1 // SPDX-License-Identifier: GPL-2.0
3 * Networking over Thunderbolt cable using Apple ThunderboltIP protocol
5 * Copyright (C) 2017, Intel Corporation
6 * Authors: Amir Levy <amir.jer.levy@intel.com>
7 * Michael Jamet <michael.jamet@intel.com>
8 * Mika Westerberg <mika.westerberg@linux.intel.com>
11 #include <linux/atomic.h>
12 #include <linux/highmem.h>
13 #include <linux/if_vlan.h>
14 #include <linux/jhash.h>
15 #include <linux/module.h>
16 #include <linux/etherdevice.h>
17 #include <linux/rtnetlink.h>
18 #include <linux/sizes.h>
19 #include <linux/thunderbolt.h>
20 #include <linux/uuid.h>
21 #include <linux/workqueue.h>
23 #include <net/ip6_checksum.h>
25 /* Protocol timeouts in ms */
26 #define TBNET_LOGIN_DELAY 4500
27 #define TBNET_LOGIN_TIMEOUT 500
28 #define TBNET_LOGOUT_TIMEOUT 100
30 #define TBNET_RING_SIZE 256
31 #define TBNET_LOCAL_PATH 0xf
32 #define TBNET_LOGIN_RETRIES 60
33 #define TBNET_LOGOUT_RETRIES 5
34 #define TBNET_MATCH_FRAGS_ID BIT(1)
35 #define TBNET_MAX_MTU SZ_64K
36 #define TBNET_FRAME_SIZE SZ_4K
37 #define TBNET_MAX_PAYLOAD_SIZE \
38 (TBNET_FRAME_SIZE - sizeof(struct thunderbolt_ip_frame_header))
39 /* Rx packets need to hold space for skb_shared_info */
40 #define TBNET_RX_MAX_SIZE \
41 (TBNET_FRAME_SIZE + SKB_DATA_ALIGN(sizeof(struct skb_shared_info)))
42 #define TBNET_RX_PAGE_ORDER get_order(TBNET_RX_MAX_SIZE)
43 #define TBNET_RX_PAGE_SIZE (PAGE_SIZE << TBNET_RX_PAGE_ORDER)
45 #define TBNET_L0_PORT_NUM(route) ((route) & GENMASK(5, 0))
48 * struct thunderbolt_ip_frame_header - Header for each Thunderbolt frame
49 * @frame_size: size of the data with the frame
50 * @frame_index: running index on the frames
51 * @frame_id: ID of the frame to match frames to specific packet
52 * @frame_count: how many frames assembles a full packet
54 * Each data frame passed to the high-speed DMA ring has this header. If
55 * the XDomain network directory announces that %TBNET_MATCH_FRAGS_ID is
56 * supported then @frame_id is filled, otherwise it stays %0.
58 struct thunderbolt_ip_frame_header {
65 enum thunderbolt_ip_frame_pdf {
66 TBIP_PDF_FRAME_START = 1,
70 enum thunderbolt_ip_type {
77 struct thunderbolt_ip_header {
82 uuid_t initiator_uuid;
88 #define TBIP_HDR_LENGTH_MASK GENMASK(5, 0)
89 #define TBIP_HDR_SN_MASK GENMASK(28, 27)
90 #define TBIP_HDR_SN_SHIFT 27
92 struct thunderbolt_ip_login {
93 struct thunderbolt_ip_header hdr;
99 #define TBIP_LOGIN_PROTO_VERSION 1
101 struct thunderbolt_ip_login_response {
102 struct thunderbolt_ip_header hdr;
105 u32 receiver_mac_len;
109 struct thunderbolt_ip_logout {
110 struct thunderbolt_ip_header hdr;
113 struct thunderbolt_ip_status {
114 struct thunderbolt_ip_header hdr;
125 u64 rx_length_errors;
128 u64 rx_missed_errors;
132 struct net_device *dev;
134 struct ring_frame frame;
138 struct tbnet_frame frames[TBNET_RING_SIZE];
141 struct tb_ring *ring;
145 * struct tbnet - ThunderboltIP network driver private data
146 * @svc: XDomain service the driver is bound to
147 * @xd: XDomain the service blongs to
148 * @handler: ThunderboltIP configuration protocol handler
149 * @dev: Networking device
150 * @napi: NAPI structure for Rx polling
151 * @stats: Network statistics
152 * @skb: Network packet that is currently processed on Rx path
153 * @command_id: ID used for next configuration protocol packet
154 * @login_sent: ThunderboltIP login message successfully sent
155 * @login_received: ThunderboltIP login message received from the remote
157 * @transmit_path: HopID the other end needs to use building the
158 * opposite side path.
159 * @connection_lock: Lock serializing access to @login_sent,
160 * @login_received and @transmit_path.
161 * @login_retries: Number of login retries currently done
162 * @login_work: Worker to send ThunderboltIP login packets
163 * @connected_work: Worker that finalizes the ThunderboltIP connection
164 * setup and enables DMA paths for high speed data
166 * @disconnect_work: Worker that handles tearing down the ThunderboltIP
168 * @rx_hdr: Copy of the currently processed Rx frame. Used when a
169 * network packet consists of multiple Thunderbolt frames.
170 * In host byte order.
