1 // SPDX-License-Identifier: GPL-2.0-only
3 * Copyright (c) 2009, Microsoft Corporation.
6 * Haiyang Zhang <haiyangz@microsoft.com>
7 * Hank Janssen <hjanssen@microsoft.com>
9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11 #include <linux/kernel.h>
12 #include <linux/sched.h>
13 #include <linux/wait.h>
15 #include <linux/delay.h>
17 #include <linux/slab.h>
18 #include <linux/netdevice.h>
19 #include <linux/if_ether.h>
20 #include <linux/vmalloc.h>
21 #include <linux/rtnetlink.h>
22 #include <linux/prefetch.h>
23 #include <linux/filter.h>
25 #include <asm/sync_bitops.h>
26 #include <asm/mshyperv.h>
28 #include "hyperv_net.h"
29 #include "netvsc_trace.h"
32 * Switch the data path from the synthetic interface to the VF
35 int netvsc_switch_datapath(struct net_device *ndev, bool vf)
37 struct net_device_context *net_device_ctx = netdev_priv(ndev);
38 struct hv_device *dev = net_device_ctx->device_ctx;
39 struct netvsc_device *nv_dev = rtnl_dereference(net_device_ctx->nvdev);
40 struct nvsp_message *init_pkt = &nv_dev->channel_init_pkt;
43 /* Block sending traffic to VF if it's about to be gone */
45 net_device_ctx->data_path_is_vf = vf;
47 memset(init_pkt, 0, sizeof(struct nvsp_message));
48 init_pkt->hdr.msg_type = NVSP_MSG4_TYPE_SWITCH_DATA_PATH;
50 init_pkt->msg.v4_msg.active_dp.active_datapath =
53 init_pkt->msg.v4_msg.active_dp.active_datapath =
54 NVSP_DATAPATH_SYNTHETIC;
57 trace_nvsp_send(ndev, init_pkt);
59 ret = vmbus_sendpacket(dev->channel, init_pkt,
60 sizeof(struct nvsp_message),
61 (unsigned long)init_pkt, VM_PKT_DATA_INBAND,
62 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
64 /* If failed to switch to/from VF, let data_path_is_vf stay false,
65 * so we use synthetic path to send data.
70 "Unable to send sw datapath msg, err: %d\n",
75 if (retry++ < RETRY_MAX) {
76 usleep_range(RETRY_US_LO, RETRY_US_HI);
81 "Retry failed to send sw datapath msg, err: %d\n",
87 wait_for_completion(&nv_dev->channel_init_wait);
88 net_device_ctx->data_path_is_vf = vf;
93 /* Worker to setup sub channels on initial setup
94 * Initial hotplug event occurs in softirq context
95 * and can't wait for channels.
97 static void netvsc_subchan_work(struct work_struct *w)
99 struct netvsc_device *nvdev =
100 container_of(w, struct netvsc_device, subchan_work);
101 struct rndis_device *rdev;
104 /* Avoid deadlock with device removal already under RTNL */
105 if (!rtnl_trylock()) {
110 rdev = nvdev->extension;
112 ret = rndis_set_subchannel(rdev->ndev, nvdev, NULL);
114 netif_device_attach(rdev->ndev);
116 /* fallback to only primary channel */
117 for (i = 1; i < nvdev->num_chn; i++)
118 netif_napi_del(&nvdev->chan_table[i].napi);
128 static struct netvsc_device *alloc_net_device(void)
130 struct netvsc_device *net_device;
132 net_device = kzalloc(sizeof(struct netvsc_device), GFP_KERNEL);
136 init_waitqueue_head(&net_device->wait_drain);
137 net_device->destroy = false;
138 net_device->tx_disable = true;
140 net_device->max_pkt = RNDIS_MAX_PKT_DEFAULT;
141 net_device->pkt_align = RNDIS_PKT_ALIGN_DEFAULT;
143 init_completion(&net_device->channel_init_wait);
144 init_waitqueue_head(&net_device->subchan_open);
145 INIT_WORK(&net_device->subchan_work, netvsc_subchan_work);
150 static void free_netvsc_device(struct rcu_head *head)
152 struct netvsc_device *nvdev
153 = container_of(head, struct netvsc_device, rcu);
156 kfree(nvdev->extension);
158 if (!nvdev->recv_buf_gpadl_handle.decrypted)
159 vfree(nvdev->recv_buf);
160 if (!nvdev->send_buf_gpadl_handle.decrypted)
161 vfree(nvdev->send_buf);
162 bitmap_free(nvdev->send_section_map);
164 for (i = 0; i < VRSS_CHANNEL_MAX; i++) {
165 xdp_rxq_info_unreg(&nvdev->chan_table[i].xdp_rxq);
166 kfree(nvdev->chan_table[i].recv_buf);
167 vfree(nvdev->chan_table[i].mrc.slots);
173 static void free_netvsc_device_rcu(struct netvsc_device *nvdev)
175 call_rcu(&nvdev->rcu, free_netvsc_device);
178 static void netvsc_revoke_recv_buf(struct hv_device *device,
179 struct netvsc_device *net_device,
180 struct net_device *ndev)
182 struct nvsp_message *revoke_packet;
186 * If we got a section count, it means we received a
187 * SendReceiveBufferComplete msg (ie sent
188 * NvspMessage1TypeSendReceiveBuffer msg) therefore, we need
189 * to send a revoke msg here
191 if (net_device->recv_section_cnt) {
192 /* Send the revoke receive buffer */
193 revoke_packet = &net_device->revoke_packet;
194 memset(revoke_packet, 0, sizeof(struct nvsp_message));
196 revoke_packet->hdr.msg_type =
197 NVSP_MSG1_TYPE_REVOKE_RECV_BUF;
198 revoke_packet->msg.v1_msg.
199 revoke_recv_buf.id = NETVSC_RECEIVE_BUFFER_ID;
201 trace_nvsp_send(ndev, revoke_packet);
203 ret = vmbus_sendpacket(device->channel,
205 sizeof(struct nvsp_message),
206 VMBUS_RQST_ID_NO_RESPONSE,
207 VM_PKT_DATA_INBAND, 0);
208 /* If the failure is because the channel is rescinded;
209 * ignore the failure since we cannot send on a rescinded
210 * channel. This would allow us to properly cleanup
211 * even when the channel is rescinded.
213 if (device->channel->rescind)
216 * If we failed here, we might as well return and
217 * have a leak rather than continue and a bugchk
220 netdev_err(ndev, "unable to send "
221 "revoke receive buffer to netvsp\n");
224 net_device->recv_section_cnt = 0;
228 static void netvsc_revoke_send_buf(struct hv_device *device,
229 struct netvsc_device *net_device,
230 struct net_device *ndev)
232 struct nvsp_message *revoke_packet;
235 /* Deal with the send buffer we may have setup.
236 * If we got a send section size, it means we received a
237 * NVSP_MSG1_TYPE_SEND_SEND_BUF_COMPLETE msg (ie sent
238 * NVSP_MSG1_TYPE_SEND_SEND_BUF msg) therefore, we need
239 * to send a revoke msg here
241 if (net_device->send_section_cnt) {
242 /* Send the revoke receive buffer */
243 revoke_packet = &net_device->revoke_packet;
244 memset(revoke_packet, 0, sizeof(struct nvsp_message));
246 revoke_packet->hdr.msg_type =
247 NVSP_MSG1_TYPE_REVOKE_SEND_BUF;
248 revoke_packet->msg.v1_msg.revoke_send_buf.id =
249 NETVSC_SEND_BUFFER_ID;
251 trace_nvsp_send(ndev, revoke_packet);
253 ret = vmbus_sendpacket(device->channel,
255 sizeof(struct nvsp_message),
256 VMBUS_RQST_ID_NO_RESPONSE,
257 VM_PKT_DATA_INBAND, 0);
259 /* If the failure is because the channel is rescinded;
260 * ignore the failure since we cannot send on a rescinded
261 * channel. This would allow us to properly cleanup
262 * even when the channel is rescinded.
