lustre/kernel_patches/patches/tcp-zero-copy-2.4.19-pre1.patch

   1 Index: linux-2.4.19-pre1/include/linux/skbuff.h
   2 ===================================================================
   3 --- linux-2.4.19-pre1.orig/include/linux/skbuff.h       2001-11-22 22:46:26.000000000 +0300
   4 +++ linux-2.4.19-pre1/include/linux/skbuff.h    2004-01-14 01:15:13.000000000 +0300
   5 @@ -116,6 +116,30 @@
   6         __u16 size;
   7  };
   8
   9 +/* Support for callback when skb data has been released */
  10 +typedef struct zccd                            /* Zero Copy Callback Descriptor */
  11 +{                                              /* (embed as first member of custom struct) */
  12 +       atomic_t        zccd_count;             /* reference count */
  13 +       void           (*zccd_destructor)(struct zccd *); /* callback when refcount reaches zero */
  14 +} zccd_t;
  15 +
  16 +static inline void zccd_init (zccd_t *d, void (*callback)(zccd_t *))
  17 +{
  18 +       atomic_set (&d->zccd_count, 1);
  19 +       d->zccd_destructor = callback;
  20 +}
  21 +
  22 +static inline void zccd_get (zccd_t *d)                /* take a reference */
  23 +{
  24 +       atomic_inc (&d->zccd_count);
  25 +}
  26 +
  27 +static inline void zccd_put (zccd_t *d)                /* release a reference */
  28 +{
  29 +       if (atomic_dec_and_test (&d->zccd_count))
  30 +               (d->zccd_destructor)(d);
  31 +}
  32 +
  33  /* This data is invariant across clones and lives at
  34   * the end of the header data, ie. at skb->end.
  35   */
  36 @@ -123,6 +147,12 @@
  37         atomic_t        dataref;
  38         unsigned int    nr_frags;
  39         struct sk_buff  *frag_list;
  40 +       zccd_t          *zccd;                  /* zero copy descriptor */
  41 +       zccd_t          *zccd2;                 /* 2nd zero copy descriptor */
  42 +       /* NB we expect zero-copy data to be at least 1 packet, so
  43 +        * having 2 zccds means we don't unneccessarily split the packet
  44 +        * where consecutive zero-copy sends abutt.
  45 +        */
  46         skb_frag_t      frags[MAX_SKB_FRAGS];
  47  };
  48
  49 Index: linux-2.4.19-pre1/include/net/tcp.h
  50 ===================================================================
  51 --- linux-2.4.19-pre1.orig/include/net/tcp.h    2001-11-22 22:47:22.000000000 +0300
  52 +++ linux-2.4.19-pre1/include/net/tcp.h 2004-01-14 01:15:13.000000000 +0300
  53 @@ -640,6 +640,8 @@
  54
  55  extern int                     tcp_sendmsg(struct sock *sk, struct msghdr *msg, int size);
  56  extern ssize_t                 tcp_sendpage(struct socket *sock, struct page *page, int offset, size_t size, int flags);
  57 +extern ssize_t                 tcp_sendpage_zccd(struct socket *sock, struct page *page, int offset, size_t size,
  58 +                                                 int flags, zccd_t *zccd);
  59
  60  extern int                     tcp_ioctl(struct sock *sk,
  61                                           int cmd,
  62 @@ -733,6 +735,9 @@
  63                                             struct msghdr *msg,
  64                                             int len, int nonblock,
  65                                             int flags, int *addr_len);
  66 +extern int                     tcp_recvpackets(struct sock *sk,
  67 +                                               struct sk_buff_head *packets,
  68 +                                               int len, int nonblock);
  69
  70  extern int                     tcp_listen_start(struct sock *sk);
  71
  72 Index: linux-2.4.19-pre1/net/netsyms.c
  73 ===================================================================
