1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2010-2014 Intel Corporation
3 */
4
5 #ifndef _VIRTQUEUE_H_
6 #define _VIRTQUEUE_H_
7
8 #include <stdint.h>
9
10 #include <rte_atomic.h>
11 #include <rte_memory.h>
12 #include <rte_mempool.h>
13 #include <rte_net.h>
14
15 #include "virtio.h"
16 #include "virtio_ring.h"
17 #include "virtio_logs.h"
18 #include "virtio_rxtx.h"
19
20 struct rte_mbuf;
21
22 #define DEFAULT_TX_FREE_THRESH 32
23 #define DEFAULT_RX_FREE_THRESH 32
24
25 #define VIRTIO_MBUF_BURST_SZ 64
26 /*
27 * Per virtio_ring.h in Linux.
28 * For virtio_pci on SMP, we don't need to order with respect to MMIO
29 * accesses through relaxed memory I/O windows, so thread_fence is
30 * sufficient.
31 *
32 * For using virtio to talk to real devices (eg. vDPA) we do need real
33 * barriers.
34 */
35 static inline void
virtio_mb(uint8_t weak_barriers)36 virtio_mb(uint8_t weak_barriers)
37 {
38 if (weak_barriers)
39 rte_atomic_thread_fence(__ATOMIC_SEQ_CST);
40 else
41 rte_mb();
42 }
43
44 static inline void
virtio_rmb(uint8_t weak_barriers)45 virtio_rmb(uint8_t weak_barriers)
46 {
47 if (weak_barriers)
48 rte_atomic_thread_fence(__ATOMIC_ACQUIRE);
49 else
50 rte_io_rmb();
51 }
52
53 static inline void
virtio_wmb(uint8_t weak_barriers)54 virtio_wmb(uint8_t weak_barriers)
55 {
56 if (weak_barriers)
57 rte_atomic_thread_fence(__ATOMIC_RELEASE);
58 else
59 rte_io_wmb();
60 }
61
62 static inline uint16_t
virtqueue_fetch_flags_packed(struct vring_packed_desc * dp,uint8_t weak_barriers)63 virtqueue_fetch_flags_packed(struct vring_packed_desc *dp,
64 uint8_t weak_barriers)
65 {
66 uint16_t flags;
67
68 if (weak_barriers) {
69 /* x86 prefers to using rte_io_rmb over __atomic_load_n as it reports
70 * a better perf(~1.5%), which comes from the saved branch by the compiler.
71 * The if and else branch are identical on the platforms except Arm.
72 */
73 #ifdef RTE_ARCH_ARM
74 flags = __atomic_load_n(&dp->flags, __ATOMIC_ACQUIRE);
75 #else
76 flags = dp->flags;
77 rte_io_rmb();
78 #endif
79 } else {
80 flags = dp->flags;
81 rte_io_rmb();
82 }
83
84 return flags;
85 }
86
87 static inline void
virtqueue_store_flags_packed(struct vring_packed_desc * dp,uint16_t flags,uint8_t weak_barriers)88 virtqueue_store_flags_packed(struct vring_packed_desc *dp,
89 uint16_t flags, uint8_t weak_barriers)
90 {
91 if (weak_barriers) {
92 /* x86 prefers to using rte_io_wmb over __atomic_store_n as it reports
93 * a better perf(~1.5%), which comes from the saved branch by the compiler.
94 * The if and else branch are identical on the platforms except Arm.
95 */
96 #ifdef RTE_ARCH_ARM
97 __atomic_store_n(&dp->flags, flags, __ATOMIC_RELEASE);
98 #else
99 rte_io_wmb();
100 dp->flags = flags;
101 #endif
102 } else {
103 rte_io_wmb();
104 dp->flags = flags;
105 }
106 }
107
108 #ifdef RTE_PMD_PACKET_PREFETCH
109 #define rte_packet_prefetch(p) rte_prefetch1(p)
110 #else
111 #define rte_packet_prefetch(p) do {} while(0)
112 #endif
113
114 #define VIRTQUEUE_MAX_NAME_SZ 32
115
116 /**
117 * Return the IOVA (or virtual address in case of virtio-user) of mbuf
118 * data buffer.
119 *
120 * The address is firstly casted to the word size (sizeof(uintptr_t))
121 * before casting it to uint64_t. This is to make it work with different
122 * combination of word size (64 bit and 32 bit) and virtio device
123 * (virtio-pci and virtio-user).
124 */
125 #define VIRTIO_MBUF_ADDR(mb, vq) \
126 ((uint64_t)(*(uintptr_t *)((uintptr_t)(mb) + (vq)->mbuf_addr_offset)))
127
128 /**
129 * Return the physical address (or virtual address in case of
130 * virtio-user) of mbuf data buffer, taking care of mbuf data offset
131 */
132 #define VIRTIO_MBUF_DATA_DMA_ADDR(mb, vq) \
133 (VIRTIO_MBUF_ADDR(mb, vq) + (mb)->data_off)
134
135 #define VTNET_SQ_RQ_QUEUE_IDX 0
136 #define VTNET_SQ_TQ_QUEUE_IDX 1
137 #define VTNET_SQ_CQ_QUEUE_IDX 2
138
139 enum { VTNET_RQ = 0, VTNET_TQ = 1, VTNET_CQ = 2 };
140 /**
141 * The maximum virtqueue size is 2^15. Use that value as the end of
142 * descriptor chain terminator since it will never be a valid index
143 * in the descriptor table. This is used to verify we are correctly
144 * handling vq_free_cnt.
