2 * Copyright (c) 2004, 2005 Topspin Communications. All rights reserved.
3 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
4 * Copyright (c) 2005, 2006, 2007, 2008 Mellanox Technologies. All rights reserved.
5 * Copyright (c) 2006, 2007 Cisco Systems, Inc. All rights reserved.
7 * This software is available to you under a choice of one of two
8 * licenses. You may choose to be licensed under the terms of the GNU
9 * General Public License (GPL) Version 2, available from the file
10 * COPYING in the main directory of this source tree, or the
11 * OpenIB.org BSD license below:
13 * Redistribution and use in source and binary forms, with or
14 * without modification, are permitted provided that the following
17 * - Redistributions of source code must retain the above
18 * copyright notice, this list of conditions and the following
21 * - Redistributions in binary form must reproduce the above
22 * copyright notice, this list of conditions and the following
23 * disclaimer in the documentation and/or other materials
24 * provided with the distribution.
26 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
27 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
28 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
29 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
30 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
31 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
32 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
36 #include <linux/module.h>
37 #include <linux/init.h>
38 #include <linux/errno.h>
39 #include <linux/pci.h>
40 #include <linux/dma-mapping.h>
41 #include <linux/slab.h>
42 #include <linux/io-mapping.h>
43 #include <linux/delay.h>
44 #include <linux/kmod.h>
46 #include <linux/mlx4/device.h>
47 #include <linux/mlx4/doorbell.h>
53 MODULE_AUTHOR("Roland Dreier");
54 MODULE_DESCRIPTION("Mellanox ConnectX HCA low-level driver");
55 MODULE_LICENSE("Dual BSD/GPL");
56 MODULE_VERSION(DRV_VERSION);
58 struct workqueue_struct *mlx4_wq;
60 #ifdef CONFIG_MLX4_DEBUG
62 int mlx4_debug_level = 0;
63 module_param_named(debug_level, mlx4_debug_level, int, 0644);
64 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
66 #endif /* CONFIG_MLX4_DEBUG */
71 module_param(msi_x, int, 0444);
72 MODULE_PARM_DESC(msi_x, "attempt to use MSI-X if nonzero");
74 #else /* CONFIG_PCI_MSI */
78 #endif /* CONFIG_PCI_MSI */
80 static uint8_t num_vfs[3] = {0, 0, 0};
81 static int num_vfs_argc;
82 module_param_array(num_vfs, byte , &num_vfs_argc, 0444);
83 MODULE_PARM_DESC(num_vfs, "enable #num_vfs functions if num_vfs > 0\n"
84 "num_vfs=port1,port2,port1+2");
86 static uint8_t probe_vf[3] = {0, 0, 0};
87 static int probe_vfs_argc;
88 module_param_array(probe_vf, byte, &probe_vfs_argc, 0444);
89 MODULE_PARM_DESC(probe_vf, "number of vfs to probe by pf driver (num_vfs > 0)\n"
90 "probe_vf=port1,port2,port1+2");
92 int mlx4_log_num_mgm_entry_size = MLX4_DEFAULT_MGM_LOG_ENTRY_SIZE;
93 module_param_named(log_num_mgm_entry_size,
94 mlx4_log_num_mgm_entry_size, int, 0444);
95 MODULE_PARM_DESC(log_num_mgm_entry_size, "log mgm size, that defines the num"
96 " of qp per mcg, for example:"
97 " 10 gives 248.range: 7 <="
98 " log_num_mgm_entry_size <= 12."
99 " To activate device managed"
100 " flow steering when available, set to -1");
102 static bool enable_64b_cqe_eqe = true;
103 module_param(enable_64b_cqe_eqe, bool, 0444);
104 MODULE_PARM_DESC(enable_64b_cqe_eqe,
105 "Enable 64 byte CQEs/EQEs when the FW supports this (default: True)");
107 #define PF_CONTEXT_BEHAVIOUR_MASK (MLX4_FUNC_CAP_64B_EQE_CQE | \
108 MLX4_FUNC_CAP_EQE_CQE_STRIDE | \
109 MLX4_FUNC_CAP_DMFS_A0_STATIC)
111 #define RESET_PERSIST_MASK_FLAGS (MLX4_FLAG_SRIOV)
113 static char mlx4_version[] =
114 DRV_NAME ": Mellanox ConnectX core driver v"
115 DRV_VERSION " (" DRV_RELDATE ")\n";
117 static struct mlx4_profile default_profile = {
120 .rdmarc_per_qp = 1 << 4,
124 .num_mtt = 1 << 20, /* It is really num mtt segements */
127 static struct mlx4_profile low_mem_profile = {
130 .rdmarc_per_qp = 1 << 4,
137 static int log_num_mac = 7;
138 module_param_named(log_num_mac, log_num_mac, int, 0444);
139 MODULE_PARM_DESC(log_num_mac, "Log2 max number of MACs per ETH port (1-7)");
141 static int log_num_vlan;
142 module_param_named(log_num_vlan, log_num_vlan, int, 0444);
143 MODULE_PARM_DESC(log_num_vlan, "Log2 max number of VLANs per ETH port (0-7)");
144 /* Log2 max number of VLANs per ETH port (0-7) */
145 #define MLX4_LOG_NUM_VLANS 7
146 #define MLX4_MIN_LOG_NUM_VLANS 0
147 #define MLX4_MIN_LOG_NUM_MAC 1
149 static bool use_prio;
150 module_param_named(use_prio, use_prio, bool, 0444);
151 MODULE_PARM_DESC(use_prio, "Enable steering by VLAN priority on ETH ports (deprecated)");
153 int log_mtts_per_seg = ilog2(MLX4_MTT_ENTRY_PER_SEG);
154 module_param_named(log_mtts_per_seg, log_mtts_per_seg, int, 0444);
155 MODULE_PARM_DESC(log_mtts_per_seg, "Log2 number of MTT entries per segment (1-7)");
157 static int port_type_array[2] = {MLX4_PORT_TYPE_NONE, MLX4_PORT_TYPE_NONE};
158 static int arr_argc = 2;
159 module_param_array(port_type_array, int, &arr_argc, 0444);
160 MODULE_PARM_DESC(port_type_array, "Array of port types: HW_DEFAULT (0) is default "
161 "1 for IB, 2 for Ethernet");
163 struct mlx4_port_config {
164 struct list_head list;
165 enum mlx4_port_type port_type[MLX4_MAX_PORTS + 1];
166 struct pci_dev *pdev;
169 static atomic_t pf_loading = ATOMIC_INIT(0);
171 int mlx4_check_port_params(struct mlx4_dev *dev,
172 enum mlx4_port_type *port_type)
176 if (!(dev->caps.flags & MLX4_DEV_CAP_FLAG_DPDP)) {
177 for (i = 0; i < dev->caps.num_ports - 1; i++) {
178 if (port_type[i] != port_type[i + 1]) {
179 mlx4_err(dev, "Only same port types supported on this HCA, aborting\n");
185 for (i = 0; i < dev->caps.num_ports; i++) {
186 if (!(port_type[i] & dev->caps.supported_type[i+1])) {
187 mlx4_err(dev, "Requested port type for port %d is not supported on this HCA\n",
195 static void mlx4_set_port_mask(struct mlx4_dev *dev)
199 for (i = 1; i <= dev->caps.num_ports; ++i)
200 dev->caps.port_mask[i] = dev->caps.port_type[i];
204 MLX4_QUERY_FUNC_NUM_SYS_EQS = 1 << 0,
207 static int mlx4_query_func(struct mlx4_dev *dev, struct mlx4_dev_cap *dev_cap)
210 struct mlx4_func func;
212 if (dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_SYS_EQS) {
213 err = mlx4_QUERY_FUNC(dev, &func, 0);
215 mlx4_err(dev, "QUERY_DEV_CAP command failed, aborting.\n");
218 dev_cap->max_eqs = func.max_eq;
219 dev_cap->reserved_eqs = func.rsvd_eqs;
220 dev_cap->reserved_uars = func.rsvd_uars;
221 err |= MLX4_QUERY_FUNC_NUM_SYS_EQS;
226 static void mlx4_enable_cqe_eqe_stride(struct mlx4_dev *dev)
228 struct mlx4_caps *dev_cap = &dev->caps;
230 /* FW not supporting or cancelled by user */
231 if (!(dev_cap->flags2 & MLX4_DEV_CAP_FLAG2_EQE_STRIDE) ||
232 !(dev_cap->flags2 & MLX4_DEV_CAP_FLAG2_CQE_STRIDE))
235 /* Must have 64B CQE_EQE enabled by FW to use bigger stride
236 * When FW has NCSI it may decide not to report 64B CQE/EQEs
238 if (!(dev_cap->flags & MLX4_DEV_CAP_FLAG_64B_EQE) ||
239 !(dev_cap->flags & MLX4_DEV_CAP_FLAG_64B_CQE)) {
240 dev_cap->flags2 &= ~MLX4_DEV_CAP_FLAG2_CQE_STRIDE;
241 dev_cap->flags2 &= ~MLX4_DEV_CAP_FLAG2_EQE_STRIDE;
245 if (cache_line_size() == 128 || cache_line_size() == 256) {
246 mlx4_dbg(dev, "Enabling CQE stride cacheLine supported\n");
247 /* Changing the real data inside CQE size to 32B */
248 dev_cap->flags &= ~MLX4_DEV_CAP_FLAG_64B_CQE;
249 dev_cap->flags &= ~MLX4_DEV_CAP_FLAG_64B_EQE;
251 if (mlx4_is_master(dev))
252 dev_cap->function_caps |= MLX4_FUNC_CAP_EQE_CQE_STRIDE;
254 if (cache_line_size() != 32 && cache_line_size() != 64)
255 mlx4_dbg(dev, "Disabling CQE stride, cacheLine size unsupported\n");
256 dev_cap->flags2 &= ~MLX4_DEV_CAP_FLAG2_CQE_STRIDE;
257 dev_cap->flags2 &= ~MLX4_DEV_CAP_FLAG2_EQE_STRIDE;
261 static int _mlx4_dev_port(struct mlx4_dev *dev, int port,
262 struct mlx4_port_cap *port_cap)
264 dev->caps.vl_cap[port] = port_cap->max_vl;
265 dev->caps.ib_mtu_cap[port] = port_cap->ib_mtu;
266 dev->phys_caps.gid_phys_table_len[port] = port_cap->max_gids;
267 dev->phys_caps.pkey_phys_table_len[port] = port_cap->max_pkeys;
268 /* set gid and pkey table operating lengths by default
269 * to non-sriov values
271 dev->caps.gid_table_len[port] = port_cap->max_gids;
272 dev->caps.pkey_table_len[port] = port_cap->max_pkeys;
273 dev->caps.port_width_cap[port] = port_cap->max_port_width;
274 dev->caps.eth_mtu_cap[port] = port_cap->eth_mtu;
275 dev->caps.def_mac[port] = port_cap->def_mac;
276 dev->caps.supported_type[port] = port_cap->supported_port_types;
277 dev->caps.suggested_type[port] = port_cap->suggested_type;
278 dev->caps.default_sense[port] = port_cap->default_sense;
279 dev->caps.trans_type[port] = port_cap->trans_type;
280 dev->caps.vendor_oui[port] = port_cap->vendor_oui;
281 dev->caps.wavelength[port] = port_cap->wavelength;
282 dev->caps.trans_code[port] = port_cap->trans_code;
287 static int mlx4_dev_port(struct mlx4_dev *dev, int port,
288 struct mlx4_port_cap *port_cap)
292 err = mlx4_QUERY_PORT(dev, port, port_cap);
295 mlx4_err(dev, "QUERY_PORT command failed.\n");
300 static inline void mlx4_enable_ignore_fcs(struct mlx4_dev *dev)
302 if (!(dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_IGNORE_FCS))
305 if (mlx4_is_mfunc(dev)) {
306 mlx4_dbg(dev, "SRIOV mode - Disabling Ignore FCS");
307 dev->caps.flags2 &= ~MLX4_DEV_CAP_FLAG2_IGNORE_FCS;
311 if (!(dev->caps.flags & MLX4_DEV_CAP_FLAG_FCS_KEEP)) {
313 "Keep FCS is not supported - Disabling Ignore FCS");
314 dev->caps.flags2 &= ~MLX4_DEV_CAP_FLAG2_IGNORE_FCS;
319 #define MLX4_A0_STEERING_TABLE_SIZE 256
320 static int mlx4_dev_cap(struct mlx4_dev *dev, struct mlx4_dev_cap *dev_cap)
325 err = mlx4_QUERY_DEV_CAP(dev, dev_cap);
327 mlx4_err(dev, "QUERY_DEV_CAP command failed, aborting\n");
330 mlx4_dev_cap_dump(dev, dev_cap);
332 if (dev_cap->min_page_sz > PAGE_SIZE) {
333 mlx4_err(dev, "HCA minimum page size of %d bigger than kernel PAGE_SIZE of %ld, aborting\n",
334 dev_cap->min_page_sz, PAGE_SIZE);
337 if (dev_cap->num_ports > MLX4_MAX_PORTS) {
338 mlx4_err(dev, "HCA has %d ports, but we only support %d, aborting\n",
339 dev_cap->num_ports, MLX4_MAX_PORTS);
343 if (dev_cap->uar_size > pci_resource_len(dev->persist->pdev, 2)) {
344 mlx4_err(dev, "HCA reported UAR size of 0x%x bigger than PCI resource 2 size of 0x%llx, aborting\n",
347 pci_resource_len(dev->persist->pdev, 2));
351 dev->caps.num_ports = dev_cap->num_ports;
352 dev->caps.num_sys_eqs = dev_cap->num_sys_eqs;
353 dev->phys_caps.num_phys_eqs = dev_cap->flags2 & MLX4_DEV_CAP_FLAG2_SYS_EQS ?
