xref: /dpdk/drivers/net/virtio/virtqueue.h (revision e12a0166c80f65e35408f4715b2f3a60763c3741)
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 #include "virtio_cvq.h"
20 
21 struct rte_mbuf;
22 
23 #define DEFAULT_TX_FREE_THRESH 32
24 #define DEFAULT_RX_FREE_THRESH 32
25 
26 #define VIRTIO_MBUF_BURST_SZ 64
27 /*
28  * Per virtio_ring.h in Linux.
29  *     For virtio_pci on SMP, we don't need to order with respect to MMIO
30  *     accesses through relaxed memory I/O windows, so thread_fence is
31  *     sufficient.
32  *
33  *     For using virtio to talk to real devices (eg. vDPA) we do need real
34  *     barriers.
35  */
36 static inline void
virtio_mb(uint8_t weak_barriers)37 virtio_mb(uint8_t weak_barriers)
38 {
39 	if (weak_barriers)
40 		rte_atomic_thread_fence(rte_memory_order_seq_cst);
41 	else
42 		rte_mb();
43 }
44 
45 static inline void
virtio_rmb(uint8_t weak_barriers)46 virtio_rmb(uint8_t weak_barriers)
47 {
48 	if (weak_barriers)
49 		rte_atomic_thread_fence(rte_memory_order_acquire);
50 	else
51 		rte_io_rmb();
52 }
53 
54 static inline void
virtio_wmb(uint8_t weak_barriers)55 virtio_wmb(uint8_t weak_barriers)
56 {
57 	if (weak_barriers)
58 		rte_atomic_thread_fence(rte_memory_order_release);
59 	else
60 		rte_io_wmb();
61 }
62 
63 static inline uint16_t
virtqueue_fetch_flags_packed(struct vring_packed_desc * dp,uint8_t weak_barriers)64 virtqueue_fetch_flags_packed(struct vring_packed_desc *dp,
65 			      uint8_t weak_barriers)
66 {
67 	uint16_t flags;
68 
69 	if (weak_barriers) {
70 /* x86 prefers to using rte_io_rmb over rte_atomic_load_explicit as it reports
71  * a better perf(~1.5%), which comes from the saved branch by the compiler.
72  * The if and else branch are identical  on the platforms except Arm.
73  */
74 #ifdef RTE_ARCH_ARM
75 		flags = rte_atomic_load_explicit(&dp->flags, rte_memory_order_acquire);
76 #else
77 		flags = dp->flags;
78 		rte_io_rmb();
79 #endif
80 	} else {
81 		flags = dp->flags;
82 		rte_io_rmb();
83 	}
84 
85 	return flags;
86 }
87 
88 static inline void
virtqueue_store_flags_packed(struct vring_packed_desc * dp,uint16_t flags,uint8_t weak_barriers)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_io_wmb over rte_atomic_store_explicit 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 on the platforms except Arm.
96  */
97 #ifdef RTE_ARCH_ARM
98 		rte_atomic_store_explicit(&dp->flags, flags, rte_memory_order_release);
99 #else
100 		rte_io_wmb();
101 		dp->flags = flags;
102 #endif
103 	} else {
104 		rte_io_wmb();
105 		dp->flags = flags;
106 	}
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_ARCH_32
118 #define VIRTIO_MBUF_ADDR_MASK(vq) ((vq)->mbuf_addr_mask)
119 #else
120 #define VIRTIO_MBUF_ADDR_MASK(vq) UINT64_MAX
121 #endif
122 
123 /**
124  * Return the IOVA (or virtual address in case of virtio-user) of mbuf
125  * data buffer.
126  *
127  * The address is firstly casted to the word size (sizeof(uintptr_t))
128  * before casting it to uint64_t. It is then masked with the expected
129  * address length (64 bits for virtio-pci, word size for virtio-user).
130  *
131  * This is to make it work with different combination of word size (64
132  * bit and 32 bit) and virtio device (virtio-pci and virtio-user).
133  */
134 #define VIRTIO_MBUF_ADDR(mb, vq) \
135 	((*(uint64_t *)((uintptr_t)(mb) + (vq)->mbuf_addr_offset)) & \
136 		VIRTIO_MBUF_ADDR_MASK(vq))
137 
138 /**
139  * Return the physical address (or virtual address in case of
140  * virtio-user) of mbuf data buffer, taking care of mbuf data offset
141  */
142 #define VIRTIO_MBUF_DATA_DMA_ADDR(mb, vq) \
143 	(VIRTIO_MBUF_ADDR(mb, vq) + (mb)->data_off)
144 
145 #define VTNET_SQ_RQ_QUEUE_IDX 0
146 #define VTNET_SQ_TQ_QUEUE_IDX 1
147 #define VTNET_SQ_CQ_QUEUE_IDX 2
148 
149 enum { VTNET_RQ = 0, VTNET_TQ = 1, VTNET_CQ = 2 };
150 /**
151  * The maximum virtqueue size is 2^15. Use that value as the end of
152  * descriptor chain terminator since it will never be a valid index
153  * in the descriptor table. This is used to verify we are correctly
154  * handling vq_free_cnt.
