1/*************************************************************************
2 * myri10ge.c: Myricom Myri-10G Ethernet driver.
3 *
4 * Copyright (C) 2005 - 2011 Myricom, Inc.
5 * All rights reserved.
6 *
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
9 * are met:
10 * 1. Redistributions of source code must retain the above copyright
11 *    notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 *    notice, this list of conditions and the following disclaimer in the
14 *    documentation and/or other materials provided with the distribution.
15 * 3. Neither the name of Myricom, Inc. nor the names of its contributors
16 *    may be used to endorse or promote products derived from this software
17 *    without specific prior written permission.
18 *
19 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
20 * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22 * ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
23 * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29 * POSSIBILITY OF SUCH DAMAGE.
30 *
31 *
32 * If the eeprom on your board is not recent enough, you will need to get a
33 * newer firmware image at:
34 *   http://www.myri.com/scs/download-Myri10GE.html
35 *
36 * Contact Information:
37 *   <help@myri.com>
38 *   Myricom, Inc., 325N Santa Anita Avenue, Arcadia, CA 91006
39 *************************************************************************/
40
41#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
42
43#include <linux/tcp.h>
44#include <linux/netdevice.h>
45#include <linux/skbuff.h>
46#include <linux/string.h>
47#include <linux/module.h>
48#include <linux/pci.h>
49#include <linux/dma-mapping.h>
50#include <linux/etherdevice.h>
51#include <linux/if_ether.h>
52#include <linux/if_vlan.h>
53#include <linux/dca.h>
54#include <linux/ip.h>
55#include <linux/inet.h>
56#include <linux/in.h>
57#include <linux/ethtool.h>
58#include <linux/firmware.h>
59#include <linux/delay.h>
60#include <linux/timer.h>
61#include <linux/vmalloc.h>
62#include <linux/crc32.h>
63#include <linux/moduleparam.h>
64#include <linux/io.h>
65#include <linux/log2.h>
66#include <linux/slab.h>
67#include <linux/prefetch.h>
68#include <net/checksum.h>
69#include <net/ip.h>
70#include <net/tcp.h>
71#include <asm/byteorder.h>
72#include <asm/processor.h>
73#include <net/busy_poll.h>
74
75#include "myri10ge_mcp.h"
76#include "myri10ge_mcp_gen_header.h"
77
78#define MYRI10GE_VERSION_STR "1.5.3-1.534"
79
80MODULE_DESCRIPTION("Myricom 10G driver (10GbE)");
81MODULE_AUTHOR("Maintainer: help@myri.com");
82MODULE_VERSION(MYRI10GE_VERSION_STR);
83MODULE_LICENSE("Dual BSD/GPL");
84
85#define MYRI10GE_MAX_ETHER_MTU 9014
86
87#define MYRI10GE_ETH_STOPPED 0
88#define MYRI10GE_ETH_STOPPING 1
89#define MYRI10GE_ETH_STARTING 2
90#define MYRI10GE_ETH_RUNNING 3
91#define MYRI10GE_ETH_OPEN_FAILED 4
92
93#define MYRI10GE_EEPROM_STRINGS_SIZE 256
94#define MYRI10GE_MAX_SEND_DESC_TSO ((65536 / 2048) * 2)
95
96#define MYRI10GE_NO_CONFIRM_DATA htonl(0xffffffff)
97#define MYRI10GE_NO_RESPONSE_RESULT 0xffffffff
98
99#define MYRI10GE_ALLOC_ORDER 0
100#define MYRI10GE_ALLOC_SIZE ((1 << MYRI10GE_ALLOC_ORDER) * PAGE_SIZE)
101#define MYRI10GE_MAX_FRAGS_PER_FRAME (MYRI10GE_MAX_ETHER_MTU/MYRI10GE_ALLOC_SIZE + 1)
102
103#define MYRI10GE_MAX_SLICES 32
104
105struct myri10ge_rx_buffer_state {
106	struct page *page;
107	int page_offset;
108	DEFINE_DMA_UNMAP_ADDR(bus);
109	DEFINE_DMA_UNMAP_LEN(len);
110};
111
112struct myri10ge_tx_buffer_state {
113	struct sk_buff *skb;
114	int last;
115	DEFINE_DMA_UNMAP_ADDR(bus);
116	DEFINE_DMA_UNMAP_LEN(len);
117};
118
119struct myri10ge_cmd {
120	u32 data0;
121	u32 data1;
122	u32 data2;
123};
124
125struct myri10ge_rx_buf {
126	struct mcp_kreq_ether_recv __iomem *lanai;	/* lanai ptr for recv ring */
127	struct mcp_kreq_ether_recv *shadow;	/* host shadow of recv ring */
128	struct myri10ge_rx_buffer_state *info;
129	struct page *page;
130	dma_addr_t bus;
131	int page_offset;
132	int cnt;
133	int fill_cnt;
134	int alloc_fail;
135	int mask;		/* number of rx slots -1 */
136	int watchdog_needed;
137};
138
139struct myri10ge_tx_buf {
140	struct mcp_kreq_ether_send __iomem *lanai;	/* lanai ptr for sendq */
141	__be32 __iomem *send_go;	/* "go" doorbell ptr */
142	__be32 __iomem *send_stop;	/* "stop" doorbell ptr */
143	struct mcp_kreq_ether_send *req_list;	/* host shadow of sendq */
144	char *req_bytes;
145	struct myri10ge_tx_buffer_state *info;
146	int mask;		/* number of transmit slots -1  */
147	int req ____cacheline_aligned;	/* transmit slots submitted     */
148	int pkt_start;		/* packets started */
149	int stop_queue;
150	int linearized;
151	int done ____cacheline_aligned;	/* transmit slots completed     */
152	int pkt_done;		/* packets completed */
153	int wake_queue;
154	int queue_active;
155};
156
157struct myri10ge_rx_done {
158	struct mcp_slot *entry;
159	dma_addr_t bus;
160	int cnt;
161	int idx;
162};
163
164struct myri10ge_slice_netstats {
165	unsigned long rx_packets;
166	unsigned long tx_packets;
167	unsigned long rx_bytes;
168	unsigned long tx_bytes;
169	unsigned long rx_dropped;
170	unsigned long tx_dropped;
171};
172
173struct myri10ge_slice_state {
174	struct myri10ge_tx_buf tx;	/* transmit ring        */
175	struct myri10ge_rx_buf rx_small;
176	struct myri10ge_rx_buf rx_big;
177	struct myri10ge_rx_done rx_done;
178	struct net_device *dev;
179	struct napi_struct napi;
180	struct myri10ge_priv *mgp;
181	struct myri10ge_slice_netstats stats;
182	__be32 __iomem *irq_claim;
183	struct mcp_irq_data *fw_stats;
184	dma_addr_t fw_stats_bus;
185	int watchdog_tx_done;
186	int watchdog_tx_req;
187	int watchdog_rx_done;
188	int stuck;
189#ifdef CONFIG_MYRI10GE_DCA
190	int cached_dca_tag;
191	int cpu;
192	__be32 __iomem *dca_tag;
193#endif
194#ifdef CONFIG_NET_RX_BUSY_POLL
195	unsigned int state;
196#define SLICE_STATE_IDLE	0
197#define SLICE_STATE_NAPI	1	/* NAPI owns this slice */
198#define SLICE_STATE_POLL	2	/* poll owns this slice */
199#define SLICE_LOCKED (SLICE_STATE_NAPI | SLICE_STATE_POLL)
200#define SLICE_STATE_NAPI_YIELD	4	/* NAPI yielded this slice */
201#define SLICE_STATE_POLL_YIELD	8	/* poll yielded this slice */
202#define SLICE_USER_PEND (SLICE_STATE_POLL | SLICE_STATE_POLL_YIELD)
203	spinlock_t lock;
204	unsigned long lock_napi_yield;
205	unsigned long lock_poll_yield;
206	unsigned long busy_poll_miss;
207	unsigned long busy_poll_cnt;
208#endif /* CONFIG_NET_RX_BUSY_POLL */
209	char irq_desc[32];
210};
211
212struct myri10ge_priv {
213	struct myri10ge_slice_state *ss;
214	int tx_boundary;	/* boundary transmits cannot cross */
215	int num_slices;
216	int running;		/* running?             */
217	int small_bytes;
218	int big_bytes;
219	int max_intr_slots;
220	struct net_device *dev;
221	u8 __iomem *sram;
222	int sram_size;
223	unsigned long board_span;
224	unsigned long iomem_base;
225	__be32 __iomem *irq_deassert;
226	char *mac_addr_string;
227	struct mcp_cmd_response *cmd;
228	dma_addr_t cmd_bus;
229	struct pci_dev *pdev;
230	int msi_enabled;
231	int msix_enabled;
232	struct msix_entry *msix_vectors;
233#ifdef CONFIG_MYRI10GE_DCA
234	int dca_enabled;
235	int relaxed_order;
236#endif
237	u32 link_state;
238	unsigned int rdma_tags_available;
239	int intr_coal_delay;
240	__be32 __iomem *intr_coal_delay_ptr;
241	int wc_cookie;
242	int down_cnt;
243	wait_queue_head_t down_wq;
244	struct work_struct watchdog_work;
245	struct timer_list watchdog_timer;
246	int watchdog_resets;
247	int watchdog_pause;
248	int pause;
249	bool fw_name_allocated;
250	char *fw_name;
251	char eeprom_strings[MYRI10GE_EEPROM_STRINGS_SIZE];
252	char *product_code_string;
253	char fw_version[128];
254	int fw_ver_major;
255	int fw_ver_minor;
256	int fw_ver_tiny;
257	int adopted_rx_filter_bug;
258	u8 mac_addr[ETH_ALEN];		/* eeprom mac address */
259	unsigned long serial_number;
260	int vendor_specific_offset;
261	int fw_multicast_support;
262	u32 features;
263	u32 max_tso6;
264	u32 read_dma;
265	u32 write_dma;
266	u32 read_write_dma;
267	u32 link_changes;
268	u32 msg_enable;
269	unsigned int board_number;
270	int rebooted;
271};
272
273static char *myri10ge_fw_unaligned = "myri10ge_ethp_z8e.dat";
274static char *myri10ge_fw_aligned = "myri10ge_eth_z8e.dat";
275static char *myri10ge_fw_rss_unaligned = "myri10ge_rss_ethp_z8e.dat";
276static char *myri10ge_fw_rss_aligned = "myri10ge_rss_eth_z8e.dat";
277MODULE_FIRMWARE("myri10ge_ethp_z8e.dat");
278MODULE_FIRMWARE("myri10ge_eth_z8e.dat");
279MODULE_FIRMWARE("myri10ge_rss_ethp_z8e.dat");
280MODULE_FIRMWARE("myri10ge_rss_eth_z8e.dat");
281
282/* Careful: must be accessed under kparam_block_sysfs_write */
283static char *myri10ge_fw_name = NULL;
284module_param(myri10ge_fw_name, charp, S_IRUGO | S_IWUSR);
285MODULE_PARM_DESC(myri10ge_fw_name, "Firmware image name");
286
287#define MYRI10GE_MAX_BOARDS 8
288static char *myri10ge_fw_names[MYRI10GE_MAX_BOARDS] =
289    {[0 ... (MYRI10GE_MAX_BOARDS - 1)] = NULL };
290module_param_array_named(myri10ge_fw_names, myri10ge_fw_names, charp, NULL,
291			 0444);
292MODULE_PARM_DESC(myri10ge_fw_name, "Firmware image names per board");
293
294static int myri10ge_ecrc_enable = 1;
295module_param(myri10ge_ecrc_enable, int, S_IRUGO);
296MODULE_PARM_DESC(myri10ge_ecrc_enable, "Enable Extended CRC on PCI-E");
297
298static int myri10ge_small_bytes = -1;	/* -1 == auto */
299module_param(myri10ge_small_bytes, int, S_IRUGO | S_IWUSR);
300MODULE_PARM_DESC(myri10ge_small_bytes, "Threshold of small packets");
301
302static int myri10ge_msi = 1;	/* enable msi by default */
303module_param(myri10ge_msi, int, S_IRUGO | S_IWUSR);
304MODULE_PARM_DESC(myri10ge_msi, "Enable Message Signalled Interrupts");
305
306static int myri10ge_intr_coal_delay = 75;
307module_param(myri10ge_intr_coal_delay, int, S_IRUGO);
308MODULE_PARM_DESC(myri10ge_intr_coal_delay, "Interrupt coalescing delay");
309
310static int myri10ge_flow_control = 1;
311module_param(myri10ge_flow_control, int, S_IRUGO);
312MODULE_PARM_DESC(myri10ge_flow_control, "Pause parameter");
313
314static int myri10ge_deassert_wait = 1;
315module_param(myri10ge_deassert_wait, int, S_IRUGO | S_IWUSR);
316MODULE_PARM_DESC(myri10ge_deassert_wait,
317		 "Wait when deasserting legacy interrupts");
318
319static int myri10ge_force_firmware = 0;
320module_param(myri10ge_force_firmware, int, S_IRUGO);
321MODULE_PARM_DESC(myri10ge_force_firmware,
322		 "Force firmware to assume aligned completions");
323
324static int myri10ge_initial_mtu = MYRI10GE_MAX_ETHER_MTU - ETH_HLEN;
325module_param(myri10ge_initial_mtu, int, S_IRUGO);
326MODULE_PARM_DESC(myri10ge_initial_mtu, "Initial MTU");
327
328static int myri10ge_napi_weight = 64;
329module_param(myri10ge_napi_weight, int, S_IRUGO);
330MODULE_PARM_DESC(myri10ge_napi_weight, "Set NAPI weight");
331
332static int myri10ge_watchdog_timeout = 1;
333module_param(myri10ge_watchdog_timeout, int, S_IRUGO);
334MODULE_PARM_DESC(myri10ge_watchdog_timeout, "Set watchdog timeout");
335
336static int myri10ge_max_irq_loops = 1048576;
337module_param(myri10ge_max_irq_loops, int, S_IRUGO);
338MODULE_PARM_DESC(myri10ge_max_irq_loops,
339		 "Set stuck legacy IRQ detection threshold");
340
341#define MYRI10GE_MSG_DEFAULT NETIF_MSG_LINK
342
343static int myri10ge_debug = -1;	/* defaults above */
344module_param(myri10ge_debug, int, 0);
345MODULE_PARM_DESC(myri10ge_debug, "Debug level (0=none,...,16=all)");
346
347static int myri10ge_fill_thresh = 256;
348module_param(myri10ge_fill_thresh, int, S_IRUGO | S_IWUSR);
349MODULE_PARM_DESC(myri10ge_fill_thresh, "Number of empty rx slots allowed");
350
351static int myri10ge_reset_recover = 1;
352
353static int myri10ge_max_slices = 1;
354module_param(myri10ge_max_slices, int, S_IRUGO);
355MODULE_PARM_DESC(myri10ge_max_slices, "Max tx/rx queues");
356
357static int myri10ge_rss_hash = MXGEFW_RSS_HASH_TYPE_SRC_DST_PORT;
358module_param(myri10ge_rss_hash, int, S_IRUGO);
359MODULE_PARM_DESC(myri10ge_rss_hash, "Type of RSS hashing to do");
360
361static int myri10ge_dca = 1;
362module_param(myri10ge_dca, int, S_IRUGO);
363MODULE_PARM_DESC(myri10ge_dca, "Enable DCA if possible");
364
365#define MYRI10GE_FW_OFFSET 1024*1024
366#define MYRI10GE_HIGHPART_TO_U32(X) \
367(sizeof (X) == 8) ? ((u32)((u64)(X) >> 32)) : (0)
368#define MYRI10GE_LOWPART_TO_U32(X) ((u32)(X))
369
370#define myri10ge_pio_copy(to,from,size) __iowrite64_copy(to,from,size/8)
371
372static void myri10ge_set_multicast_list(struct net_device *dev);
373static netdev_tx_t myri10ge_sw_tso(struct sk_buff *skb,
374					 struct net_device *dev);
375
376static inline void put_be32(__be32 val, __be32 __iomem * p)
377{
378	__raw_writel((__force __u32) val, (__force void __iomem *)p);
379}
380
381static struct rtnl_link_stats64 *myri10ge_get_stats(struct net_device *dev,
382						    struct rtnl_link_stats64 *stats);
383
384static void set_fw_name(struct myri10ge_priv *mgp, char *name, bool allocated)
385{
386	if (mgp->fw_name_allocated)
387		kfree(mgp->fw_name);
388	mgp->fw_name = name;
389	mgp->fw_name_allocated = allocated;
390}
391
392static int
393myri10ge_send_cmd(struct myri10ge_priv *mgp, u32 cmd,
394		  struct myri10ge_cmd *data, int atomic)
395{
396	struct mcp_cmd *buf;
397	char buf_bytes[sizeof(*buf) + 8];
398	struct mcp_cmd_response *response = mgp->cmd;
399	char __iomem *cmd_addr = mgp->sram + MXGEFW_ETH_CMD;
400	u32 dma_low, dma_high, result, value;
401	int sleep_total = 0;
402
403	/* ensure buf is aligned to 8 bytes */
404	buf = (struct mcp_cmd *)ALIGN((unsigned long)buf_bytes, 8);
405
406	buf->data0 = htonl(data->data0);
407	buf->data1 = htonl(data->data1);
408	buf->data2 = htonl(data->data2);
409	buf->cmd = htonl(cmd);
410	dma_low = MYRI10GE_LOWPART_TO_U32(mgp->cmd_bus);
411	dma_high = MYRI10GE_HIGHPART_TO_U32(mgp->cmd_bus);
412
413	buf->response_addr.low = htonl(dma_low);
414	buf->response_addr.high = htonl(dma_high);
415	response->result = htonl(MYRI10GE_NO_RESPONSE_RESULT);
416	mb();
417	myri10ge_pio_copy(cmd_addr, buf, sizeof(*buf));
418
419	/* wait up to 15ms. Longest command is the DMA benchmark,
420	 * which is capped at 5ms, but runs from a timeout handler
421	 * that runs every 7.8ms. So a 15ms timeout leaves us with
422	 * a 2.2ms margin
423	 */
424	if (atomic) {
425		/* if atomic is set, do not sleep,
426		 * and try to get the completion quickly
427		 * (1ms will be enough for those commands) */
428		for (sleep_total = 0;
429		     sleep_total < 1000 &&
430		     response->result == htonl(MYRI10GE_NO_RESPONSE_RESULT);
431		     sleep_total += 10) {
432			udelay(10);
433			mb();
434		}
435	} else {
436		/* use msleep for most command */
437		for (sleep_total = 0;
438		     sleep_total < 15 &&
439		     response->result == htonl(MYRI10GE_NO_RESPONSE_RESULT);
440		     sleep_total++)
441			msleep(1);
442	}
443
444	result = ntohl(response->result);
445	value = ntohl(response->data);
446	if (result != MYRI10GE_NO_RESPONSE_RESULT) {
447		if (result == 0) {
448			data->data0 = value;
449			return 0;
450		} else if (result == MXGEFW_CMD_UNKNOWN) {
451			return -ENOSYS;
452		} else if (result == MXGEFW_CMD_ERROR_UNALIGNED) {
453			return -E2BIG;
454		} else if (result == MXGEFW_CMD_ERROR_RANGE &&
455			   cmd == MXGEFW_CMD_ENABLE_RSS_QUEUES &&
456			   (data->
457			    data1 & MXGEFW_SLICE_ENABLE_MULTIPLE_TX_QUEUES) !=
458			   0) {
459			return -ERANGE;
460		} else {
461			dev_err(&mgp->pdev->dev,
462				"command %d failed, result = %d\n",
463				cmd, result);
464			return -ENXIO;
465		}
466	}
467
468	dev_err(&mgp->pdev->dev, "command %d timed out, result = %d\n",
469		cmd, result);
470	return -EAGAIN;
471}
472
473/*
474 * The eeprom strings on the lanaiX have the format
475 * SN=x\0
476 * MAC=x:x:x:x:x:x\0
477 * PT:ddd mmm xx xx:xx:xx xx\0
478 * PV:ddd mmm xx xx:xx:xx xx\0
479 */
480static int myri10ge_read_mac_addr(struct myri10ge_priv *mgp)
481{
482	char *ptr, *limit;
483	int i;
484
485	ptr = mgp->eeprom_strings;
486	limit = mgp->eeprom_strings + MYRI10GE_EEPROM_STRINGS_SIZE;
487
488	while (*ptr != '\0' && ptr < limit) {
489		if (memcmp(ptr, "MAC=", 4) == 0) {
490			ptr += 4;
491			mgp->mac_addr_string = ptr;
492			for (i = 0; i < 6; i++) {
493				if ((ptr + 2) > limit)
494					goto abort;
495				mgp->mac_addr[i] =
496				    simple_strtoul(ptr, &ptr, 16);
497				ptr += 1;
498			}
499		}
500		if (memcmp(ptr, "PC=", 3) == 0) {
501			ptr += 3;
502			mgp->product_code_string = ptr;
503		}
504		if (memcmp((const void *)ptr, "SN=", 3) == 0) {
505			ptr += 3;
506			mgp->serial_number = simple_strtoul(ptr, &ptr, 10);
507		}
508		while (ptr < limit && *ptr++) ;
509	}
510
511	return 0;
512
513abort:
514	dev_err(&mgp->pdev->dev, "failed to parse eeprom_strings\n");
515	return -ENXIO;
516}
517
518/*
519 * Enable or disable periodic RDMAs from the host to make certain
