1 // SPDX-License-Identifier: GPL-2.0
3 * Networking over Thunderbolt/USB4 cables using USB4NET protocol
4 * (formerly Apple ThunderboltIP).
6 * Copyright (C) 2017, Intel Corporation
12 #include <linux/atomic.h>
13 #include <linux/highmem.h>
14 #include <linux/if_vlan.h>
15 #include <linux/jhash.h>
16 #include <linux/module.h>
17 #include <linux/etherdevice.h>
18 #include <linux/rtnetlink.h>
19 #include <linux/sizes.h>
20 #include <linux/thunderbolt.h>
21 #include <linux/uuid.h>
22 #include <linux/workqueue.h>
24 #include <net/ip6_checksum.h>
28 /* Protocol timeouts in ms */
29 #define TBNET_LOGIN_DELAY 4500
30 #define TBNET_LOGIN_TIMEOUT 500
31 #define TBNET_LOGOUT_TIMEOUT 1000
33 #define TBNET_RING_SIZE 256
34 #define TBNET_LOGIN_RETRIES 60
35 #define TBNET_LOGOUT_RETRIES 10
36 #define TBNET_E2E BIT(0)
37 #define TBNET_MATCH_FRAGS_ID BIT(1)
38 #define TBNET_64K_FRAMES BIT(2)
39 #define TBNET_MAX_MTU SZ_64K
40 #define TBNET_FRAME_SIZE SZ_4K
41 #define TBNET_MAX_PAYLOAD_SIZE \
42 (TBNET_FRAME_SIZE - sizeof(struct thunderbolt_ip_frame_header))
43 /* Rx packets need to hold space for skb_shared_info */
44 #define TBNET_RX_MAX_SIZE \
45 (TBNET_FRAME_SIZE + SKB_DATA_ALIGN(sizeof(struct skb_shared_info)))
46 #define TBNET_RX_PAGE_ORDER get_order(TBNET_RX_MAX_SIZE)
47 #define TBNET_RX_PAGE_SIZE (PAGE_SIZE << TBNET_RX_PAGE_ORDER)
49 #define TBNET_L0_PORT_NUM(route) ((route) & GENMASK(5, 0))
52 * struct thunderbolt_ip_frame_header - Header for each Thunderbolt frame
53 * @frame_size: size of the data with the frame
54 * @frame_index: running index on the frames
55 * @frame_id: ID of the frame to match frames to specific packet
56 * @frame_count: how many frames assembles a full packet
58 * Each data frame passed to the high-speed DMA ring has this header. If
59 * the XDomain network directory announces that %TBNET_MATCH_FRAGS_ID is
60 * supported then @frame_id is filled, otherwise it stays %0.
62 struct thunderbolt_ip_frame_header {
69 enum thunderbolt_ip_frame_pdf {
70 TBIP_PDF_FRAME_START = 1,
74 enum thunderbolt_ip_type {
81 struct thunderbolt_ip_header {
86 uuid_t initiator_uuid;
92 #define TBIP_HDR_LENGTH_MASK GENMASK(5, 0)
93 #define TBIP_HDR_SN_MASK GENMASK(28, 27)
94 #define TBIP_HDR_SN_SHIFT 27
96 struct thunderbolt_ip_login {
97 struct thunderbolt_ip_header hdr;
103 #define TBIP_LOGIN_PROTO_VERSION 1
105 struct thunderbolt_ip_login_response {
106 struct thunderbolt_ip_header hdr;
109 u32 receiver_mac_len;
113 struct thunderbolt_ip_logout {
114 struct thunderbolt_ip_header hdr;
117 struct thunderbolt_ip_status {
118 struct thunderbolt_ip_header hdr;
129 u64 rx_length_errors;
132 u64 rx_missed_errors;
136 struct net_device *dev;
138 struct ring_frame frame;
142 struct tbnet_frame frames[TBNET_RING_SIZE];
145 struct tb_ring *ring;
149 * struct tbnet - ThunderboltIP network driver private data
150 * @svc: XDomain service the driver is bound to
151 * @xd: XDomain the service belongs to
152 * @handler: ThunderboltIP configuration protocol handler
153 * @dev: Networking device
154 * @napi: NAPI structure for Rx polling
155 * @stats: Network statistics
156 * @skb: Network packet that is currently processed on Rx path
157 * @command_id: ID used for next configuration protocol packet
158 * @login_sent: ThunderboltIP login message successfully sent
159 * @login_received: ThunderboltIP login message received from the remote
161 * @local_transmit_path: HopID we are using to send out packets
162 * @remote_transmit_path: HopID the other end is using to send packets to us
163 * @connection_lock: Lock serializing access to @login_sent,
164 * @login_received and @transmit_path.
165 * @login_retries: Number of login retries currently done
166 * @login_work: Worker to send ThunderboltIP login packets
167 * @connected_work: Worker that finalizes the ThunderboltIP connection
168 * setup and enables DMA paths for high speed data
170 * @disconnect_work: Worker that handles tearing down the ThunderboltIP
172 * @rx_hdr: Copy of the currently processed Rx frame. Used when a
173 * network packet consists of multiple Thunderbolt frames.
174 * In host byte order.