354 dev->caps.num_sys_eqs :
356 for (i = 1; i <= dev->caps.num_ports; ++i) {
357 err = _mlx4_dev_port(dev, i, dev_cap->port_cap + i);
359 mlx4_err(dev, "QUERY_PORT command failed, aborting\n");
364 dev->caps.uar_page_size = PAGE_SIZE;
365 dev->caps.num_uars = dev_cap->uar_size / PAGE_SIZE;
366 dev->caps.local_ca_ack_delay = dev_cap->local_ca_ack_delay;
367 dev->caps.bf_reg_size = dev_cap->bf_reg_size;
368 dev->caps.bf_regs_per_page = dev_cap->bf_regs_per_page;
369 dev->caps.max_sq_sg = dev_cap->max_sq_sg;
370 dev->caps.max_rq_sg = dev_cap->max_rq_sg;
371 dev->caps.max_wqes = dev_cap->max_qp_sz;
372 dev->caps.max_qp_init_rdma = dev_cap->max_requester_per_qp;
373 dev->caps.max_srq_wqes = dev_cap->max_srq_sz;
374 dev->caps.max_srq_sge = dev_cap->max_rq_sg - 1;
375 dev->caps.reserved_srqs = dev_cap->reserved_srqs;
376 dev->caps.max_sq_desc_sz = dev_cap->max_sq_desc_sz;
377 dev->caps.max_rq_desc_sz = dev_cap->max_rq_desc_sz;
379 * Subtract 1 from the limit because we need to allocate a
380 * spare CQE so the HCA HW can tell the difference between an
381 * empty CQ and a full CQ.
383 dev->caps.max_cqes = dev_cap->max_cq_sz - 1;
384 dev->caps.reserved_cqs = dev_cap->reserved_cqs;
385 dev->caps.reserved_eqs = dev_cap->reserved_eqs;
386 dev->caps.reserved_mtts = dev_cap->reserved_mtts;
387 dev->caps.reserved_mrws = dev_cap->reserved_mrws;
389 /* The first 128 UARs are used for EQ doorbells */
390 dev->caps.reserved_uars = max_t(int, 128, dev_cap->reserved_uars);
391 dev->caps.reserved_pds = dev_cap->reserved_pds;
392 dev->caps.reserved_xrcds = (dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC) ?
393 dev_cap->reserved_xrcds : 0;
394 dev->caps.max_xrcds = (dev->caps.flags & MLX4_DEV_CAP_FLAG_XRC) ?
395 dev_cap->max_xrcds : 0;
396 dev->caps.mtt_entry_sz = dev_cap->mtt_entry_sz;
398 dev->caps.max_msg_sz = dev_cap->max_msg_sz;
399 dev->caps.page_size_cap = ~(u32) (dev_cap->min_page_sz - 1);
400 dev->caps.flags = dev_cap->flags;
401 dev->caps.flags2 = dev_cap->flags2;
402 dev->caps.bmme_flags = dev_cap->bmme_flags;
403 dev->caps.reserved_lkey = dev_cap->reserved_lkey;
404 dev->caps.stat_rate_support = dev_cap->stat_rate_support;
405 dev->caps.max_gso_sz = dev_cap->max_gso_sz;
406 dev->caps.max_rss_tbl_sz = dev_cap->max_rss_tbl_sz;
408 if (dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_PHV_EN) {
409 struct mlx4_init_hca_param hca_param;
411 memset(&hca_param, 0, sizeof(hca_param));
412 err = mlx4_QUERY_HCA(dev, &hca_param);
413 /* Turn off PHV_EN flag in case phv_check_en is set.
414 * phv_check_en is a HW check that parse the packet and verify
415 * phv bit was reported correctly in the wqe. To allow QinQ
416 * PHV_EN flag should be set and phv_check_en must be cleared
417 * otherwise QinQ packets will be drop by the HW.
419 if (err || hca_param.phv_check_en)
420 dev->caps.flags2 &= ~MLX4_DEV_CAP_FLAG2_PHV_EN;
423 /* Sense port always allowed on supported devices for ConnectX-1 and -2 */
424 if (mlx4_priv(dev)->pci_dev_data & MLX4_PCI_DEV_FORCE_SENSE_PORT)
425 dev->caps.flags |= MLX4_DEV_CAP_FLAG_SENSE_SUPPORT;
426 /* Don't do sense port on multifunction devices (for now at least) */
427 if (mlx4_is_mfunc(dev))
428 dev->caps.flags &= ~MLX4_DEV_CAP_FLAG_SENSE_SUPPORT;
430 if (mlx4_low_memory_profile()) {
431 dev->caps.log_num_macs = MLX4_MIN_LOG_NUM_MAC;
432 dev->caps.log_num_vlans = MLX4_MIN_LOG_NUM_VLANS;
434 dev->caps.log_num_macs = log_num_mac;
435 dev->caps.log_num_vlans = MLX4_LOG_NUM_VLANS;
438 for (i = 1; i <= dev->caps.num_ports; ++i) {
439 dev->caps.port_type[i] = MLX4_PORT_TYPE_NONE;
440 if (dev->caps.supported_type[i]) {
441 /* if only ETH is supported - assign ETH */
442 if (dev->caps.supported_type[i] == MLX4_PORT_TYPE_ETH)
443 dev->caps.port_type[i] = MLX4_PORT_TYPE_ETH;
444 /* if only IB is supported, assign IB */
445 else if (dev->caps.supported_type[i] ==
447 dev->caps.port_type[i] = MLX4_PORT_TYPE_IB;
449 /* if IB and ETH are supported, we set the port
450 * type according to user selection of port type;
451 * if user selected none, take the FW hint */
452 if (port_type_array[i - 1] == MLX4_PORT_TYPE_NONE)
453 dev->caps.port_type[i] = dev->caps.suggested_type[i] ?
454 MLX4_PORT_TYPE_ETH : MLX4_PORT_TYPE_IB;
456 dev->caps.port_type[i] = port_type_array[i - 1];
460 * Link sensing is allowed on the port if 3 conditions are true:
461 * 1. Both protocols are supported on the port.
462 * 2. Different types are supported on the port
463 * 3. FW declared that it supports link sensing
465 mlx4_priv(dev)->sense.sense_allowed[i] =
466 ((dev->caps.supported_type[i] == MLX4_PORT_TYPE_AUTO) &&
467 (dev->caps.flags & MLX4_DEV_CAP_FLAG_DPDP) &&
468 (dev->caps.flags & MLX4_DEV_CAP_FLAG_SENSE_SUPPORT));
471 * If "default_sense" bit is set, we move the port to "AUTO" mode
472 * and perform sense_port FW command to try and set the correct
473 * port type from beginning
475 if (mlx4_priv(dev)->sense.sense_allowed[i] && dev->caps.default_sense[i]) {
476 enum mlx4_port_type sensed_port = MLX4_PORT_TYPE_NONE;
477 dev->caps.possible_type[i] = MLX4_PORT_TYPE_AUTO;
478 mlx4_SENSE_PORT(dev, i, &sensed_port);
479 if (sensed_port != MLX4_PORT_TYPE_NONE)
480 dev->caps.port_type[i] = sensed_port;
482 dev->caps.possible_type[i] = dev->caps.port_type[i];
485 if (dev->caps.log_num_macs > dev_cap->port_cap[i].log_max_macs) {
486 dev->caps.log_num_macs = dev_cap->port_cap[i].log_max_macs;
487 mlx4_warn(dev, "Requested number of MACs is too much for port %d, reducing to %d\n",
488 i, 1 << dev->caps.log_num_macs);
490 if (dev->caps.log_num_vlans > dev_cap->port_cap[i].log_max_vlans) {
491 dev->caps.log_num_vlans = dev_cap->port_cap[i].log_max_vlans;
492 mlx4_warn(dev, "Requested number of VLANs is too much for port %d, reducing to %d\n",
493 i, 1 << dev->caps.log_num_vlans);
497 if (mlx4_is_master(dev) && (dev->caps.num_ports == 2) &&
498 (port_type_array[0] == MLX4_PORT_TYPE_IB) &&
499 (port_type_array[1] == MLX4_PORT_TYPE_ETH)) {
501 "Granular QoS per VF not supported with IB/Eth configuration\n");
502 dev->caps.flags2 &= ~MLX4_DEV_CAP_FLAG2_QOS_VPP;
505 dev->caps.max_counters = dev_cap->max_counters;
507 dev->caps.reserved_qps_cnt[MLX4_QP_REGION_FW] = dev_cap->reserved_qps;
508 dev->caps.reserved_qps_cnt[MLX4_QP_REGION_ETH_ADDR] =
509 dev->caps.reserved_qps_cnt[MLX4_QP_REGION_FC_ADDR] =
510 (1 << dev->caps.log_num_macs) *
511 (1 << dev->caps.log_num_vlans) *
513 dev->caps.reserved_qps_cnt[MLX4_QP_REGION_FC_EXCH] = MLX4_NUM_FEXCH;
515 if (dev_cap->dmfs_high_rate_qpn_base > 0 &&
516 dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_FS_EN)
517 dev->caps.dmfs_high_rate_qpn_base = dev_cap->dmfs_high_rate_qpn_base;
519 dev->caps.dmfs_high_rate_qpn_base =
520 dev->caps.reserved_qps_cnt[MLX4_QP_REGION_FW];
522 if (dev_cap->dmfs_high_rate_qpn_range > 0 &&
523 dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_FS_EN) {
524 dev->caps.dmfs_high_rate_qpn_range = dev_cap->dmfs_high_rate_qpn_range;
525 dev->caps.dmfs_high_steer_mode = MLX4_STEERING_DMFS_A0_DEFAULT;
526 dev->caps.flags2 |= MLX4_DEV_CAP_FLAG2_FS_A0;
528 dev->caps.dmfs_high_steer_mode = MLX4_STEERING_DMFS_A0_NOT_SUPPORTED;
