[go: nahoru, domu]

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/io.h>
73#include <asm/processor.h>
74#ifdef CONFIG_MTRR
75#include <asm/mtrr.h>
76#endif
77#include <net/busy_poll.h>
78
79#include "myri10ge_mcp.h"
80#include "myri10ge_mcp_gen_header.h"
81
82#define MYRI10GE_VERSION_STR "1.5.3-1.534"
83
84MODULE_DESCRIPTION("Myricom 10G driver (10GbE)");
85MODULE_AUTHOR("Maintainer: help@myri.com");
86MODULE_VERSION(MYRI10GE_VERSION_STR);
87MODULE_LICENSE("Dual BSD/GPL");
88
89#define MYRI10GE_MAX_ETHER_MTU 9014
90
91#define MYRI10GE_ETH_STOPPED 0
92#define MYRI10GE_ETH_STOPPING 1
93#define MYRI10GE_ETH_STARTING 2
94#define MYRI10GE_ETH_RUNNING 3
95#define MYRI10GE_ETH_OPEN_FAILED 4
96
97#define MYRI10GE_EEPROM_STRINGS_SIZE 256
98#define MYRI10GE_MAX_SEND_DESC_TSO ((65536 / 2048) * 2)
99
100#define MYRI10GE_NO_CONFIRM_DATA htonl(0xffffffff)
101#define MYRI10GE_NO_RESPONSE_RESULT 0xffffffff
102
103#define MYRI10GE_ALLOC_ORDER 0
104#define MYRI10GE_ALLOC_SIZE ((1 << MYRI10GE_ALLOC_ORDER) * PAGE_SIZE)
105#define MYRI10GE_MAX_FRAGS_PER_FRAME (MYRI10GE_MAX_ETHER_MTU/MYRI10GE_ALLOC_SIZE + 1)
106
107#define MYRI10GE_MAX_SLICES 32
108
109struct myri10ge_rx_buffer_state {
110	struct page *page;
111	int page_offset;
112	DEFINE_DMA_UNMAP_ADDR(bus);
113	DEFINE_DMA_UNMAP_LEN(len);
114};
115
116struct myri10ge_tx_buffer_state {
117	struct sk_buff *skb;
118	int last;
119	DEFINE_DMA_UNMAP_ADDR(bus);
120	DEFINE_DMA_UNMAP_LEN(len);
121};
122
123struct myri10ge_cmd {
124	u32 data0;
125	u32 data1;
126	u32 data2;
127};
128
129struct myri10ge_rx_buf {
130	struct mcp_kreq_ether_recv __iomem *lanai;	/* lanai ptr for recv ring */
131	struct mcp_kreq_ether_recv *shadow;	/* host shadow of recv ring */
132	struct myri10ge_rx_buffer_state *info;
133	struct page *page;
134	dma_addr_t bus;
135	int page_offset;
136	int cnt;
137	int fill_cnt;
138	int alloc_fail;
139	int mask;		/* number of rx slots -1 */
140	int watchdog_needed;
141};
142
143struct myri10ge_tx_buf {
144	struct mcp_kreq_ether_send __iomem *lanai;	/* lanai ptr for sendq */
145	__be32 __iomem *send_go;	/* "go" doorbell ptr */
146	__be32 __iomem *send_stop;	/* "stop" doorbell ptr */
147	struct mcp_kreq_ether_send *req_list;	/* host shadow of sendq */
148	char *req_bytes;
149	struct myri10ge_tx_buffer_state *info;
150	int mask;		/* number of transmit slots -1  */
151	int req ____cacheline_aligned;	/* transmit slots submitted     */
152	int pkt_start;		/* packets started */
153	int stop_queue;
154	int linearized;
155	int done ____cacheline_aligned;	/* transmit slots completed     */
156	int pkt_done;		/* packets completed */
157	int wake_queue;
158	int queue_active;
159};
160
161struct myri10ge_rx_done {
162	struct mcp_slot *entry;
163	dma_addr_t bus;
164	int cnt;
165	int idx;
166};
167
168struct myri10ge_slice_netstats {
169	unsigned long rx_packets;
170	unsigned long tx_packets;
171	unsigned long rx_bytes;
172	unsigned long tx_bytes;
173	unsigned long rx_dropped;
174	unsigned long tx_dropped;
175};
176
177struct myri10ge_slice_state {
178	struct myri10ge_tx_buf tx;	/* transmit ring        */
179	struct myri10ge_rx_buf rx_small;
180	struct myri10ge_rx_buf rx_big;
181	struct myri10ge_rx_done rx_done;
182	struct net_device *dev;
183	struct napi_struct napi;
184	struct myri10ge_priv *mgp;
185	struct myri10ge_slice_netstats stats;
186	__be32 __iomem *irq_claim;
187	struct mcp_irq_data *fw_stats;
188	dma_addr_t fw_stats_bus;
189	int watchdog_tx_done;
190	int watchdog_tx_req;
191	int watchdog_rx_done;
192	int stuck;
193#ifdef CONFIG_MYRI10GE_DCA
194	int cached_dca_tag;
195	int cpu;
196	__be32 __iomem *dca_tag;
197#endif
198#ifdef CONFIG_NET_RX_BUSY_POLL
199	unsigned int state;
200#define SLICE_STATE_IDLE	0
201#define SLICE_STATE_NAPI	1	/* NAPI owns this slice */
202#define SLICE_STATE_POLL	2	/* poll owns this slice */
203#define SLICE_LOCKED (SLICE_STATE_NAPI | SLICE_STATE_POLL)
204#define SLICE_STATE_NAPI_YIELD	4	/* NAPI yielded this slice */
205#define SLICE_STATE_POLL_YIELD	8	/* poll yielded this slice */
206#define SLICE_USER_PEND (SLICE_STATE_POLL | SLICE_STATE_POLL_YIELD)
207	spinlock_t lock;
208	unsigned long lock_napi_yield;
209	unsigned long lock_poll_yield;
210	unsigned long busy_poll_miss;
211	unsigned long busy_poll_cnt;
212#endif /* CONFIG_NET_RX_BUSY_POLL */
213	char irq_desc[32];
214};
215
216struct myri10ge_priv {
217	struct myri10ge_slice_state *ss;
218	int tx_boundary;	/* boundary transmits cannot cross */
219	int num_slices;
220	int running;		/* running?             */
221	int small_bytes;
222	int big_bytes;
223	int max_intr_slots;
224	struct net_device *dev;
225	u8 __iomem *sram;
226	int sram_size;
227	unsigned long board_span;
228	unsigned long iomem_base;
229	__be32 __iomem *irq_deassert;
230	char *mac_addr_string;
231	struct mcp_cmd_response *cmd;
232	dma_addr_t cmd_bus;
233	struct pci_dev *pdev;
234	int msi_enabled;
235	int msix_enabled;
236	struct msix_entry *msix_vectors;
237#ifdef CONFIG_MYRI10GE_DCA
238	int dca_enabled;
239	int relaxed_order;
240#endif
241	u32 link_state;
242	unsigned int rdma_tags_available;
243	int intr_coal_delay;
244	__be32 __iomem *intr_coal_delay_ptr;
245	int mtrr;
246	int wc_enabled;
247	int down_cnt;
248	wait_queue_head_t down_wq;
249	struct work_struct watchdog_work;
250	struct timer_list watchdog_timer;
251	int watchdog_resets;
252	int watchdog_pause;
253	int pause;
254	bool fw_name_allocated;
255	char *fw_name;
256	char eeprom_strings[MYRI10GE_EEPROM_STRINGS_SIZE];
257	char *product_code_string;
258	char fw_version[128];
259	int fw_ver_major;
260	int fw_ver_minor;
261	int fw_ver_tiny;
262	int adopted_rx_filter_bug;
263	u8 mac_addr[ETH_ALEN];		/* eeprom mac address */
264	unsigned long serial_number;
265	int vendor_specific_offset;
266	int fw_multicast_support;
267	u32 features;
268	u32 max_tso6;
269	u32 read_dma;
270	u32 write_dma;
271	u32 read_write_dma;
272	u32 link_changes;
273	u32 msg_enable;
274	unsigned int board_number;
275	int rebooted;
276};
277
278static char *myri10ge_fw_unaligned = "myri10ge_ethp_z8e.dat";
279static char *myri10ge_fw_aligned = "myri10ge_eth_z8e.dat";
280static char *myri10ge_fw_rss_unaligned = "myri10ge_rss_ethp_z8e.dat";
281static char *myri10ge_fw_rss_aligned = "myri10ge_rss_eth_z8e.dat";
282MODULE_FIRMWARE("myri10ge_ethp_z8e.dat");
283MODULE_FIRMWARE("myri10ge_eth_z8e.dat");
284MODULE_FIRMWARE("myri10ge_rss_ethp_z8e.dat");
285MODULE_FIRMWARE("myri10ge_rss_eth_z8e.dat");
286
287/* Careful: must be accessed under kparam_block_sysfs_write */
288static char *myri10ge_fw_name = NULL;
289module_param(myri10ge_fw_name, charp, S_IRUGO | S_IWUSR);
290MODULE_PARM_DESC(myri10ge_fw_name, "Firmware image name");
291
292#define MYRI10GE_MAX_BOARDS 8
293static char *myri10ge_fw_names[MYRI10GE_MAX_BOARDS] =
294    {[0 ... (MYRI10GE_MAX_BOARDS - 1)] = NULL };
295module_param_array_named(myri10ge_fw_names, myri10ge_fw_names, charp, NULL,
296			 0444);
297MODULE_PARM_DESC(myri10ge_fw_name, "Firmware image names per board");
298
299static int myri10ge_ecrc_enable = 1;
300module_param(myri10ge_ecrc_enable, int, S_IRUGO);
301MODULE_PARM_DESC(myri10ge_ecrc_enable, "Enable Extended CRC on PCI-E");
302
303static int myri10ge_small_bytes = -1;	/* -1 == auto */
304module_param(myri10ge_small_bytes, int, S_IRUGO | S_IWUSR);
305MODULE_PARM_DESC(myri10ge_small_bytes, "Threshold of small packets");
306
307static int myri10ge_msi = 1;	/* enable msi by default */
308module_param(myri10ge_msi, int, S_IRUGO | S_IWUSR);
309MODULE_PARM_DESC(myri10ge_msi, "Enable Message Signalled Interrupts");
310
311static int myri10ge_intr_coal_delay = 75;
312module_param(myri10ge_intr_coal_delay, int, S_IRUGO);
313MODULE_PARM_DESC(myri10ge_intr_coal_delay, "Interrupt coalescing delay");
314
315static int myri10ge_flow_control = 1;
316module_param(myri10ge_flow_control, int, S_IRUGO);
317MODULE_PARM_DESC(myri10ge_flow_control, "Pause parameter");
318
319static int myri10ge_deassert_wait = 1;
320module_param(myri10ge_deassert_wait, int, S_IRUGO | S_IWUSR);
321MODULE_PARM_DESC(myri10ge_deassert_wait,
322		 "Wait when deasserting legacy interrupts");
323
324static int myri10ge_force_firmware = 0;
325module_param(myri10ge_force_firmware, int, S_IRUGO);
326MODULE_PARM_DESC(myri10ge_force_firmware,
327		 "Force firmware to assume aligned completions");
328
329static int myri10ge_initial_mtu = MYRI10GE_MAX_ETHER_MTU - ETH_HLEN;
330module_param(myri10ge_initial_mtu, int, S_IRUGO);
331MODULE_PARM_DESC(myri10ge_initial_mtu, "Initial MTU");
332
333static int myri10ge_napi_weight = 64;
334module_param(myri10ge_napi_weight, int, S_IRUGO);
335MODULE_PARM_DESC(myri10ge_napi_weight, "Set NAPI weight");
336
337static int myri10ge_watchdog_timeout = 1;
338module_param(myri10ge_watchdog_timeout, int, S_IRUGO);
339MODULE_PARM_DESC(myri10ge_watchdog_timeout, "Set watchdog timeout");
340
341static int myri10ge_max_irq_loops = 1048576;
342module_param(myri10ge_max_irq_loops, int, S_IRUGO);
343MODULE_PARM_DESC(myri10ge_max_irq_loops,
344		 "Set stuck legacy IRQ detection threshold");
345
346#define MYRI10GE_MSG_DEFAULT NETIF_MSG_LINK
347
348static int myri10ge_debug = -1;	/* defaults above */
349module_param(myri10ge_debug, int, 0);
350MODULE_PARM_DESC(myri10ge_debug, "Debug level (0=none,...,16=all)");
351
352static int myri10ge_fill_thresh = 256;
353module_param(myri10ge_fill_thresh, int, S_IRUGO | S_IWUSR);
354MODULE_PARM_DESC(myri10ge_fill_thresh, "Number of empty rx slots allowed");
355
356static int myri10ge_reset_recover = 1;
357
358static int myri10ge_max_slices = 1;
359module_param(myri10ge_max_slices, int, S_IRUGO);
360MODULE_PARM_DESC(myri10ge_max_slices, "Max tx/rx queues");
361
362static int myri10ge_rss_hash = MXGEFW_RSS_HASH_TYPE_SRC_DST_PORT;
363module_param(myri10ge_rss_hash, int, S_IRUGO);
364MODULE_PARM_DESC(myri10ge_rss_hash, "Type of RSS hashing to do");
365
366static int myri10ge_dca = 1;
367module_param(myri10ge_dca, int, S_IRUGO);
368MODULE_PARM_DESC(myri10ge_dca, "Enable DCA if possible");
369
370#define MYRI10GE_FW_OFFSET 1024*1024
371#define MYRI10GE_HIGHPART_TO_U32(X) \
372(sizeof (X) == 8) ? ((u32)((u64)(X) >> 32)) : (0)
373#define MYRI10GE_LOWPART_TO_U32(X) ((u32)(X))
374
375#define myri10ge_pio_copy(to,from,size) __iowrite64_copy(to,from,size/8)
376
377static void myri10ge_set_multicast_list(struct net_device *dev);
378static netdev_tx_t myri10ge_sw_tso(struct sk_buff *skb,
379					 struct net_device *dev);
380
381static inline void put_be32(__be32 val, __be32 __iomem * p)
382{
383	__raw_writel((__force __u32) val, (__force void __iomem *)p);
384}
385
386static struct rtnl_link_stats64 *myri10ge_get_stats(struct net_device *dev,
387						    struct rtnl_link_stats64 *stats);
388
389static void set_fw_name(struct myri10ge_priv *mgp, char *name, bool allocated)
390{
391	if (mgp->fw_name_allocated)
392		kfree(mgp->fw_name);
393	mgp->fw_name = name;
394	mgp->fw_name_allocated = allocated;
395}
396
397static int
398myri10ge_send_cmd(struct myri10ge_priv *mgp, u32 cmd,
399		  struct myri10ge_cmd *data, int atomic)
400{
401	struct mcp_cmd *buf;
402	char buf_bytes[sizeof(*buf) + 8];
403	struct mcp_cmd_response *response = mgp->cmd;
404	char __iomem *cmd_addr = mgp->sram + MXGEFW_ETH_CMD;
405	u32 dma_low, dma_high, result, value;
406	int sleep_total = 0;
407
408	/* ensure buf is aligned to 8 bytes */
409	buf = (struct mcp_cmd *)ALIGN((unsigned long)buf_bytes, 8);
410
411	buf->data0 = htonl(data->data0);
412	buf->data1 = htonl(data->data1);
413	buf->data2 = htonl(data->data2);
414	buf->cmd = htonl(cmd);
415	dma_low = MYRI10GE_LOWPART_TO_U32(mgp->cmd_bus);
416	dma_high = MYRI10GE_HIGHPART_TO_U32(mgp->cmd_bus);
417
418	buf->response_addr.low = htonl(dma_low);
419	buf->response_addr.high = htonl(dma_high);
420	response->result = htonl(MYRI10GE_NO_RESPONSE_RESULT);
421	mb();
422	myri10ge_pio_copy(cmd_addr, buf, sizeof(*buf));
423
424	/* wait up to 15ms. Longest command is the DMA benchmark,
425	 * which is capped at 5ms, but runs from a timeout handler
426	 * that runs every 7.8ms. So a 15ms timeout leaves us with
427	 * a 2.2ms margin
428	 */
429	if (atomic) {
430		/* if atomic is set, do not sleep,
431		 * and try to get the completion quickly
432		 * (1ms will be enough for those commands) */
433		for (sleep_total = 0;
434		     sleep_total < 1000 &&
435		     response->result == htonl(MYRI10GE_NO_RESPONSE_RESULT);
436		     sleep_total += 10) {
437			udelay(10);
438			mb();
439		}
440	} else {
441		/* use msleep for most command */
442		for (sleep_total = 0;
443		     sleep_total < 15 &&
444		     response->result == htonl(MYRI10GE_NO_RESPONSE_RESULT);
445		     sleep_total++)
446			msleep(1);
447	}
448
449	result = ntohl(response->result);
450	value = ntohl(response->data);
451	if (result != MYRI10GE_NO_RESPONSE_RESULT) {
452		if (result == 0) {
453			data->data0 = value;
454			return 0;
455		} else if (result == MXGEFW_CMD_UNKNOWN) {
456			return -ENOSYS;
457		} else if (result == MXGEFW_CMD_ERROR_UNALIGNED) {
458			return -E2BIG;
459		} else if (result == MXGEFW_CMD_ERROR_RANGE &&
460			   cmd == MXGEFW_CMD_ENABLE_RSS_QUEUES &&
461			   (data->
462			    data1 & MXGEFW_SLICE_ENABLE_MULTIPLE_TX_QUEUES) !=
463			   0) {
464			return -ERANGE;
465		} else {
466			dev_err(&mgp->pdev->dev,
467				"command %d failed, result = %d\n",
468				cmd, result);
469			return -ENXIO;
470		}
471	}
472
473	dev_err(&mgp->pdev->dev, "command %d timed out, result = %d\n",
474		cmd, result);
475	return -EAGAIN;
476}
477
478/*
479 * The eeprom strings on the lanaiX have the format
480 * SN=x\0
481 * MAC=x:x:x:x:x:x\0
482 * PT:ddd mmm xx xx:xx:xx xx\0
483 * PV:ddd mmm xx xx:xx:xx xx\0
484 */
485static int myri10ge_read_mac_addr(struct myri10ge_priv *mgp)
486{
487	char *ptr, *limit;
488	int i;
489
490	ptr = mgp->eeprom_strings;
491	limit = mgp->eeprom_strings + MYRI10GE_EEPROM_STRINGS_SIZE;
492
493	while (*ptr != '\0' && ptr < limit) {
494		if (memcmp(ptr, "MAC=", 4) == 0) {
495			ptr += 4;
496			mgp->mac_addr_string = ptr;
497			for (i = 0; i < 6; i++) {
498				if ((ptr + 2) > limit)
499					goto abort;
500				mgp->mac_addr[i] =
501				    simple_strtoul(ptr, &ptr, 16);
502				ptr += 1;
503			}
504		}
505		if (memcmp(ptr, "PC=", 3) == 0) {
506			ptr += 3;
507			mgp->product_code_string = ptr;
508		}
509		if (memcmp((const void *)ptr, "SN=", 3) == 0) {
510			ptr += 3;
511			mgp->serial_number = simple_strtoul(ptr, &ptr, 10);
512		}
513		while (ptr < limit && *ptr++) ;
514	}
515
516	return 0;
517
518abort:
519	dev_err(&mgp->pdev->dev, "failed to parse eeprom_strings\n");
520	return -ENXIO;
521}
522
523/*
524 * Enable or disable periodic RDMAs from the host to make certain
525 * chipsets resend dropped PCIe messages
526 */
527
528static void myri10ge_dummy_rdma(struct myri10ge_priv *mgp, int enable)
529{
530	char __iomem *submit;
531	__be32 buf[16] __attribute__ ((__aligned__(8)));
532	u32 dma_low, dma_high;
533	int i;
534
535	/* clear confirmation addr */
536	mgp->cmd->data = 0;
537	mb();
538
539	/* send a rdma command to the PCIe engine, and wait for the
540	 * response in the confirmation address.  The firmware should
541	 * write a -1 there to indicate it is alive and well
542	 */
543	dma_low = MYRI10GE_LOWPART_TO_U32(mgp->cmd_bus);
544	dma_high = MYRI10GE_HIGHPART_TO_U32(mgp->cmd_bus);
545
546	buf[0] = htonl(dma_high);	/* confirm addr MSW */
547	buf[1] = htonl(dma_low);	/* confirm addr LSW */
548	buf[2] = MYRI10GE_NO_CONFIRM_DATA;	/* confirm data */
549	buf[3] = htonl(dma_high);	/* dummy addr MSW */
550	buf[4] = htonl(dma_low);	/* dummy addr LSW */
551	buf[5] = htonl(enable);	/* enable? */
552
553	submit = mgp->sram + MXGEFW_BOOT_DUMMY_RDMA;
554
555	myri10ge_pio_copy(submit, &buf, sizeof(buf));
