xref: /openbmc/linux/drivers/dax/device.c (revision 9a29f5fc)
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 
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 */
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 
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 
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 
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
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
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 
231 static vm_fault_t dev_dax_huge_fault(struct vm_fault *vmf,
232 		enum page_entry_size pe_size)
233 {
234 	struct file *filp = vmf->vma->vm_file;
235 	vm_fault_t rc = VM_FAULT_SIGBUS;
236 	int id;
237 	struct dev_dax *dev_dax = filp->private_data;
238 
239 	dev_dbg(&dev_dax->dev, "%s: %s (%#lx - %#lx) size = %d\n", current->comm,
240 			(vmf->flags & FAULT_FLAG_WRITE) ? "write" : "read",
241 			vmf->vma->vm_start, vmf->vma->vm_end, pe_size);
242 
243 	id = dax_read_lock();
244 	switch (pe_size) {
245 	case PE_SIZE_PTE:
246 		rc = __dev_dax_pte_fault(dev_dax, vmf);
247 		break;
248 	case PE_SIZE_PMD:
249 		rc = __dev_dax_pmd_fault(dev_dax, vmf);
250 		break;
251 	case PE_SIZE_PUD:
252 		rc = __dev_dax_pud_fault(dev_dax, vmf);
253 		break;
254 	default:
255 		rc = VM_FAULT_SIGBUS;
256 	}
257 
258 	dax_read_unlock(id);
259 
260 	return rc;
261 }
262 
263 static vm_fault_t dev_dax_fault(struct vm_fault *vmf)
264 {
265 	return dev_dax_huge_fault(vmf, PE_SIZE_PTE);
266 }
267 
268 static int dev_dax_may_split(struct vm_area_struct *vma, unsigned long addr)
269 {
270 	struct file *filp = vma->vm_file;
271 	struct dev_dax *dev_dax = filp->private_data;
272 
273 	if (!IS_ALIGNED(addr, dev_dax->align))
274 		return -EINVAL;
275 	return 0;
276 }
277 
278 static unsigned long dev_dax_pagesize(struct vm_area_struct *vma)
279 {
280 	struct file *filp = vma->vm_file;
281 	struct dev_dax *dev_dax = filp->private_data;
282 
283 	return dev_dax->align;
284 }
285 
286 static const struct vm_operations_struct dax_vm_ops = {
287 	.fault = dev_dax_fault,
288 	.huge_fault = dev_dax_huge_fault,
289 	.may_split = dev_dax_may_split,
290 	.pagesize = dev_dax_pagesize,
291 };
292 
293 static int dax_mmap(struct file *filp, struct vm_area_struct *vma)
294 {
295 	struct dev_dax *dev_dax = filp->private_data;
296 	int rc, id;
297 
298 	dev_dbg(&dev_dax->dev, "trace\n");
299 
300 	/*
301 	 * We lock to check dax_dev liveness and will re-check at
302 	 * fault time.
303 	 */
304 	id = dax_read_lock();
305 	rc = check_vma(dev_dax, vma, __func__);
306 	dax_read_unlock(id);
307 	if (rc)
308 		return rc;
309 
310 	vma->vm_ops = &dax_vm_ops;
311 	vma->vm_flags |= VM_HUGEPAGE;
312 	return 0;
313 }
314 
315 /* return an unmapped area aligned to the dax region specified alignment */
316 static unsigned long dax_get_unmapped_area(struct file *filp,
317 		unsigned long addr, unsigned long len, unsigned long pgoff,
318 		unsigned long flags)
319 {
320 	unsigned long off, off_end, off_align, len_align, addr_align, align;
321 	struct dev_dax *dev_dax = filp ? filp->private_data : NULL;
322 
323 	if (!dev_dax || addr)
324 		goto out;
325 
326 	align = dev_dax->align;
327 	off = pgoff << PAGE_SHIFT;
328 	off_end = off + len;
329 	off_align = round_up(off, align);
330 
331 	if ((off_end <= off_align) || ((off_end - off_align) < align))
332 		goto out;
333 
334 	len_align = len + align;
335 	if ((off + len_align) < off)
336 		goto out;
337 
338 	addr_align = current->mm->get_unmapped_area(filp, addr, len_align,
339 			pgoff, flags);
340 	if (!IS_ERR_VALUE(addr_align)) {
341 		addr_align += (off - addr_align) & (align - 1);
342 		return addr_align;
343 	}
344  out:
345 	return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
346 }
347 
348 static const struct address_space_operations dev_dax_aops = {
349 	.dirty_folio	= noop_dirty_folio,
350 };
351 
352 static int dax_open(struct inode *inode, struct file *filp)
353 {
