xref: /dpdk/drivers/net/virtio/virtqueue.h (revision 8809f78c7dd9f33a44a4f89c58fc91ded34296ed)
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_pci.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 smp_mb() et al are
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_smp_mb();
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_smp_rmb();
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_smp_wmb();
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_smp_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 with the smp and io barriers both
72  * defined as compiler barriers on x86.
73  */
74 #ifdef RTE_ARCH_X86_64
75 		flags = dp->flags;
76 		rte_smp_rmb();
77 #else
78 		flags = __atomic_load_n(&dp->flags, __ATOMIC_ACQUIRE);
79 #endif
80 	} else {
81 		flags = dp->flags;
82 		rte_io_rmb();
83 	}
84 
85 	return flags;
86 }
87 
88 static inline void
89 virtqueue_store_flags_packed(struct vring_packed_desc *dp,
90 			      uint16_t flags, uint8_t weak_barriers)
91 {
92 	if (weak_barriers) {
93 /* x86 prefers to using rte_smp_wmb over __atomic_store_n as it reports
94  * a better perf(~1.5%), which comes from the saved branch by the compiler.
95  * The if and else branch are identical with the smp and io barriers both
96  * defined as compiler barriers on x86.
97  */
98 #ifdef RTE_ARCH_X86_64
99 		rte_smp_wmb();
100 		dp->flags = flags;
101 #else
102 		__atomic_store_n(&dp->flags, flags, __ATOMIC_RELEASE);
103 #endif
104 	} else {
105 		rte_io_wmb();
106 		dp->flags = flags;
107 	}
108 }
109 #ifdef RTE_PMD_PACKET_PREFETCH
110 #define rte_packet_prefetch(p)  rte_prefetch1(p)
111 #else
112 #define rte_packet_prefetch(p)  do {} while(0)
113 #endif
114 
115 #define VIRTQUEUE_MAX_NAME_SZ 32
116 
117 #ifdef RTE_VIRTIO_USER
118 /**
119  * Return the physical address (or virtual address in case of
120  * virtio-user) of mbuf data buffer.
121  *
122  * The address is firstly casted to the word size (sizeof(uintptr_t))
123  * before casting it to uint64_t. This is to make it work with different
124  * combination of word size (64 bit and 32 bit) and virtio device
125  * (virtio-pci and virtio-user).
126  */
127 #define VIRTIO_MBUF_ADDR(mb, vq) \
128 	((uint64_t)(*(uintptr_t *)((uintptr_t)(mb) + (vq)->offset)))
129 #else
130 #define VIRTIO_MBUF_ADDR(mb, vq) ((mb)->buf_iova)
131 #endif
132 
133 /**
134  * Return the physical address (or virtual address in case of
135  * virtio-user) of mbuf data buffer, taking care of mbuf data offset
136  */
137 #define VIRTIO_MBUF_DATA_DMA_ADDR(mb, vq) \
138 	(VIRTIO_MBUF_ADDR(mb, vq) + (mb)->data_off)
139 
140 #define VTNET_SQ_RQ_QUEUE_IDX 0
141 #define VTNET_SQ_TQ_QUEUE_IDX 1
142 #define VTNET_SQ_CQ_QUEUE_IDX 2
143 
144 enum { VTNET_RQ = 0, VTNET_TQ = 1, VTNET_CQ = 2 };
145 /**
146  * The maximum virtqueue size is 2^15. Use that value as the end of
147  * descriptor chain terminator since it will never be a valid index
148  * in the descriptor table. This is used to verify we are correctly
149  * handling vq_free_cnt.
150  */
151 #define VQ_RING_DESC_CHAIN_END 32768
152 
153 /**
154  * Control the RX mode, ie. promiscuous, allmulti, etc...
155  * All commands require an "out" sg entry containing a 1 byte
156  * state value, zero = disable, non-zero = enable.  Commands
157  * 0 and 1 are supported with the VIRTIO_NET_F_CTRL_RX feature.
158  * Commands 2-5 are added with VIRTIO_NET_F_CTRL_RX_EXTRA.
159  */
160 #define VIRTIO_NET_CTRL_RX              0
161 #define VIRTIO_NET_CTRL_RX_PROMISC      0
162 #define VIRTIO_NET_CTRL_RX_ALLMULTI     1
163 #define VIRTIO_NET_CTRL_RX_ALLUNI       2
164 #define VIRTIO_NET_CTRL_RX_NOMULTI      3
165 #define VIRTIO_NET_CTRL_RX_NOUNI        4
166 #define VIRTIO_NET_CTRL_RX_NOBCAST      5
167 
168 /**
169  * Control the MAC
170  *
171  * The MAC filter table is managed by the hypervisor, the guest should
172  * assume the size is infinite.  Filtering should be considered
173  * non-perfect, ie. based on hypervisor resources, the guest may
174  * received packets from sources not specified in the filter list.
