Root/net/rds/iw_send.c

1/*
2 * Copyright (c) 2006 Oracle. All rights reserved.
3 *
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
22 *
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30 * SOFTWARE.
31 *
32 */
33#include <linux/kernel.h>
34#include <linux/in.h>
35#include <linux/device.h>
36#include <linux/dmapool.h>
37
38#include "rds.h"
39#include "rdma.h"
40#include "iw.h"
41
42static void rds_iw_send_rdma_complete(struct rds_message *rm,
43                      int wc_status)
44{
45    int notify_status;
46
47    switch (wc_status) {
48    case IB_WC_WR_FLUSH_ERR:
49        return;
50
51    case IB_WC_SUCCESS:
52        notify_status = RDS_RDMA_SUCCESS;
53        break;
54
55    case IB_WC_REM_ACCESS_ERR:
56        notify_status = RDS_RDMA_REMOTE_ERROR;
57        break;
58
59    default:
60        notify_status = RDS_RDMA_OTHER_ERROR;
61        break;
62    }
63    rds_rdma_send_complete(rm, notify_status);
64}
65
66static void rds_iw_send_unmap_rdma(struct rds_iw_connection *ic,
67                   struct rds_rdma_op *op)
68{
69    if (op->r_mapped) {
70        ib_dma_unmap_sg(ic->i_cm_id->device,
71            op->r_sg, op->r_nents,
72            op->r_write ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
73        op->r_mapped = 0;
74    }
75}
76
77static void rds_iw_send_unmap_rm(struct rds_iw_connection *ic,
78              struct rds_iw_send_work *send,
79              int wc_status)
80{
81    struct rds_message *rm = send->s_rm;
82
83    rdsdebug("ic %p send %p rm %p\n", ic, send, rm);
84
85    ib_dma_unmap_sg(ic->i_cm_id->device,
86             rm->m_sg, rm->m_nents,
87             DMA_TO_DEVICE);
88
89    if (rm->m_rdma_op != NULL) {
90        rds_iw_send_unmap_rdma(ic, rm->m_rdma_op);
91
92        /* If the user asked for a completion notification on this
93         * message, we can implement three different semantics:
94         * 1. Notify when we received the ACK on the RDS message
95         * that was queued with the RDMA. This provides reliable
96         * notification of RDMA status at the expense of a one-way
97         * packet delay.
98         * 2. Notify when the IB stack gives us the completion event for
99         * the RDMA operation.
100         * 3. Notify when the IB stack gives us the completion event for
101         * the accompanying RDS messages.
102         * Here, we implement approach #3. To implement approach #2,
103         * call rds_rdma_send_complete from the cq_handler. To implement #1,
104         * don't call rds_rdma_send_complete at all, and fall back to the notify
105         * handling in the ACK processing code.
106         *
107         * Note: There's no need to explicitly sync any RDMA buffers using
108         * ib_dma_sync_sg_for_cpu - the completion for the RDMA
109         * operation itself unmapped the RDMA buffers, which takes care
110         * of synching.
111         */
112        rds_iw_send_rdma_complete(rm, wc_status);
113
114        if (rm->m_rdma_op->r_write)
115            rds_stats_add(s_send_rdma_bytes, rm->m_rdma_op->r_bytes);
116        else
117            rds_stats_add(s_recv_rdma_bytes, rm->m_rdma_op->r_bytes);
118    }
119
120    /* If anyone waited for this message to get flushed out, wake
121     * them up now */
122    rds_message_unmapped(rm);
123
124    rds_message_put(rm);
125    send->s_rm = NULL;
126}
127
128void rds_iw_send_init_ring(struct rds_iw_connection *ic)
129{
130    struct rds_iw_send_work *send;
131    u32 i;
132
133    for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
134        struct ib_sge *sge;
135
136        send->s_rm = NULL;
137        send->s_op = NULL;
138        send->s_mapping = NULL;
139
140        send->s_wr.next = NULL;
141        send->s_wr.wr_id = i;
142        send->s_wr.sg_list = send->s_sge;
143        send->s_wr.num_sge = 1;
144        send->s_wr.opcode = IB_WR_SEND;
145        send->s_wr.send_flags = 0;
146        send->s_wr.ex.imm_data = 0;
147
148        sge = rds_iw_data_sge(ic, send->s_sge);
149        sge->lkey = 0;
150
151        sge = rds_iw_header_sge(ic, send->s_sge);
152        sge->addr = ic->i_send_hdrs_dma + (i * sizeof(struct rds_header));
153        sge->length = sizeof(struct rds_header);
154        sge->lkey = 0;
155
156        send->s_mr = ib_alloc_fast_reg_mr(ic->i_pd, fastreg_message_size);
157        if (IS_ERR(send->s_mr)) {
158            printk(KERN_WARNING "RDS/IW: ib_alloc_fast_reg_mr failed\n");
159            break;
160        }
161
162        send->s_page_list = ib_alloc_fast_reg_page_list(
163            ic->i_cm_id->device, fastreg_message_size);
164        if (IS_ERR(send->s_page_list)) {
165            printk(KERN_WARNING "RDS/IW: ib_alloc_fast_reg_page_list failed\n");
166            break;
167        }
168    }
169}
170
171void rds_iw_send_clear_ring(struct rds_iw_connection *ic)
172{
173    struct rds_iw_send_work *send;
174    u32 i;
175
176    for (i = 0, send = ic->i_sends; i < ic->i_send_ring.w_nr; i++, send++) {
177        BUG_ON(!send->s_mr);
178        ib_dereg_mr(send->s_mr);
179        BUG_ON(!send->s_page_list);
180        ib_free_fast_reg_page_list(send->s_page_list);
181        if (send->s_wr.opcode == 0xdead)
182            continue;
183        if (send->s_rm)
184            rds_iw_send_unmap_rm(ic, send, IB_WC_WR_FLUSH_ERR);
185        if (send->s_op)
186            rds_iw_send_unmap_rdma(ic, send->s_op);
187    }
188}
189
190/*
191 * The _oldest/_free ring operations here race cleanly with the alloc/unalloc
192 * operations performed in the send path. As the sender allocs and potentially
193 * unallocs the next free entry in the ring it doesn't alter which is
194 * the next to be freed, which is what this is concerned with.
195 */
196void rds_iw_send_cq_comp_handler(struct ib_cq *cq, void *context)
197{
198    struct rds_connection *conn = context;
199    struct rds_iw_connection *ic = conn->c_transport_data;
200    struct ib_wc wc;
201    struct rds_iw_send_work *send;
202    u32 completed;
203    u32 oldest;
204    u32 i;
205    int ret;
206
207    rdsdebug("cq %p conn %p\n", cq, conn);
208    rds_iw_stats_inc(s_iw_tx_cq_call);
209    ret = ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
210    if (ret)
211        rdsdebug("ib_req_notify_cq send failed: %d\n", ret);
212
213    while (ib_poll_cq(cq, 1, &wc) > 0) {
214        rdsdebug("wc wr_id 0x%llx status %u byte_len %u imm_data %u\n",
215             (unsigned long long)wc.wr_id, wc.status, wc.byte_len,
216             be32_to_cpu(wc.ex.imm_data));
217        rds_iw_stats_inc(s_iw_tx_cq_event);
218
219        if (wc.status != IB_WC_SUCCESS) {
220            printk(KERN_ERR "WC Error: status = %d opcode = %d\n", wc.status, wc.opcode);
221            break;
222        }
223
224        if (wc.opcode == IB_WC_LOCAL_INV && wc.wr_id == RDS_IW_LOCAL_INV_WR_ID) {
225            ic->i_fastreg_posted = 0;
226            continue;
227        }
228
229        if (wc.opcode == IB_WC_FAST_REG_MR && wc.wr_id == RDS_IW_FAST_REG_WR_ID) {
230            ic->i_fastreg_posted = 1;
231            continue;
232        }
233
234        if (wc.wr_id == RDS_IW_ACK_WR_ID) {
235            if (ic->i_ack_queued + HZ/2 < jiffies)
236                rds_iw_stats_inc(s_iw_tx_stalled);
237            rds_iw_ack_send_complete(ic);
238            continue;
239        }
240
241        oldest = rds_iw_ring_oldest(&ic->i_send_ring);
242
243        completed = rds_iw_ring_completed(&ic->i_send_ring, wc.wr_id, oldest);
244
245        for (i = 0; i < completed; i++) {
246            send = &ic->i_sends[oldest];
247
248            /* In the error case, wc.opcode sometimes contains garbage */
249            switch (send->s_wr.opcode) {
250            case IB_WR_SEND:
251                if (send->s_rm)
252                    rds_iw_send_unmap_rm(ic, send, wc.status);
253                break;
254            case IB_WR_FAST_REG_MR:
255            case IB_WR_RDMA_WRITE:
256            case IB_WR_RDMA_READ:
257            case IB_WR_RDMA_READ_WITH_INV:
258                /* Nothing to be done - the SG list will be unmapped
259                 * when the SEND completes. */
260                break;
261            default:
262                if (printk_ratelimit())
263                    printk(KERN_NOTICE
264                        "RDS/IW: %s: unexpected opcode 0x%x in WR!\n",
265                        __func__, send->s_wr.opcode);
266                break;
267            }
268
269            send->s_wr.opcode = 0xdead;
270            send->s_wr.num_sge = 1;
271            if (send->s_queued + HZ/2 < jiffies)
272                rds_iw_stats_inc(s_iw_tx_stalled);
273
274            /* If a RDMA operation produced an error, signal this right
275             * away. If we don't, the subsequent SEND that goes with this
276             * RDMA will be canceled with ERR_WFLUSH, and the application
277             * never learn that the RDMA failed. */
278            if (unlikely(wc.status == IB_WC_REM_ACCESS_ERR && send->s_op)) {
279                struct rds_message *rm;
280
281                rm = rds_send_get_message(conn, send->s_op);
282                if (rm)
283                    rds_iw_send_rdma_complete(rm, wc.status);
284            }
285
286            oldest = (oldest + 1) % ic->i_send_ring.w_nr;
287        }
288
289        rds_iw_ring_free(&ic->i_send_ring, completed);
290
291        if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags) ||
292            test_bit(0, &conn->c_map_queued))
293            queue_delayed_work(rds_wq, &conn->c_send_w, 0);
294
295        /* We expect errors as the qp is drained during shutdown */
296        if (wc.status != IB_WC_SUCCESS && rds_conn_up(conn)) {
297            rds_iw_conn_error(conn,
298                "send completion on %pI4 "
299                "had status %u, disconnecting and reconnecting\n",
300                &conn->c_faddr, wc.status);
301        }
302    }
303}
304
305/*
306 * This is the main function for allocating credits when sending
307 * messages.
