arch/powerpc/mm/tlb_hash64.c

   1 /*
   2  * This file contains the routines for flushing entries from the
   3  * TLB and MMU hash table.
   4  *
   5  *  Derived from arch/ppc64/mm/init.c:
   6  *    Copyright (C) 1995-1996 Gary Thomas (gdt@linuxppc.org)
   7  *
   8  *  Modifications by Paul Mackerras (PowerMac) (paulus@cs.anu.edu.au)
   9  *  and Cort Dougan (PReP) (cort@cs.nmt.edu)
  10  *    Copyright (C) 1996 Paul Mackerras
  11  *
  12  *  Derived from "arch/i386/mm/init.c"
  13  *    Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
  14  *
  15  *  Dave Engebretsen <engebret@us.ibm.com>
  16  *      Rework for PPC64 port.
  17  *
  18  *  This program is free software; you can redistribute it and/or
  19  *  modify it under the terms of the GNU General Public License
  20  *  as published by the Free Software Foundation; either version
  21  *  2 of the License, or (at your option) any later version.
  22  */
  23
  24 #include <linux/kernel.h>
  25 #include <linux/mm.h>
  26 #include <linux/percpu.h>
  27 #include <linux/hardirq.h>
  28 #include <asm/pgalloc.h>
  29 #include <asm/tlbflush.h>
  30 #include <asm/tlb.h>
  31 #include <asm/bug.h>
  32
  33 #include <trace/events/thp.h>
  34
  35 DEFINE_PER_CPU(struct ppc64_tlb_batch, ppc64_tlb_batch);
  36
  37 /*
  38  * A linux PTE was changed and the corresponding hash table entry
  39  * neesd to be flushed. This function will either perform the flush
  40  * immediately or will batch it up if the current CPU has an active
  41  * batch on it.
  42  */
  43 void hpte_need_flush(struct mm_struct *mm, unsigned long addr,
  44                      pte_t *ptep, unsigned long pte, int huge)
  45 {
  46         unsigned long vpn;
  47         struct ppc64_tlb_batch *batch = &get_cpu_var(ppc64_tlb_batch);
  48         unsigned long vsid;
  49         unsigned int psize;
  50         int ssize;
  51         real_pte_t rpte;
  52         int i;
  53
  54         i = batch->index;
  55
  56         /* Get page size (maybe move back to caller).
  57          *
  58          * NOTE: when using special 64K mappings in 4K environment like
  59          * for SPEs, we obtain the page size from the slice, which thus
  60          * must still exist (and thus the VMA not reused) at the time
  61          * of this call
  62          */
  63         if (huge) {
  64 #ifdef CONFIG_HUGETLB_PAGE
  65                 psize = get_slice_psize(mm, addr);
  66                 /* Mask the address for the correct page size */
  67                 addr &= ~((1UL << mmu_psize_defs[psize].shift) - 1);
  68 #else
  69                 BUG();
  70                 psize = pte_pagesize_index(mm, addr, pte); /* shutup gcc */
  71 #endif
  72         } else {
  73                 psize = pte_pagesize_index(mm, addr, pte);
  74                 /* Mask the address for the standard page size.  If we
  75                  * have a 64k page kernel, but the hardware does not
  76                  * support 64k pages, this might be different from the
  77                  * hardware page size encoded in the slice table. */
  78                 addr &= PAGE_MASK;
  79         }
  80
  81
  82         /* Build full vaddr */
  83         if (!is_kernel_addr(addr)) {
  84                 ssize = user_segment_size(addr);
  85                 vsid = get_vsid(mm->context.id, addr, ssize);
  86         } else {
  87                 vsid = get_kernel_vsid(addr, mmu_kernel_ssize);
  88                 ssize = mmu_kernel_ssize;
  89         }
  90         WARN_ON(vsid == 0);
  91         vpn = hpt_vpn(addr, vsid, ssize);
  92         rpte = __real_pte(__pte(pte), ptep);
  93
  94         /*
  95          * Check if we have an active batch on this CPU. If not, just
  96          * flush now and return. For now, we don global invalidates
  97          * in that case, might be worth testing the mm cpu mask though
  98          * and decide to use local invalidates instead...
  99          */
 100         if (!batch->active) {
 101                 flush_hash_page(vpn, rpte, psize, ssize, 0);
 102                 put_cpu_var(ppc64_tlb_batch);
 103                 return;
 104         }
 105
 106         /*
 107          * This can happen when we are in the middle of a TLB batch and
 108          * we encounter memory pressure (eg copy_page_range when it tries
 109          * to allocate a new pte). If we have to reclaim memory and end
 110          * up scanning and resetting referenced bits then our batch context
 111          * will change mid stream.
 112          *
 113          * We also need to ensure only one page size is present in a given
 114          * batch
 115          */
 116         if (i != 0 && (mm != batch->mm || batch->psize != psize ||
 117                        batch->ssize != ssize)) {
 118                 __flush_tlb_pending(batch);
 119                 i = 0;
 120         }
 121         if (i == 0) {
 122                 batch->mm = mm;
 123                 batch->psize = psize;
 124                 batch->ssize = ssize;
 125         }
 126         batch->pte[i] = rpte;
 127         batch->vpn[i] = vpn;
 128         batch->index = ++i;
 129         if (i >= PPC64_TLB_BATCH_NR)
 130                 __flush_tlb_pending(batch);
 131         put_cpu_var(ppc64_tlb_batch);
 132 }
 133
 134 /*
 135  * This function is called when terminating an mmu batch or when a batch
 136  * is full. It will perform the flush of all the entries currently stored
 137  * in a batch.