556	for (i = 0; mgp->cmd->data != MYRI10GE_NO_CONFIRM_DATA && i < 20; i++)
557		msleep(1);
558	if (mgp->cmd->data != MYRI10GE_NO_CONFIRM_DATA)
559		dev_err(&mgp->pdev->dev, "dummy rdma %s failed\n",
560			(enable ? "enable" : "disable"));
561}
562
563static int
564myri10ge_validate_firmware(struct myri10ge_priv *mgp,
565			   struct mcp_gen_header *hdr)
566{
567	struct device *dev = &mgp->pdev->dev;
568
569	/* check firmware type */
570	if (ntohl(hdr->mcp_type) != MCP_TYPE_ETH) {
571		dev_err(dev, "Bad firmware type: 0x%x\n", ntohl(hdr->mcp_type));
572		return -EINVAL;
573	}
574
575	/* save firmware version for ethtool */
576	strncpy(mgp->fw_version, hdr->version, sizeof(mgp->fw_version));
577	mgp->fw_version[sizeof(mgp->fw_version) - 1] = '\0';
578
579	sscanf(mgp->fw_version, "%d.%d.%d", &mgp->fw_ver_major,
580	       &mgp->fw_ver_minor, &mgp->fw_ver_tiny);
581
582	if (!(mgp->fw_ver_major == MXGEFW_VERSION_MAJOR &&
583	      mgp->fw_ver_minor == MXGEFW_VERSION_MINOR)) {
584		dev_err(dev, "Found firmware version %s\n", mgp->fw_version);
585		dev_err(dev, "Driver needs %d.%d\n", MXGEFW_VERSION_MAJOR,
586			MXGEFW_VERSION_MINOR);
587		return -EINVAL;
588	}
589	return 0;
590}
591
592static int myri10ge_load_hotplug_firmware(struct myri10ge_priv *mgp, u32 * size)
593{
594	unsigned crc, reread_crc;
595	const struct firmware *fw;
596	struct device *dev = &mgp->pdev->dev;
597	unsigned char *fw_readback;
598	struct mcp_gen_header *hdr;
599	size_t hdr_offset;
600	int status;
601	unsigned i;
602
603	if ((status = request_firmware(&fw, mgp->fw_name, dev)) < 0) {
604		dev_err(dev, "Unable to load %s firmware image via hotplug\n",
605			mgp->fw_name);
606		status = -EINVAL;
607		goto abort_with_nothing;
608	}
609
610	/* check size */
611
612	if (fw->size >= mgp->sram_size - MYRI10GE_FW_OFFSET ||
613	    fw->size < MCP_HEADER_PTR_OFFSET + 4) {
614		dev_err(dev, "Firmware size invalid:%d\n", (int)fw->size);
615		status = -EINVAL;
616		goto abort_with_fw;
617	}
618
619	/* check id */
620	hdr_offset = ntohl(*(__be32 *) (fw->data + MCP_HEADER_PTR_OFFSET));
621	if ((hdr_offset & 3) || hdr_offset + sizeof(*hdr) > fw->size) {
622		dev_err(dev, "Bad firmware file\n");
623		status = -EINVAL;
624		goto abort_with_fw;
625	}
626	hdr = (void *)(fw->data + hdr_offset);
627
628	status = myri10ge_validate_firmware(mgp, hdr);
629	if (status != 0)
630		goto abort_with_fw;
631
632	crc = crc32(~0, fw->data, fw->size);
633	for (i = 0; i < fw->size; i += 256) {
634		myri10ge_pio_copy(mgp->sram + MYRI10GE_FW_OFFSET + i,
635				  fw->data + i,
636				  min(256U, (unsigned)(fw->size - i)));
637		mb();
638		readb(mgp->sram);
639	}
640	fw_readback = vmalloc(fw->size);
641	if (!fw_readback) {
642		status = -ENOMEM;
643		goto abort_with_fw;
644	}
645	/* corruption checking is good for parity recovery and buggy chipset */
646	memcpy_fromio(fw_readback, mgp->sram + MYRI10GE_FW_OFFSET, fw->size);
647	reread_crc = crc32(~0, fw_readback, fw->size);
648	vfree(fw_readback);
649	if (crc != reread_crc) {
650		dev_err(dev, "CRC failed(fw-len=%u), got 0x%x (expect 0x%x)\n",
651			(unsigned)fw->size, reread_crc, crc);
652		status = -EIO;
653		goto abort_with_fw;
654	}
655	*size = (u32) fw->size;
656
657abort_with_fw:
658	release_firmware(fw);
659
660abort_with_nothing:
661	return status;
662}
663
664static int myri10ge_adopt_running_firmware(struct myri10ge_priv *mgp)
665{
666	struct mcp_gen_header *hdr;
667	struct device *dev = &mgp->pdev->dev;
668	const size_t bytes = sizeof(struct mcp_gen_header);
669	size_t hdr_offset;
670	int status;
671
672	/* find running firmware header */
673	hdr_offset = swab32(readl(mgp->sram + MCP_HEADER_PTR_OFFSET));
674
675	if ((hdr_offset & 3) || hdr_offset + sizeof(*hdr) > mgp->sram_size) {
676		dev_err(dev, "Running firmware has bad header offset (%d)\n",
677			(int)hdr_offset);
678		return -EIO;
679	}
680
681	/* copy header of running firmware from SRAM to host memory to
682	 * validate firmware */
683	hdr = kmalloc(bytes, GFP_KERNEL);
684	if (hdr == NULL)
685		return -ENOMEM;
686
687	memcpy_fromio(hdr, mgp->sram + hdr_offset, bytes);
688	status = myri10ge_validate_firmware(mgp, hdr);
689	kfree(hdr);
690
691	/* check to see if adopted firmware has bug where adopting
692	 * it will cause broadcasts to be filtered unless the NIC
693	 * is kept in ALLMULTI mode */
694	if (mgp->fw_ver_major == 1 && mgp->fw_ver_minor == 4 &&
695	    mgp->fw_ver_tiny >= 4 && mgp->fw_ver_tiny <= 11) {
696		mgp->adopted_rx_filter_bug = 1;
697		dev_warn(dev, "Adopting fw %d.%d.%d: "
698			 "working around rx filter bug\n",
699			 mgp->fw_ver_major, mgp->fw_ver_minor,
700			 mgp->fw_ver_tiny);
701	}
702	return status;
703}
704
705static int myri10ge_get_firmware_capabilities(struct myri10ge_priv *mgp)
706{
707	struct myri10ge_cmd cmd;
708	int status;
709
710	/* probe for IPv6 TSO support */
711	mgp->features = NETIF_F_SG | NETIF_F_HW_CSUM | NETIF_F_TSO;
712	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_MAX_TSO6_HDR_SIZE,
713				   &cmd, 0);
714	if (status == 0) {
715		mgp->max_tso6 = cmd.data0;
716		mgp->features |= NETIF_F_TSO6;
717	}
718
719	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_RX_RING_SIZE, &cmd, 0);
720	if (status != 0) {
721		dev_err(&mgp->pdev->dev,
722			"failed MXGEFW_CMD_GET_RX_RING_SIZE\n");
723		return -ENXIO;
724	}
725
726	mgp->max_intr_slots = 2 * (cmd.data0 / sizeof(struct mcp_dma_addr));
727
728	return 0;
729}
730
731static int myri10ge_load_firmware(struct myri10ge_priv *mgp, int adopt)
732{
733	char __iomem *submit;
734	__be32 buf[16] __attribute__ ((__aligned__(8)));
735	u32 dma_low, dma_high, size;
736	int status, i;
737
738	size = 0;
739	status = myri10ge_load_hotplug_firmware(mgp, &size);
740	if (status) {
741		if (!adopt)
742			return status;
743		dev_warn(&mgp->pdev->dev, "hotplug firmware loading failed\n");
744
745		/* Do not attempt to adopt firmware if there
746		 * was a bad crc */
747		if (status == -EIO)
748			return status;
749
750		status = myri10ge_adopt_running_firmware(mgp);
751		if (status != 0) {
752			dev_err(&mgp->pdev->dev,
753				"failed to adopt running firmware\n");
754			return status;
755		}
756		dev_info(&mgp->pdev->dev,
757			 "Successfully adopted running firmware\n");
758		if (mgp->tx_boundary == 4096) {
759			dev_warn(&mgp->pdev->dev,
760				 "Using firmware currently running on NIC"
761				 ".  For optimal\n");
762			dev_warn(&mgp->pdev->dev,
763				 "performance consider loading optimized "
764				 "firmware\n");
765			dev_warn(&mgp->pdev->dev, "via hotplug\n");
766		}
767
768		set_fw_name(mgp, "adopted", false);
769		mgp->tx_boundary = 2048;
770		myri10ge_dummy_rdma(mgp, 1);
771		status = myri10ge_get_firmware_capabilities(mgp);
772		return status;
773	}
774
775	/* clear confirmation addr */
776	mgp->cmd->data = 0;
777	mb();
778
779	/* send a reload command to the bootstrap MCP, and wait for the
780	 *  response in the confirmation address.  The firmware should
781	 * write a -1 there to indicate it is alive and well
782	 */
783	dma_low = MYRI10GE_LOWPART_TO_U32(mgp->cmd_bus);
784	dma_high = MYRI10GE_HIGHPART_TO_U32(mgp->cmd_bus);
785
786	buf[0] = htonl(dma_high);	/* confirm addr MSW */
787	buf[1] = htonl(dma_low);	/* confirm addr LSW */
788	buf[2] = MYRI10GE_NO_CONFIRM_DATA;	/* confirm data */
789
790	/* FIX: All newest firmware should un-protect the bottom of
791	 * the sram before handoff. However, the very first interfaces
792	 * do not. Therefore the handoff copy must skip the first 8 bytes
793	 */
794	buf[3] = htonl(MYRI10GE_FW_OFFSET + 8);	/* where the code starts */
795	buf[4] = htonl(size - 8);	/* length of code */
796	buf[5] = htonl(8);	/* where to copy to */
797	buf[6] = htonl(0);	/* where to jump to */
798
799	submit = mgp->sram + MXGEFW_BOOT_HANDOFF;
800
801	myri10ge_pio_copy(submit, &buf, sizeof(buf));
802	mb();
803	msleep(1);
804	mb();
805	i = 0;
806	while (mgp->cmd->data != MYRI10GE_NO_CONFIRM_DATA && i < 9) {
807		msleep(1 << i);
808		i++;
809	}
810	if (mgp->cmd->data != MYRI10GE_NO_CONFIRM_DATA) {
811		dev_err(&mgp->pdev->dev, "handoff failed\n");
812		return -ENXIO;
813	}
814	myri10ge_dummy_rdma(mgp, 1);
815	status = myri10ge_get_firmware_capabilities(mgp);
816
817	return status;
818}
819
820static int myri10ge_update_mac_address(struct myri10ge_priv *mgp, u8 * addr)
821{
822	struct myri10ge_cmd cmd;
823	int status;
824
825	cmd.data0 = ((addr[0] << 24) | (addr[1] << 16)
826		     | (addr[2] << 8) | addr[3]);
827
828	cmd.data1 = ((addr[4] << 8) | (addr[5]));
829
830	status = myri10ge_send_cmd(mgp, MXGEFW_SET_MAC_ADDRESS, &cmd, 0);
831	return status;
832}
833
834static int myri10ge_change_pause(struct myri10ge_priv *mgp, int pause)
835{
836	struct myri10ge_cmd cmd;
837	int status, ctl;
838
839	ctl = pause ? MXGEFW_ENABLE_FLOW_CONTROL : MXGEFW_DISABLE_FLOW_CONTROL;
840	status = myri10ge_send_cmd(mgp, ctl, &cmd, 0);
841
842	if (status) {
843		netdev_err(mgp->dev, "Failed to set flow control mode\n");
844		return status;
845	}
846	mgp->pause = pause;
847	return 0;
848}
849
850static void
851myri10ge_change_promisc(struct myri10ge_priv *mgp, int promisc, int atomic)
852{
853	struct myri10ge_cmd cmd;
854	int status, ctl;
855
856	ctl = promisc ? MXGEFW_ENABLE_PROMISC : MXGEFW_DISABLE_PROMISC;
857	status = myri10ge_send_cmd(mgp, ctl, &cmd, atomic);
858	if (status)
859		netdev_err(mgp->dev, "Failed to set promisc mode\n");
860}
861
862static int myri10ge_dma_test(struct myri10ge_priv *mgp, int test_type)
863{
864	struct myri10ge_cmd cmd;
865	int status;
866	u32 len;
867	struct page *dmatest_page;
868	dma_addr_t dmatest_bus;
869	char *test = " ";
870
871	dmatest_page = alloc_page(GFP_KERNEL);
872	if (!dmatest_page)
873		return -ENOMEM;
874	dmatest_bus = pci_map_page(mgp->pdev, dmatest_page, 0, PAGE_SIZE,
875				   DMA_BIDIRECTIONAL);
876	if (unlikely(pci_dma_mapping_error(mgp->pdev, dmatest_bus))) {
877		__free_page(dmatest_page);
878		return -ENOMEM;
879	}
880
881	/* Run a small DMA test.
882	 * The magic multipliers to the length tell the firmware
883	 * to do DMA read, write, or read+write tests.  The
884	 * results are returned in cmd.data0.  The upper 16
885	 * bits or the return is the number of transfers completed.
886	 * The lower 16 bits is the time in 0.5us ticks that the
887	 * transfers took to complete.
888	 */
889
890	len = mgp->tx_boundary;
891
892	cmd.data0 = MYRI10GE_LOWPART_TO_U32(dmatest_bus);
893	cmd.data1 = MYRI10GE_HIGHPART_TO_U32(dmatest_bus);
894	cmd.data2 = len * 0x10000;
895	status = myri10ge_send_cmd(mgp, test_type, &cmd, 0);
896	if (status != 0) {
897		test = "read";
898		goto abort;
899	}
900	mgp->read_dma = ((cmd.data0 >> 16) * len * 2) / (cmd.data0 & 0xffff);
901	cmd.data0 = MYRI10GE_LOWPART_TO_U32(dmatest_bus);
902	cmd.data1 = MYRI10GE_HIGHPART_TO_U32(dmatest_bus);
903	cmd.data2 = len * 0x1;
904	status = myri10ge_send_cmd(mgp, test_type, &cmd, 0);
905	if (status != 0) {
906		test = "write";
907		goto abort;
908	}
909	mgp->write_dma = ((cmd.data0 >> 16) * len * 2) / (cmd.data0 & 0xffff);
910
911	cmd.data0 = MYRI10GE_LOWPART_TO_U32(dmatest_bus);
912	cmd.data1 = MYRI10GE_HIGHPART_TO_U32(dmatest_bus);
913	cmd.data2 = len * 0x10001;
914	status = myri10ge_send_cmd(mgp, test_type, &cmd, 0);
915	if (status != 0) {
916		test = "read/write";
917		goto abort;
918	}
919	mgp->read_write_dma = ((cmd.data0 >> 16) * len * 2 * 2) /
920	    (cmd.data0 & 0xffff);
921
922abort:
923	pci_unmap_page(mgp->pdev, dmatest_bus, PAGE_SIZE, DMA_BIDIRECTIONAL);
924	put_page(dmatest_page);
925
926	if (status != 0 && test_type != MXGEFW_CMD_UNALIGNED_TEST)
927		dev_warn(&mgp->pdev->dev, "DMA %s benchmark failed: %d\n",
928			 test, status);
929
930	return status;
931}
932
933#ifdef CONFIG_NET_RX_BUSY_POLL
934static inline void myri10ge_ss_init_lock(struct myri10ge_slice_state *ss)
935{
936	spin_lock_init(&ss->lock);
937	ss->state = SLICE_STATE_IDLE;
938}
939
940static inline bool myri10ge_ss_lock_napi(struct myri10ge_slice_state *ss)
941{
942	bool rc = true;
943	spin_lock(&ss->lock);
944	if ((ss->state & SLICE_LOCKED)) {
945		WARN_ON((ss->state & SLICE_STATE_NAPI));
946		ss->state |= SLICE_STATE_NAPI_YIELD;
947		rc = false;
948		ss->lock_napi_yield++;
949	} else
950		ss->state = SLICE_STATE_NAPI;
951	spin_unlock(&ss->lock);
952	return rc;
953}
954
955static inline void myri10ge_ss_unlock_napi(struct myri10ge_slice_state *ss)
956{
957	spin_lock(&ss->lock);
958	WARN_ON((ss->state & (SLICE_STATE_POLL | SLICE_STATE_NAPI_YIELD)));
959	ss->state = SLICE_STATE_IDLE;
960	spin_unlock(&ss->lock);
961}
962
963static inline bool myri10ge_ss_lock_poll(struct myri10ge_slice_state *ss)
964{
965	bool rc = true;
966	spin_lock_bh(&ss->lock);
967	if ((ss->state & SLICE_LOCKED)) {
968		ss->state |= SLICE_STATE_POLL_YIELD;
969		rc = false;
970		ss->lock_poll_yield++;
971	} else
972		ss->state |= SLICE_STATE_POLL;
973	spin_unlock_bh(&ss->lock);
974	return rc;
975}
976
977static inline void myri10ge_ss_unlock_poll(struct myri10ge_slice_state *ss)
978{
979	spin_lock_bh(&ss->lock);
980	WARN_ON((ss->state & SLICE_STATE_NAPI));
981	ss->state = SLICE_STATE_IDLE;
982	spin_unlock_bh(&ss->lock);
983}
984
985static inline bool myri10ge_ss_busy_polling(struct myri10ge_slice_state *ss)
986{
987	WARN_ON(!(ss->state & SLICE_LOCKED));
988	return (ss->state & SLICE_USER_PEND);
989}
990#else /* CONFIG_NET_RX_BUSY_POLL */
991static inline void myri10ge_ss_init_lock(struct myri10ge_slice_state *ss)
992{
993}
994
995static inline bool myri10ge_ss_lock_napi(struct myri10ge_slice_state *ss)
996{
997	return false;
998}
999
1000static inline void myri10ge_ss_unlock_napi(struct myri10ge_slice_state *ss)
1001{
1002}
1003
1004static inline bool myri10ge_ss_lock_poll(struct myri10ge_slice_state *ss)
1005{
1006	return false;
1007}
1008
1009static inline void myri10ge_ss_unlock_poll(struct myri10ge_slice_state *ss)
1010{
1011}
1012
1013static inline bool myri10ge_ss_busy_polling(struct myri10ge_slice_state *ss)
1014{
1015	return false;
1016}
1017#endif
1018
1019static int myri10ge_reset(struct myri10ge_priv *mgp)
1020{
1021	struct myri10ge_cmd cmd;
1022	struct myri10ge_slice_state *ss;
1023	int i, status;
1024	size_t bytes;
1025#ifdef CONFIG_MYRI10GE_DCA
1026	unsigned long dca_tag_off;
1027#endif
1028
1029	/* try to send a reset command to the card to see if it
1030	 * is alive */
1031	memset(&cmd, 0, sizeof(cmd));
1032	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_RESET, &cmd, 0);
1033	if (status != 0) {
1034		dev_err(&mgp->pdev->dev, "failed reset\n");
1035		return -ENXIO;
1036	}
1037
1038	(void)myri10ge_dma_test(mgp, MXGEFW_DMA_TEST);
1039	/*
1040	 * Use non-ndis mcp_slot (eg, 4 bytes total,
1041	 * no toeplitz hash value returned.  Older firmware will
1042	 * not understand this command, but will use the correct
1043	 * sized mcp_slot, so we ignore error returns
1044	 */
1045	cmd.data0 = MXGEFW_RSS_MCP_SLOT_TYPE_MIN;
1046	(void)myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_RSS_MCP_SLOT_TYPE, &cmd, 0);
1047
1048	/* Now exchange information about interrupts  */
1049
1050	bytes = mgp->max_intr_slots * sizeof(*mgp->ss[0].rx_done.entry);
1051	cmd.data0 = (u32) bytes;
1052	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_INTRQ_SIZE, &cmd, 0);
1053
1054	/*
1055	 * Even though we already know how many slices are supported
1056	 * via myri10ge_probe_slices() MXGEFW_CMD_GET_MAX_RSS_QUEUES
1057	 * has magic side effects, and must be called after a reset.
1058	 * It must be called prior to calling any RSS related cmds,
1059	 * including assigning an interrupt queue for anything but
1060	 * slice 0.  It must also be called *after*
1061	 * MXGEFW_CMD_SET_INTRQ_SIZE, since the intrq size is used by
1062	 * the firmware to compute offsets.