  74 --- linux-2.4.19-pre1.orig/net/netsyms.c        2004-01-14 01:10:37.000000000 +0300
  75 +++ linux-2.4.19-pre1/net/netsyms.c     2004-01-14 01:15:54.000000000 +0300
  76 @@ -409,6 +409,9 @@
  77
  78  #endif
  79
  80 +EXPORT_SYMBOL(tcp_sendpage_zccd);
  81 +EXPORT_SYMBOL(tcp_recvpackets);
  82 +
  83  EXPORT_SYMBOL(netlink_set_err);
  84  EXPORT_SYMBOL(netlink_broadcast);
  85  EXPORT_SYMBOL(netlink_unicast);
  86 Index: linux-2.4.19-pre1/net/core/skbuff.c
  87 ===================================================================
  88 --- linux-2.4.19-pre1.orig/net/core/skbuff.c    2001-12-21 20:42:05.000000000 +0300
  89 +++ linux-2.4.19-pre1/net/core/skbuff.c 2004-01-14 01:15:13.000000000 +0300
  90 @@ -208,6 +208,8 @@
  91         atomic_set(&(skb_shinfo(skb)->dataref), 1);
  92         skb_shinfo(skb)->nr_frags = 0;
  93         skb_shinfo(skb)->frag_list = NULL;
  94 +       skb_shinfo(skb)->zccd = NULL;           /* skbuffs kick off with NO user zero copy descriptors */
  95 +       skb_shinfo(skb)->zccd2 = NULL;
  96         return skb;
  97
  98  nodata:
  99 @@ -276,6 +278,10 @@
 100  {
 101         if (!skb->cloned ||
 102             atomic_dec_and_test(&(skb_shinfo(skb)->dataref))) {
 103 +               if (skb_shinfo(skb)->zccd != NULL) /* zero copy callback descriptor? */
 104 +                       zccd_put (skb_shinfo(skb)->zccd); /* release hold */
 105 +               if (skb_shinfo(skb)->zccd2 != NULL) /* 2nd zero copy callback descriptor? */
 106 +                       zccd_put (skb_shinfo(skb)->zccd2); /* release hold */
 107                 if (skb_shinfo(skb)->nr_frags) {
 108                         int i;
 109                         for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
 110 @@ -532,6 +538,8 @@
 111         atomic_set(&(skb_shinfo(skb)->dataref), 1);
 112         skb_shinfo(skb)->nr_frags = 0;
 113         skb_shinfo(skb)->frag_list = NULL;
 114 +       skb_shinfo(skb)->zccd = NULL;           /* copied data => no user zero copy descriptor */
 115 +       skb_shinfo(skb)->zccd2 = NULL;
 116
 117         /* We are no longer a clone, even if we were. */
 118         skb->cloned = 0;
 119 @@ -578,6 +586,14 @@
 120         n->data_len = skb->data_len;
 121         n->len = skb->len;
 122
 123 +       if (skb_shinfo(skb)->zccd != NULL)      /* user zero copy descriptor? */
 124 +               zccd_get (skb_shinfo(skb)->zccd); /* 1 more ref (pages are shared) */
 125 +       skb_shinfo(n)->zccd = skb_shinfo(skb)->zccd;
 126 +
 127 +       if (skb_shinfo(skb)->zccd2 != NULL)     /* 2nd user zero copy descriptor? */
 128 +               zccd_get (skb_shinfo(skb)->zccd2); /* 1 more ref (pages are shared) */
 129 +       skb_shinfo(n)->zccd2 = skb_shinfo(skb)->zccd2;
 130 +
 131         if (skb_shinfo(skb)->nr_frags) {
 132                 int i;
 133
 134 @@ -620,6 +636,8 @@
 135         u8 *data;
 136         int size = nhead + (skb->end - skb->head) + ntail;
 137         long off;
 138 +       zccd_t *zccd = skb_shinfo(skb)->zccd;   /* stash user zero copy descriptor */
 139 +       zccd_t *zccd2 = skb_shinfo(skb)->zccd2; /* stash 2nd user zero copy descriptor */
 140
 141         if (skb_shared(skb))
 142                 BUG();
 143 @@ -641,6 +659,11 @@
 144         if (skb_shinfo(skb)->frag_list)
 145                 skb_clone_fraglist(skb);
 146
 147 +       if (zccd != NULL)                       /* user zero copy descriptor? */
 148 +               zccd_get (zccd);                /* extra ref (pages are shared) */
 149 +       if (zccd2 != NULL)                      /* 2nd user zero copy descriptor? */
 150 +               zccd_get (zccd2);               /* extra ref (pages are shared) */
 151 +
 152         skb_release_data(skb);
