lustre/kernel_patches/patches/tcp-zero-copy-2.4.22-rh.patch

   1  include/linux/skbuff.h |   30 +++++
   2  include/net/tcp.h      |    5
   3  net/core/skbuff.c      |   25 ++++
   4  net/ipv4/tcp.c         |  252 ++++++++++++++++++++++++++++++++++++++++++++++++-
   5  net/netsyms.c          |    2
   6  5 files changed, 311 insertions(+), 3 deletions(-)
   7
   8 Index: linux-2.4.22-vanilla/include/linux/skbuff.h
   9 ===================================================================
  10 --- linux-2.4.22-vanilla.orig/include/linux/skbuff.h    2003-11-03 23:22:13.000000000 +0300
  11 +++ linux-2.4.22-vanilla/include/linux/skbuff.h 2003-12-02 23:56:35.000000000 +0300
  12 @@ -116,6 +116,30 @@
  13         __u16 size;
  14  };
  15
  16 +/* Support for callback when skb data has been released */
  17 +typedef struct zccd                            /* Zero Copy Callback Descriptor */
  18 +{                                              /* (embed as first member of custom struct) */
  19 +       atomic_t        zccd_count;             /* reference count */
  20 +       void           (*zccd_destructor)(struct zccd *); /* callback when refcount reaches zero */
  21 +} zccd_t;
  22 +
  23 +static inline void zccd_init (zccd_t *d, void (*callback)(zccd_t *))
  24 +{
  25 +       atomic_set (&d->zccd_count, 1);
  26 +       d->zccd_destructor = callback;
  27 +}
  28 +
  29 +static inline void zccd_get (zccd_t *d)                /* take a reference */
  30 +{
  31 +       atomic_inc (&d->zccd_count);
  32 +}
  33 +
  34 +static inline void zccd_put (zccd_t *d)                /* release a reference */
  35 +{
  36 +       if (atomic_dec_and_test (&d->zccd_count))
  37 +               (d->zccd_destructor)(d);
  38 +}
  39 +
  40  /* This data is invariant across clones and lives at
  41   * the end of the header data, ie. at skb->end.
  42   */
  43 @@ -123,6 +147,12 @@
  44         atomic_t        dataref;
  45         unsigned int    nr_frags;
  46         struct sk_buff  *frag_list;
  47 +       zccd_t          *zccd;                  /* zero copy descriptor */
  48 +       zccd_t          *zccd2;                 /* 2nd zero copy descriptor */
  49 +       /* NB we expect zero-copy data to be at least 1 packet, so
  50 +        * having 2 zccds means we don't unneccessarily split the packet
  51 +        * where consecutive zero-copy sends abutt.
  52 +        */
  53         skb_frag_t      frags[MAX_SKB_FRAGS];
  54  };
  55
  56 Index: linux-2.4.22-vanilla/include/net/tcp.h
  57 ===================================================================
  58 --- linux-2.4.22-vanilla.orig/include/net/tcp.h 2003-11-03 23:22:13.000000000 +0300
  59 +++ linux-2.4.22-vanilla/include/net/tcp.h      2003-12-02 23:58:10.000000000 +0300
  60 @@ -643,6 +643,8 @@
  61
  62  extern int                     tcp_sendmsg(struct sock *sk, struct msghdr *msg, int size);
  63  extern ssize_t                 tcp_sendpage(struct socket *sock, struct page *page, int offset, size_t size, int flags);
  64 +extern ssize_t                 tcp_sendpage_zccd(struct socket *sock, struct page *page, int offset, size_t size,
  65 +                                                 int flags, zccd_t *zccd);
  66
  67  extern int                     tcp_ioctl(struct sock *sk,
  68                                           int cmd,
  69 @@ -737,6 +739,9 @@
  70                                             struct msghdr *msg,
  71                                             int len, int nonblock,
  72                                             int flags, int *addr_len);
  73 +extern int                     tcp_recvpackets(struct sock *sk,
