1 // SPDX-License-Identifier: GPL-2.0-or-later 1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /* 2 /*
3 * Copyright 2013 Red Hat Inc. 3 * Copyright 2013 Red Hat Inc.
4 * 4 *
5 * Authors: Jérôme Glisse <jglisse@redhat.co 5 * Authors: Jérôme Glisse <jglisse@redhat.com>
6 */ 6 */
7 /* 7 /*
8 * Refer to include/linux/hmm.h for informatio 8 * Refer to include/linux/hmm.h for information about heterogeneous memory
9 * management or HMM for short. 9 * management or HMM for short.
10 */ 10 */
11 #include <linux/pagewalk.h> 11 #include <linux/pagewalk.h>
12 #include <linux/hmm.h> 12 #include <linux/hmm.h>
13 #include <linux/init.h> 13 #include <linux/init.h>
14 #include <linux/rmap.h> 14 #include <linux/rmap.h>
15 #include <linux/swap.h> 15 #include <linux/swap.h>
16 #include <linux/slab.h> 16 #include <linux/slab.h>
17 #include <linux/sched.h> 17 #include <linux/sched.h>
18 #include <linux/mmzone.h> 18 #include <linux/mmzone.h>
19 #include <linux/pagemap.h> 19 #include <linux/pagemap.h>
20 #include <linux/swapops.h> 20 #include <linux/swapops.h>
21 #include <linux/hugetlb.h> 21 #include <linux/hugetlb.h>
22 #include <linux/memremap.h> 22 #include <linux/memremap.h>
23 #include <linux/sched/mm.h> 23 #include <linux/sched/mm.h>
24 #include <linux/jump_label.h> 24 #include <linux/jump_label.h>
25 #include <linux/dma-mapping.h> 25 #include <linux/dma-mapping.h>
26 #include <linux/mmu_notifier.h> 26 #include <linux/mmu_notifier.h>
27 #include <linux/memory_hotplug.h> 27 #include <linux/memory_hotplug.h>
28 28
29 #include "internal.h" <<
30 <<
31 struct hmm_vma_walk { 29 struct hmm_vma_walk {
32 struct hmm_range *range; 30 struct hmm_range *range;
33 unsigned long last; 31 unsigned long last;
34 }; 32 };
35 33
36 enum { 34 enum {
37 HMM_NEED_FAULT = 1 << 0, 35 HMM_NEED_FAULT = 1 << 0,
38 HMM_NEED_WRITE_FAULT = 1 << 1, 36 HMM_NEED_WRITE_FAULT = 1 << 1,
39 HMM_NEED_ALL_BITS = HMM_NEED_FAULT | H 37 HMM_NEED_ALL_BITS = HMM_NEED_FAULT | HMM_NEED_WRITE_FAULT,
40 }; 38 };
41 39
42 static int hmm_pfns_fill(unsigned long addr, u 40 static int hmm_pfns_fill(unsigned long addr, unsigned long end,
43 struct hmm_range *ran 41 struct hmm_range *range, unsigned long cpu_flags)
44 { 42 {
45 unsigned long i = (addr - range->start 43 unsigned long i = (addr - range->start) >> PAGE_SHIFT;
46 44
47 for (; addr < end; addr += PAGE_SIZE, 45 for (; addr < end; addr += PAGE_SIZE, i++)
48 range->hmm_pfns[i] = cpu_flags 46 range->hmm_pfns[i] = cpu_flags;
49 return 0; 47 return 0;
50 } 48 }
51 49
52 /* 50 /*
53 * hmm_vma_fault() - fault in a range lacking 51 * hmm_vma_fault() - fault in a range lacking valid pmd or pte(s)
54 * @addr: range virtual start address (inclusi 52 * @addr: range virtual start address (inclusive)
55 * @end: range virtual end address (exclusive) 53 * @end: range virtual end address (exclusive)
56 * @required_fault: HMM_NEED_* flags 54 * @required_fault: HMM_NEED_* flags
57 * @walk: mm_walk structure 55 * @walk: mm_walk structure
58 * Return: -EBUSY after page fault, or page fa 56 * Return: -EBUSY after page fault, or page fault error
59 * 57 *
60 * This function will be called whenever pmd_n 58 * This function will be called whenever pmd_none() or pte_none() returns true,
61 * or whenever there is no page directory cove 59 * or whenever there is no page directory covering the virtual address range.