520 * chipsets resend dropped PCIe messages
521 */
522
523static void myri10ge_dummy_rdma(struct myri10ge_priv *mgp, int enable)
524{
525	char __iomem *submit;
526	__be32 buf[16] __attribute__ ((__aligned__(8)));
527	u32 dma_low, dma_high;
528	int i;
529
530	/* clear confirmation addr */
531	mgp->cmd->data = 0;
532	mb();
533
534	/* send a rdma command to the PCIe engine, and wait for the
535	 * response in the confirmation address.  The firmware should
536	 * write a -1 there to indicate it is alive and well
537	 */
538	dma_low = MYRI10GE_LOWPART_TO_U32(mgp->cmd_bus);
539	dma_high = MYRI10GE_HIGHPART_TO_U32(mgp->cmd_bus);
540
541	buf[0] = htonl(dma_high);	/* confirm addr MSW */
542	buf[1] = htonl(dma_low);	/* confirm addr LSW */
543	buf[2] = MYRI10GE_NO_CONFIRM_DATA;	/* confirm data */
544	buf[3] = htonl(dma_high);	/* dummy addr MSW */
545	buf[4] = htonl(dma_low);	/* dummy addr LSW */
546	buf[5] = htonl(enable);	/* enable? */
547
548	submit = mgp->sram + MXGEFW_BOOT_DUMMY_RDMA;
549
550	myri10ge_pio_copy(submit, &buf, sizeof(buf));
551	for (i = 0; mgp->cmd->data != MYRI10GE_NO_CONFIRM_DATA && i < 20; i++)
552		msleep(1);
553	if (mgp->cmd->data != MYRI10GE_NO_CONFIRM_DATA)
554		dev_err(&mgp->pdev->dev, "dummy rdma %s failed\n",
555			(enable ? "enable" : "disable"));
556}
557
558static int
559myri10ge_validate_firmware(struct myri10ge_priv *mgp,
560			   struct mcp_gen_header *hdr)
561{
562	struct device *dev = &mgp->pdev->dev;
563
564	/* check firmware type */
565	if (ntohl(hdr->mcp_type) != MCP_TYPE_ETH) {
566		dev_err(dev, "Bad firmware type: 0x%x\n", ntohl(hdr->mcp_type));
567		return -EINVAL;
568	}
569
570	/* save firmware version for ethtool */
571	strncpy(mgp->fw_version, hdr->version, sizeof(mgp->fw_version));
572	mgp->fw_version[sizeof(mgp->fw_version) - 1] = '\0';
573
574	sscanf(mgp->fw_version, "%d.%d.%d", &mgp->fw_ver_major,
575	       &mgp->fw_ver_minor, &mgp->fw_ver_tiny);
576
577	if (!(mgp->fw_ver_major == MXGEFW_VERSION_MAJOR &&
578	      mgp->fw_ver_minor == MXGEFW_VERSION_MINOR)) {
579		dev_err(dev, "Found firmware version %s\n", mgp->fw_version);
580		dev_err(dev, "Driver needs %d.%d\n", MXGEFW_VERSION_MAJOR,
581			MXGEFW_VERSION_MINOR);
582		return -EINVAL;
583	}
584	return 0;
585}
586
587static int myri10ge_load_hotplug_firmware(struct myri10ge_priv *mgp, u32 * size)
588{
589	unsigned crc, reread_crc;
590	const struct firmware *fw;
591	struct device *dev = &mgp->pdev->dev;
592	unsigned char *fw_readback;
593	struct mcp_gen_header *hdr;
594	size_t hdr_offset;
595	int status;
596	unsigned i;
597
598	if ((status = request_firmware(&fw, mgp->fw_name, dev)) < 0) {
599		dev_err(dev, "Unable to load %s firmware image via hotplug\n",
600			mgp->fw_name);
601		status = -EINVAL;
602		goto abort_with_nothing;
603	}
604
605	/* check size */
606
607	if (fw->size >= mgp->sram_size - MYRI10GE_FW_OFFSET ||
608	    fw->size < MCP_HEADER_PTR_OFFSET + 4) {
609		dev_err(dev, "Firmware size invalid:%d\n", (int)fw->size);
610		status = -EINVAL;
611		goto abort_with_fw;
612	}
613
614	/* check id */
615	hdr_offset = ntohl(*(__be32 *) (fw->data + MCP_HEADER_PTR_OFFSET));
616	if ((hdr_offset & 3) || hdr_offset + sizeof(*hdr) > fw->size) {
617		dev_err(dev, "Bad firmware file\n");
618		status = -EINVAL;
619		goto abort_with_fw;
620	}
621	hdr = (void *)(fw->data + hdr_offset);
622
623	status = myri10ge_validate_firmware(mgp, hdr);
624	if (status != 0)
625		goto abort_with_fw;
626
627	crc = crc32(~0, fw->data, fw->size);
628	for (i = 0; i < fw->size; i += 256) {
629		myri10ge_pio_copy(mgp->sram + MYRI10GE_FW_OFFSET + i,
630				  fw->data + i,
631				  min(256U, (unsigned)(fw->size - i)));
632		mb();
633		readb(mgp->sram);
634	}
635	fw_readback = vmalloc(fw->size);
636	if (!fw_readback) {
637		status = -ENOMEM;
638		goto abort_with_fw;
639	}
640	/* corruption checking is good for parity recovery and buggy chipset */
641	memcpy_fromio(fw_readback, mgp->sram + MYRI10GE_FW_OFFSET, fw->size);
642	reread_crc = crc32(~0, fw_readback, fw->size);
643	vfree(fw_readback);
644	if (crc != reread_crc) {
645		dev_err(dev, "CRC failed(fw-len=%u), got 0x%x (expect 0x%x)\n",
646			(unsigned)fw->size, reread_crc, crc);
647		status = -EIO;
648		goto abort_with_fw;
649	}
650	*size = (u32) fw->size;
651
652abort_with_fw:
653	release_firmware(fw);
654
655abort_with_nothing:
656	return status;
657}
658
659static int myri10ge_adopt_running_firmware(struct myri10ge_priv *mgp)
660{
661	struct mcp_gen_header *hdr;
662	struct device *dev = &mgp->pdev->dev;
663	const size_t bytes = sizeof(struct mcp_gen_header);
664	size_t hdr_offset;
665	int status;
666
667	/* find running firmware header */
668	hdr_offset = swab32(readl(mgp->sram + MCP_HEADER_PTR_OFFSET));
669
670	if ((hdr_offset & 3) || hdr_offset + sizeof(*hdr) > mgp->sram_size) {
671		dev_err(dev, "Running firmware has bad header offset (%d)\n",
672			(int)hdr_offset);
673		return -EIO;
674	}
675
676	/* copy header of running firmware from SRAM to host memory to
677	 * validate firmware */
678	hdr = kmalloc(bytes, GFP_KERNEL);
679	if (hdr == NULL)
680		return -ENOMEM;
681
682	memcpy_fromio(hdr, mgp->sram + hdr_offset, bytes);
683	status = myri10ge_validate_firmware(mgp, hdr);
684	kfree(hdr);
685
686	/* check to see if adopted firmware has bug where adopting
687	 * it will cause broadcasts to be filtered unless the NIC
688	 * is kept in ALLMULTI mode */
689	if (mgp->fw_ver_major == 1 && mgp->fw_ver_minor == 4 &&
690	    mgp->fw_ver_tiny >= 4 && mgp->fw_ver_tiny <= 11) {
691		mgp->adopted_rx_filter_bug = 1;
692		dev_warn(dev, "Adopting fw %d.%d.%d: "
693			 "working around rx filter bug\n",
694			 mgp->fw_ver_major, mgp->fw_ver_minor,
695			 mgp->fw_ver_tiny);
696	}
697	return status;
698}
699
700static int myri10ge_get_firmware_capabilities(struct myri10ge_priv *mgp)
701{
702	struct myri10ge_cmd cmd;
703	int status;
704
705	/* probe for IPv6 TSO support */
706	mgp->features = NETIF_F_SG | NETIF_F_HW_CSUM | NETIF_F_TSO;
707	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_MAX_TSO6_HDR_SIZE,
708				   &cmd, 0);
709	if (status == 0) {
710		mgp->max_tso6 = cmd.data0;
711		mgp->features |= NETIF_F_TSO6;
712	}
713
714	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_RX_RING_SIZE, &cmd, 0);
715	if (status != 0) {
716		dev_err(&mgp->pdev->dev,
717			"failed MXGEFW_CMD_GET_RX_RING_SIZE\n");
718		return -ENXIO;
719	}
720
721	mgp->max_intr_slots = 2 * (cmd.data0 / sizeof(struct mcp_dma_addr));
722
723	return 0;
724}
725
726static int myri10ge_load_firmware(struct myri10ge_priv *mgp, int adopt)
727{
728	char __iomem *submit;
729	__be32 buf[16] __attribute__ ((__aligned__(8)));
730	u32 dma_low, dma_high, size;
731	int status, i;
732
733	size = 0;
734	status = myri10ge_load_hotplug_firmware(mgp, &size);
735	if (status) {
736		if (!adopt)
737			return status;
738		dev_warn(&mgp->pdev->dev, "hotplug firmware loading failed\n");
739
740		/* Do not attempt to adopt firmware if there
741		 * was a bad crc */
742		if (status == -EIO)
743			return status;
744
745		status = myri10ge_adopt_running_firmware(mgp);
746		if (status != 0) {
747			dev_err(&mgp->pdev->dev,
748				"failed to adopt running firmware\n");
749			return status;
750		}
751		dev_info(&mgp->pdev->dev,
752			 "Successfully adopted running firmware\n");
753		if (mgp->tx_boundary == 4096) {
754			dev_warn(&mgp->pdev->dev,
755				 "Using firmware currently running on NIC"
756				 ".  For optimal\n");
757			dev_warn(&mgp->pdev->dev,
758				 "performance consider loading optimized "
759				 "firmware\n");
760			dev_warn(&mgp->pdev->dev, "via hotplug\n");
761		}
762
763		set_fw_name(mgp, "adopted", false);
764		mgp->tx_boundary = 2048;
765		myri10ge_dummy_rdma(mgp, 1);
766		status = myri10ge_get_firmware_capabilities(mgp);
767		return status;
768	}
769
770	/* clear confirmation addr */
771	mgp->cmd->data = 0;
772	mb();
773
774	/* send a reload command to the bootstrap MCP, and wait for the
775	 *  response in the confirmation address.  The firmware should
776	 * write a -1 there to indicate it is alive and well
777	 */
778	dma_low = MYRI10GE_LOWPART_TO_U32(mgp->cmd_bus);
779	dma_high = MYRI10GE_HIGHPART_TO_U32(mgp->cmd_bus);
780
781	buf[0] = htonl(dma_high);	/* confirm addr MSW */
782	buf[1] = htonl(dma_low);	/* confirm addr LSW */
783	buf[2] = MYRI10GE_NO_CONFIRM_DATA;	/* confirm data */
784
785	/* FIX: All newest firmware should un-protect the bottom of
786	 * the sram before handoff. However, the very first interfaces
787	 * do not. Therefore the handoff copy must skip the first 8 bytes
788	 */
789	buf[3] = htonl(MYRI10GE_FW_OFFSET + 8);	/* where the code starts */
790	buf[4] = htonl(size - 8);	/* length of code */
791	buf[5] = htonl(8);	/* where to copy to */
792	buf[6] = htonl(0);	/* where to jump to */
793
794	submit = mgp->sram + MXGEFW_BOOT_HANDOFF;
795
796	myri10ge_pio_copy(submit, &buf, sizeof(buf));
797	mb();
798	msleep(1);
799	mb();
800	i = 0;
801	while (mgp->cmd->data != MYRI10GE_NO_CONFIRM_DATA && i < 9) {
802		msleep(1 << i);
803		i++;
804	}
805	if (mgp->cmd->data != MYRI10GE_NO_CONFIRM_DATA) {
806		dev_err(&mgp->pdev->dev, "handoff failed\n");
807		return -ENXIO;
808	}
809	myri10ge_dummy_rdma(mgp, 1);
810	status = myri10ge_get_firmware_capabilities(mgp);
811
812	return status;
813}
814
815static int myri10ge_update_mac_address(struct myri10ge_priv *mgp, u8 * addr)
816{
817	struct myri10ge_cmd cmd;
818	int status;
819
820	cmd.data0 = ((addr[0] << 24) | (addr[1] << 16)
821		     | (addr[2] << 8) | addr[3]);
822
823	cmd.data1 = ((addr[4] << 8) | (addr[5]));
824
825	status = myri10ge_send_cmd(mgp, MXGEFW_SET_MAC_ADDRESS, &cmd, 0);
826	return status;
827}
828
829static int myri10ge_change_pause(struct myri10ge_priv *mgp, int pause)
830{
831	struct myri10ge_cmd cmd;
832	int status, ctl;
833
834	ctl = pause ? MXGEFW_ENABLE_FLOW_CONTROL : MXGEFW_DISABLE_FLOW_CONTROL;
835	status = myri10ge_send_cmd(mgp, ctl, &cmd, 0);
836
837	if (status) {
838		netdev_err(mgp->dev, "Failed to set flow control mode\n");
839		return status;
840	}
841	mgp->pause = pause;
842	return 0;
843}
844
845static void
846myri10ge_change_promisc(struct myri10ge_priv *mgp, int promisc, int atomic)
847{
848	struct myri10ge_cmd cmd;
849	int status, ctl;
850
851	ctl = promisc ? MXGEFW_ENABLE_PROMISC : MXGEFW_DISABLE_PROMISC;
852	status = myri10ge_send_cmd(mgp, ctl, &cmd, atomic);
853	if (status)
854		netdev_err(mgp->dev, "Failed to set promisc mode\n");
855}
856
857static int myri10ge_dma_test(struct myri10ge_priv *mgp, int test_type)
858{
859	struct myri10ge_cmd cmd;
860	int status;
861	u32 len;
862	struct page *dmatest_page;
863	dma_addr_t dmatest_bus;
864	char *test = " ";
865
866	dmatest_page = alloc_page(GFP_KERNEL);
867	if (!dmatest_page)
868		return -ENOMEM;
869	dmatest_bus = pci_map_page(mgp->pdev, dmatest_page, 0, PAGE_SIZE,
870				   DMA_BIDIRECTIONAL);
871	if (unlikely(pci_dma_mapping_error(mgp->pdev, dmatest_bus))) {
872		__free_page(dmatest_page);
873		return -ENOMEM;
874	}
875
876	/* Run a small DMA test.
877	 * The magic multipliers to the length tell the firmware
878	 * to do DMA read, write, or read+write tests.  The
879	 * results are returned in cmd.data0.  The upper 16
880	 * bits or the return is the number of transfers completed.
881	 * The lower 16 bits is the time in 0.5us ticks that the
882	 * transfers took to complete.
883	 */
884
885	len = mgp->tx_boundary;
886
887	cmd.data0 = MYRI10GE_LOWPART_TO_U32(dmatest_bus);
888	cmd.data1 = MYRI10GE_HIGHPART_TO_U32(dmatest_bus);
889	cmd.data2 = len * 0x10000;
890	status = myri10ge_send_cmd(mgp, test_type, &cmd, 0);
891	if (status != 0) {
892		test = "read";
893		goto abort;
894	}
895	mgp->read_dma = ((cmd.data0 >> 16) * len * 2) / (cmd.data0 & 0xffff);
896	cmd.data0 = MYRI10GE_LOWPART_TO_U32(dmatest_bus);
897	cmd.data1 = MYRI10GE_HIGHPART_TO_U32(dmatest_bus);
898	cmd.data2 = len * 0x1;
899	status = myri10ge_send_cmd(mgp, test_type, &cmd, 0);
900	if (status != 0) {
901		test = "write";
902		goto abort;
903	}
904	mgp->write_dma = ((cmd.data0 >> 16) * len * 2) / (cmd.data0 & 0xffff);
905
906	cmd.data0 = MYRI10GE_LOWPART_TO_U32(dmatest_bus);
907	cmd.data1 = MYRI10GE_HIGHPART_TO_U32(dmatest_bus);
908	cmd.data2 = len * 0x10001;
909	status = myri10ge_send_cmd(mgp, test_type, &cmd, 0);
910	if (status != 0) {
911		test = "read/write";
912		goto abort;
913	}
914	mgp->read_write_dma = ((cmd.data0 >> 16) * len * 2 * 2) /
915	    (cmd.data0 & 0xffff);
916
917abort:
918	pci_unmap_page(mgp->pdev, dmatest_bus, PAGE_SIZE, DMA_BIDIRECTIONAL);
919	put_page(dmatest_page);
920
921	if (status != 0 && test_type != MXGEFW_CMD_UNALIGNED_TEST)
922		dev_warn(&mgp->pdev->dev, "DMA %s benchmark failed: %d\n",
923			 test, status);
924
925	return status;
926}
927
928#ifdef CONFIG_NET_RX_BUSY_POLL
929static inline void myri10ge_ss_init_lock(struct myri10ge_slice_state *ss)
930{
931	spin_lock_init(&ss->lock);
932	ss->state = SLICE_STATE_IDLE;
933}
934
935static inline bool myri10ge_ss_lock_napi(struct myri10ge_slice_state *ss)
936{
937	bool rc = true;
938	spin_lock(&ss->lock);
939	if ((ss->state & SLICE_LOCKED)) {
940		WARN_ON((ss->state & SLICE_STATE_NAPI));
941		ss->state |= SLICE_STATE_NAPI_YIELD;
942		rc = false;
943		ss->lock_napi_yield++;
944	} else
945		ss->state = SLICE_STATE_NAPI;
946	spin_unlock(&ss->lock);
947	return rc;
948}
949
950static inline void myri10ge_ss_unlock_napi(struct myri10ge_slice_state *ss)
951{
952	spin_lock(&ss->lock);
953	WARN_ON((ss->state & (SLICE_STATE_POLL | SLICE_STATE_NAPI_YIELD)));
954	ss->state = SLICE_STATE_IDLE;
955	spin_unlock(&ss->lock);
956}
957
958static inline bool myri10ge_ss_lock_poll(struct myri10ge_slice_state *ss)
959{
960	bool rc = true;
961	spin_lock_bh(&ss->lock);
962	if ((ss->state & SLICE_LOCKED)) {
963		ss->state |= SLICE_STATE_POLL_YIELD;
964		rc = false;
965		ss->lock_poll_yield++;
966	} else
967		ss->state |= SLICE_STATE_POLL;
968	spin_unlock_bh(&ss->lock);
969	return rc;
970}
971
972static inline void myri10ge_ss_unlock_poll(struct myri10ge_slice_state *ss)
973{
974	spin_lock_bh(&ss->lock);
975	WARN_ON((ss->state & SLICE_STATE_NAPI));
976	ss->state = SLICE_STATE_IDLE;
977	spin_unlock_bh(&ss->lock);
978}
979
980static inline bool myri10ge_ss_busy_polling(struct myri10ge_slice_state *ss)
981{
982	WARN_ON(!(ss->state & SLICE_LOCKED));
983	return (ss->state & SLICE_USER_PEND);
984}
985#else /* CONFIG_NET_RX_BUSY_POLL */
986static inline void myri10ge_ss_init_lock(struct myri10ge_slice_state *ss)
987{
988}
989
990static inline bool myri10ge_ss_lock_napi(struct myri10ge_slice_state *ss)
991{
992	return false;
993}
994
995static inline void myri10ge_ss_unlock_napi(struct myri10ge_slice_state *ss)
996{
997}
998
999static inline bool myri10ge_ss_lock_poll(struct myri10ge_slice_state *ss)
1000{
1001	return false;
1002}
1003
1004static inline void myri10ge_ss_unlock_poll(struct myri10ge_slice_state *ss)
1005{
1006}
1007
1008static inline bool myri10ge_ss_busy_polling(struct myri10ge_slice_state *ss)
1009{
1010	return false;
1011}
1012#endif
1013
1014static int myri10ge_reset(struct myri10ge_priv *mgp)
1015{
1016	struct myri10ge_cmd cmd;
1017	struct myri10ge_slice_state *ss;
1018	int i, status;
1019	size_t bytes;
1020#ifdef CONFIG_MYRI10GE_DCA
1021	unsigned long dca_tag_off;
1022#endif
1023
1024	/* try to send a reset command to the card to see if it
1025	 * is alive */
1026	memset(&cmd, 0, sizeof(cmd));
1027	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_RESET, &cmd, 0);
1028	if (status != 0) {
1029		dev_err(&mgp->pdev->dev, "failed reset\n");
1030		return -ENXIO;
1031	}
1032
1033	(void)myri10ge_dma_test(mgp, MXGEFW_DMA_TEST);
1034	/*
1035	 * Use non-ndis mcp_slot (eg, 4 bytes total,
1036	 * no toeplitz hash value returned.  Older firmware will
1037	 * not understand this command, but will use the correct
1038	 * sized mcp_slot, so we ignore error returns
1039	 */
1040	cmd.data0 = MXGEFW_RSS_MCP_SLOT_TYPE_MIN;
1041	(void)myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_RSS_MCP_SLOT_TYPE, &cmd, 0);
1042
1043	/* Now exchange information about interrupts  */
1044
1045	bytes = mgp->max_intr_slots * sizeof(*mgp->ss[0].rx_done.entry);
1046	cmd.data0 = (u32) bytes;
1047	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_INTRQ_SIZE, &cmd, 0);
1048
1049	/*
1050	 * Even though we already know how many slices are supported
1051	 * via myri10ge_probe_slices() MXGEFW_CMD_GET_MAX_RSS_QUEUES
1052	 * has magic side effects, and must be called after a reset.
1053	 * It must be called prior to calling any RSS related cmds,
1054	 * including assigning an interrupt queue for anything but
1055	 * slice 0.  It must also be called *after*
1056	 * MXGEFW_CMD_SET_INTRQ_SIZE, since the intrq size is used by
1057	 * the firmware to compute offsets.
