arch/tile/lib/memcpy_tile64.c

   1 /*
   2  * Copyright 2010 Tilera Corporation. All Rights Reserved.
   3  *
   4  *   This program is free software; you can redistribute it and/or
   5  *   modify it under the terms of the GNU General Public License
   6  *   as published by the Free Software Foundation, version 2.
   7  *
   8  *   This program is distributed in the hope that it will be useful, but
   9  *   WITHOUT ANY WARRANTY; without even the implied warranty of
  10  *   MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or
  11  *   NON INFRINGEMENT.  See the GNU General Public License for
  12  *   more details.
  13  */
  14
  15 #include <linux/string.h>
  16 #include <linux/smp.h>
  17 #include <linux/module.h>
  18 #include <linux/uaccess.h>
  19 #include <asm/fixmap.h>
  20 #include <asm/kmap_types.h>
  21 #include <asm/tlbflush.h>
  22 #include <hv/hypervisor.h>
  23 #include <arch/chip.h>
  24
  25
  26 #if !CHIP_HAS_COHERENT_LOCAL_CACHE()
  27
  28 /* Defined in memcpy.S */
  29 extern unsigned long __memcpy_asm(void *to, const void *from, unsigned long n);
  30 extern unsigned long __copy_to_user_inatomic_asm(
  31         void __user *to, const void *from, unsigned long n);
  32 extern unsigned long __copy_from_user_inatomic_asm(
  33         void *to, const void __user *from, unsigned long n);
  34 extern unsigned long __copy_from_user_zeroing_asm(
  35         void *to, const void __user *from, unsigned long n);
  36
  37 typedef unsigned long (*memcpy_t)(void *, const void *, unsigned long);
  38
  39 /* Size above which to consider TLB games for performance */
  40 #define LARGE_COPY_CUTOFF 2048
  41
  42 /* Communicate to the simulator what we are trying to do. */
  43 #define sim_allow_multiple_caching(b) \
  44   __insn_mtspr(SPR_SIM_CONTROL, \
  45    SIM_CONTROL_ALLOW_MULTIPLE_CACHING | ((b) << _SIM_CONTROL_OPERATOR_BITS))
  46
  47 /*
  48  * Copy memory by briefly enabling incoherent cacheline-at-a-time mode.
  49  *
  50  * We set up our own source and destination PTEs that we fully control.
  51  * This is the only way to guarantee that we don't race with another
  52  * thread that is modifying the PTE; we can't afford to try the
  53  * copy_{to,from}_user() technique of catching the interrupt, since
  54  * we must run with interrupts disabled to avoid the risk of some
  55  * other code seeing the incoherent data in our cache.  (Recall that
  56  * our cache is indexed by PA, so even if the other code doesn't use
  57  * our kmap_atomic virtual addresses, they'll still hit in cache using
  58  * the normal VAs that aren't supposed to hit in cache.)
  59  */
  60 static void memcpy_multicache(void *dest, const void *source,
  61                               pte_t dst_pte, pte_t src_pte, int len)
  62 {
  63         int idx;
  64         unsigned long flags, newsrc, newdst;
  65         pmd_t *pmdp;
  66         pte_t *ptep;
  67         int type0, type1;
  68         int cpu = get_cpu();
  69
  70         /*
  71          * Disable interrupts so that we don't recurse into memcpy()
  72          * in an interrupt handler, nor accidentally reference
  73          * the PA of the source from an interrupt routine.  Also
  74          * notify the simulator that we're playing games so we don't
  75          * generate spurious coherency warnings.