175  *
176  * In addition to the class/cmd header, the TABLE_SET command requires
177  * two out scatterlists.  Each contains a 4 byte count of entries followed
178  * by a concatenated byte stream of the ETH_ALEN MAC addresses.  The
179  * first sg list contains unicast addresses, the second is for multicast.
180  * This functionality is present if the VIRTIO_NET_F_CTRL_RX feature
181  * is available.
182  *
183  * The ADDR_SET command requests one out scatterlist, it contains a
184  * 6 bytes MAC address. This functionality is present if the
185  * VIRTIO_NET_F_CTRL_MAC_ADDR feature is available.
186  */
187 struct virtio_net_ctrl_mac {
188 	uint32_t entries;
189 	uint8_t macs[][RTE_ETHER_ADDR_LEN];
190 } __rte_packed;
191 
192 #define VIRTIO_NET_CTRL_MAC    1
193 #define VIRTIO_NET_CTRL_MAC_TABLE_SET        0
194 #define VIRTIO_NET_CTRL_MAC_ADDR_SET         1
195 
196 /**
197  * Control VLAN filtering
198  *
199  * The VLAN filter table is controlled via a simple ADD/DEL interface.
200  * VLAN IDs not added may be filtered by the hypervisor.  Del is the
201  * opposite of add.  Both commands expect an out entry containing a 2
202  * byte VLAN ID.  VLAN filtering is available with the
203  * VIRTIO_NET_F_CTRL_VLAN feature bit.
204  */
205 #define VIRTIO_NET_CTRL_VLAN     2
206 #define VIRTIO_NET_CTRL_VLAN_ADD 0
207 #define VIRTIO_NET_CTRL_VLAN_DEL 1
208 
209 /*
210  * Control link announce acknowledgement
211  *
212  * The command VIRTIO_NET_CTRL_ANNOUNCE_ACK is used to indicate that
213  * driver has recevied the notification; device would clear the
214  * VIRTIO_NET_S_ANNOUNCE bit in the status field after it receives
215  * this command.
216  */
217 #define VIRTIO_NET_CTRL_ANNOUNCE     3
218 #define VIRTIO_NET_CTRL_ANNOUNCE_ACK 0
219 
220 struct virtio_net_ctrl_hdr {
221 	uint8_t class;
222 	uint8_t cmd;
223 } __rte_packed;
224 
225 typedef uint8_t virtio_net_ctrl_ack;
226 
227 #define VIRTIO_NET_OK     0
228 #define VIRTIO_NET_ERR    1
229 
230 #define VIRTIO_MAX_CTRL_DATA 2048
231 
232 struct virtio_pmd_ctrl {
233 	struct virtio_net_ctrl_hdr hdr;
234 	virtio_net_ctrl_ack status;
235 	uint8_t data[VIRTIO_MAX_CTRL_DATA];
236 };
237 
238 struct vq_desc_extra {
239 	void *cookie;
240 	uint16_t ndescs;
241 	uint16_t next;
242 };
243 
244 struct virtqueue {
245 	struct virtio_hw  *hw; /**< virtio_hw structure pointer. */
246 	union {
247 		struct {
248 			/**< vring keeping desc, used and avail */
249 			struct vring ring;
250 		} vq_split;
251 
252 		struct {
253 			/**< vring keeping descs and events */
254 			struct vring_packed ring;
255 			bool used_wrap_counter;
256 			uint16_t cached_flags; /**< cached flags for descs */
257 			uint16_t event_flags_shadow;
258 		} vq_packed;
259 	};
260 
261 	uint16_t vq_used_cons_idx; /**< last consumed descriptor */
262 	uint16_t vq_nentries;  /**< vring desc numbers */
263 	uint16_t vq_free_cnt;  /**< num of desc available */
264 	uint16_t vq_avail_idx; /**< sync until needed */
265 	uint16_t vq_free_thresh; /**< free threshold */
266 
267 	void *vq_ring_virt_mem;  /**< linear address of vring*/
268 	unsigned int vq_ring_size;
269 
270 	union {
271 		struct virtnet_rx rxq;
272 		struct virtnet_tx txq;
273 		struct virtnet_ctl cq;
274 	};
275 
276 	rte_iova_t vq_ring_mem; /**< physical address of vring,
277 	                         * or virtual address for virtio_user. */
278 
279 	/**
280 	 * Head of the free chain in the descriptor table. If
281 	 * there are no free descriptors, this will be set to
282 	 * VQ_RING_DESC_CHAIN_END.
