xref: /dpdk/drivers/net/virtio/virtqueue.h (revision 68a03efeed657e6e05f281479b33b51102797e15)
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
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
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
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
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
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 #define VTNET_SQ_RQ_QUEUE_IDX 0
117 #define VTNET_SQ_TQ_QUEUE_IDX 1
118 #define VTNET_SQ_CQ_QUEUE_IDX 2
119 
120 enum { VTNET_RQ = 0, VTNET_TQ = 1, VTNET_CQ = 2 };
121 /**
122  * The maximum virtqueue size is 2^15. Use that value as the end of
123  * descriptor chain terminator since it will never be a valid index
124  * in the descriptor table. This is used to verify we are correctly
125  * handling vq_free_cnt.
126  */
127 #define VQ_RING_DESC_CHAIN_END 32768
128 
129 /**
130  * Control the RX mode, ie. promiscuous, allmulti, etc...
131  * All commands require an "out" sg entry containing a 1 byte
132  * state value, zero = disable, non-zero = enable.  Commands
133  * 0 and 1 are supported with the VIRTIO_NET_F_CTRL_RX feature.
134  * Commands 2-5 are added with VIRTIO_NET_F_CTRL_RX_EXTRA.
135  */
136 #define VIRTIO_NET_CTRL_RX              0
137 #define VIRTIO_NET_CTRL_RX_PROMISC      0
138 #define VIRTIO_NET_CTRL_RX_ALLMULTI     1
139 #define VIRTIO_NET_CTRL_RX_ALLUNI       2
140 #define VIRTIO_NET_CTRL_RX_NOMULTI      3
141 #define VIRTIO_NET_CTRL_RX_NOUNI        4
142 #define VIRTIO_NET_CTRL_RX_NOBCAST      5
143 
144 /**
145  * Control the MAC
146  *
147  * The MAC filter table is managed by the hypervisor, the guest should
148  * assume the size is infinite.  Filtering should be considered
149  * non-perfect, ie. based on hypervisor resources, the guest may
150  * received packets from sources not specified in the filter list.
151  *
152  * In addition to the class/cmd header, the TABLE_SET command requires
153  * two out scatterlists.  Each contains a 4 byte count of entries followed
154  * by a concatenated byte stream of the ETH_ALEN MAC addresses.  The
155  * first sg list contains unicast addresses, the second is for multicast.
156  * This functionality is present if the VIRTIO_NET_F_CTRL_RX feature
157  * is available.
158  *
159  * The ADDR_SET command requests one out scatterlist, it contains a
160  * 6 bytes MAC address. This functionality is present if the
161  * VIRTIO_NET_F_CTRL_MAC_ADDR feature is available.
162  */
163 struct virtio_net_ctrl_mac {
164 	uint32_t entries;
165 	uint8_t macs[][RTE_ETHER_ADDR_LEN];
166 } __rte_packed;
167 
168 #define VIRTIO_NET_CTRL_MAC    1
169 #define VIRTIO_NET_CTRL_MAC_TABLE_SET        0
170 #define VIRTIO_NET_CTRL_MAC_ADDR_SET         1
171 
172 /**
173  * Control VLAN filtering
174  *
175  * The VLAN filter table is controlled via a simple ADD/DEL interface.
176  * VLAN IDs not added may be filtered by the hypervisor.  Del is the
177  * opposite of add.  Both commands expect an out entry containing a 2
178  * byte VLAN ID.  VLAN filtering is available with the
179  * VIRTIO_NET_F_CTRL_VLAN feature bit.
180  */
181 #define VIRTIO_NET_CTRL_VLAN     2
182 #define VIRTIO_NET_CTRL_VLAN_ADD 0
183 #define VIRTIO_NET_CTRL_VLAN_DEL 1
184 
185 /*
186  * Control link announce acknowledgement
187  *
188  * The command VIRTIO_NET_CTRL_ANNOUNCE_ACK is used to indicate that
189  * driver has recevied the notification; device would clear the
190  * VIRTIO_NET_S_ANNOUNCE bit in the status field after it receives
191  * this command.