155  */
156 #define VQ_RING_DESC_CHAIN_END 32768
157 
158 #define VIRTIO_NET_OK     0
159 #define VIRTIO_NET_ERR    1
160 
161 struct vq_desc_extra {
162 	void *cookie;
163 	uint16_t ndescs;
164 	uint16_t next;
165 };
166 
167 #define virtnet_rxq_to_vq(rxvq) container_of(rxvq, struct virtqueue, rxq)
168 #define virtnet_txq_to_vq(txvq) container_of(txvq, struct virtqueue, txq)
169 #define virtnet_cq_to_vq(cvq) container_of(cvq, struct virtqueue, cq)
170 
171 struct virtqueue {
172 	struct virtio_hw  *hw; /**< virtio_hw structure pointer. */
173 	union {
174 		struct {
175 			/**< vring keeping desc, used and avail */
176 			struct vring ring;
177 		} vq_split;
178 
179 		struct {
180 			/**< vring keeping descs and events */
181 			struct vring_packed ring;
182 			bool used_wrap_counter;
183 			uint16_t cached_flags; /**< cached flags for descs */
184 			uint16_t event_flags_shadow;
185 		} vq_packed;
186 	};
187 
188 	uint16_t vq_used_cons_idx; /**< last consumed descriptor */
189 	uint16_t vq_nentries;  /**< vring desc numbers */
190 	uint16_t vq_free_cnt;  /**< num of desc available */
191 	uint16_t vq_avail_idx; /**< sync until needed */
192 	uint16_t vq_free_thresh; /**< free threshold */
193 
194 	/**
195 	 * Head of the free chain in the descriptor table. If
196 	 * there are no free descriptors, this will be set to
197 	 * VQ_RING_DESC_CHAIN_END.
198 	 */
199 	uint16_t  vq_desc_head_idx;
200 	uint16_t  vq_desc_tail_idx;
201 	uint16_t  vq_queue_index;   /**< PCI queue index */
202 
203 	void *vq_ring_virt_mem;  /**< linear address of vring*/
204 	unsigned int vq_ring_size;
205 	uint16_t mbuf_addr_offset;
206 	uint64_t mbuf_addr_mask;
207 
208 	union {
209 		struct virtnet_rx rxq;
210 		struct virtnet_tx txq;
211 		struct virtnet_ctl cq;
212 	};
213 
214 	const struct rte_memzone *mz; /**< mem zone to populate ring. */
215 	rte_iova_t vq_ring_mem; /**< physical address of vring,
216 	                         * or virtual address for virtio_user. */
217 
218 	uint16_t  *notify_addr;
219 	struct vq_desc_extra vq_descx[];
220 };
221 
222 /* If multiqueue is provided by host, then we support it. */
223 #define VIRTIO_NET_CTRL_MQ   4
224 
225 #define VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET        0
226 #define VIRTIO_NET_CTRL_MQ_RSS_CONFIG          1
227 
228 #define VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN        1
229 #define VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX        0x8000
230 
231 /**
232  * This is the first element of the scatter-gather list.  If you don't
233  * specify GSO or CSUM features, you can simply ignore the header.
234  */
235 struct virtio_net_hdr {
236 #define VIRTIO_NET_HDR_F_NEEDS_CSUM 1    /**< Use csum_start,csum_offset*/
237 #define VIRTIO_NET_HDR_F_DATA_VALID 2    /**< Checksum is valid */
238 	uint8_t flags;
239 #define VIRTIO_NET_HDR_GSO_NONE     0    /**< Not a GSO frame */
240 #define VIRTIO_NET_HDR_GSO_TCPV4    1    /**< GSO frame, IPv4 TCP (TSO) */
241 #define VIRTIO_NET_HDR_GSO_UDP      3    /**< GSO frame, IPv4 UDP (UFO) */
242 #define VIRTIO_NET_HDR_GSO_TCPV6    4    /**< GSO frame, IPv6 TCP */
243 #define VIRTIO_NET_HDR_GSO_ECN      0x80 /**< TCP has ECN set */
244 	uint8_t gso_type;
245 	uint16_t hdr_len;     /**< Ethernet + IP + tcp/udp hdrs */
246 	uint16_t gso_size;    /**< Bytes to append to hdr_len per frame */
247 	uint16_t csum_start;  /**< Position to start checksumming from */
248 	uint16_t csum_offset; /**< Offset after that to place checksum */
249 };
250 
251 /**
252  * This is the version of the header to use when the MRG_RXBUF
253  * feature has been negotiated.