283 	 */
284 	uint16_t  vq_desc_head_idx;
285 	uint16_t  vq_desc_tail_idx;
286 	uint16_t  vq_queue_index;   /**< PCI queue index */
287 	uint16_t offset; /**< relative offset to obtain addr in mbuf */
288 	uint16_t  *notify_addr;
289 	struct rte_mbuf **sw_ring;  /**< RX software ring. */
290 	struct vq_desc_extra vq_descx[0];
291 };
292 
293 /* If multiqueue is provided by host, then we suppport it. */
294 #define VIRTIO_NET_CTRL_MQ   4
295 #define VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET        0
296 #define VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN        1
297 #define VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX        0x8000
298 
299 /**
300  * This is the first element of the scatter-gather list.  If you don't
301  * specify GSO or CSUM features, you can simply ignore the header.
302  */
303 struct virtio_net_hdr {
304 #define VIRTIO_NET_HDR_F_NEEDS_CSUM 1    /**< Use csum_start,csum_offset*/
305 #define VIRTIO_NET_HDR_F_DATA_VALID 2    /**< Checksum is valid */
306 	uint8_t flags;
307 #define VIRTIO_NET_HDR_GSO_NONE     0    /**< Not a GSO frame */
308 #define VIRTIO_NET_HDR_GSO_TCPV4    1    /**< GSO frame, IPv4 TCP (TSO) */
309 #define VIRTIO_NET_HDR_GSO_UDP      3    /**< GSO frame, IPv4 UDP (UFO) */
310 #define VIRTIO_NET_HDR_GSO_TCPV6    4    /**< GSO frame, IPv6 TCP */
311 #define VIRTIO_NET_HDR_GSO_ECN      0x80 /**< TCP has ECN set */
312 	uint8_t gso_type;
313 	uint16_t hdr_len;     /**< Ethernet + IP + tcp/udp hdrs */
314 	uint16_t gso_size;    /**< Bytes to append to hdr_len per frame */
315 	uint16_t csum_start;  /**< Position to start checksumming from */
316 	uint16_t csum_offset; /**< Offset after that to place checksum */
317 };
318 
319 /**
320  * This is the version of the header to use when the MRG_RXBUF
321  * feature has been negotiated.
322  */
323 struct virtio_net_hdr_mrg_rxbuf {
324 	struct   virtio_net_hdr hdr;
325 	uint16_t num_buffers; /**< Number of merged rx buffers */
326 };
327 
328 /* Region reserved to allow for transmit header and indirect ring */
329 #define VIRTIO_MAX_TX_INDIRECT 8
330 struct virtio_tx_region {
331 	struct virtio_net_hdr_mrg_rxbuf tx_hdr;
332 	union {
333 		struct vring_desc tx_indir[VIRTIO_MAX_TX_INDIRECT];
334 		struct vring_packed_desc
335 			tx_packed_indir[VIRTIO_MAX_TX_INDIRECT];
336 	} __rte_aligned(16);
337 };
338 
339 static inline int
340 desc_is_used(struct vring_packed_desc *desc, struct virtqueue *vq)
341 {
342 	uint16_t used, avail, flags;
343 
344 	flags = virtqueue_fetch_flags_packed(desc, vq->hw->weak_barriers);
345 	used = !!(flags & VRING_PACKED_DESC_F_USED);
346 	avail = !!(flags & VRING_PACKED_DESC_F_AVAIL);
347 
348 	return avail == used && used == vq->vq_packed.used_wrap_counter;
349 }
350 
351 static inline void
352 vring_desc_init_packed(struct virtqueue *vq, int n)
353 {
354 	int i;
355 	for (i = 0; i < n - 1; i++) {
356 		vq->vq_packed.ring.desc[i].id = i;
357 		vq->vq_descx[i].next = i + 1;
358 	}
359 	vq->vq_packed.ring.desc[i].id = i;
360 	vq->vq_descx[i].next = VQ_RING_DESC_CHAIN_END;
361 }
362 
363 /* Chain all the descriptors in the ring with an END */
364 static inline void
365 vring_desc_init_split(struct vring_desc *dp, uint16_t n)
366 {
367 	uint16_t i;
368 
369 	for (i = 0; i < n - 1; i++)
370 		dp[i].next = (uint16_t)(i + 1);
371 	dp[i].next = VQ_RING_DESC_CHAIN_END;
372 }
373 
374 static inline void
375 vring_desc_init_indirect_packed(struct vring_packed_desc *dp, int n)
376 {
377 	int i;
378 	for (i = 0; i < n; i++) {
379 		dp[i].id = (uint16_t)i;
380 		dp[i].flags = VRING_DESC_F_WRITE;
381 	}
382 }
383 
384 /**
385  * Tell the backend not to interrupt us. Implementation for packed virtqueues.
