10 #include <sys/malloc.h>
11 #include <sys/kernel.h>
12 #include <sys/sysctl.h>
15 #include <sys/rwlock.h>
16 #include <sys/socket.h>
17 #include <sys/socketvar.h>
18 #include <sys/protosw.h>
21 #include <sys/domain.h>
28 #include <net/route.h>
31 #include <netinet/in.h>
32 #include <netinet/in_systm.h>
33 #include <netinet/in_var.h>
34 #include <netinet/in_pcb.h>
35 #include <netinet/tcp.h>
36 #include <netinet/tcp_fsm.h>
37 #include <netinet/tcp_timer.h>
38 #include <netinet/tcp_var.h>
40 #include <linux/device.h>
41 #include <linux/err.h>
42 #include <linux/sched.h>
43 #include <linux/workqueue.h>
44 #include <linux/wait.h>
45 #include <linux/module.h>
46 #include <linux/moduleparam.h>
47 #include <linux/pci.h>
49 #include <rdma/ib_verbs.h>
50 #include <rdma/rdma_cm.h>
51 #include <rdma/ib_cm.h>
52 #include <rdma/sdp_socket.h>
53 #include <rdma/ib_fmr_pool.h>
56 #define CONFIG_INFINIBAND_SDP_DEBUG
62 /* From sys/queue.h */
63 #define LIST_HEAD(name, type) \
65 struct type *lh_first; /* first element */ \
68 /* Interval between sucessive polls in the Tx routine when polling is used
69 instead of interrupts (in per-core Tx rings) - should be power of 2 */
70 #define SDP_TX_POLL_MODER 16
71 #define SDP_TX_POLL_TIMEOUT (HZ / 20)
72 #define SDP_NAGLE_TIMEOUT (HZ / 10)
74 #define SDP_SRCAVAIL_CANCEL_TIMEOUT (HZ * 5)
75 #define SDP_SRCAVAIL_ADV_TIMEOUT (1 * HZ)
76 #define SDP_SRCAVAIL_PAYLOAD_LEN 1
78 #define SDP_RESOLVE_TIMEOUT 1000
79 #define SDP_ROUTE_TIMEOUT 1000
80 #define SDP_RETRY_COUNT 5
81 #define SDP_KEEPALIVE_TIME (120 * 60 * HZ)
82 #define SDP_FIN_WAIT_TIMEOUT (60 * HZ) /* like TCP_FIN_TIMEOUT */
84 #define SDP_TX_SIZE 0x40
85 #define SDP_RX_SIZE 0x40
87 #define SDP_FMR_SIZE (MIN(0x1000, PAGE_SIZE) / sizeof(u64))
88 #define SDP_FMR_POOL_SIZE 1024
89 #define SDP_FMR_DIRTY_SIZE ( SDP_FMR_POOL_SIZE / 4 )
91 #define SDP_MAX_RDMA_READ_LEN (PAGE_SIZE * (SDP_FMR_SIZE - 2))
93 /* mb inlined data len - rest will be rx'ed into frags */
94 #define SDP_HEAD_SIZE (sizeof(struct sdp_bsdh))
96 /* limit tx payload len, if the sink supports bigger buffers than the source
98 * or rx fragment size (limited by sge->length size) */
99 #define SDP_MAX_PACKET (1 << 16)
100 #define SDP_MAX_PAYLOAD (SDP_MAX_PACKET - SDP_HEAD_SIZE)
102 #define SDP_MAX_RECV_SGES (SDP_MAX_PACKET / MCLBYTES)
103 #define SDP_MAX_SEND_SGES (SDP_MAX_PACKET / MCLBYTES) + 2
107 #define SDP_DEF_ZCOPY_THRESH 64*1024
108 #define SDP_MIN_ZCOPY_THRESH PAGE_SIZE
109 #define SDP_MAX_ZCOPY_THRESH 1048576
111 #define SDP_OP_RECV 0x800000000LL
112 #define SDP_OP_SEND 0x400000000LL
113 #define SDP_OP_RDMA 0x200000000LL
114 #define SDP_OP_NOP 0x100000000LL
116 /* how long (in jiffies) to block sender till tx completion*/
117 #define SDP_BZCOPY_POLL_TIMEOUT (HZ / 10)
119 #define SDP_AUTO_CONF 0xffff
120 #define AUTO_MOD_DELAY (HZ / 4)
123 __u32 seq; /* Starting sequence number */
124 struct bzcopy_state *bz;
125 struct rx_srcavail_state *rx_sa;
126 struct tx_srcavail_state *tx_sa;
129 #define M_PUSH M_PROTO1 /* Do a 'push'. */
130 #define M_URG M_PROTO2 /* Mark as urgent (oob). */
