xref: /linux/drivers/dax/device.c (revision 662fa3d6099374c4615bf64d06895e3573b935b2)
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 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 
106 static 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
147 static 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
188 static 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 
195 static 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 
254 static vm_fault_t dev_dax_fault(struct vm_fault *vmf)
255 {
256 	return dev_dax_huge_fault(vmf, PE_SIZE_PTE);
257 }
258 
259 static 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 
269 static 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 
277 static 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 
284 static 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 */
307 static 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 
339 static const struct address_space_operations dev_dax_aops = {
340 	.set_page_dirty		= __set_page_dirty_no_writeback,
341 	.invalidatepage		= noop_invalidatepage,
342 };
343 
344 static 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 
363 static 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 
371 static 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 
381 static void dev_dax_cdev_del(void *cdev)
382 {
383 	cdev_del(cdev);
384 }
385 
386 static void dev_dax_kill(void *dev_dax)
387 {
388 	kill_dev_dax(dev_dax);
389 }
390 
391 int 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 }
453 EXPORT_SYMBOL_GPL(dev_dax_probe);
454 
455 static 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 
461 static int __init dax_init(void)
462 {
463 	return dax_driver_register(&device_dax_driver);
464 }
465 
466 static void __exit dax_exit(void)
467 {
468 	dax_driver_unregister(&device_dax_driver);
469 }
470 
471 MODULE_AUTHOR("Intel Corporation");
472 MODULE_LICENSE("GPL v2");
473 module_init(dax_init);
474 module_exit(dax_exit);
475 MODULE_ALIAS_DAX_DEVICE(0);
476