1063	 */
1064
1065	if (mgp->num_slices > 1) {
1066
1067		/* ask the maximum number of slices it supports */
1068		status = myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_MAX_RSS_QUEUES,
1069					   &cmd, 0);
1070		if (status != 0) {
1071			dev_err(&mgp->pdev->dev,
1072				"failed to get number of slices\n");
1073		}
1074
1075		/*
1076		 * MXGEFW_CMD_ENABLE_RSS_QUEUES must be called prior
1077		 * to setting up the interrupt queue DMA
1078		 */
1079
1080		cmd.data0 = mgp->num_slices;
1081		cmd.data1 = MXGEFW_SLICE_INTR_MODE_ONE_PER_SLICE;
1082		if (mgp->dev->real_num_tx_queues > 1)
1083			cmd.data1 |= MXGEFW_SLICE_ENABLE_MULTIPLE_TX_QUEUES;
1084		status = myri10ge_send_cmd(mgp, MXGEFW_CMD_ENABLE_RSS_QUEUES,
1085					   &cmd, 0);
1086
1087		/* Firmware older than 1.4.32 only supports multiple
1088		 * RX queues, so if we get an error, first retry using a
1089		 * single TX queue before giving up */
1090		if (status != 0 && mgp->dev->real_num_tx_queues > 1) {
1091			netif_set_real_num_tx_queues(mgp->dev, 1);
1092			cmd.data0 = mgp->num_slices;
1093			cmd.data1 = MXGEFW_SLICE_INTR_MODE_ONE_PER_SLICE;
1094			status = myri10ge_send_cmd(mgp,
1095						   MXGEFW_CMD_ENABLE_RSS_QUEUES,
1096						   &cmd, 0);
1097		}
1098
1099		if (status != 0) {
1100			dev_err(&mgp->pdev->dev,
1101				"failed to set number of slices\n");
1102
1103			return status;
1104		}
1105	}
1106	for (i = 0; i < mgp->num_slices; i++) {
1107		ss = &mgp->ss[i];
1108		cmd.data0 = MYRI10GE_LOWPART_TO_U32(ss->rx_done.bus);
1109		cmd.data1 = MYRI10GE_HIGHPART_TO_U32(ss->rx_done.bus);
1110		cmd.data2 = i;
1111		status |= myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_INTRQ_DMA,
1112					    &cmd, 0);
1113	}
1114
1115	status |=
1116	    myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_IRQ_ACK_OFFSET, &cmd, 0);
1117	for (i = 0; i < mgp->num_slices; i++) {
1118		ss = &mgp->ss[i];
1119		ss->irq_claim =
1120		    (__iomem __be32 *) (mgp->sram + cmd.data0 + 8 * i);
1121	}
1122	status |= myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_IRQ_DEASSERT_OFFSET,
1123				    &cmd, 0);
1124	mgp->irq_deassert = (__iomem __be32 *) (mgp->sram + cmd.data0);
1125
1126	status |= myri10ge_send_cmd
1127	    (mgp, MXGEFW_CMD_GET_INTR_COAL_DELAY_OFFSET, &cmd, 0);
1128	mgp->intr_coal_delay_ptr = (__iomem __be32 *) (mgp->sram + cmd.data0);
1129	if (status != 0) {
1130		dev_err(&mgp->pdev->dev, "failed set interrupt parameters\n");
1131		return status;
1132	}
1133	put_be32(htonl(mgp->intr_coal_delay), mgp->intr_coal_delay_ptr);
1134
1135#ifdef CONFIG_MYRI10GE_DCA
1136	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_DCA_OFFSET, &cmd, 0);
1137	dca_tag_off = cmd.data0;
1138	for (i = 0; i < mgp->num_slices; i++) {
1139		ss = &mgp->ss[i];
1140		if (status == 0) {
1141			ss->dca_tag = (__iomem __be32 *)
1142			    (mgp->sram + dca_tag_off + 4 * i);
1143		} else {
1144			ss->dca_tag = NULL;
1145		}
1146	}
1147#endif				/* CONFIG_MYRI10GE_DCA */
1148
1149	/* reset mcp/driver shared state back to 0 */
1150
1151	mgp->link_changes = 0;
1152	for (i = 0; i < mgp->num_slices; i++) {
1153		ss = &mgp->ss[i];
1154
1155		memset(ss->rx_done.entry, 0, bytes);
1156		ss->tx.req = 0;
1157		ss->tx.done = 0;
1158		ss->tx.pkt_start = 0;
1159		ss->tx.pkt_done = 0;
1160		ss->rx_big.cnt = 0;
1161		ss->rx_small.cnt = 0;
1162		ss->rx_done.idx = 0;
1163		ss->rx_done.cnt = 0;
1164		ss->tx.wake_queue = 0;
1165		ss->tx.stop_queue = 0;
1166	}
1167
1168	status = myri10ge_update_mac_address(mgp, mgp->dev->dev_addr);
1169	myri10ge_change_pause(mgp, mgp->pause);
1170	myri10ge_set_multicast_list(mgp->dev);
1171	return status;
1172}
1173
1174#ifdef CONFIG_MYRI10GE_DCA
1175static int myri10ge_toggle_relaxed(struct pci_dev *pdev, int on)
1176{
1177	int ret;
1178	u16 ctl;
1179
1180	pcie_capability_read_word(pdev, PCI_EXP_DEVCTL, &ctl);
1181
1182	ret = (ctl & PCI_EXP_DEVCTL_RELAX_EN) >> 4;
1183	if (ret != on) {
1184		ctl &= ~PCI_EXP_DEVCTL_RELAX_EN;
1185		ctl |= (on << 4);
1186		pcie_capability_write_word(pdev, PCI_EXP_DEVCTL, ctl);
1187	}
1188	return ret;
1189}
1190
1191static void
1192myri10ge_write_dca(struct myri10ge_slice_state *ss, int cpu, int tag)
1193{
1194	ss->cached_dca_tag = tag;
1195	put_be32(htonl(tag), ss->dca_tag);
1196}
1197
1198static inline void myri10ge_update_dca(struct myri10ge_slice_state *ss)
1199{
1200	int cpu = get_cpu();
1201	int tag;
1202
1203	if (cpu != ss->cpu) {
1204		tag = dca3_get_tag(&ss->mgp->pdev->dev, cpu);
1205		if (ss->cached_dca_tag != tag)
1206			myri10ge_write_dca(ss, cpu, tag);
1207		ss->cpu = cpu;
1208	}
1209	put_cpu();
1210}
1211
1212static void myri10ge_setup_dca(struct myri10ge_priv *mgp)
1213{
1214	int err, i;
1215	struct pci_dev *pdev = mgp->pdev;
1216
1217	if (mgp->ss[0].dca_tag == NULL || mgp->dca_enabled)
1218		return;
1219	if (!myri10ge_dca) {
1220		dev_err(&pdev->dev, "dca disabled by administrator\n");
1221		return;
1222	}
1223	err = dca_add_requester(&pdev->dev);
1224	if (err) {
1225		if (err != -ENODEV)
1226			dev_err(&pdev->dev,
1227				"dca_add_requester() failed, err=%d\n", err);
1228		return;
1229	}
1230	mgp->relaxed_order = myri10ge_toggle_relaxed(pdev, 0);
1231	mgp->dca_enabled = 1;
1232	for (i = 0; i < mgp->num_slices; i++) {
1233		mgp->ss[i].cpu = -1;
1234		mgp->ss[i].cached_dca_tag = -1;
1235		myri10ge_update_dca(&mgp->ss[i]);
1236	}
1237}
1238
1239static void myri10ge_teardown_dca(struct myri10ge_priv *mgp)
1240{
1241	struct pci_dev *pdev = mgp->pdev;
1242
1243	if (!mgp->dca_enabled)
1244		return;
1245	mgp->dca_enabled = 0;
1246	if (mgp->relaxed_order)
1247		myri10ge_toggle_relaxed(pdev, 1);
1248	dca_remove_requester(&pdev->dev);
1249}
1250
1251static int myri10ge_notify_dca_device(struct device *dev, void *data)
1252{
1253	struct myri10ge_priv *mgp;
1254	unsigned long event;
1255
1256	mgp = dev_get_drvdata(dev);
1257	event = *(unsigned long *)data;
1258
1259	if (event == DCA_PROVIDER_ADD)
1260		myri10ge_setup_dca(mgp);
1261	else if (event == DCA_PROVIDER_REMOVE)
1262		myri10ge_teardown_dca(mgp);
1263	return 0;
1264}
1265#endif				/* CONFIG_MYRI10GE_DCA */
1266
1267static inline void
1268myri10ge_submit_8rx(struct mcp_kreq_ether_recv __iomem * dst,
1269		    struct mcp_kreq_ether_recv *src)
1270{
1271	__be32 low;
1272
1273	low = src->addr_low;
1274	src->addr_low = htonl(DMA_BIT_MASK(32));
1275	myri10ge_pio_copy(dst, src, 4 * sizeof(*src));
1276	mb();
1277	myri10ge_pio_copy(dst + 4, src + 4, 4 * sizeof(*src));
1278	mb();
1279	src->addr_low = low;
1280	put_be32(low, &dst->addr_low);
1281	mb();
1282}
1283
1284static inline void myri10ge_vlan_ip_csum(struct sk_buff *skb, __wsum hw_csum)
1285{
1286	struct vlan_hdr *vh = (struct vlan_hdr *)(skb->data);
1287
1288	if ((skb->protocol == htons(ETH_P_8021Q)) &&
1289	    (vh->h_vlan_encapsulated_proto == htons(ETH_P_IP) ||
1290	     vh->h_vlan_encapsulated_proto == htons(ETH_P_IPV6))) {
1291		skb->csum = hw_csum;
1292		skb->ip_summed = CHECKSUM_COMPLETE;
1293	}
1294}
1295
1296static void
1297myri10ge_alloc_rx_pages(struct myri10ge_priv *mgp, struct myri10ge_rx_buf *rx,
1298			int bytes, int watchdog)
1299{
1300	struct page *page;
1301	dma_addr_t bus;
1302	int idx;
1303#if MYRI10GE_ALLOC_SIZE > 4096
1304	int end_offset;
1305#endif
1306
1307	if (unlikely(rx->watchdog_needed && !watchdog))
1308		return;
1309
1310	/* try to refill entire ring */
1311	while (rx->fill_cnt != (rx->cnt + rx->mask + 1)) {
1312		idx = rx->fill_cnt & rx->mask;
1313		if (rx->page_offset + bytes <= MYRI10GE_ALLOC_SIZE) {
1314			/* we can use part of previous page */
1315			get_page(rx->page);
1316		} else {
1317			/* we need a new page */
1318			page =
1319			    alloc_pages(GFP_ATOMIC | __GFP_COMP,
1320					MYRI10GE_ALLOC_ORDER);
1321			if (unlikely(page == NULL)) {
1322				if (rx->fill_cnt - rx->cnt < 16)
1323					rx->watchdog_needed = 1;
1324				return;
1325			}
1326
1327			bus = pci_map_page(mgp->pdev, page, 0,
1328					   MYRI10GE_ALLOC_SIZE,
1329					   PCI_DMA_FROMDEVICE);
1330			if (unlikely(pci_dma_mapping_error(mgp->pdev, bus))) {
1331				__free_pages(page, MYRI10GE_ALLOC_ORDER);
1332				if (rx->fill_cnt - rx->cnt < 16)
1333					rx->watchdog_needed = 1;
1334				return;
1335			}
1336
1337			rx->page = page;
1338			rx->page_offset = 0;
1339			rx->bus = bus;
1340
1341		}
1342		rx->info[idx].page = rx->page;
1343		rx->info[idx].page_offset = rx->page_offset;
1344		/* note that this is the address of the start of the
1345		 * page */
1346		dma_unmap_addr_set(&rx->info[idx], bus, rx->bus);
1347		rx->shadow[idx].addr_low =
1348		    htonl(MYRI10GE_LOWPART_TO_U32(rx->bus) + rx->page_offset);
1349		rx->shadow[idx].addr_high =
1350		    htonl(MYRI10GE_HIGHPART_TO_U32(rx->bus));
1351
1352		/* start next packet on a cacheline boundary */
1353		rx->page_offset += SKB_DATA_ALIGN(bytes);
1354
1355#if MYRI10GE_ALLOC_SIZE > 4096
1356		/* don't cross a 4KB boundary */
1357		end_offset = rx->page_offset + bytes - 1;
1358		if ((unsigned)(rx->page_offset ^ end_offset) > 4095)
1359			rx->page_offset = end_offset & ~4095;
1360#endif
1361		rx->fill_cnt++;
1362
1363		/* copy 8 descriptors to the firmware at a time */
1364		if ((idx & 7) == 7) {
1365			myri10ge_submit_8rx(&rx->lanai[idx - 7],
1366					    &rx->shadow[idx - 7]);
1367		}
1368	}
1369}
1370
1371static inline void
1372myri10ge_unmap_rx_page(struct pci_dev *pdev,
1373		       struct myri10ge_rx_buffer_state *info, int bytes)
1374{
1375	/* unmap the recvd page if we're the only or last user of it */
1376	if (bytes >= MYRI10GE_ALLOC_SIZE / 2 ||
1377	    (info->page_offset + 2 * bytes) > MYRI10GE_ALLOC_SIZE) {
1378		pci_unmap_page(pdev, (dma_unmap_addr(info, bus)
1379				      & ~(MYRI10GE_ALLOC_SIZE - 1)),
1380			       MYRI10GE_ALLOC_SIZE, PCI_DMA_FROMDEVICE);
1381	}
1382}
1383
1384/*
1385 * GRO does not support acceleration of tagged vlan frames, and
1386 * this NIC does not support vlan tag offload, so we must pop
1387 * the tag ourselves to be able to achieve GRO performance that
1388 * is comparable to LRO.
1389 */
1390
1391static inline void
1392myri10ge_vlan_rx(struct net_device *dev, void *addr, struct sk_buff *skb)
1393{
1394	u8 *va;
1395	struct vlan_ethhdr *veh;
1396	struct skb_frag_struct *frag;
1397	__wsum vsum;
1398
1399	va = addr;
1400	va += MXGEFW_PAD;
1401	veh = (struct vlan_ethhdr *)va;
1402	if ((dev->features & NETIF_F_HW_VLAN_CTAG_RX) ==
1403	    NETIF_F_HW_VLAN_CTAG_RX &&
1404	    veh->h_vlan_proto == htons(ETH_P_8021Q)) {
1405		/* fixup csum if needed */
1406		if (skb->ip_summed == CHECKSUM_COMPLETE) {
1407			vsum = csum_partial(va + ETH_HLEN, VLAN_HLEN, 0);
1408			skb->csum = csum_sub(skb->csum, vsum);
1409		}
1410		/* pop tag */
1411		__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), ntohs(veh->h_vlan_TCI));
1412		memmove(va + VLAN_HLEN, va, 2 * ETH_ALEN);
1413		skb->len -= VLAN_HLEN;
1414		skb->data_len -= VLAN_HLEN;
1415		frag = skb_shinfo(skb)->frags;
1416		frag->page_offset += VLAN_HLEN;
1417		skb_frag_size_set(frag, skb_frag_size(frag) - VLAN_HLEN);
1418	}
1419}
1420
1421#define MYRI10GE_HLEN 64 /* Bytes to copy from page to skb linear memory */
1422
1423static inline int
1424myri10ge_rx_done(struct myri10ge_slice_state *ss, int len, __wsum csum)
1425{
1426	struct myri10ge_priv *mgp = ss->mgp;
1427	struct sk_buff *skb;
1428	struct skb_frag_struct *rx_frags;
1429	struct myri10ge_rx_buf *rx;
1430	int i, idx, remainder, bytes;
1431	struct pci_dev *pdev = mgp->pdev;
1432	struct net_device *dev = mgp->dev;
1433	u8 *va;
1434	bool polling;
1435
1436	if (len <= mgp->small_bytes) {
1437		rx = &ss->rx_small;
1438		bytes = mgp->small_bytes;
1439	} else {
1440		rx = &ss->rx_big;
1441		bytes = mgp->big_bytes;
1442	}
1443
1444	len += MXGEFW_PAD;
1445	idx = rx->cnt & rx->mask;
1446	va = page_address(rx->info[idx].page) + rx->info[idx].page_offset;
1447	prefetch(va);
1448
1449	/* When busy polling in user context, allocate skb and copy headers to
1450	 * skb's linear memory ourselves.  When not busy polling, use the napi
1451	 * gro api.
1452	 */
1453	polling = myri10ge_ss_busy_polling(ss);
1454	if (polling)
1455		skb = netdev_alloc_skb(dev, MYRI10GE_HLEN + 16);
1456	else
1457		skb = napi_get_frags(&ss->napi);
1458	if (unlikely(skb == NULL)) {
1459		ss->stats.rx_dropped++;
1460		for (i = 0, remainder = len; remainder > 0; i++) {
1461			myri10ge_unmap_rx_page(pdev, &rx->info[idx], bytes);
1462			put_page(rx->info[idx].page);
1463			rx->cnt++;
1464			idx = rx->cnt & rx->mask;
1465			remainder -= MYRI10GE_ALLOC_SIZE;
1466		}
1467		return 0;
1468	}
1469	rx_frags = skb_shinfo(skb)->frags;
1470	/* Fill skb_frag_struct(s) with data from our receive */
1471	for (i = 0, remainder = len; remainder > 0; i++) {
1472		myri10ge_unmap_rx_page(pdev, &rx->info[idx], bytes);
1473		skb_fill_page_desc(skb, i, rx->info[idx].page,
1474				   rx->info[idx].page_offset,
1475				   remainder < MYRI10GE_ALLOC_SIZE ?
1476				   remainder : MYRI10GE_ALLOC_SIZE);
1477		rx->cnt++;
1478		idx = rx->cnt & rx->mask;
1479		remainder -= MYRI10GE_ALLOC_SIZE;
1480	}
1481
1482	/* remove padding */
1483	rx_frags[0].page_offset += MXGEFW_PAD;
1484	rx_frags[0].size -= MXGEFW_PAD;
1485	len -= MXGEFW_PAD;
1486
1487	skb->len = len;
1488	skb->data_len = len;
1489	skb->truesize += len;
1490	if (dev->features & NETIF_F_RXCSUM) {
1491		skb->ip_summed = CHECKSUM_COMPLETE;
1492		skb->csum = csum;
1493	}
1494	myri10ge_vlan_rx(mgp->dev, va, skb);
1495	skb_record_rx_queue(skb, ss - &mgp->ss[0]);
1496	skb_mark_napi_id(skb, &ss->napi);
1497
1498	if (polling) {
1499		int hlen;
1500
1501		/* myri10ge_vlan_rx might have moved the header, so compute
1502		 * length and address again.
1503		 */
1504		hlen = MYRI10GE_HLEN > skb->len ? skb->len : MYRI10GE_HLEN;
1505		va = page_address(skb_frag_page(&rx_frags[0])) +
1506			rx_frags[0].page_offset;
1507		/* Copy header into the skb linear memory */
1508		skb_copy_to_linear_data(skb, va, hlen);
1509		rx_frags[0].page_offset += hlen;
1510		rx_frags[0].size -= hlen;
1511		skb->data_len -= hlen;
1512		skb->tail += hlen;
1513		skb->protocol = eth_type_trans(skb, dev);
1514		netif_receive_skb(skb);
1515	}
1516	else
1517		napi_gro_frags(&ss->napi);
1518
1519	return 1;
1520}
1521
1522static inline void
1523myri10ge_tx_done(struct myri10ge_slice_state *ss, int mcp_index)
1524{
1525	struct pci_dev *pdev = ss->mgp->pdev;
1526	struct myri10ge_tx_buf *tx = &ss->tx;
1527	struct netdev_queue *dev_queue;
1528	struct sk_buff *skb;
1529	int idx, len;
1530
1531	while (tx->pkt_done != mcp_index) {
1532		idx = tx->done & tx->mask;
1533		skb = tx->info[idx].skb;
1534
1535		/* Mark as free */
1536		tx->info[idx].skb = NULL;
1537		if (tx->info[idx].last) {
1538			tx->pkt_done++;
1539			tx->info[idx].last = 0;
1540		}
1541		tx->done++;
1542		len = dma_unmap_len(&tx->info[idx], len);
1543		dma_unmap_len_set(&tx->info[idx], len, 0);
1544		if (skb) {
1545			ss->stats.tx_bytes += skb->len;
1546			ss->stats.tx_packets++;
1547			dev_kfree_skb_irq(skb);
1548			if (len)
1549				pci_unmap_single(pdev,
1550						 dma_unmap_addr(&tx->info[idx],
1551								bus), len,
1552						 PCI_DMA_TODEVICE);
1553		} else {
1554			if (len)
1555				pci_unmap_page(pdev,
1556					       dma_unmap_addr(&tx->info[idx],
1557							      bus), len,
1558					       PCI_DMA_TODEVICE);
1559		}
1560	}
1561
1562	dev_queue = netdev_get_tx_queue(ss->dev, ss - ss->mgp->ss);
1563	/*
1564	 * Make a minimal effort to prevent the NIC from polling an
1565	 * idle tx queue.  If we can't get the lock we leave the queue
1566	 * active. In this case, either a thread was about to start
1567	 * using the queue anyway, or we lost a race and the NIC will
1568	 * waste some of its resources polling an inactive queue for a
1569	 * while.
1570	 */
1571
1572	if ((ss->mgp->dev->real_num_tx_queues > 1) &&
1573	    __netif_tx_trylock(dev_queue)) {
1574		if (tx->req == tx->done) {
1575			tx->queue_active = 0;
1576			put_be32(htonl(1), tx->send_stop);
1577			mb();
1578			mmiowb();
1579		}
1580		__netif_tx_unlock(dev_queue);
1581	}
1582
1583	/* start the queue if we've stopped it */
1584	if (netif_tx_queue_stopped(dev_queue) &&
1585	    tx->req - tx->done < (tx->mask >> 1) &&
1586	    ss->mgp->running == MYRI10GE_ETH_RUNNING) {
1587		tx->wake_queue++;
1588		netif_tx_wake_queue(dev_queue);
1589	}
1590}
1591
1592static inline int
1593myri10ge_clean_rx_done(struct myri10ge_slice_state *ss, int budget)
1594{
1595	struct myri10ge_rx_done *rx_done = &ss->rx_done;
1596	struct myri10ge_priv *mgp = ss->mgp;
1597	unsigned long rx_bytes = 0;
1598	unsigned long rx_packets = 0;
1599	unsigned long rx_ok;
1600	int idx = rx_done->idx;
1601	int cnt = rx_done->cnt;
1602	int work_done = 0;
1603	u16 length;
1604	__wsum checksum;
1605
1606	while (rx_done->entry[idx].length != 0 && work_done < budget) {
1607		length = ntohs(rx_done->entry[idx].length);
1608		rx_done->entry[idx].length = 0;
1609		checksum = csum_unfold(rx_done->entry[idx].checksum);
1610		rx_ok = myri10ge_rx_done(ss, length, checksum);
1611		rx_packets += rx_ok;
1612		rx_bytes += rx_ok * (unsigned long)length;
1613		cnt++;
1614		idx = cnt & (mgp->max_intr_slots - 1);
1615		work_done++;
1616	}
1617	rx_done->idx = idx;
1618	rx_done->cnt = cnt;
1619	ss->stats.rx_packets += rx_packets;
1620	ss->stats.rx_bytes += rx_bytes;
1621
1622	/* restock receive rings if needed */
1623	if (ss->rx_small.fill_cnt - ss->rx_small.cnt < myri10ge_fill_thresh)
1624		myri10ge_alloc_rx_pages(mgp, &ss->rx_small,
1625					mgp->small_bytes + MXGEFW_PAD, 0);
1626	if (ss->rx_big.fill_cnt - ss->rx_big.cnt < myri10ge_fill_thresh)
1627		myri10ge_alloc_rx_pages(mgp, &ss->rx_big, mgp->big_bytes, 0);
1628
1629	return work_done;
1630}
1631
1632static inline void myri10ge_check_statblock(struct myri10ge_priv *mgp)
1633{
1634	struct mcp_irq_data *stats = mgp->ss[0].fw_stats;
1635
1636	if (unlikely(stats->stats_updated)) {
1637		unsigned link_up = ntohl(stats->link_up);
1638		if (mgp->link_state != link_up) {
1639			mgp->link_state = link_up;
1640
1641			if (mgp->link_state == MXGEFW_LINK_UP) {
1642				netif_info(mgp, link, mgp->dev, "link up\n");
1643				netif_carrier_on(mgp->dev);
1644				mgp->link_changes++;
1645			} else {
1646				netif_info(mgp, link, mgp->dev, "link %s\n",
1647					   (link_up == MXGEFW_LINK_MYRINET ?