132 #define SDP_SKB_CB(__mb) ((struct sdp_mb_cb *)&((__mb)->cb[0]))
133 #define BZCOPY_STATE(mb) (SDP_SKB_CB(mb)->bz)
134 #define RX_SRCAVAIL_STATE(mb) (SDP_SKB_CB(mb)->rx_sa)
135 #define TX_SRCAVAIL_STATE(mb) (SDP_SKB_CB(mb)->tx_sa)
138 #define MIN(a, b) (a < b ? a : b)
141 #define ring_head(ring) (atomic_read(&(ring).head))
142 #define ring_tail(ring) (atomic_read(&(ring).tail))
143 #define ring_posted(ring) (ring_head(ring) - ring_tail(ring))
145 #define rx_ring_posted(ssk) ring_posted(ssk->rx_ring)
147 #define tx_ring_posted(ssk) (ring_posted(ssk->tx_ring) + \
148 (ssk->tx_ring.rdma_inflight ? ssk->tx_ring.rdma_inflight->busy : 0))
150 #define tx_ring_posted(ssk) ring_posted(ssk->tx_ring)
153 extern int sdp_zcopy_thresh;
154 extern int rcvbuf_initial_size;
155 extern struct workqueue_struct *rx_comp_wq;
156 extern struct ib_client sdp_client;
160 SDP_MID_HELLO_ACK = 0x1,
161 SDP_MID_DISCONN = 0x2,
163 SDP_MID_SENDSM = 0x4,
164 SDP_MID_RDMARDCOMPL = 0x6,
165 SDP_MID_SRCAVAIL_CANCEL = 0x8,
166 SDP_MID_CHRCVBUF = 0xB,
167 SDP_MID_CHRCVBUF_ACK = 0xC,
168 SDP_MID_SINKAVAIL = 0xFD,
169 SDP_MID_SRCAVAIL = 0xFE,
174 SDP_OOB_PRES = 1 << 0,
175 SDP_OOB_PEND = 1 << 1,
179 SDP_MIN_TX_CREDITS = 2
196 } __attribute__((__packed__));
204 } __attribute__((__packed__));
206 /* TODO: too much? Can I avoid having the src/dst and port here? */
208 struct sdp_bsdh bsdh;
217 union cma_ip_addr src_addr;
218 union cma_ip_addr dst_addr;
219 u8 rsvd3[IB_CM_REQ_PRIVATE_DATA_SIZE - sizeof(struct sdp_bsdh) - 48];
220 } __attribute__((__packed__));
223 struct sdp_bsdh bsdh;
229 u8 rsvd2[IB_CM_REP_PRIVATE_DATA_SIZE - sizeof(struct sdp_bsdh) - 8];
230 } __attribute__((__packed__));
234 } __attribute__((__packed__));
240 } __attribute__((__packed__));
244 u64 mapping[SDP_MAX_SEND_SGES];
245 } __attribute__((__packed__));
247 struct sdp_chrecvbuf {
249 } __attribute__((__packed__));
251 /* Context used for synchronous zero copy bcopy (BZCOPY) */
252 struct bzcopy_state {
253 unsigned char __user *u_base;
260 struct sdp_sock *ssk;
270 TX_SA_CROSS_SEND = 0x02,
271 TX_SA_INTRRUPTED = 0x04,
272 TX_SA_TIMEDOUT = 0x08,
276 struct rx_srcavail_state {
277 /* Advertised buffer stuff */
286 struct ib_umem *umem;
287 struct ib_pool_fmr *fmr;
291 enum rx_sa_flag flags;
294 struct tx_srcavail_state {
295 /* Data below 'busy' will be reset */
298 struct ib_umem *umem;
299 struct ib_pool_fmr *fmr;
304 enum tx_sa_flag abort_flags;
312 struct rx_srcavail_state *rdma_inflight;
314 struct sdp_buf *buffer;
320 #define tx_credits(ssk) (atomic_read(&ssk->tx_ring.credits))
322 struct callout timer;
327 struct sdp_buf *buffer;
333 struct rwlock destroyed_lock;
339 struct ib_fmr_pool *fmr_pool;
342 struct sdp_moderation {
343 unsigned long last_moder_packets;
344 unsigned long last_moder_tx_packets;
345 unsigned long last_moder_bytes;
346 unsigned long last_moder_jiffies;
356 u16 adaptive_rx_coal;
363 /* These are flags fields. */
364 #define SDP_TIMEWAIT 0x0001 /* In ssk timewait state. */
365 #define SDP_DROPPED 0x0002 /* Socket has been dropped. */