1058	 */
1059
1060	if (mgp->num_slices > 1) {
1061
1062		/* ask the maximum number of slices it supports */
1063		status = myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_MAX_RSS_QUEUES,
1064					   &cmd, 0);
1065		if (status != 0) {
1066			dev_err(&mgp->pdev->dev,
1067				"failed to get number of slices\n");
1068		}
1069
1070		/*
1071		 * MXGEFW_CMD_ENABLE_RSS_QUEUES must be called prior
1072		 * to setting up the interrupt queue DMA
1073		 */
1074
1075		cmd.data0 = mgp->num_slices;
1076		cmd.data1 = MXGEFW_SLICE_INTR_MODE_ONE_PER_SLICE;
1077		if (mgp->dev->real_num_tx_queues > 1)
1078			cmd.data1 |= MXGEFW_SLICE_ENABLE_MULTIPLE_TX_QUEUES;
1079		status = myri10ge_send_cmd(mgp, MXGEFW_CMD_ENABLE_RSS_QUEUES,
1080					   &cmd, 0);
1081
1082		/* Firmware older than 1.4.32 only supports multiple
1083		 * RX queues, so if we get an error, first retry using a
1084		 * single TX queue before giving up */
1085		if (status != 0 && mgp->dev->real_num_tx_queues > 1) {
1086			netif_set_real_num_tx_queues(mgp->dev, 1);
1087			cmd.data0 = mgp->num_slices;
1088			cmd.data1 = MXGEFW_SLICE_INTR_MODE_ONE_PER_SLICE;
1089			status = myri10ge_send_cmd(mgp,
1090						   MXGEFW_CMD_ENABLE_RSS_QUEUES,
1091						   &cmd, 0);
1092		}
1093
1094		if (status != 0) {
1095			dev_err(&mgp->pdev->dev,
1096				"failed to set number of slices\n");
1097
1098			return status;
1099		}
1100	}
1101	for (i = 0; i < mgp->num_slices; i++) {
1102		ss = &mgp->ss[i];
1103		cmd.data0 = MYRI10GE_LOWPART_TO_U32(ss->rx_done.bus);
1104		cmd.data1 = MYRI10GE_HIGHPART_TO_U32(ss->rx_done.bus);
1105		cmd.data2 = i;
1106		status |= myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_INTRQ_DMA,
1107					    &cmd, 0);
1108	}
1109
1110	status |=
1111	    myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_IRQ_ACK_OFFSET, &cmd, 0);
1112	for (i = 0; i < mgp->num_slices; i++) {
1113		ss = &mgp->ss[i];
1114		ss->irq_claim =
1115		    (__iomem __be32 *) (mgp->sram + cmd.data0 + 8 * i);
1116	}
1117	status |= myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_IRQ_DEASSERT_OFFSET,
1118				    &cmd, 0);
1119	mgp->irq_deassert = (__iomem __be32 *) (mgp->sram + cmd.data0);
1120
1121	status |= myri10ge_send_cmd
1122	    (mgp, MXGEFW_CMD_GET_INTR_COAL_DELAY_OFFSET, &cmd, 0);
1123	mgp->intr_coal_delay_ptr = (__iomem __be32 *) (mgp->sram + cmd.data0);
1124	if (status != 0) {
1125		dev_err(&mgp->pdev->dev, "failed set interrupt parameters\n");
1126		return status;
1127	}
1128	put_be32(htonl(mgp->intr_coal_delay), mgp->intr_coal_delay_ptr);
1129
1130#ifdef CONFIG_MYRI10GE_DCA
1131	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_DCA_OFFSET, &cmd, 0);
1132	dca_tag_off = cmd.data0;
1133	for (i = 0; i < mgp->num_slices; i++) {
1134		ss = &mgp->ss[i];
1135		if (status == 0) {
1136			ss->dca_tag = (__iomem __be32 *)
1137			    (mgp->sram + dca_tag_off + 4 * i);
1138		} else {
1139			ss->dca_tag = NULL;
1140		}
1141	}
1142#endif				/* CONFIG_MYRI10GE_DCA */
1143
1144	/* reset mcp/driver shared state back to 0 */
1145
1146	mgp->link_changes = 0;
1147	for (i = 0; i < mgp->num_slices; i++) {
1148		ss = &mgp->ss[i];
1149
1150		memset(ss->rx_done.entry, 0, bytes);
1151		ss->tx.req = 0;
1152		ss->tx.done = 0;
1153		ss->tx.pkt_start = 0;
1154		ss->tx.pkt_done = 0;
1155		ss->rx_big.cnt = 0;
1156		ss->rx_small.cnt = 0;
1157		ss->rx_done.idx = 0;
1158		ss->rx_done.cnt = 0;
1159		ss->tx.wake_queue = 0;
1160		ss->tx.stop_queue = 0;
1161	}
1162
1163	status = myri10ge_update_mac_address(mgp, mgp->dev->dev_addr);
1164	myri10ge_change_pause(mgp, mgp->pause);
1165	myri10ge_set_multicast_list(mgp->dev);
1166	return status;
1167}
1168
1169#ifdef CONFIG_MYRI10GE_DCA
1170static int myri10ge_toggle_relaxed(struct pci_dev *pdev, int on)
1171{
1172	int ret;
1173	u16 ctl;
1174
1175	pcie_capability_read_word(pdev, PCI_EXP_DEVCTL, &ctl);
1176
1177	ret = (ctl & PCI_EXP_DEVCTL_RELAX_EN) >> 4;
1178	if (ret != on) {
1179		ctl &= ~PCI_EXP_DEVCTL_RELAX_EN;
1180		ctl |= (on << 4);
1181		pcie_capability_write_word(pdev, PCI_EXP_DEVCTL, ctl);
1182	}
1183	return ret;
1184}
1185
1186static void
1187myri10ge_write_dca(struct myri10ge_slice_state *ss, int cpu, int tag)
1188{
1189	ss->cached_dca_tag = tag;
1190	put_be32(htonl(tag), ss->dca_tag);
1191}
1192
1193static inline void myri10ge_update_dca(struct myri10ge_slice_state *ss)
1194{
1195	int cpu = get_cpu();
1196	int tag;
1197
1198	if (cpu != ss->cpu) {
1199		tag = dca3_get_tag(&ss->mgp->pdev->dev, cpu);
1200		if (ss->cached_dca_tag != tag)
1201			myri10ge_write_dca(ss, cpu, tag);
1202		ss->cpu = cpu;
1203	}
1204	put_cpu();
1205}
1206
1207static void myri10ge_setup_dca(struct myri10ge_priv *mgp)
1208{
1209	int err, i;
1210	struct pci_dev *pdev = mgp->pdev;
1211
1212	if (mgp->ss[0].dca_tag == NULL || mgp->dca_enabled)
1213		return;
1214	if (!myri10ge_dca) {
1215		dev_err(&pdev->dev, "dca disabled by administrator\n");
1216		return;
1217	}
1218	err = dca_add_requester(&pdev->dev);
1219	if (err) {
1220		if (err != -ENODEV)
1221			dev_err(&pdev->dev,
1222				"dca_add_requester() failed, err=%d\n", err);
1223		return;
1224	}
1225	mgp->relaxed_order = myri10ge_toggle_relaxed(pdev, 0);
1226	mgp->dca_enabled = 1;
1227	for (i = 0; i < mgp->num_slices; i++) {
1228		mgp->ss[i].cpu = -1;
1229		mgp->ss[i].cached_dca_tag = -1;
1230		myri10ge_update_dca(&mgp->ss[i]);
1231	}
1232}
1233
1234static void myri10ge_teardown_dca(struct myri10ge_priv *mgp)
1235{
1236	struct pci_dev *pdev = mgp->pdev;
1237
1238	if (!mgp->dca_enabled)
1239		return;
1240	mgp->dca_enabled = 0;
1241	if (mgp->relaxed_order)
1242		myri10ge_toggle_relaxed(pdev, 1);
1243	dca_remove_requester(&pdev->dev);
1244}
1245
1246static int myri10ge_notify_dca_device(struct device *dev, void *data)
1247{
1248	struct myri10ge_priv *mgp;
1249	unsigned long event;
1250
1251	mgp = dev_get_drvdata(dev);
1252	event = *(unsigned long *)data;
1253
1254	if (event == DCA_PROVIDER_ADD)
1255		myri10ge_setup_dca(mgp);
1256	else if (event == DCA_PROVIDER_REMOVE)
1257		myri10ge_teardown_dca(mgp);
1258	return 0;
1259}
1260#endif				/* CONFIG_MYRI10GE_DCA */
1261
1262static inline void
1263myri10ge_submit_8rx(struct mcp_kreq_ether_recv __iomem * dst,
1264		    struct mcp_kreq_ether_recv *src)
1265{
1266	__be32 low;
1267
1268	low = src->addr_low;
1269	src->addr_low = htonl(DMA_BIT_MASK(32));
1270	myri10ge_pio_copy(dst, src, 4 * sizeof(*src));
1271	mb();
1272	myri10ge_pio_copy(dst + 4, src + 4, 4 * sizeof(*src));
1273	mb();
1274	src->addr_low = low;
1275	put_be32(low, &dst->addr_low);
1276	mb();
1277}
1278
1279static inline void myri10ge_vlan_ip_csum(struct sk_buff *skb, __wsum hw_csum)
1280{
1281	struct vlan_hdr *vh = (struct vlan_hdr *)(skb->data);
1282
1283	if ((skb->protocol == htons(ETH_P_8021Q)) &&
1284	    (vh->h_vlan_encapsulated_proto == htons(ETH_P_IP) ||
1285	     vh->h_vlan_encapsulated_proto == htons(ETH_P_IPV6))) {
1286		skb->csum = hw_csum;
1287		skb->ip_summed = CHECKSUM_COMPLETE;
1288	}
1289}
1290
1291static void
1292myri10ge_alloc_rx_pages(struct myri10ge_priv *mgp, struct myri10ge_rx_buf *rx,
1293			int bytes, int watchdog)
1294{
1295	struct page *page;
1296	dma_addr_t bus;
1297	int idx;
1298#if MYRI10GE_ALLOC_SIZE > 4096
1299	int end_offset;
1300#endif
1301
1302	if (unlikely(rx->watchdog_needed && !watchdog))
1303		return;
1304
1305	/* try to refill entire ring */
1306	while (rx->fill_cnt != (rx->cnt + rx->mask + 1)) {
1307		idx = rx->fill_cnt & rx->mask;
1308		if (rx->page_offset + bytes <= MYRI10GE_ALLOC_SIZE) {
1309			/* we can use part of previous page */
1310			get_page(rx->page);
1311		} else {
1312			/* we need a new page */
1313			page =
1314			    alloc_pages(GFP_ATOMIC | __GFP_COMP,
1315					MYRI10GE_ALLOC_ORDER);
1316			if (unlikely(page == NULL)) {
1317				if (rx->fill_cnt - rx->cnt < 16)
1318					rx->watchdog_needed = 1;
1319				return;
1320			}
1321
1322			bus = pci_map_page(mgp->pdev, page, 0,
1323					   MYRI10GE_ALLOC_SIZE,
1324					   PCI_DMA_FROMDEVICE);
1325			if (unlikely(pci_dma_mapping_error(mgp->pdev, bus))) {
1326				__free_pages(page, MYRI10GE_ALLOC_ORDER);
1327				if (rx->fill_cnt - rx->cnt < 16)
1328					rx->watchdog_needed = 1;
1329				return;
1330			}
1331
1332			rx->page = page;
1333			rx->page_offset = 0;
1334			rx->bus = bus;
1335
1336		}
1337		rx->info[idx].page = rx->page;
1338		rx->info[idx].page_offset = rx->page_offset;
1339		/* note that this is the address of the start of the
1340		 * page */
1341		dma_unmap_addr_set(&rx->info[idx], bus, rx->bus);
1342		rx->shadow[idx].addr_low =
1343		    htonl(MYRI10GE_LOWPART_TO_U32(rx->bus) + rx->page_offset);
1344		rx->shadow[idx].addr_high =
1345		    htonl(MYRI10GE_HIGHPART_TO_U32(rx->bus));
1346
1347		/* start next packet on a cacheline boundary */
1348		rx->page_offset += SKB_DATA_ALIGN(bytes);
1349
1350#if MYRI10GE_ALLOC_SIZE > 4096
1351		/* don't cross a 4KB boundary */
1352		end_offset = rx->page_offset + bytes - 1;
1353		if ((unsigned)(rx->page_offset ^ end_offset) > 4095)
1354			rx->page_offset = end_offset & ~4095;
1355#endif
1356		rx->fill_cnt++;
1357
1358		/* copy 8 descriptors to the firmware at a time */
1359		if ((idx & 7) == 7) {
1360			myri10ge_submit_8rx(&rx->lanai[idx - 7],
1361					    &rx->shadow[idx - 7]);
1362		}
1363	}
1364}
1365
1366static inline void
1367myri10ge_unmap_rx_page(struct pci_dev *pdev,
1368		       struct myri10ge_rx_buffer_state *info, int bytes)
1369{
1370	/* unmap the recvd page if we're the only or last user of it */
1371	if (bytes >= MYRI10GE_ALLOC_SIZE / 2 ||
1372	    (info->page_offset + 2 * bytes) > MYRI10GE_ALLOC_SIZE) {
1373		pci_unmap_page(pdev, (dma_unmap_addr(info, bus)
1374				      & ~(MYRI10GE_ALLOC_SIZE - 1)),
1375			       MYRI10GE_ALLOC_SIZE, PCI_DMA_FROMDEVICE);
1376	}
1377}
1378
1379/*
1380 * GRO does not support acceleration of tagged vlan frames, and
1381 * this NIC does not support vlan tag offload, so we must pop
1382 * the tag ourselves to be able to achieve GRO performance that
1383 * is comparable to LRO.
1384 */
1385
1386static inline void
1387myri10ge_vlan_rx(struct net_device *dev, void *addr, struct sk_buff *skb)
1388{
1389	u8 *va;
1390	struct vlan_ethhdr *veh;
1391	struct skb_frag_struct *frag;
1392	__wsum vsum;
1393
1394	va = addr;
1395	va += MXGEFW_PAD;
1396	veh = (struct vlan_ethhdr *)va;
1397	if ((dev->features & NETIF_F_HW_VLAN_CTAG_RX) ==
1398	    NETIF_F_HW_VLAN_CTAG_RX &&
1399	    veh->h_vlan_proto == htons(ETH_P_8021Q)) {
1400		/* fixup csum if needed */
1401		if (skb->ip_summed == CHECKSUM_COMPLETE) {
1402			vsum = csum_partial(va + ETH_HLEN, VLAN_HLEN, 0);
1403			skb->csum = csum_sub(skb->csum, vsum);
1404		}
1405		/* pop tag */
1406		__vlan_hwaccel_put_tag(skb, htons(ETH_P_8021Q), ntohs(veh->h_vlan_TCI));
1407		memmove(va + VLAN_HLEN, va, 2 * ETH_ALEN);
1408		skb->len -= VLAN_HLEN;
1409		skb->data_len -= VLAN_HLEN;
1410		frag = skb_shinfo(skb)->frags;
1411		frag->page_offset += VLAN_HLEN;
1412		skb_frag_size_set(frag, skb_frag_size(frag) - VLAN_HLEN);
1413	}
1414}
1415
1416#define MYRI10GE_HLEN 64 /* Bytes to copy from page to skb linear memory */
1417
1418static inline int
1419myri10ge_rx_done(struct myri10ge_slice_state *ss, int len, __wsum csum)
1420{
1421	struct myri10ge_priv *mgp = ss->mgp;
1422	struct sk_buff *skb;
1423	struct skb_frag_struct *rx_frags;
1424	struct myri10ge_rx_buf *rx;
1425	int i, idx, remainder, bytes;
1426	struct pci_dev *pdev = mgp->pdev;
1427	struct net_device *dev = mgp->dev;
1428	u8 *va;
1429	bool polling;
1430
1431	if (len <= mgp->small_bytes) {
1432		rx = &ss->rx_small;
1433		bytes = mgp->small_bytes;
1434	} else {
1435		rx = &ss->rx_big;
1436		bytes = mgp->big_bytes;
1437	}
1438
1439	len += MXGEFW_PAD;
1440	idx = rx->cnt & rx->mask;
1441	va = page_address(rx->info[idx].page) + rx->info[idx].page_offset;
1442	prefetch(va);
1443
1444	/* When busy polling in user context, allocate skb and copy headers to
1445	 * skb's linear memory ourselves.  When not busy polling, use the napi
1446	 * gro api.
1447	 */
1448	polling = myri10ge_ss_busy_polling(ss);
1449	if (polling)
1450		skb = netdev_alloc_skb(dev, MYRI10GE_HLEN + 16);
1451	else
1452		skb = napi_get_frags(&ss->napi);
1453	if (unlikely(skb == NULL)) {
1454		ss->stats.rx_dropped++;
1455		for (i = 0, remainder = len; remainder > 0; i++) {
1456			myri10ge_unmap_rx_page(pdev, &rx->info[idx], bytes);
1457			put_page(rx->info[idx].page);
1458			rx->cnt++;
1459			idx = rx->cnt & rx->mask;
1460			remainder -= MYRI10GE_ALLOC_SIZE;
1461		}
1462		return 0;
1463	}
1464	rx_frags = skb_shinfo(skb)->frags;
1465	/* Fill skb_frag_struct(s) with data from our receive */
1466	for (i = 0, remainder = len; remainder > 0; i++) {
1467		myri10ge_unmap_rx_page(pdev, &rx->info[idx], bytes);
1468		skb_fill_page_desc(skb, i, rx->info[idx].page,
1469				   rx->info[idx].page_offset,
1470				   remainder < MYRI10GE_ALLOC_SIZE ?
1471				   remainder : MYRI10GE_ALLOC_SIZE);
1472		rx->cnt++;
1473		idx = rx->cnt & rx->mask;
1474		remainder -= MYRI10GE_ALLOC_SIZE;
1475	}
1476
1477	/* remove padding */
1478	rx_frags[0].page_offset += MXGEFW_PAD;
1479	rx_frags[0].size -= MXGEFW_PAD;
1480	len -= MXGEFW_PAD;
1481
1482	skb->len = len;
1483	skb->data_len = len;
1484	skb->truesize += len;
1485	if (dev->features & NETIF_F_RXCSUM) {
1486		skb->ip_summed = CHECKSUM_COMPLETE;
1487		skb->csum = csum;
1488	}
1489	myri10ge_vlan_rx(mgp->dev, va, skb);
1490	skb_record_rx_queue(skb, ss - &mgp->ss[0]);
1491	skb_mark_napi_id(skb, &ss->napi);
1492
1493	if (polling) {
1494		int hlen;
1495
1496		/* myri10ge_vlan_rx might have moved the header, so compute
1497		 * length and address again.
1498		 */
1499		hlen = MYRI10GE_HLEN > skb->len ? skb->len : MYRI10GE_HLEN;
1500		va = page_address(skb_frag_page(&rx_frags[0])) +
1501			rx_frags[0].page_offset;
1502		/* Copy header into the skb linear memory */
1503		skb_copy_to_linear_data(skb, va, hlen);
1504		rx_frags[0].page_offset += hlen;
1505		rx_frags[0].size -= hlen;
1506		skb->data_len -= hlen;
1507		skb->tail += hlen;
1508		skb->protocol = eth_type_trans(skb, dev);
1509		netif_receive_skb(skb);
1510	}
1511	else
1512		napi_gro_frags(&ss->napi);
1513
1514	return 1;
1515}
1516
1517static inline void
1518myri10ge_tx_done(struct myri10ge_slice_state *ss, int mcp_index)
1519{
1520	struct pci_dev *pdev = ss->mgp->pdev;
1521	struct myri10ge_tx_buf *tx = &ss->tx;
1522	struct netdev_queue *dev_queue;
1523	struct sk_buff *skb;
1524	int idx, len;
1525
1526	while (tx->pkt_done != mcp_index) {
1527		idx = tx->done & tx->mask;
1528		skb = tx->info[idx].skb;
1529
1530		/* Mark as free */
1531		tx->info[idx].skb = NULL;
1532		if (tx->info[idx].last) {
1533			tx->pkt_done++;
1534			tx->info[idx].last = 0;
1535		}
1536		tx->done++;
1537		len = dma_unmap_len(&tx->info[idx], len);
1538		dma_unmap_len_set(&tx->info[idx], len, 0);
1539		if (skb) {
1540			ss->stats.tx_bytes += skb->len;
1541			ss->stats.tx_packets++;
1542			dev_kfree_skb_irq(skb);
1543			if (len)
1544				pci_unmap_single(pdev,
1545						 dma_unmap_addr(&tx->info[idx],
1546								bus), len,
1547						 PCI_DMA_TODEVICE);
1548		} else {
1549			if (len)
1550				pci_unmap_page(pdev,
1551					       dma_unmap_addr(&tx->info[idx],
1552							      bus), len,
1553					       PCI_DMA_TODEVICE);
1554		}
1555	}
1556
1557	dev_queue = netdev_get_tx_queue(ss->dev, ss - ss->mgp->ss);
1558	/*
1559	 * Make a minimal effort to prevent the NIC from polling an
1560	 * idle tx queue.  If we can't get the lock we leave the queue
1561	 * active. In this case, either a thread was about to start
1562	 * using the queue anyway, or we lost a race and the NIC will
1563	 * waste some of its resources polling an inactive queue for a
1564	 * while.
1565	 */
1566
1567	if ((ss->mgp->dev->real_num_tx_queues > 1) &&
1568	    __netif_tx_trylock(dev_queue)) {
1569		if (tx->req == tx->done) {
1570			tx->queue_active = 0;
1571			put_be32(htonl(1), tx->send_stop);
1572			mb();
1573			mmiowb();
1574		}
1575		__netif_tx_unlock(dev_queue);
1576	}
1577
1578	/* start the queue if we've stopped it */
1579	if (netif_tx_queue_stopped(dev_queue) &&
1580	    tx->req - tx->done < (tx->mask >> 1) &&
1581	    ss->mgp->running == MYRI10GE_ETH_RUNNING) {
1582		tx->wake_queue++;
1583		netif_tx_wake_queue(dev_queue);
1584	}
1585}
1586
1587static inline int
1588myri10ge_clean_rx_done(struct myri10ge_slice_state *ss, int budget)
1589{
1590	struct myri10ge_rx_done *rx_done = &ss->rx_done;
1591	struct myri10ge_priv *mgp = ss->mgp;
1592	unsigned long rx_bytes = 0;
1593	unsigned long rx_packets = 0;
1594	unsigned long rx_ok;
1595	int idx = rx_done->idx;
1596	int cnt = rx_done->cnt;
1597	int work_done = 0;
1598	u16 length;
1599	__wsum checksum;
1600
1601	while (rx_done->entry[idx].length != 0 && work_done < budget) {
1602		length = ntohs(rx_done->entry[idx].length);
1603		rx_done->entry[idx].length = 0;
1604		checksum = csum_unfold(rx_done->entry[idx].checksum);
1605		rx_ok = myri10ge_rx_done(ss, length, checksum);
1606		rx_packets += rx_ok;
1607		rx_bytes += rx_ok * (unsigned long)length;
1608		cnt++;
1609		idx = cnt & (mgp->max_intr_slots - 1);
1610		work_done++;
1611	}
1612	rx_done->idx = idx;
1613	rx_done->cnt = cnt;
1614	ss->stats.rx_packets += rx_packets;
1615	ss->stats.rx_bytes += rx_bytes;
1616
1617	/* restock receive rings if needed */
1618	if (ss->rx_small.fill_cnt - ss->rx_small.cnt < myri10ge_fill_thresh)
1619		myri10ge_alloc_rx_pages(mgp, &ss->rx_small,
1620					mgp->small_bytes + MXGEFW_PAD, 0);
1621	if (ss->rx_big.fill_cnt - ss->rx_big.cnt < myri10ge_fill_thresh)
1622		myri10ge_alloc_rx_pages(mgp, &ss->rx_big, mgp->big_bytes, 0);
1623
1624	return work_done;
1625}
1626
1627static inline void myri10ge_check_statblock(struct myri10ge_priv *mgp)
1628{
1629	struct mcp_irq_data *stats = mgp->ss[0].fw_stats;
1630
1631	if (unlikely(stats->stats_updated)) {
1632		unsigned link_up = ntohl(stats->link_up);
1633		if (mgp->link_state != link_up) {
1634			mgp->link_state = link_up;
1635
1636			if (mgp->link_state == MXGEFW_LINK_UP) {
1637				netif_info(mgp, link, mgp->dev, "link up\n");
1638				netif_carrier_on(mgp->dev);
1639				mgp->link_changes++;
1640			} else {
1641				netif_info(mgp, link, mgp->dev, "link %s\n",
1642					   (link_up == MXGEFW_LINK_MYRINET ?
