xref: /openbmc/linux/drivers/dax/device.c (revision 367e5927)
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 struct dev_dax *ref_to_dev_dax(struct percpu_ref *ref)
18 {
19 	return container_of(ref, struct dev_dax, ref);
20 }
21 
22 static void dev_dax_percpu_release(struct percpu_ref *ref)
23 {
24 	struct dev_dax *dev_dax = ref_to_dev_dax(ref);
25 
26 	dev_dbg(&dev_dax->dev, "%s\n", __func__);
27 	complete(&dev_dax->cmp);
28 }
29 
30 static void dev_dax_percpu_exit(void *data)
31 {
32 	struct percpu_ref *ref = data;
33 	struct dev_dax *dev_dax = ref_to_dev_dax(ref);
34 
35 	dev_dbg(&dev_dax->dev, "%s\n", __func__);
36 	wait_for_completion(&dev_dax->cmp);
37 	percpu_ref_exit(ref);
38 }
39 
40 static void dev_dax_percpu_kill(struct percpu_ref *data)
41 {
42 	struct percpu_ref *ref = data;
43 	struct dev_dax *dev_dax = ref_to_dev_dax(ref);
44 
45 	dev_dbg(&dev_dax->dev, "%s\n", __func__);
46 	percpu_ref_kill(ref);
47 }
48 
49 static int check_vma(struct dev_dax *dev_dax, struct vm_area_struct *vma,
50 		const char *func)
51 {
52 	struct dax_region *dax_region = dev_dax->region;
53 	struct device *dev = &dev_dax->dev;
54 	unsigned long mask;
55 
56 	if (!dax_alive(dev_dax->dax_dev))
57 		return -ENXIO;
58 
59 	/* prevent private mappings from being established */
60 	if ((vma->vm_flags & VM_MAYSHARE) != VM_MAYSHARE) {
61 		dev_info_ratelimited(dev,
62 				"%s: %s: fail, attempted private mapping\n",
63 				current->comm, func);
64 		return -EINVAL;
65 	}
66 
67 	mask = dax_region->align - 1;
68 	if (vma->vm_start & mask || vma->vm_end & mask) {
69 		dev_info_ratelimited(dev,
70 				"%s: %s: fail, unaligned vma (%#lx - %#lx, %#lx)\n",
71 				current->comm, func, vma->vm_start, vma->vm_end,
72 				mask);
73 		return -EINVAL;
74 	}
75 
76 	if ((dax_region->pfn_flags & (PFN_DEV|PFN_MAP)) == PFN_DEV
77 			&& (vma->vm_flags & VM_DONTCOPY) == 0) {
78 		dev_info_ratelimited(dev,
79 				"%s: %s: fail, dax range requires MADV_DONTFORK\n",
80 				current->comm, func);
81 		return -EINVAL;
82 	}
83 
84 	if (!vma_is_dax(vma)) {
85 		dev_info_ratelimited(dev,
86 				"%s: %s: fail, vma is not DAX capable\n",
87 				current->comm, func);
88 		return -EINVAL;
89 	}
90 
91 	return 0;
92 }
93 
94 /* see "strong" declaration in tools/testing/nvdimm/dax-dev.c */
95 __weak phys_addr_t dax_pgoff_to_phys(struct dev_dax *dev_dax, pgoff_t pgoff,
96 		unsigned long size)
97 {
98 	struct resource *res = &dev_dax->region->res;
99 	phys_addr_t phys;
100 
101 	phys = pgoff * PAGE_SIZE + res->start;
102 	if (phys >= res->start && phys <= res->end) {
103 		if (phys + size - 1 <= res->end)
104 			return phys;
105 	}
106 
107 	return -1;
108 }
109 
110 static vm_fault_t __dev_dax_pte_fault(struct dev_dax *dev_dax,
111 				struct vm_fault *vmf, pfn_t *pfn)
112 {
