2 * This file contains the routines for TLB flushing.
3 * On machines where the MMU does not use a hash table to store virtual to
4 * physical translations (ie, SW loaded TLBs or Book3E compilant processors,
5 * this does -not- include 603 however which shares the implementation with
6 * hash based processors)
10 * Copyright 2008,2009 Ben Herrenschmidt <benh@kernel.crashing.org>
13 * Derived from arch/ppc/mm/init.c:
14 * Copyright (C) 1995-1996 Gary Thomas (gdt@linuxppc.org)
16 * Modifications by Paul Mackerras (PowerMac) (paulus@cs.anu.edu.au)
17 * and Cort Dougan (PReP) (cort@cs.nmt.edu)
18 * Copyright (C) 1996 Paul Mackerras
20 * Derived from "arch/i386/mm/init.c"
21 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
23 * This program is free software; you can redistribute it and/or
24 * modify it under the terms of the GNU General Public License
25 * as published by the Free Software Foundation; either version
26 * 2 of the License, or (at your option) any later version.
30 #include <linux/kernel.h>
31 #include <linux/export.h>
33 #include <linux/init.h>
34 #include <linux/highmem.h>
35 #include <linux/pagemap.h>
36 #include <linux/preempt.h>
37 #include <linux/spinlock.h>
38 #include <linux/memblock.h>
39 #include <linux/of_fdt.h>
40 #include <linux/hugetlb.h>
42 #include <asm/tlbflush.h>
44 #include <asm/code-patching.h>
45 #include <asm/hugetlb.h>
51 * This struct lists the sw-supported page sizes. The hardawre MMU may support
52 * other sizes not listed here. The .ind field is only used on MMUs that have
53 * indirect page table entries.
55 #ifdef CONFIG_PPC_BOOK3E_MMU
56 #ifdef CONFIG_PPC_FSL_BOOK3E
57 struct mmu_psize_def mmu_psize_defs
[MMU_PAGE_COUNT
] = {
60 .enc
= BOOK3E_PAGESZ_4K
,
64 .enc
= BOOK3E_PAGESZ_2M
,
68 .enc
= BOOK3E_PAGESZ_4M
,
72 .enc
= BOOK3E_PAGESZ_16M
,
76 .enc
= BOOK3E_PAGESZ_64M
,
80 .enc
= BOOK3E_PAGESZ_256M
,
84 .enc
= BOOK3E_PAGESZ_1GB
,
88 struct mmu_psize_def mmu_psize_defs
[MMU_PAGE_COUNT
] = {
92 .enc
= BOOK3E_PAGESZ_4K
,
96 .enc
= BOOK3E_PAGESZ_16K
,
101 .enc
= BOOK3E_PAGESZ_64K
,
105 .enc
= BOOK3E_PAGESZ_1M
,
110 .enc
= BOOK3E_PAGESZ_16M
,
114 .enc
= BOOK3E_PAGESZ_256M
,
118 .enc
= BOOK3E_PAGESZ_1GB
,
121 #endif /* CONFIG_FSL_BOOKE */
123 static inline int mmu_get_tsize(int psize
)
125 return mmu_psize_defs
[psize
].enc
;
128 static inline int mmu_get_tsize(int psize
)
130 /* This isn't used on !Book3E for now */
133 #endif /* CONFIG_PPC_BOOK3E_MMU */
135 /* The variables below are currently only used on 64-bit Book3E
136 * though this will probably be made common with other nohash
137 * implementations at some point
141 int mmu_linear_psize
; /* Page size used for the linear mapping */
142 int mmu_pte_psize
; /* Page size used for PTE pages */
143 int mmu_vmemmap_psize
; /* Page size used for the virtual mem map */
144 int book3e_htw_mode
; /* HW tablewalk? Value is PPC_HTW_* */
145 unsigned long linear_map_top
; /* Top of linear mapping */
149 * Number of bytes to add to SPRN_SPRG_TLB_EXFRAME on crit/mcheck/debug
150 * exceptions. This is used for bolted and e6500 TLB miss handlers which
151 * do not modify this SPRG in the TLB miss code; for other TLB miss handlers,
152 * this is set to zero.