1643					    "mismatch (Myrinet detected)" :
1644					    "down"));
1645				netif_carrier_off(mgp->dev);
1646				mgp->link_changes++;
1647			}
1648		}
1649		if (mgp->rdma_tags_available !=
1650		    ntohl(stats->rdma_tags_available)) {
1651			mgp->rdma_tags_available =
1652			    ntohl(stats->rdma_tags_available);
1653			netdev_warn(mgp->dev, "RDMA timed out! %d tags left\n",
1654				    mgp->rdma_tags_available);
1655		}
1656		mgp->down_cnt += stats->link_down;
1657		if (stats->link_down)
1658			wake_up(&mgp->down_wq);
1659	}
1660}
1661
1662static int myri10ge_poll(struct napi_struct *napi, int budget)
1663{
1664	struct myri10ge_slice_state *ss =
1665	    container_of(napi, struct myri10ge_slice_state, napi);
1666	int work_done;
1667
1668#ifdef CONFIG_MYRI10GE_DCA
1669	if (ss->mgp->dca_enabled)
1670		myri10ge_update_dca(ss);
1671#endif
1672	/* Try later if the busy_poll handler is running. */
1673	if (!myri10ge_ss_lock_napi(ss))
1674		return budget;
1675
1676	/* process as many rx events as NAPI will allow */
1677	work_done = myri10ge_clean_rx_done(ss, budget);
1678
1679	myri10ge_ss_unlock_napi(ss);
1680	if (work_done < budget) {
1681		napi_complete(napi);
1682		put_be32(htonl(3), ss->irq_claim);
1683	}
1684	return work_done;
1685}
1686
1687#ifdef CONFIG_NET_RX_BUSY_POLL
1688static int myri10ge_busy_poll(struct napi_struct *napi)
1689{
1690	struct myri10ge_slice_state *ss =
1691	    container_of(napi, struct myri10ge_slice_state, napi);
1692	struct myri10ge_priv *mgp = ss->mgp;
1693	int work_done;
1694
1695	/* Poll only when the link is up */
1696	if (mgp->link_state != MXGEFW_LINK_UP)
1697		return LL_FLUSH_FAILED;
1698
1699	if (!myri10ge_ss_lock_poll(ss))
1700		return LL_FLUSH_BUSY;
1701
1702	/* Process a small number of packets */
1703	work_done = myri10ge_clean_rx_done(ss, 4);
1704	if (work_done)
1705		ss->busy_poll_cnt += work_done;
1706	else
1707		ss->busy_poll_miss++;
1708
1709	myri10ge_ss_unlock_poll(ss);
1710
1711	return work_done;
1712}
1713#endif /* CONFIG_NET_RX_BUSY_POLL */
1714
1715static irqreturn_t myri10ge_intr(int irq, void *arg)
1716{
1717	struct myri10ge_slice_state *ss = arg;
1718	struct myri10ge_priv *mgp = ss->mgp;
1719	struct mcp_irq_data *stats = ss->fw_stats;
1720	struct myri10ge_tx_buf *tx = &ss->tx;
1721	u32 send_done_count;
1722	int i;
1723
1724	/* an interrupt on a non-zero receive-only slice is implicitly
1725	 * valid  since MSI-X irqs are not shared */
1726	if ((mgp->dev->real_num_tx_queues == 1) && (ss != mgp->ss)) {
1727		napi_schedule(&ss->napi);
1728		return IRQ_HANDLED;
1729	}
1730
1731	/* make sure it is our IRQ, and that the DMA has finished */
1732	if (unlikely(!stats->valid))
1733		return IRQ_NONE;
1734
1735	/* low bit indicates receives are present, so schedule
1736	 * napi poll handler */
1737	if (stats->valid & 1)
1738		napi_schedule(&ss->napi);
1739
1740	if (!mgp->msi_enabled && !mgp->msix_enabled) {
1741		put_be32(0, mgp->irq_deassert);
1742		if (!myri10ge_deassert_wait)
1743			stats->valid = 0;
1744		mb();
1745	} else
1746		stats->valid = 0;
1747
1748	/* Wait for IRQ line to go low, if using INTx */
1749	i = 0;
1750	while (1) {
1751		i++;
1752		/* check for transmit completes and receives */
1753		send_done_count = ntohl(stats->send_done_count);
1754		if (send_done_count != tx->pkt_done)
1755			myri10ge_tx_done(ss, (int)send_done_count);
1756		if (unlikely(i > myri10ge_max_irq_loops)) {
1757			netdev_warn(mgp->dev, "irq stuck?\n");
1758			stats->valid = 0;
1759			schedule_work(&mgp->watchdog_work);
1760		}
1761		if (likely(stats->valid == 0))
1762			break;
1763		cpu_relax();
1764		barrier();
1765	}
1766
1767	/* Only slice 0 updates stats */
1768	if (ss == mgp->ss)
1769		myri10ge_check_statblock(mgp);
1770
1771	put_be32(htonl(3), ss->irq_claim + 1);
1772	return IRQ_HANDLED;
1773}
1774
1775static int
1776myri10ge_get_settings(struct net_device *netdev, struct ethtool_cmd *cmd)
1777{
1778	struct myri10ge_priv *mgp = netdev_priv(netdev);
1779	char *ptr;
1780	int i;
1781
1782	cmd->autoneg = AUTONEG_DISABLE;
1783	ethtool_cmd_speed_set(cmd, SPEED_10000);
1784	cmd->duplex = DUPLEX_FULL;
1785
1786	/*
1787	 * parse the product code to deterimine the interface type
1788	 * (CX4, XFP, Quad Ribbon Fiber) by looking at the character
1789	 * after the 3rd dash in the driver's cached copy of the
1790	 * EEPROM's product code string.
1791	 */
1792	ptr = mgp->product_code_string;
1793	if (ptr == NULL) {
1794		netdev_err(netdev, "Missing product code\n");
1795		return 0;
1796	}
1797	for (i = 0; i < 3; i++, ptr++) {
1798		ptr = strchr(ptr, '-');
1799		if (ptr == NULL) {
1800			netdev_err(netdev, "Invalid product code %s\n",
1801				   mgp->product_code_string);
1802			return 0;
1803		}
1804	}
1805	if (*ptr == '2')
1806		ptr++;
1807	if (*ptr == 'R' || *ptr == 'Q' || *ptr == 'S') {
1808		/* We've found either an XFP, quad ribbon fiber, or SFP+ */
1809		cmd->port = PORT_FIBRE;
1810		cmd->supported |= SUPPORTED_FIBRE;
1811		cmd->advertising |= ADVERTISED_FIBRE;
1812	} else {
1813		cmd->port = PORT_OTHER;
1814	}
1815	if (*ptr == 'R' || *ptr == 'S')
1816		cmd->transceiver = XCVR_EXTERNAL;
1817	else
1818		cmd->transceiver = XCVR_INTERNAL;
1819
1820	return 0;
1821}
1822
1823static void
1824myri10ge_get_drvinfo(struct net_device *netdev, struct ethtool_drvinfo *info)
1825{
1826	struct myri10ge_priv *mgp = netdev_priv(netdev);
1827
1828	strlcpy(info->driver, "myri10ge", sizeof(info->driver));
1829	strlcpy(info->version, MYRI10GE_VERSION_STR, sizeof(info->version));
1830	strlcpy(info->fw_version, mgp->fw_version, sizeof(info->fw_version));
1831	strlcpy(info->bus_info, pci_name(mgp->pdev), sizeof(info->bus_info));
1832}
1833
1834static int
1835myri10ge_get_coalesce(struct net_device *netdev, struct ethtool_coalesce *coal)
1836{
1837	struct myri10ge_priv *mgp = netdev_priv(netdev);
1838
1839	coal->rx_coalesce_usecs = mgp->intr_coal_delay;
1840	return 0;
1841}
1842
1843static int
1844myri10ge_set_coalesce(struct net_device *netdev, struct ethtool_coalesce *coal)
1845{
1846	struct myri10ge_priv *mgp = netdev_priv(netdev);
1847
1848	mgp->intr_coal_delay = coal->rx_coalesce_usecs;
1849	put_be32(htonl(mgp->intr_coal_delay), mgp->intr_coal_delay_ptr);
1850	return 0;
1851}
1852
1853static void
1854myri10ge_get_pauseparam(struct net_device *netdev,
1855			struct ethtool_pauseparam *pause)
1856{
1857	struct myri10ge_priv *mgp = netdev_priv(netdev);
1858
1859	pause->autoneg = 0;
1860	pause->rx_pause = mgp->pause;
1861	pause->tx_pause = mgp->pause;
1862}
1863
1864static int
1865myri10ge_set_pauseparam(struct net_device *netdev,
1866			struct ethtool_pauseparam *pause)
1867{
1868	struct myri10ge_priv *mgp = netdev_priv(netdev);
1869
1870	if (pause->tx_pause != mgp->pause)
1871		return myri10ge_change_pause(mgp, pause->tx_pause);
1872	if (pause->rx_pause != mgp->pause)
1873		return myri10ge_change_pause(mgp, pause->rx_pause);
1874	if (pause->autoneg != 0)
1875		return -EINVAL;
1876	return 0;
1877}
1878
1879static void
1880myri10ge_get_ringparam(struct net_device *netdev,
1881		       struct ethtool_ringparam *ring)
1882{
1883	struct myri10ge_priv *mgp = netdev_priv(netdev);
1884
1885	ring->rx_mini_max_pending = mgp->ss[0].rx_small.mask + 1;
1886	ring->rx_max_pending = mgp->ss[0].rx_big.mask + 1;
1887	ring->rx_jumbo_max_pending = 0;
1888	ring->tx_max_pending = mgp->ss[0].tx.mask + 1;
1889	ring->rx_mini_pending = ring->rx_mini_max_pending;
1890	ring->rx_pending = ring->rx_max_pending;
1891	ring->rx_jumbo_pending = ring->rx_jumbo_max_pending;
1892	ring->tx_pending = ring->tx_max_pending;
1893}
1894
1895static const char myri10ge_gstrings_main_stats[][ETH_GSTRING_LEN] = {
1896	"rx_packets", "tx_packets", "rx_bytes", "tx_bytes", "rx_errors",
1897	"tx_errors", "rx_dropped", "tx_dropped", "multicast", "collisions",
1898	"rx_length_errors", "rx_over_errors", "rx_crc_errors",
1899	"rx_frame_errors", "rx_fifo_errors", "rx_missed_errors",
1900	"tx_aborted_errors", "tx_carrier_errors", "tx_fifo_errors",
1901	"tx_heartbeat_errors", "tx_window_errors",
1902	/* device-specific stats */
1903	"tx_boundary", "irq", "MSI", "MSIX",
1904	"read_dma_bw_MBs", "write_dma_bw_MBs", "read_write_dma_bw_MBs",
1905	"serial_number", "watchdog_resets",
1906#ifdef CONFIG_MYRI10GE_DCA
1907	"dca_capable_firmware", "dca_device_present",
1908#endif
1909	"link_changes", "link_up", "dropped_link_overflow",
1910	"dropped_link_error_or_filtered",
1911	"dropped_pause", "dropped_bad_phy", "dropped_bad_crc32",
1912	"dropped_unicast_filtered", "dropped_multicast_filtered",
1913	"dropped_runt", "dropped_overrun", "dropped_no_small_buffer",
1914	"dropped_no_big_buffer"
1915};
1916
1917static const char myri10ge_gstrings_slice_stats[][ETH_GSTRING_LEN] = {
1918	"----------- slice ---------",
1919	"tx_pkt_start", "tx_pkt_done", "tx_req", "tx_done",
1920	"rx_small_cnt", "rx_big_cnt",
1921	"wake_queue", "stop_queue", "tx_linearized",
1922#ifdef CONFIG_NET_RX_BUSY_POLL
1923	"rx_lock_napi_yield", "rx_lock_poll_yield", "rx_busy_poll_miss",
1924	"rx_busy_poll_cnt",
1925#endif
1926};
1927
1928#define MYRI10GE_NET_STATS_LEN      21
1929#define MYRI10GE_MAIN_STATS_LEN  ARRAY_SIZE(myri10ge_gstrings_main_stats)
1930#define MYRI10GE_SLICE_STATS_LEN  ARRAY_SIZE(myri10ge_gstrings_slice_stats)
1931
1932static void
1933myri10ge_get_strings(struct net_device *netdev, u32 stringset, u8 * data)
1934{
1935	struct myri10ge_priv *mgp = netdev_priv(netdev);
1936	int i;
1937
1938	switch (stringset) {
1939	case ETH_SS_STATS:
1940		memcpy(data, *myri10ge_gstrings_main_stats,
1941		       sizeof(myri10ge_gstrings_main_stats));
1942		data += sizeof(myri10ge_gstrings_main_stats);
1943		for (i = 0; i < mgp->num_slices; i++) {
1944			memcpy(data, *myri10ge_gstrings_slice_stats,
1945			       sizeof(myri10ge_gstrings_slice_stats));
1946			data += sizeof(myri10ge_gstrings_slice_stats);
1947		}
1948		break;
1949	}
1950}
1951
1952static int myri10ge_get_sset_count(struct net_device *netdev, int sset)
1953{
1954	struct myri10ge_priv *mgp = netdev_priv(netdev);
1955
1956	switch (sset) {
1957	case ETH_SS_STATS:
1958		return MYRI10GE_MAIN_STATS_LEN +
1959		    mgp->num_slices * MYRI10GE_SLICE_STATS_LEN;
1960	default:
1961		return -EOPNOTSUPP;
1962	}
1963}
1964
1965static void
1966myri10ge_get_ethtool_stats(struct net_device *netdev,
1967			   struct ethtool_stats *stats, u64 * data)
1968{
1969	struct myri10ge_priv *mgp = netdev_priv(netdev);
1970	struct myri10ge_slice_state *ss;
1971	struct rtnl_link_stats64 link_stats;
1972	int slice;
1973	int i;
1974
1975	/* force stats update */
1976	memset(&link_stats, 0, sizeof(link_stats));
1977	(void)myri10ge_get_stats(netdev, &link_stats);
1978	for (i = 0; i < MYRI10GE_NET_STATS_LEN; i++)
1979		data[i] = ((u64 *)&link_stats)[i];
1980
1981	data[i++] = (unsigned int)mgp->tx_boundary;
1982	data[i++] = (unsigned int)mgp->pdev->irq;
1983	data[i++] = (unsigned int)mgp->msi_enabled;
1984	data[i++] = (unsigned int)mgp->msix_enabled;
1985	data[i++] = (unsigned int)mgp->read_dma;
1986	data[i++] = (unsigned int)mgp->write_dma;
1987	data[i++] = (unsigned int)mgp->read_write_dma;
1988	data[i++] = (unsigned int)mgp->serial_number;
1989	data[i++] = (unsigned int)mgp->watchdog_resets;
1990#ifdef CONFIG_MYRI10GE_DCA
1991	data[i++] = (unsigned int)(mgp->ss[0].dca_tag != NULL);
1992	data[i++] = (unsigned int)(mgp->dca_enabled);
1993#endif
1994	data[i++] = (unsigned int)mgp->link_changes;
1995
1996	/* firmware stats are useful only in the first slice */
1997	ss = &mgp->ss[0];
1998	data[i++] = (unsigned int)ntohl(ss->fw_stats->link_up);
1999	data[i++] = (unsigned int)ntohl(ss->fw_stats->dropped_link_overflow);
2000	data[i++] =
2001	    (unsigned int)ntohl(ss->fw_stats->dropped_link_error_or_filtered);
2002	data[i++] = (unsigned int)ntohl(ss->fw_stats->dropped_pause);
2003	data[i++] = (unsigned int)ntohl(ss->fw_stats->dropped_bad_phy);
2004	data[i++] = (unsigned int)ntohl(ss->fw_stats->dropped_bad_crc32);
2005	data[i++] = (unsigned int)ntohl(ss->fw_stats->dropped_unicast_filtered);
2006	data[i++] =
2007	    (unsigned int)ntohl(ss->fw_stats->dropped_multicast_filtered);
2008	data[i++] = (unsigned int)ntohl(ss->fw_stats->dropped_runt);
2009	data[i++] = (unsigned int)ntohl(ss->fw_stats->dropped_overrun);
2010	data[i++] = (unsigned int)ntohl(ss->fw_stats->dropped_no_small_buffer);
2011	data[i++] = (unsigned int)ntohl(ss->fw_stats->dropped_no_big_buffer);
2012
2013	for (slice = 0; slice < mgp->num_slices; slice++) {
2014		ss = &mgp->ss[slice];
2015		data[i++] = slice;
2016		data[i++] = (unsigned int)ss->tx.pkt_start;
2017		data[i++] = (unsigned int)ss->tx.pkt_done;
2018		data[i++] = (unsigned int)ss->tx.req;
2019		data[i++] = (unsigned int)ss->tx.done;
2020		data[i++] = (unsigned int)ss->rx_small.cnt;
2021		data[i++] = (unsigned int)ss->rx_big.cnt;
2022		data[i++] = (unsigned int)ss->tx.wake_queue;
2023		data[i++] = (unsigned int)ss->tx.stop_queue;
2024		data[i++] = (unsigned int)ss->tx.linearized;
2025#ifdef CONFIG_NET_RX_BUSY_POLL
2026		data[i++] = ss->lock_napi_yield;
2027		data[i++] = ss->lock_poll_yield;
2028		data[i++] = ss->busy_poll_miss;
2029		data[i++] = ss->busy_poll_cnt;
2030#endif
2031	}
2032}
2033
2034static void myri10ge_set_msglevel(struct net_device *netdev, u32 value)
2035{
2036	struct myri10ge_priv *mgp = netdev_priv(netdev);
2037	mgp->msg_enable = value;
2038}
2039
2040static u32 myri10ge_get_msglevel(struct net_device *netdev)
2041{
2042	struct myri10ge_priv *mgp = netdev_priv(netdev);
2043	return mgp->msg_enable;
2044}
2045
2046/*
2047 * Use a low-level command to change the LED behavior. Rather than
2048 * blinking (which is the normal case), when identify is used, the
2049 * yellow LED turns solid.