  74 +                                               struct sk_buff_head *packets,
  75 +                                               int len, int nonblock);
  76
  77  extern int                     tcp_listen_start(struct sock *sk);
  78
  79 Index: linux-2.4.22-vanilla/net/core/skbuff.c
  80 ===================================================================
  81 --- linux-2.4.22-vanilla.orig/net/core/skbuff.c 2003-11-03 23:22:13.000000000 +0300
  82 +++ linux-2.4.22-vanilla/net/core/skbuff.c      2003-12-02 23:56:15.000000000 +0300
  83 @@ -208,6 +208,8 @@
  84         atomic_set(&(skb_shinfo(skb)->dataref), 1);
  85         skb_shinfo(skb)->nr_frags = 0;
  86         skb_shinfo(skb)->frag_list = NULL;
  87 +       skb_shinfo(skb)->zccd = NULL;           /* skbuffs kick off with NO user zero copy descriptors */
  88 +       skb_shinfo(skb)->zccd2 = NULL;
  89         return skb;
  90
  91  nodata:
  92 @@ -277,6 +279,10 @@
  93  {
  94         if (!skb->cloned ||
  95             atomic_dec_and_test(&(skb_shinfo(skb)->dataref))) {
  96 +               if (skb_shinfo(skb)->zccd != NULL) /* zero copy callback descriptor? */
  97 +                       zccd_put (skb_shinfo(skb)->zccd); /* release hold */
  98 +               if (skb_shinfo(skb)->zccd2 != NULL) /* 2nd zero copy callback descriptor? */
  99 +                       zccd_put (skb_shinfo(skb)->zccd2); /* release hold */
 100                 if (skb_shinfo(skb)->nr_frags) {
 101                         int i;
 102                         for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
 103 @@ -535,6 +541,8 @@
 104         atomic_set(&(skb_shinfo(skb)->dataref), 1);
 105         skb_shinfo(skb)->nr_frags = 0;
 106         skb_shinfo(skb)->frag_list = NULL;
 107 +       skb_shinfo(skb)->zccd = NULL;           /* copied data => no user zero copy descriptor */
 108 +       skb_shinfo(skb)->zccd2 = NULL;
 109
 110         /* We are no longer a clone, even if we were. */
 111         skb->cloned = 0;
 112 @@ -581,6 +589,14 @@
 113         n->data_len = skb->data_len;
 114         n->len = skb->len;
 115
 116 +       if (skb_shinfo(skb)->zccd != NULL)      /* user zero copy descriptor? */
 117 +               zccd_get (skb_shinfo(skb)->zccd); /* 1 more ref (pages are shared) */
 118 +       skb_shinfo(n)->zccd = skb_shinfo(skb)->zccd;
 119 +
 120 +       if (skb_shinfo(skb)->zccd2 != NULL)     /* 2nd user zero copy descriptor? */
 121 +               zccd_get (skb_shinfo(skb)->zccd2); /* 1 more ref (pages are shared) */
 122 +       skb_shinfo(n)->zccd2 = skb_shinfo(skb)->zccd2;
 123 +
 124         if (skb_shinfo(skb)->nr_frags) {
 125                 int i;
 126
 127 @@ -623,6 +639,8 @@
 128         u8 *data;
 129         int size = nhead + (skb->end - skb->head) + ntail;
 130         long off;
 131 +       zccd_t *zccd = skb_shinfo(skb)->zccd;   /* stash user zero copy descriptor */
 132 +       zccd_t *zccd2 = skb_shinfo(skb)->zccd2; /* stash 2nd user zero copy descriptor */
 133
 134         if (skb_shared(skb))
 135                 BUG();
 136 @@ -644,6 +662,11 @@
 137         if (skb_shinfo(skb)->frag_list)
 138                 skb_clone_fraglist(skb);
 139
 140 +       if (zccd != NULL)                       /* user zero copy descriptor? */
 141 +               zccd_get (zccd);                /* extra ref (pages are shared) */
 142 +       if (zccd2 != NULL)                      /* 2nd user zero copy descriptor? */
 143 +               zccd_get (zccd2);               /* extra ref (pages are shared) */
 144 +
 145         skb_release_data(skb);
 146
 147         off = (data+nhead) - skb->head;
 148 @@ -658,6 +681,8 @@