171 * @rx_ring: Software ring holding Rx frames
172 * @frame_id: Frame ID use for next Tx packet
173 * (if %TBNET_MATCH_FRAGS_ID is supported in both ends)
174 * @tx_ring: Software ring holding Tx frames
177 const struct tb_service *svc;
178 struct tb_xdomain *xd;
179 struct tb_protocol_handler handler;
180 struct net_device *dev;
181 struct napi_struct napi;
182 struct tbnet_stats stats;
188 struct mutex connection_lock;
190 struct delayed_work login_work;
191 struct work_struct connected_work;
192 struct work_struct disconnect_work;
193 struct thunderbolt_ip_frame_header rx_hdr;
194 struct tbnet_ring rx_ring;
196 struct tbnet_ring tx_ring;
199 /* Network property directory UUID: c66189ca-1cce-4195-bdb8-49592e5f5a4f */
200 static const uuid_t tbnet_dir_uuid =
201 UUID_INIT(0xc66189ca, 0x1cce, 0x4195,
202 0xbd, 0xb8, 0x49, 0x59, 0x2e, 0x5f, 0x5a, 0x4f);
204 /* ThunderboltIP protocol UUID: 798f589e-3616-8a47-97c6-5664a920c8dd */
205 static const uuid_t tbnet_svc_uuid =
206 UUID_INIT(0x798f589e, 0x3616, 0x8a47,
207 0x97, 0xc6, 0x56, 0x64, 0xa9, 0x20, 0xc8, 0xdd);
209 static struct tb_property_dir *tbnet_dir;
211 static void tbnet_fill_header(struct thunderbolt_ip_header *hdr, u64 route,
212 u8 sequence, const uuid_t *initiator_uuid, const uuid_t *target_uuid,
213 enum thunderbolt_ip_type type, size_t size, u32 command_id)
217 /* Length does not include route_hi/lo and length_sn fields */
218 length_sn = (size - 3 * 4) / 4;
219 length_sn |= (sequence << TBIP_HDR_SN_SHIFT) & TBIP_HDR_SN_MASK;
221 hdr->route_hi = upper_32_bits(route);
222 hdr->route_lo = lower_32_bits(route);
223 hdr->length_sn = length_sn;
224 uuid_copy(&hdr->uuid, &tbnet_svc_uuid);
225 uuid_copy(&hdr->initiator_uuid, initiator_uuid);
226 uuid_copy(&hdr->target_uuid, target_uuid);
228 hdr->command_id = command_id;
231 static int tbnet_login_response(struct tbnet *net, u64 route, u8 sequence,
234 struct thunderbolt_ip_login_response reply;
235 struct tb_xdomain *xd = net->xd;
237 memset(&reply, 0, sizeof(reply));
238 tbnet_fill_header(&reply.hdr, route, sequence, xd->local_uuid,
239 xd->remote_uuid, TBIP_LOGIN_RESPONSE, sizeof(reply),
241 memcpy(reply.receiver_mac, net->dev->dev_addr, ETH_ALEN);
242 reply.receiver_mac_len = ETH_ALEN;
244 return tb_xdomain_response(xd, &reply, sizeof(reply),
245 TB_CFG_PKG_XDOMAIN_RESP);
248 static int tbnet_login_request(struct tbnet *net, u8 sequence)
250 struct thunderbolt_ip_login_response reply;
251 struct thunderbolt_ip_login request;
252 struct tb_xdomain *xd = net->xd;
254 memset(&request, 0, sizeof(request));
255 tbnet_fill_header(&request.hdr, xd->route, sequence, xd->local_uuid,
256 xd->remote_uuid, TBIP_LOGIN, sizeof(request),
257 atomic_inc_return(&net->command_id));
259 request.proto_version = TBIP_LOGIN_PROTO_VERSION;
260 request.transmit_path = TBNET_LOCAL_PATH;
262 return tb_xdomain_request(xd, &request, sizeof(request),
263 TB_CFG_PKG_XDOMAIN_RESP, &reply,
264 sizeof(reply), TB_CFG_PKG_XDOMAIN_RESP,
265 TBNET_LOGIN_TIMEOUT);
268 static int tbnet_logout_response(struct tbnet *net, u64 route, u8 sequence,
271 struct thunderbolt_ip_status reply;
272 struct tb_xdomain *xd = net->xd;
274 memset(&reply, 0, sizeof(reply));
275 tbnet_fill_header(&reply.hdr, route, sequence, xd->local_uuid,
276 xd->remote_uuid, TBIP_STATUS, sizeof(reply),
277 atomic_inc_return(&net->command_id));
278 return tb_xdomain_response(xd, &reply, sizeof(reply),
279 TB_CFG_PKG_XDOMAIN_RESP);
282 static int tbnet_logout_request(struct tbnet *net)
284 struct thunderbolt_ip_logout request;
285 struct thunderbolt_ip_status reply;
286 struct tb_xdomain *xd = net->xd;
288 memset(&request, 0, sizeof(request));
289 tbnet_fill_header(&request.hdr, xd->route, 0, xd->local_uuid,
290 xd->remote_uuid, TBIP_LOGOUT, sizeof(request),
291 atomic_inc_return(&net->command_id));
293 return tb_xdomain_request(xd, &request, sizeof(request),
294 TB_CFG_PKG_XDOMAIN_RESP, &reply,
