~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/arch/powerpc/mm/hugetlbpage.c

Version: ~ [ linux-5.3-rc4 ] ~ [ linux-5.2.8 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.66 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.138 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.189 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.189 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.71 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-3.9.11 ] ~ [ linux-3.8.13 ] ~ [ linux-3.7.10 ] ~ [ linux-3.6.11 ] ~ [ linux-3.5.7 ] ~ [ linux-3.4.113 ] ~ [ linux-3.3.8 ] ~ [ linux-3.2.102 ] ~ [ linux-3.1.10 ] ~ [ linux-3.0.101 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /*
  2  * PPC Huge TLB Page Support for Kernel.
  3  *
  4  * Copyright (C) 2003 David Gibson, IBM Corporation.
  5  * Copyright (C) 2011 Becky Bruce, Freescale Semiconductor
  6  *
  7  * Based on the IA-32 version:
  8  * Copyright (C) 2002, Rohit Seth <rohit.seth@intel.com>
  9  */
 10 
 11 #include <linux/mm.h>
 12 #include <linux/io.h>
 13 #include <linux/slab.h>
 14 #include <linux/hugetlb.h>
 15 #include <linux/export.h>
 16 #include <linux/of_fdt.h>
 17 #include <linux/memblock.h>
 18 #include <linux/bootmem.h>
 19 #include <linux/moduleparam.h>
 20 #include <asm/pgtable.h>
 21 #include <asm/pgalloc.h>
 22 #include <asm/tlb.h>
 23 #include <asm/setup.h>
 24 #include <asm/hugetlb.h>
 25 
 26 #ifdef CONFIG_HUGETLB_PAGE
 27 
 28 #define PAGE_SHIFT_64K  16
 29 #define PAGE_SHIFT_16M  24
 30 #define PAGE_SHIFT_16G  34
 31 
 32 unsigned int HPAGE_SHIFT;
 33 
 34 /*
 35  * Tracks gpages after the device tree is scanned and before the
 36  * huge_boot_pages list is ready.  On non-Freescale implementations, this is
 37  * just used to track 16G pages and so is a single array.  FSL-based
 38  * implementations may have more than one gpage size, so we need multiple
 39  * arrays
 40  */
 41 #ifdef CONFIG_PPC_FSL_BOOK3E
 42 #define MAX_NUMBER_GPAGES       128
 43 struct psize_gpages {
 44         u64 gpage_list[MAX_NUMBER_GPAGES];
 45         unsigned int nr_gpages;
 46 };
 47 static struct psize_gpages gpage_freearray[MMU_PAGE_COUNT];
 48 #else
 49 #define MAX_NUMBER_GPAGES       1024
 50 static u64 gpage_freearray[MAX_NUMBER_GPAGES];
 51 static unsigned nr_gpages;
 52 #endif
 53 
 54 #define hugepd_none(hpd)        ((hpd).pd == 0)
 55 
 56 #ifdef CONFIG_PPC_BOOK3S_64
 57 /*
 58  * At this point we do the placement change only for BOOK3S 64. This would
 59  * possibly work on other subarchs.
 60  */
 61 
 62 /*
 63  * We have PGD_INDEX_SIZ = 12 and PTE_INDEX_SIZE = 8, so that we can have
 64  * 16GB hugepage pte in PGD and 16MB hugepage pte at PMD;
 65  *
 66  * Defined in such a way that we can optimize away code block at build time
 67  * if CONFIG_HUGETLB_PAGE=n.
 68  */
 69 int pmd_huge(pmd_t pmd)
 70 {
 71         /*
 72          * leaf pte for huge page, bottom two bits != 00
 73          */
 74         return ((pmd_val(pmd) & 0x3) != 0x0);
 75 }
 76 
 77 int pud_huge(pud_t pud)
 78 {
 79         /*
 80          * leaf pte for huge page, bottom two bits != 00
 81          */
 82         return ((pud_val(pud) & 0x3) != 0x0);
 83 }
 84 
 85 int pgd_huge(pgd_t pgd)
 86 {
 87         /*
 88          * leaf pte for huge page, bottom two bits != 00
 89          */
 90         return ((pgd_val(pgd) & 0x3) != 0x0);
 91 }
 92 #else
 93 int pmd_huge(pmd_t pmd)
 94 {
 95         return 0;
 96 }
 97 
 98 int pud_huge(pud_t pud)
 99 {
100         return 0;
101 }
102 
103 int pgd_huge(pgd_t pgd)
104 {
105         return 0;
106 }
107 #endif
108 
109 pte_t *huge_pte_offset(struct mm_struct *mm, unsigned long addr)
110 {
111         /* Only called for hugetlbfs pages, hence can ignore THP */
112         return __find_linux_pte_or_hugepte(mm->pgd, addr, NULL);
113 }
114 
115 static int __hugepte_alloc(struct mm_struct *mm, hugepd_t *hpdp,
116                            unsigned long address, unsigned pdshift, unsigned pshift)
117 {
118         struct kmem_cache *cachep;
119         pte_t *new;
120 
121 #ifdef CONFIG_PPC_FSL_BOOK3E
122         int i;
123         int num_hugepd = 1 << (pshift - pdshift);
124         cachep = hugepte_cache;
125 #else
126         cachep = PGT_CACHE(pdshift - pshift);
127 #endif
128 
129         new = kmem_cache_zalloc(cachep, GFP_KERNEL|__GFP_REPEAT);
130 
131         BUG_ON(pshift > HUGEPD_SHIFT_MASK);
132         BUG_ON((unsigned long)new & HUGEPD_SHIFT_MASK);
133 
134         if (! new)
135                 return -ENOMEM;
136 
137         spin_lock(&mm->page_table_lock);
138 #ifdef CONFIG_PPC_FSL_BOOK3E
139         /*
140          * We have multiple higher-level entries that point to the same
141          * actual pte location.  Fill in each as we go and backtrack on error.