386  */
387 static inline void
388 virtqueue_disable_intr_packed(struct virtqueue *vq)
389 {
390 	if (vq->vq_packed.event_flags_shadow != RING_EVENT_FLAGS_DISABLE) {
391 		vq->vq_packed.event_flags_shadow = RING_EVENT_FLAGS_DISABLE;
392 		vq->vq_packed.ring.driver->desc_event_flags =
393 			vq->vq_packed.event_flags_shadow;
394 	}
395 }
396 
397 /**
398  * Tell the backend not to interrupt us. Implementation for split virtqueues.
399  */
400 static inline void
401 virtqueue_disable_intr_split(struct virtqueue *vq)
402 {
403 	vq->vq_split.ring.avail->flags |= VRING_AVAIL_F_NO_INTERRUPT;
404 }
405 
406 /**
407  * Tell the backend not to interrupt us.
408  */
409 static inline void
410 virtqueue_disable_intr(struct virtqueue *vq)
411 {
412 	if (vtpci_packed_queue(vq->hw))
413 		virtqueue_disable_intr_packed(vq);
414 	else
415 		virtqueue_disable_intr_split(vq);
416 }
417 
418 /**
419  * Tell the backend to interrupt. Implementation for packed virtqueues.
420  */
421 static inline void
422 virtqueue_enable_intr_packed(struct virtqueue *vq)
423 {
424 	if (vq->vq_packed.event_flags_shadow == RING_EVENT_FLAGS_DISABLE) {
425 		vq->vq_packed.event_flags_shadow = RING_EVENT_FLAGS_ENABLE;
426 		vq->vq_packed.ring.driver->desc_event_flags =
427 			vq->vq_packed.event_flags_shadow;
428 	}
429 }
430 
431 /**
432  * Tell the backend to interrupt. Implementation for split virtqueues.
433  */
434 static inline void
435 virtqueue_enable_intr_split(struct virtqueue *vq)
436 {
437 	vq->vq_split.ring.avail->flags &= (~VRING_AVAIL_F_NO_INTERRUPT);
438 }
439 
440 /**
441  * Tell the backend to interrupt us.
442  */
443 static inline void
444 virtqueue_enable_intr(struct virtqueue *vq)
445 {
446 	if (vtpci_packed_queue(vq->hw))
447 		virtqueue_enable_intr_packed(vq);
448 	else
449 		virtqueue_enable_intr_split(vq);
450 }
451 
452 /**
453  *  Dump virtqueue internal structures, for debug purpose only.
454  */
455 void virtqueue_dump(struct virtqueue *vq);
456 /**
457  *  Get all mbufs to be freed.
458  */
459 struct rte_mbuf *virtqueue_detach_unused(struct virtqueue *vq);
460 
461 /* Flush the elements in the used ring. */
462 void virtqueue_rxvq_flush(struct virtqueue *vq);
463 
464 int virtqueue_rxvq_reset_packed(struct virtqueue *vq);
465 
466 int virtqueue_txvq_reset_packed(struct virtqueue *vq);
467 
468 static inline int
469 virtqueue_full(const struct virtqueue *vq)
470 {
471 	return vq->vq_free_cnt == 0;
472 }
473 
474 static inline int
475 virtio_get_queue_type(struct virtio_hw *hw, uint16_t vtpci_queue_idx)
476 {
477 	if (vtpci_queue_idx == hw->max_queue_pairs * 2)
478 		return VTNET_CQ;
479 	else if (vtpci_queue_idx % 2 == 0)
480 		return VTNET_RQ;
481 	else
482 		return VTNET_TQ;
483 }
484 
485 /* virtqueue_nused has load-acquire or rte_io_rmb insed */
486 static inline uint16_t
487 virtqueue_nused(const struct virtqueue *vq)
488 {
489 	uint16_t idx;
490 
491 	if (vq->hw->weak_barriers) {
492 	/**
493 	 * x86 prefers to using rte_smp_rmb over __atomic_load_n as it
494 	 * reports a slightly better perf, which comes from the saved
495 	 * branch by the compiler.
496 	 * The if and else branches are identical with the smp and io
497 	 * barriers both defined as compiler barriers on x86.