62 */ 60 */
63 static int hmm_vma_fault(unsigned long addr, u 61 static int hmm_vma_fault(unsigned long addr, unsigned long end,
64 unsigned int required 62 unsigned int required_fault, struct mm_walk *walk)
65 { 63 {
66 struct hmm_vma_walk *hmm_vma_walk = wa 64 struct hmm_vma_walk *hmm_vma_walk = walk->private;
67 struct vm_area_struct *vma = walk->vma 65 struct vm_area_struct *vma = walk->vma;
68 unsigned int fault_flags = FAULT_FLAG_ 66 unsigned int fault_flags = FAULT_FLAG_REMOTE;
69 67
70 WARN_ON_ONCE(!required_fault); 68 WARN_ON_ONCE(!required_fault);
71 hmm_vma_walk->last = addr; 69 hmm_vma_walk->last = addr;
72 70
73 if (required_fault & HMM_NEED_WRITE_FA 71 if (required_fault & HMM_NEED_WRITE_FAULT) {
74 if (!(vma->vm_flags & VM_WRITE 72 if (!(vma->vm_flags & VM_WRITE))
75 return -EPERM; 73 return -EPERM;
76 fault_flags |= FAULT_FLAG_WRIT 74 fault_flags |= FAULT_FLAG_WRITE;
77 } 75 }
78 76
79 for (; addr < end; addr += PAGE_SIZE) 77 for (; addr < end; addr += PAGE_SIZE)
80 if (handle_mm_fault(vma, addr, 78 if (handle_mm_fault(vma, addr, fault_flags, NULL) &
81 VM_FAULT_ERROR) 79 VM_FAULT_ERROR)
82 return -EFAULT; 80 return -EFAULT;
83 return -EBUSY; 81 return -EBUSY;
84 } 82 }
85 83
86 static unsigned int hmm_pte_need_fault(const s 84 static unsigned int hmm_pte_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
87 unsigne 85 unsigned long pfn_req_flags,
88 unsigne 86 unsigned long cpu_flags)
89 { 87 {
90 struct hmm_range *range = hmm_vma_walk 88 struct hmm_range *range = hmm_vma_walk->range;
91 89
92 /* 90 /*
93 * So we not only consider the individ 91 * So we not only consider the individual per page request we also
94 * consider the default flags requeste 92 * consider the default flags requested for the range. The API can
95 * be used 2 ways. The first one where 93 * be used 2 ways. The first one where the HMM user coalesces
96 * multiple page faults into one reque 94 * multiple page faults into one request and sets flags per pfn for
97 * those faults. The second one where 95 * those faults. The second one where the HMM user wants to pre-
98 * fault a range with specific flags. 96 * fault a range with specific flags. For the latter one it is a
99 * waste to have the user pre-fill the 97 * waste to have the user pre-fill the pfn arrays with a default
100 * flags value. 98 * flags value.
101 */ 99 */
102 pfn_req_flags &= range->pfn_flags_mask 100 pfn_req_flags &= range->pfn_flags_mask;
103 pfn_req_flags |= range->default_flags; 101 pfn_req_flags |= range->default_flags;
104 102
105 /* We aren't ask to do anything ... */ 103 /* We aren't ask to do anything ... */
106 if (!(pfn_req_flags & HMM_PFN_REQ_FAUL 104 if (!(pfn_req_flags & HMM_PFN_REQ_FAULT))
107 return 0; 105 return 0;
108 106
109 /* Need to write fault ? */ 107 /* Need to write fault ? */
110 if ((pfn_req_flags & HMM_PFN_REQ_WRITE 108 if ((pfn_req_flags & HMM_PFN_REQ_WRITE) &&
111 !(cpu_flags & HMM_PFN_WRITE)) 109 !(cpu_flags & HMM_PFN_WRITE))
112 return HMM_NEED_FAULT | HMM_NE 110 return HMM_NEED_FAULT | HMM_NEED_WRITE_FAULT;
113 111
114 /* If CPU page table is not valid then 112 /* If CPU page table is not valid then we need to fault */
115 if (!(cpu_flags & HMM_PFN_VALID)) 113 if (!(cpu_flags & HMM_PFN_VALID))
116 return HMM_NEED_FAULT; 114 return HMM_NEED_FAULT;
117 return 0; 115 return 0;
118 } 116 }
119 117
120 static unsigned int 118 static unsigned int
121 hmm_range_need_fault(const struct hmm_vma_walk 119 hmm_range_need_fault(const struct hmm_vma_walk *hmm_vma_walk,
122 const unsigned long hmm_p 120 const unsigned long hmm_pfns[], unsigned long npages,
123 unsigned long cpu_flags) 121 unsigned long cpu_flags)
124 { 122 {
125 struct hmm_range *range = hmm_vma_walk 123 struct hmm_range *range = hmm_vma_walk->range;
126 unsigned int required_fault = 0; 124 unsigned int required_fault = 0;
127 unsigned long i; 125 unsigned long i;
128 126
129 /* 127 /*
130 * If the default flags do not request 128 * If the default flags do not request to fault pages, and the mask does
131 * not allow for individual pages to b 129 * not allow for individual pages to be faulted, then
132 * hmm_pte_need_fault() will always re 130 * hmm_pte_need_fault() will always return 0.