156 #endif /* CONFIG_PPC64 */
158 #ifdef CONFIG_PPC_FSL_BOOK3E
159 /* next_tlbcam_idx is used to round-robin tlbcam entry assignment */
160 DEFINE_PER_CPU(int, next_tlbcam_idx
);
161 EXPORT_PER_CPU_SYMBOL(next_tlbcam_idx
);
165 * Base TLB flushing operations:
167 * - flush_tlb_mm(mm) flushes the specified mm context TLB's
168 * - flush_tlb_page(vma, vmaddr) flushes one page
169 * - flush_tlb_range(vma, start, end) flushes a range of pages
170 * - flush_tlb_kernel_range(start, end) flushes kernel pages
172 * - local_* variants of page and mm only apply to the current
177 * These are the base non-SMP variants of page and mm flushing
179 void local_flush_tlb_mm(struct mm_struct
*mm
)
184 pid
= mm
->context
.id
;
185 if (pid
!= MMU_NO_CONTEXT
)
189 EXPORT_SYMBOL(local_flush_tlb_mm
);
191 void __local_flush_tlb_page(struct mm_struct
*mm
, unsigned long vmaddr
,
197 pid
= mm
? mm
->context
.id
: 0;
198 if (pid
!= MMU_NO_CONTEXT
)
199 _tlbil_va(vmaddr
, pid
, tsize
, ind
);
203 void local_flush_tlb_page(struct vm_area_struct
*vma
, unsigned long vmaddr
)
205 __local_flush_tlb_page(vma
? vma
->vm_mm
: NULL
, vmaddr
,
206 mmu_get_tsize(mmu_virtual_psize
), 0);
208 EXPORT_SYMBOL(local_flush_tlb_page
);
211 * And here are the SMP non-local implementations
215 static DEFINE_RAW_SPINLOCK(tlbivax_lock
);
217 static int mm_is_core_local(struct mm_struct
*mm
)
219 return cpumask_subset(mm_cpumask(mm
),
220 topology_thread_cpumask(smp_processor_id()));
223 struct tlb_flush_param
{
230 static void do_flush_tlb_mm_ipi(void *param
)
232 struct tlb_flush_param
*p
= param
;
234 _tlbil_pid(p
? p
->pid
: 0);
237 static void do_flush_tlb_page_ipi(void *param
)
239 struct tlb_flush_param
*p
= param
;
241 _tlbil_va(p
->addr
, p
->pid
, p
->tsize
, p
->ind
);
245 /* Note on invalidations and PID:
247 * We snapshot the PID with preempt disabled. At this point, it can still
248 * change either because:
249 * - our context is being stolen (PID -> NO_CONTEXT) on another CPU
250 * - we are invaliating some target that isn't currently running here
251 * and is concurrently acquiring a new PID on another CPU
252 * - some other CPU is re-acquiring a lost PID for this mm
255 * However, this shouldn't be a problem as we only guarantee
256 * invalidation of TLB entries present prior to this call, so we
257 * don't care about the PID changing, and invalidating a stale PID
258 * is generally harmless.