113 	struct device *dev = &dev_dax->dev;
114 	struct dax_region *dax_region;
115 	phys_addr_t phys;
116 	unsigned int fault_size = PAGE_SIZE;
117 
118 	if (check_vma(dev_dax, vmf->vma, __func__))
119 		return VM_FAULT_SIGBUS;
120 
121 	dax_region = dev_dax->region;
122 	if (dax_region->align > PAGE_SIZE) {
123 		dev_dbg(dev, "alignment (%#x) > fault size (%#x)\n",
124 			dax_region->align, fault_size);
125 		return VM_FAULT_SIGBUS;
126 	}
127 
128 	if (fault_size != dax_region->align)
129 		return VM_FAULT_SIGBUS;
130 
131 	phys = dax_pgoff_to_phys(dev_dax, vmf->pgoff, PAGE_SIZE);
132 	if (phys == -1) {
133 		dev_dbg(dev, "pgoff_to_phys(%#lx) failed\n", vmf->pgoff);
134 		return VM_FAULT_SIGBUS;
135 	}
136 
137 	*pfn = phys_to_pfn_t(phys, dax_region->pfn_flags);
138 
139 	return vmf_insert_mixed(vmf->vma, vmf->address, *pfn);
140 }
141 
142 static vm_fault_t __dev_dax_pmd_fault(struct dev_dax *dev_dax,
143 				struct vm_fault *vmf, pfn_t *pfn)
144 {
145 	unsigned long pmd_addr = vmf->address & PMD_MASK;
146 	struct device *dev = &dev_dax->dev;
147 	struct dax_region *dax_region;
148 	phys_addr_t phys;
149 	pgoff_t pgoff;
150 	unsigned int fault_size = PMD_SIZE;
151 
152 	if (check_vma(dev_dax, vmf->vma, __func__))
153 		return VM_FAULT_SIGBUS;
154 
155 	dax_region = dev_dax->region;
156 	if (dax_region->align > PMD_SIZE) {
157 		dev_dbg(dev, "alignment (%#x) > fault size (%#x)\n",
158 			dax_region->align, fault_size);
159 		return VM_FAULT_SIGBUS;
160 	}
161 
162 	/* dax pmd mappings require pfn_t_devmap() */
163 	if ((dax_region->pfn_flags & (PFN_DEV|PFN_MAP)) != (PFN_DEV|PFN_MAP)) {
164 		dev_dbg(dev, "region lacks devmap flags\n");
165 		return VM_FAULT_SIGBUS;
166 	}
167 
168 	if (fault_size < dax_region->align)
169 		return VM_FAULT_SIGBUS;
170 	else if (fault_size > dax_region->align)
171 		return VM_FAULT_FALLBACK;
172 
173 	/* if we are outside of the VMA */
174 	if (pmd_addr < vmf->vma->vm_start ||
175 			(pmd_addr + PMD_SIZE) > vmf->vma->vm_end)
176 		return VM_FAULT_SIGBUS;
177 
178 	pgoff = linear_page_index(vmf->vma, pmd_addr);
179 	phys = dax_pgoff_to_phys(dev_dax, pgoff, PMD_SIZE);
180 	if (phys == -1) {
181 		dev_dbg(dev, "pgoff_to_phys(%#lx) failed\n", pgoff);
182 		return VM_FAULT_SIGBUS;
183 	}
184 
185 	*pfn = phys_to_pfn_t(phys, dax_region->pfn_flags);
186 
187 	return vmf_insert_pfn_pmd(vmf->vma, vmf->address, vmf->pmd, *pfn,
188 			vmf->flags & FAULT_FLAG_WRITE);
189 }
190 
191 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
192 static vm_fault_t __dev_dax_pud_fault(struct dev_dax *dev_dax,
193 				struct vm_fault *vmf, pfn_t *pfn)
194 {
195 	unsigned long pud_addr = vmf->address & PUD_MASK;
196 	struct device *dev = &dev_dax->dev;
197 	struct dax_region *dax_region;
198 	phys_addr_t phys;
199 	pgoff_t pgoff;