142          * We need all of these so the DTLB pgtable walk code can find the
143          * right higher-level entry without knowing if it's a hugepage or not.
144          */
145         for (i = 0; i < num_hugepd; i++, hpdp++) {
146                 if (unlikely(!hugepd_none(*hpdp)))
147                         break;
148                 else
149                         /* We use the old format for PPC_FSL_BOOK3E */
150                         hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift;
151         }
152         /* If we bailed from the for loop early, an error occurred, clean up */
153         if (i < num_hugepd) {
154                 for (i = i - 1 ; i >= 0; i--, hpdp--)
155                         hpdp->pd = 0;
156                 kmem_cache_free(cachep, new);
157         }
158 #else
159         if (!hugepd_none(*hpdp))
160                 kmem_cache_free(cachep, new);
161         else {
162 #ifdef CONFIG_PPC_BOOK3S_64
163                 hpdp->pd = (unsigned long)new |
164                             (shift_to_mmu_psize(pshift) << 2);
165 #else
166                 hpdp->pd = ((unsigned long)new & ~PD_HUGE) | pshift;
167 #endif
168         }
169 #endif
170         spin_unlock(&mm->page_table_lock);
171         return 0;
172 }
173 
174 /*
175  * These macros define how to determine which level of the page table holds
176  * the hpdp.
177  */
178 #ifdef CONFIG_PPC_FSL_BOOK3E
179 #define HUGEPD_PGD_SHIFT PGDIR_SHIFT
180 #define HUGEPD_PUD_SHIFT PUD_SHIFT
181 #else
182 #define HUGEPD_PGD_SHIFT PUD_SHIFT
183 #define HUGEPD_PUD_SHIFT PMD_SHIFT
184 #endif
185 
186 #ifdef CONFIG_PPC_BOOK3S_64
187 /*
188  * At this point we do the placement change only for BOOK3S 64. This would
189  * possibly work on other subarchs.
190  */
191 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
192 {
193         pgd_t *pg;
194         pud_t *pu;
195         pmd_t *pm;
196         hugepd_t *hpdp = NULL;
197         unsigned pshift = __ffs(sz);
198         unsigned pdshift = PGDIR_SHIFT;
199 
200         addr &= ~(sz-1);
201         pg = pgd_offset(mm, addr);
202 
203         if (pshift == PGDIR_SHIFT)
204                 /* 16GB huge page */
205                 return (pte_t *) pg;
206         else if (pshift > PUD_SHIFT)
207                 /*
208                  * We need to use hugepd table
209                  */
210                 hpdp = (hugepd_t *)pg;
211         else {
212                 pdshift = PUD_SHIFT;
213                 pu = pud_alloc(mm, pg, addr);
214                 if (pshift == PUD_SHIFT)
215                         return (pte_t *)pu;
216                 else if (pshift > PMD_SHIFT)
217                         hpdp = (hugepd_t *)pu;
218                 else {
219                         pdshift = PMD_SHIFT;
220                         pm = pmd_alloc(mm, pu, addr);
221                         if (pshift == PMD_SHIFT)
222                                 /* 16MB hugepage */
223                                 return (pte_t *)pm;
224                         else
225                                 hpdp = (hugepd_t *)pm;
226                 }
227         }
228         if (!hpdp)
229                 return NULL;
230 
231         BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
232 
233         if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
234                 return NULL;
235 
236         return hugepte_offset(*hpdp, addr, pdshift);
237 }
238 
239 #else
240 
241 pte_t *huge_pte_alloc(struct mm_struct *mm, unsigned long addr, unsigned long sz)
242 {
243         pgd_t *pg;
244         pud_t *pu;
245         pmd_t *pm;
246         hugepd_t *hpdp = NULL;
247         unsigned pshift = __ffs(sz);
248         unsigned pdshift = PGDIR_SHIFT;
249 
250         addr &= ~(sz-1);
251 
252         pg = pgd_offset(mm, addr);
253 
254         if (pshift >= HUGEPD_PGD_SHIFT) {
255                 hpdp = (hugepd_t *)pg;
256         } else {
257                 pdshift = PUD_SHIFT;
258                 pu = pud_alloc(mm, pg, addr);
259                 if (pshift >= HUGEPD_PUD_SHIFT) {
260                         hpdp = (hugepd_t *)pu;
261                 } else {
262                         pdshift = PMD_SHIFT;
263                         pm = pmd_alloc(mm, pu, addr);
264                         hpdp = (hugepd_t *)pm;
265                 }
266         }
267 
268         if (!hpdp)
269                 return NULL;
270 
271         BUG_ON(!hugepd_none(*hpdp) && !hugepd_ok(*hpdp));
272 
273         if (hugepd_none(*hpdp) && __hugepte_alloc(mm, hpdp, addr, pdshift, pshift))
274                 return NULL;
275 
276         return hugepte_offset(*hpdp, addr, pdshift);
277 }
278 #endif
279 
280 #ifdef CONFIG_PPC_FSL_BOOK3E
281 /* Build list of addresses of gigantic pages.  This function is used in early
282  * boot before the buddy allocator is setup.