308 *
309 * Conceptually, we have two counters:
310 * - send credits: this tells us how many WRs we're allowed
311 * to submit without overruning the reciever's queue. For
312 * each SEND WR we post, we decrement this by one.
313 *
314 * - posted credits: this tells us how many WRs we recently
315 * posted to the receive queue. This value is transferred
316 * to the peer as a "credit update" in a RDS header field.
317 * Every time we transmit credits to the peer, we subtract
318 * the amount of transferred credits from this counter.
319 *
320 * It is essential that we avoid situations where both sides have
321 * exhausted their send credits, and are unable to send new credits
322 * to the peer. We achieve this by requiring that we send at least
323 * one credit update to the peer before exhausting our credits.
324 * When new credits arrive, we subtract one credit that is withheld
325 * until we've posted new buffers and are ready to transmit these
326 * credits (see rds_iw_send_add_credits below).
327 *
328 * The RDS send code is essentially single-threaded; rds_send_xmit
329 * grabs c_send_lock to ensure exclusive access to the send ring.
330 * However, the ACK sending code is independent and can race with
331 * message SENDs.
332 *
333 * In the send path, we need to update the counters for send credits
334 * and the counter of posted buffers atomically - when we use the
335 * last available credit, we cannot allow another thread to race us
336 * and grab the posted credits counter. Hence, we have to use a
337 * spinlock to protect the credit counter, or use atomics.
338 *
339 * Spinlocks shared between the send and the receive path are bad,
340 * because they create unnecessary delays. An early implementation
341 * using a spinlock showed a 5% degradation in throughput at some
342 * loads.
343 *
344 * This implementation avoids spinlocks completely, putting both
345 * counters into a single atomic, and updating that atomic using
346 * atomic_add (in the receive path, when receiving fresh credits),
347 * and using atomic_cmpxchg when updating the two counters.
348 */
349int rds_iw_send_grab_credits(struct rds_iw_connection *ic,
350                 u32 wanted, u32 *adv_credits, int need_posted, int max_posted)
351{
352    unsigned int avail, posted, got = 0, advertise;
353    long oldval, newval;
354
355    *adv_credits = 0;
356    if (!ic->i_flowctl)
357        return wanted;
358
359try_again:
360    advertise = 0;
361    oldval = newval = atomic_read(&ic->i_credits);
362    posted = IB_GET_POST_CREDITS(oldval);
363    avail = IB_GET_SEND_CREDITS(oldval);
364
365    rdsdebug("rds_iw_send_grab_credits(%u): credits=%u posted=%u\n",
366            wanted, avail, posted);
367
368    /* The last credit must be used to send a credit update. */
369    if (avail && !posted)
370        avail--;
371
372    if (avail < wanted) {
373        struct rds_connection *conn = ic->i_cm_id->context;
374
375        /* Oops, there aren't that many credits left! */
376        set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
377        got = avail;
378    } else {
379        /* Sometimes you get what you want, lalala. */
380        got = wanted;
381    }
382    newval -= IB_SET_SEND_CREDITS(got);
383
384    /*
385     * If need_posted is non-zero, then the caller wants
386     * the posted regardless of whether any send credits are
387     * available.