  76          */
  77         local_irq_save(flags);
  78         sim_allow_multiple_caching(1);
  79
  80         /* Set up the new dest mapping */
  81         type0 = kmap_atomic_idx_push();
  82         idx = FIX_KMAP_BEGIN + (KM_TYPE_NR * cpu) + type0;
  83         newdst = __fix_to_virt(idx) + ((unsigned long)dest & (PAGE_SIZE-1));
  84         pmdp = pmd_offset(pud_offset(pgd_offset_k(newdst), newdst), newdst);
  85         ptep = pte_offset_kernel(pmdp, newdst);
  86         if (pte_val(*ptep) != pte_val(dst_pte)) {
  87                 set_pte(ptep, dst_pte);
  88                 local_flush_tlb_page(NULL, newdst, PAGE_SIZE);
  89         }
  90
  91         /* Set up the new source mapping */
  92         type1 = kmap_atomic_idx_push();
  93         idx += (type0 - type1);
  94         src_pte = hv_pte_set_nc(src_pte);
  95         src_pte = hv_pte_clear_writable(src_pte);  /* be paranoid */
  96         newsrc = __fix_to_virt(idx) + ((unsigned long)source & (PAGE_SIZE-1));
  97         pmdp = pmd_offset(pud_offset(pgd_offset_k(newsrc), newsrc), newsrc);
  98         ptep = pte_offset_kernel(pmdp, newsrc);
  99         __set_pte(ptep, src_pte);   /* set_pte() would be confused by this */
 100         local_flush_tlb_page(NULL, newsrc, PAGE_SIZE);
 101
 102         /* Actually move the data. */
 103         __memcpy_asm((void *)newdst, (const void *)newsrc, len);
 104
 105         /*
 106          * Remap the source as locally-cached and not OLOC'ed so that
 107          * we can inval without also invaling the remote cpu's cache.
 108          * This also avoids known errata with inv'ing cacheable oloc data.
 109          */
 110         src_pte = hv_pte_set_mode(src_pte, HV_PTE_MODE_CACHE_NO_L3);
 111         src_pte = hv_pte_set_writable(src_pte); /* need write access for inv */
 112         __set_pte(ptep, src_pte);   /* set_pte() would be confused by this */
 113         local_flush_tlb_page(NULL, newsrc, PAGE_SIZE);
 114
 115         /*
 116          * Do the actual invalidation, covering the full L2 cache line
 117          * at the end since __memcpy_asm() is somewhat aggressive.
 118          */
 119         __inv_buffer((void *)newsrc, len);
 120
 121         /*
 122          * We're done: notify the simulator that all is back to normal,
 123          * and re-enable interrupts and pre-emption.
 124          */
 125         kmap_atomic_idx_pop();
 126         kmap_atomic_idx_pop();
 127         sim_allow_multiple_caching(0);
 128         local_irq_restore(flags);
 129         put_cpu();
 130 }
 131
 132 /*
 133  * Identify large copies from remotely-cached memory, and copy them
 134  * via memcpy_multicache() if they look good, otherwise fall back
 135  * to the particular kind of copying passed as the memcpy_t function.
 136  */
 137 static unsigned long fast_copy(void *dest, const void *source, int len,
 138                                memcpy_t func)
 139 {
 140         /*
 141          * Check if it's big enough to bother with.  We may end up doing a
 142          * small copy via TLB manipulation if we're near a page boundary,
 143          * but presumably we'll make it up when we hit the second page.
 144          */
 145         while (len >= LARGE_COPY_CUTOFF) {
 146                 int copy_size, bytes_left_on_page;
 147                 pte_t *src_ptep, *dst_ptep;
 148                 pte_t src_pte, dst_pte;
 149                 struct page *src_page, *dst_page;
 150
 151                 /* Is the source page oloc'ed to a remote cpu? */
 152 retry_source:
 153                 src_ptep = virt_to_pte(current->mm, (unsigned long)source);
 154                 if (src_ptep == NULL)
 155                         break;
 156                 src_pte = *src_ptep;
 157                 if (!hv_pte_get_present(src_pte) ||
 158                     !hv_pte_get_readable(src_pte) ||
 159                     hv_pte_get_mode(src_pte) != HV_PTE_MODE_CACHE_TILE_L3)
 160                         break;
 161                 if (get_remote_cache_cpu(src_pte) == smp_processor_id())
 162                         break;
 163                 src_page = pfn_to_page(pte_pfn(src_pte));
 164                 get_page(src_page);
 165                 if (pte_val(src_pte) != pte_val(*src_ptep)) {
 166                         put_page(src_page);
 167                         goto retry_source;
 168                 }
 169                 if (pte_huge(src_pte)) {
 170                         /* Adjust the PTE to correspond to a small page */
 171                         int pfn = pte_pfn(src_pte);
 172                         pfn += (((unsigned long)source & (HPAGE_SIZE-1))
 173                                 >> PAGE_SHIFT);
 174                         src_pte = pfn_pte(pfn, src_pte);
 175                         src_pte = pte_mksmall(src_pte);
 176                 }
 177
 178                 /* Is the destination page writable? */
 179 retry_dest:
 180                 dst_ptep = virt_to_pte(current->mm, (unsigned long)dest);
 181                 if (dst_ptep == NULL) {
 182                         put_page(src_page);
 183                         break;
 184                 }
 185                 dst_pte = *dst_ptep;
 186                 if (!hv_pte_get_present(dst_pte) ||
 187                     !hv_pte_get_writable(dst_pte)) {
 188                         put_page(src_page);
 189                         break;
 190                 }
 191                 dst_page = pfn_to_page(pte_pfn(dst_pte));
 192                 if (dst_page == src_page) {
 193                         /*
 194                          * Source and dest are on the same page; this
 195                          * potentially exposes us to incoherence if any
 196                          * part of src and dest overlap on a cache line.