2050 */
2051static int myri10ge_led(struct myri10ge_priv *mgp, int on)
2052{
2053	struct mcp_gen_header *hdr;
2054	struct device *dev = &mgp->pdev->dev;
2055	size_t hdr_off, pattern_off, hdr_len;
2056	u32 pattern = 0xfffffffe;
2057
2058	/* find running firmware header */
2059	hdr_off = swab32(readl(mgp->sram + MCP_HEADER_PTR_OFFSET));
2060	if ((hdr_off & 3) || hdr_off + sizeof(*hdr) > mgp->sram_size) {
2061		dev_err(dev, "Running firmware has bad header offset (%d)\n",
2062			(int)hdr_off);
2063		return -EIO;
2064	}
2065	hdr_len = swab32(readl(mgp->sram + hdr_off +
2066			       offsetof(struct mcp_gen_header, header_length)));
2067	pattern_off = hdr_off + offsetof(struct mcp_gen_header, led_pattern);
2068	if (pattern_off >= (hdr_len + hdr_off)) {
2069		dev_info(dev, "Firmware does not support LED identification\n");
2070		return -EINVAL;
2071	}
2072	if (!on)
2073		pattern = swab32(readl(mgp->sram + pattern_off + 4));
2074	writel(swab32(pattern), mgp->sram + pattern_off);
2075	return 0;
2076}
2077
2078static int
2079myri10ge_phys_id(struct net_device *netdev, enum ethtool_phys_id_state state)
2080{
2081	struct myri10ge_priv *mgp = netdev_priv(netdev);
2082	int rc;
2083
2084	switch (state) {
2085	case ETHTOOL_ID_ACTIVE:
2086		rc = myri10ge_led(mgp, 1);
2087		break;
2088
2089	case ETHTOOL_ID_INACTIVE:
2090		rc =  myri10ge_led(mgp, 0);
2091		break;
2092
2093	default:
2094		rc = -EINVAL;
2095	}
2096
2097	return rc;
2098}
2099
2100static const struct ethtool_ops myri10ge_ethtool_ops = {
2101	.get_settings = myri10ge_get_settings,
2102	.get_drvinfo = myri10ge_get_drvinfo,
2103	.get_coalesce = myri10ge_get_coalesce,
2104	.set_coalesce = myri10ge_set_coalesce,
2105	.get_pauseparam = myri10ge_get_pauseparam,
2106	.set_pauseparam = myri10ge_set_pauseparam,
2107	.get_ringparam = myri10ge_get_ringparam,
2108	.get_link = ethtool_op_get_link,
2109	.get_strings = myri10ge_get_strings,
2110	.get_sset_count = myri10ge_get_sset_count,
2111	.get_ethtool_stats = myri10ge_get_ethtool_stats,
2112	.set_msglevel = myri10ge_set_msglevel,
2113	.get_msglevel = myri10ge_get_msglevel,
2114	.set_phys_id = myri10ge_phys_id,
2115};
2116
2117static int myri10ge_allocate_rings(struct myri10ge_slice_state *ss)
2118{
2119	struct myri10ge_priv *mgp = ss->mgp;
2120	struct myri10ge_cmd cmd;
2121	struct net_device *dev = mgp->dev;
2122	int tx_ring_size, rx_ring_size;
2123	int tx_ring_entries, rx_ring_entries;
2124	int i, slice, status;
2125	size_t bytes;
2126
2127	/* get ring sizes */
2128	slice = ss - mgp->ss;
2129	cmd.data0 = slice;
2130	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_SEND_RING_SIZE, &cmd, 0);
2131	tx_ring_size = cmd.data0;
2132	cmd.data0 = slice;
2133	status |= myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_RX_RING_SIZE, &cmd, 0);
2134	if (status != 0)
2135		return status;
2136	rx_ring_size = cmd.data0;
2137
2138	tx_ring_entries = tx_ring_size / sizeof(struct mcp_kreq_ether_send);
2139	rx_ring_entries = rx_ring_size / sizeof(struct mcp_dma_addr);
2140	ss->tx.mask = tx_ring_entries - 1;
2141	ss->rx_small.mask = ss->rx_big.mask = rx_ring_entries - 1;
2142
2143	status = -ENOMEM;
2144
2145	/* allocate the host shadow rings */
2146
2147	bytes = 8 + (MYRI10GE_MAX_SEND_DESC_TSO + 4)
2148	    * sizeof(*ss->tx.req_list);
2149	ss->tx.req_bytes = kzalloc(bytes, GFP_KERNEL);
2150	if (ss->tx.req_bytes == NULL)
2151		goto abort_with_nothing;
2152
2153	/* ensure req_list entries are aligned to 8 bytes */
2154	ss->tx.req_list = (struct mcp_kreq_ether_send *)
2155	    ALIGN((unsigned long)ss->tx.req_bytes, 8);
2156	ss->tx.queue_active = 0;
2157
2158	bytes = rx_ring_entries * sizeof(*ss->rx_small.shadow);
2159	ss->rx_small.shadow = kzalloc(bytes, GFP_KERNEL);
2160	if (ss->rx_small.shadow == NULL)
2161		goto abort_with_tx_req_bytes;
2162
2163	bytes = rx_ring_entries * sizeof(*ss->rx_big.shadow);
2164	ss->rx_big.shadow = kzalloc(bytes, GFP_KERNEL);
2165	if (ss->rx_big.shadow == NULL)
2166		goto abort_with_rx_small_shadow;
2167
2168	/* allocate the host info rings */
2169
2170	bytes = tx_ring_entries * sizeof(*ss->tx.info);
2171	ss->tx.info = kzalloc(bytes, GFP_KERNEL);
2172	if (ss->tx.info == NULL)
2173		goto abort_with_rx_big_shadow;
2174
2175	bytes = rx_ring_entries * sizeof(*ss->rx_small.info);
2176	ss->rx_small.info = kzalloc(bytes, GFP_KERNEL);
2177	if (ss->rx_small.info == NULL)
2178		goto abort_with_tx_info;
2179
2180	bytes = rx_ring_entries * sizeof(*ss->rx_big.info);
2181	ss->rx_big.info = kzalloc(bytes, GFP_KERNEL);
2182	if (ss->rx_big.info == NULL)
2183		goto abort_with_rx_small_info;
2184
2185	/* Fill the receive rings */
2186	ss->rx_big.cnt = 0;
2187	ss->rx_small.cnt = 0;
2188	ss->rx_big.fill_cnt = 0;
2189	ss->rx_small.fill_cnt = 0;
2190	ss->rx_small.page_offset = MYRI10GE_ALLOC_SIZE;
2191	ss->rx_big.page_offset = MYRI10GE_ALLOC_SIZE;
2192	ss->rx_small.watchdog_needed = 0;
2193	ss->rx_big.watchdog_needed = 0;
2194	if (mgp->small_bytes == 0) {
2195		ss->rx_small.fill_cnt = ss->rx_small.mask + 1;
2196	} else {
2197		myri10ge_alloc_rx_pages(mgp, &ss->rx_small,
2198					mgp->small_bytes + MXGEFW_PAD, 0);
2199	}
2200
2201	if (ss->rx_small.fill_cnt < ss->rx_small.mask + 1) {
2202		netdev_err(dev, "slice-%d: alloced only %d small bufs\n",
2203			   slice, ss->rx_small.fill_cnt);
2204		goto abort_with_rx_small_ring;
2205	}
2206
2207	myri10ge_alloc_rx_pages(mgp, &ss->rx_big, mgp->big_bytes, 0);
2208	if (ss->rx_big.fill_cnt < ss->rx_big.mask + 1) {
2209		netdev_err(dev, "slice-%d: alloced only %d big bufs\n",
2210			   slice, ss->rx_big.fill_cnt);
2211		goto abort_with_rx_big_ring;
2212	}
2213
2214	return 0;
2215
2216abort_with_rx_big_ring:
2217	for (i = ss->rx_big.cnt; i < ss->rx_big.fill_cnt; i++) {
2218		int idx = i & ss->rx_big.mask;
2219		myri10ge_unmap_rx_page(mgp->pdev, &ss->rx_big.info[idx],
2220				       mgp->big_bytes);
2221		put_page(ss->rx_big.info[idx].page);
2222	}
2223
2224abort_with_rx_small_ring:
2225	if (mgp->small_bytes == 0)
2226		ss->rx_small.fill_cnt = ss->rx_small.cnt;
2227	for (i = ss->rx_small.cnt; i < ss->rx_small.fill_cnt; i++) {
2228		int idx = i & ss->rx_small.mask;
2229		myri10ge_unmap_rx_page(mgp->pdev, &ss->rx_small.info[idx],
2230				       mgp->small_bytes + MXGEFW_PAD);
2231		put_page(ss->rx_small.info[idx].page);
2232	}
2233
2234	kfree(ss->rx_big.info);
2235
2236abort_with_rx_small_info:
2237	kfree(ss->rx_small.info);
2238
2239abort_with_tx_info:
2240	kfree(ss->tx.info);
2241
2242abort_with_rx_big_shadow:
2243	kfree(ss->rx_big.shadow);
2244
2245abort_with_rx_small_shadow:
2246	kfree(ss->rx_small.shadow);
2247
2248abort_with_tx_req_bytes:
2249	kfree(ss->tx.req_bytes);
2250	ss->tx.req_bytes = NULL;
2251	ss->tx.req_list = NULL;
2252
2253abort_with_nothing:
2254	return status;
2255}
2256
2257static void myri10ge_free_rings(struct myri10ge_slice_state *ss)
2258{
2259	struct myri10ge_priv *mgp = ss->mgp;
2260	struct sk_buff *skb;
2261	struct myri10ge_tx_buf *tx;
2262	int i, len, idx;
2263
2264	/* If not allocated, skip it */
2265	if (ss->tx.req_list == NULL)
2266		return;
2267
2268	for (i = ss->rx_big.cnt; i < ss->rx_big.fill_cnt; i++) {
2269		idx = i & ss->rx_big.mask;
2270		if (i == ss->rx_big.fill_cnt - 1)
2271			ss->rx_big.info[idx].page_offset = MYRI10GE_ALLOC_SIZE;
2272		myri10ge_unmap_rx_page(mgp->pdev, &ss->rx_big.info[idx],
2273				       mgp->big_bytes);
2274		put_page(ss->rx_big.info[idx].page);
2275	}
2276
2277	if (mgp->small_bytes == 0)
2278		ss->rx_small.fill_cnt = ss->rx_small.cnt;
2279	for (i = ss->rx_small.cnt; i < ss->rx_small.fill_cnt; i++) {
2280		idx = i & ss->rx_small.mask;
2281		if (i == ss->rx_small.fill_cnt - 1)
2282			ss->rx_small.info[idx].page_offset =
2283			    MYRI10GE_ALLOC_SIZE;
2284		myri10ge_unmap_rx_page(mgp->pdev, &ss->rx_small.info[idx],
2285				       mgp->small_bytes + MXGEFW_PAD);
2286		put_page(ss->rx_small.info[idx].page);
2287	}
2288	tx = &ss->tx;
2289	while (tx->done != tx->req) {
2290		idx = tx->done & tx->mask;
2291		skb = tx->info[idx].skb;
2292
2293		/* Mark as free */
2294		tx->info[idx].skb = NULL;
2295		tx->done++;
2296		len = dma_unmap_len(&tx->info[idx], len);
2297		dma_unmap_len_set(&tx->info[idx], len, 0);
2298		if (skb) {
2299			ss->stats.tx_dropped++;
2300			dev_kfree_skb_any(skb);
2301			if (len)
2302				pci_unmap_single(mgp->pdev,
2303						 dma_unmap_addr(&tx->info[idx],
2304								bus), len,
2305						 PCI_DMA_TODEVICE);
2306		} else {
2307			if (len)
2308				pci_unmap_page(mgp->pdev,
2309					       dma_unmap_addr(&tx->info[idx],
2310							      bus), len,
2311					       PCI_DMA_TODEVICE);
2312		}
2313	}
2314	kfree(ss->rx_big.info);
2315
2316	kfree(ss->rx_small.info);
2317
2318	kfree(ss->tx.info);
2319
2320	kfree(ss->rx_big.shadow);
2321
2322	kfree(ss->rx_small.shadow);
2323
2324	kfree(ss->tx.req_bytes);
2325	ss->tx.req_bytes = NULL;
2326	ss->tx.req_list = NULL;
2327}
2328
2329static int myri10ge_request_irq(struct myri10ge_priv *mgp)
2330{
2331	struct pci_dev *pdev = mgp->pdev;
2332	struct myri10ge_slice_state *ss;
2333	struct net_device *netdev = mgp->dev;
2334	int i;
2335	int status;
2336
2337	mgp->msi_enabled = 0;
2338	mgp->msix_enabled = 0;
2339	status = 0;
2340	if (myri10ge_msi) {
2341		if (mgp->num_slices > 1) {
2342			status = pci_enable_msix_range(pdev, mgp->msix_vectors,
2343					mgp->num_slices, mgp->num_slices);
2344			if (status < 0) {
2345				dev_err(&pdev->dev,
2346					"Error %d setting up MSI-X\n", status);
2347				return status;
2348			}
2349			mgp->msix_enabled = 1;
2350		}
2351		if (mgp->msix_enabled == 0) {
2352			status = pci_enable_msi(pdev);
2353			if (status != 0) {
2354				dev_err(&pdev->dev,
2355					"Error %d setting up MSI; falling back to xPIC\n",
2356					status);
2357			} else {
2358				mgp->msi_enabled = 1;
2359			}
2360		}
2361	}
2362	if (mgp->msix_enabled) {
2363		for (i = 0; i < mgp->num_slices; i++) {
2364			ss = &mgp->ss[i];
2365			snprintf(ss->irq_desc, sizeof(ss->irq_desc),
2366				 "%s:slice-%d", netdev->name, i);
2367			status = request_irq(mgp->msix_vectors[i].vector,
2368					     myri10ge_intr, 0, ss->irq_desc,
2369					     ss);
2370			if (status != 0) {
2371				dev_err(&pdev->dev,
2372					"slice %d failed to allocate IRQ\n", i);
2373				i--;
2374				while (i >= 0) {
2375					free_irq(mgp->msix_vectors[i].vector,
2376						 &mgp->ss[i]);
2377					i--;
2378				}
2379				pci_disable_msix(pdev);
2380				return status;
2381			}
2382		}
2383	} else {
2384		status = request_irq(pdev->irq, myri10ge_intr, IRQF_SHARED,
2385				     mgp->dev->name, &mgp->ss[0]);
2386		if (status != 0) {
2387			dev_err(&pdev->dev, "failed to allocate IRQ\n");
2388			if (mgp->msi_enabled)
2389				pci_disable_msi(pdev);
2390		}
2391	}
2392	return status;
2393}
2394
2395static void myri10ge_free_irq(struct myri10ge_priv *mgp)
2396{
2397	struct pci_dev *pdev = mgp->pdev;
2398	int i;
2399
2400	if (mgp->msix_enabled) {
2401		for (i = 0; i < mgp->num_slices; i++)
2402			free_irq(mgp->msix_vectors[i].vector, &mgp->ss[i]);
2403	} else {
2404		free_irq(pdev->irq, &mgp->ss[0]);
2405	}
2406	if (mgp->msi_enabled)
2407		pci_disable_msi(pdev);
2408	if (mgp->msix_enabled)
2409		pci_disable_msix(pdev);
2410}
2411
2412static int myri10ge_get_txrx(struct myri10ge_priv *mgp, int slice)
2413{
2414	struct myri10ge_cmd cmd;
2415	struct myri10ge_slice_state *ss;
2416	int status;
2417
2418	ss = &mgp->ss[slice];
2419	status = 0;
2420	if (slice == 0 || (mgp->dev->real_num_tx_queues > 1)) {
2421		cmd.data0 = slice;
2422		status = myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_SEND_OFFSET,
2423					   &cmd, 0);
2424		ss->tx.lanai = (struct mcp_kreq_ether_send __iomem *)
2425		    (mgp->sram + cmd.data0);
2426	}
2427	cmd.data0 = slice;
2428	status |= myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_SMALL_RX_OFFSET,
2429				    &cmd, 0);
2430	ss->rx_small.lanai = (struct mcp_kreq_ether_recv __iomem *)
2431	    (mgp->sram + cmd.data0);
2432
2433	cmd.data0 = slice;
2434	status |= myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_BIG_RX_OFFSET, &cmd, 0);
2435	ss->rx_big.lanai = (struct mcp_kreq_ether_recv __iomem *)
2436	    (mgp->sram + cmd.data0);
2437
2438	ss->tx.send_go = (__iomem __be32 *)
2439	    (mgp->sram + MXGEFW_ETH_SEND_GO + 64 * slice);
2440	ss->tx.send_stop = (__iomem __be32 *)
2441	    (mgp->sram + MXGEFW_ETH_SEND_STOP + 64 * slice);
2442	return status;
2443
2444}
2445
2446static int myri10ge_set_stats(struct myri10ge_priv *mgp, int slice)
2447{
2448	struct myri10ge_cmd cmd;
2449	struct myri10ge_slice_state *ss;
2450	int status;
2451
2452	ss = &mgp->ss[slice];
2453	cmd.data0 = MYRI10GE_LOWPART_TO_U32(ss->fw_stats_bus);
2454	cmd.data1 = MYRI10GE_HIGHPART_TO_U32(ss->fw_stats_bus);
2455	cmd.data2 = sizeof(struct mcp_irq_data) | (slice << 16);
2456	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_STATS_DMA_V2, &cmd, 0);
2457	if (status == -ENOSYS) {
2458		dma_addr_t bus = ss->fw_stats_bus;
2459		if (slice != 0)
2460			return -EINVAL;
2461		bus += offsetof(struct mcp_irq_data, send_done_count);
2462		cmd.data0 = MYRI10GE_LOWPART_TO_U32(bus);
2463		cmd.data1 = MYRI10GE_HIGHPART_TO_U32(bus);
2464		status = myri10ge_send_cmd(mgp,
2465					   MXGEFW_CMD_SET_STATS_DMA_OBSOLETE,
2466					   &cmd, 0);
2467		/* Firmware cannot support multicast without STATS_DMA_V2 */
2468		mgp->fw_multicast_support = 0;
2469	} else {
2470		mgp->fw_multicast_support = 1;
2471	}
2472	return 0;
2473}
2474
2475static int myri10ge_open(struct net_device *dev)
2476{
2477	struct myri10ge_slice_state *ss;
2478	struct myri10ge_priv *mgp = netdev_priv(dev);
2479	struct myri10ge_cmd cmd;
2480	int i, status, big_pow2, slice;
2481	u8 __iomem *itable;
2482
2483	if (mgp->running != MYRI10GE_ETH_STOPPED)
2484		return -EBUSY;
2485
2486	mgp->running = MYRI10GE_ETH_STARTING;
2487	status = myri10ge_reset(mgp);
2488	if (status != 0) {
2489		netdev_err(dev, "failed reset\n");
2490		goto abort_with_nothing;
2491	}
2492
2493	if (mgp->num_slices > 1) {
2494		cmd.data0 = mgp->num_slices;
2495		cmd.data1 = MXGEFW_SLICE_INTR_MODE_ONE_PER_SLICE;
2496		if (mgp->dev->real_num_tx_queues > 1)
2497			cmd.data1 |= MXGEFW_SLICE_ENABLE_MULTIPLE_TX_QUEUES;
2498		status = myri10ge_send_cmd(mgp, MXGEFW_CMD_ENABLE_RSS_QUEUES,
2499					   &cmd, 0);
2500		if (status != 0) {
2501			netdev_err(dev, "failed to set number of slices\n");
2502			goto abort_with_nothing;
2503		}
2504		/* setup the indirection table */
2505		cmd.data0 = mgp->num_slices;
2506		status = myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_RSS_TABLE_SIZE,
2507					   &cmd, 0);
2508
2509		status |= myri10ge_send_cmd(mgp,
2510					    MXGEFW_CMD_GET_RSS_TABLE_OFFSET,
2511					    &cmd, 0);
2512		if (status != 0) {
2513			netdev_err(dev, "failed to setup rss tables\n");
2514			goto abort_with_nothing;
2515		}
2516
2517		/* just enable an identity mapping */
2518		itable = mgp->sram + cmd.data0;
2519		for (i = 0; i < mgp->num_slices; i++)
2520			__raw_writeb(i, &itable[i]);
2521
2522		cmd.data0 = 1;
2523		cmd.data1 = myri10ge_rss_hash;
2524		status = myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_RSS_ENABLE,
2525					   &cmd, 0);
2526		if (status != 0) {
2527			netdev_err(dev, "failed to enable slices\n");
2528			goto abort_with_nothing;
2529		}
2530	}
2531
2532	status = myri10ge_request_irq(mgp);
2533	if (status != 0)
2534		goto abort_with_nothing;
2535
2536	/* decide what small buffer size to use.  For good TCP rx
2537	 * performance, it is important to not receive 1514 byte
2538	 * frames into jumbo buffers, as it confuses the socket buffer
2539	 * accounting code, leading to drops and erratic performance.
2540	 */
2541
2542	if (dev->mtu <= ETH_DATA_LEN)
2543		/* enough for a TCP header */
2544		mgp->small_bytes = (128 > SMP_CACHE_BYTES)
2545		    ? (128 - MXGEFW_PAD)
2546		    : (SMP_CACHE_BYTES - MXGEFW_PAD);
2547	else
2548		/* enough for a vlan encapsulated ETH_DATA_LEN frame */
2549		mgp->small_bytes = VLAN_ETH_FRAME_LEN;
2550
2551	/* Override the small buffer size? */
2552	if (myri10ge_small_bytes >= 0)
2553		mgp->small_bytes = myri10ge_small_bytes;
2554
2555	/* Firmware needs the big buff size as a power of 2.  Lie and
2556	 * tell him the buffer is larger, because we only use 1
2557	 * buffer/pkt, and the mtu will prevent overruns.
2558	 */
2559	big_pow2 = dev->mtu + ETH_HLEN + VLAN_HLEN + MXGEFW_PAD;
2560	if (big_pow2 < MYRI10GE_ALLOC_SIZE / 2) {
2561		while (!is_power_of_2(big_pow2))
2562			big_pow2++;
2563		mgp->big_bytes = dev->mtu + ETH_HLEN + VLAN_HLEN + MXGEFW_PAD;
2564	} else {
2565		big_pow2 = MYRI10GE_ALLOC_SIZE;
2566		mgp->big_bytes = big_pow2;
2567	}
2568
2569	/* setup the per-slice data structures */
2570	for (slice = 0; slice < mgp->num_slices; slice++) {
2571		ss = &mgp->ss[slice];
2572
2573		status = myri10ge_get_txrx(mgp, slice);
2574		if (status != 0) {
2575			netdev_err(dev, "failed to get ring sizes or locations\n");
2576			goto abort_with_rings;
2577		}
2578		status = myri10ge_allocate_rings(ss);
2579		if (status != 0)
2580			goto abort_with_rings;
2581
2582		/* only firmware which supports multiple TX queues
2583		 * supports setting up the tx stats on non-zero
2584		 * slices */
2585		if (slice == 0 || mgp->dev->real_num_tx_queues > 1)
2586			status = myri10ge_set_stats(mgp, slice);
2587		if (status) {
2588			netdev_err(dev, "Couldn't set stats DMA\n");
2589			goto abort_with_rings;
2590		}
2591
2592		/* Initialize the slice spinlock and state used for polling */
2593		myri10ge_ss_init_lock(ss);
2594
2595		/* must happen prior to any irq */
2596		napi_enable(&(ss)->napi);
2597	}
2598
2599	/* now give firmware buffers sizes, and MTU */
2600	cmd.data0 = dev->mtu + ETH_HLEN + VLAN_HLEN;
2601	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_MTU, &cmd, 0);
2602	cmd.data0 = mgp->small_bytes;
2603	status |=
2604	    myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_SMALL_BUFFER_SIZE, &cmd, 0);
2605	cmd.data0 = big_pow2;
2606	status |=
2607	    myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_BIG_BUFFER_SIZE, &cmd, 0);
2608	if (status) {
2609		netdev_err(dev, "Couldn't set buffer sizes\n");
2610		goto abort_with_rings;
2611	}
2612
2613	/*
2614	 * Set Linux style TSO mode; this is needed only on newer
2615	 *  firmware versions.  Older versions default to Linux
2616	 *  style TSO
2617	 */
2618	cmd.data0 = 0;
2619	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_TSO_MODE, &cmd, 0);
2620	if (status && status != -ENOSYS) {
2621		netdev_err(dev, "Couldn't set TSO mode\n");
2622		goto abort_with_rings;
2623	}
2624
2625	mgp->link_state = ~0U;
2626	mgp->rdma_tags_available = 15;
2627
2628	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_ETHERNET_UP, &cmd, 0);
2629	if (status) {
2630		netdev_err(dev, "Couldn't bring up link\n");
2631		goto abort_with_rings;
2632	}
2633
2634	mgp->running = MYRI10GE_ETH_RUNNING;
2635	mgp->watchdog_timer.expires = jiffies + myri10ge_watchdog_timeout * HZ;
2636	add_timer(&mgp->watchdog_timer);
2637	netif_tx_wake_all_queues(dev);
2638
2639	return 0;
2640
2641abort_with_rings:
2642	while (slice) {
2643		slice--;
2644		napi_disable(&mgp->ss[slice].napi);
2645	}
2646	for (i = 0; i < mgp->num_slices; i++)
2647		myri10ge_free_rings(&mgp->ss[i]);
2648
2649	myri10ge_free_irq(mgp);
2650
2651abort_with_nothing:
2652	mgp->running = MYRI10GE_ETH_STOPPED;
2653	return -ENOMEM;
2654}
2655
2656static int myri10ge_close(struct net_device *dev)
2657{
2658	struct myri10ge_priv *mgp = netdev_priv(dev);
2659	struct myri10ge_cmd cmd;
2660	int status, old_down_cnt;
2661	int i;
2662
2663	if (mgp->running != MYRI10GE_ETH_RUNNING)
2664		return 0;
2665
2666	if (mgp->ss[0].tx.req_bytes == NULL)
2667		return 0;
2668
2669	del_timer_sync(&mgp->watchdog_timer);
2670	mgp->running = MYRI10GE_ETH_STOPPING;
2671	local_bh_disable(); /* myri10ge_ss_lock_napi needs bh disabled */
2672	for (i = 0; i < mgp->num_slices; i++) {
2673		napi_disable(&mgp->ss[i].napi);
2674		/* Lock the slice to prevent the busy_poll handler from
2675		 * accessing it.  Later when we bring the NIC up, myri10ge_open
2676		 * resets the slice including this lock.