 153
 154         off = (data+nhead) - skb->head;
 155 @@ -655,6 +678,8 @@
 156         skb->nh.raw += off;
 157         skb->cloned = 0;
 158         atomic_set(&skb_shinfo(skb)->dataref, 1);
 159 +       skb_shinfo(skb)->zccd = zccd;
 160 +       skb_shinfo(skb)->zccd2 = zccd2;
 161         return 0;
 162
 163  nodata:
 164 Index: linux-2.4.19-pre1/net/ipv4/tcp.c
 165 ===================================================================
 166 --- linux-2.4.19-pre1.orig/net/ipv4/tcp.c       2001-12-21 20:42:05.000000000 +0300
 167 +++ linux-2.4.19-pre1/net/ipv4/tcp.c    2004-01-14 01:15:13.000000000 +0300
 168 @@ -744,7 +744,7 @@
 169         goto out;
 170  }
 171
 172 -ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset, size_t psize, int flags);
 173 +ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset, size_t psize, int flags, zccd_t *zccd);
 174
 175  static inline int
 176  can_coalesce(struct sk_buff *skb, int i, struct page *page, int off)
 177 @@ -823,7 +823,8 @@
 178         return err;
 179  }
 180
 181 -ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset, size_t psize, int flags)
 182 +/* Extra parameter: user zero copy descriptor (or NULL if not doing that) */
 183 +ssize_t do_tcp_sendpages(struct sock *sk, struct page **pages, int poffset, size_t psize, int flags, zccd_t *zccd)
 184  {
 185         struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
 186         int mss_now;
 187 @@ -871,6 +872,17 @@
 188                         copy = size;
 189
 190                 i = skb_shinfo(skb)->nr_frags;
 191 +
 192 +               if (zccd != NULL &&             /* this is a zcc I/O */
 193 +                   skb_shinfo(skb)->zccd != NULL && /* skb is part of a zcc I/O */
 194 +                   skb_shinfo(skb)->zccd2 != NULL &&
 195 +                   skb_shinfo(skb)->zccd != zccd && /* not the same one */
 196 +                   skb_shinfo(skb)->zccd2 != zccd)
 197 +               {
 198 +                       tcp_mark_push (tp, skb);
 199 +                       goto new_segment;
 200 +               }
 201 +
 202                 if (can_coalesce(skb, i, page, offset)) {
 203                         skb_shinfo(skb)->frags[i-1].size += copy;
 204                 } else if (i < MAX_SKB_FRAGS) {
 205 @@ -881,6 +893,20 @@
 206                         goto new_segment;
 207                 }
 208
 209 +               if (zccd != NULL &&     /* this is a zcc I/O */
 210 +                   skb_shinfo(skb)->zccd != zccd && /* not already referencing this zccd */
 211 +                   skb_shinfo(skb)->zccd2 != zccd)
 212 +               {
 213 +                       zccd_get (zccd);        /* bump ref count */
 214 +
 215 +                       BUG_TRAP (skb_shinfo(skb)->zccd2 == NULL);
 216 +
 217 +                       if (skb_shinfo(skb)->zccd == NULL) /* reference this zccd */
 218 +                               skb_shinfo(skb)->zccd = zccd;
 219 +                       else
 220 +                               skb_shinfo(skb)->zccd2 = zccd;
 221 +               }
 222 +
 223                 skb->len += copy;
 224                 skb->data_len += copy;
 225                 skb->ip_summed = CHECKSUM_HW;
 226 @@ -944,7 +970,31 @@
 227
 228         lock_sock(sk);
 229         TCP_CHECK_TIMER(sk);
 230 -       res = do_tcp_sendpages(sk, &page, offset, size, flags);
 231 +       res = do_tcp_sendpages(sk, &page, offset, size, flags, NULL);
 232 +       TCP_CHECK_TIMER(sk);
 233 +       release_sock(sk);
 234 +       return res;
 235 +}
 236 +
 237 +ssize_t tcp_sendpage_zccd(struct socket *sock, struct page *page, int offset, size_t size,
 238 +                         int flags, zccd_t *zccd)
 239 +{