 138  *
 139  * Must be called from within some kind of spinlock/non-preempt region...
 140  */
 141 void __flush_tlb_pending(struct ppc64_tlb_batch *batch)
 142 {
 143         const struct cpumask *tmp;
 144         int i, local = 0;
 145
 146         i = batch->index;
 147         tmp = cpumask_of(smp_processor_id());
 148         if (cpumask_equal(mm_cpumask(batch->mm), tmp))
 149                 local = 1;
 150         if (i == 1)
 151                 flush_hash_page(batch->vpn[0], batch->pte[0],
 152                                 batch->psize, batch->ssize, local);
 153         else
 154                 flush_hash_range(i, local);
 155         batch->index = 0;
 156 }
 157
 158 void tlb_flush(struct mmu_gather *tlb)
 159 {
 160         struct ppc64_tlb_batch *tlbbatch = &get_cpu_var(ppc64_tlb_batch);
 161
 162         /* If there's a TLB batch pending, then we must flush it because the
 163          * pages are going to be freed and we really don't want to have a CPU
 164          * access a freed page because it has a stale TLB
 165          */
 166         if (tlbbatch->index)
 167                 __flush_tlb_pending(tlbbatch);
 168
 169         put_cpu_var(ppc64_tlb_batch);
 170 }
 171
 172 /**
 173  * __flush_hash_table_range - Flush all HPTEs for a given address range
 174  *                            from the hash table (and the TLB). But keeps
 175  *                            the linux PTEs intact.
 176  *
 177  * @mm          : mm_struct of the target address space (generally init_mm)
 178  * @start       : starting address
 179  * @end         : ending address (not included in the flush)
 180  *
 181  * This function is mostly to be used by some IO hotplug code in order
 182  * to remove all hash entries from a given address range used to map IO
 183  * space on a removed PCI-PCI bidge without tearing down the full mapping
 184  * since 64K pages may overlap with other bridges when using 64K pages
 185  * with 4K HW pages on IO space.
 186  *
 187  * Because of that usage pattern, it is implemented for small size rather
 188  * than speed.
 189  */
 190 void __flush_hash_table_range(struct mm_struct *mm, unsigned long start,
 191                               unsigned long end)
 192 {
 193         int hugepage_shift;
 194         unsigned long flags;
 195
 196         start = _ALIGN_DOWN(start, PAGE_SIZE);
 197         end = _ALIGN_UP(end, PAGE_SIZE);
 198
 199         BUG_ON(!mm->pgd);
 200
 201         /* Note: Normally, we should only ever use a batch within a
 202          * PTE locked section. This violates the rule, but will work
 203          * since we don't actually modify the PTEs, we just flush the
 204          * hash while leaving the PTEs intact (including their reference
 205          * to being hashed). This is not the most performance oriented
 206          * way to do things but is fine for our needs here.
 207          */
 208         local_irq_save(flags);
 209         arch_enter_lazy_mmu_mode();
 210         for (; start < end; start += PAGE_SIZE) {
 211                 pte_t *ptep = find_linux_pte_or_hugepte(mm->pgd, start,
 212                                                         &hugepage_shift);
 213                 unsigned long pte;
 214
 215                 if (ptep == NULL)
 216                         continue;
 217                 pte = pte_val(*ptep);
 218                 if (hugepage_shift)
 219                         trace_hugepage_invalidate(start, pte_val(pte));
 220                 if (!(pte & _PAGE_HASHPTE))
 221                         continue;
 222                 if (unlikely(hugepage_shift && pmd_trans_huge(*(pmd_t *)pte)))
 223                         hpte_do_hugepage_flush(mm, start, (pmd_t *)ptep, pte);
 224                 else
 225                         hpte_need_flush(mm, start, ptep, pte, 0);
 226         }
 227         arch_leave_lazy_mmu_mode();
 228         local_irq_restore(flags);
 229 }
 230
 231 void flush_tlb_pmd_range(struct mm_struct *mm, pmd_t *pmd, unsigned long addr)
 232 {
 233         pte_t *pte;
 234         pte_t *start_pte;
 235         unsigned long flags;
 236
 237         addr = _ALIGN_DOWN(addr, PMD_SIZE);
 238         /* Note: Normally, we should only ever use a batch within a
 239          * PTE locked section. This violates the rule, but will work
 240          * since we don't actually modify the PTEs, we just flush the
 241          * hash while leaving the PTEs intact (including their reference
 242          * to being hashed). This is not the most performance oriented
 243          * way to do things but is fine for our needs here.
 244          */
 245         local_irq_save(flags);
 246         arch_enter_lazy_mmu_mode();
 247         start_pte = pte_offset_map(pmd, addr);
 248         for (pte = start_pte; pte < start_pte + PTRS_PER_PTE; pte++) {
 249                 unsigned long pteval = pte_val(*pte);
 250                 if (pteval & _PAGE_HASHPTE)
 251                         hpte_need_flush(mm, addr, pte, pteval, 0);
 252                 addr += PAGE_SIZE;
 253         }
 254         arch_leave_lazy_mmu_mode();
 255         local_irq_restore(flags);
 256 }