2677		 */
2678		while (!myri10ge_ss_lock_napi(&mgp->ss[i])) {
2679			pr_info("Slice %d locked\n", i);
2680			mdelay(1);
2681		}
2682	}
2683	local_bh_enable();
2684	netif_carrier_off(dev);
2685
2686	netif_tx_stop_all_queues(dev);
2687	if (mgp->rebooted == 0) {
2688		old_down_cnt = mgp->down_cnt;
2689		mb();
2690		status =
2691		    myri10ge_send_cmd(mgp, MXGEFW_CMD_ETHERNET_DOWN, &cmd, 0);
2692		if (status)
2693			netdev_err(dev, "Couldn't bring down link\n");
2694
2695		wait_event_timeout(mgp->down_wq, old_down_cnt != mgp->down_cnt,
2696				   HZ);
2697		if (old_down_cnt == mgp->down_cnt)
2698			netdev_err(dev, "never got down irq\n");
2699	}
2700	netif_tx_disable(dev);
2701	myri10ge_free_irq(mgp);
2702	for (i = 0; i < mgp->num_slices; i++)
2703		myri10ge_free_rings(&mgp->ss[i]);
2704
2705	mgp->running = MYRI10GE_ETH_STOPPED;
2706	return 0;
2707}
2708
2709/* copy an array of struct mcp_kreq_ether_send's to the mcp.  Copy
2710 * backwards one at a time and handle ring wraps */
2711
2712static inline void
2713myri10ge_submit_req_backwards(struct myri10ge_tx_buf *tx,
2714			      struct mcp_kreq_ether_send *src, int cnt)
2715{
2716	int idx, starting_slot;
2717	starting_slot = tx->req;
2718	while (cnt > 1) {
2719		cnt--;
2720		idx = (starting_slot + cnt) & tx->mask;
2721		myri10ge_pio_copy(&tx->lanai[idx], &src[cnt], sizeof(*src));
2722		mb();
2723	}
2724}
2725
2726/*
2727 * copy an array of struct mcp_kreq_ether_send's to the mcp.  Copy
2728 * at most 32 bytes at a time, so as to avoid involving the software
2729 * pio handler in the nic.   We re-write the first segment's flags
2730 * to mark them valid only after writing the entire chain.
2731 */
2732
2733static inline void
2734myri10ge_submit_req(struct myri10ge_tx_buf *tx, struct mcp_kreq_ether_send *src,
2735		    int cnt)
2736{
2737	int idx, i;
2738	struct mcp_kreq_ether_send __iomem *dstp, *dst;
2739	struct mcp_kreq_ether_send *srcp;
2740	u8 last_flags;
2741
2742	idx = tx->req & tx->mask;
2743
2744	last_flags = src->flags;
2745	src->flags = 0;
2746	mb();
2747	dst = dstp = &tx->lanai[idx];
2748	srcp = src;
2749
2750	if ((idx + cnt) < tx->mask) {
2751		for (i = 0; i < (cnt - 1); i += 2) {
2752			myri10ge_pio_copy(dstp, srcp, 2 * sizeof(*src));
2753			mb();	/* force write every 32 bytes */
2754			srcp += 2;
2755			dstp += 2;
2756		}
2757	} else {
2758		/* submit all but the first request, and ensure
2759		 * that it is submitted below */
2760		myri10ge_submit_req_backwards(tx, src, cnt);
2761		i = 0;
2762	}
2763	if (i < cnt) {
2764		/* submit the first request */
2765		myri10ge_pio_copy(dstp, srcp, sizeof(*src));
2766		mb();		/* barrier before setting valid flag */
2767	}
2768
2769	/* re-write the last 32-bits with the valid flags */
2770	src->flags = last_flags;
2771	put_be32(*((__be32 *) src + 3), (__be32 __iomem *) dst + 3);
2772	tx->req += cnt;
2773	mb();
2774}
2775
2776static void myri10ge_unmap_tx_dma(struct myri10ge_priv *mgp,
2777				  struct myri10ge_tx_buf *tx, int idx)
2778{
2779	unsigned int len;
2780	int last_idx;
2781
2782	/* Free any DMA resources we've alloced and clear out the skb slot */
2783	last_idx = (idx + 1) & tx->mask;
2784	idx = tx->req & tx->mask;
2785	do {
2786		len = dma_unmap_len(&tx->info[idx], len);
2787		if (len) {
2788			if (tx->info[idx].skb != NULL)
2789				pci_unmap_single(mgp->pdev,
2790						 dma_unmap_addr(&tx->info[idx],
2791								bus), len,
2792						 PCI_DMA_TODEVICE);
2793			else
2794				pci_unmap_page(mgp->pdev,
2795					       dma_unmap_addr(&tx->info[idx],
2796							      bus), len,
2797					       PCI_DMA_TODEVICE);
2798			dma_unmap_len_set(&tx->info[idx], len, 0);
2799			tx->info[idx].skb = NULL;
2800		}
2801		idx = (idx + 1) & tx->mask;
2802	} while (idx != last_idx);
2803}
2804
2805/*
2806 * Transmit a packet.  We need to split the packet so that a single
2807 * segment does not cross myri10ge->tx_boundary, so this makes segment
2808 * counting tricky.  So rather than try to count segments up front, we
2809 * just give up if there are too few segments to hold a reasonably
2810 * fragmented packet currently available.  If we run
2811 * out of segments while preparing a packet for DMA, we just linearize
2812 * it and try again.
2813 */
2814
2815static netdev_tx_t myri10ge_xmit(struct sk_buff *skb,
2816				       struct net_device *dev)
2817{
2818	struct myri10ge_priv *mgp = netdev_priv(dev);
2819	struct myri10ge_slice_state *ss;
2820	struct mcp_kreq_ether_send *req;
2821	struct myri10ge_tx_buf *tx;
2822	struct skb_frag_struct *frag;
2823	struct netdev_queue *netdev_queue;
2824	dma_addr_t bus;
2825	u32 low;
2826	__be32 high_swapped;
2827	unsigned int len;
2828	int idx, avail, frag_cnt, frag_idx, count, mss, max_segments;
2829	u16 pseudo_hdr_offset, cksum_offset, queue;
2830	int cum_len, seglen, boundary, rdma_count;
2831	u8 flags, odd_flag;
2832
2833	queue = skb_get_queue_mapping(skb);
2834	ss = &mgp->ss[queue];
2835	netdev_queue = netdev_get_tx_queue(mgp->dev, queue);
2836	tx = &ss->tx;
2837
2838again:
2839	req = tx->req_list;
2840	avail = tx->mask - 1 - (tx->req - tx->done);
2841
2842	mss = 0;
2843	max_segments = MXGEFW_MAX_SEND_DESC;
2844
2845	if (skb_is_gso(skb)) {
2846		mss = skb_shinfo(skb)->gso_size;
2847		max_segments = MYRI10GE_MAX_SEND_DESC_TSO;
2848	}
2849
2850	if ((unlikely(avail < max_segments))) {
2851		/* we are out of transmit resources */
2852		tx->stop_queue++;
2853		netif_tx_stop_queue(netdev_queue);
2854		return NETDEV_TX_BUSY;
2855	}
2856
2857	/* Setup checksum offloading, if needed */
2858	cksum_offset = 0;
2859	pseudo_hdr_offset = 0;
2860	odd_flag = 0;
2861	flags = (MXGEFW_FLAGS_NO_TSO | MXGEFW_FLAGS_FIRST);
2862	if (likely(skb->ip_summed == CHECKSUM_PARTIAL)) {
2863		cksum_offset = skb_checksum_start_offset(skb);
2864		pseudo_hdr_offset = cksum_offset + skb->csum_offset;
2865		/* If the headers are excessively large, then we must
2866		 * fall back to a software checksum */
2867		if (unlikely(!mss && (cksum_offset > 255 ||
2868				      pseudo_hdr_offset > 127))) {
2869			if (skb_checksum_help(skb))
2870				goto drop;
2871			cksum_offset = 0;
2872			pseudo_hdr_offset = 0;
2873		} else {
2874			odd_flag = MXGEFW_FLAGS_ALIGN_ODD;
2875			flags |= MXGEFW_FLAGS_CKSUM;
2876		}
2877	}
2878
2879	cum_len = 0;
2880
2881	if (mss) {		/* TSO */
2882		/* this removes any CKSUM flag from before */
2883		flags = (MXGEFW_FLAGS_TSO_HDR | MXGEFW_FLAGS_FIRST);
2884
2885		/* negative cum_len signifies to the
2886		 * send loop that we are still in the
2887		 * header portion of the TSO packet.
2888		 * TSO header can be at most 1KB long */
2889		cum_len = -(skb_transport_offset(skb) + tcp_hdrlen(skb));
2890
2891		/* for IPv6 TSO, the checksum offset stores the
2892		 * TCP header length, to save the firmware from
2893		 * the need to parse the headers */
2894		if (skb_is_gso_v6(skb)) {
2895			cksum_offset = tcp_hdrlen(skb);
2896			/* Can only handle headers <= max_tso6 long */
2897			if (unlikely(-cum_len > mgp->max_tso6))
2898				return myri10ge_sw_tso(skb, dev);
2899		}
2900		/* for TSO, pseudo_hdr_offset holds mss.
2901		 * The firmware figures out where to put
2902		 * the checksum by parsing the header. */
2903		pseudo_hdr_offset = mss;
2904	} else
2905		/* Mark small packets, and pad out tiny packets */
2906	if (skb->len <= MXGEFW_SEND_SMALL_SIZE) {
2907		flags |= MXGEFW_FLAGS_SMALL;
2908
2909		/* pad frames to at least ETH_ZLEN bytes */
2910		if (eth_skb_pad(skb)) {
2911			/* The packet is gone, so we must
2912			 * return 0 */
2913			ss->stats.tx_dropped += 1;
2914			return NETDEV_TX_OK;
2915		}
2916	}
2917
2918	/* map the skb for DMA */
2919	len = skb_headlen(skb);
2920	bus = pci_map_single(mgp->pdev, skb->data, len, PCI_DMA_TODEVICE);
2921	if (unlikely(pci_dma_mapping_error(mgp->pdev, bus)))
2922		goto drop;
2923
2924	idx = tx->req & tx->mask;
2925	tx->info[idx].skb = skb;
2926	dma_unmap_addr_set(&tx->info[idx], bus, bus);
2927	dma_unmap_len_set(&tx->info[idx], len, len);
2928
2929	frag_cnt = skb_shinfo(skb)->nr_frags;
2930	frag_idx = 0;
2931	count = 0;
2932	rdma_count = 0;
2933
2934	/* "rdma_count" is the number of RDMAs belonging to the
2935	 * current packet BEFORE the current send request. For
2936	 * non-TSO packets, this is equal to "count".
2937	 * For TSO packets, rdma_count needs to be reset
2938	 * to 0 after a segment cut.
2939	 *
2940	 * The rdma_count field of the send request is
2941	 * the number of RDMAs of the packet starting at
2942	 * that request. For TSO send requests with one ore more cuts
2943	 * in the middle, this is the number of RDMAs starting
2944	 * after the last cut in the request. All previous
2945	 * segments before the last cut implicitly have 1 RDMA.
2946	 *
2947	 * Since the number of RDMAs is not known beforehand,
2948	 * it must be filled-in retroactively - after each
2949	 * segmentation cut or at the end of the entire packet.
2950	 */
2951
2952	while (1) {
2953		/* Break the SKB or Fragment up into pieces which
2954		 * do not cross mgp->tx_boundary */
2955		low = MYRI10GE_LOWPART_TO_U32(bus);
2956		high_swapped = htonl(MYRI10GE_HIGHPART_TO_U32(bus));
2957		while (len) {
2958			u8 flags_next;
2959			int cum_len_next;
2960
2961			if (unlikely(count == max_segments))
2962				goto abort_linearize;
2963
2964			boundary =
2965			    (low + mgp->tx_boundary) & ~(mgp->tx_boundary - 1);
2966			seglen = boundary - low;
2967			if (seglen > len)
2968				seglen = len;
2969			flags_next = flags & ~MXGEFW_FLAGS_FIRST;
2970			cum_len_next = cum_len + seglen;
2971			if (mss) {	/* TSO */
2972				(req - rdma_count)->rdma_count = rdma_count + 1;
2973
2974				if (likely(cum_len >= 0)) {	/* payload */
2975					int next_is_first, chop;
2976
2977					chop = (cum_len_next > mss);
2978					cum_len_next = cum_len_next % mss;
2979					next_is_first = (cum_len_next == 0);
2980					flags |= chop * MXGEFW_FLAGS_TSO_CHOP;
2981					flags_next |= next_is_first *
2982					    MXGEFW_FLAGS_FIRST;
2983					rdma_count |= -(chop | next_is_first);
2984					rdma_count += chop & ~next_is_first;
2985				} else if (likely(cum_len_next >= 0)) {	/* header ends */
2986					int small;
2987
2988					rdma_count = -1;
2989					cum_len_next = 0;
2990					seglen = -cum_len;
2991					small = (mss <= MXGEFW_SEND_SMALL_SIZE);
2992					flags_next = MXGEFW_FLAGS_TSO_PLD |
2993					    MXGEFW_FLAGS_FIRST |
2994					    (small * MXGEFW_FLAGS_SMALL);
2995				}
2996			}
2997			req->addr_high = high_swapped;
2998			req->addr_low = htonl(low);
2999			req->pseudo_hdr_offset = htons(pseudo_hdr_offset);
3000			req->pad = 0;	/* complete solid 16-byte block; does this matter? */
3001			req->rdma_count = 1;
3002			req->length = htons(seglen);
3003			req->cksum_offset = cksum_offset;
3004			req->flags = flags | ((cum_len & 1) * odd_flag);
3005
3006			low += seglen;
3007			len -= seglen;
3008			cum_len = cum_len_next;
3009			flags = flags_next;
3010			req++;
3011			count++;
3012			rdma_count++;
3013			if (cksum_offset != 0 && !(mss && skb_is_gso_v6(skb))) {
3014				if (unlikely(cksum_offset > seglen))
3015					cksum_offset -= seglen;
3016				else
3017					cksum_offset = 0;
3018			}
3019		}
3020		if (frag_idx == frag_cnt)
3021			break;
3022
3023		/* map next fragment for DMA */
3024		frag = &skb_shinfo(skb)->frags[frag_idx];
3025		frag_idx++;
3026		len = skb_frag_size(frag);
3027		bus = skb_frag_dma_map(&mgp->pdev->dev, frag, 0, len,
3028				       DMA_TO_DEVICE);
3029		if (unlikely(pci_dma_mapping_error(mgp->pdev, bus))) {
3030			myri10ge_unmap_tx_dma(mgp, tx, idx);
3031			goto drop;
3032		}
3033		idx = (count + tx->req) & tx->mask;
3034		dma_unmap_addr_set(&tx->info[idx], bus, bus);
3035		dma_unmap_len_set(&tx->info[idx], len, len);
3036	}
3037
3038	(req - rdma_count)->rdma_count = rdma_count;
3039	if (mss)
3040		do {
3041			req--;
3042			req->flags |= MXGEFW_FLAGS_TSO_LAST;
3043		} while (!(req->flags & (MXGEFW_FLAGS_TSO_CHOP |
3044					 MXGEFW_FLAGS_FIRST)));
3045	idx = ((count - 1) + tx->req) & tx->mask;
3046	tx->info[idx].last = 1;
3047	myri10ge_submit_req(tx, tx->req_list, count);
3048	/* if using multiple tx queues, make sure NIC polls the
3049	 * current slice */
3050	if ((mgp->dev->real_num_tx_queues > 1) && tx->queue_active == 0) {
3051		tx->queue_active = 1;
3052		put_be32(htonl(1), tx->send_go);
3053		mb();
3054		mmiowb();
3055	}
3056	tx->pkt_start++;
3057	if ((avail - count) < MXGEFW_MAX_SEND_DESC) {
3058		tx->stop_queue++;
3059		netif_tx_stop_queue(netdev_queue);
3060	}
3061	return NETDEV_TX_OK;
3062
3063abort_linearize:
3064	myri10ge_unmap_tx_dma(mgp, tx, idx);
3065
3066	if (skb_is_gso(skb)) {
3067		netdev_err(mgp->dev, "TSO but wanted to linearize?!?!?\n");
3068		goto drop;
3069	}
3070
3071	if (skb_linearize(skb))
3072		goto drop;
3073
3074	tx->linearized++;
3075	goto again;
3076
3077drop:
3078	dev_kfree_skb_any(skb);
3079	ss->stats.tx_dropped += 1;
3080	return NETDEV_TX_OK;
3081
3082}
3083
3084static netdev_tx_t myri10ge_sw_tso(struct sk_buff *skb,
3085					 struct net_device *dev)
3086{
3087	struct sk_buff *segs, *curr;
3088	struct myri10ge_priv *mgp = netdev_priv(dev);
3089	struct myri10ge_slice_state *ss;
3090	netdev_tx_t status;
3091
3092	segs = skb_gso_segment(skb, dev->features & ~NETIF_F_TSO6);
3093	if (IS_ERR(segs))
3094		goto drop;
3095
3096	while (segs) {
3097		curr = segs;
3098		segs = segs->next;
3099		curr->next = NULL;
3100		status = myri10ge_xmit(curr, dev);
3101		if (status != 0) {
3102			dev_kfree_skb_any(curr);
3103			if (segs != NULL) {
3104				curr = segs;
3105				segs = segs->next;
3106				curr->next = NULL;
3107				dev_kfree_skb_any(segs);
3108			}
3109			goto drop;
3110		}
3111	}
3112	dev_kfree_skb_any(skb);
3113	return NETDEV_TX_OK;
3114
3115drop:
3116	ss = &mgp->ss[skb_get_queue_mapping(skb)];
3117	dev_kfree_skb_any(skb);
3118	ss->stats.tx_dropped += 1;
3119	return NETDEV_TX_OK;
3120}
3121
3122static struct rtnl_link_stats64 *myri10ge_get_stats(struct net_device *dev,
3123						    struct rtnl_link_stats64 *stats)
3124{
3125	const struct myri10ge_priv *mgp = netdev_priv(dev);
3126	const struct myri10ge_slice_netstats *slice_stats;
3127	int i;
3128
3129	for (i = 0; i < mgp->num_slices; i++) {
3130		slice_stats = &mgp->ss[i].stats;
3131		stats->rx_packets += slice_stats->rx_packets;
3132		stats->tx_packets += slice_stats->tx_packets;
3133		stats->rx_bytes += slice_stats->rx_bytes;
3134		stats->tx_bytes += slice_stats->tx_bytes;
3135		stats->rx_dropped += slice_stats->rx_dropped;
3136		stats->tx_dropped += slice_stats->tx_dropped;
3137	}
3138	return stats;
3139}
3140
3141static void myri10ge_set_multicast_list(struct net_device *dev)
3142{
3143	struct myri10ge_priv *mgp = netdev_priv(dev);
3144	struct myri10ge_cmd cmd;
3145	struct netdev_hw_addr *ha;
3146	__be32 data[2] = { 0, 0 };
3147	int err;
3148
3149	/* can be called from atomic contexts,
3150	 * pass 1 to force atomicity in myri10ge_send_cmd() */
3151	myri10ge_change_promisc(mgp, dev->flags & IFF_PROMISC, 1);
3152
3153	/* This firmware is known to not support multicast */
3154	if (!mgp->fw_multicast_support)
3155		return;
3156
3157	/* Disable multicast filtering */
3158
3159	err = myri10ge_send_cmd(mgp, MXGEFW_ENABLE_ALLMULTI, &cmd, 1);
3160	if (err != 0) {
3161		netdev_err(dev, "Failed MXGEFW_ENABLE_ALLMULTI, error status: %d\n",
3162			   err);
3163		goto abort;
3164	}
3165
3166	if ((dev->flags & IFF_ALLMULTI) || mgp->adopted_rx_filter_bug) {
3167		/* request to disable multicast filtering, so quit here */
3168		return;
3169	}
3170
3171	/* Flush the filters */
3172
3173	err = myri10ge_send_cmd(mgp, MXGEFW_LEAVE_ALL_MULTICAST_GROUPS,
3174				&cmd, 1);
3175	if (err != 0) {
3176		netdev_err(dev, "Failed MXGEFW_LEAVE_ALL_MULTICAST_GROUPS, error status: %d\n",
3177			   err);
3178		goto abort;
3179	}
3180
3181	/* Walk the multicast list, and add each address */
3182	netdev_for_each_mc_addr(ha, dev) {
3183		memcpy(data, &ha->addr, ETH_ALEN);
3184		cmd.data0 = ntohl(data[0]);
3185		cmd.data1 = ntohl(data[1]);
3186		err = myri10ge_send_cmd(mgp, MXGEFW_JOIN_MULTICAST_GROUP,
3187					&cmd, 1);
3188
3189		if (err != 0) {
3190			netdev_err(dev, "Failed MXGEFW_JOIN_MULTICAST_GROUP, error status:%d %pM\n",
3191				   err, ha->addr);
3192			goto abort;
3193		}
3194	}
3195	/* Enable multicast filtering */
3196	err = myri10ge_send_cmd(mgp, MXGEFW_DISABLE_ALLMULTI, &cmd, 1);
3197	if (err != 0) {
3198		netdev_err(dev, "Failed MXGEFW_DISABLE_ALLMULTI, error status: %d\n",
3199			   err);
3200		goto abort;
3201	}
3202
3203	return;
3204
3205abort:
3206	return;
3207}
3208
3209static int myri10ge_set_mac_address(struct net_device *dev, void *addr)
3210{
3211	struct sockaddr *sa = addr;
3212	struct myri10ge_priv *mgp = netdev_priv(dev);
3213	int status;
3214
3215	if (!is_valid_ether_addr(sa->sa_data))
3216		return -EADDRNOTAVAIL;
3217
3218	status = myri10ge_update_mac_address(mgp, sa->sa_data);
3219	if (status != 0) {
3220		netdev_err(dev, "changing mac address failed with %d\n",
3221			   status);
3222		return status;
3223	}
3224
3225	/* change the dev structure */
3226	memcpy(dev->dev_addr, sa->sa_data, ETH_ALEN);
3227	return 0;
3228}
3229
3230static int myri10ge_change_mtu(struct net_device *dev, int new_mtu)
3231{
3232	struct myri10ge_priv *mgp = netdev_priv(dev);
3233	int error = 0;
3234
3235	if ((new_mtu < 68) || (ETH_HLEN + new_mtu > MYRI10GE_MAX_ETHER_MTU)) {
3236		netdev_err(dev, "new mtu (%d) is not valid\n", new_mtu);
3237		return -EINVAL;
3238	}
3239	netdev_info(dev, "changing mtu from %d to %d\n", dev->mtu, new_mtu);
3240	if (mgp->running) {
3241		/* if we change the mtu on an active device, we must
3242		 * reset the device so the firmware sees the change */
3243		myri10ge_close(dev);
3244		dev->mtu = new_mtu;
3245		myri10ge_open(dev);
3246	} else
3247		dev->mtu = new_mtu;
3248
3249	return error;
3250}
3251
3252/*
3253 * Enable ECRC to align PCI-E Completion packets on an 8-byte boundary.
