]> git.proxmox.com Git - mirror_ubuntu-zesty-kernel.git/blob - arch/powerpc/mm/pgtable_64.c
Merge tag 'efi-next' of git://git.kernel.org/pub/scm/linux/kernel/git/mfleming/efi...
[mirror_ubuntu-zesty-kernel.git] / arch / powerpc / mm / pgtable_64.c
1 /*
2 * This file contains ioremap and related functions for 64-bit machines.
3 *
4 * Derived from arch/ppc64/mm/init.c
5 * Copyright (C) 1995-1996 Gary Thomas (gdt@linuxppc.org)
6 *
7 * Modifications by Paul Mackerras (PowerMac) (paulus@samba.org)
8 * and Cort Dougan (PReP) (cort@cs.nmt.edu)
9 * Copyright (C) 1996 Paul Mackerras
10 *
11 * Derived from "arch/i386/mm/init.c"
12 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
13 *
14 * Dave Engebretsen <engebret@us.ibm.com>
15 * Rework for PPC64 port.
16 *
17 * This program is free software; you can redistribute it and/or
18 * modify it under the terms of the GNU General Public License
19 * as published by the Free Software Foundation; either version
20 * 2 of the License, or (at your option) any later version.
21 *
22 */
23
24 #include <linux/signal.h>
25 #include <linux/sched.h>
26 #include <linux/kernel.h>
27 #include <linux/errno.h>
28 #include <linux/string.h>
29 #include <linux/export.h>
30 #include <linux/types.h>
31 #include <linux/mman.h>
32 #include <linux/mm.h>
33 #include <linux/swap.h>
34 #include <linux/stddef.h>
35 #include <linux/vmalloc.h>
36 #include <linux/memblock.h>
37 #include <linux/slab.h>
38 #include <linux/hugetlb.h>
39
40 #include <asm/pgalloc.h>
41 #include <asm/page.h>
42 #include <asm/prom.h>
43 #include <asm/io.h>
44 #include <asm/mmu_context.h>
45 #include <asm/pgtable.h>
46 #include <asm/mmu.h>
47 #include <asm/smp.h>
48 #include <asm/machdep.h>
49 #include <asm/tlb.h>
50 #include <asm/processor.h>
51 #include <asm/cputable.h>
52 #include <asm/sections.h>
53 #include <asm/firmware.h>
54 #include <asm/dma.h>
55
56 #include "mmu_decl.h"
57
58 #define CREATE_TRACE_POINTS
59 #include <trace/events/thp.h>
60
61 /* Some sanity checking */
62 #if TASK_SIZE_USER64 > PGTABLE_RANGE
63 #error TASK_SIZE_USER64 exceeds pagetable range
64 #endif
65
66 #ifdef CONFIG_PPC_STD_MMU_64
67 #if TASK_SIZE_USER64 > (1UL << (ESID_BITS + SID_SHIFT))
68 #error TASK_SIZE_USER64 exceeds user VSID range
69 #endif
70 #endif
71
72 unsigned long ioremap_bot = IOREMAP_BASE;
73
74 #ifdef CONFIG_PPC_MMU_NOHASH
75 static __ref void *early_alloc_pgtable(unsigned long size)
76 {
77 void *pt;
78
79 pt = __va(memblock_alloc_base(size, size, __pa(MAX_DMA_ADDRESS)));
80 memset(pt, 0, size);
81
82 return pt;
83 }
84 #endif /* CONFIG_PPC_MMU_NOHASH */
85
86 /*
87 * map_kernel_page currently only called by __ioremap
88 * map_kernel_page adds an entry to the ioremap page table
89 * and adds an entry to the HPT, possibly bolting it
90 */
91 int map_kernel_page(unsigned long ea, unsigned long pa, int flags)
92 {
93 pgd_t *pgdp;
94 pud_t *pudp;
95 pmd_t *pmdp;
96 pte_t *ptep;
97
98 if (slab_is_available()) {
99 pgdp = pgd_offset_k(ea);
