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         addr = kmap(page);
  60         if (to_user) {
  61                 rds_stats_add(s_copy_to_user, bytes);
  62                 ret = copy_to_user(ptr, addr + offset, bytes);
  63         } else {
  64                 rds_stats_add(s_copy_from_user, bytes);
  65                 ret = copy_from_user(addr + offset, ptr, bytes);
  66         }
  67         kunmap(page);
  68
  69         return ret ? -EFAULT : 0;
  70 }
  71 EXPORT_SYMBOL_GPL(rds_page_copy_user);
  72
  73 /*
  74  * Message allocation uses this to build up regions of a message.
  75  *
  76  * @bytes - the number of bytes needed.
  77  * @gfp - the waiting behaviour of the allocation
  78  *
  79  * @gfp is always ored with __GFP_HIGHMEM.  Callers must be prepared to
  80  * kmap the pages, etc.
  81  *
  82  * If @bytes is at least a full page then this just returns a page from
  83  * alloc_page().
  84  *
  85  * If @bytes is a partial page then this stores the unused region of the
  86  * page in a per-cpu structure.  Future partial-page allocations may be
  87  * satisfied from that cached region.  This lets us waste less memory on
  88  * small allocations with minimal complexity.  It works because the transmit
  89  * path passes read-only page regions down to devices.  They hold a page
  90  * reference until they are done with the region.
  91  */
  92 int rds_page_remainder_alloc(struct scatterlist *scat, unsigned long bytes,
  93                              gfp_t gfp)
  94 {
  95         struct rds_page_remainder *rem;
  96         unsigned long flags;
  97         struct page *page;
  98         int ret;
  99
 100         gfp |= __GFP_HIGHMEM;
 101
 102         /* jump straight to allocation if we're trying for a huge page */
 103         if (bytes >= PAGE_SIZE) {
 104                 page = alloc_page(gfp);
 105                 if (page == NULL) {
 106                         ret = -ENOMEM;
 107                 } else {
 108                         sg_set_page(scat, page, PAGE_SIZE, 0);
 109                         ret = 0;
 110                 }
 111                 goto out;
 112         }
 113
 114         rem = &per_cpu(rds_page_remainders, get_cpu());
 115         local_irq_save(flags);
 116
 117         while (1) {
 118                 /* avoid a tiny region getting stuck by tossing it */
 119                 if (rem->r_page && bytes > (PAGE_SIZE - rem->r_offset)) {
 120                         rds_stats_inc(s_page_remainder_miss);
 121                         __free_page(rem->r_page);
 122                         rem->r_page = NULL;
 123                 }
 124
 125                 /* hand out a fragment from the cached page */
 126                 if (rem->r_page && bytes <= (PAGE_SIZE - rem->r_offset)) {
 127                         sg_set_page(scat, rem->r_page, bytes, rem->r_offset);
 128                         get_page(sg_page(scat));
 129
 130                         if (rem->r_offset != 0)
 131                                 rds_stats_inc(s_page_remainder_hit);
 132
 133                         rem->r_offset += bytes;
 134                         if (rem->r_offset == PAGE_SIZE) {
 135                                 __free_page(rem->r_page);
 136                                 rem->r_page = NULL;
 137                         }
 138                         ret = 0;
 139                         break;
 140                 }
 141
 142                 /* alloc if there is nothing for us to use */
 143                 local_irq_restore(flags);
 144                 put_cpu();
 145
 146                 page = alloc_page(gfp);
 147
 148                 rem = &per_cpu(rds_page_remainders, get_cpu());
 149                 local_irq_save(flags);
 150
 151                 if (page == NULL) {
 152                         ret = -ENOMEM;
 153                         break;
 154                 }
 155
 156                 /* did someone race to fill the remainder before us? */
 157                 if (rem->r_page) {
 158                         __free_page(page);
 159                         continue;
 160                 }
 161
 162                 /* otherwise install our page and loop around to alloc */
 163                 rem->r_page = page;
 164                 rem->r_offset = 0;
 165         }
 166
 167         local_irq_restore(flags);
 168         put_cpu();
 169 out:
 170         rdsdebug("bytes %lu ret %d %p %u %u\n", bytes, ret,
 171                  ret ? NULL : sg_page(scat), ret ? 0 : scat->offset,
 172                  ret ? 0 : scat->length);
 173         return ret;
 174 }
 175
 176 static int rds_page_remainder_cpu_notify(struct notifier_block *self,
 177                                          unsigned long action, void *hcpu)
 178 {
 179         struct rds_page_remainder *rem;
 180         long cpu = (long)hcpu;
 181
 182         rem = &per_cpu(rds_page_remainders, cpu);
 183
 184         rdsdebug("cpu %ld action 0x%lx\n", cpu, action);
 185
 186         switch (action) {
 187         case CPU_DEAD:
 188                 if (rem->r_page)
 189                         __free_page(rem->r_page);
 190                 rem->r_page = NULL;
 191                 break;
 192         }
 193
 194         return 0;
 195 }
 196
 197 static struct notifier_block rds_page_remainder_nb = {
 198         .notifier_call = rds_page_remainder_cpu_notify,
 199 };
 200
 201 void rds_page_exit(void)
 202 {
 203         int i;
 204
 205         for_each_possible_cpu(i)
 206                 rds_page_remainder_cpu_notify(&rds_page_remainder_nb,
 207                                               (unsigned long)CPU_DEAD,
 208                                               (void *)(long)i);
 209 }