3254 * Only do it if the bridge is a root port since we don't want to disturb
3255 * any other device, except if forced with myri10ge_ecrc_enable > 1.
3256 */
3257
3258static void myri10ge_enable_ecrc(struct myri10ge_priv *mgp)
3259{
3260	struct pci_dev *bridge = mgp->pdev->bus->self;
3261	struct device *dev = &mgp->pdev->dev;
3262	int cap;
3263	unsigned err_cap;
3264	int ret;
3265
3266	if (!myri10ge_ecrc_enable || !bridge)
3267		return;
3268
3269	/* check that the bridge is a root port */
3270	if (pci_pcie_type(bridge) != PCI_EXP_TYPE_ROOT_PORT) {
3271		if (myri10ge_ecrc_enable > 1) {
3272			struct pci_dev *prev_bridge, *old_bridge = bridge;
3273
3274			/* Walk the hierarchy up to the root port
3275			 * where ECRC has to be enabled */
3276			do {
3277				prev_bridge = bridge;
3278				bridge = bridge->bus->self;
3279				if (!bridge || prev_bridge == bridge) {
3280					dev_err(dev,
3281						"Failed to find root port"
3282						" to force ECRC\n");
3283					return;
3284				}
3285			} while (pci_pcie_type(bridge) !=
3286				 PCI_EXP_TYPE_ROOT_PORT);
3287
3288			dev_info(dev,
3289				 "Forcing ECRC on non-root port %s"
3290				 " (enabling on root port %s)\n",
3291				 pci_name(old_bridge), pci_name(bridge));
3292		} else {
3293			dev_err(dev,
3294				"Not enabling ECRC on non-root port %s\n",
3295				pci_name(bridge));
3296			return;
3297		}
3298	}
3299
3300	cap = pci_find_ext_capability(bridge, PCI_EXT_CAP_ID_ERR);
3301	if (!cap)
3302		return;
3303
3304	ret = pci_read_config_dword(bridge, cap + PCI_ERR_CAP, &err_cap);
3305	if (ret) {
3306		dev_err(dev, "failed reading ext-conf-space of %s\n",
3307			pci_name(bridge));
3308		dev_err(dev, "\t pci=nommconf in use? "
3309			"or buggy/incomplete/absent ACPI MCFG attr?\n");
3310		return;
3311	}
3312	if (!(err_cap & PCI_ERR_CAP_ECRC_GENC))
3313		return;
3314
3315	err_cap |= PCI_ERR_CAP_ECRC_GENE;
3316	pci_write_config_dword(bridge, cap + PCI_ERR_CAP, err_cap);
3317	dev_info(dev, "Enabled ECRC on upstream bridge %s\n", pci_name(bridge));
3318}
3319
3320/*
3321 * The Lanai Z8E PCI-E interface achieves higher Read-DMA throughput
3322 * when the PCI-E Completion packets are aligned on an 8-byte
3323 * boundary.  Some PCI-E chip sets always align Completion packets; on
3324 * the ones that do not, the alignment can be enforced by enabling
3325 * ECRC generation (if supported).
3326 *
3327 * When PCI-E Completion packets are not aligned, it is actually more
3328 * efficient to limit Read-DMA transactions to 2KB, rather than 4KB.
3329 *
3330 * If the driver can neither enable ECRC nor verify that it has
3331 * already been enabled, then it must use a firmware image which works
3332 * around unaligned completion packets (myri10ge_rss_ethp_z8e.dat), and it
3333 * should also ensure that it never gives the device a Read-DMA which is
3334 * larger than 2KB by setting the tx_boundary to 2KB.  If ECRC is
3335 * enabled, then the driver should use the aligned (myri10ge_rss_eth_z8e.dat)
3336 * firmware image, and set tx_boundary to 4KB.
3337 */
3338
3339static void myri10ge_firmware_probe(struct myri10ge_priv *mgp)
3340{
3341	struct pci_dev *pdev = mgp->pdev;
3342	struct device *dev = &pdev->dev;
3343	int status;
3344
3345	mgp->tx_boundary = 4096;
3346	/*
3347	 * Verify the max read request size was set to 4KB
3348	 * before trying the test with 4KB.
3349	 */
3350	status = pcie_get_readrq(pdev);
3351	if (status < 0) {
3352		dev_err(dev, "Couldn't read max read req size: %d\n", status);
3353		goto abort;
3354	}
3355	if (status != 4096) {
3356		dev_warn(dev, "Max Read Request size != 4096 (%d)\n", status);
3357		mgp->tx_boundary = 2048;
3358	}
3359	/*
3360	 * load the optimized firmware (which assumes aligned PCIe
3361	 * completions) in order to see if it works on this host.
3362	 */
3363	set_fw_name(mgp, myri10ge_fw_aligned, false);
3364	status = myri10ge_load_firmware(mgp, 1);
3365	if (status != 0) {
3366		goto abort;
3367	}
3368
3369	/*
3370	 * Enable ECRC if possible
3371	 */
3372	myri10ge_enable_ecrc(mgp);
3373
3374	/*
3375	 * Run a DMA test which watches for unaligned completions and
3376	 * aborts on the first one seen.
3377	 */
3378
3379	status = myri10ge_dma_test(mgp, MXGEFW_CMD_UNALIGNED_TEST);
3380	if (status == 0)
3381		return;		/* keep the aligned firmware */
3382
3383	if (status != -E2BIG)
3384		dev_warn(dev, "DMA test failed: %d\n", status);
3385	if (status == -ENOSYS)
3386		dev_warn(dev, "Falling back to ethp! "
3387			 "Please install up to date fw\n");
3388abort:
3389	/* fall back to using the unaligned firmware */
3390	mgp->tx_boundary = 2048;
3391	set_fw_name(mgp, myri10ge_fw_unaligned, false);
3392}
3393
3394static void myri10ge_select_firmware(struct myri10ge_priv *mgp)
3395{
3396	int overridden = 0;
3397
3398	if (myri10ge_force_firmware == 0) {
3399		int link_width;
3400		u16 lnk;
3401
3402		pcie_capability_read_word(mgp->pdev, PCI_EXP_LNKSTA, &lnk);
3403		link_width = (lnk >> 4) & 0x3f;
3404
3405		/* Check to see if Link is less than 8 or if the
3406		 * upstream bridge is known to provide aligned
3407		 * completions */
3408		if (link_width < 8) {
3409			dev_info(&mgp->pdev->dev, "PCIE x%d Link\n",
3410				 link_width);
3411			mgp->tx_boundary = 4096;
3412			set_fw_name(mgp, myri10ge_fw_aligned, false);
3413		} else {
3414			myri10ge_firmware_probe(mgp);
3415		}
3416	} else {
3417		if (myri10ge_force_firmware == 1) {
3418			dev_info(&mgp->pdev->dev,
3419				 "Assuming aligned completions (forced)\n");
3420			mgp->tx_boundary = 4096;
3421			set_fw_name(mgp, myri10ge_fw_aligned, false);
3422		} else {
3423			dev_info(&mgp->pdev->dev,
3424				 "Assuming unaligned completions (forced)\n");
3425			mgp->tx_boundary = 2048;
3426			set_fw_name(mgp, myri10ge_fw_unaligned, false);
3427		}
3428	}
3429
3430	kparam_block_sysfs_write(myri10ge_fw_name);
3431	if (myri10ge_fw_name != NULL) {
3432		char *fw_name = kstrdup(myri10ge_fw_name, GFP_KERNEL);
3433		if (fw_name) {
3434			overridden = 1;
3435			set_fw_name(mgp, fw_name, true);
3436		}
3437	}
3438	kparam_unblock_sysfs_write(myri10ge_fw_name);
3439
3440	if (mgp->board_number < MYRI10GE_MAX_BOARDS &&
3441	    myri10ge_fw_names[mgp->board_number] != NULL &&
3442	    strlen(myri10ge_fw_names[mgp->board_number])) {
3443		set_fw_name(mgp, myri10ge_fw_names[mgp->board_number], false);
3444		overridden = 1;
3445	}
3446	if (overridden)
3447		dev_info(&mgp->pdev->dev, "overriding firmware to %s\n",
3448			 mgp->fw_name);
3449}
3450
3451static void myri10ge_mask_surprise_down(struct pci_dev *pdev)
3452{
3453	struct pci_dev *bridge = pdev->bus->self;
3454	int cap;
3455	u32 mask;
3456
3457	if (bridge == NULL)
3458		return;
3459
3460	cap = pci_find_ext_capability(bridge, PCI_EXT_CAP_ID_ERR);
3461	if (cap) {
3462		/* a sram parity error can cause a surprise link
3463		 * down; since we expect and can recover from sram
3464		 * parity errors, mask surprise link down events */
3465		pci_read_config_dword(bridge, cap + PCI_ERR_UNCOR_MASK, &mask);
3466		mask |= 0x20;
3467		pci_write_config_dword(bridge, cap + PCI_ERR_UNCOR_MASK, mask);
3468	}
3469}
3470
3471#ifdef CONFIG_PM
3472static int myri10ge_suspend(struct pci_dev *pdev, pm_message_t state)
3473{
3474	struct myri10ge_priv *mgp;
3475	struct net_device *netdev;
3476
3477	mgp = pci_get_drvdata(pdev);
3478	if (mgp == NULL)
3479		return -EINVAL;
3480	netdev = mgp->dev;
3481
3482	netif_device_detach(netdev);
3483	if (netif_running(netdev)) {
3484		netdev_info(netdev, "closing\n");
3485		rtnl_lock();
3486		myri10ge_close(netdev);
3487		rtnl_unlock();
3488	}
3489	myri10ge_dummy_rdma(mgp, 0);
3490	pci_save_state(pdev);
3491	pci_disable_device(pdev);
3492
3493	return pci_set_power_state(pdev, pci_choose_state(pdev, state));
3494}
3495
3496static int myri10ge_resume(struct pci_dev *pdev)
3497{
3498	struct myri10ge_priv *mgp;
3499	struct net_device *netdev;
3500	int status;
3501	u16 vendor;
3502
3503	mgp = pci_get_drvdata(pdev);
3504	if (mgp == NULL)
3505		return -EINVAL;
3506	netdev = mgp->dev;
3507	pci_set_power_state(pdev, PCI_D0);	/* zeros conf space as a side effect */
3508	msleep(5);		/* give card time to respond */
3509	pci_read_config_word(mgp->pdev, PCI_VENDOR_ID, &vendor);
3510	if (vendor == 0xffff) {
3511		netdev_err(mgp->dev, "device disappeared!\n");
3512		return -EIO;
3513	}
3514
3515	pci_restore_state(pdev);
3516
3517	status = pci_enable_device(pdev);
3518	if (status) {
3519		dev_err(&pdev->dev, "failed to enable device\n");
3520		return status;
3521	}
3522
3523	pci_set_master(pdev);
3524
3525	myri10ge_reset(mgp);
3526	myri10ge_dummy_rdma(mgp, 1);
3527
3528	/* Save configuration space to be restored if the
3529	 * nic resets due to a parity error */
3530	pci_save_state(pdev);
3531
3532	if (netif_running(netdev)) {
3533		rtnl_lock();
3534		status = myri10ge_open(netdev);
3535		rtnl_unlock();
3536		if (status != 0)
3537			goto abort_with_enabled;
3538
3539	}
3540	netif_device_attach(netdev);
3541
3542	return 0;
3543
3544abort_with_enabled:
3545	pci_disable_device(pdev);
3546	return -EIO;
3547
3548}
3549#endif				/* CONFIG_PM */
3550
3551static u32 myri10ge_read_reboot(struct myri10ge_priv *mgp)
3552{
3553	struct pci_dev *pdev = mgp->pdev;
3554	int vs = mgp->vendor_specific_offset;
3555	u32 reboot;
3556
3557	/*enter read32 mode */
3558	pci_write_config_byte(pdev, vs + 0x10, 0x3);
3559
3560	/*read REBOOT_STATUS (0xfffffff0) */
3561	pci_write_config_dword(pdev, vs + 0x18, 0xfffffff0);
3562	pci_read_config_dword(pdev, vs + 0x14, &reboot);
3563	return reboot;
3564}
3565
3566static void
3567myri10ge_check_slice(struct myri10ge_slice_state *ss, int *reset_needed,
3568		     int *busy_slice_cnt, u32 rx_pause_cnt)
3569{
3570	struct myri10ge_priv *mgp = ss->mgp;
3571	int slice = ss - mgp->ss;
3572
3573	if (ss->tx.req != ss->tx.done &&
3574	    ss->tx.done == ss->watchdog_tx_done &&
3575	    ss->watchdog_tx_req != ss->watchdog_tx_done) {
3576		/* nic seems like it might be stuck.. */
3577		if (rx_pause_cnt != mgp->watchdog_pause) {
3578			if (net_ratelimit())
3579				netdev_warn(mgp->dev, "slice %d: TX paused, "
3580					    "check link partner\n", slice);
3581		} else {
3582			netdev_warn(mgp->dev,
3583				    "slice %d: TX stuck %d %d %d %d %d %d\n",
3584				    slice, ss->tx.queue_active, ss->tx.req,
3585				    ss->tx.done, ss->tx.pkt_start,
3586				    ss->tx.pkt_done,
3587				    (int)ntohl(mgp->ss[slice].fw_stats->
3588					       send_done_count));
3589			*reset_needed = 1;
3590			ss->stuck = 1;
3591		}
3592	}
3593	if (ss->watchdog_tx_done != ss->tx.done ||
3594	    ss->watchdog_rx_done != ss->rx_done.cnt) {
3595		*busy_slice_cnt += 1;
3596	}
3597	ss->watchdog_tx_done = ss->tx.done;
3598	ss->watchdog_tx_req = ss->tx.req;
3599	ss->watchdog_rx_done = ss->rx_done.cnt;
3600}
3601
3602/*
3603 * This watchdog is used to check whether the board has suffered
3604 * from a parity error and needs to be recovered.
3605 */
3606static void myri10ge_watchdog(struct work_struct *work)
3607{
3608	struct myri10ge_priv *mgp =
3609	    container_of(work, struct myri10ge_priv, watchdog_work);
3610	struct myri10ge_slice_state *ss;
3611	u32 reboot, rx_pause_cnt;
3612	int status, rebooted;
3613	int i;
3614	int reset_needed = 0;
3615	int busy_slice_cnt = 0;
3616	u16 cmd, vendor;
3617
3618	mgp->watchdog_resets++;
3619	pci_read_config_word(mgp->pdev, PCI_COMMAND, &cmd);
3620	rebooted = 0;
3621	if ((cmd & PCI_COMMAND_MASTER) == 0) {
3622		/* Bus master DMA disabled?  Check to see
3623		 * if the card rebooted due to a parity error
3624		 * For now, just report it */
3625		reboot = myri10ge_read_reboot(mgp);
3626		netdev_err(mgp->dev, "NIC rebooted (0x%x),%s resetting\n",
3627			   reboot, myri10ge_reset_recover ? "" : " not");
3628		if (myri10ge_reset_recover == 0)
3629			return;
3630		rtnl_lock();
3631		mgp->rebooted = 1;
3632		rebooted = 1;
3633		myri10ge_close(mgp->dev);
3634		myri10ge_reset_recover--;
3635		mgp->rebooted = 0;
3636		/*
3637		 * A rebooted nic will come back with config space as
3638		 * it was after power was applied to PCIe bus.
3639		 * Attempt to restore config space which was saved
3640		 * when the driver was loaded, or the last time the
3641		 * nic was resumed from power saving mode.
3642		 */
3643		pci_restore_state(mgp->pdev);
3644
3645		/* save state again for accounting reasons */
3646		pci_save_state(mgp->pdev);
3647
3648	} else {
3649		/* if we get back -1's from our slot, perhaps somebody
3650		 * powered off our card.  Don't try to reset it in
3651		 * this case */
3652		if (cmd == 0xffff) {
3653			pci_read_config_word(mgp->pdev, PCI_VENDOR_ID, &vendor);
3654			if (vendor == 0xffff) {
3655				netdev_err(mgp->dev, "device disappeared!\n");
3656				return;
3657			}
3658		}
3659		/* Perhaps it is a software error. See if stuck slice
3660		 * has recovered, reset if not */
3661		rx_pause_cnt = ntohl(mgp->ss[0].fw_stats->dropped_pause);
3662		for (i = 0; i < mgp->num_slices; i++) {
3663			ss = mgp->ss;
3664			if (ss->stuck) {
3665				myri10ge_check_slice(ss, &reset_needed,
3666						     &busy_slice_cnt,
3667						     rx_pause_cnt);
3668				ss->stuck = 0;
3669			}
3670		}
3671		if (!reset_needed) {
3672			netdev_dbg(mgp->dev, "not resetting\n");
3673			return;
3674		}
3675
3676		netdev_err(mgp->dev, "device timeout, resetting\n");
3677	}
3678
3679	if (!rebooted) {
3680		rtnl_lock();
3681		myri10ge_close(mgp->dev);
3682	}
3683	status = myri10ge_load_firmware(mgp, 1);
3684	if (status != 0)
3685		netdev_err(mgp->dev, "failed to load firmware\n");
3686	else
3687		myri10ge_open(mgp->dev);
3688	rtnl_unlock();
3689}
3690
3691/*
3692 * We use our own timer routine rather than relying upon
3693 * netdev->tx_timeout because we have a very large hardware transmit
3694 * queue.  Due to the large queue, the netdev->tx_timeout function
3695 * cannot detect a NIC with a parity error in a timely fashion if the
3696 * NIC is lightly loaded.
3697 */
3698static void myri10ge_watchdog_timer(unsigned long arg)
3699{
3700	struct myri10ge_priv *mgp;
3701	struct myri10ge_slice_state *ss;
3702	int i, reset_needed, busy_slice_cnt;
3703	u32 rx_pause_cnt;
3704	u16 cmd;
3705
3706	mgp = (struct myri10ge_priv *)arg;
3707
3708	rx_pause_cnt = ntohl(mgp->ss[0].fw_stats->dropped_pause);
3709	busy_slice_cnt = 0;
3710	for (i = 0, reset_needed = 0;
3711	     i < mgp->num_slices && reset_needed == 0; ++i) {
3712
3713		ss = &mgp->ss[i];
3714		if (ss->rx_small.watchdog_needed) {
3715			myri10ge_alloc_rx_pages(mgp, &ss->rx_small,
3716						mgp->small_bytes + MXGEFW_PAD,
3717						1);
3718			if (ss->rx_small.fill_cnt - ss->rx_small.cnt >=
3719			    myri10ge_fill_thresh)
3720				ss->rx_small.watchdog_needed = 0;
3721		}
3722		if (ss->rx_big.watchdog_needed) {
3723			myri10ge_alloc_rx_pages(mgp, &ss->rx_big,
3724						mgp->big_bytes, 1);
3725			if (ss->rx_big.fill_cnt - ss->rx_big.cnt >=
3726			    myri10ge_fill_thresh)
3727				ss->rx_big.watchdog_needed = 0;
3728		}
3729		myri10ge_check_slice(ss, &reset_needed, &busy_slice_cnt,
3730				     rx_pause_cnt);
3731	}
3732	/* if we've sent or received no traffic, poll the NIC to
3733	 * ensure it is still there.  Otherwise, we risk not noticing
3734	 * an error in a timely fashion */
3735	if (busy_slice_cnt == 0) {
3736		pci_read_config_word(mgp->pdev, PCI_COMMAND, &cmd);
3737		if ((cmd & PCI_COMMAND_MASTER) == 0) {
3738			reset_needed = 1;
3739		}
3740	}
3741	mgp->watchdog_pause = rx_pause_cnt;
3742
3743	if (reset_needed) {
3744		schedule_work(&mgp->watchdog_work);
3745	} else {
3746		/* rearm timer */
3747		mod_timer(&mgp->watchdog_timer,
3748			  jiffies + myri10ge_watchdog_timeout * HZ);
3749	}
3750}
3751
3752static void myri10ge_free_slices(struct myri10ge_priv *mgp)
3753{
3754	struct myri10ge_slice_state *ss;
3755	struct pci_dev *pdev = mgp->pdev;
3756	size_t bytes;
3757	int i;
3758
3759	if (mgp->ss == NULL)
3760		return;
3761
3762	for (i = 0; i < mgp->num_slices; i++) {
3763		ss = &mgp->ss[i];
3764		if (ss->rx_done.entry != NULL) {
3765			bytes = mgp->max_intr_slots *
3766			    sizeof(*ss->rx_done.entry);
3767			dma_free_coherent(&pdev->dev, bytes,
3768					  ss->rx_done.entry, ss->rx_done.bus);
3769			ss->rx_done.entry = NULL;
3770		}
3771		if (ss->fw_stats != NULL) {
3772			bytes = sizeof(*ss->fw_stats);
3773			dma_free_coherent(&pdev->dev, bytes,
3774					  ss->fw_stats, ss->fw_stats_bus);
3775			ss->fw_stats = NULL;
3776		}
3777		napi_hash_del(&ss->napi);
3778		netif_napi_del(&ss->napi);
3779	}
3780	/* Wait till napi structs are no longer used, and then free ss. */
3781	synchronize_rcu();
3782	kfree(mgp->ss);
3783	mgp->ss = NULL;
3784}
3785
3786static int myri10ge_alloc_slices(struct myri10ge_priv *mgp)
3787{
3788	struct myri10ge_slice_state *ss;
3789	struct pci_dev *pdev = mgp->pdev;
3790	size_t bytes;
3791	int i;
3792
3793	bytes = sizeof(*mgp->ss) * mgp->num_slices;
3794	mgp->ss = kzalloc(bytes, GFP_KERNEL);
3795	if (mgp->ss == NULL) {
3796		return -ENOMEM;
3797	}
3798
3799	for (i = 0; i < mgp->num_slices; i++) {
3800		ss = &mgp->ss[i];
3801		bytes = mgp->max_intr_slots * sizeof(*ss->rx_done.entry);
3802		ss->rx_done.entry = dma_zalloc_coherent(&pdev->dev, bytes,
3803							&ss->rx_done.bus,
3804							GFP_KERNEL);
3805		if (ss->rx_done.entry == NULL)
3806			goto abort;
3807		bytes = sizeof(*ss->fw_stats);
3808		ss->fw_stats = dma_alloc_coherent(&pdev->dev, bytes,
3809						  &ss->fw_stats_bus,
3810						  GFP_KERNEL);
3811		if (ss->fw_stats == NULL)
3812			goto abort;
3813		ss->mgp = mgp;
3814		ss->dev = mgp->dev;
3815		netif_napi_add(ss->dev, &ss->napi, myri10ge_poll,
3816			       myri10ge_napi_weight);
3817		napi_hash_add(&ss->napi);
3818	}
3819	return 0;
3820abort:
3821	myri10ge_free_slices(mgp);
3822	return -ENOMEM;
3823}
3824
3825/*
3826 * This function determines the number of slices supported.