145 */
146 #define VQ_RING_DESC_CHAIN_END 32768
147
148 /**
149 * Control the RX mode, ie. promiscuous, allmulti, etc...
150 * All commands require an "out" sg entry containing a 1 byte
151 * state value, zero = disable, non-zero = enable. Commands
152 * 0 and 1 are supported with the VIRTIO_NET_F_CTRL_RX feature.
153 * Commands 2-5 are added with VIRTIO_NET_F_CTRL_RX_EXTRA.
154 */
155 #define VIRTIO_NET_CTRL_RX 0
156 #define VIRTIO_NET_CTRL_RX_PROMISC 0
157 #define VIRTIO_NET_CTRL_RX_ALLMULTI 1
158 #define VIRTIO_NET_CTRL_RX_ALLUNI 2
159 #define VIRTIO_NET_CTRL_RX_NOMULTI 3
160 #define VIRTIO_NET_CTRL_RX_NOUNI 4
161 #define VIRTIO_NET_CTRL_RX_NOBCAST 5
162
163 /**
164 * Control the MAC
165 *
166 * The MAC filter table is managed by the hypervisor, the guest should
167 * assume the size is infinite. Filtering should be considered
168 * non-perfect, ie. based on hypervisor resources, the guest may
169 * received packets from sources not specified in the filter list.
170 *
171 * In addition to the class/cmd header, the TABLE_SET command requires
172 * two out scatterlists. Each contains a 4 byte count of entries followed
173 * by a concatenated byte stream of the ETH_ALEN MAC addresses. The
174 * first sg list contains unicast addresses, the second is for multicast.
175 * This functionality is present if the VIRTIO_NET_F_CTRL_RX feature
176 * is available.
177 *
178 * The ADDR_SET command requests one out scatterlist, it contains a
179 * 6 bytes MAC address. This functionality is present if the
180 * VIRTIO_NET_F_CTRL_MAC_ADDR feature is available.
181 */
182 struct virtio_net_ctrl_mac {
183 uint32_t entries;
184 uint8_t macs[][RTE_ETHER_ADDR_LEN];
185 } __rte_packed;
186
187 #define VIRTIO_NET_CTRL_MAC 1
188 #define VIRTIO_NET_CTRL_MAC_TABLE_SET 0
189 #define VIRTIO_NET_CTRL_MAC_ADDR_SET 1
190
191 /**
192 * Control VLAN filtering
193 *
194 * The VLAN filter table is controlled via a simple ADD/DEL interface.
195 * VLAN IDs not added may be filtered by the hypervisor. Del is the
196 * opposite of add. Both commands expect an out entry containing a 2
197 * byte VLAN ID. VLAN filtering is available with the
198 * VIRTIO_NET_F_CTRL_VLAN feature bit.
199 */
200 #define VIRTIO_NET_CTRL_VLAN 2
201 #define VIRTIO_NET_CTRL_VLAN_ADD 0
202 #define VIRTIO_NET_CTRL_VLAN_DEL 1
203
204 /**
205 * RSS control
206 *
207 * The RSS feature configuration message is sent by the driver when
208 * VIRTIO_NET_F_RSS has been negotiated. It provides the device with
209 * hash types to use, hash key and indirection table. In this
210 * implementation, the driver only supports fixed key length (40B)
211 * and indirection table size (128 entries).
212 */
213 #define VIRTIO_NET_RSS_RETA_SIZE 128
214 #define VIRTIO_NET_RSS_KEY_SIZE 40
215
216 struct virtio_net_ctrl_rss {
217 uint32_t hash_types;
218 uint16_t indirection_table_mask;
219 uint16_t unclassified_queue;
220 uint16_t indirection_table[VIRTIO_NET_RSS_RETA_SIZE];
221 uint16_t max_tx_vq;
222 uint8_t hash_key_length;
223 uint8_t hash_key_data[VIRTIO_NET_RSS_KEY_SIZE];
224 };
225
226 /*
227 * Control link announce acknowledgement
228 *
229 * The command VIRTIO_NET_CTRL_ANNOUNCE_ACK is used to indicate that
230 * driver has received the notification; device would clear the
231 * VIRTIO_NET_S_ANNOUNCE bit in the status field after it receives
232 * this command.