100 pudp = pud_alloc(&init_mm, pgdp, ea);
101 if (!pudp)
102 return -ENOMEM;
103 pmdp = pmd_alloc(&init_mm, pudp, ea);
104 if (!pmdp)
105 return -ENOMEM;
106 ptep = pte_alloc_kernel(pmdp, ea);
107 if (!ptep)
108 return -ENOMEM;
109 set_pte_at(&init_mm, ea, ptep, pfn_pte(pa >> PAGE_SHIFT,
110 __pgprot(flags)));
111 } else {
112 #ifdef CONFIG_PPC_MMU_NOHASH
113 pgdp = pgd_offset_k(ea);
114 #ifdef PUD_TABLE_SIZE
115 if (pgd_none(*pgdp)) {
116 pudp = early_alloc_pgtable(PUD_TABLE_SIZE);
117 BUG_ON(pudp == NULL);
118 pgd_populate(&init_mm, pgdp, pudp);
119 }
120 #endif /* PUD_TABLE_SIZE */
121 pudp = pud_offset(pgdp, ea);
122 if (pud_none(*pudp)) {
123 pmdp = early_alloc_pgtable(PMD_TABLE_SIZE);
124 BUG_ON(pmdp == NULL);
125 pud_populate(&init_mm, pudp, pmdp);
126 }
127 pmdp = pmd_offset(pudp, ea);
128 if (!pmd_present(*pmdp)) {
129 ptep = early_alloc_pgtable(PAGE_SIZE);
130 BUG_ON(ptep == NULL);
131 pmd_populate_kernel(&init_mm, pmdp, ptep);
132 }
133 ptep = pte_offset_kernel(pmdp, ea);
134 set_pte_at(&init_mm, ea, ptep, pfn_pte(pa >> PAGE_SHIFT,
135 __pgprot(flags)));
136 #else /* CONFIG_PPC_MMU_NOHASH */
137 /*
138 * If the mm subsystem is not fully up, we cannot create a
139 * linux page table entry for this mapping. Simply bolt an
140 * entry in the hardware page table.
141 *
142 */
143 if (htab_bolt_mapping(ea, ea + PAGE_SIZE, pa, flags,
144 mmu_io_psize, mmu_kernel_ssize)) {
145 printk(KERN_ERR "Failed to do bolted mapping IO "
146 "memory at %016lx !\n", pa);
147 return -ENOMEM;
148 }
149 #endif /* !CONFIG_PPC_MMU_NOHASH */
150 }
151
152 #ifdef CONFIG_PPC_BOOK3E_64
153 /*
154 * With hardware tablewalk, a sync is needed to ensure that
155 * subsequent accesses see the PTE we just wrote. Unlike userspace
156 * mappings, we can't tolerate spurious faults, so make sure
157 * the new PTE will be seen the first time.
158 */
159 mb();
160 #else
161 smp_wmb();
162 #endif
163 return 0;
164 }
165
166
167 /**
168 * __ioremap_at - Low level function to establish the page tables
169 * for an IO mapping
170 */
171 void __iomem * __ioremap_at(phys_addr_t pa, void *ea, unsigned long size,
172 unsigned long flags)
173 {
174 unsigned long i;
175
176 /* Make sure we have the base flags */
177 if ((flags & _PAGE_PRESENT) == 0)
178 flags |= pgprot_val(PAGE_KERNEL);
179
180 /* Non-cacheable page cannot be coherent */
181 if (flags & _PAGE_NO_CACHE)
182 flags &= ~_PAGE_COHERENT;
183
184 /* We don't support the 4K PFN hack with ioremap */
185 if (flags & _PAGE_4K_PFN)
186 return NULL;
187
188 WARN_ON(pa & ~PAGE_MASK);
189 WARN_ON(((unsigned long)ea) & ~PAGE_MASK);
190 WARN_ON(size & ~PAGE_MASK);
191
192 for (i = 0; i < size; i += PAGE_SIZE)
193 if (map_kernel_page((unsigned long)ea+i, pa+i, flags))
194 return NULL;
195
196 return (void __iomem *)ea;
197 }
198
199 /**
200 * __iounmap_from - Low level function to tear down the page tables
201 * for an IO mapping. This is used for mappings that
202 * are manipulated manually, like partial unmapping of
203 * PCI IOs or ISA space.