192  */
193 #define VIRTIO_NET_CTRL_ANNOUNCE     3
194 #define VIRTIO_NET_CTRL_ANNOUNCE_ACK 0
195 
196 struct virtio_net_ctrl_hdr {
197 	uint8_t class;
198 	uint8_t cmd;
199 } __rte_packed;
200 
201 typedef uint8_t virtio_net_ctrl_ack;
202 
203 #define VIRTIO_NET_OK     0
204 #define VIRTIO_NET_ERR    1
205 
206 #define VIRTIO_MAX_CTRL_DATA 2048
207 
208 struct virtio_pmd_ctrl {
209 	struct virtio_net_ctrl_hdr hdr;
210 	virtio_net_ctrl_ack status;
211 	uint8_t data[VIRTIO_MAX_CTRL_DATA];
212 };
213 
214 struct vq_desc_extra {
215 	void *cookie;
216 	uint16_t ndescs;
217 	uint16_t next;
218 };
219 
220 #define virtnet_rxq_to_vq(rxvq) container_of(rxvq, struct virtqueue, rxq)
221 #define virtnet_txq_to_vq(txvq) container_of(txvq, struct virtqueue, txq)
222 #define virtnet_cq_to_vq(cvq) container_of(cvq, struct virtqueue, cq)
223 
224 struct virtqueue {
225 	struct virtio_hw  *hw; /**< virtio_hw structure pointer. */
226 	union {
227 		struct {
228 			/**< vring keeping desc, used and avail */
229 			struct vring ring;
230 		} vq_split;
231 
232 		struct {
233 			/**< vring keeping descs and events */
234 			struct vring_packed ring;
235 			bool used_wrap_counter;
236 			uint16_t cached_flags; /**< cached flags for descs */
237 			uint16_t event_flags_shadow;
238 		} vq_packed;
239 	};
240 
241 	uint16_t vq_used_cons_idx; /**< last consumed descriptor */
242 	uint16_t vq_nentries;  /**< vring desc numbers */
243 	uint16_t vq_free_cnt;  /**< num of desc available */
244 	uint16_t vq_avail_idx; /**< sync until needed */
245 	uint16_t vq_free_thresh; /**< free threshold */
246 
247 	/**
248 	 * Head of the free chain in the descriptor table. If
249 	 * there are no free descriptors, this will be set to
250 	 * VQ_RING_DESC_CHAIN_END.
251 	 */
252 	uint16_t  vq_desc_head_idx;
253 	uint16_t  vq_desc_tail_idx;
254 	uint16_t  vq_queue_index;   /**< PCI queue index */
255 
256 	void *vq_ring_virt_mem;  /**< linear address of vring*/
257 	unsigned int vq_ring_size;
258 
259 	union {
260 		struct virtnet_rx rxq;
261 		struct virtnet_tx txq;
262 		struct virtnet_ctl cq;
263 	};
264 
265 	rte_iova_t vq_ring_mem; /**< physical address of vring,
266 	                         * or virtual address for virtio_user. */
267 
268 	uint16_t  *notify_addr;
269 	struct rte_mbuf **sw_ring;  /**< RX software ring. */
270 	struct vq_desc_extra vq_descx[0];
271 };
272 
273 /* If multiqueue is provided by host, then we suppport it. */
274 #define VIRTIO_NET_CTRL_MQ   4
275 #define VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET        0
276 #define VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN        1
277 #define VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX        0x8000
278 
279 /**
280  * This is the first element of the scatter-gather list.  If you don't
281  * specify GSO or CSUM features, you can simply ignore the header.
282  */
283 struct virtio_net_hdr {
284 #define VIRTIO_NET_HDR_F_NEEDS_CSUM 1    /**< Use csum_start,csum_offset*/
285 #define VIRTIO_NET_HDR_F_DATA_VALID 2    /**< Checksum is valid */
286 	uint8_t flags;
287 #define VIRTIO_NET_HDR_GSO_NONE     0    /**< Not a GSO frame */
288 #define VIRTIO_NET_HDR_GSO_TCPV4    1    /**< GSO frame, IPv4 TCP (TSO) */
289 #define VIRTIO_NET_HDR_GSO_UDP      3    /**< GSO frame, IPv4 UDP (UFO) */
290 #define VIRTIO_NET_HDR_GSO_TCPV6    4    /**< GSO frame, IPv6 TCP */
291 #define VIRTIO_NET_HDR_GSO_ECN      0x80 /**< TCP has ECN set */
292 	uint8_t gso_type;
293 	uint16_t hdr_len;     /**< Ethernet + IP + tcp/udp hdrs */
294 	uint16_t gso_size;    /**< Bytes to append to hdr_len per frame */
295 	uint16_t csum_start;  /**< Position to start checksumming from */
296 	uint16_t csum_offset; /**< Offset after that to place checksum */
297 };
298 
299 /**
300  * This is the version of the header to use when the MRG_RXBUF
301  * feature has been negotiated.