233 */
234 #define VIRTIO_NET_CTRL_ANNOUNCE 3
235 #define VIRTIO_NET_CTRL_ANNOUNCE_ACK 0
236
237 struct virtio_net_ctrl_hdr {
238 uint8_t class;
239 uint8_t cmd;
240 } __rte_packed;
241
242 typedef uint8_t virtio_net_ctrl_ack;
243
244 #define VIRTIO_NET_OK 0
245 #define VIRTIO_NET_ERR 1
246
247 #define VIRTIO_MAX_CTRL_DATA 2048
248
249 struct virtio_pmd_ctrl {
250 struct virtio_net_ctrl_hdr hdr;
251 virtio_net_ctrl_ack status;
252 uint8_t data[VIRTIO_MAX_CTRL_DATA];
253 };
254
255 struct vq_desc_extra {
256 void *cookie;
257 uint16_t ndescs;
258 uint16_t next;
259 };
260
261 #define virtnet_rxq_to_vq(rxvq) container_of(rxvq, struct virtqueue, rxq)
262 #define virtnet_txq_to_vq(txvq) container_of(txvq, struct virtqueue, txq)
263 #define virtnet_cq_to_vq(cvq) container_of(cvq, struct virtqueue, cq)
264
265 struct virtqueue {
266 struct virtio_hw *hw; /**< virtio_hw structure pointer. */
267 union {
268 struct {
269 /**< vring keeping desc, used and avail */
270 struct vring ring;
271 } vq_split;
272
273 struct {
274 /**< vring keeping descs and events */
275 struct vring_packed ring;
276 bool used_wrap_counter;
277 uint16_t cached_flags; /**< cached flags for descs */
278 uint16_t event_flags_shadow;
279 } vq_packed;
280 };
281
282 uint16_t vq_used_cons_idx; /**< last consumed descriptor */
283 uint16_t vq_nentries; /**< vring desc numbers */
284 uint16_t vq_free_cnt; /**< num of desc available */
285 uint16_t vq_avail_idx; /**< sync until needed */
286 uint16_t vq_free_thresh; /**< free threshold */
287
288 /**
289 * Head of the free chain in the descriptor table. If
290 * there are no free descriptors, this will be set to
291 * VQ_RING_DESC_CHAIN_END.
292 */
293 uint16_t vq_desc_head_idx;
294 uint16_t vq_desc_tail_idx;
295 uint16_t vq_queue_index; /**< PCI queue index */
296
297 void *vq_ring_virt_mem; /**< linear address of vring*/
298 unsigned int vq_ring_size;
299 uint16_t mbuf_addr_offset;
300
301 union {
302 struct virtnet_rx rxq;
303 struct virtnet_tx txq;
304 struct virtnet_ctl cq;
305 };
306
307 rte_iova_t vq_ring_mem; /**< physical address of vring,
308 * or virtual address for virtio_user. */
309
310 uint16_t *notify_addr;
311 struct rte_mbuf **sw_ring; /**< RX software ring. */
312 struct vq_desc_extra vq_descx[0];
313 };
314
315 /* If multiqueue is provided by host, then we support it. */
316 #define VIRTIO_NET_CTRL_MQ 4
317
318 #define VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET 0
319 #define VIRTIO_NET_CTRL_MQ_RSS_CONFIG 1
320
321 #define VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN 1
322 #define VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX 0x8000
323
324 /**
325 * This is the first element of the scatter-gather list. If you don't
326 * specify GSO or CSUM features, you can simply ignore the header.
327 */
328 struct virtio_net_hdr {
329 #define VIRTIO_NET_HDR_F_NEEDS_CSUM 1 /**< Use csum_start,csum_offset*/
330 #define VIRTIO_NET_HDR_F_DATA_VALID 2 /**< Checksum is valid */
331 uint8_t flags;
332 #define VIRTIO_NET_HDR_GSO_NONE 0 /**< Not a GSO frame */
333 #define VIRTIO_NET_HDR_GSO_TCPV4 1 /**< GSO frame, IPv4 TCP (TSO) */
334 #define VIRTIO_NET_HDR_GSO_UDP 3 /**< GSO frame, IPv4 UDP (UFO) */
335 #define VIRTIO_NET_HDR_GSO_TCPV6 4 /**< GSO frame, IPv6 TCP */
336 #define VIRTIO_NET_HDR_GSO_ECN 0x80 /**< TCP has ECN set */
337 uint8_t gso_type;
338 uint16_t hdr_len; /**< Ethernet + IP + tcp/udp hdrs */
339 uint16_t gso_size; /**< Bytes to append to hdr_len per frame */
340 uint16_t csum_start; /**< Position to start checksumming from */
341 uint16_t csum_offset; /**< Offset after that to place checksum */
342 };
343
344 /**
345 * This is the version of the header to use when the MRG_RXBUF
346 * feature has been negotiated.