388     */
389    if (posted && (got || need_posted)) {
390        advertise = min_t(unsigned int, posted, max_posted);
391        newval -= IB_SET_POST_CREDITS(advertise);
392    }
393
394    /* Finally bill everything */
395    if (atomic_cmpxchg(&ic->i_credits, oldval, newval) != oldval)
396        goto try_again;
397
398    *adv_credits = advertise;
399    return got;
400}
401
402void rds_iw_send_add_credits(struct rds_connection *conn, unsigned int credits)
403{
404    struct rds_iw_connection *ic = conn->c_transport_data;
405
406    if (credits == 0)
407        return;
408
409    rdsdebug("rds_iw_send_add_credits(%u): current=%u%s\n",
410            credits,
411            IB_GET_SEND_CREDITS(atomic_read(&ic->i_credits)),
412            test_bit(RDS_LL_SEND_FULL, &conn->c_flags) ? ", ll_send_full" : "");
413
414    atomic_add(IB_SET_SEND_CREDITS(credits), &ic->i_credits);
415    if (test_and_clear_bit(RDS_LL_SEND_FULL, &conn->c_flags))
416        queue_delayed_work(rds_wq, &conn->c_send_w, 0);
417
418    WARN_ON(IB_GET_SEND_CREDITS(credits) >= 16384);
419
420    rds_iw_stats_inc(s_iw_rx_credit_updates);
421}
422
423void rds_iw_advertise_credits(struct rds_connection *conn, unsigned int posted)
424{
425    struct rds_iw_connection *ic = conn->c_transport_data;
426
427    if (posted == 0)
428        return;
429
430    atomic_add(IB_SET_POST_CREDITS(posted), &ic->i_credits);
431
432    /* Decide whether to send an update to the peer now.
433     * If we would send a credit update for every single buffer we
434     * post, we would end up with an ACK storm (ACK arrives,
435     * consumes buffer, we refill the ring, send ACK to remote
436     * advertising the newly posted buffer... ad inf)
437     *
438     * Performance pretty much depends on how often we send
439     * credit updates - too frequent updates mean lots of ACKs.
440     * Too infrequent updates, and the peer will run out of
441     * credits and has to throttle.
442     * For the time being, 16 seems to be a good compromise.
443     */
444    if (IB_GET_POST_CREDITS(atomic_read(&ic->i_credits)) >= 16)
445        set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
446}
447
448static inline void
449rds_iw_xmit_populate_wr(struct rds_iw_connection *ic,
450        struct rds_iw_send_work *send, unsigned int pos,
451        unsigned long buffer, unsigned int length,
452        int send_flags)
453{
454    struct ib_sge *sge;
455
456    WARN_ON(pos != send - ic->i_sends);
457
458    send->s_wr.send_flags = send_flags;
459    send->s_wr.opcode = IB_WR_SEND;
460    send->s_wr.num_sge = 2;
461    send->s_wr.next = NULL;
462    send->s_queued = jiffies;
463    send->s_op = NULL;
464
465    if (length != 0) {
466        sge = rds_iw_data_sge(ic, send->s_sge);
467        sge->addr = buffer;
468        sge->length = length;
469        sge->lkey = rds_iw_local_dma_lkey(ic);
470
471        sge = rds_iw_header_sge(ic, send->s_sge);
472    } else {
473        /* We're sending a packet with no payload. There is only
474         * one SGE */
475        send->s_wr.num_sge = 1;
476        sge = &send->s_sge[0];
477    }
478
479    sge->addr = ic->i_send_hdrs_dma + (pos * sizeof(struct rds_header));
480    sge->length = sizeof(struct rds_header);
481    sge->lkey = rds_iw_local_dma_lkey(ic);
482}
483
484/*
485 * This can be called multiple times for a given message. The first time
486 * we see a message we map its scatterlist into the IB device so that
487 * we can provide that mapped address to the IB scatter gather entries
488 * in the IB work requests. We translate the scatterlist into a series
489 * of work requests that fragment the message. These work requests complete
490 * in order so we pass ownership of the message to the completion handler
491 * once we send the final fragment.
492 *
493 * The RDS core uses the c_send_lock to only enter this function once
494 * per connection. This makes sure that the tx ring alloc/unalloc pairs
495 * don't get out of sync and confuse the ring.