261 void flush_tlb_mm(struct mm_struct
*mm
)
266 pid
= mm
->context
.id
;
267 if (unlikely(pid
== MMU_NO_CONTEXT
))
269 if (!mm_is_core_local(mm
)) {
270 struct tlb_flush_param p
= { .pid
= pid
};
271 /* Ignores smp_processor_id() even if set. */
272 smp_call_function_many(mm_cpumask(mm
),
273 do_flush_tlb_mm_ipi
, &p
, 1);
279 EXPORT_SYMBOL(flush_tlb_mm
);
281 void __flush_tlb_page(struct mm_struct
*mm
, unsigned long vmaddr
,
284 struct cpumask
*cpu_mask
;
288 * This function as well as __local_flush_tlb_page() must only be called
291 if (unlikely(WARN_ON(!mm
)))
295 pid
= mm
->context
.id
;
296 if (unlikely(pid
== MMU_NO_CONTEXT
))
298 cpu_mask
= mm_cpumask(mm
);
299 if (!mm_is_core_local(mm
)) {
300 /* If broadcast tlbivax is supported, use it */
301 if (mmu_has_feature(MMU_FTR_USE_TLBIVAX_BCAST
)) {
302 int lock
= mmu_has_feature(MMU_FTR_LOCK_BCAST_INVAL
);
304 raw_spin_lock(&tlbivax_lock
);
305 _tlbivax_bcast(vmaddr
, pid
, tsize
, ind
);
307 raw_spin_unlock(&tlbivax_lock
);
310 struct tlb_flush_param p
= {
316 /* Ignores smp_processor_id() even if set in cpu_mask */
317 smp_call_function_many(cpu_mask
,
318 do_flush_tlb_page_ipi
, &p
, 1);
321 _tlbil_va(vmaddr
, pid
, tsize
, ind
);
326 void flush_tlb_page(struct vm_area_struct
*vma
, unsigned long vmaddr
)
328 #ifdef CONFIG_HUGETLB_PAGE
329 if (vma
&& is_vm_hugetlb_page(vma
))
330 flush_hugetlb_page(vma
, vmaddr
);
333 __flush_tlb_page(vma
? vma
->vm_mm
: NULL
, vmaddr
,
334 mmu_get_tsize(mmu_virtual_psize
), 0);
336 EXPORT_SYMBOL(flush_tlb_page
);
338 #endif /* CONFIG_SMP */
340 #ifdef CONFIG_PPC_47x
341 void __init
early_init_mmu_47x(void)
344 unsigned long root
= of_get_flat_dt_root();
345 if (of_get_flat_dt_prop(root
, "cooperative-partition", NULL
))
346 mmu_clear_feature(MMU_FTR_USE_TLBIVAX_BCAST
);
347 #endif /* CONFIG_SMP */
349 #endif /* CONFIG_PPC_47x */
352 * Flush kernel TLB entries in the given range
354 void flush_tlb_kernel_range(unsigned long start
, unsigned long end
)
358 smp_call_function(do_flush_tlb_mm_ipi
, NULL
, 1);
365 EXPORT_SYMBOL(flush_tlb_kernel_range
);
368 * Currently, for range flushing, we just do a full mm flush. This should
369 * be optimized based on a threshold on the size of the range, since
370 * some implementation can stack multiple tlbivax before a tlbsync but
371 * for now, we keep it that way
373 void flush_tlb_range(struct vm_area_struct
*vma
, unsigned long start
,
377 flush_tlb_mm(vma
->vm_mm
);
379 EXPORT_SYMBOL(flush_tlb_range
);
381 void tlb_flush(struct mmu_gather
*tlb
)
383 flush_tlb_mm(tlb
->mm
);
387 * Below are functions specific to the 64-bit variant of Book3E though that
388 * may change in the future
394 * Handling of virtual linear page tables or indirect TLB entries
395 * flushing when PTE pages are freed
397 void tlb_flush_pgtable(struct mmu_gather
*tlb
, unsigned long address
)
399 int tsize
= mmu_psize_defs
[mmu_pte_psize
].enc
;
401 if (book3e_htw_mode
!= PPC_HTW_NONE
) {
402 unsigned long start
= address
& PMD_MASK