347 */
348 struct virtio_net_hdr_mrg_rxbuf {
349 struct virtio_net_hdr hdr;
350 uint16_t num_buffers; /**< Number of merged rx buffers */
351 };
352
353 /* Region reserved to allow for transmit header and indirect ring */
354 #define VIRTIO_MAX_TX_INDIRECT 8
355 struct virtio_tx_region {
356 struct virtio_net_hdr_mrg_rxbuf tx_hdr;
357 union {
358 struct vring_desc tx_indir[VIRTIO_MAX_TX_INDIRECT];
359 struct vring_packed_desc
360 tx_packed_indir[VIRTIO_MAX_TX_INDIRECT];
361 } __rte_aligned(16);
362 };
363
364 static inline int
desc_is_used(struct vring_packed_desc * desc,struct virtqueue * vq)365 desc_is_used(struct vring_packed_desc *desc, struct virtqueue *vq)
366 {
367 uint16_t used, avail, flags;
368
369 flags = virtqueue_fetch_flags_packed(desc, vq->hw->weak_barriers);
370 used = !!(flags & VRING_PACKED_DESC_F_USED);
371 avail = !!(flags & VRING_PACKED_DESC_F_AVAIL);
372
373 return avail == used && used == vq->vq_packed.used_wrap_counter;
374 }
375
376 static inline void
vring_desc_init_packed(struct virtqueue * vq,int n)377 vring_desc_init_packed(struct virtqueue *vq, int n)
378 {
379 int i;
380 for (i = 0; i < n - 1; i++) {
381 vq->vq_packed.ring.desc[i].id = i;
382 vq->vq_descx[i].next = i + 1;
383 }
384 vq->vq_packed.ring.desc[i].id = i;
385 vq->vq_descx[i].next = VQ_RING_DESC_CHAIN_END;
386 }
387
388 /* Chain all the descriptors in the ring with an END */
389 static inline void
vring_desc_init_split(struct vring_desc * dp,uint16_t n)390 vring_desc_init_split(struct vring_desc *dp, uint16_t n)
391 {
392 uint16_t i;
393
394 for (i = 0; i < n - 1; i++)
395 dp[i].next = (uint16_t)(i + 1);
396 dp[i].next = VQ_RING_DESC_CHAIN_END;
397 }
398
399 static inline void
vring_desc_init_indirect_packed(struct vring_packed_desc * dp,int n)400 vring_desc_init_indirect_packed(struct vring_packed_desc *dp, int n)
401 {
402 int i;
403 for (i = 0; i < n; i++) {
404 dp[i].id = (uint16_t)i;
405 dp[i].flags = VRING_DESC_F_WRITE;
406 }
407 }
408
409 /**
410 * Tell the backend not to interrupt us. Implementation for packed virtqueues.
411 */
412 static inline void
virtqueue_disable_intr_packed(struct virtqueue * vq)413 virtqueue_disable_intr_packed(struct virtqueue *vq)
414 {
415 if (vq->vq_packed.event_flags_shadow != RING_EVENT_FLAGS_DISABLE) {
416 vq->vq_packed.event_flags_shadow = RING_EVENT_FLAGS_DISABLE;
417 vq->vq_packed.ring.driver->desc_event_flags =
418 vq->vq_packed.event_flags_shadow;
419 }
420 }
421
422 /**
423 * Tell the backend not to interrupt us. Implementation for split virtqueues.
424 */
425 static inline void
virtqueue_disable_intr_split(struct virtqueue * vq)426 virtqueue_disable_intr_split(struct virtqueue *vq)
427 {
428 vq->vq_split.ring.avail->flags |= VRING_AVAIL_F_NO_INTERRUPT;
429 }
430
431 /**
432 * Tell the backend not to interrupt us.
433 */
434 static inline void
virtqueue_disable_intr(struct virtqueue * vq)435 virtqueue_disable_intr(struct virtqueue *vq)
436 {
437 if (virtio_with_packed_queue(vq->hw))
438 virtqueue_disable_intr_packed(vq);
439 else
440 virtqueue_disable_intr_split(vq);
441 }
442
443 /**
444 * Tell the backend to interrupt. Implementation for packed virtqueues.
445 */
446 static inline void
virtqueue_enable_intr_packed(struct virtqueue * vq)447 virtqueue_enable_intr_packed(struct virtqueue *vq)
448 {
449 if (vq->vq_packed.event_flags_shadow == RING_EVENT_FLAGS_DISABLE) {
450 vq->vq_packed.event_flags_shadow = RING_EVENT_FLAGS_ENABLE;
451 vq->vq_packed.ring.driver->desc_event_flags =
452 vq->vq_packed.event_flags_shadow;
453 }
454 }
455
456 /**
457 * Tell the backend to interrupt. Implementation for split virtqueues.
458 */
459 static inline void
virtqueue_enable_intr_split(struct virtqueue * vq)460 virtqueue_enable_intr_split(struct virtqueue *vq)
461 {
462 vq->vq_split.ring.avail->flags &= (~VRING_AVAIL_F_NO_INTERRUPT);
463 }
464
465 /**
466 * Tell the backend to interrupt us.
467 */
468 static inline void
virtqueue_enable_intr(struct virtqueue * vq)469 virtqueue_enable_intr(struct virtqueue *vq)
470 {
471 if (virtio_with_packed_queue(vq->hw))
472 virtqueue_enable_intr_packed(vq);
473 else
474 virtqueue_enable_intr_split(vq);
475 }
476
477 /**
478 * Dump virtqueue internal structures, for debug purpose only.
479 */
480 void virtqueue_dump(struct virtqueue *vq);
481 /**
482 * Get all mbufs to be freed.