283  */
284 void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
285 {
286         unsigned int idx = shift_to_mmu_psize(__ffs(page_size));
287         int i;
288 
289         if (addr == 0)
290                 return;
291 
292         gpage_freearray[idx].nr_gpages = number_of_pages;
293 
294         for (i = 0; i < number_of_pages; i++) {
295                 gpage_freearray[idx].gpage_list[i] = addr;
296                 addr += page_size;
297         }
298 }
299 
300 /*
301  * Moves the gigantic page addresses from the temporary list to the
302  * huge_boot_pages list.
303  */
304 int alloc_bootmem_huge_page(struct hstate *hstate)
305 {
306         struct huge_bootmem_page *m;
307         int idx = shift_to_mmu_psize(huge_page_shift(hstate));
308         int nr_gpages = gpage_freearray[idx].nr_gpages;
309 
310         if (nr_gpages == 0)
311                 return 0;
312 
313 #ifdef CONFIG_HIGHMEM
314         /*
315          * If gpages can be in highmem we can't use the trick of storing the
316          * data structure in the page; allocate space for this
317          */
318         m = memblock_virt_alloc(sizeof(struct huge_bootmem_page), 0);
319         m->phys = gpage_freearray[idx].gpage_list[--nr_gpages];
320 #else
321         m = phys_to_virt(gpage_freearray[idx].gpage_list[--nr_gpages]);
322 #endif
323 
324         list_add(&m->list, &huge_boot_pages);
325         gpage_freearray[idx].nr_gpages = nr_gpages;
326         gpage_freearray[idx].gpage_list[nr_gpages] = 0;
327         m->hstate = hstate;
328 
329         return 1;
330 }
331 /*
332  * Scan the command line hugepagesz= options for gigantic pages; store those in
333  * a list that we use to allocate the memory once all options are parsed.
334  */
335 
336 unsigned long gpage_npages[MMU_PAGE_COUNT];
337 
338 static int __init do_gpage_early_setup(char *param, char *val,
339                                        const char *unused, void *arg)
340 {
341         static phys_addr_t size;
342         unsigned long npages;
343 
344         /*
345          * The hugepagesz and hugepages cmdline options are interleaved.  We
346          * use the size variable to keep track of whether or not this was done
347          * properly and skip over instances where it is incorrect.  Other
348          * command-line parsing code will issue warnings, so we don't need to.
349          *
350          */
351         if ((strcmp(param, "default_hugepagesz") == 0) ||
352             (strcmp(param, "hugepagesz") == 0)) {
353                 size = memparse(val, NULL);
354         } else if (strcmp(param, "hugepages") == 0) {
355                 if (size != 0) {
356                         if (sscanf(val, "%lu", &npages) <= 0)
357                                 npages = 0;
358                         if (npages > MAX_NUMBER_GPAGES) {
359                                 pr_warn("MMU: %lu pages requested for page "
360                                         "size %llu KB, limiting to "
361                                         __stringify(MAX_NUMBER_GPAGES) "\n",
362                                         npages, size / 1024);
363                                 npages = MAX_NUMBER_GPAGES;
364                         }
365                         gpage_npages[shift_to_mmu_psize(__ffs(size))] = npages;
366                         size = 0;
367                 }
368         }
369         return 0;
370 }
371 
372 
373 /*
374  * This function allocates physical space for pages that are larger than the
375  * buddy allocator can handle.  We want to allocate these in highmem because
376  * the amount of lowmem is limited.  This means that this function MUST be
377  * called before lowmem_end_addr is set up in MMU_init() in order for the lmb
378  * allocate to grab highmem.
379  */
380 void __init reserve_hugetlb_gpages(void)
381 {
382         static __initdata char cmdline[COMMAND_LINE_SIZE];
383         phys_addr_t size, base;
384         int i;
385 
386         strlcpy(cmdline, boot_command_line, COMMAND_LINE_SIZE);
387         parse_args("hugetlb gpages", cmdline, NULL, 0, 0, 0,
388                         NULL, &do_gpage_early_setup);
389 
390         /*
391          * Walk gpage list in reverse, allocating larger page sizes first.
392          * Skip over unsupported sizes, or sizes that have 0 gpages allocated.
393          * When we reach the point in the list where pages are no longer
394          * considered gpages, we're done.
