xref: /openbmc/linux/drivers/dax/device.c (revision fb574682)
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 dax_region *dax_region = dev_dax->region;
21 	struct device *dev = &dev_dax->dev;
22 	unsigned long mask;
23 
24 	if (!dax_alive(dev_dax->dax_dev))
25 		return -ENXIO;
26 
27 	/* prevent private mappings from being established */
28 	if ((vma->vm_flags & VM_MAYSHARE) != VM_MAYSHARE) {
29 		dev_info_ratelimited(dev,
30 				"%s: %s: fail, attempted private mapping\n",
31 				current->comm, func);
32 		return -EINVAL;
33 	}
34 
35 	mask = dax_region->align - 1;
36 	if (vma->vm_start & mask || vma->vm_end & mask) {
37 		dev_info_ratelimited(dev,
38 				"%s: %s: fail, unaligned vma (%#lx - %#lx, %#lx)\n",
39 				current->comm, func, vma->vm_start, vma->vm_end,
40 				mask);
41 		return -EINVAL;
42 	}
43 
44 	if ((dax_region->pfn_flags & (PFN_DEV|PFN_MAP)) == PFN_DEV
45 			&& (vma->vm_flags & VM_DONTCOPY) == 0) {
46 		dev_info_ratelimited(dev,
47 				"%s: %s: fail, dax range requires MADV_DONTFORK\n",
48 				current->comm, func);
49 		return -EINVAL;
50 	}
51 
52 	if (!vma_is_dax(vma)) {
53 		dev_info_ratelimited(dev,
54 				"%s: %s: fail, vma is not DAX capable\n",
55 				current->comm, func);
56 		return -EINVAL;
57 	}
58 
59 	return 0;
60 }
61 
62 /* see "strong" declaration in tools/testing/nvdimm/dax-dev.c */
63 __weak phys_addr_t dax_pgoff_to_phys(struct dev_dax *dev_dax, pgoff_t pgoff,
64 		unsigned long size)
65 {
66 	struct resource *res = &dev_dax->region->res;
67 	phys_addr_t phys;
68 
69 	phys = pgoff * PAGE_SIZE + res->start;
70 	if (phys >= res->start && phys <= res->end) {
71 		if (phys + size - 1 <= res->end)
72 			return phys;
73 	}
74 
75 	return -1;
76 }
77 
78 static vm_fault_t __dev_dax_pte_fault(struct dev_dax *dev_dax,
79 				struct vm_fault *vmf, pfn_t *pfn)
80 {
81 	struct device *dev = &dev_dax->dev;
82 	struct dax_region *dax_region;
83 	phys_addr_t phys;
84 	unsigned int fault_size = PAGE_SIZE;
85 
86 	if (check_vma(dev_dax, vmf->vma, __func__))
87 		return VM_FAULT_SIGBUS;
88 
89 	dax_region = dev_dax->region;
90 	if (dax_region->align > PAGE_SIZE) {
91 		dev_dbg(dev, "alignment (%#x) > fault size (%#x)\n",
92 			dax_region->align, fault_size);
93 		return VM_FAULT_SIGBUS;
94 	}
95 
96 	if (fault_size != dax_region->align)
97 		return VM_FAULT_SIGBUS;
98 
99 	phys = dax_pgoff_to_phys(dev_dax, vmf->pgoff, PAGE_SIZE);
100 	if (phys == -1) {
101 		dev_dbg(dev, "pgoff_to_phys(%#lx) failed\n", vmf->pgoff);
102 		return VM_FAULT_SIGBUS;
103 	}
104 
105 	*pfn = phys_to_pfn_t(phys, dax_region->pfn_flags);
106 
107 	return vmf_insert_mixed(vmf->vma, vmf->address, *pfn);
108 }
109 
110 static vm_fault_t __dev_dax_pmd_fault(struct dev_dax *dev_dax,
111 				struct vm_fault *vmf, pfn_t *pfn)
112 {
113 	unsigned long pmd_addr = vmf->address & PMD_MASK;
114 	struct device *dev = &dev_dax->dev;
115 	struct dax_region *dax_region;
116 	phys_addr_t phys;
117 	pgoff_t pgoff;
118 	unsigned int fault_size = PMD_SIZE;
119 
120 	if (check_vma(dev_dax, vmf->vma, __func__))
121 		return VM_FAULT_SIGBUS;
122 
123 	dax_region = dev_dax->region;
124 	if (dax_region->align > PMD_SIZE) {
125 		dev_dbg(dev, "alignment (%#x) > fault size (%#x)\n",
126 			dax_region->align, fault_size);
127 		return VM_FAULT_SIGBUS;
128 	}
129 
130 	/* dax pmd mappings require pfn_t_devmap() */
131 	if ((dax_region->pfn_flags & (PFN_DEV|PFN_MAP)) != (PFN_DEV|PFN_MAP)) {
132 		dev_dbg(dev, "region lacks devmap flags\n");
133 		return VM_FAULT_SIGBUS;
134 	}
135 
136 	if (fault_size < dax_region->align)
137 		return VM_FAULT_SIGBUS;
138 	else if (fault_size > dax_region->align)
139 		return VM_FAULT_FALLBACK;
140 