483 */
484 struct rte_mbuf *virtqueue_detach_unused(struct virtqueue *vq);
485
486 /* Flush the elements in the used ring. */
487 void virtqueue_rxvq_flush(struct virtqueue *vq);
488
489 int virtqueue_rxvq_reset_packed(struct virtqueue *vq);
490
491 int virtqueue_txvq_reset_packed(struct virtqueue *vq);
492
493 static inline int
virtqueue_full(const struct virtqueue * vq)494 virtqueue_full(const struct virtqueue *vq)
495 {
496 return vq->vq_free_cnt == 0;
497 }
498
499 static inline int
virtio_get_queue_type(struct virtio_hw * hw,uint16_t vq_idx)500 virtio_get_queue_type(struct virtio_hw *hw, uint16_t vq_idx)
501 {
502 if (vq_idx == hw->max_queue_pairs * 2)
503 return VTNET_CQ;
504 else if (vq_idx % 2 == 0)
505 return VTNET_RQ;
506 else
507 return VTNET_TQ;
508 }
509
510 /* virtqueue_nused has load-acquire or rte_io_rmb insed */
511 static inline uint16_t
virtqueue_nused(const struct virtqueue * vq)512 virtqueue_nused(const struct virtqueue *vq)
513 {
514 uint16_t idx;
515
516 if (vq->hw->weak_barriers) {
517 /**
518 * x86 prefers to using rte_smp_rmb over __atomic_load_n as it
519 * reports a slightly better perf, which comes from the saved
520 * branch by the compiler.
521 * The if and else branches are identical with the smp and io
522 * barriers both defined as compiler barriers on x86.
523 */
524 #ifdef RTE_ARCH_X86_64
525 idx = vq->vq_split.ring.used->idx;
526 rte_smp_rmb();
527 #else
528 idx = __atomic_load_n(&(vq)->vq_split.ring.used->idx,
529 __ATOMIC_ACQUIRE);
530 #endif
531 } else {
532 idx = vq->vq_split.ring.used->idx;
533 rte_io_rmb();
534 }
535 return idx - vq->vq_used_cons_idx;
536 }
537
538 void vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx);
539 void vq_ring_free_chain_packed(struct virtqueue *vq, uint16_t used_idx);
540 void vq_ring_free_inorder(struct virtqueue *vq, uint16_t desc_idx,
541 uint16_t num);
542
543 static inline void
vq_update_avail_idx(struct virtqueue * vq)544 vq_update_avail_idx(struct virtqueue *vq)
545 {
546 if (vq->hw->weak_barriers) {
547 /* x86 prefers to using rte_smp_wmb over __atomic_store_n as
548 * it reports a slightly better perf, which comes from the
549 * saved branch by the compiler.
550 * The if and else branches are identical with the smp and
551 * io barriers both defined as compiler barriers on x86.
552 */
553 #ifdef RTE_ARCH_X86_64
554 rte_smp_wmb();
555 vq->vq_split.ring.avail->idx = vq->vq_avail_idx;
556 #else
557 __atomic_store_n(&vq->vq_split.ring.avail->idx,
558 vq->vq_avail_idx, __ATOMIC_RELEASE);
559 #endif
560 } else {
561 rte_io_wmb();
562 vq->vq_split.ring.avail->idx = vq->vq_avail_idx;
563 }
564 }
565
566 static inline void
vq_update_avail_ring(struct virtqueue * vq,uint16_t desc_idx)567 vq_update_avail_ring(struct virtqueue *vq, uint16_t desc_idx)
568 {
569 uint16_t avail_idx;
570 /*
571 * Place the head of the descriptor chain into the next slot and make
572 * it usable to the host. The chain is made available now rather than
573 * deferring to virtqueue_notify() in the hopes that if the host is
574 * currently running on another CPU, we can keep it processing the new
575 * descriptor.
576 */
577 avail_idx = (uint16_t)(vq->vq_avail_idx & (vq->vq_nentries - 1));
578 if (unlikely(vq->vq_split.ring.avail->ring[avail_idx] != desc_idx))
579 vq->vq_split.ring.avail->ring[avail_idx] = desc_idx;
580 vq->vq_avail_idx++;
581 }
582
583 static inline int
virtqueue_kick_prepare(struct virtqueue * vq)584 virtqueue_kick_prepare(struct virtqueue *vq)
585 {
586 /*
587 * Ensure updated avail->idx is visible to vhost before reading
588 * the used->flags.
589 */
590 virtio_mb(vq->hw->weak_barriers);
591 return !(vq->vq_split.ring.used->flags & VRING_USED_F_NO_NOTIFY);
592 }
593
594 static inline int
virtqueue_kick_prepare_packed(struct virtqueue * vq)595 virtqueue_kick_prepare_packed(struct virtqueue *vq)
596 {
597 uint16_t flags;
598
599 /*
600 * Ensure updated data is visible to vhost before reading the flags.
601 */
602 virtio_mb(vq->hw->weak_barriers);
603 flags = vq->vq_packed.ring.device->desc_event_flags;
604
605 return flags != RING_EVENT_FLAGS_DISABLE;
606 }
607
608 /*
609 * virtqueue_kick_prepare*() or the virtio_wmb() should be called
610 * before this function to be sure that all the data is visible to vhost.