133 */ 131 */
134 if (!((range->default_flags | range->p 132 if (!((range->default_flags | range->pfn_flags_mask) &
135 HMM_PFN_REQ_FAULT)) 133 HMM_PFN_REQ_FAULT))
136 return 0; 134 return 0;
137 135
138 for (i = 0; i < npages; ++i) { 136 for (i = 0; i < npages; ++i) {
139 required_fault |= hmm_pte_need 137 required_fault |= hmm_pte_need_fault(hmm_vma_walk, hmm_pfns[i],
140 138 cpu_flags);
141 if (required_fault == HMM_NEED 139 if (required_fault == HMM_NEED_ALL_BITS)
142 return required_fault; 140 return required_fault;
143 } 141 }
144 return required_fault; 142 return required_fault;
145 } 143 }
146 144
147 static int hmm_vma_walk_hole(unsigned long add 145 static int hmm_vma_walk_hole(unsigned long addr, unsigned long end,
148 __always_unused i 146 __always_unused int depth, struct mm_walk *walk)
149 { 147 {
150 struct hmm_vma_walk *hmm_vma_walk = wa 148 struct hmm_vma_walk *hmm_vma_walk = walk->private;
151 struct hmm_range *range = hmm_vma_walk 149 struct hmm_range *range = hmm_vma_walk->range;
152 unsigned int required_fault; 150 unsigned int required_fault;
153 unsigned long i, npages; 151 unsigned long i, npages;
154 unsigned long *hmm_pfns; 152 unsigned long *hmm_pfns;
155 153
156 i = (addr - range->start) >> PAGE_SHIF 154 i = (addr - range->start) >> PAGE_SHIFT;
157 npages = (end - addr) >> PAGE_SHIFT; 155 npages = (end - addr) >> PAGE_SHIFT;
158 hmm_pfns = &range->hmm_pfns[i]; 156 hmm_pfns = &range->hmm_pfns[i];
159 required_fault = 157 required_fault =
160 hmm_range_need_fault(hmm_vma_w 158 hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, 0);
161 if (!walk->vma) { 159 if (!walk->vma) {
162 if (required_fault) 160 if (required_fault)
163 return -EFAULT; 161 return -EFAULT;
164 return hmm_pfns_fill(addr, end 162 return hmm_pfns_fill(addr, end, range, HMM_PFN_ERROR);
165 } 163 }
166 if (required_fault) 164 if (required_fault)
167 return hmm_vma_fault(addr, end 165 return hmm_vma_fault(addr, end, required_fault, walk);
168 return hmm_pfns_fill(addr, end, range, 166 return hmm_pfns_fill(addr, end, range, 0);
169 } 167 }
170 168
171 static inline unsigned long hmm_pfn_flags_orde 169 static inline unsigned long hmm_pfn_flags_order(unsigned long order)
172 { 170 {
173 return order << HMM_PFN_ORDER_SHIFT; 171 return order << HMM_PFN_ORDER_SHIFT;
174 } 172 }
175 173
176 static inline unsigned long pmd_to_hmm_pfn_fla 174 static inline unsigned long pmd_to_hmm_pfn_flags(struct hmm_range *range,
177 175 pmd_t pmd)
178 { 176 {
179 if (pmd_protnone(pmd)) 177 if (pmd_protnone(pmd))
180 return 0; 178 return 0;
181 return (pmd_write(pmd) ? (HMM_PFN_VALI 179 return (pmd_write(pmd) ? (HMM_PFN_VALID | HMM_PFN_WRITE) :
182 HMM_PFN_VALID 180 HMM_PFN_VALID) |
183 hmm_pfn_flags_order(PMD_SHIFT - 181 hmm_pfn_flags_order(PMD_SHIFT - PAGE_SHIFT);
184 } 182 }
185 183
186 #ifdef CONFIG_TRANSPARENT_HUGEPAGE 184 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
187 static int hmm_vma_handle_pmd(struct mm_walk * 185 static int hmm_vma_handle_pmd(struct mm_walk *walk, unsigned long addr,
188 unsigned long en 186 unsigned long end, unsigned long hmm_pfns[],
189 pmd_t pmd) 187 pmd_t pmd)
190 { 188 {
191 struct hmm_vma_walk *hmm_vma_walk = wa 189 struct hmm_vma_walk *hmm_vma_walk = walk->private;
192 struct hmm_range *range = hmm_vma_walk 190 struct hmm_range *range = hmm_vma_walk->range;
193 unsigned long pfn, npages, i; 191 unsigned long pfn, npages, i;
194 unsigned int required_fault; 192 unsigned int required_fault;
195 unsigned long cpu_flags; 193 unsigned long cpu_flags;
196 194
197 npages = (end - addr) >> PAGE_SHIFT; 195 npages = (end - addr) >> PAGE_SHIFT;
198 cpu_flags = pmd_to_hmm_pfn_flags(range 196 cpu_flags = pmd_to_hmm_pfn_flags(range, pmd);
199 required_fault = 197 required_fault =
200 hmm_range_need_fault(hmm_vma_w 198 hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, cpu_flags);
201 if (required_fault) 199 if (required_fault)
202 return hmm_vma_fault(addr, end 200 return hmm_vma_fault(addr, end, required_fault, walk);
203 201
204 pfn = pmd_pfn(pmd) + ((addr & ~PMD_MAS 202 pfn = pmd_pfn(pmd) + ((addr & ~PMD_MASK) >> PAGE_SHIFT);
205 for (i = 0; addr < end; addr += PAGE_S 203 for (i = 0; addr < end; addr += PAGE_SIZE, i++, pfn++)
206 hmm_pfns[i] = pfn | cpu_flags; 204 hmm_pfns[i] = pfn | cpu_flags;
207 return 0; 205 return 0;
208 } 206 }
209 #else /* CONFIG_TRANSPARENT_HUGEPAGE */ 207 #else /* CONFIG_TRANSPARENT_HUGEPAGE */
210 /* stub to allow the code below to compile */ 208 /* stub to allow the code below to compile */
