47
48#include <sys/types.h>
49#include <sys/mbuf.h>
50#include <sys/smp.h>
51#include <sys/socket.h>
52#include <sys/sysctl.h>
53#include <sys/syslog.h>
54
55#include <net/bpf.h>
56#include <net/ethernet.h>
57#include <net/if.h>
58#include <net/if_vlan_var.h>
59
60#include <netinet/in.h>
61#include <netinet/ip.h>
62#include <netinet/ip6.h>
63#include <netinet/tcp.h>
64
65#include "common/efx.h"
66
67#include "sfxge.h"
68#include "sfxge_tx.h"
69
70/* Set the block level to ensure there is space to generate a
71 * large number of descriptors for TSO. With minimum MSS and
72 * maximum mbuf length we might need more than a ring-ful of
73 * descriptors, but this should not happen in practice except
74 * due to deliberate attack. In that case we will truncate
75 * the output at a packet boundary.
76 */
77#define SFXGE_TSO_MAX_DESC \
78 (SFXGE_TSO_MAX_SEGS * 2 + SFXGE_TX_MAPPING_MAX_SEG - 1)
79#define SFXGE_TXQ_BLOCK_LEVEL(_entries) ((_entries) - SFXGE_TSO_MAX_DESC)
80
81#ifdef SFXGE_HAVE_MQ
82
83#define SFXGE_PARAM_TX_DPL_GET_MAX SFXGE_PARAM(tx_dpl_get_max)
84static int sfxge_tx_dpl_get_max = SFXGE_TX_DPL_GET_PKT_LIMIT_DEFAULT;
85TUNABLE_INT(SFXGE_PARAM_TX_DPL_GET_MAX, &sfxge_tx_dpl_get_max);
86SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_get_max, CTLFLAG_RDTUN,
87 &sfxge_tx_dpl_get_max, 0,
88 "Maximum number of any packets in deferred packet get-list");
89
90#define SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX \
91 SFXGE_PARAM(tx_dpl_get_non_tcp_max)
92static int sfxge_tx_dpl_get_non_tcp_max =
93 SFXGE_TX_DPL_GET_NON_TCP_PKT_LIMIT_DEFAULT;
94TUNABLE_INT(SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX, &sfxge_tx_dpl_get_non_tcp_max);
95SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_get_non_tcp_max, CTLFLAG_RDTUN,
96 &sfxge_tx_dpl_get_non_tcp_max, 0,
97 "Maximum number of non-TCP packets in deferred packet get-list");
98
99#define SFXGE_PARAM_TX_DPL_PUT_MAX SFXGE_PARAM(tx_dpl_put_max)
100static int sfxge_tx_dpl_put_max = SFXGE_TX_DPL_PUT_PKT_LIMIT_DEFAULT;
101TUNABLE_INT(SFXGE_PARAM_TX_DPL_PUT_MAX, &sfxge_tx_dpl_put_max);
102SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_put_max, CTLFLAG_RDTUN,
103 &sfxge_tx_dpl_put_max, 0,
104 "Maximum number of any packets in deferred packet put-list");
105
106#endif
107
108
109/* Forward declarations. */
110static void sfxge_tx_qdpl_service(struct sfxge_txq *txq);
111static void sfxge_tx_qlist_post(struct sfxge_txq *txq);
112static void sfxge_tx_qunblock(struct sfxge_txq *txq);
113static int sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf,
114 const bus_dma_segment_t *dma_seg, int n_dma_seg);
115
116void
117sfxge_tx_qcomplete(struct sfxge_txq *txq, struct sfxge_evq *evq)
118{
119 unsigned int completed;
120
121 SFXGE_EVQ_LOCK_ASSERT_OWNED(evq);
122
123 completed = txq->completed;
124 while (completed != txq->pending) {
125 struct sfxge_tx_mapping *stmp;
126 unsigned int id;
127
128 id = completed++ & txq->ptr_mask;
129
130 stmp = &txq->stmp[id];
131 if (stmp->flags & TX_BUF_UNMAP) {
132 bus_dmamap_unload(txq->packet_dma_tag, stmp->map);
133 if (stmp->flags & TX_BUF_MBUF) {
134 struct mbuf *m = stmp->u.mbuf;
135 do
136 m = m_free(m);
137 while (m != NULL);
138 } else {
139 free(stmp->u.heap_buf, M_SFXGE);
140 }
141 stmp->flags = 0;
142 }
143 }
144 txq->completed = completed;
145
146 /* Check whether we need to unblock the queue. */
147 mb();
148 if (txq->blocked) {
149 unsigned int level;
150
151 level = txq->added - txq->completed;
152 if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries))
153 sfxge_tx_qunblock(txq);
154 }
155}
156
157#ifdef SFXGE_HAVE_MQ
158
159static unsigned int
160sfxge_is_mbuf_non_tcp(struct mbuf *mbuf)
161{
162 /* Absense of TCP checksum flags does not mean that it is non-TCP
163 * but it should be true if user wants to achieve high throughput.
164 */
165 return (!(mbuf->m_pkthdr.csum_flags & (CSUM_IP_TCP | CSUM_IP6_TCP)));
166}
167
168/*
169 * Reorder the put list and append it to the get list.
170 */
171static void
172sfxge_tx_qdpl_swizzle(struct sfxge_txq *txq)
173{
174 struct sfxge_tx_dpl *stdp;
175 struct mbuf *mbuf, *get_next, **get_tailp;
176 volatile uintptr_t *putp;
177 uintptr_t put;
178 unsigned int count;
179 unsigned int non_tcp_count;
180
181 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
182
183 stdp = &txq->dpl;
184
185 /* Acquire the put list. */
186 putp = &stdp->std_put;
187 put = atomic_readandclear_ptr(putp);
188 mbuf = (void *)put;
189
190 if (mbuf == NULL)
191 return;
192
193 /* Reverse the put list. */
194 get_tailp = &mbuf->m_nextpkt;
195 get_next = NULL;
196
197 count = 0;
198 non_tcp_count = 0;
199 do {
200 struct mbuf *put_next;
201
202 non_tcp_count += sfxge_is_mbuf_non_tcp(mbuf);
203 put_next = mbuf->m_nextpkt;
204 mbuf->m_nextpkt = get_next;
205 get_next = mbuf;
206 mbuf = put_next;
207
208 count++;
209 } while (mbuf != NULL);
210
211 /* Append the reversed put list to the get list. */
212 KASSERT(*get_tailp == NULL, ("*get_tailp != NULL"));
213 *stdp->std_getp = get_next;
214 stdp->std_getp = get_tailp;
215 stdp->std_get_count += count;
216 stdp->std_get_non_tcp_count += non_tcp_count;
217}
218
219#endif /* SFXGE_HAVE_MQ */
220
221static void
222sfxge_tx_qreap(struct sfxge_txq *txq)
223{
224 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
225
226 txq->reaped = txq->completed;
227}
228
229static void
230sfxge_tx_qlist_post(struct sfxge_txq *txq)
231{
232 unsigned int old_added;
233 unsigned int level;
234 int rc;
235
236 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
237
238 KASSERT(txq->n_pend_desc != 0, ("txq->n_pend_desc == 0"));
239 KASSERT(txq->n_pend_desc <= SFXGE_TSO_MAX_DESC,
240 ("txq->n_pend_desc too large"));
241 KASSERT(!txq->blocked, ("txq->blocked"));
242
243 old_added = txq->added;
244
245 /* Post the fragment list. */
246 rc = efx_tx_qpost(txq->common, txq->pend_desc, txq->n_pend_desc,
247 txq->reaped, &txq->added);
248 KASSERT(rc == 0, ("efx_tx_qpost() failed"));
249
250 /* If efx_tx_qpost() had to refragment, our information about
251 * buffers to free may be associated with the wrong
252 * descriptors.
253 */
254 KASSERT(txq->added - old_added == txq->n_pend_desc,
255 ("efx_tx_qpost() refragmented descriptors"));
256
257 level = txq->added - txq->reaped;
258 KASSERT(level <= txq->entries, ("overfilled TX queue"));
259
260 /* Clear the fragment list. */
261 txq->n_pend_desc = 0;
262
263 /* Have we reached the block level? */
264 if (level < SFXGE_TXQ_BLOCK_LEVEL(txq->entries))
265 return;
266
267 /* Reap, and check again */
268 sfxge_tx_qreap(txq);
269 level = txq->added - txq->reaped;
270 if (level < SFXGE_TXQ_BLOCK_LEVEL(txq->entries))
271 return;
272
273 txq->blocked = 1;
274
275 /*
276 * Avoid a race with completion interrupt handling that could leave
277 * the queue blocked.
278 */
279 mb();
280 sfxge_tx_qreap(txq);
281 level = txq->added - txq->reaped;
282 if (level < SFXGE_TXQ_BLOCK_LEVEL(txq->entries)) {
283 mb();
284 txq->blocked = 0;
285 }
286}
287
288static int sfxge_tx_queue_mbuf(struct sfxge_txq *txq, struct mbuf *mbuf)
289{
290 bus_dmamap_t *used_map;
291 bus_dmamap_t map;
292 bus_dma_segment_t dma_seg[SFXGE_TX_MAPPING_MAX_SEG];
293 unsigned int id;
294 struct sfxge_tx_mapping *stmp;
295 efx_buffer_t *desc;
296 int n_dma_seg;
297 int rc;
298 int i;
299
300 KASSERT(!txq->blocked, ("txq->blocked"));
301
302 if (mbuf->m_pkthdr.csum_flags & CSUM_TSO)
303 prefetch_read_many(mbuf->m_data);
304
305 if (txq->init_state != SFXGE_TXQ_STARTED) {
306 rc = EINTR;
307 goto reject;
308 }
309
310 /* Load the packet for DMA. */
311 id = txq->added & txq->ptr_mask;
312 stmp = &txq->stmp[id];
313 rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag, stmp->map,
314 mbuf, dma_seg, &n_dma_seg, 0);
315 if (rc == EFBIG) {
316 /* Try again. */
317 struct mbuf *new_mbuf = m_collapse(mbuf, M_NOWAIT,
318 SFXGE_TX_MAPPING_MAX_SEG);
319 if (new_mbuf == NULL)
320 goto reject;
321 ++txq->collapses;
322 mbuf = new_mbuf;
323 rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag,
324 stmp->map, mbuf,
325 dma_seg, &n_dma_seg, 0);
326 }
327 if (rc != 0)
328 goto reject;
329
330 /* Make the packet visible to the hardware. */
331 bus_dmamap_sync(txq->packet_dma_tag, stmp->map, BUS_DMASYNC_PREWRITE);
332
333 used_map = &stmp->map;
334
335 if (mbuf->m_pkthdr.csum_flags & CSUM_TSO) {
336 rc = sfxge_tx_queue_tso(txq, mbuf, dma_seg, n_dma_seg);
337 if (rc < 0)
338 goto reject_mapped;
339 stmp = &txq->stmp[rc];
340 } else {
341 /* Add the mapping to the fragment list, and set flags
342 * for the buffer.
343 */
344 i = 0;
345 for (;;) {
346 desc = &txq->pend_desc[i];
347 desc->eb_addr = dma_seg[i].ds_addr;
348 desc->eb_size = dma_seg[i].ds_len;
349 if (i == n_dma_seg - 1) {
350 desc->eb_eop = 1;
351 break;
352 }
353 desc->eb_eop = 0;
354 i++;
355
356 stmp->flags = 0;
357 if (__predict_false(stmp ==
358 &txq->stmp[txq->ptr_mask]))
359 stmp = &txq->stmp[0];
360 else
361 stmp++;
362 }
363 txq->n_pend_desc = n_dma_seg;
364 }
365
366 /*
367 * If the mapping required more than one descriptor
368 * then we need to associate the DMA map with the last
369 * descriptor, not the first.
370 */
371 if (used_map != &stmp->map) {
372 map = stmp->map;
373 stmp->map = *used_map;
374 *used_map = map;
375 }
376
377 stmp->u.mbuf = mbuf;
378 stmp->flags = TX_BUF_UNMAP | TX_BUF_MBUF;
379
380 /* Post the fragment list. */
381 sfxge_tx_qlist_post(txq);
382
383 return (0);
384
385reject_mapped:
386 bus_dmamap_unload(txq->packet_dma_tag, *used_map);
387reject:
388 /* Drop the packet on the floor. */
389 m_freem(mbuf);
390 ++txq->drops;
391
392 return (rc);
393}
394
395#ifdef SFXGE_HAVE_MQ
396
397/*
398 * Drain the deferred packet list into the transmit queue.
399 */
400static void
401sfxge_tx_qdpl_drain(struct sfxge_txq *txq)
402{
403 struct sfxge_softc *sc;
404 struct sfxge_tx_dpl *stdp;
405 struct mbuf *mbuf, *next;
406 unsigned int count;
407 unsigned int non_tcp_count;
408 unsigned int pushed;
409 int rc;
410
411 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
412
413 sc = txq->sc;
414 stdp = &txq->dpl;
415 pushed = txq->added;
416
417 prefetch_read_many(sc->enp);
418 prefetch_read_many(txq->common);
419
420 mbuf = stdp->std_get;
421 count = stdp->std_get_count;
422 non_tcp_count = stdp->std_get_non_tcp_count;
423
424 if (count > stdp->std_get_hiwat)
425 stdp->std_get_hiwat = count;
426
427 while (count != 0) {
428 KASSERT(mbuf != NULL, ("mbuf == NULL"));
429
430 next = mbuf->m_nextpkt;
431 mbuf->m_nextpkt = NULL;
432
433 ETHER_BPF_MTAP(sc->ifnet, mbuf); /* packet capture */
434
435 if (next != NULL)
436 prefetch_read_many(next);
437
438 rc = sfxge_tx_queue_mbuf(txq, mbuf);
439 --count;
440 non_tcp_count -= sfxge_is_mbuf_non_tcp(mbuf);
441 mbuf = next;
442 if (rc != 0)
443 continue;
444
445 if (txq->blocked)
446 break;
447
448 /* Push the fragments to the hardware in batches. */
449 if (txq->added - pushed >= SFXGE_TX_BATCH) {
450 efx_tx_qpush(txq->common, txq->added);
451 pushed = txq->added;
452 }
453 }
454
455 if (count == 0) {
456 KASSERT(mbuf == NULL, ("mbuf != NULL"));
457 KASSERT(non_tcp_count == 0,
458 ("inconsistent TCP/non-TCP detection"));
459 stdp->std_get = NULL;
460 stdp->std_get_count = 0;
461 stdp->std_get_non_tcp_count = 0;
462 stdp->std_getp = &stdp->std_get;
463 } else {
464 stdp->std_get = mbuf;
465 stdp->std_get_count = count;
466 stdp->std_get_non_tcp_count = non_tcp_count;
467 }
468
469 if (txq->added != pushed)
470 efx_tx_qpush(txq->common, txq->added);
471
472 KASSERT(txq->blocked || stdp->std_get_count == 0,
473 ("queue unblocked but count is non-zero"));
474}
475
476#define SFXGE_TX_QDPL_PENDING(_txq) \
477 ((_txq)->dpl.std_put != 0)
478
479/*
480 * Service the deferred packet list.