;
403 unsigned long end
= address
+ PMD_SIZE
;
404 unsigned long size
= 1UL << mmu_psize_defs
[mmu_pte_psize
].shift
;
406 /* This isn't the most optimal, ideally we would factor out the
407 * while preempt & CPU mask mucking around, or even the IPI but
410 while (start
< end
) {
411 __flush_tlb_page(tlb
->mm
, start
, tsize
, 1);
415 unsigned long rmask
= 0xf000000000000000ul
;
416 unsigned long rid
= (address
& rmask
) | 0x1000000000000000ul
;
417 unsigned long vpte
= address
& ~rmask
;
419 #ifdef CONFIG_PPC_64K_PAGES
420 vpte
= (vpte
>> (PAGE_SHIFT
- 4)) & ~0xfffful
;
422 vpte
= (vpte
>> (PAGE_SHIFT
- 3)) & ~0xffful
;
425 __flush_tlb_page(tlb
->mm
, vpte
, tsize
, 0);
429 static void setup_page_sizes(void)
431 unsigned int tlb0cfg
;
436 #ifdef CONFIG_PPC_FSL_BOOK3E
437 unsigned int mmucfg
= mfspr(SPRN_MMUCFG
);
438 int fsl_mmu
= mmu_has_feature(MMU_FTR_TYPE_FSL_E
);
440 if (fsl_mmu
&& (mmucfg
& MMUCFG_MAVN
) == MMUCFG_MAVN_V1
) {
441 unsigned int tlb1cfg
= mfspr(SPRN_TLB1CFG
);
442 unsigned int min_pg
, max_pg
;
444 min_pg
= (tlb1cfg
& TLBnCFG_MINSIZE
) >> TLBnCFG_MINSIZE_SHIFT
;
445 max_pg
= (tlb1cfg
& TLBnCFG_MAXSIZE
) >> TLBnCFG_MAXSIZE_SHIFT
;
447 for (psize
= 0; psize
< MMU_PAGE_COUNT
; ++psize
) {
448 struct mmu_psize_def
*def
;
451 def
= &mmu_psize_defs
[psize
];
454 if (shift
== 0 || shift
& 1)
457 /* adjust to be in terms of 4^shift Kb */
458 shift
= (shift
- 10) >> 1;
460 if ((shift
>= min_pg
) && (shift
<= max_pg
))
461 def
->flags
|= MMU_PAGE_SIZE_DIRECT
;
467 if (fsl_mmu
&& (mmucfg
& MMUCFG_MAVN
) == MMUCFG_MAVN_V2
) {
470 tlb0cfg
= mfspr(SPRN_TLB0CFG
);
471 tlb1cfg
= mfspr(SPRN_TLB1CFG
);
472 tlb1ps
= mfspr(SPRN_TLB1PS
);
473 eptcfg
= mfspr(SPRN_EPTCFG
);
475 if ((tlb1cfg
& TLBnCFG_IND
) && (tlb0cfg
& TLBnCFG_PT
))
476 book3e_htw_mode
= PPC_HTW_E6500
;
479 * We expect 4K subpage size and unrestricted indirect size.
480 * The lack of a restriction on indirect size is a Freescale
481 * extension, indicated by PSn = 0 but SPSn != 0.
484 book3e_htw_mode
= PPC_HTW_NONE
;
486 for (psize
= 0; psize
< MMU_PAGE_COUNT
; ++psize
) {
487 struct mmu_psize_def
*def
= &mmu_psize_defs
[psize
];
489 if (tlb1ps
& (1U << (def
->shift
- 10))) {
490 def
->flags
|= MMU_PAGE_SIZE_DIRECT
;
492 if (book3e_htw_mode
&& psize
== MMU_PAGE_2M
)
493 def
->flags
|= MMU_PAGE_SIZE_INDIRECT
;
501 tlb0cfg
= mfspr(SPRN_TLB0CFG
);
502 tlb0ps
= mfspr(SPRN_TLB0PS
);
503 eptcfg
= mfspr(SPRN_EPTCFG
);
505 /* Look for supported direct sizes */
506 for (psize
= 0; psize
< MMU_PAGE_COUNT
; ++psize
) {
507 struct mmu_psize_def
*def
= &mmu_psize_defs
[psize
];
509 if (tlb0ps
& (1U << (def
->shift
- 10)))
510 def
->flags
|= MMU_PAGE_SIZE_DIRECT
;
513 /* Indirect page sizes supported ? */
514 if ((tlb0cfg
& TLBnCFG_IND
) == 0 ||
515 (tlb0cfg
& TLBnCFG_PT
) == 0)
518 book3e_htw_mode
= PPC_HTW_IBM
;
520 /* Now, we only deal with one IND page size for each
521 * direct size. Hopefully all implementations today are
522 * unambiguous, but we might want to be careful in the