529 dev->caps.dmfs_high_rate_qpn_base =
530 dev->caps.reserved_qps_cnt[MLX4_QP_REGION_FW];
531 dev->caps.dmfs_high_rate_qpn_range = MLX4_A0_STEERING_TABLE_SIZE;
534 dev->caps.rl_caps = dev_cap->rl_caps;
536 dev->caps.reserved_qps_cnt[MLX4_QP_REGION_RSS_RAW_ETH] =
537 dev->caps.dmfs_high_rate_qpn_range;
539 dev->caps.reserved_qps = dev->caps.reserved_qps_cnt[MLX4_QP_REGION_FW] +
540 dev->caps.reserved_qps_cnt[MLX4_QP_REGION_ETH_ADDR] +
541 dev->caps.reserved_qps_cnt[MLX4_QP_REGION_FC_ADDR] +
542 dev->caps.reserved_qps_cnt[MLX4_QP_REGION_FC_EXCH];
544 dev->caps.sqp_demux = (mlx4_is_master(dev)) ? MLX4_MAX_NUM_SLAVES : 0;
546 if (!enable_64b_cqe_eqe && !mlx4_is_slave(dev)) {
548 (MLX4_DEV_CAP_FLAG_64B_CQE | MLX4_DEV_CAP_FLAG_64B_EQE)) {
549 mlx4_warn(dev, "64B EQEs/CQEs supported by the device but not enabled\n");
550 dev->caps.flags &= ~MLX4_DEV_CAP_FLAG_64B_CQE;
551 dev->caps.flags &= ~MLX4_DEV_CAP_FLAG_64B_EQE;
554 if (dev_cap->flags2 &
555 (MLX4_DEV_CAP_FLAG2_CQE_STRIDE |
556 MLX4_DEV_CAP_FLAG2_EQE_STRIDE)) {
557 mlx4_warn(dev, "Disabling EQE/CQE stride per user request\n");
558 dev_cap->flags2 &= ~MLX4_DEV_CAP_FLAG2_CQE_STRIDE;
559 dev_cap->flags2 &= ~MLX4_DEV_CAP_FLAG2_EQE_STRIDE;
563 if ((dev->caps.flags &
564 (MLX4_DEV_CAP_FLAG_64B_CQE | MLX4_DEV_CAP_FLAG_64B_EQE)) &&
566 dev->caps.function_caps |= MLX4_FUNC_CAP_64B_EQE_CQE;
568 if (!mlx4_is_slave(dev)) {
569 mlx4_enable_cqe_eqe_stride(dev);
570 dev->caps.alloc_res_qp_mask =
571 (dev->caps.bf_reg_size ? MLX4_RESERVE_ETH_BF_QP : 0) |
574 if (!(dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_ETS_CFG) &&
575 dev->caps.flags & MLX4_DEV_CAP_FLAG_SET_ETH_SCHED) {
576 mlx4_warn(dev, "Old device ETS support detected\n");
577 mlx4_warn(dev, "Consider upgrading device FW.\n");
578 dev->caps.flags2 |= MLX4_DEV_CAP_FLAG2_ETS_CFG;
582 dev->caps.alloc_res_qp_mask = 0;
585 mlx4_enable_ignore_fcs(dev);
590 static int mlx4_get_pcie_dev_link_caps(struct mlx4_dev *dev,
591 enum pci_bus_speed *speed,
592 enum pcie_link_width *width)
594 u32 lnkcap1, lnkcap2;
597 #define PCIE_MLW_CAP_SHIFT 4 /* start of MLW mask in link capabilities */
599 *speed = PCI_SPEED_UNKNOWN;
600 *width = PCIE_LNK_WIDTH_UNKNOWN;
602 err1 = pcie_capability_read_dword(dev->persist->pdev, PCI_EXP_LNKCAP,
604 err2 = pcie_capability_read_dword(dev->persist->pdev, PCI_EXP_LNKCAP2,
606 if (!err2 && lnkcap2) { /* PCIe r3.0-compliant */
607 if (lnkcap2 & PCI_EXP_LNKCAP2_SLS_8_0GB)
608 *speed = PCIE_SPEED_8_0GT;
609 else if (lnkcap2 & PCI_EXP_LNKCAP2_SLS_5_0GB)
610 *speed = PCIE_SPEED_5_0GT;
611 else if (lnkcap2 & PCI_EXP_LNKCAP2_SLS_2_5GB)
612 *speed = PCIE_SPEED_2_5GT;
615 *width = (lnkcap1 & PCI_EXP_LNKCAP_MLW) >> PCIE_MLW_CAP_SHIFT;
616 if (!lnkcap2) { /* pre-r3.0 */
617 if (lnkcap1 & PCI_EXP_LNKCAP_SLS_5_0GB)
618 *speed = PCIE_SPEED_5_0GT;
619 else if (lnkcap1 & PCI_EXP_LNKCAP_SLS_2_5GB)
620 *speed = PCIE_SPEED_2_5GT;
624 if (*speed == PCI_SPEED_UNKNOWN || *width == PCIE_LNK_WIDTH_UNKNOWN) {
626 err2 ? err2 : -EINVAL;
631 static void mlx4_check_pcie_caps(struct mlx4_dev *dev)
633 enum pcie_link_width width, width_cap;
634 enum pci_bus_speed speed, speed_cap;
637 #define PCIE_SPEED_STR(speed) \
638 (speed == PCIE_SPEED_8_0GT ? "8.0GT/s" : \
639 speed == PCIE_SPEED_5_0GT ? "5.0GT/s" : \
640 speed == PCIE_SPEED_2_5GT ? "2.5GT/s" : \
643 err = mlx4_get_pcie_dev_link_caps(dev, &speed_cap, &width_cap);
646 "Unable to determine PCIe device BW capabilities\n");
650 err = pcie_get_minimum_link(dev->persist->pdev, &speed, &width);
651 if (err || speed == PCI_SPEED_UNKNOWN ||
652 width == PCIE_LNK_WIDTH_UNKNOWN) {
654 "Unable to determine PCI device chain minimum BW\n");
658 if (width != width_cap || speed != speed_cap)
660 "PCIe BW is different than device's capability\n");
662 mlx4_info(dev, "PCIe link speed is %s, device supports %s\n",
663 PCIE_SPEED_STR(speed), PCIE_SPEED_STR(speed_cap));
664 mlx4_info(dev, "PCIe link width is x%d, device supports x%d\n",
669 /*The function checks if there are live vf, return the num of them*/
670 static int mlx4_how_many_lives_vf(struct mlx4_dev *dev)
672 struct mlx4_priv *priv = mlx4_priv(dev);
673 struct mlx4_slave_state *s_state;
677 for (i = 1/*the ppf is 0*/; i < dev->num_slaves; ++i) {
678 s_state = &priv->mfunc.master.slave_state[i];
679 if (s_state->active && s_state->last_cmd !=
680 MLX4_COMM_CMD_RESET) {
681 mlx4_warn(dev, "%s: slave: %d is still active\n",
689 int mlx4_get_parav_qkey(struct mlx4_dev *dev, u32 qpn, u32 *qkey)
691 u32 qk = MLX4_RESERVED_QKEY_BASE;
693 if (qpn >= dev->phys_caps.base_tunnel_sqpn + 8 * MLX4_MFUNC_MAX ||
694 qpn < dev->phys_caps.base_proxy_sqpn)
697 if (qpn >= dev->phys_caps.base_tunnel_sqpn)
699 qk += qpn - dev->phys_caps.base_tunnel_sqpn;
701 qk += qpn - dev->phys_caps.base_proxy_sqpn;
705 EXPORT_SYMBOL(mlx4_get_parav_qkey);
707 void mlx4_sync_pkey_table(struct mlx4_dev *dev, int slave, int port, int i, int val)
709 struct mlx4_priv *priv = container_of(dev, struct mlx4_priv, dev);
711 if (!mlx4_is_master(dev))
714 priv->virt2phys_pkey[slave][port - 1][i] = val;
716 EXPORT_SYMBOL(mlx4_sync_pkey_table);
718 void mlx4_put_slave_node_guid(struct mlx4_dev *dev, int slave, __be64 guid)
720 struct mlx4_priv *priv = container_of(dev, struct mlx4_priv, dev);
722 if (!mlx4_is_master(dev))
725 priv->slave_node_guids[slave] = guid;
727 EXPORT_SYMBOL(mlx4_put_slave_node_guid);
729 __be64 mlx4_get_slave_node_guid(struct mlx4_dev *dev, int slave)
731 struct mlx4_priv *priv = container_of(dev, struct mlx4_priv, dev);
733 if (!mlx4_is_master(dev))
736 return priv->slave_node_guids[slave];
738 EXPORT_SYMBOL(mlx4_get_slave_node_guid);
740 int mlx4_is_slave_active(struct mlx4_dev *dev, int slave)
742 struct mlx4_priv *priv = mlx4_priv(dev);
743 struct mlx4_slave_state *s_slave;
745 if (!mlx4_is_master(dev))
748 s_slave = &priv->mfunc.master.slave_state[slave];
749 return !!s_slave->active;
751 EXPORT_SYMBOL(mlx4_is_slave_active);
753 static void slave_adjust_steering_mode(struct mlx4_dev *dev,
754 struct mlx4_dev_cap *dev_cap,
755 struct mlx4_init_hca_param *hca_param)
757 dev->caps.steering_mode = hca_param->steering_mode;
758 if (dev->caps.steering_mode == MLX4_STEERING_MODE_DEVICE_MANAGED) {
759 dev->caps.num_qp_per_mgm = dev_cap->fs_max_num_qp_per_entry;
760 dev->caps.fs_log_max_ucast_qp_range_size =
761 dev_cap->fs_log_max_ucast_qp_range_size;
763 dev->caps.num_qp_per_mgm =
764 4 * ((1 << hca_param->log_mc_entry_sz)/16 - 2);
766 mlx4_dbg(dev, "Steering mode is: %s\n",
767 mlx4_steering_mode_str(dev->caps.steering_mode));
770 static int mlx4_slave_cap(struct mlx4_dev *dev)
774 struct mlx4_dev_cap dev_cap;
775 struct mlx4_func_cap func_cap;
776 struct mlx4_init_hca_param hca_param;
779 memset(&hca_param, 0, sizeof(hca_param));
780 err = mlx4_QUERY_HCA(dev, &hca_param);
782 mlx4_err(dev, "QUERY_HCA command failed, aborting\n");
786 /* fail if the hca has an unknown global capability
787 * at this time global_caps should be always zeroed
789 if (hca_param.global_caps) {
790 mlx4_err(dev, "Unknown hca global capabilities\n");