254  */
255 struct virtio_net_hdr_mrg_rxbuf {
256 	struct   virtio_net_hdr hdr;
257 	uint16_t num_buffers; /**< Number of merged rx buffers */
258 };
259 
260 /* Region reserved to allow for transmit header and indirect ring */
261 #define VIRTIO_MAX_TX_INDIRECT 8
262 struct virtio_tx_region {
263 	struct virtio_net_hdr_mrg_rxbuf tx_hdr;
264 	union __rte_aligned(16) {
265 		struct vring_desc tx_indir[VIRTIO_MAX_TX_INDIRECT];
266 		struct vring_packed_desc
267 			tx_packed_indir[VIRTIO_MAX_TX_INDIRECT];
268 	};
269 };
270 
271 static inline int
desc_is_used(struct vring_packed_desc * desc,struct virtqueue * vq)272 desc_is_used(struct vring_packed_desc *desc, struct virtqueue *vq)
273 {
274 	uint16_t used, avail, flags;
275 
276 	flags = virtqueue_fetch_flags_packed(desc, vq->hw->weak_barriers);
277 	used = !!(flags & VRING_PACKED_DESC_F_USED);
278 	avail = !!(flags & VRING_PACKED_DESC_F_AVAIL);
279 
280 	return avail == used && used == vq->vq_packed.used_wrap_counter;
281 }
282 
283 static inline void
vring_desc_init_packed(struct virtqueue * vq,int n)284 vring_desc_init_packed(struct virtqueue *vq, int n)
285 {
286 	int i;
287 	for (i = 0; i < n - 1; i++) {
288 		vq->vq_packed.ring.desc[i].id = i;
289 		vq->vq_descx[i].next = i + 1;
290 	}
291 	vq->vq_packed.ring.desc[i].id = i;
292 	vq->vq_descx[i].next = VQ_RING_DESC_CHAIN_END;
293 }
294 
295 /* Chain all the descriptors in the ring with an END */
296 static inline void
vring_desc_init_split(struct vring_desc * dp,uint16_t n)297 vring_desc_init_split(struct vring_desc *dp, uint16_t n)
298 {
299 	uint16_t i;
300 
301 	for (i = 0; i < n - 1; i++)
302 		dp[i].next = (uint16_t)(i + 1);
303 	dp[i].next = VQ_RING_DESC_CHAIN_END;
304 }
305 
306 static inline void
vring_desc_init_indirect_packed(struct vring_packed_desc * dp,int n)307 vring_desc_init_indirect_packed(struct vring_packed_desc *dp, int n)
308 {
309 	int i;
310 	for (i = 0; i < n; i++) {
311 		dp[i].id = (uint16_t)i;
312 		dp[i].flags = VRING_DESC_F_WRITE;
313 	}
314 }
315 
316 /**
317  * Tell the backend not to interrupt us. Implementation for packed virtqueues.
318  */
319 static inline void
virtqueue_disable_intr_packed(struct virtqueue * vq)320 virtqueue_disable_intr_packed(struct virtqueue *vq)
321 {
322 	if (vq->vq_packed.event_flags_shadow != RING_EVENT_FLAGS_DISABLE) {
323 		vq->vq_packed.event_flags_shadow = RING_EVENT_FLAGS_DISABLE;
324 		vq->vq_packed.ring.driver->desc_event_flags =
325 			vq->vq_packed.event_flags_shadow;
326 	}
327 }
328 
329 /**
330  * Tell the backend not to interrupt us. Implementation for split virtqueues.
331  */
332 static inline void
virtqueue_disable_intr_split(struct virtqueue * vq)333 virtqueue_disable_intr_split(struct virtqueue *vq)
334 {
335 	vq->vq_split.ring.avail->flags |= VRING_AVAIL_F_NO_INTERRUPT;
336 }
337 
338 /**
339  * Tell the backend not to interrupt us.
340  */
341 static inline void
virtqueue_disable_intr(struct virtqueue * vq)342 virtqueue_disable_intr(struct virtqueue *vq)
343 {
344 	if (virtio_with_packed_queue(vq->hw))
345 		virtqueue_disable_intr_packed(vq);
346 	else
347 		virtqueue_disable_intr_split(vq);
348 }
349 
350 /**
351  * Tell the backend to interrupt. Implementation for packed virtqueues.
352  */
353 static inline void
virtqueue_enable_intr_packed(struct virtqueue * vq)354 virtqueue_enable_intr_packed(struct virtqueue *vq)
355 {
356 	if (vq->vq_packed.event_flags_shadow == RING_EVENT_FLAGS_DISABLE) {
357 		vq->vq_packed.event_flags_shadow = RING_EVENT_FLAGS_ENABLE;
358 		vq->vq_packed.ring.driver->desc_event_flags =
359 			vq->vq_packed.event_flags_shadow;
360 	}
361 }
362 
363 /**
364  * Tell the backend to interrupt. Implementation for split virtqueues.
