src/pcap_sg.c

   1 /*
   2  * netsniff-ng - the packet sniffing beast
   3  * By Daniel Borkmann <daniel@netsniff-ng.org>
   4  * Copyright 2011 Daniel Borkmann.
   5  * Subject to the GPL, version 2.
   6  */
   7
   8 #include <stdio.h>
   9 #include <sys/types.h>
  10 #include <sys/stat.h>
  11 #include <fcntl.h>
  12 #include <sys/uio.h>
  13 #include <unistd.h>
  14
  15 #include "pcap.h"
  16 #include "xmalloc.h"
  17 #include "xio.h"
  18 #include "xsys.h"
  19 #include "locking.h"
  20 #include "built_in.h"
  21
  22 #define PAGE_SIZE         (getpagesize())
  23 #define PAGE_MASK         (~(PAGE_SIZE - 1))
  24 #define PAGE_ALIGN(addr)  (((addr) + PAGE_SIZE - 1) & PAGE_MASK)
  25
  26 #define IOVSIZ   1000
  27 #define ALLSIZ          (PAGE_SIZE * 3)
  28 #define ALLSIZ_2K       (PAGE_SIZE * 3)  // 12K max
  29 #define ALLSIZ_JUMBO    (PAGE_SIZE * 16) // 64K max
  30
  31 static struct iovec iov[IOVSIZ];
  32 static unsigned long c = 0;
  33 static struct spinlock lock;
  34 static ssize_t avail, used, iov_used;
  35
  36 static int pcap_sg_pull_file_header(int fd)
  37 {
  38         ssize_t ret;
  39         struct pcap_filehdr hdr;
  40
  41         ret = read(fd, &hdr, sizeof(hdr));
  42         if (unlikely(ret != sizeof(hdr)))
  43                 return -EIO;
  44
  45         ret = pcap_validate_header(&hdr);
  46         if (ret < 0)
  47                 lseek(fd, -sizeof(hdr), SEEK_CUR);
  48
  49         return 0;
  50 }
  51
  52 static int pcap_sg_push_file_header(int fd)
  53 {
  54         ssize_t ret;
  55         struct pcap_filehdr hdr;
  56
  57         fmemset(&hdr, 0, sizeof(hdr));
  58         pcap_prepare_header(&hdr, LINKTYPE_EN10MB, 0,
  59                             PCAP_DEFAULT_SNAPSHOT_LEN);
  60         ret = write_or_die(fd, &hdr, sizeof(hdr));
  61         if (unlikely(ret != sizeof(hdr))) {
  62                 whine("Failed to write pkt file header!\n");
  63                 return -EIO;
  64         }
  65
  66         return 0;
  67 }
  68
  69 static int pcap_sg_prepare_writing_pcap(int fd)
  70 {
  71         set_ioprio_be();
  72         return 0;
  73 }
  74
  75 static ssize_t pcap_sg_write_pcap_pkt(int fd, struct pcap_pkthdr *hdr,
  76                                       uint8_t *packet, size_t len)
  77 {
  78         ssize_t ret;
  79
  80         spinlock_lock(&lock);
  81
  82         if (unlikely(c == IOVSIZ)) {
  83                 ret = writev(fd, iov, IOVSIZ);
  84                 if (ret < 0)
  85                         panic("writev I/O error!\n");
  86
  87                 c = 0;
  88         }
  89
  90         iov[c].iov_len = 0;
  91         fmemcpy(iov[c].iov_base, hdr, sizeof(*hdr));
  92
  93         iov[c].iov_len += sizeof(*hdr);
  94         fmemcpy(iov[c].iov_base + iov[c].iov_len, packet, len);
  95
  96         iov[c].iov_len += len;
  97         ret = iov[c].iov_len;
  98
  99         c++;
 100
 101         spinlock_unlock(&lock);
 102
 103         return ret;
 104 }
 105
 106 static int pcap_sg_prepare_reading_pcap(int fd)
 107 {
 108         set_ioprio_be();
 109
 110         spinlock_lock(&lock);
 111
 112         avail = readv(fd, iov, IOVSIZ);
 113         if (avail <= 0)
 114                 return -EIO;
 115
 116         used = iov_used = 0;
 117         c = 0;
 118
 119         spinlock_unlock(&lock);
 120
 121         return 0;
 122 }
 123
 124 static ssize_t pcap_sg_read_pcap_pkt(int fd, struct pcap_pkthdr *hdr,
 125                                      uint8_t *packet, size_t len)
 126 {
 127         /* In contrast to writing, reading gets really ugly ... */
 128         spinlock_lock(&lock);
 129
 130         if (likely(avail - used >= sizeof(*hdr) &&
 131                    iov[c].iov_len - iov_used >= sizeof(*hdr))) {
 132
 133                 fmemcpy(hdr, iov[c].iov_base + iov_used, sizeof(*hdr));
 134                 iov_used += sizeof(*hdr);
 135
 136                 used += sizeof(*hdr);
 137
 138         } else {
 139                 size_t remainder, offset = 0;
 140
 141                 if (avail - used < sizeof(*hdr))
 142                         return -ENOMEM;
 143
 144                 offset = iov[c].iov_len - iov_used;
 145                 remainder = sizeof(*hdr) - offset;
 146