794 dev->caps.hca_core_clock = hca_param.hca_core_clock;
796 memset(&dev_cap, 0, sizeof(dev_cap));
797 dev->caps.max_qp_dest_rdma = 1 << hca_param.log_rd_per_qp;
798 err = mlx4_dev_cap(dev, &dev_cap);
800 mlx4_err(dev, "QUERY_DEV_CAP command failed, aborting\n");
804 err = mlx4_QUERY_FW(dev);
806 mlx4_err(dev, "QUERY_FW command failed: could not get FW version\n");
808 page_size = ~dev->caps.page_size_cap + 1;
809 mlx4_warn(dev, "HCA minimum page size:%d\n", page_size);
810 if (page_size > PAGE_SIZE) {
811 mlx4_err(dev, "HCA minimum page size of %d bigger than kernel PAGE_SIZE of %ld, aborting\n",
812 page_size, PAGE_SIZE);
816 /* slave gets uar page size from QUERY_HCA fw command */
817 dev->caps.uar_page_size = 1 << (hca_param.uar_page_sz + 12);
819 /* TODO: relax this assumption */
820 if (dev->caps.uar_page_size != PAGE_SIZE) {
821 mlx4_err(dev, "UAR size:%d != kernel PAGE_SIZE of %ld\n",
822 dev->caps.uar_page_size, PAGE_SIZE);
826 memset(&func_cap, 0, sizeof(func_cap));
827 err = mlx4_QUERY_FUNC_CAP(dev, 0, &func_cap);
829 mlx4_err(dev, "QUERY_FUNC_CAP general command failed, aborting (%d)\n",
834 if ((func_cap.pf_context_behaviour | PF_CONTEXT_BEHAVIOUR_MASK) !=
835 PF_CONTEXT_BEHAVIOUR_MASK) {
836 mlx4_err(dev, "Unknown pf context behaviour %x known flags %x\n",
837 func_cap.pf_context_behaviour, PF_CONTEXT_BEHAVIOUR_MASK);
841 dev->caps.num_ports = func_cap.num_ports;
842 dev->quotas.qp = func_cap.qp_quota;
843 dev->quotas.srq = func_cap.srq_quota;
844 dev->quotas.cq = func_cap.cq_quota;
845 dev->quotas.mpt = func_cap.mpt_quota;
846 dev->quotas.mtt = func_cap.mtt_quota;
847 dev->caps.num_qps = 1 << hca_param.log_num_qps;
848 dev->caps.num_srqs = 1 << hca_param.log_num_srqs;
849 dev->caps.num_cqs = 1 << hca_param.log_num_cqs;
850 dev->caps.num_mpts = 1 << hca_param.log_mpt_sz;
851 dev->caps.num_eqs = func_cap.max_eq;
852 dev->caps.reserved_eqs = func_cap.reserved_eq;
853 dev->caps.reserved_lkey = func_cap.reserved_lkey;
854 dev->caps.num_pds = MLX4_NUM_PDS;
855 dev->caps.num_mgms = 0;
856 dev->caps.num_amgms = 0;
858 if (dev->caps.num_ports > MLX4_MAX_PORTS) {
859 mlx4_err(dev, "HCA has %d ports, but we only support %d, aborting\n",
860 dev->caps.num_ports, MLX4_MAX_PORTS);
864 mlx4_replace_zero_macs(dev);
866 dev->caps.qp0_qkey = kcalloc(dev->caps.num_ports, sizeof(u32), GFP_KERNEL);
867 dev->caps.qp0_tunnel = kcalloc(dev->caps.num_ports, sizeof (u32), GFP_KERNEL);
868 dev->caps.qp0_proxy = kcalloc(dev->caps.num_ports, sizeof (u32), GFP_KERNEL);
869 dev->caps.qp1_tunnel = kcalloc(dev->caps.num_ports, sizeof (u32), GFP_KERNEL);
870 dev->caps.qp1_proxy = kcalloc(dev->caps.num_ports, sizeof (u32), GFP_KERNEL);
872 if (!dev->caps.qp0_tunnel || !dev->caps.qp0_proxy ||
873 !dev->caps.qp1_tunnel || !dev->caps.qp1_proxy ||
874 !dev->caps.qp0_qkey) {
879 for (i = 1; i <= dev->caps.num_ports; ++i) {
880 err = mlx4_QUERY_FUNC_CAP(dev, i, &func_cap);
882 mlx4_err(dev, "QUERY_FUNC_CAP port command failed for port %d, aborting (%d)\n",
886 dev->caps.qp0_qkey[i - 1] = func_cap.qp0_qkey;
887 dev->caps.qp0_tunnel[i - 1] = func_cap.qp0_tunnel_qpn;
888 dev->caps.qp0_proxy[i - 1] = func_cap.qp0_proxy_qpn;
889 dev->caps.qp1_tunnel[i - 1] = func_cap.qp1_tunnel_qpn;
890 dev->caps.qp1_proxy[i - 1] = func_cap.qp1_proxy_qpn;
891 dev->caps.port_mask[i] = dev->caps.port_type[i];
892 dev->caps.phys_port_id[i] = func_cap.phys_port_id;
893 err = mlx4_get_slave_pkey_gid_tbl_len(dev, i,
894 &dev->caps.gid_table_len[i],
895 &dev->caps.pkey_table_len[i]);
900 if (dev->caps.uar_page_size * (dev->caps.num_uars -
901 dev->caps.reserved_uars) >
902 pci_resource_len(dev->persist->pdev,
904 mlx4_err(dev, "HCA reported UAR region size of 0x%x bigger than PCI resource 2 size of 0x%llx, aborting\n",
905 dev->caps.uar_page_size * dev->caps.num_uars,
907 pci_resource_len(dev->persist->pdev, 2));
912 if (hca_param.dev_cap_enabled & MLX4_DEV_CAP_64B_EQE_ENABLED) {
913 dev->caps.eqe_size = 64;
914 dev->caps.eqe_factor = 1;
916 dev->caps.eqe_size = 32;
917 dev->caps.eqe_factor = 0;
920 if (hca_param.dev_cap_enabled & MLX4_DEV_CAP_64B_CQE_ENABLED) {
921 dev->caps.cqe_size = 64;
922 dev->caps.userspace_caps |= MLX4_USER_DEV_CAP_LARGE_CQE;
924 dev->caps.cqe_size = 32;
927 if (hca_param.dev_cap_enabled & MLX4_DEV_CAP_EQE_STRIDE_ENABLED) {
928 dev->caps.eqe_size = hca_param.eqe_size;
929 dev->caps.eqe_factor = 0;
932 if (hca_param.dev_cap_enabled & MLX4_DEV_CAP_CQE_STRIDE_ENABLED) {
933 dev->caps.cqe_size = hca_param.cqe_size;
934 /* User still need to know when CQE > 32B */
935 dev->caps.userspace_caps |= MLX4_USER_DEV_CAP_LARGE_CQE;
938 dev->caps.flags2 &= ~MLX4_DEV_CAP_FLAG2_TS;
939 mlx4_warn(dev, "Timestamping is not supported in slave mode\n");
941 slave_adjust_steering_mode(dev, &dev_cap, &hca_param);
942 mlx4_dbg(dev, "RSS support for IP fragments is %s\n",
943 hca_param.rss_ip_frags ? "on" : "off");
945 if (func_cap.extra_flags & MLX4_QUERY_FUNC_FLAGS_BF_RES_QP &&
946 dev->caps.bf_reg_size)
947 dev->caps.alloc_res_qp_mask |= MLX4_RESERVE_ETH_BF_QP;
949 if (func_cap.extra_flags & MLX4_QUERY_FUNC_FLAGS_A0_RES_QP)
950 dev->caps.alloc_res_qp_mask |= MLX4_RESERVE_A0_QP;
955 kfree(dev->caps.qp0_qkey);
956 kfree(dev->caps.qp0_tunnel);
957 kfree(dev->caps.qp0_proxy);
958 kfree(dev->caps.qp1_tunnel);
959 kfree(dev->caps.qp1_proxy);
960 dev->caps.qp0_qkey = NULL;
961 dev->caps.qp0_tunnel = NULL;
962 dev->caps.qp0_proxy = NULL;
963 dev->caps.qp1_tunnel = NULL;
964 dev->caps.qp1_proxy = NULL;
969 static void mlx4_request_modules(struct mlx4_dev *dev)
972 int has_ib_port = false;
973 int has_eth_port = false;
974 #define EN_DRV_NAME "mlx4_en"
975 #define IB_DRV_NAME "mlx4_ib"
977 for (port = 1; port <= dev->caps.num_ports; port++) {
978 if (dev->caps.port_type[port] == MLX4_PORT_TYPE_IB)
980 else if (dev->caps.port_type[port] == MLX4_PORT_TYPE_ETH)
985 request_module_nowait(EN_DRV_NAME);
986 if (has_ib_port || (dev->caps.flags & MLX4_DEV_CAP_FLAG_IBOE))
987 request_module_nowait(IB_DRV_NAME);
991 * Change the port configuration of the device.
992 * Every user of this function must hold the port mutex.
994 int mlx4_change_port_types(struct mlx4_dev *dev,
995 enum mlx4_port_type *port_types)
1001 for (port = 0; port < dev->caps.num_ports; port++) {
1002 /* Change the port type only if the new type is different
1003 * from the current, and not set to Auto */
1004 if (port_types[port] != dev->caps.port_type[port + 1])
1008 mlx4_unregister_device(dev);
1009 for (port = 1; port <= dev->caps.num_ports; port++) {
1010 mlx4_CLOSE_PORT(dev, port);
1011 dev->caps.port_type[port] = port_types[port - 1];
1012 err = mlx4_SET_PORT(dev, port, -1);
1014 mlx4_err(dev, "Failed to set port %d, aborting\n",
1019 mlx4_set_port_mask(dev);
1020 err = mlx4_register_device(dev);
1022 mlx4_err(dev, "Failed to register device\n");
1025 mlx4_request_modules(dev);
1032 static ssize_t show_port_type(struct device *dev,
1033 struct device_attribute *attr,
1036 struct mlx4_port_info *info = container_of(attr, struct mlx4_port_info,
1038 struct mlx4_dev *mdev = info->dev;
1042 (mdev->caps.port_type[info->port] == MLX4_PORT_TYPE_IB) ?