1648					    "mismatch (Myrinet detected)" :
1649					    "down"));
1650				netif_carrier_off(mgp->dev);
1651				mgp->link_changes++;
1652			}
1653		}
1654		if (mgp->rdma_tags_available !=
1655		    ntohl(stats->rdma_tags_available)) {
1656			mgp->rdma_tags_available =
1657			    ntohl(stats->rdma_tags_available);
1658			netdev_warn(mgp->dev, "RDMA timed out! %d tags left\n",
1659				    mgp->rdma_tags_available);
1660		}
1661		mgp->down_cnt += stats->link_down;
1662		if (stats->link_down)
1663			wake_up(&mgp->down_wq);
1664	}
1665}
1666
1667static int myri10ge_poll(struct napi_struct *napi, int budget)
1668{
1669	struct myri10ge_slice_state *ss =
1670	    container_of(napi, struct myri10ge_slice_state, napi);
1671	int work_done;
1672
1673#ifdef CONFIG_MYRI10GE_DCA
1674	if (ss->mgp->dca_enabled)
1675		myri10ge_update_dca(ss);
1676#endif
1677	/* Try later if the busy_poll handler is running. */
1678	if (!myri10ge_ss_lock_napi(ss))
1679		return budget;
1680
1681	/* process as many rx events as NAPI will allow */
1682	work_done = myri10ge_clean_rx_done(ss, budget);
1683
1684	myri10ge_ss_unlock_napi(ss);
1685	if (work_done < budget) {
1686		napi_complete(napi);
1687		put_be32(htonl(3), ss->irq_claim);
1688	}
1689	return work_done;
1690}
1691
1692#ifdef CONFIG_NET_RX_BUSY_POLL
1693static int myri10ge_busy_poll(struct napi_struct *napi)
1694{
1695	struct myri10ge_slice_state *ss =
1696	    container_of(napi, struct myri10ge_slice_state, napi);
1697	struct myri10ge_priv *mgp = ss->mgp;
1698	int work_done;
1699
1700	/* Poll only when the link is up */
1701	if (mgp->link_state != MXGEFW_LINK_UP)
1702		return LL_FLUSH_FAILED;
1703
1704	if (!myri10ge_ss_lock_poll(ss))
1705		return LL_FLUSH_BUSY;
1706
1707	/* Process a small number of packets */
1708	work_done = myri10ge_clean_rx_done(ss, 4);
1709	if (work_done)
1710		ss->busy_poll_cnt += work_done;
1711	else
1712		ss->busy_poll_miss++;
1713
1714	myri10ge_ss_unlock_poll(ss);
1715
1716	return work_done;
1717}
1718#endif /* CONFIG_NET_RX_BUSY_POLL */
1719
1720static irqreturn_t myri10ge_intr(int irq, void *arg)
1721{
1722	struct myri10ge_slice_state *ss = arg;
1723	struct myri10ge_priv *mgp = ss->mgp;
1724	struct mcp_irq_data *stats = ss->fw_stats;
1725	struct myri10ge_tx_buf *tx = &ss->tx;
1726	u32 send_done_count;
1727	int i;
1728
1729	/* an interrupt on a non-zero receive-only slice is implicitly
1730	 * valid  since MSI-X irqs are not shared */
1731	if ((mgp->dev->real_num_tx_queues == 1) && (ss != mgp->ss)) {
1732		napi_schedule(&ss->napi);
1733		return IRQ_HANDLED;
1734	}
1735
1736	/* make sure it is our IRQ, and that the DMA has finished */
1737	if (unlikely(!stats->valid))
1738		return IRQ_NONE;
1739
1740	/* low bit indicates receives are present, so schedule
1741	 * napi poll handler */
1742	if (stats->valid & 1)
1743		napi_schedule(&ss->napi);
1744
1745	if (!mgp->msi_enabled && !mgp->msix_enabled) {
1746		put_be32(0, mgp->irq_deassert);
1747		if (!myri10ge_deassert_wait)
1748			stats->valid = 0;
1749		mb();
1750	} else
1751		stats->valid = 0;
1752
1753	/* Wait for IRQ line to go low, if using INTx */
1754	i = 0;
1755	while (1) {
1756		i++;
1757		/* check for transmit completes and receives */
1758		send_done_count = ntohl(stats->send_done_count);
1759		if (send_done_count != tx->pkt_done)
1760			myri10ge_tx_done(ss, (int)send_done_count);
1761		if (unlikely(i > myri10ge_max_irq_loops)) {
1762			netdev_warn(mgp->dev, "irq stuck?\n");
1763			stats->valid = 0;
1764			schedule_work(&mgp->watchdog_work);
1765		}
1766		if (likely(stats->valid == 0))
1767			break;
1768		cpu_relax();
1769		barrier();
1770	}
1771
1772	/* Only slice 0 updates stats */
1773	if (ss == mgp->ss)
1774		myri10ge_check_statblock(mgp);
1775
1776	put_be32(htonl(3), ss->irq_claim + 1);
1777	return IRQ_HANDLED;
1778}
1779
1780static int
1781myri10ge_get_settings(struct net_device *netdev, struct ethtool_cmd *cmd)
1782{
1783	struct myri10ge_priv *mgp = netdev_priv(netdev);
1784	char *ptr;
1785	int i;
1786
1787	cmd->autoneg = AUTONEG_DISABLE;
1788	ethtool_cmd_speed_set(cmd, SPEED_10000);
1789	cmd->duplex = DUPLEX_FULL;
1790
1791	/*
1792	 * parse the product code to deterimine the interface type
1793	 * (CX4, XFP, Quad Ribbon Fiber) by looking at the character
1794	 * after the 3rd dash in the driver's cached copy of the
1795	 * EEPROM's product code string.
1796	 */
1797	ptr = mgp->product_code_string;
1798	if (ptr == NULL) {
1799		netdev_err(netdev, "Missing product code\n");
1800		return 0;
1801	}
1802	for (i = 0; i < 3; i++, ptr++) {
1803		ptr = strchr(ptr, '-');
1804		if (ptr == NULL) {
1805			netdev_err(netdev, "Invalid product code %s\n",
1806				   mgp->product_code_string);
1807			return 0;
1808		}
1809	}
1810	if (*ptr == '2')
1811		ptr++;
1812	if (*ptr == 'R' || *ptr == 'Q' || *ptr == 'S') {
1813		/* We've found either an XFP, quad ribbon fiber, or SFP+ */
1814		cmd->port = PORT_FIBRE;
1815		cmd->supported |= SUPPORTED_FIBRE;
1816		cmd->advertising |= ADVERTISED_FIBRE;
1817	} else {
1818		cmd->port = PORT_OTHER;
1819	}
1820	if (*ptr == 'R' || *ptr == 'S')
1821		cmd->transceiver = XCVR_EXTERNAL;
1822	else
1823		cmd->transceiver = XCVR_INTERNAL;
1824
1825	return 0;
1826}
1827
1828static void
1829myri10ge_get_drvinfo(struct net_device *netdev, struct ethtool_drvinfo *info)
1830{
1831	struct myri10ge_priv *mgp = netdev_priv(netdev);
1832
1833	strlcpy(info->driver, "myri10ge", sizeof(info->driver));
1834	strlcpy(info->version, MYRI10GE_VERSION_STR, sizeof(info->version));
1835	strlcpy(info->fw_version, mgp->fw_version, sizeof(info->fw_version));
1836	strlcpy(info->bus_info, pci_name(mgp->pdev), sizeof(info->bus_info));
1837}
1838
1839static int
1840myri10ge_get_coalesce(struct net_device *netdev, struct ethtool_coalesce *coal)
1841{
1842	struct myri10ge_priv *mgp = netdev_priv(netdev);
1843
1844	coal->rx_coalesce_usecs = mgp->intr_coal_delay;
1845	return 0;
1846}
1847
1848static int
1849myri10ge_set_coalesce(struct net_device *netdev, struct ethtool_coalesce *coal)
1850{
1851	struct myri10ge_priv *mgp = netdev_priv(netdev);
1852
1853	mgp->intr_coal_delay = coal->rx_coalesce_usecs;
1854	put_be32(htonl(mgp->intr_coal_delay), mgp->intr_coal_delay_ptr);
1855	return 0;
1856}
1857
1858static void
1859myri10ge_get_pauseparam(struct net_device *netdev,
1860			struct ethtool_pauseparam *pause)
1861{
1862	struct myri10ge_priv *mgp = netdev_priv(netdev);
1863
1864	pause->autoneg = 0;
1865	pause->rx_pause = mgp->pause;
1866	pause->tx_pause = mgp->pause;
1867}
1868
1869static int
1870myri10ge_set_pauseparam(struct net_device *netdev,
1871			struct ethtool_pauseparam *pause)
1872{
1873	struct myri10ge_priv *mgp = netdev_priv(netdev);
1874
1875	if (pause->tx_pause != mgp->pause)
1876		return myri10ge_change_pause(mgp, pause->tx_pause);
1877	if (pause->rx_pause != mgp->pause)
1878		return myri10ge_change_pause(mgp, pause->rx_pause);
1879	if (pause->autoneg != 0)
1880		return -EINVAL;
1881	return 0;
1882}
1883
1884static void
1885myri10ge_get_ringparam(struct net_device *netdev,
1886		       struct ethtool_ringparam *ring)
1887{
1888	struct myri10ge_priv *mgp = netdev_priv(netdev);
1889
1890	ring->rx_mini_max_pending = mgp->ss[0].rx_small.mask + 1;
1891	ring->rx_max_pending = mgp->ss[0].rx_big.mask + 1;
1892	ring->rx_jumbo_max_pending = 0;
1893	ring->tx_max_pending = mgp->ss[0].tx.mask + 1;
1894	ring->rx_mini_pending = ring->rx_mini_max_pending;
1895	ring->rx_pending = ring->rx_max_pending;
1896	ring->rx_jumbo_pending = ring->rx_jumbo_max_pending;
1897	ring->tx_pending = ring->tx_max_pending;
1898}
1899
1900static const char myri10ge_gstrings_main_stats[][ETH_GSTRING_LEN] = {
1901	"rx_packets", "tx_packets", "rx_bytes", "tx_bytes", "rx_errors",
1902	"tx_errors", "rx_dropped", "tx_dropped", "multicast", "collisions",
1903	"rx_length_errors", "rx_over_errors", "rx_crc_errors",
1904	"rx_frame_errors", "rx_fifo_errors", "rx_missed_errors",
1905	"tx_aborted_errors", "tx_carrier_errors", "tx_fifo_errors",
1906	"tx_heartbeat_errors", "tx_window_errors",
1907	/* device-specific stats */
1908	"tx_boundary", "WC", "irq", "MSI", "MSIX",
1909	"read_dma_bw_MBs", "write_dma_bw_MBs", "read_write_dma_bw_MBs",
1910	"serial_number", "watchdog_resets",
1911#ifdef CONFIG_MYRI10GE_DCA
1912	"dca_capable_firmware", "dca_device_present",
1913#endif
1914	"link_changes", "link_up", "dropped_link_overflow",
1915	"dropped_link_error_or_filtered",
1916	"dropped_pause", "dropped_bad_phy", "dropped_bad_crc32",
1917	"dropped_unicast_filtered", "dropped_multicast_filtered",
1918	"dropped_runt", "dropped_overrun", "dropped_no_small_buffer",
1919	"dropped_no_big_buffer"
1920};
1921
1922static const char myri10ge_gstrings_slice_stats[][ETH_GSTRING_LEN] = {
1923	"----------- slice ---------",
1924	"tx_pkt_start", "tx_pkt_done", "tx_req", "tx_done",
1925	"rx_small_cnt", "rx_big_cnt",
1926	"wake_queue", "stop_queue", "tx_linearized",
1927#ifdef CONFIG_NET_RX_BUSY_POLL
1928	"rx_lock_napi_yield", "rx_lock_poll_yield", "rx_busy_poll_miss",
1929	"rx_busy_poll_cnt",
1930#endif
1931};
1932
1933#define MYRI10GE_NET_STATS_LEN      21
1934#define MYRI10GE_MAIN_STATS_LEN  ARRAY_SIZE(myri10ge_gstrings_main_stats)
1935#define MYRI10GE_SLICE_STATS_LEN  ARRAY_SIZE(myri10ge_gstrings_slice_stats)
1936
1937static void
1938myri10ge_get_strings(struct net_device *netdev, u32 stringset, u8 * data)
1939{
1940	struct myri10ge_priv *mgp = netdev_priv(netdev);
1941	int i;
1942
1943	switch (stringset) {
1944	case ETH_SS_STATS:
1945		memcpy(data, *myri10ge_gstrings_main_stats,
1946		       sizeof(myri10ge_gstrings_main_stats));
1947		data += sizeof(myri10ge_gstrings_main_stats);
1948		for (i = 0; i < mgp->num_slices; i++) {
1949			memcpy(data, *myri10ge_gstrings_slice_stats,
1950			       sizeof(myri10ge_gstrings_slice_stats));
1951			data += sizeof(myri10ge_gstrings_slice_stats);
1952		}
1953		break;
1954	}
1955}
1956
1957static int myri10ge_get_sset_count(struct net_device *netdev, int sset)
1958{
1959	struct myri10ge_priv *mgp = netdev_priv(netdev);
1960
1961	switch (sset) {
1962	case ETH_SS_STATS:
1963		return MYRI10GE_MAIN_STATS_LEN +
1964		    mgp->num_slices * MYRI10GE_SLICE_STATS_LEN;
1965	default:
1966		return -EOPNOTSUPP;
1967	}
1968}
1969
1970static void
1971myri10ge_get_ethtool_stats(struct net_device *netdev,
1972			   struct ethtool_stats *stats, u64 * data)
1973{
1974	struct myri10ge_priv *mgp = netdev_priv(netdev);
1975	struct myri10ge_slice_state *ss;
1976	struct rtnl_link_stats64 link_stats;
1977	int slice;
1978	int i;
1979
1980	/* force stats update */
1981	memset(&link_stats, 0, sizeof(link_stats));
1982	(void)myri10ge_get_stats(netdev, &link_stats);
1983	for (i = 0; i < MYRI10GE_NET_STATS_LEN; i++)
1984		data[i] = ((u64 *)&link_stats)[i];
1985
1986	data[i++] = (unsigned int)mgp->tx_boundary;
1987	data[i++] = (unsigned int)mgp->wc_enabled;
1988	data[i++] = (unsigned int)mgp->pdev->irq;
1989	data[i++] = (unsigned int)mgp->msi_enabled;
1990	data[i++] = (unsigned int)mgp->msix_enabled;
1991	data[i++] = (unsigned int)mgp->read_dma;
1992	data[i++] = (unsigned int)mgp->write_dma;
1993	data[i++] = (unsigned int)mgp->read_write_dma;
1994	data[i++] = (unsigned int)mgp->serial_number;
1995	data[i++] = (unsigned int)mgp->watchdog_resets;
1996#ifdef CONFIG_MYRI10GE_DCA
1997	data[i++] = (unsigned int)(mgp->ss[0].dca_tag != NULL);
1998	data[i++] = (unsigned int)(mgp->dca_enabled);
1999#endif
2000	data[i++] = (unsigned int)mgp->link_changes;
2001
2002	/* firmware stats are useful only in the first slice */
2003	ss = &mgp->ss[0];
2004	data[i++] = (unsigned int)ntohl(ss->fw_stats->link_up);
2005	data[i++] = (unsigned int)ntohl(ss->fw_stats->dropped_link_overflow);
2006	data[i++] =
2007	    (unsigned int)ntohl(ss->fw_stats->dropped_link_error_or_filtered);
2008	data[i++] = (unsigned int)ntohl(ss->fw_stats->dropped_pause);
2009	data[i++] = (unsigned int)ntohl(ss->fw_stats->dropped_bad_phy);
2010	data[i++] = (unsigned int)ntohl(ss->fw_stats->dropped_bad_crc32);
2011	data[i++] = (unsigned int)ntohl(ss->fw_stats->dropped_unicast_filtered);
2012	data[i++] =
2013	    (unsigned int)ntohl(ss->fw_stats->dropped_multicast_filtered);
2014	data[i++] = (unsigned int)ntohl(ss->fw_stats->dropped_runt);
2015	data[i++] = (unsigned int)ntohl(ss->fw_stats->dropped_overrun);
2016	data[i++] = (unsigned int)ntohl(ss->fw_stats->dropped_no_small_buffer);
2017	data[i++] = (unsigned int)ntohl(ss->fw_stats->dropped_no_big_buffer);
2018
2019	for (slice = 0; slice < mgp->num_slices; slice++) {
2020		ss = &mgp->ss[slice];
2021		data[i++] = slice;
2022		data[i++] = (unsigned int)ss->tx.pkt_start;
2023		data[i++] = (unsigned int)ss->tx.pkt_done;
2024		data[i++] = (unsigned int)ss->tx.req;
2025		data[i++] = (unsigned int)ss->tx.done;
2026		data[i++] = (unsigned int)ss->rx_small.cnt;
2027		data[i++] = (unsigned int)ss->rx_big.cnt;
2028		data[i++] = (unsigned int)ss->tx.wake_queue;
2029		data[i++] = (unsigned int)ss->tx.stop_queue;
2030		data[i++] = (unsigned int)ss->tx.linearized;
2031#ifdef CONFIG_NET_RX_BUSY_POLL
2032		data[i++] = ss->lock_napi_yield;
2033		data[i++] = ss->lock_poll_yield;
2034		data[i++] = ss->busy_poll_miss;
2035		data[i++] = ss->busy_poll_cnt;
2036#endif
2037	}
2038}
2039
2040static void myri10ge_set_msglevel(struct net_device *netdev, u32 value)
2041{
2042	struct myri10ge_priv *mgp = netdev_priv(netdev);
2043	mgp->msg_enable = value;
2044}
2045
2046static u32 myri10ge_get_msglevel(struct net_device *netdev)
2047{
2048	struct myri10ge_priv *mgp = netdev_priv(netdev);
2049	return mgp->msg_enable;
2050}
2051
2052/*
2053 * Use a low-level command to change the LED behavior. Rather than
2054 * blinking (which is the normal case), when identify is used, the
2055 * yellow LED turns solid.