211 int hmm_vma_handle_pmd(struct mm_walk *walk, u 209 int hmm_vma_handle_pmd(struct mm_walk *walk, unsigned long addr,
212 unsigned long end, unsigned lo 210 unsigned long end, unsigned long hmm_pfns[], pmd_t pmd);
213 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */ 211 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
214 212
>> 213 static inline bool hmm_is_device_private_entry(struct hmm_range *range,
>> 214 swp_entry_t entry)
>> 215 {
>> 216 return is_device_private_entry(entry) &&
>> 217 device_private_entry_to_page(entry)->pgmap->owner ==
>> 218 range->dev_private_owner;
>> 219 }
>> 220
215 static inline unsigned long pte_to_hmm_pfn_fla 221 static inline unsigned long pte_to_hmm_pfn_flags(struct hmm_range *range,
216 222 pte_t pte)
217 { 223 {
218 if (pte_none(pte) || !pte_present(pte) 224 if (pte_none(pte) || !pte_present(pte) || pte_protnone(pte))
219 return 0; 225 return 0;
220 return pte_write(pte) ? (HMM_PFN_VALID 226 return pte_write(pte) ? (HMM_PFN_VALID | HMM_PFN_WRITE) : HMM_PFN_VALID;
221 } 227 }
222 228
223 static int hmm_vma_handle_pte(struct mm_walk * 229 static int hmm_vma_handle_pte(struct mm_walk *walk, unsigned long addr,
224 unsigned long en 230 unsigned long end, pmd_t *pmdp, pte_t *ptep,
225 unsigned long *h 231 unsigned long *hmm_pfn)
226 { 232 {
227 struct hmm_vma_walk *hmm_vma_walk = wa 233 struct hmm_vma_walk *hmm_vma_walk = walk->private;
228 struct hmm_range *range = hmm_vma_walk 234 struct hmm_range *range = hmm_vma_walk->range;
229 unsigned int required_fault; 235 unsigned int required_fault;
230 unsigned long cpu_flags; 236 unsigned long cpu_flags;
231 pte_t pte = ptep_get(ptep); !! 237 pte_t pte = *ptep;
232 uint64_t pfn_req_flags = *hmm_pfn; 238 uint64_t pfn_req_flags = *hmm_pfn;
233 239
234 if (pte_none_mostly(pte)) { !! 240 if (pte_none(pte)) {
235 required_fault = 241 required_fault =
236 hmm_pte_need_fault(hmm 242 hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, 0);
237 if (required_fault) 243 if (required_fault)
238 goto fault; 244 goto fault;
239 *hmm_pfn = 0; 245 *hmm_pfn = 0;
240 return 0; 246 return 0;
241 } 247 }
242 248
243 if (!pte_present(pte)) { 249 if (!pte_present(pte)) {
244 swp_entry_t entry = pte_to_swp 250 swp_entry_t entry = pte_to_swp_entry(pte);
245 251
246 /* 252 /*
247 * Don't fault in device priva !! 253 * Never fault in device private pages, but just report
248 * just report the PFN. !! 254 * the PFN even if not present.
249 */ 255 */
250 if (is_device_private_entry(en !! 256 if (hmm_is_device_private_entry(range, entry)) {
251 pfn_swap_entry_to_page(ent <<
252 range->dev_private_owner) <<
253 cpu_flags = HMM_PFN_VA 257 cpu_flags = HMM_PFN_VALID;
254 if (is_writable_device !! 258 if (is_write_device_private_entry(entry))
255 cpu_flags |= H 259 cpu_flags |= HMM_PFN_WRITE;
256 *hmm_pfn = swp_offset_ !! 260 *hmm_pfn = device_private_entry_to_pfn(entry) |
>> 261 cpu_flags;
257 return 0; 262 return 0;
258 } 263 }
259 264
260 required_fault = 265 required_fault =
261 hmm_pte_need_fault(hmm 266 hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, 0);
262 if (!required_fault) { 267 if (!required_fault) {
263 *hmm_pfn = 0; 268 *hmm_pfn = 0;
264 return 0; 269 return 0;
265 } 270 }
266 271
267 if (!non_swap_entry(entry)) 272 if (!non_swap_entry(entry))
268 goto fault; 273 goto fault;
269 274
270 if (is_device_private_entry(en <<
271 goto fault; <<
272 <<
273 if (is_device_exclusive_entry( <<
274 goto fault; <<
275 <<
276 if (is_migration_entry(entry)) 275 if (is_migration_entry(entry)) {
277 pte_unmap(ptep); 276 pte_unmap(ptep);
278 hmm_vma_walk->last = a 277 hmm_vma_walk->last = addr;
279 migration_entry_wait(w 278 migration_entry_wait(walk->mm, pmdp, addr);
280 return -EBUSY; 279 return -EBUSY;
281 } 280 }
282 281
283 /* Report error for everything 282 /* Report error for everything else */
284 pte_unmap(ptep); 283 pte_unmap(ptep);
285 return -EFAULT; 284 return -EFAULT;
286 } 285 }
287 286
288 cpu_flags = pte_to_hmm_pfn_flags(range 287 cpu_flags = pte_to_hmm_pfn_flags(range, pte);
289 required_fault = 288 required_fault =
290 hmm_pte_need_fault(hmm_vma_wal 289 hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, cpu_flags);
291 if (required_fault) 290 if (required_fault)
292 goto fault; 291 goto fault;
293 292
294 /* 293 /*
295 * Bypass devmap pte such as DAX page <<
296 * flags(pfn_req_flags) are fulfilled. <<
297 * Since each architecture defines a s 294 * Since each architecture defines a struct page for the zero page, just
298 * fall through and treat it like a no 295 * fall through and treat it like a normal page.
