1 include/linux/skbuff.h | 30 +++++
3 net/core/skbuff.c | 25 ++++
4 net/ipv4/tcp.c | 252 ++++++++++++++++++++++++++++++++++++++++++++++++-
6 5 files changed, 311 insertions(+), 3 deletions(-)
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
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 */
23 +static inline void zccd_init (zccd_t *d, void (*callback)(zccd_t *))
25 + atomic_set (&d->zccd_count, 1);
26 + d->zccd_destructor = callback;
29 +static inline void zccd_get (zccd_t *d) /* take a reference */
31 + atomic_inc (&d->zccd_count);
34 +static inline void zccd_put (zccd_t *d) /* release a reference */
36 + if (atomic_dec_and_test (&d->zccd_count))
37 + (d->zccd_destructor)(d);
40 /* This data is invariant across clones and lives at
41 * the end of the header data, ie. at skb->end.
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.
53 skb_frag_t frags[MAX_SKB_FRAGS];
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
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);
67 extern int tcp_ioctl(struct sock *sk,
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);
77 extern int tcp_listen_start(struct sock *sk);
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
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;
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) {
102 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++)
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;
110 /* We are no longer a clone, even if we were. */
113 n->data_len = skb->data_len;
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;
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;
124 if (skb_shinfo(skb)->nr_frags) {
129 int size = nhead + (skb->end - skb->head) + ntail;
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 */
137 if (skb_shinfo(skb)->frag_list)
138 skb_clone_fraglist(skb);
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) */
145 skb_release_data(skb);
147 off = (data+nhead) - skb->head;
151 atomic_set(&skb_shinfo(skb)->dataref, 1);
152 + skb_shinfo(skb)->zccd = zccd;
153 + skb_shinfo(skb)->zccd2 = zccd2;
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
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);
169 can_coalesce(struct sk_buff *skb, int i, struct page *page, int off)
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)
178 struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
183 i = skb_shinfo(skb)->nr_frags;
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)
191 + tcp_mark_push (tp, skb);
195 if (can_coalesce(skb, i, page, offset)) {
196 skb_shinfo(skb)->frags[i-1].size += copy;
197 } else if (i < MAX_SKB_FRAGS) {
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)
206 + zccd_get (zccd); /* bump ref count */
208 + BUG_TRAP (skb_shinfo(skb)->zccd2 == NULL);
210 + if (skb_shinfo(skb)->zccd == NULL) /* reference this zccd */
211 + skb_shinfo(skb)->zccd = zccd;
213 + skb_shinfo(skb)->zccd2 = zccd;
217 skb->data_len += copy;
218 skb->ip_summed = CHECKSUM_HW;
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);
230 +ssize_t tcp_sendpage_zccd(struct socket *sock, struct page *page, int offset, size_t size,
231 + int flags, zccd_t *zccd)
234 + struct sock *sk = sock->sk;
236 +#define TCP_ZC_CSUM_FLAGS (NETIF_F_IP_CSUM|NETIF_F_NO_CSUM|NETIF_F_HW_CSUM)
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 */
242 +#undef TCP_ZC_CSUM_FLAGS
245 + TCP_CHECK_TIMER(sk);
247 + res = do_tcp_sendpages(sk, &page, offset, size, flags, zccd);
252 @@ -1771,6 +1821,202 @@
256 +int tcp_recvpackets (struct sock *sk, struct sk_buff_head *packets,
257 + int len, int nonblock)
259 + struct tcp_opt *tp = &(sk->tp_pinfo.af_tcp);
263 + BUG_TRAP (len > 0);
264 + /*BUG_TRAP ((flags & (MSG_OOB | MSG_PEEK | MSG_TRUNC)) == 0);*/
268 + TCP_CHECK_TIMER(sk);
270 + copied = -ENOTCONN;
271 + if (sk->state == TCP_LISTEN)
275 + timeo = sock_rcvtimeo(sk, nonblock);
278 + struct sk_buff * skb;
280 + unsigned long used;
284 + /* Are we at urgent data? Stop if we have read anything. */
285 + if (copied && tp->urg_data && tp->urg_seq == tp->copied_seq)
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
292 + if (signal_pending(current)) {
295 + copied = timeo ? sock_intr_errno(timeo) : -EAGAIN;
299 + /* Next get a buffer. */
301 + skb = skb_peek(&sk->receive_queue);
303 + if (skb == NULL) /* nothing ready */
307 + sk->state == TCP_CLOSE ||
308 + (sk->shutdown & RCV_SHUTDOWN) ||
317 + copied = sock_error(sk);
321 + if (sk->shutdown & RCV_SHUTDOWN)
324 + if (sk->state == TCP_CLOSE) {
326 + /* This occurs when user tries to read
327 + * from never connected socket.
329 + copied = -ENOTCONN;
341 + cleanup_rbuf(sk, copied);
342 + timeo = tcp_data_wait(sk, timeo);
346 + BUG_TRAP (atomic_read (&skb->users) == 1);
348 + exhausted = eaten = 0;
350 + offset = tp->copied_seq - TCP_SKB_CB(skb)->seq;
351 + if (skb->h.th->syn)
354 + used = skb->len - offset;
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 */
365 + } else /* truncate read */
370 + BUG_TRAP (used >= 0);
378 + if (skb_is_nonlinear (skb))
380 + int rc = skb_linearize (skb, GFP_KERNEL);
382 + printk ("tcp_recvpackets(): linearising: %d\n", rc);
392 + if ((offset + used) == skb->len) /* consuming the whole packet */
394 + __skb_unlink (skb, &sk->receive_queue);
395 + dst_release (skb->dst);
397 + __skb_pull (skb, offset);
398 + __skb_queue_tail (packets, skb);
399 + exhausted = eaten = 1;
401 + else /* consuming only part of the packet */
403 + struct sk_buff *skb2 = skb_clone (skb, GFP_KERNEL);
412 + dst_release (skb2->dst);
413 + __skb_pull (skb2, offset);
414 + __skb_trim (skb2, used);
415 + __skb_queue_tail (packets, skb2);
418 + tp->copied_seq += used;
423 + if (tp->urg_data && after(tp->copied_seq,tp->urg_seq)) {
425 + tcp_fast_path_check(sk, tp);
431 + if (skb->h.th->fin)
435 + tcp_eat_skb (sk, skb);
440 + tcp_eat_skb (sk, skb);
445 + /* Clean up data we have read: This will do ACK frames. */
446 + cleanup_rbuf(sk, copied);
447 + TCP_CHECK_TIMER(sk);
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
463 +EXPORT_SYMBOL(tcp_sendpage_zccd);
464 +EXPORT_SYMBOL(tcp_recvpackets);
465 EXPORT_SYMBOL(tcp_read_sock);
467 EXPORT_SYMBOL(netlink_set_err);