2056 */
2057static int myri10ge_led(struct myri10ge_priv *mgp, int on)
2058{
2059	struct mcp_gen_header *hdr;
2060	struct device *dev = &mgp->pdev->dev;
2061	size_t hdr_off, pattern_off, hdr_len;
2062	u32 pattern = 0xfffffffe;
2063
2064	/* find running firmware header */
2065	hdr_off = swab32(readl(mgp->sram + MCP_HEADER_PTR_OFFSET));
2066	if ((hdr_off & 3) || hdr_off + sizeof(*hdr) > mgp->sram_size) {
2067		dev_err(dev, "Running firmware has bad header offset (%d)\n",
2068			(int)hdr_off);
2069		return -EIO;
2070	}
2071	hdr_len = swab32(readl(mgp->sram + hdr_off +
2072			       offsetof(struct mcp_gen_header, header_length)));
2073	pattern_off = hdr_off + offsetof(struct mcp_gen_header, led_pattern);
2074	if (pattern_off >= (hdr_len + hdr_off)) {
2075		dev_info(dev, "Firmware does not support LED identification\n");
2076		return -EINVAL;
2077	}
2078	if (!on)
2079		pattern = swab32(readl(mgp->sram + pattern_off + 4));
2080	writel(swab32(pattern), mgp->sram + pattern_off);
2081	return 0;
2082}
2083
2084static int
2085myri10ge_phys_id(struct net_device *netdev, enum ethtool_phys_id_state state)
2086{
2087	struct myri10ge_priv *mgp = netdev_priv(netdev);
2088	int rc;
2089
2090	switch (state) {
2091	case ETHTOOL_ID_ACTIVE:
2092		rc = myri10ge_led(mgp, 1);
2093		break;
2094
2095	case ETHTOOL_ID_INACTIVE:
2096		rc =  myri10ge_led(mgp, 0);
2097		break;
2098
2099	default:
2100		rc = -EINVAL;
2101	}
2102
2103	return rc;
2104}
2105
2106static const struct ethtool_ops myri10ge_ethtool_ops = {
2107	.get_settings = myri10ge_get_settings,
2108	.get_drvinfo = myri10ge_get_drvinfo,
2109	.get_coalesce = myri10ge_get_coalesce,
2110	.set_coalesce = myri10ge_set_coalesce,
2111	.get_pauseparam = myri10ge_get_pauseparam,
2112	.set_pauseparam = myri10ge_set_pauseparam,
2113	.get_ringparam = myri10ge_get_ringparam,
2114	.get_link = ethtool_op_get_link,
2115	.get_strings = myri10ge_get_strings,
2116	.get_sset_count = myri10ge_get_sset_count,
2117	.get_ethtool_stats = myri10ge_get_ethtool_stats,
2118	.set_msglevel = myri10ge_set_msglevel,
2119	.get_msglevel = myri10ge_get_msglevel,
2120	.set_phys_id = myri10ge_phys_id,
2121};
2122
2123static int myri10ge_allocate_rings(struct myri10ge_slice_state *ss)
2124{
2125	struct myri10ge_priv *mgp = ss->mgp;
2126	struct myri10ge_cmd cmd;
2127	struct net_device *dev = mgp->dev;
2128	int tx_ring_size, rx_ring_size;
2129	int tx_ring_entries, rx_ring_entries;
2130	int i, slice, status;
2131	size_t bytes;
2132
2133	/* get ring sizes */
2134	slice = ss - mgp->ss;
2135	cmd.data0 = slice;
2136	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_SEND_RING_SIZE, &cmd, 0);
2137	tx_ring_size = cmd.data0;
2138	cmd.data0 = slice;
2139	status |= myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_RX_RING_SIZE, &cmd, 0);
2140	if (status != 0)
2141		return status;
2142	rx_ring_size = cmd.data0;
2143
2144	tx_ring_entries = tx_ring_size / sizeof(struct mcp_kreq_ether_send);
2145	rx_ring_entries = rx_ring_size / sizeof(struct mcp_dma_addr);
2146	ss->tx.mask = tx_ring_entries - 1;
2147	ss->rx_small.mask = ss->rx_big.mask = rx_ring_entries - 1;
2148
2149	status = -ENOMEM;
2150
2151	/* allocate the host shadow rings */
2152
2153	bytes = 8 + (MYRI10GE_MAX_SEND_DESC_TSO + 4)
2154	    * sizeof(*ss->tx.req_list);
2155	ss->tx.req_bytes = kzalloc(bytes, GFP_KERNEL);
2156	if (ss->tx.req_bytes == NULL)
2157		goto abort_with_nothing;
2158
2159	/* ensure req_list entries are aligned to 8 bytes */
2160	ss->tx.req_list = (struct mcp_kreq_ether_send *)
2161	    ALIGN((unsigned long)ss->tx.req_bytes, 8);
2162	ss->tx.queue_active = 0;
2163
2164	bytes = rx_ring_entries * sizeof(*ss->rx_small.shadow);
2165	ss->rx_small.shadow = kzalloc(bytes, GFP_KERNEL);
2166	if (ss->rx_small.shadow == NULL)
2167		goto abort_with_tx_req_bytes;
2168
2169	bytes = rx_ring_entries * sizeof(*ss->rx_big.shadow);
2170	ss->rx_big.shadow = kzalloc(bytes, GFP_KERNEL);
2171	if (ss->rx_big.shadow == NULL)
2172		goto abort_with_rx_small_shadow;
2173
2174	/* allocate the host info rings */
2175
2176	bytes = tx_ring_entries * sizeof(*ss->tx.info);
2177	ss->tx.info = kzalloc(bytes, GFP_KERNEL);
2178	if (ss->tx.info == NULL)
2179		goto abort_with_rx_big_shadow;
2180
2181	bytes = rx_ring_entries * sizeof(*ss->rx_small.info);
2182	ss->rx_small.info = kzalloc(bytes, GFP_KERNEL);
2183	if (ss->rx_small.info == NULL)
2184		goto abort_with_tx_info;
2185
2186	bytes = rx_ring_entries * sizeof(*ss->rx_big.info);
2187	ss->rx_big.info = kzalloc(bytes, GFP_KERNEL);
2188	if (ss->rx_big.info == NULL)
2189		goto abort_with_rx_small_info;
2190
2191	/* Fill the receive rings */
2192	ss->rx_big.cnt = 0;
2193	ss->rx_small.cnt = 0;
2194	ss->rx_big.fill_cnt = 0;
2195	ss->rx_small.fill_cnt = 0;
2196	ss->rx_small.page_offset = MYRI10GE_ALLOC_SIZE;
2197	ss->rx_big.page_offset = MYRI10GE_ALLOC_SIZE;
2198	ss->rx_small.watchdog_needed = 0;
2199	ss->rx_big.watchdog_needed = 0;
2200	if (mgp->small_bytes == 0) {
2201		ss->rx_small.fill_cnt = ss->rx_small.mask + 1;
2202	} else {
2203		myri10ge_alloc_rx_pages(mgp, &ss->rx_small,
2204					mgp->small_bytes + MXGEFW_PAD, 0);
2205	}
2206
2207	if (ss->rx_small.fill_cnt < ss->rx_small.mask + 1) {
2208		netdev_err(dev, "slice-%d: alloced only %d small bufs\n",
2209			   slice, ss->rx_small.fill_cnt);
2210		goto abort_with_rx_small_ring;
2211	}
2212
2213	myri10ge_alloc_rx_pages(mgp, &ss->rx_big, mgp->big_bytes, 0);
2214	if (ss->rx_big.fill_cnt < ss->rx_big.mask + 1) {
2215		netdev_err(dev, "slice-%d: alloced only %d big bufs\n",
2216			   slice, ss->rx_big.fill_cnt);
2217		goto abort_with_rx_big_ring;
2218	}
2219
2220	return 0;
2221
2222abort_with_rx_big_ring:
2223	for (i = ss->rx_big.cnt; i < ss->rx_big.fill_cnt; i++) {
2224		int idx = i & ss->rx_big.mask;
2225		myri10ge_unmap_rx_page(mgp->pdev, &ss->rx_big.info[idx],
2226				       mgp->big_bytes);
2227		put_page(ss->rx_big.info[idx].page);
2228	}
2229
2230abort_with_rx_small_ring:
2231	if (mgp->small_bytes == 0)
2232		ss->rx_small.fill_cnt = ss->rx_small.cnt;
2233	for (i = ss->rx_small.cnt; i < ss->rx_small.fill_cnt; i++) {
2234		int idx = i & ss->rx_small.mask;
2235		myri10ge_unmap_rx_page(mgp->pdev, &ss->rx_small.info[idx],
2236				       mgp->small_bytes + MXGEFW_PAD);
2237		put_page(ss->rx_small.info[idx].page);
2238	}
2239
2240	kfree(ss->rx_big.info);
2241
2242abort_with_rx_small_info:
2243	kfree(ss->rx_small.info);
2244
2245abort_with_tx_info:
2246	kfree(ss->tx.info);
2247
2248abort_with_rx_big_shadow:
2249	kfree(ss->rx_big.shadow);
2250
2251abort_with_rx_small_shadow:
2252	kfree(ss->rx_small.shadow);
2253
2254abort_with_tx_req_bytes:
2255	kfree(ss->tx.req_bytes);
2256	ss->tx.req_bytes = NULL;
2257	ss->tx.req_list = NULL;
2258
2259abort_with_nothing:
2260	return status;
2261}
2262
2263static void myri10ge_free_rings(struct myri10ge_slice_state *ss)
2264{
2265	struct myri10ge_priv *mgp = ss->mgp;
2266	struct sk_buff *skb;
2267	struct myri10ge_tx_buf *tx;
2268	int i, len, idx;
2269
2270	/* If not allocated, skip it */
2271	if (ss->tx.req_list == NULL)
2272		return;
2273
2274	for (i = ss->rx_big.cnt; i < ss->rx_big.fill_cnt; i++) {
2275		idx = i & ss->rx_big.mask;
2276		if (i == ss->rx_big.fill_cnt - 1)
2277			ss->rx_big.info[idx].page_offset = MYRI10GE_ALLOC_SIZE;
2278		myri10ge_unmap_rx_page(mgp->pdev, &ss->rx_big.info[idx],
2279				       mgp->big_bytes);
2280		put_page(ss->rx_big.info[idx].page);
2281	}
2282
2283	if (mgp->small_bytes == 0)
2284		ss->rx_small.fill_cnt = ss->rx_small.cnt;
2285	for (i = ss->rx_small.cnt; i < ss->rx_small.fill_cnt; i++) {
2286		idx = i & ss->rx_small.mask;
2287		if (i == ss->rx_small.fill_cnt - 1)
2288			ss->rx_small.info[idx].page_offset =
2289			    MYRI10GE_ALLOC_SIZE;
2290		myri10ge_unmap_rx_page(mgp->pdev, &ss->rx_small.info[idx],
2291				       mgp->small_bytes + MXGEFW_PAD);
2292		put_page(ss->rx_small.info[idx].page);
2293	}
2294	tx = &ss->tx;
2295	while (tx->done != tx->req) {
2296		idx = tx->done & tx->mask;
2297		skb = tx->info[idx].skb;
2298
2299		/* Mark as free */
2300		tx->info[idx].skb = NULL;
2301		tx->done++;
2302		len = dma_unmap_len(&tx->info[idx], len);
2303		dma_unmap_len_set(&tx->info[idx], len, 0);
2304		if (skb) {
2305			ss->stats.tx_dropped++;
2306			dev_kfree_skb_any(skb);
2307			if (len)
2308				pci_unmap_single(mgp->pdev,
2309						 dma_unmap_addr(&tx->info[idx],
2310								bus), len,
2311						 PCI_DMA_TODEVICE);
2312		} else {
2313			if (len)
2314				pci_unmap_page(mgp->pdev,
2315					       dma_unmap_addr(&tx->info[idx],
2316							      bus), len,
2317					       PCI_DMA_TODEVICE);
2318		}
2319	}
2320	kfree(ss->rx_big.info);
2321
2322	kfree(ss->rx_small.info);
2323
2324	kfree(ss->tx.info);
2325
2326	kfree(ss->rx_big.shadow);
2327
2328	kfree(ss->rx_small.shadow);
2329
2330	kfree(ss->tx.req_bytes);
2331	ss->tx.req_bytes = NULL;
2332	ss->tx.req_list = NULL;
2333}
2334
2335static int myri10ge_request_irq(struct myri10ge_priv *mgp)
2336{
2337	struct pci_dev *pdev = mgp->pdev;
2338	struct myri10ge_slice_state *ss;
2339	struct net_device *netdev = mgp->dev;
2340	int i;
2341	int status;
2342
2343	mgp->msi_enabled = 0;
2344	mgp->msix_enabled = 0;
2345	status = 0;
2346	if (myri10ge_msi) {
2347		if (mgp->num_slices > 1) {
2348			status = pci_enable_msix_range(pdev, mgp->msix_vectors,
2349					mgp->num_slices, mgp->num_slices);
2350			if (status < 0) {
2351				dev_err(&pdev->dev,
2352					"Error %d setting up MSI-X\n", status);
2353				return status;
2354			}
2355			mgp->msix_enabled = 1;
2356		}
2357		if (mgp->msix_enabled == 0) {
2358			status = pci_enable_msi(pdev);
2359			if (status != 0) {
2360				dev_err(&pdev->dev,
2361					"Error %d setting up MSI; falling back to xPIC\n",
2362					status);
2363			} else {
2364				mgp->msi_enabled = 1;
2365			}
2366		}
2367	}
2368	if (mgp->msix_enabled) {
2369		for (i = 0; i < mgp->num_slices; i++) {
2370			ss = &mgp->ss[i];
2371			snprintf(ss->irq_desc, sizeof(ss->irq_desc),
2372				 "%s:slice-%d", netdev->name, i);
2373			status = request_irq(mgp->msix_vectors[i].vector,
2374					     myri10ge_intr, 0, ss->irq_desc,
2375					     ss);
2376			if (status != 0) {
2377				dev_err(&pdev->dev,
2378					"slice %d failed to allocate IRQ\n", i);
2379				i--;
2380				while (i >= 0) {
2381					free_irq(mgp->msix_vectors[i].vector,
2382						 &mgp->ss[i]);
2383					i--;
2384				}
2385				pci_disable_msix(pdev);
2386				return status;
2387			}
2388		}
2389	} else {
2390		status = request_irq(pdev->irq, myri10ge_intr, IRQF_SHARED,
2391				     mgp->dev->name, &mgp->ss[0]);
2392		if (status != 0) {
2393			dev_err(&pdev->dev, "failed to allocate IRQ\n");
2394			if (mgp->msi_enabled)
2395				pci_disable_msi(pdev);
2396		}
2397	}
2398	return status;
2399}
2400
2401static void myri10ge_free_irq(struct myri10ge_priv *mgp)
2402{
2403	struct pci_dev *pdev = mgp->pdev;
2404	int i;
2405
2406	if (mgp->msix_enabled) {
2407		for (i = 0; i < mgp->num_slices; i++)
2408			free_irq(mgp->msix_vectors[i].vector, &mgp->ss[i]);
2409	} else {
2410		free_irq(pdev->irq, &mgp->ss[0]);
2411	}
2412	if (mgp->msi_enabled)
2413		pci_disable_msi(pdev);
2414	if (mgp->msix_enabled)
2415		pci_disable_msix(pdev);
2416}
2417
2418static int myri10ge_get_txrx(struct myri10ge_priv *mgp, int slice)
2419{
2420	struct myri10ge_cmd cmd;
2421	struct myri10ge_slice_state *ss;
2422	int status;
2423
2424	ss = &mgp->ss[slice];
2425	status = 0;
2426	if (slice == 0 || (mgp->dev->real_num_tx_queues > 1)) {
2427		cmd.data0 = slice;
2428		status = myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_SEND_OFFSET,
2429					   &cmd, 0);
2430		ss->tx.lanai = (struct mcp_kreq_ether_send __iomem *)
2431		    (mgp->sram + cmd.data0);
2432	}
2433	cmd.data0 = slice;
2434	status |= myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_SMALL_RX_OFFSET,
2435				    &cmd, 0);
2436	ss->rx_small.lanai = (struct mcp_kreq_ether_recv __iomem *)
2437	    (mgp->sram + cmd.data0);
2438
2439	cmd.data0 = slice;
2440	status |= myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_BIG_RX_OFFSET, &cmd, 0);
2441	ss->rx_big.lanai = (struct mcp_kreq_ether_recv __iomem *)
2442	    (mgp->sram + cmd.data0);
2443
2444	ss->tx.send_go = (__iomem __be32 *)
2445	    (mgp->sram + MXGEFW_ETH_SEND_GO + 64 * slice);
2446	ss->tx.send_stop = (__iomem __be32 *)
2447	    (mgp->sram + MXGEFW_ETH_SEND_STOP + 64 * slice);
2448	return status;
2449
2450}
2451
2452static int myri10ge_set_stats(struct myri10ge_priv *mgp, int slice)
2453{
2454	struct myri10ge_cmd cmd;
2455	struct myri10ge_slice_state *ss;
2456	int status;
2457
2458	ss = &mgp->ss[slice];
2459	cmd.data0 = MYRI10GE_LOWPART_TO_U32(ss->fw_stats_bus);
2460	cmd.data1 = MYRI10GE_HIGHPART_TO_U32(ss->fw_stats_bus);
2461	cmd.data2 = sizeof(struct mcp_irq_data) | (slice << 16);
2462	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_STATS_DMA_V2, &cmd, 0);
2463	if (status == -ENOSYS) {
2464		dma_addr_t bus = ss->fw_stats_bus;
2465		if (slice != 0)
2466			return -EINVAL;
2467		bus += offsetof(struct mcp_irq_data, send_done_count);
2468		cmd.data0 = MYRI10GE_LOWPART_TO_U32(bus);
2469		cmd.data1 = MYRI10GE_HIGHPART_TO_U32(bus);
2470		status = myri10ge_send_cmd(mgp,
2471					   MXGEFW_CMD_SET_STATS_DMA_OBSOLETE,
2472					   &cmd, 0);
2473		/* Firmware cannot support multicast without STATS_DMA_V2 */
2474		mgp->fw_multicast_support = 0;
2475	} else {
2476		mgp->fw_multicast_support = 1;
2477	}
2478	return 0;
2479}
2480
2481static int myri10ge_open(struct net_device *dev)
2482{
2483	struct myri10ge_slice_state *ss;
2484	struct myri10ge_priv *mgp = netdev_priv(dev);
2485	struct myri10ge_cmd cmd;
2486	int i, status, big_pow2, slice;
2487	u8 __iomem *itable;
2488
2489	if (mgp->running != MYRI10GE_ETH_STOPPED)
2490		return -EBUSY;
2491
2492	mgp->running = MYRI10GE_ETH_STARTING;
2493	status = myri10ge_reset(mgp);
2494	if (status != 0) {
2495		netdev_err(dev, "failed reset\n");
2496		goto abort_with_nothing;
2497	}
2498
2499	if (mgp->num_slices > 1) {
2500		cmd.data0 = mgp->num_slices;
2501		cmd.data1 = MXGEFW_SLICE_INTR_MODE_ONE_PER_SLICE;
2502		if (mgp->dev->real_num_tx_queues > 1)
2503			cmd.data1 |= MXGEFW_SLICE_ENABLE_MULTIPLE_TX_QUEUES;
2504		status = myri10ge_send_cmd(mgp, MXGEFW_CMD_ENABLE_RSS_QUEUES,
2505					   &cmd, 0);
2506		if (status != 0) {
2507			netdev_err(dev, "failed to set number of slices\n");
2508			goto abort_with_nothing;
2509		}
2510		/* setup the indirection table */
2511		cmd.data0 = mgp->num_slices;
2512		status = myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_RSS_TABLE_SIZE,
2513					   &cmd, 0);
2514
2515		status |= myri10ge_send_cmd(mgp,
2516					    MXGEFW_CMD_GET_RSS_TABLE_OFFSET,
2517					    &cmd, 0);
2518		if (status != 0) {
2519			netdev_err(dev, "failed to setup rss tables\n");
2520			goto abort_with_nothing;
2521		}
2522
2523		/* just enable an identity mapping */
2524		itable = mgp->sram + cmd.data0;
2525		for (i = 0; i < mgp->num_slices; i++)
2526			__raw_writeb(i, &itable[i]);
2527
2528		cmd.data0 = 1;
2529		cmd.data1 = myri10ge_rss_hash;
2530		status = myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_RSS_ENABLE,
2531					   &cmd, 0);
2532		if (status != 0) {
2533			netdev_err(dev, "failed to enable slices\n");
2534			goto abort_with_nothing;
2535		}
2536	}
2537
2538	status = myri10ge_request_irq(mgp);
2539	if (status != 0)
2540		goto abort_with_nothing;
2541
2542	/* decide what small buffer size to use.  For good TCP rx
2543	 * performance, it is important to not receive 1514 byte
2544	 * frames into jumbo buffers, as it confuses the socket buffer
2545	 * accounting code, leading to drops and erratic performance.
2546	 */
2547
2548	if (dev->mtu <= ETH_DATA_LEN)
2549		/* enough for a TCP header */
2550		mgp->small_bytes = (128 > SMP_CACHE_BYTES)
2551		    ? (128 - MXGEFW_PAD)
2552		    : (SMP_CACHE_BYTES - MXGEFW_PAD);
2553	else
2554		/* enough for a vlan encapsulated ETH_DATA_LEN frame */
2555		mgp->small_bytes = VLAN_ETH_FRAME_LEN;
2556
2557	/* Override the small buffer size? */
2558	if (myri10ge_small_bytes >= 0)
2559		mgp->small_bytes = myri10ge_small_bytes;
2560
2561	/* Firmware needs the big buff size as a power of 2.  Lie and
2562	 * tell him the buffer is larger, because we only use 1
2563	 * buffer/pkt, and the mtu will prevent overruns.
2564	 */
2565	big_pow2 = dev->mtu + ETH_HLEN + VLAN_HLEN + MXGEFW_PAD;
2566	if (big_pow2 < MYRI10GE_ALLOC_SIZE / 2) {
2567		while (!is_power_of_2(big_pow2))
2568			big_pow2++;
2569		mgp->big_bytes = dev->mtu + ETH_HLEN + VLAN_HLEN + MXGEFW_PAD;
2570	} else {
2571		big_pow2 = MYRI10GE_ALLOC_SIZE;
2572		mgp->big_bytes = big_pow2;
2573	}
2574
2575	/* setup the per-slice data structures */
2576	for (slice = 0; slice < mgp->num_slices; slice++) {
2577		ss = &mgp->ss[slice];
2578
2579		status = myri10ge_get_txrx(mgp, slice);
2580		if (status != 0) {
2581			netdev_err(dev, "failed to get ring sizes or locations\n");
2582			goto abort_with_rings;
2583		}
2584		status = myri10ge_allocate_rings(ss);
2585		if (status != 0)
2586			goto abort_with_rings;
2587
2588		/* only firmware which supports multiple TX queues
2589		 * supports setting up the tx stats on non-zero
2590		 * slices */
2591		if (slice == 0 || mgp->dev->real_num_tx_queues > 1)
2592			status = myri10ge_set_stats(mgp, slice);
2593		if (status) {
2594			netdev_err(dev, "Couldn't set stats DMA\n");
2595			goto abort_with_rings;
2596		}
2597
2598		/* Initialize the slice spinlock and state used for polling */
2599		myri10ge_ss_init_lock(ss);
2600
2601		/* must happen prior to any irq */
2602		napi_enable(&(ss)->napi);
2603	}
2604
2605	/* now give firmware buffers sizes, and MTU */
2606	cmd.data0 = dev->mtu + ETH_HLEN + VLAN_HLEN;
2607	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_MTU, &cmd, 0);
2608	cmd.data0 = mgp->small_bytes;
2609	status |=
2610	    myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_SMALL_BUFFER_SIZE, &cmd, 0);
2611	cmd.data0 = big_pow2;
2612	status |=
2613	    myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_BIG_BUFFER_SIZE, &cmd, 0);
2614	if (status) {
2615		netdev_err(dev, "Couldn't set buffer sizes\n");
2616		goto abort_with_rings;
2617	}
2618
2619	/*
2620	 * Set Linux style TSO mode; this is needed only on newer
2621	 *  firmware versions.  Older versions default to Linux
2622	 *  style TSO
2623	 */
2624	cmd.data0 = 0;
2625	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_TSO_MODE, &cmd, 0);
2626	if (status && status != -ENOSYS) {
2627		netdev_err(dev, "Couldn't set TSO mode\n");
2628		goto abort_with_rings;
2629	}
2630
2631	mgp->link_state = ~0U;
2632	mgp->rdma_tags_available = 15;
2633
2634	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_ETHERNET_UP, &cmd, 0);
2635	if (status) {
2636		netdev_err(dev, "Couldn't bring up link\n");
2637		goto abort_with_rings;
2638	}
2639
2640	mgp->running = MYRI10GE_ETH_RUNNING;
2641	mgp->watchdog_timer.expires = jiffies + myri10ge_watchdog_timeout * HZ;
2642	add_timer(&mgp->watchdog_timer);
2643	netif_tx_wake_all_queues(dev);
2644
2645	return 0;
2646
2647abort_with_rings:
2648	while (slice) {
2649		slice--;
2650		napi_disable(&mgp->ss[slice].napi);
2651	}
2652	for (i = 0; i < mgp->num_slices; i++)
2653		myri10ge_free_rings(&mgp->ss[i]);
2654
2655	myri10ge_free_irq(mgp);
2656
2657abort_with_nothing:
2658	mgp->running = MYRI10GE_ETH_STOPPED;
2659	return -ENOMEM;
2660}
2661
2662static int myri10ge_close(struct net_device *dev)
2663{
2664	struct myri10ge_priv *mgp = netdev_priv(dev);
2665	struct myri10ge_cmd cmd;
2666	int status, old_down_cnt;
2667	int i;
2668
2669	if (mgp->running != MYRI10GE_ETH_RUNNING)
2670		return 0;
2671
2672	if (mgp->ss[0].tx.req_bytes == NULL)
2673		return 0;
2674
2675	del_timer_sync(&mgp->watchdog_timer);
2676	mgp->running = MYRI10GE_ETH_STOPPING;
2677	local_bh_disable(); /* myri10ge_ss_lock_napi needs bh disabled */
2678	for (i = 0; i < mgp->num_slices; i++) {
2679		napi_disable(&mgp->ss[i].napi);
2680		/* Lock the slice to prevent the busy_poll handler from
2681		 * accessing it.  Later when we bring the NIC up, myri10ge_open
2682		 * resets the slice including this lock.
2683		 */
2684		while (!myri10ge_ss_lock_napi(&mgp->ss[i])) {
2685			pr_info("Slice %d locked\n", i);
2686			mdelay(1);
2687		}
2688	}
2689	local_bh_enable();
2690	netif_carrier_off(dev);
2691
2692	netif_tx_stop_all_queues(dev);
2693	if (mgp->rebooted == 0) {
2694		old_down_cnt = mgp->down_cnt;
2695		mb();
2696		status =
2697		    myri10ge_send_cmd(mgp, MXGEFW_CMD_ETHERNET_DOWN, &cmd, 0);
2698		if (status)
2699			netdev_err(dev, "Couldn't bring down link\n");
2700
2701		wait_event_timeout(mgp->down_wq, old_down_cnt != mgp->down_cnt,
2702				   HZ);
2703		if (old_down_cnt == mgp->down_cnt)
2704			netdev_err(dev, "never got down irq\n");
2705	}
2706	netif_tx_disable(dev);
2707	myri10ge_free_irq(mgp);
2708	for (i = 0; i < mgp->num_slices; i++)
2709		myri10ge_free_rings(&mgp->ss[i]);
2710
2711	mgp->running = MYRI10GE_ETH_STOPPED;
2712	return 0;
2713}
2714
2715/* copy an array of struct mcp_kreq_ether_send's to the mcp.  Copy
2716 * backwards one at a time and handle ring wraps */
2717
2718static inline void
2719myri10ge_submit_req_backwards(struct myri10ge_tx_buf *tx,
2720			      struct mcp_kreq_ether_send *src, int cnt)
2721{
2722	int idx, starting_slot;
2723	starting_slot = tx->req;
2724	while (cnt > 1) {
2725		cnt--;
2726		idx = (starting_slot + cnt) & tx->mask;
2727		myri10ge_pio_copy(&tx->lanai[idx], &src[cnt], sizeof(*src));
2728		mb();
2729	}
2730}
2731
2732/*
2733 * copy an array of struct mcp_kreq_ether_send's to the mcp.  Copy
2734 * at most 32 bytes at a time, so as to avoid involving the software
2735 * pio handler in the nic.   We re-write the first segment's flags
2736 * to mark them valid only after writing the entire chain.
2737 */
2738
2739static inline void
2740myri10ge_submit_req(struct myri10ge_tx_buf *tx, struct mcp_kreq_ether_send *src,
2741		    int cnt)
2742{
2743	int idx, i;
2744	struct mcp_kreq_ether_send __iomem *dstp, *dst;
2745	struct mcp_kreq_ether_send *srcp;
2746	u8 last_flags;
2747
2748	idx = tx->req & tx->mask;
2749
2750	last_flags = src->flags;
2751	src->flags = 0;
2752	mb();
2753	dst = dstp = &tx->lanai[idx];
2754	srcp = src;
2755
2756	if ((idx + cnt) < tx->mask) {
2757		for (i = 0; i < (cnt - 1); i += 2) {
2758			myri10ge_pio_copy(dstp, srcp, 2 * sizeof(*src));
2759			mb();	/* force write every 32 bytes */
2760			srcp += 2;
2761			dstp += 2;
2762		}
2763	} else {
2764		/* submit all but the first request, and ensure
2765		 * that it is submitted below */
2766		myri10ge_submit_req_backwards(tx, src, cnt);
2767		i = 0;
2768	}
2769	if (i < cnt) {
2770		/* submit the first request */
2771		myri10ge_pio_copy(dstp, srcp, sizeof(*src));
2772		mb();		/* barrier before setting valid flag */
2773	}
2774
2775	/* re-write the last 32-bits with the valid flags */
2776	src->flags = last_flags;
2777	put_be32(*((__be32 *) src + 3), (__be32 __iomem *) dst + 3);
2778	tx->req += cnt;
2779	mb();
2780}
2781
2782static void myri10ge_unmap_tx_dma(struct myri10ge_priv *mgp,
2783				  struct myri10ge_tx_buf *tx, int idx)
2784{
2785	unsigned int len;
2786	int last_idx;
2787
2788	/* Free any DMA resources we've alloced and clear out the skb slot */
2789	last_idx = (idx + 1) & tx->mask;
2790	idx = tx->req & tx->mask;
2791	do {
2792		len = dma_unmap_len(&tx->info[idx], len);
2793		if (len) {
2794			if (tx->info[idx].skb != NULL)
2795				pci_unmap_single(mgp->pdev,
2796						 dma_unmap_addr(&tx->info[idx],
2797								bus), len,
2798						 PCI_DMA_TODEVICE);
2799			else
2800				pci_unmap_page(mgp->pdev,
2801					       dma_unmap_addr(&tx->info[idx],
2802							      bus), len,
2803					       PCI_DMA_TODEVICE);
2804			dma_unmap_len_set(&tx->info[idx], len, 0);
2805			tx->info[idx].skb = NULL;
2806		}
2807		idx = (idx + 1) & tx->mask;
2808	} while (idx != last_idx);
2809}
2810
2811/*
2812 * Transmit a packet.  We need to split the packet so that a single
2813 * segment does not cross myri10ge->tx_boundary, so this makes segment
2814 * counting tricky.  So rather than try to count segments up front, we
2815 * just give up if there are too few segments to hold a reasonably
2816 * fragmented packet currently available.  If we run
2817 * out of segments while preparing a packet for DMA, we just linearize
2818 * it and try again.
2819 */
2820
2821static netdev_tx_t myri10ge_xmit(struct sk_buff *skb,
2822				       struct net_device *dev)
2823{
2824	struct myri10ge_priv *mgp = netdev_priv(dev);
2825	struct myri10ge_slice_state *ss;
2826	struct mcp_kreq_ether_send *req;
2827	struct myri10ge_tx_buf *tx;
2828	struct skb_frag_struct *frag;
2829	struct netdev_queue *netdev_queue;
2830	dma_addr_t bus;
2831	u32 low;
2832	__be32 high_swapped;
2833	unsigned int len;
2834	int idx, avail, frag_cnt, frag_idx, count, mss, max_segments;
2835	u16 pseudo_hdr_offset, cksum_offset, queue;
2836	int cum_len, seglen, boundary, rdma_count;
2837	u8 flags, odd_flag;
2838
2839	queue = skb_get_queue_mapping(skb);
2840	ss = &mgp->ss[queue];
2841	netdev_queue = netdev_get_tx_queue(mgp->dev, queue);
2842	tx = &ss->tx;
2843
2844again:
2845	req = tx->req_list;
2846	avail = tx->mask - 1 - (tx->req - tx->done);
2847
2848	mss = 0;
2849	max_segments = MXGEFW_MAX_SEND_DESC;
2850
2851	if (skb_is_gso(skb)) {
2852		mss = skb_shinfo(skb)->gso_size;
2853		max_segments = MYRI10GE_MAX_SEND_DESC_TSO;
2854	}
2855
2856	if ((unlikely(avail < max_segments))) {
2857		/* we are out of transmit resources */
2858		tx->stop_queue++;
2859		netif_tx_stop_queue(netdev_queue);
2860		return NETDEV_TX_BUSY;
2861	}
2862
2863	/* Setup checksum offloading, if needed */
2864	cksum_offset = 0;
2865	pseudo_hdr_offset = 0;
2866	odd_flag = 0;
2867	flags = (MXGEFW_FLAGS_NO_TSO | MXGEFW_FLAGS_FIRST);
2868	if (likely(skb->ip_summed == CHECKSUM_PARTIAL)) {
2869		cksum_offset = skb_checksum_start_offset(skb);
2870		pseudo_hdr_offset = cksum_offset + skb->csum_offset;
2871		/* If the headers are excessively large, then we must
2872		 * fall back to a software checksum */
2873		if (unlikely(!mss && (cksum_offset > 255 ||
2874				      pseudo_hdr_offset > 127))) {
2875			if (skb_checksum_help(skb))
2876				goto drop;
2877			cksum_offset = 0;
2878			pseudo_hdr_offset = 0;
2879		} else {
2880			odd_flag = MXGEFW_FLAGS_ALIGN_ODD;
2881			flags |= MXGEFW_FLAGS_CKSUM;
2882		}
2883	}
2884
2885	cum_len = 0;
2886
2887	if (mss) {		/* TSO */
2888		/* this removes any CKSUM flag from before */
2889		flags = (MXGEFW_FLAGS_TSO_HDR | MXGEFW_FLAGS_FIRST);
2890
2891		/* negative cum_len signifies to the
2892		 * send loop that we are still in the
2893		 * header portion of the TSO packet.
2894		 * TSO header can be at most 1KB long */
2895		cum_len = -(skb_transport_offset(skb) + tcp_hdrlen(skb));
2896
2897		/* for IPv6 TSO, the checksum offset stores the
2898		 * TCP header length, to save the firmware from
2899		 * the need to parse the headers */
2900		if (skb_is_gso_v6(skb)) {
2901			cksum_offset = tcp_hdrlen(skb);
2902			/* Can only handle headers <= max_tso6 long */
2903			if (unlikely(-cum_len > mgp->max_tso6))
2904				return myri10ge_sw_tso(skb, dev);
2905		}
2906		/* for TSO, pseudo_hdr_offset holds mss.
2907		 * The firmware figures out where to put
2908		 * the checksum by parsing the header. */
2909		pseudo_hdr_offset = mss;
2910	} else
2911		/* Mark small packets, and pad out tiny packets */
2912	if (skb->len <= MXGEFW_SEND_SMALL_SIZE) {
2913		flags |= MXGEFW_FLAGS_SMALL;
2914
2915		/* pad frames to at least ETH_ZLEN bytes */
2916		if (unlikely(skb->len < ETH_ZLEN)) {
2917			if (skb_padto(skb, ETH_ZLEN)) {
2918				/* The packet is gone, so we must
2919				 * return 0 */
2920				ss->stats.tx_dropped += 1;
2921				return NETDEV_TX_OK;
2922			}
2923			/* adjust the len to account for the zero pad
2924			 * so that the nic can know how long it is */
2925			skb->len = ETH_ZLEN;
2926		}
2927	}
2928
2929	/* map the skb for DMA */
2930	len = skb_headlen(skb);
2931	bus = pci_map_single(mgp->pdev, skb->data, len, PCI_DMA_TODEVICE);
2932	if (unlikely(pci_dma_mapping_error(mgp->pdev, bus)))
2933		goto drop;
2934
2935	idx = tx->req & tx->mask;
2936	tx->info[idx].skb = skb;
2937	dma_unmap_addr_set(&tx->info[idx], bus, bus);
2938	dma_unmap_len_set(&tx->info[idx], len, len);
2939
2940	frag_cnt = skb_shinfo(skb)->nr_frags;
2941	frag_idx = 0;
2942	count = 0;
2943	rdma_count = 0;
2944
2945	/* "rdma_count" is the number of RDMAs belonging to the
2946	 * current packet BEFORE the current send request. For
2947	 * non-TSO packets, this is equal to "count".
2948	 * For TSO packets, rdma_count needs to be reset
2949	 * to 0 after a segment cut.
2950	 *
2951	 * The rdma_count field of the send request is
2952	 * the number of RDMAs of the packet starting at
2953	 * that request. For TSO send requests with one ore more cuts
2954	 * in the middle, this is the number of RDMAs starting
2955	 * after the last cut in the request. All previous
2956	 * segments before the last cut implicitly have 1 RDMA.
2957	 *
2958	 * Since the number of RDMAs is not known beforehand,
2959	 * it must be filled-in retroactively - after each
2960	 * segmentation cut or at the end of the entire packet.
2961	 */
2962
2963	while (1) {
2964		/* Break the SKB or Fragment up into pieces which
2965		 * do not cross mgp->tx_boundary */
2966		low = MYRI10GE_LOWPART_TO_U32(bus);
2967		high_swapped = htonl(MYRI10GE_HIGHPART_TO_U32(bus));
2968		while (len) {
2969			u8 flags_next;
2970			int cum_len_next;
2971
2972			if (unlikely(count == max_segments))
2973				goto abort_linearize;
2974
2975			boundary =
2976			    (low + mgp->tx_boundary) & ~(mgp->tx_boundary - 1);
2977			seglen = boundary - low;
2978			if (seglen > len)
2979				seglen = len;
2980			flags_next = flags & ~MXGEFW_FLAGS_FIRST;
2981			cum_len_next = cum_len + seglen;
2982			if (mss) {	/* TSO */
2983				(req - rdma_count)->rdma_count = rdma_count + 1;
2984
2985				if (likely(cum_len >= 0)) {	/* payload */
2986					int next_is_first, chop;
2987
2988					chop = (cum_len_next > mss);
2989					cum_len_next = cum_len_next % mss;
2990					next_is_first = (cum_len_next == 0);
2991					flags |= chop * MXGEFW_FLAGS_TSO_CHOP;
2992					flags_next |= next_is_first *
2993					    MXGEFW_FLAGS_FIRST;
2994					rdma_count |= -(chop | next_is_first);
2995					rdma_count += chop & ~next_is_first;
2996				} else if (likely(cum_len_next >= 0)) {	/* header ends */
2997					int small;
2998
2999					rdma_count = -1;
3000					cum_len_next = 0;
3001					seglen = -cum_len;
3002					small = (mss <= MXGEFW_SEND_SMALL_SIZE);
3003					flags_next = MXGEFW_FLAGS_TSO_PLD |
3004					    MXGEFW_FLAGS_FIRST |
3005					    (small * MXGEFW_FLAGS_SMALL);
3006				}
3007			}
3008			req->addr_high = high_swapped;
3009			req->addr_low = htonl(low);
3010			req->pseudo_hdr_offset = htons(pseudo_hdr_offset);
3011			req->pad = 0;	/* complete solid 16-byte block; does this matter? */
3012			req->rdma_count = 1;
3013			req->length = htons(seglen);
3014			req->cksum_offset = cksum_offset;
3015			req->flags = flags | ((cum_len & 1) * odd_flag);
3016
3017			low += seglen;
3018			len -= seglen;
3019			cum_len = cum_len_next;
3020			flags = flags_next;
3021			req++;
3022			count++;
3023			rdma_count++;
3024			if (cksum_offset != 0 && !(mss && skb_is_gso_v6(skb))) {
3025				if (unlikely(cksum_offset > seglen))
3026					cksum_offset -= seglen;
3027				else
3028					cksum_offset = 0;
3029			}
3030		}
3031		if (frag_idx == frag_cnt)
3032			break;
3033
3034		/* map next fragment for DMA */
3035		frag = &skb_shinfo(skb)->frags[frag_idx];
3036		frag_idx++;
3037		len = skb_frag_size(frag);
3038		bus = skb_frag_dma_map(&mgp->pdev->dev, frag, 0, len,
3039				       DMA_TO_DEVICE);
3040		if (unlikely(pci_dma_mapping_error(mgp->pdev, bus))) {
3041			myri10ge_unmap_tx_dma(mgp, tx, idx);
3042			goto drop;
3043		}
3044		idx = (count + tx->req) & tx->mask;
3045		dma_unmap_addr_set(&tx->info[idx], bus, bus);
3046		dma_unmap_len_set(&tx->info[idx], len, len);
3047	}
3048
3049	(req - rdma_count)->rdma_count = rdma_count;
3050	if (mss)
3051		do {
3052			req--;
3053			req->flags |= MXGEFW_FLAGS_TSO_LAST;
3054		} while (!(req->flags & (MXGEFW_FLAGS_TSO_CHOP |
3055					 MXGEFW_FLAGS_FIRST)));
3056	idx = ((count - 1) + tx->req) & tx->mask;
3057	tx->info[idx].last = 1;
3058	myri10ge_submit_req(tx, tx->req_list, count);
3059	/* if using multiple tx queues, make sure NIC polls the
3060	 * current slice */
3061	if ((mgp->dev->real_num_tx_queues > 1) && tx->queue_active == 0) {
3062		tx->queue_active = 1;
3063		put_be32(htonl(1), tx->send_go);
3064		mb();
3065		mmiowb();
3066	}
3067	tx->pkt_start++;
3068	if ((avail - count) < MXGEFW_MAX_SEND_DESC) {
3069		tx->stop_queue++;
3070		netif_tx_stop_queue(netdev_queue);
3071	}
3072	return NETDEV_TX_OK;
3073
3074abort_linearize:
3075	myri10ge_unmap_tx_dma(mgp, tx, idx);
3076
3077	if (skb_is_gso(skb)) {
3078		netdev_err(mgp->dev, "TSO but wanted to linearize?!?!?\n");
3079		goto drop;
3080	}
3081
3082	if (skb_linearize(skb))
3083		goto drop;
3084
3085	tx->linearized++;
3086	goto again;
3087
3088drop:
3089	dev_kfree_skb_any(skb);
3090	ss->stats.tx_dropped += 1;
3091	return NETDEV_TX_OK;
3092
3093}
3094
3095static netdev_tx_t myri10ge_sw_tso(struct sk_buff *skb,
3096					 struct net_device *dev)
3097{
3098	struct sk_buff *segs, *curr;
3099	struct myri10ge_priv *mgp = netdev_priv(dev);
3100	struct myri10ge_slice_state *ss;
3101	netdev_tx_t status;
3102
3103	segs = skb_gso_segment(skb, dev->features & ~NETIF_F_TSO6);
3104	if (IS_ERR(segs))
3105		goto drop;
3106
3107	while (segs) {
3108		curr = segs;
3109		segs = segs->next;
3110		curr->next = NULL;
3111		status = myri10ge_xmit(curr, dev);
3112		if (status != 0) {
3113			dev_kfree_skb_any(curr);
3114			if (segs != NULL) {
3115				curr = segs;
3116				segs = segs->next;
3117				curr->next = NULL;
3118				dev_kfree_skb_any(segs);
3119			}
3120			goto drop;
3121		}
3122	}
3123	dev_kfree_skb_any(skb);
3124	return NETDEV_TX_OK;
3125
3126drop:
3127	ss = &mgp->ss[skb_get_queue_mapping(skb)];
3128	dev_kfree_skb_any(skb);
3129	ss->stats.tx_dropped += 1;
3130	return NETDEV_TX_OK;
3131}
3132
3133static struct rtnl_link_stats64 *myri10ge_get_stats(struct net_device *dev,
3134						    struct rtnl_link_stats64 *stats)
3135{
3136	const struct myri10ge_priv *mgp = netdev_priv(dev);
3137	const struct myri10ge_slice_netstats *slice_stats;
3138	int i;
3139
3140	for (i = 0; i < mgp->num_slices; i++) {
3141		slice_stats = &mgp->ss[i].stats;
3142		stats->rx_packets += slice_stats->rx_packets;
3143		stats->tx_packets += slice_stats->tx_packets;
3144		stats->rx_bytes += slice_stats->rx_bytes;
3145		stats->tx_bytes += slice_stats->tx_bytes;
3146		stats->rx_dropped += slice_stats->rx_dropped;
3147		stats->tx_dropped += slice_stats->tx_dropped;
3148	}
3149	return stats;
3150}
3151
3152static void myri10ge_set_multicast_list(struct net_device *dev)
3153{
3154	struct myri10ge_priv *mgp = netdev_priv(dev);
3155	struct myri10ge_cmd cmd;
3156	struct netdev_hw_addr *ha;
3157	__be32 data[2] = { 0, 0 };
3158	int err;
3159
3160	/* can be called from atomic contexts,
3161	 * pass 1 to force atomicity in myri10ge_send_cmd() */
3162	myri10ge_change_promisc(mgp, dev->flags & IFF_PROMISC, 1);
3163
3164	/* This firmware is known to not support multicast */
3165	if (!mgp->fw_multicast_support)
3166		return;
3167
3168	/* Disable multicast filtering */
3169
3170	err = myri10ge_send_cmd(mgp, MXGEFW_ENABLE_ALLMULTI, &cmd, 1);
3171	if (err != 0) {
3172		netdev_err(dev, "Failed MXGEFW_ENABLE_ALLMULTI, error status: %d\n",
3173			   err);
3174		goto abort;
3175	}
3176
3177	if ((dev->flags & IFF_ALLMULTI) || mgp->adopted_rx_filter_bug) {
3178		/* request to disable multicast filtering, so quit here */
3179		return;
3180	}
3181
3182	/* Flush the filters */
3183
3184	err = myri10ge_send_cmd(mgp, MXGEFW_LEAVE_ALL_MULTICAST_GROUPS,
3185				&cmd, 1);
3186	if (err != 0) {
3187		netdev_err(dev, "Failed MXGEFW_LEAVE_ALL_MULTICAST_GROUPS, error status: %d\n",
3188			   err);
3189		goto abort;
3190	}
3191
3192	/* Walk the multicast list, and add each address */
3193	netdev_for_each_mc_addr(ha, dev) {
3194		memcpy(data, &ha->addr, ETH_ALEN);
3195		cmd.data0 = ntohl(data[0]);
3196		cmd.data1 = ntohl(data[1]);
3197		err = myri10ge_send_cmd(mgp, MXGEFW_JOIN_MULTICAST_GROUP,
3198					&cmd, 1);
3199
3200		if (err != 0) {
3201			netdev_err(dev, "Failed MXGEFW_JOIN_MULTICAST_GROUP, error status:%d %pM\n",
3202				   err, ha->addr);
3203			goto abort;
3204		}
3205	}
3206	/* Enable multicast filtering */
3207	err = myri10ge_send_cmd(mgp, MXGEFW_DISABLE_ALLMULTI, &cmd, 1);
3208	if (err != 0) {
3209		netdev_err(dev, "Failed MXGEFW_DISABLE_ALLMULTI, error status: %d\n",
3210			   err);
3211		goto abort;
3212	}
3213
3214	return;
3215
3216abort:
3217	return;
3218}
3219
3220static int myri10ge_set_mac_address(struct net_device *dev, void *addr)
3221{
3222	struct sockaddr *sa = addr;
3223	struct myri10ge_priv *mgp = netdev_priv(dev);
3224	int status;
3225
3226	if (!is_valid_ether_addr(sa->sa_data))
3227		return -EADDRNOTAVAIL;
3228
3229	status = myri10ge_update_mac_address(mgp, sa->sa_data);
3230	if (status != 0) {
3231		netdev_err(dev, "changing mac address failed with %d\n",
3232			   status);
3233		return status;
3234	}
3235
3236	/* change the dev structure */
3237	memcpy(dev->dev_addr, sa->sa_data, ETH_ALEN);
3238	return 0;
3239}
3240
3241static int myri10ge_change_mtu(struct net_device *dev, int new_mtu)
3242{
3243	struct myri10ge_priv *mgp = netdev_priv(dev);
3244	int error = 0;
3245
3246	if ((new_mtu < 68) || (ETH_HLEN + new_mtu > MYRI10GE_MAX_ETHER_MTU)) {
3247		netdev_err(dev, "new mtu (%d) is not valid\n", new_mtu);
3248		return -EINVAL;
3249	}
3250	netdev_info(dev, "changing mtu from %d to %d\n", dev->mtu, new_mtu);
3251	if (mgp->running) {
3252		/* if we change the mtu on an active device, we must
3253		 * reset the device so the firmware sees the change */
3254		myri10ge_close(dev);
3255		dev->mtu = new_mtu;
3256		myri10ge_open(dev);
3257	} else
3258		dev->mtu = new_mtu;
3259
3260	return error;
3261}
3262
3263/*
3264 * Enable ECRC to align PCI-E Completion packets on an 8-byte boundary.