 240 +       ssize_t res;
 241 +       struct sock *sk = sock->sk;
 242 +
 243 +#define TCP_ZC_CSUM_FLAGS (NETIF_F_IP_CSUM|NETIF_F_NO_CSUM|NETIF_F_HW_CSUM)
 244 +
 245 +       if (!(sk->route_caps & NETIF_F_SG) ||   /* caller shouldn't waste her time */
 246 +           !(sk->route_caps & TCP_ZC_CSUM_FLAGS)) /* on double mapping */
 247 +               BUG ();
 248 +
 249 +#undef TCP_ZC_CSUM_FLAGS
 250 +
 251 +       lock_sock(sk);
 252 +       TCP_CHECK_TIMER(sk);
 253 +
 254 +       res = do_tcp_sendpages(sk, &page, offset, size, flags, zccd);
 255 +
 256         TCP_CHECK_TIMER(sk);
 257         release_sock(sk);
 258         return res;
 259 @@ -1683,6 +1733,202 @@
 260         goto out;
 261  }
 262
 263 +int tcp_recvpackets (struct sock *sk, struct sk_buff_head *packets,
 264 +                    int len, int nonblock)
 265 +{
 266 +       struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
 267 +       int copied;
 268 +       long timeo;
 269 +
 270 +       BUG_TRAP (len > 0);
 271 +       /*BUG_TRAP ((flags & (MSG_OOB | MSG_PEEK | MSG_TRUNC)) == 0);*/
 272 +
 273 +       lock_sock(sk);
 274 +
 275 +       TCP_CHECK_TIMER(sk);
 276 +
 277 +       copied = -ENOTCONN;
 278 +       if (sk->state == TCP_LISTEN)
 279 +               goto out;
 280 +
 281 +       copied = 0;
 282 +       timeo = sock_rcvtimeo(sk, nonblock);
 283 +
 284 +       do {
 285 +               struct sk_buff * skb;
 286 +               u32 offset;
 287 +               unsigned long used;
 288 +               int exhausted;
 289 +               int eaten;
 290 +
 291 +               /* Are we at urgent data? Stop if we have read anything. */
 292 +               if (copied && tp->urg_data && tp->urg_seq == tp->copied_seq)
 293 +                       break;
 294 +
 295 +               /* We need to check signals first, to get correct SIGURG
 296 +                * handling. FIXME: Need to check this doesnt impact 1003.1g
 297 +                * and move it down to the bottom of the loop
 298 +                */
 299 +               if (signal_pending(current)) {
 300 +                       if (copied)
 301 +                               break;
 302 +                       copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
 303 +                       break;
 304 +               }
 305 +
 306 +               /* Next get a buffer. */
 307 +
 308 +               skb = skb_peek(&sk->receive_queue);
 309 +
 310 +               if (skb == NULL)                /* nothing ready */
 311 +               {
 312 +                       if (copied) {
 313 +                               if (sk->err ||
 314 +                                   sk->state == TCP_CLOSE ||
 315 +                                   (sk->shutdown & RCV_SHUTDOWN) ||
 316 +                                   !timeo ||
 317 +                                   (0))
 318 +                                       break;
 319 +                       } else {
 320 +                               if (sk->done)
 321 +                                       break;
 322 +
 323 +                               if (sk->err) {
 324 +                                       copied = sock_error(sk);
 325 +                                       break;
 326 +                               }
 327 +
 328 +                               if (sk->shutdown & RCV_SHUTDOWN)
 329 +                                       break;
 330 +
 331 +                               if (sk->state == TCP_CLOSE) {
 332 +                                       if (!sk->done) {
 333 +                                               /* This occurs when user tries to read
 334 +                                                * from never connected socket.