204 */
205 void __iounmap_at(void *ea, unsigned long size)
206 {
207 WARN_ON(((unsigned long)ea) & ~PAGE_MASK);
208 WARN_ON(size & ~PAGE_MASK);
209
210 unmap_kernel_range((unsigned long)ea, size);
211 }
212
213 void __iomem * __ioremap_caller(phys_addr_t addr, unsigned long size,
214 unsigned long flags, void *caller)
215 {
216 phys_addr_t paligned;
217 void __iomem *ret;
218
219 /*
220 * Choose an address to map it to.
221 * Once the imalloc system is running, we use it.
222 * Before that, we map using addresses going
223 * up from ioremap_bot. imalloc will use
224 * the addresses from ioremap_bot through
225 * IMALLOC_END
226 *
227 */
228 paligned = addr & PAGE_MASK;
229 size = PAGE_ALIGN(addr + size) - paligned;
230
231 if ((size == 0) || (paligned == 0))
232 return NULL;
233
234 if (slab_is_available()) {
235 struct vm_struct *area;
236
237 area = __get_vm_area_caller(size, VM_IOREMAP,
238 ioremap_bot, IOREMAP_END,
239 caller);
240 if (area == NULL)
241 return NULL;
242
243 area->phys_addr = paligned;
244 ret = __ioremap_at(paligned, area->addr, size, flags);
245 if (!ret)
246 vunmap(area->addr);
247 } else {
248 ret = __ioremap_at(paligned, (void *)ioremap_bot, size, flags);
249 if (ret)
250 ioremap_bot += size;
251 }
252
253 if (ret)
254 ret += addr & ~PAGE_MASK;
255 return ret;
256 }
257
258 void __iomem * __ioremap(phys_addr_t addr, unsigned long size,
259 unsigned long flags)
260 {
261 return __ioremap_caller(addr, size, flags, __builtin_return_address(0));
262 }
263
264 void __iomem * ioremap(phys_addr_t addr, unsigned long size)
265 {
266 unsigned long flags = _PAGE_NO_CACHE | _PAGE_GUARDED;
267 void *caller = __builtin_return_address(0);
268
269 if (ppc_md.ioremap)
270 return ppc_md.ioremap(addr, size, flags, caller);
271 return __ioremap_caller(addr, size, flags, caller);
272 }
273
274 void __iomem * ioremap_wc(phys_addr_t addr, unsigned long size)
275 {
276 unsigned long flags = _PAGE_NO_CACHE;
277 void *caller = __builtin_return_address(0);
278
279 if (ppc_md.ioremap)
280 return ppc_md.ioremap(addr, size, flags, caller);
281 return __ioremap_caller(addr, size, flags, caller);
282 }
283
284 void __iomem * ioremap_prot(phys_addr_t addr, unsigned long size,
285 unsigned long flags)
286 {
287 void *caller = __builtin_return_address(0);
288
289 /* writeable implies dirty for kernel addresses */
290 if (flags & _PAGE_RW)
291 flags |= _PAGE_DIRTY;
292
293 /* we don't want to let _PAGE_USER and _PAGE_EXEC leak out */
294 flags &= ~(_PAGE_USER | _PAGE_EXEC);
295
296 #ifdef _PAGE_BAP_SR
297 /* _PAGE_USER contains _PAGE_BAP_SR on BookE using the new PTE format
298 * which means that we just cleared supervisor access... oops ;-) This
299 * restores it
300 */
301 flags |= _PAGE_BAP_SR;
302 #endif
303
304 if (ppc_md.ioremap)
305 return ppc_md.ioremap(addr, size, flags, caller);
306 return __ioremap_caller(addr, size, flags, caller);
307 }
308
309
310 /*
311 * Unmap an IO region and remove it from imalloc'd list.
312 * Access to IO memory should be serialized by driver.