395          */
396         for (i = MMU_PAGE_COUNT-1; i >= 0; i--) {
397                 if (mmu_psize_defs[i].shift == 0 || gpage_npages[i] == 0)
398                         continue;
399                 else if (mmu_psize_to_shift(i) < (MAX_ORDER + PAGE_SHIFT))
400                         break;
401 
402                 size = (phys_addr_t)(1ULL << mmu_psize_to_shift(i));
403                 base = memblock_alloc_base(size * gpage_npages[i], size,
404                                            MEMBLOCK_ALLOC_ANYWHERE);
405                 add_gpage(base, size, gpage_npages[i]);
406         }
407 }
408 
409 #else /* !PPC_FSL_BOOK3E */
410 
411 /* Build list of addresses of gigantic pages.  This function is used in early
412  * boot before the buddy allocator is setup.
413  */
414 void add_gpage(u64 addr, u64 page_size, unsigned long number_of_pages)
415 {
416         if (!addr)
417                 return;
418         while (number_of_pages > 0) {
419                 gpage_freearray[nr_gpages] = addr;
420                 nr_gpages++;
421                 number_of_pages--;
422                 addr += page_size;
423         }
424 }
425 
426 /* Moves the gigantic page addresses from the temporary list to the
427  * huge_boot_pages list.
428  */
429 int alloc_bootmem_huge_page(struct hstate *hstate)
430 {
431         struct huge_bootmem_page *m;
432         if (nr_gpages == 0)
433                 return 0;
434         m = phys_to_virt(gpage_freearray[--nr_gpages]);
435         gpage_freearray[nr_gpages] = 0;
436         list_add(&m->list, &huge_boot_pages);
437         m->hstate = hstate;
438         return 1;
439 }
440 #endif
441 
442 #ifdef CONFIG_PPC_FSL_BOOK3E
443 #define HUGEPD_FREELIST_SIZE \
444         ((PAGE_SIZE - sizeof(struct hugepd_freelist)) / sizeof(pte_t))
445 
446 struct hugepd_freelist {
447         struct rcu_head rcu;
448         unsigned int index;
449         void *ptes[0];
450 };
451 
452 static DEFINE_PER_CPU(struct hugepd_freelist *, hugepd_freelist_cur);
453 
454 static void hugepd_free_rcu_callback(struct rcu_head *head)
455 {
456         struct hugepd_freelist *batch =
457                 container_of(head, struct hugepd_freelist, rcu);
458         unsigned int i;
459 
460         for (i = 0; i < batch->index; i++)
461                 kmem_cache_free(hugepte_cache, batch->ptes[i]);
462 
463         free_page((unsigned long)batch);
464 }
465 
466 static void hugepd_free(struct mmu_gather *tlb, void *hugepte)
467 {
468         struct hugepd_freelist **batchp;
469 
470         batchp = this_cpu_ptr(&hugepd_freelist_cur);
471 
472         if (atomic_read(&tlb->mm->mm_users) < 2 ||
473             cpumask_equal(mm_cpumask(tlb->mm),
474                           cpumask_of(smp_processor_id()))) {
475                 kmem_cache_free(hugepte_cache, hugepte);
476         put_cpu_var(hugepd_freelist_cur);
477                 return;
478         }
479 
480         if (*batchp == NULL) {
481                 *batchp = (struct hugepd_freelist *)__get_free_page(GFP_ATOMIC);
482                 (*batchp)->index = 0;
483         }
484 
485         (*batchp)->ptes[(*batchp)->index++] = hugepte;
486         if ((*batchp)->index == HUGEPD_FREELIST_SIZE) {
487                 call_rcu_sched(&(*batchp)->rcu, hugepd_free_rcu_callback);
488                 *batchp = NULL;
489         }
490         put_cpu_var(hugepd_freelist_cur);
491 }
492 #endif
493 
494 static void free_hugepd_range(struct mmu_gather *tlb, hugepd_t *hpdp, int pdshift,
495                               unsigned long start, unsigned long end,
496                               unsigned long floor, unsigned long ceiling)
497 {
498         pte_t *hugepte = hugepd_page(*hpdp);
499         int i;
500 
501         unsigned long pdmask = ~((1UL << pdshift) - 1);
502         unsigned int num_hugepd = 1;
503 
504 #ifdef CONFIG_PPC_FSL_BOOK3E
505         /* Note: On fsl the hpdp may be the first of several */
506         num_hugepd = (1 << (hugepd_shift(*hpdp) - pdshift));
507 #else
508         unsigned int shift = hugepd_shift(*hpdp);
509 #endif
510 
511         start &= pdmask;
512         if (start < floor)
513                 return;
514         if (ceiling) {
515                 ceiling &= pdmask;
516                 if (! ceiling)
517                         return;
518         }
519         if (end - 1 > ceiling - 1)
520                 return;
521 
522         for (i = 0; i < num_hugepd; i++, hpdp++)
523                 hpdp->pd = 0;
524 
525 #ifdef CONFIG_PPC_FSL_BOOK3E
526         hugepd_free(tlb, hugepte);
527 #else
528         pgtable_free_tlb(tlb, hugepte, pdshift - shift);
529 #endif
530 }
531 
532 static void hugetlb_free_pmd_range(struct mmu_gather *tlb, pud_t *pud,
533                                    unsigned long addr, unsigned long end,
534                                    unsigned long floor, unsigned long ceiling)
535 {
536         pmd_t *pmd;
537         unsigned long next;
538         unsigned long start;
539 
540         start = addr;
541         do {
542                 pmd = pmd_offset(pud, addr);
543                 next = pmd_addr_end(addr, end);
544                 if (!is_hugepd(__hugepd(pmd_val(*pmd)))) {
545                         /*
546                          * if it is not hugepd pointer, we should already find
547                          * it cleared.