 335 +                                                */
 336 +                                               copied = -ENOTCONN;
 337 +                                               break;
 338 +                                       }
 339 +                                       break;
 340 +                               }
 341 +
 342 +                               if (!timeo) {
 343 +                                       copied = -EAGAIN;
 344 +                                       break;
 345 +                               }
 346 +                       }
 347 +
 348 +                       cleanup_rbuf(sk, copied);
 349 +                       timeo = tcp_data_wait(sk, timeo);
 350 +                       continue;
 351 +               }
 352 +
 353 +               BUG_TRAP (atomic_read (&skb->users) == 1);
 354 +
 355 +               exhausted = eaten = 0;
 356 +
 357 +               offset = tp->copied_seq - TCP_SKB_CB(skb)->seq;
 358 +               if (skb->h.th->syn)
 359 +                       offset--;
 360 +
 361 +               used = skb->len - offset;
 362 +
 363 +               if (tp->urg_data) {
 364 +                       u32 urg_offset = tp->urg_seq - tp->copied_seq;
 365 +                       if (urg_offset < used) {
 366 +                               if (!urg_offset) { /* at urgent date */
 367 +                                       if (!sk->urginline) {
 368 +                                               tp->copied_seq++; /* discard the single byte of urgent data */
 369 +                                               offset++;
 370 +                                               used--;
 371 +                                       }
 372 +                               } else          /* truncate read */
 373 +                                       used = urg_offset;
 374 +                       }
 375 +               }
 376 +
 377 +               BUG_TRAP (used >= 0);
 378 +               if (len < used)
 379 +                       used = len;
 380 +
 381 +               if (used == 0)
 382 +                       exhausted = 1;
 383 +               else
 384 +               {
 385 +                       if (skb_is_nonlinear (skb))
 386 +                       {
 387 +                               int   rc = skb_linearize (skb, GFP_KERNEL);
 388 +
 389 +                               printk ("tcp_recvpackets(): linearising: %d\n", rc);
 390 +
 391 +                               if (rc)
 392 +                               {
 393 +                                       if (!copied)
 394 +                                               copied = rc;
 395 +                                       break;
 396 +                               }
 397 +                       }
 398 +
 399 +                       if ((offset + used) == skb->len) /* consuming the whole packet */
 400 +                       {
 401 +                               __skb_unlink (skb, &sk->receive_queue);
 402 +                               dst_release (skb->dst);
 403 +                               skb_orphan (skb);
 404 +                               __skb_pull (skb, offset);
 405 +                               __skb_queue_tail (packets, skb);
 406 +                               exhausted = eaten = 1;
 407 +                       }
 408 +                       else                    /* consuming only part of the packet */
 409 +                       {
 410 +                               struct sk_buff *skb2 = skb_clone (skb, GFP_KERNEL);
 411 +
 412 +                               if (skb2 == NULL)
 413 +                               {
 414 +                                       if (!copied)
 415 +                                               copied = -ENOMEM;
 416 +                                       break;
 417 +                               }
 418 +
 419 +                               dst_release (skb2->dst);
 420 +                               __skb_pull (skb2, offset);
 421 +                               __skb_trim (skb2, used);
 422 +                               __skb_queue_tail (packets, skb2);
 423 +                       }
 424 +
 425 +                       tp->copied_seq += used;
 426 +                       copied += used;
 427 +                       len -= used;
 428 +               }
 429 +
 430 +               if (tp->urg_data && after(tp->copied_seq,tp->urg_seq)) {
 431 +                       tp->urg_data = 0;
 432 +                       tcp_fast_path_check(sk, tp);
 433 +               }
 434 +
 435 +               if (!exhausted)
 436 +                       continue;
 437 +
 438 +               if (skb->h.th->fin)
 439 +               {
 440 +                       tp->copied_seq++;
 441 +                       if (!eaten)
 442 +                               tcp_eat_skb (sk, skb);
 443 +                       break;
 444 +               }
 445 +
 446 +               if (!eaten)
 447 +                       tcp_eat_skb (sk, skb);
 448 +
 449 +       } while (len > 0);
 450 +
 451 + out:
 452 +       /* Clean up data we have read: This will do ACK frames. */
 453 +       cleanup_rbuf(sk, copied);
 454 +       TCP_CHECK_TIMER(sk);
 455 +       release_sock(sk);
 456 +       return copied;
 457 +}
 458 +
 459  /*
 460   *     State processing on a close. This implements the state shift for
 461   *     sending our FIN frame. Note that we only send a FIN for some