1044 if (mdev->caps.possible_type[info->port] == MLX4_PORT_TYPE_AUTO)
1045 sprintf(buf, "auto (%s)\n", type);
1047 sprintf(buf, "%s\n", type);
1052 static ssize_t set_port_type(struct device *dev,
1053 struct device_attribute *attr,
1054 const char *buf, size_t count)
1056 struct mlx4_port_info *info = container_of(attr, struct mlx4_port_info,
1058 struct mlx4_dev *mdev = info->dev;
1059 struct mlx4_priv *priv = mlx4_priv(mdev);
1060 enum mlx4_port_type types[MLX4_MAX_PORTS];
1061 enum mlx4_port_type new_types[MLX4_MAX_PORTS];
1062 static DEFINE_MUTEX(set_port_type_mutex);
1066 mutex_lock(&set_port_type_mutex);
1068 if (!strcmp(buf, "ib\n"))
1069 info->tmp_type = MLX4_PORT_TYPE_IB;
1070 else if (!strcmp(buf, "eth\n"))
1071 info->tmp_type = MLX4_PORT_TYPE_ETH;
1072 else if (!strcmp(buf, "auto\n"))
1073 info->tmp_type = MLX4_PORT_TYPE_AUTO;
1075 mlx4_err(mdev, "%s is not supported port type\n", buf);
1080 mlx4_stop_sense(mdev);
1081 mutex_lock(&priv->port_mutex);
1082 /* Possible type is always the one that was delivered */
1083 mdev->caps.possible_type[info->port] = info->tmp_type;
1085 for (i = 0; i < mdev->caps.num_ports; i++) {
1086 types[i] = priv->port[i+1].tmp_type ? priv->port[i+1].tmp_type :
1087 mdev->caps.possible_type[i+1];
1088 if (types[i] == MLX4_PORT_TYPE_AUTO)
1089 types[i] = mdev->caps.port_type[i+1];
1092 if (!(mdev->caps.flags & MLX4_DEV_CAP_FLAG_DPDP) &&
1093 !(mdev->caps.flags & MLX4_DEV_CAP_FLAG_SENSE_SUPPORT)) {
1094 for (i = 1; i <= mdev->caps.num_ports; i++) {
1095 if (mdev->caps.possible_type[i] == MLX4_PORT_TYPE_AUTO) {
1096 mdev->caps.possible_type[i] = mdev->caps.port_type[i];
1102 mlx4_err(mdev, "Auto sensing is not supported on this HCA. Set only 'eth' or 'ib' for both ports (should be the same)\n");
1106 mlx4_do_sense_ports(mdev, new_types, types);
1108 err = mlx4_check_port_params(mdev, new_types);
1112 /* We are about to apply the changes after the configuration
1113 * was verified, no need to remember the temporary types
1115 for (i = 0; i < mdev->caps.num_ports; i++)
1116 priv->port[i + 1].tmp_type = 0;
1118 err = mlx4_change_port_types(mdev, new_types);
1121 mlx4_start_sense(mdev);
1122 mutex_unlock(&priv->port_mutex);
1124 mutex_unlock(&set_port_type_mutex);
1126 return err ? err : count;
1137 static inline int int_to_ibta_mtu(int mtu)
1140 case 256: return IB_MTU_256;
1141 case 512: return IB_MTU_512;
1142 case 1024: return IB_MTU_1024;
1143 case 2048: return IB_MTU_2048;
1144 case 4096: return IB_MTU_4096;
1149 static inline int ibta_mtu_to_int(enum ibta_mtu mtu)
1152 case IB_MTU_256: return 256;
1153 case IB_MTU_512: return 512;
1154 case IB_MTU_1024: return 1024;
1155 case IB_MTU_2048: return 2048;
1156 case IB_MTU_4096: return 4096;
1161 static ssize_t show_port_ib_mtu(struct device *dev,
1162 struct device_attribute *attr,
1165 struct mlx4_port_info *info = container_of(attr, struct mlx4_port_info,
1167 struct mlx4_dev *mdev = info->dev;
1169 if (mdev->caps.port_type[info->port] == MLX4_PORT_TYPE_ETH)
1170 mlx4_warn(mdev, "port level mtu is only used for IB ports\n");
1172 sprintf(buf, "%d\n",
1173 ibta_mtu_to_int(mdev->caps.port_ib_mtu[info->port]));
1177 static ssize_t set_port_ib_mtu(struct device *dev,
1178 struct device_attribute *attr,
1179 const char *buf, size_t count)
1181 struct mlx4_port_info *info = container_of(attr, struct mlx4_port_info,
1183 struct mlx4_dev *mdev = info->dev;
1184 struct mlx4_priv *priv = mlx4_priv(mdev);
1185 int err, port, mtu, ibta_mtu = -1;
1187 if (mdev->caps.port_type[info->port] == MLX4_PORT_TYPE_ETH) {
1188 mlx4_warn(mdev, "port level mtu is only used for IB ports\n");
1192 err = kstrtoint(buf, 0, &mtu);
1194 ibta_mtu = int_to_ibta_mtu(mtu);
1196 if (err || ibta_mtu < 0) {
1197 mlx4_err(mdev, "%s is invalid IBTA mtu\n", buf);
1201 mdev->caps.port_ib_mtu[info->port] = ibta_mtu;
1203 mlx4_stop_sense(mdev);
1204 mutex_lock(&priv->port_mutex);
1205 mlx4_unregister_device(mdev);
1206 for (port = 1; port <= mdev->caps.num_ports; port++) {
1207 mlx4_CLOSE_PORT(mdev, port);
1208 err = mlx4_SET_PORT(mdev, port, -1);
1210 mlx4_err(mdev, "Failed to set port %d, aborting\n",
1215 err = mlx4_register_device(mdev);
1217 mutex_unlock(&priv->port_mutex);
1218 mlx4_start_sense(mdev);
1219 return err ? err : count;
1222 int mlx4_bond(struct mlx4_dev *dev)
1225 struct mlx4_priv *priv = mlx4_priv(dev);
1227 mutex_lock(&priv->bond_mutex);
1229 if (!mlx4_is_bonded(dev))
1230 ret = mlx4_do_bond(dev, true);
1234 mutex_unlock(&priv->bond_mutex);
1236 mlx4_err(dev, "Failed to bond device: %d\n", ret);
1238 mlx4_dbg(dev, "Device is bonded\n");
1241 EXPORT_SYMBOL_GPL(mlx4_bond);
1243 int mlx4_unbond(struct mlx4_dev *dev)
1246 struct mlx4_priv *priv = mlx4_priv(dev);
1248 mutex_lock(&priv->bond_mutex);
1250 if (mlx4_is_bonded(dev))
1251 ret = mlx4_do_bond(dev, false);
1253 mutex_unlock(&priv->bond_mutex);
1255 mlx4_err(dev, "Failed to unbond device: %d\n", ret);
1257 mlx4_dbg(dev, "Device is unbonded\n");
1260 EXPORT_SYMBOL_GPL(mlx4_unbond);
1263 int mlx4_port_map_set(struct mlx4_dev *dev, struct mlx4_port_map *v2p)
1265 u8 port1 = v2p->port1;
1266 u8 port2 = v2p->port2;
1267 struct mlx4_priv *priv = mlx4_priv(dev);
1270 if (!(dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_PORT_REMAP))
1273 mutex_lock(&priv->bond_mutex);
1275 /* zero means keep current mapping for this port */
1277 port1 = priv->v2p.port1;
1279 port2 = priv->v2p.port2;
1281 if ((port1 < 1) || (port1 > MLX4_MAX_PORTS) ||
1282 (port2 < 1) || (port2 > MLX4_MAX_PORTS) ||
1283 (port1 == 2 && port2 == 1)) {
1284 /* besides boundary checks cross mapping makes
1285 * no sense and therefore not allowed */
1287 } else if ((port1 == priv->v2p.port1) &&
1288 (port2 == priv->v2p.port2)) {
1291 err = mlx4_virt2phy_port_map(dev, port1, port2);
1293 mlx4_dbg(dev, "port map changed: [%d][%d]\n",
1295 priv->v2p.port1 = port1;
1296 priv->v2p.port2 = port2;
1298 mlx4_err(dev, "Failed to change port mape: %d\n", err);
1302 mutex_unlock(&priv->bond_mutex);
1305 EXPORT_SYMBOL_GPL(mlx4_port_map_set);
1307 static int mlx4_load_fw(struct mlx4_dev *dev)
1309 struct mlx4_priv *priv = mlx4_priv(dev);
1312 priv->fw.fw_icm = mlx4_alloc_icm(dev, priv->fw.fw_pages,
1313 GFP_HIGHUSER | __GFP_NOWARN, 0);
1314 if (!priv->fw.fw_icm) {
1315 mlx4_err(dev, "Couldn't allocate FW area, aborting\n");
1319 err = mlx4_MAP_FA(dev, priv->fw.fw_icm);
1321 mlx4_err(dev, "MAP_FA command failed, aborting\n");
1325 err = mlx4_RUN_FW(dev);
1327 mlx4_err(dev, "RUN_FW command failed, aborting\n");
1337 mlx4_free_icm(dev, priv->fw.fw_icm, 0);
1341 static int mlx4_init_cmpt_table(struct mlx4_dev *dev, u64 cmpt_base,
1344 struct mlx4_priv *priv = mlx4_priv(dev);
1348 err = mlx4_init_icm_table(dev, &priv->qp_table.cmpt_table,
1350 ((u64) (MLX4_CMPT_TYPE_QP *
1351 cmpt_entry_sz) << MLX4_CMPT_SHIFT),
1352 cmpt_entry_sz, dev->caps.num_qps,
1353 dev->caps.reserved_qps_cnt[MLX4_QP_REGION_FW],
1358 err = mlx4_init_icm_table(dev, &priv->srq_table.cmpt_table,
1360 ((u64) (MLX4_CMPT_TYPE_SRQ *
1361 cmpt_entry_sz) << MLX4_CMPT_SHIFT),
1362 cmpt_entry_sz, dev->caps.num_srqs,
1363 dev->caps.reserved_srqs, 0, 0);
1367 err = mlx4_init_icm_table(dev, &priv->cq_table.cmpt_table,
1369 ((u64) (MLX4_CMPT_TYPE_CQ *
1370 cmpt_entry_sz) << MLX4_CMPT_SHIFT),
1371 cmpt_entry_sz, dev->caps.num_cqs,
1372 dev->caps.reserved_cqs, 0, 0);
1376 num_eqs = dev->phys_caps.num_phys_eqs;
1377 err = mlx4_init_icm_table(dev, &priv->eq_table.cmpt_table,
1379 ((u64) (MLX4_CMPT_TYPE_EQ *
1380 cmpt_entry_sz) << MLX4_CMPT_SHIFT),
1381 cmpt_entry_sz, num_eqs, num_eqs, 0, 0);
1388 mlx4_cleanup_icm_table(dev, &priv->cq_table.cmpt_table);
1391 mlx4_cleanup_icm_table(dev, &priv->srq_table.cmpt_table);
1394 mlx4_cleanup_icm_table(dev, &priv->qp_table.cmpt_table);
1400 static int mlx4_init_icm(struct mlx4_dev *dev, struct mlx4_dev_cap *dev_cap,
1401 struct mlx4_init_hca_param *init_hca, u64 icm_size)
1403 struct mlx4_priv *priv = mlx4_priv(dev);
1408 err = mlx4_SET_ICM_SIZE(dev, icm_size, &aux_pages);
1410 mlx4_err(dev, "SET_ICM_SIZE command failed, aborting\n");
1414 mlx4_dbg(dev, "%lld KB of HCA context requires %lld KB aux memory\n",
1415 (unsigned long long) icm_size >> 10,
1416 (unsigned long long) aux_pages << 2);
1418 priv->fw.aux_icm = mlx4_alloc_icm(dev, aux_pages,
1419 GFP_HIGHUSER | __GFP_NOWARN, 0);
1420 if (!priv->fw.aux_icm) {
1421 mlx4_err(dev, "Couldn't allocate aux memory, aborting\n");
1425 err = mlx4_MAP_ICM_AUX(dev, priv->fw.aux_icm);
1427 mlx4_err(dev, "MAP_ICM_AUX command failed, aborting\n");
1431 err = mlx4_init_cmpt_table(dev, init_hca->cmpt_base, dev_cap->cmpt_entry_sz);
1433 mlx4_err(dev, "Failed to map cMPT context memory, aborting\n");
1438 num_eqs = dev->phys_caps.num_phys_eqs;
1439 err = mlx4_init_icm_table(dev, &priv->eq_table.table,
1440 init_hca->eqc_base, dev_cap->eqc_entry_sz,
1441 num_eqs, num_eqs, 0, 0);
1443 mlx4_err(dev, "Failed to map EQ context memory, aborting\n");
1444 goto err_unmap_cmpt;
1448 * Reserved MTT entries must be aligned up to a cacheline
1449 * boundary, since the FW will write to them, while the driver
1450 * writes to all other MTT entries. (The variable
1451 * dev->caps.mtt_entry_sz below is really the MTT segment
1452 * size, not the raw entry size)
1454 dev->caps.reserved_mtts =
1455 ALIGN(dev->caps.reserved_mtts * dev->caps.mtt_entry_sz,
1456 dma_get_cache_alignment()) / dev->caps.mtt_entry_sz;
1458 err = mlx4_init_icm_table(dev, &priv->mr_table.mtt_table,
1460 dev->caps.mtt_entry_sz,
1462 dev->caps.reserved_mtts, 1, 0);
1464 mlx4_err(dev, "Failed to map MTT context memory, aborting\n");
1468 err = mlx4_init_icm_table(dev, &priv->mr_table.dmpt_table,
1469 init_hca->dmpt_base,
1470 dev_cap->dmpt_entry_sz,
1472 dev->caps.reserved_mrws, 1, 1);
1474 mlx4_err(dev, "Failed to map dMPT context memory, aborting\n");
1478 err = mlx4_init_icm_table(dev, &priv->qp_table.qp_table,
1480 dev_cap->qpc_entry_sz,
1482 dev->caps.reserved_qps_cnt[MLX4_QP_REGION_FW],
1485 mlx4_err(dev, "Failed to map QP context memory, aborting\n");
1486 goto err_unmap_dmpt;
1489 err = mlx4_init_icm_table(dev, &priv->qp_table.auxc_table,
1490 init_hca->auxc_base,
1491 dev_cap->aux_entry_sz,
1493 dev->caps.reserved_qps_cnt[MLX4_QP_REGION_FW],
1496 mlx4_err(dev, "Failed to map AUXC context memory, aborting\n");
1500 err = mlx4_init_icm_table(dev, &priv->qp_table.altc_table,
1501 init_hca->altc_base,
1502 dev_cap->altc_entry_sz,
1504 dev->caps.reserved_qps_cnt[MLX4_QP_REGION_FW],
1507 mlx4_err(dev, "Failed to map ALTC context memory, aborting\n");
1508 goto err_unmap_auxc;
1511 err = mlx4_init_icm_table(dev, &priv->qp_table.rdmarc_table,
1512 init_hca->rdmarc_base,
1513 dev_cap->rdmarc_entry_sz << priv->qp_table.rdmarc_shift,
1515 dev->caps.reserved_qps_cnt[MLX4_QP_REGION_FW],
1518 mlx4_err(dev, "Failed to map RDMARC context memory, aborting\n");
1519 goto err_unmap_altc;
1522 err = mlx4_init_icm_table(dev, &priv->cq_table.table,
1524 dev_cap->cqc_entry_sz,
1526 dev->caps.reserved_cqs, 0, 0);
1528 mlx4_err(dev, "Failed to map CQ context memory, aborting\n");
1529 goto err_unmap_rdmarc;
1532 err = mlx4_init_icm_table(dev, &priv->srq_table.table,
1533 init_hca->srqc_base,
1534 dev_cap->srq_entry_sz,
1536 dev->caps.reserved_srqs, 0, 0);
1538 mlx4_err(dev, "Failed to map SRQ context memory, aborting\n");
1543 * For flow steering device managed mode it is required to use
1544 * mlx4_init_icm_table. For B0 steering mode it's not strictly
1545 * required, but for simplicity just map the whole multicast
1546 * group table now. The table isn't very big and it's a lot
1547 * easier than trying to track ref counts.
