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 "xutils.h"
  19 #include "locking.h"
  20 #include "built_in.h"
  21
  22 #define IOVSIZ          1000
  23 #define ALLSIZ          (PAGE_SIZE * 3)
  24 #define ALLSIZ_2K       (PAGE_SIZE * 3)  // 12K max
  25 #define ALLSIZ_JUMBO    (PAGE_SIZE * 16) // 64K max
  26
  27 static struct iovec iov[IOVSIZ];
  28 static unsigned long c = 0;
  29 static struct spinlock lock;
  30 static ssize_t iov_used;
  31
  32 static int pcap_sg_pull_file_header(int fd, uint32_t *linktype)
  33 {
  34         ssize_t ret;
  35         struct pcap_filehdr hdr;
  36
  37         ret = read(fd, &hdr, sizeof(hdr));
  38         if (unlikely(ret != sizeof(hdr)))
  39                 return -EIO;
  40
  41         pcap_validate_header(&hdr);
  42
  43         *linktype = hdr.linktype;
  44
  45         return 0;
  46 }
  47
  48 static int pcap_sg_push_file_header(int fd, uint32_t linktype)
  49 {
  50         ssize_t ret;
  51         struct pcap_filehdr hdr;
  52
  53         fmemset(&hdr, 0, sizeof(hdr));
  54         pcap_prepare_header(&hdr, linktype, 0, PCAP_DEFAULT_SNAPSHOT_LEN);
  55
  56         ret = write_or_die(fd, &hdr, sizeof(hdr));
  57         if (unlikely(ret != sizeof(hdr))) {
  58                 whine("Failed to write pkt file header!\n");
  59                 return -EIO;
  60         }
  61
  62         return 0;
  63 }
  64
  65 static ssize_t pcap_sg_write_pcap_pkt(int fd, struct pcap_pkthdr *hdr,
  66                                       uint8_t *packet, size_t len)
  67 {
  68         ssize_t ret;
  69
  70         spinlock_lock(&lock);
  71
  72         if (unlikely(c == IOVSIZ)) {
  73                 ret = writev(fd, iov, IOVSIZ);
  74                 if (ret < 0)
  75                         panic("writev I/O error!\n");
  76
  77                 c = 0;
  78         }
  79
  80         iov[c].iov_len = 0;
  81         fmemcpy(iov[c].iov_base, hdr, sizeof(*hdr));
  82
  83         iov[c].iov_len += sizeof(*hdr);
  84         fmemcpy(iov[c].iov_base + iov[c].iov_len, packet, len);
  85
  86         iov[c].iov_len += len;
  87         ret = iov[c].iov_len;
  88
  89         c++;
  90
  91         spinlock_unlock(&lock);
  92
  93         return ret;
  94 }
  95
  96 static int pcap_sg_prepare_reading_pcap(int fd)
  97 {
  98         spinlock_lock(&lock);
  99         if (readv(fd, iov, IOVSIZ) <= 0)
 100                 return -EIO;
 101
 102         iov_used = 0;
 103         c = 0;
 104         spinlock_unlock(&lock);
 105
 106         return 0;
 107 }
 108
 109 static ssize_t pcap_sg_read_pcap_pkt(int fd, struct pcap_pkthdr *hdr,
 110                                      uint8_t *packet, size_t len)
 111 {
 112         ssize_t ret = 0;
 113
 114         /* In contrast to writing, reading gets really ugly ... */
 115         spinlock_lock(&lock);
 116
 117         if (likely(iov[c].iov_len - iov_used >= sizeof(*hdr))) {
 118                 fmemcpy(hdr, iov[c].iov_base + iov_used, sizeof(*hdr));
 119                 iov_used += sizeof(*hdr);
 120         } else {
 121                 size_t offset = 0;
 122                 ssize_t remainder;
 123
 124                 offset = iov[c].iov_len - iov_used;
 125                 remainder = sizeof(*hdr) - offset;
 126                 if (remainder < 0)
 127                         remainder = 0;
 128
 129                 bug_on(offset + remainder != sizeof(*hdr));
 130
 131                 fmemcpy(hdr, iov[c].iov_base + iov_used, offset);
 132
 133                 iov_used = 0;
 134                 c++;
 135
 136                 if (c == IOVSIZ) {
 137                         /* We need to refetch! */
 138                         c = 0;
 139                         if (readv(fd, iov, IOVSIZ) <= 0) {
 140                                 ret = -EIO;
