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 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_DOWN(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
103static 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
136static 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
180static 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
224static 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
231static 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: op=%s addr=%#lx order=%d\n", current->comm,
239 (vmf->flags & FAULT_FLAG_WRITE) ? "write" : "read",
240 vmf->address & ~((1UL << (order + PAGE_SHIFT)) - 1), 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
257static vm_fault_t dev_dax_fault(struct vm_fault *vmf)
258{
259 return dev_dax_huge_fault(vmf, 0);
260}
261
262static 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
272static 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
280static 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
287static 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 */
310static 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 = mm_get_unmapped_area(current->mm, 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 mm_get_unmapped_area(current->mm, filp, addr, len, pgoff, flags);
340}
341
342static const struct address_space_operations dev_dax_aops = {
343 .dirty_folio = noop_dirty_folio,
344};
345
346static 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
365static 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
373static 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 .fop_flags = FOP_MMAP_SYNC,
381};
382
383static void dev_dax_cdev_del(void *cdev)
384{
385 cdev_del(cdev);
386}
387
388static void dev_dax_kill(void *dev_dax)
389{
390 kill_dev_dax(dev_dax);
391}
392
393static 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
469static struct dax_device_driver device_dax_driver = {
470 .probe = dev_dax_probe,
471 .type = DAXDRV_DEVICE_TYPE,
472};
473
474static int __init dax_init(void)
475{
476 return dax_driver_register(&device_dax_driver);
477}
478
479static void __exit dax_exit(void)
480{
481 dax_driver_unregister(&device_dax_driver);
482}
483
484MODULE_AUTHOR("Intel Corporation");
485MODULE_DESCRIPTION("Device DAX: direct access device driver");
486MODULE_LICENSE("GPL v2");
487module_init(dax_init);
488module_exit(dax_exit);
489MODULE_ALIAS_DAX_DEVICE(0);