313 */
314 void __iounmap(volatile void __iomem *token)
315 {
316 void *addr;
317
318 if (!slab_is_available())
319 return;
320
321 addr = (void *) ((unsigned long __force)
322 PCI_FIX_ADDR(token) & PAGE_MASK);
323 if ((unsigned long)addr < ioremap_bot) {
324 printk(KERN_WARNING "Attempt to iounmap early bolted mapping"
325 " at 0x%p\n", addr);
326 return;
327 }
328 vunmap(addr);
329 }
330
331 void iounmap(volatile void __iomem *token)
332 {
333 if (ppc_md.iounmap)
334 ppc_md.iounmap(token);
335 else
336 __iounmap(token);
337 }
338
339 EXPORT_SYMBOL(ioremap);
340 EXPORT_SYMBOL(ioremap_wc);
341 EXPORT_SYMBOL(ioremap_prot);
342 EXPORT_SYMBOL(__ioremap);
343 EXPORT_SYMBOL(__ioremap_at);
344 EXPORT_SYMBOL(iounmap);
345 EXPORT_SYMBOL(__iounmap);
346 EXPORT_SYMBOL(__iounmap_at);
347
348 #ifndef __PAGETABLE_PUD_FOLDED
349 /* 4 level page table */
350 struct page *pgd_page(pgd_t pgd)
351 {
352 if (pgd_huge(pgd))
353 return pte_page(pgd_pte(pgd));
354 return virt_to_page(pgd_page_vaddr(pgd));
355 }
356 #endif
357
358 struct page *pud_page(pud_t pud)
359 {
360 if (pud_huge(pud))
361 return pte_page(pud_pte(pud));
362 return virt_to_page(pud_page_vaddr(pud));
363 }
364
365 /*
366 * For hugepage we have pfn in the pmd, we use PTE_RPN_SHIFT bits for flags
367 * For PTE page, we have a PTE_FRAG_SIZE (4K) aligned virtual address.
368 */
369 struct page *pmd_page(pmd_t pmd)
370 {
371 if (pmd_trans_huge(pmd) || pmd_huge(pmd))
372 return pfn_to_page(pmd_pfn(pmd));
373 return virt_to_page(pmd_page_vaddr(pmd));
374 }
375
376 #ifdef CONFIG_PPC_64K_PAGES
377 static pte_t *get_from_cache(struct mm_struct *mm)
378 {
379 void *pte_frag, *ret;
380
381 spin_lock(&mm->page_table_lock);
382 ret = mm->context.pte_frag;
383 if (ret) {
384 pte_frag = ret + PTE_FRAG_SIZE;
385 /*
386 * If we have taken up all the fragments mark PTE page NULL
387 */
388 if (((unsigned long)pte_frag & ~PAGE_MASK) == 0)
389 pte_frag = NULL;
390 mm->context.pte_frag = pte_frag;
391 }
392 spin_unlock(&mm->page_table_lock);
393 return (pte_t *)ret;
394 }
395
396 static pte_t *__alloc_for_cache(struct mm_struct *mm, int kernel)
397 {
398 void *ret = NULL;
399 struct page *page = alloc_page(GFP_KERNEL | __GFP_NOTRACK |
400 __GFP_REPEAT | __GFP_ZERO);
401 if (!page)
402 return NULL;
403 if (!kernel && !pgtable_page_ctor(page)) {
404 __free_page(page);
405 return NULL;
406 }
407
408 ret = page_address(page);
409 spin_lock(&mm->page_table_lock);
410 /*
411 * If we find pgtable_page set, we return
412 * the allocated page with single fragement
413 * count.