 141                                 goto out_err;
 142                         }
 143                 }
 144
 145                 /* Now we copy the remainder and go on with business ... */
 146                 fmemcpy((uint8_t *) hdr + offset,
 147                         iov[c].iov_base + iov_used, remainder);
 148                 iov_used += remainder;
 149         }
 150
 151         /* header read completed */
 152
 153         if (unlikely(hdr->caplen == 0 || hdr->caplen > len)) {
 154                 ret = -EINVAL; /* Bogus packet */
 155                 goto out_err;
 156         }
 157
 158         /* now we read data ... */
 159
 160         if (likely(iov[c].iov_len - iov_used >= hdr->caplen)) {
 161                 fmemcpy(packet, iov[c].iov_base + iov_used, hdr->caplen);
 162                 iov_used += hdr->caplen;
 163         } else {
 164                 size_t offset = 0;
 165                 ssize_t remainder;
 166
 167                 offset = iov[c].iov_len - iov_used;
 168                 remainder = hdr->caplen - offset;
 169                 if (remainder < 0)
 170                         remainder = 0;
 171
 172                 bug_on(offset + remainder != hdr->caplen);
 173
 174                 fmemcpy(packet, iov[c].iov_base + iov_used, offset);
 175
 176                 iov_used = 0;
 177                 c++;
 178
 179                 if (c == IOVSIZ) {
 180                         /* We need to refetch! */
 181                         c = 0;
 182                         if (readv(fd, iov, IOVSIZ) <= 0) {
 183                                 ret = -EIO;
 184                                 goto out_err;
 185                         }
 186                 }
 187
 188                 /* Now we copy the remainder and go on with business ... */
 189                 fmemcpy(packet + offset, iov[c].iov_base + iov_used, remainder);
 190                 iov_used += remainder;
 191         }
 192
 193         spinlock_unlock(&lock);
 194
 195         return sizeof(*hdr) + hdr->caplen;
 196
 197 out_err:
 198         spinlock_unlock(&lock);
 199         return ret;
 200 }
 201
 202 static void pcap_sg_fsync_pcap(int fd)
 203 {
 204         ssize_t ret;
 205
 206         spinlock_lock(&lock);
 207         ret = writev(fd, iov, c);
 208         if (ret < 0)
 209                 panic("writev I/O error!\n");
 210
 211         c = 0;
 212
 213         fdatasync(fd);
 214         spinlock_unlock(&lock);
 215 }
 216
 217 const struct pcap_file_ops pcap_sg_ops = {
 218         .name = "scatter-gather",
 219         .pull_file_header = pcap_sg_pull_file_header,
 220         .push_file_header = pcap_sg_push_file_header,
 221         .write_pcap_pkt = pcap_sg_write_pcap_pkt,
 222         .prepare_reading_pcap =  pcap_sg_prepare_reading_pcap,
 223         .read_pcap_pkt = pcap_sg_read_pcap_pkt,
 224         .fsync_pcap = pcap_sg_fsync_pcap,
 225 };
 226
 227 int init_pcap_sg(int jumbo_support)
 228 {
 229         unsigned long i;
 230         size_t allocsz = 0;
 231
 232         c = 0;
 233
 234         fmemset(iov, 0, sizeof(iov));
 235
 236         if (jumbo_support)
 237                 allocsz = ALLSIZ_JUMBO;
 238         else
 239                 allocsz = ALLSIZ_2K;
 240
 241         for (i = 0; i < IOVSIZ; ++i) {
 242                 iov[i].iov_base = xzmalloc_aligned(allocsz, 64);
 243                 iov[i].iov_len = allocsz;
 244         }
 245
 246         spinlock_init(&lock);
 247
 248         set_ioprio_rt();
 249
 250         return pcap_ops_group_register(&pcap_sg_ops, PCAP_OPS_SG);
 251 }
 252
 253 void cleanup_pcap_sg(void)
 254 {
 255         unsigned long i;
 256
 257         spinlock_destroy(&lock);
 258
 259         for (i = 0; i < IOVSIZ; ++i)
 260                 xfree(iov[i].iov_base);
 261
 262         pcap_ops_group_unregister(PCAP_OPS_SG);
 263 }