141 	/* if we are outside of the VMA */
142 	if (pmd_addr < vmf->vma->vm_start ||
143 			(pmd_addr + PMD_SIZE) > vmf->vma->vm_end)
144 		return VM_FAULT_SIGBUS;
145 
146 	pgoff = linear_page_index(vmf->vma, pmd_addr);
147 	phys = dax_pgoff_to_phys(dev_dax, pgoff, PMD_SIZE);
148 	if (phys == -1) {
149 		dev_dbg(dev, "pgoff_to_phys(%#lx) failed\n", pgoff);
150 		return VM_FAULT_SIGBUS;
151 	}
152 
153 	*pfn = phys_to_pfn_t(phys, dax_region->pfn_flags);
154 
155 	return vmf_insert_pfn_pmd(vmf, *pfn, vmf->flags & FAULT_FLAG_WRITE);
156 }
157 
158 #ifdef CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD
159 static vm_fault_t __dev_dax_pud_fault(struct dev_dax *dev_dax,
160 				struct vm_fault *vmf, pfn_t *pfn)
161 {
162 	unsigned long pud_addr = vmf->address & PUD_MASK;
163 	struct device *dev = &dev_dax->dev;
164 	struct dax_region *dax_region;
165 	phys_addr_t phys;
166 	pgoff_t pgoff;
167 	unsigned int fault_size = PUD_SIZE;
168 
169 
170 	if (check_vma(dev_dax, vmf->vma, __func__))
171 		return VM_FAULT_SIGBUS;
172 
173 	dax_region = dev_dax->region;
174 	if (dax_region->align > PUD_SIZE) {
175 		dev_dbg(dev, "alignment (%#x) > fault size (%#x)\n",
176 			dax_region->align, fault_size);
177 		return VM_FAULT_SIGBUS;
178 	}
179 
180 	/* dax pud mappings require pfn_t_devmap() */
181 	if ((dax_region->pfn_flags & (PFN_DEV|PFN_MAP)) != (PFN_DEV|PFN_MAP)) {
182 		dev_dbg(dev, "region lacks devmap flags\n");
183 		return VM_FAULT_SIGBUS;
184 	}
185 
186 	if (fault_size < dax_region->align)
187 		return VM_FAULT_SIGBUS;
188 	else if (fault_size > dax_region->align)
189 		return VM_FAULT_FALLBACK;
190 
191 	/* if we are outside of the VMA */
192 	if (pud_addr < vmf->vma->vm_start ||
193 			(pud_addr + PUD_SIZE) > vmf->vma->vm_end)
194 		return VM_FAULT_SIGBUS;
195 
196 	pgoff = linear_page_index(vmf->vma, pud_addr);
197 	phys = dax_pgoff_to_phys(dev_dax, pgoff, PUD_SIZE);
198 	if (phys == -1) {
199 		dev_dbg(dev, "pgoff_to_phys(%#lx) failed\n", pgoff);
200 		return VM_FAULT_SIGBUS;
201 	}
202 
203 	*pfn = phys_to_pfn_t(phys, dax_region->pfn_flags);
204 
205 	return vmf_insert_pfn_pud(vmf, *pfn, vmf->flags & FAULT_FLAG_WRITE);
206 }
207 #else
208 static vm_fault_t __dev_dax_pud_fault(struct dev_dax *dev_dax,
209 				struct vm_fault *vmf, pfn_t *pfn)
210 {
211 	return VM_FAULT_FALLBACK;
212 }
213 #endif /* !CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD */
214 
215 static vm_fault_t dev_dax_huge_fault(struct vm_fault *vmf,
216 		enum page_entry_size pe_size)
217 {
218 	struct file *filp = vmf->vma->vm_file;
219 	unsigned long fault_size;
220 	vm_fault_t rc = VM_FAULT_SIGBUS;
221 	int id;
222 	pfn_t pfn;
223 	struct dev_dax *dev_dax = filp->private_data;
224 
225 	dev_dbg(&dev_dax->dev, "%s: %s (%#lx - %#lx) size = %d\n", current->comm,
226 			(vmf->flags & FAULT_FLAG_WRITE) ? "write" : "read",
227 			vmf->vma->vm_start, vmf->vma->vm_end, pe_size);
228 
229 	id = dax_read_lock();
230 	switch (pe_size) {
231 	case PE_SIZE_PTE:
232 		fault_size = PAGE_SIZE;
233 		rc = __dev_dax_pte_fault(dev_dax, vmf, &pfn);
234 		break;
235 	case PE_SIZE_PMD:
236 		fault_size = PMD_SIZE;
237 		rc = __dev_dax_pmd_fault(dev_dax, vmf, &pfn);
238 		break;
239 	case PE_SIZE_PUD:
240 		fault_size = PUD_SIZE;
241 		rc = __dev_dax_pud_fault(dev_dax, vmf, &pfn);
242 		break;
243 	default:
244 		rc = VM_FAULT_SIGBUS;
245 	}
246 
247 	if (rc == VM_FAULT_NOPAGE) {
248 		unsigned long i;
249 		pgoff_t pgoff;
250 
251 		/*
252 		 * In the device-dax case the only possibility for a
253 		 * VM_FAULT_NOPAGE result is when device-dax capacity is
254 		 * mapped. No need to consider the zero page, or racing
255 		 * conflicting mappings.