 197                          * Just give up rather than trying to be precise.
 198                          */
 199                         put_page(src_page);
 200                         break;
 201                 }
 202                 get_page(dst_page);
 203                 if (pte_val(dst_pte) != pte_val(*dst_ptep)) {
 204                         put_page(dst_page);
 205                         goto retry_dest;
 206                 }
 207                 if (pte_huge(dst_pte)) {
 208                         /* Adjust the PTE to correspond to a small page */
 209                         int pfn = pte_pfn(dst_pte);
 210                         pfn += (((unsigned long)dest & (HPAGE_SIZE-1))
 211                                 >> PAGE_SHIFT);
 212                         dst_pte = pfn_pte(pfn, dst_pte);
 213                         dst_pte = pte_mksmall(dst_pte);
 214                 }
 215
 216                 /* All looks good: create a cachable PTE and copy from it */
 217                 copy_size = len;
 218                 bytes_left_on_page =
 219                         PAGE_SIZE - (((int)source) & (PAGE_SIZE-1));
 220                 if (copy_size > bytes_left_on_page)
 221                         copy_size = bytes_left_on_page;
 222                 bytes_left_on_page =
 223                         PAGE_SIZE - (((int)dest) & (PAGE_SIZE-1));
 224                 if (copy_size > bytes_left_on_page)
 225                         copy_size = bytes_left_on_page;
 226                 memcpy_multicache(dest, source, dst_pte, src_pte, copy_size);
 227
 228                 /* Release the pages */
 229                 put_page(dst_page);
 230                 put_page(src_page);
 231
 232                 /* Continue on the next page */
 233                 dest += copy_size;
 234                 source += copy_size;
 235                 len -= copy_size;
 236         }
 237
 238         return func(dest, source, len);
 239 }
 240
 241 void *memcpy(void *to, const void *from, __kernel_size_t n)
 242 {
 243         if (n < LARGE_COPY_CUTOFF)
 244                 return (void *)__memcpy_asm(to, from, n);
 245         else
 246                 return (void *)fast_copy(to, from, n, __memcpy_asm);
 247 }
 248
 249 unsigned long __copy_to_user_inatomic(void __user *to, const void *from,
 250                                       unsigned long n)
 251 {
 252         if (n < LARGE_COPY_CUTOFF)
 253                 return __copy_to_user_inatomic_asm(to, from, n);
 254         else
 255                 return fast_copy(to, from, n, __copy_to_user_inatomic_asm);
 256 }
 257
 258 unsigned long __copy_from_user_inatomic(void *to, const void __user *from,
 259                                         unsigned long n)
 260 {
 261         if (n < LARGE_COPY_CUTOFF)
 262                 return __copy_from_user_inatomic_asm(to, from, n);
 263         else
 264                 return fast_copy(to, from, n, __copy_from_user_inatomic_asm);
 265 }
 266
 267 unsigned long __copy_from_user_zeroing(void *to, const void __user *from,
 268                                        unsigned long n)
 269 {
 270         if (n < LARGE_COPY_CUTOFF)
 271                 return __copy_from_user_zeroing_asm(to, from, n);
 272         else
 273                 return fast_copy(to, from, n, __copy_from_user_zeroing_asm);
 274 }
 275
 276 #endif /* !CHIP_HAS_COHERENT_LOCAL_CACHE() */