302  */
303 struct virtio_net_hdr_mrg_rxbuf {
304 	struct   virtio_net_hdr hdr;
305 	uint16_t num_buffers; /**< Number of merged rx buffers */
306 };
307 
308 /* Region reserved to allow for transmit header and indirect ring */
309 #define VIRTIO_MAX_TX_INDIRECT 8
310 struct virtio_tx_region {
311 	struct virtio_net_hdr_mrg_rxbuf tx_hdr;
312 	union {
313 		struct vring_desc tx_indir[VIRTIO_MAX_TX_INDIRECT];
314 		struct vring_packed_desc
315 			tx_packed_indir[VIRTIO_MAX_TX_INDIRECT];
316 	} __rte_aligned(16);
317 };
318 
319 static inline int
320 desc_is_used(struct vring_packed_desc *desc, struct virtqueue *vq)
321 {
322 	uint16_t used, avail, flags;
323 
324 	flags = virtqueue_fetch_flags_packed(desc, vq->hw->weak_barriers);
325 	used = !!(flags & VRING_PACKED_DESC_F_USED);
326 	avail = !!(flags & VRING_PACKED_DESC_F_AVAIL);
327 
328 	return avail == used && used == vq->vq_packed.used_wrap_counter;
329 }
330 
331 static inline void
332 vring_desc_init_packed(struct virtqueue *vq, int n)
333 {
334 	int i;
335 	for (i = 0; i < n - 1; i++) {
336 		vq->vq_packed.ring.desc[i].id = i;
337 		vq->vq_descx[i].next = i + 1;
338 	}
339 	vq->vq_packed.ring.desc[i].id = i;
340 	vq->vq_descx[i].next = VQ_RING_DESC_CHAIN_END;
341 }
342 
343 /* Chain all the descriptors in the ring with an END */
344 static inline void
345 vring_desc_init_split(struct vring_desc *dp, uint16_t n)
346 {
347 	uint16_t i;
348 
349 	for (i = 0; i < n - 1; i++)
350 		dp[i].next = (uint16_t)(i + 1);
351 	dp[i].next = VQ_RING_DESC_CHAIN_END;
352 }
353 
354 static inline void
355 vring_desc_init_indirect_packed(struct vring_packed_desc *dp, int n)
356 {
357 	int i;
358 	for (i = 0; i < n; i++) {
359 		dp[i].id = (uint16_t)i;
360 		dp[i].flags = VRING_DESC_F_WRITE;
361 	}
362 }
363 
364 /**
365  * Tell the backend not to interrupt us. Implementation for packed virtqueues.
366  */
367 static inline void
368 virtqueue_disable_intr_packed(struct virtqueue *vq)
369 {
370 	if (vq->vq_packed.event_flags_shadow != RING_EVENT_FLAGS_DISABLE) {
371 		vq->vq_packed.event_flags_shadow = RING_EVENT_FLAGS_DISABLE;
372 		vq->vq_packed.ring.driver->desc_event_flags =
373 			vq->vq_packed.event_flags_shadow;
374 	}
375 }
376 
377 /**
378  * Tell the backend not to interrupt us. Implementation for split virtqueues.
379  */
380 static inline void
381 virtqueue_disable_intr_split(struct virtqueue *vq)
382 {
383 	vq->vq_split.ring.avail->flags |= VRING_AVAIL_F_NO_INTERRUPT;
384 }
385 
386 /**
387  * Tell the backend not to interrupt us.
388  */
389 static inline void
390 virtqueue_disable_intr(struct virtqueue *vq)
391 {
392 	if (virtio_with_packed_queue(vq->hw))
393 		virtqueue_disable_intr_packed(vq);
394 	else
395 		virtqueue_disable_intr_split(vq);
396 }
397 
398 /**
399  * Tell the backend to interrupt. Implementation for packed virtqueues.