548                          */
549                         WARN_ON(!pmd_none_or_clear_bad(pmd));
550                         continue;
551                 }
552 #ifdef CONFIG_PPC_FSL_BOOK3E
553                 /*
554                  * Increment next by the size of the huge mapping since
555                  * there may be more than one entry at this level for a
556                  * single hugepage, but all of them point to
557                  * the same kmem cache that holds the hugepte.
558                  */
559                 next = addr + (1 << hugepd_shift(*(hugepd_t *)pmd));
560 #endif
561                 free_hugepd_range(tlb, (hugepd_t *)pmd, PMD_SHIFT,
562                                   addr, next, floor, ceiling);
563         } while (addr = next, addr != end);
564 
565         start &= PUD_MASK;
566         if (start < floor)
567                 return;
568         if (ceiling) {
569                 ceiling &= PUD_MASK;
570                 if (!ceiling)
571                         return;
572         }
573         if (end - 1 > ceiling - 1)
574                 return;
575 
576         pmd = pmd_offset(pud, start);
577         pud_clear(pud);
578         pmd_free_tlb(tlb, pmd, start);
579         mm_dec_nr_pmds(tlb->mm);
580 }
581 
582 static void hugetlb_free_pud_range(struct mmu_gather *tlb, pgd_t *pgd,
583                                    unsigned long addr, unsigned long end,
584                                    unsigned long floor, unsigned long ceiling)
585 {
586         pud_t *pud;
587         unsigned long next;
588         unsigned long start;
589 
590         start = addr;
591         do {
592                 pud = pud_offset(pgd, addr);
593                 next = pud_addr_end(addr, end);
594                 if (!is_hugepd(__hugepd(pud_val(*pud)))) {
595                         if (pud_none_or_clear_bad(pud))
596                                 continue;
597                         hugetlb_free_pmd_range(tlb, pud, addr, next, floor,
598                                                ceiling);
599                 } else {
600 #ifdef CONFIG_PPC_FSL_BOOK3E
601                         /*
602                          * Increment next by the size of the huge mapping since
603                          * there may be more than one entry at this level for a
604                          * single hugepage, but all of them point to
605                          * the same kmem cache that holds the hugepte.
606                          */
607                         next = addr + (1 << hugepd_shift(*(hugepd_t *)pud));
608 #endif
609                         free_hugepd_range(tlb, (hugepd_t *)pud, PUD_SHIFT,
610                                           addr, next, floor, ceiling);
611                 }
612         } while (addr = next, addr != end);
613 
614         start &= PGDIR_MASK;
615         if (start < floor)
616                 return;
617         if (ceiling) {
618                 ceiling &= PGDIR_MASK;
619                 if (!ceiling)
620                         return;
621         }
622         if (end - 1 > ceiling - 1)
623                 return;
624 
625         pud = pud_offset(pgd, start);
626         pgd_clear(pgd);
627         pud_free_tlb(tlb, pud, start);
628 }
629 
630 /*
631  * This function frees user-level page tables of a process.
632  */
633 void hugetlb_free_pgd_range(struct mmu_gather *tlb,
634                             unsigned long addr, unsigned long end,
635                             unsigned long floor, unsigned long ceiling)
636 {
637         pgd_t *pgd;
638         unsigned long next;
639 
640         /*
641          * Because there are a number of different possible pagetable
642          * layouts for hugepage ranges, we limit knowledge of how
643          * things should be laid out to the allocation path
644          * (huge_pte_alloc(), above).  Everything else works out the
645          * structure as it goes from information in the hugepd
646          * pointers.  That means that we can't here use the
647          * optimization used in the normal page free_pgd_range(), of
648          * checking whether we're actually covering a large enough
649          * range to have to do anything at the top level of the walk
650          * instead of at the bottom.
651          *
652          * To make sense of this, you should probably go read the big
653          * block comment at the top of the normal free_pgd_range(),
654          * too.
655          */
656 
657         do {
658                 next = pgd_addr_end(addr, end);
659                 pgd = pgd_offset(tlb->mm, addr);
660                 if (!is_hugepd(__hugepd(pgd_val(*pgd)))) {
661                         if (pgd_none_or_clear_bad(pgd))
662                                 continue;
663                         hugetlb_free_pud_range(tlb, pgd, addr, next, floor, ceiling);
664                 } else {
665 #ifdef CONFIG_PPC_FSL_BOOK3E
666                         /*
667                          * Increment next by the size of the huge mapping since
668                          * there may be more than one entry at the pgd level
669                          * for a single hugepage, but all of them point to the
670                          * same kmem cache that holds the hugepte.