366 #define SDP_SOCKREF 0x0004 /* Holding a sockref for close. */
367 #define SDP_NODELAY 0x0008 /* Disble nagle. */
368 #define SDP_NEEDFIN 0x0010 /* Send a fin on the next tx. */
369 #define SDP_DREQWAIT 0x0020 /* Waiting on DREQ. */
370 #define SDP_DESTROY 0x0040 /* Being destroyed. */
371 #define SDP_DISCON 0x0080 /* rdma_disconnect is owed. */
373 /* These are oobflags */
374 #define SDP_HADOOB 0x0001 /* Had OOB data. */
375 #define SDP_HAVEOOB 0x0002 /* Have OOB data. */
378 LIST_ENTRY(sdp_sock) list;
379 struct socket *socket;
380 struct rdma_cm_id *id;
381 struct ib_device *ib_device;
382 struct sdp_device *sdp_dev;
385 struct callout keep2msl; /* 2msl and keepalive timer. */
386 struct callout nagle_timer; /* timeout waiting for ack */
387 struct ib_ucontext context;
393 int oobflags; /* protected by rx lock. */
396 int recv_bytes; /* Bytes per recv. buf including header */
400 struct sdp_rx_ring rx_ring;
401 struct sdp_tx_ring tx_ring;
403 struct mbuf *rx_ctl_q;
404 struct mbuf *rx_ctl_tail;
406 int qp_active; /* XXX Flag. */
408 struct work_struct rx_comp_work;
409 #define rcv_nxt(ssk) atomic_read(&(ssk->rcv_nxt))
414 #define mseq_ack(ssk) (atomic_read(&ssk->mseq_ack))
415 unsigned max_bufs; /* Initial buffers offered by other side */
416 unsigned min_bufs; /* Low water mark to wake senders */
418 unsigned long nagle_last_unacked; /* mseq of lastest unacked packet */
420 atomic_t remote_credits;
421 #define remote_credits(ssk) (atomic_read(&ssk->remote_credits))
425 int recv_request_head; /* mark the rx_head when the resize request
427 int recv_request; /* XXX flag if request to resize was recieved */
429 unsigned long tx_packets;
430 unsigned long rx_packets;
431 unsigned long tx_bytes;
432 unsigned long rx_bytes;
433 struct sdp_moderation auto_mod;
434 struct task shutdown_task;
436 struct tx_srcavail_state *tx_sa;
437 struct rx_srcavail_state *rx_sa;
438 spinlock_t tx_sa_lock;
439 struct delayed_work srcavail_cancel_work;
440 int srcavail_cancel_mseq;
441 /* ZCOPY data: -1:use global; 0:disable zcopy; >0: zcopy threshold */
446 #define sdp_sk(so) ((struct sdp_sock *)(so->so_pcb))
448 #define SDP_RLOCK(ssk) rw_rlock(&(ssk)->lock)
449 #define SDP_WLOCK(ssk) rw_wlock(&(ssk)->lock)
450 #define SDP_RUNLOCK(ssk) rw_runlock(&(ssk)->lock)
451 #define SDP_WUNLOCK(ssk) rw_wunlock(&(ssk)->lock)
452 #define SDP_WLOCK_ASSERT(ssk) rw_assert(&(ssk)->lock, RA_WLOCKED)
453 #define SDP_RLOCK_ASSERT(ssk) rw_assert(&(ssk)->lock, RA_RLOCKED)
454 #define SDP_LOCK_ASSERT(ssk) rw_assert(&(ssk)->lock, RA_LOCKED)
456 static inline void tx_sa_reset(struct tx_srcavail_state *tx_sa)
458 memset((void *)&tx_sa->busy, 0,
459 sizeof(*tx_sa) - offsetof(typeof(*tx_sa), busy));
462 static inline void rx_ring_unlock(struct sdp_rx_ring *rx_ring)
464 rw_runlock(&rx_ring->destroyed_lock);
467 static inline int rx_ring_trylock(struct sdp_rx_ring *rx_ring)
469 rw_rlock(&rx_ring->destroyed_lock);
470 if (rx_ring->destroyed) {
471 rx_ring_unlock(rx_ring);
477 static inline void rx_ring_destroy_lock(struct sdp_rx_ring *rx_ring)