299 */ 296 */
300 if (!vm_normal_page(walk->vma, addr, p !! 297 if (pte_special(pte) && !is_zero_pfn(pte_pfn(pte))) {
301 !pte_devmap(pte) && <<
302 !is_zero_pfn(pte_pfn(pte))) { <<
303 if (hmm_pte_need_fault(hmm_vma 298 if (hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, 0)) {
304 pte_unmap(ptep); 299 pte_unmap(ptep);
305 return -EFAULT; 300 return -EFAULT;
306 } 301 }
307 *hmm_pfn = HMM_PFN_ERROR; 302 *hmm_pfn = HMM_PFN_ERROR;
308 return 0; 303 return 0;
309 } 304 }
310 305
311 *hmm_pfn = pte_pfn(pte) | cpu_flags; 306 *hmm_pfn = pte_pfn(pte) | cpu_flags;
312 return 0; 307 return 0;
313 308
314 fault: 309 fault:
315 pte_unmap(ptep); 310 pte_unmap(ptep);
316 /* Fault any virtual address we were a 311 /* Fault any virtual address we were asked to fault */
317 return hmm_vma_fault(addr, end, requir 312 return hmm_vma_fault(addr, end, required_fault, walk);
318 } 313 }
319 314
320 static int hmm_vma_walk_pmd(pmd_t *pmdp, 315 static int hmm_vma_walk_pmd(pmd_t *pmdp,
321 unsigned long star 316 unsigned long start,
322 unsigned long end, 317 unsigned long end,
323 struct mm_walk *wa 318 struct mm_walk *walk)
324 { 319 {
325 struct hmm_vma_walk *hmm_vma_walk = wa 320 struct hmm_vma_walk *hmm_vma_walk = walk->private;
326 struct hmm_range *range = hmm_vma_walk 321 struct hmm_range *range = hmm_vma_walk->range;
327 unsigned long *hmm_pfns = 322 unsigned long *hmm_pfns =
328 &range->hmm_pfns[(start - rang 323 &range->hmm_pfns[(start - range->start) >> PAGE_SHIFT];
329 unsigned long npages = (end - start) > 324 unsigned long npages = (end - start) >> PAGE_SHIFT;
330 unsigned long addr = start; 325 unsigned long addr = start;
331 pte_t *ptep; 326 pte_t *ptep;
332 pmd_t pmd; 327 pmd_t pmd;
333 328
334 again: 329 again:
335 pmd = pmdp_get_lockless(pmdp); !! 330 pmd = READ_ONCE(*pmdp);
336 if (pmd_none(pmd)) 331 if (pmd_none(pmd))
337 return hmm_vma_walk_hole(start 332 return hmm_vma_walk_hole(start, end, -1, walk);
338 333
339 if (thp_migration_supported() && is_pm 334 if (thp_migration_supported() && is_pmd_migration_entry(pmd)) {
340 if (hmm_range_need_fault(hmm_v 335 if (hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, 0)) {
341 hmm_vma_walk->last = a 336 hmm_vma_walk->last = addr;
342 pmd_migration_entry_wa 337 pmd_migration_entry_wait(walk->mm, pmdp);
343 return -EBUSY; 338 return -EBUSY;
344 } 339 }
345 return hmm_pfns_fill(start, en 340 return hmm_pfns_fill(start, end, range, 0);
346 } 341 }
347 342
348 if (!pmd_present(pmd)) { 343 if (!pmd_present(pmd)) {
349 if (hmm_range_need_fault(hmm_v 344 if (hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, 0))
350 return -EFAULT; 345 return -EFAULT;
351 return hmm_pfns_fill(start, en 346 return hmm_pfns_fill(start, end, range, HMM_PFN_ERROR);
352 } 347 }
353 348
354 if (pmd_devmap(pmd) || pmd_trans_huge( 349 if (pmd_devmap(pmd) || pmd_trans_huge(pmd)) {
355 /* 350 /*
356 * No need to take pmd_lock he 351 * No need to take pmd_lock here, even if some other thread
357 * is splitting the huge pmd w 352 * is splitting the huge pmd we will get that event through
358 * mmu_notifier callback. 353 * mmu_notifier callback.
359 * 354 *
360 * So just read pmd value and 355 * So just read pmd value and check again it's a transparent
361 * huge or device mapping one 356 * huge or device mapping one and compute corresponding pfn
362 * values. 357 * values.
363 */ 358 */
364 pmd = pmdp_get_lockless(pmdp); !! 359 pmd = pmd_read_atomic(pmdp);
>> 360 barrier();
365 if (!pmd_devmap(pmd) && !pmd_t 361 if (!pmd_devmap(pmd) && !pmd_trans_huge(pmd))
366 goto again; 362 goto again;
367 363
368 return hmm_vma_handle_pmd(walk 364 return hmm_vma_handle_pmd(walk, addr, end, hmm_pfns, pmd);
369 } 365 }
370 366
371 /* 367 /*
372 * We have handled all the valid cases 368 * We have handled all the valid cases above ie either none, migration,
373 * huge or transparent huge. At this p 369 * huge or transparent huge. At this point either it is a valid pmd
374 * entry pointing to pte directory or 370 * entry pointing to pte directory or it is a bad pmd that will not
375 * recover. 371 * recover.