671                          */
672                         next = addr + (1 << hugepd_shift(*(hugepd_t *)pgd));
673 #endif
674                         free_hugepd_range(tlb, (hugepd_t *)pgd, PGDIR_SHIFT,
675                                           addr, next, floor, ceiling);
676                 }
677         } while (addr = next, addr != end);
678 }
679 
680 /*
681  * We are holding mmap_sem, so a parallel huge page collapse cannot run.
682  * To prevent hugepage split, disable irq.
683  */
684 struct page *
685 follow_huge_addr(struct mm_struct *mm, unsigned long address, int write)
686 {
687         pte_t *ptep, pte;
688         unsigned shift;
689         unsigned long mask, flags;
690         struct page *page = ERR_PTR(-EINVAL);
691 
692         local_irq_save(flags);
693         ptep = find_linux_pte_or_hugepte(mm->pgd, address, &shift);
694         if (!ptep)
695                 goto no_page;
696         pte = READ_ONCE(*ptep);
697         /*
698          * Verify it is a huge page else bail.
699          * Transparent hugepages are handled by generic code. We can skip them
700          * here.
701          */
702         if (!shift || pmd_trans_huge(__pmd(pte_val(pte))))
703                 goto no_page;
704 
705         if (!pte_present(pte)) {
706                 page = NULL;
707                 goto no_page;
708         }
709         mask = (1UL << shift) - 1;
710         page = pte_page(pte);
711         if (page)
712                 page += (address & mask) / PAGE_SIZE;
713 
714 no_page:
715         local_irq_restore(flags);
716         return page;
717 }
718 
719 struct page *
720 follow_huge_pmd(struct mm_struct *mm, unsigned long address,
721                 pmd_t *pmd, int write)
722 {
723         BUG();
724         return NULL;
725 }
726 
727 struct page *
728 follow_huge_pud(struct mm_struct *mm, unsigned long address,
729                 pud_t *pud, int write)
730 {
731         BUG();
732         return NULL;
733 }
734 
735 static unsigned long hugepte_addr_end(unsigned long addr, unsigned long end,
736                                       unsigned long sz)
737 {
738         unsigned long __boundary = (addr + sz) & ~(sz-1);
739         return (__boundary - 1 < end - 1) ? __boundary : end;
740 }
741 
742 int gup_huge_pd(hugepd_t hugepd, unsigned long addr, unsigned pdshift,
743                 unsigned long end, int write, struct page **pages, int *nr)
744 {
745         pte_t *ptep;
746         unsigned long sz = 1UL << hugepd_shift(hugepd);
747         unsigned long next;
748 
749         ptep = hugepte_offset(hugepd, addr, pdshift);
750         do {
751                 next = hugepte_addr_end(addr, end, sz);
752                 if (!gup_hugepte(ptep, sz, addr, end, write, pages, nr))
753                         return 0;
754         } while (ptep++, addr = next, addr != end);
755 
756         return 1;
757 }
758 
759 #ifdef CONFIG_PPC_MM_SLICES
760 unsigned long hugetlb_get_unmapped_area(struct file *file, unsigned long addr,
761                                         unsigned long len, unsigned long pgoff,
762                                         unsigned long flags)
763 {
764         struct hstate *hstate = hstate_file(file);
765         int mmu_psize = shift_to_mmu_psize(huge_page_shift(hstate));
766 
767         return slice_get_unmapped_area(addr, len, flags, mmu_psize, 1);
768 }
769 #endif
770 
771 unsigned long vma_mmu_pagesize(struct vm_area_struct *vma)
772 {
773 #ifdef CONFIG_PPC_MM_SLICES
774         unsigned int psize = get_slice_psize(vma->vm_mm, vma->vm_start);
775 
776         return 1UL << mmu_psize_to_shift(psize);
777 #else
778         if (!is_vm_hugetlb_page(vma))
779                 return PAGE_SIZE;
780 
781         return huge_page_size(hstate_vma(vma));
782 #endif
783 }
784 
785 static inline bool is_power_of_4(unsigned long x)
786 {
787         if (is_power_of_2(x))
788                 return (__ilog2(x) % 2) ? false : true;
789         return false;
790 }
791 
792 static int __init add_huge_page_size(unsigned long long size)
793 {
794         int shift = __ffs(size);
795         int mmu_psize;
796 
797         /* Check that it is a page size supported by the hardware and
798          * that it fits within pagetable and slice limits. */
799 #ifdef CONFIG_PPC_FSL_BOOK3E
800         if ((size < PAGE_SIZE) || !is_power_of_4(size))
801                 return -EINVAL;
802 #else
803         if (!is_power_of_2(size)
804             || (shift > SLICE_HIGH_SHIFT) || (shift <= PAGE_SHIFT))
805                 return -EINVAL;
806 #endif
807 
808         if ((mmu_psize = shift_to_mmu_psize(shift)) < 0)
809                 return -EINVAL;
810 
811 #ifdef CONFIG_SPU_FS_64K_LS
812         /* Disable support for 64K huge pages when 64K SPU local store
813          * support is enabled as the current implementation conflicts.
