1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * PowerPC64 port by Mike Corrigan and Dave Engebretsen
4 * {mikejc|engebret}@us.ibm.com
6 * Copyright (c) 2000 Mike Corrigan <mikejc@us.ibm.com>
8 * SMP scalability work:
9 * Copyright (C) 2001 Anton Blanchard <anton@au.ibm.com>, IBM
14 * PowerPC Hashed Page Table functions
20 #define pr_fmt(fmt) "hash-mmu: " fmt
21 #include <linux/spinlock.h>
22 #include <linux/errno.h>
23 #include <linux/sched/mm.h>
24 #include <linux/proc_fs.h>
25 #include <linux/stat.h>
26 #include <linux/sysctl.h>
27 #include <linux/export.h>
28 #include <linux/ctype.h>
29 #include <linux/cache.h>
30 #include <linux/init.h>
31 #include <linux/signal.h>
32 #include <linux/memblock.h>
33 #include <linux/context_tracking.h>
34 #include <linux/libfdt.h>
35 #include <linux/pkeys.h>
36 #include <linux/hugetlb.h>
37 #include <linux/cpu.h>
38 #include <linux/pgtable.h>
40 #include <asm/debugfs.h>
41 #include <asm/processor.h>
43 #include <asm/mmu_context.h>
45 #include <asm/types.h>
46 #include <linux/uaccess.h>
47 #include <asm/machdep.h>
52 #include <asm/cacheflush.h>
53 #include <asm/cputable.h>
54 #include <asm/sections.h>
55 #include <asm/copro.h>
57 #include <asm/code-patching.h>
58 #include <asm/fadump.h>
59 #include <asm/firmware.h>
61 #include <asm/trace.h>
63 #include <asm/pte-walk.h>
64 #include <asm/asm-prototypes.h>
65 #include <asm/ultravisor.h>
67 #include <mm/mmu_decl.h>
73 #define DBG(fmt...) udbg_printf(fmt)
79 #define DBG_LOW(fmt...) udbg_printf(fmt)
81 #define DBG_LOW(fmt...)
89 * Note: pte --> Linux PTE
90 * HPTE --> PowerPC Hashed Page Table Entry
93 * htab_initialize is called with the MMU off (of course), but
94 * the kernel has been copied down to zero so it can directly
95 * reference global data. At this point it is very difficult
96 * to print debug info.
100 static unsigned long _SDR1;
101 struct mmu_psize_def mmu_psize_defs[MMU_PAGE_COUNT];
102 EXPORT_SYMBOL_GPL(mmu_psize_defs);
104 u8 hpte_page_sizes[1 << LP_BITS];
105 EXPORT_SYMBOL_GPL(hpte_page_sizes);
107 struct hash_pte *htab_address;
108 unsigned long htab_size_bytes;
109 unsigned long htab_hash_mask;
110 EXPORT_SYMBOL_GPL(htab_hash_mask);
111 int mmu_linear_psize = MMU_PAGE_4K;
112 EXPORT_SYMBOL_GPL(mmu_linear_psize);
113 int mmu_virtual_psize = MMU_PAGE_4K;
114 int mmu_vmalloc_psize = MMU_PAGE_4K;
115 #ifdef CONFIG_SPARSEMEM_VMEMMAP
116 int mmu_vmemmap_psize = MMU_PAGE_4K;
118 int mmu_io_psize = MMU_PAGE_4K;
119 int mmu_kernel_ssize = MMU_SEGSIZE_256M;
120 EXPORT_SYMBOL_GPL(mmu_kernel_ssize);
121 int mmu_highuser_ssize = MMU_SEGSIZE_256M;
122 u16 mmu_slb_size = 64;
123 EXPORT_SYMBOL_GPL(mmu_slb_size);
124 #ifdef CONFIG_PPC_64K_PAGES
125 int mmu_ci_restrictions;
127 #ifdef CONFIG_DEBUG_PAGEALLOC
128 static u8 *linear_map_hash_slots;
129 static unsigned long linear_map_hash_count;
130 static DEFINE_SPINLOCK(linear_map_hash_lock);
131 #endif /* CONFIG_DEBUG_PAGEALLOC */
132 struct mmu_hash_ops mmu_hash_ops;
133 EXPORT_SYMBOL(mmu_hash_ops);
136 * These are definitions of page sizes arrays to be used when none
137 * is provided by the firmware.
141 * Fallback (4k pages only)
143 static struct mmu_psize_def mmu_psize_defaults[] = {
147 .penc = {[MMU_PAGE_4K] = 0, [1 ... MMU_PAGE_COUNT - 1] = -1},
154 * POWER4, GPUL, POWER5
156 * Support for 16Mb large pages
158 static struct mmu_psize_def mmu_psize_defaults_gp[] = {
162 .penc = {[MMU_PAGE_4K] = 0, [1 ... MMU_PAGE_COUNT - 1] = -1},
169 .penc = {[0 ... MMU_PAGE_16M - 1] = -1, [MMU_PAGE_16M] = 0,
170 [MMU_PAGE_16M + 1 ... MMU_PAGE_COUNT - 1] = -1 },
177 * 'R' and 'C' update notes:
178 * - Under pHyp or KVM, the updatepp path will not set C, thus it *will*
179 * create writeable HPTEs without C set, because the hcall H_PROTECT
180 * that we use in that case will not update C
181 * - The above is however not a problem, because we also don't do that
182 * fancy "no flush" variant of eviction and we use H_REMOVE which will
183 * do the right thing and thus we don't have the race I described earlier
185 * - Under bare metal, we do have the race, so we need R and C set
186 * - We make sure R is always set and never lost
187 * - C is _PAGE_DIRTY, and *should* always be set for a writeable mapping
189 unsigned long htab_convert_pte_flags(unsigned long pteflags)
191 unsigned long rflags = 0;
193 /* _PAGE_EXEC -> NOEXEC */
194 if ((pteflags & _PAGE_EXEC) == 0)
198 * Linux uses slb key 0 for kernel and 1 for user.
199 * kernel RW areas are mapped with PPP=0b000
200 * User area is mapped with PPP=0b010 for read/write
201 * or PPP=0b011 for read-only (including writeable but clean pages).
203 if (pteflags & _PAGE_PRIVILEGED) {
205 * Kernel read only mapped with ppp bits 0b110
207 if (!(pteflags & _PAGE_WRITE)) {
208 if (mmu_has_feature(MMU_FTR_KERNEL_RO))
209 rflags |= (HPTE_R_PP0 | 0x2);
214 if (pteflags & _PAGE_RWX)
216 if (!((pteflags & _PAGE_WRITE) && (pteflags & _PAGE_DIRTY)))
220 * We can't allow hardware to update hpte bits. Hence always
221 * set 'R' bit and set 'C' if it is a write fault
225 if (pteflags & _PAGE_DIRTY)
231 if ((pteflags & _PAGE_CACHE_CTL) == _PAGE_TOLERANT)
233 else if ((pteflags & _PAGE_CACHE_CTL) == _PAGE_NON_IDEMPOTENT)
234 rflags |= (HPTE_R_I | HPTE_R_G);
235 else if ((pteflags & _PAGE_CACHE_CTL) == _PAGE_SAO)
236 rflags |= (HPTE_R_W | HPTE_R_I | HPTE_R_M);
239 * Add memory coherence if cache inhibited is not set
243 rflags |= pte_to_hpte_pkey_bits(pteflags);
247 int htab_bolt_mapping(unsigned long vstart, unsigned long vend,
248 unsigned long pstart, unsigned long prot,
249 int psize, int ssize)
251 unsigned long vaddr, paddr;
252 unsigned int step, shift;
255 shift = mmu_psize_defs[psize].shift;
258 prot = htab_convert_pte_flags(prot);
260 DBG("htab_bolt_mapping(%lx..%lx -> %lx (%lx,%d,%d)\n",
261 vstart, vend, pstart, prot, psize, ssize);
263 /* Carefully map only the possible range */
264 vaddr = ALIGN(vstart, step);
265 paddr = ALIGN(pstart, step);
266 vend = ALIGN_DOWN(vend, step);
268 for (; vaddr < vend; vaddr += step, paddr += step) {
269 unsigned long hash, hpteg;
270 unsigned long vsid = get_kernel_vsid(vaddr, ssize);
271 unsigned long vpn = hpt_vpn(vaddr, vsid, ssize);
272 unsigned long tprot = prot;
273 bool secondary_hash = false;
276 * If we hit a bad address return error.