295 sizeof(reply), TB_CFG_PKG_XDOMAIN_RESP,
296 TBNET_LOGOUT_TIMEOUT);
299 static void start_login(struct tbnet *net)
301 mutex_lock(&net->connection_lock);
302 net->login_sent = false;
303 net->login_received = false;
304 mutex_unlock(&net->connection_lock);
306 queue_delayed_work(system_long_wq, &net->login_work,
307 msecs_to_jiffies(1000));
310 static void stop_login(struct tbnet *net)
312 cancel_delayed_work_sync(&net->login_work);
313 cancel_work_sync(&net->connected_work);
316 static inline unsigned int tbnet_frame_size(const struct tbnet_frame *tf)
318 return tf->frame.size ? : TBNET_FRAME_SIZE;
321 static void tbnet_free_buffers(struct tbnet_ring *ring)
325 for (i = 0; i < TBNET_RING_SIZE; i++) {
326 struct device *dma_dev = tb_ring_dma_device(ring->ring);
327 struct tbnet_frame *tf = &ring->frames[i];
328 enum dma_data_direction dir;
335 if (ring->ring->is_tx) {
338 size = TBNET_FRAME_SIZE;
340 dir = DMA_FROM_DEVICE;
341 order = TBNET_RX_PAGE_ORDER;
342 size = TBNET_RX_PAGE_SIZE;
345 if (tf->frame.buffer_phy)
346 dma_unmap_page(dma_dev, tf->frame.buffer_phy, size,
349 __free_pages(tf->page, order);
357 static void tbnet_tear_down(struct tbnet *net, bool send_logout)
359 netif_carrier_off(net->dev);
360 netif_stop_queue(net->dev);
364 mutex_lock(&net->connection_lock);
366 if (net->login_sent && net->login_received) {
367 int retries = TBNET_LOGOUT_RETRIES;
369 while (send_logout && retries-- > 0) {
370 int ret = tbnet_logout_request(net);
371 if (ret != -ETIMEDOUT)
375 tb_ring_stop(net->rx_ring.ring);
376 tb_ring_stop(net->tx_ring.ring);
377 tbnet_free_buffers(&net->rx_ring);
378 tbnet_free_buffers(&net->tx_ring);
380 if (tb_xdomain_disable_paths(net->xd))
381 netdev_warn(net->dev, "failed to disable DMA paths\n");
384 net->login_retries = 0;
385 net->login_sent = false;
386 net->login_received = false;
388 mutex_unlock(&net->connection_lock);
391 static int tbnet_handle_packet(const void *buf, size_t size, void *data)
393 const struct thunderbolt_ip_login *pkg = buf;
394 struct tbnet *net = data;
400 /* Make sure the packet is for us */
401 if (size < sizeof(struct thunderbolt_ip_header))
403 if (!uuid_equal(&pkg->hdr.initiator_uuid, net->xd->remote_uuid))
405 if (!uuid_equal(&pkg->hdr.target_uuid, net->xd->local_uuid))
408 route = ((u64)pkg->hdr.route_hi << 32) | pkg->hdr.route_lo;
409 route &= ~BIT_ULL(63);
410 if (route != net->xd->route)
413 sequence = pkg->hdr.length_sn & TBIP_HDR_SN_MASK;
414 sequence >>= TBIP_HDR_SN_SHIFT;
415 command_id = pkg->hdr.command_id;
417 switch (pkg->hdr.type) {
419 if (!netif_running(net->dev))
422 ret = tbnet_login_response(net, route, sequence,
423 pkg->hdr.command_id);
425 mutex_lock(&net->connection_lock);
426 net->login_received = true;
427 net->transmit_path = pkg->transmit_path;
429 /* If we reached the number of max retries or
430 * previous logout, schedule another round of
433 if (net->login_retries >= TBNET_LOGIN_RETRIES ||
435 net->login_retries = 0;
436 queue_delayed_work(system_long_wq,
437 &net->login_work, 0);
439 mutex_unlock(&net->connection_lock);
441 queue_work(system_long_wq, &net->connected_work);
446 ret = tbnet_logout_response(net, route, sequence, command_id);
448 queue_work(system_long_wq, &net->disconnect_work);
456 netdev_warn(net->dev, "failed to send ThunderboltIP response\n");
461 static unsigned int tbnet_available_buffers(const struct tbnet_ring *ring)
463 return ring->prod - ring->cons;
466 static int tbnet_alloc_rx_buffers(struct tbnet *net, unsigned int nbuffers)
468 struct tbnet_ring *ring = &net->rx_ring;
472 struct device *dma_dev = tb_ring_dma_device(ring->ring);
473 unsigned int index = ring->prod & (TBNET_RING_SIZE - 1);
474 struct tbnet_frame *tf = &ring->frames[index];
480 /* Allocate page (order > 0) so that it can hold maximum
481 * ThunderboltIP frame (4kB) and the additional room for
482 * SKB shared info required by build_skb().
484 tf->page = dev_alloc_pages(TBNET_RX_PAGE_ORDER);
490 dma_addr = dma_map_page(dma_dev, tf->page, 0,
491 TBNET_RX_PAGE_SIZE, DMA_FROM_DEVICE);
492 if (dma_mapping_error(dma_dev, dma_addr)) {
497 tf->frame.buffer_phy = dma_addr;
500 tb_ring_rx(ring->ring, &tf->frame);
508 tbnet_free_buffers(ring);
512 static struct tbnet_frame *tbnet_get_tx_buffer(struct tbnet *net)
514 struct tbnet_ring *ring = &net->tx_ring;
515 struct device *dma_dev = tb_ring_dma_device(ring->ring);
516 struct tbnet_frame *tf;
519 if (!tbnet_available_buffers(ring))
522 index = ring->cons++ & (TBNET_RING_SIZE - 1);
524 tf = &ring->frames[index];
527 dma_sync_single_for_cpu(dma_dev, tf->frame.buffer_phy,
528 tbnet_frame_size(tf), DMA_TO_DEVICE);
533 static void tbnet_tx_callback(struct tb_ring *ring, struct ring_frame *frame,
536 struct tbnet_frame *tf = container_of(frame, typeof(*tf), frame);
537 struct tbnet *net = netdev_priv(tf->dev);
539 /* Return buffer to the ring */
542 if (tbnet_available_buffers(&net->tx_ring) >= TBNET_RING_SIZE / 2)
543 netif_wake_queue(net->dev);
546 static int tbnet_alloc_tx_buffers(struct tbnet *net)
548 struct tbnet_ring *ring = &net->tx_ring;
549 struct device *dma_dev = tb_ring_dma_device(ring->ring);
552 for (i = 0; i < TBNET_RING_SIZE; i++) {
553 struct tbnet_frame *tf = &ring->frames[i];
556 tf->page = alloc_page(GFP_KERNEL);
558 tbnet_free_buffers(ring);
562 dma_addr = dma_map_page(dma_dev, tf->page, 0, TBNET_FRAME_SIZE,
564 if (dma_mapping_error(dma_dev, dma_addr)) {
565 __free_page(tf->page);
567 tbnet_free_buffers(ring);
572 tf->frame.buffer_phy = dma_addr;
573 tf->frame.callback = tbnet_tx_callback;
574 tf->frame.sof = TBIP_PDF_FRAME_START;
575 tf->frame.eof = TBIP_PDF_FRAME_END;
579 ring->prod = TBNET_RING_SIZE - 1;
584 static void tbnet_connected_work(struct work_struct *work)
586 struct tbnet *net = container_of(work, typeof(*net), connected_work);
590 if (netif_carrier_ok(net->dev))
593 mutex_lock(&net->connection_lock);
594 connected = net->login_sent && net->login_received;
595 mutex_unlock(&net->connection_lock);
600 /* Both logins successful so enable the high-speed DMA paths and
601 * start the network device queue.
603 ret = tb_xdomain_enable_paths(net->xd, TBNET_LOCAL_PATH,
604 net->rx_ring.ring->hop,
606 net->tx_ring.ring->hop);
608 netdev_err(net->dev, "failed to enable DMA paths\n");
612 tb_ring_start(net->tx_ring.ring);
613 tb_ring_start(net->rx_ring.ring);
615 ret = tbnet_alloc_rx_buffers(net, TBNET_RING_SIZE);
619 ret = tbnet_alloc_tx_buffers(net);
621 goto err_free_rx_buffers;
623 netif_carrier_on(net->dev);
624 netif_start_queue(net->dev);
628 tbnet_free_buffers(&net->rx_ring);
630 tb_ring_stop(net->rx_ring.ring);
631 tb_ring_stop(net->tx_ring.ring);
634 static void tbnet_login_work(struct work_struct *work)
636 struct tbnet *net = container_of(work, typeof(*net), login_work.work);
637 unsigned long delay = msecs_to_jiffies(TBNET_LOGIN_DELAY);
640 if (netif_carrier_ok(net->dev))
643 ret = tbnet_login_request(net, net->login_retries % 4);
645 if (net->login_retries++ < TBNET_LOGIN_RETRIES) {
646 queue_delayed_work(system_long_wq, &net->login_work,
649 netdev_info(net->dev, "ThunderboltIP login timed out\n");
652 net->login_retries = 0;
654 mutex_lock(&net->connection_lock);
655 net->login_sent = true;
656 mutex_unlock(&net->connection_lock);
658 queue_work(system_long_wq, &net->connected_work);
662 static void tbnet_disconnect_work(struct work_struct *work)
664 struct tbnet *net = container_of(work, typeof(*net), disconnect_work);
666 tbnet_tear_down(net, false);
669 static bool tbnet_check_frame(struct tbnet *net, const struct tbnet_frame *tf,
670 const struct thunderbolt_ip_frame_header *hdr)
672 u32 frame_id, frame_count, frame_size, frame_index;
675 if (tf->frame.flags & RING_DESC_CRC_ERROR) {
676 net->stats.rx_crc_errors++;
678 } else if (tf->frame.flags & RING_DESC_BUFFER_OVERRUN) {
679 net->stats.rx_over_errors++;
683 /* Should be greater than just header i.e. contains data */
684 size = tbnet_frame_size(tf);
685 if (size <= sizeof(*hdr)) {
686 net->stats.rx_length_errors++;
690 frame_count = le32_to_cpu(hdr->frame_count);
691 frame_size = le32_to_cpu(hdr->frame_size);
692 frame_index = le16_to_cpu(hdr->frame_index);
693 frame_id = le16_to_cpu(hdr->frame_id);
695 if ((frame_size > size - sizeof(*hdr)) || !frame_size) {
696 net->stats.rx_length_errors++;
700 /* In case we're in the middle of packet, validate the frame
701 * header based on first fragment of the packet.
703 if (net->skb && net->rx_hdr.frame_count) {
704 /* Check the frame count fits the count field */
705 if (frame_count != net->rx_hdr.frame_count) {
706 net->stats.rx_length_errors++;
710 /* Check the frame identifiers are incremented correctly,
711 * and id is matching.
713 if (frame_index != net->rx_hdr.frame_index + 1 ||
714 frame_id != net->rx_hdr.frame_id) {
715 net->stats.rx_missed_errors++;
719 if (net->skb->len + frame_size > TBNET_MAX_MTU) {
720 net->stats.rx_length_errors++;
727 /* Start of packet, validate the frame header */
728 if (frame_count == 0 || frame_count > TBNET_RING_SIZE / 4) {
729 net->stats.rx_length_errors++;
732 if (frame_index != 0) {
733 net->stats.rx_missed_errors++;
740 static int tbnet_poll(struct napi_struct *napi, int budget)
742 struct tbnet *net = container_of(napi, struct tbnet, napi);
743 unsigned int cleaned_count = tbnet_available_buffers(&net->rx_ring);
744 struct device *dma_dev = tb_ring_dma_device(net->rx_ring.ring);
745 unsigned int rx_packets = 0;
747 while (rx_packets < budget) {
748 const struct thunderbolt_ip_frame_header *hdr;
749 unsigned int hdr_size = sizeof(*hdr);
750 struct sk_buff *skb = NULL;
751 struct ring_frame *frame;
752 struct tbnet_frame *tf;
757 /* Return some buffers to hardware, one at a time is too
758 * slow so allocate MAX_SKB_FRAGS buffers at the same
761 if (cleaned_count >= MAX_SKB_FRAGS) {
762 tbnet_alloc_rx_buffers(net, cleaned_count);
766 frame = tb_ring_poll(net->rx_ring.ring);
770 dma_unmap_page(dma_dev, frame->buffer_phy,
771 TBNET_RX_PAGE_SIZE, DMA_FROM_DEVICE);
773 tf = container_of(frame, typeof(*tf), frame);
780 hdr = page_address(page);
781 if (!tbnet_check_frame(net, tf, hdr)) {
782 __free_pages(page, TBNET_RX_PAGE_ORDER);
783 dev_kfree_skb_any(net->skb);
788 frame_size = le32_to_cpu(hdr->frame_size);
792 skb = build_skb(page_address(page),
795 __free_pages(page, TBNET_RX_PAGE_ORDER);
796 net->stats.rx_errors++;
800 skb_reserve(skb, hdr_size);
801 skb_put(skb, frame_size);
805 skb_add_rx_frag(skb, skb_shinfo(skb)->nr_frags,
806 page, hdr_size, frame_size,
807 TBNET_RX_PAGE_SIZE - hdr_size);
810 net->rx_hdr.frame_size = frame_size;
811 net->rx_hdr.frame_count = le32_to_cpu(hdr->frame_count);
812 net->rx_hdr.frame_index = le16_to_cpu(hdr->frame_index);
813 net->rx_hdr.frame_id = le16_to_cpu(hdr->frame_id);
814 last = net->rx_hdr.frame_index == net->rx_hdr.frame_count - 1;
817 net->stats.rx_bytes += frame_size;
820 skb->protocol = eth_type_trans(skb, net->dev);
821 napi_gro_receive(&net->napi, skb);
826 net->stats.rx_packets += rx_packets;
829 tbnet_alloc_rx_buffers(net, cleaned_count);
831 if (rx_packets >= budget)
834 napi_complete_done(napi, rx_packets);
835 /* Re-enable the ring interrupt */
836 tb_ring_poll_complete(net->rx_ring.ring);
841 static void tbnet_start_poll(void *data)
843 struct tbnet *net = data;
845 napi_schedule(&net->napi);
848 static int tbnet_open(struct net_device *dev)
850 struct tbnet *net = netdev_priv(dev);
851 struct tb_xdomain *xd = net->xd;
852 u16 sof_mask, eof_mask;
853 struct tb_ring *ring;
855 netif_carrier_off(dev);
857 ring = tb_ring_alloc_tx(xd->tb->nhi, -1, TBNET_RING_SIZE,
860 netdev_err(dev, "failed to allocate Tx ring\n");
863 net->tx_ring.ring = ring;
865 sof_mask = BIT(TBIP_PDF_FRAME_START);
866 eof_mask = BIT(TBIP_PDF_FRAME_END);
868 ring = tb_ring_alloc_rx(xd->tb->nhi, -1, TBNET_RING_SIZE,
869 RING_FLAG_FRAME, sof_mask, eof_mask,
870 tbnet_start_poll, net);
872 netdev_err(dev, "failed to allocate Rx ring\n");
873 tb_ring_free(net->tx_ring.ring);
874 net->tx_ring.ring = NULL;
877 net->rx_ring.ring = ring;
879 napi_enable(&net->napi);
885 static int tbnet_stop(struct net_device *dev)
887 struct tbnet *net = netdev_priv(dev);
889 napi_disable(&net->napi);
891 cancel_work_sync(&net->disconnect_work);
892 tbnet_tear_down(net, true);
894 tb_ring_free(net->rx_ring.ring);
895 net->rx_ring.ring = NULL;
896 tb_ring_free(net->tx_ring.ring);
897 net->tx_ring.ring = NULL;
902 static bool tbnet_xmit_csum_and_map(struct tbnet *net, struct sk_buff *skb,
903 struct tbnet_frame **frames, u32 frame_count)
905 struct thunderbolt_ip_frame_header *hdr = page_address(frames[0]->page);
906 struct device *dma_dev = tb_ring_dma_device(net->tx_ring.ring);
907 __wsum wsum = htonl(skb->len - skb_transport_offset(skb));
908 unsigned int i, len, offset = skb_transport_offset(skb);
909 __be16 protocol = skb->protocol;
910 void *data = skb->data;
911 void *dest = hdr + 1;
914 if (skb->ip_summed != CHECKSUM_PARTIAL) {
915 /* No need to calculate checksum so we just update the
916 * total frame count and sync the frames for DMA.
918 for (i = 0; i < frame_count; i++) {
919 hdr = page_address(frames[i]->page);
920 hdr->frame_count = cpu_to_le32(frame_count);
921 dma_sync_single_for_device(dma_dev,
922 frames[i]->frame.buffer_phy,
923 tbnet_frame_size(frames[i]), DMA_TO_DEVICE);
929 if (protocol == htons(ETH_P_8021Q)) {
930 struct vlan_hdr *vhdr, vh;
932 vhdr = skb_header_pointer(skb, ETH_HLEN, sizeof(vh), &vh);
936 protocol = vhdr->h_vlan_encapsulated_proto;
939 /* Data points on the beginning of packet.
940 * Check is the checksum absolute place in the packet.
941 * ipcso will update IP checksum.
942 * tucso will update TCP/UPD checksum.
944 if (protocol == htons(ETH_P_IP)) {
945 __sum16 *ipcso = dest + ((void *)&(ip_hdr(skb)->check) - data);
948 *ipcso = ip_fast_csum(dest + skb_network_offset(skb),
951 if (ip_hdr(skb)->protocol == IPPROTO_TCP)
952 tucso = dest + ((void *)&(tcp_hdr(skb)->check) - data);
953 else if (ip_hdr(skb)->protocol == IPPROTO_UDP)
954 tucso = dest + ((void *)&(udp_hdr(skb)->check) - data);
958 *tucso = ~csum_tcpudp_magic(ip_hdr(skb)->saddr,
959 ip_hdr(skb)->daddr, 0,
960 ip_hdr(skb)->protocol, 0);
961 } else if (skb_is_gso(skb) && skb_is_gso_v6(skb)) {
962 tucso = dest + ((void *)&(tcp_hdr(skb)->check) - data);
963 *tucso = ~csum_ipv6_magic(&ipv6_hdr(skb)->saddr,
964 &ipv6_hdr(skb)->daddr, 0,
966 } else if (protocol == htons(ETH_P_IPV6)) {
967 tucso = dest + skb_checksum_start_offset(skb) + skb->csum_offset;
968 *tucso = ~csum_ipv6_magic(&ipv6_hdr(skb)->saddr,
969 &ipv6_hdr(skb)->daddr, 0,
970 ipv6_hdr(skb)->nexthdr, 0);
975 /* First frame was headers, rest of the frames contain data.
976 * Calculate checksum over each frame.
978 for (i = 0; i < frame_count; i++) {
979 hdr = page_address(frames[i]->page);
980 dest = (void *)(hdr + 1) + offset;
981 len = le32_to_cpu(hdr->frame_size) - offset;
982 wsum = csum_partial(dest, len, wsum);
983 hdr->frame_count = cpu_to_le32(frame_count);
988 *tucso = csum_fold(wsum);
990 /* Checksum is finally calculated and we don't touch the memory
991 * anymore, so DMA sync the frames now.
993 for (i = 0; i < frame_count; i++) {
994 dma_sync_single_for_device(dma_dev, frames[i]->frame.buffer_phy,
995 tbnet_frame_size(frames[i]), DMA_TO_DEVICE);
1001 static void *tbnet_kmap_frag(struct sk_buff *skb, unsigned int frag_num,
1004 const skb_frag_t *frag = &skb_shinfo(skb)->frags[frag_num];
1006 *len = skb_frag_size(frag);
1007 return kmap_atomic(skb_frag_page(frag)) + skb_frag_off(frag);
1010 static netdev_tx_t tbnet_start_xmit(struct sk_buff *skb,
1011 struct net_device *dev)
1013 struct tbnet *net = netdev_priv(dev);
1014 struct tbnet_frame *frames[MAX_SKB_FRAGS];
1015 u16 frame_id = atomic_read(&net->frame_id);
1016 struct thunderbolt_ip_frame_header *hdr;
1017 unsigned int len = skb_headlen(skb);
1018 unsigned int data_len = skb->len;
1019 unsigned int nframes, i;
1020 unsigned int frag = 0;
1021 void *src = skb->data;
1022 u32 frame_index = 0;
1026 nframes = DIV_ROUND_UP(data_len, TBNET_MAX_PAYLOAD_SIZE);
1027 if (tbnet_available_buffers(&net->tx_ring) < nframes) {
1028 netif_stop_queue(net->dev);
1029 return NETDEV_TX_BUSY;
1032 frames[frame_index] = tbnet_get_tx_buffer(net);
1033 if (!frames[frame_index])
1036 hdr = page_address(frames[frame_index]->page);
1039 /* If overall packet is bigger than the frame data size */
1040 while (data_len > TBNET_MAX_PAYLOAD_SIZE) {
1041 unsigned int size_left = TBNET_MAX_PAYLOAD_SIZE;
1043 hdr->frame_size = cpu_to_le32(TBNET_MAX_PAYLOAD_SIZE);
1044 hdr->frame_index = cpu_to_le16(frame_index);
1045 hdr->frame_id = cpu_to_le16(frame_id);
1048 if (len > size_left) {
1049 /* Copy data onto Tx buffer data with
1050 * full frame size then break and go to
1053 memcpy(dest, src, size_left);
1060 memcpy(dest, src, len);
1069 /* Ensure all fragments have been processed */
1070 if (frag < skb_shinfo(skb)->nr_frags) {
1071 /* Map and then unmap quickly */
1072 src = tbnet_kmap_frag(skb, frag++, &len);
1074 } else if (unlikely(size_left > 0)) {
1077 } while (size_left > 0);
1079 data_len -= TBNET_MAX_PAYLOAD_SIZE;
1082 frames[frame_index] = tbnet_get_tx_buffer(net);
1083 if (!frames[frame_index])
1086 hdr = page_address(frames[frame_index]->page);
1090 hdr->frame_size = cpu_to_le32(data_len);
1091 hdr->frame_index = cpu_to_le16(frame_index);
1092 hdr->frame_id = cpu_to_le16(frame_id);
1094 frames[frame_index]->frame.size = data_len + sizeof(*hdr);
1096 /* In case the remaining data_len is smaller than a frame */
1097 while (len < data_len) {
1098 memcpy(dest, src, len);
1107 if (frag < skb_shinfo(skb)->nr_frags) {
1108 src = tbnet_kmap_frag(skb, frag++, &len);
1110 } else if (unlikely(data_len > 0)) {
1115 memcpy(dest, src, data_len);
1120 if (!tbnet_xmit_csum_and_map(net, skb, frames, frame_index + 1))
1123 for (i = 0; i < frame_index + 1; i++)
1124 tb_ring_tx(net->tx_ring.ring, &frames[i]->frame);
1126 if (net->svc->prtcstns & TBNET_MATCH_FRAGS_ID)
1127 atomic_inc(&net->frame_id);
1129 net->stats.tx_packets++;
1130 net->stats.tx_bytes += skb->len;
1132 dev_consume_skb_any(skb);
1134 return NETDEV_TX_OK;
1137 /* We can re-use the buffers */
1138 net->tx_ring.cons -= frame_index;
1140 dev_kfree_skb_any(skb);
1141 net->stats.tx_errors++;
1143 return NETDEV_TX_OK;
1146 static void tbnet_get_stats64(struct net_device *dev,
1147 struct rtnl_link_stats64 *stats)
1149 struct tbnet *net = netdev_priv(dev);
1151 stats->tx_packets = net->stats.tx_packets;
1152 stats->rx_packets = net->stats.rx_packets;
1153 stats->tx_bytes = net->stats.tx_bytes;
1154 stats->rx_bytes = net->stats.rx_bytes;
1155 stats->rx_errors = net->stats.rx_errors + net->stats.rx_length_errors +
1156 net->stats.rx_over_errors + net->stats.rx_crc_errors +
1157 net->stats.rx_missed_errors;
1158 stats->tx_errors = net->stats.tx_errors;
1159 stats->rx_length_errors = net->stats.rx_length_errors;
1160 stats->rx_over_errors = net->stats.rx_over_errors;
1161 stats->rx_crc_errors = net->stats.rx_crc_errors;
1162 stats->rx_missed_errors = net->stats.rx_missed_errors;
1165 static const struct net_device_ops tbnet_netdev_ops = {
1166 .ndo_open = tbnet_open,
1167 .ndo_stop = tbnet_stop,
1168 .ndo_start_xmit = tbnet_start_xmit,
1169 .ndo_get_stats64 = tbnet_get_stats64,
1172 static void tbnet_generate_mac(struct net_device *dev)
1174 const struct tbnet *net = netdev_priv(dev);
1175 const struct tb_xdomain *xd = net->xd;
1179 phy_port = tb_phy_port_from_link(TBNET_L0_PORT_NUM(xd->route));
1181 /* Unicast and locally administered MAC */
1182 dev->dev_addr[0] = phy_port << 4 | 0x02;
1183 hash = jhash2((u32 *)xd->local_uuid, 4, 0);
1184 memcpy(dev->dev_addr + 1, &hash, sizeof(hash));
1185 hash = jhash2((u32 *)xd->local_uuid, 4, hash);
1186 dev->dev_addr[5] = hash & 0xff;
1189 static int tbnet_probe(struct tb_service *svc, const struct tb_service_id *id)
1191 struct tb_xdomain *xd = tb_service_parent(svc);
1192 struct net_device *dev;
1196 dev = alloc_etherdev(sizeof(*net));
1200 SET_NETDEV_DEV(dev, &svc->dev);
1202 net = netdev_priv(dev);
1203 INIT_DELAYED_WORK(&net->login_work, tbnet_login_work);
1204 INIT_WORK(&net->connected_work, tbnet_connected_work);
1205 INIT_WORK(&net->disconnect_work, tbnet_disconnect_work);
1206 mutex_init(&net->connection_lock);
1207 atomic_set(&net->command_id, 0);
1208 atomic_set(&net->frame_id, 0);
1213 tbnet_generate_mac(dev);
1215 strcpy(dev->name, "thunderbolt%d");
1216 dev->netdev_ops = &tbnet_netdev_ops;
1218 /* ThunderboltIP takes advantage of TSO packets but instead of
1219 * segmenting them we just split the packet into Thunderbolt
1220 * frames (maximum payload size of each frame is 4084 bytes) and
1221 * calculate checksum over the whole packet here.
1223 * The receiving side does the opposite if the host OS supports
1224 * LRO, otherwise it needs to split the large packet into MTU
1225 * sized smaller packets.
1227 * In order to receive large packets from the networking stack,
1228 * we need to announce support for most of the offloading
1231 dev->hw_features = NETIF_F_SG | NETIF_F_ALL_TSO | NETIF_F_GRO |
1232 NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM;
1233 dev->features = dev->hw_features | NETIF_F_HIGHDMA;
1234 dev->hard_header_len += sizeof(struct thunderbolt_ip_frame_header);
1236 netif_napi_add(dev, &net->napi, tbnet_poll, NAPI_POLL_WEIGHT);
1238 /* MTU range: 68 - 65522 */
1239 dev->min_mtu = ETH_MIN_MTU;
1240 dev->max_mtu = TBNET_MAX_MTU - ETH_HLEN;
1242 net->handler.uuid = &tbnet_svc_uuid;
1243 net->handler.callback = tbnet_handle_packet,
1244 net->handler.data = net;
1245 tb_register_protocol_handler(&net->handler);
1247 tb_service_set_drvdata(svc, net);
1249 ret = register_netdev(dev);
1251 tb_unregister_protocol_handler(&net->handler);
1259 static void tbnet_remove(struct tb_service *svc)
1261 struct tbnet *net = tb_service_get_drvdata(svc);
1263 unregister_netdev(net->dev);
1264 tb_unregister_protocol_handler(&net->handler);
1265 free_netdev(net->dev);
1268 static void tbnet_shutdown(struct tb_service *svc)
1270 tbnet_tear_down(tb_service_get_drvdata(svc), true);
1273 static int __maybe_unused tbnet_suspend(struct device *dev)
1275 struct tb_service *svc = tb_to_service(dev);
1276 struct tbnet *net = tb_service_get_drvdata(svc);
1279 if (netif_running(net->dev)) {
1280 netif_device_detach(net->dev);
1281 tbnet_tear_down(net, true);
1284 tb_unregister_protocol_handler(&net->handler);
1288 static int __maybe_unused tbnet_resume(struct device *dev)
1290 struct tb_service *svc = tb_to_service(dev);
1291 struct tbnet *net = tb_service_get_drvdata(svc);
1293 tb_register_protocol_handler(&net->handler);
1295 netif_carrier_off(net->dev);
1296 if (netif_running(net->dev)) {
1297 netif_device_attach(net->dev);
1304 static const struct dev_pm_ops tbnet_pm_ops = {
1305 SET_SYSTEM_SLEEP_PM_OPS(tbnet_suspend, tbnet_resume)
1308 static const struct tb_service_id tbnet_ids[] = {
1309 { TB_SERVICE("network", 1) },
1312 MODULE_DEVICE_TABLE(tbsvc, tbnet_ids);
1314 static struct tb_service_driver tbnet_driver = {
1316 .owner = THIS_MODULE,
1317 .name = "thunderbolt-net",
1318 .pm = &tbnet_pm_ops,
1320 .probe = tbnet_probe,
1321 .remove = tbnet_remove,
1322 .shutdown = tbnet_shutdown,
1323 .id_table = tbnet_ids,
1326 static int __init tbnet_init(void)
1330 tbnet_dir = tb_property_create_dir(&tbnet_dir_uuid);
1334 tb_property_add_immediate(tbnet_dir, "prtcid", 1);
1335 tb_property_add_immediate(tbnet_dir, "prtcvers", 1);
1336 tb_property_add_immediate(tbnet_dir, "prtcrevs", 1);
1337 /* Currently only announce support for match frags ID (bit 1). Bit 0
1338 * is reserved for full E2E flow control which we do not support at
1341 tb_property_add_immediate(tbnet_dir, "prtcstns",
1342 TBNET_MATCH_FRAGS_ID);
1344 ret = tb_register_property_dir("network", tbnet_dir);
1348 ret = tb_register_service_driver(&tbnet_driver);
1350 goto err_unregister;
1355 tb_unregister_property_dir("network", tbnet_dir);
1357 tb_property_free_dir(tbnet_dir);
1361 module_init(tbnet_init);
1363 static void __exit tbnet_exit(void)
1365 tb_unregister_service_driver(&tbnet_driver);
1366 tb_unregister_property_dir("network", tbnet_dir);
1367 tb_property_free_dir(tbnet_dir);
1369 module_exit(tbnet_exit);
1371 MODULE_AUTHOR("Amir Levy <amir.jer.levy@intel.com>");
1372 MODULE_AUTHOR("Michael Jamet <michael.jamet@intel.com>");
1373 MODULE_AUTHOR("Mika Westerberg <mika.westerberg@linux.intel.com>");
1374 MODULE_DESCRIPTION("Thunderbolt network driver");
1375 MODULE_LICENSE("GPL v2");