400  */
401 static inline void
402 virtqueue_enable_intr_packed(struct virtqueue *vq)
403 {
404 	if (vq->vq_packed.event_flags_shadow == RING_EVENT_FLAGS_DISABLE) {
405 		vq->vq_packed.event_flags_shadow = RING_EVENT_FLAGS_ENABLE;
406 		vq->vq_packed.ring.driver->desc_event_flags =
407 			vq->vq_packed.event_flags_shadow;
408 	}
409 }
410 
411 /**
412  * Tell the backend to interrupt. Implementation for split virtqueues.
413  */
414 static inline void
415 virtqueue_enable_intr_split(struct virtqueue *vq)
416 {
417 	vq->vq_split.ring.avail->flags &= (~VRING_AVAIL_F_NO_INTERRUPT);
418 }
419 
420 /**
421  * Tell the backend to interrupt us.
422  */
423 static inline void
424 virtqueue_enable_intr(struct virtqueue *vq)
425 {
426 	if (virtio_with_packed_queue(vq->hw))
427 		virtqueue_enable_intr_packed(vq);
428 	else
429 		virtqueue_enable_intr_split(vq);
430 }
431 
432 /**
433  *  Dump virtqueue internal structures, for debug purpose only.
434  */
435 void virtqueue_dump(struct virtqueue *vq);
436 /**
437  *  Get all mbufs to be freed.
438  */
439 struct rte_mbuf *virtqueue_detach_unused(struct virtqueue *vq);
440 
441 /* Flush the elements in the used ring. */
442 void virtqueue_rxvq_flush(struct virtqueue *vq);
443 
444 int virtqueue_rxvq_reset_packed(struct virtqueue *vq);
445 
446 int virtqueue_txvq_reset_packed(struct virtqueue *vq);
447 
448 static inline int
449 virtqueue_full(const struct virtqueue *vq)
450 {
451 	return vq->vq_free_cnt == 0;
452 }
453 
454 static inline int
455 virtio_get_queue_type(struct virtio_hw *hw, uint16_t vq_idx)
456 {
457 	if (vq_idx == hw->max_queue_pairs * 2)
458 		return VTNET_CQ;
459 	else if (vq_idx % 2 == 0)
460 		return VTNET_RQ;
461 	else
462 		return VTNET_TQ;
463 }
464 
465 /* virtqueue_nused has load-acquire or rte_io_rmb insed */
466 static inline uint16_t
467 virtqueue_nused(const struct virtqueue *vq)
468 {
469 	uint16_t idx;
470 
471 	if (vq->hw->weak_barriers) {
472 	/**
473 	 * x86 prefers to using rte_smp_rmb over __atomic_load_n as it
474 	 * reports a slightly better perf, which comes from the saved
475 	 * branch by the compiler.
476 	 * The if and else branches are identical with the smp and io
477 	 * barriers both defined as compiler barriers on x86.
478 	 */
479 #ifdef RTE_ARCH_X86_64
480 		idx = vq->vq_split.ring.used->idx;
481 		rte_smp_rmb();
482 #else
483 		idx = __atomic_load_n(&(vq)->vq_split.ring.used->idx,
484 				__ATOMIC_ACQUIRE);
485 #endif
486 	} else {
487 		idx = vq->vq_split.ring.used->idx;
488 		rte_io_rmb();
489 	}
490 	return idx - vq->vq_used_cons_idx;
491 }
492 
493 void vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx);
494 void vq_ring_free_chain_packed(struct virtqueue *vq, uint16_t used_idx);
495 void vq_ring_free_inorder(struct virtqueue *vq, uint16_t desc_idx,
496 			  uint16_t num);
497 
498 static inline void
499 vq_update_avail_idx(struct virtqueue *vq)
500 {
501 	if (vq->hw->weak_barriers) {
502 	/* x86 prefers to using rte_smp_wmb over __atomic_store_n as
503 	 * it reports a slightly better perf, which comes from the
504 	 * saved branch by the compiler.
505 	 * The if and else branches are identical with the smp and
506 	 * io barriers both defined as compiler barriers on x86.