414 */
415 if (likely(!mm->context.pte_frag)) {
416 atomic_set(&page->_count, PTE_FRAG_NR);
417 mm->context.pte_frag = ret + PTE_FRAG_SIZE;
418 }
419 spin_unlock(&mm->page_table_lock);
420
421 return (pte_t *)ret;
422 }
423
424 pte_t *page_table_alloc(struct mm_struct *mm, unsigned long vmaddr, int kernel)
425 {
426 pte_t *pte;
427
428 pte = get_from_cache(mm);
429 if (pte)
430 return pte;
431
432 return __alloc_for_cache(mm, kernel);
433 }
434
435 void page_table_free(struct mm_struct *mm, unsigned long *table, int kernel)
436 {
437 struct page *page = virt_to_page(table);
438 if (put_page_testzero(page)) {
439 if (!kernel)
440 pgtable_page_dtor(page);
441 free_hot_cold_page(page, 0);
442 }
443 }
444
445 #ifdef CONFIG_SMP
446 static void page_table_free_rcu(void *table)
447 {
448 struct page *page = virt_to_page(table);
449 if (put_page_testzero(page)) {
450 pgtable_page_dtor(page);
451 free_hot_cold_page(page, 0);
452 }
453 }
454
455 void pgtable_free_tlb(struct mmu_gather *tlb, void *table, int shift)
456 {
457 unsigned long pgf = (unsigned long)table;
458
459 BUG_ON(shift > MAX_PGTABLE_INDEX_SIZE);
460 pgf |= shift;
461 tlb_remove_table(tlb, (void *)pgf);
462 }
463
464 void __tlb_remove_table(void *_table)
465 {
466 void *table = (void *)((unsigned long)_table & ~MAX_PGTABLE_INDEX_SIZE);
467 unsigned shift = (unsigned long)_table & MAX_PGTABLE_INDEX_SIZE;
468
469 if (!shift)
470 /* PTE page needs special handling */
471 page_table_free_rcu(table);
472 else {
473 BUG_ON(shift > MAX_PGTABLE_INDEX_SIZE);
474 kmem_cache_free(PGT_CACHE(shift), table);
475 }
476 }
477 #else
478 void pgtable_free_tlb(struct mmu_gather *tlb, void *table, int shift)
479 {
480 if (!shift) {
481 /* PTE page needs special handling */
482 struct page *page = virt_to_page(table);
483 if (put_page_testzero(page)) {
484 pgtable_page_dtor(page);
485 free_hot_cold_page(page, 0);
486 }
487 } else {
488 BUG_ON(shift > MAX_PGTABLE_INDEX_SIZE);
489 kmem_cache_free(PGT_CACHE(shift), table);
490 }
491 }
492 #endif
493 #endif /* CONFIG_PPC_64K_PAGES */
494
495 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
496
497 /*
498 * This is called when relaxing access to a hugepage. It's also called in the page
499 * fault path when we don't hit any of the major fault cases, ie, a minor
500 * update of _PAGE_ACCESSED, _PAGE_DIRTY, etc... The generic code will have
501 * handled those two for us, we additionally deal with missing execute
502 * permission here on some processors
503 */
504 int pmdp_set_access_flags(struct vm_area_struct *vma, unsigned long address,
505 pmd_t *pmdp, pmd_t entry, int dirty)
506 {
507 int changed;
508 #ifdef CONFIG_DEBUG_VM
509 WARN_ON(!pmd_trans_huge(*pmdp));
510 assert_spin_locked(&vma->vm_mm->page_table_lock);
511 #endif
512 changed = !pmd_same(*(pmdp), entry);
513 if (changed) {
514 __ptep_set_access_flags(pmdp_ptep(pmdp), pmd_pte(entry));
515 /*
516 * Since we are not supporting SW TLB systems, we don't
517 * have any thing similar to flush_tlb_page_nohash()
518 */
519 }
520 return changed;
521 }
522
523 unsigned long pmd_hugepage_update(struct mm_struct *mm, unsigned long addr,
524 pmd_t *pmdp, unsigned long clr,
525 unsigned long set)
526 {
527
528 unsigned long old, tmp;
529
530 #ifdef CONFIG_DEBUG_VM
531 WARN_ON(!pmd_trans_huge(*pmdp));
532 assert_spin_locked(&mm->page_table_lock);
533 #endif
534
535 #ifdef PTE_ATOMIC_UPDATES
536 __asm__ __volatile__(
537 "1: ldarx %0,0,%3\n\
538 andi. %1,%0,%6\n\
539 bne- 1b \n\
540 andc %1,%0,%4 \n\
541 or %1,%1,%7\n\
542 stdcx. %1,0,%3 \n\
543 bne- 1b"
544 : "=&r" (old), "=&r" (tmp), "=m" (*pmdp)
545 : "r" (pmdp), "r" (clr), "m" (*pmdp), "i" (_PAGE_BUSY), "r" (set)
546 : "cc" );
547 #else
548 old = pmd_val(*pmdp);
549 *pmdp = __pmd((old & ~clr) | set);
550 #endif
551 trace_hugepage_update(addr, old, clr, set);
552 if (old & _PAGE_HASHPTE)
553 hpte_do_hugepage_flush(mm, addr, pmdp, old);
554 return old;
555 }
556
557 pmd_t pmdp_clear_flush(struct vm_area_struct *vma, unsigned long address,
558 pmd_t *pmdp)
559 {
560 pmd_t pmd;
561
562 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
563 if (pmd_trans_huge(*pmdp)) {
564 pmd = pmdp_get_and_clear(vma->vm_mm, address, pmdp);
565 } else {
566 /*
567 * khugepaged calls this for normal pmd
568 */
569 pmd = *pmdp;
570 pmd_clear(pmdp);
571 /*
572 * Wait for all pending hash_page to finish. This is needed
573 * in case of subpage collapse. When we collapse normal pages
574 * to hugepage, we first clear the pmd, then invalidate all
575 * the PTE entries. The assumption here is that any low level
576 * page fault will see a none pmd and take the slow path that
577 * will wait on mmap_sem. But we could very well be in a
578 * hash_page with local ptep pointer value. Such a hash page
579 * can result in adding new HPTE entries for normal subpages.