498 	 */
499 #ifdef RTE_ARCH_X86_64
500 		idx = vq->vq_split.ring.used->idx;
501 		rte_smp_rmb();
502 #else
503 		idx = __atomic_load_n(&(vq)->vq_split.ring.used->idx,
504 				__ATOMIC_ACQUIRE);
505 #endif
506 	} else {
507 		idx = vq->vq_split.ring.used->idx;
508 		rte_io_rmb();
509 	}
510 	return idx - vq->vq_used_cons_idx;
511 }
512 
513 void vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx);
514 void vq_ring_free_chain_packed(struct virtqueue *vq, uint16_t used_idx);
515 void vq_ring_free_inorder(struct virtqueue *vq, uint16_t desc_idx,
516 			  uint16_t num);
517 
518 static inline void
519 vq_update_avail_idx(struct virtqueue *vq)
520 {
521 	if (vq->hw->weak_barriers) {
522 	/* x86 prefers to using rte_smp_wmb over __atomic_store_n as
523 	 * it reports a slightly better perf, which comes from the
524 	 * saved branch by the compiler.
525 	 * The if and else branches are identical with the smp and
526 	 * io barriers both defined as compiler barriers on x86.
527 	 */
528 #ifdef RTE_ARCH_X86_64
529 		rte_smp_wmb();
530 		vq->vq_split.ring.avail->idx = vq->vq_avail_idx;
531 #else
532 		__atomic_store_n(&vq->vq_split.ring.avail->idx,
533 				 vq->vq_avail_idx, __ATOMIC_RELEASE);
534 #endif
535 	} else {
536 		rte_io_wmb();
537 		vq->vq_split.ring.avail->idx = vq->vq_avail_idx;
538 	}
539 }
540 
541 static inline void
542 vq_update_avail_ring(struct virtqueue *vq, uint16_t desc_idx)
543 {
544 	uint16_t avail_idx;
545 	/*
546 	 * Place the head of the descriptor chain into the next slot and make
547 	 * it usable to the host. The chain is made available now rather than
548 	 * deferring to virtqueue_notify() in the hopes that if the host is
549 	 * currently running on another CPU, we can keep it processing the new
550 	 * descriptor.
551 	 */
552 	avail_idx = (uint16_t)(vq->vq_avail_idx & (vq->vq_nentries - 1));
553 	if (unlikely(vq->vq_split.ring.avail->ring[avail_idx] != desc_idx))
554 		vq->vq_split.ring.avail->ring[avail_idx] = desc_idx;
555 	vq->vq_avail_idx++;
556 }
557 
558 static inline int
559 virtqueue_kick_prepare(struct virtqueue *vq)
560 {
561 	/*
562 	 * Ensure updated avail->idx is visible to vhost before reading
563 	 * the used->flags.
564 	 */
565 	virtio_mb(vq->hw->weak_barriers);
566 	return !(vq->vq_split.ring.used->flags & VRING_USED_F_NO_NOTIFY);
567 }
568 
569 static inline int
570 virtqueue_kick_prepare_packed(struct virtqueue *vq)
571 {
572 	uint16_t flags;
573 
574 	/*
575 	 * Ensure updated data is visible to vhost before reading the flags.
576 	 */
577 	virtio_mb(vq->hw->weak_barriers);
578 	flags = vq->vq_packed.ring.device->desc_event_flags;
579 
580 	return flags != RING_EVENT_FLAGS_DISABLE;
581 }
582 
583 /*
584  * virtqueue_kick_prepare*() or the virtio_wmb() should be called
585  * before this function to be sure that all the data is visible to vhost.
586  */
587 static inline void
588 virtqueue_notify(struct virtqueue *vq)
589 {
590 	VTPCI_OPS(vq->hw)->notify_queue(vq->hw, vq);
591 }
592 
593 #ifdef RTE_LIBRTE_VIRTIO_DEBUG_DUMP
594 #define VIRTQUEUE_DUMP(vq) do { \
595 	uint16_t used_idx, nused; \
596 	used_idx = __atomic_load_n(&(vq)->vq_split.ring.used->idx, \
597 				   __ATOMIC_RELAXED); \
598 	nused = (uint16_t)(used_idx - (vq)->vq_used_cons_idx); \
599 	if (vtpci_packed_queue((vq)->hw)) { \
600 		PMD_INIT_LOG(DEBUG, \
601 		"VQ: - size=%d; free=%d; used_cons_idx=%d; avail_idx=%d;" \
602 		" cached_flags=0x%x; used_wrap_counter=%d", \
603 		(vq)->vq_nentries, (vq)->vq_free_cnt, (vq)->vq_used_cons_idx, \
604 		(vq)->vq_avail_idx, (vq)->vq_packed.cached_flags, \
605 		(vq)->vq_packed.used_wrap_counter); \
606 		break; \
607 	} \
608 	PMD_INIT_LOG(DEBUG, \
609 	  "VQ: - size=%d; free=%d; used=%d; desc_head_idx=%d;" \
610 	  " avail.idx=%d; used_cons_idx=%d; used.idx=%d;" \
611 	  " avail.flags=0x%x; used.flags=0x%x", \
612 	  (vq)->vq_nentries, (vq)->vq_free_cnt, nused, (vq)->vq_desc_head_idx, \
613 	  (vq)->vq_split.ring.avail->idx, (vq)->vq_used_cons_idx, \