496 */
497int rds_iw_xmit(struct rds_connection *conn, struct rds_message *rm,
498        unsigned int hdr_off, unsigned int sg, unsigned int off)
499{
500    struct rds_iw_connection *ic = conn->c_transport_data;
501    struct ib_device *dev = ic->i_cm_id->device;
502    struct rds_iw_send_work *send = NULL;
503    struct rds_iw_send_work *first;
504    struct rds_iw_send_work *prev;
505    struct ib_send_wr *failed_wr;
506    struct scatterlist *scat;
507    u32 pos;
508    u32 i;
509    u32 work_alloc;
510    u32 credit_alloc;
511    u32 posted;
512    u32 adv_credits = 0;
513    int send_flags = 0;
514    int sent;
515    int ret;
516    int flow_controlled = 0;
517
518    BUG_ON(off % RDS_FRAG_SIZE);
519    BUG_ON(hdr_off != 0 && hdr_off != sizeof(struct rds_header));
520
521    /* Fastreg support */
522    if (rds_rdma_cookie_key(rm->m_rdma_cookie) && !ic->i_fastreg_posted) {
523        ret = -EAGAIN;
524        goto out;
525    }
526
527    /* FIXME we may overallocate here */
528    if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0)
529        i = 1;
530    else
531        i = ceil(be32_to_cpu(rm->m_inc.i_hdr.h_len), RDS_FRAG_SIZE);
532
533    work_alloc = rds_iw_ring_alloc(&ic->i_send_ring, i, &pos);
534    if (work_alloc == 0) {
535        set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
536        rds_iw_stats_inc(s_iw_tx_ring_full);
537        ret = -ENOMEM;
538        goto out;
539    }
540
541    credit_alloc = work_alloc;
542    if (ic->i_flowctl) {
543        credit_alloc = rds_iw_send_grab_credits(ic, work_alloc, &posted, 0, RDS_MAX_ADV_CREDIT);
544        adv_credits += posted;
545        if (credit_alloc < work_alloc) {
546            rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc - credit_alloc);
547            work_alloc = credit_alloc;
548            flow_controlled++;
549        }
550        if (work_alloc == 0) {
551            set_bit(RDS_LL_SEND_FULL, &conn->c_flags);
552            rds_iw_stats_inc(s_iw_tx_throttle);
553            ret = -ENOMEM;
554            goto out;
555        }
556    }
557
558    /* map the message the first time we see it */
559    if (ic->i_rm == NULL) {
560        /*
561        printk(KERN_NOTICE "rds_iw_xmit prep msg dport=%u flags=0x%x len=%d\n",
562                be16_to_cpu(rm->m_inc.i_hdr.h_dport),
563                rm->m_inc.i_hdr.h_flags,
564                be32_to_cpu(rm->m_inc.i_hdr.h_len));
565           */
566        if (rm->m_nents) {
567            rm->m_count = ib_dma_map_sg(dev,
568                     rm->m_sg, rm->m_nents, DMA_TO_DEVICE);
569            rdsdebug("ic %p mapping rm %p: %d\n", ic, rm, rm->m_count);
570            if (rm->m_count == 0) {
571                rds_iw_stats_inc(s_iw_tx_sg_mapping_failure);
572                rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
573                ret = -ENOMEM; /* XXX ? */
574                goto out;
575            }
576        } else {
577            rm->m_count = 0;
578        }
579
580        ic->i_unsignaled_wrs = rds_iw_sysctl_max_unsig_wrs;
581        ic->i_unsignaled_bytes = rds_iw_sysctl_max_unsig_bytes;
582        rds_message_addref(rm);
583        ic->i_rm = rm;
584
585        /* Finalize the header */
586        if (test_bit(RDS_MSG_ACK_REQUIRED, &rm->m_flags))
587            rm->m_inc.i_hdr.h_flags |= RDS_FLAG_ACK_REQUIRED;
588        if (test_bit(RDS_MSG_RETRANSMITTED, &rm->m_flags))
589            rm->m_inc.i_hdr.h_flags |= RDS_FLAG_RETRANSMITTED;
590
591        /* If it has a RDMA op, tell the peer we did it. This is
592         * used by the peer to release use-once RDMA MRs. */
593        if (rm->m_rdma_op) {
594            struct rds_ext_header_rdma ext_hdr;
595
596            ext_hdr.h_rdma_rkey = cpu_to_be32(rm->m_rdma_op->r_key);
597            rds_message_add_extension(&rm->m_inc.i_hdr,
598                    RDS_EXTHDR_RDMA, &ext_hdr, sizeof(ext_hdr));
599        }
600        if (rm->m_rdma_cookie) {
601            rds_message_add_rdma_dest_extension(&rm->m_inc.i_hdr,
602                    rds_rdma_cookie_key(rm->m_rdma_cookie),
603                    rds_rdma_cookie_offset(rm->m_rdma_cookie));
604        }
605
606        /* Note - rds_iw_piggyb_ack clears the ACK_REQUIRED bit, so
607         * we should not do this unless we have a chance of at least
608         * sticking the header into the send ring. Which is why we
609         * should call rds_iw_ring_alloc first. */
610        rm->m_inc.i_hdr.h_ack = cpu_to_be64(rds_iw_piggyb_ack(ic));
611        rds_message_make_checksum(&rm->m_inc.i_hdr);
612
613        /*
614         * Update adv_credits since we reset the ACK_REQUIRED bit.
