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