814          */
815         if (shift == PAGE_SHIFT_64K)
816                 return -EINVAL;
817 #endif /* CONFIG_SPU_FS_64K_LS */
818 
819         BUG_ON(mmu_psize_defs[mmu_psize].shift != shift);
820 
821         /* Return if huge page size has already been setup */
822         if (size_to_hstate(size))
823                 return 0;
824 
825         hugetlb_add_hstate(shift - PAGE_SHIFT);
826 
827         return 0;
828 }
829 
830 static int __init hugepage_setup_sz(char *str)
831 {
832         unsigned long long size;
833 
834         size = memparse(str, &str);
835 
836         if (add_huge_page_size(size) != 0)
837                 printk(KERN_WARNING "Invalid huge page size specified(%llu)\n", size);
838 
839         return 1;
840 }
841 __setup("hugepagesz=", hugepage_setup_sz);
842 
843 #ifdef CONFIG_PPC_FSL_BOOK3E
844 struct kmem_cache *hugepte_cache;
845 static int __init hugetlbpage_init(void)
846 {
847         int psize;
848 
849         for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
850                 unsigned shift;
851 
852                 if (!mmu_psize_defs[psize].shift)
853                         continue;
854 
855                 shift = mmu_psize_to_shift(psize);
856 
857                 /* Don't treat normal page sizes as huge... */
858                 if (shift != PAGE_SHIFT)
859                         if (add_huge_page_size(1ULL << shift) < 0)
860                                 continue;
861         }
862 
863         /*
864          * Create a kmem cache for hugeptes.  The bottom bits in the pte have
865          * size information encoded in them, so align them to allow this
866          */
867         hugepte_cache =  kmem_cache_create("hugepte-cache", sizeof(pte_t),
868                                            HUGEPD_SHIFT_MASK + 1, 0, NULL);
869         if (hugepte_cache == NULL)
870                 panic("%s: Unable to create kmem cache for hugeptes\n",
871                       __func__);
872 
873         /* Default hpage size = 4M */
874         if (mmu_psize_defs[MMU_PAGE_4M].shift)
875                 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_4M].shift;
876         else
877                 panic("%s: Unable to set default huge page size\n", __func__);
878 
879 
880         return 0;
881 }
882 #else
883 static int __init hugetlbpage_init(void)
884 {
885         int psize;
886 
887         if (!mmu_has_feature(MMU_FTR_16M_PAGE))
888                 return -ENODEV;
889 
890         for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
891                 unsigned shift;
892                 unsigned pdshift;
893 
894                 if (!mmu_psize_defs[psize].shift)
895                         continue;
896 
897                 shift = mmu_psize_to_shift(psize);
898 
899                 if (add_huge_page_size(1ULL << shift) < 0)
900                         continue;
901 
902                 if (shift < PMD_SHIFT)
903                         pdshift = PMD_SHIFT;
904                 else if (shift < PUD_SHIFT)
905                         pdshift = PUD_SHIFT;
906                 else
907                         pdshift = PGDIR_SHIFT;
908                 /*
909                  * if we have pdshift and shift value same, we don't
910                  * use pgt cache for hugepd.
911                  */
912                 if (pdshift != shift) {
913                         pgtable_cache_add(pdshift - shift, NULL);
914                         if (!PGT_CACHE(pdshift - shift))
915                                 panic("hugetlbpage_init(): could not create "
916                                       "pgtable cache for %d bit pagesize\n", shift);
917                 }
918         }
919 
920         /* Set default large page size. Currently, we pick 16M or 1M
921          * depending on what is available
922          */
923         if (mmu_psize_defs[MMU_PAGE_16M].shift)
924                 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_16M].shift;
925         else if (mmu_psize_defs[MMU_PAGE_1M].shift)
926                 HPAGE_SHIFT = mmu_psize_defs[MMU_PAGE_1M].shift;
927 
928         return 0;
929 }
930 #endif
931 arch_initcall(hugetlbpage_init);
932 
933 void flush_dcache_icache_hugepage(struct page *page)
934 {
935         int i;
936         void *start;
937 
938         BUG_ON(!PageCompound(page));
939 
940         for (i = 0; i < (1UL << compound_order(page)); i++) {
941                 if (!PageHighMem(page)) {
942                         __flush_dcache_icache(page_address(page+i));
943                 } else {
944                         start = kmap_atomic(page+i);
945                         __flush_dcache_icache(start);
946                         kunmap_atomic(start);
947                 }
948         }
949 }
950 
951 #endif /* CONFIG_HUGETLB_PAGE */
952 
953 /*
954  * We have 4 cases for pgds and pmds:
955  * (1) invalid (all zeroes)
956  * (2) pointer to next table, as normal; bottom 6 bits == 0
957  * (3) leaf pte for huge page, bottom two bits != 00
958  * (4) hugepd pointer, bottom two bits == 00, next 4 bits indicate size of table
959  *
960  * So long as we atomically load page table pointers we are safe against teardown,
961  * we can follow the address down to the the page and take a ref on it.