365  */
366 static inline void
virtqueue_enable_intr_split(struct virtqueue * vq)367 virtqueue_enable_intr_split(struct virtqueue *vq)
368 {
369 	vq->vq_split.ring.avail->flags &= (~VRING_AVAIL_F_NO_INTERRUPT);
370 }
371 
372 /**
373  * Tell the backend to interrupt us.
374  */
375 static inline void
virtqueue_enable_intr(struct virtqueue * vq)376 virtqueue_enable_intr(struct virtqueue *vq)
377 {
378 	if (virtio_with_packed_queue(vq->hw))
379 		virtqueue_enable_intr_packed(vq);
380 	else
381 		virtqueue_enable_intr_split(vq);
382 }
383 
384 /**
385  *  Get all mbufs to be freed.
386  */
387 struct rte_mbuf *virtqueue_detach_unused(struct virtqueue *vq);
388 
389 /* Flush the elements in the used ring. */
390 void virtqueue_rxvq_flush(struct virtqueue *vq);
391 
392 int virtqueue_rxvq_reset_packed(struct virtqueue *vq);
393 
394 int virtqueue_txvq_reset_packed(struct virtqueue *vq);
395 
396 void virtqueue_txq_indirect_headers_init(struct virtqueue *vq);
397 
398 struct virtqueue *virtqueue_alloc(struct virtio_hw *hw, uint16_t index,
399 		uint16_t num, int type, int node, const char *name);
400 
401 void virtqueue_free(struct virtqueue *vq);
402 
403 static inline int
virtqueue_full(const struct virtqueue * vq)404 virtqueue_full(const struct virtqueue *vq)
405 {
406 	return vq->vq_free_cnt == 0;
407 }
408 
409 static inline int
virtio_get_queue_type(struct virtio_hw * hw,uint16_t vq_idx)410 virtio_get_queue_type(struct virtio_hw *hw, uint16_t vq_idx)
411 {
412 	if (vq_idx == hw->max_queue_pairs * 2)
413 		return VTNET_CQ;
414 	else if (vq_idx % 2 == 0)
415 		return VTNET_RQ;
416 	else
417 		return VTNET_TQ;
418 }
419 
420 /* virtqueue_nused has load-acquire or rte_io_rmb insed */
421 static inline uint16_t
virtqueue_nused(const struct virtqueue * vq)422 virtqueue_nused(const struct virtqueue *vq)
423 {
424 	uint16_t idx;
425 
426 	if (vq->hw->weak_barriers) {
427 	/**
428 	 * x86 prefers to using rte_smp_rmb over rte_atomic_load_explicit as it
429 	 * reports a slightly better perf, which comes from the saved
430 	 * branch by the compiler.
431 	 * The if and else branches are identical with the smp and io
432 	 * barriers both defined as compiler barriers on x86.
433 	 */
434 #ifdef RTE_ARCH_X86_64
435 		idx = vq->vq_split.ring.used->idx;
436 		rte_smp_rmb();
437 #else
438 		idx = rte_atomic_load_explicit(&(vq)->vq_split.ring.used->idx,
439 				rte_memory_order_acquire);
440 #endif
441 	} else {
442 		idx = vq->vq_split.ring.used->idx;
443 		rte_io_rmb();
444 	}
445 	return idx - vq->vq_used_cons_idx;
446 }
447 
448 void vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx);
449 void vq_ring_free_chain_packed(struct virtqueue *vq, uint16_t used_idx);
450 void vq_ring_free_inorder(struct virtqueue *vq, uint16_t desc_idx,
451 			  uint16_t num);
452 
453 static inline void
vq_update_avail_idx(struct virtqueue * vq)454 vq_update_avail_idx(struct virtqueue *vq)
455 {
456 	if (vq->hw->weak_barriers) {
457 	/* x86 prefers to using rte_smp_wmb over rte_atomic_store_explicit as
458 	 * it reports a slightly better perf, which comes from the
459 	 * saved branch by the compiler.
460 	 * The if and else branches are identical with the smp and
461 	 * io barriers both defined as compiler barriers on x86.
462 	 */
463 #ifdef RTE_ARCH_X86_64
464 		rte_smp_wmb();
465 		vq->vq_split.ring.avail->idx = vq->vq_avail_idx;
466 #else
467 		rte_atomic_store_explicit(&vq->vq_split.ring.avail->idx,
468 				 vq->vq_avail_idx, rte_memory_order_release);
469 #endif
470 	} else {
471 		rte_io_wmb();
472 		vq->vq_split.ring.avail->idx = vq->vq_avail_idx;
473 	}
474 }
475 
476 static inline void
vq_update_avail_ring(struct virtqueue * vq,uint16_t desc_idx)477 vq_update_avail_ring(struct virtqueue *vq, uint16_t desc_idx)
478 {
479 	uint16_t avail_idx;
480 	/*
481 	 * Place the head of the descriptor chain into the next slot and make
482 	 * it usable to the host. The chain is made available now rather than
483 	 * deferring to virtqueue_notify() in the hopes that if the host is
484 	 * currently running on another CPU, we can keep it processing the new
485 	 * descriptor.