200 	unsigned int fault_size = PUD_SIZE;
201 
202 
203 	if (check_vma(dev_dax, vmf->vma, __func__))
204 		return VM_FAULT_SIGBUS;
205 
206 	dax_region = dev_dax->region;
207 	if (dax_region->align > PUD_SIZE) {
208 		dev_dbg(dev, "alignment (%#x) > fault size (%#x)\n",
209 			dax_region->align, fault_size);
210 		return VM_FAULT_SIGBUS;
211 	}
212 
213 	/* dax pud mappings require pfn_t_devmap() */
214 	if ((dax_region->pfn_flags & (PFN_DEV|PFN_MAP)) != (PFN_DEV|PFN_MAP)) {
215 		dev_dbg(dev, "region lacks devmap flags\n");
216 		return VM_FAULT_SIGBUS;
217 	}
218 
219 	if (fault_size < dax_region->align)
220 		return VM_FAULT_SIGBUS;
221 	else if (fault_size > dax_region->align)
222 		return VM_FAULT_FALLBACK;
223 
224 	/* if we are outside of the VMA */
225 	if (pud_addr < vmf->vma->vm_start ||
226 			(pud_addr + PUD_SIZE) > vmf->vma->vm_end)
227 		return VM_FAULT_SIGBUS;
228 
229 	pgoff = linear_page_index(vmf->vma, pud_addr);
230 	phys = dax_pgoff_to_phys(dev_dax, pgoff, PUD_SIZE);
231 	if (phys == -1) {
232 		dev_dbg(dev, "pgoff_to_phys(%#lx) failed\n", pgoff);
233 		return VM_FAULT_SIGBUS;
234 	}
235 
236 	*pfn = phys_to_pfn_t(phys, dax_region->pfn_flags);
237 
238 	return vmf_insert_pfn_pud(vmf->vma, vmf->address, vmf->pud, *pfn,
239 			vmf->flags & FAULT_FLAG_WRITE);
240 }
241 #else
242 static vm_fault_t __dev_dax_pud_fault(struct dev_dax *dev_dax,
243 				struct vm_fault *vmf, pfn_t *pfn)
244 {
245 	return VM_FAULT_FALLBACK;
246 }
247 #endif /* !CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
248 
249 static vm_fault_t dev_dax_huge_fault(struct vm_fault *vmf,
250 		enum page_entry_size pe_size)
251 {
252 	struct file *filp = vmf->vma->vm_file;
253 	unsigned long fault_size;
254 	vm_fault_t rc = VM_FAULT_SIGBUS;
255 	int id;
256 	pfn_t pfn;
257 	struct dev_dax *dev_dax = filp->private_data;
258 
259 	dev_dbg(&dev_dax->dev, "%s: %s (%#lx - %#lx) size = %d\n", current->comm,
260 			(vmf->flags & FAULT_FLAG_WRITE) ? "write" : "read",
261 			vmf->vma->vm_start, vmf->vma->vm_end, pe_size);
262 
263 	id = dax_read_lock();
264 	switch (pe_size) {
265 	case PE_SIZE_PTE:
266 		fault_size = PAGE_SIZE;
267 		rc = __dev_dax_pte_fault(dev_dax, vmf, &pfn);
268 		break;
269 	case PE_SIZE_PMD:
270 		fault_size = PMD_SIZE;
271 		rc = __dev_dax_pmd_fault(dev_dax, vmf, &pfn);
272 		break;
273 	case PE_SIZE_PUD:
274 		fault_size = PUD_SIZE;
275 		rc = __dev_dax_pud_fault(dev_dax, vmf, &pfn);
276 		break;
277 	default:
278 		rc = VM_FAULT_SIGBUS;
279 	}
280 
281 	if (rc == VM_FAULT_NOPAGE) {
282 		unsigned long i;
283 		pgoff_t pgoff;
284 
285 		/*
286 		 * In the device-dax case the only possibility for a
287 		 * VM_FAULT_NOPAGE result is when device-dax capacity is
288 		 * mapped. No need to consider the zero page, or racing
289 		 * conflicting mappings.