481 *
482 * NOTE: drops the txq mutex!
483 */
484static void
485sfxge_tx_qdpl_service(struct sfxge_txq *txq)
486{
487 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
488
489 do {
490 if (SFXGE_TX_QDPL_PENDING(txq))
491 sfxge_tx_qdpl_swizzle(txq);
492
493 if (!txq->blocked)
494 sfxge_tx_qdpl_drain(txq);
495
496 SFXGE_TXQ_UNLOCK(txq);
497 } while (SFXGE_TX_QDPL_PENDING(txq) &&
498 SFXGE_TXQ_TRYLOCK(txq));
499}
500
501/*
502 * Put a packet on the deferred packet list.
503 *
504 * If we are called with the txq lock held, we put the packet on the "get
505 * list", otherwise we atomically push it on the "put list". The swizzle
506 * function takes care of ordering.
507 *
| 47
48#include <sys/types.h>
49#include <sys/mbuf.h>
50#include <sys/smp.h>
51#include <sys/socket.h>
52#include <sys/sysctl.h>
53#include <sys/syslog.h>
54
55#include <net/bpf.h>
56#include <net/ethernet.h>
57#include <net/if.h>
58#include <net/if_vlan_var.h>
59
60#include <netinet/in.h>
61#include <netinet/ip.h>
62#include <netinet/ip6.h>
63#include <netinet/tcp.h>
64
65#include "common/efx.h"
66
67#include "sfxge.h"
68#include "sfxge_tx.h"
69
70/* Set the block level to ensure there is space to generate a
71 * large number of descriptors for TSO. With minimum MSS and
72 * maximum mbuf length we might need more than a ring-ful of
73 * descriptors, but this should not happen in practice except
74 * due to deliberate attack. In that case we will truncate
75 * the output at a packet boundary.
76 */
77#define SFXGE_TSO_MAX_DESC \
78 (SFXGE_TSO_MAX_SEGS * 2 + SFXGE_TX_MAPPING_MAX_SEG - 1)
79#define SFXGE_TXQ_BLOCK_LEVEL(_entries) ((_entries) - SFXGE_TSO_MAX_DESC)
80
81#ifdef SFXGE_HAVE_MQ
82
83#define SFXGE_PARAM_TX_DPL_GET_MAX SFXGE_PARAM(tx_dpl_get_max)
84static int sfxge_tx_dpl_get_max = SFXGE_TX_DPL_GET_PKT_LIMIT_DEFAULT;
85TUNABLE_INT(SFXGE_PARAM_TX_DPL_GET_MAX, &sfxge_tx_dpl_get_max);
86SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_get_max, CTLFLAG_RDTUN,
87 &sfxge_tx_dpl_get_max, 0,
88 "Maximum number of any packets in deferred packet get-list");
89
90#define SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX \
91 SFXGE_PARAM(tx_dpl_get_non_tcp_max)
92static int sfxge_tx_dpl_get_non_tcp_max =
93 SFXGE_TX_DPL_GET_NON_TCP_PKT_LIMIT_DEFAULT;
94TUNABLE_INT(SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX, &sfxge_tx_dpl_get_non_tcp_max);
95SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_get_non_tcp_max, CTLFLAG_RDTUN,
96 &sfxge_tx_dpl_get_non_tcp_max, 0,
97 "Maximum number of non-TCP packets in deferred packet get-list");
98
99#define SFXGE_PARAM_TX_DPL_PUT_MAX SFXGE_PARAM(tx_dpl_put_max)
100static int sfxge_tx_dpl_put_max = SFXGE_TX_DPL_PUT_PKT_LIMIT_DEFAULT;
101TUNABLE_INT(SFXGE_PARAM_TX_DPL_PUT_MAX, &sfxge_tx_dpl_put_max);
102SYSCTL_INT(_hw_sfxge, OID_AUTO, tx_dpl_put_max, CTLFLAG_RDTUN,
103 &sfxge_tx_dpl_put_max, 0,
104 "Maximum number of any packets in deferred packet put-list");
105
106#endif
107
108
109/* Forward declarations. */
110static void sfxge_tx_qdpl_service(struct sfxge_txq *txq);
111static void sfxge_tx_qlist_post(struct sfxge_txq *txq);
112static void sfxge_tx_qunblock(struct sfxge_txq *txq);
113static int sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf,
114 const bus_dma_segment_t *dma_seg, int n_dma_seg);
115
116void
117sfxge_tx_qcomplete(struct sfxge_txq *txq, struct sfxge_evq *evq)
118{
119 unsigned int completed;
120
121 SFXGE_EVQ_LOCK_ASSERT_OWNED(evq);
122
123 completed = txq->completed;
124 while (completed != txq->pending) {
125 struct sfxge_tx_mapping *stmp;
126 unsigned int id;
127
128 id = completed++ & txq->ptr_mask;
129
130 stmp = &txq->stmp[id];
131 if (stmp->flags & TX_BUF_UNMAP) {
132 bus_dmamap_unload(txq->packet_dma_tag, stmp->map);
133 if (stmp->flags & TX_BUF_MBUF) {
134 struct mbuf *m = stmp->u.mbuf;
135 do
136 m = m_free(m);
137 while (m != NULL);
138 } else {
139 free(stmp->u.heap_buf, M_SFXGE);
140 }
141 stmp->flags = 0;
142 }
143 }
144 txq->completed = completed;
145
146 /* Check whether we need to unblock the queue. */
147 mb();
148 if (txq->blocked) {
149 unsigned int level;
150
151 level = txq->added - txq->completed;
152 if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries))
153 sfxge_tx_qunblock(txq);
154 }
155}
156
157#ifdef SFXGE_HAVE_MQ
158
159static unsigned int
160sfxge_is_mbuf_non_tcp(struct mbuf *mbuf)
161{
162 /* Absense of TCP checksum flags does not mean that it is non-TCP
163 * but it should be true if user wants to achieve high throughput.
164 */
165 return (!(mbuf->m_pkthdr.csum_flags & (CSUM_IP_TCP | CSUM_IP6_TCP)));
166}
167
168/*
169 * Reorder the put list and append it to the get list.
170 */
171static void
172sfxge_tx_qdpl_swizzle(struct sfxge_txq *txq)
173{
174 struct sfxge_tx_dpl *stdp;
175 struct mbuf *mbuf, *get_next, **get_tailp;
176 volatile uintptr_t *putp;
177 uintptr_t put;
178 unsigned int count;
179 unsigned int non_tcp_count;
180
181 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
182
183 stdp = &txq->dpl;
184
185 /* Acquire the put list. */
186 putp = &stdp->std_put;
187 put = atomic_readandclear_ptr(putp);
188 mbuf = (void *)put;
189
190 if (mbuf == NULL)
191 return;
192
193 /* Reverse the put list. */
194 get_tailp = &mbuf->m_nextpkt;
195 get_next = NULL;
196
197 count = 0;
198 non_tcp_count = 0;
199 do {
200 struct mbuf *put_next;
201
202 non_tcp_count += sfxge_is_mbuf_non_tcp(mbuf);
203 put_next = mbuf->m_nextpkt;
204 mbuf->m_nextpkt = get_next;
205 get_next = mbuf;
206 mbuf = put_next;
207
208 count++;
209 } while (mbuf != NULL);
210
211 /* Append the reversed put list to the get list. */
212 KASSERT(*get_tailp == NULL, ("*get_tailp != NULL"));
213 *stdp->std_getp = get_next;
214 stdp->std_getp = get_tailp;
215 stdp->std_get_count += count;
216 stdp->std_get_non_tcp_count += non_tcp_count;
217}
218
219#endif /* SFXGE_HAVE_MQ */
220
221static void
222sfxge_tx_qreap(struct sfxge_txq *txq)
223{
224 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
225
226 txq->reaped = txq->completed;
227}
228
229static void
230sfxge_tx_qlist_post(struct sfxge_txq *txq)
231{
232 unsigned int old_added;
233 unsigned int level;
234 int rc;
235
236 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
237
238 KASSERT(txq->n_pend_desc != 0, ("txq->n_pend_desc == 0"));
239 KASSERT(txq->n_pend_desc <= SFXGE_TSO_MAX_DESC,
240 ("txq->n_pend_desc too large"));
241 KASSERT(!txq->blocked, ("txq->blocked"));
242
243 old_added = txq->added;
244
245 /* Post the fragment list. */
246 rc = efx_tx_qpost(txq->common, txq->pend_desc, txq->n_pend_desc,
247 txq->reaped, &txq->added);
248 KASSERT(rc == 0, ("efx_tx_qpost() failed"));
249
250 /* If efx_tx_qpost() had to refragment, our information about
251 * buffers to free may be associated with the wrong
252 * descriptors.
253 */
254 KASSERT(txq->added - old_added == txq->n_pend_desc,
255 ("efx_tx_qpost() refragmented descriptors"));
256
257 level = txq->added - txq->reaped;
258 KASSERT(level <= txq->entries, ("overfilled TX queue"));
259
260 /* Clear the fragment list. */
261 txq->n_pend_desc = 0;
262
263 /* Have we reached the block level? */
264 if (level < SFXGE_TXQ_BLOCK_LEVEL(txq->entries))
265 return;
266
267 /* Reap, and check again */
268 sfxge_tx_qreap(txq);
269 level = txq->added - txq->reaped;
270 if (level < SFXGE_TXQ_BLOCK_LEVEL(txq->entries))
271 return;
272
273 txq->blocked = 1;
274
275 /*
276 * Avoid a race with completion interrupt handling that could leave
277 * the queue blocked.
278 */
279 mb();
280 sfxge_tx_qreap(txq);
281 level = txq->added - txq->reaped;
282 if (level < SFXGE_TXQ_BLOCK_LEVEL(txq->entries)) {
283 mb();
284 txq->blocked = 0;
285 }
286}
287
288static int sfxge_tx_queue_mbuf(struct sfxge_txq *txq, struct mbuf *mbuf)
289{
290 bus_dmamap_t *used_map;
291 bus_dmamap_t map;
292 bus_dma_segment_t dma_seg[SFXGE_TX_MAPPING_MAX_SEG];
293 unsigned int id;
294 struct sfxge_tx_mapping *stmp;
295 efx_buffer_t *desc;
296 int n_dma_seg;
297 int rc;
298 int i;
299
300 KASSERT(!txq->blocked, ("txq->blocked"));
301
302 if (mbuf->m_pkthdr.csum_flags & CSUM_TSO)
303 prefetch_read_many(mbuf->m_data);
304
305 if (txq->init_state != SFXGE_TXQ_STARTED) {
306 rc = EINTR;
307 goto reject;
308 }
309
310 /* Load the packet for DMA. */
311 id = txq->added & txq->ptr_mask;
312 stmp = &txq->stmp[id];
313 rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag, stmp->map,
314 mbuf, dma_seg, &n_dma_seg, 0);
315 if (rc == EFBIG) {
316 /* Try again. */
317 struct mbuf *new_mbuf = m_collapse(mbuf, M_NOWAIT,
318 SFXGE_TX_MAPPING_MAX_SEG);
319 if (new_mbuf == NULL)
320 goto reject;
321 ++txq->collapses;
322 mbuf = new_mbuf;
323 rc = bus_dmamap_load_mbuf_sg(txq->packet_dma_tag,
324 stmp->map, mbuf,
325 dma_seg, &n_dma_seg, 0);
326 }
327 if (rc != 0)
328 goto reject;
329
330 /* Make the packet visible to the hardware. */
331 bus_dmamap_sync(txq->packet_dma_tag, stmp->map, BUS_DMASYNC_PREWRITE);
332
333 used_map = &stmp->map;
334
335 if (mbuf->m_pkthdr.csum_flags & CSUM_TSO) {
336 rc = sfxge_tx_queue_tso(txq, mbuf, dma_seg, n_dma_seg);
337 if (rc < 0)
338 goto reject_mapped;
339 stmp = &txq->stmp[rc];
340 } else {
341 /* Add the mapping to the fragment list, and set flags
342 * for the buffer.
343 */
344 i = 0;
345 for (;;) {
346 desc = &txq->pend_desc[i];
347 desc->eb_addr = dma_seg[i].ds_addr;
348 desc->eb_size = dma_seg[i].ds_len;
349 if (i == n_dma_seg - 1) {
350 desc->eb_eop = 1;
351 break;
352 }
353 desc->eb_eop = 0;
354 i++;
355
356 stmp->flags = 0;
357 if (__predict_false(stmp ==
358 &txq->stmp[txq->ptr_mask]))
359 stmp = &txq->stmp[0];
360 else
361 stmp++;
362 }
363 txq->n_pend_desc = n_dma_seg;
364 }
365
366 /*
367 * If the mapping required more than one descriptor
368 * then we need to associate the DMA map with the last
369 * descriptor, not the first.
370 */
371 if (used_map != &stmp->map) {
372 map = stmp->map;
373 stmp->map = *used_map;
374 *used_map = map;
375 }
376
377 stmp->u.mbuf = mbuf;
378 stmp->flags = TX_BUF_UNMAP | TX_BUF_MBUF;
379
380 /* Post the fragment list. */
381 sfxge_tx_qlist_post(txq);
382
383 return (0);
384
385reject_mapped:
386 bus_dmamap_unload(txq->packet_dma_tag, *used_map);
387reject:
388 /* Drop the packet on the floor. */
389 m_freem(mbuf);
390 ++txq->drops;
391
392 return (rc);
393}
394
395#ifdef SFXGE_HAVE_MQ
396
397/*
398 * Drain the deferred packet list into the transmit queue.