614 	  __atomic_load_n(&(vq)->vq_split.ring.used->idx, __ATOMIC_RELAXED), \
615 	  (vq)->vq_split.ring.avail->flags, (vq)->vq_split.ring.used->flags); \
616 } while (0)
617 #else
618 #define VIRTQUEUE_DUMP(vq) do { } while (0)
619 #endif
620 
621 /* avoid write operation when necessary, to lessen cache issues */
622 #define ASSIGN_UNLESS_EQUAL(var, val) do {	\
623 	typeof(var) *const var_ = &(var);	\
624 	typeof(val)  const val_ = (val);	\
625 	if (*var_ != val_)			\
626 		*var_ = val_;			\
627 } while (0)
628 
629 #define virtqueue_clear_net_hdr(hdr) do {		\
630 	typeof(hdr) hdr_ = (hdr);			\
631 	ASSIGN_UNLESS_EQUAL((hdr_)->csum_start, 0);	\
632 	ASSIGN_UNLESS_EQUAL((hdr_)->csum_offset, 0);	\
633 	ASSIGN_UNLESS_EQUAL((hdr_)->flags, 0);		\
634 	ASSIGN_UNLESS_EQUAL((hdr_)->gso_type, 0);	\
635 	ASSIGN_UNLESS_EQUAL((hdr_)->gso_size, 0);	\
636 	ASSIGN_UNLESS_EQUAL((hdr_)->hdr_len, 0);	\
637 } while (0)
638 
639 static inline void
640 virtqueue_xmit_offload(struct virtio_net_hdr *hdr,
641 			struct rte_mbuf *cookie,
642 			bool offload)
643 {
644 	if (offload) {
645 		if (cookie->ol_flags & PKT_TX_TCP_SEG)
646 			cookie->ol_flags |= PKT_TX_TCP_CKSUM;
647 
648 		switch (cookie->ol_flags & PKT_TX_L4_MASK) {
649 		case PKT_TX_UDP_CKSUM:
650 			hdr->csum_start = cookie->l2_len + cookie->l3_len;
651 			hdr->csum_offset = offsetof(struct rte_udp_hdr,
652 				dgram_cksum);
653 			hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
654 			break;
655 
656 		case PKT_TX_TCP_CKSUM:
657 			hdr->csum_start = cookie->l2_len + cookie->l3_len;
658 			hdr->csum_offset = offsetof(struct rte_tcp_hdr, cksum);
659 			hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
660 			break;
661 
662 		default:
663 			ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
664 			ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
665 			ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
666 			break;
667 		}
668 
669 		/* TCP Segmentation Offload */
670 		if (cookie->ol_flags & PKT_TX_TCP_SEG) {
671 			hdr->gso_type = (cookie->ol_flags & PKT_TX_IPV6) ?
672 				VIRTIO_NET_HDR_GSO_TCPV6 :
673 				VIRTIO_NET_HDR_GSO_TCPV4;
674 			hdr->gso_size = cookie->tso_segsz;
675 			hdr->hdr_len =
676 				cookie->l2_len +
677 				cookie->l3_len +
678 				cookie->l4_len;
679 		} else {
680 			ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
681 			ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
682 			ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
683 		}
684 	}
685 }
686 
687 static inline void
688 virtqueue_enqueue_xmit_packed(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
689 			      uint16_t needed, int use_indirect, int can_push,
690 			      int in_order)
691 {
692 	struct virtio_tx_region *txr = txvq->virtio_net_hdr_mz->addr;
693 	struct vq_desc_extra *dxp;
694 	struct virtqueue *vq = txvq->vq;
695 	struct vring_packed_desc *start_dp, *head_dp;
696 	uint16_t idx, id, head_idx, head_flags;
697 	int16_t head_size = vq->hw->vtnet_hdr_size;
698 	struct virtio_net_hdr *hdr;
699 	uint16_t prev;
700 	bool prepend_header = false;
701 	uint16_t seg_num = cookie->nb_segs;
702 
703 	id = in_order ? vq->vq_avail_idx : vq->vq_desc_head_idx;
704 
705 	dxp = &vq->vq_descx[id];
706 	dxp->ndescs = needed;
707 	dxp->cookie = cookie;
708 
709 	head_idx = vq->vq_avail_idx;
710 	idx = head_idx;
711 	prev = head_idx;
712 	start_dp = vq->vq_packed.ring.desc;
713 
714 	head_dp = &vq->vq_packed.ring.desc[idx];
715 	head_flags = cookie->next ? VRING_DESC_F_NEXT : 0;
716 	head_flags |= vq->vq_packed.cached_flags;
717 
718 	if (can_push) {
719 		/* prepend cannot fail, checked by caller */
720 		hdr = rte_pktmbuf_mtod_offset(cookie, struct virtio_net_hdr *,
721 					      -head_size);
722 		prepend_header = true;
723 
724 		/* if offload disabled, it is not zeroed below, do it now */
725 		if (!vq->hw->has_tx_offload)
726 			virtqueue_clear_net_hdr(hdr);
727 	} else if (use_indirect) {
728 		/* setup tx ring slot to point to indirect
729 		 * descriptor list stored in reserved region.