962  * This function need to be called with interrupts disabled. We use this variant
963  * when we have MSR[EE] = 0 but the paca->soft_enabled = 1
964  */
965 
966 pte_t *__find_linux_pte_or_hugepte(pgd_t *pgdir, unsigned long ea,
967                                    unsigned *shift)
968 {
969         pgd_t pgd, *pgdp;
970         pud_t pud, *pudp;
971         pmd_t pmd, *pmdp;
972         pte_t *ret_pte;
973         hugepd_t *hpdp = NULL;
974         unsigned pdshift = PGDIR_SHIFT;
975 
976         if (shift)
977                 *shift = 0;
978 
979         pgdp = pgdir + pgd_index(ea);
980         pgd  = READ_ONCE(*pgdp);
981         /*
982          * Always operate on the local stack value. This make sure the
983          * value don't get updated by a parallel THP split/collapse,
984          * page fault or a page unmap. The return pte_t * is still not
985          * stable. So should be checked there for above conditions.
986          */
987         if (pgd_none(pgd))
988                 return NULL;
989         else if (pgd_huge(pgd)) {
990                 ret_pte = (pte_t *) pgdp;
991                 goto out;
992         } else if (is_hugepd(__hugepd(pgd_val(pgd))))
993                 hpdp = (hugepd_t *)&pgd;
994         else {
995                 /*
996                  * Even if we end up with an unmap, the pgtable will not
997                  * be freed, because we do an rcu free and here we are
998                  * irq disabled
999                  */
1000                 pdshift = PUD_SHIFT;
1001                 pudp = pud_offset(&pgd, ea);
1002                 pud  = READ_ONCE(*pudp);
1003 
1004                 if (pud_none(pud))
1005                         return NULL;
1006                 else if (pud_huge(pud)) {
1007                         ret_pte = (pte_t *) pudp;
1008                         goto out;
1009                 } else if (is_hugepd(__hugepd(pud_val(pud))))
1010                         hpdp = (hugepd_t *)&pud;
1011                 else {
1012                         pdshift = PMD_SHIFT;
1013                         pmdp = pmd_offset(&pud, ea);
1014                         pmd  = READ_ONCE(*pmdp);
1015                         /*
1016                          * A hugepage collapse is captured by pmd_none, because
1017                          * it mark the pmd none and do a hpte invalidate.
1018                          *
1019                          * We don't worry about pmd_trans_splitting here, The
1020                          * caller if it needs to handle the splitting case
1021                          * should check for that.
1022                          */
1023                         if (pmd_none(pmd))
1024                                 return NULL;
1025 
1026                         if (pmd_huge(pmd) || pmd_large(pmd)) {
1027                                 ret_pte = (pte_t *) pmdp;
1028                                 goto out;
1029                         } else if (is_hugepd(__hugepd(pmd_val(pmd))))
1030                                 hpdp = (hugepd_t *)&pmd;
1031                         else
1032                                 return pte_offset_kernel(&pmd, ea);
1033                 }
1034         }
1035         if (!hpdp)
1036                 return NULL;
1037 
1038         ret_pte = hugepte_offset(*hpdp, ea, pdshift);
1039         pdshift = hugepd_shift(*hpdp);
1040 out:
1041         if (shift)
1042                 *shift = pdshift;
1043         return ret_pte;
1044 }
1045 EXPORT_SYMBOL_GPL(__find_linux_pte_or_hugepte);
1046 
1047 int gup_hugepte(pte_t *ptep, unsigned long sz, unsigned long addr,
1048                 unsigned long end, int write, struct page **pages, int *nr)
1049 {
1050         unsigned long mask;
1051         unsigned long pte_end;
1052         struct page *head, *page, *tail;
1053         pte_t pte;
1054         int refs;
1055 
1056         pte_end = (addr + sz) & ~(sz-1);
1057         if (pte_end < end)
1058                 end = pte_end;
1059 
1060         pte = READ_ONCE(*ptep);
1061         mask = _PAGE_PRESENT | _PAGE_USER;
1062         if (write)
1063                 mask |= _PAGE_RW;
1064 
1065         if ((pte_val(pte) & mask) != mask)
1066                 return 0;
1067 
1068         /* hugepages are never "special" */
1069         VM_BUG_ON(!pfn_valid(pte_pfn(pte)));
1070 
1071         refs = 0;
1072         head = pte_page(pte);
1073 
1074         page = head + ((addr & (sz-1)) >> PAGE_SHIFT);
1075         tail = page;
1076         do {
1077                 VM_BUG_ON(compound_head(page) != head);
1078                 pages[*nr] = page;
1079                 (*nr)++;
1080                 page++;
1081                 refs++;
1082         } while (addr += PAGE_SIZE, addr != end);
1083 
1084         if (!page_cache_add_speculative(head, refs)) {
1085                 *nr -= refs;
1086                 return 0;
1087         }
1088 
1089         if (unlikely(pte_val(pte) != pte_val(*ptep))) {
1090                 /* Could be optimized better */
1091                 *nr -= refs;
1092                 while (refs--)
1093                         put_page(head);
1094                 return 0;
1095         }
1096 
1097         /*
1098          * Any tail page need their mapcount reference taken before we
1099          * return.
1100          */
1101         while (refs--) {
1102                 if (PageTail(tail))
1103                         get_huge_page_tail(tail);
1104                 tail++;
1105         }
1106 
1107         return 1;
1108 }
1109 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp