2 * PowerPC64 SLB support.
5 * Based on earlier code written by:
6 * Dave Engebretsen and Mike Corrigan {engebret|mikejc}@us.ibm.com
7 * Copyright (c) 2001 Dave Engebretsen
11 * This program is free software; you can redistribute it and/or
12 * modify it under the terms of the GNU General Public License
13 * as published by the Free Software Foundation; either version
14 * 2 of the License, or (at your option) any later version.
17 #include <asm/asm-prototypes.h>
18 #include <asm/pgtable.h>
20 #include <asm/mmu_context.h>
22 #include <asm/ppc-opcode.h>
23 #include <asm/cputable.h>
24 #include <asm/cacheflush.h>
26 #include <linux/compiler.h>
27 #include <linux/context_tracking.h>
28 #include <linux/mm_types.h>
31 #include <asm/code-patching.h>
34 LINEAR_INDEX = 0, /* Kernel linear map (0xc000000000000000) */
35 KSTACK_INDEX = 1, /* Kernel stack map */
38 static long slb_allocate_user(struct mm_struct *mm, unsigned long ea);
40 #define slb_esid_mask(ssize) \
41 (((ssize) == MMU_SEGSIZE_256M)? ESID_MASK: ESID_MASK_1T)
43 static inline unsigned long mk_esid_data(unsigned long ea, int ssize,
46 return (ea & slb_esid_mask(ssize)) | SLB_ESID_V | index;
49 static inline unsigned long __mk_vsid_data(unsigned long vsid, int ssize,
52 return (vsid << slb_vsid_shift(ssize)) | flags |
53 ((unsigned long) ssize << SLB_VSID_SSIZE_SHIFT);
56 static inline unsigned long mk_vsid_data(unsigned long ea, int ssize,
59 return __mk_vsid_data(get_kernel_vsid(ea, ssize), ssize, flags);
62 static void assert_slb_presence(bool present, unsigned long ea)
64 #ifdef CONFIG_DEBUG_VM
67 WARN_ON_ONCE(mfmsr() & MSR_EE);
69 if (!cpu_has_feature(CPU_FTR_ARCH_206))
72 asm volatile(__PPC_SLBFEE_DOT(%0, %1) : "=r"(tmp) : "r"(ea) : "cr0");
74 WARN_ON(present == (tmp == 0));
78 static inline void slb_shadow_update(unsigned long ea, int ssize,
82 struct slb_shadow *p = get_slb_shadow();
85 * Clear the ESID first so the entry is not valid while we are
86 * updating it. No write barriers are needed here, provided
87 * we only update the current CPU's SLB shadow buffer.
89 WRITE_ONCE(p->save_area[index].esid, 0);
90 WRITE_ONCE(p->save_area[index].vsid, cpu_to_be64(mk_vsid_data(ea, ssize, flags)));
91 WRITE_ONCE(p->save_area[index].esid, cpu_to_be64(mk_esid_data(ea, ssize, index)));
94 static inline void slb_shadow_clear(enum slb_index index)
96 WRITE_ONCE(get_slb_shadow()->save_area[index].esid, cpu_to_be64(index));
99 static inline void create_shadowed_slbe(unsigned long ea, int ssize,
101 enum slb_index index)
104 * Updating the shadow buffer before writing the SLB ensures
105 * we don't get a stale entry here if we get preempted by PHYP
106 * between these two statements.
108 slb_shadow_update(ea, ssize, flags, index);
110 assert_slb_presence(false, ea);
111 asm volatile("slbmte %0,%1" :
112 : "r" (mk_vsid_data(ea, ssize, flags)),
113 "r" (mk_esid_data(ea, ssize, index))
118 * Insert bolted entries into SLB (which may not be empty, so don't clear
121 void __slb_restore_bolted_realmode(void)
123 struct slb_shadow *p = get_slb_shadow();
124 enum slb_index index;
126 /* No isync needed because realmode. */
127 for (index = 0; index < SLB_NUM_BOLTED; index++) {
128 asm volatile("slbmte %0,%1" :
129 : "r" (be64_to_cpu(p->save_area[index].vsid)),
130 "r" (be64_to_cpu(p->save_area[index].esid)));
133 assert_slb_presence(true, local_paca->kstack);
137 * Insert the bolted entries into an empty SLB.
139 void slb_restore_bolted_realmode(void)
141 __slb_restore_bolted_realmode();
142 get_paca()->slb_cache_ptr = 0;
144 get_paca()->slb_kern_bitmap = (1U << SLB_NUM_BOLTED) - 1;
145 get_paca()->slb_used_bitmap = get_paca()->slb_kern_bitmap;
149 * This flushes all SLB entries including 0, so it must be realmode.
151 void slb_flush_all_realmode(void)
153 asm volatile("slbmte %0,%0; slbia" : : "r" (0));
157 * This flushes non-bolted entries, it can be run in virtual mode. Must
158 * be called with interrupts disabled.
160 void slb_flush_and_restore_bolted(void)
162 struct slb_shadow *p = get_slb_shadow();
164 BUILD_BUG_ON(SLB_NUM_BOLTED != 2);
166 WARN_ON(!irqs_disabled());
169 * We can't take a PMU exception in the following code, so hard
170 * disable interrupts.
174 asm volatile("isync\n"
178 :: "r" (be64_to_cpu(p->save_area[KSTACK_INDEX].vsid)),
179 "r" (be64_to_cpu(p->save_area[KSTACK_INDEX].esid))
181 assert_slb_presence(true, get_paca()->kstack);
183 get_paca()->slb_cache_ptr = 0;
185 get_paca()->slb_kern_bitmap = (1U << SLB_NUM_BOLTED) - 1;
186 get_paca()->slb_used_bitmap = get_paca()->slb_kern_bitmap;
189 void slb_save_contents(struct slb_entry *slb_ptr)
194 /* Save slb_cache_ptr value. */
195 get_paca()->slb_save_cache_ptr = get_paca()->slb_cache_ptr;
200 for (i = 0; i < mmu_slb_size; i++) {
201 asm volatile("slbmfee %0,%1" : "=r" (e) : "r" (i));
202 asm volatile("slbmfev %0,%1" : "=r" (v) : "r" (i));
209 void slb_dump_contents(struct slb_entry *slb_ptr)
218 pr_err("SLB contents of cpu 0x%x\n", smp_processor_id());
219 pr_err("Last SLB entry inserted at slot %d\n", get_paca()->stab_rr);
221 for (i = 0; i < mmu_slb_size; i++) {
229 pr_err("%02d %016lx %016lx\n", i, e, v);
231 if (!(e & SLB_ESID_V)) {
235 llp = v & SLB_VSID_LLP;
236 if (v & SLB_VSID_B_1T) {
237 pr_err(" 1T ESID=%9lx VSID=%13lx LLP:%3lx\n",
239 (v & ~SLB_VSID_B) >> SLB_VSID_SHIFT_1T, llp);
241 pr_err(" 256M ESID=%9lx VSID=%13lx LLP:%3lx\n",
243 (v & ~SLB_VSID_B) >> SLB_VSID_SHIFT, llp);
246 pr_err("----------------------------------\n");
248 /* Dump slb cache entires as well. */
249 pr_err("SLB cache ptr value = %d\n", get_paca()->slb_save_cache_ptr);
250 pr_err("Valid SLB cache entries:\n");
251 n = min_t(int, get_paca()->slb_save_cache_ptr, SLB_CACHE_ENTRIES);
252 for (i = 0; i < n; i++)
253 pr_err("%02d EA[0-35]=%9x\n", i, get_paca()->slb_cache[i]);
254 pr_err("Rest of SLB cache entries:\n");
255 for (i = n; i < SLB_CACHE_ENTRIES; i++)
256 pr_err("%02d EA[0-35]=%9x\n", i, get_paca()->slb_cache[i]);
259 void slb_vmalloc_update(void)
262 * vmalloc is not bolted, so just have to flush non-bolted.