280 /* Make kernel text executable */
281 if (overlaps_kernel_text(vaddr, vaddr + step))
285 * If relocatable, check if it overlaps interrupt vectors that
286 * are copied down to real 0. For relocatable kernel
287 * (e.g. kdump case) we copy interrupt vectors down to real
288 * address 0. Mark that region as executable. This is
289 * because on p8 system with relocation on exception feature
290 * enabled, exceptions are raised with MMU (IR=DR=1) ON. Hence
291 * in order to execute the interrupt handlers in virtual
292 * mode the vector region need to be marked as executable.
294 if ((PHYSICAL_START > MEMORY_START) &&
295 overlaps_interrupt_vector_text(vaddr, vaddr + step))
298 hash = hpt_hash(vpn, shift, ssize);
299 hpteg = ((hash & htab_hash_mask) * HPTES_PER_GROUP);
301 BUG_ON(!mmu_hash_ops.hpte_insert);
303 ret = mmu_hash_ops.hpte_insert(hpteg, vpn, paddr, tprot,
304 HPTE_V_BOLTED, psize, psize,
308 * Try to to keep bolted entries in primary.
309 * Remove non bolted entries and try insert again
311 ret = mmu_hash_ops.hpte_remove(hpteg);
313 ret = mmu_hash_ops.hpte_insert(hpteg, vpn, paddr, tprot,
314 HPTE_V_BOLTED, psize, psize,
316 if (ret == -1 && !secondary_hash) {
317 secondary_hash = true;
318 hpteg = ((~hash & htab_hash_mask) * HPTES_PER_GROUP);
327 #ifdef CONFIG_DEBUG_PAGEALLOC
328 if (debug_pagealloc_enabled() &&
329 (paddr >> PAGE_SHIFT) < linear_map_hash_count)
330 linear_map_hash_slots[paddr >> PAGE_SHIFT] = ret | 0x80;
331 #endif /* CONFIG_DEBUG_PAGEALLOC */
333 return ret < 0 ? ret : 0;
336 int htab_remove_mapping(unsigned long vstart, unsigned long vend,
337 int psize, int ssize)
340 unsigned int step, shift;
344 shift = mmu_psize_defs[psize].shift;
347 if (!mmu_hash_ops.hpte_removebolted)
350 /* Unmap the full range specificied */
351 vaddr = ALIGN_DOWN(vstart, step);
352 for (;vaddr < vend; vaddr += step) {
353 rc = mmu_hash_ops.hpte_removebolted(vaddr, psize, ssize);
365 static bool disable_1tb_segments = false;
367 static int __init parse_disable_1tb_segments(char *p)
369 disable_1tb_segments = true;
372 early_param("disable_1tb_segments", parse_disable_1tb_segments);
374 static int __init htab_dt_scan_seg_sizes(unsigned long node,
375 const char *uname, int depth,
378 const char *type = of_get_flat_dt_prop(node, "device_type", NULL);
382 /* We are scanning "cpu" nodes only */
383 if (type == NULL || strcmp(type, "cpu") != 0)
386 prop = of_get_flat_dt_prop(node, "ibm,processor-segment-sizes", &size);
389 for (; size >= 4; size -= 4, ++prop) {
390 if (be32_to_cpu(prop[0]) == 40) {
391 DBG("1T segment support detected\n");
393 if (disable_1tb_segments) {
394 DBG("1T segments disabled by command line\n");
398 cur_cpu_spec->mmu_features |= MMU_FTR_1T_SEGMENT;
402 cur_cpu_spec->mmu_features &= ~MMU_FTR_NO_SLBIE_B;
406 static int __init get_idx_from_shift(unsigned int shift)
430 static int __init htab_dt_scan_page_sizes(unsigned long node,
431 const char *uname, int depth,
434 const char *type = of_get_flat_dt_prop(node, "device_type", NULL);
438 /* We are scanning "cpu" nodes only */
439 if (type == NULL || strcmp(type, "cpu") != 0)
442 prop = of_get_flat_dt_prop(node, "ibm,segment-page-sizes", &size);
446 pr_info("Page sizes from device-tree:\n");
448 cur_cpu_spec->mmu_features &= ~(MMU_FTR_16M_PAGE);
450 unsigned int base_shift = be32_to_cpu(prop[0]);
451 unsigned int slbenc = be32_to_cpu(prop[1]);
452 unsigned int lpnum = be32_to_cpu(prop[2]);
453 struct mmu_psize_def *def;
456 size -= 3; prop += 3;
457 base_idx = get_idx_from_shift(base_shift);
459 /* skip the pte encoding also */
460 prop += lpnum * 2; size -= lpnum * 2;
463 def = &mmu_psize_defs[base_idx];
464 if (base_idx == MMU_PAGE_16M)
465 cur_cpu_spec->mmu_features |= MMU_FTR_16M_PAGE;
467 def->shift = base_shift;
468 if (base_shift <= 23)
471 def->avpnm = (1 << (base_shift - 23)) - 1;
474 * We don't know for sure what's up with tlbiel, so
475 * for now we only set it for 4K and 64K pages
477 if (base_idx == MMU_PAGE_4K || base_idx == MMU_PAGE_64K)
482 while (size > 0 && lpnum) {
483 unsigned int shift = be32_to_cpu(prop[0]);
484 int penc = be32_to_cpu(prop[1]);
486 prop += 2; size -= 2;
489 idx = get_idx_from_shift(shift);
494 pr_err("Invalid penc for base_shift=%d "
495 "shift=%d\n", base_shift, shift);
497 def->penc[idx] = penc;
498 pr_info("base_shift=%d: shift=%d, sllp=0x%04lx,"
499 " avpnm=0x%08lx, tlbiel=%d, penc=%d\n",
500 base_shift, shift, def->sllp,
501 def->avpnm, def->tlbiel, def->penc[idx]);
508 #ifdef CONFIG_HUGETLB_PAGE
510 * Scan for 16G memory blocks that have been set aside for huge pages
511 * and reserve those blocks for 16G huge pages.
513 static int __init htab_dt_scan_hugepage_blocks(unsigned long node,
514 const char *uname, int depth,
516 const char *type = of_get_flat_dt_prop(node, "device_type", NULL);
517 const __be64 *addr_prop;
518 const __be32 *page_count_prop;
519 unsigned int expected_pages;
520 long unsigned int phys_addr;
521 long unsigned int block_size;
523 /* We are scanning "memory" nodes only */
524 if (type == NULL || strcmp(type, "memory") != 0)
528 * This property is the log base 2 of the number of virtual pages that
529 * will represent this memory block.
531 page_count_prop = of_get_flat_dt_prop(node, "ibm,expected#pages", NULL);
532 if (page_count_prop == NULL)
534 expected_pages = (1 << be32_to_cpu(page_count_prop[0]));
535 addr_prop = of_get_flat_dt_prop(node, "reg", NULL);
536 if (addr_prop == NULL)
538 phys_addr = be64_to_cpu(addr_prop[0]);
539 block_size = be64_to_cpu(addr_prop[1]);
540 if (block_size != (16 * GB))
542 printk(KERN_INFO "Huge page(16GB) memory: "
543 "addr = 0x%lX size = 0x%lX pages = %d\n",
544 phys_addr, block_size, expected_pages);
545 if (phys_addr + block_size * expected_pages <= memblock_end_of_DRAM()) {
546 memblock_reserve(phys_addr, block_size * expected_pages);
547 pseries_add_gpage(phys_addr, block_size, expected_pages);
551 #endif /* CONFIG_HUGETLB_PAGE */
553 static void mmu_psize_set_default_penc(void)
556 for (bpsize = 0; bpsize < MMU_PAGE_COUNT; bpsize++)
557 for (apsize = 0; apsize < MMU_PAGE_COUNT; apsize++)
558 mmu_psize_defs[bpsize].penc[apsize] = -1;
561 #ifdef CONFIG_PPC_64K_PAGES
563 static bool might_have_hea(void)
566 * The HEA ethernet adapter requires awareness of the
567 * GX bus. Without that awareness we can easily assume
568 * we will never see an HEA ethernet device.
570 #ifdef CONFIG_IBMEBUS
571 return !cpu_has_feature(CPU_FTR_ARCH_207S) &&
572 firmware_has_feature(FW_FEATURE_SPLPAR);
578 #endif /* #ifdef CONFIG_PPC_64K_PAGES */
580 static void __init htab_scan_page_sizes(void)
584 /* se the invalid penc to -1 */
585 mmu_psize_set_default_penc();
587 /* Default to 4K pages only */
588 memcpy(mmu_psize_defs, mmu_psize_defaults,
589 sizeof(mmu_psize_defaults));
592 * Try to find the available page sizes in the device-tree
594 rc = of_scan_flat_dt(htab_dt_scan_page_sizes, NULL);
595 if (rc == 0 && early_mmu_has_feature(MMU_FTR_16M_PAGE)) {
597 * Nothing in the device-tree, but the CPU supports 16M pages,
598 * so let's fallback on a known size list for 16M capable CPUs.
600 memcpy(mmu_psize_defs, mmu_psize_defaults_gp,
601 sizeof(mmu_psize_defaults_gp));
604 #ifdef CONFIG_HUGETLB_PAGE
605 if (!hugetlb_disabled && !early_radix_enabled() ) {
606 /* Reserve 16G huge page memory sections for huge pages */
607 of_scan_flat_dt(htab_dt_scan_hugepage_blocks, NULL);
609 #endif /* CONFIG_HUGETLB_PAGE */
613 * Fill in the hpte_page_sizes[] array.
614 * We go through the mmu_psize_defs[] array looking for all the
615 * supported base/actual page size combinations. Each combination
616 * has a unique pagesize encoding (penc) value in the low bits of
617 * the LP field of the HPTE. For actual page sizes less than 1MB,
618 * some of the upper LP bits are used for RPN bits, meaning that
619 * we need to fill in several entries in hpte_page_sizes[].
621 * In diagrammatic form, with r = RPN bits and z = page size bits:
622 * PTE LP actual page size
629 * The zzzz bits are implementation-specific but are chosen so that
630 * no encoding for a larger page size uses the same value in its
631 * low-order N bits as the encoding for the 2^(12+N) byte page size
634 static void init_hpte_page_sizes(void)
637 long int shift, penc;
639 for (bp = 0; bp < MMU_PAGE_COUNT; ++bp) {
640 if (!mmu_psize_defs[bp].shift)
641 continue; /* not a supported page size */
642 for (ap = bp; ap < MMU_PAGE_COUNT; ++ap) {
643 penc = mmu_psize_defs[bp].penc[ap];
644 if (penc == -1 || !mmu_psize_defs[ap].shift)
646 shift = mmu_psize_defs[ap].shift - LP_SHIFT;
648 continue; /* should never happen */
650 * For page sizes less than 1MB, this loop
651 * replicates the entry for all possible values
654 while (penc < (1 << LP_BITS)) {
655 hpte_page_sizes[penc] = (ap << 4) | bp;
662 static void __init htab_init_page_sizes(void)
665 init_hpte_page_sizes();
667 if (!debug_pagealloc_enabled()) {
669 * Pick a size for the linear mapping. Currently, we only
670 * support 16M, 1M and 4K which is the default
672 if (IS_ENABLED(CONFIG_STRICT_KERNEL_RWX) &&
673 (unsigned long)_stext % 0x1000000) {
674 if (mmu_psize_defs[MMU_PAGE_16M].shift)
675 pr_warn("Kernel not 16M aligned, disabling 16M linear map alignment\n");
679 if (mmu_psize_defs[MMU_PAGE_16M].shift && aligned)
680 mmu_linear_psize = MMU_PAGE_16M;
681 else if (mmu_psize_defs[MMU_PAGE_1M].shift)
682 mmu_linear_psize = MMU_PAGE_1M;
685 #ifdef CONFIG_PPC_64K_PAGES
687 * Pick a size for the ordinary pages. Default is 4K, we support
688 * 64K for user mappings and vmalloc if supported by the processor.
689 * We only use 64k for ioremap if the processor
690 * (and firmware) support cache-inhibited large pages.
691 * If not, we use 4k and set mmu_ci_restrictions so that
692 * hash_page knows to switch processes that use cache-inhibited
693 * mappings to 4k pages.
695 if (mmu_psize_defs[MMU_PAGE_64K].shift) {
696 mmu_virtual_psize = MMU_PAGE_64K;
697 mmu_vmalloc_psize = MMU_PAGE_64K;
698 if (mmu_linear_psize == MMU_PAGE_4K)
699 mmu_linear_psize = MMU_PAGE_64K;
700 if (mmu_has_feature(MMU_FTR_CI_LARGE_PAGE)) {
702 * When running on pSeries using 64k pages for ioremap
703 * would stop us accessing the HEA ethernet. So if we
704 * have the chance of ever seeing one, stay at 4k.
706 if (!might_have_hea())
707 mmu_io_psize = MMU_PAGE_64K;
709 mmu_ci_restrictions = 1;
711 #endif /* CONFIG_PPC_64K_PAGES */
713 #ifdef CONFIG_SPARSEMEM_VMEMMAP
715 * We try to use 16M pages for vmemmap if that is supported
716 * and we have at least 1G of RAM at boot
718 if (mmu_psize_defs[MMU_PAGE_16M].shift &&
719 memblock_phys_mem_size() >= 0x40000000)
720 mmu_vmemmap_psize = MMU_PAGE_16M;
722 mmu_vmemmap_psize = mmu_virtual_psize;
723 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
725 printk(KERN_DEBUG "Page orders: linear mapping = %d, "
726 "virtual = %d, io = %d"
727 #ifdef CONFIG_SPARSEMEM_VMEMMAP
731 mmu_psize_defs[mmu_linear_psize].shift,
732 mmu_psize_defs[mmu_virtual_psize].shift,
733 mmu_psize_defs[mmu_io_psize].shift
734 #ifdef CONFIG_SPARSEMEM_VMEMMAP
735 ,mmu_psize_defs[mmu_vmemmap_psize].shift
740 static int __init htab_dt_scan_pftsize(unsigned long node,
741 const char *uname, int depth,
744 const char *type = of_get_flat_dt_prop(node, "device_type", NULL);
747 /* We are scanning "cpu" nodes only */
748 if (type == NULL || strcmp(type, "cpu") != 0)
751 prop = of_get_flat_dt_prop(node, "ibm,pft-size", NULL);
753 /* pft_size[0] is the NUMA CEC cookie */
754 ppc64_pft_size = be32_to_cpu(prop[1]);
760 unsigned htab_shift_for_mem_size(unsigned long mem_size)
762 unsigned memshift = __ilog2(mem_size);
763 unsigned pshift = mmu_psize_defs[mmu_virtual_psize].shift;
766 /* round mem_size up to next power of 2 */
767 if ((1UL << memshift) < mem_size)
770 /* aim for 2 pages / pteg */
771 pteg_shift = memshift - (pshift + 1);
774 * 2^11 PTEGS of 128 bytes each, ie. 2^18 bytes is the minimum htab
775 * size permitted by the architecture.
777 return max(pteg_shift + 7, 18U);
780 static unsigned long __init htab_get_table_size(void)
783 * If hash size isn't already provided by the platform, we try to
784 * retrieve it from the device-tree. If it's not there neither, we
785 * calculate it now based on the total RAM size
787 if (ppc64_pft_size == 0)
788 of_scan_flat_dt(htab_dt_scan_pftsize, NULL);
790 return 1UL << ppc64_pft_size;
792 return 1UL << htab_shift_for_mem_size(memblock_phys_mem_size());
795 #ifdef CONFIG_MEMORY_HOTPLUG
796 static int resize_hpt_for_hotplug(unsigned long new_mem_size)
798 unsigned target_hpt_shift;
800 if (!mmu_hash_ops.resize_hpt)
803 target_hpt_shift = htab_shift_for_mem_size(new_mem_size);
806 * To avoid lots of HPT resizes if memory size is fluctuating
807 * across a boundary, we deliberately have some hysterisis
808 * here: we immediately increase the HPT size if the target
809 * shift exceeds the current shift, but we won't attempt to
810 * reduce unless the target shift is at least 2 below the
813 if (target_hpt_shift > ppc64_pft_size ||
814 target_hpt_shift < ppc64_pft_size - 1)
815 return mmu_hash_ops.resize_hpt(target_hpt_shift);
820 int hash__create_section_mapping(unsigned long start, unsigned long end,
821 int nid, pgprot_t prot)
825 if (end >= H_VMALLOC_START) {
826 pr_warn("Outside the supported range\n");
830 resize_hpt_for_hotplug(memblock_phys_mem_size());
832 rc = htab_bolt_mapping(start, end, __pa(start),
833 pgprot_val(prot), mmu_linear_psize,
837 int rc2 = htab_remove_mapping(start, end, mmu_linear_psize,
839 BUG_ON(rc2 && (rc2 != -ENOENT));
844 int hash__remove_section_mapping(unsigned long start, unsigned long end)
846 int rc = htab_remove_mapping(start, end, mmu_linear_psize,
850 if (resize_hpt_for_hotplug(memblock_phys_mem_size()) == -ENOSPC)
851 pr_warn("Hash collision while resizing HPT\n");
855 #endif /* CONFIG_MEMORY_HOTPLUG */
857 static void __init hash_init_partition_table(phys_addr_t hash_table,
858 unsigned long htab_size)
860 mmu_partition_table_init();
863 * PS field (VRMA page size) is not used for LPID 0, hence set to 0.
864 * For now, UPRT is 0 and we have no segment table.
866 htab_size = __ilog2(htab_size) - 18;
867 mmu_partition_table_set_entry(0, hash_table | htab_size, 0, false);
868 pr_info("Partition table %p\n", partition_tb);
871 static void __init htab_initialize(void)
874 unsigned long pteg_count;
876 phys_addr_t base = 0, size = 0, end;
879 DBG(" -> htab_initialize()\n");
881 if (mmu_has_feature(MMU_FTR_1T_SEGMENT)) {
882 mmu_kernel_ssize = MMU_SEGSIZE_1T;
883 mmu_highuser_ssize = MMU_SEGSIZE_1T;
884 printk(KERN_INFO "Using 1TB segments\n");
887 if (stress_slb_enabled)
888 static_branch_enable(&stress_slb_key);
891 * Calculate the required size of the htab. We want the number of
892 * PTEGs to equal one half the number of real pages.
894 htab_size_bytes = htab_get_table_size();
895 pteg_count = htab_size_bytes >> 7;
897 htab_hash_mask = pteg_count - 1;
899 if (firmware_has_feature(FW_FEATURE_LPAR) ||
900 firmware_has_feature(FW_FEATURE_PS3_LV1)) {
901 /* Using a hypervisor which owns the htab */
904 #ifdef CONFIG_FA_DUMP
906 * If firmware assisted dump is active firmware preserves
907 * the contents of htab along with entire partition memory.
908 * Clear the htab if firmware assisted dump is active so
909 * that we dont end up using old mappings.
911 if (is_fadump_active() && mmu_hash_ops.hpte_clear_all)
912 mmu_hash_ops.hpte_clear_all();
915 unsigned long limit = MEMBLOCK_ALLOC_ANYWHERE;
917 #ifdef CONFIG_PPC_CELL
919 * Cell may require the hash table down low when using the
920 * Axon IOMMU in order to fit the dynamic region over it, see
921 * comments in cell/iommu.c
923 if (fdt_subnode_offset(initial_boot_params, 0, "axon") > 0) {
925 pr_info("Hash table forced below 2G for Axon IOMMU\n");
927 #endif /* CONFIG_PPC_CELL */
929 table = memblock_phys_alloc_range(htab_size_bytes,
933 panic("ERROR: Failed to allocate %pa bytes below %pa\n",
934 &htab_size_bytes, &limit);
936 DBG("Hash table allocated at %lx, size: %lx\n", table,
939 htab_address = __va(table);
941 /* htab absolute addr + encoded htabsize */
942 _SDR1 = table + __ilog2(htab_size_bytes) - 18;
944 /* Initialize the HPT with no entries */
945 memset((void *)table, 0, htab_size_bytes);
947 if (!cpu_has_feature(CPU_FTR_ARCH_300))
949 mtspr(SPRN_SDR1, _SDR1);
951 hash_init_partition_table(table, htab_size_bytes);
954 prot = pgprot_val(PAGE_KERNEL);
956 #ifdef CONFIG_DEBUG_PAGEALLOC
957 if (debug_pagealloc_enabled()) {
958 linear_map_hash_count = memblock_end_of_DRAM() >> PAGE_SHIFT;
959 linear_map_hash_slots = memblock_alloc_try_nid(
960 linear_map_hash_count, 1, MEMBLOCK_LOW_LIMIT,
961 ppc64_rma_size, NUMA_NO_NODE);
962 if (!linear_map_hash_slots)
963 panic("%s: Failed to allocate %lu bytes max_addr=%pa\n",
964 __func__, linear_map_hash_count, &ppc64_rma_size);
966 #endif /* CONFIG_DEBUG_PAGEALLOC */
968 /* create bolted the linear mapping in the hash table */
969 for_each_mem_range(i, &base, &end) {
971 base = (unsigned long)__va(base);
973 DBG("creating mapping for region: %lx..%lx (prot: %lx)\n",
976 if ((base + size) >= H_VMALLOC_START) {
977 pr_warn("Outside the supported range\n");
981 BUG_ON(htab_bolt_mapping(base, base + size, __pa(base),
982 prot, mmu_linear_psize, mmu_kernel_ssize));
984 memblock_set_current_limit(MEMBLOCK_ALLOC_ANYWHERE);
987 * If we have a memory_limit and we've allocated TCEs then we need to
988 * explicitly map the TCE area at the top of RAM. We also cope with the
989 * case that the TCEs start below memory_limit.
990 * tce_alloc_start/end are 16MB aligned so the mapping should work
991 * for either 4K or 16MB pages.
993 if (tce_alloc_start) {
994 tce_alloc_start = (unsigned long)__va(tce_alloc_start);
995 tce_alloc_end = (unsigned long)__va(tce_alloc_end);
997 if (base + size >= tce_alloc_start)
998 tce_alloc_start = base + size + 1;
1000 BUG_ON(htab_bolt_mapping(tce_alloc_start, tce_alloc_end,
1001 __pa(tce_alloc_start), prot,
1002 mmu_linear_psize, mmu_kernel_ssize));
1006 DBG(" <- htab_initialize()\n");
1011 void __init hash__early_init_devtree(void)
1013 /* Initialize segment sizes */
1014 of_scan_flat_dt(htab_dt_scan_seg_sizes, NULL);
1016 /* Initialize page sizes */
1017 htab_scan_page_sizes();
1020 static struct hash_mm_context init_hash_mm_context;
1021 void __init hash__early_init_mmu(void)
1023 #ifndef CONFIG_PPC_64K_PAGES
1025 * We have code in __hash_page_4K() and elsewhere, which assumes it can
1027 * new_pte |= (slot << H_PAGE_F_GIX_SHIFT) & (H_PAGE_F_SECOND | H_PAGE_F_GIX);
1029 * Where the slot number is between 0-15, and values of 8-15 indicate
1030 * the secondary bucket. For that code to work H_PAGE_F_SECOND and
1031 * H_PAGE_F_GIX must occupy four contiguous bits in the PTE, and
1032 * H_PAGE_F_SECOND must be placed above H_PAGE_F_GIX. Assert that here
1033 * with a BUILD_BUG_ON().
1035 BUILD_BUG_ON(H_PAGE_F_SECOND != (1ul << (H_PAGE_F_GIX_SHIFT + 3)));
1036 #endif /* CONFIG_PPC_64K_PAGES */
1038 htab_init_page_sizes();
1041 * initialize page table size
1043 __pte_frag_nr = H_PTE_FRAG_NR;
1044 __pte_frag_size_shift = H_PTE_FRAG_SIZE_SHIFT;
1045 __pmd_frag_nr = H_PMD_FRAG_NR;
1046 __pmd_frag_size_shift = H_PMD_FRAG_SIZE_SHIFT;
1048 __pte_index_size = H_PTE_INDEX_SIZE;
1049 __pmd_index_size = H_PMD_INDEX_SIZE;
1050 __pud_index_size = H_PUD_INDEX_SIZE;
1051 __pgd_index_size = H_PGD_INDEX_SIZE;
1052 __pud_cache_index = H_PUD_CACHE_INDEX;
1053 __pte_table_size = H_PTE_TABLE_SIZE;
1054 __pmd_table_size = H_PMD_TABLE_SIZE;
1055 __pud_table_size = H_PUD_TABLE_SIZE;
1056 __pgd_table_size = H_PGD_TABLE_SIZE;
1058 * 4k use hugepd format, so for hash set then to
1061 __pmd_val_bits = HASH_PMD_VAL_BITS;
1062 __pud_val_bits = HASH_PUD_VAL_BITS;
1063 __pgd_val_bits = HASH_PGD_VAL_BITS;
1065 __kernel_virt_start = H_KERN_VIRT_START;
1066 __vmalloc_start = H_VMALLOC_START;
1067 __vmalloc_end = H_VMALLOC_END;
1068 __kernel_io_start = H_KERN_IO_START;
1069 __kernel_io_end = H_KERN_IO_END;
1070 vmemmap = (struct page *)H_VMEMMAP_START;
1071 ioremap_bot = IOREMAP_BASE;
1074 pci_io_base = ISA_IO_BASE;
1077 /* Select appropriate backend */
1078 if (firmware_has_feature(FW_FEATURE_PS3_LV1))
1079 ps3_early_mm_init();
1080 else if (firmware_has_feature(FW_FEATURE_LPAR))
1081 hpte_init_pseries();
1082 else if (IS_ENABLED(CONFIG_PPC_NATIVE))
1085 if (!mmu_hash_ops.hpte_insert)
1086 panic("hash__early_init_mmu: No MMU hash ops defined!\n");
1089 * Initialize the MMU Hash table and create the linear mapping
1090 * of memory. Has to be done before SLB initialization as this is
1091 * currently where the page size encoding is obtained.
1095 init_mm.context.hash_context = &init_hash_mm_context;
1096 mm_ctx_set_slb_addr_limit(&init_mm.context, SLB_ADDR_LIMIT_DEFAULT);
1098 pr_info("Initializing hash mmu with SLB\n");
1099 /* Initialize SLB management */
1102 if (cpu_has_feature(CPU_FTR_ARCH_206)
1103 && cpu_has_feature(CPU_FTR_HVMODE))
1108 void hash__early_init_mmu_secondary(void)
1110 /* Initialize hash table for that CPU */
1111 if (!firmware_has_feature(FW_FEATURE_LPAR)) {
1113 if (!cpu_has_feature(CPU_FTR_ARCH_300))
1114 mtspr(SPRN_SDR1, _SDR1);
1116 set_ptcr_when_no_uv(__pa(partition_tb) |
1117 (PATB_SIZE_SHIFT - 12));
1119 /* Initialize SLB */
1122 if (cpu_has_feature(CPU_FTR_ARCH_206)
1123 && cpu_has_feature(CPU_FTR_HVMODE))
1126 #ifdef CONFIG_PPC_MEM_KEYS
1127 if (mmu_has_feature(MMU_FTR_PKEY))
1128 mtspr(SPRN_UAMOR, default_uamor);
1131 #endif /* CONFIG_SMP */
1134 * Called by asm hashtable.S for doing lazy icache flush
1136 unsigned int hash_page_do_lazy_icache(unsigned int pp, pte_t pte, int trap)
1140 if (!pfn_valid(pte_pfn(pte)))
1143 page = pte_page(pte);
1146 if (!test_bit(PG_arch_1, &page->flags) && !PageReserved(page)) {
1147 if (trap == 0x400) {
1148 flush_dcache_icache_page(page);
1149 set_bit(PG_arch_1, &page->flags);
1156 #ifdef CONFIG_PPC_MM_SLICES
1157 static unsigned int get_paca_psize(unsigned long addr)
1159 unsigned char *psizes;
1160 unsigned long index, mask_index;
1162 if (addr < SLICE_LOW_TOP) {
1163 psizes = get_paca()->mm_ctx_low_slices_psize;
1164 index = GET_LOW_SLICE_INDEX(addr);
1166 psizes = get_paca()->mm_ctx_high_slices_psize;
1167 index = GET_HIGH_SLICE_INDEX(addr);
1169 mask_index = index & 0x1;
1170 return (psizes[index >> 1] >> (mask_index * 4)) & 0xF;
1174 unsigned int get_paca_psize(unsigned long addr)
1176 return get_paca()->mm_ctx_user_psize;
1181 * Demote a segment to using 4k pages.
1182 * For now this makes the whole process use 4k pages.
1184 #ifdef CONFIG_PPC_64K_PAGES
1185 void demote_segment_4k(struct mm_struct *mm, unsigned long addr)
1187 if (get_slice_psize(mm, addr) == MMU_PAGE_4K)
1189 slice_set_range_psize(mm, addr, 1, MMU_PAGE_4K);
1190 copro_flush_all_slbs(mm);
1191 if ((get_paca_psize(addr) != MMU_PAGE_4K) && (current->mm == mm)) {
1193 copy_mm_to_paca(mm);
1194 slb_flush_and_restore_bolted();
1197 #endif /* CONFIG_PPC_64K_PAGES */
1199 #ifdef CONFIG_PPC_SUBPAGE_PROT
1201 * This looks up a 2-bit protection code for a 4k subpage of a 64k page.
1202 * Userspace sets the subpage permissions using the subpage_prot system call.
1204 * Result is 0: full permissions, _PAGE_RW: read-only,
1205 * _PAGE_RWX: no access.
1207 static int subpage_protection(struct mm_struct *mm, unsigned long ea)
1209 struct subpage_prot_table *spt = mm_ctx_subpage_prot(&mm->context);
1216 if (ea >= spt->maxaddr)
1218 if (ea < 0x100000000UL) {
1219 /* addresses below 4GB use spt->low_prot */
1220 sbpm = spt->low_prot;
1222 sbpm = spt->protptrs[ea >> SBP_L3_SHIFT];
1226 sbpp = sbpm[(ea >> SBP_L2_SHIFT) & (SBP_L2_COUNT - 1)];
1229 spp = sbpp[(ea >> PAGE_SHIFT) & (SBP_L1_COUNT - 1)];
1231 /* extract 2-bit bitfield for this 4k subpage */
1232 spp >>= 30 - 2 * ((ea >> 12) & 0xf);
1235 * 0 -> full premission
1238 * We return the flag that need to be cleared.
1240 spp = ((spp & 2) ? _PAGE_RWX : 0) | ((spp & 1) ? _PAGE_WRITE : 0);
1244 #else /* CONFIG_PPC_SUBPAGE_PROT */
1245 static inline int subpage_protection(struct mm_struct *mm, unsigned long ea)
1251 void hash_failure_debug(unsigned long ea, unsigned long access,
1252 unsigned long vsid, unsigned long trap,
1253 int ssize, int psize, int lpsize, unsigned long pte)
1255 if (!printk_ratelimit())
1257 pr_info("mm: Hashing failure ! EA=0x%lx access=0x%lx current=%s\n",
1258 ea, access, current->comm);
1259 pr_info(" trap=0x%lx vsid=0x%lx ssize=%d base psize=%d psize %d pte=0x%lx\n",
1260 trap, vsid, ssize, psize, lpsize, pte);
1263 static void check_paca_psize(unsigned long ea, struct mm_struct *mm,
1264 int psize, bool user_region)
1267 if (psize != get_paca_psize(ea)) {
1268 copy_mm_to_paca(mm);
1269 slb_flush_and_restore_bolted();
1271 } else if (get_paca()->vmalloc_sllp !=
1272 mmu_psize_defs[mmu_vmalloc_psize].sllp) {
1273 get_paca()->vmalloc_sllp =
1274 mmu_psize_defs[mmu_vmalloc_psize].sllp;
1275 slb_vmalloc_update();
1282 * 1 - normal page fault
1283 * -1 - critical hash insertion error
1284 * -2 - access not permitted by subpage protection mechanism
1286 int hash_page_mm(struct mm_struct *mm, unsigned long ea,
1287 unsigned long access, unsigned long trap,
1288 unsigned long flags)
1291 enum ctx_state prev_state = exception_enter();
1296 int rc, user_region = 0;
1299 DBG_LOW("hash_page(ea=%016lx, access=%lx, trap=%lx\n",
1301 trace_hash_fault(ea, access, trap);
1303 /* Get region & vsid */
1304 switch (get_region_id(ea)) {
1305 case USER_REGION_ID:
1308 DBG_LOW(" user region with no mm !\n");
1312 psize = get_slice_psize(mm, ea);
1313 ssize = user_segment_size(ea);
1314 vsid = get_user_vsid(&mm->context, ea, ssize);
1316 case VMALLOC_REGION_ID:
1317 vsid = get_kernel_vsid(ea, mmu_kernel_ssize);
1318 psize = mmu_vmalloc_psize;
1319 ssize = mmu_kernel_ssize;
1323 vsid = get_kernel_vsid(ea, mmu_kernel_ssize);
1324 psize = mmu_io_psize;
1325 ssize = mmu_kernel_ssize;
1330 * Send the problem up to do_page_fault()
1335 DBG_LOW(" mm=%p, mm->pgdir=%p, vsid=%016lx\n", mm, mm->pgd, vsid);
1339 DBG_LOW("Bad address!\n");
1345 if (pgdir == NULL) {
1350 /* Check CPU locality */
1351 if (user_region && mm_is_thread_local(mm))
1352 flags |= HPTE_LOCAL_UPDATE;
1354 #ifndef CONFIG_PPC_64K_PAGES
1356 * If we use 4K pages and our psize is not 4K, then we might
1357 * be hitting a special driver mapping, and need to align the
1358 * address before we fetch the PTE.
1360 * It could also be a hugepage mapping, in which case this is
1361 * not necessary, but it's not harmful, either.
1363 if (psize != MMU_PAGE_4K)
1364 ea &= ~((1ul << mmu_psize_defs[psize].shift) - 1);
1365 #endif /* CONFIG_PPC_64K_PAGES */
1367 /* Get PTE and page size from page tables */
1368 ptep = find_linux_pte(pgdir, ea, &is_thp, &hugeshift);
1369 if (ptep == NULL || !pte_present(*ptep)) {
1370 DBG_LOW(" no PTE !\n");
1376 * Add _PAGE_PRESENT to the required access perm. If there are parallel
1377 * updates to the pte that can possibly clear _PAGE_PTE, catch that too.
1379 * We can safely use the return pte address in rest of the function
1380 * because we do set H_PAGE_BUSY which prevents further updates to pte
1381 * from generic code.
1383 access |= _PAGE_PRESENT | _PAGE_PTE;
1386 * Pre-check access permissions (will be re-checked atomically
1387 * in __hash_page_XX but this pre-check is a fast path
1389 if (!check_pte_access(access, pte_val(*ptep))) {
1390 DBG_LOW(" no access !\n");
1397 rc = __hash_page_thp(ea, access, vsid, (pmd_t *)ptep,
1398 trap, flags, ssize, psize);
1399 #ifdef CONFIG_HUGETLB_PAGE
1401 rc = __hash_page_huge(ea, access, vsid, ptep, trap,
1402 flags, ssize, hugeshift, psize);
1406 * if we have hugeshift, and is not transhuge with
1407 * hugetlb disabled, something is really wrong.
1413 if (current->mm == mm)
1414 check_paca_psize(ea, mm, psize, user_region);
1419 #ifndef CONFIG_PPC_64K_PAGES
1420 DBG_LOW(" i-pte: %016lx\n", pte_val(*ptep));
1422 DBG_LOW(" i-pte: %016lx %016lx\n", pte_val(*ptep),
1423 pte_val(*(ptep + PTRS_PER_PTE)));
1425 /* Do actual hashing */
1426 #ifdef CONFIG_PPC_64K_PAGES
1427 /* If H_PAGE_4K_PFN is set, make sure this is a 4k segment */
1428 if ((pte_val(*ptep) & H_PAGE_4K_PFN) && psize == MMU_PAGE_64K) {
1429 demote_segment_4k(mm, ea);
1430 psize = MMU_PAGE_4K;
1434 * If this PTE is non-cacheable and we have restrictions on
1435 * using non cacheable large pages, then we switch to 4k
1437 if (mmu_ci_restrictions && psize == MMU_PAGE_64K && pte_ci(*ptep)) {
1439 demote_segment_4k(mm, ea);
1440 psize = MMU_PAGE_4K;
1441 } else if (ea < VMALLOC_END) {
1443 * some driver did a non-cacheable mapping
1444 * in vmalloc space, so switch vmalloc
1447 printk(KERN_ALERT "Reducing vmalloc segment "
1448 "to 4kB pages because of "
1449 "non-cacheable mapping\n");
1450 psize = mmu_vmalloc_psize = MMU_PAGE_4K;
1451 copro_flush_all_slbs(mm);
1455 #endif /* CONFIG_PPC_64K_PAGES */
1457 if (current->mm == mm)
1458 check_paca_psize(ea, mm, psize, user_region);
1460 #ifdef CONFIG_PPC_64K_PAGES
1461 if (psize == MMU_PAGE_64K)
1462 rc = __hash_page_64K(ea, access, vsid, ptep, trap,
1465 #endif /* CONFIG_PPC_64K_PAGES */
1467 int spp = subpage_protection(mm, ea);
1471 rc = __hash_page_4K(ea, access, vsid, ptep, trap,
1476 * Dump some info in case of hash insertion failure, they should
1477 * never happen so it is really useful to know if/when they do
1480 hash_failure_debug(ea, access, vsid, trap, ssize, psize,
1481 psize, pte_val(*ptep));
1482 #ifndef CONFIG_PPC_64K_PAGES
1483 DBG_LOW(" o-pte: %016lx\n", pte_val(*ptep));
1485 DBG_LOW(" o-pte: %016lx %016lx\n", pte_val(*ptep),
1486 pte_val(*(ptep + PTRS_PER_PTE)));
1488 DBG_LOW(" -> rc=%d\n", rc);
1491 exception_exit(prev_state);
1494 EXPORT_SYMBOL_GPL(hash_page_mm);
1496 int hash_page(unsigned long ea, unsigned long access, unsigned long trap,
1497 unsigned long dsisr)
1499 unsigned long flags = 0;
1500 struct mm_struct *mm = current->mm;
1502 if ((get_region_id(ea) == VMALLOC_REGION_ID) ||
1503 (get_region_id(ea) == IO_REGION_ID))
1506 if (dsisr & DSISR_NOHPTE)
1507 flags |= HPTE_NOHPTE_UPDATE;
1509 return hash_page_mm(mm, ea, access, trap, flags);
1511 EXPORT_SYMBOL_GPL(hash_page);
1513 int __hash_page(unsigned long trap, unsigned long ea, unsigned long dsisr,
1516 unsigned long access = _PAGE_PRESENT | _PAGE_READ;
1517 unsigned long flags = 0;
1518 struct mm_struct *mm = current->mm;
1519 unsigned int region_id = get_region_id(ea);
1521 if ((region_id == VMALLOC_REGION_ID) || (region_id == IO_REGION_ID))
1524 if (dsisr & DSISR_NOHPTE)
1525 flags |= HPTE_NOHPTE_UPDATE;
1527 if (dsisr & DSISR_ISSTORE)
1528 access |= _PAGE_WRITE;
1530 * We set _PAGE_PRIVILEGED only when
1531 * kernel mode access kernel space.
1533 * _PAGE_PRIVILEGED is NOT set
1534 * 1) when kernel mode access user space
1535 * 2) user space access kernel space.
1537 access |= _PAGE_PRIVILEGED;
1538 if ((msr & MSR_PR) || (region_id == USER_REGION_ID))
1539 access &= ~_PAGE_PRIVILEGED;
1542 access |= _PAGE_EXEC;
1544 return hash_page_mm(mm, ea, access, trap, flags);
1547 #ifdef CONFIG_PPC_MM_SLICES
1548 static bool should_hash_preload(struct mm_struct *mm, unsigned long ea)
1550 int psize = get_slice_psize(mm, ea);
1552 /* We only prefault standard pages for now */
1553 if (unlikely(psize != mm_ctx_user_psize(&mm->context)))
1557 * Don't prefault if subpage protection is enabled for the EA.