615         */
616        rds_iw_send_grab_credits(ic, 0, &posted, 1, RDS_MAX_ADV_CREDIT - adv_credits);
617        adv_credits += posted;
618        BUG_ON(adv_credits > 255);
619    } else if (ic->i_rm != rm)
620        BUG();
621
622    send = &ic->i_sends[pos];
623    first = send;
624    prev = NULL;
625    scat = &rm->m_sg[sg];
626    sent = 0;
627    i = 0;
628
629    /* Sometimes you want to put a fence between an RDMA
630     * READ and the following SEND.
631     * We could either do this all the time
632     * or when requested by the user. Right now, we let
633     * the application choose.
634     */
635    if (rm->m_rdma_op && rm->m_rdma_op->r_fence)
636        send_flags = IB_SEND_FENCE;
637
638    /*
639     * We could be copying the header into the unused tail of the page.
640     * That would need to be changed in the future when those pages might
641     * be mapped userspace pages or page cache pages. So instead we always
642     * use a second sge and our long-lived ring of mapped headers. We send
643     * the header after the data so that the data payload can be aligned on
644     * the receiver.
645     */
646
647    /* handle a 0-len message */
648    if (be32_to_cpu(rm->m_inc.i_hdr.h_len) == 0) {
649        rds_iw_xmit_populate_wr(ic, send, pos, 0, 0, send_flags);
650        goto add_header;
651    }
652
653    /* if there's data reference it with a chain of work reqs */
654    for (; i < work_alloc && scat != &rm->m_sg[rm->m_count]; i++) {
655        unsigned int len;
656
657        send = &ic->i_sends[pos];
658
659        len = min(RDS_FRAG_SIZE, ib_sg_dma_len(dev, scat) - off);
660        rds_iw_xmit_populate_wr(ic, send, pos,
661                ib_sg_dma_address(dev, scat) + off, len,
662                send_flags);
663
664        /*
665         * We want to delay signaling completions just enough to get
666         * the batching benefits but not so much that we create dead time
667         * on the wire.
668         */
669        if (ic->i_unsignaled_wrs-- == 0) {
670            ic->i_unsignaled_wrs = rds_iw_sysctl_max_unsig_wrs;
671            send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
672        }
673
674        ic->i_unsignaled_bytes -= len;
675        if (ic->i_unsignaled_bytes <= 0) {
676            ic->i_unsignaled_bytes = rds_iw_sysctl_max_unsig_bytes;
677            send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
678        }
679
680        /*
681         * Always signal the last one if we're stopping due to flow control.
682         */
683        if (flow_controlled && i == (work_alloc-1))
684            send->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
685
686        rdsdebug("send %p wr %p num_sge %u next %p\n", send,
687             &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
688
689        sent += len;
690        off += len;
691        if (off == ib_sg_dma_len(dev, scat)) {
692            scat++;
693            off = 0;
694        }
695
696add_header:
697        /* Tack on the header after the data. The header SGE should already
698         * have been set up to point to the right header buffer. */
699        memcpy(&ic->i_send_hdrs[pos], &rm->m_inc.i_hdr, sizeof(struct rds_header));
700
701        if (0) {
702            struct rds_header *hdr = &ic->i_send_hdrs[pos];
703
704            printk(KERN_NOTICE "send WR dport=%u flags=0x%x len=%d\n",
705                be16_to_cpu(hdr->h_dport),
706                hdr->h_flags,
707                be32_to_cpu(hdr->h_len));
708        }
709        if (adv_credits) {
710            struct rds_header *hdr = &ic->i_send_hdrs[pos];
711
712            /* add credit and redo the header checksum */
713            hdr->h_credit = adv_credits;
714            rds_message_make_checksum(hdr);
715            adv_credits = 0;
716            rds_iw_stats_inc(s_iw_tx_credit_updates);
717        }
718
719        if (prev)
720            prev->s_wr.next = &send->s_wr;
721        prev = send;
722
723        pos = (pos + 1) % ic->i_send_ring.w_nr;
724    }
725
726    /* Account the RDS header in the number of bytes we sent, but just once.