264 slb_flush_and_restore_bolted();
267 static bool preload_hit(struct thread_info *ti, unsigned long esid)
271 for (i = 0; i < ti->slb_preload_nr; i++) {
274 idx = (ti->slb_preload_tail + i) % SLB_PRELOAD_NR;
275 if (esid == ti->slb_preload_esid[idx])
281 static bool preload_add(struct thread_info *ti, unsigned long ea)
286 if (mmu_has_feature(MMU_FTR_1T_SEGMENT)) {
287 /* EAs are stored >> 28 so 256MB segments don't need clearing */
288 if (ea & ESID_MASK_1T)
292 esid = ea >> SID_SHIFT;
294 if (preload_hit(ti, esid))
297 idx = (ti->slb_preload_tail + ti->slb_preload_nr) % SLB_PRELOAD_NR;
298 ti->slb_preload_esid[idx] = esid;
299 if (ti->slb_preload_nr == SLB_PRELOAD_NR)
300 ti->slb_preload_tail = (ti->slb_preload_tail + 1) % SLB_PRELOAD_NR;
302 ti->slb_preload_nr++;
307 static void preload_age(struct thread_info *ti)
309 if (!ti->slb_preload_nr)
311 ti->slb_preload_nr--;
312 ti->slb_preload_tail = (ti->slb_preload_tail + 1) % SLB_PRELOAD_NR;
315 void slb_setup_new_exec(void)
317 struct thread_info *ti = current_thread_info();
318 struct mm_struct *mm = current->mm;
319 unsigned long exec = 0x10000000;
321 WARN_ON(irqs_disabled());
324 * preload cache can only be used to determine whether a SLB
325 * entry exists if it does not start to overflow.
327 if (ti->slb_preload_nr + 2 > SLB_PRELOAD_NR)
333 * We have no good place to clear the slb preload cache on exec,
334 * flush_thread is about the earliest arch hook but that happens
335 * after we switch to the mm and have aleady preloaded the SLBEs.
337 * For the most part that's probably okay to use entries from the
338 * previous exec, they will age out if unused. It may turn out to
339 * be an advantage to clear the cache before switching to it,
344 * preload some userspace segments into the SLB.
345 * Almost all 32 and 64bit PowerPC executables are linked at
346 * 0x10000000 so it makes sense to preload this segment.
348 if (!is_kernel_addr(exec)) {
349 if (preload_add(ti, exec))
350 slb_allocate_user(mm, exec);
353 /* Libraries and mmaps. */
354 if (!is_kernel_addr(mm->mmap_base)) {
355 if (preload_add(ti, mm->mmap_base))
356 slb_allocate_user(mm, mm->mmap_base);
360 asm volatile("isync" : : : "memory");
365 void preload_new_slb_context(unsigned long start, unsigned long sp)
367 struct thread_info *ti = current_thread_info();
368 struct mm_struct *mm = current->mm;
369 unsigned long heap = mm->start_brk;
371 WARN_ON(irqs_disabled());
374 if (ti->slb_preload_nr + 3 > SLB_PRELOAD_NR)
379 /* Userspace entry address. */
380 if (!is_kernel_addr(start)) {
381 if (preload_add(ti, start))
382 slb_allocate_user(mm, start);
385 /* Top of stack, grows down. */
386 if (!is_kernel_addr(sp)) {
387 if (preload_add(ti, sp))
388 slb_allocate_user(mm, sp);
391 /* Bottom of heap, grows up. */
392 if (heap && !is_kernel_addr(heap)) {
393 if (preload_add(ti, heap))
394 slb_allocate_user(mm, heap);
398 asm volatile("isync" : : : "memory");
404 /* Flush all user entries from the segment table of the current processor. */
405 void switch_slb(struct task_struct *tsk, struct mm_struct *mm)
407 struct thread_info *ti = task_thread_info(tsk);
411 * We need interrupts hard-disabled here, not just soft-disabled,
412 * so that a PMU interrupt can't occur, which might try to access
413 * user memory (to get a stack trace) and possible cause an SLB miss
414 * which would update the slb_cache/slb_cache_ptr fields in the PACA.
417 asm volatile("isync" : : : "memory");
418 if (cpu_has_feature(CPU_FTR_ARCH_300)) {
420 * SLBIA IH=3 invalidates all Class=1 SLBEs and their
421 * associated lookaside structures, which matches what
422 * switch_slb wants. So ARCH_300 does not use the slb
425 asm volatile(PPC_SLBIA(3));
427 unsigned long offset = get_paca()->slb_cache_ptr;
429 if (!mmu_has_feature(MMU_FTR_NO_SLBIE_B) &&
430 offset <= SLB_CACHE_ENTRIES) {
431 unsigned long slbie_data = 0;
433 for (i = 0; i < offset; i++) {
437 get_paca()->slb_cache[i] << SID_SHIFT;
439 * Could assert_slb_presence(true) here, but
440 * hypervisor or machine check could have come
441 * in and removed the entry at this point.