1559 if (unlikely((psize == MMU_PAGE_4K) && subpage_protection(mm, ea)))
1565 static bool should_hash_preload(struct mm_struct *mm, unsigned long ea)
1571 static void hash_preload(struct mm_struct *mm, pte_t *ptep, unsigned long ea,
1572 bool is_exec, unsigned long trap)
1576 int rc, ssize, update_flags = 0;
1577 unsigned long access = _PAGE_PRESENT | _PAGE_READ | (is_exec ? _PAGE_EXEC : 0);
1578 unsigned long flags;
1580 BUG_ON(get_region_id(ea) != USER_REGION_ID);
1582 if (!should_hash_preload(mm, ea))
1585 DBG_LOW("hash_preload(mm=%p, mm->pgdir=%p, ea=%016lx, access=%lx,"
1586 " trap=%lx\n", mm, mm->pgd, ea, access, trap);
1588 /* Get Linux PTE if available */
1594 ssize = user_segment_size(ea);
1595 vsid = get_user_vsid(&mm->context, ea, ssize);
1599 #ifdef CONFIG_PPC_64K_PAGES
1600 /* If either H_PAGE_4K_PFN or cache inhibited is set (and we are on
1601 * a 64K kernel), then we don't preload, hash_page() will take
1602 * care of it once we actually try to access the page.
1603 * That way we don't have to duplicate all of the logic for segment
1604 * page size demotion here
1605 * Called with PTL held, hence can be sure the value won't change in
1608 if ((pte_val(*ptep) & H_PAGE_4K_PFN) || pte_ci(*ptep))
1610 #endif /* CONFIG_PPC_64K_PAGES */
1613 * __hash_page_* must run with interrupts off, as it sets the
1614 * H_PAGE_BUSY bit. It's possible for perf interrupts to hit at any
1615 * time and may take a hash fault reading the user stack, see
1616 * read_user_stack_slow() in the powerpc/perf code.
1618 * If that takes a hash fault on the same page as we lock here, it
1619 * will bail out when seeing H_PAGE_BUSY set, and retry the access
1620 * leading to an infinite loop.
1622 * Disabling interrupts here does not prevent perf interrupts, but it
1623 * will prevent them taking hash faults (see the NMI test in
1624 * do_hash_page), then read_user_stack's copy_from_user_nofault will
1625 * fail and perf will fall back to read_user_stack_slow(), which
1626 * walks the Linux page tables.
1628 * Interrupts must also be off for the duration of the
1629 * mm_is_thread_local test and update, to prevent preempt running the
1630 * mm on another CPU (XXX: this may be racy vs kthread_use_mm).
1632 local_irq_save(flags);
1634 /* Is that local to this CPU ? */
1635 if (mm_is_thread_local(mm))
1636 update_flags |= HPTE_LOCAL_UPDATE;
1639 #ifdef CONFIG_PPC_64K_PAGES
1640 if (mm_ctx_user_psize(&mm->context) == MMU_PAGE_64K)
1641 rc = __hash_page_64K(ea, access, vsid, ptep, trap,
1642 update_flags, ssize);
1644 #endif /* CONFIG_PPC_64K_PAGES */
1645 rc = __hash_page_4K(ea, access, vsid, ptep, trap, update_flags,
1646 ssize, subpage_protection(mm, ea));
1648 /* Dump some info in case of hash insertion failure, they should
1649 * never happen so it is really useful to know if/when they do
1652 hash_failure_debug(ea, access, vsid, trap, ssize,
1653 mm_ctx_user_psize(&mm->context),
1654 mm_ctx_user_psize(&mm->context),
1657 local_irq_restore(flags);
1661 * This is called at the end of handling a user page fault, when the
1662 * fault has been handled by updating a PTE in the linux page tables.
1663 * We use it to preload an HPTE into the hash table corresponding to
1664 * the updated linux PTE.
1666 * This must always be called with the pte lock held.
1668 void update_mmu_cache(struct vm_area_struct *vma, unsigned long address,
1672 * We don't need to worry about _PAGE_PRESENT here because we are
1673 * called with either mm->page_table_lock held or ptl lock held
1678 if (radix_enabled())
1681 /* We only want HPTEs for linux PTEs that have _PAGE_ACCESSED set */
1682 if (!pte_young(*ptep) || address >= TASK_SIZE)
1686 * We try to figure out if we are coming from an instruction
1687 * access fault and pass that down to __hash_page so we avoid
1688 * double-faulting on execution of fresh text. We have to test
1689 * for regs NULL since init will get here first thing at boot.
1691 * We also avoid filling the hash if not coming from a fault.
1694 trap = current->thread.regs ? TRAP(current->thread.regs) : 0UL;
1706 hash_preload(vma->vm_mm, ptep, address, is_exec, trap);
1709 #ifdef CONFIG_PPC_TRANSACTIONAL_MEM
1710 static inline void tm_flush_hash_page(int local)
1713 * Transactions are not aborted by tlbiel, only tlbie. Without, syncing a
1714 * page back to a block device w/PIO could pick up transactional data
1715 * (bad!) so we force an abort here. Before the sync the page will be
1716 * made read-only, which will flush_hash_page. BIG ISSUE here: if the
1717 * kernel uses a page from userspace without unmapping it first, it may
1718 * see the speculated version.
1720 if (local && cpu_has_feature(CPU_FTR_TM) && current->thread.regs &&
1721 MSR_TM_ACTIVE(current->thread.regs->msr)) {
1723 tm_abort(TM_CAUSE_TLBI);
1727 static inline void tm_flush_hash_page(int local)
1733 * Return the global hash slot, corresponding to the given PTE, which contains
1736 unsigned long pte_get_hash_gslot(unsigned long vpn, unsigned long shift,
1737 int ssize, real_pte_t rpte, unsigned int subpg_index)
1739 unsigned long hash, gslot, hidx;
1741 hash = hpt_hash(vpn, shift, ssize);
1742 hidx = __rpte_to_hidx(rpte, subpg_index);
1743 if (hidx & _PTEIDX_SECONDARY)
1745 gslot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1746 gslot += hidx & _PTEIDX_GROUP_IX;
1750 void flush_hash_page(unsigned long vpn, real_pte_t pte, int psize, int ssize,
1751 unsigned long flags)
1753 unsigned long index, shift, gslot;
1754 int local = flags & HPTE_LOCAL_UPDATE;
1756 DBG_LOW("flush_hash_page(vpn=%016lx)\n", vpn);
1757 pte_iterate_hashed_subpages(pte, psize, vpn, index, shift) {
1758 gslot = pte_get_hash_gslot(vpn, shift, ssize, pte, index);
1759 DBG_LOW(" sub %ld: gslot=%lx\n", index, gslot);
1761 * We use same base page size and actual psize, because we don't
1762 * use these functions for hugepage
1764 mmu_hash_ops.hpte_invalidate(gslot, vpn, psize, psize,
1766 } pte_iterate_hashed_end();
1768 tm_flush_hash_page(local);
1771 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
1772 void flush_hash_hugepage(unsigned long vsid, unsigned long addr,
1773 pmd_t *pmdp, unsigned int psize, int ssize,
1774 unsigned long flags)
1776 int i, max_hpte_count, valid;
1777 unsigned long s_addr;
1778 unsigned char *hpte_slot_array;
1779 unsigned long hidx, shift, vpn, hash, slot;
1780 int local = flags & HPTE_LOCAL_UPDATE;
1782 s_addr = addr & HPAGE_PMD_MASK;
1783 hpte_slot_array = get_hpte_slot_array(pmdp);
1785 * IF we try to do a HUGE PTE update after a withdraw is done.