525 for (i
= 0; i
< 3; i
++) {
526 unsigned int ps
, sps
;
534 for (psize
= 0; psize
< MMU_PAGE_COUNT
; psize
++) {
535 struct mmu_psize_def
*def
= &mmu_psize_defs
[psize
];
537 if (ps
== (def
->shift
- 10))
538 def
->flags
|= MMU_PAGE_SIZE_INDIRECT
;
539 if (sps
== (def
->shift
- 10))
545 /* Cleanup array and print summary */
546 pr_info("MMU: Supported page sizes\n");
547 for (psize
= 0; psize
< MMU_PAGE_COUNT
; ++psize
) {
548 struct mmu_psize_def
*def
= &mmu_psize_defs
[psize
];
549 const char *__page_type_names
[] = {
555 if (def
->flags
== 0) {
559 pr_info(" %8ld KB as %s\n", 1ul << (def
->shift
- 10),
560 __page_type_names
[def
->flags
& 0x3]);
564 static void setup_mmu_htw(void)
567 * If we want to use HW tablewalk, enable it by patching the TLB miss
568 * handlers to branch to the one dedicated to it.
571 switch (book3e_htw_mode
) {
573 patch_exception(0x1c0, exc_data_tlb_miss_htw_book3e
);
574 patch_exception(0x1e0, exc_instruction_tlb_miss_htw_book3e
);
576 #ifdef CONFIG_PPC_FSL_BOOK3E
578 extlb_level_exc
= EX_TLB_SIZE
;
579 patch_exception(0x1c0, exc_data_tlb_miss_e6500_book3e
);
580 patch_exception(0x1e0, exc_instruction_tlb_miss_e6500_book3e
);
584 pr_info("MMU: Book3E HW tablewalk %s\n",
585 book3e_htw_mode
!= PPC_HTW_NONE
? "enabled" : "not supported");
589 * Early initialization of the MMU TLB code
591 static void early_init_this_mmu(void)
595 /* Set MAS4 based on page table setting */
597 mas4
= 0x4 << MAS4_WIMGED_SHIFT
;
598 switch (book3e_htw_mode
) {
601 mas4
|= BOOK3E_PAGESZ_2M
<< MAS4_TSIZED_SHIFT
;
602 mas4
|= MAS4_TLBSELD(1);
603 mmu_pte_psize
= MMU_PAGE_2M
;
608 #ifdef CONFIG_PPC_64K_PAGES
609 mas4
|= BOOK3E_PAGESZ_256M
<< MAS4_TSIZED_SHIFT
;
610 mmu_pte_psize
= MMU_PAGE_256M
;
612 mas4
|= BOOK3E_PAGESZ_1M
<< MAS4_TSIZED_SHIFT
;
613 mmu_pte_psize
= MMU_PAGE_1M
;
618 #ifdef CONFIG_PPC_64K_PAGES
619 mas4
|= BOOK3E_PAGESZ_64K
<< MAS4_TSIZED_SHIFT
;
621 mas4
|= BOOK3E_PAGESZ_4K
<< MAS4_TSIZED_SHIFT
;
623 mmu_pte_psize
= mmu_virtual_psize
;
626 mtspr(SPRN_MAS4
, mas4
);
628 #ifdef CONFIG_PPC_FSL_BOOK3E
629 if (mmu_has_feature(MMU_FTR_TYPE_FSL_E
)) {
630 unsigned int num_cams
;
632 /* use a quarter of the TLBCAM for bolted linear map */
633 num_cams
= (mfspr(SPRN_TLB1CFG
) & TLBnCFG_N_ENTRY
) / 4;
634 linear_map_top
= map_mem_in_cams(linear_map_top
, num_cams
);
638 /* A sync won't hurt us after mucking around with
639 * the MMU configuration
644 static void __init
early_init_mmu_global(void)
646 /* XXX This will have to be decided at runtime, but right
647 * now our boot and TLB miss code hard wires it. Ideally
648 * we should find out a suitable page size and patch the
649 * TLB miss code (either that or use the PACA to store
652 mmu_linear_psize
= MMU_PAGE_1G
;
654 /* XXX This should be decided at runtime based on supported
655 * page sizes in the TLB, but for now let's assume 16M is
656 * always there and a good fit (which it probably is)
658 * Freescale booke only supports 4K pages in TLB0, so use that.