486 	 */
487 	avail_idx = (uint16_t)(vq->vq_avail_idx & (vq->vq_nentries - 1));
488 	if (unlikely(vq->vq_split.ring.avail->ring[avail_idx] != desc_idx))
489 		vq->vq_split.ring.avail->ring[avail_idx] = desc_idx;
490 	vq->vq_avail_idx++;
491 }
492 
493 static inline int
virtqueue_kick_prepare(struct virtqueue * vq)494 virtqueue_kick_prepare(struct virtqueue *vq)
495 {
496 	/*
497 	 * Ensure updated avail->idx is visible to vhost before reading
498 	 * the used->flags.
499 	 */
500 	virtio_mb(vq->hw->weak_barriers);
501 	return !(vq->vq_split.ring.used->flags & VRING_USED_F_NO_NOTIFY);
502 }
503 
504 static inline int
virtqueue_kick_prepare_packed(struct virtqueue * vq)505 virtqueue_kick_prepare_packed(struct virtqueue *vq)
506 {
507 	uint16_t flags;
508 
509 	/*
510 	 * Ensure updated data is visible to vhost before reading the flags.
511 	 */
512 	virtio_mb(vq->hw->weak_barriers);
513 	flags = vq->vq_packed.ring.device->desc_event_flags;
514 
515 	return flags != RING_EVENT_FLAGS_DISABLE;
516 }
517 
518 /*
519  * virtqueue_kick_prepare*() or the virtio_wmb() should be called
520  * before this function to be sure that all the data is visible to vhost.
521  */
522 static inline void
virtqueue_notify(struct virtqueue * vq)523 virtqueue_notify(struct virtqueue *vq)
524 {
525 	VIRTIO_OPS(vq->hw)->notify_queue(vq->hw, vq);
526 }
527 
528 #ifdef RTE_LIBRTE_VIRTIO_DEBUG_DUMP
529 #define VIRTQUEUE_DUMP(vq) do { \
530 	uint16_t used_idx, nused; \
531 	used_idx = rte_atomic_load_explicit(&(vq)->vq_split.ring.used->idx, \
532 				   rte_memory_order_relaxed); \
533 	nused = (uint16_t)(used_idx - (vq)->vq_used_cons_idx); \
534 	if (virtio_with_packed_queue((vq)->hw)) { \
535 		PMD_INIT_LOG(DEBUG, \
536 		"VQ: - size=%d; free=%d; used_cons_idx=%d; avail_idx=%d;" \
537 		" cached_flags=0x%x; used_wrap_counter=%d", \
538 		(vq)->vq_nentries, (vq)->vq_free_cnt, (vq)->vq_used_cons_idx, \
539 		(vq)->vq_avail_idx, (vq)->vq_packed.cached_flags, \
540 		(vq)->vq_packed.used_wrap_counter); \
541 		break; \
542 	} \
543 	PMD_INIT_LOG(DEBUG, \
544 	  "VQ: - size=%d; free=%d; used=%d; desc_head_idx=%d;" \
545 	  " avail.idx=%d; used_cons_idx=%d; used.idx=%d;" \
546 	  " avail.flags=0x%x; used.flags=0x%x", \
547 	  (vq)->vq_nentries, (vq)->vq_free_cnt, nused, (vq)->vq_desc_head_idx, \
548 	  (vq)->vq_split.ring.avail->idx, (vq)->vq_used_cons_idx, \
549 	  rte_atomic_load_explicit(&(vq)->vq_split.ring.used->idx, rte_memory_order_relaxed), \
550 	  (vq)->vq_split.ring.avail->flags, (vq)->vq_split.ring.used->flags); \
551 } while (0)
552 #else
553 #define VIRTQUEUE_DUMP(vq) do { } while (0)
554 #endif
555 
556 /* avoid write operation when necessary, to lessen cache issues */
557 #define ASSIGN_UNLESS_EQUAL(var, val) do {	\
558 	typeof(var) *const var_ = &(var);	\
559 	typeof(val)  const val_ = (val);	\
560 	if (*var_ != val_)			\
561 		*var_ = val_;			\
562 } while (0)
563 
564 #define virtqueue_clear_net_hdr(hdr) do {		\
565 	typeof(hdr) hdr_ = (hdr);			\
566 	ASSIGN_UNLESS_EQUAL((hdr_)->csum_start, 0);	\
567 	ASSIGN_UNLESS_EQUAL((hdr_)->csum_offset, 0);	\
568 	ASSIGN_UNLESS_EQUAL((hdr_)->flags, 0);		\
569 	ASSIGN_UNLESS_EQUAL((hdr_)->gso_type, 0);	\
570 	ASSIGN_UNLESS_EQUAL((hdr_)->gso_size, 0);	\
571 	ASSIGN_UNLESS_EQUAL((hdr_)->hdr_len, 0);	\
572 } while (0)
573 
574 static inline void
virtqueue_xmit_offload(struct virtio_net_hdr * hdr,struct rte_mbuf * cookie)575 virtqueue_xmit_offload(struct virtio_net_hdr *hdr, struct rte_mbuf *cookie)
576 {
577 	uint64_t csum_l4 = cookie->ol_flags & RTE_MBUF_F_TX_L4_MASK;
578 	uint16_t o_l23_len = (cookie->ol_flags & RTE_MBUF_F_TX_TUNNEL_MASK) ?