399 */
400static void
401sfxge_tx_qdpl_drain(struct sfxge_txq *txq)
402{
403 struct sfxge_softc *sc;
404 struct sfxge_tx_dpl *stdp;
405 struct mbuf *mbuf, *next;
406 unsigned int count;
407 unsigned int non_tcp_count;
408 unsigned int pushed;
409 int rc;
410
411 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
412
413 sc = txq->sc;
414 stdp = &txq->dpl;
415 pushed = txq->added;
416
417 prefetch_read_many(sc->enp);
418 prefetch_read_many(txq->common);
419
420 mbuf = stdp->std_get;
421 count = stdp->std_get_count;
422 non_tcp_count = stdp->std_get_non_tcp_count;
423
424 if (count > stdp->std_get_hiwat)
425 stdp->std_get_hiwat = count;
426
427 while (count != 0) {
428 KASSERT(mbuf != NULL, ("mbuf == NULL"));
429
430 next = mbuf->m_nextpkt;
431 mbuf->m_nextpkt = NULL;
432
433 ETHER_BPF_MTAP(sc->ifnet, mbuf); /* packet capture */
434
435 if (next != NULL)
436 prefetch_read_many(next);
437
438 rc = sfxge_tx_queue_mbuf(txq, mbuf);
439 --count;
440 non_tcp_count -= sfxge_is_mbuf_non_tcp(mbuf);
441 mbuf = next;
442 if (rc != 0)
443 continue;
444
445 if (txq->blocked)
446 break;
447
448 /* Push the fragments to the hardware in batches. */
449 if (txq->added - pushed >= SFXGE_TX_BATCH) {
450 efx_tx_qpush(txq->common, txq->added);
451 pushed = txq->added;
452 }
453 }
454
455 if (count == 0) {
456 KASSERT(mbuf == NULL, ("mbuf != NULL"));
457 KASSERT(non_tcp_count == 0,
458 ("inconsistent TCP/non-TCP detection"));
459 stdp->std_get = NULL;
460 stdp->std_get_count = 0;
461 stdp->std_get_non_tcp_count = 0;
462 stdp->std_getp = &stdp->std_get;
463 } else {
464 stdp->std_get = mbuf;
465 stdp->std_get_count = count;
466 stdp->std_get_non_tcp_count = non_tcp_count;
467 }
468
469 if (txq->added != pushed)
470 efx_tx_qpush(txq->common, txq->added);
471
472 KASSERT(txq->blocked || stdp->std_get_count == 0,
473 ("queue unblocked but count is non-zero"));
474}
475
476#define SFXGE_TX_QDPL_PENDING(_txq) \
477 ((_txq)->dpl.std_put != 0)
478
479/*
480 * Service the deferred packet list.
481 *
482 * NOTE: drops the txq mutex!
483 */
484static void
485sfxge_tx_qdpl_service(struct sfxge_txq *txq)
486{
487 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
488
489 do {
490 if (SFXGE_TX_QDPL_PENDING(txq))
491 sfxge_tx_qdpl_swizzle(txq);
492
493 if (!txq->blocked)
494 sfxge_tx_qdpl_drain(txq);
495
496 SFXGE_TXQ_UNLOCK(txq);
497 } while (SFXGE_TX_QDPL_PENDING(txq) &&
498 SFXGE_TXQ_TRYLOCK(txq));
499}
500
501/*
502 * Put a packet on the deferred packet list.
503 *
504 * If we are called with the txq lock held, we put the packet on the "get
505 * list", otherwise we atomically push it on the "put list". The swizzle
506 * function takes care of ordering.
507 *
|
573 */
574int
575sfxge_tx_packet_add(struct sfxge_txq *txq, struct mbuf *m)
576{
577 int locked;
578 int rc;
579
580 if (!SFXGE_LINK_UP(txq->sc)) {
581 rc = ENETDOWN;
582 atomic_add_long(&txq->netdown_drops, 1);
583 goto fail;
584 }
585
586 /*
587 * Try to grab the txq lock. If we are able to get the lock,
588 * the packet will be appended to the "get list" of the deferred
589 * packet list. Otherwise, it will be pushed on the "put list".
590 */
591 locked = SFXGE_TXQ_TRYLOCK(txq);
592
593 if (sfxge_tx_qdpl_put(txq, m, locked) != 0) {
594 if (locked)
595 SFXGE_TXQ_UNLOCK(txq);
596 rc = ENOBUFS;
597 goto fail;
598 }
599
600 /*
601 * Try to grab the lock again.
602 *
603 * If we are able to get the lock, we need to process the deferred
604 * packet list. If we are not able to get the lock, another thread
605 * is processing the list.
606 */
607 if (!locked)
608 locked = SFXGE_TXQ_TRYLOCK(txq);
609
610 if (locked) {
611 /* Try to service the list. */
612 sfxge_tx_qdpl_service(txq);
613 /* Lock has been dropped. */
614 }
615
616 return (0);
617
618fail:
619 m_freem(m);
620 return (rc);
621}
622
623static void
624sfxge_tx_qdpl_flush(struct sfxge_txq *txq)
625{
626 struct sfxge_tx_dpl *stdp = &txq->dpl;
627 struct mbuf *mbuf, *next;
628
629 SFXGE_TXQ_LOCK(txq);
630
631 sfxge_tx_qdpl_swizzle(txq);
632 for (mbuf = stdp->std_get; mbuf != NULL; mbuf = next) {
633 next = mbuf->m_nextpkt;
634 m_freem(mbuf);
635 }
636 stdp->std_get = NULL;
637 stdp->std_get_count = 0;
638 stdp->std_get_non_tcp_count = 0;
639 stdp->std_getp = &stdp->std_get;
640
641 SFXGE_TXQ_UNLOCK(txq);
642}
643
644void
645sfxge_if_qflush(struct ifnet *ifp)
646{
647 struct sfxge_softc *sc;
648 int i;
649
650 sc = ifp->if_softc;
651
652 for (i = 0; i < sc->txq_count; i++)
653 sfxge_tx_qdpl_flush(sc->txq[i]);
654}
655
656/*
657 * TX start -- called by the stack.
658 */
659int
660sfxge_if_transmit(struct ifnet *ifp, struct mbuf *m)
661{
662 struct sfxge_softc *sc;
663 struct sfxge_txq *txq;
664 int rc;
665
666 sc = (struct sfxge_softc *)ifp->if_softc;
667
668 KASSERT(ifp->if_flags & IFF_UP, ("interface not up"));
669
670 /* Pick the desired transmit queue. */
671 if (m->m_pkthdr.csum_flags & (CSUM_DELAY_DATA | CSUM_TSO)) {
672 int index = 0;
673
674 /* check if flowid is set */
675 if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
676 uint32_t hash = m->m_pkthdr.flowid;
677
678 index = sc->rx_indir_table[hash % SFXGE_RX_SCALE_MAX];
679 }
680 txq = sc->txq[SFXGE_TXQ_IP_TCP_UDP_CKSUM + index];
681 } else if (m->m_pkthdr.csum_flags & CSUM_DELAY_IP) {
682 txq = sc->txq[SFXGE_TXQ_IP_CKSUM];
683 } else {
684 txq = sc->txq[SFXGE_TXQ_NON_CKSUM];
685 }
686
687 rc = sfxge_tx_packet_add(txq, m);
688
689 return (rc);
690}
691
692#else /* !SFXGE_HAVE_MQ */
693
694static void sfxge_if_start_locked(struct ifnet *ifp)
695{
696 struct sfxge_softc *sc = ifp->if_softc;
697 struct sfxge_txq *txq;
698 struct mbuf *mbuf;
699 unsigned int pushed[SFXGE_TXQ_NTYPES];
700 unsigned int q_index;
701
702 if ((ifp->if_drv_flags & (IFF_DRV_RUNNING|IFF_DRV_OACTIVE)) !=
703 IFF_DRV_RUNNING)
704 return;
705
706 if (!sc->port.link_up)
707 return;
708
709 for (q_index = 0; q_index < SFXGE_TXQ_NTYPES; q_index++) {
710 txq = sc->txq[q_index];
711 pushed[q_index] = txq->added;
712 }
713
714 while (!IFQ_DRV_IS_EMPTY(&ifp->if_snd)) {
715 IFQ_DRV_DEQUEUE(&ifp->if_snd, mbuf);
716 if (mbuf == NULL)
717 break;
718
719 ETHER_BPF_MTAP(ifp, mbuf); /* packet capture */
720
721 /* Pick the desired transmit queue. */
722 if (mbuf->m_pkthdr.csum_flags & (CSUM_DELAY_DATA | CSUM_TSO))
723 q_index = SFXGE_TXQ_IP_TCP_UDP_CKSUM;
724 else if (mbuf->m_pkthdr.csum_flags & CSUM_DELAY_IP)
725 q_index = SFXGE_TXQ_IP_CKSUM;
726 else
727 q_index = SFXGE_TXQ_NON_CKSUM;
728 txq = sc->txq[q_index];
729
730 if (sfxge_tx_queue_mbuf(txq, mbuf) != 0)
731 continue;
732
733 if (txq->blocked) {
734 ifp->if_drv_flags |= IFF_DRV_OACTIVE;
735 break;
736 }
737
738 /* Push the fragments to the hardware in batches. */
739 if (txq->added - pushed[q_index] >= SFXGE_TX_BATCH) {
740 efx_tx_qpush(txq->common, txq->added);
741 pushed[q_index] = txq->added;
742 }
743 }
744
745 for (q_index = 0; q_index < SFXGE_TXQ_NTYPES; q_index++) {
746 txq = sc->txq[q_index];
747 if (txq->added != pushed[q_index])
748 efx_tx_qpush(txq->common, txq->added);
749 }
750}
751
752void sfxge_if_start(struct ifnet *ifp)
753{
754 struct sfxge_softc *sc = ifp->if_softc;
755
756 SFXGE_TXQ_LOCK(sc->txq[0]);
757 sfxge_if_start_locked(ifp);
758 SFXGE_TXQ_UNLOCK(sc->txq[0]);
759}
760
761static void
762sfxge_tx_qdpl_service(struct sfxge_txq *txq)
763{
764 struct ifnet *ifp = txq->sc->ifnet;
765
766 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
767 ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
768 sfxge_if_start_locked(ifp);
769 SFXGE_TXQ_UNLOCK(txq);
770}
771
772#endif /* SFXGE_HAVE_MQ */
773
774/*
775 * Software "TSO". Not quite as good as doing it in hardware, but
776 * still faster than segmenting in the stack.
777 */
778
779struct sfxge_tso_state {
780 /* Output position */
781 unsigned out_len; /* Remaining length in current segment */
782 unsigned seqnum; /* Current sequence number */
783 unsigned packet_space; /* Remaining space in current packet */
784
785 /* Input position */
786 uint64_t dma_addr; /* DMA address of current position */
787 unsigned in_len; /* Remaining length in current mbuf */
788
789 const struct mbuf *mbuf; /* Input mbuf (head of chain) */
790 u_short protocol; /* Network protocol (after VLAN decap) */
791 ssize_t nh_off; /* Offset of network header */
792 ssize_t tcph_off; /* Offset of TCP header */
793 unsigned header_len; /* Number of bytes of header */
794 unsigned seg_size; /* TCP segment size */
795};
796
797static const struct ip *tso_iph(const struct sfxge_tso_state *tso)
798{
799 KASSERT(tso->protocol == htons(ETHERTYPE_IP),
800 ("tso_iph() in non-IPv4 state"));
801 return (const struct ip *)(tso->mbuf->m_data + tso->nh_off);
802}
803static __unused const struct ip6_hdr *tso_ip6h(const struct sfxge_tso_state *tso)
804{
805 KASSERT(tso->protocol == htons(ETHERTYPE_IPV6),
806 ("tso_ip6h() in non-IPv6 state"));
807 return (const struct ip6_hdr *)(tso->mbuf->m_data + tso->nh_off);
808}
809static const struct tcphdr *tso_tcph(const struct sfxge_tso_state *tso)
810{
811 return (const struct tcphdr *)(tso->mbuf->m_data + tso->tcph_off);
812}
813
814/* Size of preallocated TSO header buffers. Larger blocks must be
815 * allocated from the heap.
816 */
817#define TSOH_STD_SIZE 128
818
819/* At most half the descriptors in the queue at any time will refer to
820 * a TSO header buffer, since they must always be followed by a
821 * payload descriptor referring to an mbuf.