 149         skb->nh.raw += off;
 150         skb->cloned = 0;
 151         atomic_set(&skb_shinfo(skb)->dataref, 1);
 152 +       skb_shinfo(skb)->zccd = zccd;
 153 +       skb_shinfo(skb)->zccd2 = zccd2;
 154         return 0;
 155
 156  nodata:
 157 Index: linux-2.4.22-vanilla/net/ipv4/tcp.c
 158 ===================================================================
 159 --- linux-2.4.22-vanilla.orig/net/ipv4/tcp.c    2003-11-03 23:22:13.000000000 +0300
 160 +++ linux-2.4.22-vanilla/net/ipv4/tcp.c 2003-12-02 23:56:15.000000000 +0300
 161 @@ -747,7 +747,7 @@
 162         goto out;
 163  }
 164
 165 -ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset, size_t psize, int flags);
 166 +ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset, size_t psize, int flags, zccd_t *zccd);
 167
 168  static inline int
 169  can_coalesce(struct sk_buff *skb, int i, struct page *page, int off)
 170 @@ -826,7 +826,8 @@
 171         return err;
 172  }
 173
 174 -ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset, size_t psize, int flags)
 175 +/* Extra parameter: user zero copy descriptor (or NULL if not doing that) */
 176 +ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset, size_t psize, int flags, zccd_t *zccd)
 177  {
 178         struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
 179         int mss_now;
 180 @@ -874,6 +875,17 @@
 181                         copy = size;
 182
 183                 i = skb_shinfo(skb)->nr_frags;
 184 +
 185 +               if (zccd != NULL &&             /* this is a zcc I/O */
 186 +                   skb_shinfo(skb)->zccd != NULL && /* skb is part of a zcc I/O */
 187 +                   skb_shinfo(skb)->zccd2 != NULL &&
 188 +                   skb_shinfo(skb)->zccd != zccd && /* not the same one */
 189 +                   skb_shinfo(skb)->zccd2 != zccd)
 190 +               {
 191 +                       tcp_mark_push (tp, skb);
 192 +                       goto new_segment;
 193 +               }
 194 +
 195                 if (can_coalesce(skb, i, page, offset)) {
 196                         skb_shinfo(skb)->frags[i-1].size += copy;
 197                 } else if (i < MAX_SKB_FRAGS) {
 198 @@ -884,6 +896,20 @@
 199                         goto new_segment;
 200                 }
 201
 202 +               if (zccd != NULL &&     /* this is a zcc I/O */
 203 +                   skb_shinfo(skb)->zccd != zccd && /* not already referencing this zccd */
 204 +                   skb_shinfo(skb)->zccd2 != zccd)
 205 +               {
 206 +                       zccd_get (zccd);        /* bump ref count */
 207 +
 208 +                       BUG_TRAP (skb_shinfo(skb)->zccd2 == NULL);
 209 +
 210 +                       if (skb_shinfo(skb)->zccd == NULL) /* reference this zccd */
 211 +                               skb_shinfo(skb)->zccd = zccd;
 212 +                       else
 213 +                               skb_shinfo(skb)->zccd2 = zccd;
 214 +               }
 215 +
 216                 skb->len += copy;
 217                 skb->data_len += copy;
 218                 skb->ip_summed = CHECKSUM_HW;
 219 @@ -947,7 +973,31 @@
 220
 221         lock_sock(sk);
 222         TCP_CHECK_TIMER(sk);
 223 -       res = do_tcp_sendpages(sk, &page, offset, size, flags);
 224 +       res = do_tcp_sendpages(sk, &page, offset, size, flags, NULL);
 225 +       TCP_CHECK_TIMER(sk);
 226 +       release_sock(sk);
 227 +       return res;
 228 +}
 229 +
 230 +ssize_t tcp_sendpage_zccd(struct socket *sock, struct page *page, int offset, size_t size,
 231 +                         int flags, zccd_t *zccd)
 232 +{
 233 +       ssize_t res;
 234 +       struct sock *sk = sock->sk;