 147                 bug_on(offset + remainder != sizeof(*hdr));
 148
 149                 fmemcpy(hdr, iov[c].iov_base + iov_used, offset);
 150                 used += offset;
 151
 152                 iov_used = 0;
 153                 c++;
 154
 155                 if (c == IOVSIZ) {
 156                         /* We need to refetch! */
 157                         c = 0;
 158
 159                         avail = readv(fd, iov, IOVSIZ);
 160                         if (avail < 0)
 161                                 return -EIO;
 162
 163                         used = 0;
 164                 }
 165
 166                 /* Now we copy the remainder and go on with business ... */
 167                 fmemcpy(hdr, iov[c].iov_base + iov_used, remainder);
 168                 iov_used += remainder;
 169
 170                 used += remainder;
 171         }
 172         if (likely(avail - used >= hdr->len &&
 173                    iov[c].iov_len - iov_used >= hdr->len)) {
 174
 175                 fmemcpy(packet, iov[c].iov_base + iov_used, hdr->len);
 176                 iov_used += hdr->len;
 177
 178                 used += hdr->len;
 179
 180         } else {
 181                 size_t remainder, offset = 0;
 182
 183                 if (avail - used < hdr->len)
 184                         return -ENOMEM;
 185
 186                 offset = iov[c].iov_len - iov_used;
 187                 remainder = hdr->len - offset;
 188
 189                 bug_on(offset + remainder != hdr->len);
 190
 191                 fmemcpy(packet, iov[c].iov_base + iov_used, offset);
 192                 used += offset;
 193
 194                 iov_used = 0;
 195                 c++;
 196
 197                 if (c == IOVSIZ) {
 198                         /* We need to refetch! */
 199                         c = 0;
 200
 201                         avail = readv(fd, iov, IOVSIZ);
 202                         if (avail < 0)
 203                                 return -EIO;
 204
 205                         used = 0;
 206                 }
 207
 208                 /* Now we copy the remainder and go on with business ... */
 209                 fmemcpy(packet, iov[c].iov_base + iov_used, remainder);
 210                 iov_used += remainder;
 211
 212                 used += remainder;
 213         }
 214
 215         spinlock_unlock(&lock);
 216
 217         if (unlikely(hdr->len == 0))
 218                 return -EINVAL; /* Bogus packet */
 219
 220         return sizeof(*hdr) + hdr->len;
 221 }
 222
 223 static void pcap_sg_fsync_pcap(int fd)
 224 {
 225         ssize_t ret;
 226
 227         spinlock_lock(&lock);
 228
 229         ret = writev(fd, iov, c);
 230         if (ret < 0)
 231                 panic("writev I/O error!\n");
 232
 233         c = 0;
 234
 235         fdatasync(fd);
 236
 237         spinlock_unlock(&lock);
 238 }
 239
 240 struct pcap_file_ops pcap_sg_ops __read_mostly = {
 241         .name = "scatter-gather",
 242         .pull_file_header = pcap_sg_pull_file_header,
 243         .push_file_header = pcap_sg_push_file_header,
 244         .write_pcap_pkt = pcap_sg_write_pcap_pkt,
 245         .prepare_reading_pcap =  pcap_sg_prepare_reading_pcap,
 246         .prepare_writing_pcap =  pcap_sg_prepare_writing_pcap,
 247         .read_pcap_pkt = pcap_sg_read_pcap_pkt,
 248         .fsync_pcap = pcap_sg_fsync_pcap,
 249 };
 250
 251 int init_pcap_sg(int jumbo_support)
 252 {
 253         unsigned long i;
 254         size_t allocsz = 0;
 255
 256         c = 0;
 257
 258         fmemset(iov, 0, sizeof(iov));
 259
 260         if (jumbo_support)
 261                 allocsz = ALLSIZ_JUMBO;
 262         else
 263                 allocsz = ALLSIZ_2K;
 264
 265         for (i = 0; i < IOVSIZ; ++i) {
 266                 iov[i].iov_base = xmalloc_aligned(allocsz, 64);
 267                 iov[i].iov_len = allocsz;
 268         }
 269
 270         spinlock_init(&lock);
 271
 272         return pcap_ops_group_register(&pcap_sg_ops, PCAP_OPS_SG);
 273 }
 274
 275 void cleanup_pcap_sg(void)
 276 {
 277         unsigned long i;
 278
 279         spinlock_destroy(&lock);
 280
 281         for (i = 0; i < IOVSIZ; ++i)
 282                 xfree(iov[i].iov_base);
 283
 284         pcap_ops_group_unregister(PCAP_OPS_SG);
 285 }