822 */
823#define TSOH_COUNT(_txq_entries) ((_txq_entries) / 2u)
824#define TSOH_PER_PAGE (PAGE_SIZE / TSOH_STD_SIZE)
825#define TSOH_PAGE_COUNT(_txq_entries) \
826 ((TSOH_COUNT(_txq_entries) + TSOH_PER_PAGE - 1) / TSOH_PER_PAGE)
827
828static int tso_init(struct sfxge_txq *txq)
829{
830 struct sfxge_softc *sc = txq->sc;
831 unsigned int tsoh_page_count = TSOH_PAGE_COUNT(sc->txq_entries);
832 int i, rc;
833
834 /* Allocate TSO header buffers */
835 txq->tsoh_buffer = malloc(tsoh_page_count * sizeof(txq->tsoh_buffer[0]),
836 M_SFXGE, M_WAITOK);
837
838 for (i = 0; i < tsoh_page_count; i++) {
839 rc = sfxge_dma_alloc(sc, PAGE_SIZE, &txq->tsoh_buffer[i]);
840 if (rc != 0)
841 goto fail;
842 }
843
844 return (0);
845
846fail:
847 while (i-- > 0)
848 sfxge_dma_free(&txq->tsoh_buffer[i]);
849 free(txq->tsoh_buffer, M_SFXGE);
850 txq->tsoh_buffer = NULL;
851 return (rc);
852}
853
854static void tso_fini(struct sfxge_txq *txq)
855{
856 int i;
857
858 if (txq->tsoh_buffer != NULL) {
859 for (i = 0; i < TSOH_PAGE_COUNT(txq->sc->txq_entries); i++)
860 sfxge_dma_free(&txq->tsoh_buffer[i]);
861 free(txq->tsoh_buffer, M_SFXGE);
862 }
863}
864
865static void tso_start(struct sfxge_tso_state *tso, struct mbuf *mbuf)
866{
867 struct ether_header *eh = mtod(mbuf, struct ether_header *);
868 const struct tcphdr *th;
869 struct tcphdr th_copy;
870
871 tso->mbuf = mbuf;
872
873 /* Find network protocol and header */
874 tso->protocol = eh->ether_type;
875 if (tso->protocol == htons(ETHERTYPE_VLAN)) {
876 struct ether_vlan_header *veh =
877 mtod(mbuf, struct ether_vlan_header *);
878 tso->protocol = veh->evl_proto;
879 tso->nh_off = sizeof(*veh);
880 } else {
881 tso->nh_off = sizeof(*eh);
882 }
883
884 /* Find TCP header */
885 if (tso->protocol == htons(ETHERTYPE_IP)) {
886 KASSERT(tso_iph(tso)->ip_p == IPPROTO_TCP,
887 ("TSO required on non-TCP packet"));
888 tso->tcph_off = tso->nh_off + 4 * tso_iph(tso)->ip_hl;
889 } else {
890 KASSERT(tso->protocol == htons(ETHERTYPE_IPV6),
891 ("TSO required on non-IP packet"));
892 KASSERT(tso_ip6h(tso)->ip6_nxt == IPPROTO_TCP,
893 ("TSO required on non-TCP packet"));
894 tso->tcph_off = tso->nh_off + sizeof(struct ip6_hdr);
895 }
896
897 KASSERT(mbuf->m_len >= tso->tcph_off,
898 ("network header is fragmented in mbuf"));
899 /* We need TCP header including flags (window is the next) */
900 if (mbuf->m_len < tso->tcph_off + offsetof(struct tcphdr, th_win)) {
901 m_copydata(tso->mbuf, tso->tcph_off, sizeof(th_copy),
902 (caddr_t)&th_copy);
903 th = &th_copy;
904 } else {
905 th = tso_tcph(tso);
906 }
907
908 tso->header_len = tso->tcph_off + 4 * th->th_off;
909 tso->seg_size = mbuf->m_pkthdr.tso_segsz;
910
911 tso->seqnum = ntohl(th->th_seq);
912
913 /* These flags must not be duplicated */
914 KASSERT(!(th->th_flags & (TH_URG | TH_SYN | TH_RST)),
915 ("incompatible TCP flag on TSO packet"));
916
917 tso->out_len = mbuf->m_pkthdr.len - tso->header_len;
918}
919
920/*
921 * tso_fill_packet_with_fragment - form descriptors for the current fragment
922 *
923 * Form descriptors for the current fragment, until we reach the end
924 * of fragment or end-of-packet. Return 0 on success, 1 if not enough
925 * space.
926 */
927static void tso_fill_packet_with_fragment(struct sfxge_txq *txq,
928 struct sfxge_tso_state *tso)
929{
930 efx_buffer_t *desc;
931 int n;
932
933 if (tso->in_len == 0 || tso->packet_space == 0)
934 return;
935
936 KASSERT(tso->in_len > 0, ("TSO input length went negative"));
937 KASSERT(tso->packet_space > 0, ("TSO packet space went negative"));
938
939 n = min(tso->in_len, tso->packet_space);
940
941 tso->packet_space -= n;
942 tso->out_len -= n;
943 tso->in_len -= n;
944
945 desc = &txq->pend_desc[txq->n_pend_desc++];
946 desc->eb_addr = tso->dma_addr;
947 desc->eb_size = n;
948 desc->eb_eop = tso->out_len == 0 || tso->packet_space == 0;
949
950 tso->dma_addr += n;
951}
952
953/* Callback from bus_dmamap_load() for long TSO headers. */
954static void tso_map_long_header(void *dma_addr_ret,
955 bus_dma_segment_t *segs, int nseg,
956 int error)
957{
958 *(uint64_t *)dma_addr_ret = ((__predict_true(error == 0) &&
959 __predict_true(nseg == 1)) ?
960 segs->ds_addr : 0);
961}
962
963/*
964 * tso_start_new_packet - generate a new header and prepare for the new packet
965 *
966 * Generate a new header and prepare for the new packet. Return 0 on
967 * success, or an error code if failed to alloc header.
968 */
969static int tso_start_new_packet(struct sfxge_txq *txq,
970 struct sfxge_tso_state *tso,
971 unsigned int id)
972{
973 struct sfxge_tx_mapping *stmp = &txq->stmp[id];
974 struct tcphdr *tsoh_th;
975 unsigned ip_length;
976 caddr_t header;
977 uint64_t dma_addr;
978 bus_dmamap_t map;
979 efx_buffer_t *desc;
980 int rc;
981
982 /* Allocate a DMA-mapped header buffer. */
983 if (__predict_true(tso->header_len <= TSOH_STD_SIZE)) {
984 unsigned int page_index = (id / 2) / TSOH_PER_PAGE;
985 unsigned int buf_index = (id / 2) % TSOH_PER_PAGE;
986
987 header = (txq->tsoh_buffer[page_index].esm_base +
988 buf_index * TSOH_STD_SIZE);
989 dma_addr = (txq->tsoh_buffer[page_index].esm_addr +
990 buf_index * TSOH_STD_SIZE);
991 map = txq->tsoh_buffer[page_index].esm_map;
992
993 stmp->flags = 0;
994 } else {
995 /* We cannot use bus_dmamem_alloc() as that may sleep */
996 header = malloc(tso->header_len, M_SFXGE, M_NOWAIT);
997 if (__predict_false(!header))
998 return (ENOMEM);
999 rc = bus_dmamap_load(txq->packet_dma_tag, stmp->map,
1000 header, tso->header_len,
1001 tso_map_long_header, &dma_addr,
1002 BUS_DMA_NOWAIT);
1003 if (__predict_false(dma_addr == 0)) {
1004 if (rc == 0) {
1005 /* Succeeded but got >1 segment */
1006 bus_dmamap_unload(txq->packet_dma_tag,
1007 stmp->map);
1008 rc = EINVAL;
1009 }
1010 free(header, M_SFXGE);
1011 return (rc);
1012 }
1013 map = stmp->map;
1014
1015 txq->tso_long_headers++;
1016 stmp->u.heap_buf = header;
1017 stmp->flags = TX_BUF_UNMAP;
1018 }
1019
1020 tsoh_th = (struct tcphdr *)(header + tso->tcph_off);
1021
1022 /* Copy and update the headers. */
1023 m_copydata(tso->mbuf, 0, tso->header_len, header);
1024
1025 tsoh_th->th_seq = htonl(tso->seqnum);
1026 tso->seqnum += tso->seg_size;
1027 if (tso->out_len > tso->seg_size) {
1028 /* This packet will not finish the TSO burst. */
1029 ip_length = tso->header_len - tso->nh_off + tso->seg_size;
1030 tsoh_th->th_flags &= ~(TH_FIN | TH_PUSH);
1031 } else {
1032 /* This packet will be the last in the TSO burst. */
1033 ip_length = tso->header_len - tso->nh_off + tso->out_len;
1034 }
1035
1036 if (tso->protocol == htons(ETHERTYPE_IP)) {
1037 struct ip *tsoh_iph = (struct ip *)(header + tso->nh_off);
1038 tsoh_iph->ip_len = htons(ip_length);
1039 /* XXX We should increment ip_id, but FreeBSD doesn't
1040 * currently allocate extra IDs for multiple segments.
1041 */
1042 } else {
1043 struct ip6_hdr *tsoh_iph =
1044 (struct ip6_hdr *)(header + tso->nh_off);
1045 tsoh_iph->ip6_plen = htons(ip_length - sizeof(*tsoh_iph));
1046 }
1047
1048 /* Make the header visible to the hardware. */
1049 bus_dmamap_sync(txq->packet_dma_tag, map, BUS_DMASYNC_PREWRITE);
1050
1051 tso->packet_space = tso->seg_size;
1052 txq->tso_packets++;
1053
1054 /* Form a descriptor for this header. */
1055 desc = &txq->pend_desc[txq->n_pend_desc++];
1056 desc->eb_addr = dma_addr;
1057 desc->eb_size = tso->header_len;
1058 desc->eb_eop = 0;
1059
1060 return (0);
1061}
1062
1063static int
1064sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf,
1065 const bus_dma_segment_t *dma_seg, int n_dma_seg)
1066{
1067 struct sfxge_tso_state tso;
1068 unsigned int id, next_id;
1069 unsigned skipped = 0;
1070
1071 tso_start(&tso, mbuf);
1072
1073 while (dma_seg->ds_len + skipped <= tso.header_len) {
1074 skipped += dma_seg->ds_len;
1075 --n_dma_seg;
1076 KASSERT(n_dma_seg, ("no payload found in TSO packet"));
1077 ++dma_seg;
1078 }
1079 tso.in_len = dma_seg->ds_len + (tso.header_len - skipped);
1080 tso.dma_addr = dma_seg->ds_addr + (tso.header_len - skipped);
1081
1082 id = txq->added & txq->ptr_mask;
1083 if (__predict_false(tso_start_new_packet(txq, &tso, id)))
1084 return (-1);
1085
1086 while (1) {
1087 id = (id + 1) & txq->ptr_mask;
1088 tso_fill_packet_with_fragment(txq, &tso);
1089
1090 /* Move onto the next fragment? */
1091 if (tso.in_len == 0) {
1092 --n_dma_seg;
1093 if (n_dma_seg == 0)
1094 break;
1095 ++dma_seg;
1096 tso.in_len = dma_seg->ds_len;
1097 tso.dma_addr = dma_seg->ds_addr;
1098 }
1099
1100 /* End of packet? */
1101 if (tso.packet_space == 0) {
1102 /* If the queue is now full due to tiny MSS,
1103 * or we can't create another header, discard
1104 * the remainder of the input mbuf but do not
1105 * roll back the work we have done.
1106 */
1107 if (txq->n_pend_desc >
1108 SFXGE_TSO_MAX_DESC - (1 + SFXGE_TX_MAPPING_MAX_SEG)) {
1109 txq->tso_pdrop_too_many++;
1110 break;
1111 }
1112 next_id = (id + 1) & txq->ptr_mask;
1113 if (__predict_false(tso_start_new_packet(txq, &tso,
1114 next_id))) {
1115 txq->tso_pdrop_no_rsrc++;
1116 break;
1117 }
1118 id = next_id;
1119 }
1120 }
1121
1122 txq->tso_bursts++;
1123 return (id);
1124}
1125
1126static void
1127sfxge_tx_qunblock(struct sfxge_txq *txq)
1128{
1129 struct sfxge_softc *sc;
1130 struct sfxge_evq *evq;
1131
1132 sc = txq->sc;
1133 evq = sc->evq[txq->evq_index];
1134
1135 SFXGE_EVQ_LOCK_ASSERT_OWNED(evq);
1136
1137 if (txq->init_state != SFXGE_TXQ_STARTED)
1138 return;
1139
1140 SFXGE_TXQ_LOCK(txq);
1141
1142 if (txq->blocked) {
1143 unsigned int level;
1144
1145 level = txq->added - txq->completed;
1146 if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries)) {
1147 /* reaped must be in sync with blocked */
1148 sfxge_tx_qreap(txq);
1149 txq->blocked = 0;
1150 }
1151 }
1152
1153 sfxge_tx_qdpl_service(txq);
1154 /* note: lock has been dropped */
1155}
1156
1157void
1158sfxge_tx_qflush_done(struct sfxge_txq *txq)
1159{
1160
1161 txq->flush_state = SFXGE_FLUSH_DONE;
1162}
1163
1164static void
1165sfxge_tx_qstop(struct sfxge_softc *sc, unsigned int index)
1166{
1167 struct sfxge_txq *txq;
1168 struct sfxge_evq *evq;
1169 unsigned int count;
1170
1171 txq = sc->txq[index];
1172 evq = sc->evq[txq->evq_index];
1173
1174 SFXGE_TXQ_LOCK(txq);
1175
1176 KASSERT(txq->init_state == SFXGE_TXQ_STARTED,
1177 ("txq->init_state != SFXGE_TXQ_STARTED"));
1178
1179 txq->init_state = SFXGE_TXQ_INITIALIZED;
1180 txq->flush_state = SFXGE_FLUSH_PENDING;
1181
1182 /* Flush the transmit queue. */
1183 efx_tx_qflush(txq->common);
1184
1185 SFXGE_TXQ_UNLOCK(txq);
1186
1187 count = 0;
1188 do {
1189 /* Spin for 100ms. */
1190 DELAY(100000);
1191
1192 if (txq->flush_state != SFXGE_FLUSH_PENDING)
1193 break;
1194 } while (++count < 20);
1195
1196 SFXGE_EVQ_LOCK(evq);
1197 SFXGE_TXQ_LOCK(txq);
1198
1199 KASSERT(txq->flush_state != SFXGE_FLUSH_FAILED,
1200 ("txq->flush_state == SFXGE_FLUSH_FAILED"));
1201
1202 txq->flush_state = SFXGE_FLUSH_DONE;
1203
1204 txq->blocked = 0;
1205 txq->pending = txq->added;
1206
1207 sfxge_tx_qcomplete(txq, evq);
1208 KASSERT(txq->completed == txq->added,
1209 ("txq->completed != txq->added"));
1210
1211 sfxge_tx_qreap(txq);
1212 KASSERT(txq->reaped == txq->completed,
1213 ("txq->reaped != txq->completed"));
1214
1215 txq->added = 0;
1216 txq->pending = 0;
1217 txq->completed = 0;
1218 txq->reaped = 0;
1219
1220 /* Destroy the common code transmit queue. */
1221 efx_tx_qdestroy(txq->common);
1222 txq->common = NULL;
1223
1224 efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id,
1225 EFX_TXQ_NBUFS(sc->txq_entries));
1226
1227 SFXGE_EVQ_UNLOCK(evq);
1228 SFXGE_TXQ_UNLOCK(txq);
1229}
1230
1231static int
1232sfxge_tx_qstart(struct sfxge_softc *sc, unsigned int index)
1233{
1234 struct sfxge_txq *txq;
1235 efsys_mem_t *esmp;
1236 uint16_t flags;
1237 struct sfxge_evq *evq;
1238 int rc;
1239
1240 txq = sc->txq[index];
1241 esmp = &txq->mem;
1242 evq = sc->evq[txq->evq_index];
1243
1244 KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED,
1245 ("txq->init_state != SFXGE_TXQ_INITIALIZED"));
1246 KASSERT(evq->init_state == SFXGE_EVQ_STARTED,
1247 ("evq->init_state != SFXGE_EVQ_STARTED"));
1248
1249 /* Program the buffer table. */
1250 if ((rc = efx_sram_buf_tbl_set(sc->enp, txq->buf_base_id, esmp,
1251 EFX_TXQ_NBUFS(sc->txq_entries))) != 0)
1252 return (rc);
1253
1254 /* Determine the kind of queue we are creating. */
1255 switch (txq->type) {
1256 case SFXGE_TXQ_NON_CKSUM:
1257 flags = 0;
1258 break;
1259 case SFXGE_TXQ_IP_CKSUM:
1260 flags = EFX_CKSUM_IPV4;
1261 break;
1262 case SFXGE_TXQ_IP_TCP_UDP_CKSUM:
1263 flags = EFX_CKSUM_IPV4 | EFX_CKSUM_TCPUDP;
1264 break;
1265 default:
1266 KASSERT(0, ("Impossible TX queue"));
1267 flags = 0;
1268 break;
1269 }
1270
1271 /* Create the common code transmit queue. */
1272 if ((rc = efx_tx_qcreate(sc->enp, index, txq->type, esmp,
1273 sc->txq_entries, txq->buf_base_id, flags, evq->common,
1274 &txq->common)) != 0)
1275 goto fail;
1276
1277 SFXGE_TXQ_LOCK(txq);
1278
1279 /* Enable the transmit queue. */
1280 efx_tx_qenable(txq->common);
1281
1282 txq->init_state = SFXGE_TXQ_STARTED;
1283
1284 SFXGE_TXQ_UNLOCK(txq);
1285
1286 return (0);
1287
1288fail:
1289 efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id,
1290 EFX_TXQ_NBUFS(sc->txq_entries));
1291 return (rc);
1292}
1293
1294void
1295sfxge_tx_stop(struct sfxge_softc *sc)
1296{
1297 int index;
1298
1299 index = sc->txq_count;
1300 while (--index >= 0)
1301 sfxge_tx_qstop(sc, index);
1302
1303 /* Tear down the transmit module */
1304 efx_tx_fini(sc->enp);
1305}
1306
1307int
1308sfxge_tx_start(struct sfxge_softc *sc)
1309{
1310 int index;
1311 int rc;
1312
1313 /* Initialize the common code transmit module. */
1314 if ((rc = efx_tx_init(sc->enp)) != 0)
1315 return (rc);
1316
1317 for (index = 0; index < sc->txq_count; index++) {
1318 if ((rc = sfxge_tx_qstart(sc, index)) != 0)
1319 goto fail;
1320 }
1321
1322 return (0);
1323
1324fail:
1325 while (--index >= 0)
1326 sfxge_tx_qstop(sc, index);
1327
1328 efx_tx_fini(sc->enp);
1329
1330 return (rc);
1331}
1332
1333/**
1334 * Destroy a transmit queue.