727     * The caller has no concept of fragmentation. */
728    if (hdr_off == 0)
729        sent += sizeof(struct rds_header);
730
731    /* if we finished the message then send completion owns it */
732    if (scat == &rm->m_sg[rm->m_count]) {
733        prev->s_rm = ic->i_rm;
734        prev->s_wr.send_flags |= IB_SEND_SIGNALED | IB_SEND_SOLICITED;
735        ic->i_rm = NULL;
736    }
737
738    if (i < work_alloc) {
739        rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc - i);
740        work_alloc = i;
741    }
742    if (ic->i_flowctl && i < credit_alloc)
743        rds_iw_send_add_credits(conn, credit_alloc - i);
744
745    /* XXX need to worry about failed_wr and partial sends. */
746    failed_wr = &first->s_wr;
747    ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
748    rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
749         first, &first->s_wr, ret, failed_wr);
750    BUG_ON(failed_wr != &first->s_wr);
751    if (ret) {
752        printk(KERN_WARNING "RDS/IW: ib_post_send to %pI4 "
753               "returned %d\n", &conn->c_faddr, ret);
754        rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
755        if (prev->s_rm) {
756            ic->i_rm = prev->s_rm;
757            prev->s_rm = NULL;
758        }
759        goto out;
760    }
761
762    ret = sent;
763out:
764    BUG_ON(adv_credits);
765    return ret;
766}
767
768static void rds_iw_build_send_fastreg(struct rds_iw_device *rds_iwdev, struct rds_iw_connection *ic, struct rds_iw_send_work *send, int nent, int len, u64 sg_addr)
769{
770    BUG_ON(nent > send->s_page_list->max_page_list_len);
771    /*
772     * Perform a WR for the fast_reg_mr. Each individual page
773     * in the sg list is added to the fast reg page list and placed
774     * inside the fast_reg_mr WR.
775     */
776    send->s_wr.opcode = IB_WR_FAST_REG_MR;
777    send->s_wr.wr.fast_reg.length = len;
778    send->s_wr.wr.fast_reg.rkey = send->s_mr->rkey;
779    send->s_wr.wr.fast_reg.page_list = send->s_page_list;
780    send->s_wr.wr.fast_reg.page_list_len = nent;
781    send->s_wr.wr.fast_reg.page_shift = PAGE_SHIFT;
782    send->s_wr.wr.fast_reg.access_flags = IB_ACCESS_REMOTE_WRITE;
783    send->s_wr.wr.fast_reg.iova_start = sg_addr;
784
785    ib_update_fast_reg_key(send->s_mr, send->s_remap_count++);
786}
787
788int rds_iw_xmit_rdma(struct rds_connection *conn, struct rds_rdma_op *op)
789{
790    struct rds_iw_connection *ic = conn->c_transport_data;
791    struct rds_iw_send_work *send = NULL;
792    struct rds_iw_send_work *first;
793    struct rds_iw_send_work *prev;
794    struct ib_send_wr *failed_wr;
795    struct rds_iw_device *rds_iwdev;
796    struct scatterlist *scat;
797    unsigned long len;
798    u64 remote_addr = op->r_remote_addr;
799    u32 pos, fr_pos;
800    u32 work_alloc;
801    u32 i;
802    u32 j;
803    int sent;
804    int ret;
805    int num_sge;
806
807    rds_iwdev = ib_get_client_data(ic->i_cm_id->device, &rds_iw_client);
808
809    /* map the message the first time we see it */
810    if (!op->r_mapped) {
811        op->r_count = ib_dma_map_sg(ic->i_cm_id->device,
812                    op->r_sg, op->r_nents, (op->r_write) ?
813                    DMA_TO_DEVICE : DMA_FROM_DEVICE);
814        rdsdebug("ic %p mapping op %p: %d\n", ic, op, op->r_count);
815        if (op->r_count == 0) {
816            rds_iw_stats_inc(s_iw_tx_sg_mapping_failure);
817            ret = -ENOMEM; /* XXX ? */
818            goto out;
819        }
820
821        op->r_mapped = 1;
822    }
823
824    if (!op->r_write) {
825        /* Alloc space on the send queue for the fastreg */
826        work_alloc = rds_iw_ring_alloc(&ic->i_send_ring, 1, &fr_pos);
827        if (work_alloc != 1) {
828            rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
829            rds_iw_stats_inc(s_iw_tx_ring_full);
830            ret = -ENOMEM;
831            goto out;
832        }
833    }
834
835    /*
836     * Instead of knowing how to return a partial rdma read/write we insist that there
837     * be enough work requests to send the entire message.
838     */
839    i = ceil(op->r_count, rds_iwdev->max_sge);
840
841    work_alloc = rds_iw_ring_alloc(&ic->i_send_ring, i, &pos);
842    if (work_alloc != i) {
843        rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
844        rds_iw_stats_inc(s_iw_tx_ring_full);
845        ret = -ENOMEM;
846        goto out;
847    }
848
849    send = &ic->i_sends[pos];
850    if (!op->r_write) {
851        first = prev = &ic->i_sends[fr_pos];
852    } else {
853        first = send;
854        prev = NULL;
855    }
856    scat = &op->r_sg[0];
857    sent = 0;
858    num_sge = op->r_count;
859
860    for (i = 0; i < work_alloc && scat != &op->r_sg[op->r_count]; i++) {
861        send->s_wr.send_flags = 0;
862        send->s_queued = jiffies;
863
864        /*
865         * We want to delay signaling completions just enough to get
866         * the batching benefits but not so much that we create dead time on the wire.
