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)
13 * Derived from arch/ppc/mm/init.c:
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/cputhreads.h>
46 #include <asm/hugetlb.h>
52 * This struct lists the sw-supported page sizes. The hardawre MMU may support
53 * other sizes not listed here. The .ind field is only used on MMUs that have
54 * indirect page table entries.
56 #if defined(CONFIG_PPC_BOOK3E_MMU) || defined(CONFIG_PPC_8xx)
57 #ifdef CONFIG_PPC_FSL_BOOK3E
58 struct mmu_psize_def mmu_psize_defs[MMU_PAGE_COUNT] = {
61 .enc = BOOK3E_PAGESZ_4K,
65 .enc = BOOK3E_PAGESZ_2M,
69 .enc = BOOK3E_PAGESZ_4M,
73 .enc = BOOK3E_PAGESZ_16M,
77 .enc = BOOK3E_PAGESZ_64M,
81 .enc = BOOK3E_PAGESZ_256M,
85 .enc = BOOK3E_PAGESZ_1GB,
88 #elif defined(CONFIG_PPC_8xx)
89 struct mmu_psize_def mmu_psize_defs[MMU_PAGE_COUNT] = {
90 /* we only manage 4k and 16k pages as normal pages */
91 #ifdef CONFIG_PPC_4K_PAGES
108 struct mmu_psize_def mmu_psize_defs[MMU_PAGE_COUNT] = {
112 .enc = BOOK3E_PAGESZ_4K,
116 .enc = BOOK3E_PAGESZ_16K,
121 .enc = BOOK3E_PAGESZ_64K,
125 .enc = BOOK3E_PAGESZ_1M,
130 .enc = BOOK3E_PAGESZ_16M,
134 .enc = BOOK3E_PAGESZ_256M,
138 .enc = BOOK3E_PAGESZ_1GB,
141 #endif /* CONFIG_FSL_BOOKE */
143 static inline int mmu_get_tsize(int psize)
145 return mmu_psize_defs[psize].enc;
148 static inline int mmu_get_tsize(int psize)
150 /* This isn't used on !Book3E for now */
153 #endif /* CONFIG_PPC_BOOK3E_MMU */
155 /* The variables below are currently only used on 64-bit Book3E
156 * though this will probably be made common with other nohash
157 * implementations at some point
161 int mmu_linear_psize; /* Page size used for the linear mapping */
162 int mmu_pte_psize; /* Page size used for PTE pages */
163 int mmu_vmemmap_psize; /* Page size used for the virtual mem map */
164 int book3e_htw_mode; /* HW tablewalk? Value is PPC_HTW_* */
165 unsigned long linear_map_top; /* Top of linear mapping */
169 * Number of bytes to add to SPRN_SPRG_TLB_EXFRAME on crit/mcheck/debug
170 * exceptions. This is used for bolted and e6500 TLB miss handlers which
171 * do not modify this SPRG in the TLB miss code; for other TLB miss handlers,
172 * this is set to zero.
176 #endif /* CONFIG_PPC64 */
178 #ifdef CONFIG_PPC_FSL_BOOK3E
179 /* next_tlbcam_idx is used to round-robin tlbcam entry assignment */
180 DEFINE_PER_CPU(int, next_tlbcam_idx);
181 EXPORT_PER_CPU_SYMBOL(next_tlbcam_idx);
185 * Base TLB flushing operations:
187 * - flush_tlb_mm(mm) flushes the specified mm context TLB's
188 * - flush_tlb_page(vma, vmaddr) flushes one page
189 * - flush_tlb_range(vma, start, end) flushes a range of pages
190 * - flush_tlb_kernel_range(start, end) flushes kernel pages
192 * - local_* variants of page and mm only apply to the current
197 * These are the base non-SMP variants of page and mm flushing
199 void local_flush_tlb_mm(struct mm_struct *mm)
204 pid = mm->context.id;
205 if (pid != MMU_NO_CONTEXT)
209 EXPORT_SYMBOL(local_flush_tlb_mm);
211 void __local_flush_tlb_page(struct mm_struct *mm, unsigned long vmaddr,
217 pid = mm ? mm->context.id : 0;
218 if (pid != MMU_NO_CONTEXT)
219 _tlbil_va(vmaddr, pid, tsize, ind);
223 void local_flush_tlb_page(struct vm_area_struct *vma, unsigned long vmaddr)
225 __local_flush_tlb_page(vma ? vma->vm_mm : NULL, vmaddr,
226 mmu_get_tsize(mmu_virtual_psize), 0);
228 EXPORT_SYMBOL(local_flush_tlb_page);
231 * And here are the SMP non-local implementations
235 static DEFINE_RAW_SPINLOCK(tlbivax_lock);
237 struct tlb_flush_param {
244 static void do_flush_tlb_mm_ipi(void *param)
246 struct tlb_flush_param *p = param;
248 _tlbil_pid(p ? p->pid : 0);
251 static void do_flush_tlb_page_ipi(void *param)
253 struct tlb_flush_param *p = param;
255 _tlbil_va(p->addr, p->pid, p->tsize, p->ind);
259 /* Note on invalidations and PID:
261 * We snapshot the PID with preempt disabled. At this point, it can still
262 * change either because:
263 * - our context is being stolen (PID -> NO_CONTEXT) on another CPU
264 * - we are invaliating some target that isn't currently running here
265 * and is concurrently acquiring a new PID on another CPU
266 * - some other CPU is re-acquiring a lost PID for this mm
269 * However, this shouldn't be a problem as we only guarantee
270 * invalidation of TLB entries present prior to this call, so we
271 * don't care about the PID changing, and invalidating a stale PID
272 * is generally harmless.
275 void flush_tlb_mm(struct mm_struct *mm)
280 pid = mm->context.id;
281 if (unlikely(pid == MMU_NO_CONTEXT))
283 if (!mm_is_core_local(mm)) {
284 struct tlb_flush_param p = { .pid = pid };
285 /* Ignores smp_processor_id() even if set. */
286 smp_call_function_many(mm_cpumask(mm),
287 do_flush_tlb_mm_ipi, &p, 1);
293 EXPORT_SYMBOL(flush_tlb_mm);
295 void __flush_tlb_page(struct mm_struct *mm, unsigned long vmaddr,
298 struct cpumask *cpu_mask;
302 * This function as well as __local_flush_tlb_page() must only be called
305 if (unlikely(WARN_ON(!mm)))
309 pid = mm->context.id;
310 if (unlikely(pid == MMU_NO_CONTEXT))
312 cpu_mask = mm_cpumask(mm);
313 if (!mm_is_core_local(mm)) {
314 /* If broadcast tlbivax is supported, use it */
315 if (mmu_has_feature(MMU_FTR_USE_TLBIVAX_BCAST)) {
316 int lock = mmu_has_feature(MMU_FTR_LOCK_BCAST_INVAL);
318 raw_spin_lock(&tlbivax_lock);
319 _tlbivax_bcast(vmaddr, pid, tsize, ind);
321 raw_spin_unlock(&tlbivax_lock);
324 struct tlb_flush_param p = {
330 /* Ignores smp_processor_id() even if set in cpu_mask */
331 smp_call_function_many(cpu_mask,
332 do_flush_tlb_page_ipi, &p, 1);
335 _tlbil_va(vmaddr, pid, tsize, ind);
340 void flush_tlb_page(struct vm_area_struct *vma, unsigned long vmaddr)
342 #ifdef CONFIG_HUGETLB_PAGE
343 if (vma && is_vm_hugetlb_page(vma))
344 flush_hugetlb_page(vma, vmaddr);
347 __flush_tlb_page(vma ? vma->vm_mm : NULL, vmaddr,
348 mmu_get_tsize(mmu_virtual_psize), 0);
350 EXPORT_SYMBOL(flush_tlb_page);
352 #endif /* CONFIG_SMP */
354 #ifdef CONFIG_PPC_47x
355 void __init early_init_mmu_47x(void)
358 unsigned long root = of_get_flat_dt_root();
359 if (of_get_flat_dt_prop(root, "cooperative-partition", NULL))
360 mmu_clear_feature(MMU_FTR_USE_TLBIVAX_BCAST);
361 #endif /* CONFIG_SMP */
363 #endif /* CONFIG_PPC_47x */
366 * Flush kernel TLB entries in the given range
368 void flush_tlb_kernel_range(unsigned long start, unsigned long end)
372 smp_call_function(do_flush_tlb_mm_ipi, NULL, 1);
379 EXPORT_SYMBOL(flush_tlb_kernel_range);
382 * Currently, for range flushing, we just do a full mm flush. This should
383 * be optimized based on a threshold on the size of the range, since
384 * some implementation can stack multiple tlbivax before a tlbsync but
385 * for now, we keep it that way
387 void flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
391 flush_tlb_mm(vma->vm_mm);
393 EXPORT_SYMBOL(flush_tlb_range);
395 void tlb_flush(struct mmu_gather *tlb)
397 flush_tlb_mm(tlb->mm);
401 * Below are functions specific to the 64-bit variant of Book3E though that
402 * may change in the future
408 * Handling of virtual linear page tables or indirect TLB entries
409 * flushing when PTE pages are freed
411 void tlb_flush_pgtable(struct mmu_gather *tlb, unsigned long address)
413 int tsize = mmu_psize_defs[mmu_pte_psize].enc;
415 if (book3e_htw_mode != PPC_HTW_NONE) {
416 unsigned long start = address & PMD_MASK;
417 unsigned long end = address + PMD_SIZE;
418 unsigned long size = 1UL << mmu_psize_defs[mmu_pte_psize].shift;
420 /* This isn't the most optimal, ideally we would factor out the
421 * while preempt & CPU mask mucking around, or even the IPI but
424 while (start < end) {
425 __flush_tlb_page(tlb->mm, start, tsize, 1);
429 unsigned long rmask = 0xf000000000000000ul;
430 unsigned long rid = (address & rmask) | 0x1000000000000000ul;
431 unsigned long vpte = address & ~rmask;
433 #ifdef CONFIG_PPC_64K_PAGES
434 vpte = (vpte >> (PAGE_SHIFT - 4)) & ~0xfffful;
436 vpte = (vpte >> (PAGE_SHIFT - 3)) & ~0xffful;
439 __flush_tlb_page(tlb->mm, vpte, tsize, 0);
443 static void setup_page_sizes(void)
445 unsigned int tlb0cfg;
450 #ifdef CONFIG_PPC_FSL_BOOK3E
451 unsigned int mmucfg = mfspr(SPRN_MMUCFG);
452 int fsl_mmu = mmu_has_feature(MMU_FTR_TYPE_FSL_E);
454 if (fsl_mmu && (mmucfg & MMUCFG_MAVN) == MMUCFG_MAVN_V1) {
455 unsigned int tlb1cfg = mfspr(SPRN_TLB1CFG);
456 unsigned int min_pg, max_pg;
458 min_pg = (tlb1cfg & TLBnCFG_MINSIZE) >> TLBnCFG_MINSIZE_SHIFT;
459 max_pg = (tlb1cfg & TLBnCFG_MAXSIZE) >> TLBnCFG_MAXSIZE_SHIFT;
461 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
462 struct mmu_psize_def *def;
465 def = &mmu_psize_defs[psize];
468 if (shift == 0 || shift & 1)
471 /* adjust to be in terms of 4^shift Kb */
472 shift = (shift - 10) >> 1;
474 if ((shift >= min_pg) && (shift <= max_pg))
475 def->flags |= MMU_PAGE_SIZE_DIRECT;
481 if (fsl_mmu && (mmucfg & MMUCFG_MAVN) == MMUCFG_MAVN_V2) {
484 tlb0cfg = mfspr(SPRN_TLB0CFG);
485 tlb1cfg = mfspr(SPRN_TLB1CFG);
486 tlb1ps = mfspr(SPRN_TLB1PS);
487 eptcfg = mfspr(SPRN_EPTCFG);
489 if ((tlb1cfg & TLBnCFG_IND) && (tlb0cfg & TLBnCFG_PT))
490 book3e_htw_mode = PPC_HTW_E6500;
493 * We expect 4K subpage size and unrestricted indirect size.
494 * The lack of a restriction on indirect size is a Freescale
495 * extension, indicated by PSn = 0 but SPSn != 0.
498 book3e_htw_mode = PPC_HTW_NONE;
500 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
501 struct mmu_psize_def *def = &mmu_psize_defs[psize];
503 if (tlb1ps & (1U << (def->shift - 10))) {
504 def->flags |= MMU_PAGE_SIZE_DIRECT;
506 if (book3e_htw_mode && psize == MMU_PAGE_2M)
507 def->flags |= MMU_PAGE_SIZE_INDIRECT;
515 tlb0cfg = mfspr(SPRN_TLB0CFG);
516 tlb0ps = mfspr(SPRN_TLB0PS);
517 eptcfg = mfspr(SPRN_EPTCFG);
519 /* Look for supported direct sizes */
520 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
521 struct mmu_psize_def *def = &mmu_psize_defs[psize];
523 if (tlb0ps & (1U << (def->shift - 10)))
524 def->flags |= MMU_PAGE_SIZE_DIRECT;
527 /* Indirect page sizes supported ? */
528 if ((tlb0cfg & TLBnCFG_IND) == 0 ||
529 (tlb0cfg & TLBnCFG_PT) == 0)
532 book3e_htw_mode = PPC_HTW_IBM;
534 /* Now, we only deal with one IND page size for each
535 * direct size. Hopefully all implementations today are
536 * unambiguous, but we might want to be careful in the
539 for (i = 0; i < 3; i++) {
540 unsigned int ps, sps;
548 for (psize = 0; psize < MMU_PAGE_COUNT; psize++) {
549 struct mmu_psize_def *def = &mmu_psize_defs[psize];
551 if (ps == (def->shift - 10))
552 def->flags |= MMU_PAGE_SIZE_INDIRECT;
553 if (sps == (def->shift - 10))
559 /* Cleanup array and print summary */
560 pr_info("MMU: Supported page sizes\n");
561 for (psize = 0; psize < MMU_PAGE_COUNT; ++psize) {
562 struct mmu_psize_def *def = &mmu_psize_defs[psize];
563 const char *__page_type_names[] = {
569 if (def->flags == 0) {
573 pr_info(" %8ld KB as %s\n", 1ul << (def->shift - 10),
574 __page_type_names[def->flags & 0x3]);
578 static void setup_mmu_htw(void)
581 * If we want to use HW tablewalk, enable it by patching the TLB miss
582 * handlers to branch to the one dedicated to it.
585 switch (book3e_htw_mode) {
587 patch_exception(0x1c0, exc_data_tlb_miss_htw_book3e);
588 patch_exception(0x1e0, exc_instruction_tlb_miss_htw_book3e);
590 #ifdef CONFIG_PPC_FSL_BOOK3E
592 extlb_level_exc = EX_TLB_SIZE;
593 patch_exception(0x1c0, exc_data_tlb_miss_e6500_book3e);
594 patch_exception(0x1e0, exc_instruction_tlb_miss_e6500_book3e);
598 pr_info("MMU: Book3E HW tablewalk %s\n",
599 book3e_htw_mode != PPC_HTW_NONE ? "enabled" : "not supported");
603 * Early initialization of the MMU TLB code
605 static void early_init_this_mmu(void)
609 /* Set MAS4 based on page table setting */
611 mas4 = 0x4 << MAS4_WIMGED_SHIFT;
612 switch (book3e_htw_mode) {
615 mas4 |= BOOK3E_PAGESZ_2M << MAS4_TSIZED_SHIFT;
616 mas4 |= MAS4_TLBSELD(1);
617 mmu_pte_psize = MMU_PAGE_2M;
622 #ifdef CONFIG_PPC_64K_PAGES
623 mas4 |= BOOK3E_PAGESZ_256M << MAS4_TSIZED_SHIFT;
624 mmu_pte_psize = MMU_PAGE_256M;
626 mas4 |= BOOK3E_PAGESZ_1M << MAS4_TSIZED_SHIFT;
627 mmu_pte_psize = MMU_PAGE_1M;
632 #ifdef CONFIG_PPC_64K_PAGES
633 mas4 |= BOOK3E_PAGESZ_64K << MAS4_TSIZED_SHIFT;
635 mas4 |= BOOK3E_PAGESZ_4K << MAS4_TSIZED_SHIFT;
637 mmu_pte_psize = mmu_virtual_psize;
640 mtspr(SPRN_MAS4, mas4);
642 #ifdef CONFIG_PPC_FSL_BOOK3E
643 if (mmu_has_feature(MMU_FTR_TYPE_FSL_E)) {
644 unsigned int num_cams;
645 int __maybe_unused cpu = smp_processor_id();
648 /* use a quarter of the TLBCAM for bolted linear map */
649 num_cams = (mfspr(SPRN_TLB1CFG) & TLBnCFG_N_ENTRY) / 4;
652 * Only do the mapping once per core, or else the
653 * transient mapping would cause problems.