507 	 */
508 #ifdef RTE_ARCH_X86_64
509 		rte_smp_wmb();
510 		vq->vq_split.ring.avail->idx = vq->vq_avail_idx;
511 #else
512 		__atomic_store_n(&vq->vq_split.ring.avail->idx,
513 				 vq->vq_avail_idx, __ATOMIC_RELEASE);
514 #endif
515 	} else {
516 		rte_io_wmb();
517 		vq->vq_split.ring.avail->idx = vq->vq_avail_idx;
518 	}
519 }
520 
521 static inline void
522 vq_update_avail_ring(struct virtqueue *vq, uint16_t desc_idx)
523 {
524 	uint16_t avail_idx;
525 	/*
526 	 * Place the head of the descriptor chain into the next slot and make
527 	 * it usable to the host. The chain is made available now rather than
528 	 * deferring to virtqueue_notify() in the hopes that if the host is
529 	 * currently running on another CPU, we can keep it processing the new
530 	 * descriptor.
531 	 */
532 	avail_idx = (uint16_t)(vq->vq_avail_idx & (vq->vq_nentries - 1));
533 	if (unlikely(vq->vq_split.ring.avail->ring[avail_idx] != desc_idx))
534 		vq->vq_split.ring.avail->ring[avail_idx] = desc_idx;
535 	vq->vq_avail_idx++;
536 }
537 
538 static inline int
539 virtqueue_kick_prepare(struct virtqueue *vq)
540 {
541 	/*
542 	 * Ensure updated avail->idx is visible to vhost before reading
543 	 * the used->flags.
544 	 */
545 	virtio_mb(vq->hw->weak_barriers);
546 	return !(vq->vq_split.ring.used->flags & VRING_USED_F_NO_NOTIFY);
547 }
548 
549 static inline int
550 virtqueue_kick_prepare_packed(struct virtqueue *vq)
551 {
552 	uint16_t flags;
553 
554 	/*
555 	 * Ensure updated data is visible to vhost before reading the flags.
556 	 */
557 	virtio_mb(vq->hw->weak_barriers);
558 	flags = vq->vq_packed.ring.device->desc_event_flags;
559 
560 	return flags != RING_EVENT_FLAGS_DISABLE;
561 }
562 
563 /*
564  * virtqueue_kick_prepare*() or the virtio_wmb() should be called
565  * before this function to be sure that all the data is visible to vhost.
566  */
567 static inline void
568 virtqueue_notify(struct virtqueue *vq)
569 {
570 	VIRTIO_OPS(vq->hw)->notify_queue(vq->hw, vq);
571 }
572 
573 #ifdef RTE_LIBRTE_VIRTIO_DEBUG_DUMP
574 #define VIRTQUEUE_DUMP(vq) do { \
575 	uint16_t used_idx, nused; \
576 	used_idx = __atomic_load_n(&(vq)->vq_split.ring.used->idx, \
577 				   __ATOMIC_RELAXED); \
578 	nused = (uint16_t)(used_idx - (vq)->vq_used_cons_idx); \
579 	if (virtio_with_packed_queue((vq)->hw)) { \
580 		PMD_INIT_LOG(DEBUG, \
581 		"VQ: - size=%d; free=%d; used_cons_idx=%d; avail_idx=%d;" \
582 		" cached_flags=0x%x; used_wrap_counter=%d", \
583 		(vq)->vq_nentries, (vq)->vq_free_cnt, (vq)->vq_used_cons_idx, \
584 		(vq)->vq_avail_idx, (vq)->vq_packed.cached_flags, \
585 		(vq)->vq_packed.used_wrap_counter); \
586 		break; \
587 	} \
588 	PMD_INIT_LOG(DEBUG, \
589 	  "VQ: - size=%d; free=%d; used=%d; desc_head_idx=%d;" \
590 	  " avail.idx=%d; used_cons_idx=%d; used.idx=%d;" \
591 	  " avail.flags=0x%x; used.flags=0x%x", \
592 	  (vq)->vq_nentries, (vq)->vq_free_cnt, nused, (vq)->vq_desc_head_idx, \
593 	  (vq)->vq_split.ring.avail->idx, (vq)->vq_used_cons_idx, \
594 	  __atomic_load_n(&(vq)->vq_split.ring.used->idx, __ATOMIC_RELAXED), \
595 	  (vq)->vq_split.ring.avail->flags, (vq)->vq_split.ring.used->flags); \
596 } while (0)
597 #else
598 #define VIRTQUEUE_DUMP(vq) do { } while (0)
599 #endif
600 
601 /* avoid write operation when necessary, to lessen cache issues */
602 #define ASSIGN_UNLESS_EQUAL(var, val) do {	\
603 	typeof(var) *const var_ = &(var);	\
604 	typeof(val)  const val_ = (val);	\
605 	if (*var_ != val_)			\
606 		*var_ = val_;			\
607 } while (0)
608 