1786 * we will find the below NULL. This happens when we do
1789 if (!hpte_slot_array)
1792 if (mmu_hash_ops.hugepage_invalidate) {
1793 mmu_hash_ops.hugepage_invalidate(vsid, s_addr, hpte_slot_array,
1794 psize, ssize, local);
1798 * No bluk hpte removal support, invalidate each entry
1800 shift = mmu_psize_defs[psize].shift;
1801 max_hpte_count = HPAGE_PMD_SIZE >> shift;
1802 for (i = 0; i < max_hpte_count; i++) {
1804 * 8 bits per each hpte entries
1805 * 000| [ secondary group (one bit) | hidx (3 bits) | valid bit]
1807 valid = hpte_valid(hpte_slot_array, i);
1810 hidx = hpte_hash_index(hpte_slot_array, i);
1813 addr = s_addr + (i * (1ul << shift));
1814 vpn = hpt_vpn(addr, vsid, ssize);
1815 hash = hpt_hash(vpn, shift, ssize);
1816 if (hidx & _PTEIDX_SECONDARY)
1819 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1820 slot += hidx & _PTEIDX_GROUP_IX;
1821 mmu_hash_ops.hpte_invalidate(slot, vpn, psize,
1822 MMU_PAGE_16M, ssize, local);
1825 tm_flush_hash_page(local);
1827 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
1829 void flush_hash_range(unsigned long number, int local)
1831 if (mmu_hash_ops.flush_hash_range)
1832 mmu_hash_ops.flush_hash_range(number, local);
1835 struct ppc64_tlb_batch *batch =
1836 this_cpu_ptr(&ppc64_tlb_batch);
1838 for (i = 0; i < number; i++)
1839 flush_hash_page(batch->vpn[i], batch->pte[i],
1840 batch->psize, batch->ssize, local);
1845 * low_hash_fault is called when we the low level hash code failed
1846 * to instert a PTE due to an hypervisor error
1848 void low_hash_fault(struct pt_regs *regs, unsigned long address, int rc)
1850 enum ctx_state prev_state = exception_enter();
1852 if (user_mode(regs)) {
1853 #ifdef CONFIG_PPC_SUBPAGE_PROT
1855 _exception(SIGSEGV, regs, SEGV_ACCERR, address);
1858 _exception(SIGBUS, regs, BUS_ADRERR, address);
1860 bad_page_fault(regs, address, SIGBUS);
1862 exception_exit(prev_state);
1865 long hpte_insert_repeating(unsigned long hash, unsigned long vpn,
1866 unsigned long pa, unsigned long rflags,
1867 unsigned long vflags, int psize, int ssize)
1869 unsigned long hpte_group;
1873 hpte_group = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1875 /* Insert into the hash table, primary slot */
1876 slot = mmu_hash_ops.hpte_insert(hpte_group, vpn, pa, rflags, vflags,
1877 psize, psize, ssize);
1879 /* Primary is full, try the secondary */
1880 if (unlikely(slot == -1)) {
1881 hpte_group = (~hash & htab_hash_mask) * HPTES_PER_GROUP;
1882 slot = mmu_hash_ops.hpte_insert(hpte_group, vpn, pa, rflags,
1883 vflags | HPTE_V_SECONDARY,
1884 psize, psize, ssize);
1887 hpte_group = (hash & htab_hash_mask) *
1890 mmu_hash_ops.hpte_remove(hpte_group);
1898 #ifdef CONFIG_DEBUG_PAGEALLOC
1899 static void kernel_map_linear_page(unsigned long vaddr, unsigned long lmi)
1902 unsigned long vsid = get_kernel_vsid(vaddr, mmu_kernel_ssize);
1903 unsigned long vpn = hpt_vpn(vaddr, vsid, mmu_kernel_ssize);
1904 unsigned long mode = htab_convert_pte_flags(pgprot_val(PAGE_KERNEL));
1907 hash = hpt_hash(vpn, PAGE_SHIFT, mmu_kernel_ssize);
1909 /* Don't create HPTE entries for bad address */
1913 ret = hpte_insert_repeating(hash, vpn, __pa(vaddr), mode,
1915 mmu_linear_psize, mmu_kernel_ssize);
1918 spin_lock(&linear_map_hash_lock);
1919 BUG_ON(linear_map_hash_slots[lmi] & 0x80);
1920 linear_map_hash_slots[lmi] = ret | 0x80;
1921 spin_unlock(&linear_map_hash_lock);
1924 static void kernel_unmap_linear_page(unsigned long vaddr, unsigned long lmi)
1926 unsigned long hash, hidx, slot;
1927 unsigned long vsid = get_kernel_vsid(vaddr, mmu_kernel_ssize);
1928 unsigned long vpn = hpt_vpn(vaddr, vsid, mmu_kernel_ssize);
1930 hash = hpt_hash(vpn, PAGE_SHIFT, mmu_kernel_ssize);
1931 spin_lock(&linear_map_hash_lock);
1932 BUG_ON(!(linear_map_hash_slots[lmi] & 0x80));
1933 hidx = linear_map_hash_slots[lmi] & 0x7f;
1934 linear_map_hash_slots[lmi] = 0;
1935 spin_unlock(&linear_map_hash_lock);
1936 if (hidx & _PTEIDX_SECONDARY)
1938 slot = (hash & htab_hash_mask) * HPTES_PER_GROUP;
1939 slot += hidx & _PTEIDX_GROUP_IX;
1940 mmu_hash_ops.hpte_invalidate(slot, vpn, mmu_linear_psize,
1942 mmu_kernel_ssize, 0);
1945 void __kernel_map_pages(struct page *page, int numpages, int enable)
1947 unsigned long flags, vaddr, lmi;
1950 local_irq_save(flags);
1951 for (i = 0; i < numpages; i++, page++) {
1952 vaddr = (unsigned long)page_address(page);
1953 lmi = __pa(vaddr) >> PAGE_SHIFT;
1954 if (lmi >= linear_map_hash_count)
1957 kernel_map_linear_page(vaddr, lmi);
1959 kernel_unmap_linear_page(vaddr, lmi);
1961 local_irq_restore(flags);
1963 #endif /* CONFIG_DEBUG_PAGEALLOC */
1965 void hash__setup_initial_memory_limit(phys_addr_t first_memblock_base,
1966 phys_addr_t first_memblock_size)
1969 * We don't currently support the first MEMBLOCK not mapping 0
1970 * physical on those processors
1972 BUG_ON(first_memblock_base != 0);
1975 * On virtualized systems the first entry is our RMA region aka VRMA,
1976 * non-virtualized 64-bit hash MMU systems don't have a limitation
1977 * on real mode access.
1979 * For guests on platforms before POWER9, we clamp the it limit to 1G
1980 * to avoid some funky things such as RTAS bugs etc...
1982 * On POWER9 we limit to 1TB in case the host erroneously told us that
1983 * the RMA was >1TB. Effective address bits 0:23 are treated as zero
1984 * (meaning the access is aliased to zero i.e. addr = addr % 1TB)
1985 * for virtual real mode addressing and so it doesn't make sense to
1986 * have an area larger than 1TB as it can't be addressed.
1988 if (!early_cpu_has_feature(CPU_FTR_HVMODE)) {
1989 ppc64_rma_size = first_memblock_size;
1990 if (!early_cpu_has_feature(CPU_FTR_ARCH_300))
1991 ppc64_rma_size = min_t(u64, ppc64_rma_size, 0x40000000);
1993 ppc64_rma_size = min_t(u64, ppc64_rma_size,
1994 1UL << SID_SHIFT_1T);
1996 /* Finally limit subsequent allocations */
1997 memblock_set_current_limit(ppc64_rma_size);
1999 ppc64_rma_size = ULONG_MAX;
2003 #ifdef CONFIG_DEBUG_FS
2005 static int hpt_order_get(void *data, u64 *val)
2007 *val = ppc64_pft_size;
2011 static int hpt_order_set(void *data, u64 val)
2015 if (!mmu_hash_ops.resize_hpt)
2019 ret = mmu_hash_ops.resize_hpt(val);
2025 DEFINE_DEBUGFS_ATTRIBUTE(fops_hpt_order, hpt_order_get, hpt_order_set, "%llu\n");
2027 static int __init hash64_debugfs(void)
2029 debugfs_create_file("hpt_order", 0600, powerpc_debugfs_root, NULL,
2033 machine_device_initcall(pseries, hash64_debugfs);
2034 #endif /* CONFIG_DEBUG_FS */
2036 void __init print_system_hash_info(void)
2038 pr_info("ppc64_pft_size = 0x%llx\n", ppc64_pft_size);
2041 pr_info("htab_hash_mask = 0x%lx\n", htab_hash_mask);