376 */ 372 */
377 if (pmd_bad(pmd)) { 373 if (pmd_bad(pmd)) {
378 if (hmm_range_need_fault(hmm_v 374 if (hmm_range_need_fault(hmm_vma_walk, hmm_pfns, npages, 0))
379 return -EFAULT; 375 return -EFAULT;
380 return hmm_pfns_fill(start, en 376 return hmm_pfns_fill(start, end, range, HMM_PFN_ERROR);
381 } 377 }
382 378
383 ptep = pte_offset_map(pmdp, addr); 379 ptep = pte_offset_map(pmdp, addr);
384 if (!ptep) <<
385 goto again; <<
386 for (; addr < end; addr += PAGE_SIZE, 380 for (; addr < end; addr += PAGE_SIZE, ptep++, hmm_pfns++) {
387 int r; 381 int r;
388 382
389 r = hmm_vma_handle_pte(walk, a 383 r = hmm_vma_handle_pte(walk, addr, end, pmdp, ptep, hmm_pfns);
390 if (r) { 384 if (r) {
391 /* hmm_vma_handle_pte( 385 /* hmm_vma_handle_pte() did pte_unmap() */
392 return r; 386 return r;
393 } 387 }
394 } 388 }
395 pte_unmap(ptep - 1); 389 pte_unmap(ptep - 1);
396 return 0; 390 return 0;
397 } 391 }
398 392
399 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && \ 393 #if defined(CONFIG_ARCH_HAS_PTE_DEVMAP) && \
400 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEP 394 defined(CONFIG_HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD)
401 static inline unsigned long pud_to_hmm_pfn_fla 395 static inline unsigned long pud_to_hmm_pfn_flags(struct hmm_range *range,
402 396 pud_t pud)
403 { 397 {
404 if (!pud_present(pud)) 398 if (!pud_present(pud))
405 return 0; 399 return 0;
406 return (pud_write(pud) ? (HMM_PFN_VALI 400 return (pud_write(pud) ? (HMM_PFN_VALID | HMM_PFN_WRITE) :
407 HMM_PFN_VALID 401 HMM_PFN_VALID) |
408 hmm_pfn_flags_order(PUD_SHIFT - 402 hmm_pfn_flags_order(PUD_SHIFT - PAGE_SHIFT);
409 } 403 }
410 404
411 static int hmm_vma_walk_pud(pud_t *pudp, unsig 405 static int hmm_vma_walk_pud(pud_t *pudp, unsigned long start, unsigned long end,
412 struct mm_walk *walk) 406 struct mm_walk *walk)
413 { 407 {
414 struct hmm_vma_walk *hmm_vma_walk = wa 408 struct hmm_vma_walk *hmm_vma_walk = walk->private;
415 struct hmm_range *range = hmm_vma_walk 409 struct hmm_range *range = hmm_vma_walk->range;
416 unsigned long addr = start; 410 unsigned long addr = start;
417 pud_t pud; 411 pud_t pud;
>> 412 int ret = 0;
418 spinlock_t *ptl = pud_trans_huge_lock( 413 spinlock_t *ptl = pud_trans_huge_lock(pudp, walk->vma);
419 414
420 if (!ptl) 415 if (!ptl)
421 return 0; 416 return 0;
422 417
423 /* Normally we don't want to split the 418 /* Normally we don't want to split the huge page */
424 walk->action = ACTION_CONTINUE; 419 walk->action = ACTION_CONTINUE;
425 420
426 pud = READ_ONCE(*pudp); 421 pud = READ_ONCE(*pudp);
427 if (!pud_present(pud)) { !! 422 if (pud_none(pud)) {
428 spin_unlock(ptl); 423 spin_unlock(ptl);
429 return hmm_vma_walk_hole(start 424 return hmm_vma_walk_hole(start, end, -1, walk);
430 } 425 }
431 426
432 if (pud_leaf(pud) && pud_devmap(pud)) !! 427 if (pud_huge(pud) && pud_devmap(pud)) {
433 unsigned long i, npages, pfn; 428 unsigned long i, npages, pfn;
434 unsigned int required_fault; 429 unsigned int required_fault;
435 unsigned long *hmm_pfns; 430 unsigned long *hmm_pfns;
436 unsigned long cpu_flags; 431 unsigned long cpu_flags;
437 432
>> 433 if (!pud_present(pud)) {
>> 434 spin_unlock(ptl);
>> 435 return hmm_vma_walk_hole(start, end, -1, walk);
>> 436 }
>> 437
438 i = (addr - range->start) >> P 438 i = (addr - range->start) >> PAGE_SHIFT;
439 npages = (end - addr) >> PAGE_ 439 npages = (end - addr) >> PAGE_SHIFT;
440 hmm_pfns = &range->hmm_pfns[i] 440 hmm_pfns = &range->hmm_pfns[i];
441 441
442 cpu_flags = pud_to_hmm_pfn_fla 442 cpu_flags = pud_to_hmm_pfn_flags(range, pud);
443 required_fault = hmm_range_nee 443 required_fault = hmm_range_need_fault(hmm_vma_walk, hmm_pfns,
444 444 npages, cpu_flags);
445 if (required_fault) { 445 if (required_fault) {
446 spin_unlock(ptl); 446 spin_unlock(ptl);
447 return hmm_vma_fault(a 447 return hmm_vma_fault(addr, end, required_fault, walk);
448 } 448 }
449 449
450 pfn = pud_pfn(pud) + ((addr & 450 pfn = pud_pfn(pud) + ((addr & ~PUD_MASK) >> PAGE_SHIFT);
451 for (i = 0; i < npages; ++i, + 451 for (i = 0; i < npages; ++i, ++pfn)