580 * That means we could be modifying the page content as we
581 * copy them to a huge page. So wait for parallel hash_page
582 * to finish before invalidating HPTE entries. We can do this
583 * by sending an IPI to all the cpus and executing a dummy
584 * function there.
585 */
586 kick_all_cpus_sync();
587 /*
588 * Now invalidate the hpte entries in the range
589 * covered by pmd. This make sure we take a
590 * fault and will find the pmd as none, which will
591 * result in a major fault which takes mmap_sem and
592 * hence wait for collapse to complete. Without this
593 * the __collapse_huge_page_copy can result in copying
594 * the old content.
595 */
596 flush_tlb_pmd_range(vma->vm_mm, &pmd, address);
597 }
598 return pmd;
599 }
600
601 int pmdp_test_and_clear_young(struct vm_area_struct *vma,
602 unsigned long address, pmd_t *pmdp)
603 {
604 return __pmdp_test_and_clear_young(vma->vm_mm, address, pmdp);
605 }
606
607 /*
608 * We currently remove entries from the hashtable regardless of whether
609 * the entry was young or dirty. The generic routines only flush if the
610 * entry was young or dirty which is not good enough.
611 *
612 * We should be more intelligent about this but for the moment we override
613 * these functions and force a tlb flush unconditionally
614 */
615 int pmdp_clear_flush_young(struct vm_area_struct *vma,
616 unsigned long address, pmd_t *pmdp)
617 {
618 return __pmdp_test_and_clear_young(vma->vm_mm, address, pmdp);
619 }
620
621 /*
622 * We mark the pmd splitting and invalidate all the hpte
623 * entries for this hugepage.
624 */
625 void pmdp_splitting_flush(struct vm_area_struct *vma,
626 unsigned long address, pmd_t *pmdp)
627 {
628 unsigned long old, tmp;
629
630 VM_BUG_ON(address & ~HPAGE_PMD_MASK);
631
632 #ifdef CONFIG_DEBUG_VM
633 WARN_ON(!pmd_trans_huge(*pmdp));
634 assert_spin_locked(&vma->vm_mm->page_table_lock);
635 #endif
636
637 #ifdef PTE_ATOMIC_UPDATES
638
639 __asm__ __volatile__(
640 "1: ldarx %0,0,%3\n\
641 andi. %1,%0,%6\n\
642 bne- 1b \n\
643 ori %1,%0,%4 \n\
644 stdcx. %1,0,%3 \n\
645 bne- 1b"
646 : "=&r" (old), "=&r" (tmp), "=m" (*pmdp)
647 : "r" (pmdp), "i" (_PAGE_SPLITTING), "m" (*pmdp), "i" (_PAGE_BUSY)
648 : "cc" );
649 #else
650 old = pmd_val(*pmdp);
651 *pmdp = __pmd(old | _PAGE_SPLITTING);
652 #endif
653 /*
654 * If we didn't had the splitting flag set, go and flush the
655 * HPTE entries.
656 */
657 trace_hugepage_splitting(address, old);
658 if (!(old & _PAGE_SPLITTING)) {
659 /* We need to flush the hpte */
660 if (old & _PAGE_HASHPTE)
661 hpte_do_hugepage_flush(vma->vm_mm, address, pmdp, old);
662 }
663 /*
664 * This ensures that generic code that rely on IRQ disabling
665 * to prevent a parallel THP split work as expected.