 235 +
 236 +#define TCP_ZC_CSUM_FLAGS (NETIF_F_IP_CSUM|NETIF_F_NO_CSUM|NETIF_F_HW_CSUM)
 237 +
 238 +       if (!(sk->route_caps & NETIF_F_SG) ||   /* caller shouldn't waste her time */
 239 +           !(sk->route_caps & TCP_ZC_CSUM_FLAGS)) /* on double mapping */
 240 +               BUG ();
 241 +
 242 +#undef TCP_ZC_CSUM_FLAGS
 243 +
 244 +       lock_sock(sk);
 245 +       TCP_CHECK_TIMER(sk);
 246 +
 247 +       res = do_tcp_sendpages(sk, &page, offset, size, flags, zccd);
 248 +
 249         TCP_CHECK_TIMER(sk);
 250         release_sock(sk);
 251         return res;
 252 @@ -1771,6 +1821,202 @@
 253         goto out;
 254  }
 255
 256 +int tcp_recvpackets (struct sock *sk, struct sk_buff_head *packets,
 257 +                    int len, int nonblock)
 258 +{
 259 +       struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
 260 +       int copied;
 261 +       long timeo;
 262 +
 263 +       BUG_TRAP (len > 0);
 264 +       /*BUG_TRAP ((flags & (MSG_OOB | MSG_PEEK | MSG_TRUNC)) == 0);*/
 265 +
 266 +       lock_sock(sk);
 267 +
 268 +       TCP_CHECK_TIMER(sk);
 269 +
 270 +       copied = -ENOTCONN;
 271 +       if (sk->state == TCP_LISTEN)
 272 +               goto out;
 273 +
 274 +       copied = 0;
 275 +       timeo = sock_rcvtimeo(sk, nonblock);
 276 +
 277 +       do {
 278 +               struct sk_buff * skb;
 279 +               u32 offset;
 280 +               unsigned long used;
 281 +               int exhausted;
 282 +               int eaten;
 283 +
 284 +               /* Are we at urgent data? Stop if we have read anything. */
 285 +               if (copied && tp->urg_data && tp->urg_seq == tp->copied_seq)
 286 +                       break;
 287 +
 288 +               /* We need to check signals first, to get correct SIGURG
 289 +                * handling. FIXME: Need to check this doesnt impact 1003.1g
 290 +                * and move it down to the bottom of the loop
 291 +                */
 292 +               if (signal_pending(current)) {
 293 +                       if (copied)
 294 +                               break;
 295 +                       copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
 296 +                       break;
 297 +               }
 298 +
 299 +               /* Next get a buffer. */
 300 +
 301 +               skb = skb_peek(&sk->receive_queue);
 302 +
 303 +               if (skb == NULL)                /* nothing ready */
 304 +               {
 305 +                       if (copied) {
 306 +                               if (sk->err ||
 307 +                                   sk->state == TCP_CLOSE ||
 308 +                                   (sk->shutdown & RCV_SHUTDOWN) ||
 309 +                                   !timeo ||
 310 +                                   (0))
 311 +                                       break;
 312 +                       } else {
 313 +                               if (sk->done)
 314 +                                       break;
 315 +
 316 +                               if (sk->err) {
 317 +                                       copied = sock_error(sk);
 318 +                                       break;
 319 +                               }
 320 +
 321 +                               if (sk->shutdown & RCV_SHUTDOWN)
 322 +                                       break;
 323 +
 324 +                               if (sk->state == TCP_CLOSE) {
 325 +                                       if (!sk->done) {
 326 +                                               /* This occurs when user tries to read
 327 +                                                * from never connected socket.