609 #define virtqueue_clear_net_hdr(hdr) do {		\
610 	typeof(hdr) hdr_ = (hdr);			\
611 	ASSIGN_UNLESS_EQUAL((hdr_)->csum_start, 0);	\
612 	ASSIGN_UNLESS_EQUAL((hdr_)->csum_offset, 0);	\
613 	ASSIGN_UNLESS_EQUAL((hdr_)->flags, 0);		\
614 	ASSIGN_UNLESS_EQUAL((hdr_)->gso_type, 0);	\
615 	ASSIGN_UNLESS_EQUAL((hdr_)->gso_size, 0);	\
616 	ASSIGN_UNLESS_EQUAL((hdr_)->hdr_len, 0);	\
617 } while (0)
618 
619 static inline void
620 virtqueue_xmit_offload(struct virtio_net_hdr *hdr,
621 			struct rte_mbuf *cookie,
622 			uint8_t offload)
623 {
624 	if (offload) {
625 		if (cookie->ol_flags & PKT_TX_TCP_SEG)
626 			cookie->ol_flags |= PKT_TX_TCP_CKSUM;
627 
628 		switch (cookie->ol_flags & PKT_TX_L4_MASK) {
629 		case PKT_TX_UDP_CKSUM:
630 			hdr->csum_start = cookie->l2_len + cookie->l3_len;
631 			hdr->csum_offset = offsetof(struct rte_udp_hdr,
632 				dgram_cksum);
633 			hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
634 			break;
635 
636 		case PKT_TX_TCP_CKSUM:
637 			hdr->csum_start = cookie->l2_len + cookie->l3_len;
638 			hdr->csum_offset = offsetof(struct rte_tcp_hdr, cksum);
639 			hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
640 			break;
641 
642 		default:
643 			ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
644 			ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
645 			ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
646 			break;
647 		}
648 
649 		/* TCP Segmentation Offload */
650 		if (cookie->ol_flags & PKT_TX_TCP_SEG) {
651 			hdr->gso_type = (cookie->ol_flags & PKT_TX_IPV6) ?
652 				VIRTIO_NET_HDR_GSO_TCPV6 :
653 				VIRTIO_NET_HDR_GSO_TCPV4;
654 			hdr->gso_size = cookie->tso_segsz;
655 			hdr->hdr_len =
656 				cookie->l2_len +
657 				cookie->l3_len +
658 				cookie->l4_len;
659 		} else {
660 			ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
661 			ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
662 			ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
663 		}
664 	}
665 }
666 
667 static inline void
668 virtqueue_enqueue_xmit_packed(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
669 			      uint16_t needed, int use_indirect, int can_push,
670 			      int in_order)
671 {
672 	struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
673 	struct vq_desc_extra *dxp;
674 	struct virtqueue *vq = virtnet_txq_to_vq(txvq);
675 	struct vring_packed_desc *start_dp, *head_dp;
676 	uint16_t idx, id, head_idx, head_flags;
677 	int16_t head_size = vq->hw->vtnet_hdr_size;
678 	struct virtio_net_hdr *hdr;
679 	uint16_t prev;
680 	bool prepend_header = false;
681 	uint16_t seg_num = cookie->nb_segs;
682 
683 	id = in_order ? vq->vq_avail_idx : vq->vq_desc_head_idx;
684 
685 	dxp = &vq->vq_descx[id];
686 	dxp->ndescs = needed;
687 	dxp->cookie = cookie;
688 
689 	head_idx = vq->vq_avail_idx;
690 	idx = head_idx;
691 	prev = head_idx;
692 	start_dp = vq->vq_packed.ring.desc;
693 
694 	head_dp = &vq->vq_packed.ring.desc[idx];
695 	head_flags = cookie->next ? VRING_DESC_F_NEXT : 0;
696 	head_flags |= vq->vq_packed.cached_flags;
697 
698 	if (can_push) {
699 		/* prepend cannot fail, checked by caller */
700 		hdr = rte_pktmbuf_mtod_offset(cookie, struct virtio_net_hdr *,
701 					      -head_size);
702 		prepend_header = true;
703 
704 		/* if offload disabled, it is not zeroed below, do it now */
705 		if (!vq->hw->has_tx_offload)
706 			virtqueue_clear_net_hdr(hdr);
707 	} else if (use_indirect) {
708 		/* setup tx ring slot to point to indirect
709 		 * descriptor list stored in reserved region.
