src/patches/suse-2.6.27.31/patches.arch/x86-fix-kmap-contig.patch

   1 From: Nick Piggin <npiggin@suse.de>
   2 Subject: x86: contiguous kmap fix
   3 References: bnc#449812
   4 Patch-mainline: perhaps 2.6.30 (maybe .29)
   5
   6 The early fixmap pmd entry inserted at the very top of the KVA is casing the
   7 subsequent fixmap mapping code to not provide physically linear pte pages over
   8 the kmap atomic portion of the fixmap (which relies on said property to calculate
   9 pte address).
  10
  11 This has caused weird boot failures in kmap_atomic much later in the boot
  12 process (initial userspace faults) on a 32-bit PAE system with a larger number
  13 of CPUs (smaller CPU counts tend not to run over into the next page so don't
  14 show up the problem).
  15
  16 Solve this by attempting to clear out the page table, and copy any of its
  17 entries to the new one. Also, add a bug if a nonlinear condition is encountered
  18 and can't be resolved, which might save some hours of debugging if this fragile
  19 scheme ever breaks again...
  20
  21 Signed-off-by: Nick Piggin <npiggin@suse.de>
  22
  23 Once we have such logic, we can also use it to eliminate the early ioremap
  24 trickery around the page table setup for the fixmap area. This also fixes
  25 potential issues with FIX_* entries sharing the leaf page table with the early
  26 ioremap ones getting discarded by early_ioremap_clear() and not restored by
  27 early_ioremap_reset(). It at once eliminates the temporary (and configuration,
  28 namely NR_CPUS, dependent) unavailability of early fixed mappings during the
  29 time the fixmap area page tables get constructed.
  30
  31 Finally, also replace the hard coded calculation of the initial table space
  32 needed for the fixmap area with a proper one, allowing kernels configured for
  33 large CPU counts to actually boot.
  34
  35 Signed-off-by: Jan Beulich <jbeulich@novell.com>
  36
  37 ---
  38  arch/x86/mm/init_32.c |   49 ++++++++++++++++++++++++++++++++++++++++++++++---
  39  arch/x86/mm/ioremap.c |   25 -------------------------
  40  include/asm-x86/io.h  |    2 --
  41  3 files changed, 46 insertions(+), 30 deletions(-)
  42
  43 --- a/arch/x86/mm/init_32.c
  44 +++ b/arch/x86/mm/init_32.c
  45 @@ -137,6 +137,48 @@ static pte_t * __init one_page_table_ini
  46         return pte_offset_kernel(pmd, 0);
  47  }
  48
  49 +static pte_t *__init page_table_kmap_check(pte_t *pte, pmd_t *pmd,
  50 +                                          unsigned long vaddr, pte_t *lastpte)
  51 +{
  52 +#ifdef CONFIG_HIGHMEM
  53 +       /*
  54 +        * Something (early fixmap) may already have put a pte
  55 +        * page here, which causes the page table allocation
  56 +        * to become nonlinear. Attempt to fix it, and if it
  57 +        * is still nonlinear then we have to bug.
  58 +        */
  59 +       int pmd_idx_kmap_begin = fix_to_virt(FIX_KMAP_END) >> PMD_SHIFT;
  60 +       int pmd_idx_kmap_end = fix_to_virt(FIX_KMAP_BEGIN) >> PMD_SHIFT;
  61 +
  62 +       if (pmd_idx_kmap_begin != pmd_idx_kmap_end
  63 +           && (vaddr >> PMD_SHIFT) >= pmd_idx_kmap_begin
  64 +           && (vaddr >> PMD_SHIFT) <= pmd_idx_kmap_end
  65 +           && ((__pa(pte) >> PAGE_SHIFT) < table_start
  66 +               || (__pa(pte) >> PAGE_SHIFT) >= table_end)) {
  67 +               pte_t *newpte;
  68 +               unsigned long phys;
  69 +               int i;
  70 +
  71 +               BUG_ON(after_init_bootmem);
  72 +               newpte = alloc_low_page(&phys);
  73 +               for (i = 0; i < PTRS_PER_PTE; i++)
  74 +                       set_pte(newpte + i, pte[i]);
  75 +
  76 +               paravirt_alloc_pte(&init_mm, __pa(newpte) >> PAGE_SHIFT);
  77 +               set_pmd(pmd, __pmd(__pa(newpte)|_PAGE_TABLE));
  78 +               BUG_ON(newpte != pte_offset_kernel(pmd, 0));
  79 +               __flush_tlb_all();
  80 +
  81 +               paravirt_release_pte(__pa(pte) >> PAGE_SHIFT);
  82 +               pte = newpte;
  83 +       }