354 	struct dax_device *dax_dev = inode_dax(inode);
355 	struct inode *__dax_inode = dax_inode(dax_dev);
356 	struct dev_dax *dev_dax = dax_get_private(dax_dev);
357 
358 	dev_dbg(&dev_dax->dev, "trace\n");
359 	inode->i_mapping = __dax_inode->i_mapping;
360 	inode->i_mapping->host = __dax_inode;
361 	inode->i_mapping->a_ops = &dev_dax_aops;
362 	filp->f_mapping = inode->i_mapping;
363 	filp->f_wb_err = filemap_sample_wb_err(filp->f_mapping);
364 	filp->f_sb_err = file_sample_sb_err(filp);
365 	filp->private_data = dev_dax;
366 	inode->i_flags = S_DAX;
367 
368 	return 0;
369 }
370 
371 static int dax_release(struct inode *inode, struct file *filp)
372 {
373 	struct dev_dax *dev_dax = filp->private_data;
374 
375 	dev_dbg(&dev_dax->dev, "trace\n");
376 	return 0;
377 }
378 
379 static const struct file_operations dax_fops = {
380 	.llseek = noop_llseek,
381 	.owner = THIS_MODULE,
382 	.open = dax_open,
383 	.release = dax_release,
384 	.get_unmapped_area = dax_get_unmapped_area,
385 	.mmap = dax_mmap,
386 	.mmap_supported_flags = MAP_SYNC,
387 };
388 
389 static void dev_dax_cdev_del(void *cdev)
390 {
391 	cdev_del(cdev);
392 }
393 
394 static void dev_dax_kill(void *dev_dax)
395 {
396 	kill_dev_dax(dev_dax);
397 }
398 
399 int dev_dax_probe(struct dev_dax *dev_dax)
400 {
401 	struct dax_device *dax_dev = dev_dax->dax_dev;
402 	struct device *dev = &dev_dax->dev;
403 	struct dev_pagemap *pgmap;
404 	struct inode *inode;
405 	struct cdev *cdev;
406 	void *addr;
407 	int rc, i;
408 
409 	if (static_dev_dax(dev_dax))  {
410 		if (dev_dax->nr_range > 1) {
411 			dev_warn(dev,
412 				"static pgmap / multi-range device conflict\n");
413 			return -EINVAL;
414 		}
415 
416 		pgmap = dev_dax->pgmap;
417 	} else {
418 		if (dev_dax->pgmap) {
419 			dev_warn(dev,
420 				 "dynamic-dax with pre-populated page map\n");
421 			return -EINVAL;
422 		}
423 
424 		pgmap = devm_kzalloc(dev,
425                        struct_size(pgmap, ranges, dev_dax->nr_range - 1),
426                        GFP_KERNEL);
427 		if (!pgmap)
428 			return -ENOMEM;
429 
430 		pgmap->nr_range = dev_dax->nr_range;
431 		dev_dax->pgmap = pgmap;
432 
433 		for (i = 0; i < dev_dax->nr_range; i++) {
434 			struct range *range = &dev_dax->ranges[i].range;
435 			pgmap->ranges[i] = *range;
436 		}
437 	}
438 
439 	for (i = 0; i < dev_dax->nr_range; i++) {
440 		struct range *range = &dev_dax->ranges[i].range;
441 
442 		if (!devm_request_mem_region(dev, range->start,
443 					range_len(range), dev_name(dev))) {
444 			dev_warn(dev, "mapping%d: %#llx-%#llx could not reserve range\n",
445 					i, range->start, range->end);
446 			return -EBUSY;
447 		}
448 	}
449 
450 	pgmap->type = MEMORY_DEVICE_GENERIC;
451 	if (dev_dax->align > PAGE_SIZE)
452 		pgmap->vmemmap_shift =
453 			order_base_2(dev_dax->align >> PAGE_SHIFT);
454 	addr = devm_memremap_pages(dev, pgmap);
455 	if (IS_ERR(addr))
456 		return PTR_ERR(addr);
457 
458 	inode = dax_inode(dax_dev);
459 	cdev = inode->i_cdev;
460 	cdev_init(cdev, &dax_fops);
461 	cdev->owner = dev->driver->owner;
462 	cdev_set_parent(cdev, &dev->kobj);
463 	rc = cdev_add(cdev, dev->devt, 1);
464 	if (rc)
465 		return rc;
466 
467 	rc = devm_add_action_or_reset(dev, dev_dax_cdev_del, cdev);
468 	if (rc)
469 		return rc;
470 
471 	run_dax(dax_dev);
472 	return devm_add_action_or_reset(dev, dev_dax_kill, dev_dax);
473 }
474 EXPORT_SYMBOL_GPL(dev_dax_probe);
475 
476 static struct dax_device_driver device_dax_driver = {
477 	.probe = dev_dax_probe,
478 	/* all probe actions are unwound by devm, so .remove isn't necessary */
479 	.match_always = 1,
480 };
481 
482 static int __init dax_init(void)
483 {
484 	return dax_driver_register(&device_dax_driver);
485 }
486 
487 static void __exit dax_exit(void)
488 {
489 	dax_driver_unregister(&device_dax_driver);
490 }
491 
492 MODULE_AUTHOR("Intel Corporation");
493 MODULE_LICENSE("GPL v2");
494 module_init(dax_init);
495 module_exit(dax_exit);
496 MODULE_ALIAS_DAX_DEVICE(0);
497