3265 * Only do it if the bridge is a root port since we don't want to disturb
3266 * any other device, except if forced with myri10ge_ecrc_enable > 1.
3267 */
3268
3269static void myri10ge_enable_ecrc(struct myri10ge_priv *mgp)
3270{
3271	struct pci_dev *bridge = mgp->pdev->bus->self;
3272	struct device *dev = &mgp->pdev->dev;
3273	int cap;
3274	unsigned err_cap;
3275	int ret;
3276
3277	if (!myri10ge_ecrc_enable || !bridge)
3278		return;
3279
3280	/* check that the bridge is a root port */
3281	if (pci_pcie_type(bridge) != PCI_EXP_TYPE_ROOT_PORT) {
3282		if (myri10ge_ecrc_enable > 1) {
3283			struct pci_dev *prev_bridge, *old_bridge = bridge;
3284
3285			/* Walk the hierarchy up to the root port
3286			 * where ECRC has to be enabled */
3287			do {
3288				prev_bridge = bridge;
3289				bridge = bridge->bus->self;
3290				if (!bridge || prev_bridge == bridge) {
3291					dev_err(dev,
3292						"Failed to find root port"
3293						" to force ECRC\n");
3294					return;
3295				}
3296			} while (pci_pcie_type(bridge) !=
3297				 PCI_EXP_TYPE_ROOT_PORT);
3298
3299			dev_info(dev,
3300				 "Forcing ECRC on non-root port %s"
3301				 " (enabling on root port %s)\n",
3302				 pci_name(old_bridge), pci_name(bridge));
3303		} else {
3304			dev_err(dev,
3305				"Not enabling ECRC on non-root port %s\n",
3306				pci_name(bridge));
3307			return;
3308		}
3309	}
3310
3311	cap = pci_find_ext_capability(bridge, PCI_EXT_CAP_ID_ERR);
3312	if (!cap)
3313		return;
3314
3315	ret = pci_read_config_dword(bridge, cap + PCI_ERR_CAP, &err_cap);
3316	if (ret) {
3317		dev_err(dev, "failed reading ext-conf-space of %s\n",
3318			pci_name(bridge));
3319		dev_err(dev, "\t pci=nommconf in use? "
3320			"or buggy/incomplete/absent ACPI MCFG attr?\n");
3321		return;
3322	}
3323	if (!(err_cap & PCI_ERR_CAP_ECRC_GENC))
3324		return;
3325
3326	err_cap |= PCI_ERR_CAP_ECRC_GENE;
3327	pci_write_config_dword(bridge, cap + PCI_ERR_CAP, err_cap);
3328	dev_info(dev, "Enabled ECRC on upstream bridge %s\n", pci_name(bridge));
3329}
3330
3331/*
3332 * The Lanai Z8E PCI-E interface achieves higher Read-DMA throughput
3333 * when the PCI-E Completion packets are aligned on an 8-byte
3334 * boundary.  Some PCI-E chip sets always align Completion packets; on
3335 * the ones that do not, the alignment can be enforced by enabling
3336 * ECRC generation (if supported).
3337 *
3338 * When PCI-E Completion packets are not aligned, it is actually more
3339 * efficient to limit Read-DMA transactions to 2KB, rather than 4KB.
3340 *
3341 * If the driver can neither enable ECRC nor verify that it has
3342 * already been enabled, then it must use a firmware image which works
3343 * around unaligned completion packets (myri10ge_rss_ethp_z8e.dat), and it
3344 * should also ensure that it never gives the device a Read-DMA which is
3345 * larger than 2KB by setting the tx_boundary to 2KB.  If ECRC is
3346 * enabled, then the driver should use the aligned (myri10ge_rss_eth_z8e.dat)
3347 * firmware image, and set tx_boundary to 4KB.
3348 */
3349
3350static void myri10ge_firmware_probe(struct myri10ge_priv *mgp)
3351{
3352	struct pci_dev *pdev = mgp->pdev;
3353	struct device *dev = &pdev->dev;
3354	int status;
3355
3356	mgp->tx_boundary = 4096;
3357	/*
3358	 * Verify the max read request size was set to 4KB
3359	 * before trying the test with 4KB.
3360	 */
3361	status = pcie_get_readrq(pdev);
3362	if (status < 0) {
3363		dev_err(dev, "Couldn't read max read req size: %d\n", status);
3364		goto abort;
3365	}
3366	if (status != 4096) {
3367		dev_warn(dev, "Max Read Request size != 4096 (%d)\n", status);
3368		mgp->tx_boundary = 2048;
3369	}
3370	/*
3371	 * load the optimized firmware (which assumes aligned PCIe
3372	 * completions) in order to see if it works on this host.
3373	 */
3374	set_fw_name(mgp, myri10ge_fw_aligned, false);
3375	status = myri10ge_load_firmware(mgp, 1);
3376	if (status != 0) {
3377		goto abort;
3378	}
3379
3380	/*
3381	 * Enable ECRC if possible
3382	 */
3383	myri10ge_enable_ecrc(mgp);
3384
3385	/*
3386	 * Run a DMA test which watches for unaligned completions and
3387	 * aborts on the first one seen.
3388	 */
3389
3390	status = myri10ge_dma_test(mgp, MXGEFW_CMD_UNALIGNED_TEST);
3391	if (status == 0)
3392		return;		/* keep the aligned firmware */
3393
3394	if (status != -E2BIG)
3395		dev_warn(dev, "DMA test failed: %d\n", status);
3396	if (status == -ENOSYS)
3397		dev_warn(dev, "Falling back to ethp! "
3398			 "Please install up to date fw\n");
3399abort:
3400	/* fall back to using the unaligned firmware */
3401	mgp->tx_boundary = 2048;
3402	set_fw_name(mgp, myri10ge_fw_unaligned, false);
3403}
3404
3405static void myri10ge_select_firmware(struct myri10ge_priv *mgp)
3406{
3407	int overridden = 0;
3408
3409	if (myri10ge_force_firmware == 0) {
3410		int link_width;
3411		u16 lnk;
3412
3413		pcie_capability_read_word(mgp->pdev, PCI_EXP_LNKSTA, &lnk);
3414		link_width = (lnk >> 4) & 0x3f;
3415
3416		/* Check to see if Link is less than 8 or if the
3417		 * upstream bridge is known to provide aligned
3418		 * completions */
3419		if (link_width < 8) {
3420			dev_info(&mgp->pdev->dev, "PCIE x%d Link\n",
3421				 link_width);
3422			mgp->tx_boundary = 4096;
3423			set_fw_name(mgp, myri10ge_fw_aligned, false);
3424		} else {
3425			myri10ge_firmware_probe(mgp);
3426		}
3427	} else {
3428		if (myri10ge_force_firmware == 1) {
3429			dev_info(&mgp->pdev->dev,
3430				 "Assuming aligned completions (forced)\n");
3431			mgp->tx_boundary = 4096;
3432			set_fw_name(mgp, myri10ge_fw_aligned, false);
3433		} else {
3434			dev_info(&mgp->pdev->dev,
3435				 "Assuming unaligned completions (forced)\n");
3436			mgp->tx_boundary = 2048;
3437			set_fw_name(mgp, myri10ge_fw_unaligned, false);
3438		}
3439	}
3440
3441	kparam_block_sysfs_write(myri10ge_fw_name);
3442	if (myri10ge_fw_name != NULL) {
3443		char *fw_name = kstrdup(myri10ge_fw_name, GFP_KERNEL);
3444		if (fw_name) {
3445			overridden = 1;
3446			set_fw_name(mgp, fw_name, true);
3447		}
3448	}
3449	kparam_unblock_sysfs_write(myri10ge_fw_name);
3450
3451	if (mgp->board_number < MYRI10GE_MAX_BOARDS &&
3452	    myri10ge_fw_names[mgp->board_number] != NULL &&
3453	    strlen(myri10ge_fw_names[mgp->board_number])) {
3454		set_fw_name(mgp, myri10ge_fw_names[mgp->board_number], false);
3455		overridden = 1;
3456	}
3457	if (overridden)
3458		dev_info(&mgp->pdev->dev, "overriding firmware to %s\n",
3459			 mgp->fw_name);
3460}
3461
3462static void myri10ge_mask_surprise_down(struct pci_dev *pdev)
3463{
3464	struct pci_dev *bridge = pdev->bus->self;
3465	int cap;
3466	u32 mask;
3467
3468	if (bridge == NULL)
3469		return;
3470
3471	cap = pci_find_ext_capability(bridge, PCI_EXT_CAP_ID_ERR);
3472	if (cap) {
3473		/* a sram parity error can cause a surprise link
3474		 * down; since we expect and can recover from sram
3475		 * parity errors, mask surprise link down events */
3476		pci_read_config_dword(bridge, cap + PCI_ERR_UNCOR_MASK, &mask);
3477		mask |= 0x20;
3478		pci_write_config_dword(bridge, cap + PCI_ERR_UNCOR_MASK, mask);
3479	}
3480}
3481
3482#ifdef CONFIG_PM
3483static int myri10ge_suspend(struct pci_dev *pdev, pm_message_t state)
3484{
3485	struct myri10ge_priv *mgp;
3486	struct net_device *netdev;
3487
3488	mgp = pci_get_drvdata(pdev);
3489	if (mgp == NULL)
3490		return -EINVAL;
3491	netdev = mgp->dev;
3492
3493	netif_device_detach(netdev);
3494	if (netif_running(netdev)) {
3495		netdev_info(netdev, "closing\n");
3496		rtnl_lock();
3497		myri10ge_close(netdev);
3498		rtnl_unlock();
3499	}
3500	myri10ge_dummy_rdma(mgp, 0);
3501	pci_save_state(pdev);
3502	pci_disable_device(pdev);
3503
3504	return pci_set_power_state(pdev, pci_choose_state(pdev, state));
3505}
3506
3507static int myri10ge_resume(struct pci_dev *pdev)
3508{
3509	struct myri10ge_priv *mgp;
3510	struct net_device *netdev;
3511	int status;
3512	u16 vendor;
3513
3514	mgp = pci_get_drvdata(pdev);
3515	if (mgp == NULL)
3516		return -EINVAL;
3517	netdev = mgp->dev;
3518	pci_set_power_state(pdev, PCI_D0);	/* zeros conf space as a side effect */
3519	msleep(5);		/* give card time to respond */
3520	pci_read_config_word(mgp->pdev, PCI_VENDOR_ID, &vendor);
3521	if (vendor == 0xffff) {
3522		netdev_err(mgp->dev, "device disappeared!\n");
3523		return -EIO;
3524	}
3525
3526	pci_restore_state(pdev);
3527
3528	status = pci_enable_device(pdev);
3529	if (status) {
3530		dev_err(&pdev->dev, "failed to enable device\n");
3531		return status;
3532	}
3533
3534	pci_set_master(pdev);
3535
3536	myri10ge_reset(mgp);
3537	myri10ge_dummy_rdma(mgp, 1);
3538
3539	/* Save configuration space to be restored if the
3540	 * nic resets due to a parity error */
3541	pci_save_state(pdev);
3542
3543	if (netif_running(netdev)) {
3544		rtnl_lock();
3545		status = myri10ge_open(netdev);
3546		rtnl_unlock();
3547		if (status != 0)
3548			goto abort_with_enabled;
3549
3550	}
3551	netif_device_attach(netdev);
3552
3553	return 0;
3554
3555abort_with_enabled:
3556	pci_disable_device(pdev);
3557	return -EIO;
3558
3559}
3560#endif				/* CONFIG_PM */
3561
3562static u32 myri10ge_read_reboot(struct myri10ge_priv *mgp)
3563{
3564	struct pci_dev *pdev = mgp->pdev;
3565	int vs = mgp->vendor_specific_offset;
3566	u32 reboot;
3567
3568	/*enter read32 mode */
3569	pci_write_config_byte(pdev, vs + 0x10, 0x3);
3570
3571	/*read REBOOT_STATUS (0xfffffff0) */
3572	pci_write_config_dword(pdev, vs + 0x18, 0xfffffff0);
3573	pci_read_config_dword(pdev, vs + 0x14, &reboot);
3574	return reboot;
3575}
3576
3577static void
3578myri10ge_check_slice(struct myri10ge_slice_state *ss, int *reset_needed,
3579		     int *busy_slice_cnt, u32 rx_pause_cnt)
3580{
3581	struct myri10ge_priv *mgp = ss->mgp;
3582	int slice = ss - mgp->ss;
3583
3584	if (ss->tx.req != ss->tx.done &&
3585	    ss->tx.done == ss->watchdog_tx_done &&
3586	    ss->watchdog_tx_req != ss->watchdog_tx_done) {
3587		/* nic seems like it might be stuck.. */
3588		if (rx_pause_cnt != mgp->watchdog_pause) {
3589			if (net_ratelimit())
3590				netdev_warn(mgp->dev, "slice %d: TX paused, "
3591					    "check link partner\n", slice);
3592		} else {
3593			netdev_warn(mgp->dev,
3594				    "slice %d: TX stuck %d %d %d %d %d %d\n",
3595				    slice, ss->tx.queue_active, ss->tx.req,
3596				    ss->tx.done, ss->tx.pkt_start,
3597				    ss->tx.pkt_done,
3598				    (int)ntohl(mgp->ss[slice].fw_stats->
3599					       send_done_count));
3600			*reset_needed = 1;
3601			ss->stuck = 1;
3602		}
3603	}
3604	if (ss->watchdog_tx_done != ss->tx.done ||
3605	    ss->watchdog_rx_done != ss->rx_done.cnt) {
3606		*busy_slice_cnt += 1;
3607	}
3608	ss->watchdog_tx_done = ss->tx.done;
3609	ss->watchdog_tx_req = ss->tx.req;
3610	ss->watchdog_rx_done = ss->rx_done.cnt;
3611}
3612
3613/*
3614 * This watchdog is used to check whether the board has suffered
3615 * from a parity error and needs to be recovered.
3616 */
3617static void myri10ge_watchdog(struct work_struct *work)
3618{
3619	struct myri10ge_priv *mgp =
3620	    container_of(work, struct myri10ge_priv, watchdog_work);
3621	struct myri10ge_slice_state *ss;
3622	u32 reboot, rx_pause_cnt;
3623	int status, rebooted;
3624	int i;
3625	int reset_needed = 0;
3626	int busy_slice_cnt = 0;
3627	u16 cmd, vendor;
3628
3629	mgp->watchdog_resets++;
3630	pci_read_config_word(mgp->pdev, PCI_COMMAND, &cmd);
3631	rebooted = 0;
3632	if ((cmd & PCI_COMMAND_MASTER) == 0) {
3633		/* Bus master DMA disabled?  Check to see
3634		 * if the card rebooted due to a parity error
3635		 * For now, just report it */
3636		reboot = myri10ge_read_reboot(mgp);
3637		netdev_err(mgp->dev, "NIC rebooted (0x%x),%s resetting\n",
3638			   reboot, myri10ge_reset_recover ? "" : " not");
3639		if (myri10ge_reset_recover == 0)
3640			return;
3641		rtnl_lock();
3642		mgp->rebooted = 1;
3643		rebooted = 1;
3644		myri10ge_close(mgp->dev);
3645		myri10ge_reset_recover--;
3646		mgp->rebooted = 0;
3647		/*
3648		 * A rebooted nic will come back with config space as
3649		 * it was after power was applied to PCIe bus.
3650		 * Attempt to restore config space which was saved
3651		 * when the driver was loaded, or the last time the
3652		 * nic was resumed from power saving mode.
3653		 */
3654		pci_restore_state(mgp->pdev);
3655
3656		/* save state again for accounting reasons */
3657		pci_save_state(mgp->pdev);
3658
3659	} else {
3660		/* if we get back -1's from our slot, perhaps somebody
3661		 * powered off our card.  Don't try to reset it in
3662		 * this case */
3663		if (cmd == 0xffff) {
3664			pci_read_config_word(mgp->pdev, PCI_VENDOR_ID, &vendor);
3665			if (vendor == 0xffff) {
3666				netdev_err(mgp->dev, "device disappeared!\n");
3667				return;
3668			}
3669		}
3670		/* Perhaps it is a software error. See if stuck slice
3671		 * has recovered, reset if not */
3672		rx_pause_cnt = ntohl(mgp->ss[0].fw_stats->dropped_pause);
3673		for (i = 0; i < mgp->num_slices; i++) {
3674			ss = mgp->ss;
3675			if (ss->stuck) {
3676				myri10ge_check_slice(ss, &reset_needed,
3677						     &busy_slice_cnt,
3678						     rx_pause_cnt);
3679				ss->stuck = 0;
3680			}
3681		}
3682		if (!reset_needed) {
3683			netdev_dbg(mgp->dev, "not resetting\n");
3684			return;
3685		}
3686
3687		netdev_err(mgp->dev, "device timeout, resetting\n");
3688	}
3689
3690	if (!rebooted) {
3691		rtnl_lock();
3692		myri10ge_close(mgp->dev);
3693	}
3694	status = myri10ge_load_firmware(mgp, 1);
3695	if (status != 0)
3696		netdev_err(mgp->dev, "failed to load firmware\n");
3697	else
3698		myri10ge_open(mgp->dev);
3699	rtnl_unlock();
3700}
3701
3702/*
3703 * We use our own timer routine rather than relying upon
3704 * netdev->tx_timeout because we have a very large hardware transmit
3705 * queue.  Due to the large queue, the netdev->tx_timeout function
3706 * cannot detect a NIC with a parity error in a timely fashion if the
3707 * NIC is lightly loaded.
3708 */
3709static void myri10ge_watchdog_timer(unsigned long arg)
3710{
3711	struct myri10ge_priv *mgp;
3712	struct myri10ge_slice_state *ss;
3713	int i, reset_needed, busy_slice_cnt;
3714	u32 rx_pause_cnt;
3715	u16 cmd;
3716
3717	mgp = (struct myri10ge_priv *)arg;
3718
3719	rx_pause_cnt = ntohl(mgp->ss[0].fw_stats->dropped_pause);
3720	busy_slice_cnt = 0;
3721	for (i = 0, reset_needed = 0;
3722	     i < mgp->num_slices && reset_needed == 0; ++i) {
3723
3724		ss = &mgp->ss[i];
3725		if (ss->rx_small.watchdog_needed) {
3726			myri10ge_alloc_rx_pages(mgp, &ss->rx_small,
3727						mgp->small_bytes + MXGEFW_PAD,
3728						1);
3729			if (ss->rx_small.fill_cnt - ss->rx_small.cnt >=
3730			    myri10ge_fill_thresh)
3731				ss->rx_small.watchdog_needed = 0;
3732		}
3733		if (ss->rx_big.watchdog_needed) {
3734			myri10ge_alloc_rx_pages(mgp, &ss->rx_big,
3735						mgp->big_bytes, 1);
3736			if (ss->rx_big.fill_cnt - ss->rx_big.cnt >=
3737			    myri10ge_fill_thresh)
3738				ss->rx_big.watchdog_needed = 0;
3739		}
3740		myri10ge_check_slice(ss, &reset_needed, &busy_slice_cnt,
3741				     rx_pause_cnt);
3742	}
3743	/* if we've sent or received no traffic, poll the NIC to
3744	 * ensure it is still there.  Otherwise, we risk not noticing
3745	 * an error in a timely fashion */
3746	if (busy_slice_cnt == 0) {
3747		pci_read_config_word(mgp->pdev, PCI_COMMAND, &cmd);
3748		if ((cmd & PCI_COMMAND_MASTER) == 0) {
3749			reset_needed = 1;
3750		}
3751	}
3752	mgp->watchdog_pause = rx_pause_cnt;
3753
3754	if (reset_needed) {
3755		schedule_work(&mgp->watchdog_work);
3756	} else {
3757		/* rearm timer */
3758		mod_timer(&mgp->watchdog_timer,
3759			  jiffies + myri10ge_watchdog_timeout * HZ);
3760	}
3761}
3762
3763static void myri10ge_free_slices(struct myri10ge_priv *mgp)
3764{
3765	struct myri10ge_slice_state *ss;
3766	struct pci_dev *pdev = mgp->pdev;
3767	size_t bytes;
3768	int i;
3769
3770	if (mgp->ss == NULL)
3771		return;
3772
3773	for (i = 0; i < mgp->num_slices; i++) {
3774		ss = &mgp->ss[i];
3775		if (ss->rx_done.entry != NULL) {
3776			bytes = mgp->max_intr_slots *
3777			    sizeof(*ss->rx_done.entry);
3778			dma_free_coherent(&pdev->dev, bytes,
3779					  ss->rx_done.entry, ss->rx_done.bus);
3780			ss->rx_done.entry = NULL;
3781		}
3782		if (ss->fw_stats != NULL) {
3783			bytes = sizeof(*ss->fw_stats);
3784			dma_free_coherent(&pdev->dev, bytes,
3785					  ss->fw_stats, ss->fw_stats_bus);
3786			ss->fw_stats = NULL;
3787		}
3788		napi_hash_del(&ss->napi);
3789		netif_napi_del(&ss->napi);
3790	}
3791	/* Wait till napi structs are no longer used, and then free ss. */
3792	synchronize_rcu();
3793	kfree(mgp->ss);
3794	mgp->ss = NULL;
3795}
3796
3797static int myri10ge_alloc_slices(struct myri10ge_priv *mgp)
3798{
3799	struct myri10ge_slice_state *ss;
3800	struct pci_dev *pdev = mgp->pdev;
3801	size_t bytes;
3802	int i;
3803
3804	bytes = sizeof(*mgp->ss) * mgp->num_slices;
3805	mgp->ss = kzalloc(bytes, GFP_KERNEL);
3806	if (mgp->ss == NULL) {
3807		return -ENOMEM;
3808	}
3809
3810	for (i = 0; i < mgp->num_slices; i++) {
3811		ss = &mgp->ss[i];
3812		bytes = mgp->max_intr_slots * sizeof(*ss->rx_done.entry);
3813		ss->rx_done.entry = dma_zalloc_coherent(&pdev->dev, bytes,
3814							&ss->rx_done.bus,
3815							GFP_KERNEL);
3816		if (ss->rx_done.entry == NULL)
3817			goto abort;
3818		bytes = sizeof(*ss->fw_stats);
3819		ss->fw_stats = dma_alloc_coherent(&pdev->dev, bytes,
3820						  &ss->fw_stats_bus,
3821						  GFP_KERNEL);
3822		if (ss->fw_stats == NULL)
3823			goto abort;
3824		ss->mgp = mgp;
3825		ss->dev = mgp->dev;
3826		netif_napi_add(ss->dev, &ss->napi, myri10ge_poll,
3827			       myri10ge_napi_weight);
3828		napi_hash_add(&ss->napi);
3829	}
3830	return 0;
3831abort:
3832	myri10ge_free_slices(mgp);
3833	return -ENOMEM;
3834}
3835
3836/*
3837 * This function determines the number of slices supported.