710 		 *
711 		 * the first slot in indirect ring is already preset
712 		 * to point to the header in reserved region
713 		 */
714 		start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
715 			RTE_PTR_DIFF(&txr[idx].tx_packed_indir, txr);
716 		start_dp[idx].len   = (seg_num + 1) *
717 			sizeof(struct vring_packed_desc);
718 		/* reset flags for indirect desc */
719 		head_flags = VRING_DESC_F_INDIRECT;
720 		head_flags |= vq->vq_packed.cached_flags;
721 		hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
722 
723 		/* loop below will fill in rest of the indirect elements */
724 		start_dp = txr[idx].tx_packed_indir;
725 		idx = 1;
726 	} else {
727 		/* setup first tx ring slot to point to header
728 		 * stored in reserved region.
729 		 */
730 		start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
731 			RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
732 		start_dp[idx].len   = vq->hw->vtnet_hdr_size;
733 		hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
734 		idx++;
735 		if (idx >= vq->vq_nentries) {
736 			idx -= vq->vq_nentries;
737 			vq->vq_packed.cached_flags ^=
738 				VRING_PACKED_DESC_F_AVAIL_USED;
739 		}
740 	}
741 
742 	virtqueue_xmit_offload(hdr, cookie, vq->hw->has_tx_offload);
743 
744 	do {
745 		uint16_t flags;
746 
747 		start_dp[idx].addr = rte_mbuf_data_iova(cookie);
748 		start_dp[idx].len  = cookie->data_len;
749 		if (prepend_header) {
750 			start_dp[idx].addr -= head_size;
751 			start_dp[idx].len += head_size;
752 			prepend_header = false;
753 		}
754 
755 		if (likely(idx != head_idx)) {
756 			flags = cookie->next ? VRING_DESC_F_NEXT : 0;
757 			flags |= vq->vq_packed.cached_flags;
758 			start_dp[idx].flags = flags;
759 		}
760 		prev = idx;
761 		idx++;
762 		if (idx >= vq->vq_nentries) {
763 			idx -= vq->vq_nentries;
764 			vq->vq_packed.cached_flags ^=
765 				VRING_PACKED_DESC_F_AVAIL_USED;
766 		}
767 	} while ((cookie = cookie->next) != NULL);
768 
769 	start_dp[prev].id = id;
770 
771 	if (use_indirect) {
772 		idx = head_idx;
773 		if (++idx >= vq->vq_nentries) {
774 			idx -= vq->vq_nentries;
775 			vq->vq_packed.cached_flags ^=
776 				VRING_PACKED_DESC_F_AVAIL_USED;
777 		}
778 	}
779 
780 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
781 	vq->vq_avail_idx = idx;
782 
783 	if (!in_order) {
784 		vq->vq_desc_head_idx = dxp->next;
785 		if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
786 			vq->vq_desc_tail_idx = VQ_RING_DESC_CHAIN_END;
787 	}
788 
789 	virtqueue_store_flags_packed(head_dp, head_flags,
790 				     vq->hw->weak_barriers);
791 }
792 
793 static void
794 vq_ring_free_id_packed(struct virtqueue *vq, uint16_t id)
795 {
796 	struct vq_desc_extra *dxp;
797 
798 	dxp = &vq->vq_descx[id];
799 	vq->vq_free_cnt += dxp->ndescs;
800 
801 	if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END)
802 		vq->vq_desc_head_idx = id;
803 	else
804 		vq->vq_descx[vq->vq_desc_tail_idx].next = id;
805 
806 	vq->vq_desc_tail_idx = id;
807 	dxp->next = VQ_RING_DESC_CHAIN_END;
808 }
809 
810 static void
811 virtio_xmit_cleanup_inorder_packed(struct virtqueue *vq, int num)
812 {
813 	uint16_t used_idx, id, curr_id, free_cnt = 0;
814 	uint16_t size = vq->vq_nentries;
815 	struct vring_packed_desc *desc = vq->vq_packed.ring.desc;
816 	struct vq_desc_extra *dxp;
817 
818 	used_idx = vq->vq_used_cons_idx;
819 	/* desc_is_used has a load-acquire or rte_io_rmb inside
820 	 * and wait for used desc in virtqueue.