579 			     cookie->outer_l2_len + cookie->outer_l3_len : 0;
580 
581 	if (cookie->ol_flags & RTE_MBUF_F_TX_TCP_SEG)
582 		csum_l4 |= RTE_MBUF_F_TX_TCP_CKSUM;
583 
584 	switch (csum_l4) {
585 	case RTE_MBUF_F_TX_UDP_CKSUM:
586 		hdr->csum_start = o_l23_len + cookie->l2_len + cookie->l3_len;
587 		hdr->csum_offset = offsetof(struct rte_udp_hdr, dgram_cksum);
588 		hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
589 		break;
590 
591 	case RTE_MBUF_F_TX_TCP_CKSUM:
592 		hdr->csum_start = o_l23_len + cookie->l2_len + cookie->l3_len;
593 		hdr->csum_offset = offsetof(struct rte_tcp_hdr, cksum);
594 		hdr->flags = VIRTIO_NET_HDR_F_NEEDS_CSUM;
595 		break;
596 
597 	default:
598 		ASSIGN_UNLESS_EQUAL(hdr->csum_start, 0);
599 		ASSIGN_UNLESS_EQUAL(hdr->csum_offset, 0);
600 		ASSIGN_UNLESS_EQUAL(hdr->flags, 0);
601 		break;
602 	}
603 
604 	/* TCP Segmentation Offload */
605 	if (cookie->ol_flags & RTE_MBUF_F_TX_TCP_SEG) {
606 		hdr->gso_type = (cookie->ol_flags & RTE_MBUF_F_TX_IPV6) ?
607 			VIRTIO_NET_HDR_GSO_TCPV6 :
608 			VIRTIO_NET_HDR_GSO_TCPV4;
609 		hdr->gso_size = cookie->tso_segsz;
610 		hdr->hdr_len = o_l23_len + cookie->l2_len + cookie->l3_len +
611 			       cookie->l4_len;
612 	} else {
613 		ASSIGN_UNLESS_EQUAL(hdr->gso_type, 0);
614 		ASSIGN_UNLESS_EQUAL(hdr->gso_size, 0);
615 		ASSIGN_UNLESS_EQUAL(hdr->hdr_len, 0);
616 	}
617 }
618 
619 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)620 virtqueue_enqueue_xmit_packed(struct virtnet_tx *txvq, struct rte_mbuf *cookie,
621 			      uint16_t needed, int use_indirect, int can_push,
622 			      int in_order)
623 {
624 	struct virtio_tx_region *txr = txvq->hdr_mz->addr;
625 	struct vq_desc_extra *dxp;
626 	struct virtqueue *vq = virtnet_txq_to_vq(txvq);
627 	struct vring_packed_desc *start_dp, *head_dp;
628 	uint16_t idx, id, head_idx, head_flags;
629 	int16_t head_size = vq->hw->vtnet_hdr_size;
630 	struct virtio_net_hdr *hdr;
631 	uint16_t prev;
632 	bool prepend_header = false;
633 	uint16_t seg_num = cookie->nb_segs;
634 
635 	id = in_order ? vq->vq_avail_idx : vq->vq_desc_head_idx;
636 
637 	dxp = &vq->vq_descx[id];
638 	dxp->ndescs = needed;
639 	dxp->cookie = cookie;
640 
641 	head_idx = vq->vq_avail_idx;
642 	idx = head_idx;
643 	prev = head_idx;
644 	start_dp = vq->vq_packed.ring.desc;
645 
646 	head_dp = &vq->vq_packed.ring.desc[idx];
647 	head_flags = cookie->next ? VRING_DESC_F_NEXT : 0;
648 	head_flags |= vq->vq_packed.cached_flags;
649 
650 	if (can_push) {
651 		/* prepend cannot fail, checked by caller */
652 		hdr = rte_pktmbuf_mtod_offset(cookie, struct virtio_net_hdr *,
653 					      -head_size);
654 		prepend_header = true;
655 
656 		/* if offload disabled, it is not zeroed below, do it now */
657 		if (!vq->hw->has_tx_offload)
658 			virtqueue_clear_net_hdr(hdr);
659 	} else if (use_indirect) {
660 		/* setup tx ring slot to point to indirect
661 		 * descriptor list stored in reserved region.