1335 */
1336static void
1337sfxge_tx_qfini(struct sfxge_softc *sc, unsigned int index)
1338{
1339 struct sfxge_txq *txq;
1340 unsigned int nmaps;
1341
1342 txq = sc->txq[index];
1343
1344 KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED,
1345 ("txq->init_state != SFXGE_TXQ_INITIALIZED"));
1346
1347 if (txq->type == SFXGE_TXQ_IP_TCP_UDP_CKSUM)
1348 tso_fini(txq);
1349
1350 /* Free the context arrays. */
1351 free(txq->pend_desc, M_SFXGE);
1352 nmaps = sc->txq_entries;
1353 while (nmaps-- != 0)
1354 bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map);
1355 free(txq->stmp, M_SFXGE);
1356
1357 /* Release DMA memory mapping. */
1358 sfxge_dma_free(&txq->mem);
1359
1360 sc->txq[index] = NULL;
1361
1362#ifdef SFXGE_HAVE_MQ
1363 SFXGE_TXQ_LOCK_DESTROY(txq);
1364#endif
1365
1366 free(txq, M_SFXGE);
1367}
1368
1369static int
1370sfxge_tx_qinit(struct sfxge_softc *sc, unsigned int txq_index,
1371 enum sfxge_txq_type type, unsigned int evq_index)
1372{
1373 char name[16];
1374 struct sysctl_oid *txq_node;
1375 struct sfxge_txq *txq;
1376 struct sfxge_evq *evq;
1377#ifdef SFXGE_HAVE_MQ
1378 struct sfxge_tx_dpl *stdp;
1379#endif
1380 efsys_mem_t *esmp;
1381 unsigned int nmaps;
1382 int rc;
1383
1384 txq = malloc(sizeof(struct sfxge_txq), M_SFXGE, M_ZERO | M_WAITOK);
1385 txq->sc = sc;
1386 txq->entries = sc->txq_entries;
1387 txq->ptr_mask = txq->entries - 1;
1388
1389 sc->txq[txq_index] = txq;
1390 esmp = &txq->mem;
1391
1392 evq = sc->evq[evq_index];
1393
1394 /* Allocate and zero DMA space for the descriptor ring. */
1395 if ((rc = sfxge_dma_alloc(sc, EFX_TXQ_SIZE(sc->txq_entries), esmp)) != 0)
1396 return (rc);
1397 (void)memset(esmp->esm_base, 0, EFX_TXQ_SIZE(sc->txq_entries));
1398
1399 /* Allocate buffer table entries. */
1400 sfxge_sram_buf_tbl_alloc(sc, EFX_TXQ_NBUFS(sc->txq_entries),
1401 &txq->buf_base_id);
1402
1403 /* Create a DMA tag for packet mappings. */
1404 if (bus_dma_tag_create(sc->parent_dma_tag, 1, 0x1000,
1405 MIN(0x3FFFFFFFFFFFUL, BUS_SPACE_MAXADDR), BUS_SPACE_MAXADDR, NULL,
1406 NULL, 0x11000, SFXGE_TX_MAPPING_MAX_SEG, 0x1000, 0, NULL, NULL,
1407 &txq->packet_dma_tag) != 0) {
1408 device_printf(sc->dev, "Couldn't allocate txq DMA tag\n");
1409 rc = ENOMEM;
1410 goto fail;
1411 }
1412
1413 /* Allocate pending descriptor array for batching writes. */
1414 txq->pend_desc = malloc(sizeof(efx_buffer_t) * sc->txq_entries,
1415 M_SFXGE, M_ZERO | M_WAITOK);
1416
1417 /* Allocate and initialise mbuf DMA mapping array. */
1418 txq->stmp = malloc(sizeof(struct sfxge_tx_mapping) * sc->txq_entries,
1419 M_SFXGE, M_ZERO | M_WAITOK);
1420 for (nmaps = 0; nmaps < sc->txq_entries; nmaps++) {
1421 rc = bus_dmamap_create(txq->packet_dma_tag, 0,
1422 &txq->stmp[nmaps].map);
1423 if (rc != 0)
1424 goto fail2;
1425 }
1426
1427 snprintf(name, sizeof(name), "%u", txq_index);
1428 txq_node = SYSCTL_ADD_NODE(
1429 device_get_sysctl_ctx(sc->dev),
1430 SYSCTL_CHILDREN(sc->txqs_node),
1431 OID_AUTO, name, CTLFLAG_RD, NULL, "");
1432 if (txq_node == NULL) {
1433 rc = ENOMEM;
1434 goto fail_txq_node;
1435 }
1436
1437 if (type == SFXGE_TXQ_IP_TCP_UDP_CKSUM &&
1438 (rc = tso_init(txq)) != 0)
1439 goto fail3;
1440
1441#ifdef SFXGE_HAVE_MQ
1442 if (sfxge_tx_dpl_get_max <= 0) {
1443 log(LOG_ERR, "%s=%d must be greater than 0",
1444 SFXGE_PARAM_TX_DPL_GET_MAX, sfxge_tx_dpl_get_max);
1445 rc = EINVAL;
1446 goto fail_tx_dpl_get_max;
1447 }
1448 if (sfxge_tx_dpl_get_non_tcp_max <= 0) {
1449 log(LOG_ERR, "%s=%d must be greater than 0",
1450 SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX,
1451 sfxge_tx_dpl_get_non_tcp_max);
1452 rc = EINVAL;
1453 goto fail_tx_dpl_get_max;
1454 }
1455 if (sfxge_tx_dpl_put_max < 0) {
1456 log(LOG_ERR, "%s=%d must be greater or equal to 0",
1457 SFXGE_PARAM_TX_DPL_PUT_MAX, sfxge_tx_dpl_put_max);
1458 rc = EINVAL;
1459 goto fail_tx_dpl_put_max;
1460 }
1461
1462 /* Initialize the deferred packet list. */
1463 stdp = &txq->dpl;
1464 stdp->std_put_max = sfxge_tx_dpl_put_max;
1465 stdp->std_get_max = sfxge_tx_dpl_get_max;
1466 stdp->std_get_non_tcp_max = sfxge_tx_dpl_get_non_tcp_max;
1467 stdp->std_getp = &stdp->std_get;
1468
1469 SFXGE_TXQ_LOCK_INIT(txq, device_get_nameunit(sc->dev), txq_index);
1470
1471 SYSCTL_ADD_UINT(device_get_sysctl_ctx(sc->dev),
1472 SYSCTL_CHILDREN(txq_node), OID_AUTO,
1473 "dpl_get_count", CTLFLAG_RD | CTLFLAG_STATS,
1474 &stdp->std_get_count, 0, "");
1475 SYSCTL_ADD_UINT(device_get_sysctl_ctx(sc->dev),
1476 SYSCTL_CHILDREN(txq_node), OID_AUTO,
1477 "dpl_get_non_tcp_count", CTLFLAG_RD | CTLFLAG_STATS,
1478 &stdp->std_get_non_tcp_count, 0, "");
1479 SYSCTL_ADD_UINT(device_get_sysctl_ctx(sc->dev),
1480 SYSCTL_CHILDREN(txq_node), OID_AUTO,
1481 "dpl_get_hiwat", CTLFLAG_RD | CTLFLAG_STATS,
1482 &stdp->std_get_hiwat, 0, "");
1483#endif
1484
1485 txq->type = type;
1486 txq->evq_index = evq_index;
1487 txq->txq_index = txq_index;
1488 txq->init_state = SFXGE_TXQ_INITIALIZED;
1489
1490 return (0);
1491
1492fail_tx_dpl_put_max:
1493fail_tx_dpl_get_max:
1494fail3:
1495fail_txq_node:
1496 free(txq->pend_desc, M_SFXGE);
1497fail2:
1498 while (nmaps-- != 0)
1499 bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map);
1500 free(txq->stmp, M_SFXGE);
1501 bus_dma_tag_destroy(txq->packet_dma_tag);
1502
1503fail:
1504 sfxge_dma_free(esmp);
1505
1506 return (rc);
1507}
1508
1509static const struct {
1510 const char *name;
1511 size_t offset;
1512} sfxge_tx_stats[] = {
1513#define SFXGE_TX_STAT(name, member) \
1514 { #name, offsetof(struct sfxge_txq, member) }
1515 SFXGE_TX_STAT(tso_bursts, tso_bursts),
1516 SFXGE_TX_STAT(tso_packets, tso_packets),
1517 SFXGE_TX_STAT(tso_long_headers, tso_long_headers),
1518 SFXGE_TX_STAT(tso_pdrop_too_many, tso_pdrop_too_many),
1519 SFXGE_TX_STAT(tso_pdrop_no_rsrc, tso_pdrop_no_rsrc),
1520 SFXGE_TX_STAT(tx_collapses, collapses),
1521 SFXGE_TX_STAT(tx_drops, drops),
1522 SFXGE_TX_STAT(tx_get_overflow, get_overflow),
1523 SFXGE_TX_STAT(tx_get_non_tcp_overflow, get_non_tcp_overflow),
1524 SFXGE_TX_STAT(tx_put_overflow, put_overflow),
1525 SFXGE_TX_STAT(tx_netdown_drops, netdown_drops),
1526};
1527
1528static int
1529sfxge_tx_stat_handler(SYSCTL_HANDLER_ARGS)
1530{
1531 struct sfxge_softc *sc = arg1;
1532 unsigned int id = arg2;
1533 unsigned long sum;
1534 unsigned int index;
1535
1536 /* Sum across all TX queues */
1537 sum = 0;
1538 for (index = 0; index < sc->txq_count; index++)
1539 sum += *(unsigned long *)((caddr_t)sc->txq[index] +
1540 sfxge_tx_stats[id].offset);
1541
1542 return (SYSCTL_OUT(req, &sum, sizeof(sum)));
1543}
1544
1545static void
1546sfxge_tx_stat_init(struct sfxge_softc *sc)
1547{
1548 struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev);
1549 struct sysctl_oid_list *stat_list;
1550 unsigned int id;
1551
1552 stat_list = SYSCTL_CHILDREN(sc->stats_node);
1553
1554 for (id = 0; id < nitems(sfxge_tx_stats); id++) {
1555 SYSCTL_ADD_PROC(
1556 ctx, stat_list,
1557 OID_AUTO, sfxge_tx_stats[id].name,
1558 CTLTYPE_ULONG|CTLFLAG_RD,
1559 sc, id, sfxge_tx_stat_handler, "LU",
1560 "");
1561 }
1562}
1563
1564void
1565sfxge_tx_fini(struct sfxge_softc *sc)
1566{
1567 int index;
1568
1569 index = sc->txq_count;
1570 while (--index >= 0)
1571 sfxge_tx_qfini(sc, index);
1572
1573 sc->txq_count = 0;
1574}
1575
1576
1577int
1578sfxge_tx_init(struct sfxge_softc *sc)
1579{
1580 struct sfxge_intr *intr;
1581 int index;
1582 int rc;
1583
1584 intr = &sc->intr;
1585
1586 KASSERT(intr->state == SFXGE_INTR_INITIALIZED,
1587 ("intr->state != SFXGE_INTR_INITIALIZED"));
1588
1589#ifdef SFXGE_HAVE_MQ
1590 sc->txq_count = SFXGE_TXQ_NTYPES - 1 + sc->intr.n_alloc;
1591#else
1592 sc->txq_count = SFXGE_TXQ_NTYPES;
1593#endif
1594
1595 sc->txqs_node = SYSCTL_ADD_NODE(
1596 device_get_sysctl_ctx(sc->dev),
1597 SYSCTL_CHILDREN(device_get_sysctl_tree(sc->dev)),
1598 OID_AUTO, "txq", CTLFLAG_RD, NULL, "Tx queues");
1599 if (sc->txqs_node == NULL) {
1600 rc = ENOMEM;
1601 goto fail_txq_node;
1602 }
1603
1604 /* Initialize the transmit queues */
1605 if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NON_CKSUM,
1606 SFXGE_TXQ_NON_CKSUM, 0)) != 0)
1607 goto fail;
1608
1609 if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_IP_CKSUM,
1610 SFXGE_TXQ_IP_CKSUM, 0)) != 0)
1611 goto fail2;
1612
1613 for (index = 0;
1614 index < sc->txq_count - SFXGE_TXQ_NTYPES + 1;
1615 index++) {
1616 if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NTYPES - 1 + index,
1617 SFXGE_TXQ_IP_TCP_UDP_CKSUM, index)) != 0)
1618 goto fail3;
1619 }
1620
1621 sfxge_tx_stat_init(sc);
1622
1623 return (0);
1624
1625fail3:
1626 while (--index >= 0)
1627 sfxge_tx_qfini(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index);
1628
1629 sfxge_tx_qfini(sc, SFXGE_TXQ_IP_CKSUM);
1630
1631fail2:
1632 sfxge_tx_qfini(sc, SFXGE_TXQ_NON_CKSUM);
1633
1634fail:
1635fail_txq_node:
1636 sc->txq_count = 0;
1637 return (rc);
1638}
| 573 */
574int
575sfxge_tx_packet_add(struct sfxge_txq *txq, struct mbuf *m)
576{
577 int locked;
578 int rc;
579
580 if (!SFXGE_LINK_UP(txq->sc)) {
581 rc = ENETDOWN;
582 atomic_add_long(&txq->netdown_drops, 1);
583 goto fail;
584 }
585
586 /*
587 * Try to grab the txq lock. If we are able to get the lock,
588 * the packet will be appended to the "get list" of the deferred
589 * packet list. Otherwise, it will be pushed on the "put list".
