xref: /dpdk/drivers/net/virtio/virtqueue.h (revision 1f41d98c207aee8982ced709864c96c463d4503a)
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 
14 #include "virtio_pci.h"
15 #include "virtio_ring.h"
16 #include "virtio_logs.h"
17 #include "virtio_rxtx.h"
18 
19 struct rte_mbuf;
20 
21 /*
22  * Per virtio_ring.h in Linux.
23  *     For virtio_pci on SMP, we don't need to order with respect to MMIO
24  *     accesses through relaxed memory I/O windows, so smp_mb() et al are
25  *     sufficient.
26  *
27  *     For using virtio to talk to real devices (eg. vDPA) we do need real
28  *     barriers.
29  */
30 static inline void
31 virtio_mb(uint8_t weak_barriers)
32 {
33 	if (weak_barriers)
34 		rte_smp_mb();
35 	else
36 		rte_mb();
37 }
38 
39 static inline void
40 virtio_rmb(uint8_t weak_barriers)
41 {
42 	if (weak_barriers)
43 		rte_smp_rmb();
44 	else
45 		rte_cio_rmb();
46 }
47 
48 static inline void
49 virtio_wmb(uint8_t weak_barriers)
50 {
51 	if (weak_barriers)
52 		rte_smp_wmb();
53 	else
54 		rte_cio_wmb();
55 }
56 
57 static inline uint16_t
58 virtqueue_fetch_flags_packed(struct vring_packed_desc *dp,
59 			      uint8_t weak_barriers)
60 {
61 	uint16_t flags;
62 
63 	if (weak_barriers) {
64 /* x86 prefers to using rte_smp_rmb over __atomic_load_n as it reports
65  * a better perf(~1.5%), which comes from the saved branch by the compiler.
66  * The if and else branch are identical with the smp and cio barriers both
67  * defined as compiler barriers on x86.
68  */
69 #ifdef RTE_ARCH_X86_64
70 		flags = dp->flags;
71 		rte_smp_rmb();
72 #else
73 		flags = __atomic_load_n(&dp->flags, __ATOMIC_ACQUIRE);
74 #endif
75 	} else {
76 		flags = dp->flags;
77 		rte_cio_rmb();
78 	}
79 
80 	return flags;
81 }
82 
83 static inline void
84 virtqueue_store_flags_packed(struct vring_packed_desc *dp,
85 			      uint16_t flags, uint8_t weak_barriers)
86 {
87 	if (weak_barriers) {
88 /* x86 prefers to using rte_smp_wmb over __atomic_store_n as it reports
89  * a better perf(~1.5%), which comes from the saved branch by the compiler.
90  * The if and else branch are identical with the smp and cio barriers both
91  * defined as compiler barriers on x86.
92  */
93 #ifdef RTE_ARCH_X86_64
94 		rte_smp_wmb();
95 		dp->flags = flags;
96 #else
97 		__atomic_store_n(&dp->flags, flags, __ATOMIC_RELEASE);
98 #endif
99 	} else {
100 		rte_cio_wmb();
101 		dp->flags = flags;
102 	}
103 }
104 #ifdef RTE_PMD_PACKET_PREFETCH
105 #define rte_packet_prefetch(p)  rte_prefetch1(p)
106 #else
107 #define rte_packet_prefetch(p)  do {} while(0)
108 #endif
109 
110 #define VIRTQUEUE_MAX_NAME_SZ 32
111 
112 #ifdef RTE_VIRTIO_USER
113 /**
114  * Return the physical address (or virtual address in case of
115  * virtio-user) of mbuf data buffer.
116  *
117  * The address is firstly casted to the word size (sizeof(uintptr_t))
118  * before casting it to uint64_t. This is to make it work with different
119  * combination of word size (64 bit and 32 bit) and virtio device
120  * (virtio-pci and virtio-user).
