfn) + i);
243 if (page->mapping)
244 continue;
245 page->mapping = filp->f_mapping;
246 page->index = pgoff + i;
247 }
248 }
249 dax_read_unlock(id);
250
251 return rc;
252}
253
254static vm_fault_t dev_dax_fault(struct vm_fault *vmf)
255{
256 return dev_dax_huge_fault(vmf, PE_SIZE_PTE);
257}
258
259static int dev_dax_may_split(struct vm_area_struct *vma, unsigned long addr)
260{
261 struct file *filp = vma->vm_file;
262 struct dev_dax *dev_dax = filp->private_data;
263
264 if (!IS_ALIGNED(addr, dev_dax->align))
265 return -EINVAL;
266 return 0;
267}
268
269static unsigned long dev_dax_pagesize(struct vm_area_struct *vma)
270{
271 struct file *filp = vma->vm_file;
272 struct dev_dax *dev_dax = filp->private_data;
273
274 return dev_dax->align;
275}
276
277static const struct vm_operations_struct dax_vm_ops = {
278 .fault = dev_dax_fault,
279 .huge_fault = dev_dax_huge_fault,
280 .may_split = dev_dax_may_split,
281 .pagesize = dev_dax_pagesize,
282};
283
284static int dax_mmap(struct file *filp, struct vm_area_struct *vma)
285{
286 struct dev_dax *dev_dax = filp->private_data;
287 int rc, id;
288
289 dev_dbg(&dev_dax->dev, "trace\n");
290
291 /*
292 * We lock to check dax_dev liveness and will re-check at
293 * fault time.
294 */
295 id = dax_read_lock();
296 rc = check_vma(dev_dax, vma, __func__);
297 dax_read_unlock(id);
298 if (rc)
299 return rc;
300
301 vma->vm_ops = &dax_vm_ops;
302 vma->vm_flags |= VM_HUGEPAGE;
303 return 0;
304}
305
306/* return an unmapped area aligned to the dax region specified alignment */
307static unsigned long dax_get_unmapped_area(struct file *filp,
308 unsigned long addr, unsigned long len, unsigned long pgoff,
309 unsigned long flags)
310{
311 unsigned long off, off_end, off_align, len_align, addr_align, align;
312 struct dev_dax *dev_dax = filp ? filp->private_data : NULL;
313
314 if (!dev_dax || addr)
315 goto out;
316
317 align = dev_dax->align;
318 off = pgoff << PAGE_SHIFT;
319 off_end = off + len;
320 off_align = round_up(off, align);
321
322 if ((off_end <= off_align) || ((off_end - off_align) < align))
323 goto out;
324
325 len_align = len + align;
326 if ((off + len_align) < off)
327 goto out;
328
329 addr_align = current->mm->get_unmapped_area(filp, addr, len_align,
330 pgoff, flags);
331 if (!IS_ERR_VALUE(addr_align)) {
332 addr_align += (off - addr_align) & (align - 1);
333 return addr_align;
334 }
335 out:
336 return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
337}
338
339static const struct address_space_operations dev_dax_aops = {
340 .set_page_dirty = __set_page_dirty_no_writeback,
341 .invalidatepage = noop_invalidatepage,
342};
343
344static int dax_open(struct inode *inode, struct file *filp)
345{
346 struct dax_device *dax_dev = inode_dax(inode);
347 struct inode *__dax_inode = dax_inode(dax_dev);
348 struct dev_dax *dev_dax = dax_get_private(dax_dev);
349
350 dev_dbg(&dev_dax->dev, "trace\n");
351 inode->i_mapping = __dax_inode->i_mapping;
352 inode->i_mapping->host = __dax_inode;
353 inode->i_mapping->a_ops = &dev_dax_aops;
354 filp->f_mapping = inode->i_mapping;
355 filp->f_wb_err = filemap_sample_wb_err(filp->f_mapping);