662 		 *
663 		 * the first slot in indirect ring is already preset
664 		 * to point to the header in reserved region
665 		 */
666 		start_dp[idx].addr = txvq->hdr_mem + RTE_PTR_DIFF(&txr[idx].tx_packed_indir, txr);
667 		start_dp[idx].len = (seg_num + 1) * sizeof(struct vring_packed_desc);
668 		/* Packed descriptor id needs to be restored when inorder. */
669 		if (in_order)
670 			start_dp[idx].id = idx;
671 		/* reset flags for indirect desc */
672 		head_flags = VRING_DESC_F_INDIRECT;
673 		head_flags |= vq->vq_packed.cached_flags;
674 		hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
675 
676 		/* loop below will fill in rest of the indirect elements */
677 		start_dp = txr[idx].tx_packed_indir;
678 		idx = 1;
679 	} else {
680 		/* setup first tx ring slot to point to header
681 		 * stored in reserved region.
682 		 */
683 		start_dp[idx].addr = txvq->hdr_mem + RTE_PTR_DIFF(&txr[idx].tx_hdr, txr);
684 		start_dp[idx].len = vq->hw->vtnet_hdr_size;
685 		head_flags |= VRING_DESC_F_NEXT;
686 		hdr = (struct virtio_net_hdr *)&txr[idx].tx_hdr;
687 		idx++;
688 		if (idx >= vq->vq_nentries) {
689 			idx -= vq->vq_nentries;
690 			vq->vq_packed.cached_flags ^=
691 				VRING_PACKED_DESC_F_AVAIL_USED;
692 		}
693 	}
694 
695 	if (vq->hw->has_tx_offload)
696 		virtqueue_xmit_offload(hdr, cookie);
697 
698 	do {
699 		uint16_t flags;
700 
701 		start_dp[idx].addr = VIRTIO_MBUF_DATA_DMA_ADDR(cookie, vq);
702 		start_dp[idx].len  = cookie->data_len;
703 		if (prepend_header) {
704 			start_dp[idx].addr -= head_size;
705 			start_dp[idx].len += head_size;
706 			prepend_header = false;
707 		}
708 
709 		if (likely(idx != head_idx)) {
710 			flags = cookie->next ? VRING_DESC_F_NEXT : 0;
711 			flags |= vq->vq_packed.cached_flags;
712 			start_dp[idx].flags = flags;
713 		}
714 		prev = idx;
715 		idx++;
716 		if (idx >= vq->vq_nentries) {
717 			idx -= vq->vq_nentries;
718 			vq->vq_packed.cached_flags ^=
719 				VRING_PACKED_DESC_F_AVAIL_USED;
720 		}
721 	} while ((cookie = cookie->next) != NULL);
722 
723 	start_dp[prev].id = id;
724 
725 	if (use_indirect) {
726 		idx = head_idx;
727 		if (++idx >= vq->vq_nentries) {
728 			idx -= vq->vq_nentries;
729 			vq->vq_packed.cached_flags ^=
730 				VRING_PACKED_DESC_F_AVAIL_USED;
731 		}
732 	}
733 
734 	vq->vq_free_cnt = (uint16_t)(vq->vq_free_cnt - needed);
735 	vq->vq_avail_idx = idx;
736 
737 	if (!in_order) {
738 		vq->vq_desc_head_idx = dxp->next;
739 		if (vq->vq_desc_head_idx == VQ_RING_DESC_CHAIN_END)
740 			vq->vq_desc_tail_idx = VQ_RING_DESC_CHAIN_END;
741 	}
742 
743 	virtqueue_store_flags_packed(head_dp, head_flags,
744 				     vq->hw->weak_barriers);
745 }
746 
747 static void
vq_ring_free_id_packed(struct virtqueue * vq,uint16_t id)748 vq_ring_free_id_packed(struct virtqueue *vq, uint16_t id)
749 {
750 	struct vq_desc_extra *dxp;
751 
752 	dxp = &vq->vq_descx[id];
753 	vq->vq_free_cnt += dxp->ndescs;
754 
755 	if (vq->vq_desc_tail_idx == VQ_RING_DESC_CHAIN_END)
756 		vq->vq_desc_head_idx = id;
757 	else
758 		vq->vq_descx[vq->vq_desc_tail_idx].next = id;
759 
760 	vq->vq_desc_tail_idx = id;
761 	dxp->next = VQ_RING_DESC_CHAIN_END;
762 }
763 
764 static void
virtio_xmit_cleanup_inorder_packed(struct virtqueue * vq,uint16_t num)765 virtio_xmit_cleanup_inorder_packed(struct virtqueue *vq, uint16_t num)
766 {
767 	uint16_t used_idx, id, curr_id, free_cnt = 0;
768 	uint16_t size = vq->vq_nentries;
769 	struct vring_packed_desc *desc = vq->vq_packed.ring.desc;
770 	struct vq_desc_extra *dxp;
771 	int nb = num;
772 
773 	used_idx = vq->vq_used_cons_idx;
774 	/* desc_is_used has a load-acquire or rte_io_rmb inside
775 	 * and wait for used desc in virtqueue.