452 hmm_pfns[i] = pfn | cp 452 hmm_pfns[i] = pfn | cpu_flags;
453 goto out_unlock; 453 goto out_unlock;
454 } 454 }
455 455
456 /* Ask for the PUD to be split */ 456 /* Ask for the PUD to be split */
457 walk->action = ACTION_SUBTREE; 457 walk->action = ACTION_SUBTREE;
458 458
459 out_unlock: 459 out_unlock:
460 spin_unlock(ptl); 460 spin_unlock(ptl);
461 return 0; !! 461 return ret;
462 } 462 }
463 #else 463 #else
464 #define hmm_vma_walk_pud NULL 464 #define hmm_vma_walk_pud NULL
465 #endif 465 #endif
466 466
467 #ifdef CONFIG_HUGETLB_PAGE 467 #ifdef CONFIG_HUGETLB_PAGE
468 static int hmm_vma_walk_hugetlb_entry(pte_t *p 468 static int hmm_vma_walk_hugetlb_entry(pte_t *pte, unsigned long hmask,
469 unsigned 469 unsigned long start, unsigned long end,
470 struct m 470 struct mm_walk *walk)
471 { 471 {
472 unsigned long addr = start, i, pfn; 472 unsigned long addr = start, i, pfn;
473 struct hmm_vma_walk *hmm_vma_walk = wa 473 struct hmm_vma_walk *hmm_vma_walk = walk->private;
474 struct hmm_range *range = hmm_vma_walk 474 struct hmm_range *range = hmm_vma_walk->range;
475 struct vm_area_struct *vma = walk->vma 475 struct vm_area_struct *vma = walk->vma;
476 unsigned int required_fault; 476 unsigned int required_fault;
477 unsigned long pfn_req_flags; 477 unsigned long pfn_req_flags;
478 unsigned long cpu_flags; 478 unsigned long cpu_flags;
479 spinlock_t *ptl; 479 spinlock_t *ptl;
480 pte_t entry; 480 pte_t entry;
481 481
482 ptl = huge_pte_lock(hstate_vma(vma), w 482 ptl = huge_pte_lock(hstate_vma(vma), walk->mm, pte);
483 entry = huge_ptep_get(walk->mm, addr, !! 483 entry = huge_ptep_get(pte);
484 484
485 i = (start - range->start) >> PAGE_SHI 485 i = (start - range->start) >> PAGE_SHIFT;
486 pfn_req_flags = range->hmm_pfns[i]; 486 pfn_req_flags = range->hmm_pfns[i];
487 cpu_flags = pte_to_hmm_pfn_flags(range 487 cpu_flags = pte_to_hmm_pfn_flags(range, entry) |
488 hmm_pfn_flags_order(huge_p 488 hmm_pfn_flags_order(huge_page_order(hstate_vma(vma)));
489 required_fault = 489 required_fault =
490 hmm_pte_need_fault(hmm_vma_wal 490 hmm_pte_need_fault(hmm_vma_walk, pfn_req_flags, cpu_flags);
491 if (required_fault) { 491 if (required_fault) {
492 int ret; <<
493 <<
494 spin_unlock(ptl); 492 spin_unlock(ptl);
495 hugetlb_vma_unlock_read(vma); !! 493 return hmm_vma_fault(addr, end, required_fault, walk);
496 /* <<
497 * Avoid deadlock: drop the vm <<
498 * hmm_vma_fault(), which will <<
499 * drop the vma lock. This is <<
500 * protection point of view, b <<
501 * use here of either pte or p <<
502 * lock. <<
503 */ <<
504 ret = hmm_vma_fault(addr, end, <<
505 hugetlb_vma_lock_read(vma); <<
506 return ret; <<
507 } 494 }
508 495
509 pfn = pte_pfn(entry) + ((start & ~hmas 496 pfn = pte_pfn(entry) + ((start & ~hmask) >> PAGE_SHIFT);
510 for (; addr < end; addr += PAGE_SIZE, 497 for (; addr < end; addr += PAGE_SIZE, i++, pfn++)
511 range->hmm_pfns[i] = pfn | cpu 498 range->hmm_pfns[i] = pfn | cpu_flags;
512 499
513 spin_unlock(ptl); 500 spin_unlock(ptl);
514 return 0; 501 return 0;
515 } 502 }
516 #else 503 #else
517 #define hmm_vma_walk_hugetlb_entry NULL 504 #define hmm_vma_walk_hugetlb_entry NULL
518 #endif /* CONFIG_HUGETLB_PAGE */ 505 #endif /* CONFIG_HUGETLB_PAGE */
519 506
520 static int hmm_vma_walk_test(unsigned long sta 507 static int hmm_vma_walk_test(unsigned long start, unsigned long end,
521 struct mm_walk *w 508 struct mm_walk *walk)
522 { 509 {
523 struct hmm_vma_walk *hmm_vma_walk = wa 510 struct hmm_vma_walk *hmm_vma_walk = walk->private;
524 struct hmm_range *range = hmm_vma_walk 511 struct hmm_range *range = hmm_vma_walk->range;
525 struct vm_area_struct *vma = walk->vma 512 struct vm_area_struct *vma = walk->vma;
526 513
527 if (!(vma->vm_flags & (VM_IO | VM_PFNM !! 514 if (!(vma->vm_flags & (VM_IO | VM_PFNMAP | VM_MIXEDMAP)) &&
528 vma->vm_flags & VM_READ) 515 vma->vm_flags & VM_READ)
529 return 0; 516 return 0;
530 517
531 /* 518 /*
532 * vma ranges that don't have struct p 519 * vma ranges that don't have struct page backing them or map I/O
533 * devices directly cannot be handled 520 * devices directly cannot be handled by hmm_range_fault().