121  */
122 #define VIRTIO_MBUF_ADDR(mb, vq) \
123 	((uint64_t)(*(uintptr_t *)((uintptr_t)(mb) + (vq)->offset)))
124 #else
125 #define VIRTIO_MBUF_ADDR(mb, vq) ((mb)->buf_iova)
126 #endif
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 } __attribute__((__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  * Control link announce acknowledgement
206  *
207  * The command VIRTIO_NET_CTRL_ANNOUNCE_ACK is used to indicate that
208  * driver has recevied the notification; device would clear the
209  * VIRTIO_NET_S_ANNOUNCE bit in the status field after it receives
210  * this command.
211  */
212 #define VIRTIO_NET_CTRL_ANNOUNCE     3
213 #define VIRTIO_NET_CTRL_ANNOUNCE_ACK 0
214 
215 struct virtio_net_ctrl_hdr {
216 	uint8_t class;
217 	uint8_t cmd;
218 } __attribute__((packed));
219 
220 typedef uint8_t virtio_net_ctrl_ack;
221 
222 #define VIRTIO_NET_OK     0
223 #define VIRTIO_NET_ERR    1
224 
225 #define VIRTIO_MAX_CTRL_DATA 2048
226 
227 struct virtio_pmd_ctrl {
228 	struct virtio_net_ctrl_hdr hdr;
229 	virtio_net_ctrl_ack status;
230 	uint8_t data[VIRTIO_MAX_CTRL_DATA];
231 };
232 
233 struct vq_desc_extra {
234 	void *cookie;
235 	uint16_t ndescs;
236 	uint16_t next;
237 };
238 
239 struct virtqueue {
240 	struct virtio_hw  *hw; /**< virtio_hw structure pointer. */
241 	union {
242 		struct {
243 			/**< vring keeping desc, used and avail */
244 			struct vring ring;
245 		} vq_split;
246 
247 		struct {
248 			/**< vring keeping descs and events */
249 			struct vring_packed ring;
250 			bool used_wrap_counter;
251 			uint16_t cached_flags; /**< cached flags for descs */
252 			uint16_t event_flags_shadow;
253 		} vq_packed;
254 	};
255 
256 	uint16_t vq_used_cons_idx; /**< last consumed descriptor */
257 	uint16_t vq_nentries;  /**< vring desc numbers */
258 	uint16_t vq_free_cnt;  /**< num of desc available */
259 	uint16_t vq_avail_idx; /**< sync until needed */
260 	uint16_t vq_free_thresh; /**< free threshold */
261 
262 	void *vq_ring_virt_mem;  /**< linear address of vring*/
263 	unsigned int vq_ring_size;
264 
265 	union {
266 		struct virtnet_rx rxq;
267 		struct virtnet_tx txq;
268 		struct virtnet_ctl cq;
269 	};
270 
271 	rte_iova_t vq_ring_mem; /**< physical address of vring,
272 	                         * or virtual address for virtio_user. */
273 
274 	/**
275 	 * Head of the free chain in the descriptor table. If
276 	 * there are no free descriptors, this will be set to
277 	 * VQ_RING_DESC_CHAIN_END.
278 	 */
279 	uint16_t  vq_desc_head_idx;
280 	uint16_t  vq_desc_tail_idx;
281 	uint16_t  vq_queue_index;   /**< PCI queue index */
282 	uint16_t offset; /**< relative offset to obtain addr in mbuf */
283 	uint16_t  *notify_addr;
284 	struct rte_mbuf **sw_ring;  /**< RX software ring. */
285 	struct vq_desc_extra vq_descx[0];
286 };
287 
288 /* If multiqueue is provided by host, then we suppport it. */
289 #define VIRTIO_NET_CTRL_MQ   4
290 #define VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET        0
291 #define VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MIN        1
292 #define VIRTIO_NET_CTRL_MQ_VQ_PAIRS_MAX        0x8000
293 
294 /**
295  * This is the first element of the scatter-gather list.  If you don't
296  * specify GSO or CSUM features, you can simply ignore the header.
297  */
298 struct virtio_net_hdr {
299 #define VIRTIO_NET_HDR_F_NEEDS_CSUM 1    /**< Use csum_start,csum_offset*/
300 #define VIRTIO_NET_HDR_F_DATA_VALID 2    /**< Checksum is valid */
301 	uint8_t flags;
302 #define VIRTIO_NET_HDR_GSO_NONE     0    /**< Not a GSO frame */
303 #define VIRTIO_NET_HDR_GSO_TCPV4    1    /**< GSO frame, IPv4 TCP (TSO) */
304 #define VIRTIO_NET_HDR_GSO_UDP      3    /**< GSO frame, IPv4 UDP (UFO) */
305 #define VIRTIO_NET_HDR_GSO_TCPV6    4    /**< GSO frame, IPv6 TCP */
306 #define VIRTIO_NET_HDR_GSO_ECN      0x80 /**< TCP has ECN set */
307 	uint8_t gso_type;
308 	uint16_t hdr_len;     /**< Ethernet + IP + tcp/udp hdrs */
309 	uint16_t gso_size;    /**< Bytes to append to hdr_len per frame */
310 	uint16_t csum_start;  /**< Position to start checksumming from */
311 	uint16_t csum_offset; /**< Offset after that to place checksum */
312 };
313 
314 /**
315  * This is the version of the header to use when the MRG_RXBUF
316  * feature has been negotiated.
317  */
318 struct virtio_net_hdr_mrg_rxbuf {
319 	struct   virtio_net_hdr hdr;
320 	uint16_t num_buffers; /**< Number of merged rx buffers */
321 };
322 
323 /* Region reserved to allow for transmit header and indirect ring */
324 #define VIRTIO_MAX_TX_INDIRECT 8
325 struct virtio_tx_region {
326 	struct virtio_net_hdr_mrg_rxbuf tx_hdr;
327 	struct vring_desc tx_indir[VIRTIO_MAX_TX_INDIRECT]
328 		__attribute__((__aligned__(16)));
329 };
330 
331 static inline int
332 desc_is_used(struct vring_packed_desc *desc, struct virtqueue *vq)
333 {
334 	uint16_t used, avail, flags;
335 
336 	flags = virtqueue_fetch_flags_packed(desc, vq->hw->weak_barriers);
337 	used = !!(flags & VRING_PACKED_DESC_F_USED);
338 	avail = !!(flags & VRING_PACKED_DESC_F_AVAIL);
339 
340 	return avail == used && used == vq->vq_packed.used_wrap_counter;
341 }
342 
343 static inline void
344 vring_desc_init_packed(struct virtqueue *vq, int n)
345 {
346 	int i;
347 	for (i = 0; i < n - 1; i++) {
348 		vq->vq_packed.ring.desc[i].id = i;
349 		vq->vq_descx[i].next = i + 1;
350 	}
351 	vq->vq_packed.ring.desc[i].id = i;
352 	vq->vq_descx[i].next = VQ_RING_DESC_CHAIN_END;
353 }
354 
355 /* Chain all the descriptors in the ring with an END */
356 static inline void
357 vring_desc_init_split(struct vring_desc *dp, uint16_t n)
358 {
359 	uint16_t i;
360 
361 	for (i = 0; i < n - 1; i++)
362 		dp[i].next = (uint16_t)(i + 1);
363 	dp[i].next = VQ_RING_DESC_CHAIN_END;
364 }
365 
366 /**
367  * Tell the backend not to interrupt us. Implementation for packed virtqueues.
368  */
369 static inline void
370 virtqueue_disable_intr_packed(struct virtqueue *vq)
371 {
372 	if (vq->vq_packed.event_flags_shadow != RING_EVENT_FLAGS_DISABLE) {
373 		vq->vq_packed.event_flags_shadow = RING_EVENT_FLAGS_DISABLE;
374 		vq->vq_packed.ring.driver->desc_event_flags =
375 			vq->vq_packed.event_flags_shadow;
376 	}
377 }
378 
379 /**
380  * Tell the backend not to interrupt us. Implementation for split virtqueues.
381  */
382 static inline void
383 virtqueue_disable_intr_split(struct virtqueue *vq)
384 {
385 	vq->vq_split.ring.avail->flags |= VRING_AVAIL_F_NO_INTERRUPT;
386 }
387 
388 /**
389  * Tell the backend not to interrupt us.
390  */
391 static inline void
392 virtqueue_disable_intr(struct virtqueue *vq)
393 {
394 	if (vtpci_packed_queue(vq->hw))
395 		virtqueue_disable_intr_packed(vq);
396 	else
397 		virtqueue_disable_intr_split(vq);
398 }
399 
400 /**
401  * Tell the backend to interrupt. Implementation for packed virtqueues.
402  */
403 static inline void
404 virtqueue_enable_intr_packed(struct virtqueue *vq)
405 {
406 	if (vq->vq_packed.event_flags_shadow == RING_EVENT_FLAGS_DISABLE) {
407 		vq->vq_packed.event_flags_shadow = RING_EVENT_FLAGS_ENABLE;
408 		vq->vq_packed.ring.driver->desc_event_flags =
409 			vq->vq_packed.event_flags_shadow;
410 	}
411 }
412 
413 /**
414  * Tell the backend to interrupt. Implementation for split virtqueues.
415  */
416 static inline void
417 virtqueue_enable_intr_split(struct virtqueue *vq)
418 {
419 	vq->vq_split.ring.avail->flags &= (~VRING_AVAIL_F_NO_INTERRUPT);
420 }
421 
422 /**
423  * Tell the backend to interrupt us.
424  */
425 static inline void
426 virtqueue_enable_intr(struct virtqueue *vq)
427 {
428 	if (vtpci_packed_queue(vq->hw))
429 		virtqueue_enable_intr_packed(vq);
430 	else
431 		virtqueue_enable_intr_split(vq);
432 }
433 
434 /**
435  *  Dump virtqueue internal structures, for debug purpose only.
436  */
437 void virtqueue_dump(struct virtqueue *vq);
438 /**
439  *  Get all mbufs to be freed.
440  */
441 struct rte_mbuf *virtqueue_detach_unused(struct virtqueue *vq);
442 
443 /* Flush the elements in the used ring. */
444 void virtqueue_rxvq_flush(struct virtqueue *vq);
445 
446 static inline int
447 virtqueue_full(const struct virtqueue *vq)
448 {
449 	return vq->vq_free_cnt == 0;
450 }
451 
452 static inline int
453 virtio_get_queue_type(struct virtio_hw *hw, uint16_t vtpci_queue_idx)
454 {
455 	if (vtpci_queue_idx == hw->max_queue_pairs * 2)
456 		return VTNET_CQ;
457 	else if (vtpci_queue_idx % 2 == 0)
458 		return VTNET_RQ;
459 	else
460 		return VTNET_TQ;
461 }
462 
463 #define VIRTQUEUE_NUSED(vq) ((uint16_t)((vq)->vq_split.ring.used->idx - \
464 					(vq)->vq_used_cons_idx))
465 
466 void vq_ring_free_chain(struct virtqueue *vq, uint16_t desc_idx);
467 void vq_ring_free_chain_packed(struct virtqueue *vq, uint16_t used_idx);
468 void vq_ring_free_inorder(struct virtqueue *vq, uint16_t desc_idx,
469 			  uint16_t num);
470 
471 static inline void
472 vq_update_avail_idx(struct virtqueue *vq)
473 {
474 	virtio_wmb(vq->hw->weak_barriers);
475 	vq->vq_split.ring.avail->idx = vq->vq_avail_idx;
476 }
477 
478 static inline void
479 vq_update_avail_ring(struct virtqueue *vq, uint16_t desc_idx)
480 {
481 	uint16_t avail_idx;
482 	/*
483 	 * Place the head of the descriptor chain into the next slot and make
484 	 * it usable to the host. The chain is made available now rather than
485 	 * deferring to virtqueue_notify() in the hopes that if the host is
486 	 * currently running on another CPU, we can keep it processing the new
487 	 * descriptor.
488 	 */
489 	avail_idx = (uint16_t)(vq->vq_avail_idx & (vq->vq_nentries - 1));
490 	if (unlikely(vq->vq_split.ring.avail->ring[avail_idx] != desc_idx))
491 		vq->vq_split.ring.avail->ring[avail_idx] = desc_idx;
492 	vq->vq_avail_idx++;
493 }
494 
495 static inline int
496 virtqueue_kick_prepare(struct virtqueue *vq)
497 {
498 	/*
499 	 * Ensure updated avail->idx is visible to vhost before reading
500 	 * the used->flags.
501 	 */
502 	virtio_mb(vq->hw->weak_barriers);
503 	return !(vq->vq_split.ring.used->flags & VRING_USED_F_NO_NOTIFY);
504 }
505 
506 static inline int
507 virtqueue_kick_prepare_packed(struct virtqueue *vq)
508 {
509 	uint16_t flags;
510 
511 	/*
512 	 * Ensure updated data is visible to vhost before reading the flags.
513 	 */
514 	virtio_mb(vq->hw->weak_barriers);
515 	flags = vq->vq_packed.ring.device->desc_event_flags;
516 
517 	return flags != RING_EVENT_FLAGS_DISABLE;
518 }
519 
520 /*
521  * virtqueue_kick_prepare*() or the virtio_wmb() should be called
522  * before this function to be sure that all the data is visible to vhost.
523  */
524 static inline void
525 virtqueue_notify(struct virtqueue *vq)
526 {
527 	VTPCI_OPS(vq->hw)->notify_queue(vq->hw, vq);
528 }
529 
530 #ifdef RTE_LIBRTE_VIRTIO_DEBUG_DUMP
531 #define VIRTQUEUE_DUMP(vq) do { \
532 	uint16_t used_idx, nused; \
533 	used_idx = (vq)->vq_split.ring.used->idx; \
534 	nused = (uint16_t)(used_idx - (vq)->vq_used_cons_idx); \
535 	if (vtpci_packed_queue((vq)->hw)) { \
536 		PMD_INIT_LOG(DEBUG, \
537 		"VQ: - size=%d; free=%d; used_cons_idx=%d; avail_idx=%d;" \
538 		" cached_flags=0x%x; used_wrap_counter=%d", \
539 		(vq)->vq_nentries, (vq)->vq_free_cnt, (vq)->vq_used_cons_idx, \
540 		(vq)->vq_avail_idx, (vq)->vq_packed.cached_flags, \
541 		(vq)->vq_packed.used_wrap_counter); \
542 		break; \
543 	} \
544 	PMD_INIT_LOG(DEBUG, \
545 	  "VQ: - size=%d; free=%d; used=%d; desc_head_idx=%d;" \
546 	  " avail.idx=%d; used_cons_idx=%d; used.idx=%d;" \
547 	  " avail.flags=0x%x; used.flags=0x%x", \
548 	  (vq)->vq_nentries, (vq)->vq_free_cnt, nused, \
549 	  (vq)->vq_desc_head_idx, (vq)->vq_split.ring.avail->idx, \
550 	  (vq)->vq_used_cons_idx, (vq)->vq_split.ring.used->idx, \
551 	  (vq)->vq_split.ring.avail->flags, (vq)->vq_split.ring.used->flags); \
552 } while (0)
553 #else
554 #define VIRTQUEUE_DUMP(vq) do { } while (0)
555 #endif
556 
557 #endif /* _VIRTQUEUE_H_ */
558