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 iov_used;
  35
  36 static int pcap_sg_pull_file_header(int fd, uint32_t *linktype)
  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         pcap_validate_header(&hdr);
  46
  47         *linktype = hdr.linktype;
  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_rt();
  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_rt();
 109
 110         spinlock_lock(&lock);
 111         if (readv(fd, iov, IOVSIZ) <= 0)
 112                 return -EIO;
 113
 114         iov_used = 0;
 115         c = 0;
 116         spinlock_unlock(&lock);
 117
 118         return 0;
 119 }
 120
 121 static ssize_t pcap_sg_read_pcap_pkt(int fd, struct pcap_pkthdr *hdr,
 122                                      uint8_t *packet, size_t len)
 123 {
 124         ssize_t ret = 0;
 125
 126         /* In contrast to writing, reading gets really ugly ... */
 127         spinlock_lock(&lock);
 128
 129         if (likely(iov[c].iov_len - iov_used >= sizeof(*hdr))) {
 130                 fmemcpy(hdr, iov[c].iov_base + iov_used, sizeof(*hdr));
 131                 iov_used += sizeof(*hdr);
 132         } else {
 133                 size_t offset = 0;
 134                 ssize_t remainder;
 135
 136                 offset = iov[c].iov_len - iov_used;
 137                 remainder = sizeof(*hdr) - offset;
 138                 if (remainder < 0)
 139                         remainder = 0;
 140
 141                 bug_on(offset + remainder != sizeof(*hdr));
 142
 143                 fmemcpy(hdr, iov[c].iov_base + iov_used, offset);
 144
 145                 iov_used = 0;
 146                 c++;
 147
 148                 if (c == IOVSIZ) {
 149                         /* We need to refetch! */
 150                         c = 0;
 151                         if (readv(fd, iov, IOVSIZ) <= 0) {
 152                                 ret = -EIO;
 153                                 goto out_err;
 154                         }
 155                 }
 156
 157                 /* Now we copy the remainder and go on with business ... */
 158                 fmemcpy((uint8_t *) hdr + offset,
 159                         iov[c].iov_base + iov_used, remainder);
 160                 iov_used += remainder;
 161         }
 162
 163         /* header read completed */
 164
 165         if (unlikely(hdr->caplen == 0 || hdr->caplen > len)) {
 166                 ret = -EINVAL; /* Bogus packet */
 167                 goto out_err;
 168         }
 169
 170         /* now we read data ... */
 171
 172         if (likely(iov[c].iov_len - iov_used >= hdr->caplen)) {
 173                 fmemcpy(packet, iov[c].iov_base + iov_used, hdr->caplen);
 174                 iov_used += hdr->caplen;
 175         } else {
 176                 size_t offset = 0;
 177                 ssize_t remainder;
 178
 179                 offset = iov[c].iov_len - iov_used;
 180                 remainder = hdr->caplen - offset;
 181                 if (remainder < 0)
 182                         remainder = 0;
 183
 184                 bug_on(offset + remainder != hdr->caplen);
 185
 186                 fmemcpy(packet, iov[c].iov_base + iov_used, offset);
 187
 188                 iov_used = 0;
 189                 c++;
 190
 191                 if (c == IOVSIZ) {
 192                         /* We need to refetch! */
 193                         c = 0;
 194                         if (readv(fd, iov, IOVSIZ)) {
 195                                 ret = -EIO;
 196                                 goto out_err;
 197                         }
 198                 }
 199
 200                 /* Now we copy the remainder and go on with business ... */
 201                 fmemcpy(packet + offset, iov[c].iov_base + iov_used, remainder);
 202                 iov_used += remainder;
 203         }
 204
 205         spinlock_unlock(&lock);
 206
 207         return sizeof(*hdr) + hdr->caplen;
 208
 209 out_err:
 210         spinlock_unlock(&lock);
 211         return ret;
 212 }
 213
 214 static void pcap_sg_fsync_pcap(int fd)
 215 {
 216         ssize_t ret;
 217
 218         spinlock_lock(&lock);
 219         ret = writev(fd, iov, c);
 220         if (ret < 0)
 221                 panic("writev I/O error!\n");
 222
 223         c = 0;
 224
 225         fdatasync(fd);
 226         spinlock_unlock(&lock);
 227 }
 228
 229 struct pcap_file_ops pcap_sg_ops __read_mostly = {
 230         .name = "scatter-gather",
 231         .pull_file_header = pcap_sg_pull_file_header,
 232         .push_file_header = pcap_sg_push_file_header,
 233         .write_pcap_pkt = pcap_sg_write_pcap_pkt,
 234         .prepare_reading_pcap =  pcap_sg_prepare_reading_pcap,
 235         .prepare_writing_pcap =  pcap_sg_prepare_writing_pcap,
 236         .read_pcap_pkt = pcap_sg_read_pcap_pkt,
 237         .fsync_pcap = pcap_sg_fsync_pcap,
 238 };
 239
 240 int init_pcap_sg(int jumbo_support)
 241 {
 242         unsigned long i;
 243         size_t allocsz = 0;
 244
 245         c = 0;
 246
 247         fmemset(iov, 0, sizeof(iov));
 248
 249         if (jumbo_support)
 250                 allocsz = ALLSIZ_JUMBO;
 251         else
 252                 allocsz = ALLSIZ_2K;
 253
 254         for (i = 0; i < IOVSIZ; ++i) {
 255                 iov[i].iov_base = xzmalloc_aligned(allocsz, 64);
 256                 iov[i].iov_len = allocsz;
 257         }
 258
 259         spinlock_init(&lock);
 260
 261         return pcap_ops_group_register(&pcap_sg_ops, PCAP_OPS_SG);
 262 }
 263
 264 void cleanup_pcap_sg(void)
 265 {
 266         unsigned long i;
 267
 268         spinlock_destroy(&lock);
 269
 270         for (i = 0; i < IOVSIZ; ++i)
 271                 xfree(iov[i].iov_base);
 272
 273         pcap_ops_group_unregister(PCAP_OPS_SG);
 274 }