590 */
591 locked = SFXGE_TXQ_TRYLOCK(txq);
592
593 if (sfxge_tx_qdpl_put(txq, m, locked) != 0) {
594 if (locked)
595 SFXGE_TXQ_UNLOCK(txq);
596 rc = ENOBUFS;
597 goto fail;
598 }
599
600 /*
601 * Try to grab the lock again.
602 *
603 * If we are able to get the lock, we need to process the deferred
604 * packet list. If we are not able to get the lock, another thread
605 * is processing the list.
606 */
607 if (!locked)
608 locked = SFXGE_TXQ_TRYLOCK(txq);
609
610 if (locked) {
611 /* Try to service the list. */
612 sfxge_tx_qdpl_service(txq);
613 /* Lock has been dropped. */
614 }
615
616 return (0);
617
618fail:
619 m_freem(m);
620 return (rc);
621}
622
623static void
624sfxge_tx_qdpl_flush(struct sfxge_txq *txq)
625{
626 struct sfxge_tx_dpl *stdp = &txq->dpl;
627 struct mbuf *mbuf, *next;
628
629 SFXGE_TXQ_LOCK(txq);
630
631 sfxge_tx_qdpl_swizzle(txq);
632 for (mbuf = stdp->std_get; mbuf != NULL; mbuf = next) {
633 next = mbuf->m_nextpkt;
634 m_freem(mbuf);
635 }
636 stdp->std_get = NULL;
637 stdp->std_get_count = 0;
638 stdp->std_get_non_tcp_count = 0;
639 stdp->std_getp = &stdp->std_get;
640
641 SFXGE_TXQ_UNLOCK(txq);
642}
643
644void
645sfxge_if_qflush(struct ifnet *ifp)
646{
647 struct sfxge_softc *sc;
648 int i;
649
650 sc = ifp->if_softc;
651
652 for (i = 0; i < sc->txq_count; i++)
653 sfxge_tx_qdpl_flush(sc->txq[i]);
654}
655
656/*
657 * TX start -- called by the stack.
658 */
659int
660sfxge_if_transmit(struct ifnet *ifp, struct mbuf *m)
661{
662 struct sfxge_softc *sc;
663 struct sfxge_txq *txq;
664 int rc;
665
666 sc = (struct sfxge_softc *)ifp->if_softc;
667
668 KASSERT(ifp->if_flags & IFF_UP, ("interface not up"));
669
670 /* Pick the desired transmit queue. */
671 if (m->m_pkthdr.csum_flags & (CSUM_DELAY_DATA | CSUM_TSO)) {
672 int index = 0;
673
674 /* check if flowid is set */
675 if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) {
676 uint32_t hash = m->m_pkthdr.flowid;
677
678 index = sc->rx_indir_table[hash % SFXGE_RX_SCALE_MAX];
679 }
680 txq = sc->txq[SFXGE_TXQ_IP_TCP_UDP_CKSUM + index];
681 } else if (m->m_pkthdr.csum_flags & CSUM_DELAY_IP) {
682 txq = sc->txq[SFXGE_TXQ_IP_CKSUM];
683 } else {
684 txq = sc->txq[SFXGE_TXQ_NON_CKSUM];
685 }
686
687 rc = sfxge_tx_packet_add(txq, m);
688
689 return (rc);
690}
691
692#else /* !SFXGE_HAVE_MQ */
693
694static void sfxge_if_start_locked(struct ifnet *ifp)
695{
696 struct sfxge_softc *sc = ifp->if_softc;
697 struct sfxge_txq *txq;
698 struct mbuf *mbuf;
699 unsigned int pushed[SFXGE_TXQ_NTYPES];
700 unsigned int q_index;
701
702 if ((ifp->if_drv_flags & (IFF_DRV_RUNNING|IFF_DRV_OACTIVE)) !=
703 IFF_DRV_RUNNING)
704 return;
705
706 if (!sc->port.link_up)
707 return;
708
709 for (q_index = 0; q_index < SFXGE_TXQ_NTYPES; q_index++) {
710 txq = sc->txq[q_index];
711 pushed[q_index] = txq->added;
712 }
713
714 while (!IFQ_DRV_IS_EMPTY(&ifp->if_snd)) {
715 IFQ_DRV_DEQUEUE(&ifp->if_snd, mbuf);
716 if (mbuf == NULL)
717 break;
718
719 ETHER_BPF_MTAP(ifp, mbuf); /* packet capture */
720
721 /* Pick the desired transmit queue. */
722 if (mbuf->m_pkthdr.csum_flags & (CSUM_DELAY_DATA | CSUM_TSO))
723 q_index = SFXGE_TXQ_IP_TCP_UDP_CKSUM;
724 else if (mbuf->m_pkthdr.csum_flags & CSUM_DELAY_IP)
725 q_index = SFXGE_TXQ_IP_CKSUM;
726 else
727 q_index = SFXGE_TXQ_NON_CKSUM;
728 txq = sc->txq[q_index];
729
730 if (sfxge_tx_queue_mbuf(txq, mbuf) != 0)
731 continue;
732
733 if (txq->blocked) {
734 ifp->if_drv_flags |= IFF_DRV_OACTIVE;
735 break;
736 }
737
738 /* Push the fragments to the hardware in batches. */
739 if (txq->added - pushed[q_index] >= SFXGE_TX_BATCH) {
740 efx_tx_qpush(txq->common, txq->added);
741 pushed[q_index] = txq->added;
742 }
743 }
744
745 for (q_index = 0; q_index < SFXGE_TXQ_NTYPES; q_index++) {
746 txq = sc->txq[q_index];
747 if (txq->added != pushed[q_index])
748 efx_tx_qpush(txq->common, txq->added);
749 }
750}
751
752void sfxge_if_start(struct ifnet *ifp)
753{
754 struct sfxge_softc *sc = ifp->if_softc;
755
756 SFXGE_TXQ_LOCK(sc->txq[0]);
757 sfxge_if_start_locked(ifp);
758 SFXGE_TXQ_UNLOCK(sc->txq[0]);
759}
760
761static void
762sfxge_tx_qdpl_service(struct sfxge_txq *txq)
763{
764 struct ifnet *ifp = txq->sc->ifnet;
765
766 SFXGE_TXQ_LOCK_ASSERT_OWNED(txq);
767 ifp->if_drv_flags &= ~IFF_DRV_OACTIVE;
768 sfxge_if_start_locked(ifp);
769 SFXGE_TXQ_UNLOCK(txq);
770}
771
772#endif /* SFXGE_HAVE_MQ */
773
774/*
775 * Software "TSO". Not quite as good as doing it in hardware, but
776 * still faster than segmenting in the stack.
777 */
778
779struct sfxge_tso_state {
780 /* Output position */
781 unsigned out_len; /* Remaining length in current segment */
782 unsigned seqnum; /* Current sequence number */
783 unsigned packet_space; /* Remaining space in current packet */
784
785 /* Input position */
786 uint64_t dma_addr; /* DMA address of current position */
787 unsigned in_len; /* Remaining length in current mbuf */
788
789 const struct mbuf *mbuf; /* Input mbuf (head of chain) */
790 u_short protocol; /* Network protocol (after VLAN decap) */
791 ssize_t nh_off; /* Offset of network header */
792 ssize_t tcph_off; /* Offset of TCP header */
793 unsigned header_len; /* Number of bytes of header */
794 unsigned seg_size; /* TCP segment size */
795};
796
797static const struct ip *tso_iph(const struct sfxge_tso_state *tso)
798{
799 KASSERT(tso->protocol == htons(ETHERTYPE_IP),
800 ("tso_iph() in non-IPv4 state"));
801 return (const struct ip *)(tso->mbuf->m_data + tso->nh_off);
802}
803static __unused const struct ip6_hdr *tso_ip6h(const struct sfxge_tso_state *tso)
804{
805 KASSERT(tso->protocol == htons(ETHERTYPE_IPV6),
806 ("tso_ip6h() in non-IPv6 state"));
807 return (const struct ip6_hdr *)(tso->mbuf->m_data + tso->nh_off);
808}
809static const struct tcphdr *tso_tcph(const struct sfxge_tso_state *tso)
810{
811 return (const struct tcphdr *)(tso->mbuf->m_data + tso->tcph_off);
812}
813
814/* Size of preallocated TSO header buffers. Larger blocks must be
815 * allocated from the heap.
816 */
817#define TSOH_STD_SIZE 128
818
819/* At most half the descriptors in the queue at any time will refer to
820 * a TSO header buffer, since they must always be followed by a
821 * payload descriptor referring to an mbuf.
822 */
823#define TSOH_COUNT(_txq_entries) ((_txq_entries) / 2u)
824#define TSOH_PER_PAGE (PAGE_SIZE / TSOH_STD_SIZE)
825#define TSOH_PAGE_COUNT(_txq_entries) \
826 ((TSOH_COUNT(_txq_entries) + TSOH_PER_PAGE - 1) / TSOH_PER_PAGE)
827
828static int tso_init(struct sfxge_txq *txq)
829{
830 struct sfxge_softc *sc = txq->sc;
831 unsigned int tsoh_page_count = TSOH_PAGE_COUNT(sc->txq_entries);
832 int i, rc;
833
834 /* Allocate TSO header buffers */
835 txq->tsoh_buffer = malloc(tsoh_page_count * sizeof(txq->tsoh_buffer[0]),
836 M_SFXGE, M_WAITOK);
837
838 for (i = 0; i < tsoh_page_count; i++) {
839 rc = sfxge_dma_alloc(sc, PAGE_SIZE, &txq->tsoh_buffer[i]);
840 if (rc != 0)
841 goto fail;
842 }
843
844 return (0);
845
846fail:
847 while (i-- > 0)
848 sfxge_dma_free(&txq->tsoh_buffer[i]);
849 free(txq->tsoh_buffer, M_SFXGE);
850 txq->tsoh_buffer = NULL;
851 return (rc);
852}
853
854static void tso_fini(struct sfxge_txq *txq)
855{
856 int i;
857
858 if (txq->tsoh_buffer != NULL) {
859 for (i = 0; i < TSOH_PAGE_COUNT(txq->sc->txq_entries); i++)
860 sfxge_dma_free(&txq->tsoh_buffer[i]);
861 free(txq->tsoh_buffer, M_SFXGE);
862 }
863}
864
865static void tso_start(struct sfxge_tso_state *tso, struct mbuf *mbuf)
866{
867 struct ether_header *eh = mtod(mbuf, struct ether_header *);
868 const struct tcphdr *th;
869 struct tcphdr th_copy;
870
871 tso->mbuf = mbuf;
872
873 /* Find network protocol and header */
874 tso->protocol = eh->ether_type;
875 if (tso->protocol == htons(ETHERTYPE_VLAN)) {
876 struct ether_vlan_header *veh =
877 mtod(mbuf, struct ether_vlan_header *);
878 tso->protocol = veh->evl_proto;
879 tso->nh_off = sizeof(*veh);
880 } else {
881 tso->nh_off = sizeof(*eh);
882 }
883
884 /* Find TCP header */
885 if (tso->protocol == htons(ETHERTYPE_IP)) {
886 KASSERT(tso_iph(tso)->ip_p == IPPROTO_TCP,
887 ("TSO required on non-TCP packet"));
888 tso->tcph_off = tso->nh_off + 4 * tso_iph(tso)->ip_hl;
889 } else {
890 KASSERT(tso->protocol == htons(ETHERTYPE_IPV6),
891 ("TSO required on non-IP packet"));
892 KASSERT(tso_ip6h(tso)->ip6_nxt == IPPROTO_TCP,
893 ("TSO required on non-TCP packet"));
894 tso->tcph_off = tso->nh_off + sizeof(struct ip6_hdr);
895 }
896
897 KASSERT(mbuf->m_len >= tso->tcph_off,
898 ("network header is fragmented in mbuf"));
899 /* We need TCP header including flags (window is the next) */
900 if (mbuf->m_len < tso->tcph_off + offsetof(struct tcphdr, th_win)) {
901 m_copydata(tso->mbuf, tso->tcph_off, sizeof(th_copy),
902 (caddr_t)&th_copy);
903 th = &th_copy;
904 } else {
905 th = tso_tcph(tso);
906 }
907
908 tso->header_len = tso->tcph_off + 4 * th->th_off;
909 tso->seg_size = mbuf->m_pkthdr.tso_segsz;
910
911 tso->seqnum = ntohl(th->th_seq);
912
913 /* These flags must not be duplicated */
914 KASSERT(!(th->th_flags & (TH_URG | TH_SYN | TH_RST)),
915 ("incompatible TCP flag on TSO packet"));
916
917 tso->out_len = mbuf->m_pkthdr.len - tso->header_len;
918}
919
920/*
921 * tso_fill_packet_with_fragment - form descriptors for the current fragment
922 *
923 * Form descriptors for the current fragment, until we reach the end
924 * of fragment or end-of-packet. Return 0 on success, 1 if not enough
925 * space.