256 		 */
257 		pgoff = linear_page_index(vmf->vma, vmf->address
258 				& ~(fault_size - 1));
259 		for (i = 0; i < fault_size / PAGE_SIZE; i++) {
260 			struct page *page;
261 
262 			page = pfn_to_page(pfn_t_to_pfn(pfn) + i);
263 			if (page->mapping)
264 				continue;
265 			page->mapping = filp->f_mapping;
266 			page->index = pgoff + i;
267 		}
268 	}
269 	dax_read_unlock(id);
270 
271 	return rc;
272 }
273 
274 static vm_fault_t dev_dax_fault(struct vm_fault *vmf)
275 {
276 	return dev_dax_huge_fault(vmf, PE_SIZE_PTE);
277 }
278 
279 static int dev_dax_split(struct vm_area_struct *vma, unsigned long addr)
280 {
281 	struct file *filp = vma->vm_file;
282 	struct dev_dax *dev_dax = filp->private_data;
283 	struct dax_region *dax_region = dev_dax->region;
284 
285 	if (!IS_ALIGNED(addr, dax_region->align))
286 		return -EINVAL;
287 	return 0;
288 }
289 
290 static unsigned long dev_dax_pagesize(struct vm_area_struct *vma)
291 {
292 	struct file *filp = vma->vm_file;
293 	struct dev_dax *dev_dax = filp->private_data;
294 	struct dax_region *dax_region = dev_dax->region;
295 
296 	return dax_region->align;
297 }
298 
299 static const struct vm_operations_struct dax_vm_ops = {
300 	.fault = dev_dax_fault,
301 	.huge_fault = dev_dax_huge_fault,
302 	.split = dev_dax_split,
303 	.pagesize = dev_dax_pagesize,
304 };
305 
306 static int dax_mmap(struct file *filp, struct vm_area_struct *vma)
307 {
308 	struct dev_dax *dev_dax = filp->private_data;
309 	int rc, id;
310 
311 	dev_dbg(&dev_dax->dev, "trace\n");
312 
313 	/*
314 	 * We lock to check dax_dev liveness and will re-check at
315 	 * fault time.
316 	 */
317 	id = dax_read_lock();
318 	rc = check_vma(dev_dax, vma, __func__);
319 	dax_read_unlock(id);
320 	if (rc)
321 		return rc;
322 
323 	vma->vm_ops = &dax_vm_ops;
324 	vma->vm_flags |= VM_HUGEPAGE;
325 	return 0;
326 }
327 
328 /* return an unmapped area aligned to the dax region specified alignment */
329 static unsigned long dax_get_unmapped_area(struct file *filp,
330 		unsigned long addr, unsigned long len, unsigned long pgoff,
331 		unsigned long flags)
332 {
333 	unsigned long off, off_end, off_align, len_align, addr_align, align;
334 	struct dev_dax *dev_dax = filp ? filp->private_data : NULL;
335 	struct dax_region *dax_region;
336 
337 	if (!dev_dax || addr)
338 		goto out;
339 
340 	dax_region = dev_dax->region;
341 	align = dax_region->align;
342 	off = pgoff << PAGE_SHIFT;
343 	off_end = off + len;
344 	off_align = round_up(off, align);
345 
346 	if ((off_end <= off_align) || ((off_end - off_align) < align))
347 		goto out;
348 
349 	len_align = len + align;
350 	if ((off + len_align) < off)
351 		goto out;
352 
353 	addr_align = current->mm->get_unmapped_area(filp, addr, len_align,
354 			pgoff, flags);
355 	if (!IS_ERR_VALUE(addr_align)) {
356 		addr_align += (off - addr_align) & (align - 1);
357 		return addr_align;
358 	}
359  out:
360 	return current->mm->get_unmapped_area(filp, addr, len, pgoff, flags);
361 }
362 
363 static const struct address_space_operations dev_dax_aops = {
364 	.set_page_dirty		= noop_set_page_dirty,
365 	.invalidatepage		= noop_invalidatepage,
366 };
367 
368 static int dax_open(struct inode *inode, struct file *filp)
369 {