611 */
612 static inline void
virtqueue_notify(struct virtqueue * vq)613 virtqueue_notify(struct virtqueue *vq)
614 {
615 VIRTIO_OPS(vq->hw)->notify_queue(vq->hw, vq);
616 }
617
618 #ifdef RTE_LIBRTE_VIRTIO_DEBUG_DUMP
619 #define VIRTQUEUE_DUMP(vq) do { \
620 uint16_t used_idx, nused; \
621 used_idx = __atomic_load_n(&(vq)->vq_split.ring.used->idx, \
622 __ATOMIC_RELAXED); \
623 nused = (uint16_t)(used_idx - (vq)->vq_used_cons_idx); \
624 if (virtio_with_packed_queue((vq)->hw)) { \
625 PMD_INIT_LOG(DEBUG, \
626 "VQ: - size=%d; free=%d; used_cons_idx=%d; avail_idx=%d;" \
627 " cached_flags=0x%x; used_wrap_counter=%d", \
628 (vq)->vq_nentries, (vq)->vq_free_cnt, (vq)->vq_used_cons_idx, \
629 (vq)->vq_avail_idx, (vq)->vq_packed.cached_flags, \
630 (vq)->vq_packed.used_wrap_counter); \
631 break; \
632 } \
633 PMD_INIT_LOG(DEBUG, \
634 "VQ: - size=%d; free=%d; used=%d; desc_head_idx=%d;" \
635 " avail.idx=%d; used_cons_idx=%d; used.idx=%d;" \
636 " avail.flags=0x%x; used.flags=0x%x", \
637 (vq)->vq_nentries, (vq)->vq_free_cnt, nused, (vq)->vq_desc_head_idx, \
638 (vq)->vq_split.ring.avail->idx, (vq)->vq_used_cons_idx, \
639 __atomic_load_n(&(vq)->vq_split.ring.used->idx, __ATOMIC_RELAXED), \
640 (vq)->vq_split.ring.avail->flags, (vq)->vq_split.ring.used->flags); \
641 } while (0)
642 #else
643 #define VIRTQUEUE_DUMP(vq) do { } while (0)
644 #endif
645
646 /* avoid write operation when necessary, to lessen cache issues */
647 #define ASSIGN_UNLESS_EQUAL(var, val) do { \
648 typeof(var) *const var_ = &(var); \
649 typeof(val) const val_ = (val); \
650 if (*var_ != val_) \
651 *var_ = val_; \
652 } while (0)
653
654 #define virtqueue_clear_net_hdr(hdr) do { \
655 typeof(hdr) hdr_ = (hdr); \
656 ASSIGN_UNLESS_EQUAL((hdr_)->csum_start, 0); \
657 ASSIGN_UNLESS_EQUAL((hdr_)->csum_offset, 0); \
658 ASSIGN_UNLESS_EQUAL((hdr_)->flags, 0); \
659 ASSIGN_UNLESS_EQUAL((hdr_)->gso_type, 0); \
660 ASSIGN_UNLESS_EQUAL((hdr_)->gso_size, 0); \
661 ASSIGN_UNLESS_EQUAL((hdr_)->hdr_len, 0); \
662 } while (0)
663
664 static inline void
virtqueue_xmit_offload(struct virtio_net_hdr * hdr,struct rte_mbuf * cookie)665 virtqueue_xmit_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *cookie)
666 {
667 uint64_t csum_l4 = cookie->ol_flags & RTE_MBUF_F_TX_L4_MASK;
668 uint16_t o_l23_len = (cookie->ol_flags & RTE_MBUF_F_TX_TUNNEL_MASK) ?
669 cookie->outer_l2_len + cookie->outer_l3_len : 0;
670
671 if (cookie->ol_flags & RTE_MBUF_F_TX_TCP_SEG)
672 csum_l4 |= RTE_MBUF_F_TX_TCP_CKSUM;
673
674 switch (csum_l4) {
675 case RTE_MBUF_F_TX_UDP_CKSUM:
676 hdr->csum_start = o_l23_len + cookie->l2_len + cookie->l3_len;
677 hdr->csum_offset = offsetof(struct rte_udp_hdr, dgram_cksum);
678 hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
679 break;
680
681 case RTE_MBUF_F_TX_TCP_CKSUM:
682 hdr->csum_start = o_l23_len + cookie->l2_len + cookie->l3_len;
683 hdr->csum_offset = offsetof(struct rte_tcp_hdr, cksum);
684 hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
685 break;
686
687 default:
688 ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
689 ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
690 ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
691 break;
692 }
693
694 /* TCP Segmentation Offload */
695 if (cookie->ol_flags & RTE_MBUF_F_TX_TCP_SEG) {
696 hdr->gso_type = (cookie->ol_flags & RTE_MBUF_F_TX_IPV6) ?