290 		 */
291 		pgoff = linear_page_index(vmf->vma, vmf->address
292 				& ~(fault_size - 1));
293 		for (i = 0; i < fault_size / PAGE_SIZE; i++) {
294 			struct page *page;
295 
296 			page = pfn_to_page(pfn_t_to_pfn(pfn) + i);
297 			if (page->mapping)
298 				continue;
299 			page->mapping = filp->f_mapping;
300 			page->index = pgoff + i;
301 		}
302 	}
303 	dax_read_unlock(id);
304 
305 	return rc;
306 }
307 
308 static vm_fault_t dev_dax_fault(struct vm_fault *vmf)
309 {
310 	return dev_dax_huge_fault(vmf, PE_SIZE_PTE);
311 }
312 
313 static int dev_dax_split(struct vm_area_struct *vma, unsigned long addr)
314 {
315 	struct file *filp = vma->vm_file;
316 	struct dev_dax *dev_dax = filp->private_data;
317 	struct dax_region *dax_region = dev_dax->region;
318 
319 	if (!IS_ALIGNED(addr, dax_region->align))
320 		return -EINVAL;
321 	return 0;
322 }
323 
324 static unsigned long dev_dax_pagesize(struct vm_area_struct *vma)
325 {
326 	struct file *filp = vma->vm_file;
327 	struct dev_dax *dev_dax = filp->private_data;
328 	struct dax_region *dax_region = dev_dax->region;
329 
330 	return dax_region->align;
331 }
332 
333 static const struct vm_operations_struct dax_vm_ops = {
334 	.fault = dev_dax_fault,
335 	.huge_fault = dev_dax_huge_fault,
336 	.split = dev_dax_split,
337 	.pagesize = dev_dax_pagesize,
338 };
339 
340 static int dax_mmap(struct file *filp, struct vm_area_struct *vma)
341 {
342 	struct dev_dax *dev_dax = filp->private_data;
343 	int rc, id;
344 
345 	dev_dbg(&dev_dax->dev, "trace\n");
346 
347 	/*
348 	 * We lock to check dax_dev liveness and will re-check at
349 	 * fault time.
350 	 */
351 	id = dax_read_lock();
352 	rc = check_vma(dev_dax, vma, __func__);
353 	dax_read_unlock(id);
354 	if (rc)
355 		return rc;
356 
357 	vma->vm_ops = &dax_vm_ops;
358 	vma->vm_flags |= VM_HUGEPAGE;
359 	return 0;
360 }
361 
362 /* return an unmapped area aligned to the dax region specified alignment */
363 static unsigned long dax_get_unmapped_area(struct file *filp,
364 		unsigned long addr, unsigned long len, unsigned long pgoff,
365 		unsigned long flags)
366 {
367 	unsigned long off, off_end, off_align, len_align, addr_align, align;
368 	struct dev_dax *dev_dax = filp ? filp->private_data : NULL;
369 	struct dax_region *dax_region;
370 
371 	if (!dev_dax || addr)
372 		goto out;
373 
374 	dax_region = dev_dax->region;
375 	align = dax_region->align;
376 	off = pgoff << PAGE_SHIFT;
377 	off_end = off + len;
378 	off_align = round_up(off, align);
379 
380 	if ((off_end <= off_align) || ((off_end - off_align) < align))
381 		goto out;
382 
383 	len_align = len + align;
384 	if ((off + len_align) < off)
385 		goto out;
386 
387 	addr_align = current->mm->get_unmapped_area(filp, addr, len_align,
388 			pgoff, flags);
389 	if (!IS_ERR_VALUE(addr_align)) {
390 		addr_align += (off - addr_align) & (align - 1);
391 		return addr_align;
392 	}
393  out:
394 	return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
395 }
396 
397 static const struct address_space_operations dev_dax_aops = {
398 	.set_page_dirty		= noop_set_page_dirty,
399 	.invalidatepage		= noop_invalidatepage,
400 };
401 
402 static int dax_open(struct inode *inode, struct file *filp)
403 {
404 	struct dax_device *dax_dev = inode_dax(inode);
405 	struct inode *__dax_inode = dax_inode(dax_dev);
406 	struct dev_dax *dev_dax = dax_get_private(dax_dev);
407 
408 	dev_dbg(&dev_dax->dev, "trace\n");