666 */
667 kick_all_cpus_sync();
668 }
669
670 /*
671 * We want to put the pgtable in pmd and use pgtable for tracking
672 * the base page size hptes
673 */
674 void pgtable_trans_huge_deposit(struct mm_struct *mm, pmd_t *pmdp,
675 pgtable_t pgtable)
676 {
677 pgtable_t *pgtable_slot;
678 assert_spin_locked(&mm->page_table_lock);
679 /*
680 * we store the pgtable in the second half of PMD
681 */
682 pgtable_slot = (pgtable_t *)pmdp + PTRS_PER_PMD;
683 *pgtable_slot = pgtable;
684 /*
685 * expose the deposited pgtable to other cpus.
686 * before we set the hugepage PTE at pmd level
687 * hash fault code looks at the deposted pgtable
688 * to store hash index values.
689 */
690 smp_wmb();
691 }
692
693 pgtable_t pgtable_trans_huge_withdraw(struct mm_struct *mm, pmd_t *pmdp)
694 {
695 pgtable_t pgtable;
696 pgtable_t *pgtable_slot;
697
698 assert_spin_locked(&mm->page_table_lock);
699 pgtable_slot = (pgtable_t *)pmdp + PTRS_PER_PMD;
700 pgtable = *pgtable_slot;
701 /*
702 * Once we withdraw, mark the entry NULL.
703 */
704 *pgtable_slot = NULL;
705 /*
706 * We store HPTE information in the deposited PTE fragment.
707 * zero out the content on withdraw.
708 */
709 memset(pgtable, 0, PTE_FRAG_SIZE);
710 return pgtable;
711 }
712
713 /*
714 * set a new huge pmd. We should not be called for updating
715 * an existing pmd entry. That should go via pmd_hugepage_update.
716 */
717 void set_pmd_at(struct mm_struct *mm, unsigned long addr,
718 pmd_t *pmdp, pmd_t pmd)
719 {
720 #ifdef CONFIG_DEBUG_VM
721 WARN_ON((pmd_val(*pmdp) & (_PAGE_PRESENT | _PAGE_USER)) ==
722 (_PAGE_PRESENT | _PAGE_USER));
723 assert_spin_locked(&mm->page_table_lock);
724 WARN_ON(!pmd_trans_huge(pmd));
725 #endif
726 trace_hugepage_set_pmd(addr, pmd_val(pmd));
727 return set_pte_at(mm, addr, pmdp_ptep(pmdp), pmd_pte(pmd));
728 }
729
730 void pmdp_invalidate(struct vm_area_struct *vma, unsigned long address,
731 pmd_t *pmdp)
732 {
733 pmd_hugepage_update(vma->vm_mm, address, pmdp, _PAGE_PRESENT, 0);
734 }
735
736 /*
737 * A linux hugepage PMD was changed and the corresponding hash table entries
738 * neesd to be flushed.
739 */
740 void hpte_do_hugepage_flush(struct mm_struct *mm, unsigned long addr,
741 pmd_t *pmdp, unsigned long old_pmd)
742 {
743 int ssize;
744 unsigned int psize;
745 unsigned long vsid;
746 unsigned long flags = 0;
747 const struct cpumask *tmp;
748
749 /* get the base page size,vsid and segment size */
750 #ifdef CONFIG_DEBUG_VM
751 psize = get_slice_psize(mm, addr);
752 BUG_ON(psize == MMU_PAGE_16M);
753 #endif
754 if (old_pmd & _PAGE_COMBO)
755 psize = MMU_PAGE_4K;
756 else
757 psize = MMU_PAGE_64K;
758
759 if (!is_kernel_addr(addr)) {
760 ssize = user_segment_size(addr);
761 vsid = get_vsid(mm->context.id, addr, ssize);
762 WARN_ON(vsid == 0);
763 } else {
764 vsid = get_kernel_vsid(addr, mmu_kernel_ssize);
765 ssize = mmu_kernel_ssize;
766 }
767
768 tmp = cpumask_of(smp_processor_id());
769 if (cpumask_equal(mm_cpumask(mm), tmp))
770 flags |= HPTE_LOCAL_UPDATE;
771
772 return flush_hash_hugepage(vsid, addr, pmdp, psize, ssize, flags);
773 }
774
775 static pmd_t pmd_set_protbits(pmd_t pmd, pgprot_t pgprot)
776 {
777 pmd_val(pmd) |= pgprot_val(pgprot);
778 return pmd;
779 }
780
781 pmd_t pfn_pmd(unsigned long pfn, pgprot_t pgprot)
782 {
783 pmd_t pmd;
784 /*
785 * For a valid pte, we would have _PAGE_PRESENT always
786 * set. We use this to check THP page at pmd level.