776 	 */
777 	while (nb > 0 && desc_is_used(&desc[used_idx], vq)) {
778 		id = desc[used_idx].id;
779 		do {
780 			curr_id = used_idx;
781 			dxp = &vq->vq_descx[used_idx];
782 			used_idx += dxp->ndescs;
783 			free_cnt += dxp->ndescs;
784 			nb -= dxp->ndescs;
785 			if (used_idx >= size) {
786 				used_idx -= size;
787 				vq->vq_packed.used_wrap_counter ^= 1;
788 			}
789 			if (dxp->cookie != NULL) {
790 				rte_pktmbuf_free(dxp->cookie);
791 				dxp->cookie = NULL;
792 			}
793 		} while (curr_id != id);
794 	}
795 	vq->vq_used_cons_idx = used_idx;
796 	vq->vq_free_cnt += free_cnt;
797 }
798 
799 static void
virtio_xmit_cleanup_normal_packed(struct virtqueue * vq,uint16_t num)800 virtio_xmit_cleanup_normal_packed(struct virtqueue *vq, uint16_t num)
801 {
802 	uint16_t used_idx, id;
803 	uint16_t size = vq->vq_nentries;
804 	struct vring_packed_desc *desc = vq->vq_packed.ring.desc;
805 	struct vq_desc_extra *dxp;
806 
807 	used_idx = vq->vq_used_cons_idx;
808 	/* desc_is_used has a load-acquire or rte_io_rmb inside
809 	 * and wait for used desc in virtqueue.
810 	 */
811 	while (num-- && desc_is_used(&desc[used_idx], vq)) {
812 		id = desc[used_idx].id;
813 		dxp = &vq->vq_descx[id];
814 		vq->vq_used_cons_idx += dxp->ndescs;
815 		if (vq->vq_used_cons_idx >= size) {
816 			vq->vq_used_cons_idx -= size;
817 			vq->vq_packed.used_wrap_counter ^= 1;
818 		}
819 		vq_ring_free_id_packed(vq, id);
820 		if (dxp->cookie != NULL) {
821 			rte_pktmbuf_free(dxp->cookie);
822 			dxp->cookie = NULL;
823 		}
824 		used_idx = vq->vq_used_cons_idx;
825 	}
826 }
827 
828 /* Cleanup from completed transmits. */
829 static inline void
virtio_xmit_cleanup_packed(struct virtqueue * vq,uint16_t num,int in_order)830 virtio_xmit_cleanup_packed(struct virtqueue *vq, uint16_t num, int in_order)
831 {
832 	if (in_order)
833 		virtio_xmit_cleanup_inorder_packed(vq, num);
834 	else
835 		virtio_xmit_cleanup_normal_packed(vq, num);
836 }
837 
838 static inline void
virtio_xmit_cleanup(struct virtqueue * vq,uint16_t num)839 virtio_xmit_cleanup(struct virtqueue *vq, uint16_t num)
840 {
841 	uint16_t i, used_idx, desc_idx;
842 	for (i = 0; i < num; i++) {
843 		struct vring_used_elem *uep;
844 		struct vq_desc_extra *dxp;
845 
846 		used_idx = (uint16_t)(vq->vq_used_cons_idx &
847 				(vq->vq_nentries - 1));
848 		uep = &vq->vq_split.ring.used->ring[used_idx];
849 
850 		desc_idx = (uint16_t)uep->id;
851 		dxp = &vq->vq_descx[desc_idx];
852 		vq->vq_used_cons_idx++;
853 		vq_ring_free_chain(vq, desc_idx);
854 
855 		if (dxp->cookie != NULL) {
856 			rte_pktmbuf_free(dxp->cookie);
857 			dxp->cookie = NULL;
858 		}
859 	}
860 }
861 
862 /* Cleanup from completed inorder transmits. */
863 static __rte_always_inline void
virtio_xmit_cleanup_inorder(struct virtqueue * vq,uint16_t num)864 virtio_xmit_cleanup_inorder(struct virtqueue *vq, uint16_t num)
865 {
866 	uint16_t i, idx = vq->vq_used_cons_idx;
867 	int16_t free_cnt = 0;
868 	struct vq_desc_extra *dxp = NULL;
869 
870 	if (unlikely(num == 0))
871 		return;
872 
873 	for (i = 0; i < num; i++) {
874 		dxp = &vq->vq_descx[idx++ & (vq->vq_nentries - 1)];
875 		free_cnt += dxp->ndescs;
876 		if (dxp->cookie != NULL) {
877 			rte_pktmbuf_free(dxp->cookie);
878 			dxp->cookie = NULL;
879 		}
880 	}
881 
882 	vq->vq_free_cnt += free_cnt;
883 	vq->vq_used_cons_idx = idx;
884 }
885 #endif /* _VIRTQUEUE_H_ */
886