697 VIRTIO_NET_HDR_GSO_TCPV6 :
698 VIRTIO_NET_HDR_GSO_TCPV4;
699 hdr->gso_size = cookie->tso_segsz;
700 hdr->hdr_len = o_l23_len + cookie->l2_len + cookie->l3_len +
701 cookie->l4_len;
702 } else {
703 ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
704 ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
705 ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
706 }
707 }
708
709 static inline void
virtqueue_enqueue_xmit_packed(struct virtnet_tx * txvq,struct rte_mbuf * cookie,uint16_t needed,int use_indirect,int can_push,int in_order)710 virtqueue_enqueue_xmit_packed(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
711 uint16_t needed, int use_indirect, int can_push,
712 int in_order)
713 {
714 struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
715 struct vq_desc_extra *dxp;
716 struct virtqueue *vq = virtnet_txq_to_vq(txvq);
717 struct vring_packed_desc *start_dp, *head_dp;
718 uint16_t idx, id, head_idx, head_flags;
719 int16_t head_size = vq->hw->vtnet_hdr_size;
720 struct virtio_net_hdr *hdr;
721 uint16_t prev;
722 bool prepend_header = false;
723 uint16_t seg_num = cookie->nb_segs;
724
725 id = in_order ? vq->vq_avail_idx : vq->vq_desc_head_idx;
726
727 dxp = &vq->vq_descx[id];
728 dxp->ndescs = needed;
729 dxp->cookie = cookie;
730
731 head_idx = vq->vq_avail_idx;
732 idx = head_idx;
733 prev = head_idx;
734 start_dp = vq->vq_packed.ring.desc;
735
736 head_dp = &vq->vq_packed.ring.desc[idx];
737 head_flags = cookie->next ? VRING_DESC_F_NEXT : 0;
738 head_flags |= vq->vq_packed.cached_flags;
739
740 if (can_push) {
741 /* prepend cannot fail, checked by caller */
742 hdr = rte_pktmbuf_mtod_offset(cookie, struct virtio_net_hdr *,
743 -head_size);
744 prepend_header = true;
745
746 /* if offload disabled, it is not zeroed below, do it now */
747 if (!vq->hw->has_tx_offload)
748 virtqueue_clear_net_hdr(hdr);
749 } else if (use_indirect) {
750 /* setup tx ring slot to point to indirect
751 * descriptor list stored in reserved region.
752 *
753 * the first slot in indirect ring is already preset
754 * to point to the header in reserved region
755 */
756 start_dp[idx].addr = txvq->virtio_net_hdr_mem +
757 RTE_PTR_DIFF(&txr[idx].tx_packed_indir, txr);
758 start_dp[idx].len = (seg_num + 1) *
759 sizeof(struct vring_packed_desc);
760 /* Packed descriptor id needs to be restored when inorder. */
761 if (in_order)
762 start_dp[idx].id = idx;
763 /* reset flags for indirect desc */
764 head_flags = VRING_DESC_F_INDIRECT;
765 head_flags |= vq->vq_packed.cached_flags;
766 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
767
768 /* loop below will fill in rest of the indirect elements */
769 start_dp = txr[idx].tx_packed_indir;
770 idx = 1;
771 } else {
772 /* setup first tx ring slot to point to header
773 * stored in reserved region.
774 */
775 start_dp[idx].addr = txvq->virtio_net_hdr_mem +
776 RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
777 start_dp[idx].len = vq->hw->vtnet_hdr_size;
778 hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
779 idx++;
780 if (idx >= vq->vq_nentries) {
781 idx -= vq->vq_nentries;
782 vq->vq_packed.cached_flags ^=
783 VRING_PACKED_DESC_F_AVAIL_USED;
784 }
785 }
786
787 if (vq->hw->has_tx_offload)
788 virtqueue_xmit_offload(hdr, cookie);
789
790 do {
791 uint16_t flags;
792
793 start_dp[idx].addr = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
794 start_dp[idx].len = cookie->data_len;
795 if (prepend_header) {
796 start_dp[idx].addr -= head_size;
797 start_dp[idx].len += head_size;
798 prepend_header = false;
799 }
800
801 if (likely(idx != head_idx)) {
802 flags = cookie->next ? VRING_DESC_F_NEXT : 0;
803 flags |= vq->vq_packed.cached_flags;
804 start_dp[idx].flags = flags;
805 }
806 prev = idx;
807 idx++;
808 if (idx >= vq->vq_nentries) {
809 idx -= vq->vq_nentries;
810 vq->vq_packed.cached_flags ^=
811 VRING_PACKED_DESC_F_AVAIL_USED;
812 }
813 } while ((cookie = cookie->next) != NULL);
814
815 start_dp[prev].id = id;
816
817 if (use_indirect) {
818 idx = head_idx;
819 if (++idx >= vq->vq_nentries) {
820 idx -= vq->vq_nentries;
821 vq->vq_packed.cached_flags ^=
822 VRING_PACKED_DESC_F_AVAIL_USED;
823 }
824 }
825
826 vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
827 vq->vq_avail_idx = idx;
828
829 if (!in_order) {
830 vq->vq_desc_head_idx = dxp->next;
831 if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
832 vq->vq_desc_tail_idx = VQ_RING_DESC_CHAIN_END;
833 }
834
835 virtqueue_store_flags_packed(head_dp, head_flags,
836 vq->hw->weak_barriers);
837 }
838
839 static void
vq_ring_free_id_packed(struct virtqueue * vq,uint16_t id)840 vq_ring_free_id_packed(struct virtqueue *vq, uint16_t id)
841 {
842 struct vq_desc_extra *dxp;
843
844 dxp = &vq->vq_descx[id];
845 vq->vq_free_cnt += dxp->ndescs;
846
847 if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END)
848 vq->vq_desc_head_idx = id;
849 else
850 vq->vq_descx[vq->vq_desc_tail_idx].next = id;
851
852 vq->vq_desc_tail_idx = id;
853 dxp->next = VQ_RING_DESC_CHAIN_END;
854 }
855
856 static void
virtio_xmit_cleanup_inorder_packed(struct virtqueue * vq,uint16_t num)857 virtio_xmit_cleanup_inorder_packed(struct virtqueue *vq, uint16_t num)
858 {
859 uint16_t used_idx, id, curr_id, free_cnt = 0;
860 uint16_t size = vq->vq_nentries;
861 struct vring_packed_desc *desc = vq->vq_packed.ring.desc;
862 struct vq_desc_extra *dxp;
863 int nb = num;
864
865 used_idx = vq->vq_used_cons_idx;
866 /* desc_is_used has a load-acquire or rte_io_rmb inside
867 * and wait for used desc in virtqueue.