479 rw_wlock(&rx_ring->destroyed_lock);
480 rx_ring->destroyed = 1;
481 rw_wunlock(&rx_ring->destroyed_lock);
484 static inline void sdp_arm_rx_cq(struct sdp_sock *ssk)
486 sdp_prf(ssk->socket, NULL, "Arming RX cq");
487 sdp_dbg_data(ssk->socket, "Arming RX cq\n");
489 ib_req_notify_cq(ssk->rx_ring.cq, IB_CQ_NEXT_COMP);
492 static inline void sdp_arm_tx_cq(struct sdp_sock *ssk)
494 sdp_prf(ssk->socket, NULL, "Arming TX cq");
495 sdp_dbg_data(ssk->socket, "Arming TX cq. credits: %d, posted: %d\n",
496 tx_credits(ssk), tx_ring_posted(ssk));
498 ib_req_notify_cq(ssk->tx_ring.cq, IB_CQ_NEXT_COMP);
501 /* return the min of:
503 * - free slots in tx_ring (not including SDP_MIN_TX_CREDITS
505 static inline int tx_slots_free(struct sdp_sock *ssk)
509 min_free = MIN(tx_credits(ssk),
510 SDP_TX_SIZE - tx_ring_posted(ssk));
511 if (min_free < SDP_MIN_TX_CREDITS)
514 return min_free - SDP_MIN_TX_CREDITS;
518 static inline char *mid2str(int mid)
520 #define ENUM2STR(e) [e] = #e
521 static char *mid2str[] = {
522 ENUM2STR(SDP_MID_HELLO),
523 ENUM2STR(SDP_MID_HELLO_ACK),
524 ENUM2STR(SDP_MID_ABORT),
525 ENUM2STR(SDP_MID_DISCONN),
526 ENUM2STR(SDP_MID_SENDSM),
527 ENUM2STR(SDP_MID_RDMARDCOMPL),
528 ENUM2STR(SDP_MID_SRCAVAIL_CANCEL),
529 ENUM2STR(SDP_MID_CHRCVBUF),
530 ENUM2STR(SDP_MID_CHRCVBUF_ACK),
531 ENUM2STR(SDP_MID_DATA),
532 ENUM2STR(SDP_MID_SRCAVAIL),
533 ENUM2STR(SDP_MID_SINKAVAIL),
536 if (mid >= ARRAY_SIZE(mid2str))
542 static inline struct mbuf *
543 sdp_alloc_mb(struct socket *sk, u8 mid, int size, int wait)
548 MGETHDR(mb, wait, MT_DATA);
551 mb->m_pkthdr.len = mb->m_len = sizeof(struct sdp_bsdh);
552 h = mtod(mb, struct sdp_bsdh *);
557 static inline struct mbuf *
558 sdp_alloc_mb_data(struct socket *sk, int wait)
560 return sdp_alloc_mb(sk, SDP_MID_DATA, 0, wait);
563 static inline struct mbuf *
564 sdp_alloc_mb_disconnect(struct socket *sk, int wait)
566 return sdp_alloc_mb(sk, SDP_MID_DISCONN, 0, wait);
570 mb_put(struct mbuf *mb, int len)
580 static inline struct mbuf *
581 sdp_alloc_mb_chrcvbuf_ack(struct socket *sk, int size, int wait)
584 struct sdp_chrecvbuf *resp_size;
586 mb = sdp_alloc_mb(sk, SDP_MID_CHRCVBUF_ACK, sizeof(*resp_size), wait);
589 resp_size = (struct sdp_chrecvbuf *)mb_put(mb, sizeof *resp_size);
590 resp_size->size = htonl(size);
595 static inline struct mbuf *
596 sdp_alloc_mb_srcavail(struct socket *sk, u32 len, u32 rkey, u64 vaddr, int wait)
599 struct sdp_srcah *srcah;
601 mb = sdp_alloc_mb(sk, SDP_MID_SRCAVAIL, sizeof(*srcah), wait);
604 srcah = (struct sdp_srcah *)mb_put(mb, sizeof(*srcah));
605 srcah->len = htonl(len);
606 srcah->rkey = htonl(rkey);
607 srcah->vaddr = cpu_to_be64(vaddr);
612 static inline struct mbuf *
613 sdp_alloc_mb_srcavail_cancel(struct socket *sk, int wait)
615 return sdp_alloc_mb(sk, SDP_MID_SRCAVAIL_CANCEL, 0, wait);
618 static inline struct mbuf *
619 sdp_alloc_mb_rdmardcompl(struct socket *sk, u32 len, int wait)
622 struct sdp_rrch *rrch;
624 mb = sdp_alloc_mb(sk, SDP_MID_RDMARDCOMPL, sizeof(*rrch), wait);