264 if (device->channel->rescind)
267 /* If we failed here, we might as well return and
268 * have a leak rather than continue and a bugchk
271 netdev_err(ndev, "unable to send "
272 "revoke send buffer to netvsp\n");
275 net_device->send_section_cnt = 0;
279 static void netvsc_teardown_recv_gpadl(struct hv_device *device,
280 struct netvsc_device *net_device,
281 struct net_device *ndev)
285 if (net_device->recv_buf_gpadl_handle.gpadl_handle) {
286 ret = vmbus_teardown_gpadl(device->channel,
287 &net_device->recv_buf_gpadl_handle);
289 /* If we failed here, we might as well return and have a leak
290 * rather than continue and a bugchk
294 "unable to teardown receive buffer's gpadl\n");
300 static void netvsc_teardown_send_gpadl(struct hv_device *device,
301 struct netvsc_device *net_device,
302 struct net_device *ndev)
306 if (net_device->send_buf_gpadl_handle.gpadl_handle) {
307 ret = vmbus_teardown_gpadl(device->channel,
308 &net_device->send_buf_gpadl_handle);
310 /* If we failed here, we might as well return and have a leak
311 * rather than continue and a bugchk
315 "unable to teardown send buffer's gpadl\n");
321 int netvsc_alloc_recv_comp_ring(struct netvsc_device *net_device, u32 q_idx)
323 struct netvsc_channel *nvchan = &net_device->chan_table[q_idx];
324 int node = cpu_to_node(nvchan->channel->target_cpu);
327 size = net_device->recv_completion_cnt * sizeof(struct recv_comp_data);
328 nvchan->mrc.slots = vzalloc_node(size, node);
329 if (!nvchan->mrc.slots)
330 nvchan->mrc.slots = vzalloc(size);
332 return nvchan->mrc.slots ? 0 : -ENOMEM;
335 static int netvsc_init_buf(struct hv_device *device,
336 struct netvsc_device *net_device,
337 const struct netvsc_device_info *device_info)
339 struct nvsp_1_message_send_receive_buffer_complete *resp;
340 struct net_device *ndev = hv_get_drvdata(device);
341 struct nvsp_message *init_packet;
342 unsigned int buf_size;
345 /* Get receive buffer area. */
346 buf_size = device_info->recv_sections * device_info->recv_section_size;
347 buf_size = roundup(buf_size, PAGE_SIZE);
349 /* Legacy hosts only allow smaller receive buffer */
350 if (net_device->nvsp_version <= NVSP_PROTOCOL_VERSION_2)
351 buf_size = min_t(unsigned int, buf_size,
352 NETVSC_RECEIVE_BUFFER_SIZE_LEGACY);
354 net_device->recv_buf = vzalloc(buf_size);
355 if (!net_device->recv_buf) {
357 "unable to allocate receive buffer of size %u\n",
363 net_device->recv_buf_size = buf_size;
366 * Establish the gpadl handle for this buffer on this
367 * channel. Note: This call uses the vmbus connection rather
368 * than the channel to establish the gpadl handle.
370 ret = vmbus_establish_gpadl(device->channel, net_device->recv_buf,
372 &net_device->recv_buf_gpadl_handle);
375 "unable to establish receive buffer's gpadl\n");
379 /* Notify the NetVsp of the gpadl handle */
380 init_packet = &net_device->channel_init_pkt;
381 memset(init_packet, 0, sizeof(struct nvsp_message));
382 init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_RECV_BUF;
383 init_packet->msg.v1_msg.send_recv_buf.
384 gpadl_handle = net_device->recv_buf_gpadl_handle.gpadl_handle;
385 init_packet->msg.v1_msg.
386 send_recv_buf.id = NETVSC_RECEIVE_BUFFER_ID;
388 trace_nvsp_send(ndev, init_packet);
390 /* Send the gpadl notification request */
391 ret = vmbus_sendpacket(device->channel, init_packet,
392 sizeof(struct nvsp_message),
393 (unsigned long)init_packet,
395 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
398 "unable to send receive buffer's gpadl to netvsp\n");
402 wait_for_completion(&net_device->channel_init_wait);
404 /* Check the response */
405 resp = &init_packet->msg.v1_msg.send_recv_buf_complete;
406 if (resp->status != NVSP_STAT_SUCCESS) {
408 "Unable to complete receive buffer initialization with NetVsp - status %d\n",
414 /* Parse the response */
415 netdev_dbg(ndev, "Receive sections: %u sub_allocs: size %u count: %u\n",
416 resp->num_sections, resp->sections[0].sub_alloc_size,
417 resp->sections[0].num_sub_allocs);
419 /* There should only be one section for the entire receive buffer */
420 if (resp->num_sections != 1 || resp->sections[0].offset != 0) {
425 net_device->recv_section_size = resp->sections[0].sub_alloc_size;
426 net_device->recv_section_cnt = resp->sections[0].num_sub_allocs;
428 /* Ensure buffer will not overflow */
429 if (net_device->recv_section_size < NETVSC_MTU_MIN || (u64)net_device->recv_section_size *
430 (u64)net_device->recv_section_cnt > (u64)buf_size) {
431 netdev_err(ndev, "invalid recv_section_size %u\n",
432 net_device->recv_section_size);
437 for (i = 0; i < VRSS_CHANNEL_MAX; i++) {
438 struct netvsc_channel *nvchan = &net_device->chan_table[i];
440 nvchan->recv_buf = kzalloc(net_device->recv_section_size, GFP_KERNEL);
441 if (nvchan->recv_buf == NULL) {
447 /* Setup receive completion ring.
448 * Add 1 to the recv_section_cnt because at least one entry in a
449 * ring buffer has to be empty.
451 net_device->recv_completion_cnt = net_device->recv_section_cnt + 1;
452 ret = netvsc_alloc_recv_comp_ring(net_device, 0);
456 /* Now setup the send buffer. */
457 buf_size = device_info->send_sections * device_info->send_section_size;
458 buf_size = round_up(buf_size, PAGE_SIZE);
460 net_device->send_buf = vzalloc(buf_size);
461 if (!net_device->send_buf) {
462 netdev_err(ndev, "unable to allocate send buffer of size %u\n",
467 net_device->send_buf_size = buf_size;
469 /* Establish the gpadl handle for this buffer on this
470 * channel. Note: This call uses the vmbus connection rather
471 * than the channel to establish the gpadl handle.
473 ret = vmbus_establish_gpadl(device->channel, net_device->send_buf,
475 &net_device->send_buf_gpadl_handle);
478 "unable to establish send buffer's gpadl\n");
482 /* Notify the NetVsp of the gpadl handle */
483 init_packet = &net_device->channel_init_pkt;
484 memset(init_packet, 0, sizeof(struct nvsp_message));
485 init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_SEND_BUF;
486 init_packet->msg.v1_msg.send_send_buf.gpadl_handle =
487 net_device->send_buf_gpadl_handle.gpadl_handle;
488 init_packet->msg.v1_msg.send_send_buf.id = NETVSC_SEND_BUFFER_ID;
490 trace_nvsp_send(ndev, init_packet);
492 /* Send the gpadl notification request */
493 ret = vmbus_sendpacket(device->channel, init_packet,
494 sizeof(struct nvsp_message),
495 (unsigned long)init_packet,
497 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
500 "unable to send send buffer's gpadl to netvsp\n");
504 wait_for_completion(&net_device->channel_init_wait);
506 /* Check the response */
507 if (init_packet->msg.v1_msg.
508 send_send_buf_complete.status != NVSP_STAT_SUCCESS) {
509 netdev_err(ndev, "Unable to complete send buffer "
510 "initialization with NetVsp - status %d\n",
511 init_packet->msg.v1_msg.
512 send_send_buf_complete.status);
517 /* Parse the response */
518 net_device->send_section_size = init_packet->msg.
519 v1_msg.send_send_buf_complete.section_size;
520 if (net_device->send_section_size < NETVSC_MTU_MIN) {
521 netdev_err(ndev, "invalid send_section_size %u\n",
522 net_device->send_section_size);
527 /* Section count is simply the size divided by the section size. */
528 net_device->send_section_cnt = buf_size / net_device->send_section_size;
530 netdev_dbg(ndev, "Send section size: %d, Section count:%d\n",
531 net_device->send_section_size, net_device->send_section_cnt);
533 /* Setup state for managing the send buffer. */
534 net_device->send_section_map = bitmap_zalloc(net_device->send_section_cnt,
536 if (!net_device->send_section_map) {
544 netvsc_revoke_recv_buf(device, net_device, ndev);
545 netvsc_revoke_send_buf(device, net_device, ndev);
546 netvsc_teardown_recv_gpadl(device, net_device, ndev);
547 netvsc_teardown_send_gpadl(device, net_device, ndev);
553 /* Negotiate NVSP protocol version */
554 static int negotiate_nvsp_ver(struct hv_device *device,
555 struct netvsc_device *net_device,
556 struct nvsp_message *init_packet,
559 struct net_device *ndev = hv_get_drvdata(device);
562 memset(init_packet, 0, sizeof(struct nvsp_message));
563 init_packet->hdr.msg_type = NVSP_MSG_TYPE_INIT;
564 init_packet->msg.init_msg.init.min_protocol_ver = nvsp_ver;
565 init_packet->msg.init_msg.init.max_protocol_ver = nvsp_ver;
566 trace_nvsp_send(ndev, init_packet);
568 /* Send the init request */
569 ret = vmbus_sendpacket(device->channel, init_packet,
570 sizeof(struct nvsp_message),
571 (unsigned long)init_packet,
573 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
578 wait_for_completion(&net_device->channel_init_wait);
580 if (init_packet->msg.init_msg.init_complete.status !=
584 if (nvsp_ver == NVSP_PROTOCOL_VERSION_1)
587 /* NVSPv2 or later: Send NDIS config */
588 memset(init_packet, 0, sizeof(struct nvsp_message));
589 init_packet->hdr.msg_type = NVSP_MSG2_TYPE_SEND_NDIS_CONFIG;
590 init_packet->msg.v2_msg.send_ndis_config.mtu = ndev->mtu + ETH_HLEN;
591 init_packet->msg.v2_msg.send_ndis_config.capability.ieee8021q = 1;
593 if (nvsp_ver >= NVSP_PROTOCOL_VERSION_5) {
594 if (hv_is_isolation_supported())
595 netdev_info(ndev, "SR-IOV not advertised by guests on the host supporting isolation\n");
597 init_packet->msg.v2_msg.send_ndis_config.capability.sriov = 1;
599 /* Teaming bit is needed to receive link speed updates */
600 init_packet->msg.v2_msg.send_ndis_config.capability.teaming = 1;
603 if (nvsp_ver >= NVSP_PROTOCOL_VERSION_61)
604 init_packet->msg.v2_msg.send_ndis_config.capability.rsc = 1;
606 trace_nvsp_send(ndev, init_packet);
608 ret = vmbus_sendpacket(device->channel, init_packet,
609 sizeof(struct nvsp_message),
610 VMBUS_RQST_ID_NO_RESPONSE,
611 VM_PKT_DATA_INBAND, 0);
616 static int netvsc_connect_vsp(struct hv_device *device,
617 struct netvsc_device *net_device,
618 const struct netvsc_device_info *device_info)
620 struct net_device *ndev = hv_get_drvdata(device);
621 static const u32 ver_list[] = {
622 NVSP_PROTOCOL_VERSION_1, NVSP_PROTOCOL_VERSION_2,
623 NVSP_PROTOCOL_VERSION_4, NVSP_PROTOCOL_VERSION_5,
624 NVSP_PROTOCOL_VERSION_6, NVSP_PROTOCOL_VERSION_61
626 struct nvsp_message *init_packet;
627 int ndis_version, i, ret;
629 init_packet = &net_device->channel_init_pkt;
631 /* Negotiate the latest NVSP protocol supported */
632 for (i = ARRAY_SIZE(ver_list) - 1; i >= 0; i--)
633 if (negotiate_nvsp_ver(device, net_device, init_packet,
635 net_device->nvsp_version = ver_list[i];
644 if (hv_is_isolation_supported() && net_device->nvsp_version < NVSP_PROTOCOL_VERSION_61) {
645 netdev_err(ndev, "Invalid NVSP version 0x%x (expected >= 0x%x) from the host supporting isolation\n",
646 net_device->nvsp_version, NVSP_PROTOCOL_VERSION_61);
651 pr_debug("Negotiated NVSP version:%x\n", net_device->nvsp_version);
653 /* Send the ndis version */
654 memset(init_packet, 0, sizeof(struct nvsp_message));
656 if (net_device->nvsp_version <= NVSP_PROTOCOL_VERSION_4)
657 ndis_version = 0x00060001;
659 ndis_version = 0x0006001e;
661 init_packet->hdr.msg_type = NVSP_MSG1_TYPE_SEND_NDIS_VER;
662 init_packet->msg.v1_msg.
663 send_ndis_ver.ndis_major_ver =
664 (ndis_version & 0xFFFF0000) >> 16;
665 init_packet->msg.v1_msg.
666 send_ndis_ver.ndis_minor_ver =
667 ndis_version & 0xFFFF;
669 trace_nvsp_send(ndev, init_packet);
671 /* Send the init request */
672 ret = vmbus_sendpacket(device->channel, init_packet,
673 sizeof(struct nvsp_message),
674 VMBUS_RQST_ID_NO_RESPONSE,
675 VM_PKT_DATA_INBAND, 0);
680 ret = netvsc_init_buf(device, net_device, device_info);
687 * netvsc_device_remove - Callback when the root bus device is removed
689 void netvsc_device_remove(struct hv_device *device)
691 struct net_device *ndev = hv_get_drvdata(device);
692 struct net_device_context *net_device_ctx = netdev_priv(ndev);
693 struct netvsc_device *net_device
694 = rtnl_dereference(net_device_ctx->nvdev);
698 * Revoke receive buffer. If host is pre-Win2016 then tear down
699 * receive buffer GPADL. Do the same for send buffer.
701 netvsc_revoke_recv_buf(device, net_device, ndev);
702 if (vmbus_proto_version < VERSION_WIN10)
703 netvsc_teardown_recv_gpadl(device, net_device, ndev);
705 netvsc_revoke_send_buf(device, net_device, ndev);
706 if (vmbus_proto_version < VERSION_WIN10)
707 netvsc_teardown_send_gpadl(device, net_device, ndev);
709 RCU_INIT_POINTER(net_device_ctx->nvdev, NULL);
711 /* Disable NAPI and disassociate its context from the device. */
712 for (i = 0; i < net_device->num_chn; i++) {
713 /* See also vmbus_reset_channel_cb(). */
714 /* only disable enabled NAPI channel */
715 if (i < ndev->real_num_rx_queues)
716 napi_disable(&net_device->chan_table[i].napi);
718 netif_napi_del(&net_device->chan_table[i].napi);
722 * At this point, no one should be accessing net_device
725 netdev_dbg(ndev, "net device safe to remove\n");
727 /* Now, we can close the channel safely */
728 vmbus_close(device->channel);
731 * If host is Win2016 or higher then we do the GPADL tear down
732 * here after VMBus is closed.
734 if (vmbus_proto_version >= VERSION_WIN10) {
735 netvsc_teardown_recv_gpadl(device, net_device, ndev);
736 netvsc_teardown_send_gpadl(device, net_device, ndev);
739 /* Release all resources */
740 free_netvsc_device_rcu(net_device);
743 #define RING_AVAIL_PERCENT_HIWATER 20
744 #define RING_AVAIL_PERCENT_LOWATER 10
746 static inline void netvsc_free_send_slot(struct netvsc_device *net_device,
749 sync_change_bit(index, net_device->send_section_map);
752 static void netvsc_send_tx_complete(struct net_device *ndev,
753 struct netvsc_device *net_device,
754 struct vmbus_channel *channel,
755 const struct vmpacket_descriptor *desc,
758 struct net_device_context *ndev_ctx = netdev_priv(ndev);
764 cmd_rqst = channel->request_addr_callback(channel, desc->trans_id);
765 if (cmd_rqst == VMBUS_RQST_ERROR) {
766 netdev_err(ndev, "Invalid transaction ID %llx\n", desc->trans_id);
770 skb = (struct sk_buff *)(unsigned long)cmd_rqst;
772 /* Notify the layer above us */
774 struct hv_netvsc_packet *packet
775 = (struct hv_netvsc_packet *)skb->cb;
776 u32 send_index = packet->send_buf_index;
777 struct netvsc_stats_tx *tx_stats;
779 if (send_index != NETVSC_INVALID_INDEX)
780 netvsc_free_send_slot(net_device, send_index);
781 q_idx = packet->q_idx;
783 tx_stats = &net_device->chan_table[q_idx].tx_stats;
785 u64_stats_update_begin(&tx_stats->syncp);
786 tx_stats->packets += packet->total_packets;
787 tx_stats->bytes += packet->total_bytes;
788 u64_stats_update_end(&tx_stats->syncp);
790 netvsc_dma_unmap(ndev_ctx->device_ctx, packet);
791 napi_consume_skb(skb, budget);
795 atomic_dec_return(&net_device->chan_table[q_idx].queue_sends);
797 if (unlikely(net_device->destroy)) {
798 if (queue_sends == 0)
799 wake_up(&net_device->wait_drain);
801 struct netdev_queue *txq = netdev_get_tx_queue(ndev, q_idx);
803 if (netif_tx_queue_stopped(txq) && !net_device->tx_disable &&
804 (hv_get_avail_to_write_percent(&channel->outbound) >
805 RING_AVAIL_PERCENT_HIWATER || queue_sends < 1)) {
806 netif_tx_wake_queue(txq);
807 ndev_ctx->eth_stats.wake_queue++;
812 static void netvsc_send_completion(struct net_device *ndev,
813 struct netvsc_device *net_device,
814 struct vmbus_channel *incoming_channel,
815 const struct vmpacket_descriptor *desc,
818 const struct nvsp_message *nvsp_packet;
819 u32 msglen = hv_pkt_datalen(desc);
820 struct nvsp_message *pkt_rqst;
824 /* First check if this is a VMBUS completion without data payload */
826 cmd_rqst = incoming_channel->request_addr_callback(incoming_channel,
828 if (cmd_rqst == VMBUS_RQST_ERROR) {
829 netdev_err(ndev, "Invalid transaction ID %llx\n", desc->trans_id);
833 pkt_rqst = (struct nvsp_message *)(uintptr_t)cmd_rqst;
834 switch (pkt_rqst->hdr.msg_type) {
835 case NVSP_MSG4_TYPE_SWITCH_DATA_PATH:
836 complete(&net_device->channel_init_wait);
840 netdev_err(ndev, "Unexpected VMBUS completion!!\n");
845 /* Ensure packet is big enough to read header fields */
846 if (msglen < sizeof(struct nvsp_message_header)) {
847 netdev_err(ndev, "nvsp_message length too small: %u\n", msglen);
851 nvsp_packet = hv_pkt_data(desc);
852 switch (nvsp_packet->hdr.msg_type) {
853 case NVSP_MSG_TYPE_INIT_COMPLETE:
854 if (msglen < sizeof(struct nvsp_message_header) +
855 sizeof(struct nvsp_message_init_complete)) {
856 netdev_err(ndev, "nvsp_msg length too small: %u\n",
862 case NVSP_MSG1_TYPE_SEND_RECV_BUF_COMPLETE:
863 if (msglen < sizeof(struct nvsp_message_header) +
864 sizeof(struct nvsp_1_message_send_receive_buffer_complete)) {
865 netdev_err(ndev, "nvsp_msg1 length too small: %u\n",
871 case NVSP_MSG1_TYPE_SEND_SEND_BUF_COMPLETE:
872 if (msglen < sizeof(struct nvsp_message_header) +
873 sizeof(struct nvsp_1_message_send_send_buffer_complete)) {
874 netdev_err(ndev, "nvsp_msg1 length too small: %u\n",
880 case NVSP_MSG5_TYPE_SUBCHANNEL:
881 if (msglen < sizeof(struct nvsp_message_header) +
882 sizeof(struct nvsp_5_subchannel_complete)) {
883 netdev_err(ndev, "nvsp_msg5 length too small: %u\n",
889 case NVSP_MSG1_TYPE_SEND_RNDIS_PKT_COMPLETE:
890 if (msglen < sizeof(struct nvsp_message_header) +
891 sizeof(struct nvsp_1_message_send_rndis_packet_complete)) {
893 netdev_err(ndev, "nvsp_rndis_pkt_complete length too small: %u\n",
898 /* If status indicates an error, output a message so we know
899 * there's a problem. But process the completion anyway so the
900 * resources are released.
902 status = nvsp_packet->msg.v1_msg.send_rndis_pkt_complete.status;
903 if (status != NVSP_STAT_SUCCESS && net_ratelimit())
904 netdev_err(ndev, "nvsp_rndis_pkt_complete error status: %x\n",
907 netvsc_send_tx_complete(ndev, net_device, incoming_channel,
913 "Unknown send completion type %d received!!\n",
914 nvsp_packet->hdr.msg_type);
918 /* Copy the response back */
919 memcpy(&net_device->channel_init_pkt, nvsp_packet,
920 sizeof(struct nvsp_message));
921 complete(&net_device->channel_init_wait);
924 static u32 netvsc_get_next_send_section(struct netvsc_device *net_device)
926 unsigned long *map_addr = net_device->send_section_map;
929 for_each_clear_bit(i, map_addr, net_device->send_section_cnt) {
930 if (sync_test_and_set_bit(i, map_addr) == 0)
934 return NETVSC_INVALID_INDEX;
937 static void netvsc_copy_to_send_buf(struct netvsc_device *net_device,
938 unsigned int section_index,
940 struct hv_netvsc_packet *packet,
941 struct rndis_message *rndis_msg,
942 struct hv_page_buffer *pb,
945 char *start = net_device->send_buf;
946 char *dest = start + (section_index * net_device->send_section_size)
950 u32 page_count = packet->cp_partial ? packet->rmsg_pgcnt :
951 packet->page_buf_cnt;
955 remain = packet->total_data_buflen & (net_device->pkt_align - 1);
956 if (xmit_more && remain) {
957 padding = net_device->pkt_align - remain;
958 rndis_msg->msg_len += padding;
959 packet->total_data_buflen += padding;
962 for (i = 0; i < page_count; i++) {
963 char *src = phys_to_virt(pb[i].pfn << HV_HYP_PAGE_SHIFT);
964 u32 offset = pb[i].offset;
967 memcpy(dest, (src + offset), len);
972 memset(dest, 0, padding);
975 void netvsc_dma_unmap(struct hv_device *hv_dev,
976 struct hv_netvsc_packet *packet)
980 if (!hv_is_isolation_supported())
983 if (!packet->dma_range)
986 for (i = 0; i < packet->page_buf_cnt; i++)
987 dma_unmap_single(&hv_dev->device, packet->dma_range[i].dma,
988 packet->dma_range[i].mapping_size,
991 kfree(packet->dma_range);
994 /* netvsc_dma_map - Map swiotlb bounce buffer with data page of
995 * packet sent by vmbus_sendpacket_pagebuffer() in the Isolation
998 * In isolation VM, netvsc send buffer has been marked visible to
999 * host and so the data copied to send buffer doesn't need to use
1000 * bounce buffer. The data pages handled by vmbus_sendpacket_pagebuffer()
1001 * may not be copied to send buffer and so these pages need to be
1002 * mapped with swiotlb bounce buffer. netvsc_dma_map() is to do
1003 * that. The pfns in the struct hv_page_buffer need to be converted
1004 * to bounce buffer's pfn. The loop here is necessary because the
1005 * entries in the page buffer array are not necessarily full
1006 * pages of data. Each entry in the array has a separate offset and
1007 * len that may be non-zero, even for entries in the middle of the
1008 * array. And the entries are not physically contiguous. So each
1009 * entry must be individually mapped rather than as a contiguous unit.
1010 * So not use dma_map_sg() here.
1012 static int netvsc_dma_map(struct hv_device *hv_dev,
1013 struct hv_netvsc_packet *packet,
1014 struct hv_page_buffer *pb)
1016 u32 page_count = packet->page_buf_cnt;
1020 if (!hv_is_isolation_supported())
1023 packet->dma_range = kcalloc(page_count,
1024 sizeof(*packet->dma_range),
1026 if (!packet->dma_range)
1029 for (i = 0; i < page_count; i++) {
1030 char *src = phys_to_virt((pb[i].pfn << HV_HYP_PAGE_SHIFT)
1032 u32 len = pb[i].len;
1034 dma = dma_map_single(&hv_dev->device, src, len,
1036 if (dma_mapping_error(&hv_dev->device, dma)) {
1037 kfree(packet->dma_range);
1041 /* pb[].offset and pb[].len are not changed during dma mapping
1042 * and so not reassign.
1044 packet->dma_range[i].dma = dma;
1045 packet->dma_range[i].mapping_size = len;
1046 pb[i].pfn = dma >> HV_HYP_PAGE_SHIFT;
1052 static inline int netvsc_send_pkt(
1053 struct hv_device *device,
1054 struct hv_netvsc_packet *packet,
1055 struct netvsc_device *net_device,
1056 struct hv_page_buffer *pb,
1057 struct sk_buff *skb)
1059 struct nvsp_message nvmsg;
1060 struct nvsp_1_message_send_rndis_packet *rpkt =
1061 &nvmsg.msg.v1_msg.send_rndis_pkt;
1062 struct netvsc_channel * const nvchan =
1063 &net_device->chan_table[packet->q_idx];
1064 struct vmbus_channel *out_channel = nvchan->channel;
1065 struct net_device *ndev = hv_get_drvdata(device);
1066 struct net_device_context *ndev_ctx = netdev_priv(ndev);
1067 struct netdev_queue *txq = netdev_get_tx_queue(ndev, packet->q_idx);
1070 u32 ring_avail = hv_get_avail_to_write_percent(&out_channel->outbound);
1072 memset(&nvmsg, 0, sizeof(struct nvsp_message));
1073 nvmsg.hdr.msg_type = NVSP_MSG1_TYPE_SEND_RNDIS_PKT;
1075 rpkt->channel_type = 0; /* 0 is RMC_DATA */
1077 rpkt->channel_type = 1; /* 1 is RMC_CONTROL */
1079 rpkt->send_buf_section_index = packet->send_buf_index;
1080 if (packet->send_buf_index == NETVSC_INVALID_INDEX)
1081 rpkt->send_buf_section_size = 0;
1083 rpkt->send_buf_section_size = packet->total_data_buflen;
1085 req_id = (ulong)skb;
1087 if (out_channel->rescind)
1090 trace_nvsp_send_pkt(ndev, out_channel, rpkt);
1092 packet->dma_range = NULL;
1093 if (packet->page_buf_cnt) {
1094 if (packet->cp_partial)
1095 pb += packet->rmsg_pgcnt;
1097 ret = netvsc_dma_map(ndev_ctx->device_ctx, packet, pb);
1103 ret = vmbus_sendpacket_pagebuffer(out_channel,
1104 pb, packet->page_buf_cnt,
1105 &nvmsg, sizeof(nvmsg),
1109 netvsc_dma_unmap(ndev_ctx->device_ctx, packet);
1111 ret = vmbus_sendpacket(out_channel,
1112 &nvmsg, sizeof(nvmsg),
1113 req_id, VM_PKT_DATA_INBAND,
1114 VMBUS_DATA_PACKET_FLAG_COMPLETION_REQUESTED);
1119 atomic_inc_return(&nvchan->queue_sends);
1121 if (ring_avail < RING_AVAIL_PERCENT_LOWATER) {
1122 netif_tx_stop_queue(txq);
1123 ndev_ctx->eth_stats.stop_queue++;
1125 } else if (ret == -EAGAIN) {
1126 netif_tx_stop_queue(txq);
1127 ndev_ctx->eth_stats.stop_queue++;
1130 "Unable to send packet pages %u len %u, ret %d\n",
1131 packet->page_buf_cnt, packet->total_data_buflen,
1135 if (netif_tx_queue_stopped(txq) &&
1136 atomic_read(&nvchan->queue_sends) < 1 &&
1137 !net_device->tx_disable) {
1138 netif_tx_wake_queue(txq);
1139 ndev_ctx->eth_stats.wake_queue++;
1147 /* Move packet out of multi send data (msd), and clear msd */
1148 static inline void move_pkt_msd(struct hv_netvsc_packet **msd_send,
1149 struct sk_buff **msd_skb,
1150 struct multi_send_data *msdp)
1152 *msd_skb = msdp->skb;
1153 *msd_send = msdp->pkt;
1159 /* RCU already held by caller */
1160 /* Batching/bouncing logic is designed to attempt to optimize
1163 * For small, non-LSO packets we copy the packet to a send buffer
1164 * which is pre-registered with the Hyper-V side. This enables the
1165 * hypervisor to avoid remapping the aperture to access the packet
1166 * descriptor and data.
1168 * If we already started using a buffer and the netdev is transmitting
1169 * a burst of packets, keep on copying into the buffer until it is
1170 * full or we are done collecting a burst. If there is an existing
1171 * buffer with space for the RNDIS descriptor but not the packet, copy
1172 * the RNDIS descriptor to the buffer, keeping the packet in place.
1174 * If we do batching and send more than one packet using a single
1175 * NetVSC message, free the SKBs of the packets copied, except for the
1176 * last packet. This is done to streamline the handling of the case
1177 * where the last packet only had the RNDIS descriptor copied to the
1178 * send buffer, with the data pointers included in the NetVSC message.
1180 int netvsc_send(struct net_device *ndev,
1181 struct hv_netvsc_packet *packet,
1182 struct rndis_message *rndis_msg,
1183 struct hv_page_buffer *pb,
1184 struct sk_buff *skb,
1187 struct net_device_context *ndev_ctx = netdev_priv(ndev);
1188 struct netvsc_device *net_device
1189 = rcu_dereference_bh(ndev_ctx->nvdev);
1190 struct hv_device *device = ndev_ctx->device_ctx;
1192 struct netvsc_channel *nvchan;
1193 u32 pktlen = packet->total_data_buflen, msd_len = 0;
1194 unsigned int section_index = NETVSC_INVALID_INDEX;
1195 struct multi_send_data *msdp;
1196 struct hv_netvsc_packet *msd_send = NULL, *cur_send = NULL;
1197 struct sk_buff *msd_skb = NULL;
1198 bool try_batch, xmit_more;
1200 /* If device is rescinded, return error and packet will get dropped. */
1201 if (unlikely(!net_device || net_device->destroy))
1204 nvchan = &net_device->chan_table[packet->q_idx];
1205 packet->send_buf_index = NETVSC_INVALID_INDEX;
1206 packet->cp_partial = false;
1208 /* Send a control message or XDP packet directly without accessing
1209 * msd (Multi-Send Data) field which may be changed during data packet
1213 return netvsc_send_pkt(device, packet, net_device, pb, skb);
1215 /* batch packets in send buffer if possible */
1216 msdp = &nvchan->msd;
1218 msd_len = msdp->pkt->total_data_buflen;
1220 try_batch = msd_len > 0 && msdp->count < net_device->max_pkt;
1221 if (try_batch && msd_len + pktlen + net_device->pkt_align <
1222 net_device->send_section_size) {
1223 section_index = msdp->pkt->send_buf_index;
1225 } else if (try_batch && msd_len + packet->rmsg_size <
1226 net_device->send_section_size) {
1227 section_index = msdp->pkt->send_buf_index;
1228 packet->cp_partial = true;
1230 } else if (pktlen + net_device->pkt_align <
1231 net_device->send_section_size) {
1232 section_index = netvsc_get_next_send_section(net_device);
1233 if (unlikely(section_index == NETVSC_INVALID_INDEX)) {
1234 ++ndev_ctx->eth_stats.tx_send_full;
1236 move_pkt_msd(&msd_send, &msd_skb, msdp);
1241 /* Keep aggregating only if stack says more data is coming
1242 * and not doing mixed modes send and not flow blocked
1244 xmit_more = netdev_xmit_more() &&
1245 !packet->cp_partial &&
1246 !netif_xmit_stopped(netdev_get_tx_queue(ndev, packet->q_idx));
1248 if (section_index != NETVSC_INVALID_INDEX) {
1249 netvsc_copy_to_send_buf(net_device,
1250 section_index, msd_len,
1251 packet, rndis_msg, pb, xmit_more);
1253 packet->send_buf_index = section_index;
1255 if (packet->cp_partial) {
1256 packet->page_buf_cnt -= packet->rmsg_pgcnt;
1257 packet->total_data_buflen = msd_len + packet->rmsg_size;
1259 packet->page_buf_cnt = 0;
1260 packet->total_data_buflen += msd_len;
1264 packet->total_packets += msdp->pkt->total_packets;
1265 packet->total_bytes += msdp->pkt->total_bytes;
1269 dev_consume_skb_any(msdp->skb);
1282 move_pkt_msd(&msd_send, &msd_skb, msdp);
1287 int m_ret = netvsc_send_pkt(device, msd_send, net_device,
1291 netvsc_free_send_slot(net_device,
1292 msd_send->send_buf_index);
1293 dev_kfree_skb_any(msd_skb);
1298 ret = netvsc_send_pkt(device, cur_send, net_device, pb, skb);
1300 if (ret != 0 && section_index != NETVSC_INVALID_INDEX)
1301 netvsc_free_send_slot(net_device, section_index);
1306 /* Send pending recv completions */
1307 static int send_recv_completions(struct net_device *ndev,
1308 struct netvsc_device *nvdev,
1309 struct netvsc_channel *nvchan)
1311 struct multi_recv_comp *mrc = &nvchan->mrc;
1312 struct recv_comp_msg {
1313 struct nvsp_message_header hdr;
1316 struct recv_comp_msg msg = {
1317 .hdr.msg_type = NVSP_MSG1_TYPE_SEND_RNDIS_PKT_COMPLETE,
1321 while (mrc->first != mrc->next) {
1322 const struct recv_comp_data *rcd
1323 = mrc->slots + mrc->first;
1325 msg.status = rcd->status;
1326 ret = vmbus_sendpacket(nvchan->channel, &msg, sizeof(msg),
1327 rcd->tid, VM_PKT_COMP, 0);
1328 if (unlikely(ret)) {
1329 struct net_device_context *ndev_ctx = netdev_priv(ndev);
1331 ++ndev_ctx->eth_stats.rx_comp_busy;
1335 if (++mrc->first == nvdev->recv_completion_cnt)
1339 /* receive completion ring has been emptied */
1340 if (unlikely(nvdev->destroy))
1341 wake_up(&nvdev->wait_drain);
1346 /* Count how many receive completions are outstanding */
1347 static void recv_comp_slot_avail(const struct netvsc_device *nvdev,
1348 const struct multi_recv_comp *mrc,
1349 u32 *filled, u32 *avail)
1351 u32 count = nvdev->recv_completion_cnt;
1353 if (mrc->next >= mrc->first)
1354 *filled = mrc->next - mrc->first;
1356 *filled = (count - mrc->first) + mrc->next;
1358 *avail = count - *filled - 1;
1361 /* Add receive complete to ring to send to host. */
1362 static void enq_receive_complete(struct net_device *ndev,
1363 struct netvsc_device *nvdev, u16 q_idx,
1364 u64 tid, u32 status)
1366 struct netvsc_channel *nvchan = &nvdev->chan_table[q_idx];
1367 struct multi_recv_comp *mrc = &nvchan->mrc;
1368 struct recv_comp_data *rcd;
1371 recv_comp_slot_avail(nvdev, mrc, &filled, &avail);
1373 if (unlikely(filled > NAPI_POLL_WEIGHT)) {
1374 send_recv_completions(ndev, nvdev, nvchan);
1375 recv_comp_slot_avail(nvdev, mrc, &filled, &avail);
1378 if (unlikely(!avail)) {
1379 netdev_err(ndev, "Recv_comp full buf q:%hd, tid:%llx\n",
1384 rcd = mrc->slots + mrc->next;
1386 rcd->status = status;
1388 if (++mrc->next == nvdev->recv_completion_cnt)
1392 static int netvsc_receive(struct net_device *ndev,
1393 struct netvsc_device *net_device,
1394 struct netvsc_channel *nvchan,
1395 const struct vmpacket_descriptor *desc)
1397 struct net_device_context *net_device_ctx = netdev_priv(ndev);
1398 struct vmbus_channel *channel = nvchan->channel;
1399 const struct vmtransfer_page_packet_header *vmxferpage_packet
1400 = container_of(desc, const struct vmtransfer_page_packet_header, d);
1401 const struct nvsp_message *nvsp = hv_pkt_data(desc);
1402 u32 msglen = hv_pkt_datalen(desc);
1403 u16 q_idx = channel->offermsg.offer.sub_channel_index;
1404 char *recv_buf = net_device->recv_buf;
1405 u32 status = NVSP_STAT_SUCCESS;
1409 /* Ensure packet is big enough to read header fields */
1410 if (msglen < sizeof(struct nvsp_message_header)) {
1411 netif_err(net_device_ctx, rx_err, ndev,
1412 "invalid nvsp header, length too small: %u\n",
1417 /* Make sure this is a valid nvsp packet */
1418 if (unlikely(nvsp->hdr.msg_type != NVSP_MSG1_TYPE_SEND_RNDIS_PKT)) {
1419 netif_err(net_device_ctx, rx_err, ndev,
1420 "Unknown nvsp packet type received %u\n",
1421 nvsp->hdr.msg_type);
1425 /* Validate xfer page pkt header */
1426 if ((desc->offset8 << 3) < sizeof(struct vmtransfer_page_packet_header)) {
1427 netif_err(net_device_ctx, rx_err, ndev,
1428 "Invalid xfer page pkt, offset too small: %u\n",
1429 desc->offset8 << 3);
1433 if (unlikely(vmxferpage_packet->xfer_pageset_id != NETVSC_RECEIVE_BUFFER_ID)) {
1434 netif_err(net_device_ctx, rx_err, ndev,
1435 "Invalid xfer page set id - expecting %x got %x\n",
1436 NETVSC_RECEIVE_BUFFER_ID,
1437 vmxferpage_packet->xfer_pageset_id);
1441 count = vmxferpage_packet->range_cnt;
1443 /* Check count for a valid value */
1444 if (NETVSC_XFER_HEADER_SIZE(count) > desc->offset8 << 3) {
1445 netif_err(net_device_ctx, rx_err, ndev,
1446 "Range count is not valid: %d\n",
1451 /* Each range represents 1 RNDIS pkt that contains 1 ethernet frame */
1452 for (i = 0; i < count; i++) {
1453 u32 offset = vmxferpage_packet->ranges[i].byte_offset;
1454 u32 buflen = vmxferpage_packet->ranges[i].byte_count;
1458 if (unlikely(offset > net_device->recv_buf_size ||
1459 buflen > net_device->recv_buf_size - offset)) {
1460 nvchan->rsc.cnt = 0;
1461 status = NVSP_STAT_FAIL;
1462 netif_err(net_device_ctx, rx_err, ndev,
1463 "Packet offset:%u + len:%u too big\n",
1469 /* We're going to copy (sections of) the packet into nvchan->recv_buf;
1470 * make sure that nvchan->recv_buf is large enough to hold the packet.
1472 if (unlikely(buflen > net_device->recv_section_size)) {
1473 nvchan->rsc.cnt = 0;
1474 status = NVSP_STAT_FAIL;
1475 netif_err(net_device_ctx, rx_err, ndev,
1476 "Packet too big: buflen=%u recv_section_size=%u\n",
1477 buflen, net_device->recv_section_size);
1482 data = recv_buf + offset;
1484 nvchan->rsc.is_last = (i == count - 1);
1486 trace_rndis_recv(ndev, q_idx, data);
1488 /* Pass it to the upper layer */
1489 ret = rndis_filter_receive(ndev, net_device,
1490 nvchan, data, buflen);
1492 if (unlikely(ret != NVSP_STAT_SUCCESS)) {
1493 /* Drop incomplete packet */
1494 nvchan->rsc.cnt = 0;
1495 status = NVSP_STAT_FAIL;
1499 enq_receive_complete(ndev, net_device, q_idx,
1500 vmxferpage_packet->d.trans_id, status);
1505 static void netvsc_send_table(struct net_device *ndev,
1506 struct netvsc_device *nvscdev,
1507 const struct nvsp_message *nvmsg,
1510 struct net_device_context *net_device_ctx = netdev_priv(ndev);
1511 u32 count, offset, *tab;
1514 /* Ensure packet is big enough to read send_table fields */
1515 if (msglen < sizeof(struct nvsp_message_header) +
1516 sizeof(struct nvsp_5_send_indirect_table)) {
1517 netdev_err(ndev, "nvsp_v5_msg length too small: %u\n", msglen);
1521 count = nvmsg->msg.v5_msg.send_table.count;
1522 offset = nvmsg->msg.v5_msg.send_table.offset;
1524 if (count != VRSS_SEND_TAB_SIZE) {
1525 netdev_err(ndev, "Received wrong send-table size:%u\n", count);
1529 /* If negotiated version <= NVSP_PROTOCOL_VERSION_6, the offset may be
1530 * wrong due to a host bug. So fix the offset here.
1532 if (nvscdev->nvsp_version <= NVSP_PROTOCOL_VERSION_6 &&
1533 msglen >= sizeof(struct nvsp_message_header) +
1534 sizeof(union nvsp_6_message_uber) + count * sizeof(u32))
1535 offset = sizeof(struct nvsp_message_header) +
1536 sizeof(union nvsp_6_message_uber);
1538 /* Boundary check for all versions */
1539 if (msglen < count * sizeof(u32) || offset > msglen - count * sizeof(u32)) {
1540 netdev_err(ndev, "Received send-table offset too big:%u\n",
1545 tab = (void *)nvmsg + offset;
1547 for (i = 0; i < count; i++)
1548 net_device_ctx->tx_table[i] = tab[i];
1551 static void netvsc_send_vf(struct net_device *ndev,
1552 const struct nvsp_message *nvmsg,
1555 struct net_device_context *net_device_ctx = netdev_priv(ndev);
1557 /* Ensure packet is big enough to read its fields */
1558 if (msglen < sizeof(struct nvsp_message_header) +
1559 sizeof(struct nvsp_4_send_vf_association)) {
1560 netdev_err(ndev, "nvsp_v4_msg length too small: %u\n", msglen);
1564 net_device_ctx->vf_alloc = nvmsg->msg.v4_msg.vf_assoc.allocated;
1565 net_device_ctx->vf_serial = nvmsg->msg.v4_msg.vf_assoc.serial;
1567 if (net_device_ctx->vf_alloc)
1568 complete(&net_device_ctx->vf_add);
1570 netdev_info(ndev, "VF slot %u %s\n",
1571 net_device_ctx->vf_serial,
1572 net_device_ctx->vf_alloc ? "added" : "removed");
1575 static void netvsc_receive_inband(struct net_device *ndev,
1576 struct netvsc_device *nvscdev,
1577 const struct vmpacket_descriptor *desc)
1579 const struct nvsp_message *nvmsg = hv_pkt_data(desc);
1580 u32 msglen = hv_pkt_datalen(desc);
1582 /* Ensure packet is big enough to read header fields */
1583 if (msglen < sizeof(struct nvsp_message_header)) {
1584 netdev_err(ndev, "inband nvsp_message length too small: %u\n", msglen);
1588 switch (nvmsg->hdr.msg_type) {
1589 case NVSP_MSG5_TYPE_SEND_INDIRECTION_TABLE:
1590 netvsc_send_table(ndev, nvscdev, nvmsg, msglen);
1593 case NVSP_MSG4_TYPE_SEND_VF_ASSOCIATION:
1594 if (hv_is_isolation_supported())
1595 netdev_err(ndev, "Ignore VF_ASSOCIATION msg from the host supporting isolation\n");
1597 netvsc_send_vf(ndev, nvmsg, msglen);
1602 static int netvsc_process_raw_pkt(struct hv_device *device,
1603 struct netvsc_channel *nvchan,
1604 struct netvsc_device *net_device,
1605 struct net_device *ndev,
1606 const struct vmpacket_descriptor *desc,
1609 struct vmbus_channel *channel = nvchan->channel;
1610 const struct nvsp_message *nvmsg = hv_pkt_data(desc);
1612 trace_nvsp_recv(ndev, channel, nvmsg);
1614 switch (desc->type) {
1616 netvsc_send_completion(ndev, net_device, channel, desc, budget);
1619 case VM_PKT_DATA_USING_XFER_PAGES:
1620 return netvsc_receive(ndev, net_device, nvchan, desc);
1622 case VM_PKT_DATA_INBAND:
1623 netvsc_receive_inband(ndev, net_device, desc);
1627 netdev_err(ndev, "unhandled packet type %d, tid %llx\n",
1628 desc->type, desc->trans_id);
1635 static struct hv_device *netvsc_channel_to_device(struct vmbus_channel *channel)
1637 struct vmbus_channel *primary = channel->primary_channel;
1639 return primary ? primary->device_obj : channel->device_obj;
1642 /* Network processing softirq
1643 * Process data in incoming ring buffer from host
1644 * Stops when ring is empty or budget is met or exceeded.
1646 int netvsc_poll(struct napi_struct *napi, int budget)
1648 struct netvsc_channel *nvchan
1649 = container_of(napi, struct netvsc_channel, napi);
1650 struct netvsc_device *net_device = nvchan->net_device;
1651 struct vmbus_channel *channel = nvchan->channel;
1652 struct hv_device *device = netvsc_channel_to_device(channel);
1653 struct net_device *ndev = hv_get_drvdata(device);
1657 /* If starting a new interval */
1659 nvchan->desc = hv_pkt_iter_first(channel);
1661 nvchan->xdp_flush = false;
1663 while (nvchan->desc && work_done < budget) {
1664 work_done += netvsc_process_raw_pkt(device, nvchan, net_device,
1665 ndev, nvchan->desc, budget);
1666 nvchan->desc = hv_pkt_iter_next(channel, nvchan->desc);
1669 if (nvchan->xdp_flush)
1672 /* Send any pending receive completions */
1673 ret = send_recv_completions(ndev, net_device, nvchan);
1675 /* If it did not exhaust NAPI budget this time
1676 * and not doing busy poll
1677 * then re-enable host interrupts
1678 * and reschedule if ring is not empty
1679 * or sending receive completion failed.
1681 if (work_done < budget &&
1682 napi_complete_done(napi, work_done) &&
1683 (ret || hv_end_read(&channel->inbound)) &&
1684 napi_schedule_prep(napi)) {
1685 hv_begin_read(&channel->inbound);
1686 __napi_schedule(napi);
1689 /* Driver may overshoot since multiple packets per descriptor */
1690 return min(work_done, budget);
1693 /* Call back when data is available in host ring buffer.
1694 * Processing is deferred until network softirq (NAPI)
1696 void netvsc_channel_cb(void *context)
1698 struct netvsc_channel *nvchan = context;
1699 struct vmbus_channel *channel = nvchan->channel;
1700 struct hv_ring_buffer_info *rbi = &channel->inbound;
1702 /* preload first vmpacket descriptor */
1703 prefetch(hv_get_ring_buffer(rbi) + rbi->priv_read_index);
1705 if (napi_schedule_prep(&nvchan->napi)) {
1706 /* disable interrupts from host */
1709 __napi_schedule_irqoff(&nvchan->napi);
1714 * netvsc_device_add - Callback when the device belonging to this
1717 struct netvsc_device *netvsc_device_add(struct hv_device *device,
1718 const struct netvsc_device_info *device_info)
1721 struct netvsc_device *net_device;
1722 struct net_device *ndev = hv_get_drvdata(device);
1723 struct net_device_context *net_device_ctx = netdev_priv(ndev);
1725 net_device = alloc_net_device();
1727 return ERR_PTR(-ENOMEM);
1729 for (i = 0; i < VRSS_SEND_TAB_SIZE; i++)
1730 net_device_ctx->tx_table[i] = 0;
1732 /* Because the device uses NAPI, all the interrupt batching and
1733 * control is done via Net softirq, not the channel handling
1735 set_channel_read_mode(device->channel, HV_CALL_ISR);
1737 /* If we're reopening the device we may have multiple queues, fill the
1738 * chn_table with the default channel to use it before subchannels are
1740 * Initialize the channel state before we open;
1741 * we can be interrupted as soon as we open the channel.
1744 for (i = 0; i < VRSS_CHANNEL_MAX; i++) {
1745 struct netvsc_channel *nvchan = &net_device->chan_table[i];
1747 nvchan->channel = device->channel;
1748 nvchan->net_device = net_device;
1749 u64_stats_init(&nvchan->tx_stats.syncp);
1750 u64_stats_init(&nvchan->rx_stats.syncp);
1752 ret = xdp_rxq_info_reg(&nvchan->xdp_rxq, ndev, i, 0);
1755 netdev_err(ndev, "xdp_rxq_info_reg fail: %d\n", ret);
1759 ret = xdp_rxq_info_reg_mem_model(&nvchan->xdp_rxq,
1760 MEM_TYPE_PAGE_SHARED, NULL);
1763 netdev_err(ndev, "xdp reg_mem_model fail: %d\n", ret);
1768 /* Enable NAPI handler before init callbacks */
1769 netif_napi_add(ndev, &net_device->chan_table[0].napi, netvsc_poll);
1771 /* Open the channel */
1772 device->channel->next_request_id_callback = vmbus_next_request_id;
1773 device->channel->request_addr_callback = vmbus_request_addr;
1774 device->channel->rqstor_size = netvsc_rqstor_size(netvsc_ring_bytes);
1775 device->channel->max_pkt_size = NETVSC_MAX_PKT_SIZE;
1777 ret = vmbus_open(device->channel, netvsc_ring_bytes,
1778 netvsc_ring_bytes, NULL, 0,
1779 netvsc_channel_cb, net_device->chan_table);
1782 netdev_err(ndev, "unable to open channel: %d\n", ret);
1786 /* Channel is opened */
1787 netdev_dbg(ndev, "hv_netvsc channel opened successfully\n");
1789 napi_enable(&net_device->chan_table[0].napi);
1791 /* Connect with the NetVsp */
1792 ret = netvsc_connect_vsp(device, net_device, device_info);
1795 "unable to connect to NetVSP - %d\n", ret);
1799 /* Writing nvdev pointer unlocks netvsc_send(), make sure chn_table is
1802 rcu_assign_pointer(net_device_ctx->nvdev, net_device);
1807 RCU_INIT_POINTER(net_device_ctx->nvdev, NULL);
1808 napi_disable(&net_device->chan_table[0].napi);
1810 /* Now, we can close the channel safely */
1811 vmbus_close(device->channel);
1814 netif_napi_del(&net_device->chan_table[0].napi);
1817 free_netvsc_device(&net_device->rcu);
1819 return ERR_PTR(ret);