534 * 521 *
535 * If the vma does not allow read acce 522 * If the vma does not allow read access, then assume that it does not
536 * allow write access either. HMM does 523 * allow write access either. HMM does not support architectures that
537 * allow write without read. 524 * allow write without read.
538 * 525 *
539 * If a fault is requested for an unsu 526 * If a fault is requested for an unsupported range then it is a hard
540 * failure. 527 * failure.
541 */ 528 */
542 if (hmm_range_need_fault(hmm_vma_walk, 529 if (hmm_range_need_fault(hmm_vma_walk,
543 range->hmm_pf 530 range->hmm_pfns +
544 ((sta 531 ((start - range->start) >> PAGE_SHIFT),
545 (end - start) 532 (end - start) >> PAGE_SHIFT, 0))
546 return -EFAULT; 533 return -EFAULT;
547 534
548 hmm_pfns_fill(start, end, range, HMM_P 535 hmm_pfns_fill(start, end, range, HMM_PFN_ERROR);
549 536
550 /* Skip this vma and continue processi 537 /* Skip this vma and continue processing the next vma. */
551 return 1; 538 return 1;
552 } 539 }
553 540
554 static const struct mm_walk_ops hmm_walk_ops = 541 static const struct mm_walk_ops hmm_walk_ops = {
555 .pud_entry = hmm_vma_walk_pud, 542 .pud_entry = hmm_vma_walk_pud,
556 .pmd_entry = hmm_vma_walk_pmd, 543 .pmd_entry = hmm_vma_walk_pmd,
557 .pte_hole = hmm_vma_walk_hole, 544 .pte_hole = hmm_vma_walk_hole,
558 .hugetlb_entry = hmm_vma_walk_hugetlb 545 .hugetlb_entry = hmm_vma_walk_hugetlb_entry,
559 .test_walk = hmm_vma_walk_test, 546 .test_walk = hmm_vma_walk_test,
560 .walk_lock = PGWALK_RDLOCK, <<
561 }; 547 };
562 548
563 /** 549 /**
564 * hmm_range_fault - try to fault some address 550 * hmm_range_fault - try to fault some address in a virtual address range
565 * @range: argument structure 551 * @range: argument structure
566 * 552 *
567 * Returns 0 on success or one of the followin 553 * Returns 0 on success or one of the following error codes:
568 * 554 *
569 * -EINVAL: Invalid arguments or mm or vir 555 * -EINVAL: Invalid arguments or mm or virtual address is in an invalid vma
570 * (e.g., device file vma). 556 * (e.g., device file vma).
571 * -ENOMEM: Out of memory. 557 * -ENOMEM: Out of memory.
572 * -EPERM: Invalid permission (e.g., aski 558 * -EPERM: Invalid permission (e.g., asking for write and range is read
573 * only). 559 * only).
574 * -EBUSY: The range has been invalidated 560 * -EBUSY: The range has been invalidated and the caller needs to wait for
575 * the invalidation to finish. 561 * the invalidation to finish.
576 * -EFAULT: A page was requested to be val 562 * -EFAULT: A page was requested to be valid and could not be made valid
577 * ie it has no backing VMA or it 563 * ie it has no backing VMA or it is illegal to access
578 * 564 *
579 * This is similar to get_user_pages(), except 565 * This is similar to get_user_pages(), except that it can read the page tables
580 * without mutating them (ie causing faults). 566 * without mutating them (ie causing faults).
581 */ 567 */
582 int hmm_range_fault(struct hmm_range *range) 568 int hmm_range_fault(struct hmm_range *range)
583 { 569 {
584 struct hmm_vma_walk hmm_vma_walk = { 570 struct hmm_vma_walk hmm_vma_walk = {
585 .range = range, 571 .range = range,
586 .last = range->start, 572 .last = range->start,
587 }; 573 };
588 struct mm_struct *mm = range->notifier 574 struct mm_struct *mm = range->notifier->mm;
589 int ret; 575 int ret;
590 576
591 mmap_assert_locked(mm); 577 mmap_assert_locked(mm);
592 578
593 do { 579 do {
594 /* If range is no longer valid 580 /* If range is no longer valid force retry. */
595 if (mmu_interval_check_retry(r 581 if (mmu_interval_check_retry(range->notifier,
596 r 582 range->notifier_seq))
597 return -EBUSY; 583 return -EBUSY;
598 ret = walk_page_range(mm, hmm_ 584 ret = walk_page_range(mm, hmm_vma_walk.last, range->end,
599 &hmm_wal 585 &hmm_walk_ops, &hmm_vma_walk);
600 /* 586 /*
601 * When -EBUSY is returned the 587 * When -EBUSY is returned the loop restarts with
602 * hmm_vma_walk.last set to an 588 * hmm_vma_walk.last set to an address that has not been stored
603 * in pfns. All entries < last 589 * in pfns. All entries < last in the pfn array are set to their
604 * output, and all >= are stil 590 * output, and all >= are still at their input values.
605 */ 591 */
606 } while (ret == -EBUSY); 592 } while (ret == -EBUSY);
607 return ret; 593 return ret;
608 } 594 }
609 EXPORT_SYMBOL(hmm_range_fault); 595 EXPORT_SYMBOL(hmm_range_fault);
610 596
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.