627 rrch = (struct sdp_rrch *)mb_put(mb, sizeof(*rrch));
628 rrch->len = htonl(len);
633 static inline struct mbuf *
634 sdp_alloc_mb_sendsm(struct socket *sk, int wait)
636 return sdp_alloc_mb(sk, SDP_MID_SENDSM, 0, wait);
638 static inline int sdp_tx_ring_slots_left(struct sdp_sock *ssk)
640 return SDP_TX_SIZE - tx_ring_posted(ssk);
643 static inline int credit_update_needed(struct sdp_sock *ssk)
647 c = remote_credits(ssk);
648 if (likely(c > SDP_MIN_TX_CREDITS))
650 return unlikely(c < rx_ring_posted(ssk)) &&
651 likely(tx_credits(ssk) > 0) &&
652 likely(sdp_tx_ring_slots_left(ssk));
656 #define SDPSTATS_COUNTER_INC(stat)
657 #define SDPSTATS_COUNTER_ADD(stat, val)
658 #define SDPSTATS_COUNTER_MID_INC(stat, mid)
659 #define SDPSTATS_HIST_LINEAR(stat, size)
660 #define SDPSTATS_HIST(stat, size)
663 sdp_cleanup_sdp_buf(struct sdp_sock *ssk, struct sdp_buf *sbuf,
664 enum dma_data_direction dir)
666 struct ib_device *dev;
670 dev = ssk->ib_device;
671 for (i = 0, mb = sbuf->mb; mb != NULL; mb = mb->m_next, i++)
672 ib_dma_unmap_single(dev, sbuf->mapping[i], mb->m_len, dir);
676 void sdp_set_default_moderation(struct sdp_sock *ssk);
677 void sdp_start_keepalive_timer(struct socket *sk);
678 void sdp_urg(struct sdp_sock *ssk, struct mbuf *mb);
679 void sdp_cancel_dreq_wait_timeout(struct sdp_sock *ssk);
680 void sdp_abort(struct socket *sk);
681 struct sdp_sock *sdp_notify(struct sdp_sock *ssk, int error);
685 int sdp_cma_handler(struct rdma_cm_id *, struct rdma_cm_event *);
688 int sdp_tx_ring_create(struct sdp_sock *ssk, struct ib_device *device);
689 void sdp_tx_ring_destroy(struct sdp_sock *ssk);
690 int sdp_xmit_poll(struct sdp_sock *ssk, int force);
691 void sdp_post_send(struct sdp_sock *ssk, struct mbuf *mb);
692 void sdp_post_sends(struct sdp_sock *ssk, int wait);
693 void sdp_post_keepalive(struct sdp_sock *ssk);
696 void sdp_rx_ring_init(struct sdp_sock *ssk);
697 int sdp_rx_ring_create(struct sdp_sock *ssk, struct ib_device *device);
698 void sdp_rx_ring_destroy(struct sdp_sock *ssk);
699 int sdp_resize_buffers(struct sdp_sock *ssk, u32 new_size);
700 int sdp_init_buffers(struct sdp_sock *ssk, u32 new_size);
701 void sdp_do_posts(struct sdp_sock *ssk);
702 void sdp_rx_comp_full(struct sdp_sock *ssk);
705 int sdp_sendmsg_zcopy(struct kiocb *iocb, struct socket *sk, struct iovec *iov);
706 int sdp_handle_srcavail(struct sdp_sock *ssk, struct sdp_srcah *srcah);
707 void sdp_handle_sendsm(struct sdp_sock *ssk, u32 mseq_ack);
708 void sdp_handle_rdma_read_compl(struct sdp_sock *ssk, u32 mseq_ack,
709 u32 bytes_completed);
710 int sdp_handle_rdma_read_cqe(struct sdp_sock *ssk);
711 int sdp_rdma_to_iovec(struct socket *sk, struct iovec *iov, struct mbuf *mb,
712 unsigned long *used);
713 int sdp_post_rdma_rd_compl(struct sdp_sock *ssk,
714 struct rx_srcavail_state *rx_sa);
715 int sdp_post_sendsm(struct socket *sk);
716 void srcavail_cancel_timeout(struct work_struct *work);
717 void sdp_abort_srcavail(struct socket *sk);
718 void sdp_abort_rdma_read(struct socket *sk);
719 int sdp_process_rx(struct sdp_sock *ssk);