730 		 *
731 		 * the first slot in indirect ring is already preset
732 		 * to point to the header in reserved region
733 		 */
734 		start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
735 			RTE_PTR_DIFF(&txr[idx].tx_packed_indir, txr);
736 		start_dp[idx].len   = (seg_num + 1) *
737 			sizeof(struct vring_packed_desc);
738 		/* reset flags for indirect desc */
739 		head_flags = VRING_DESC_F_INDIRECT;
740 		head_flags |= vq->vq_packed.cached_flags;
741 		hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
742 
743 		/* loop below will fill in rest of the indirect elements */
744 		start_dp = txr[idx].tx_packed_indir;
745 		idx = 1;
746 	} else {
747 		/* setup first tx ring slot to point to header
748 		 * stored in reserved region.
749 		 */
750 		start_dp[idx].addr  = txvq->virtio_net_hdr_mem +
751 			RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
752 		start_dp[idx].len   = vq->hw->vtnet_hdr_size;
753 		hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
754 		idx++;
755 		if (idx >= vq->vq_nentries) {
756 			idx -= vq->vq_nentries;
757 			vq->vq_packed.cached_flags ^=
758 				VRING_PACKED_DESC_F_AVAIL_USED;
759 		}
760 	}
761 
762 	virtqueue_xmit_offload(hdr, cookie, vq->hw->has_tx_offload);
763 
764 	do {
765 		uint16_t flags;
766 
767 		start_dp[idx].addr = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
768 		start_dp[idx].len  = cookie->data_len;
769 		if (prepend_header) {
770 			start_dp[idx].addr -= head_size;
771 			start_dp[idx].len += head_size;
772 			prepend_header = false;
773 		}
774 
775 		if (likely(idx != head_idx)) {
776 			flags = cookie->next ? VRING_DESC_F_NEXT : 0;
777 			flags |= vq->vq_packed.cached_flags;
778 			start_dp[idx].flags = flags;
779 		}
780 		prev = idx;
781 		idx++;
782 		if (idx >= vq->vq_nentries) {
783 			idx -= vq->vq_nentries;
784 			vq->vq_packed.cached_flags ^=
785 				VRING_PACKED_DESC_F_AVAIL_USED;
786 		}
787 	} while ((cookie = cookie->next) != NULL);
788 
789 	start_dp[prev].id = id;
790 
791 	if (use_indirect) {
792 		idx = head_idx;
793 		if (++idx >= vq->vq_nentries) {
794 			idx -= vq->vq_nentries;
795 			vq->vq_packed.cached_flags ^=
796 				VRING_PACKED_DESC_F_AVAIL_USED;
797 		}
798 	}
799 
800 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
801 	vq->vq_avail_idx = idx;
802 
803 	if (!in_order) {
804 		vq->vq_desc_head_idx = dxp->next;
805 		if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
806 			vq->vq_desc_tail_idx = VQ_RING_DESC_CHAIN_END;
807 	}
808 
809 	virtqueue_store_flags_packed(head_dp, head_flags,
810 				     vq->hw->weak_barriers);
811 }
812 
813 static void
814 vq_ring_free_id_packed(struct virtqueue *vq, uint16_t id)
815 {
816 	struct vq_desc_extra *dxp;
817 
818 	dxp = &vq->vq_descx[id];
819 	vq->vq_free_cnt += dxp->ndescs;
820 
821 	if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END)
822 		vq->vq_desc_head_idx = id;
823 	else
824 		vq->vq_descx[vq->vq_desc_tail_idx].next = id;
825 
826 	vq->vq_desc_tail_idx = id;
827 	dxp->next = VQ_RING_DESC_CHAIN_END;
828 }
829 
830 static void
831 virtio_xmit_cleanup_inorder_packed(struct virtqueue *vq, int num)
832 {
833 	uint16_t used_idx, id, curr_id, free_cnt = 0;
834 	uint16_t size = vq->vq_nentries;
835 	struct vring_packed_desc *desc = vq->vq_packed.ring.desc;
836 	struct vq_desc_extra *dxp;
837 
838 	used_idx = vq->vq_used_cons_idx;
839 	/* desc_is_used has a load-acquire or rte_io_rmb inside
840 	 * and wait for used desc in virtqueue.
