net/rds/page.c

   1 /*
   2  * Copyright (c) 2006 Oracle.  All rights reserved.
   3  *
   4  * This software is available to you under a choice of one of two
   5  * licenses.  You may choose to be licensed under the terms of the GNU
   6  * General Public License (GPL) Version 2, available from the file
   7  * COPYING in the main directory of this source tree, or the
   8  * OpenIB.org BSD license below:
   9  *
  10  *     Redistribution and use in source and binary forms, with or
  11  *     without modification, are permitted provided that the following
  12  *     conditions are met:
  13  *
  14  *      - Redistributions of source code must retain the above
  15  *        copyright notice, this list of conditions and the following
  16  *        disclaimer.
  17  *
  18  *      - Redistributions in binary form must reproduce the above
  19  *        copyright notice, this list of conditions and the following
  20  *        disclaimer in the documentation and/or other materials
  21  *        provided with the distribution.
  22  *
  23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  30  * SOFTWARE.
  31  *
  32  */
  33 #include <linux/highmem.h>
  34
  35 #include "rds.h"
  36
  37 struct rds_page_remainder {
  38         struct page     *r_page;
  39         unsigned long   r_offset;
  40 };
  41
  42 DEFINE_PER_CPU_SHARED_ALIGNED(struct rds_page_remainder, rds_page_remainders);
  43
  44 /*
  45  * returns 0 on success or -errno on failure.
  46  *
  47  * We don't have to worry about flush_dcache_page() as this only works
  48  * with private pages.  If, say, we were to do directed receive to pinned
  49  * user pages we'd have to worry more about cache coherence.  (Though
  50  * the flush_dcache_page() in get_user_pages() would probably be enough).
  51  */
  52 int rds_page_copy_user(struct page *page, unsigned long offset,
  53                        void __user *ptr, unsigned long bytes,
  54                        int to_user)
  55 {
  56         unsigned long ret;
  57         void *addr;
  58
  59         if (to_user)
  60                 rds_stats_add(s_copy_to_user, bytes);
  61         else
  62                 rds_stats_add(s_copy_from_user, bytes);
  63
  64         addr = kmap_atomic(page, KM_USER0);
  65         if (to_user)
  66                 ret = __copy_to_user_inatomic(ptr, addr + offset, bytes);
  67         else
  68                 ret = __copy_from_user_inatomic(addr + offset, ptr, bytes);
  69         kunmap_atomic(addr, KM_USER0);
  70
  71         if (ret) {
  72                 addr = kmap(page);
  73                 if (to_user)
  74                         ret = copy_to_user(ptr, addr + offset, bytes);
  75                 else
  76                         ret = copy_from_user(addr + offset, ptr, bytes);
  77                 kunmap(page);
  78                 if (ret)
  79                         return -EFAULT;
  80         }
  81
  82         return 0;
  83 }
  84 EXPORT_SYMBOL_GPL(rds_page_copy_user);
  85
  86 /*
  87  * Message allocation uses this to build up regions of a message.
  88  *
  89  * @bytes - the number of bytes needed.
  90  * @gfp - the waiting behaviour of the allocation
  91  *
  92  * @gfp is always ored with __GFP_HIGHMEM.  Callers must be prepared to
  93  * kmap the pages, etc.
  94  *
  95  * If @bytes is at least a full page then this just returns a page from
  96  * alloc_page().
  97  *
  98  * If @bytes is a partial page then this stores the unused region of the
  99  * page in a per-cpu structure.  Future partial-page allocations may be
 100  * satisfied from that cached region.  This lets us waste less memory on
 101  * small allocations with minimal complexity.  It works because the transmit
 102  * path passes read-only page regions down to devices.  They hold a page
 103  * reference until they are done with the region.