3827 * The number slices is the minimum of the number of CPUS,
3828 * the number of MSI-X irqs supported, the number of slices
3829 * supported by the firmware
3830 */
3831static void myri10ge_probe_slices(struct myri10ge_priv *mgp)
3832{
3833	struct myri10ge_cmd cmd;
3834	struct pci_dev *pdev = mgp->pdev;
3835	char *old_fw;
3836	bool old_allocated;
3837	int i, status, ncpus;
3838
3839	mgp->num_slices = 1;
3840	ncpus = netif_get_num_default_rss_queues();
3841
3842	if (myri10ge_max_slices == 1 || !pdev->msix_cap ||
3843	    (myri10ge_max_slices == -1 && ncpus < 2))
3844		return;
3845
3846	/* try to load the slice aware rss firmware */
3847	old_fw = mgp->fw_name;
3848	old_allocated = mgp->fw_name_allocated;
3849	/* don't free old_fw if we override it. */
3850	mgp->fw_name_allocated = false;
3851
3852	if (myri10ge_fw_name != NULL) {
3853		dev_info(&mgp->pdev->dev, "overriding rss firmware to %s\n",
3854			 myri10ge_fw_name);
3855		set_fw_name(mgp, myri10ge_fw_name, false);
3856	} else if (old_fw == myri10ge_fw_aligned)
3857		set_fw_name(mgp, myri10ge_fw_rss_aligned, false);
3858	else
3859		set_fw_name(mgp, myri10ge_fw_rss_unaligned, false);
3860	status = myri10ge_load_firmware(mgp, 0);
3861	if (status != 0) {
3862		dev_info(&pdev->dev, "Rss firmware not found\n");
3863		if (old_allocated)
3864			kfree(old_fw);
3865		return;
3866	}
3867
3868	/* hit the board with a reset to ensure it is alive */
3869	memset(&cmd, 0, sizeof(cmd));
3870	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_RESET, &cmd, 0);
3871	if (status != 0) {
3872		dev_err(&mgp->pdev->dev, "failed reset\n");
3873		goto abort_with_fw;
3874	}
3875
3876	mgp->max_intr_slots = cmd.data0 / sizeof(struct mcp_slot);
3877
3878	/* tell it the size of the interrupt queues */
3879	cmd.data0 = mgp->max_intr_slots * sizeof(struct mcp_slot);
3880	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_SET_INTRQ_SIZE, &cmd, 0);
3881	if (status != 0) {
3882		dev_err(&mgp->pdev->dev, "failed MXGEFW_CMD_SET_INTRQ_SIZE\n");
3883		goto abort_with_fw;
3884	}
3885
3886	/* ask the maximum number of slices it supports */
3887	status = myri10ge_send_cmd(mgp, MXGEFW_CMD_GET_MAX_RSS_QUEUES, &cmd, 0);
3888	if (status != 0)
3889		goto abort_with_fw;
3890	else
3891		mgp->num_slices = cmd.data0;
3892
3893	/* Only allow multiple slices if MSI-X is usable */
3894	if (!myri10ge_msi) {
3895		goto abort_with_fw;
3896	}
3897
3898	/* if the admin did not specify a limit to how many
3899	 * slices we should use, cap it automatically to the
3900	 * number of CPUs currently online */
3901	if (myri10ge_max_slices == -1)
3902		myri10ge_max_slices = ncpus;
3903
3904	if (mgp->num_slices > myri10ge_max_slices)
3905		mgp->num_slices = myri10ge_max_slices;
3906
3907	/* Now try to allocate as many MSI-X vectors as we have
3908	 * slices. We give up on MSI-X if we can only get a single
3909	 * vector. */
3910
3911	mgp->msix_vectors = kcalloc(mgp->num_slices, sizeof(*mgp->msix_vectors),
3912				    GFP_KERNEL);
3913	if (mgp->msix_vectors == NULL)
3914		goto no_msix;
3915	for (i = 0; i < mgp->num_slices; i++) {
3916		mgp->msix_vectors[i].entry = i;
3917	}
3918
3919	while (mgp->num_slices > 1) {
3920		mgp->num_slices = rounddown_pow_of_two(mgp->num_slices);
3921		if (mgp->num_slices == 1)
3922			goto no_msix;
3923		status = pci_enable_msix_range(pdev,
3924					       mgp->msix_vectors,
3925					       mgp->num_slices,
3926					       mgp->num_slices);
3927		if (status < 0)
3928			goto no_msix;
3929
3930		pci_disable_msix(pdev);
3931
3932		if (status == mgp->num_slices) {
3933			if (old_allocated)
3934				kfree(old_fw);
3935			return;
3936		} else {
3937			mgp->num_slices = status;
3938		}
3939	}
3940
3941no_msix:
3942	if (mgp->msix_vectors != NULL) {
3943		kfree(mgp->msix_vectors);
3944		mgp->msix_vectors = NULL;
3945	}
3946
3947abort_with_fw:
3948	mgp->num_slices = 1;
3949	set_fw_name(mgp, old_fw, old_allocated);
3950	myri10ge_load_firmware(mgp, 0);
3951}
3952
3953static const struct net_device_ops myri10ge_netdev_ops = {
3954	.ndo_open		= myri10ge_open,
3955	.ndo_stop		= myri10ge_close,
3956	.ndo_start_xmit		= myri10ge_xmit,
3957	.ndo_get_stats64	= myri10ge_get_stats,
3958	.ndo_validate_addr	= eth_validate_addr,
3959	.ndo_change_mtu		= myri10ge_change_mtu,
3960	.ndo_set_rx_mode	= myri10ge_set_multicast_list,
3961	.ndo_set_mac_address	= myri10ge_set_mac_address,
3962#ifdef CONFIG_NET_RX_BUSY_POLL
3963	.ndo_busy_poll		= myri10ge_busy_poll,
3964#endif
3965};
3966
3967static int myri10ge_probe(struct pci_dev *pdev, const struct pci_device_id *ent)
3968{
3969	struct net_device *netdev;
3970	struct myri10ge_priv *mgp;
3971	struct device *dev = &pdev->dev;
3972	int i;
3973	int status = -ENXIO;
3974	int dac_enabled;
3975	unsigned hdr_offset, ss_offset;
3976	static int board_number;
3977
3978	netdev = alloc_etherdev_mq(sizeof(*mgp), MYRI10GE_MAX_SLICES);
3979	if (netdev == NULL)
3980		return -ENOMEM;
3981
3982	SET_NETDEV_DEV(netdev, &pdev->dev);
3983
3984	mgp = netdev_priv(netdev);
3985	mgp->dev = netdev;
3986	mgp->pdev = pdev;
3987	mgp->pause = myri10ge_flow_control;
3988	mgp->intr_coal_delay = myri10ge_intr_coal_delay;
3989	mgp->msg_enable = netif_msg_init(myri10ge_debug, MYRI10GE_MSG_DEFAULT);
3990	mgp->board_number = board_number;
3991	init_waitqueue_head(&mgp->down_wq);
3992
3993	if (pci_enable_device(pdev)) {
3994		dev_err(&pdev->dev, "pci_enable_device call failed\n");
3995		status = -ENODEV;
3996		goto abort_with_netdev;
3997	}
3998
3999	/* Find the vendor-specific cap so we can check
4000	 * the reboot register later on */
4001	mgp->vendor_specific_offset
4002	    = pci_find_capability(pdev, PCI_CAP_ID_VNDR);
4003
4004	/* Set our max read request to 4KB */
4005	status = pcie_set_readrq(pdev, 4096);
4006	if (status != 0) {
4007		dev_err(&pdev->dev, "Error %d writing PCI_EXP_DEVCTL\n",
4008			status);
4009		goto abort_with_enabled;
4010	}
4011
4012	myri10ge_mask_surprise_down(pdev);
4013	pci_set_master(pdev);
4014	dac_enabled = 1;
4015	status = pci_set_dma_mask(pdev, DMA_BIT_MASK(64));
4016	if (status != 0) {
4017		dac_enabled = 0;
4018		dev_err(&pdev->dev,
4019			"64-bit pci address mask was refused, "
4020			"trying 32-bit\n");
4021		status = pci_set_dma_mask(pdev, DMA_BIT_MASK(32));
4022	}
4023	if (status != 0) {
4024		dev_err(&pdev->dev, "Error %d setting DMA mask\n", status);
4025		goto abort_with_enabled;
4026	}
4027	(void)pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64));
4028	mgp->cmd = dma_alloc_coherent(&pdev->dev, sizeof(*mgp->cmd),
4029				      &mgp->cmd_bus, GFP_KERNEL);
4030	if (!mgp->cmd) {
4031		status = -ENOMEM;
4032		goto abort_with_enabled;
4033	}
4034
4035	mgp->board_span = pci_resource_len(pdev, 0);
4036	mgp->iomem_base = pci_resource_start(pdev, 0);
4037	mgp->wc_cookie = arch_phys_wc_add(mgp->iomem_base, mgp->board_span);
4038	mgp->sram = ioremap_wc(mgp->iomem_base, mgp->board_span);
4039	if (mgp->sram == NULL) {
4040		dev_err(&pdev->dev, "ioremap failed for %ld bytes at 0x%lx\n",
4041			mgp->board_span, mgp->iomem_base);
4042		status = -ENXIO;
4043		goto abort_with_mtrr;
4044	}
4045	hdr_offset =
4046	    swab32(readl(mgp->sram + MCP_HEADER_PTR_OFFSET)) & 0xffffc;
4047	ss_offset = hdr_offset + offsetof(struct mcp_gen_header, string_specs);
4048	mgp->sram_size = swab32(readl(mgp->sram + ss_offset));
4049	if (mgp->sram_size > mgp->board_span ||
4050	    mgp->sram_size <= MYRI10GE_FW_OFFSET) {
4051		dev_err(&pdev->dev,
4052			"invalid sram_size %dB or board span %ldB\n",
4053			mgp->sram_size, mgp->board_span);
4054		goto abort_with_ioremap;
4055	}
4056	memcpy_fromio(mgp->eeprom_strings,
4057		      mgp->sram + mgp->sram_size, MYRI10GE_EEPROM_STRINGS_SIZE);
4058	memset(mgp->eeprom_strings + MYRI10GE_EEPROM_STRINGS_SIZE - 2, 0, 2);
4059	status = myri10ge_read_mac_addr(mgp);
4060	if (status)
4061		goto abort_with_ioremap;
4062
4063	for (i = 0; i < ETH_ALEN; i++)
4064		netdev->dev_addr[i] = mgp->mac_addr[i];
4065
4066	myri10ge_select_firmware(mgp);
4067
4068	status = myri10ge_load_firmware(mgp, 1);
4069	if (status != 0) {
4070		dev_err(&pdev->dev, "failed to load firmware\n");
4071		goto abort_with_ioremap;
4072	}
4073	myri10ge_probe_slices(mgp);
4074	status = myri10ge_alloc_slices(mgp);
4075	if (status != 0) {
4076		dev_err(&pdev->dev, "failed to alloc slice state\n");
4077		goto abort_with_firmware;
4078	}
4079	netif_set_real_num_tx_queues(netdev, mgp->num_slices);
4080	netif_set_real_num_rx_queues(netdev, mgp->num_slices);
4081	status = myri10ge_reset(mgp);
4082	if (status != 0) {
4083		dev_err(&pdev->dev, "failed reset\n");
4084		goto abort_with_slices;
4085	}
4086#ifdef CONFIG_MYRI10GE_DCA
4087	myri10ge_setup_dca(mgp);
4088#endif
4089	pci_set_drvdata(pdev, mgp);
4090	if ((myri10ge_initial_mtu + ETH_HLEN) > MYRI10GE_MAX_ETHER_MTU)
4091		myri10ge_initial_mtu = MYRI10GE_MAX_ETHER_MTU - ETH_HLEN;
4092	if ((myri10ge_initial_mtu + ETH_HLEN) < 68)
4093		myri10ge_initial_mtu = 68;
4094
4095	netdev->netdev_ops = &myri10ge_netdev_ops;
4096	netdev->mtu = myri10ge_initial_mtu;
4097	netdev->hw_features = mgp->features | NETIF_F_RXCSUM;
4098
4099	/* fake NETIF_F_HW_VLAN_CTAG_RX for good GRO performance */
4100	netdev->hw_features |= NETIF_F_HW_VLAN_CTAG_RX;
4101
4102	netdev->features = netdev->hw_features;
4103
4104	if (dac_enabled)
4105		netdev->features |= NETIF_F_HIGHDMA;
4106
4107	netdev->vlan_features |= mgp->features;
4108	if (mgp->fw_ver_tiny < 37)
4109		netdev->vlan_features &= ~NETIF_F_TSO6;
4110	if (mgp->fw_ver_tiny < 32)
4111		netdev->vlan_features &= ~NETIF_F_TSO;
4112
4113	/* make sure we can get an irq, and that MSI can be
4114	 * setup (if available). */
4115	status = myri10ge_request_irq(mgp);
4116	if (status != 0)
4117		goto abort_with_firmware;
4118	myri10ge_free_irq(mgp);
4119
4120	/* Save configuration space to be restored if the
4121	 * nic resets due to a parity error */
4122	pci_save_state(pdev);
4123
4124	/* Setup the watchdog timer */
4125	setup_timer(&mgp->watchdog_timer, myri10ge_watchdog_timer,
4126		    (unsigned long)mgp);
4127
4128	netdev->ethtool_ops = &myri10ge_ethtool_ops;
4129	INIT_WORK(&mgp->watchdog_work, myri10ge_watchdog);
4130	status = register_netdev(netdev);
4131	if (status != 0) {
4132		dev_err(&pdev->dev, "register_netdev failed: %d\n", status);
4133		goto abort_with_state;
4134	}
4135	if (mgp->msix_enabled)
4136		dev_info(dev, "%d MSI-X IRQs, tx bndry %d, fw %s, MTRR %s, WC Enabled\n",
4137			 mgp->num_slices, mgp->tx_boundary, mgp->fw_name,
4138			 (mgp->wc_cookie > 0 ? "Enabled" : "Disabled"));
4139	else
4140		dev_info(dev, "%s IRQ %d, tx bndry %d, fw %s, MTRR %s, WC Enabled\n",
4141			 mgp->msi_enabled ? "MSI" : "xPIC",
4142			 pdev->irq, mgp->tx_boundary, mgp->fw_name,
4143			 (mgp->wc_cookie > 0 ? "Enabled" : "Disabled"));
4144
4145	board_number++;
4146	return 0;
4147
4148abort_with_state:
4149	pci_restore_state(pdev);
4150
4151abort_with_slices:
4152	myri10ge_free_slices(mgp);
4153
4154abort_with_firmware:
4155	myri10ge_dummy_rdma(mgp, 0);
4156
4157abort_with_ioremap:
4158	if (mgp->mac_addr_string != NULL)
4159		dev_err(&pdev->dev,
4160			"myri10ge_probe() failed: MAC=%s, SN=%ld\n",
4161			mgp->mac_addr_string, mgp->serial_number);
4162	iounmap(mgp->sram);
4163
4164abort_with_mtrr:
4165	arch_phys_wc_del(mgp->wc_cookie);
4166	dma_free_coherent(&pdev->dev, sizeof(*mgp->cmd),
4167			  mgp->cmd, mgp->cmd_bus);
4168
4169abort_with_enabled:
4170	pci_disable_device(pdev);
4171
4172abort_with_netdev:
4173	set_fw_name(mgp, NULL, false);
4174	free_netdev(netdev);
4175	return status;
4176}
4177
4178/*
4179 * myri10ge_remove
4180 *
4181 * Does what is necessary to shutdown one Myrinet device. Called
4182 *   once for each Myrinet card by the kernel when a module is
4183 *   unloaded.
4184 */
4185static void myri10ge_remove(struct pci_dev *pdev)
4186{
4187	struct myri10ge_priv *mgp;
4188	struct net_device *netdev;
4189
4190	mgp = pci_get_drvdata(pdev);
4191	if (mgp == NULL)
4192		return;
4193
4194	cancel_work_sync(&mgp->watchdog_work);
4195	netdev = mgp->dev;
4196	unregister_netdev(netdev);
4197
4198#ifdef CONFIG_MYRI10GE_DCA
4199	myri10ge_teardown_dca(mgp);
4200#endif
4201	myri10ge_dummy_rdma(mgp, 0);
4202
4203	/* avoid a memory leak */
4204	pci_restore_state(pdev);
4205
4206	iounmap(mgp->sram);
4207	arch_phys_wc_del(mgp->wc_cookie);
4208	myri10ge_free_slices(mgp);
4209	kfree(mgp->msix_vectors);
4210	dma_free_coherent(&pdev->dev, sizeof(*mgp->cmd),
4211			  mgp->cmd, mgp->cmd_bus);
4212
4213	set_fw_name(mgp, NULL, false);
4214	free_netdev(netdev);
4215	pci_disable_device(pdev);
4216}
4217
4218#define PCI_DEVICE_ID_MYRICOM_MYRI10GE_Z8E 	0x0008
4219#define PCI_DEVICE_ID_MYRICOM_MYRI10GE_Z8E_9	0x0009
4220
4221static const struct pci_device_id myri10ge_pci_tbl[] = {
4222	{PCI_DEVICE(PCI_VENDOR_ID_MYRICOM, PCI_DEVICE_ID_MYRICOM_MYRI10GE_Z8E)},
4223	{PCI_DEVICE
4224	 (PCI_VENDOR_ID_MYRICOM, PCI_DEVICE_ID_MYRICOM_MYRI10GE_Z8E_9)},
4225	{0},
4226};
4227
4228MODULE_DEVICE_TABLE(pci, myri10ge_pci_tbl);
4229
4230static struct pci_driver myri10ge_driver = {
4231	.name = "myri10ge",
4232	.probe = myri10ge_probe,
4233	.remove = myri10ge_remove,
4234	.id_table = myri10ge_pci_tbl,
4235#ifdef CONFIG_PM
4236	.suspend = myri10ge_suspend,
4237	.resume = myri10ge_resume,
4238#endif
4239};
4240
4241#ifdef CONFIG_MYRI10GE_DCA
4242static int
4243myri10ge_notify_dca(struct notifier_block *nb, unsigned long event, void *p)
4244{
4245	int err = driver_for_each_device(&myri10ge_driver.driver,
4246					 NULL, &event,
4247					 myri10ge_notify_dca_device);
4248
4249	if (err)
4250		return NOTIFY_BAD;
4251	return NOTIFY_DONE;
4252}
4253
4254static struct notifier_block myri10ge_dca_notifier = {
4255	.notifier_call = myri10ge_notify_dca,
4256	.next = NULL,
4257	.priority = 0,
4258};
4259#endif				/* CONFIG_MYRI10GE_DCA */
4260
4261static __init int myri10ge_init_module(void)
4262{
4263	pr_info("Version %s\n", MYRI10GE_VERSION_STR);
4264
4265	if (myri10ge_rss_hash > MXGEFW_RSS_HASH_TYPE_MAX) {
4266		pr_err("Illegal rssh hash type %d, defaulting to source port\n",
4267		       myri10ge_rss_hash);
4268		myri10ge_rss_hash = MXGEFW_RSS_HASH_TYPE_SRC_PORT;
4269	}
4270#ifdef CONFIG_MYRI10GE_DCA
4271	dca_register_notify(&myri10ge_dca_notifier);
4272#endif
4273	if (myri10ge_max_slices > MYRI10GE_MAX_SLICES)
4274		myri10ge_max_slices = MYRI10GE_MAX_SLICES;
4275
4276	return pci_register_driver(&myri10ge_driver);
4277}
4278
4279module_init(myri10ge_init_module);
4280
4281static __exit void myri10ge_cleanup_module(void)
4282{
4283#ifdef CONFIG_MYRI10GE_DCA
4284	dca_unregister_notify(&myri10ge_dca_notifier);
4285#endif
4286	pci_unregister_driver(&myri10ge_driver);
4287}
4288
4289module_exit(myri10ge_cleanup_module);
4290