175 * @rx_ring: Software ring holding Rx frames
176 * @frame_id: Frame ID use for next Tx packet
177 * (if %TBNET_MATCH_FRAGS_ID is supported in both ends)
178 * @tx_ring: Software ring holding Tx frames
181 const struct tb_service *svc;
182 struct tb_xdomain *xd;
183 struct tb_protocol_handler handler;
184 struct net_device *dev;
185 struct napi_struct napi;
186 struct tbnet_stats stats;
191 int local_transmit_path;
192 int remote_transmit_path;
193 struct mutex connection_lock;
195 struct delayed_work login_work;
196 struct work_struct connected_work;
197 struct work_struct disconnect_work;
198 struct thunderbolt_ip_frame_header rx_hdr;
199 struct tbnet_ring rx_ring;
201 struct tbnet_ring tx_ring;
204 /* Network property directory UUID: c66189ca-1cce-4195-bdb8-49592e5f5a4f */
205 static const uuid_t tbnet_dir_uuid =
206 UUID_INIT(0xc66189ca, 0x1cce, 0x4195,
207 0xbd, 0xb8, 0x49, 0x59, 0x2e, 0x5f, 0x5a, 0x4f);
209 /* ThunderboltIP protocol UUID: 798f589e-3616-8a47-97c6-5664a920c8dd */
210 static const uuid_t tbnet_svc_uuid =
211 UUID_INIT(0x798f589e, 0x3616, 0x8a47,
212 0x97, 0xc6, 0x56, 0x64, 0xa9, 0x20, 0xc8, 0xdd);
214 static struct tb_property_dir *tbnet_dir;
216 static bool tbnet_e2e = true;
217 module_param_named(e2e, tbnet_e2e, bool, 0444);
218 MODULE_PARM_DESC(e2e, "USB4NET full end-to-end flow control (default: true)");
220 static void tbnet_fill_header(struct thunderbolt_ip_header *hdr, u64 route,
221 u8 sequence, const uuid_t *initiator_uuid, const uuid_t *target_uuid,
222 enum thunderbolt_ip_type type, size_t size, u32 command_id)
226 /* Length does not include route_hi/lo and length_sn fields */
227 length_sn = (size - 3 * 4) / 4;
228 length_sn |= (sequence << TBIP_HDR_SN_SHIFT) & TBIP_HDR_SN_MASK;
230 hdr->route_hi = upper_32_bits(route);
231 hdr->route_lo = lower_32_bits(route);
232 hdr->length_sn = length_sn;
233 uuid_copy(&hdr->uuid, &tbnet_svc_uuid);
234 uuid_copy(&hdr->initiator_uuid, initiator_uuid);
235 uuid_copy(&hdr->target_uuid, target_uuid);
237 hdr->command_id = command_id;
240 static int tbnet_login_response(struct tbnet *net, u64 route, u8 sequence,
243 struct thunderbolt_ip_login_response reply;
244 struct tb_xdomain *xd = net->xd;
246 memset(&reply, 0, sizeof(reply));
247 tbnet_fill_header(&reply.hdr, route, sequence, xd->local_uuid,
248 xd->remote_uuid, TBIP_LOGIN_RESPONSE, sizeof(reply),
250 memcpy(reply.receiver_mac, net->dev->dev_addr, ETH_ALEN);
251 reply.receiver_mac_len = ETH_ALEN;
253 return tb_xdomain_response(xd, &reply, sizeof(reply),
254 TB_CFG_PKG_XDOMAIN_RESP);
257 static int tbnet_login_request(struct tbnet *net, u8 sequence)
259 struct thunderbolt_ip_login_response reply;
260 struct thunderbolt_ip_login request;
261 struct tb_xdomain *xd = net->xd;
263 memset(&request, 0, sizeof(request));
264 tbnet_fill_header(&request.hdr, xd->route, sequence, xd->local_uuid,
265 xd->remote_uuid, TBIP_LOGIN, sizeof(request),
266 atomic_inc_return(&net->command_id));
268 request.proto_version = TBIP_LOGIN_PROTO_VERSION;
269 request.transmit_path = net->local_transmit_path;
271 return tb_xdomain_request(xd, &request, sizeof(request),
272 TB_CFG_PKG_XDOMAIN_RESP, &reply,
273 sizeof(reply), TB_CFG_PKG_XDOMAIN_RESP,
274 TBNET_LOGIN_TIMEOUT);
277 static int tbnet_logout_response(struct tbnet *net, u64 route, u8 sequence,
280 struct thunderbolt_ip_status reply;
281 struct tb_xdomain *xd = net->xd;
283 memset(&reply, 0, sizeof(reply));
284 tbnet_fill_header(&reply.hdr, route, sequence, xd->local_uuid,
285 xd->remote_uuid, TBIP_STATUS, sizeof(reply),
286 atomic_inc_return(&net->command_id));
287 return tb_xdomain_response(xd, &reply, sizeof(reply),
288 TB_CFG_PKG_XDOMAIN_RESP);
291 static int tbnet_logout_request(struct tbnet *net)
293 struct thunderbolt_ip_logout request;
294 struct thunderbolt_ip_status reply;
295 struct tb_xdomain *xd = net->xd;
297 memset(&request, 0, sizeof(request));
298 tbnet_fill_header(&request.hdr, xd->route, 0, xd->local_uuid,
299 xd->remote_uuid, TBIP_LOGOUT, sizeof(request),
300 atomic_inc_return(&net->command_id));
302 return tb_xdomain_request(xd, &request, sizeof(request),
303 TB_CFG_PKG_XDOMAIN_RESP, &reply,
304 sizeof(reply), TB_CFG_PKG_XDOMAIN_RESP,
305 TBNET_LOGOUT_TIMEOUT);
308 static void start_login(struct tbnet *net)
310 netdev_dbg(net->dev, "login started\n");
312 mutex_lock(&net->connection_lock);
313 net->login_sent = false;
314 net->login_received = false;
315 mutex_unlock(&net->connection_lock);
317 queue_delayed_work(system_long_wq, &net->login_work,
318 msecs_to_jiffies(1000));
321 static void stop_login(struct tbnet *net)
323 cancel_delayed_work_sync(&net->login_work);
324 cancel_work_sync(&net->connected_work);
326 netdev_dbg(net->dev, "login stopped\n");
329 static inline unsigned int tbnet_frame_size(const struct tbnet_frame *tf)
331 return tf->frame.size ? : TBNET_FRAME_SIZE;
334 static void tbnet_free_buffers(struct tbnet_ring *ring)
338 for (i = 0; i < TBNET_RING_SIZE; i++) {
339 struct device *dma_dev = tb_ring_dma_device(ring->ring);
340 struct tbnet_frame *tf = &ring->frames[i];
341 enum dma_data_direction dir;
348 if (ring->ring->is_tx) {
351 size = TBNET_FRAME_SIZE;
353 dir = DMA_FROM_DEVICE;
354 order = TBNET_RX_PAGE_ORDER;
355 size = TBNET_RX_PAGE_SIZE;
358 trace_tbnet_free_frame(i, tf->page, tf->frame.buffer_phy, dir);
360 if (tf->frame.buffer_phy)
361 dma_unmap_page(dma_dev, tf->frame.buffer_phy, size,
364 __free_pages(tf->page, order);
372 static void tbnet_tear_down(struct tbnet *net, bool send_logout)
374 netif_carrier_off(net->dev);
375 netif_stop_queue(net->dev);
379 mutex_lock(&net->connection_lock);
381 if (net->login_sent && net->login_received) {
382 int ret, retries = TBNET_LOGOUT_RETRIES;
384 while (send_logout && retries-- > 0) {
385 netdev_dbg(net->dev, "sending logout request %u\n",
387 ret = tbnet_logout_request(net);
388 if (ret != -ETIMEDOUT)
392 tb_ring_stop(net->rx_ring.ring);
393 tb_ring_stop(net->tx_ring.ring);
394 tbnet_free_buffers(&net->rx_ring);
395 tbnet_free_buffers(&net->tx_ring);
397 ret = tb_xdomain_disable_paths(net->xd,
398 net->local_transmit_path,
399 net->rx_ring.ring->hop,
400 net->remote_transmit_path,
401 net->tx_ring.ring->hop);
403 netdev_warn(net->dev, "failed to disable DMA paths\n");
405 tb_xdomain_release_in_hopid(net->xd, net->remote_transmit_path);
406 net->remote_transmit_path = 0;
409 net->login_retries = 0;
410 net->login_sent = false;
411 net->login_received = false;
413 netdev_dbg(net->dev, "network traffic stopped\n");
415 mutex_unlock(&net->connection_lock);
418 static int tbnet_handle_packet(const void *buf, size_t size, void *data)
420 const struct thunderbolt_ip_login *pkg = buf;
421 struct tbnet *net = data;
427 /* Make sure the packet is for us */
428 if (size < sizeof(struct thunderbolt_ip_header))
430 if (!uuid_equal(&pkg->hdr.initiator_uuid, net->xd->remote_uuid))
432 if (!uuid_equal(&pkg->hdr.target_uuid, net->xd->local_uuid))
435 route = ((u64)pkg->hdr.route_hi << 32) | pkg->hdr.route_lo;
436 route &= ~BIT_ULL(63);
437 if (route != net->xd->route)
440 sequence = pkg->hdr.length_sn & TBIP_HDR_SN_MASK;
441 sequence >>= TBIP_HDR_SN_SHIFT;
442 command_id = pkg->hdr.command_id;
444 switch (pkg->hdr.type) {
446 netdev_dbg(net->dev, "remote login request received\n");
447 if (!netif_running(net->dev))
450 ret = tbnet_login_response(net, route, sequence,
451 pkg->hdr.command_id);
453 netdev_dbg(net->dev, "remote login response sent\n");
455 mutex_lock(&net->connection_lock);
456 net->login_received = true;
457 net->remote_transmit_path = pkg->transmit_path;
459 /* If we reached the number of max retries or
460 * previous logout, schedule another round of
463 if (net->login_retries >= TBNET_LOGIN_RETRIES ||
465 net->login_retries = 0;
466 queue_delayed_work(system_long_wq,
467 &net->login_work, 0);
469 mutex_unlock(&net->connection_lock);
471 queue_work(system_long_wq, &net->connected_work);
476 netdev_dbg(net->dev, "remote logout request received\n");
477 ret = tbnet_logout_response(net, route, sequence, command_id);
479 netdev_dbg(net->dev, "remote logout response sent\n");
480 queue_work(system_long_wq, &net->disconnect_work);
489 netdev_warn(net->dev, "failed to send ThunderboltIP response\n");
494 static unsigned int tbnet_available_buffers(const struct tbnet_ring *ring)
496 return ring->prod - ring->cons;
499 static int tbnet_alloc_rx_buffers(struct tbnet *net, unsigned int nbuffers)
501 struct tbnet_ring *ring = &net->rx_ring;
505 struct device *dma_dev = tb_ring_dma_device(ring->ring);
506 unsigned int index = ring->prod & (TBNET_RING_SIZE - 1);
507 struct tbnet_frame *tf = &ring->frames[index];
513 /* Allocate page (order > 0) so that it can hold maximum
514 * ThunderboltIP frame (4kB) and the additional room for
515 * SKB shared info required by build_skb().
517 tf->page = dev_alloc_pages(TBNET_RX_PAGE_ORDER);
523 dma_addr = dma_map_page(dma_dev, tf->page, 0,
524 TBNET_RX_PAGE_SIZE, DMA_FROM_DEVICE);
525 if (dma_mapping_error(dma_dev, dma_addr)) {
530 tf->frame.buffer_phy = dma_addr;
533 trace_tbnet_alloc_rx_frame(index, tf->page, dma_addr,
536 tb_ring_rx(ring->ring, &tf->frame);
544 tbnet_free_buffers(ring);
548 static struct tbnet_frame *tbnet_get_tx_buffer(struct tbnet *net)
550 struct tbnet_ring *ring = &net->tx_ring;
551 struct device *dma_dev = tb_ring_dma_device(ring->ring);
552 struct tbnet_frame *tf;
555 if (!tbnet_available_buffers(ring))
558 index = ring->cons++ & (TBNET_RING_SIZE - 1);
560 tf = &ring->frames[index];
563 dma_sync_single_for_cpu(dma_dev, tf->frame.buffer_phy,
564 tbnet_frame_size(tf), DMA_TO_DEVICE);
569 static void tbnet_tx_callback(struct tb_ring *ring, struct ring_frame *frame,
572 struct tbnet_frame *tf = container_of(frame, typeof(*tf), frame);
573 struct tbnet *net = netdev_priv(tf->dev);
575 /* Return buffer to the ring */
578 if (tbnet_available_buffers(&net->tx_ring) >= TBNET_RING_SIZE / 2)
579 netif_wake_queue(net->dev);
582 static int tbnet_alloc_tx_buffers(struct tbnet *net)
584 struct tbnet_ring *ring = &net->tx_ring;
585 struct device *dma_dev = tb_ring_dma_device(ring->ring);
588 for (i = 0; i < TBNET_RING_SIZE; i++) {
589 struct tbnet_frame *tf = &ring->frames[i];
592 tf->page = alloc_page(GFP_KERNEL);
594 tbnet_free_buffers(ring);
598 dma_addr = dma_map_page(dma_dev, tf->page, 0, TBNET_FRAME_SIZE,
600 if (dma_mapping_error(dma_dev, dma_addr)) {
601 __free_page(tf->page);
603 tbnet_free_buffers(ring);
608 tf->frame.buffer_phy = dma_addr;
609 tf->frame.callback = tbnet_tx_callback;
610 tf->frame.sof = TBIP_PDF_FRAME_START;
611 tf->frame.eof = TBIP_PDF_FRAME_END;
613 trace_tbnet_alloc_tx_frame(i, tf->page, dma_addr, DMA_TO_DEVICE);
617 ring->prod = TBNET_RING_SIZE - 1;
622 static void tbnet_connected_work(struct work_struct *work)
624 struct tbnet *net = container_of(work, typeof(*net), connected_work);
628 if (netif_carrier_ok(net->dev))
631 mutex_lock(&net->connection_lock);
632 connected = net->login_sent && net->login_received;
633 mutex_unlock(&net->connection_lock);
638 netdev_dbg(net->dev, "login successful, enabling paths\n");
640 ret = tb_xdomain_alloc_in_hopid(net->xd, net->remote_transmit_path);
641 if (ret != net->remote_transmit_path) {
642 netdev_err(net->dev, "failed to allocate Rx HopID\n");
646 /* Both logins successful so enable the rings, high-speed DMA
647 * paths and start the network device queue.
649 * Note we enable the DMA paths last to make sure we have primed
650 * the Rx ring before any incoming packets are allowed to
653 tb_ring_start(net->tx_ring.ring);
654 tb_ring_start(net->rx_ring.ring);
656 ret = tbnet_alloc_rx_buffers(net, TBNET_RING_SIZE);
660 ret = tbnet_alloc_tx_buffers(net);
662 goto err_free_rx_buffers;
664 ret = tb_xdomain_enable_paths(net->xd, net->local_transmit_path,
665 net->rx_ring.ring->hop,
666 net->remote_transmit_path,
667 net->tx_ring.ring->hop);
669 netdev_err(net->dev, "failed to enable DMA paths\n");
670 goto err_free_tx_buffers;
673 netif_carrier_on(net->dev);
674 netif_start_queue(net->dev);
676 netdev_dbg(net->dev, "network traffic started\n");
680 tbnet_free_buffers(&net->tx_ring);
682 tbnet_free_buffers(&net->rx_ring);
684 tb_ring_stop(net->rx_ring.ring);
685 tb_ring_stop(net->tx_ring.ring);
686 tb_xdomain_release_in_hopid(net->xd, net->remote_transmit_path);
689 static void tbnet_login_work(struct work_struct *work)
691 struct tbnet *net = container_of(work, typeof(*net), login_work.work);
692 unsigned long delay = msecs_to_jiffies(TBNET_LOGIN_DELAY);
695 if (netif_carrier_ok(net->dev))
698 netdev_dbg(net->dev, "sending login request, retries=%u\n",
701 ret = tbnet_login_request(net, net->login_retries % 4);
703 netdev_dbg(net->dev, "sending login request failed, ret=%d\n",
705 if (net->login_retries++ < TBNET_LOGIN_RETRIES) {
706 queue_delayed_work(system_long_wq, &net->login_work,
709 netdev_info(net->dev, "ThunderboltIP login timed out\n");
712 netdev_dbg(net->dev, "received login reply\n");
714 net->login_retries = 0;
716 mutex_lock(&net->connection_lock);
717 net->login_sent = true;
718 mutex_unlock(&net->connection_lock);
720 queue_work(system_long_wq, &net->connected_work);
724 static void tbnet_disconnect_work(struct work_struct *work)
726 struct tbnet *net = container_of(work, typeof(*net), disconnect_work);
728 tbnet_tear_down(net, false);
731 static bool tbnet_check_frame(struct tbnet *net, const struct tbnet_frame *tf,
732 const struct thunderbolt_ip_frame_header *hdr)
734 u32 frame_id, frame_count, frame_size, frame_index;
737 if (tf->frame.flags & RING_DESC_CRC_ERROR) {
738 net->stats.rx_crc_errors++;
740 } else if (tf->frame.flags & RING_DESC_BUFFER_OVERRUN) {
741 net->stats.rx_over_errors++;
745 /* Should be greater than just header i.e. contains data */
746 size = tbnet_frame_size(tf);
747 if (size <= sizeof(*hdr)) {
748 net->stats.rx_length_errors++;
752 frame_count = le32_to_cpu(hdr->frame_count);
753 frame_size = le32_to_cpu(hdr->frame_size);
754 frame_index = le16_to_cpu(hdr->frame_index);
755 frame_id = le16_to_cpu(hdr->frame_id);
757 if ((frame_size > size - sizeof(*hdr)) || !frame_size) {
758 net->stats.rx_length_errors++;
762 /* In case we're in the middle of packet, validate the frame
763 * header based on first fragment of the packet.
765 if (net->skb && net->rx_hdr.frame_count) {
766 /* Check the frame count fits the count field */
767 if (frame_count != le32_to_cpu(net->rx_hdr.frame_count)) {
768 net->stats.rx_length_errors++;
772 /* Check the frame identifiers are incremented correctly,
773 * and id is matching.
775 if (frame_index != le16_to_cpu(net->rx_hdr.frame_index) + 1 ||
776 frame_id != le16_to_cpu(net->rx_hdr.frame_id)) {
777 net->stats.rx_missed_errors++;
781 if (net->skb->len + frame_size > TBNET_MAX_MTU) {
782 net->stats.rx_length_errors++;
789 /* Start of packet, validate the frame header */
790 if (frame_count == 0 || frame_count > TBNET_RING_SIZE / 4) {
791 net->stats.rx_length_errors++;
794 if (frame_index != 0) {
795 net->stats.rx_missed_errors++;
802 static int tbnet_poll(struct napi_struct *napi, int budget)
804 struct tbnet *net = container_of(napi, struct tbnet, napi);
805 unsigned int cleaned_count = tbnet_available_buffers(&net->rx_ring);
806 struct device *dma_dev = tb_ring_dma_device(net->rx_ring.ring);
807 unsigned int rx_packets = 0;
809 while (rx_packets < budget) {
810 const struct thunderbolt_ip_frame_header *hdr;
811 unsigned int hdr_size = sizeof(*hdr);
812 struct sk_buff *skb = NULL;
813 struct ring_frame *frame;
814 struct tbnet_frame *tf;
819 /* Return some buffers to hardware, one at a time is too
820 * slow so allocate MAX_SKB_FRAGS buffers at the same
823 if (cleaned_count >= MAX_SKB_FRAGS) {
824 tbnet_alloc_rx_buffers(net, cleaned_count);
828 frame = tb_ring_poll(net->rx_ring.ring);
832 dma_unmap_page(dma_dev, frame->buffer_phy,
833 TBNET_RX_PAGE_SIZE, DMA_FROM_DEVICE);
835 tf = container_of(frame, typeof(*tf), frame);
842 hdr = page_address(page);
843 if (!tbnet_check_frame(net, tf, hdr)) {
844 trace_tbnet_invalid_rx_ip_frame(hdr->frame_size,
845 hdr->frame_id, hdr->frame_index, hdr->frame_count);
846 __free_pages(page, TBNET_RX_PAGE_ORDER);
847 dev_kfree_skb_any(net->skb);
852 trace_tbnet_rx_ip_frame(hdr->frame_size, hdr->frame_id,
853 hdr->frame_index, hdr->frame_count);
854 frame_size = le32_to_cpu(hdr->frame_size);
858 skb = build_skb(page_address(page),
861 __free_pages(page, TBNET_RX_PAGE_ORDER);
862 net->stats.rx_errors++;
866 skb_reserve(skb, hdr_size);
867 skb_put(skb, frame_size);
871 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags,
872 page, hdr_size, frame_size,
873 TBNET_RX_PAGE_SIZE - hdr_size);
876 net->rx_hdr.frame_size = hdr->frame_size;
877 net->rx_hdr.frame_count = hdr->frame_count;
878 net->rx_hdr.frame_index = hdr->frame_index;
879 net->rx_hdr.frame_id = hdr->frame_id;
880 last = le16_to_cpu(net->rx_hdr.frame_index) ==
881 le32_to_cpu(net->rx_hdr.frame_count) - 1;
884 net->stats.rx_bytes += frame_size;
887 skb->protocol = eth_type_trans(skb, net->dev);
888 trace_tbnet_rx_skb(skb);
889 napi_gro_receive(&net->napi, skb);
894 net->stats.rx_packets += rx_packets;
897 tbnet_alloc_rx_buffers(net, cleaned_count);
899 if (rx_packets >= budget)
902 napi_complete_done(napi, rx_packets);
903 /* Re-enable the ring interrupt */
904 tb_ring_poll_complete(net->rx_ring.ring);
909 static void tbnet_start_poll(void *data)
911 struct tbnet *net = data;
913 napi_schedule(&net->napi);
916 static int tbnet_open(struct net_device *dev)
918 struct tbnet *net = netdev_priv(dev);
919 struct tb_xdomain *xd = net->xd;
920 u16 sof_mask, eof_mask;
921 struct tb_ring *ring;
925 netif_carrier_off(dev);
927 ring = tb_ring_alloc_tx(xd->tb->nhi, -1, TBNET_RING_SIZE,
930 netdev_err(dev, "failed to allocate Tx ring\n");
933 net->tx_ring.ring = ring;
935 hopid = tb_xdomain_alloc_out_hopid(xd, -1);
937 netdev_err(dev, "failed to allocate Tx HopID\n");
938 tb_ring_free(net->tx_ring.ring);
939 net->tx_ring.ring = NULL;
942 net->local_transmit_path = hopid;
944 sof_mask = BIT(TBIP_PDF_FRAME_START);
945 eof_mask = BIT(TBIP_PDF_FRAME_END);
947 flags = RING_FLAG_FRAME;
948 /* Only enable full E2E if the other end supports it too */
949 if (tbnet_e2e && net->svc->prtcstns & TBNET_E2E)
950 flags |= RING_FLAG_E2E;
952 ring = tb_ring_alloc_rx(xd->tb->nhi, -1, TBNET_RING_SIZE, flags,
953 net->tx_ring.ring->hop, sof_mask,
954 eof_mask, tbnet_start_poll, net);
956 netdev_err(dev, "failed to allocate Rx ring\n");
957 tb_xdomain_release_out_hopid(xd, hopid);
958 tb_ring_free(net->tx_ring.ring);
959 net->tx_ring.ring = NULL;
962 net->rx_ring.ring = ring;
964 napi_enable(&net->napi);
970 static int tbnet_stop(struct net_device *dev)
972 struct tbnet *net = netdev_priv(dev);
974 napi_disable(&net->napi);
976 cancel_work_sync(&net->disconnect_work);
977 tbnet_tear_down(net, true);
979 tb_ring_free(net->rx_ring.ring);
980 net->rx_ring.ring = NULL;
982 tb_xdomain_release_out_hopid(net->xd, net->local_transmit_path);
983 tb_ring_free(net->tx_ring.ring);
984 net->tx_ring.ring = NULL;
989 static bool tbnet_xmit_csum_and_map(struct tbnet *net, struct sk_buff *skb,
990 struct tbnet_frame **frames, u32 frame_count)
992 struct thunderbolt_ip_frame_header *hdr = page_address(frames[0]->page);
993 struct device *dma_dev = tb_ring_dma_device(net->tx_ring.ring);
994 unsigned int i, len, offset = skb_transport_offset(skb);
995 /* Remove payload length from checksum */
996 u32 paylen = skb->len - skb_transport_offset(skb);
997 __wsum wsum = (__force __wsum)htonl(paylen);
998 __be16 protocol = skb->protocol;
999 void *data = skb->data;
1000 void *dest = hdr + 1;
1003 if (skb->ip_summed != CHECKSUM_PARTIAL) {
1004 /* No need to calculate checksum so we just update the
1005 * total frame count and sync the frames for DMA.
1007 for (i = 0; i < frame_count; i++) {
1008 hdr = page_address(frames[i]->page);
1009 hdr->frame_count = cpu_to_le32(frame_count);
1010 trace_tbnet_tx_ip_frame(hdr->frame_size, hdr->frame_id,
1011 hdr->frame_index, hdr->frame_count);
1012 dma_sync_single_for_device(dma_dev,
1013 frames[i]->frame.buffer_phy,
1014 tbnet_frame_size(frames[i]), DMA_TO_DEVICE);
1020 if (protocol == htons(ETH_P_8021Q)) {
1021 struct vlan_hdr *vhdr, vh;
1023 vhdr = skb_header_pointer(skb, ETH_HLEN, sizeof(vh), &vh);
1027 protocol = vhdr->h_vlan_encapsulated_proto;
1030 /* Data points on the beginning of packet.
1031 * Check is the checksum absolute place in the packet.
1032 * ipcso will update IP checksum.
1033 * tucso will update TCP/UDP checksum.
1035 if (protocol == htons(ETH_P_IP)) {
1036 __sum16 *ipcso = dest + ((void *)&(ip_hdr(skb)->check) - data);
1039 *ipcso = ip_fast_csum(dest + skb_network_offset(skb),
1042 if (ip_hdr(skb)->protocol == IPPROTO_TCP)
1043 tucso = dest + ((void *)&(tcp_hdr(skb)->check) - data);
1044 else if (ip_hdr(skb)->protocol == IPPROTO_UDP)
1045 tucso = dest + ((void *)&(udp_hdr(skb)->check) - data);
1049 *tucso = ~csum_tcpudp_magic(ip_hdr(skb)->saddr,
1050 ip_hdr(skb)->daddr, 0,
1051 ip_hdr(skb)->protocol, 0);
1052 } else if (skb_is_gso(skb) && skb_is_gso_v6(skb)) {
1053 tucso = dest + ((void *)&(tcp_hdr(skb)->check) - data);
1054 *tucso = ~csum_ipv6_magic(&ipv6_hdr(skb)->saddr,
1055 &ipv6_hdr(skb)->daddr, 0,
1057 } else if (protocol == htons(ETH_P_IPV6)) {
1058 tucso = dest + skb_checksum_start_offset(skb) + skb->csum_offset;
1059 *tucso = ~csum_ipv6_magic(&ipv6_hdr(skb)->saddr,
1060 &ipv6_hdr(skb)->daddr, 0,
1061 ipv6_hdr(skb)->nexthdr, 0);
1066 /* First frame was headers, rest of the frames contain data.
1067 * Calculate checksum over each frame.
1069 for (i = 0; i < frame_count; i++) {
1070 hdr = page_address(frames[i]->page);
1071 dest = (void *)(hdr + 1) + offset;
1072 len = le32_to_cpu(hdr->frame_size) - offset;
1073 wsum = csum_partial(dest, len, wsum);
1074 hdr->frame_count = cpu_to_le32(frame_count);
1075 trace_tbnet_tx_ip_frame(hdr->frame_size, hdr->frame_id,
1076 hdr->frame_index, hdr->frame_count);
1081 *tucso = csum_fold(wsum);
1083 /* Checksum is finally calculated and we don't touch the memory
1084 * anymore, so DMA sync the frames now.
1086 for (i = 0; i < frame_count; i++) {
1087 dma_sync_single_for_device(dma_dev, frames[i]->frame.buffer_phy,
1088 tbnet_frame_size(frames[i]), DMA_TO_DEVICE);
1094 static void *tbnet_kmap_frag(struct sk_buff *skb, unsigned int frag_num,
1097 const skb_frag_t *frag = &skb_shinfo(skb)->frags[frag_num];
1099 *len = skb_frag_size(frag);
1100 return kmap_local_page(skb_frag_page(frag)) + skb_frag_off(frag);
1103 static netdev_tx_t tbnet_start_xmit(struct sk_buff *skb,
1104 struct net_device *dev)
1106 struct tbnet *net = netdev_priv(dev);
1107 struct tbnet_frame *frames[MAX_SKB_FRAGS];
1108 u16 frame_id = atomic_read(&net->frame_id);
1109 struct thunderbolt_ip_frame_header *hdr;
1110 unsigned int len = skb_headlen(skb);
1111 unsigned int data_len = skb->len;
1112 unsigned int nframes, i;
1113 unsigned int frag = 0;
1114 void *src = skb->data;
1115 u32 frame_index = 0;
1119 trace_tbnet_tx_skb(skb);
1121 nframes = DIV_ROUND_UP(data_len, TBNET_MAX_PAYLOAD_SIZE);
1122 if (tbnet_available_buffers(&net->tx_ring) < nframes) {
1123 netif_stop_queue(net->dev);
1124 return NETDEV_TX_BUSY;
1127 frames[frame_index] = tbnet_get_tx_buffer(net);
1128 if (!frames[frame_index])
1131 hdr = page_address(frames[frame_index]->page);
1134 /* If overall packet is bigger than the frame data size */
1135 while (data_len > TBNET_MAX_PAYLOAD_SIZE) {
1136 unsigned int size_left = TBNET_MAX_PAYLOAD_SIZE;
1138 hdr->frame_size = cpu_to_le32(TBNET_MAX_PAYLOAD_SIZE);
1139 hdr->frame_index = cpu_to_le16(frame_index);
1140 hdr->frame_id = cpu_to_le16(frame_id);
1143 if (len > size_left) {
1144 /* Copy data onto Tx buffer data with
1145 * full frame size then break and go to
1148 memcpy(dest, src, size_left);
1155 memcpy(dest, src, len);
1164 /* Ensure all fragments have been processed */
1165 if (frag < skb_shinfo(skb)->nr_frags) {
1166 /* Map and then unmap quickly */
1167 src = tbnet_kmap_frag(skb, frag++, &len);
1169 } else if (unlikely(size_left > 0)) {
1172 } while (size_left > 0);
1174 data_len -= TBNET_MAX_PAYLOAD_SIZE;
1177 frames[frame_index] = tbnet_get_tx_buffer(net);
1178 if (!frames[frame_index])
1181 hdr = page_address(frames[frame_index]->page);
1185 hdr->frame_size = cpu_to_le32(data_len);
1186 hdr->frame_index = cpu_to_le16(frame_index);
1187 hdr->frame_id = cpu_to_le16(frame_id);
1189 frames[frame_index]->frame.size = data_len + sizeof(*hdr);
1191 /* In case the remaining data_len is smaller than a frame */
1192 while (len < data_len) {
1193 memcpy(dest, src, len);
1202 if (frag < skb_shinfo(skb)->nr_frags) {
1203 src = tbnet_kmap_frag(skb, frag++, &len);
1205 } else if (unlikely(data_len > 0)) {
1210 memcpy(dest, src, data_len);
1215 if (!tbnet_xmit_csum_and_map(net, skb, frames, frame_index + 1))
1218 for (i = 0; i < frame_index + 1; i++)
1219 tb_ring_tx(net->tx_ring.ring, &frames[i]->frame);
1221 if (net->svc->prtcstns & TBNET_MATCH_FRAGS_ID)
1222 atomic_inc(&net->frame_id);
1224 net->stats.tx_packets++;
1225 net->stats.tx_bytes += skb->len;
1227 trace_tbnet_consume_skb(skb);
1228 dev_consume_skb_any(skb);
1230 return NETDEV_TX_OK;
1233 /* We can re-use the buffers */
1234 net->tx_ring.cons -= frame_index;
1236 dev_kfree_skb_any(skb);
1237 net->stats.tx_errors++;
1239 return NETDEV_TX_OK;
1242 static void tbnet_get_stats64(struct net_device *dev,
1243 struct rtnl_link_stats64 *stats)
1245 struct tbnet *net = netdev_priv(dev);
1247 stats->tx_packets = net->stats.tx_packets;
1248 stats->rx_packets = net->stats.rx_packets;
1249 stats->tx_bytes = net->stats.tx_bytes;
1250 stats->rx_bytes = net->stats.rx_bytes;
1251 stats->rx_errors = net->stats.rx_errors + net->stats.rx_length_errors +
1252 net->stats.rx_over_errors + net->stats.rx_crc_errors +
1253 net->stats.rx_missed_errors;
1254 stats->tx_errors = net->stats.tx_errors;
1255 stats->rx_length_errors = net->stats.rx_length_errors;
1256 stats->rx_over_errors = net->stats.rx_over_errors;
1257 stats->rx_crc_errors = net->stats.rx_crc_errors;
1258 stats->rx_missed_errors = net->stats.rx_missed_errors;
1261 static const struct net_device_ops tbnet_netdev_ops = {
1262 .ndo_open = tbnet_open,
1263 .ndo_stop = tbnet_stop,
1264 .ndo_start_xmit = tbnet_start_xmit,
1265 .ndo_get_stats64 = tbnet_get_stats64,
1268 static void tbnet_generate_mac(struct net_device *dev)
1270 const struct tbnet *net = netdev_priv(dev);
1271 const struct tb_xdomain *xd = net->xd;
1276 phy_port = tb_phy_port_from_link(TBNET_L0_PORT_NUM(xd->route));
1278 /* Unicast and locally administered MAC */
1279 addr[0] = phy_port << 4 | 0x02;
1280 hash = jhash2((u32 *)xd->local_uuid, 4, 0);
1281 memcpy(addr + 1, &hash, sizeof(hash));
1282 hash = jhash2((u32 *)xd->local_uuid, 4, hash);
1283 addr[5] = hash & 0xff;
1284 eth_hw_addr_set(dev, addr);
1287 static int tbnet_probe(struct tb_service *svc, const struct tb_service_id *id)
1289 struct tb_xdomain *xd = tb_service_parent(svc);
1290 struct net_device *dev;
1294 dev = alloc_etherdev(sizeof(*net));
1298 SET_NETDEV_DEV(dev, &svc->dev);
1300 net = netdev_priv(dev);
1301 INIT_DELAYED_WORK(&net->login_work, tbnet_login_work);
1302 INIT_WORK(&net->connected_work, tbnet_connected_work);
1303 INIT_WORK(&net->disconnect_work, tbnet_disconnect_work);
1304 mutex_init(&net->connection_lock);
1305 atomic_set(&net->command_id, 0);
1306 atomic_set(&net->frame_id, 0);
1311 tbnet_generate_mac(dev);
1313 strcpy(dev->name, "thunderbolt%d");
1314 dev->netdev_ops = &tbnet_netdev_ops;
1316 /* ThunderboltIP takes advantage of TSO packets but instead of
1317 * segmenting them we just split the packet into Thunderbolt
1318 * frames (maximum payload size of each frame is 4084 bytes) and
1319 * calculate checksum over the whole packet here.
1321 * The receiving side does the opposite if the host OS supports
1322 * LRO, otherwise it needs to split the large packet into MTU
1323 * sized smaller packets.
1325 * In order to receive large packets from the networking stack,
1326 * we need to announce support for most of the offloading
1329 dev->hw_features = NETIF_F_SG | NETIF_F_ALL_TSO | NETIF_F_GRO |
1330 NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM;
1331 dev->features = dev->hw_features | NETIF_F_HIGHDMA;
1332 dev->hard_header_len += sizeof(struct thunderbolt_ip_frame_header);
1334 netif_napi_add(dev, &net->napi, tbnet_poll);
1336 /* MTU range: 68 - 65522 */
1337 dev->min_mtu = ETH_MIN_MTU;
1338 dev->max_mtu = TBNET_MAX_MTU - ETH_HLEN;
1340 net->handler.uuid = &tbnet_svc_uuid;
1341 net->handler.callback = tbnet_handle_packet;
1342 net->handler.data = net;
1343 tb_register_protocol_handler(&net->handler);
1345 tb_service_set_drvdata(svc, net);
1347 ret = register_netdev(dev);
1349 tb_unregister_protocol_handler(&net->handler);
1357 static void tbnet_remove(struct tb_service *svc)
1359 struct tbnet *net = tb_service_get_drvdata(svc);
1361 unregister_netdev(net->dev);
1362 tb_unregister_protocol_handler(&net->handler);
1363 free_netdev(net->dev);
1366 static void tbnet_shutdown(struct tb_service *svc)
1368 tbnet_tear_down(tb_service_get_drvdata(svc), true);
1371 static int tbnet_suspend(struct device *dev)
1373 struct tb_service *svc = tb_to_service(dev);
1374 struct tbnet *net = tb_service_get_drvdata(svc);
1377 if (netif_running(net->dev)) {
1378 netif_device_detach(net->dev);
1379 tbnet_tear_down(net, true);
1382 tb_unregister_protocol_handler(&net->handler);
1386 static int tbnet_resume(struct device *dev)
1388 struct tb_service *svc = tb_to_service(dev);
1389 struct tbnet *net = tb_service_get_drvdata(svc);
1391 tb_register_protocol_handler(&net->handler);
1393 netif_carrier_off(net->dev);
1394 if (netif_running(net->dev)) {
1395 netif_device_attach(net->dev);
1402 static DEFINE_SIMPLE_DEV_PM_OPS(tbnet_pm_ops, tbnet_suspend, tbnet_resume);
1404 static const struct tb_service_id tbnet_ids[] = {
1405 { TB_SERVICE("network", 1) },
1408 MODULE_DEVICE_TABLE(tbsvc, tbnet_ids);
1410 static struct tb_service_driver tbnet_driver = {
1412 .owner = THIS_MODULE,
1413 .name = "thunderbolt-net",
1414 .pm = pm_sleep_ptr(&tbnet_pm_ops),
1416 .probe = tbnet_probe,
1417 .remove = tbnet_remove,
1418 .shutdown = tbnet_shutdown,
1419 .id_table = tbnet_ids,
1422 static int __init tbnet_init(void)
1427 tbnet_dir = tb_property_create_dir(&tbnet_dir_uuid);
1431 tb_property_add_immediate(tbnet_dir, "prtcid", 1);
1432 tb_property_add_immediate(tbnet_dir, "prtcvers", 1);
1433 tb_property_add_immediate(tbnet_dir, "prtcrevs", 1);
1435 flags = TBNET_MATCH_FRAGS_ID | TBNET_64K_FRAMES;
1438 tb_property_add_immediate(tbnet_dir, "prtcstns", flags);
1440 ret = tb_register_property_dir("network", tbnet_dir);
1444 ret = tb_register_service_driver(&tbnet_driver);
1446 goto err_unregister;
1451 tb_unregister_property_dir("network", tbnet_dir);
1453 tb_property_free_dir(tbnet_dir);
1457 module_init(tbnet_init);
1459 static void __exit tbnet_exit(void)
1461 tb_unregister_service_driver(&tbnet_driver);
1462 tb_unregister_property_dir("network", tbnet_dir);
1463 tb_property_free_dir(tbnet_dir);
1465 module_exit(tbnet_exit);
1470 MODULE_DESCRIPTION("Thunderbolt/USB4 network driver");
1471 MODULE_LICENSE("GPL v2");