 104  */
 105 int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
 106                              gfp_t gfp)
 107 {
 108         struct rds_page_remainder *rem;
 109         unsigned long flags;
 110         struct page *page;
 111         int ret;
 112
 113         gfp |= __GFP_HIGHMEM;
 114
 115         /* jump straight to allocation if we're trying for a huge page */
 116         if (bytes >= PAGE_SIZE) {
 117                 page = alloc_page(gfp);
 118                 if (page == NULL) {
 119                         ret = -ENOMEM;
 120                 } else {
 121                         sg_set_page(scat, page, PAGE_SIZE, 0);
 122                         ret = 0;
 123                 }
 124                 goto out;
 125         }
 126
 127         rem = &per_cpu(rds_page_remainders, get_cpu());
 128         local_irq_save(flags);
 129
 130         while (1) {
 131                 /* avoid a tiny region getting stuck by tossing it */
 132                 if (rem->r_page && bytes > (PAGE_SIZE - rem->r_offset)) {
 133                         rds_stats_inc(s_page_remainder_miss);
 134                         __free_page(rem->r_page);
 135                         rem->r_page = NULL;
 136                 }
 137
 138                 /* hand out a fragment from the cached page */
 139                 if (rem->r_page && bytes <= (PAGE_SIZE - rem->r_offset)) {
 140                         sg_set_page(scat, rem->r_page, bytes, rem->r_offset);
 141                         get_page(sg_page(scat));
 142
 143                         if (rem->r_offset != 0)
 144                                 rds_stats_inc(s_page_remainder_hit);
 145
 146                         rem->r_offset += bytes;
 147                         if (rem->r_offset == PAGE_SIZE) {
 148                                 __free_page(rem->r_page);
 149                                 rem->r_page = NULL;
 150                         }
 151                         ret = 0;
 152                         break;
 153                 }
 154
 155                 /* alloc if there is nothing for us to use */
 156                 local_irq_restore(flags);
 157                 put_cpu();
 158
 159                 page = alloc_page(gfp);
 160
 161                 rem = &per_cpu(rds_page_remainders, get_cpu());
 162                 local_irq_save(flags);
 163
 164                 if (page == NULL) {
 165                         ret = -ENOMEM;
 166                         break;
 167                 }
 168
 169                 /* did someone race to fill the remainder before us? */
 170                 if (rem->r_page) {
 171                         __free_page(page);
 172                         continue;
 173                 }
 174
 175                 /* otherwise install our page and loop around to alloc */
 176                 rem->r_page = page;
 177                 rem->r_offset = 0;
 178         }
 179
 180         local_irq_restore(flags);
 181         put_cpu();
 182 out:
 183         rdsdebug("bytes %lu ret %d %p %u %u\n", bytes, ret,
 184                  ret ? NULL : sg_page(scat), ret ? 0 : scat->offset,
 185                  ret ? 0 : scat->length);
 186         return ret;
 187 }
 188
 189 static int rds_page_remainder_cpu_notify(struct notifier_block *self,
 190                                          unsigned long action, void *hcpu)
 191 {
 192         struct rds_page_remainder *rem;
 193         long cpu = (long)hcpu;
 194
 195         rem = &per_cpu(rds_page_remainders, cpu);
 196
 197         rdsdebug("cpu %ld action 0x%lx\n", cpu, action);
 198
 199         switch (action) {
 200         case CPU_DEAD:
 201                 if (rem->r_page)
 202                         __free_page(rem->r_page);
 203                 rem->r_page = NULL;
 204                 break;
 205         }
 206
 207         return 0;
 208 }
 209
 210 static struct notifier_block rds_page_remainder_nb = {
 211         .notifier_call = rds_page_remainder_cpu_notify,
 212 };
 213
 214 void rds_page_exit(void)
 215 {
 216         int i;
 217
 218         for_each_possible_cpu(i)
 219                 rds_page_remainder_cpu_notify(&rds_page_remainder_nb,
 220                                               (unsigned long)CPU_DEAD,
 221                                               (void *)(long)i);
 222 }