656 if (hweight32(get_tensr()) > 1)
661 linear_map_top = map_mem_in_cams(linear_map_top,
666 /* A sync won't hurt us after mucking around with
667 * the MMU configuration
672 static void __init early_init_mmu_global(void)
674 /* XXX This will have to be decided at runtime, but right
675 * now our boot and TLB miss code hard wires it. Ideally
676 * we should find out a suitable page size and patch the
677 * TLB miss code (either that or use the PACA to store
680 mmu_linear_psize = MMU_PAGE_1G;
682 /* XXX This should be decided at runtime based on supported
683 * page sizes in the TLB, but for now let's assume 16M is
684 * always there and a good fit (which it probably is)
686 * Freescale booke only supports 4K pages in TLB0, so use that.
688 if (mmu_has_feature(MMU_FTR_TYPE_FSL_E))
689 mmu_vmemmap_psize = MMU_PAGE_4K;
691 mmu_vmemmap_psize = MMU_PAGE_16M;
693 /* XXX This code only checks for TLB 0 capabilities and doesn't
694 * check what page size combos are supported by the HW. It
695 * also doesn't handle the case where a separate array holds
696 * the IND entries from the array loaded by the PT.
698 /* Look for supported page sizes */
701 /* Look for HW tablewalk support */
704 #ifdef CONFIG_PPC_FSL_BOOK3E
705 if (mmu_has_feature(MMU_FTR_TYPE_FSL_E)) {
706 if (book3e_htw_mode == PPC_HTW_NONE) {
707 extlb_level_exc = EX_TLB_SIZE;
708 patch_exception(0x1c0, exc_data_tlb_miss_bolted_book3e);
709 patch_exception(0x1e0,
710 exc_instruction_tlb_miss_bolted_book3e);
715 /* Set the global containing the top of the linear mapping
716 * for use by the TLB miss code
718 linear_map_top = memblock_end_of_DRAM();
721 static void __init early_mmu_set_memory_limit(void)
723 #ifdef CONFIG_PPC_FSL_BOOK3E
724 if (mmu_has_feature(MMU_FTR_TYPE_FSL_E)) {
726 * Limit memory so we dont have linear faults.
727 * Unlike memblock_set_current_limit, which limits
728 * memory available during early boot, this permanently
729 * reduces the memory available to Linux. We need to
730 * do this because highmem is not supported on 64-bit.
732 memblock_enforce_memory_limit(linear_map_top);
736 memblock_set_current_limit(linear_map_top);
740 void __init early_init_mmu(void)
742 early_init_mmu_global();
743 early_init_this_mmu();
744 early_mmu_set_memory_limit();
747 void early_init_mmu_secondary(void)
749 early_init_this_mmu();
752 void setup_initial_memory_limit(phys_addr_t first_memblock_base,
753 phys_addr_t first_memblock_size)
755 /* On non-FSL Embedded 64-bit, we adjust the RMA size to match
756 * the bolted TLB entry. We know for now that only 1G
757 * entries are supported though that may eventually
760 * on FSL Embedded 64-bit, usually all RAM is bolted, but with
761 * unusual memory sizes it's possible for some RAM to not be mapped
762 * (such RAM is not used at all by Linux, since we don't support
763 * highmem on 64-bit). We limit ppc64_rma_size to what would be
764 * mappable if this memblock is the only one. Additional memblocks
765 * can only increase, not decrease, the amount that ends up getting
766 * mapped. We still limit max to 1G even if we'll eventually map
767 * more. This is due to what the early init code is set up to do.
769 * We crop it to the size of the first MEMBLOCK to
770 * avoid going over total available memory just in case...
772 #ifdef CONFIG_PPC_FSL_BOOK3E
773 if (mmu_has_feature(MMU_FTR_TYPE_FSL_E)) {
774 unsigned long linear_sz;
775 unsigned int num_cams;
777 /* use a quarter of the TLBCAM for bolted linear map */
778 num_cams = (mfspr(SPRN_TLB1CFG) & TLBnCFG_N_ENTRY) / 4;
780 linear_sz = map_mem_in_cams(first_memblock_size, num_cams,
783 ppc64_rma_size = min_t(u64, linear_sz, 0x40000000);
786 ppc64_rma_size = min_t(u64, first_memblock_size, 0x40000000);
788 /* Finally limit subsequent allocations */
789 memblock_set_current_limit(first_memblock_base + ppc64_rma_size);
791 #else /* ! CONFIG_PPC64 */
792 void __init early_init_mmu(void)
794 #ifdef CONFIG_PPC_47x
795 early_init_mmu_47x();
798 #endif /* CONFIG_PPC64 */