926 */
927static void tso_fill_packet_with_fragment(struct sfxge_txq *txq,
928 struct sfxge_tso_state *tso)
929{
930 efx_buffer_t *desc;
931 int n;
932
933 if (tso->in_len == 0 || tso->packet_space == 0)
934 return;
935
936 KASSERT(tso->in_len > 0, ("TSO input length went negative"));
937 KASSERT(tso->packet_space > 0, ("TSO packet space went negative"));
938
939 n = min(tso->in_len, tso->packet_space);
940
941 tso->packet_space -= n;
942 tso->out_len -= n;
943 tso->in_len -= n;
944
945 desc = &txq->pend_desc[txq->n_pend_desc++];
946 desc->eb_addr = tso->dma_addr;
947 desc->eb_size = n;
948 desc->eb_eop = tso->out_len == 0 || tso->packet_space == 0;
949
950 tso->dma_addr += n;
951}
952
953/* Callback from bus_dmamap_load() for long TSO headers. */
954static void tso_map_long_header(void *dma_addr_ret,
955 bus_dma_segment_t *segs, int nseg,
956 int error)
957{
958 *(uint64_t *)dma_addr_ret = ((__predict_true(error == 0) &&
959 __predict_true(nseg == 1)) ?
960 segs->ds_addr : 0);
961}
962
963/*
964 * tso_start_new_packet - generate a new header and prepare for the new packet
965 *
966 * Generate a new header and prepare for the new packet. Return 0 on
967 * success, or an error code if failed to alloc header.
968 */
969static int tso_start_new_packet(struct sfxge_txq *txq,
970 struct sfxge_tso_state *tso,
971 unsigned int id)
972{
973 struct sfxge_tx_mapping *stmp = &txq->stmp[id];
974 struct tcphdr *tsoh_th;
975 unsigned ip_length;
976 caddr_t header;
977 uint64_t dma_addr;
978 bus_dmamap_t map;
979 efx_buffer_t *desc;
980 int rc;
981
982 /* Allocate a DMA-mapped header buffer. */
983 if (__predict_true(tso->header_len <= TSOH_STD_SIZE)) {
984 unsigned int page_index = (id / 2) / TSOH_PER_PAGE;
985 unsigned int buf_index = (id / 2) % TSOH_PER_PAGE;
986
987 header = (txq->tsoh_buffer[page_index].esm_base +
988 buf_index * TSOH_STD_SIZE);
989 dma_addr = (txq->tsoh_buffer[page_index].esm_addr +
990 buf_index * TSOH_STD_SIZE);
991 map = txq->tsoh_buffer[page_index].esm_map;
992
993 stmp->flags = 0;
994 } else {
995 /* We cannot use bus_dmamem_alloc() as that may sleep */
996 header = malloc(tso->header_len, M_SFXGE, M_NOWAIT);
997 if (__predict_false(!header))
998 return (ENOMEM);
999 rc = bus_dmamap_load(txq->packet_dma_tag, stmp->map,
1000 header, tso->header_len,
1001 tso_map_long_header, &dma_addr,
1002 BUS_DMA_NOWAIT);
1003 if (__predict_false(dma_addr == 0)) {
1004 if (rc == 0) {
1005 /* Succeeded but got >1 segment */
1006 bus_dmamap_unload(txq->packet_dma_tag,
1007 stmp->map);
1008 rc = EINVAL;
1009 }
1010 free(header, M_SFXGE);
1011 return (rc);
1012 }
1013 map = stmp->map;
1014
1015 txq->tso_long_headers++;
1016 stmp->u.heap_buf = header;
1017 stmp->flags = TX_BUF_UNMAP;
1018 }
1019
1020 tsoh_th = (struct tcphdr *)(header + tso->tcph_off);
1021
1022 /* Copy and update the headers. */
1023 m_copydata(tso->mbuf, 0, tso->header_len, header);
1024
1025 tsoh_th->th_seq = htonl(tso->seqnum);
1026 tso->seqnum += tso->seg_size;
1027 if (tso->out_len > tso->seg_size) {
1028 /* This packet will not finish the TSO burst. */
1029 ip_length = tso->header_len - tso->nh_off + tso->seg_size;
1030 tsoh_th->th_flags &= ~(TH_FIN | TH_PUSH);
1031 } else {
1032 /* This packet will be the last in the TSO burst. */
1033 ip_length = tso->header_len - tso->nh_off + tso->out_len;
1034 }
1035
1036 if (tso->protocol == htons(ETHERTYPE_IP)) {
1037 struct ip *tsoh_iph = (struct ip *)(header + tso->nh_off);
1038 tsoh_iph->ip_len = htons(ip_length);
1039 /* XXX We should increment ip_id, but FreeBSD doesn't
1040 * currently allocate extra IDs for multiple segments.
1041 */
1042 } else {
1043 struct ip6_hdr *tsoh_iph =
1044 (struct ip6_hdr *)(header + tso->nh_off);
1045 tsoh_iph->ip6_plen = htons(ip_length - sizeof(*tsoh_iph));
1046 }
1047
1048 /* Make the header visible to the hardware. */
1049 bus_dmamap_sync(txq->packet_dma_tag, map, BUS_DMASYNC_PREWRITE);
1050
1051 tso->packet_space = tso->seg_size;
1052 txq->tso_packets++;
1053
1054 /* Form a descriptor for this header. */
1055 desc = &txq->pend_desc[txq->n_pend_desc++];
1056 desc->eb_addr = dma_addr;
1057 desc->eb_size = tso->header_len;
1058 desc->eb_eop = 0;
1059
1060 return (0);
1061}
1062
1063static int
1064sfxge_tx_queue_tso(struct sfxge_txq *txq, struct mbuf *mbuf,
1065 const bus_dma_segment_t *dma_seg, int n_dma_seg)
1066{
1067 struct sfxge_tso_state tso;
1068 unsigned int id, next_id;
1069 unsigned skipped = 0;
1070
1071 tso_start(&tso, mbuf);
1072
1073 while (dma_seg->ds_len + skipped <= tso.header_len) {
1074 skipped += dma_seg->ds_len;
1075 --n_dma_seg;
1076 KASSERT(n_dma_seg, ("no payload found in TSO packet"));
1077 ++dma_seg;
1078 }
1079 tso.in_len = dma_seg->ds_len + (tso.header_len - skipped);
1080 tso.dma_addr = dma_seg->ds_addr + (tso.header_len - skipped);
1081
1082 id = txq->added & txq->ptr_mask;
1083 if (__predict_false(tso_start_new_packet(txq, &tso, id)))
1084 return (-1);
1085
1086 while (1) {
1087 id = (id + 1) & txq->ptr_mask;
1088 tso_fill_packet_with_fragment(txq, &tso);
1089
1090 /* Move onto the next fragment? */
1091 if (tso.in_len == 0) {
1092 --n_dma_seg;
1093 if (n_dma_seg == 0)
1094 break;
1095 ++dma_seg;
1096 tso.in_len = dma_seg->ds_len;
1097 tso.dma_addr = dma_seg->ds_addr;
1098 }
1099
1100 /* End of packet? */
1101 if (tso.packet_space == 0) {
1102 /* If the queue is now full due to tiny MSS,
1103 * or we can't create another header, discard
1104 * the remainder of the input mbuf but do not
1105 * roll back the work we have done.
1106 */
1107 if (txq->n_pend_desc >
1108 SFXGE_TSO_MAX_DESC - (1 + SFXGE_TX_MAPPING_MAX_SEG)) {
1109 txq->tso_pdrop_too_many++;
1110 break;
1111 }
1112 next_id = (id + 1) & txq->ptr_mask;
1113 if (__predict_false(tso_start_new_packet(txq, &tso,
1114 next_id))) {
1115 txq->tso_pdrop_no_rsrc++;
1116 break;
1117 }
1118 id = next_id;
1119 }
1120 }
1121
1122 txq->tso_bursts++;
1123 return (id);
1124}
1125
1126static void
1127sfxge_tx_qunblock(struct sfxge_txq *txq)
1128{
1129 struct sfxge_softc *sc;
1130 struct sfxge_evq *evq;
1131
1132 sc = txq->sc;
1133 evq = sc->evq[txq->evq_index];
1134
1135 SFXGE_EVQ_LOCK_ASSERT_OWNED(evq);
1136
1137 if (txq->init_state != SFXGE_TXQ_STARTED)
1138 return;
1139
1140 SFXGE_TXQ_LOCK(txq);
1141
1142 if (txq->blocked) {
1143 unsigned int level;
1144
1145 level = txq->added - txq->completed;
1146 if (level <= SFXGE_TXQ_UNBLOCK_LEVEL(txq->entries)) {
1147 /* reaped must be in sync with blocked */
1148 sfxge_tx_qreap(txq);
1149 txq->blocked = 0;
1150 }
1151 }
1152
1153 sfxge_tx_qdpl_service(txq);
1154 /* note: lock has been dropped */
1155}
1156
1157void
1158sfxge_tx_qflush_done(struct sfxge_txq *txq)
1159{
1160
1161 txq->flush_state = SFXGE_FLUSH_DONE;
1162}
1163
1164static void
1165sfxge_tx_qstop(struct sfxge_softc *sc, unsigned int index)
1166{
1167 struct sfxge_txq *txq;
1168 struct sfxge_evq *evq;
1169 unsigned int count;
1170
1171 txq = sc->txq[index];
1172 evq = sc->evq[txq->evq_index];
1173
1174 SFXGE_TXQ_LOCK(txq);
1175
1176 KASSERT(txq->init_state == SFXGE_TXQ_STARTED,
1177 ("txq->init_state != SFXGE_TXQ_STARTED"));
1178
1179 txq->init_state = SFXGE_TXQ_INITIALIZED;
1180 txq->flush_state = SFXGE_FLUSH_PENDING;
1181
1182 /* Flush the transmit queue. */
1183 efx_tx_qflush(txq->common);
1184
1185 SFXGE_TXQ_UNLOCK(txq);
1186
1187 count = 0;
1188 do {
1189 /* Spin for 100ms. */
1190 DELAY(100000);
1191
1192 if (txq->flush_state != SFXGE_FLUSH_PENDING)
1193 break;
1194 } while (++count < 20);
1195
1196 SFXGE_EVQ_LOCK(evq);
1197 SFXGE_TXQ_LOCK(txq);
1198
1199 KASSERT(txq->flush_state != SFXGE_FLUSH_FAILED,
1200 ("txq->flush_state == SFXGE_FLUSH_FAILED"));
1201
1202 txq->flush_state = SFXGE_FLUSH_DONE;
1203
1204 txq->blocked = 0;
1205 txq->pending = txq->added;
1206
1207 sfxge_tx_qcomplete(txq, evq);
1208 KASSERT(txq->completed == txq->added,
1209 ("txq->completed != txq->added"));
1210
1211 sfxge_tx_qreap(txq);
1212 KASSERT(txq->reaped == txq->completed,
1213 ("txq->reaped != txq->completed"));
1214
1215 txq->added = 0;
1216 txq->pending = 0;
1217 txq->completed = 0;
1218 txq->reaped = 0;
1219
1220 /* Destroy the common code transmit queue. */
1221 efx_tx_qdestroy(txq->common);
1222 txq->common = NULL;
1223
1224 efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id,
1225 EFX_TXQ_NBUFS(sc->txq_entries));
1226
1227 SFXGE_EVQ_UNLOCK(evq);
1228 SFXGE_TXQ_UNLOCK(txq);
1229}
1230
1231static int
1232sfxge_tx_qstart(struct sfxge_softc *sc, unsigned int index)
1233{
1234 struct sfxge_txq *txq;
1235 efsys_mem_t *esmp;
1236 uint16_t flags;
1237 struct sfxge_evq *evq;
1238 int rc;
1239
1240 txq = sc->txq[index];
1241 esmp = &txq->mem;
1242 evq = sc->evq[txq->evq_index];
1243
1244 KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED,
1245 ("txq->init_state != SFXGE_TXQ_INITIALIZED"));
1246 KASSERT(evq->init_state == SFXGE_EVQ_STARTED,
1247 ("evq->init_state != SFXGE_EVQ_STARTED"));
1248
1249 /* Program the buffer table. */
1250 if ((rc = efx_sram_buf_tbl_set(sc->enp, txq->buf_base_id, esmp,
1251 EFX_TXQ_NBUFS(sc->txq_entries))) != 0)
1252 return (rc);
1253
1254 /* Determine the kind of queue we are creating. */
1255 switch (txq->type) {
1256 case SFXGE_TXQ_NON_CKSUM:
1257 flags = 0;
1258 break;
1259 case SFXGE_TXQ_IP_CKSUM:
1260 flags = EFX_CKSUM_IPV4;
1261 break;
1262 case SFXGE_TXQ_IP_TCP_UDP_CKSUM:
1263 flags = EFX_CKSUM_IPV4 | EFX_CKSUM_TCPUDP;
1264 break;
1265 default:
1266 KASSERT(0, ("Impossible TX queue"));
1267 flags = 0;
1268 break;
1269 }
1270
1271 /* Create the common code transmit queue. */
1272 if ((rc = efx_tx_qcreate(sc->enp, index, txq->type, esmp,
1273 sc->txq_entries, txq->buf_base_id, flags, evq->common,
1274 &txq->common)) != 0)
1275 goto fail;
1276
1277 SFXGE_TXQ_LOCK(txq);
1278
1279 /* Enable the transmit queue. */
1280 efx_tx_qenable(txq->common);
1281
1282 txq->init_state = SFXGE_TXQ_STARTED;
1283
1284 SFXGE_TXQ_UNLOCK(txq);
1285
1286 return (0);
1287
1288fail:
1289 efx_sram_buf_tbl_clear(sc->enp, txq->buf_base_id,
1290 EFX_TXQ_NBUFS(sc->txq_entries));
1291 return (rc);
1292}
1293
1294void
1295sfxge_tx_stop(struct sfxge_softc *sc)
1296{
1297 int index;
1298
1299 index = sc->txq_count;
1300 while (--index >= 0)
1301 sfxge_tx_qstop(sc, index);
1302
1303 /* Tear down the transmit module */
1304 efx_tx_fini(sc->enp);
1305}
1306
1307int
1308sfxge_tx_start(struct sfxge_softc *sc)
1309{
1310 int index;
1311 int rc;
1312
1313 /* Initialize the common code transmit module. */
1314 if ((rc = efx_tx_init(sc->enp)) != 0)
1315 return (rc);
1316
1317 for (index = 0; index < sc->txq_count; index++) {
1318 if ((rc = sfxge_tx_qstart(sc, index)) != 0)
1319 goto fail;
1320 }
1321
1322 return (0);
1323
1324fail:
1325 while (--index >= 0)
1326 sfxge_tx_qstop(sc, index);
1327
1328 efx_tx_fini(sc->enp);
1329
1330 return (rc);
1331}
1332
1333/**
1334 * Destroy a transmit queue.