445 slbie_data |= user_segment_size(slbie_data)
446 << SLBIE_SSIZE_SHIFT;
447 slbie_data |= SLBIE_C; /* user slbs have C=1 */
448 asm volatile("slbie %0" : : "r" (slbie_data));
451 /* Workaround POWER5 < DD2.1 issue */
452 if (!cpu_has_feature(CPU_FTR_ARCH_207S) && offset == 1)
453 asm volatile("slbie %0" : : "r" (slbie_data));
456 struct slb_shadow *p = get_slb_shadow();
457 unsigned long ksp_esid_data =
458 be64_to_cpu(p->save_area[KSTACK_INDEX].esid);
459 unsigned long ksp_vsid_data =
460 be64_to_cpu(p->save_area[KSTACK_INDEX].vsid);
462 asm volatile(PPC_SLBIA(1) "\n"
465 :: "r"(ksp_vsid_data),
468 get_paca()->slb_kern_bitmap = (1U << SLB_NUM_BOLTED) - 1;
471 get_paca()->slb_cache_ptr = 0;
473 get_paca()->slb_used_bitmap = get_paca()->slb_kern_bitmap;
478 * We gradually age out SLBs after a number of context switches to
479 * reduce reload overhead of unused entries (like we do with FP/VEC
480 * reload). Each time we wrap 256 switches, take an entry out of the
483 tsk->thread.load_slb++;
484 if (!tsk->thread.load_slb) {
485 unsigned long pc = KSTK_EIP(tsk);
491 for (i = 0; i < ti->slb_preload_nr; i++) {
495 idx = (ti->slb_preload_tail + i) % SLB_PRELOAD_NR;
496 ea = (unsigned long)ti->slb_preload_esid[idx] << SID_SHIFT;
498 slb_allocate_user(mm, ea);
502 * Synchronize slbmte preloads with possible subsequent user memory
503 * address accesses by the kernel (user mode won't happen until
504 * rfid, which is safe).
506 asm volatile("isync" : : : "memory");
509 void slb_set_size(u16 size)
514 void slb_initialize(void)
516 unsigned long linear_llp, vmalloc_llp, io_llp;
517 unsigned long lflags;
518 static int slb_encoding_inited;
519 #ifdef CONFIG_SPARSEMEM_VMEMMAP
520 unsigned long vmemmap_llp;
523 /* Prepare our SLB miss handler based on our page size */
524 linear_llp = mmu_psize_defs[mmu_linear_psize].sllp;
525 io_llp = mmu_psize_defs[mmu_io_psize].sllp;
526 vmalloc_llp = mmu_psize_defs[mmu_vmalloc_psize].sllp;
527 get_paca()->vmalloc_sllp = SLB_VSID_KERNEL | vmalloc_llp;
528 #ifdef CONFIG_SPARSEMEM_VMEMMAP
529 vmemmap_llp = mmu_psize_defs[mmu_vmemmap_psize].sllp;
531 if (!slb_encoding_inited) {
532 slb_encoding_inited = 1;
533 pr_devel("SLB: linear LLP = %04lx\n", linear_llp);
534 pr_devel("SLB: io LLP = %04lx\n", io_llp);
535 #ifdef CONFIG_SPARSEMEM_VMEMMAP
536 pr_devel("SLB: vmemmap LLP = %04lx\n", vmemmap_llp);
540 get_paca()->stab_rr = SLB_NUM_BOLTED - 1;
541 get_paca()->slb_kern_bitmap = (1U << SLB_NUM_BOLTED) - 1;
542 get_paca()->slb_used_bitmap = get_paca()->slb_kern_bitmap;
544 lflags = SLB_VSID_KERNEL | linear_llp;
546 /* Invalidate the entire SLB (even entry 0) & all the ERATS */
547 asm volatile("isync":::"memory");
548 asm volatile("slbmte %0,%0"::"r" (0) : "memory");
549 asm volatile("isync; slbia; isync":::"memory");
550 create_shadowed_slbe(PAGE_OFFSET, mmu_kernel_ssize, lflags, LINEAR_INDEX);
552 /* For the boot cpu, we're running on the stack in init_thread_union,
553 * which is in the first segment of the linear mapping, and also
554 * get_paca()->kstack hasn't been initialized yet.