1549 err = mlx4_init_icm_table(dev, &priv->mcg_table.table,
1551 mlx4_get_mgm_entry_size(dev),
1552 dev->caps.num_mgms + dev->caps.num_amgms,
1553 dev->caps.num_mgms + dev->caps.num_amgms,
1556 mlx4_err(dev, "Failed to map MCG context memory, aborting\n");
1563 mlx4_cleanup_icm_table(dev, &priv->srq_table.table);
1566 mlx4_cleanup_icm_table(dev, &priv->cq_table.table);
1569 mlx4_cleanup_icm_table(dev, &priv->qp_table.rdmarc_table);
1572 mlx4_cleanup_icm_table(dev, &priv->qp_table.altc_table);
1575 mlx4_cleanup_icm_table(dev, &priv->qp_table.auxc_table);
1578 mlx4_cleanup_icm_table(dev, &priv->qp_table.qp_table);
1581 mlx4_cleanup_icm_table(dev, &priv->mr_table.dmpt_table);
1584 mlx4_cleanup_icm_table(dev, &priv->mr_table.mtt_table);
1587 mlx4_cleanup_icm_table(dev, &priv->eq_table.table);
1590 mlx4_cleanup_icm_table(dev, &priv->eq_table.cmpt_table);
1591 mlx4_cleanup_icm_table(dev, &priv->cq_table.cmpt_table);
1592 mlx4_cleanup_icm_table(dev, &priv->srq_table.cmpt_table);
1593 mlx4_cleanup_icm_table(dev, &priv->qp_table.cmpt_table);
1596 mlx4_UNMAP_ICM_AUX(dev);
1599 mlx4_free_icm(dev, priv->fw.aux_icm, 0);
1604 static void mlx4_free_icms(struct mlx4_dev *dev)
1606 struct mlx4_priv *priv = mlx4_priv(dev);
1608 mlx4_cleanup_icm_table(dev, &priv->mcg_table.table);
1609 mlx4_cleanup_icm_table(dev, &priv->srq_table.table);
1610 mlx4_cleanup_icm_table(dev, &priv->cq_table.table);
1611 mlx4_cleanup_icm_table(dev, &priv->qp_table.rdmarc_table);
1612 mlx4_cleanup_icm_table(dev, &priv->qp_table.altc_table);
1613 mlx4_cleanup_icm_table(dev, &priv->qp_table.auxc_table);
1614 mlx4_cleanup_icm_table(dev, &priv->qp_table.qp_table);
1615 mlx4_cleanup_icm_table(dev, &priv->mr_table.dmpt_table);
1616 mlx4_cleanup_icm_table(dev, &priv->mr_table.mtt_table);
1617 mlx4_cleanup_icm_table(dev, &priv->eq_table.table);
1618 mlx4_cleanup_icm_table(dev, &priv->eq_table.cmpt_table);
1619 mlx4_cleanup_icm_table(dev, &priv->cq_table.cmpt_table);
1620 mlx4_cleanup_icm_table(dev, &priv->srq_table.cmpt_table);
1621 mlx4_cleanup_icm_table(dev, &priv->qp_table.cmpt_table);
1623 mlx4_UNMAP_ICM_AUX(dev);
1624 mlx4_free_icm(dev, priv->fw.aux_icm, 0);
1627 static void mlx4_slave_exit(struct mlx4_dev *dev)
1629 struct mlx4_priv *priv = mlx4_priv(dev);
1631 mutex_lock(&priv->cmd.slave_cmd_mutex);
1632 if (mlx4_comm_cmd(dev, MLX4_COMM_CMD_RESET, 0, MLX4_COMM_CMD_NA_OP,
1634 mlx4_warn(dev, "Failed to close slave function\n");
1635 mutex_unlock(&priv->cmd.slave_cmd_mutex);
1638 static int map_bf_area(struct mlx4_dev *dev)
1640 struct mlx4_priv *priv = mlx4_priv(dev);
1641 resource_size_t bf_start;
1642 resource_size_t bf_len;
1645 if (!dev->caps.bf_reg_size)
1648 bf_start = pci_resource_start(dev->persist->pdev, 2) +
1649 (dev->caps.num_uars << PAGE_SHIFT);
1650 bf_len = pci_resource_len(dev->persist->pdev, 2) -
1651 (dev->caps.num_uars << PAGE_SHIFT);
1652 priv->bf_mapping = io_mapping_create_wc(bf_start, bf_len);
1653 if (!priv->bf_mapping)
1659 static void unmap_bf_area(struct mlx4_dev *dev)
1661 if (mlx4_priv(dev)->bf_mapping)
1662 io_mapping_free(mlx4_priv(dev)->bf_mapping);
1665 cycle_t mlx4_read_clock(struct mlx4_dev *dev)
1667 u32 clockhi, clocklo, clockhi1;
1670 struct mlx4_priv *priv = mlx4_priv(dev);
1672 for (i = 0; i < 10; i++) {
1673 clockhi = swab32(readl(priv->clock_mapping));
1674 clocklo = swab32(readl(priv->clock_mapping + 4));
1675 clockhi1 = swab32(readl(priv->clock_mapping));
1676 if (clockhi == clockhi1)
1680 cycles = (u64) clockhi << 32 | (u64) clocklo;
1684 EXPORT_SYMBOL_GPL(mlx4_read_clock);
1687 static int map_internal_clock(struct mlx4_dev *dev)
1689 struct mlx4_priv *priv = mlx4_priv(dev);
1691 priv->clock_mapping =
1692 ioremap(pci_resource_start(dev->persist->pdev,
1693 priv->fw.clock_bar) +
1694 priv->fw.clock_offset, MLX4_CLOCK_SIZE);
1696 if (!priv->clock_mapping)
1702 int mlx4_get_internal_clock_params(struct mlx4_dev *dev,
1703 struct mlx4_clock_params *params)
1705 struct mlx4_priv *priv = mlx4_priv(dev);
1707 if (mlx4_is_slave(dev))
1713 params->bar = priv->fw.clock_bar;
1714 params->offset = priv->fw.clock_offset;
1715 params->size = MLX4_CLOCK_SIZE;
1719 EXPORT_SYMBOL_GPL(mlx4_get_internal_clock_params);
1721 static void unmap_internal_clock(struct mlx4_dev *dev)
1723 struct mlx4_priv *priv = mlx4_priv(dev);
1725 if (priv->clock_mapping)
1726 iounmap(priv->clock_mapping);
1729 static void mlx4_close_hca(struct mlx4_dev *dev)
1731 unmap_internal_clock(dev);
1733 if (mlx4_is_slave(dev))
1734 mlx4_slave_exit(dev);
1736 mlx4_CLOSE_HCA(dev, 0);
1737 mlx4_free_icms(dev);
1741 static void mlx4_close_fw(struct mlx4_dev *dev)
1743 if (!mlx4_is_slave(dev)) {
1745 mlx4_free_icm(dev, mlx4_priv(dev)->fw.fw_icm, 0);
1749 static int mlx4_comm_check_offline(struct mlx4_dev *dev)
1751 #define COMM_CHAN_OFFLINE_OFFSET 0x09
1756 struct mlx4_priv *priv = mlx4_priv(dev);
1758 end = msecs_to_jiffies(MLX4_COMM_OFFLINE_TIME_OUT) + jiffies;
1759 while (time_before(jiffies, end)) {
1760 comm_flags = swab32(readl((__iomem char *)priv->mfunc.comm +
1761 MLX4_COMM_CHAN_FLAGS));
1762 offline_bit = (comm_flags &
1763 (u32)(1 << COMM_CHAN_OFFLINE_OFFSET));
1767 /* If device removal has been requested,
1768 * do not continue retrying.