3838 * The number slices is the minimum of the number of CPUS,
3839 * the number of MSI-X irqs supported, the number of slices
3840 * supported by the firmware
3841 */
3842static void myri10ge_probe_slices(struct myri10ge_priv *mgp)
3843{
3844	struct myri10ge_cmd cmd;
3845	struct pci_dev *pdev = mgp->pdev;
3846	char *old_fw;
3847	bool old_allocated;
3848	int i, status, ncpus;
3849
3850	mgp->num_slices = 1;
3851	ncpus = netif_get_num_default_rss_queues();
3852
3853	if (myri10ge_max_slices == 1 || !pdev->msix_cap ||
3854	    (myri10ge_max_slices == -1 && ncpus < 2))
3855		return;
3856
3857	/* try to load the slice aware rss firmware */
3858	old_fw = mgp->fw_name;
3859	old_allocated = mgp->fw_name_allocated;
3860	/* don't free old_fw if we override it. */
3861	mgp->fw_name_allocated = false;
3862
3863	if (myri10ge_fw_name != NULL) {
3864		dev_info(&mgp->pdev->dev, "overriding rss firmware to %s\n",
3865			 myri10ge_fw_name);
3866		set_fw_name(mgp, myri10ge_fw_name, false);
3867	} else if (old_fw == myri10ge_fw_aligned)
3868		set_fw_name(mgp, myri10ge_fw_rss_aligned, false);
3869	else
3870		set_fw_name(mgp, myri10ge_fw_rss_unaligned, false);
3871	status = myri10ge_load_firmware(mgp, 0);
3872	if (status != 0) {
3873		dev_info(&pdev->dev, "Rss firmware not found\n");
3874		if (old_allocated)
3875			kfree(old_fw);
3876		return;
3877	}
3878
3879	/* hit the board with a reset to ensure it is alive */
3880	memset(&cmd, 0, sizeof(cmd));
3881	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_RESET, &cmd, 0);
3882	if (status != 0) {
3883		dev_err(&mgp->pdev->dev, "failed reset\n");
3884		goto abort_with_fw;
3885	}
3886
3887	mgp->max_intr_slots = cmd.data0 / sizeof(struct mcp_slot);
3888
3889	/* tell it the size of the interrupt queues */
3890	cmd.data0 = mgp->max_intr_slots * sizeof(struct mcp_slot);
3891	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_INTRQ_SIZE, &cmd, 0);
3892	if (status != 0) {
3893		dev_err(&mgp->pdev->dev, "failed MXGEFW_CMD_SET_INTRQ_SIZE\n");
3894		goto abort_with_fw;
3895	}
3896
3897	/* ask the maximum number of slices it supports */
3898	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_MAX_RSS_QUEUES, &cmd, 0);
3899	if (status != 0)
3900		goto abort_with_fw;
3901	else
3902		mgp->num_slices = cmd.data0;
3903
3904	/* Only allow multiple slices if MSI-X is usable */
3905	if (!myri10ge_msi) {
3906		goto abort_with_fw;
3907	}
3908
3909	/* if the admin did not specify a limit to how many
3910	 * slices we should use, cap it automatically to the
3911	 * number of CPUs currently online */
3912	if (myri10ge_max_slices == -1)
3913		myri10ge_max_slices = ncpus;
3914
3915	if (mgp->num_slices > myri10ge_max_slices)
3916		mgp->num_slices = myri10ge_max_slices;
3917
3918	/* Now try to allocate as many MSI-X vectors as we have
3919	 * slices. We give up on MSI-X if we can only get a single
3920	 * vector. */
3921
3922	mgp->msix_vectors = kcalloc(mgp->num_slices, sizeof(*mgp->msix_vectors),
3923				    GFP_KERNEL);
3924	if (mgp->msix_vectors == NULL)
3925		goto no_msix;
3926	for (i = 0; i < mgp->num_slices; i++) {
3927		mgp->msix_vectors[i].entry = i;
3928	}
3929
3930	while (mgp->num_slices > 1) {
3931		mgp->num_slices = rounddown_pow_of_two(mgp->num_slices);
3932		if (mgp->num_slices == 1)
3933			goto no_msix;
3934		status = pci_enable_msix_range(pdev,
3935					       mgp->msix_vectors,
3936					       mgp->num_slices,
3937					       mgp->num_slices);
3938		if (status < 0)
3939			goto no_msix;
3940
3941		pci_disable_msix(pdev);
3942
3943		if (status == mgp->num_slices) {
3944			if (old_allocated)
3945				kfree(old_fw);
3946			return;
3947		} else {
3948			mgp->num_slices = status;
3949		}
3950	}
3951
3952no_msix:
3953	if (mgp->msix_vectors != NULL) {
3954		kfree(mgp->msix_vectors);
3955		mgp->msix_vectors = NULL;
3956	}
3957
3958abort_with_fw:
3959	mgp->num_slices = 1;
3960	set_fw_name(mgp, old_fw, old_allocated);
3961	myri10ge_load_firmware(mgp, 0);
3962}
3963
3964static const struct net_device_ops myri10ge_netdev_ops = {
3965	.ndo_open		= myri10ge_open,
3966	.ndo_stop		= myri10ge_close,
3967	.ndo_start_xmit		= myri10ge_xmit,
3968	.ndo_get_stats64	= myri10ge_get_stats,
3969	.ndo_validate_addr	= eth_validate_addr,
3970	.ndo_change_mtu		= myri10ge_change_mtu,
3971	.ndo_set_rx_mode	= myri10ge_set_multicast_list,
3972	.ndo_set_mac_address	= myri10ge_set_mac_address,
3973#ifdef CONFIG_NET_RX_BUSY_POLL
3974	.ndo_busy_poll		= myri10ge_busy_poll,
3975#endif
3976};
3977
3978static int myri10ge_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
3979{
3980	struct net_device *netdev;
3981	struct myri10ge_priv *mgp;
3982	struct device *dev = &pdev->dev;
3983	int i;
3984	int status = -ENXIO;
3985	int dac_enabled;
3986	unsigned hdr_offset, ss_offset;
3987	static int board_number;
3988
3989	netdev = alloc_etherdev_mq(sizeof(*mgp), MYRI10GE_MAX_SLICES);
3990	if (netdev == NULL)
3991		return -ENOMEM;
3992
3993	SET_NETDEV_DEV(netdev, &pdev->dev);
3994
3995	mgp = netdev_priv(netdev);
3996	mgp->dev = netdev;
3997	mgp->pdev = pdev;
3998	mgp->pause = myri10ge_flow_control;
3999	mgp->intr_coal_delay = myri10ge_intr_coal_delay;
4000	mgp->msg_enable = netif_msg_init(myri10ge_debug, MYRI10GE_MSG_DEFAULT);
4001	mgp->board_number = board_number;
4002	init_waitqueue_head(&mgp->down_wq);
4003
4004	if (pci_enable_device(pdev)) {
4005		dev_err(&pdev->dev, "pci_enable_device call failed\n");
4006		status = -ENODEV;
4007		goto abort_with_netdev;
4008	}
4009
4010	/* Find the vendor-specific cap so we can check
4011	 * the reboot register later on */
4012	mgp->vendor_specific_offset
4013	    = pci_find_capability(pdev, PCI_CAP_ID_VNDR);
4014
4015	/* Set our max read request to 4KB */
4016	status = pcie_set_readrq(pdev, 4096);
4017	if (status != 0) {
4018		dev_err(&pdev->dev, "Error %d writing PCI_EXP_DEVCTL\n",
4019			status);
4020		goto abort_with_enabled;
4021	}
4022
4023	myri10ge_mask_surprise_down(pdev);
4024	pci_set_master(pdev);
4025	dac_enabled = 1;
4026	status = pci_set_dma_mask(pdev, DMA_BIT_MASK(64));
4027	if (status != 0) {
4028		dac_enabled = 0;
4029		dev_err(&pdev->dev,
4030			"64-bit pci address mask was refused, "
4031			"trying 32-bit\n");
4032		status = pci_set_dma_mask(pdev, DMA_BIT_MASK(32));
4033	}
4034	if (status != 0) {
4035		dev_err(&pdev->dev, "Error %d setting DMA mask\n", status);
4036		goto abort_with_enabled;
4037	}
4038	(void)pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64));
4039	mgp->cmd = dma_alloc_coherent(&pdev->dev, sizeof(*mgp->cmd),
4040				      &mgp->cmd_bus, GFP_KERNEL);
4041	if (mgp->cmd == NULL)
4042		goto abort_with_enabled;
4043
4044	mgp->board_span = pci_resource_len(pdev, 0);
4045	mgp->iomem_base = pci_resource_start(pdev, 0);
4046	mgp->mtrr = -1;
4047	mgp->wc_enabled = 0;
4048#ifdef CONFIG_MTRR
4049	mgp->mtrr = mtrr_add(mgp->iomem_base, mgp->board_span,
4050			     MTRR_TYPE_WRCOMB, 1);
4051	if (mgp->mtrr >= 0)
4052		mgp->wc_enabled = 1;
4053#endif
4054	mgp->sram = ioremap_wc(mgp->iomem_base, mgp->board_span);
4055	if (mgp->sram == NULL) {
4056		dev_err(&pdev->dev, "ioremap failed for %ld bytes at 0x%lx\n",
4057			mgp->board_span, mgp->iomem_base);
4058		status = -ENXIO;
4059		goto abort_with_mtrr;
4060	}
4061	hdr_offset =
4062	    swab32(readl(mgp->sram + MCP_HEADER_PTR_OFFSET)) & 0xffffc;
4063	ss_offset = hdr_offset + offsetof(struct mcp_gen_header, string_specs);
4064	mgp->sram_size = swab32(readl(mgp->sram + ss_offset));
4065	if (mgp->sram_size > mgp->board_span ||
4066	    mgp->sram_size <= MYRI10GE_FW_OFFSET) {
4067		dev_err(&pdev->dev,
4068			"invalid sram_size %dB or board span %ldB\n",
4069			mgp->sram_size, mgp->board_span);
4070		goto abort_with_ioremap;
4071	}
4072	memcpy_fromio(mgp->eeprom_strings,
4073		      mgp->sram + mgp->sram_size, MYRI10GE_EEPROM_STRINGS_SIZE);
4074	memset(mgp->eeprom_strings + MYRI10GE_EEPROM_STRINGS_SIZE - 2, 0, 2);
4075	status = myri10ge_read_mac_addr(mgp);
4076	if (status)
4077		goto abort_with_ioremap;
4078
4079	for (i = 0; i < ETH_ALEN; i++)
4080		netdev->dev_addr[i] = mgp->mac_addr[i];
4081
4082	myri10ge_select_firmware(mgp);
4083
4084	status = myri10ge_load_firmware(mgp, 1);
4085	if (status != 0) {
4086		dev_err(&pdev->dev, "failed to load firmware\n");
4087		goto abort_with_ioremap;
4088	}
4089	myri10ge_probe_slices(mgp);
4090	status = myri10ge_alloc_slices(mgp);
4091	if (status != 0) {
4092		dev_err(&pdev->dev, "failed to alloc slice state\n");
4093		goto abort_with_firmware;
4094	}
4095	netif_set_real_num_tx_queues(netdev, mgp->num_slices);
4096	netif_set_real_num_rx_queues(netdev, mgp->num_slices);
4097	status = myri10ge_reset(mgp);
4098	if (status != 0) {
4099		dev_err(&pdev->dev, "failed reset\n");
4100		goto abort_with_slices;
4101	}
4102#ifdef CONFIG_MYRI10GE_DCA
4103	myri10ge_setup_dca(mgp);
4104#endif
4105	pci_set_drvdata(pdev, mgp);
4106	if ((myri10ge_initial_mtu + ETH_HLEN) > MYRI10GE_MAX_ETHER_MTU)
4107		myri10ge_initial_mtu = MYRI10GE_MAX_ETHER_MTU - ETH_HLEN;
4108	if ((myri10ge_initial_mtu + ETH_HLEN) < 68)
4109		myri10ge_initial_mtu = 68;
4110
4111	netdev->netdev_ops = &myri10ge_netdev_ops;
4112	netdev->mtu = myri10ge_initial_mtu;
4113	netdev->hw_features = mgp->features | NETIF_F_RXCSUM;
4114
4115	/* fake NETIF_F_HW_VLAN_CTAG_RX for good GRO performance */
4116	netdev->hw_features |= NETIF_F_HW_VLAN_CTAG_RX;
4117
4118	netdev->features = netdev->hw_features;
4119
4120	if (dac_enabled)
4121		netdev->features |= NETIF_F_HIGHDMA;
4122
4123	netdev->vlan_features |= mgp->features;
4124	if (mgp->fw_ver_tiny < 37)
4125		netdev->vlan_features &= ~NETIF_F_TSO6;
4126	if (mgp->fw_ver_tiny < 32)
4127		netdev->vlan_features &= ~NETIF_F_TSO;
4128
4129	/* make sure we can get an irq, and that MSI can be
4130	 * setup (if available). */
4131	status = myri10ge_request_irq(mgp);
4132	if (status != 0)
4133		goto abort_with_firmware;
4134	myri10ge_free_irq(mgp);
4135
4136	/* Save configuration space to be restored if the
4137	 * nic resets due to a parity error */
4138	pci_save_state(pdev);
4139
4140	/* Setup the watchdog timer */
4141	setup_timer(&mgp->watchdog_timer, myri10ge_watchdog_timer,
4142		    (unsigned long)mgp);
4143
4144	netdev->ethtool_ops = &myri10ge_ethtool_ops;
4145	INIT_WORK(&mgp->watchdog_work, myri10ge_watchdog);
4146	status = register_netdev(netdev);
4147	if (status != 0) {
4148		dev_err(&pdev->dev, "register_netdev failed: %d\n", status);
4149		goto abort_with_state;
4150	}
4151	if (mgp->msix_enabled)
4152		dev_info(dev, "%d MSI-X IRQs, tx bndry %d, fw %s, WC %s\n",
4153			 mgp->num_slices, mgp->tx_boundary, mgp->fw_name,
4154			 (mgp->wc_enabled ? "Enabled" : "Disabled"));
4155	else
4156		dev_info(dev, "%s IRQ %d, tx bndry %d, fw %s, WC %s\n",
4157			 mgp->msi_enabled ? "MSI" : "xPIC",
4158			 pdev->irq, mgp->tx_boundary, mgp->fw_name,
4159			 (mgp->wc_enabled ? "Enabled" : "Disabled"));
4160
4161	board_number++;
4162	return 0;
4163
4164abort_with_state:
4165	pci_restore_state(pdev);
4166
4167abort_with_slices:
4168	myri10ge_free_slices(mgp);
4169
4170abort_with_firmware:
4171	myri10ge_dummy_rdma(mgp, 0);
4172
4173abort_with_ioremap:
4174	if (mgp->mac_addr_string != NULL)
4175		dev_err(&pdev->dev,
4176			"myri10ge_probe() failed: MAC=%s, SN=%ld\n",
4177			mgp->mac_addr_string, mgp->serial_number);
4178	iounmap(mgp->sram);
4179
4180abort_with_mtrr:
4181#ifdef CONFIG_MTRR
4182	if (mgp->mtrr >= 0)
4183		mtrr_del(mgp->mtrr, mgp->iomem_base, mgp->board_span);
4184#endif
4185	dma_free_coherent(&pdev->dev, sizeof(*mgp->cmd),
4186			  mgp->cmd, mgp->cmd_bus);
4187
4188abort_with_enabled:
4189	pci_disable_device(pdev);
4190
4191abort_with_netdev:
4192	set_fw_name(mgp, NULL, false);
4193	free_netdev(netdev);
4194	return status;
4195}
4196
4197/*
4198 * myri10ge_remove
4199 *
4200 * Does what is necessary to shutdown one Myrinet device. Called
4201 *   once for each Myrinet card by the kernel when a module is
4202 *   unloaded.
4203 */
4204static void myri10ge_remove(struct pci_dev *pdev)
4205{
4206	struct myri10ge_priv *mgp;
4207	struct net_device *netdev;
4208
4209	mgp = pci_get_drvdata(pdev);
4210	if (mgp == NULL)
4211		return;
4212
4213	cancel_work_sync(&mgp->watchdog_work);
4214	netdev = mgp->dev;
4215	unregister_netdev(netdev);
4216
4217#ifdef CONFIG_MYRI10GE_DCA
4218	myri10ge_teardown_dca(mgp);
4219#endif
4220	myri10ge_dummy_rdma(mgp, 0);
4221
4222	/* avoid a memory leak */
4223	pci_restore_state(pdev);
4224
4225	iounmap(mgp->sram);
4226
4227#ifdef CONFIG_MTRR
4228	if (mgp->mtrr >= 0)
4229		mtrr_del(mgp->mtrr, mgp->iomem_base, mgp->board_span);
4230#endif
4231	myri10ge_free_slices(mgp);
4232	if (mgp->msix_vectors != NULL)
4233		kfree(mgp->msix_vectors);
4234	dma_free_coherent(&pdev->dev, sizeof(*mgp->cmd),
4235			  mgp->cmd, mgp->cmd_bus);
4236
4237	set_fw_name(mgp, NULL, false);
4238	free_netdev(netdev);
4239	pci_disable_device(pdev);
4240}
4241
4242#define PCI_DEVICE_ID_MYRICOM_MYRI10GE_Z8E 	0x0008
4243#define PCI_DEVICE_ID_MYRICOM_MYRI10GE_Z8E_9	0x0009
4244
4245static const struct pci_device_id myri10ge_pci_tbl[] = {
4246	{PCI_DEVICE(PCI_VENDOR_ID_MYRICOM, PCI_DEVICE_ID_MYRICOM_MYRI10GE_Z8E)},
4247	{PCI_DEVICE
4248	 (PCI_VENDOR_ID_MYRICOM, PCI_DEVICE_ID_MYRICOM_MYRI10GE_Z8E_9)},
4249	{0},
4250};
4251
4252MODULE_DEVICE_TABLE(pci, myri10ge_pci_tbl);
4253
4254static struct pci_driver myri10ge_driver = {
4255	.name = "myri10ge",
4256	.probe = myri10ge_probe,
4257	.remove = myri10ge_remove,
4258	.id_table = myri10ge_pci_tbl,
4259#ifdef CONFIG_PM
4260	.suspend = myri10ge_suspend,
4261	.resume = myri10ge_resume,
4262#endif
4263};
4264
4265#ifdef CONFIG_MYRI10GE_DCA
4266static int
4267myri10ge_notify_dca(struct notifier_block *nb, unsigned long event, void *p)
4268{
4269	int err = driver_for_each_device(&myri10ge_driver.driver,
4270					 NULL, &event,
4271					 myri10ge_notify_dca_device);
4272
4273	if (err)
4274		return NOTIFY_BAD;
4275	return NOTIFY_DONE;
4276}
4277
4278static struct notifier_block myri10ge_dca_notifier = {
4279	.notifier_call = myri10ge_notify_dca,
4280	.next = NULL,
4281	.priority = 0,
4282};
4283#endif				/* CONFIG_MYRI10GE_DCA */
4284
4285static __init int myri10ge_init_module(void)
4286{
4287	pr_info("Version %s\n", MYRI10GE_VERSION_STR);
4288
4289	if (myri10ge_rss_hash > MXGEFW_RSS_HASH_TYPE_MAX) {
4290		pr_err("Illegal rssh hash type %d, defaulting to source port\n",
4291		       myri10ge_rss_hash);
4292		myri10ge_rss_hash = MXGEFW_RSS_HASH_TYPE_SRC_PORT;
4293	}
4294#ifdef CONFIG_MYRI10GE_DCA
4295	dca_register_notify(&myri10ge_dca_notifier);
4296#endif
4297	if (myri10ge_max_slices > MYRI10GE_MAX_SLICES)
4298		myri10ge_max_slices = MYRI10GE_MAX_SLICES;
4299
4300	return pci_register_driver(&myri10ge_driver);
4301}
4302
4303module_init(myri10ge_init_module);
4304
4305static __exit void myri10ge_cleanup_module(void)
4306{
4307#ifdef CONFIG_MYRI10GE_DCA
4308	dca_unregister_notify(&myri10ge_dca_notifier);
4309#endif
4310	pci_unregister_driver(&myri10ge_driver);
4311}
4312
4313module_exit(myri10ge_cleanup_module);
4314