555 * For secondary cpus, we need to bolt the kernel stack entry now.
557 slb_shadow_clear(KSTACK_INDEX);
558 if (raw_smp_processor_id() != boot_cpuid &&
559 (get_paca()->kstack & slb_esid_mask(mmu_kernel_ssize)) > PAGE_OFFSET)
560 create_shadowed_slbe(get_paca()->kstack,
561 mmu_kernel_ssize, lflags, KSTACK_INDEX);
563 asm volatile("isync":::"memory");
566 static void slb_cache_update(unsigned long esid_data)
570 if (cpu_has_feature(CPU_FTR_ARCH_300))
571 return; /* ISAv3.0B and later does not use slb_cache */
574 * Now update slb cache entries
576 slb_cache_index = local_paca->slb_cache_ptr;
577 if (slb_cache_index < SLB_CACHE_ENTRIES) {
579 * We have space in slb cache for optimized switch_slb().
580 * Top 36 bits from esid_data as per ISA
582 local_paca->slb_cache[slb_cache_index++] = esid_data >> 28;
583 local_paca->slb_cache_ptr++;
586 * Our cache is full and the current cache content strictly
587 * doesn't indicate the active SLB conents. Bump the ptr
588 * so that switch_slb() will ignore the cache.
590 local_paca->slb_cache_ptr = SLB_CACHE_ENTRIES + 1;
594 static enum slb_index alloc_slb_index(bool kernel)
596 enum slb_index index;
599 * The allocation bitmaps can become out of synch with the SLB
600 * when the _switch code does slbie when bolting a new stack
601 * segment and it must not be anywhere else in the SLB. This leaves
602 * a kernel allocated entry that is unused in the SLB. With very
603 * large systems or small segment sizes, the bitmaps could slowly
604 * fill with these entries. They will eventually be cleared out
605 * by the round robin allocator in that case, so it's probably not
606 * worth accounting for.
610 * SLBs beyond 32 entries are allocated with stab_rr only
611 * POWER7/8/9 have 32 SLB entries, this could be expanded if a
612 * future CPU has more.
614 if (local_paca->slb_used_bitmap != U32_MAX) {
615 index = ffz(local_paca->slb_used_bitmap);
616 local_paca->slb_used_bitmap |= 1U << index;
618 local_paca->slb_kern_bitmap |= 1U << index;
620 /* round-robin replacement of slb starting at SLB_NUM_BOLTED. */
621 index = local_paca->stab_rr;
622 if (index < (mmu_slb_size - 1))
625 index = SLB_NUM_BOLTED;
626 local_paca->stab_rr = index;
629 local_paca->slb_kern_bitmap |= 1U << index;
631 local_paca->slb_kern_bitmap &= ~(1U << index);
634 BUG_ON(index < SLB_NUM_BOLTED);
639 static long slb_insert_entry(unsigned long ea, unsigned long context,
640 unsigned long flags, int ssize, bool kernel)
643 unsigned long vsid_data, esid_data;
644 enum slb_index index;
646 vsid = get_vsid(context, ea, ssize);
651 * There must not be a kernel SLB fault in alloc_slb_index or before
652 * slbmte here or the allocation bitmaps could get out of whack with
655 * User SLB faults or preloads take this path which might get inlined
656 * into the caller, so add compiler barriers here to ensure unsafe
657 * memory accesses do not come between.
661 index = alloc_slb_index(kernel);
663 vsid_data = __mk_vsid_data(vsid, ssize, flags);
664 esid_data = mk_esid_data(ea, ssize, index);
667 * No need for an isync before or after this slbmte. The exception
668 * we enter with and the rfid we exit with are context synchronizing.
669 * User preloads should add isync afterwards in case the kernel
670 * accesses user memory before it returns to userspace with rfid.