1770 if (dev->persist->interface_state &
1771 MLX4_INTERFACE_STATE_NOWAIT)
1774 /* There are cases as part of AER/Reset flow that PF needs
1775 * around 100 msec to load. We therefore sleep for 100 msec
1776 * to allow other tasks to make use of that CPU during this
1781 mlx4_err(dev, "Communication channel is offline.\n");
1785 static void mlx4_reset_vf_support(struct mlx4_dev *dev)
1787 #define COMM_CHAN_RST_OFFSET 0x1e
1789 struct mlx4_priv *priv = mlx4_priv(dev);
1793 comm_caps = swab32(readl((__iomem char *)priv->mfunc.comm +
1794 MLX4_COMM_CHAN_CAPS));
1795 comm_rst = (comm_caps & (u32)(1 << COMM_CHAN_RST_OFFSET));
1798 dev->caps.vf_caps |= MLX4_VF_CAP_FLAG_RESET;
1801 static int mlx4_init_slave(struct mlx4_dev *dev)
1803 struct mlx4_priv *priv = mlx4_priv(dev);
1804 u64 dma = (u64) priv->mfunc.vhcr_dma;
1805 int ret_from_reset = 0;
1807 u32 cmd_channel_ver;
1809 if (atomic_read(&pf_loading)) {
1810 mlx4_warn(dev, "PF is not ready - Deferring probe\n");
1811 return -EPROBE_DEFER;
1814 mutex_lock(&priv->cmd.slave_cmd_mutex);
1815 priv->cmd.max_cmds = 1;
1816 if (mlx4_comm_check_offline(dev)) {
1817 mlx4_err(dev, "PF is not responsive, skipping initialization\n");
1821 mlx4_reset_vf_support(dev);
1822 mlx4_warn(dev, "Sending reset\n");
1823 ret_from_reset = mlx4_comm_cmd(dev, MLX4_COMM_CMD_RESET, 0,
1824 MLX4_COMM_CMD_NA_OP, MLX4_COMM_TIME);
1825 /* if we are in the middle of flr the slave will try
1826 * NUM_OF_RESET_RETRIES times before leaving.*/
1827 if (ret_from_reset) {
1828 if (MLX4_DELAY_RESET_SLAVE == ret_from_reset) {
1829 mlx4_warn(dev, "slave is currently in the middle of FLR - Deferring probe\n");
1830 mutex_unlock(&priv->cmd.slave_cmd_mutex);
1831 return -EPROBE_DEFER;
1836 /* check the driver version - the slave I/F revision
1837 * must match the master's */
1838 slave_read = swab32(readl(&priv->mfunc.comm->slave_read));
1839 cmd_channel_ver = mlx4_comm_get_version();
1841 if (MLX4_COMM_GET_IF_REV(cmd_channel_ver) !=
1842 MLX4_COMM_GET_IF_REV(slave_read)) {
1843 mlx4_err(dev, "slave driver version is not supported by the master\n");
1847 mlx4_warn(dev, "Sending vhcr0\n");
1848 if (mlx4_comm_cmd(dev, MLX4_COMM_CMD_VHCR0, dma >> 48,
1849 MLX4_COMM_CMD_NA_OP, MLX4_COMM_TIME))
1851 if (mlx4_comm_cmd(dev, MLX4_COMM_CMD_VHCR1, dma >> 32,
1852 MLX4_COMM_CMD_NA_OP, MLX4_COMM_TIME))
1854 if (mlx4_comm_cmd(dev, MLX4_COMM_CMD_VHCR2, dma >> 16,
1855 MLX4_COMM_CMD_NA_OP, MLX4_COMM_TIME))
1857 if (mlx4_comm_cmd(dev, MLX4_COMM_CMD_VHCR_EN, dma,
1858 MLX4_COMM_CMD_NA_OP, MLX4_COMM_TIME))
1861 mutex_unlock(&priv->cmd.slave_cmd_mutex);
1865 mlx4_comm_cmd(dev, MLX4_COMM_CMD_RESET, 0, MLX4_COMM_CMD_NA_OP, 0);
1867 mutex_unlock(&priv->cmd.slave_cmd_mutex);
1871 static void mlx4_parav_master_pf_caps(struct mlx4_dev *dev)
1875 for (i = 1; i <= dev->caps.num_ports; i++) {
1876 if (dev->caps.port_type[i] == MLX4_PORT_TYPE_ETH)
1877 dev->caps.gid_table_len[i] =
1878 mlx4_get_slave_num_gids(dev, 0, i);
1880 dev->caps.gid_table_len[i] = 1;
1881 dev->caps.pkey_table_len[i] =
1882 dev->phys_caps.pkey_phys_table_len[i] - 1;
1886 static int choose_log_fs_mgm_entry_size(int qp_per_entry)
1888 int i = MLX4_MIN_MGM_LOG_ENTRY_SIZE;
1890 for (i = MLX4_MIN_MGM_LOG_ENTRY_SIZE; i <= MLX4_MAX_MGM_LOG_ENTRY_SIZE;
1892 if (qp_per_entry <= 4 * ((1 << i) / 16 - 2))
1896 return (i <= MLX4_MAX_MGM_LOG_ENTRY_SIZE) ? i : -1;
1899 static const char *dmfs_high_rate_steering_mode_str(int dmfs_high_steer_mode)
1901 switch (dmfs_high_steer_mode) {
1902 case MLX4_STEERING_DMFS_A0_DEFAULT:
1903 return "default performance";
1905 case MLX4_STEERING_DMFS_A0_DYNAMIC:
1906 return "dynamic hybrid mode";
1908 case MLX4_STEERING_DMFS_A0_STATIC:
1909 return "performance optimized for limited rule configuration (static)";
1911 case MLX4_STEERING_DMFS_A0_DISABLE:
1912 return "disabled performance optimized steering";
1914 case MLX4_STEERING_DMFS_A0_NOT_SUPPORTED:
1915 return "performance optimized steering not supported";
1918 return "Unrecognized mode";
1922 #define MLX4_DMFS_A0_STEERING (1UL << 2)
1924 static void choose_steering_mode(struct mlx4_dev *dev,
1925 struct mlx4_dev_cap *dev_cap)
1927 if (mlx4_log_num_mgm_entry_size <= 0) {
1928 if ((-mlx4_log_num_mgm_entry_size) & MLX4_DMFS_A0_STEERING) {
1929 if (dev->caps.dmfs_high_steer_mode ==
1930 MLX4_STEERING_DMFS_A0_NOT_SUPPORTED)
1931 mlx4_err(dev, "DMFS high rate mode not supported\n");
1933 dev->caps.dmfs_high_steer_mode =
1934 MLX4_STEERING_DMFS_A0_STATIC;
1938 if (mlx4_log_num_mgm_entry_size <= 0 &&
1939 dev_cap->flags2 & MLX4_DEV_CAP_FLAG2_FS_EN &&
1940 (!mlx4_is_mfunc(dev) ||
1941 (dev_cap->fs_max_num_qp_per_entry >=
1942 (dev->persist->num_vfs + 1))) &&
1943 choose_log_fs_mgm_entry_size(dev_cap->fs_max_num_qp_per_entry) >=
1944 MLX4_MIN_MGM_LOG_ENTRY_SIZE) {
1945 dev->oper_log_mgm_entry_size =
1946 choose_log_fs_mgm_entry_size(dev_cap->fs_max_num_qp_per_entry);
1947 dev->caps.steering_mode = MLX4_STEERING_MODE_DEVICE_MANAGED;
1948 dev->caps.num_qp_per_mgm = dev_cap->fs_max_num_qp_per_entry;
1949 dev->caps.fs_log_max_ucast_qp_range_size =
1950 dev_cap->fs_log_max_ucast_qp_range_size;
1952 if (dev->caps.dmfs_high_steer_mode !=
1953 MLX4_STEERING_DMFS_A0_NOT_SUPPORTED)
1954 dev->caps.dmfs_high_steer_mode = MLX4_STEERING_DMFS_A0_DISABLE;
1955 if (dev->caps.flags & MLX4_DEV_CAP_FLAG_VEP_UC_STEER &&
1956 dev->caps.flags & MLX4_DEV_CAP_FLAG_VEP_MC_STEER)
1957 dev->caps.steering_mode = MLX4_STEERING_MODE_B0;
1959 dev->caps.steering_mode = MLX4_STEERING_MODE_A0;
1961 if (dev->caps.flags & MLX4_DEV_CAP_FLAG_VEP_UC_STEER ||
1962 dev->caps.flags & MLX4_DEV_CAP_FLAG_VEP_MC_STEER)
1963 mlx4_warn(dev, "Must have both UC_STEER and MC_STEER flags set to use B0 steering - falling back to A0 steering mode\n");
1965 dev->oper_log_mgm_entry_size =
1966 mlx4_log_num_mgm_entry_size > 0 ?
1967 mlx4_log_num_mgm_entry_size :
1968 MLX4_DEFAULT_MGM_LOG_ENTRY_SIZE;
1969 dev->caps.num_qp_per_mgm = mlx4_get_qp_per_mgm(dev);
1971 mlx4_dbg(dev, "Steering mode is: %s, oper_log_mgm_entry_size = %d, modparam log_num_mgm_entry_size = %d\n",
1972 mlx4_steering_mode_str(dev->caps.steering_mode),
1973 dev->oper_log_mgm_entry_size,
1974 mlx4_log_num_mgm_entry_size);
1977 static void choose_tunnel_offload_mode(struct mlx4_dev *dev,
1978 struct mlx4_dev_cap *dev_cap)
1980 if (dev->caps.steering_mode == MLX4_STEERING_MODE_DEVICE_MANAGED &&
1981 dev_cap->flags2 & MLX4_DEV_CAP_FLAG2_VXLAN_OFFLOADS)
1982 dev->caps.tunnel_offload_mode = MLX4_TUNNEL_OFFLOAD_MODE_VXLAN;
1984 dev->caps.tunnel_offload_mode = MLX4_TUNNEL_OFFLOAD_MODE_NONE;
1986 mlx4_dbg(dev, "Tunneling offload mode is: %s\n", (dev->caps.tunnel_offload_mode
1987 == MLX4_TUNNEL_OFFLOAD_MODE_VXLAN) ? "vxlan" : "none");
1990 static int mlx4_validate_optimized_steering(struct mlx4_dev *dev)
1993 struct mlx4_port_cap port_cap;
1995 if (dev->caps.dmfs_high_steer_mode == MLX4_STEERING_DMFS_A0_NOT_SUPPORTED)
1998 for (i = 1; i <= dev->caps.num_ports; i++) {
1999 if (mlx4_dev_port(dev, i, &port_cap)) {
2001 "QUERY_DEV_CAP command failed, can't veify DMFS high rate steering.\n");
2002 } else if ((dev->caps.dmfs_high_steer_mode !=
2003 MLX4_STEERING_DMFS_A0_DEFAULT) &&
2004 (port_cap.dmfs_optimized_state ==
2005 !!(dev->caps.dmfs_high_steer_mode ==
2006 MLX4_STEERING_DMFS_A0_DISABLE))) {
2008 "DMFS high rate steer mode differ, driver requested %s but %s in FW.\n",
2009 dmfs_high_rate_steering_mode_str(
2010 dev->caps.dmfs_high_steer_mode),
2011 (port_cap.dmfs_optimized_state ?
2012 "enabled" : "disabled"));
2019 static int mlx4_init_fw(struct mlx4_dev *dev)
2021 struct mlx4_mod_stat_cfg mlx4_cfg;
2024 if (!mlx4_is_slave(dev)) {
2025 err = mlx4_QUERY_FW(dev);
2028 mlx4_info(dev, "non-primary physical function, skipping\n");
2030 mlx4_err(dev, "QUERY_FW command failed, aborting\n");
2034 err = mlx4_load_fw(dev);
2036 mlx4_err(dev, "Failed to start FW, aborting\n");
2040 mlx4_cfg.log_pg_sz_m = 1;
2041 mlx4_cfg.log_pg_sz = 0;
2042 err = mlx4_MOD_STAT_CFG(dev, &mlx4_cfg);
2044 mlx4_warn(dev, "Failed to override log_pg_sz parameter\n");
2050 static int mlx4_init_hca(struct mlx4_dev *dev)
2052 struct mlx4_priv *priv = mlx4_priv(dev);
2053 struct mlx4_adapter adapter;
2054 struct mlx4_dev_cap dev_cap;
2055 struct mlx4_profile profile;
2056 struct mlx4_init_hca_param init_hca;
2058 struct mlx4_config_dev_params params;
2061 if (!mlx4_is_slave(dev)) {
2062 err = mlx4_dev_cap(dev, &dev_cap);
2064 mlx4_err(dev, "QUERY_DEV_CAP command failed, aborting\n");
2068 choose_steering_mode(dev, &dev_cap);
2069 choose_tunnel_offload_mode(dev, &dev_cap);
2071 if (dev->caps.dmfs_high_steer_mode == MLX4_STEERING_DMFS_A0_STATIC &&
2072 mlx4_is_master(dev))
2073 dev->caps.function_caps |= MLX4_FUNC_CAP_DMFS_A0_STATIC;
2075 err = mlx4_get_phys_port_id(dev);
2077 mlx4_err(dev, "Fail to get physical port id\n");
2079 if (mlx4_is_master(dev))
2080 mlx4_parav_master_pf_caps(dev);
2082 if (mlx4_low_memory_profile()) {
2083 mlx4_info(dev, "Running from within kdump kernel. Using low memory profile\n");
2084 profile = low_mem_profile;
2086 profile = default_profile;
2088 if (dev->caps.steering_mode ==
2089 MLX4_STEERING_MODE_DEVICE_MANAGED)
2090 profile.num_mcg = MLX4_FS_NUM_MCG;
2092 icm_size = mlx4_make_profile(dev, &profile, &dev_cap,
2094 if ((long long) icm_size < 0) {
2099 dev->caps.max_fmr_maps = (1 << (32 - ilog2(dev->caps.num_mpts))) - 1;