370 	struct dax_device *dax_dev = inode_dax(inode);
371 	struct inode *__dax_inode = dax_inode(dax_dev);
372 	struct dev_dax *dev_dax = dax_get_private(dax_dev);
373 
374 	dev_dbg(&dev_dax->dev, "trace\n");
375 	inode->i_mapping = __dax_inode->i_mapping;
376 	inode->i_mapping->host = __dax_inode;
377 	inode->i_mapping->a_ops = &dev_dax_aops;
378 	filp->f_mapping = inode->i_mapping;
379 	filp->f_wb_err = filemap_sample_wb_err(filp->f_mapping);
380 	filp->f_sb_err = file_sample_sb_err(filp);
381 	filp->private_data = dev_dax;
382 	inode->i_flags = S_DAX;
383 
384 	return 0;
385 }
386 
387 static int dax_release(struct inode *inode, struct file *filp)
388 {
389 	struct dev_dax *dev_dax = filp->private_data;
390 
391 	dev_dbg(&dev_dax->dev, "trace\n");
392 	return 0;
393 }
394 
395 static const struct file_operations dax_fops = {
396 	.llseek = noop_llseek,
397 	.owner = THIS_MODULE,
398 	.open = dax_open,
399 	.release = dax_release,
400 	.get_unmapped_area = dax_get_unmapped_area,
401 	.mmap = dax_mmap,
402 	.mmap_supported_flags = MAP_SYNC,
403 };
404 
405 static void dev_dax_cdev_del(void *cdev)
406 {
407 	cdev_del(cdev);
408 }
409 
410 static void dev_dax_kill(void *dev_dax)
411 {
412 	kill_dev_dax(dev_dax);
413 }
414 
415 int dev_dax_probe(struct device *dev)
416 {
417 	struct dev_dax *dev_dax = to_dev_dax(dev);
418 	struct dax_device *dax_dev = dev_dax->dax_dev;
419 	struct resource *res = &dev_dax->region->res;
420 	struct inode *inode;
421 	struct cdev *cdev;
422 	void *addr;
423 	int rc;
424 
425 	/* 1:1 map region resource range to device-dax instance range */
426 	if (!devm_request_mem_region(dev, res->start, resource_size(res),
427 				dev_name(dev))) {
428 		dev_warn(dev, "could not reserve region %pR\n", res);
429 		return -EBUSY;
430 	}
431 
432 	dev_dax->pgmap.type = MEMORY_DEVICE_DEVDAX;
433 	addr = devm_memremap_pages(dev, &dev_dax->pgmap);
434 	if (IS_ERR(addr))
435 		return PTR_ERR(addr);
436 
437 	inode = dax_inode(dax_dev);
438 	cdev = inode->i_cdev;
439 	cdev_init(cdev, &dax_fops);
440 	if (dev->class) {
441 		/* for the CONFIG_DEV_DAX_PMEM_COMPAT case */
442 		cdev->owner = dev->parent->driver->owner;
443 	} else
444 		cdev->owner = dev->driver->owner;
445 	cdev_set_parent(cdev, &dev->kobj);
446 	rc = cdev_add(cdev, dev->devt, 1);
447 	if (rc)
448 		return rc;
449 
450 	rc = devm_add_action_or_reset(dev, dev_dax_cdev_del, cdev);
451 	if (rc)
452 		return rc;
453 
454 	run_dax(dax_dev);
455 	return devm_add_action_or_reset(dev, dev_dax_kill, dev_dax);
456 }
457 EXPORT_SYMBOL_GPL(dev_dax_probe);
458 
459 static int dev_dax_remove(struct device *dev)
460 {
461 	/* all probe actions are unwound by devm */
462 	return 0;
463 }
464 
465 static struct dax_device_driver device_dax_driver = {
466 	.drv = {
467 		.probe = dev_dax_probe,
468 		.remove = dev_dax_remove,
469 	},
470 	.match_always = 1,
471 };
472 
473 static int __init dax_init(void)
474 {
475 	return dax_driver_register(&device_dax_driver);
476 }
477 
478 static void __exit dax_exit(void)
479 {
480 	dax_driver_unregister(&device_dax_driver);
481 }
482 
483 MODULE_AUTHOR("Intel Corporation");
484 MODULE_LICENSE("GPL v2");
485 module_init(dax_init);
486 module_exit(dax_exit);
487 MODULE_ALIAS_DAX_DEVICE(0);
488