868 */
869 while (nb > 0 && desc_is_used(&desc[used_idx], vq)) {
870 id = desc[used_idx].id;
871 do {
872 curr_id = used_idx;
873 dxp = &vq->vq_descx[used_idx];
874 used_idx += dxp->ndescs;
875 free_cnt += dxp->ndescs;
876 nb -= dxp->ndescs;
877 if (used_idx >= size) {
878 used_idx -= size;
879 vq->vq_packed.used_wrap_counter ^= 1;
880 }
881 if (dxp->cookie != NULL) {
882 rte_pktmbuf_free(dxp->cookie);
883 dxp->cookie = NULL;
884 }
885 } while (curr_id != id);
886 }
887 vq->vq_used_cons_idx = used_idx;
888 vq->vq_free_cnt += free_cnt;
889 }
890
891 static void
virtio_xmit_cleanup_normal_packed(struct virtqueue * vq,uint16_t num)892 virtio_xmit_cleanup_normal_packed(struct virtqueue *vq, uint16_t num)
893 {
894 uint16_t used_idx, id;
895 uint16_t size = vq->vq_nentries;
896 struct vring_packed_desc *desc = vq->vq_packed.ring.desc;
897 struct vq_desc_extra *dxp;
898
899 used_idx = vq->vq_used_cons_idx;
900 /* desc_is_used has a load-acquire or rte_io_rmb inside
901 * and wait for used desc in virtqueue.
902 */
903 while (num-- && desc_is_used(&desc[used_idx], vq)) {
904 id = desc[used_idx].id;
905 dxp = &vq->vq_descx[id];
906 vq->vq_used_cons_idx += dxp->ndescs;
907 if (vq->vq_used_cons_idx >= size) {
908 vq->vq_used_cons_idx -= size;
909 vq->vq_packed.used_wrap_counter ^= 1;
910 }
911 vq_ring_free_id_packed(vq, id);
912 if (dxp->cookie != NULL) {
913 rte_pktmbuf_free(dxp->cookie);
914 dxp->cookie = NULL;
915 }
916 used_idx = vq->vq_used_cons_idx;
917 }
918 }
919
920 /* Cleanup from completed transmits. */
921 static inline void
virtio_xmit_cleanup_packed(struct virtqueue * vq,uint16_t num,int in_order)922 virtio_xmit_cleanup_packed(struct virtqueue *vq, uint16_t num, int in_order)
923 {
924 if (in_order)
925 virtio_xmit_cleanup_inorder_packed(vq, num);
926 else
927 virtio_xmit_cleanup_normal_packed(vq, num);
928 }
929
930 static inline void
virtio_xmit_cleanup(struct virtqueue * vq,uint16_t num)931 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
932 {
933 uint16_t i, used_idx, desc_idx;
934 for (i = 0; i < num; i++) {
935 struct vring_used_elem *uep;
936 struct vq_desc_extra *dxp;
937
938 used_idx = (uint16_t)(vq->vq_used_cons_idx &
939 (vq->vq_nentries - 1));
940 uep = &vq->vq_split.ring.used->ring[used_idx];
941
942 desc_idx = (uint16_t)uep->id;
943 dxp = &vq->vq_descx[desc_idx];
944 vq->vq_used_cons_idx++;
945 vq_ring_free_chain(vq, desc_idx);
946
947 if (dxp->cookie != NULL) {
948 rte_pktmbuf_free(dxp->cookie);
949 dxp->cookie = NULL;
950 }
951 }
952 }
953
954 /* Cleanup from completed inorder transmits. */
955 static __rte_always_inline void
virtio_xmit_cleanup_inorder(struct virtqueue * vq,uint16_t num)956 virtio_xmit_cleanup_inorder(struct virtqueue *vq, uint16_t num)
957 {
958 uint16_t i, idx = vq->vq_used_cons_idx;
959 int16_t free_cnt = 0;
960 struct vq_desc_extra *dxp = NULL;
961
962 if (unlikely(num == 0))
963 return;
964
965 for (i = 0; i < num; i++) {
966 dxp = &vq->vq_descx[idx++ & (vq->vq_nentries - 1)];
967 free_cnt += dxp->ndescs;
968 if (dxp->cookie != NULL) {
969 rte_pktmbuf_free(dxp->cookie);
970 dxp->cookie = NULL;
971 }
972 }
973
974 vq->vq_free_cnt += free_cnt;
975 vq->vq_used_cons_idx = idx;
976 }
977 #endif /* _VIRTQUEUE_H_ */
978