867         */
868        if (ic->i_unsignaled_wrs-- == 0) {
869            ic->i_unsignaled_wrs = rds_iw_sysctl_max_unsig_wrs;
870            send->s_wr.send_flags = IB_SEND_SIGNALED;
871        }
872
873        /* To avoid the need to have the plumbing to invalidate the fastreg_mr used
874         * for local access after RDS is finished with it, using
875         * IB_WR_RDMA_READ_WITH_INV will invalidate it after the read has completed.
876         */
877        if (op->r_write)
878            send->s_wr.opcode = IB_WR_RDMA_WRITE;
879        else
880            send->s_wr.opcode = IB_WR_RDMA_READ_WITH_INV;
881
882        send->s_wr.wr.rdma.remote_addr = remote_addr;
883        send->s_wr.wr.rdma.rkey = op->r_key;
884        send->s_op = op;
885
886        if (num_sge > rds_iwdev->max_sge) {
887            send->s_wr.num_sge = rds_iwdev->max_sge;
888            num_sge -= rds_iwdev->max_sge;
889        } else
890            send->s_wr.num_sge = num_sge;
891
892        send->s_wr.next = NULL;
893
894        if (prev)
895            prev->s_wr.next = &send->s_wr;
896
897        for (j = 0; j < send->s_wr.num_sge && scat != &op->r_sg[op->r_count]; j++) {
898            len = ib_sg_dma_len(ic->i_cm_id->device, scat);
899
900            if (send->s_wr.opcode == IB_WR_RDMA_READ_WITH_INV)
901                send->s_page_list->page_list[j] = ib_sg_dma_address(ic->i_cm_id->device, scat);
902            else {
903                send->s_sge[j].addr = ib_sg_dma_address(ic->i_cm_id->device, scat);
904                send->s_sge[j].length = len;
905                send->s_sge[j].lkey = rds_iw_local_dma_lkey(ic);
906            }
907
908            sent += len;
909            rdsdebug("ic %p sent %d remote_addr %llu\n", ic, sent, remote_addr);
910            remote_addr += len;
911
912            scat++;
913        }
914
915        if (send->s_wr.opcode == IB_WR_RDMA_READ_WITH_INV) {
916            send->s_wr.num_sge = 1;
917            send->s_sge[0].addr = conn->c_xmit_rm->m_rs->rs_user_addr;
918            send->s_sge[0].length = conn->c_xmit_rm->m_rs->rs_user_bytes;
919            send->s_sge[0].lkey = ic->i_sends[fr_pos].s_mr->lkey;
920        }
921
922        rdsdebug("send %p wr %p num_sge %u next %p\n", send,
923            &send->s_wr, send->s_wr.num_sge, send->s_wr.next);
924
925        prev = send;
926        if (++send == &ic->i_sends[ic->i_send_ring.w_nr])
927            send = ic->i_sends;
928    }
929
930    /* if we finished the message then send completion owns it */
931    if (scat == &op->r_sg[op->r_count])
932        first->s_wr.send_flags = IB_SEND_SIGNALED;
933
934    if (i < work_alloc) {
935        rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc - i);
936        work_alloc = i;
937    }
938
939    /* On iWARP, local memory access by a remote system (ie, RDMA Read) is not
940     * recommended. Putting the lkey on the wire is a security hole, as it can
941     * allow for memory access to all of memory on the remote system. Some
942     * adapters do not allow using the lkey for this at all. To bypass this use a
943     * fastreg_mr (or possibly a dma_mr)
944     */
945    if (!op->r_write) {
946        rds_iw_build_send_fastreg(rds_iwdev, ic, &ic->i_sends[fr_pos],
947            op->r_count, sent, conn->c_xmit_rm->m_rs->rs_user_addr);
948        work_alloc++;
949    }
950
951    failed_wr = &first->s_wr;
952    ret = ib_post_send(ic->i_cm_id->qp, &first->s_wr, &failed_wr);
953    rdsdebug("ic %p first %p (wr %p) ret %d wr %p\n", ic,
954         first, &first->s_wr, ret, failed_wr);
955    BUG_ON(failed_wr != &first->s_wr);
956    if (ret) {
957        printk(KERN_WARNING "RDS/IW: rdma ib_post_send to %pI4 "
958               "returned %d\n", &conn->c_faddr, ret);
959        rds_iw_ring_unalloc(&ic->i_send_ring, work_alloc);
960        goto out;
961    }
962
963out:
964    return ret;
965}
966
967void rds_iw_xmit_complete(struct rds_connection *conn)
968{
969    struct rds_iw_connection *ic = conn->c_transport_data;
970
971    /* We may have a pending ACK or window update we were unable
972     * to send previously (due to flow control). Try again. */
973    rds_iw_attempt_ack(ic);
974}
975

Archive Download this file



interactive