356 filp->f_sb_err = file_sample_sb_err(filp);
357 filp->private_data = dev_dax;
358 inode->i_flags = S_DAX;
359
360 return 0;
361}
362
363static int dax_release(struct inode *inode, struct file *filp)
364{
365 struct dev_dax *dev_dax = filp->private_data;
366
367 dev_dbg(&dev_dax->dev, "trace\n");
368 return 0;
369}
370
371static const struct file_operations dax_fops = {
372 .llseek = noop_llseek,
373 .owner = THIS_MODULE,
374 .open = dax_open,
375 .release = dax_release,
376 .get_unmapped_area = dax_get_unmapped_area,
377 .mmap = dax_mmap,
378 .mmap_supported_flags = MAP_SYNC,
379};
380
381static void dev_dax_cdev_del(void *cdev)
382{
383 cdev_del(cdev);
384}
385
386static void dev_dax_kill(void *dev_dax)
387{
388 kill_dev_dax(dev_dax);
389}
390
391int dev_dax_probe(struct dev_dax *dev_dax)
392{
393 struct dax_device *dax_dev = dev_dax->dax_dev;
394 struct device *dev = &dev_dax->dev;
395 struct dev_pagemap *pgmap;
396 struct inode *inode;
397 struct cdev *cdev;
398 void *addr;
399 int rc, i;
400
401 pgmap = dev_dax->pgmap;
402 if (dev_WARN_ONCE(dev, pgmap && dev_dax->nr_range > 1,
403 "static pgmap / multi-range device conflict\n"))
404 return -EINVAL;
405
406 if (!pgmap) {
407 pgmap = devm_kzalloc(dev, sizeof(*pgmap) + sizeof(struct range)
408 * (dev_dax->nr_range - 1), GFP_KERNEL);
409 if (!pgmap)
410 return -ENOMEM;
411 pgmap->nr_range = dev_dax->nr_range;
412 }
413
414 for (i = 0; i < dev_dax->nr_range; i++) {
415 struct range *range = &dev_dax->ranges[i].range;
416
417 if (!devm_request_mem_region(dev, range->start,
418 range_len(range), dev_name(dev))) {
419 dev_warn(dev, "mapping%d: %#llx-%#llx could not reserve range\n",
420 i, range->start, range->end);
421 return -EBUSY;
422 }
423 /* don't update the range for static pgmap */
424 if (!dev_dax->pgmap)
425 pgmap->ranges[i] = *range;
426 }
427
428 pgmap->type = MEMORY_DEVICE_GENERIC;
429 addr = devm_memremap_pages(dev, pgmap);
430 if (IS_ERR(addr))
431 return PTR_ERR(addr);
432
433 inode = dax_inode(dax_dev);
434 cdev = inode->i_cdev;
435 cdev_init(cdev, &dax_fops);
436 if (dev->class) {
437 /* for the CONFIG_DEV_DAX_PMEM_COMPAT case */
438 cdev->owner = dev->parent->driver->owner;
439 } else
440 cdev->owner = dev->driver->owner;
441 cdev_set_parent(cdev, &dev->kobj);
442 rc = cdev_add(cdev, dev->devt, 1);
443 if (rc)
444 return rc;
445
446 rc = devm_add_action_or_reset(dev, dev_dax_cdev_del, cdev);
447 if (rc)
448 return rc;
449
450 run_dax(dax_dev);
451 return devm_add_action_or_reset(dev, dev_dax_kill, dev_dax);
452}
453EXPORT_SYMBOL_GPL(dev_dax_probe);
454
455static struct dax_device_driver device_dax_driver = {
456 .probe = dev_dax_probe,
457 /* all probe actions are unwound by devm, so .remove isn't necessary */
458 .match_always = 1,
459};
460
461static int __init dax_init(void)
462{
463 return dax_driver_register(&device_dax_driver);
464}
465
466static void __exit dax_exit(void)
467{
468 dax_driver_unregister(&device_dax_driver);
469}
470
471MODULE_AUTHOR("Intel Corporation");
472MODULE_LICENSE("GPL v2");
473module_init(dax_init);
474module_exit(dax_exit);
475MODULE_ALIAS_DAX_DEVICE(0);