  84 +       BUG_ON(vaddr < fix_to_virt(FIX_KMAP_BEGIN - 1)
  85 +              && vaddr > fix_to_virt(FIX_KMAP_END)
  86 +              && lastpte && lastpte + PTRS_PER_PTE != pte);
  87 +#endif
  88 +       return pte;
  89 +}
  90 +
  91  /*
  92   * This function initializes a certain range of kernel virtual memory
  93   * with new bootmem page tables, everywhere page tables are missing in
  94 @@ -153,6 +195,7 @@ page_table_range_init(unsigned long star
  95         unsigned long vaddr;
  96         pgd_t *pgd;
  97         pmd_t *pmd;
  98 +       pte_t *pte = NULL;
  99
 100         vaddr = start;
 101         pgd_idx = pgd_index(vaddr);
 102 @@ -164,7 +207,8 @@ page_table_range_init(unsigned long star
 103                 pmd = pmd + pmd_index(vaddr);
 104                 for (; (pmd_idx < PTRS_PER_PMD) && (vaddr != end);
 105                                                         pmd++, pmd_idx++) {
 106 -                       one_page_table_init(pmd);
 107 +                       pte = page_table_kmap_check(one_page_table_init(pmd),
 108 +                                                   pmd, vaddr, pte);
 109
 110                         vaddr += PMD_SIZE;
 111                 }
 112 @@ -447,7 +491,6 @@ static void __init early_ioremap_page_ta
 113          * Fixed mappings, only the page table structure has to be
 114          * created - mappings will be set by set_fixmap():
 115          */
 116 -       early_ioremap_clear();
 117         vaddr = __fix_to_virt(__end_of_fixed_addresses - 1) & PMD_MASK;
 118         end = (FIXADDR_TOP + PMD_SIZE - 1) & PMD_MASK;
 119         page_table_range_init(vaddr, end, pgd_base);
 120 @@ -740,7 +783,7 @@ static void __init find_early_table_spac
 121         tables += PAGE_ALIGN(ptes * sizeof(pte_t));
 122
 123         /* for fixmap */
 124 -       tables += PAGE_SIZE * 2;
 125 +       tables += PAGE_ALIGN(__end_of_fixed_addresses * sizeof(pte_t));
 126
 127         /*
 128          * RED-PEN putting page tables only on node 0 could
 129 --- a/arch/x86/mm/ioremap.c
 130 +++ b/arch/x86/mm/ioremap.c
 131 @@ -480,34 +480,9 @@ void __init early_ioremap_init(void)
 132         }
 133  }
 134
 135 -void __init early_ioremap_clear(void)
 136 -{
 137 -       pmd_t *pmd;
 138 -
 139 -       if (early_ioremap_debug)
 140 -               printk(KERN_INFO "early_ioremap_clear()\n");
 141 -
 142 -       pmd = early_ioremap_pmd(fix_to_virt(FIX_BTMAP_BEGIN));
 143 -       pmd_clear(pmd);
 144 -       paravirt_release_pte(__pa(bm_pte) >> PAGE_SHIFT);
 145 -       __flush_tlb_all();
 146 -}
 147 -
 148  void __init early_ioremap_reset(void)
 149  {
 150 -       enum fixed_addresses idx;
 151 -       unsigned long addr, phys;
 152 -       pte_t *pte;
 153 -
 154         after_paging_init = 1;
 155 -       for (idx = FIX_BTMAP_BEGIN; idx >= FIX_BTMAP_END; idx--) {
 156 -               addr = fix_to_virt(idx);
 157 -               pte = early_ioremap_pte(addr);
 158 -               if (pte_present(*pte)) {
 159 -                       phys = pte_val(*pte) & PAGE_MASK;
 160 -                       set_fixmap(idx, phys);
 161 -               }
 162 -       }
 163  }
 164
 165  static void __init __early_set_fixmap(enum fixed_addresses idx,
 166 --- a/include/asm-x86/io.h
 167 +++ b/include/asm-x86/io.h
 168 @@ -12,7 +12,6 @@
 169   */
 170  #ifndef __ASSEMBLY__
 171  extern void early_ioremap_init(void);
 172 -extern void early_ioremap_clear(void);
 173  extern void early_ioremap_reset(void);
 174  extern void *early_ioremap(unsigned long offset, unsigned long size);
 175  extern void early_iounmap(void *addr, unsigned long size);
 176 @@ -92,7 +91,6 @@ extern void __iomem *ioremap_wc(unsigned
 177   * A boot-time mapping is currently limited to at most 16 pages.
 178   */
 179  extern void early_ioremap_init(void);
 180 -extern void early_ioremap_clear(void);
 181  extern void early_ioremap_reset(void);
 182  extern void *early_ioremap(unsigned long offset, unsigned long size);
 183  extern void early_iounmap(void *addr, unsigned long size);