1335 */
1336static void
1337sfxge_tx_qfini(struct sfxge_softc *sc, unsigned int index)
1338{
1339 struct sfxge_txq *txq;
1340 unsigned int nmaps;
1341
1342 txq = sc->txq[index];
1343
1344 KASSERT(txq->init_state == SFXGE_TXQ_INITIALIZED,
1345 ("txq->init_state != SFXGE_TXQ_INITIALIZED"));
1346
1347 if (txq->type == SFXGE_TXQ_IP_TCP_UDP_CKSUM)
1348 tso_fini(txq);
1349
1350 /* Free the context arrays. */
1351 free(txq->pend_desc, M_SFXGE);
1352 nmaps = sc->txq_entries;
1353 while (nmaps-- != 0)
1354 bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map);
1355 free(txq->stmp, M_SFXGE);
1356
1357 /* Release DMA memory mapping. */
1358 sfxge_dma_free(&txq->mem);
1359
1360 sc->txq[index] = NULL;
1361
1362#ifdef SFXGE_HAVE_MQ
1363 SFXGE_TXQ_LOCK_DESTROY(txq);
1364#endif
1365
1366 free(txq, M_SFXGE);
1367}
1368
1369static int
1370sfxge_tx_qinit(struct sfxge_softc *sc, unsigned int txq_index,
1371 enum sfxge_txq_type type, unsigned int evq_index)
1372{
1373 char name[16];
1374 struct sysctl_oid *txq_node;
1375 struct sfxge_txq *txq;
1376 struct sfxge_evq *evq;
1377#ifdef SFXGE_HAVE_MQ
1378 struct sfxge_tx_dpl *stdp;
1379#endif
1380 efsys_mem_t *esmp;
1381 unsigned int nmaps;
1382 int rc;
1383
1384 txq = malloc(sizeof(struct sfxge_txq), M_SFXGE, M_ZERO | M_WAITOK);
1385 txq->sc = sc;
1386 txq->entries = sc->txq_entries;
1387 txq->ptr_mask = txq->entries - 1;
1388
1389 sc->txq[txq_index] = txq;
1390 esmp = &txq->mem;
1391
1392 evq = sc->evq[evq_index];
1393
1394 /* Allocate and zero DMA space for the descriptor ring. */
1395 if ((rc = sfxge_dma_alloc(sc, EFX_TXQ_SIZE(sc->txq_entries), esmp)) != 0)
1396 return (rc);
1397 (void)memset(esmp->esm_base, 0, EFX_TXQ_SIZE(sc->txq_entries));
1398
1399 /* Allocate buffer table entries. */
1400 sfxge_sram_buf_tbl_alloc(sc, EFX_TXQ_NBUFS(sc->txq_entries),
1401 &txq->buf_base_id);
1402
1403 /* Create a DMA tag for packet mappings. */
1404 if (bus_dma_tag_create(sc->parent_dma_tag, 1, 0x1000,
1405 MIN(0x3FFFFFFFFFFFUL, BUS_SPACE_MAXADDR), BUS_SPACE_MAXADDR, NULL,
1406 NULL, 0x11000, SFXGE_TX_MAPPING_MAX_SEG, 0x1000, 0, NULL, NULL,
1407 &txq->packet_dma_tag) != 0) {
1408 device_printf(sc->dev, "Couldn't allocate txq DMA tag\n");
1409 rc = ENOMEM;
1410 goto fail;
1411 }
1412
1413 /* Allocate pending descriptor array for batching writes. */
1414 txq->pend_desc = malloc(sizeof(efx_buffer_t) * sc->txq_entries,
1415 M_SFXGE, M_ZERO | M_WAITOK);
1416
1417 /* Allocate and initialise mbuf DMA mapping array. */
1418 txq->stmp = malloc(sizeof(struct sfxge_tx_mapping) * sc->txq_entries,
1419 M_SFXGE, M_ZERO | M_WAITOK);
1420 for (nmaps = 0; nmaps < sc->txq_entries; nmaps++) {
1421 rc = bus_dmamap_create(txq->packet_dma_tag, 0,
1422 &txq->stmp[nmaps].map);
1423 if (rc != 0)
1424 goto fail2;
1425 }
1426
1427 snprintf(name, sizeof(name), "%u", txq_index);
1428 txq_node = SYSCTL_ADD_NODE(
1429 device_get_sysctl_ctx(sc->dev),
1430 SYSCTL_CHILDREN(sc->txqs_node),
1431 OID_AUTO, name, CTLFLAG_RD, NULL, "");
1432 if (txq_node == NULL) {
1433 rc = ENOMEM;
1434 goto fail_txq_node;
1435 }
1436
1437 if (type == SFXGE_TXQ_IP_TCP_UDP_CKSUM &&
1438 (rc = tso_init(txq)) != 0)
1439 goto fail3;
1440
1441#ifdef SFXGE_HAVE_MQ
1442 if (sfxge_tx_dpl_get_max <= 0) {
1443 log(LOG_ERR, "%s=%d must be greater than 0",
1444 SFXGE_PARAM_TX_DPL_GET_MAX, sfxge_tx_dpl_get_max);
1445 rc = EINVAL;
1446 goto fail_tx_dpl_get_max;
1447 }
1448 if (sfxge_tx_dpl_get_non_tcp_max <= 0) {
1449 log(LOG_ERR, "%s=%d must be greater than 0",
1450 SFXGE_PARAM_TX_DPL_GET_NON_TCP_MAX,
1451 sfxge_tx_dpl_get_non_tcp_max);
1452 rc = EINVAL;
1453 goto fail_tx_dpl_get_max;
1454 }
1455 if (sfxge_tx_dpl_put_max < 0) {
1456 log(LOG_ERR, "%s=%d must be greater or equal to 0",
1457 SFXGE_PARAM_TX_DPL_PUT_MAX, sfxge_tx_dpl_put_max);
1458 rc = EINVAL;
1459 goto fail_tx_dpl_put_max;
1460 }
1461
1462 /* Initialize the deferred packet list. */
1463 stdp = &txq->dpl;
1464 stdp->std_put_max = sfxge_tx_dpl_put_max;
1465 stdp->std_get_max = sfxge_tx_dpl_get_max;
1466 stdp->std_get_non_tcp_max = sfxge_tx_dpl_get_non_tcp_max;
1467 stdp->std_getp = &stdp->std_get;
1468
1469 SFXGE_TXQ_LOCK_INIT(txq, device_get_nameunit(sc->dev), txq_index);
1470
1471 SYSCTL_ADD_UINT(device_get_sysctl_ctx(sc->dev),
1472 SYSCTL_CHILDREN(txq_node), OID_AUTO,
1473 "dpl_get_count", CTLFLAG_RD | CTLFLAG_STATS,
1474 &stdp->std_get_count, 0, "");
1475 SYSCTL_ADD_UINT(device_get_sysctl_ctx(sc->dev),
1476 SYSCTL_CHILDREN(txq_node), OID_AUTO,
1477 "dpl_get_non_tcp_count", CTLFLAG_RD | CTLFLAG_STATS,
1478 &stdp->std_get_non_tcp_count, 0, "");
1479 SYSCTL_ADD_UINT(device_get_sysctl_ctx(sc->dev),
1480 SYSCTL_CHILDREN(txq_node), OID_AUTO,
1481 "dpl_get_hiwat", CTLFLAG_RD | CTLFLAG_STATS,
1482 &stdp->std_get_hiwat, 0, "");
1483#endif
1484
1485 txq->type = type;
1486 txq->evq_index = evq_index;
1487 txq->txq_index = txq_index;
1488 txq->init_state = SFXGE_TXQ_INITIALIZED;
1489
1490 return (0);
1491
1492fail_tx_dpl_put_max:
1493fail_tx_dpl_get_max:
1494fail3:
1495fail_txq_node:
1496 free(txq->pend_desc, M_SFXGE);
1497fail2:
1498 while (nmaps-- != 0)
1499 bus_dmamap_destroy(txq->packet_dma_tag, txq->stmp[nmaps].map);
1500 free(txq->stmp, M_SFXGE);
1501 bus_dma_tag_destroy(txq->packet_dma_tag);
1502
1503fail:
1504 sfxge_dma_free(esmp);
1505
1506 return (rc);
1507}
1508
1509static const struct {
1510 const char *name;
1511 size_t offset;
1512} sfxge_tx_stats[] = {
1513#define SFXGE_TX_STAT(name, member) \
1514 { #name, offsetof(struct sfxge_txq, member) }
1515 SFXGE_TX_STAT(tso_bursts, tso_bursts),
1516 SFXGE_TX_STAT(tso_packets, tso_packets),
1517 SFXGE_TX_STAT(tso_long_headers, tso_long_headers),
1518 SFXGE_TX_STAT(tso_pdrop_too_many, tso_pdrop_too_many),
1519 SFXGE_TX_STAT(tso_pdrop_no_rsrc, tso_pdrop_no_rsrc),
1520 SFXGE_TX_STAT(tx_collapses, collapses),
1521 SFXGE_TX_STAT(tx_drops, drops),
1522 SFXGE_TX_STAT(tx_get_overflow, get_overflow),
1523 SFXGE_TX_STAT(tx_get_non_tcp_overflow, get_non_tcp_overflow),
1524 SFXGE_TX_STAT(tx_put_overflow, put_overflow),
1525 SFXGE_TX_STAT(tx_netdown_drops, netdown_drops),
1526};
1527
1528static int
1529sfxge_tx_stat_handler(SYSCTL_HANDLER_ARGS)
1530{
1531 struct sfxge_softc *sc = arg1;
1532 unsigned int id = arg2;
1533 unsigned long sum;
1534 unsigned int index;
1535
1536 /* Sum across all TX queues */
1537 sum = 0;
1538 for (index = 0; index < sc->txq_count; index++)
1539 sum += *(unsigned long *)((caddr_t)sc->txq[index] +
1540 sfxge_tx_stats[id].offset);
1541
1542 return (SYSCTL_OUT(req, &sum, sizeof(sum)));
1543}
1544
1545static void
1546sfxge_tx_stat_init(struct sfxge_softc *sc)
1547{
1548 struct sysctl_ctx_list *ctx = device_get_sysctl_ctx(sc->dev);
1549 struct sysctl_oid_list *stat_list;
1550 unsigned int id;
1551
1552 stat_list = SYSCTL_CHILDREN(sc->stats_node);
1553
1554 for (id = 0; id < nitems(sfxge_tx_stats); id++) {
1555 SYSCTL_ADD_PROC(
1556 ctx, stat_list,
1557 OID_AUTO, sfxge_tx_stats[id].name,
1558 CTLTYPE_ULONG|CTLFLAG_RD,
1559 sc, id, sfxge_tx_stat_handler, "LU",
1560 "");
1561 }
1562}
1563
1564void
1565sfxge_tx_fini(struct sfxge_softc *sc)
1566{
1567 int index;
1568
1569 index = sc->txq_count;
1570 while (--index >= 0)
1571 sfxge_tx_qfini(sc, index);
1572
1573 sc->txq_count = 0;
1574}
1575
1576
1577int
1578sfxge_tx_init(struct sfxge_softc *sc)
1579{
1580 struct sfxge_intr *intr;
1581 int index;
1582 int rc;
1583
1584 intr = &sc->intr;
1585
1586 KASSERT(intr->state == SFXGE_INTR_INITIALIZED,
1587 ("intr->state != SFXGE_INTR_INITIALIZED"));
1588
1589#ifdef SFXGE_HAVE_MQ
1590 sc->txq_count = SFXGE_TXQ_NTYPES - 1 + sc->intr.n_alloc;
1591#else
1592 sc->txq_count = SFXGE_TXQ_NTYPES;
1593#endif
1594
1595 sc->txqs_node = SYSCTL_ADD_NODE(
1596 device_get_sysctl_ctx(sc->dev),
1597 SYSCTL_CHILDREN(device_get_sysctl_tree(sc->dev)),
1598 OID_AUTO, "txq", CTLFLAG_RD, NULL, "Tx queues");
1599 if (sc->txqs_node == NULL) {
1600 rc = ENOMEM;
1601 goto fail_txq_node;
1602 }
1603
1604 /* Initialize the transmit queues */
1605 if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NON_CKSUM,
1606 SFXGE_TXQ_NON_CKSUM, 0)) != 0)
1607 goto fail;
1608
1609 if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_IP_CKSUM,
1610 SFXGE_TXQ_IP_CKSUM, 0)) != 0)
1611 goto fail2;
1612
1613 for (index = 0;
1614 index < sc->txq_count - SFXGE_TXQ_NTYPES + 1;
1615 index++) {
1616 if ((rc = sfxge_tx_qinit(sc, SFXGE_TXQ_NTYPES - 1 + index,
1617 SFXGE_TXQ_IP_TCP_UDP_CKSUM, index)) != 0)
1618 goto fail3;
1619 }
1620
1621 sfxge_tx_stat_init(sc);
1622
1623 return (0);
1624
1625fail3:
1626 while (--index >= 0)
1627 sfxge_tx_qfini(sc, SFXGE_TXQ_IP_TCP_UDP_CKSUM + index);
1628
1629 sfxge_tx_qfini(sc, SFXGE_TXQ_IP_CKSUM);
1630
1631fail2:
1632 sfxge_tx_qfini(sc, SFXGE_TXQ_NON_CKSUM);
1633
1634fail:
1635fail_txq_node:
1636 sc->txq_count = 0;
1637 return (rc);
1638}
|