409 	inode->i_mapping = __dax_inode->i_mapping;
410 	inode->i_mapping->host = __dax_inode;
411 	inode->i_mapping->a_ops = &dev_dax_aops;
412 	filp->f_mapping = inode->i_mapping;
413 	filp->f_wb_err = filemap_sample_wb_err(filp->f_mapping);
414 	filp->private_data = dev_dax;
415 	inode->i_flags = S_DAX;
416 
417 	return 0;
418 }
419 
420 static int dax_release(struct inode *inode, struct file *filp)
421 {
422 	struct dev_dax *dev_dax = filp->private_data;
423 
424 	dev_dbg(&dev_dax->dev, "trace\n");
425 	return 0;
426 }
427 
428 static const struct file_operations dax_fops = {
429 	.llseek = noop_llseek,
430 	.owner = THIS_MODULE,
431 	.open = dax_open,
432 	.release = dax_release,
433 	.get_unmapped_area = dax_get_unmapped_area,
434 	.mmap = dax_mmap,
435 	.mmap_supported_flags = MAP_SYNC,
436 };
437 
438 static void dev_dax_cdev_del(void *cdev)
439 {
440 	cdev_del(cdev);
441 }
442 
443 static void dev_dax_kill(void *dev_dax)
444 {
445 	kill_dev_dax(dev_dax);
446 }
447 
448 int dev_dax_probe(struct device *dev)
449 {
450 	struct dev_dax *dev_dax = to_dev_dax(dev);
451 	struct dax_device *dax_dev = dev_dax->dax_dev;
452 	struct resource *res = &dev_dax->region->res;
453 	struct inode *inode;
454 	struct cdev *cdev;
455 	void *addr;
456 	int rc;
457 
458 	/* 1:1 map region resource range to device-dax instance range */
459 	if (!devm_request_mem_region(dev, res->start, resource_size(res),
460 				dev_name(dev))) {
461 		dev_warn(dev, "could not reserve region %pR\n", res);
462 		return -EBUSY;
463 	}
464 
465 	init_completion(&dev_dax->cmp);
466 	rc = percpu_ref_init(&dev_dax->ref, dev_dax_percpu_release, 0,
467 			GFP_KERNEL);
468 	if (rc)
469 		return rc;
470 
471 	rc = devm_add_action_or_reset(dev, dev_dax_percpu_exit, &dev_dax->ref);
472 	if (rc)
473 		return rc;
474 
475 	dev_dax->pgmap.ref = &dev_dax->ref;
476 	dev_dax->pgmap.kill = dev_dax_percpu_kill;
477 	addr = devm_memremap_pages(dev, &dev_dax->pgmap);
478 	if (IS_ERR(addr)) {
479 		devm_remove_action(dev, dev_dax_percpu_exit, &dev_dax->ref);
480 		percpu_ref_exit(&dev_dax->ref);
481 		return PTR_ERR(addr);
482 	}
483 
484 	inode = dax_inode(dax_dev);
485 	cdev = inode->i_cdev;
486 	cdev_init(cdev, &dax_fops);
487 	if (dev->class) {
488 		/* for the CONFIG_DEV_DAX_PMEM_COMPAT case */
489 		cdev->owner = dev->parent->driver->owner;
490 	} else
491 		cdev->owner = dev->driver->owner;
492 	cdev_set_parent(cdev, &dev->kobj);
493 	rc = cdev_add(cdev, dev->devt, 1);
494 	if (rc)
495 		return rc;
496 
497 	rc = devm_add_action_or_reset(dev, dev_dax_cdev_del, cdev);
498 	if (rc)
499 		return rc;
500 
501 	run_dax(dax_dev);
502 	return devm_add_action_or_reset(dev, dev_dax_kill, dev_dax);
503 }
504 EXPORT_SYMBOL_GPL(dev_dax_probe);
505 
506 static int dev_dax_remove(struct device *dev)
507 {
508 	/* all probe actions are unwound by devm */
509 	return 0;
510 }
511 
512 static struct dax_device_driver device_dax_driver = {
513 	.drv = {
514 		.probe = dev_dax_probe,
515 		.remove = dev_dax_remove,
516 	},
517 	.match_always = 1,
518 };
519 
520 static int __init dax_init(void)
521 {
522 	return dax_driver_register(&device_dax_driver);
523 }
524 
525 static void __exit dax_exit(void)
526 {
527 	dax_driver_unregister(&device_dax_driver);
528 }
529 
530 MODULE_AUTHOR("Intel Corporation");
531 MODULE_LICENSE("GPL v2");
532 module_init(dax_init);
533 module_exit(dax_exit);
534 MODULE_ALIAS_DAX_DEVICE(0);
535