 328 +                                                */
 329 +                                               copied = -ENOTCONN;
 330 +                                               break;
 331 +                                       }
 332 +                                       break;
 333 +                               }
 334 +
 335 +                               if (!timeo) {
 336 +                                       copied = -EAGAIN;
 337 +                                       break;
 338 +                               }
 339 +                       }
 340 +
 341 +                       cleanup_rbuf(sk, copied);
 342 +                       timeo = tcp_data_wait(sk, timeo);
 343 +                       continue;
 344 +               }
 345 +
 346 +               BUG_TRAP (atomic_read (&skb->users) == 1);
 347 +
 348 +               exhausted = eaten = 0;
 349 +
 350 +               offset = tp->copied_seq - TCP_SKB_CB(skb)->seq;
 351 +               if (skb->h.th->syn)
 352 +                       offset--;
 353 +
 354 +               used = skb->len - offset;
 355 +
 356 +               if (tp->urg_data) {
 357 +                       u32 urg_offset = tp->urg_seq - tp->copied_seq;
 358 +                       if (urg_offset < used) {
 359 +                               if (!urg_offset) { /* at urgent date */
 360 +                                       if (!sk->urginline) {
 361 +                                               tp->copied_seq++; /* discard the single byte of urgent data */
 362 +                                               offset++;
 363 +                                               used--;
 364 +                                       }
 365 +                               } else          /* truncate read */
 366 +                                       used = urg_offset;
 367 +                       }
 368 +               }
 369 +
 370 +               BUG_TRAP (used >= 0);
 371 +               if (len < used)
 372 +                       used = len;
 373 +
 374 +               if (used == 0)
 375 +                       exhausted = 1;
 376 +               else
 377 +               {
 378 +                       if (skb_is_nonlinear (skb))
 379 +                       {
 380 +                               int   rc = skb_linearize (skb, GFP_KERNEL);
 381 +
 382 +                               printk ("tcp_recvpackets(): linearising: %d\n", rc);
 383 +
 384 +                               if (rc)
 385 +                               {
 386 +                                       if (!copied)
 387 +                                               copied = rc;
 388 +                                       break;
 389 +                               }
 390 +                       }
 391 +
 392 +                       if ((offset + used) == skb->len) /* consuming the whole packet */
 393 +                       {
 394 +                               __skb_unlink (skb, &sk->receive_queue);
 395 +                               dst_release (skb->dst);
 396 +                               skb_orphan (skb);
 397 +                               __skb_pull (skb, offset);
 398 +                               __skb_queue_tail (packets, skb);
 399 +                               exhausted = eaten = 1;
 400 +                       }
 401 +                       else                    /* consuming only part of the packet */
 402 +                       {
 403 +                               struct sk_buff *skb2 = skb_clone (skb, GFP_KERNEL);
 404 +
 405 +                               if (skb2 == NULL)
 406 +                               {
 407 +                                       if (!copied)
 408 +                                               copied = -ENOMEM;
 409 +                                       break;
 410 +                               }
 411 +
 412 +                               dst_release (skb2->dst);
 413 +                               __skb_pull (skb2, offset);
 414 +                               __skb_trim (skb2, used);
 415 +                               __skb_queue_tail (packets, skb2);
 416 +                       }
 417 +
 418 +                       tp->copied_seq += used;
 419 +                       copied += used;
 420 +                       len -= used;
 421 +               }
 422 +
 423 +               if (tp->urg_data && after(tp->copied_seq,tp->urg_seq)) {
 424 +                       tp->urg_data = 0;
 425 +                       tcp_fast_path_check(sk, tp);
 426 +               }
 427 +
 428 +               if (!exhausted)
 429 +                       continue;
 430 +
 431 +               if (skb->h.th->fin)
 432 +               {
 433 +                       tp->copied_seq++;
 434 +                       if (!eaten)
 435 +                               tcp_eat_skb (sk, skb);
 436 +                       break;
 437 +               }
 438 +
 439 +               if (!eaten)
 440 +                       tcp_eat_skb (sk, skb);
 441 +
 442 +       } while (len > 0);
 443 +
 444 + out:
 445 +       /* Clean up data we have read: This will do ACK frames. */
 446 +       cleanup_rbuf(sk, copied);
 447 +       TCP_CHECK_TIMER(sk);
 448 +       release_sock(sk);
 449 +       return copied;
 450 +}
 451 +
 452  /*
 453   *     State processing on a close. This implements the state shift for
 454   *     sending our FIN frame. Note that we only send a FIN for some
 455 Index: linux-2.4.22-vanilla/net/netsyms.c
 456 ===================================================================
 457 --- linux-2.4.22-vanilla.orig/net/netsyms.c     2003-11-03 23:22:13.000000000 +0300
 458 +++ linux-2.4.22-vanilla/net/netsyms.c  2003-12-04 20:42:50.000000000 +0300
 459 @@ -417,6 +417,8 @@
 460
 461  #endif
 462
 463 +EXPORT_SYMBOL(tcp_sendpage_zccd);
 464 +EXPORT_SYMBOL(tcp_recvpackets);
 465  EXPORT_SYMBOL(tcp_read_sock);
 466
 467  EXPORT_SYMBOL(netlink_set_err);