660 if (mmu_has_feature(MMU_FTR_TYPE_FSL_E
))
661 mmu_vmemmap_psize
= MMU_PAGE_4K
;
663 mmu_vmemmap_psize
= MMU_PAGE_16M
;
665 /* XXX This code only checks for TLB 0 capabilities and doesn't
666 * check what page size combos are supported by the HW. It
667 * also doesn't handle the case where a separate array holds
668 * the IND entries from the array loaded by the PT.
670 /* Look for supported page sizes */
673 /* Look for HW tablewalk support */
676 #ifdef CONFIG_PPC_FSL_BOOK3E
677 if (mmu_has_feature(MMU_FTR_TYPE_FSL_E
)) {
678 if (book3e_htw_mode
== PPC_HTW_NONE
) {
679 extlb_level_exc
= EX_TLB_SIZE
;
680 patch_exception(0x1c0, exc_data_tlb_miss_bolted_book3e
);
681 patch_exception(0x1e0,
682 exc_instruction_tlb_miss_bolted_book3e
);
687 /* Set the global containing the top of the linear mapping
688 * for use by the TLB miss code
690 linear_map_top
= memblock_end_of_DRAM();
693 static void __init
early_mmu_set_memory_limit(void)
695 #ifdef CONFIG_PPC_FSL_BOOK3E
696 if (mmu_has_feature(MMU_FTR_TYPE_FSL_E
)) {
698 * Limit memory so we dont have linear faults.
699 * Unlike memblock_set_current_limit, which limits
700 * memory available during early boot, this permanently
701 * reduces the memory available to Linux. We need to
702 * do this because highmem is not supported on 64-bit.
704 memblock_enforce_memory_limit(linear_map_top
);
708 memblock_set_current_limit(linear_map_top
);
712 void __init
early_init_mmu(void)
714 early_init_mmu_global();
715 early_init_this_mmu();
716 early_mmu_set_memory_limit();
719 void early_init_mmu_secondary(void)
721 early_init_this_mmu();
724 void setup_initial_memory_limit(phys_addr_t first_memblock_base
,
725 phys_addr_t first_memblock_size
)
727 /* On non-FSL Embedded 64-bit, we adjust the RMA size to match
728 * the bolted TLB entry. We know for now that only 1G
729 * entries are supported though that may eventually
732 * on FSL Embedded 64-bit, we adjust the RMA size to match the
733 * first bolted TLB entry size. We still limit max to 1G even if
734 * the TLB could cover more. This is due to what the early init
735 * code is setup to do.
737 * We crop it to the size of the first MEMBLOCK to
738 * avoid going over total available memory just in case...
740 #ifdef CONFIG_PPC_FSL_BOOK3E
741 if (mmu_has_feature(MMU_FTR_TYPE_FSL_E
)) {
742 unsigned long linear_sz
;
743 linear_sz
= calc_cam_sz(first_memblock_size
, PAGE_OFFSET
,
744 first_memblock_base
);
745 ppc64_rma_size
= min_t(u64
, linear_sz
, 0x40000000);
748 ppc64_rma_size
= min_t(u64
, first_memblock_size
, 0x40000000);
750 /* Finally limit subsequent allocations */
751 memblock_set_current_limit(first_memblock_base
+ ppc64_rma_size
);
753 #else /* ! CONFIG_PPC64 */
754 void __init
early_init_mmu(void)
756 #ifdef CONFIG_PPC_47x
757 early_init_mmu_47x();
760 #endif /* CONFIG_PPC64 */