821 	 */
822 	while (num > 0 && desc_is_used(&desc[used_idx], vq)) {
823 		id = desc[used_idx].id;
824 		do {
825 			curr_id = used_idx;
826 			dxp = &vq->vq_descx[used_idx];
827 			used_idx += dxp->ndescs;
828 			free_cnt += dxp->ndescs;
829 			num -= dxp->ndescs;
830 			if (used_idx >= size) {
831 				used_idx -= size;
832 				vq->vq_packed.used_wrap_counter ^= 1;
833 			}
834 			if (dxp->cookie != NULL) {
835 				rte_pktmbuf_free(dxp->cookie);
836 				dxp->cookie = NULL;
837 			}
838 		} while (curr_id != id);
839 	}
840 	vq->vq_used_cons_idx = used_idx;
841 	vq->vq_free_cnt += free_cnt;
842 }
843 
844 static void
845 virtio_xmit_cleanup_normal_packed(struct virtqueue *vq, int num)
846 {
847 	uint16_t used_idx, id;
848 	uint16_t size = vq->vq_nentries;
849 	struct vring_packed_desc *desc = vq->vq_packed.ring.desc;
850 	struct vq_desc_extra *dxp;
851 
852 	used_idx = vq->vq_used_cons_idx;
853 	/* desc_is_used has a load-acquire or rte_io_rmb inside
854 	 * and wait for used desc in virtqueue.
855 	 */
856 	while (num-- && desc_is_used(&desc[used_idx], vq)) {
857 		id = desc[used_idx].id;
858 		dxp = &vq->vq_descx[id];
859 		vq->vq_used_cons_idx += dxp->ndescs;
860 		if (vq->vq_used_cons_idx >= size) {
861 			vq->vq_used_cons_idx -= size;
862 			vq->vq_packed.used_wrap_counter ^= 1;
863 		}
864 		vq_ring_free_id_packed(vq, id);
865 		if (dxp->cookie != NULL) {
866 			rte_pktmbuf_free(dxp->cookie);
867 			dxp->cookie = NULL;
868 		}
869 		used_idx = vq->vq_used_cons_idx;
870 	}
871 }
872 
873 /* Cleanup from completed transmits. */
874 static inline void
875 virtio_xmit_cleanup_packed(struct virtqueue *vq, int num, int in_order)
876 {
877 	if (in_order)
878 		virtio_xmit_cleanup_inorder_packed(vq, num);
879 	else
880 		virtio_xmit_cleanup_normal_packed(vq, num);
881 }
882 
883 static inline void
884 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
885 {
886 	uint16_t i, used_idx, desc_idx;
887 	for (i = 0; i < num; i++) {
888 		struct vring_used_elem *uep;
889 		struct vq_desc_extra *dxp;
890 
891 		used_idx = (uint16_t)(vq->vq_used_cons_idx &
892 				(vq->vq_nentries - 1));
893 		uep = &vq->vq_split.ring.used->ring[used_idx];
894 
895 		desc_idx = (uint16_t)uep->id;
896 		dxp = &vq->vq_descx[desc_idx];
897 		vq->vq_used_cons_idx++;
898 		vq_ring_free_chain(vq, desc_idx);
899 
900 		if (dxp->cookie != NULL) {
901 			rte_pktmbuf_free(dxp->cookie);
902 			dxp->cookie = NULL;
903 		}
904 	}
905 }
906 
907 /* Cleanup from completed inorder transmits. */
908 static __rte_always_inline void
909 virtio_xmit_cleanup_inorder(struct virtqueue *vq, uint16_t num)
910 {
911 	uint16_t i, idx = vq->vq_used_cons_idx;
912 	int16_t free_cnt = 0;
913 	struct vq_desc_extra *dxp = NULL;
914 
915 	if (unlikely(num == 0))
916 		return;
917 
918 	for (i = 0; i < num; i++) {
919 		dxp = &vq->vq_descx[idx++ & (vq->vq_nentries - 1)];
920 		free_cnt += dxp->ndescs;
921 		if (dxp->cookie != NULL) {
922 			rte_pktmbuf_free(dxp->cookie);
923 			dxp->cookie = NULL;
924 		}
925 	}
926 
927 	vq->vq_free_cnt += free_cnt;
928 	vq->vq_used_cons_idx = idx;
929 }
930 #endif /* _VIRTQUEUE_H_ */
931