787 * leaf pte for huge page, bottom two bits != 00
788 */
789 pmd_val(pmd) = pfn << PTE_RPN_SHIFT;
790 pmd_val(pmd) |= _PAGE_THP_HUGE;
791 pmd = pmd_set_protbits(pmd, pgprot);
792 return pmd;
793 }
794
795 pmd_t mk_pmd(struct page *page, pgprot_t pgprot)
796 {
797 return pfn_pmd(page_to_pfn(page), pgprot);
798 }
799
800 pmd_t pmd_modify(pmd_t pmd, pgprot_t newprot)
801 {
802
803 pmd_val(pmd) &= _HPAGE_CHG_MASK;
804 pmd = pmd_set_protbits(pmd, newprot);
805 return pmd;
806 }
807
808 /*
809 * This is called at the end of handling a user page fault, when the
810 * fault has been handled by updating a HUGE PMD entry in the linux page tables.
811 * We use it to preload an HPTE into the hash table corresponding to
812 * the updated linux HUGE PMD entry.
813 */
814 void update_mmu_cache_pmd(struct vm_area_struct *vma, unsigned long addr,
815 pmd_t *pmd)
816 {
817 return;
818 }
819
820 pmd_t pmdp_get_and_clear(struct mm_struct *mm,
821 unsigned long addr, pmd_t *pmdp)
822 {
823 pmd_t old_pmd;
824 pgtable_t pgtable;
825 unsigned long old;
826 pgtable_t *pgtable_slot;
827
828 old = pmd_hugepage_update(mm, addr, pmdp, ~0UL, 0);
829 old_pmd = __pmd(old);
830 /*
831 * We have pmd == none and we are holding page_table_lock.
832 * So we can safely go and clear the pgtable hash
833 * index info.
834 */
835 pgtable_slot = (pgtable_t *)pmdp + PTRS_PER_PMD;
836 pgtable = *pgtable_slot;
837 /*
838 * Let's zero out old valid and hash index details
839 * hash fault look at them.
840 */
841 memset(pgtable, 0, PTE_FRAG_SIZE);
842 /*
843 * Serialize against find_linux_pte_or_hugepte which does lock-less
844 * lookup in page tables with local interrupts disabled. For huge pages
845 * it casts pmd_t to pte_t. Since format of pte_t is different from
846 * pmd_t we want to prevent transit from pmd pointing to page table
847 * to pmd pointing to huge page (and back) while interrupts are disabled.
848 * We clear pmd to possibly replace it with page table pointer in
849 * different code paths. So make sure we wait for the parallel
850 * find_linux_pte_or_hugepage to finish.
851 */
852 kick_all_cpus_sync();
853 return old_pmd;
854 }
855
856 int has_transparent_hugepage(void)
857 {
858 if (!mmu_has_feature(MMU_FTR_16M_PAGE))
859 return 0;
860 /*
861 * We support THP only if PMD_SIZE is 16MB.
862 */
863 if (mmu_psize_defs[MMU_PAGE_16M].shift != PMD_SHIFT)
864 return 0;
865 /*
866 * We need to make sure that we support 16MB hugepage in a segement
867 * with base page size 64K or 4K. We only enable THP with a PAGE_SIZE
868 * of 64K.
869 */
870 /*
871 * If we have 64K HPTE, we will be using that by default
872 */
873 if (mmu_psize_defs[MMU_PAGE_64K].shift &&
874 (mmu_psize_defs[MMU_PAGE_64K].penc[MMU_PAGE_16M] == -1))
875 return 0;
876 /*
877 * Ok we only have 4K HPTE
878 */
879 if (mmu_psize_defs[MMU_PAGE_4K].penc[MMU_PAGE_16M] == -1)
880 return 0;
881
882 return 1;
883 }
884 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */