1/*************************************************************************
2 * myri10ge.c: Myricom Myri-10G Ethernet driver.
3 *
4 * Copyright (C) 2005 - 2011 Myricom, Inc.
5 * All rights reserved.
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 * 1. Redistributions of source code must retain the above copyright
11 *    notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 *    notice, this list of conditions and the following disclaimer in the
14 *    documentation and/or other materials provided with the distribution.
15 * 3. Neither the name of Myricom, Inc. nor the names of its contributors
16 *    may be used to endorse or promote products derived from this software
17 *    without specific prior written permission.
18 *
19 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
20 * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22 * ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
23 * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29 * POSSIBILITY OF SUCH DAMAGE.
30 *
31 *
32 * If the eeprom on your board is not recent enough, you will need to get a
33 * newer firmware image at:
34 *   http://www.myri.com/scs/download-Myri10GE.html
35 *
36 * Contact Information:
37 *   <help@myri.com>
38 *   Myricom, Inc., 325N Santa Anita Avenue, Arcadia, CA 91006
39 *************************************************************************/
40
41#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
42
43#include <linux/tcp.h>
44#include <linux/netdevice.h>
45#include <linux/skbuff.h>
46#include <linux/string.h>
47#include <linux/module.h>
48#include <linux/pci.h>
49#include <linux/dma-mapping.h>
50#include <linux/etherdevice.h>
51#include <linux/if_ether.h>
52#include <linux/if_vlan.h>
53#include <linux/dca.h>
54#include <linux/ip.h>
55#include <linux/inet.h>
56#include <linux/in.h>
57#include <linux/ethtool.h>
58#include <linux/firmware.h>
59#include <linux/delay.h>
60#include <linux/timer.h>
61#include <linux/vmalloc.h>
62#include <linux/crc32.h>
63#include <linux/moduleparam.h>
64#include <linux/io.h>
65#include <linux/log2.h>
66#include <linux/slab.h>
67#include <linux/prefetch.h>
68#include <net/checksum.h>
69#include <net/ip.h>
70#include <net/tcp.h>
71#include <asm/byteorder.h>
72#include <asm/processor.h>
73
74#include "myri10ge_mcp.h"
75#include "myri10ge_mcp_gen_header.h"
76
77#define MYRI10GE_VERSION_STR "1.5.3-1.534"
78
79MODULE_DESCRIPTION("Myricom 10G driver (10GbE)");
80MODULE_AUTHOR("Maintainer: help@myri.com");
81MODULE_VERSION(MYRI10GE_VERSION_STR);
82MODULE_LICENSE("Dual BSD/GPL");
83
84#define MYRI10GE_MAX_ETHER_MTU 9014
85
86#define MYRI10GE_ETH_STOPPED 0
87#define MYRI10GE_ETH_STOPPING 1
88#define MYRI10GE_ETH_STARTING 2
89#define MYRI10GE_ETH_RUNNING 3
90#define MYRI10GE_ETH_OPEN_FAILED 4
91
92#define MYRI10GE_EEPROM_STRINGS_SIZE 256
93#define MYRI10GE_MAX_SEND_DESC_TSO ((65536 / 2048) * 2)
94
95#define MYRI10GE_NO_CONFIRM_DATA htonl(0xffffffff)
96#define MYRI10GE_NO_RESPONSE_RESULT 0xffffffff
97
98#define MYRI10GE_ALLOC_ORDER 0
99#define MYRI10GE_ALLOC_SIZE ((1 << MYRI10GE_ALLOC_ORDER) * PAGE_SIZE)
100#define MYRI10GE_MAX_FRAGS_PER_FRAME (MYRI10GE_MAX_ETHER_MTU/MYRI10GE_ALLOC_SIZE + 1)
101
102#define MYRI10GE_MAX_SLICES 32
103
104struct myri10ge_rx_buffer_state {
105	struct page *page;
106	int page_offset;
107	DEFINE_DMA_UNMAP_ADDR(bus);
108	DEFINE_DMA_UNMAP_LEN(len);
109};
110
111struct myri10ge_tx_buffer_state {
112	struct sk_buff *skb;
113	int last;
114	DEFINE_DMA_UNMAP_ADDR(bus);
115	DEFINE_DMA_UNMAP_LEN(len);
116};
117
118struct myri10ge_cmd {
119	u32 data0;
120	u32 data1;
121	u32 data2;
122};
123
124struct myri10ge_rx_buf {
125	struct mcp_kreq_ether_recv __iomem *lanai;	/* lanai ptr for recv ring */
126	struct mcp_kreq_ether_recv *shadow;	/* host shadow of recv ring */
127	struct myri10ge_rx_buffer_state *info;
128	struct page *page;
129	dma_addr_t bus;
130	int page_offset;
131	int cnt;
132	int fill_cnt;
133	int alloc_fail;
134	int mask;		/* number of rx slots -1 */
135	int watchdog_needed;
136};
137
138struct myri10ge_tx_buf {
139	struct mcp_kreq_ether_send __iomem *lanai;	/* lanai ptr for sendq */
140	__be32 __iomem *send_go;	/* "go" doorbell ptr */
141	__be32 __iomem *send_stop;	/* "stop" doorbell ptr */
142	struct mcp_kreq_ether_send *req_list;	/* host shadow of sendq */
143	char *req_bytes;
144	struct myri10ge_tx_buffer_state *info;
145	int mask;		/* number of transmit slots -1  */
146	int req ____cacheline_aligned;	/* transmit slots submitted     */
147	int pkt_start;		/* packets started */
148	int stop_queue;
149	int linearized;
150	int done ____cacheline_aligned;	/* transmit slots completed     */
151	int pkt_done;		/* packets completed */
152	int wake_queue;
153	int queue_active;
154};
155
156struct myri10ge_rx_done {
157	struct mcp_slot *entry;
158	dma_addr_t bus;
159	int cnt;
160	int idx;
161};
162
163struct myri10ge_slice_netstats {
164	unsigned long rx_packets;
165	unsigned long tx_packets;
166	unsigned long rx_bytes;
167	unsigned long tx_bytes;
168	unsigned long rx_dropped;
169	unsigned long tx_dropped;
170};
171
172struct myri10ge_slice_state {
173	struct myri10ge_tx_buf tx;	/* transmit ring        */
174	struct myri10ge_rx_buf rx_small;
175	struct myri10ge_rx_buf rx_big;
176	struct myri10ge_rx_done rx_done;
177	struct net_device *dev;
178	struct napi_struct napi;
179	struct myri10ge_priv *mgp;
180	struct myri10ge_slice_netstats stats;
181	__be32 __iomem *irq_claim;
182	struct mcp_irq_data *fw_stats;
183	dma_addr_t fw_stats_bus;
184	int watchdog_tx_done;
185	int watchdog_tx_req;
186	int watchdog_rx_done;
187	int stuck;
188#ifdef CONFIG_MYRI10GE_DCA
189	int cached_dca_tag;
190	int cpu;
191	__be32 __iomem *dca_tag;
192#endif
193	char irq_desc[32];
194};
195
196struct myri10ge_priv {
197	struct myri10ge_slice_state *ss;
198	int tx_boundary;	/* boundary transmits cannot cross */
199	int num_slices;
200	int running;		/* running?             */
201	int small_bytes;
202	int big_bytes;
203	int max_intr_slots;
204	struct net_device *dev;
205	u8 __iomem *sram;
206	int sram_size;
207	unsigned long board_span;
208	unsigned long iomem_base;
209	__be32 __iomem *irq_deassert;
210	char *mac_addr_string;
211	struct mcp_cmd_response *cmd;
212	dma_addr_t cmd_bus;
213	struct pci_dev *pdev;
214	int msi_enabled;
215	int msix_enabled;
216	struct msix_entry *msix_vectors;
217#ifdef CONFIG_MYRI10GE_DCA
218	int dca_enabled;
219	int relaxed_order;
220#endif
221	u32 link_state;
222	unsigned int rdma_tags_available;
223	int intr_coal_delay;
224	__be32 __iomem *intr_coal_delay_ptr;
225	int wc_cookie;
226	int down_cnt;
227	wait_queue_head_t down_wq;
228	struct work_struct watchdog_work;
229	struct timer_list watchdog_timer;
230	int watchdog_resets;
231	int watchdog_pause;
232	int pause;
233	bool fw_name_allocated;
234	char *fw_name;
235	char eeprom_strings[MYRI10GE_EEPROM_STRINGS_SIZE];
236	char *product_code_string;
237	char fw_version[128];
238	int fw_ver_major;
239	int fw_ver_minor;
240	int fw_ver_tiny;
241	int adopted_rx_filter_bug;
242	u8 mac_addr[ETH_ALEN];		/* eeprom mac address */
243	unsigned long serial_number;
244	int vendor_specific_offset;
245	int fw_multicast_support;
246	u32 features;
247	u32 max_tso6;
248	u32 read_dma;
249	u32 write_dma;
250	u32 read_write_dma;
251	u32 link_changes;
252	u32 msg_enable;
253	unsigned int board_number;
254	int rebooted;
255};
256
257static char *myri10ge_fw_unaligned = "myri10ge_ethp_z8e.dat";
258static char *myri10ge_fw_aligned = "myri10ge_eth_z8e.dat";
259static char *myri10ge_fw_rss_unaligned = "myri10ge_rss_ethp_z8e.dat";
260static char *myri10ge_fw_rss_aligned = "myri10ge_rss_eth_z8e.dat";
261MODULE_FIRMWARE("myri10ge_ethp_z8e.dat");
262MODULE_FIRMWARE("myri10ge_eth_z8e.dat");
263MODULE_FIRMWARE("myri10ge_rss_ethp_z8e.dat");
264MODULE_FIRMWARE("myri10ge_rss_eth_z8e.dat");
265
266/* Careful: must be accessed under kernel_param_lock() */
267static char *myri10ge_fw_name = NULL;
268module_param(myri10ge_fw_name, charp, 0644);
269MODULE_PARM_DESC(myri10ge_fw_name, "Firmware image name");
270
271#define MYRI10GE_MAX_BOARDS 8
272static char *myri10ge_fw_names[MYRI10GE_MAX_BOARDS] =
273    {[0 ... (MYRI10GE_MAX_BOARDS - 1)] = NULL };
274module_param_array_named(myri10ge_fw_names, myri10ge_fw_names, charp, NULL,
275			 0444);
276MODULE_PARM_DESC(myri10ge_fw_names, "Firmware image names per board");
277
278static int myri10ge_ecrc_enable = 1;
279module_param(myri10ge_ecrc_enable, int, 0444);
280MODULE_PARM_DESC(myri10ge_ecrc_enable, "Enable Extended CRC on PCI-E");
281
282static int myri10ge_small_bytes = -1;	/* -1 == auto */
283module_param(myri10ge_small_bytes, int, 0644);
284MODULE_PARM_DESC(myri10ge_small_bytes, "Threshold of small packets");
285
286static int myri10ge_msi = 1;	/* enable msi by default */
287module_param(myri10ge_msi, int, 0644);
288MODULE_PARM_DESC(myri10ge_msi, "Enable Message Signalled Interrupts");
289
290static int myri10ge_intr_coal_delay = 75;
291module_param(myri10ge_intr_coal_delay, int, 0444);
292MODULE_PARM_DESC(myri10ge_intr_coal_delay, "Interrupt coalescing delay");
293
294static int myri10ge_flow_control = 1;
295module_param(myri10ge_flow_control, int, 0444);
296MODULE_PARM_DESC(myri10ge_flow_control, "Pause parameter");
297
298static int myri10ge_deassert_wait = 1;
299module_param(myri10ge_deassert_wait, int, 0644);
300MODULE_PARM_DESC(myri10ge_deassert_wait,
301		 "Wait when deasserting legacy interrupts");
302
303static int myri10ge_force_firmware = 0;
304module_param(myri10ge_force_firmware, int, 0444);
305MODULE_PARM_DESC(myri10ge_force_firmware,
306		 "Force firmware to assume aligned completions");
307
308static int myri10ge_initial_mtu = MYRI10GE_MAX_ETHER_MTU - ETH_HLEN;
309module_param(myri10ge_initial_mtu, int, 0444);
310MODULE_PARM_DESC(myri10ge_initial_mtu, "Initial MTU");
311
312static int myri10ge_napi_weight = 64;
313module_param(myri10ge_napi_weight, int, 0444);
314MODULE_PARM_DESC(myri10ge_napi_weight, "Set NAPI weight");
315
316static int myri10ge_watchdog_timeout = 1;
317module_param(myri10ge_watchdog_timeout, int, 0444);
318MODULE_PARM_DESC(myri10ge_watchdog_timeout, "Set watchdog timeout");
319
320static int myri10ge_max_irq_loops = 1048576;
321module_param(myri10ge_max_irq_loops, int, 0444);
322MODULE_PARM_DESC(myri10ge_max_irq_loops,
323		 "Set stuck legacy IRQ detection threshold");
324
325#define MYRI10GE_MSG_DEFAULT NETIF_MSG_LINK
326
327static int myri10ge_debug = -1;	/* defaults above */
328module_param(myri10ge_debug, int, 0);
329MODULE_PARM_DESC(myri10ge_debug, "Debug level (0=none,...,16=all)");
330
331static int myri10ge_fill_thresh = 256;
332module_param(myri10ge_fill_thresh, int, 0644);
333MODULE_PARM_DESC(myri10ge_fill_thresh, "Number of empty rx slots allowed");
334
335static int myri10ge_reset_recover = 1;
336
337static int myri10ge_max_slices = 1;
338module_param(myri10ge_max_slices, int, 0444);
339MODULE_PARM_DESC(myri10ge_max_slices, "Max tx/rx queues");
340
341static int myri10ge_rss_hash = MXGEFW_RSS_HASH_TYPE_SRC_DST_PORT;
342module_param(myri10ge_rss_hash, int, 0444);
343MODULE_PARM_DESC(myri10ge_rss_hash, "Type of RSS hashing to do");
344
345static int myri10ge_dca = 1;
346module_param(myri10ge_dca, int, 0444);
347MODULE_PARM_DESC(myri10ge_dca, "Enable DCA if possible");
348
349#define MYRI10GE_FW_OFFSET 1024*1024
350#define MYRI10GE_HIGHPART_TO_U32(X) \
351(sizeof (X) == 8) ? ((u32)((u64)(X) >> 32)) : (0)
352#define MYRI10GE_LOWPART_TO_U32(X) ((u32)(X))
353
354#define myri10ge_pio_copy(to,from,size) __iowrite64_copy(to,from,size/8)
355
356static void myri10ge_set_multicast_list(struct net_device *dev);
357static netdev_tx_t myri10ge_sw_tso(struct sk_buff *skb,
358					 struct net_device *dev);
359
360static inline void put_be32(__be32 val, __be32 __iomem * p)
361{
362	__raw_writel((__force __u32) val, (__force void __iomem *)p);
363}
364
365static void myri10ge_get_stats(struct net_device *dev,
366			       struct rtnl_link_stats64 *stats);
367
368static void set_fw_name(struct myri10ge_priv *mgp, char *name, bool allocated)
369{
370	if (mgp->fw_name_allocated)
371		kfree(mgp->fw_name);
372	mgp->fw_name = name;
373	mgp->fw_name_allocated = allocated;
374}
375
376static int
377myri10ge_send_cmd(struct myri10ge_priv *mgp, u32 cmd,
378		  struct myri10ge_cmd *data, int atomic)
379{
380	struct mcp_cmd *buf;
381	char buf_bytes[sizeof(*buf) + 8];
382	struct mcp_cmd_response *response = mgp->cmd;
383	char __iomem *cmd_addr = mgp->sram + MXGEFW_ETH_CMD;
384	u32 dma_low, dma_high, result, value;
385	int sleep_total = 0;
386
387	/* ensure buf is aligned to 8 bytes */
388	buf = (struct mcp_cmd *)ALIGN((unsigned long)buf_bytes, 8);
389
390	buf->data0 = htonl(data->data0);
391	buf->data1 = htonl(data->data1);
392	buf->data2 = htonl(data->data2);
393	buf->cmd = htonl(cmd);
394	dma_low = MYRI10GE_LOWPART_TO_U32(mgp->cmd_bus);
395	dma_high = MYRI10GE_HIGHPART_TO_U32(mgp->cmd_bus);
396
397	buf->response_addr.low = htonl(dma_low);
398	buf->response_addr.high = htonl(dma_high);
399	response->result = htonl(MYRI10GE_NO_RESPONSE_RESULT);
400	mb();
401	myri10ge_pio_copy(cmd_addr, buf, sizeof(*buf));
402
403	/* wait up to 15ms. Longest command is the DMA benchmark,
404	 * which is capped at 5ms, but runs from a timeout handler
405	 * that runs every 7.8ms. So a 15ms timeout leaves us with
406	 * a 2.2ms margin
407	 */
408	if (atomic) {
409		/* if atomic is set, do not sleep,
410		 * and try to get the completion quickly
411		 * (1ms will be enough for those commands) */
412		for (sleep_total = 0;
413		     sleep_total < 1000 &&
414		     response->result == htonl(MYRI10GE_NO_RESPONSE_RESULT);
415		     sleep_total += 10) {
416			udelay(10);
417			mb();
418		}
419	} else {
420		/* use msleep for most command */
421		for (sleep_total = 0;
422		     sleep_total < 15 &&
423		     response->result == htonl(MYRI10GE_NO_RESPONSE_RESULT);
424		     sleep_total++)
425			msleep(1);
426	}
427
428	result = ntohl(response->result);
429	value = ntohl(response->data);
430	if (result != MYRI10GE_NO_RESPONSE_RESULT) {
431		if (result == 0) {
432			data->data0 = value;
433			return 0;
434		} else if (result == MXGEFW_CMD_UNKNOWN) {
435			return -ENOSYS;
436		} else if (result == MXGEFW_CMD_ERROR_UNALIGNED) {
437			return -E2BIG;
438		} else if (result == MXGEFW_CMD_ERROR_RANGE &&
439			   cmd == MXGEFW_CMD_ENABLE_RSS_QUEUES &&
440			   (data->
441			    data1 & MXGEFW_SLICE_ENABLE_MULTIPLE_TX_QUEUES) !=
442			   0) {
443			return -ERANGE;
444		} else {
445			dev_err(&mgp->pdev->dev,
446				"command %d failed, result = %d\n",
447				cmd, result);
448			return -ENXIO;
449		}
450	}
451
452	dev_err(&mgp->pdev->dev, "command %d timed out, result = %d\n",
453		cmd, result);
454	return -EAGAIN;
455}
456
457/*
458 * The eeprom strings on the lanaiX have the format
459 * SN=x\0
460 * MAC=x:x:x:x:x:x\0
461 * PT:ddd mmm xx xx:xx:xx xx\0
462 * PV:ddd mmm xx xx:xx:xx xx\0
463 */
464static int myri10ge_read_mac_addr(struct myri10ge_priv *mgp)
465{
466	char *ptr, *limit;
467	int i;
468
469	ptr = mgp->eeprom_strings;
470	limit = mgp->eeprom_strings + MYRI10GE_EEPROM_STRINGS_SIZE;
471
472	while (*ptr != '\0' && ptr < limit) {
473		if (memcmp(ptr, "MAC=", 4) == 0) {
474			ptr += 4;
475			mgp->mac_addr_string = ptr;
476			for (i = 0; i < 6; i++) {
477				if ((ptr + 2) > limit)
478					goto abort;
479				mgp->mac_addr[i] =
480				    simple_strtoul(ptr, &ptr, 16);
481				ptr += 1;
482			}
483		}
484		if (memcmp(ptr, "PC=", 3) == 0) {
485			ptr += 3;
486			mgp->product_code_string = ptr;
487		}
488		if (memcmp((const void *)ptr, "SN=", 3) == 0) {
489			ptr += 3;
490			mgp->serial_number = simple_strtoul(ptr, &ptr, 10);
491		}
492		while (ptr < limit && *ptr++) ;
493	}
494
495	return 0;
496
497abort:
498	dev_err(&mgp->pdev->dev, "failed to parse eeprom_strings\n");
499	return -ENXIO;
500}
501
502/*
503 * Enable or disable periodic RDMAs from the host to make certain
504 * chipsets resend dropped PCIe messages
505 */
506
507static void myri10ge_dummy_rdma(struct myri10ge_priv *mgp, int enable)
508{
509	char __iomem *submit;
510	__be32 buf[16] __attribute__ ((__aligned__(8)));
511	u32 dma_low, dma_high;
512	int i;
513
514	/* clear confirmation addr */
515	mgp->cmd->data = 0;
516	mb();
517
518	/* send a rdma command to the PCIe engine, and wait for the
519	 * response in the confirmation address.  The firmware should
520	 * write a -1 there to indicate it is alive and well
521	 */
522	dma_low = MYRI10GE_LOWPART_TO_U32(mgp->cmd_bus);
523	dma_high = MYRI10GE_HIGHPART_TO_U32(mgp->cmd_bus);
524
525	buf[0] = htonl(dma_high);	/* confirm addr MSW */
526	buf[1] = htonl(dma_low);	/* confirm addr LSW */
527	buf[2] = MYRI10GE_NO_CONFIRM_DATA;	/* confirm data */
528	buf[3] = htonl(dma_high);	/* dummy addr MSW */
529	buf[4] = htonl(dma_low);	/* dummy addr LSW */
530	buf[5] = htonl(enable);	/* enable? */
531
532	submit = mgp->sram + MXGEFW_BOOT_DUMMY_RDMA;
533
534	myri10ge_pio_copy(submit, &buf, sizeof(buf));
535	for (i = 0; mgp->cmd->data != MYRI10GE_NO_CONFIRM_DATA && i < 20; i++)
536		msleep(1);
537	if (mgp->cmd->data != MYRI10GE_NO_CONFIRM_DATA)
538		dev_err(&mgp->pdev->dev, "dummy rdma %s failed\n",
539			(enable ? "enable" : "disable"));
540}
541
542static int
543myri10ge_validate_firmware(struct myri10ge_priv *mgp,
544			   struct mcp_gen_header *hdr)
545{
546	struct device *dev = &mgp->pdev->dev;
547
548	/* check firmware type */
549	if (ntohl(hdr->mcp_type) != MCP_TYPE_ETH) {
550		dev_err(dev, "Bad firmware type: 0x%x\n", ntohl(hdr->mcp_type));
551		return -EINVAL;
552	}
553
554	/* save firmware version for ethtool */
555	strncpy(mgp->fw_version, hdr->version, sizeof(mgp->fw_version));
556	mgp->fw_version[sizeof(mgp->fw_version) - 1] = '\0';
557
558	sscanf(mgp->fw_version, "%d.%d.%d", &mgp->fw_ver_major,
559	       &mgp->fw_ver_minor, &mgp->fw_ver_tiny);
560
561	if (!(mgp->fw_ver_major == MXGEFW_VERSION_MAJOR &&
562	      mgp->fw_ver_minor == MXGEFW_VERSION_MINOR)) {
563		dev_err(dev, "Found firmware version %s\n", mgp->fw_version);
564		dev_err(dev, "Driver needs %d.%d\n", MXGEFW_VERSION_MAJOR,
565			MXGEFW_VERSION_MINOR);
566		return -EINVAL;
567	}
568	return 0;
569}
570
571static int myri10ge_load_hotplug_firmware(struct myri10ge_priv *mgp, u32 * size)
572{
573	unsigned crc, reread_crc;
574	const struct firmware *fw;
575	struct device *dev = &mgp->pdev->dev;
576	unsigned char *fw_readback;
577	struct mcp_gen_header *hdr;
578	size_t hdr_offset;
579	int status;
580	unsigned i;
581
582	if ((status = request_firmware(&fw, mgp->fw_name, dev)) < 0) {
583		dev_err(dev, "Unable to load %s firmware image via hotplug\n",
584			mgp->fw_name);
585		status = -EINVAL;
586		goto abort_with_nothing;
587	}
588
589	/* check size */
590
591	if (fw->size >= mgp->sram_size - MYRI10GE_FW_OFFSET ||
592	    fw->size < MCP_HEADER_PTR_OFFSET + 4) {
593		dev_err(dev, "Firmware size invalid:%d\n", (int)fw->size);
594		status = -EINVAL;
595		goto abort_with_fw;
596	}
597
598	/* check id */
599	hdr_offset = ntohl(*(__be32 *) (fw->data + MCP_HEADER_PTR_OFFSET));
600	if ((hdr_offset & 3) || hdr_offset + sizeof(*hdr) > fw->size) {
601		dev_err(dev, "Bad firmware file\n");
602		status = -EINVAL;
603		goto abort_with_fw;
604	}
605	hdr = (void *)(fw->data + hdr_offset);
606
607	status = myri10ge_validate_firmware(mgp, hdr);
608	if (status != 0)
609		goto abort_with_fw;
610
611	crc = crc32(~0, fw->data, fw->size);
612	for (i = 0; i < fw->size; i += 256) {
613		myri10ge_pio_copy(mgp->sram + MYRI10GE_FW_OFFSET + i,
614				  fw->data + i,
615				  min(256U, (unsigned)(fw->size - i)));
616		mb();
617		readb(mgp->sram);
618	}
619	fw_readback = vmalloc(fw->size);
620	if (!fw_readback) {
621		status = -ENOMEM;
622		goto abort_with_fw;
623	}
624	/* corruption checking is good for parity recovery and buggy chipset */
625	memcpy_fromio(fw_readback, mgp->sram + MYRI10GE_FW_OFFSET, fw->size);
626	reread_crc = crc32(~0, fw_readback, fw->size);
627	vfree(fw_readback);
628	if (crc != reread_crc) {
629		dev_err(dev, "CRC failed(fw-len=%u), got 0x%x (expect 0x%x)\n",
630			(unsigned)fw->size, reread_crc, crc);
631		status = -EIO;
632		goto abort_with_fw;
633	}
634	*size = (u32) fw->size;
635
636abort_with_fw:
637	release_firmware(fw);
638
639abort_with_nothing:
640	return status;
641}
642
643static int myri10ge_adopt_running_firmware(struct myri10ge_priv *mgp)
644{
645	struct mcp_gen_header *hdr;
646	struct device *dev = &mgp->pdev->dev;
647	const size_t bytes = sizeof(struct mcp_gen_header);
648	size_t hdr_offset;
649	int status;
650
651	/* find running firmware header */
652	hdr_offset = swab32(readl(mgp->sram + MCP_HEADER_PTR_OFFSET));
653
654	if ((hdr_offset & 3) || hdr_offset + sizeof(*hdr) > mgp->sram_size) {
655		dev_err(dev, "Running firmware has bad header offset (%d)\n",
656			(int)hdr_offset);
657		return -EIO;
658	}
659
660	/* copy header of running firmware from SRAM to host memory to
661	 * validate firmware */
662	hdr = kmalloc(bytes, GFP_KERNEL);
663	if (hdr == NULL)
664		return -ENOMEM;
665
666	memcpy_fromio(hdr, mgp->sram + hdr_offset, bytes);
667	status = myri10ge_validate_firmware(mgp, hdr);
668	kfree(hdr);
669
670	/* check to see if adopted firmware has bug where adopting
671	 * it will cause broadcasts to be filtered unless the NIC
672	 * is kept in ALLMULTI mode */
673	if (mgp->fw_ver_major == 1 && mgp->fw_ver_minor == 4 &&
674	    mgp->fw_ver_tiny >= 4 && mgp->fw_ver_tiny <= 11) {
675		mgp->adopted_rx_filter_bug = 1;
676		dev_warn(dev, "Adopting fw %d.%d.%d: "
677			 "working around rx filter bug\n",
678			 mgp->fw_ver_major, mgp->fw_ver_minor,
679			 mgp->fw_ver_tiny);
680	}
681	return status;
682}
683
684static int myri10ge_get_firmware_capabilities(struct myri10ge_priv *mgp)
685{
686	struct myri10ge_cmd cmd;
687	int status;
688
689	/* probe for IPv6 TSO support */
690	mgp->features = NETIF_F_SG | NETIF_F_HW_CSUM | NETIF_F_TSO;
691	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_MAX_TSO6_HDR_SIZE,
692				   &cmd, 0);
693	if (status == 0) {
694		mgp->max_tso6 = cmd.data0;
695		mgp->features |= NETIF_F_TSO6;
696	}
697
698	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_RX_RING_SIZE, &cmd, 0);
699	if (status != 0) {
700		dev_err(&mgp->pdev->dev,
701			"failed MXGEFW_CMD_GET_RX_RING_SIZE\n");
702		return -ENXIO;
703	}
704
705	mgp->max_intr_slots = 2 * (cmd.data0 / sizeof(struct mcp_dma_addr));
706
707	return 0;
708}
709
710static int myri10ge_load_firmware(struct myri10ge_priv *mgp, int adopt)
711{
712	char __iomem *submit;
713	__be32 buf[16] __attribute__ ((__aligned__(8)));
714	u32 dma_low, dma_high, size;
715	int status, i;
716
717	size = 0;
718	status = myri10ge_load_hotplug_firmware(mgp, &size);
719	if (status) {
720		if (!adopt)
721			return status;
722		dev_warn(&mgp->pdev->dev, "hotplug firmware loading failed\n");
723
724		/* Do not attempt to adopt firmware if there
725		 * was a bad crc */
726		if (status == -EIO)
727			return status;
728
729		status = myri10ge_adopt_running_firmware(mgp);
730		if (status != 0) {
731			dev_err(&mgp->pdev->dev,
732				"failed to adopt running firmware\n");
733			return status;
734		}
735		dev_info(&mgp->pdev->dev,
736			 "Successfully adopted running firmware\n");
737		if (mgp->tx_boundary == 4096) {
738			dev_warn(&mgp->pdev->dev,
739				 "Using firmware currently running on NIC"
740				 ".  For optimal\n");
741			dev_warn(&mgp->pdev->dev,
742				 "performance consider loading optimized "
743				 "firmware\n");
744			dev_warn(&mgp->pdev->dev, "via hotplug\n");
745		}
746
747		set_fw_name(mgp, "adopted", false);
748		mgp->tx_boundary = 2048;
749		myri10ge_dummy_rdma(mgp, 1);
750		status = myri10ge_get_firmware_capabilities(mgp);
751		return status;
752	}
753
754	/* clear confirmation addr */
755	mgp->cmd->data = 0;
756	mb();
757
758	/* send a reload command to the bootstrap MCP, and wait for the
759	 *  response in the confirmation address.  The firmware should
760	 * write a -1 there to indicate it is alive and well
761	 */
762	dma_low = MYRI10GE_LOWPART_TO_U32(mgp->cmd_bus);
763	dma_high = MYRI10GE_HIGHPART_TO_U32(mgp->cmd_bus);
764
765	buf[0] = htonl(dma_high);	/* confirm addr MSW */
766	buf[1] = htonl(dma_low);	/* confirm addr LSW */
767	buf[2] = MYRI10GE_NO_CONFIRM_DATA;	/* confirm data */
768
769	/* FIX: All newest firmware should un-protect the bottom of
770	 * the sram before handoff. However, the very first interfaces
771	 * do not. Therefore the handoff copy must skip the first 8 bytes
772	 */
773	buf[3] = htonl(MYRI10GE_FW_OFFSET + 8);	/* where the code starts */
774	buf[4] = htonl(size - 8);	/* length of code */
775	buf[5] = htonl(8);	/* where to copy to */
776	buf[6] = htonl(0);	/* where to jump to */
777
778	submit = mgp->sram + MXGEFW_BOOT_HANDOFF;
779
780	myri10ge_pio_copy(submit, &buf, sizeof(buf));
781	mb();
782	msleep(1);
783	mb();
784	i = 0;
785	while (mgp->cmd->data != MYRI10GE_NO_CONFIRM_DATA && i < 9) {
786		msleep(1 << i);
787		i++;
788	}
789	if (mgp->cmd->data != MYRI10GE_NO_CONFIRM_DATA) {
790		dev_err(&mgp->pdev->dev, "handoff failed\n");
791		return -ENXIO;
792	}
793	myri10ge_dummy_rdma(mgp, 1);
794	status = myri10ge_get_firmware_capabilities(mgp);
795
796	return status;
797}
798
799static int myri10ge_update_mac_address(struct myri10ge_priv *mgp, u8 * addr)
800{
801	struct myri10ge_cmd cmd;
802	int status;
803
804	cmd.data0 = ((addr[0] << 24) | (addr[1] << 16)
805		     | (addr[2] << 8) | addr[3]);
806
807	cmd.data1 = ((addr[4] << 8) | (addr[5]));
808
809	status = myri10ge_send_cmd(mgp, MXGEFW_SET_MAC_ADDRESS, &cmd, 0);
810	return status;
811}
812
813static int myri10ge_change_pause(struct myri10ge_priv *mgp, int pause)
814{
815	struct myri10ge_cmd cmd;
816	int status, ctl;
817
818	ctl = pause ? MXGEFW_ENABLE_FLOW_CONTROL : MXGEFW_DISABLE_FLOW_CONTROL;
819	status = myri10ge_send_cmd(mgp, ctl, &cmd, 0);
820
821	if (status) {
822		netdev_err(mgp->dev, "Failed to set flow control mode\n");
823		return status;
824	}
825	mgp->pause = pause;
826	return 0;
827}
828
829static void
830myri10ge_change_promisc(struct myri10ge_priv *mgp, int promisc, int atomic)
831{
832	struct myri10ge_cmd cmd;
833	int status, ctl;
834
835	ctl = promisc ? MXGEFW_ENABLE_PROMISC : MXGEFW_DISABLE_PROMISC;
836	status = myri10ge_send_cmd(mgp, ctl, &cmd, atomic);
837	if (status)
838		netdev_err(mgp->dev, "Failed to set promisc mode\n");
839}
840
841static int myri10ge_dma_test(struct myri10ge_priv *mgp, int test_type)
842{
843	struct myri10ge_cmd cmd;
844	int status;
845	u32 len;
846	struct page *dmatest_page;
847	dma_addr_t dmatest_bus;
848	char *test = " ";
849
850	dmatest_page = alloc_page(GFP_KERNEL);
851	if (!dmatest_page)
852		return -ENOMEM;
853	dmatest_bus = pci_map_page(mgp->pdev, dmatest_page, 0, PAGE_SIZE,
854				   DMA_BIDIRECTIONAL);
855	if (unlikely(pci_dma_mapping_error(mgp->pdev, dmatest_bus))) {
856		__free_page(dmatest_page);
857		return -ENOMEM;
858	}
859
860	/* Run a small DMA test.
861	 * The magic multipliers to the length tell the firmware
862	 * to do DMA read, write, or read+write tests.  The
863	 * results are returned in cmd.data0.  The upper 16
864	 * bits or the return is the number of transfers completed.
865	 * The lower 16 bits is the time in 0.5us ticks that the
866	 * transfers took to complete.
867	 */
868
869	len = mgp->tx_boundary;
870
871	cmd.data0 = MYRI10GE_LOWPART_TO_U32(dmatest_bus);
872	cmd.data1 = MYRI10GE_HIGHPART_TO_U32(dmatest_bus);
873	cmd.data2 = len * 0x10000;
874	status = myri10ge_send_cmd(mgp, test_type, &cmd, 0);
875	if (status != 0) {
876		test = "read";
877		goto abort;
878	}
879	mgp->read_dma = ((cmd.data0 >> 16) * len * 2) / (cmd.data0 & 0xffff);
880	cmd.data0 = MYRI10GE_LOWPART_TO_U32(dmatest_bus);
881	cmd.data1 = MYRI10GE_HIGHPART_TO_U32(dmatest_bus);
882	cmd.data2 = len * 0x1;
883	status = myri10ge_send_cmd(mgp, test_type, &cmd, 0);
884	if (status != 0) {
885		test = "write";
886		goto abort;
887	}
888	mgp->write_dma = ((cmd.data0 >> 16) * len * 2) / (cmd.data0 & 0xffff);
889
890	cmd.data0 = MYRI10GE_LOWPART_TO_U32(dmatest_bus);
891	cmd.data1 = MYRI10GE_HIGHPART_TO_U32(dmatest_bus);
892	cmd.data2 = len * 0x10001;
893	status = myri10ge_send_cmd(mgp, test_type, &cmd, 0);
894	if (status != 0) {
895		test = "read/write";
896		goto abort;
897	}
898	mgp->read_write_dma = ((cmd.data0 >> 16) * len * 2 * 2) /
899	    (cmd.data0 & 0xffff);
900
901abort:
902	pci_unmap_page(mgp->pdev, dmatest_bus, PAGE_SIZE, DMA_BIDIRECTIONAL);
903	put_page(dmatest_page);
904
905	if (status != 0 && test_type != MXGEFW_CMD_UNALIGNED_TEST)
906		dev_warn(&mgp->pdev->dev, "DMA %s benchmark failed: %d\n",
907			 test, status);
908
909	return status;
910}
911
912static int myri10ge_reset(struct myri10ge_priv *mgp)
913{
914	struct myri10ge_cmd cmd;
915	struct myri10ge_slice_state *ss;
916	int i, status;
917	size_t bytes;
918#ifdef CONFIG_MYRI10GE_DCA
919	unsigned long dca_tag_off;
920#endif
921
922	/* try to send a reset command to the card to see if it
923	 * is alive */
924	memset(&cmd, 0, sizeof(cmd));
925	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_RESET, &cmd, 0);
926	if (status != 0) {
927		dev_err(&mgp->pdev->dev, "failed reset\n");
928		return -ENXIO;
929	}
930
931	(void)myri10ge_dma_test(mgp, MXGEFW_DMA_TEST);
932	/*
933	 * Use non-ndis mcp_slot (eg, 4 bytes total,
934	 * no toeplitz hash value returned.  Older firmware will
935	 * not understand this command, but will use the correct
936	 * sized mcp_slot, so we ignore error returns
937	 */
938	cmd.data0 = MXGEFW_RSS_MCP_SLOT_TYPE_MIN;
939	(void)myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_RSS_MCP_SLOT_TYPE, &cmd, 0);
940
941	/* Now exchange information about interrupts  */
942
943	bytes = mgp->max_intr_slots * sizeof(*mgp->ss[0].rx_done.entry);
944	cmd.data0 = (u32) bytes;
945	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_INTRQ_SIZE, &cmd, 0);
946
947	/*
948	 * Even though we already know how many slices are supported
949	 * via myri10ge_probe_slices() MXGEFW_CMD_GET_MAX_RSS_QUEUES
950	 * has magic side effects, and must be called after a reset.
951	 * It must be called prior to calling any RSS related cmds,
952	 * including assigning an interrupt queue for anything but
953	 * slice 0.  It must also be called *after*
954	 * MXGEFW_CMD_SET_INTRQ_SIZE, since the intrq size is used by
955	 * the firmware to compute offsets.
956	 */
957
958	if (mgp->num_slices > 1) {
959
960		/* ask the maximum number of slices it supports */
961		status = myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_MAX_RSS_QUEUES,
962					   &cmd, 0);
963		if (status != 0) {
964			dev_err(&mgp->pdev->dev,
965				"failed to get number of slices\n");
966		}
967
968		/*
969		 * MXGEFW_CMD_ENABLE_RSS_QUEUES must be called prior
970		 * to setting up the interrupt queue DMA
971		 */
972
973		cmd.data0 = mgp->num_slices;
974		cmd.data1 = MXGEFW_SLICE_INTR_MODE_ONE_PER_SLICE;
975		if (mgp->dev->real_num_tx_queues > 1)
976			cmd.data1 |= MXGEFW_SLICE_ENABLE_MULTIPLE_TX_QUEUES;
977		status = myri10ge_send_cmd(mgp, MXGEFW_CMD_ENABLE_RSS_QUEUES,
978					   &cmd, 0);
979
980		/* Firmware older than 1.4.32 only supports multiple
981		 * RX queues, so if we get an error, first retry using a
982		 * single TX queue before giving up */
983		if (status != 0 && mgp->dev->real_num_tx_queues > 1) {
984			netif_set_real_num_tx_queues(mgp->dev, 1);
985			cmd.data0 = mgp->num_slices;
986			cmd.data1 = MXGEFW_SLICE_INTR_MODE_ONE_PER_SLICE;
987			status = myri10ge_send_cmd(mgp,
988						   MXGEFW_CMD_ENABLE_RSS_QUEUES,
989						   &cmd, 0);
990		}
991
992		if (status != 0) {
993			dev_err(&mgp->pdev->dev,
994				"failed to set number of slices\n");
995
996			return status;
997		}
998	}
999	for (i = 0; i < mgp->num_slices; i++) {
1000		ss = &mgp->ss[i];
1001		cmd.data0 = MYRI10GE_LOWPART_TO_U32(ss->rx_done.bus);
1002		cmd.data1 = MYRI10GE_HIGHPART_TO_U32(ss->rx_done.bus);
1003		cmd.data2 = i;
1004		status |= myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_INTRQ_DMA,
1005					    &cmd, 0);
1006	}
1007
1008	status |=
1009	    myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_IRQ_ACK_OFFSET, &cmd, 0);
1010	for (i = 0; i < mgp->num_slices; i++) {
1011		ss = &mgp->ss[i];
1012		ss->irq_claim =
1013		    (__iomem __be32 *) (mgp->sram + cmd.data0 + 8 * i);
1014	}
1015	status |= myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_IRQ_DEASSERT_OFFSET,
1016				    &cmd, 0);
1017	mgp->irq_deassert = (__iomem __be32 *) (mgp->sram + cmd.data0);
1018
1019	status |= myri10ge_send_cmd
1020	    (mgp, MXGEFW_CMD_GET_INTR_COAL_DELAY_OFFSET, &cmd, 0);
1021	mgp->intr_coal_delay_ptr = (__iomem __be32 *) (mgp->sram + cmd.data0);
1022	if (status != 0) {
1023		dev_err(&mgp->pdev->dev, "failed set interrupt parameters\n");
1024		return status;
1025	}
1026	put_be32(htonl(mgp->intr_coal_delay), mgp->intr_coal_delay_ptr);
1027
1028#ifdef CONFIG_MYRI10GE_DCA
1029	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_DCA_OFFSET, &cmd, 0);
1030	dca_tag_off = cmd.data0;
1031	for (i = 0; i < mgp->num_slices; i++) {
1032		ss = &mgp->ss[i];
1033		if (status == 0) {
1034			ss->dca_tag = (__iomem __be32 *)
1035			    (mgp->sram + dca_tag_off + 4 * i);
1036		} else {
1037			ss->dca_tag = NULL;
1038		}
1039	}
1040#endif				/* CONFIG_MYRI10GE_DCA */
1041
1042	/* reset mcp/driver shared state back to 0 */
1043
1044	mgp->link_changes = 0;
1045	for (i = 0; i < mgp->num_slices; i++) {
1046		ss = &mgp->ss[i];
1047
1048		memset(ss->rx_done.entry, 0, bytes);
1049		ss->tx.req = 0;
1050		ss->tx.done = 0;
1051		ss->tx.pkt_start = 0;
1052		ss->tx.pkt_done = 0;
1053		ss->rx_big.cnt = 0;
1054		ss->rx_small.cnt = 0;
1055		ss->rx_done.idx = 0;
1056		ss->rx_done.cnt = 0;
1057		ss->tx.wake_queue = 0;
1058		ss->tx.stop_queue = 0;
1059	}
1060
1061	status = myri10ge_update_mac_address(mgp, mgp->dev->dev_addr);
1062	myri10ge_change_pause(mgp, mgp->pause);
1063	myri10ge_set_multicast_list(mgp->dev);
1064	return status;
1065}
1066
1067#ifdef CONFIG_MYRI10GE_DCA
1068static int myri10ge_toggle_relaxed(struct pci_dev *pdev, int on)
1069{
1070	int ret;
1071	u16 ctl;
1072
1073	pcie_capability_read_word(pdev, PCI_EXP_DEVCTL, &ctl);
1074
1075	ret = (ctl & PCI_EXP_DEVCTL_RELAX_EN) >> 4;
1076	if (ret != on) {
1077		ctl &= ~PCI_EXP_DEVCTL_RELAX_EN;
1078		ctl |= (on << 4);
1079		pcie_capability_write_word(pdev, PCI_EXP_DEVCTL, ctl);
1080	}
1081	return ret;
1082}
1083
1084static void
1085myri10ge_write_dca(struct myri10ge_slice_state *ss, int cpu, int tag)
1086{
1087	ss->cached_dca_tag = tag;
1088	put_be32(htonl(tag), ss->dca_tag);
1089}
1090
1091static inline void myri10ge_update_dca(struct myri10ge_slice_state *ss)
1092{
1093	int cpu = get_cpu();
1094	int tag;
1095
1096	if (cpu != ss->cpu) {
1097		tag = dca3_get_tag(&ss->mgp->pdev->dev, cpu);
1098		if (ss->cached_dca_tag != tag)
1099			myri10ge_write_dca(ss, cpu, tag);
1100		ss->cpu = cpu;
1101	}
1102	put_cpu();
1103}
1104
1105static void myri10ge_setup_dca(struct myri10ge_priv *mgp)
1106{
1107	int err, i;
1108	struct pci_dev *pdev = mgp->pdev;
1109
1110	if (mgp->ss[0].dca_tag == NULL || mgp->dca_enabled)
1111		return;
1112	if (!myri10ge_dca) {
1113		dev_err(&pdev->dev, "dca disabled by administrator\n");
1114		return;
1115	}
1116	err = dca_add_requester(&pdev->dev);
1117	if (err) {
1118		if (err != -ENODEV)
1119			dev_err(&pdev->dev,
1120				"dca_add_requester() failed, err=%d\n", err);
1121		return;
1122	}
1123	mgp->relaxed_order = myri10ge_toggle_relaxed(pdev, 0);
1124	mgp->dca_enabled = 1;
1125	for (i = 0; i < mgp->num_slices; i++) {
1126		mgp->ss[i].cpu = -1;
1127		mgp->ss[i].cached_dca_tag = -1;
1128		myri10ge_update_dca(&mgp->ss[i]);
1129	}
1130}
1131
1132static void myri10ge_teardown_dca(struct myri10ge_priv *mgp)
1133{
1134	struct pci_dev *pdev = mgp->pdev;
1135
1136	if (!mgp->dca_enabled)
1137		return;
1138	mgp->dca_enabled = 0;
1139	if (mgp->relaxed_order)
1140		myri10ge_toggle_relaxed(pdev, 1);
1141	dca_remove_requester(&pdev->dev);
1142}
1143
1144static int myri10ge_notify_dca_device(struct device *dev, void *data)
1145{
1146	struct myri10ge_priv *mgp;
1147	unsigned long event;
1148
1149	mgp = dev_get_drvdata(dev);
1150	event = *(unsigned long *)data;
1151
1152	if (event == DCA_PROVIDER_ADD)
1153		myri10ge_setup_dca(mgp);
1154	else if (event == DCA_PROVIDER_REMOVE)
1155		myri10ge_teardown_dca(mgp);
1156	return 0;
1157}
1158#endif				/* CONFIG_MYRI10GE_DCA */
1159
1160static inline void
1161myri10ge_submit_8rx(struct mcp_kreq_ether_recv __iomem * dst,
1162		    struct mcp_kreq_ether_recv *src)
1163{
1164	__be32 low;
1165
1166	low = src->addr_low;
1167	src->addr_low = htonl(DMA_BIT_MASK(32));
1168	myri10ge_pio_copy(dst, src, 4 * sizeof(*src));
1169	mb();
1170	myri10ge_pio_copy(dst + 4, src + 4, 4 * sizeof(*src));
1171	mb();
1172	src->addr_low = low;
1173	put_be32(low, &dst->addr_low);
1174	mb();
1175}
1176
1177static void
1178myri10ge_alloc_rx_pages(struct myri10ge_priv *mgp, struct myri10ge_rx_buf *rx,
1179			int bytes, int watchdog)
1180{
1181	struct page *page;
1182	dma_addr_t bus;
1183	int idx;
1184#if MYRI10GE_ALLOC_SIZE > 4096
1185	int end_offset;
1186#endif
1187
1188	if (unlikely(rx->watchdog_needed && !watchdog))
1189		return;
1190
1191	/* try to refill entire ring */
1192	while (rx->fill_cnt != (rx->cnt + rx->mask + 1)) {
1193		idx = rx->fill_cnt & rx->mask;
1194		if (rx->page_offset + bytes <= MYRI10GE_ALLOC_SIZE) {
1195			/* we can use part of previous page */
1196			get_page(rx->page);
1197		} else {
1198			/* we need a new page */
1199			page =
1200			    alloc_pages(GFP_ATOMIC | __GFP_COMP,
1201					MYRI10GE_ALLOC_ORDER);
1202			if (unlikely(page == NULL)) {
1203				if (rx->fill_cnt - rx->cnt < 16)
1204					rx->watchdog_needed = 1;
1205				return;
1206			}
1207
1208			bus = pci_map_page(mgp->pdev, page, 0,
1209					   MYRI10GE_ALLOC_SIZE,
1210					   PCI_DMA_FROMDEVICE);
1211			if (unlikely(pci_dma_mapping_error(mgp->pdev, bus))) {
1212				__free_pages(page, MYRI10GE_ALLOC_ORDER);
1213				if (rx->fill_cnt - rx->cnt < 16)
1214					rx->watchdog_needed = 1;
1215				return;
1216			}
1217
1218			rx->page = page;
1219			rx->page_offset = 0;
1220			rx->bus = bus;
1221
1222		}
1223		rx->info[idx].page = rx->page;
1224		rx->info[idx].page_offset = rx->page_offset;
1225		/* note that this is the address of the start of the
1226		 * page */
1227		dma_unmap_addr_set(&rx->info[idx], bus, rx->bus);
1228		rx->shadow[idx].addr_low =
1229		    htonl(MYRI10GE_LOWPART_TO_U32(rx->bus) + rx->page_offset);
1230		rx->shadow[idx].addr_high =
1231		    htonl(MYRI10GE_HIGHPART_TO_U32(rx->bus));
1232
1233		/* start next packet on a cacheline boundary */
1234		rx->page_offset += SKB_DATA_ALIGN(bytes);
1235
1236#if MYRI10GE_ALLOC_SIZE > 4096
1237		/* don't cross a 4KB boundary */
1238		end_offset = rx->page_offset + bytes - 1;
1239		if ((unsigned)(rx->page_offset ^ end_offset) > 4095)
1240			rx->page_offset = end_offset & ~4095;
1241#endif
1242		rx->fill_cnt++;
1243
1244		/* copy 8 descriptors to the firmware at a time */
1245		if ((idx & 7) == 7) {
1246			myri10ge_submit_8rx(&rx->lanai[idx - 7],
1247					    &rx->shadow[idx - 7]);
1248		}
1249	}
1250}
1251
1252static inline void
1253myri10ge_unmap_rx_page(struct pci_dev *pdev,
1254		       struct myri10ge_rx_buffer_state *info, int bytes)
1255{
1256	/* unmap the recvd page if we're the only or last user of it */
1257	if (bytes >= MYRI10GE_ALLOC_SIZE / 2 ||
1258	    (info->page_offset + 2 * bytes) > MYRI10GE_ALLOC_SIZE) {
1259		pci_unmap_page(pdev, (dma_unmap_addr(info, bus)
1260				      & ~(MYRI10GE_ALLOC_SIZE - 1)),
1261			       MYRI10GE_ALLOC_SIZE, PCI_DMA_FROMDEVICE);
1262	}
1263}
1264
1265/*
1266 * GRO does not support acceleration of tagged vlan frames, and
1267 * this NIC does not support vlan tag offload, so we must pop
1268 * the tag ourselves to be able to achieve GRO performance that
1269 * is comparable to LRO.
1270 */
1271
1272static inline void
1273myri10ge_vlan_rx(struct net_device *dev, void *addr, struct sk_buff *skb)
1274{
1275	u8 *va;
1276	struct vlan_ethhdr *veh;
1277	skb_frag_t *frag;
1278	__wsum vsum;
1279
1280	va = addr;
1281	va += MXGEFW_PAD;
1282	veh = (struct vlan_ethhdr *)va;
1283	if ((dev->features & NETIF_F_HW_VLAN_CTAG_RX) ==
1284	    NETIF_F_HW_VLAN_CTAG_RX &&
1285	    veh->h_vlan_proto == htons(ETH_P_8021Q)) {
1286		/* fixup csum if needed */
1287		if (skb->ip_summed == CHECKSUM_COMPLETE) {
1288			vsum = csum_partial(va + ETH_HLEN, VLAN_HLEN, 0);
1289			skb->csum = csum_sub(skb->csum, vsum);
1290		}
1291		/* pop tag */
1292		__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), ntohs(veh->h_vlan_TCI));
1293		memmove(va + VLAN_HLEN, va, 2 * ETH_ALEN);
1294		skb->len -= VLAN_HLEN;
1295		skb->data_len -= VLAN_HLEN;
1296		frag = skb_shinfo(skb)->frags;
1297		skb_frag_off_add(frag, VLAN_HLEN);
1298		skb_frag_size_sub(frag, VLAN_HLEN);
1299	}
1300}
1301
1302#define MYRI10GE_HLEN 64 /* Bytes to copy from page to skb linear memory */
1303
1304static inline int
1305myri10ge_rx_done(struct myri10ge_slice_state *ss, int len, __wsum csum)
1306{
1307	struct myri10ge_priv *mgp = ss->mgp;
1308	struct sk_buff *skb;
1309	skb_frag_t *rx_frags;
1310	struct myri10ge_rx_buf *rx;
1311	int i, idx, remainder, bytes;
1312	struct pci_dev *pdev = mgp->pdev;
1313	struct net_device *dev = mgp->dev;
1314	u8 *va;
1315
1316	if (len <= mgp->small_bytes) {
1317		rx = &ss->rx_small;
1318		bytes = mgp->small_bytes;
1319	} else {
1320		rx = &ss->rx_big;
1321		bytes = mgp->big_bytes;
1322	}
1323
1324	len += MXGEFW_PAD;
1325	idx = rx->cnt & rx->mask;
1326	va = page_address(rx->info[idx].page) + rx->info[idx].page_offset;
1327	prefetch(va);
1328
1329	skb = napi_get_frags(&ss->napi);
1330	if (unlikely(skb == NULL)) {
1331		ss->stats.rx_dropped++;
1332		for (i = 0, remainder = len; remainder > 0; i++) {
1333			myri10ge_unmap_rx_page(pdev, &rx->info[idx], bytes);
1334			put_page(rx->info[idx].page);
1335			rx->cnt++;
1336			idx = rx->cnt & rx->mask;
1337			remainder -= MYRI10GE_ALLOC_SIZE;
1338		}
1339		return 0;
1340	}
1341	rx_frags = skb_shinfo(skb)->frags;
1342	/* Fill skb_frag_t(s) with data from our receive */
1343	for (i = 0, remainder = len; remainder > 0; i++) {
1344		myri10ge_unmap_rx_page(pdev, &rx->info[idx], bytes);
1345		skb_fill_page_desc(skb, i, rx->info[idx].page,
1346				   rx->info[idx].page_offset,
1347				   remainder < MYRI10GE_ALLOC_SIZE ?
1348				   remainder : MYRI10GE_ALLOC_SIZE);
1349		rx->cnt++;
1350		idx = rx->cnt & rx->mask;
1351		remainder -= MYRI10GE_ALLOC_SIZE;
1352	}
1353
1354	/* remove padding */
1355	skb_frag_off_add(&rx_frags[0], MXGEFW_PAD);
1356	skb_frag_size_sub(&rx_frags[0], MXGEFW_PAD);
1357	len -= MXGEFW_PAD;
1358
1359	skb->len = len;
1360	skb->data_len = len;
1361	skb->truesize += len;
1362	if (dev->features & NETIF_F_RXCSUM) {
1363		skb->ip_summed = CHECKSUM_COMPLETE;
1364		skb->csum = csum;
1365	}
1366	myri10ge_vlan_rx(mgp->dev, va, skb);
1367	skb_record_rx_queue(skb, ss - &mgp->ss[0]);
1368
1369	napi_gro_frags(&ss->napi);
1370
1371	return 1;
1372}
1373
1374static inline void
1375myri10ge_tx_done(struct myri10ge_slice_state *ss, int mcp_index)
1376{
1377	struct pci_dev *pdev = ss->mgp->pdev;
1378	struct myri10ge_tx_buf *tx = &ss->tx;
1379	struct netdev_queue *dev_queue;
1380	struct sk_buff *skb;
1381	int idx, len;
1382
1383	while (tx->pkt_done != mcp_index) {
1384		idx = tx->done & tx->mask;
1385		skb = tx->info[idx].skb;
1386
1387		/* Mark as free */
1388		tx->info[idx].skb = NULL;
1389		if (tx->info[idx].last) {
1390			tx->pkt_done++;
1391			tx->info[idx].last = 0;
1392		}
1393		tx->done++;
1394		len = dma_unmap_len(&tx->info[idx], len);
1395		dma_unmap_len_set(&tx->info[idx], len, 0);
1396		if (skb) {
1397			ss->stats.tx_bytes += skb->len;
1398			ss->stats.tx_packets++;
1399			dev_consume_skb_irq(skb);
1400			if (len)
1401				pci_unmap_single(pdev,
1402						 dma_unmap_addr(&tx->info[idx],
1403								bus), len,
1404						 PCI_DMA_TODEVICE);
1405		} else {
1406			if (len)
1407				pci_unmap_page(pdev,
1408					       dma_unmap_addr(&tx->info[idx],
1409							      bus), len,
1410					       PCI_DMA_TODEVICE);
1411		}
1412	}
1413
1414	dev_queue = netdev_get_tx_queue(ss->dev, ss - ss->mgp->ss);
1415	/*
1416	 * Make a minimal effort to prevent the NIC from polling an
1417	 * idle tx queue.  If we can't get the lock we leave the queue
1418	 * active. In this case, either a thread was about to start
1419	 * using the queue anyway, or we lost a race and the NIC will
1420	 * waste some of its resources polling an inactive queue for a
1421	 * while.
1422	 */
1423
1424	if ((ss->mgp->dev->real_num_tx_queues > 1) &&
1425	    __netif_tx_trylock(dev_queue)) {
1426		if (tx->req == tx->done) {
1427			tx->queue_active = 0;
1428			put_be32(htonl(1), tx->send_stop);
1429			mb();
1430		}
1431		__netif_tx_unlock(dev_queue);
1432	}
1433
1434	/* start the queue if we've stopped it */
1435	if (netif_tx_queue_stopped(dev_queue) &&
1436	    tx->req - tx->done < (tx->mask >> 1) &&
1437	    ss->mgp->running == MYRI10GE_ETH_RUNNING) {
1438		tx->wake_queue++;
1439		netif_tx_wake_queue(dev_queue);
1440	}
1441}
1442
1443static inline int
1444myri10ge_clean_rx_done(struct myri10ge_slice_state *ss, int budget)
1445{
1446	struct myri10ge_rx_done *rx_done = &ss->rx_done;
1447	struct myri10ge_priv *mgp = ss->mgp;
1448	unsigned long rx_bytes = 0;
1449	unsigned long rx_packets = 0;
1450	unsigned long rx_ok;
1451	int idx = rx_done->idx;
1452	int cnt = rx_done->cnt;
1453	int work_done = 0;
1454	u16 length;
1455	__wsum checksum;
1456
1457	while (rx_done->entry[idx].length != 0 && work_done < budget) {
1458		length = ntohs(rx_done->entry[idx].length);
1459		rx_done->entry[idx].length = 0;
1460		checksum = csum_unfold(rx_done->entry[idx].checksum);
1461		rx_ok = myri10ge_rx_done(ss, length, checksum);
1462		rx_packets += rx_ok;
1463		rx_bytes += rx_ok * (unsigned long)length;
1464		cnt++;
1465		idx = cnt & (mgp->max_intr_slots - 1);
1466		work_done++;
1467	}
1468	rx_done->idx = idx;
1469	rx_done->cnt = cnt;
1470	ss->stats.rx_packets += rx_packets;
1471	ss->stats.rx_bytes += rx_bytes;
1472
1473	/* restock receive rings if needed */
1474	if (ss->rx_small.fill_cnt - ss->rx_small.cnt < myri10ge_fill_thresh)
1475		myri10ge_alloc_rx_pages(mgp, &ss->rx_small,
1476					mgp->small_bytes + MXGEFW_PAD, 0);
1477	if (ss->rx_big.fill_cnt - ss->rx_big.cnt < myri10ge_fill_thresh)
1478		myri10ge_alloc_rx_pages(mgp, &ss->rx_big, mgp->big_bytes, 0);
1479
1480	return work_done;
1481}
1482
1483static inline void myri10ge_check_statblock(struct myri10ge_priv *mgp)
1484{
1485	struct mcp_irq_data *stats = mgp->ss[0].fw_stats;
1486
1487	if (unlikely(stats->stats_updated)) {
1488		unsigned link_up = ntohl(stats->link_up);
1489		if (mgp->link_state != link_up) {
1490			mgp->link_state = link_up;
1491
1492			if (mgp->link_state == MXGEFW_LINK_UP) {
1493				netif_info(mgp, link, mgp->dev, "link up\n");
1494				netif_carrier_on(mgp->dev);
1495				mgp->link_changes++;
1496			} else {
1497				netif_info(mgp, link, mgp->dev, "link %s\n",
1498					   (link_up == MXGEFW_LINK_MYRINET ?
1499					    "mismatch (Myrinet detected)" :
1500					    "down"));
1501				netif_carrier_off(mgp->dev);
1502				mgp->link_changes++;
1503			}
1504		}
1505		if (mgp->rdma_tags_available !=
1506		    ntohl(stats->rdma_tags_available)) {
1507			mgp->rdma_tags_available =
1508			    ntohl(stats->rdma_tags_available);
1509			netdev_warn(mgp->dev, "RDMA timed out! %d tags left\n",
1510				    mgp->rdma_tags_available);
1511		}
1512		mgp->down_cnt += stats->link_down;
1513		if (stats->link_down)
1514			wake_up(&mgp->down_wq);
1515	}
1516}
1517
1518static int myri10ge_poll(struct napi_struct *napi, int budget)
1519{
1520	struct myri10ge_slice_state *ss =
1521	    container_of(napi, struct myri10ge_slice_state, napi);
1522	int work_done;
1523
1524#ifdef CONFIG_MYRI10GE_DCA
1525	if (ss->mgp->dca_enabled)
1526		myri10ge_update_dca(ss);
1527#endif
1528	/* process as many rx events as NAPI will allow */
1529	work_done = myri10ge_clean_rx_done(ss, budget);
1530
1531	if (work_done < budget) {
1532		napi_complete_done(napi, work_done);
1533		put_be32(htonl(3), ss->irq_claim);
1534	}
1535	return work_done;
1536}
1537
1538static irqreturn_t myri10ge_intr(int irq, void *arg)
1539{
1540	struct myri10ge_slice_state *ss = arg;
1541	struct myri10ge_priv *mgp = ss->mgp;
1542	struct mcp_irq_data *stats = ss->fw_stats;
1543	struct myri10ge_tx_buf *tx = &ss->tx;
1544	u32 send_done_count;
1545	int i;
1546
1547	/* an interrupt on a non-zero receive-only slice is implicitly
1548	 * valid  since MSI-X irqs are not shared */
1549	if ((mgp->dev->real_num_tx_queues == 1) && (ss != mgp->ss)) {
1550		napi_schedule(&ss->napi);
1551		return IRQ_HANDLED;
1552	}
1553
1554	/* make sure it is our IRQ, and that the DMA has finished */
1555	if (unlikely(!stats->valid))
1556		return IRQ_NONE;
1557
1558	/* low bit indicates receives are present, so schedule
1559	 * napi poll handler */
1560	if (stats->valid & 1)
1561		napi_schedule(&ss->napi);
1562
1563	if (!mgp->msi_enabled && !mgp->msix_enabled) {
1564		put_be32(0, mgp->irq_deassert);
1565		if (!myri10ge_deassert_wait)
1566			stats->valid = 0;
1567		mb();
1568	} else
1569		stats->valid = 0;
1570
1571	/* Wait for IRQ line to go low, if using INTx */
1572	i = 0;
1573	while (1) {
1574		i++;
1575		/* check for transmit completes and receives */
1576		send_done_count = ntohl(stats->send_done_count);
1577		if (send_done_count != tx->pkt_done)
1578			myri10ge_tx_done(ss, (int)send_done_count);
1579		if (unlikely(i > myri10ge_max_irq_loops)) {
1580			netdev_warn(mgp->dev, "irq stuck?\n");
1581			stats->valid = 0;
1582			schedule_work(&mgp->watchdog_work);
1583		}
1584		if (likely(stats->valid == 0))
1585			break;
1586		cpu_relax();
1587		barrier();
1588	}
1589
1590	/* Only slice 0 updates stats */
1591	if (ss == mgp->ss)
1592		myri10ge_check_statblock(mgp);
1593
1594	put_be32(htonl(3), ss->irq_claim + 1);
1595	return IRQ_HANDLED;
1596}
1597
1598static int
1599myri10ge_get_link_ksettings(struct net_device *netdev,
1600			    struct ethtool_link_ksettings *cmd)
1601{
1602	struct myri10ge_priv *mgp = netdev_priv(netdev);
1603	char *ptr;
1604	int i;
1605
1606	cmd->base.autoneg = AUTONEG_DISABLE;
1607	cmd->base.speed = SPEED_10000;
1608	cmd->base.duplex = DUPLEX_FULL;
1609
1610	/*
1611	 * parse the product code to deterimine the interface type
1612	 * (CX4, XFP, Quad Ribbon Fiber) by looking at the character
1613	 * after the 3rd dash in the driver's cached copy of the
1614	 * EEPROM's product code string.
1615	 */
1616	ptr = mgp->product_code_string;
1617	if (ptr == NULL) {
1618		netdev_err(netdev, "Missing product code\n");
1619		return 0;
1620	}
1621	for (i = 0; i < 3; i++, ptr++) {
1622		ptr = strchr(ptr, '-');
1623		if (ptr == NULL) {
1624			netdev_err(netdev, "Invalid product code %s\n",
1625				   mgp->product_code_string);
1626			return 0;
1627		}
1628	}
1629	if (*ptr == '2')
1630		ptr++;
1631	if (*ptr == 'R' || *ptr == 'Q' || *ptr == 'S') {
1632		/* We've found either an XFP, quad ribbon fiber, or SFP+ */
1633		cmd->base.port = PORT_FIBRE;
1634		ethtool_link_ksettings_add_link_mode(cmd, supported, FIBRE);
1635		ethtool_link_ksettings_add_link_mode(cmd, advertising, FIBRE);
1636	} else {
1637		cmd->base.port = PORT_OTHER;
1638	}
1639
1640	return 0;
1641}
1642
1643static void
1644myri10ge_get_drvinfo(struct net_device *netdev, struct ethtool_drvinfo *info)
1645{
1646	struct myri10ge_priv *mgp = netdev_priv(netdev);
1647
1648	strlcpy(info->driver, "myri10ge", sizeof(info->driver));
1649	strlcpy(info->version, MYRI10GE_VERSION_STR, sizeof(info->version));
1650	strlcpy(info->fw_version, mgp->fw_version, sizeof(info->fw_version));
1651	strlcpy(info->bus_info, pci_name(mgp->pdev), sizeof(info->bus_info));
1652}
1653
1654static int
1655myri10ge_get_coalesce(struct net_device *netdev, struct ethtool_coalesce *coal)
1656{
1657	struct myri10ge_priv *mgp = netdev_priv(netdev);
1658
1659	coal->rx_coalesce_usecs = mgp->intr_coal_delay;
1660	return 0;
1661}
1662
1663static int
1664myri10ge_set_coalesce(struct net_device *netdev, struct ethtool_coalesce *coal)
1665{
1666	struct myri10ge_priv *mgp = netdev_priv(netdev);
1667
1668	mgp->intr_coal_delay = coal->rx_coalesce_usecs;
1669	put_be32(htonl(mgp->intr_coal_delay), mgp->intr_coal_delay_ptr);
1670	return 0;
1671}
1672
1673static void
1674myri10ge_get_pauseparam(struct net_device *netdev,
1675			struct ethtool_pauseparam *pause)
1676{
1677	struct myri10ge_priv *mgp = netdev_priv(netdev);
1678
1679	pause->autoneg = 0;
1680	pause->rx_pause = mgp->pause;
1681	pause->tx_pause = mgp->pause;
1682}
1683
1684static int
1685myri10ge_set_pauseparam(struct net_device *netdev,
1686			struct ethtool_pauseparam *pause)
1687{
1688	struct myri10ge_priv *mgp = netdev_priv(netdev);
1689
1690	if (pause->tx_pause != mgp->pause)
1691		return myri10ge_change_pause(mgp, pause->tx_pause);
1692	if (pause->rx_pause != mgp->pause)
1693		return myri10ge_change_pause(mgp, pause->rx_pause);
1694	if (pause->autoneg != 0)
1695		return -EINVAL;
1696	return 0;
1697}
1698
1699static void
1700myri10ge_get_ringparam(struct net_device *netdev,
1701		       struct ethtool_ringparam *ring)
1702{
1703	struct myri10ge_priv *mgp = netdev_priv(netdev);
1704
1705	ring->rx_mini_max_pending = mgp->ss[0].rx_small.mask + 1;
1706	ring->rx_max_pending = mgp->ss[0].rx_big.mask + 1;
1707	ring->rx_jumbo_max_pending = 0;
1708	ring->tx_max_pending = mgp->ss[0].tx.mask + 1;
1709	ring->rx_mini_pending = ring->rx_mini_max_pending;
1710	ring->rx_pending = ring->rx_max_pending;
1711	ring->rx_jumbo_pending = ring->rx_jumbo_max_pending;
1712	ring->tx_pending = ring->tx_max_pending;
1713}
1714
1715static const char myri10ge_gstrings_main_stats[][ETH_GSTRING_LEN] = {
1716	"rx_packets", "tx_packets", "rx_bytes", "tx_bytes", "rx_errors",
1717	"tx_errors", "rx_dropped", "tx_dropped", "multicast", "collisions",
1718	"rx_length_errors", "rx_over_errors", "rx_crc_errors",
1719	"rx_frame_errors", "rx_fifo_errors", "rx_missed_errors",
1720	"tx_aborted_errors", "tx_carrier_errors", "tx_fifo_errors",
1721	"tx_heartbeat_errors", "tx_window_errors",
1722	/* device-specific stats */
1723	"tx_boundary", "irq", "MSI", "MSIX",
1724	"read_dma_bw_MBs", "write_dma_bw_MBs", "read_write_dma_bw_MBs",
1725	"serial_number", "watchdog_resets",
1726#ifdef CONFIG_MYRI10GE_DCA
1727	"dca_capable_firmware", "dca_device_present",
1728#endif
1729	"link_changes", "link_up", "dropped_link_overflow",
1730	"dropped_link_error_or_filtered",
1731	"dropped_pause", "dropped_bad_phy", "dropped_bad_crc32",
1732	"dropped_unicast_filtered", "dropped_multicast_filtered",
1733	"dropped_runt", "dropped_overrun", "dropped_no_small_buffer",
1734	"dropped_no_big_buffer"
1735};
1736
1737static const char myri10ge_gstrings_slice_stats[][ETH_GSTRING_LEN] = {
1738	"----------- slice ---------",
1739	"tx_pkt_start", "tx_pkt_done", "tx_req", "tx_done",
1740	"rx_small_cnt", "rx_big_cnt",
1741	"wake_queue", "stop_queue", "tx_linearized",
1742};
1743
1744#define MYRI10GE_NET_STATS_LEN      21
1745#define MYRI10GE_MAIN_STATS_LEN  ARRAY_SIZE(myri10ge_gstrings_main_stats)
1746#define MYRI10GE_SLICE_STATS_LEN  ARRAY_SIZE(myri10ge_gstrings_slice_stats)
1747
1748static void
1749myri10ge_get_strings(struct net_device *netdev, u32 stringset, u8 * data)
1750{
1751	struct myri10ge_priv *mgp = netdev_priv(netdev);
1752	int i;
1753
1754	switch (stringset) {
1755	case ETH_SS_STATS:
1756		memcpy(data, *myri10ge_gstrings_main_stats,
1757		       sizeof(myri10ge_gstrings_main_stats));
1758		data += sizeof(myri10ge_gstrings_main_stats);
1759		for (i = 0; i < mgp->num_slices; i++) {
1760			memcpy(data, *myri10ge_gstrings_slice_stats,
1761			       sizeof(myri10ge_gstrings_slice_stats));
1762			data += sizeof(myri10ge_gstrings_slice_stats);
1763		}
1764		break;
1765	}
1766}
1767
1768static int myri10ge_get_sset_count(struct net_device *netdev, int sset)
1769{
1770	struct myri10ge_priv *mgp = netdev_priv(netdev);
1771
1772	switch (sset) {
1773	case ETH_SS_STATS:
1774		return MYRI10GE_MAIN_STATS_LEN +
1775		    mgp->num_slices * MYRI10GE_SLICE_STATS_LEN;
1776	default:
1777		return -EOPNOTSUPP;
1778	}
1779}
1780
1781static void
1782myri10ge_get_ethtool_stats(struct net_device *netdev,
1783			   struct ethtool_stats *stats, u64 * data)
1784{
1785	struct myri10ge_priv *mgp = netdev_priv(netdev);
1786	struct myri10ge_slice_state *ss;
1787	struct rtnl_link_stats64 link_stats;
1788	int slice;
1789	int i;
1790
1791	/* force stats update */
1792	memset(&link_stats, 0, sizeof(link_stats));
1793	(void)myri10ge_get_stats(netdev, &link_stats);
1794	for (i = 0; i < MYRI10GE_NET_STATS_LEN; i++)
1795		data[i] = ((u64 *)&link_stats)[i];
1796
1797	data[i++] = (unsigned int)mgp->tx_boundary;
1798	data[i++] = (unsigned int)mgp->pdev->irq;
1799	data[i++] = (unsigned int)mgp->msi_enabled;
1800	data[i++] = (unsigned int)mgp->msix_enabled;
1801	data[i++] = (unsigned int)mgp->read_dma;
1802	data[i++] = (unsigned int)mgp->write_dma;
1803	data[i++] = (unsigned int)mgp->read_write_dma;
1804	data[i++] = (unsigned int)mgp->serial_number;
1805	data[i++] = (unsigned int)mgp->watchdog_resets;
1806#ifdef CONFIG_MYRI10GE_DCA
1807	data[i++] = (unsigned int)(mgp->ss[0].dca_tag != NULL);
1808	data[i++] = (unsigned int)(mgp->dca_enabled);
1809#endif
1810	data[i++] = (unsigned int)mgp->link_changes;
1811
1812	/* firmware stats are useful only in the first slice */
1813	ss = &mgp->ss[0];
1814	data[i++] = (unsigned int)ntohl(ss->fw_stats->link_up);
1815	data[i++] = (unsigned int)ntohl(ss->fw_stats->dropped_link_overflow);
1816	data[i++] =
1817	    (unsigned int)ntohl(ss->fw_stats->dropped_link_error_or_filtered);
1818	data[i++] = (unsigned int)ntohl(ss->fw_stats->dropped_pause);
1819	data[i++] = (unsigned int)ntohl(ss->fw_stats->dropped_bad_phy);
1820	data[i++] = (unsigned int)ntohl(ss->fw_stats->dropped_bad_crc32);
1821	data[i++] = (unsigned int)ntohl(ss->fw_stats->dropped_unicast_filtered);
1822	data[i++] =
1823	    (unsigned int)ntohl(ss->fw_stats->dropped_multicast_filtered);
1824	data[i++] = (unsigned int)ntohl(ss->fw_stats->dropped_runt);
1825	data[i++] = (unsigned int)ntohl(ss->fw_stats->dropped_overrun);
1826	data[i++] = (unsigned int)ntohl(ss->fw_stats->dropped_no_small_buffer);
1827	data[i++] = (unsigned int)ntohl(ss->fw_stats->dropped_no_big_buffer);
1828
1829	for (slice = 0; slice < mgp->num_slices; slice++) {
1830		ss = &mgp->ss[slice];
1831		data[i++] = slice;
1832		data[i++] = (unsigned int)ss->tx.pkt_start;
1833		data[i++] = (unsigned int)ss->tx.pkt_done;
1834		data[i++] = (unsigned int)ss->tx.req;
1835		data[i++] = (unsigned int)ss->tx.done;
1836		data[i++] = (unsigned int)ss->rx_small.cnt;
1837		data[i++] = (unsigned int)ss->rx_big.cnt;
1838		data[i++] = (unsigned int)ss->tx.wake_queue;
1839		data[i++] = (unsigned int)ss->tx.stop_queue;
1840		data[i++] = (unsigned int)ss->tx.linearized;
1841	}
1842}
1843
1844static void myri10ge_set_msglevel(struct net_device *netdev, u32 value)
1845{
1846	struct myri10ge_priv *mgp = netdev_priv(netdev);
1847	mgp->msg_enable = value;
1848}
1849
1850static u32 myri10ge_get_msglevel(struct net_device *netdev)
1851{
1852	struct myri10ge_priv *mgp = netdev_priv(netdev);
1853	return mgp->msg_enable;
1854}
1855
1856/*
1857 * Use a low-level command to change the LED behavior. Rather than
1858 * blinking (which is the normal case), when identify is used, the
1859 * yellow LED turns solid.
1860 */
1861static int myri10ge_led(struct myri10ge_priv *mgp, int on)
1862{
1863	struct mcp_gen_header *hdr;
1864	struct device *dev = &mgp->pdev->dev;
1865	size_t hdr_off, pattern_off, hdr_len;
1866	u32 pattern = 0xfffffffe;
1867
1868	/* find running firmware header */
1869	hdr_off = swab32(readl(mgp->sram + MCP_HEADER_PTR_OFFSET));
1870	if ((hdr_off & 3) || hdr_off + sizeof(*hdr) > mgp->sram_size) {
1871		dev_err(dev, "Running firmware has bad header offset (%d)\n",
1872			(int)hdr_off);
1873		return -EIO;
1874	}
1875	hdr_len = swab32(readl(mgp->sram + hdr_off +
1876			       offsetof(struct mcp_gen_header, header_length)));
1877	pattern_off = hdr_off + offsetof(struct mcp_gen_header, led_pattern);
1878	if (pattern_off >= (hdr_len + hdr_off)) {
1879		dev_info(dev, "Firmware does not support LED identification\n");
1880		return -EINVAL;
1881	}
1882	if (!on)
1883		pattern = swab32(readl(mgp->sram + pattern_off + 4));
1884	writel(swab32(pattern), mgp->sram + pattern_off);
1885	return 0;
1886}
1887
1888static int
1889myri10ge_phys_id(struct net_device *netdev, enum ethtool_phys_id_state state)
1890{
1891	struct myri10ge_priv *mgp = netdev_priv(netdev);
1892	int rc;
1893
1894	switch (state) {
1895	case ETHTOOL_ID_ACTIVE:
1896		rc = myri10ge_led(mgp, 1);
1897		break;
1898
1899	case ETHTOOL_ID_INACTIVE:
1900		rc =  myri10ge_led(mgp, 0);
1901		break;
1902
1903	default:
1904		rc = -EINVAL;
1905	}
1906
1907	return rc;
1908}
1909
1910static const struct ethtool_ops myri10ge_ethtool_ops = {
1911	.supported_coalesce_params = ETHTOOL_COALESCE_RX_USECS,
1912	.get_drvinfo = myri10ge_get_drvinfo,
1913	.get_coalesce = myri10ge_get_coalesce,
1914	.set_coalesce = myri10ge_set_coalesce,
1915	.get_pauseparam = myri10ge_get_pauseparam,
1916	.set_pauseparam = myri10ge_set_pauseparam,
1917	.get_ringparam = myri10ge_get_ringparam,
1918	.get_link = ethtool_op_get_link,
1919	.get_strings = myri10ge_get_strings,
1920	.get_sset_count = myri10ge_get_sset_count,
1921	.get_ethtool_stats = myri10ge_get_ethtool_stats,
1922	.set_msglevel = myri10ge_set_msglevel,
1923	.get_msglevel = myri10ge_get_msglevel,
1924	.set_phys_id = myri10ge_phys_id,
1925	.get_link_ksettings = myri10ge_get_link_ksettings,
1926};
1927
1928static int myri10ge_allocate_rings(struct myri10ge_slice_state *ss)
1929{
1930	struct myri10ge_priv *mgp = ss->mgp;
1931	struct myri10ge_cmd cmd;
1932	struct net_device *dev = mgp->dev;
1933	int tx_ring_size, rx_ring_size;
1934	int tx_ring_entries, rx_ring_entries;
1935	int i, slice, status;
1936	size_t bytes;
1937
1938	/* get ring sizes */
1939	slice = ss - mgp->ss;
1940	cmd.data0 = slice;
1941	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_SEND_RING_SIZE, &cmd, 0);
1942	tx_ring_size = cmd.data0;
1943	cmd.data0 = slice;
1944	status |= myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_RX_RING_SIZE, &cmd, 0);
1945	if (status != 0)
1946		return status;
1947	rx_ring_size = cmd.data0;
1948
1949	tx_ring_entries = tx_ring_size / sizeof(struct mcp_kreq_ether_send);
1950	rx_ring_entries = rx_ring_size / sizeof(struct mcp_dma_addr);
1951	ss->tx.mask = tx_ring_entries - 1;
1952	ss->rx_small.mask = ss->rx_big.mask = rx_ring_entries - 1;
1953
1954	status = -ENOMEM;
1955
1956	/* allocate the host shadow rings */
1957
1958	bytes = 8 + (MYRI10GE_MAX_SEND_DESC_TSO + 4)
1959	    * sizeof(*ss->tx.req_list);
1960	ss->tx.req_bytes = kzalloc(bytes, GFP_KERNEL);
1961	if (ss->tx.req_bytes == NULL)
1962		goto abort_with_nothing;
1963
1964	/* ensure req_list entries are aligned to 8 bytes */
1965	ss->tx.req_list = (struct mcp_kreq_ether_send *)
1966	    ALIGN((unsigned long)ss->tx.req_bytes, 8);
1967	ss->tx.queue_active = 0;
1968
1969	bytes = rx_ring_entries * sizeof(*ss->rx_small.shadow);
1970	ss->rx_small.shadow = kzalloc(bytes, GFP_KERNEL);
1971	if (ss->rx_small.shadow == NULL)
1972		goto abort_with_tx_req_bytes;
1973
1974	bytes = rx_ring_entries * sizeof(*ss->rx_big.shadow);
1975	ss->rx_big.shadow = kzalloc(bytes, GFP_KERNEL);
1976	if (ss->rx_big.shadow == NULL)
1977		goto abort_with_rx_small_shadow;
1978
1979	/* allocate the host info rings */
1980
1981	bytes = tx_ring_entries * sizeof(*ss->tx.info);
1982	ss->tx.info = kzalloc(bytes, GFP_KERNEL);
1983	if (ss->tx.info == NULL)
1984		goto abort_with_rx_big_shadow;
1985
1986	bytes = rx_ring_entries * sizeof(*ss->rx_small.info);
1987	ss->rx_small.info = kzalloc(bytes, GFP_KERNEL);
1988	if (ss->rx_small.info == NULL)
1989		goto abort_with_tx_info;
1990
1991	bytes = rx_ring_entries * sizeof(*ss->rx_big.info);
1992	ss->rx_big.info = kzalloc(bytes, GFP_KERNEL);
1993	if (ss->rx_big.info == NULL)
1994		goto abort_with_rx_small_info;
1995
1996	/* Fill the receive rings */
1997	ss->rx_big.cnt = 0;
1998	ss->rx_small.cnt = 0;
1999	ss->rx_big.fill_cnt = 0;
2000	ss->rx_small.fill_cnt = 0;
2001	ss->rx_small.page_offset = MYRI10GE_ALLOC_SIZE;
2002	ss->rx_big.page_offset = MYRI10GE_ALLOC_SIZE;
2003	ss->rx_small.watchdog_needed = 0;
2004	ss->rx_big.watchdog_needed = 0;
2005	if (mgp->small_bytes == 0) {
2006		ss->rx_small.fill_cnt = ss->rx_small.mask + 1;
2007	} else {
2008		myri10ge_alloc_rx_pages(mgp, &ss->rx_small,
2009					mgp->small_bytes + MXGEFW_PAD, 0);
2010	}
2011
2012	if (ss->rx_small.fill_cnt < ss->rx_small.mask + 1) {
2013		netdev_err(dev, "slice-%d: alloced only %d small bufs\n",
2014			   slice, ss->rx_small.fill_cnt);
2015		goto abort_with_rx_small_ring;
2016	}
2017
2018	myri10ge_alloc_rx_pages(mgp, &ss->rx_big, mgp->big_bytes, 0);
2019	if (ss->rx_big.fill_cnt < ss->rx_big.mask + 1) {
2020		netdev_err(dev, "slice-%d: alloced only %d big bufs\n",
2021			   slice, ss->rx_big.fill_cnt);
2022		goto abort_with_rx_big_ring;
2023	}
2024
2025	return 0;
2026
2027abort_with_rx_big_ring:
2028	for (i = ss->rx_big.cnt; i < ss->rx_big.fill_cnt; i++) {
2029		int idx = i & ss->rx_big.mask;
2030		myri10ge_unmap_rx_page(mgp->pdev, &ss->rx_big.info[idx],
2031				       mgp->big_bytes);
2032		put_page(ss->rx_big.info[idx].page);
2033	}
2034
2035abort_with_rx_small_ring:
2036	if (mgp->small_bytes == 0)
2037		ss->rx_small.fill_cnt = ss->rx_small.cnt;
2038	for (i = ss->rx_small.cnt; i < ss->rx_small.fill_cnt; i++) {
2039		int idx = i & ss->rx_small.mask;
2040		myri10ge_unmap_rx_page(mgp->pdev, &ss->rx_small.info[idx],
2041				       mgp->small_bytes + MXGEFW_PAD);
2042		put_page(ss->rx_small.info[idx].page);
2043	}
2044
2045	kfree(ss->rx_big.info);
2046
2047abort_with_rx_small_info:
2048	kfree(ss->rx_small.info);
2049
2050abort_with_tx_info:
2051	kfree(ss->tx.info);
2052
2053abort_with_rx_big_shadow:
2054	kfree(ss->rx_big.shadow);
2055
2056abort_with_rx_small_shadow:
2057	kfree(ss->rx_small.shadow);
2058
2059abort_with_tx_req_bytes:
2060	kfree(ss->tx.req_bytes);
2061	ss->tx.req_bytes = NULL;
2062	ss->tx.req_list = NULL;
2063
2064abort_with_nothing:
2065	return status;
2066}
2067
2068static void myri10ge_free_rings(struct myri10ge_slice_state *ss)
2069{
2070	struct myri10ge_priv *mgp = ss->mgp;
2071	struct sk_buff *skb;
2072	struct myri10ge_tx_buf *tx;
2073	int i, len, idx;
2074
2075	/* If not allocated, skip it */
2076	if (ss->tx.req_list == NULL)
2077		return;
2078
2079	for (i = ss->rx_big.cnt; i < ss->rx_big.fill_cnt; i++) {
2080		idx = i & ss->rx_big.mask;
2081		if (i == ss->rx_big.fill_cnt - 1)
2082			ss->rx_big.info[idx].page_offset = MYRI10GE_ALLOC_SIZE;
2083		myri10ge_unmap_rx_page(mgp->pdev, &ss->rx_big.info[idx],
2084				       mgp->big_bytes);
2085		put_page(ss->rx_big.info[idx].page);
2086	}
2087
2088	if (mgp->small_bytes == 0)
2089		ss->rx_small.fill_cnt = ss->rx_small.cnt;
2090	for (i = ss->rx_small.cnt; i < ss->rx_small.fill_cnt; i++) {
2091		idx = i & ss->rx_small.mask;
2092		if (i == ss->rx_small.fill_cnt - 1)
2093			ss->rx_small.info[idx].page_offset =
2094			    MYRI10GE_ALLOC_SIZE;
2095		myri10ge_unmap_rx_page(mgp->pdev, &ss->rx_small.info[idx],
2096				       mgp->small_bytes + MXGEFW_PAD);
2097		put_page(ss->rx_small.info[idx].page);
2098	}
2099	tx = &ss->tx;
2100	while (tx->done != tx->req) {
2101		idx = tx->done & tx->mask;
2102		skb = tx->info[idx].skb;
2103
2104		/* Mark as free */
2105		tx->info[idx].skb = NULL;
2106		tx->done++;
2107		len = dma_unmap_len(&tx->info[idx], len);
2108		dma_unmap_len_set(&tx->info[idx], len, 0);
2109		if (skb) {
2110			ss->stats.tx_dropped++;
2111			dev_kfree_skb_any(skb);
2112			if (len)
2113				pci_unmap_single(mgp->pdev,
2114						 dma_unmap_addr(&tx->info[idx],
2115								bus), len,
2116						 PCI_DMA_TODEVICE);
2117		} else {
2118			if (len)
2119				pci_unmap_page(mgp->pdev,
2120					       dma_unmap_addr(&tx->info[idx],
2121							      bus), len,
2122					       PCI_DMA_TODEVICE);
2123		}
2124	}
2125	kfree(ss->rx_big.info);
2126
2127	kfree(ss->rx_small.info);
2128
2129	kfree(ss->tx.info);
2130
2131	kfree(ss->rx_big.shadow);
2132
2133	kfree(ss->rx_small.shadow);
2134
2135	kfree(ss->tx.req_bytes);
2136	ss->tx.req_bytes = NULL;
2137	ss->tx.req_list = NULL;
2138}
2139
2140static int myri10ge_request_irq(struct myri10ge_priv *mgp)
2141{
2142	struct pci_dev *pdev = mgp->pdev;
2143	struct myri10ge_slice_state *ss;
2144	struct net_device *netdev = mgp->dev;
2145	int i;
2146	int status;
2147
2148	mgp->msi_enabled = 0;
2149	mgp->msix_enabled = 0;
2150	status = 0;
2151	if (myri10ge_msi) {
2152		if (mgp->num_slices > 1) {
2153			status = pci_enable_msix_range(pdev, mgp->msix_vectors,
2154					mgp->num_slices, mgp->num_slices);
2155			if (status < 0) {
2156				dev_err(&pdev->dev,
2157					"Error %d setting up MSI-X\n", status);
2158				return status;
2159			}
2160			mgp->msix_enabled = 1;
2161		}
2162		if (mgp->msix_enabled == 0) {
2163			status = pci_enable_msi(pdev);
2164			if (status != 0) {
2165				dev_err(&pdev->dev,
2166					"Error %d setting up MSI; falling back to xPIC\n",
2167					status);
2168			} else {
2169				mgp->msi_enabled = 1;
2170			}
2171		}
2172	}
2173	if (mgp->msix_enabled) {
2174		for (i = 0; i < mgp->num_slices; i++) {
2175			ss = &mgp->ss[i];
2176			snprintf(ss->irq_desc, sizeof(ss->irq_desc),
2177				 "%s:slice-%d", netdev->name, i);
2178			status = request_irq(mgp->msix_vectors[i].vector,
2179					     myri10ge_intr, 0, ss->irq_desc,
2180					     ss);
2181			if (status != 0) {
2182				dev_err(&pdev->dev,
2183					"slice %d failed to allocate IRQ\n", i);
2184				i--;
2185				while (i >= 0) {
2186					free_irq(mgp->msix_vectors[i].vector,
2187						 &mgp->ss[i]);
2188					i--;
2189				}
2190				pci_disable_msix(pdev);
2191				return status;
2192			}
2193		}
2194	} else {
2195		status = request_irq(pdev->irq, myri10ge_intr, IRQF_SHARED,
2196				     mgp->dev->name, &mgp->ss[0]);
2197		if (status != 0) {
2198			dev_err(&pdev->dev, "failed to allocate IRQ\n");
2199			if (mgp->msi_enabled)
2200				pci_disable_msi(pdev);
2201		}
2202	}
2203	return status;
2204}
2205
2206static void myri10ge_free_irq(struct myri10ge_priv *mgp)
2207{
2208	struct pci_dev *pdev = mgp->pdev;
2209	int i;
2210
2211	if (mgp->msix_enabled) {
2212		for (i = 0; i < mgp->num_slices; i++)
2213			free_irq(mgp->msix_vectors[i].vector, &mgp->ss[i]);
2214	} else {
2215		free_irq(pdev->irq, &mgp->ss[0]);
2216	}
2217	if (mgp->msi_enabled)
2218		pci_disable_msi(pdev);
2219	if (mgp->msix_enabled)
2220		pci_disable_msix(pdev);
2221}
2222
2223static int myri10ge_get_txrx(struct myri10ge_priv *mgp, int slice)
2224{
2225	struct myri10ge_cmd cmd;
2226	struct myri10ge_slice_state *ss;
2227	int status;
2228
2229	ss = &mgp->ss[slice];
2230	status = 0;
2231	if (slice == 0 || (mgp->dev->real_num_tx_queues > 1)) {
2232		cmd.data0 = slice;
2233		status = myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_SEND_OFFSET,
2234					   &cmd, 0);
2235		ss->tx.lanai = (struct mcp_kreq_ether_send __iomem *)
2236		    (mgp->sram + cmd.data0);
2237	}
2238	cmd.data0 = slice;
2239	status |= myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_SMALL_RX_OFFSET,
2240				    &cmd, 0);
2241	ss->rx_small.lanai = (struct mcp_kreq_ether_recv __iomem *)
2242	    (mgp->sram + cmd.data0);
2243
2244	cmd.data0 = slice;
2245	status |= myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_BIG_RX_OFFSET, &cmd, 0);
2246	ss->rx_big.lanai = (struct mcp_kreq_ether_recv __iomem *)
2247	    (mgp->sram + cmd.data0);
2248
2249	ss->tx.send_go = (__iomem __be32 *)
2250	    (mgp->sram + MXGEFW_ETH_SEND_GO + 64 * slice);
2251	ss->tx.send_stop = (__iomem __be32 *)
2252	    (mgp->sram + MXGEFW_ETH_SEND_STOP + 64 * slice);
2253	return status;
2254
2255}
2256
2257static int myri10ge_set_stats(struct myri10ge_priv *mgp, int slice)
2258{
2259	struct myri10ge_cmd cmd;
2260	struct myri10ge_slice_state *ss;
2261	int status;
2262
2263	ss = &mgp->ss[slice];
2264	cmd.data0 = MYRI10GE_LOWPART_TO_U32(ss->fw_stats_bus);
2265	cmd.data1 = MYRI10GE_HIGHPART_TO_U32(ss->fw_stats_bus);
2266	cmd.data2 = sizeof(struct mcp_irq_data) | (slice << 16);
2267	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_STATS_DMA_V2, &cmd, 0);
2268	if (status == -ENOSYS) {
2269		dma_addr_t bus = ss->fw_stats_bus;
2270		if (slice != 0)
2271			return -EINVAL;
2272		bus += offsetof(struct mcp_irq_data, send_done_count);
2273		cmd.data0 = MYRI10GE_LOWPART_TO_U32(bus);
2274		cmd.data1 = MYRI10GE_HIGHPART_TO_U32(bus);
2275		status = myri10ge_send_cmd(mgp,
2276					   MXGEFW_CMD_SET_STATS_DMA_OBSOLETE,
2277					   &cmd, 0);
2278		/* Firmware cannot support multicast without STATS_DMA_V2 */
2279		mgp->fw_multicast_support = 0;
2280	} else {
2281		mgp->fw_multicast_support = 1;
2282	}
2283	return 0;
2284}
2285
2286static int myri10ge_open(struct net_device *dev)
2287{
2288	struct myri10ge_slice_state *ss;
2289	struct myri10ge_priv *mgp = netdev_priv(dev);
2290	struct myri10ge_cmd cmd;
2291	int i, status, big_pow2, slice;
2292	u8 __iomem *itable;
2293
2294	if (mgp->running != MYRI10GE_ETH_STOPPED)
2295		return -EBUSY;
2296
2297	mgp->running = MYRI10GE_ETH_STARTING;
2298	status = myri10ge_reset(mgp);
2299	if (status != 0) {
2300		netdev_err(dev, "failed reset\n");
2301		goto abort_with_nothing;
2302	}
2303
2304	if (mgp->num_slices > 1) {
2305		cmd.data0 = mgp->num_slices;
2306		cmd.data1 = MXGEFW_SLICE_INTR_MODE_ONE_PER_SLICE;
2307		if (mgp->dev->real_num_tx_queues > 1)
2308			cmd.data1 |= MXGEFW_SLICE_ENABLE_MULTIPLE_TX_QUEUES;
2309		status = myri10ge_send_cmd(mgp, MXGEFW_CMD_ENABLE_RSS_QUEUES,
2310					   &cmd, 0);
2311		if (status != 0) {
2312			netdev_err(dev, "failed to set number of slices\n");
2313			goto abort_with_nothing;
2314		}
2315		/* setup the indirection table */
2316		cmd.data0 = mgp->num_slices;
2317		status = myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_RSS_TABLE_SIZE,
2318					   &cmd, 0);
2319
2320		status |= myri10ge_send_cmd(mgp,
2321					    MXGEFW_CMD_GET_RSS_TABLE_OFFSET,
2322					    &cmd, 0);
2323		if (status != 0) {
2324			netdev_err(dev, "failed to setup rss tables\n");
2325			goto abort_with_nothing;
2326		}
2327
2328		/* just enable an identity mapping */
2329		itable = mgp->sram + cmd.data0;
2330		for (i = 0; i < mgp->num_slices; i++)
2331			__raw_writeb(i, &itable[i]);
2332
2333		cmd.data0 = 1;
2334		cmd.data1 = myri10ge_rss_hash;
2335		status = myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_RSS_ENABLE,
2336					   &cmd, 0);
2337		if (status != 0) {
2338			netdev_err(dev, "failed to enable slices\n");
2339			goto abort_with_nothing;
2340		}
2341	}
2342
2343	status = myri10ge_request_irq(mgp);
2344	if (status != 0)
2345		goto abort_with_nothing;
2346
2347	/* decide what small buffer size to use.  For good TCP rx
2348	 * performance, it is important to not receive 1514 byte
2349	 * frames into jumbo buffers, as it confuses the socket buffer
2350	 * accounting code, leading to drops and erratic performance.
2351	 */
2352
2353	if (dev->mtu <= ETH_DATA_LEN)
2354		/* enough for a TCP header */
2355		mgp->small_bytes = (128 > SMP_CACHE_BYTES)
2356		    ? (128 - MXGEFW_PAD)
2357		    : (SMP_CACHE_BYTES - MXGEFW_PAD);
2358	else
2359		/* enough for a vlan encapsulated ETH_DATA_LEN frame */
2360		mgp->small_bytes = VLAN_ETH_FRAME_LEN;
2361
2362	/* Override the small buffer size? */
2363	if (myri10ge_small_bytes >= 0)
2364		mgp->small_bytes = myri10ge_small_bytes;
2365
2366	/* Firmware needs the big buff size as a power of 2.  Lie and
2367	 * tell him the buffer is larger, because we only use 1
2368	 * buffer/pkt, and the mtu will prevent overruns.
2369	 */
2370	big_pow2 = dev->mtu + ETH_HLEN + VLAN_HLEN + MXGEFW_PAD;
2371	if (big_pow2 < MYRI10GE_ALLOC_SIZE / 2) {
2372		while (!is_power_of_2(big_pow2))
2373			big_pow2++;
2374		mgp->big_bytes = dev->mtu + ETH_HLEN + VLAN_HLEN + MXGEFW_PAD;
2375	} else {
2376		big_pow2 = MYRI10GE_ALLOC_SIZE;
2377		mgp->big_bytes = big_pow2;
2378	}
2379
2380	/* setup the per-slice data structures */
2381	for (slice = 0; slice < mgp->num_slices; slice++) {
2382		ss = &mgp->ss[slice];
2383
2384		status = myri10ge_get_txrx(mgp, slice);
2385		if (status != 0) {
2386			netdev_err(dev, "failed to get ring sizes or locations\n");
2387			goto abort_with_rings;
2388		}
2389		status = myri10ge_allocate_rings(ss);
2390		if (status != 0)
2391			goto abort_with_rings;
2392
2393		/* only firmware which supports multiple TX queues
2394		 * supports setting up the tx stats on non-zero
2395		 * slices */
2396		if (slice == 0 || mgp->dev->real_num_tx_queues > 1)
2397			status = myri10ge_set_stats(mgp, slice);
2398		if (status) {
2399			netdev_err(dev, "Couldn't set stats DMA\n");
2400			goto abort_with_rings;
2401		}
2402
2403		/* must happen prior to any irq */
2404		napi_enable(&(ss)->napi);
2405	}
2406
2407	/* now give firmware buffers sizes, and MTU */
2408	cmd.data0 = dev->mtu + ETH_HLEN + VLAN_HLEN;
2409	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_MTU, &cmd, 0);
2410	cmd.data0 = mgp->small_bytes;
2411	status |=
2412	    myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_SMALL_BUFFER_SIZE, &cmd, 0);
2413	cmd.data0 = big_pow2;
2414	status |=
2415	    myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_BIG_BUFFER_SIZE, &cmd, 0);
2416	if (status) {
2417		netdev_err(dev, "Couldn't set buffer sizes\n");
2418		goto abort_with_rings;
2419	}
2420
2421	/*
2422	 * Set Linux style TSO mode; this is needed only on newer
2423	 *  firmware versions.  Older versions default to Linux
2424	 *  style TSO
2425	 */
2426	cmd.data0 = 0;
2427	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_TSO_MODE, &cmd, 0);
2428	if (status && status != -ENOSYS) {
2429		netdev_err(dev, "Couldn't set TSO mode\n");
2430		goto abort_with_rings;
2431	}
2432
2433	mgp->link_state = ~0U;
2434	mgp->rdma_tags_available = 15;
2435
2436	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_ETHERNET_UP, &cmd, 0);
2437	if (status) {
2438		netdev_err(dev, "Couldn't bring up link\n");
2439		goto abort_with_rings;
2440	}
2441
2442	mgp->running = MYRI10GE_ETH_RUNNING;
2443	mgp->watchdog_timer.expires = jiffies + myri10ge_watchdog_timeout * HZ;
2444	add_timer(&mgp->watchdog_timer);
2445	netif_tx_wake_all_queues(dev);
2446
2447	return 0;
2448
2449abort_with_rings:
2450	while (slice) {
2451		slice--;
2452		napi_disable(&mgp->ss[slice].napi);
2453	}
2454	for (i = 0; i < mgp->num_slices; i++)
2455		myri10ge_free_rings(&mgp->ss[i]);
2456
2457	myri10ge_free_irq(mgp);
2458
2459abort_with_nothing:
2460	mgp->running = MYRI10GE_ETH_STOPPED;
2461	return -ENOMEM;
2462}
2463
2464static int myri10ge_close(struct net_device *dev)
2465{
2466	struct myri10ge_priv *mgp = netdev_priv(dev);
2467	struct myri10ge_cmd cmd;
2468	int status, old_down_cnt;
2469	int i;
2470
2471	if (mgp->running != MYRI10GE_ETH_RUNNING)
2472		return 0;
2473
2474	if (mgp->ss[0].tx.req_bytes == NULL)
2475		return 0;
2476
2477	del_timer_sync(&mgp->watchdog_timer);
2478	mgp->running = MYRI10GE_ETH_STOPPING;
2479	for (i = 0; i < mgp->num_slices; i++)
2480		napi_disable(&mgp->ss[i].napi);
2481
2482	netif_carrier_off(dev);
2483
2484	netif_tx_stop_all_queues(dev);
2485	if (mgp->rebooted == 0) {
2486		old_down_cnt = mgp->down_cnt;
2487		mb();
2488		status =
2489		    myri10ge_send_cmd(mgp, MXGEFW_CMD_ETHERNET_DOWN, &cmd, 0);
2490		if (status)
2491			netdev_err(dev, "Couldn't bring down link\n");
2492
2493		wait_event_timeout(mgp->down_wq, old_down_cnt != mgp->down_cnt,
2494				   HZ);
2495		if (old_down_cnt == mgp->down_cnt)
2496			netdev_err(dev, "never got down irq\n");
2497	}
2498	netif_tx_disable(dev);
2499	myri10ge_free_irq(mgp);
2500	for (i = 0; i < mgp->num_slices; i++)
2501		myri10ge_free_rings(&mgp->ss[i]);
2502
2503	mgp->running = MYRI10GE_ETH_STOPPED;
2504	return 0;
2505}
2506
2507/* copy an array of struct mcp_kreq_ether_send's to the mcp.  Copy
2508 * backwards one at a time and handle ring wraps */
2509
2510static inline void
2511myri10ge_submit_req_backwards(struct myri10ge_tx_buf *tx,
2512			      struct mcp_kreq_ether_send *src, int cnt)
2513{
2514	int idx, starting_slot;
2515	starting_slot = tx->req;
2516	while (cnt > 1) {
2517		cnt--;
2518		idx = (starting_slot + cnt) & tx->mask;
2519		myri10ge_pio_copy(&tx->lanai[idx], &src[cnt], sizeof(*src));
2520		mb();
2521	}
2522}
2523
2524/*
2525 * copy an array of struct mcp_kreq_ether_send's to the mcp.  Copy
2526 * at most 32 bytes at a time, so as to avoid involving the software
2527 * pio handler in the nic.   We re-write the first segment's flags
2528 * to mark them valid only after writing the entire chain.
2529 */
2530
2531static inline void
2532myri10ge_submit_req(struct myri10ge_tx_buf *tx, struct mcp_kreq_ether_send *src,
2533		    int cnt)
2534{
2535	int idx, i;
2536	struct mcp_kreq_ether_send __iomem *dstp, *dst;
2537	struct mcp_kreq_ether_send *srcp;
2538	u8 last_flags;
2539
2540	idx = tx->req & tx->mask;
2541
2542	last_flags = src->flags;
2543	src->flags = 0;
2544	mb();
2545	dst = dstp = &tx->lanai[idx];
2546	srcp = src;
2547
2548	if ((idx + cnt) < tx->mask) {
2549		for (i = 0; i < (cnt - 1); i += 2) {
2550			myri10ge_pio_copy(dstp, srcp, 2 * sizeof(*src));
2551			mb();	/* force write every 32 bytes */
2552			srcp += 2;
2553			dstp += 2;
2554		}
2555	} else {
2556		/* submit all but the first request, and ensure
2557		 * that it is submitted below */
2558		myri10ge_submit_req_backwards(tx, src, cnt);
2559		i = 0;
2560	}
2561	if (i < cnt) {
2562		/* submit the first request */
2563		myri10ge_pio_copy(dstp, srcp, sizeof(*src));
2564		mb();		/* barrier before setting valid flag */
2565	}
2566
2567	/* re-write the last 32-bits with the valid flags */
2568	src->flags = last_flags;
2569	put_be32(*((__be32 *) src + 3), (__be32 __iomem *) dst + 3);
2570	tx->req += cnt;
2571	mb();
2572}
2573
2574static void myri10ge_unmap_tx_dma(struct myri10ge_priv *mgp,
2575				  struct myri10ge_tx_buf *tx, int idx)
2576{
2577	unsigned int len;
2578	int last_idx;
2579
2580	/* Free any DMA resources we've alloced and clear out the skb slot */
2581	last_idx = (idx + 1) & tx->mask;
2582	idx = tx->req & tx->mask;
2583	do {
2584		len = dma_unmap_len(&tx->info[idx], len);
2585		if (len) {
2586			if (tx->info[idx].skb != NULL)
2587				pci_unmap_single(mgp->pdev,
2588						 dma_unmap_addr(&tx->info[idx],
2589								bus), len,
2590						 PCI_DMA_TODEVICE);
2591			else
2592				pci_unmap_page(mgp->pdev,
2593					       dma_unmap_addr(&tx->info[idx],
2594							      bus), len,
2595					       PCI_DMA_TODEVICE);
2596			dma_unmap_len_set(&tx->info[idx], len, 0);
2597			tx->info[idx].skb = NULL;
2598		}
2599		idx = (idx + 1) & tx->mask;
2600	} while (idx != last_idx);
2601}
2602
2603/*
2604 * Transmit a packet.  We need to split the packet so that a single
2605 * segment does not cross myri10ge->tx_boundary, so this makes segment
2606 * counting tricky.  So rather than try to count segments up front, we
2607 * just give up if there are too few segments to hold a reasonably
2608 * fragmented packet currently available.  If we run
2609 * out of segments while preparing a packet for DMA, we just linearize
2610 * it and try again.
2611 */
2612
2613static netdev_tx_t myri10ge_xmit(struct sk_buff *skb,
2614				       struct net_device *dev)
2615{
2616	struct myri10ge_priv *mgp = netdev_priv(dev);
2617	struct myri10ge_slice_state *ss;
2618	struct mcp_kreq_ether_send *req;
2619	struct myri10ge_tx_buf *tx;
2620	skb_frag_t *frag;
2621	struct netdev_queue *netdev_queue;
2622	dma_addr_t bus;
2623	u32 low;
2624	__be32 high_swapped;
2625	unsigned int len;
2626	int idx, avail, frag_cnt, frag_idx, count, mss, max_segments;
2627	u16 pseudo_hdr_offset, cksum_offset, queue;
2628	int cum_len, seglen, boundary, rdma_count;
2629	u8 flags, odd_flag;
2630
2631	queue = skb_get_queue_mapping(skb);
2632	ss = &mgp->ss[queue];
2633	netdev_queue = netdev_get_tx_queue(mgp->dev, queue);
2634	tx = &ss->tx;
2635
2636again:
2637	req = tx->req_list;
2638	avail = tx->mask - 1 - (tx->req - tx->done);
2639
2640	mss = 0;
2641	max_segments = MXGEFW_MAX_SEND_DESC;
2642
2643	if (skb_is_gso(skb)) {
2644		mss = skb_shinfo(skb)->gso_size;
2645		max_segments = MYRI10GE_MAX_SEND_DESC_TSO;
2646	}
2647
2648	if ((unlikely(avail < max_segments))) {
2649		/* we are out of transmit resources */
2650		tx->stop_queue++;
2651		netif_tx_stop_queue(netdev_queue);
2652		return NETDEV_TX_BUSY;
2653	}
2654
2655	/* Setup checksum offloading, if needed */
2656	cksum_offset = 0;
2657	pseudo_hdr_offset = 0;
2658	odd_flag = 0;
2659	flags = (MXGEFW_FLAGS_NO_TSO | MXGEFW_FLAGS_FIRST);
2660	if (likely(skb->ip_summed == CHECKSUM_PARTIAL)) {
2661		cksum_offset = skb_checksum_start_offset(skb);
2662		pseudo_hdr_offset = cksum_offset + skb->csum_offset;
2663		/* If the headers are excessively large, then we must
2664		 * fall back to a software checksum */
2665		if (unlikely(!mss && (cksum_offset > 255 ||
2666				      pseudo_hdr_offset > 127))) {
2667			if (skb_checksum_help(skb))
2668				goto drop;
2669			cksum_offset = 0;
2670			pseudo_hdr_offset = 0;
2671		} else {
2672			odd_flag = MXGEFW_FLAGS_ALIGN_ODD;
2673			flags |= MXGEFW_FLAGS_CKSUM;
2674		}
2675	}
2676
2677	cum_len = 0;
2678
2679	if (mss) {		/* TSO */
2680		/* this removes any CKSUM flag from before */
2681		flags = (MXGEFW_FLAGS_TSO_HDR | MXGEFW_FLAGS_FIRST);
2682
2683		/* negative cum_len signifies to the
2684		 * send loop that we are still in the
2685		 * header portion of the TSO packet.
2686		 * TSO header can be at most 1KB long */
2687		cum_len = -(skb_transport_offset(skb) + tcp_hdrlen(skb));
2688
2689		/* for IPv6 TSO, the checksum offset stores the
2690		 * TCP header length, to save the firmware from
2691		 * the need to parse the headers */
2692		if (skb_is_gso_v6(skb)) {
2693			cksum_offset = tcp_hdrlen(skb);
2694			/* Can only handle headers <= max_tso6 long */
2695			if (unlikely(-cum_len > mgp->max_tso6))
2696				return myri10ge_sw_tso(skb, dev);
2697		}
2698		/* for TSO, pseudo_hdr_offset holds mss.
2699		 * The firmware figures out where to put
2700		 * the checksum by parsing the header. */
2701		pseudo_hdr_offset = mss;
2702	} else
2703		/* Mark small packets, and pad out tiny packets */
2704	if (skb->len <= MXGEFW_SEND_SMALL_SIZE) {
2705		flags |= MXGEFW_FLAGS_SMALL;
2706
2707		/* pad frames to at least ETH_ZLEN bytes */
2708		if (eth_skb_pad(skb)) {
2709			/* The packet is gone, so we must
2710			 * return 0 */
2711			ss->stats.tx_dropped += 1;
2712			return NETDEV_TX_OK;
2713		}
2714	}
2715
2716	/* map the skb for DMA */
2717	len = skb_headlen(skb);
2718	bus = pci_map_single(mgp->pdev, skb->data, len, PCI_DMA_TODEVICE);
2719	if (unlikely(pci_dma_mapping_error(mgp->pdev, bus)))
2720		goto drop;
2721
2722	idx = tx->req & tx->mask;
2723	tx->info[idx].skb = skb;
2724	dma_unmap_addr_set(&tx->info[idx], bus, bus);
2725	dma_unmap_len_set(&tx->info[idx], len, len);
2726
2727	frag_cnt = skb_shinfo(skb)->nr_frags;
2728	frag_idx = 0;
2729	count = 0;
2730	rdma_count = 0;
2731
2732	/* "rdma_count" is the number of RDMAs belonging to the
2733	 * current packet BEFORE the current send request. For
2734	 * non-TSO packets, this is equal to "count".
2735	 * For TSO packets, rdma_count needs to be reset
2736	 * to 0 after a segment cut.
2737	 *
2738	 * The rdma_count field of the send request is
2739	 * the number of RDMAs of the packet starting at
2740	 * that request. For TSO send requests with one ore more cuts
2741	 * in the middle, this is the number of RDMAs starting
2742	 * after the last cut in the request. All previous
2743	 * segments before the last cut implicitly have 1 RDMA.
2744	 *
2745	 * Since the number of RDMAs is not known beforehand,
2746	 * it must be filled-in retroactively - after each
2747	 * segmentation cut or at the end of the entire packet.
2748	 */
2749
2750	while (1) {
2751		/* Break the SKB or Fragment up into pieces which
2752		 * do not cross mgp->tx_boundary */
2753		low = MYRI10GE_LOWPART_TO_U32(bus);
2754		high_swapped = htonl(MYRI10GE_HIGHPART_TO_U32(bus));
2755		while (len) {
2756			u8 flags_next;
2757			int cum_len_next;
2758
2759			if (unlikely(count == max_segments))
2760				goto abort_linearize;
2761
2762			boundary =
2763			    (low + mgp->tx_boundary) & ~(mgp->tx_boundary - 1);
2764			seglen = boundary - low;
2765			if (seglen > len)
2766				seglen = len;
2767			flags_next = flags & ~MXGEFW_FLAGS_FIRST;
2768			cum_len_next = cum_len + seglen;
2769			if (mss) {	/* TSO */
2770				(req - rdma_count)->rdma_count = rdma_count + 1;
2771
2772				if (likely(cum_len >= 0)) {	/* payload */
2773					int next_is_first, chop;
2774
2775					chop = (cum_len_next > mss);
2776					cum_len_next = cum_len_next % mss;
2777					next_is_first = (cum_len_next == 0);
2778					flags |= chop * MXGEFW_FLAGS_TSO_CHOP;
2779					flags_next |= next_is_first *
2780					    MXGEFW_FLAGS_FIRST;
2781					rdma_count |= -(chop | next_is_first);
2782					rdma_count += chop & ~next_is_first;
2783				} else if (likely(cum_len_next >= 0)) {	/* header ends */
2784					int small;
2785
2786					rdma_count = -1;
2787					cum_len_next = 0;
2788					seglen = -cum_len;
2789					small = (mss <= MXGEFW_SEND_SMALL_SIZE);
2790					flags_next = MXGEFW_FLAGS_TSO_PLD |
2791					    MXGEFW_FLAGS_FIRST |
2792					    (small * MXGEFW_FLAGS_SMALL);
2793				}
2794			}
2795			req->addr_high = high_swapped;
2796			req->addr_low = htonl(low);
2797			req->pseudo_hdr_offset = htons(pseudo_hdr_offset);
2798			req->pad = 0;	/* complete solid 16-byte block; does this matter? */
2799			req->rdma_count = 1;
2800			req->length = htons(seglen);
2801			req->cksum_offset = cksum_offset;
2802			req->flags = flags | ((cum_len & 1) * odd_flag);
2803
2804			low += seglen;
2805			len -= seglen;
2806			cum_len = cum_len_next;
2807			flags = flags_next;
2808			req++;
2809			count++;
2810			rdma_count++;
2811			if (cksum_offset != 0 && !(mss && skb_is_gso_v6(skb))) {
2812				if (unlikely(cksum_offset > seglen))
2813					cksum_offset -= seglen;
2814				else
2815					cksum_offset = 0;
2816			}
2817		}
2818		if (frag_idx == frag_cnt)
2819			break;
2820
2821		/* map next fragment for DMA */
2822		frag = &skb_shinfo(skb)->frags[frag_idx];
2823		frag_idx++;
2824		len = skb_frag_size(frag);
2825		bus = skb_frag_dma_map(&mgp->pdev->dev, frag, 0, len,
2826				       DMA_TO_DEVICE);
2827		if (unlikely(pci_dma_mapping_error(mgp->pdev, bus))) {
2828			myri10ge_unmap_tx_dma(mgp, tx, idx);
2829			goto drop;
2830		}
2831		idx = (count + tx->req) & tx->mask;
2832		dma_unmap_addr_set(&tx->info[idx], bus, bus);
2833		dma_unmap_len_set(&tx->info[idx], len, len);
2834	}
2835
2836	(req - rdma_count)->rdma_count = rdma_count;
2837	if (mss)
2838		do {
2839			req--;
2840			req->flags |= MXGEFW_FLAGS_TSO_LAST;
2841		} while (!(req->flags & (MXGEFW_FLAGS_TSO_CHOP |
2842					 MXGEFW_FLAGS_FIRST)));
2843	idx = ((count - 1) + tx->req) & tx->mask;
2844	tx->info[idx].last = 1;
2845	myri10ge_submit_req(tx, tx->req_list, count);
2846	/* if using multiple tx queues, make sure NIC polls the
2847	 * current slice */
2848	if ((mgp->dev->real_num_tx_queues > 1) && tx->queue_active == 0) {
2849		tx->queue_active = 1;
2850		put_be32(htonl(1), tx->send_go);
2851		mb();
2852	}
2853	tx->pkt_start++;
2854	if ((avail - count) < MXGEFW_MAX_SEND_DESC) {
2855		tx->stop_queue++;
2856		netif_tx_stop_queue(netdev_queue);
2857	}
2858	return NETDEV_TX_OK;
2859
2860abort_linearize:
2861	myri10ge_unmap_tx_dma(mgp, tx, idx);
2862
2863	if (skb_is_gso(skb)) {
2864		netdev_err(mgp->dev, "TSO but wanted to linearize?!?!?\n");
2865		goto drop;
2866	}
2867
2868	if (skb_linearize(skb))
2869		goto drop;
2870
2871	tx->linearized++;
2872	goto again;
2873
2874drop:
2875	dev_kfree_skb_any(skb);
2876	ss->stats.tx_dropped += 1;
2877	return NETDEV_TX_OK;
2878
2879}
2880
2881static netdev_tx_t myri10ge_sw_tso(struct sk_buff *skb,
2882					 struct net_device *dev)
2883{
2884	struct sk_buff *segs, *curr, *next;
2885	struct myri10ge_priv *mgp = netdev_priv(dev);
2886	struct myri10ge_slice_state *ss;
2887	netdev_tx_t status;
2888
2889	segs = skb_gso_segment(skb, dev->features & ~NETIF_F_TSO6);
2890	if (IS_ERR(segs))
2891		goto drop;
2892
2893	skb_list_walk_safe(segs, curr, next) {
2894		skb_mark_not_on_list(curr);
2895		status = myri10ge_xmit(curr, dev);
2896		if (status != 0) {
2897			dev_kfree_skb_any(curr);
2898			skb_list_walk_safe(next, curr, next) {
2899				curr->next = NULL;
2900				dev_kfree_skb_any(curr);
2901			}
2902			goto drop;
2903		}
2904	}
2905	dev_kfree_skb_any(skb);
2906	return NETDEV_TX_OK;
2907
2908drop:
2909	ss = &mgp->ss[skb_get_queue_mapping(skb)];
2910	dev_kfree_skb_any(skb);
2911	ss->stats.tx_dropped += 1;
2912	return NETDEV_TX_OK;
2913}
2914
2915static void myri10ge_get_stats(struct net_device *dev,
2916			       struct rtnl_link_stats64 *stats)
2917{
2918	const struct myri10ge_priv *mgp = netdev_priv(dev);
2919	const struct myri10ge_slice_netstats *slice_stats;
2920	int i;
2921
2922	for (i = 0; i < mgp->num_slices; i++) {
2923		slice_stats = &mgp->ss[i].stats;
2924		stats->rx_packets += slice_stats->rx_packets;
2925		stats->tx_packets += slice_stats->tx_packets;
2926		stats->rx_bytes += slice_stats->rx_bytes;
2927		stats->tx_bytes += slice_stats->tx_bytes;
2928		stats->rx_dropped += slice_stats->rx_dropped;
2929		stats->tx_dropped += slice_stats->tx_dropped;
2930	}
2931}
2932
2933static void myri10ge_set_multicast_list(struct net_device *dev)
2934{
2935	struct myri10ge_priv *mgp = netdev_priv(dev);
2936	struct myri10ge_cmd cmd;
2937	struct netdev_hw_addr *ha;
2938	__be32 data[2] = { 0, 0 };
2939	int err;
2940
2941	/* can be called from atomic contexts,
2942	 * pass 1 to force atomicity in myri10ge_send_cmd() */
2943	myri10ge_change_promisc(mgp, dev->flags & IFF_PROMISC, 1);
2944
2945	/* This firmware is known to not support multicast */
2946	if (!mgp->fw_multicast_support)
2947		return;
2948
2949	/* Disable multicast filtering */
2950
2951	err = myri10ge_send_cmd(mgp, MXGEFW_ENABLE_ALLMULTI, &cmd, 1);
2952	if (err != 0) {
2953		netdev_err(dev, "Failed MXGEFW_ENABLE_ALLMULTI, error status: %d\n",
2954			   err);
2955		goto abort;
2956	}
2957
2958	if ((dev->flags & IFF_ALLMULTI) || mgp->adopted_rx_filter_bug) {
2959		/* request to disable multicast filtering, so quit here */
2960		return;
2961	}
2962
2963	/* Flush the filters */
2964
2965	err = myri10ge_send_cmd(mgp, MXGEFW_LEAVE_ALL_MULTICAST_GROUPS,
2966				&cmd, 1);
2967	if (err != 0) {
2968		netdev_err(dev, "Failed MXGEFW_LEAVE_ALL_MULTICAST_GROUPS, error status: %d\n",
2969			   err);
2970		goto abort;
2971	}
2972
2973	/* Walk the multicast list, and add each address */
2974	netdev_for_each_mc_addr(ha, dev) {
2975		memcpy(data, &ha->addr, ETH_ALEN);
2976		cmd.data0 = ntohl(data[0]);
2977		cmd.data1 = ntohl(data[1]);
2978		err = myri10ge_send_cmd(mgp, MXGEFW_JOIN_MULTICAST_GROUP,
2979					&cmd, 1);
2980
2981		if (err != 0) {
2982			netdev_err(dev, "Failed MXGEFW_JOIN_MULTICAST_GROUP, error status:%d %pM\n",
2983				   err, ha->addr);
2984			goto abort;
2985		}
2986	}
2987	/* Enable multicast filtering */
2988	err = myri10ge_send_cmd(mgp, MXGEFW_DISABLE_ALLMULTI, &cmd, 1);
2989	if (err != 0) {
2990		netdev_err(dev, "Failed MXGEFW_DISABLE_ALLMULTI, error status: %d\n",
2991			   err);
2992		goto abort;
2993	}
2994
2995	return;
2996
2997abort:
2998	return;
2999}
3000
3001static int myri10ge_set_mac_address(struct net_device *dev, void *addr)
3002{
3003	struct sockaddr *sa = addr;
3004	struct myri10ge_priv *mgp = netdev_priv(dev);
3005	int status;
3006
3007	if (!is_valid_ether_addr(sa->sa_data))
3008		return -EADDRNOTAVAIL;
3009
3010	status = myri10ge_update_mac_address(mgp, sa->sa_data);
3011	if (status != 0) {
3012		netdev_err(dev, "changing mac address failed with %d\n",
3013			   status);
3014		return status;
3015	}
3016
3017	/* change the dev structure */
3018	memcpy(dev->dev_addr, sa->sa_data, ETH_ALEN);
3019	return 0;
3020}
3021
3022static int myri10ge_change_mtu(struct net_device *dev, int new_mtu)
3023{
3024	struct myri10ge_priv *mgp = netdev_priv(dev);
3025
3026	netdev_info(dev, "changing mtu from %d to %d\n", dev->mtu, new_mtu);
3027	if (mgp->running) {
3028		/* if we change the mtu on an active device, we must
3029		 * reset the device so the firmware sees the change */
3030		myri10ge_close(dev);
3031		dev->mtu = new_mtu;
3032		myri10ge_open(dev);
3033	} else
3034		dev->mtu = new_mtu;
3035
3036	return 0;
3037}
3038
3039/*
3040 * Enable ECRC to align PCI-E Completion packets on an 8-byte boundary.
3041 * Only do it if the bridge is a root port since we don't want to disturb
3042 * any other device, except if forced with myri10ge_ecrc_enable > 1.
3043 */
3044
3045static void myri10ge_enable_ecrc(struct myri10ge_priv *mgp)
3046{
3047	struct pci_dev *bridge = mgp->pdev->bus->self;
3048	struct device *dev = &mgp->pdev->dev;
3049	int cap;
3050	unsigned err_cap;
3051	int ret;
3052
3053	if (!myri10ge_ecrc_enable || !bridge)
3054		return;
3055
3056	/* check that the bridge is a root port */
3057	if (pci_pcie_type(bridge) != PCI_EXP_TYPE_ROOT_PORT) {
3058		if (myri10ge_ecrc_enable > 1) {
3059			struct pci_dev *prev_bridge, *old_bridge = bridge;
3060
3061			/* Walk the hierarchy up to the root port
3062			 * where ECRC has to be enabled */
3063			do {
3064				prev_bridge = bridge;
3065				bridge = bridge->bus->self;
3066				if (!bridge || prev_bridge == bridge) {
3067					dev_err(dev,
3068						"Failed to find root port"
3069						" to force ECRC\n");
3070					return;
3071				}
3072			} while (pci_pcie_type(bridge) !=
3073				 PCI_EXP_TYPE_ROOT_PORT);
3074
3075			dev_info(dev,
3076				 "Forcing ECRC on non-root port %s"
3077				 " (enabling on root port %s)\n",
3078				 pci_name(old_bridge), pci_name(bridge));
3079		} else {
3080			dev_err(dev,
3081				"Not enabling ECRC on non-root port %s\n",
3082				pci_name(bridge));
3083			return;
3084		}
3085	}
3086
3087	cap = pci_find_ext_capability(bridge, PCI_EXT_CAP_ID_ERR);
3088	if (!cap)
3089		return;
3090
3091	ret = pci_read_config_dword(bridge, cap + PCI_ERR_CAP, &err_cap);
3092	if (ret) {
3093		dev_err(dev, "failed reading ext-conf-space of %s\n",
3094			pci_name(bridge));
3095		dev_err(dev, "\t pci=nommconf in use? "
3096			"or buggy/incomplete/absent ACPI MCFG attr?\n");
3097		return;
3098	}
3099	if (!(err_cap & PCI_ERR_CAP_ECRC_GENC))
3100		return;
3101
3102	err_cap |= PCI_ERR_CAP_ECRC_GENE;
3103	pci_write_config_dword(bridge, cap + PCI_ERR_CAP, err_cap);
3104	dev_info(dev, "Enabled ECRC on upstream bridge %s\n", pci_name(bridge));
3105}
3106
3107/*
3108 * The Lanai Z8E PCI-E interface achieves higher Read-DMA throughput
3109 * when the PCI-E Completion packets are aligned on an 8-byte
3110 * boundary.  Some PCI-E chip sets always align Completion packets; on
3111 * the ones that do not, the alignment can be enforced by enabling
3112 * ECRC generation (if supported).
3113 *
3114 * When PCI-E Completion packets are not aligned, it is actually more
3115 * efficient to limit Read-DMA transactions to 2KB, rather than 4KB.
3116 *
3117 * If the driver can neither enable ECRC nor verify that it has
3118 * already been enabled, then it must use a firmware image which works
3119 * around unaligned completion packets (myri10ge_rss_ethp_z8e.dat), and it
3120 * should also ensure that it never gives the device a Read-DMA which is
3121 * larger than 2KB by setting the tx_boundary to 2KB.  If ECRC is
3122 * enabled, then the driver should use the aligned (myri10ge_rss_eth_z8e.dat)
3123 * firmware image, and set tx_boundary to 4KB.
3124 */
3125
3126static void myri10ge_firmware_probe(struct myri10ge_priv *mgp)
3127{
3128	struct pci_dev *pdev = mgp->pdev;
3129	struct device *dev = &pdev->dev;
3130	int status;
3131
3132	mgp->tx_boundary = 4096;
3133	/*
3134	 * Verify the max read request size was set to 4KB
3135	 * before trying the test with 4KB.
3136	 */
3137	status = pcie_get_readrq(pdev);
3138	if (status < 0) {
3139		dev_err(dev, "Couldn't read max read req size: %d\n", status);
3140		goto abort;
3141	}
3142	if (status != 4096) {
3143		dev_warn(dev, "Max Read Request size != 4096 (%d)\n", status);
3144		mgp->tx_boundary = 2048;
3145	}
3146	/*
3147	 * load the optimized firmware (which assumes aligned PCIe
3148	 * completions) in order to see if it works on this host.
3149	 */
3150	set_fw_name(mgp, myri10ge_fw_aligned, false);
3151	status = myri10ge_load_firmware(mgp, 1);
3152	if (status != 0) {
3153		goto abort;
3154	}
3155
3156	/*
3157	 * Enable ECRC if possible
3158	 */
3159	myri10ge_enable_ecrc(mgp);
3160
3161	/*
3162	 * Run a DMA test which watches for unaligned completions and
3163	 * aborts on the first one seen.
3164	 */
3165
3166	status = myri10ge_dma_test(mgp, MXGEFW_CMD_UNALIGNED_TEST);
3167	if (status == 0)
3168		return;		/* keep the aligned firmware */
3169
3170	if (status != -E2BIG)
3171		dev_warn(dev, "DMA test failed: %d\n", status);
3172	if (status == -ENOSYS)
3173		dev_warn(dev, "Falling back to ethp! "
3174			 "Please install up to date fw\n");
3175abort:
3176	/* fall back to using the unaligned firmware */
3177	mgp->tx_boundary = 2048;
3178	set_fw_name(mgp, myri10ge_fw_unaligned, false);
3179}
3180
3181static void myri10ge_select_firmware(struct myri10ge_priv *mgp)
3182{
3183	int overridden = 0;
3184
3185	if (myri10ge_force_firmware == 0) {
3186		int link_width;
3187		u16 lnk;
3188
3189		pcie_capability_read_word(mgp->pdev, PCI_EXP_LNKSTA, &lnk);
3190		link_width = (lnk >> 4) & 0x3f;
3191
3192		/* Check to see if Link is less than 8 or if the
3193		 * upstream bridge is known to provide aligned
3194		 * completions */
3195		if (link_width < 8) {
3196			dev_info(&mgp->pdev->dev, "PCIE x%d Link\n",
3197				 link_width);
3198			mgp->tx_boundary = 4096;
3199			set_fw_name(mgp, myri10ge_fw_aligned, false);
3200		} else {
3201			myri10ge_firmware_probe(mgp);
3202		}
3203	} else {
3204		if (myri10ge_force_firmware == 1) {
3205			dev_info(&mgp->pdev->dev,
3206				 "Assuming aligned completions (forced)\n");
3207			mgp->tx_boundary = 4096;
3208			set_fw_name(mgp, myri10ge_fw_aligned, false);
3209		} else {
3210			dev_info(&mgp->pdev->dev,
3211				 "Assuming unaligned completions (forced)\n");
3212			mgp->tx_boundary = 2048;
3213			set_fw_name(mgp, myri10ge_fw_unaligned, false);
3214		}
3215	}
3216
3217	kernel_param_lock(THIS_MODULE);
3218	if (myri10ge_fw_name != NULL) {
3219		char *fw_name = kstrdup(myri10ge_fw_name, GFP_KERNEL);
3220		if (fw_name) {
3221			overridden = 1;
3222			set_fw_name(mgp, fw_name, true);
3223		}
3224	}
3225	kernel_param_unlock(THIS_MODULE);
3226
3227	if (mgp->board_number < MYRI10GE_MAX_BOARDS &&
3228	    myri10ge_fw_names[mgp->board_number] != NULL &&
3229	    strlen(myri10ge_fw_names[mgp->board_number])) {
3230		set_fw_name(mgp, myri10ge_fw_names[mgp->board_number], false);
3231		overridden = 1;
3232	}
3233	if (overridden)
3234		dev_info(&mgp->pdev->dev, "overriding firmware to %s\n",
3235			 mgp->fw_name);
3236}
3237
3238static void myri10ge_mask_surprise_down(struct pci_dev *pdev)
3239{
3240	struct pci_dev *bridge = pdev->bus->self;
3241	int cap;
3242	u32 mask;
3243
3244	if (bridge == NULL)
3245		return;
3246
3247	cap = pci_find_ext_capability(bridge, PCI_EXT_CAP_ID_ERR);
3248	if (cap) {
3249		/* a sram parity error can cause a surprise link
3250		 * down; since we expect and can recover from sram
3251		 * parity errors, mask surprise link down events */
3252		pci_read_config_dword(bridge, cap + PCI_ERR_UNCOR_MASK, &mask);
3253		mask |= 0x20;
3254		pci_write_config_dword(bridge, cap + PCI_ERR_UNCOR_MASK, mask);
3255	}
3256}
3257
3258static int __maybe_unused myri10ge_suspend(struct device *dev)
3259{
3260	struct myri10ge_priv *mgp;
3261	struct net_device *netdev;
3262
3263	mgp = dev_get_drvdata(dev);
3264	if (mgp == NULL)
3265		return -EINVAL;
3266	netdev = mgp->dev;
3267
3268	netif_device_detach(netdev);
3269	if (netif_running(netdev)) {
3270		netdev_info(netdev, "closing\n");
3271		rtnl_lock();
3272		myri10ge_close(netdev);
3273		rtnl_unlock();
3274	}
3275	myri10ge_dummy_rdma(mgp, 0);
3276
3277	return 0;
3278}
3279
3280static int __maybe_unused myri10ge_resume(struct device *dev)
3281{
3282	struct pci_dev *pdev = to_pci_dev(dev);
3283	struct myri10ge_priv *mgp;
3284	struct net_device *netdev;
3285	int status;
3286	u16 vendor;
3287
3288	mgp = pci_get_drvdata(pdev);
3289	if (mgp == NULL)
3290		return -EINVAL;
3291	netdev = mgp->dev;
3292	msleep(5);		/* give card time to respond */
3293	pci_read_config_word(mgp->pdev, PCI_VENDOR_ID, &vendor);
3294	if (vendor == 0xffff) {
3295		netdev_err(mgp->dev, "device disappeared!\n");
3296		return -EIO;
3297	}
3298
3299	myri10ge_reset(mgp);
3300	myri10ge_dummy_rdma(mgp, 1);
3301
3302	if (netif_running(netdev)) {
3303		rtnl_lock();
3304		status = myri10ge_open(netdev);
3305		rtnl_unlock();
3306		if (status != 0)
3307			goto abort_with_enabled;
3308
3309	}
3310	netif_device_attach(netdev);
3311
3312	return 0;
3313
3314abort_with_enabled:
3315	return -EIO;
3316}
3317
3318static u32 myri10ge_read_reboot(struct myri10ge_priv *mgp)
3319{
3320	struct pci_dev *pdev = mgp->pdev;
3321	int vs = mgp->vendor_specific_offset;
3322	u32 reboot;
3323
3324	/*enter read32 mode */
3325	pci_write_config_byte(pdev, vs + 0x10, 0x3);
3326
3327	/*read REBOOT_STATUS (0xfffffff0) */
3328	pci_write_config_dword(pdev, vs + 0x18, 0xfffffff0);
3329	pci_read_config_dword(pdev, vs + 0x14, &reboot);
3330	return reboot;
3331}
3332
3333static void
3334myri10ge_check_slice(struct myri10ge_slice_state *ss, int *reset_needed,
3335		     int *busy_slice_cnt, u32 rx_pause_cnt)
3336{
3337	struct myri10ge_priv *mgp = ss->mgp;
3338	int slice = ss - mgp->ss;
3339
3340	if (ss->tx.req != ss->tx.done &&
3341	    ss->tx.done == ss->watchdog_tx_done &&
3342	    ss->watchdog_tx_req != ss->watchdog_tx_done) {
3343		/* nic seems like it might be stuck.. */
3344		if (rx_pause_cnt != mgp->watchdog_pause) {
3345			if (net_ratelimit())
3346				netdev_warn(mgp->dev, "slice %d: TX paused, "
3347					    "check link partner\n", slice);
3348		} else {
3349			netdev_warn(mgp->dev,
3350				    "slice %d: TX stuck %d %d %d %d %d %d\n",
3351				    slice, ss->tx.queue_active, ss->tx.req,
3352				    ss->tx.done, ss->tx.pkt_start,
3353				    ss->tx.pkt_done,
3354				    (int)ntohl(mgp->ss[slice].fw_stats->
3355					       send_done_count));
3356			*reset_needed = 1;
3357			ss->stuck = 1;
3358		}
3359	}
3360	if (ss->watchdog_tx_done != ss->tx.done ||
3361	    ss->watchdog_rx_done != ss->rx_done.cnt) {
3362		*busy_slice_cnt += 1;
3363	}
3364	ss->watchdog_tx_done = ss->tx.done;
3365	ss->watchdog_tx_req = ss->tx.req;
3366	ss->watchdog_rx_done = ss->rx_done.cnt;
3367}
3368
3369/*
3370 * This watchdog is used to check whether the board has suffered
3371 * from a parity error and needs to be recovered.
3372 */
3373static void myri10ge_watchdog(struct work_struct *work)
3374{
3375	struct myri10ge_priv *mgp =
3376	    container_of(work, struct myri10ge_priv, watchdog_work);
3377	struct myri10ge_slice_state *ss;
3378	u32 reboot, rx_pause_cnt;
3379	int status, rebooted;
3380	int i;
3381	int reset_needed = 0;
3382	int busy_slice_cnt = 0;
3383	u16 cmd, vendor;
3384
3385	mgp->watchdog_resets++;
3386	pci_read_config_word(mgp->pdev, PCI_COMMAND, &cmd);
3387	rebooted = 0;
3388	if ((cmd & PCI_COMMAND_MASTER) == 0) {
3389		/* Bus master DMA disabled?  Check to see
3390		 * if the card rebooted due to a parity error
3391		 * For now, just report it */
3392		reboot = myri10ge_read_reboot(mgp);
3393		netdev_err(mgp->dev, "NIC rebooted (0x%x),%s resetting\n",
3394			   reboot, myri10ge_reset_recover ? "" : " not");
3395		if (myri10ge_reset_recover == 0)
3396			return;
3397		rtnl_lock();
3398		mgp->rebooted = 1;
3399		rebooted = 1;
3400		myri10ge_close(mgp->dev);
3401		myri10ge_reset_recover--;
3402		mgp->rebooted = 0;
3403		/*
3404		 * A rebooted nic will come back with config space as
3405		 * it was after power was applied to PCIe bus.
3406		 * Attempt to restore config space which was saved
3407		 * when the driver was loaded, or the last time the
3408		 * nic was resumed from power saving mode.
3409		 */
3410		pci_restore_state(mgp->pdev);
3411
3412		/* save state again for accounting reasons */
3413		pci_save_state(mgp->pdev);
3414
3415	} else {
3416		/* if we get back -1's from our slot, perhaps somebody
3417		 * powered off our card.  Don't try to reset it in
3418		 * this case */
3419		if (cmd == 0xffff) {
3420			pci_read_config_word(mgp->pdev, PCI_VENDOR_ID, &vendor);
3421			if (vendor == 0xffff) {
3422				netdev_err(mgp->dev, "device disappeared!\n");
3423				return;
3424			}
3425		}
3426		/* Perhaps it is a software error. See if stuck slice
3427		 * has recovered, reset if not */
3428		rx_pause_cnt = ntohl(mgp->ss[0].fw_stats->dropped_pause);
3429		for (i = 0; i < mgp->num_slices; i++) {
3430			ss = mgp->ss;
3431			if (ss->stuck) {
3432				myri10ge_check_slice(ss, &reset_needed,
3433						     &busy_slice_cnt,
3434						     rx_pause_cnt);
3435				ss->stuck = 0;
3436			}
3437		}
3438		if (!reset_needed) {
3439			netdev_dbg(mgp->dev, "not resetting\n");
3440			return;
3441		}
3442
3443		netdev_err(mgp->dev, "device timeout, resetting\n");
3444	}
3445
3446	if (!rebooted) {
3447		rtnl_lock();
3448		myri10ge_close(mgp->dev);
3449	}
3450	status = myri10ge_load_firmware(mgp, 1);
3451	if (status != 0)
3452		netdev_err(mgp->dev, "failed to load firmware\n");
3453	else
3454		myri10ge_open(mgp->dev);
3455	rtnl_unlock();
3456}
3457
3458/*
3459 * We use our own timer routine rather than relying upon
3460 * netdev->tx_timeout because we have a very large hardware transmit
3461 * queue.  Due to the large queue, the netdev->tx_timeout function
3462 * cannot detect a NIC with a parity error in a timely fashion if the
3463 * NIC is lightly loaded.
3464 */
3465static void myri10ge_watchdog_timer(struct timer_list *t)
3466{
3467	struct myri10ge_priv *mgp;
3468	struct myri10ge_slice_state *ss;
3469	int i, reset_needed, busy_slice_cnt;
3470	u32 rx_pause_cnt;
3471	u16 cmd;
3472
3473	mgp = from_timer(mgp, t, watchdog_timer);
3474
3475	rx_pause_cnt = ntohl(mgp->ss[0].fw_stats->dropped_pause);
3476	busy_slice_cnt = 0;
3477	for (i = 0, reset_needed = 0;
3478	     i < mgp->num_slices && reset_needed == 0; ++i) {
3479
3480		ss = &mgp->ss[i];
3481		if (ss->rx_small.watchdog_needed) {
3482			myri10ge_alloc_rx_pages(mgp, &ss->rx_small,
3483						mgp->small_bytes + MXGEFW_PAD,
3484						1);
3485			if (ss->rx_small.fill_cnt - ss->rx_small.cnt >=
3486			    myri10ge_fill_thresh)
3487				ss->rx_small.watchdog_needed = 0;
3488		}
3489		if (ss->rx_big.watchdog_needed) {
3490			myri10ge_alloc_rx_pages(mgp, &ss->rx_big,
3491						mgp->big_bytes, 1);
3492			if (ss->rx_big.fill_cnt - ss->rx_big.cnt >=
3493			    myri10ge_fill_thresh)
3494				ss->rx_big.watchdog_needed = 0;
3495		}
3496		myri10ge_check_slice(ss, &reset_needed, &busy_slice_cnt,
3497				     rx_pause_cnt);
3498	}
3499	/* if we've sent or received no traffic, poll the NIC to
3500	 * ensure it is still there.  Otherwise, we risk not noticing
3501	 * an error in a timely fashion */
3502	if (busy_slice_cnt == 0) {
3503		pci_read_config_word(mgp->pdev, PCI_COMMAND, &cmd);
3504		if ((cmd & PCI_COMMAND_MASTER) == 0) {
3505			reset_needed = 1;
3506		}
3507	}
3508	mgp->watchdog_pause = rx_pause_cnt;
3509
3510	if (reset_needed) {
3511		schedule_work(&mgp->watchdog_work);
3512	} else {
3513		/* rearm timer */
3514		mod_timer(&mgp->watchdog_timer,
3515			  jiffies + myri10ge_watchdog_timeout * HZ);
3516	}
3517}
3518
3519static void myri10ge_free_slices(struct myri10ge_priv *mgp)
3520{
3521	struct myri10ge_slice_state *ss;
3522	struct pci_dev *pdev = mgp->pdev;
3523	size_t bytes;
3524	int i;
3525
3526	if (mgp->ss == NULL)
3527		return;
3528
3529	for (i = 0; i < mgp->num_slices; i++) {
3530		ss = &mgp->ss[i];
3531		if (ss->rx_done.entry != NULL) {
3532			bytes = mgp->max_intr_slots *
3533			    sizeof(*ss->rx_done.entry);
3534			dma_free_coherent(&pdev->dev, bytes,
3535					  ss->rx_done.entry, ss->rx_done.bus);
3536			ss->rx_done.entry = NULL;
3537		}
3538		if (ss->fw_stats != NULL) {
3539			bytes = sizeof(*ss->fw_stats);
3540			dma_free_coherent(&pdev->dev, bytes,
3541					  ss->fw_stats, ss->fw_stats_bus);
3542			ss->fw_stats = NULL;
3543		}
3544		__netif_napi_del(&ss->napi);
3545	}
3546	/* Wait till napi structs are no longer used, and then free ss. */
3547	synchronize_net();
3548	kfree(mgp->ss);
3549	mgp->ss = NULL;
3550}
3551
3552static int myri10ge_alloc_slices(struct myri10ge_priv *mgp)
3553{
3554	struct myri10ge_slice_state *ss;
3555	struct pci_dev *pdev = mgp->pdev;
3556	size_t bytes;
3557	int i;
3558
3559	bytes = sizeof(*mgp->ss) * mgp->num_slices;
3560	mgp->ss = kzalloc(bytes, GFP_KERNEL);
3561	if (mgp->ss == NULL) {
3562		return -ENOMEM;
3563	}
3564
3565	for (i = 0; i < mgp->num_slices; i++) {
3566		ss = &mgp->ss[i];
3567		bytes = mgp->max_intr_slots * sizeof(*ss->rx_done.entry);
3568		ss->rx_done.entry = dma_alloc_coherent(&pdev->dev, bytes,
3569						       &ss->rx_done.bus,
3570						       GFP_KERNEL);
3571		if (ss->rx_done.entry == NULL)
3572			goto abort;
3573		bytes = sizeof(*ss->fw_stats);
3574		ss->fw_stats = dma_alloc_coherent(&pdev->dev, bytes,
3575						  &ss->fw_stats_bus,
3576						  GFP_KERNEL);
3577		if (ss->fw_stats == NULL)
3578			goto abort;
3579		ss->mgp = mgp;
3580		ss->dev = mgp->dev;
3581		netif_napi_add(ss->dev, &ss->napi, myri10ge_poll,
3582			       myri10ge_napi_weight);
3583	}
3584	return 0;
3585abort:
3586	myri10ge_free_slices(mgp);
3587	return -ENOMEM;
3588}
3589
3590/*
3591 * This function determines the number of slices supported.
3592 * The number slices is the minimum of the number of CPUS,
3593 * the number of MSI-X irqs supported, the number of slices
3594 * supported by the firmware
3595 */
3596static void myri10ge_probe_slices(struct myri10ge_priv *mgp)
3597{
3598	struct myri10ge_cmd cmd;
3599	struct pci_dev *pdev = mgp->pdev;
3600	char *old_fw;
3601	bool old_allocated;
3602	int i, status, ncpus;
3603
3604	mgp->num_slices = 1;
3605	ncpus = netif_get_num_default_rss_queues();
3606
3607	if (myri10ge_max_slices == 1 || !pdev->msix_cap ||
3608	    (myri10ge_max_slices == -1 && ncpus < 2))
3609		return;
3610
3611	/* try to load the slice aware rss firmware */
3612	old_fw = mgp->fw_name;
3613	old_allocated = mgp->fw_name_allocated;
3614	/* don't free old_fw if we override it. */
3615	mgp->fw_name_allocated = false;
3616
3617	if (myri10ge_fw_name != NULL) {
3618		dev_info(&mgp->pdev->dev, "overriding rss firmware to %s\n",
3619			 myri10ge_fw_name);
3620		set_fw_name(mgp, myri10ge_fw_name, false);
3621	} else if (old_fw == myri10ge_fw_aligned)
3622		set_fw_name(mgp, myri10ge_fw_rss_aligned, false);
3623	else
3624		set_fw_name(mgp, myri10ge_fw_rss_unaligned, false);
3625	status = myri10ge_load_firmware(mgp, 0);
3626	if (status != 0) {
3627		dev_info(&pdev->dev, "Rss firmware not found\n");
3628		if (old_allocated)
3629			kfree(old_fw);
3630		return;
3631	}
3632
3633	/* hit the board with a reset to ensure it is alive */
3634	memset(&cmd, 0, sizeof(cmd));
3635	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_RESET, &cmd, 0);
3636	if (status != 0) {
3637		dev_err(&mgp->pdev->dev, "failed reset\n");
3638		goto abort_with_fw;
3639	}
3640
3641	mgp->max_intr_slots = cmd.data0 / sizeof(struct mcp_slot);
3642
3643	/* tell it the size of the interrupt queues */
3644	cmd.data0 = mgp->max_intr_slots * sizeof(struct mcp_slot);
3645	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_INTRQ_SIZE, &cmd, 0);
3646	if (status != 0) {
3647		dev_err(&mgp->pdev->dev, "failed MXGEFW_CMD_SET_INTRQ_SIZE\n");
3648		goto abort_with_fw;
3649	}
3650
3651	/* ask the maximum number of slices it supports */
3652	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_MAX_RSS_QUEUES, &cmd, 0);
3653	if (status != 0)
3654		goto abort_with_fw;
3655	else
3656		mgp->num_slices = cmd.data0;
3657
3658	/* Only allow multiple slices if MSI-X is usable */
3659	if (!myri10ge_msi) {
3660		goto abort_with_fw;
3661	}
3662
3663	/* if the admin did not specify a limit to how many
3664	 * slices we should use, cap it automatically to the
3665	 * number of CPUs currently online */
3666	if (myri10ge_max_slices == -1)
3667		myri10ge_max_slices = ncpus;
3668
3669	if (mgp->num_slices > myri10ge_max_slices)
3670		mgp->num_slices = myri10ge_max_slices;
3671
3672	/* Now try to allocate as many MSI-X vectors as we have
3673	 * slices. We give up on MSI-X if we can only get a single
3674	 * vector. */
3675
3676	mgp->msix_vectors = kcalloc(mgp->num_slices, sizeof(*mgp->msix_vectors),
3677				    GFP_KERNEL);
3678	if (mgp->msix_vectors == NULL)
3679		goto no_msix;
3680	for (i = 0; i < mgp->num_slices; i++) {
3681		mgp->msix_vectors[i].entry = i;
3682	}
3683
3684	while (mgp->num_slices > 1) {
3685		mgp->num_slices = rounddown_pow_of_two(mgp->num_slices);
3686		if (mgp->num_slices == 1)
3687			goto no_msix;
3688		status = pci_enable_msix_range(pdev,
3689					       mgp->msix_vectors,
3690					       mgp->num_slices,
3691					       mgp->num_slices);
3692		if (status < 0)
3693			goto no_msix;
3694
3695		pci_disable_msix(pdev);
3696
3697		if (status == mgp->num_slices) {
3698			if (old_allocated)
3699				kfree(old_fw);
3700			return;
3701		} else {
3702			mgp->num_slices = status;
3703		}
3704	}
3705
3706no_msix:
3707	if (mgp->msix_vectors != NULL) {
3708		kfree(mgp->msix_vectors);
3709		mgp->msix_vectors = NULL;
3710	}
3711
3712abort_with_fw:
3713	mgp->num_slices = 1;
3714	set_fw_name(mgp, old_fw, old_allocated);
3715	myri10ge_load_firmware(mgp, 0);
3716}
3717
3718static const struct net_device_ops myri10ge_netdev_ops = {
3719	.ndo_open		= myri10ge_open,
3720	.ndo_stop		= myri10ge_close,
3721	.ndo_start_xmit		= myri10ge_xmit,
3722	.ndo_get_stats64	= myri10ge_get_stats,
3723	.ndo_validate_addr	= eth_validate_addr,
3724	.ndo_change_mtu		= myri10ge_change_mtu,
3725	.ndo_set_rx_mode	= myri10ge_set_multicast_list,
3726	.ndo_set_mac_address	= myri10ge_set_mac_address,
3727};
3728
3729static int myri10ge_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
3730{
3731	struct net_device *netdev;
3732	struct myri10ge_priv *mgp;
3733	struct device *dev = &pdev->dev;
3734	int i;
3735	int status = -ENXIO;
3736	int dac_enabled;
3737	unsigned hdr_offset, ss_offset;
3738	static int board_number;
3739
3740	netdev = alloc_etherdev_mq(sizeof(*mgp), MYRI10GE_MAX_SLICES);
3741	if (netdev == NULL)
3742		return -ENOMEM;
3743
3744	SET_NETDEV_DEV(netdev, &pdev->dev);
3745
3746	mgp = netdev_priv(netdev);
3747	mgp->dev = netdev;
3748	mgp->pdev = pdev;
3749	mgp->pause = myri10ge_flow_control;
3750	mgp->intr_coal_delay = myri10ge_intr_coal_delay;
3751	mgp->msg_enable = netif_msg_init(myri10ge_debug, MYRI10GE_MSG_DEFAULT);
3752	mgp->board_number = board_number;
3753	init_waitqueue_head(&mgp->down_wq);
3754
3755	if (pci_enable_device(pdev)) {
3756		dev_err(&pdev->dev, "pci_enable_device call failed\n");
3757		status = -ENODEV;
3758		goto abort_with_netdev;
3759	}
3760
3761	/* Find the vendor-specific cap so we can check
3762	 * the reboot register later on */
3763	mgp->vendor_specific_offset
3764	    = pci_find_capability(pdev, PCI_CAP_ID_VNDR);
3765
3766	/* Set our max read request to 4KB */
3767	status = pcie_set_readrq(pdev, 4096);
3768	if (status != 0) {
3769		dev_err(&pdev->dev, "Error %d writing PCI_EXP_DEVCTL\n",
3770			status);
3771		goto abort_with_enabled;
3772	}
3773
3774	myri10ge_mask_surprise_down(pdev);
3775	pci_set_master(pdev);
3776	dac_enabled = 1;
3777	status = pci_set_dma_mask(pdev, DMA_BIT_MASK(64));
3778	if (status != 0) {
3779		dac_enabled = 0;
3780		dev_err(&pdev->dev,
3781			"64-bit pci address mask was refused, "
3782			"trying 32-bit\n");
3783		status = pci_set_dma_mask(pdev, DMA_BIT_MASK(32));
3784	}
3785	if (status != 0) {
3786		dev_err(&pdev->dev, "Error %d setting DMA mask\n", status);
3787		goto abort_with_enabled;
3788	}
3789	(void)pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64));
3790	mgp->cmd = dma_alloc_coherent(&pdev->dev, sizeof(*mgp->cmd),
3791				      &mgp->cmd_bus, GFP_KERNEL);
3792	if (!mgp->cmd) {
3793		status = -ENOMEM;
3794		goto abort_with_enabled;
3795	}
3796
3797	mgp->board_span = pci_resource_len(pdev, 0);
3798	mgp->iomem_base = pci_resource_start(pdev, 0);
3799	mgp->wc_cookie = arch_phys_wc_add(mgp->iomem_base, mgp->board_span);
3800	mgp->sram = ioremap_wc(mgp->iomem_base, mgp->board_span);
3801	if (mgp->sram == NULL) {
3802		dev_err(&pdev->dev, "ioremap failed for %ld bytes at 0x%lx\n",
3803			mgp->board_span, mgp->iomem_base);
3804		status = -ENXIO;
3805		goto abort_with_mtrr;
3806	}
3807	hdr_offset =
3808	    swab32(readl(mgp->sram + MCP_HEADER_PTR_OFFSET)) & 0xffffc;
3809	ss_offset = hdr_offset + offsetof(struct mcp_gen_header, string_specs);
3810	mgp->sram_size = swab32(readl(mgp->sram + ss_offset));
3811	if (mgp->sram_size > mgp->board_span ||
3812	    mgp->sram_size <= MYRI10GE_FW_OFFSET) {
3813		dev_err(&pdev->dev,
3814			"invalid sram_size %dB or board span %ldB\n",
3815			mgp->sram_size, mgp->board_span);
3816		status = -EINVAL;
3817		goto abort_with_ioremap;
3818	}
3819	memcpy_fromio(mgp->eeprom_strings,
3820		      mgp->sram + mgp->sram_size, MYRI10GE_EEPROM_STRINGS_SIZE);
3821	memset(mgp->eeprom_strings + MYRI10GE_EEPROM_STRINGS_SIZE - 2, 0, 2);
3822	status = myri10ge_read_mac_addr(mgp);
3823	if (status)
3824		goto abort_with_ioremap;
3825
3826	for (i = 0; i < ETH_ALEN; i++)
3827		netdev->dev_addr[i] = mgp->mac_addr[i];
3828
3829	myri10ge_select_firmware(mgp);
3830
3831	status = myri10ge_load_firmware(mgp, 1);
3832	if (status != 0) {
3833		dev_err(&pdev->dev, "failed to load firmware\n");
3834		goto abort_with_ioremap;
3835	}
3836	myri10ge_probe_slices(mgp);
3837	status = myri10ge_alloc_slices(mgp);
3838	if (status != 0) {
3839		dev_err(&pdev->dev, "failed to alloc slice state\n");
3840		goto abort_with_firmware;
3841	}
3842	netif_set_real_num_tx_queues(netdev, mgp->num_slices);
3843	netif_set_real_num_rx_queues(netdev, mgp->num_slices);
3844	status = myri10ge_reset(mgp);
3845	if (status != 0) {
3846		dev_err(&pdev->dev, "failed reset\n");
3847		goto abort_with_slices;
3848	}
3849#ifdef CONFIG_MYRI10GE_DCA
3850	myri10ge_setup_dca(mgp);
3851#endif
3852	pci_set_drvdata(pdev, mgp);
3853
3854	/* MTU range: 68 - 9000 */
3855	netdev->min_mtu = ETH_MIN_MTU;
3856	netdev->max_mtu = MYRI10GE_MAX_ETHER_MTU - ETH_HLEN;
3857
3858	if (myri10ge_initial_mtu > netdev->max_mtu)
3859		myri10ge_initial_mtu = netdev->max_mtu;
3860	if (myri10ge_initial_mtu < netdev->min_mtu)
3861		myri10ge_initial_mtu = netdev->min_mtu;
3862
3863	netdev->mtu = myri10ge_initial_mtu;
3864
3865	netdev->netdev_ops = &myri10ge_netdev_ops;
3866	netdev->hw_features = mgp->features | NETIF_F_RXCSUM;
3867
3868	/* fake NETIF_F_HW_VLAN_CTAG_RX for good GRO performance */
3869	netdev->hw_features |= NETIF_F_HW_VLAN_CTAG_RX;
3870
3871	netdev->features = netdev->hw_features;
3872
3873	if (dac_enabled)
3874		netdev->features |= NETIF_F_HIGHDMA;
3875
3876	netdev->vlan_features |= mgp->features;
3877	if (mgp->fw_ver_tiny < 37)
3878		netdev->vlan_features &= ~NETIF_F_TSO6;
3879	if (mgp->fw_ver_tiny < 32)
3880		netdev->vlan_features &= ~NETIF_F_TSO;
3881
3882	/* make sure we can get an irq, and that MSI can be
3883	 * setup (if available). */
3884	status = myri10ge_request_irq(mgp);
3885	if (status != 0)
3886		goto abort_with_slices;
3887	myri10ge_free_irq(mgp);
3888
3889	/* Save configuration space to be restored if the
3890	 * nic resets due to a parity error */
3891	pci_save_state(pdev);
3892
3893	/* Setup the watchdog timer */
3894	timer_setup(&mgp->watchdog_timer, myri10ge_watchdog_timer, 0);
3895
3896	netdev->ethtool_ops = &myri10ge_ethtool_ops;
3897	INIT_WORK(&mgp->watchdog_work, myri10ge_watchdog);
3898	status = register_netdev(netdev);
3899	if (status != 0) {
3900		dev_err(&pdev->dev, "register_netdev failed: %d\n", status);
3901		goto abort_with_state;
3902	}
3903	if (mgp->msix_enabled)
3904		dev_info(dev, "%d MSI-X IRQs, tx bndry %d, fw %s, MTRR %s, WC Enabled\n",
3905			 mgp->num_slices, mgp->tx_boundary, mgp->fw_name,
3906			 (mgp->wc_cookie > 0 ? "Enabled" : "Disabled"));
3907	else
3908		dev_info(dev, "%s IRQ %d, tx bndry %d, fw %s, MTRR %s, WC Enabled\n",
3909			 mgp->msi_enabled ? "MSI" : "xPIC",
3910			 pdev->irq, mgp->tx_boundary, mgp->fw_name,
3911			 (mgp->wc_cookie > 0 ? "Enabled" : "Disabled"));
3912
3913	board_number++;
3914	return 0;
3915
3916abort_with_state:
3917	pci_restore_state(pdev);
3918
3919abort_with_slices:
3920	myri10ge_free_slices(mgp);
3921
3922abort_with_firmware:
3923	kfree(mgp->msix_vectors);
3924	myri10ge_dummy_rdma(mgp, 0);
3925
3926abort_with_ioremap:
3927	if (mgp->mac_addr_string != NULL)
3928		dev_err(&pdev->dev,
3929			"myri10ge_probe() failed: MAC=%s, SN=%ld\n",
3930			mgp->mac_addr_string, mgp->serial_number);
3931	iounmap(mgp->sram);
3932
3933abort_with_mtrr:
3934	arch_phys_wc_del(mgp->wc_cookie);
3935	dma_free_coherent(&pdev->dev, sizeof(*mgp->cmd),
3936			  mgp->cmd, mgp->cmd_bus);
3937
3938abort_with_enabled:
3939	pci_disable_device(pdev);
3940
3941abort_with_netdev:
3942	set_fw_name(mgp, NULL, false);
3943	free_netdev(netdev);
3944	return status;
3945}
3946
3947/*
3948 * myri10ge_remove
3949 *
3950 * Does what is necessary to shutdown one Myrinet device. Called
3951 *   once for each Myrinet card by the kernel when a module is
3952 *   unloaded.
3953 */
3954static void myri10ge_remove(struct pci_dev *pdev)
3955{
3956	struct myri10ge_priv *mgp;
3957	struct net_device *netdev;
3958
3959	mgp = pci_get_drvdata(pdev);
3960	if (mgp == NULL)
3961		return;
3962
3963	cancel_work_sync(&mgp->watchdog_work);
3964	netdev = mgp->dev;
3965	unregister_netdev(netdev);
3966
3967#ifdef CONFIG_MYRI10GE_DCA
3968	myri10ge_teardown_dca(mgp);
3969#endif
3970	myri10ge_dummy_rdma(mgp, 0);
3971
3972	/* avoid a memory leak */
3973	pci_restore_state(pdev);
3974
3975	iounmap(mgp->sram);
3976	arch_phys_wc_del(mgp->wc_cookie);
3977	myri10ge_free_slices(mgp);
3978	kfree(mgp->msix_vectors);
3979	dma_free_coherent(&pdev->dev, sizeof(*mgp->cmd),
3980			  mgp->cmd, mgp->cmd_bus);
3981
3982	set_fw_name(mgp, NULL, false);
3983	free_netdev(netdev);
3984	pci_disable_device(pdev);
3985}
3986
3987#define PCI_DEVICE_ID_MYRICOM_MYRI10GE_Z8E 	0x0008
3988#define PCI_DEVICE_ID_MYRICOM_MYRI10GE_Z8E_9	0x0009
3989
3990static const struct pci_device_id myri10ge_pci_tbl[] = {
3991	{PCI_DEVICE(PCI_VENDOR_ID_MYRICOM, PCI_DEVICE_ID_MYRICOM_MYRI10GE_Z8E)},
3992	{PCI_DEVICE
3993	 (PCI_VENDOR_ID_MYRICOM, PCI_DEVICE_ID_MYRICOM_MYRI10GE_Z8E_9)},
3994	{0},
3995};
3996
3997MODULE_DEVICE_TABLE(pci, myri10ge_pci_tbl);
3998
3999static SIMPLE_DEV_PM_OPS(myri10ge_pm_ops, myri10ge_suspend, myri10ge_resume);
4000
4001static struct pci_driver myri10ge_driver = {
4002	.name = "myri10ge",
4003	.probe = myri10ge_probe,
4004	.remove = myri10ge_remove,
4005	.id_table = myri10ge_pci_tbl,
4006	.driver.pm = &myri10ge_pm_ops,
4007};
4008
4009#ifdef CONFIG_MYRI10GE_DCA
4010static int
4011myri10ge_notify_dca(struct notifier_block *nb, unsigned long event, void *p)
4012{
4013	int err = driver_for_each_device(&myri10ge_driver.driver,
4014					 NULL, &event,
4015					 myri10ge_notify_dca_device);
4016
4017	if (err)
4018		return NOTIFY_BAD;
4019	return NOTIFY_DONE;
4020}
4021
4022static struct notifier_block myri10ge_dca_notifier = {
4023	.notifier_call = myri10ge_notify_dca,
4024	.next = NULL,
4025	.priority = 0,
4026};
4027#endif				/* CONFIG_MYRI10GE_DCA */
4028
4029static __init int myri10ge_init_module(void)
4030{
4031	pr_info("Version %s\n", MYRI10GE_VERSION_STR);
4032
4033	if (myri10ge_rss_hash > MXGEFW_RSS_HASH_TYPE_MAX) {
4034		pr_err("Illegal rssh hash type %d, defaulting to source port\n",
4035		       myri10ge_rss_hash);
4036		myri10ge_rss_hash = MXGEFW_RSS_HASH_TYPE_SRC_PORT;
4037	}
4038#ifdef CONFIG_MYRI10GE_DCA
4039	dca_register_notify(&myri10ge_dca_notifier);
4040#endif
4041	if (myri10ge_max_slices > MYRI10GE_MAX_SLICES)
4042		myri10ge_max_slices = MYRI10GE_MAX_SLICES;
4043
4044	return pci_register_driver(&myri10ge_driver);
4045}
4046
4047module_init(myri10ge_init_module);
4048
4049static __exit void myri10ge_cleanup_module(void)
4050{
4051#ifdef CONFIG_MYRI10GE_DCA
4052	dca_unregister_notify(&myri10ge_dca_notifier);
4053#endif
4054	pci_unregister_driver(&myri10ge_driver);
4055}
4056
4057module_exit(myri10ge_cleanup_module);
4058