672 assert_slb_presence(false, ea);
673 asm volatile("slbmte %0, %1" : : "r" (vsid_data), "r" (esid_data));
678 slb_cache_update(esid_data);
683 static long slb_allocate_kernel(unsigned long ea, unsigned long id)
685 unsigned long context;
689 if (id == KERNEL_REGION_ID) {
691 /* We only support upto MAX_PHYSMEM_BITS */
692 if ((ea & ~REGION_MASK) > (1UL << MAX_PHYSMEM_BITS))
695 flags = SLB_VSID_KERNEL | mmu_psize_defs[mmu_linear_psize].sllp;
697 #ifdef CONFIG_SPARSEMEM_VMEMMAP
698 } else if (id == VMEMMAP_REGION_ID) {
700 if ((ea & ~REGION_MASK) >= (1ULL << MAX_EA_BITS_PER_CONTEXT))
703 flags = SLB_VSID_KERNEL | mmu_psize_defs[mmu_vmemmap_psize].sllp;
705 } else if (id == VMALLOC_REGION_ID) {
707 if ((ea & ~REGION_MASK) >= (1ULL << MAX_EA_BITS_PER_CONTEXT))
710 if (ea < H_VMALLOC_END)
711 flags = local_paca->vmalloc_sllp;
713 flags = SLB_VSID_KERNEL | mmu_psize_defs[mmu_io_psize].sllp;
718 ssize = MMU_SEGSIZE_1T;
719 if (!mmu_has_feature(MMU_FTR_1T_SEGMENT))
720 ssize = MMU_SEGSIZE_256M;
722 context = get_kernel_context(ea);
723 return slb_insert_entry(ea, context, flags, ssize, true);
726 static long slb_allocate_user(struct mm_struct *mm, unsigned long ea)
728 unsigned long context;
734 * consider this as bad access if we take a SLB miss
735 * on an address above addr limit.
737 if (ea >= mm->context.slb_addr_limit)
740 context = get_user_context(&mm->context, ea);
744 if (unlikely(ea >= H_PGTABLE_RANGE)) {
749 ssize = user_segment_size(ea);
751 bpsize = get_slice_psize(mm, ea);
752 flags = SLB_VSID_USER | mmu_psize_defs[bpsize].sllp;
754 return slb_insert_entry(ea, context, flags, ssize, false);
757 long do_slb_fault(struct pt_regs *regs, unsigned long ea)
759 unsigned long id = REGION_ID(ea);
761 /* IRQs are not reconciled here, so can't check irqs_disabled */
762 VM_WARN_ON(mfmsr() & MSR_EE);
764 if (unlikely(!(regs->msr & MSR_RI)))
768 * SLB kernel faults must be very careful not to touch anything
769 * that is not bolted. E.g., PACA and global variables are okay,
770 * mm->context stuff is not.
772 * SLB user faults can access all of kernel memory, but must be
773 * careful not to touch things like IRQ state because it is not
774 * "reconciled" here. The difficulty is that we must use
775 * fast_exception_return to return from kernel SLB faults without
776 * looking at possible non-bolted memory. We could test user vs
777 * kernel faults in the interrupt handler asm and do a full fault,
778 * reconcile, ret_from_except for user faults which would make them
779 * first class kernel code. But for performance it's probably nicer
780 * if they go via fast_exception_return too.
782 if (id >= KERNEL_REGION_ID) {
784 #ifdef CONFIG_DEBUG_VM
785 /* Catch recursive kernel SLB faults. */
786 BUG_ON(local_paca->in_kernel_slb_handler);
787 local_paca->in_kernel_slb_handler = 1;
789 err = slb_allocate_kernel(ea, id);
790 #ifdef CONFIG_DEBUG_VM
791 local_paca->in_kernel_slb_handler = 0;
795 struct mm_struct *mm = current->mm;
801 err = slb_allocate_user(mm, ea);
803 preload_add(current_thread_info(), ea);
809 void do_bad_slb_fault(struct pt_regs *regs, unsigned long ea, long err)
811 if (err == -EFAULT) {
813 _exception(SIGSEGV, regs, SEGV_BNDERR, ea);
815 bad_page_fault(regs, ea, SIGSEGV);
816 } else if (err == -EINVAL) {
817 unrecoverable_exception(regs);