841 	 */
842 	while (num > 0 && desc_is_used(&desc[used_idx], vq)) {
843 		id = desc[used_idx].id;
844 		do {
845 			curr_id = used_idx;
846 			dxp = &vq->vq_descx[used_idx];
847 			used_idx += dxp->ndescs;
848 			free_cnt += dxp->ndescs;
849 			num -= dxp->ndescs;
850 			if (used_idx >= size) {
851 				used_idx -= size;
852 				vq->vq_packed.used_wrap_counter ^= 1;
853 			}
854 			if (dxp->cookie != NULL) {
855 				rte_pktmbuf_free(dxp->cookie);
856 				dxp->cookie = NULL;
857 			}
858 		} while (curr_id != id);
859 	}
860 	vq->vq_used_cons_idx = used_idx;
861 	vq->vq_free_cnt += free_cnt;
862 }
863 
864 static void
865 virtio_xmit_cleanup_normal_packed(struct virtqueue *vq, int num)
866 {
867 	uint16_t used_idx, id;
868 	uint16_t size = vq->vq_nentries;
869 	struct vring_packed_desc *desc = vq->vq_packed.ring.desc;
870 	struct vq_desc_extra *dxp;
871 
872 	used_idx = vq->vq_used_cons_idx;
873 	/* desc_is_used has a load-acquire or rte_io_rmb inside
874 	 * and wait for used desc in virtqueue.
875 	 */
876 	while (num-- && desc_is_used(&desc[used_idx], vq)) {
877 		id = desc[used_idx].id;
878 		dxp = &vq->vq_descx[id];
879 		vq->vq_used_cons_idx += dxp->ndescs;
880 		if (vq->vq_used_cons_idx >= size) {
881 			vq->vq_used_cons_idx -= size;
882 			vq->vq_packed.used_wrap_counter ^= 1;
883 		}
884 		vq_ring_free_id_packed(vq, id);
885 		if (dxp->cookie != NULL) {
886 			rte_pktmbuf_free(dxp->cookie);
887 			dxp->cookie = NULL;
888 		}
889 		used_idx = vq->vq_used_cons_idx;
890 	}
891 }
892 
893 /* Cleanup from completed transmits. */
894 static inline void
895 virtio_xmit_cleanup_packed(struct virtqueue *vq, int num, int in_order)
896 {
897 	if (in_order)
898 		virtio_xmit_cleanup_inorder_packed(vq, num);
899 	else
900 		virtio_xmit_cleanup_normal_packed(vq, num);
901 }
902 
903 static inline void
904 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
905 {
906 	uint16_t i, used_idx, desc_idx;
907 	for (i = 0; i < num; i++) {
908 		struct vring_used_elem *uep;
909 		struct vq_desc_extra *dxp;
910 
911 		used_idx = (uint16_t)(vq->vq_used_cons_idx &
912 				(vq->vq_nentries - 1));
913 		uep = &vq->vq_split.ring.used->ring[used_idx];
914 
915 		desc_idx = (uint16_t)uep->id;
916 		dxp = &vq->vq_descx[desc_idx];
917 		vq->vq_used_cons_idx++;
918 		vq_ring_free_chain(vq, desc_idx);
919 
920 		if (dxp->cookie != NULL) {
921 			rte_pktmbuf_free(dxp->cookie);
922 			dxp->cookie = NULL;
923 		}
924 	}
925 }
926 
927 /* Cleanup from completed inorder transmits. */
928 static __rte_always_inline void
929 virtio_xmit_cleanup_inorder(struct virtqueue *vq, uint16_t num)
930 {
931 	uint16_t i, idx = vq->vq_used_cons_idx;
932 	int16_t free_cnt = 0;
933 	struct vq_desc_extra *dxp = NULL;
934 
935 	if (unlikely(num == 0))
936 		return;
937 
938 	for (i = 0; i < num; i++) {
939 		dxp = &vq->vq_descx[idx++ & (vq->vq_nentries - 1)];
940 		free_cnt += dxp->ndescs;
941 		if (dxp->cookie != NULL) {
942 			rte_pktmbuf_free(dxp->cookie);
943 			dxp->cookie = NULL;
944 		}
945 	}
946 
947 	vq->vq_free_cnt += free_cnt;
948 	vq->vq_used_cons_idx = idx;
949 }
950 #endif /* _VIRTQUEUE_H_ */
951