2101 init_hca.log_uar_sz = ilog2(dev->caps.num_uars);
2102 init_hca.uar_page_sz = PAGE_SHIFT - 12;
2103 init_hca.mw_enabled = 0;
2104 if (dev->caps.flags & MLX4_DEV_CAP_FLAG_MEM_WINDOW ||
2105 dev->caps.bmme_flags & MLX4_BMME_FLAG_TYPE_2_WIN)
2106 init_hca.mw_enabled = INIT_HCA_TPT_MW_ENABLE;
2108 err = mlx4_init_icm(dev, &dev_cap, &init_hca, icm_size);
2112 err = mlx4_INIT_HCA(dev, &init_hca);
2114 mlx4_err(dev, "INIT_HCA command failed, aborting\n");
2118 if (dev_cap.flags2 & MLX4_DEV_CAP_FLAG2_SYS_EQS) {
2119 err = mlx4_query_func(dev, &dev_cap);
2121 mlx4_err(dev, "QUERY_FUNC command failed, aborting.\n");
2123 } else if (err & MLX4_QUERY_FUNC_NUM_SYS_EQS) {
2124 dev->caps.num_eqs = dev_cap.max_eqs;
2125 dev->caps.reserved_eqs = dev_cap.reserved_eqs;
2126 dev->caps.reserved_uars = dev_cap.reserved_uars;
2131 * If TS is supported by FW
2132 * read HCA frequency by QUERY_HCA command
2134 if (dev->caps.flags2 & MLX4_DEV_CAP_FLAG2_TS) {
2135 memset(&init_hca, 0, sizeof(init_hca));
2136 err = mlx4_QUERY_HCA(dev, &init_hca);
2138 mlx4_err(dev, "QUERY_HCA command failed, disable timestamp\n");
2139 dev->caps.flags2 &= ~MLX4_DEV_CAP_FLAG2_TS;
2141 dev->caps.hca_core_clock =
2142 init_hca.hca_core_clock;
2145 /* In case we got HCA frequency 0 - disable timestamping
2146 * to avoid dividing by zero
2148 if (!dev->caps.hca_core_clock) {
2149 dev->caps.flags2 &= ~MLX4_DEV_CAP_FLAG2_TS;
2151 "HCA frequency is 0 - timestamping is not supported\n");
2152 } else if (map_internal_clock(dev)) {
2154 * Map internal clock,
2155 * in case of failure disable timestamping
2157 dev->caps.flags2 &= ~MLX4_DEV_CAP_FLAG2_TS;
2158 mlx4_err(dev, "Failed to map internal clock. Timestamping is not supported\n");
2162 if (dev->caps.dmfs_high_steer_mode !=
2163 MLX4_STEERING_DMFS_A0_NOT_SUPPORTED) {
2164 if (mlx4_validate_optimized_steering(dev))
2165 mlx4_warn(dev, "Optimized steering validation failed\n");
2167 if (dev->caps.dmfs_high_steer_mode ==
2168 MLX4_STEERING_DMFS_A0_DISABLE) {
2169 dev->caps.dmfs_high_rate_qpn_base =
2170 dev->caps.reserved_qps_cnt[MLX4_QP_REGION_FW];
2171 dev->caps.dmfs_high_rate_qpn_range =
2172 MLX4_A0_STEERING_TABLE_SIZE;
2175 mlx4_dbg(dev, "DMFS high rate steer mode is: %s\n",
2176 dmfs_high_rate_steering_mode_str(
2177 dev->caps.dmfs_high_steer_mode));
2180 err = mlx4_init_slave(dev);
2182 if (err != -EPROBE_DEFER)
2183 mlx4_err(dev, "Failed to initialize slave\n");
2187 err = mlx4_slave_cap(dev);
2189 mlx4_err(dev, "Failed to obtain slave caps\n");
2194 if (map_bf_area(dev))
2195 mlx4_dbg(dev, "Failed to map blue flame area\n");
2197 /*Only the master set the ports, all the rest got it from it.*/
2198 if (!mlx4_is_slave(dev))
2199 mlx4_set_port_mask(dev);
2201 err = mlx4_QUERY_ADAPTER(dev, &adapter);
2203 mlx4_err(dev, "QUERY_ADAPTER command failed, aborting\n");
2207 /* Query CONFIG_DEV parameters */
2208 err = mlx4_config_dev_retrieval(dev, ¶ms);
2209 if (err && err != -ENOTSUPP) {
2210 mlx4_err(dev, "Failed to query CONFIG_DEV parameters\n");
2212 dev->caps.rx_checksum_flags_port[1] = params.rx_csum_flags_port_1;
2213 dev->caps.rx_checksum_flags_port[2] = params.rx_csum_flags_port_2;
2215 priv->eq_table.inta_pin = adapter.inta_pin;
2216 memcpy(dev->board_id, adapter.board_id, sizeof dev->board_id);
2221 unmap_internal_clock(dev);
2224 if (mlx4_is_slave(dev)) {
2225 kfree(dev->caps.qp0_qkey);
2226 kfree(dev->caps.qp0_tunnel);
2227 kfree(dev->caps.qp0_proxy);
2228 kfree(dev->caps.qp1_tunnel);
2229 kfree(dev->caps.qp1_proxy);
2233 if (mlx4_is_slave(dev))
2234 mlx4_slave_exit(dev);
2236 mlx4_CLOSE_HCA(dev, 0);
2239 if (!mlx4_is_slave(dev))
2240 mlx4_free_icms(dev);
2245 static int mlx4_init_counters_table(struct mlx4_dev *dev)
2247 struct mlx4_priv *priv = mlx4_priv(dev);
2250 if (!(dev->caps.flags & MLX4_DEV_CAP_FLAG_COUNTERS))
2253 if (!dev->caps.max_counters)
2256 nent_pow2 = roundup_pow_of_two(dev->caps.max_counters);
2257 /* reserve last counter index for sink counter */
2258 return mlx4_bitmap_init(&priv->counters_bitmap, nent_pow2,
2260 nent_pow2 - dev->caps.max_counters + 1);
2263 static void mlx4_cleanup_counters_table(struct mlx4_dev *dev)
2265 if (!(dev->caps.flags & MLX4_DEV_CAP_FLAG_COUNTERS))
2268 if (!dev->caps.max_counters)
2271 mlx4_bitmap_cleanup(&mlx4_priv(dev)->counters_bitmap);
2274 static void mlx4_cleanup_default_counters(struct mlx4_dev *dev)
2276 struct mlx4_priv *priv = mlx4_priv(dev);
2279 for (port = 0; port < dev->caps.num_ports; port++)
2280 if (priv->def_counter[port] != -1)
2281 mlx4_counter_free(dev, priv->def_counter[port]);
2284 static int mlx4_allocate_default_counters(struct mlx4_dev *dev)
2286 struct mlx4_priv *priv = mlx4_priv(dev);
2290 for (port = 0; port < dev->caps.num_ports; port++)
2291 priv->def_counter[port] = -1;
2293 for (port = 0; port < dev->caps.num_ports; port++) {
2294 err = mlx4_counter_alloc(dev, &idx);
2296 if (!err || err == -ENOSPC) {
2297 priv->def_counter[port] = idx;
2299 } else if (err == -ENOENT) {
2302 } else if (mlx4_is_slave(dev) && err == -EINVAL) {
2303 priv->def_counter[port] = MLX4_SINK_COUNTER_INDEX(dev);
2304 mlx4_warn(dev, "can't allocate counter from old PF driver, using index %d\n",
2305 MLX4_SINK_COUNTER_INDEX(dev));
2308 mlx4_err(dev, "%s: failed to allocate default counter port %d err %d\n",
2309 __func__, port + 1, err);
2310 mlx4_cleanup_default_counters(dev);
2314 mlx4_dbg(dev, "%s: default counter index %d for port %d\n",
2315 __func__, priv->def_counter[port], port + 1);
2321 int __mlx4_counter_alloc(struct mlx4_dev *dev, u32 *idx)
2323 struct mlx4_priv *priv = mlx4_priv(dev);
2325 if (!(dev->caps.flags & MLX4_DEV_CAP_FLAG_COUNTERS))
2328 *idx = mlx4_bitmap_alloc(&priv->counters_bitmap);
2330 *idx = MLX4_SINK_COUNTER_INDEX(dev);
2337 int mlx4_counter_alloc(struct mlx4_dev *dev, u32 *idx)
2342 if (mlx4_is_mfunc(dev)) {
2343 err = mlx4_cmd_imm(dev, 0, &out_param, RES_COUNTER,
2344 RES_OP_RESERVE, MLX4_CMD_ALLOC_RES,
2345 MLX4_CMD_TIME_CLASS_A, MLX4_CMD_WRAPPED);
2347 *idx = get_param_l(&out_param);
2348 if (WARN_ON(err == -ENOSPC))
2352 return __mlx4_counter_alloc(dev, idx);
2354 EXPORT_SYMBOL_GPL(mlx4_counter_alloc);
2356 static int __mlx4_clear_if_stat(struct mlx4_dev *dev,
2359 struct mlx4_cmd_mailbox *if_stat_mailbox;
2361 u32 if_stat_in_mod = (counter_index & 0xff) | MLX4_QUERY_IF_STAT_RESET;
2363 if_stat_mailbox = mlx4_alloc_cmd_mailbox(dev);
2364 if (IS_ERR(if_stat_mailbox))
2365 return PTR_ERR(if_stat_mailbox);
2367 err = mlx4_cmd_box(dev, 0, if_stat_mailbox->dma, if_stat_in_mod, 0,
2368 MLX4_CMD_QUERY_IF_STAT, MLX4_CMD_TIME_CLASS_C,
2371 mlx4_free_cmd_mailbox(dev, if_stat_mailbox);
2375 void __mlx4_counter_free(struct mlx4_dev *dev, u32 idx)
2377 if (!(dev->caps.flags & MLX4_DEV_CAP_FLAG_COUNTERS))
2380 if (idx == MLX4_SINK_COUNTER_INDEX(dev))
2383 __mlx4_clear_if_stat(dev, idx);
2385 mlx4_bitmap_free(&mlx4_priv(dev)->counters_bitmap, idx, MLX4_USE_RR);
2389 void mlx4_counter_free(struct mlx4_dev *dev, u32 idx)
2393 if (mlx4_is_mfunc(dev)) {
2394 set_param_l(&in_param, idx);
2395 mlx4_cmd(dev, in_param, RES_COUNTER, RES_OP_RESERVE,
2396 MLX4_CMD_FREE_RES, MLX4_CMD_TIME_CLASS_A,
2400 __mlx4_counter_free(dev, idx);
2402 EXPORT_SYMBOL_GPL(mlx4_counter_free);
2404 int mlx4_get_default_counter_index(struct mlx4_dev *dev, int port)
2406 struct mlx4_priv *priv = mlx4_priv(dev);
2408 return priv->def_counter[port - 1];
2410 EXPORT_SYMBOL_GPL(mlx4_get_default_counter_index);
2412 void mlx4_set_admin_guid(struct mlx4_dev *dev, __be64 guid, int entry, int port)
2414 struct mlx4_priv *priv = mlx4_priv(dev);
2416 priv->mfunc.master.vf_admin[entry].vport[port].guid = guid;
2418 EXPORT_SYMBOL_GPL(mlx4_set_admin_guid);
2420 __be64 mlx4_get_admin_guid(struct mlx4_dev *dev, int entry, int port)
2422 struct mlx4_priv *priv = mlx4_priv(dev);
2424 return priv->mfunc.master.vf_admin[entry].vport[port].guid;
2426 EXPORT_SYMBOL_GPL(mlx4_get_admin_guid);
2428 void mlx4_set_random_admin_guid(struct mlx4_dev *dev, int entry, int port)
2430 struct mlx4_priv *priv = mlx4_priv(dev);
2437 get_random_bytes((char *)&guid, sizeof(guid));
2438 guid &= ~(cpu_to_be64(1ULL << 56));
2439 guid |= cpu_to_be64(1ULL << 57);
2440 priv->mfunc.master.vf_admin[entry].vport[port].guid = guid;
2443 static int mlx4_setup_hca(struct mlx4_dev *dev)
2445 struct mlx4_priv *priv = mlx4_priv(dev);
2448 __be32 ib_port_default_caps;
2450 err = mlx4_init_uar_table(dev);
2452 mlx4_err(dev, "Failed to initialize user access region table, aborting\n");
2456 err = mlx4_uar_alloc(dev, &priv->driver_uar);
2458 mlx4_err(dev, "Failed to allocate driver access region, aborting\n");
2459 goto err_uar_table_free;
2462 priv->kar = ioremap((phys_addr_t) priv->driver_uar.pfn << PAGE_SHIFT, PAGE_SIZE);
2464 mlx4_err(dev, "Couldn't map kernel access region, aborting\n");
2469 err = mlx4_init_pd_table(dev);
2471 mlx4_err(dev, "Failed to initialize protection domain table, aborting\n");
2475 err = mlx4_init_xrcd_table(dev);
2477 mlx4_err(dev, "Failed to initialize reliable connection domain table, aborting\n");
2478 goto err_pd_table_free;
2481 err = mlx4_init_mr_table(dev);
2483 mlx4_err(dev, "Failed to initialize memory region table, aborting\n");
2484 goto err_xrcd_table_free;
2487 if (!mlx4_is_slave(dev)) {
2488 err = mlx4_init_mcg_table(dev);
2490 mlx4_err(dev, "Failed to initialize multicast group table, aborting\n");
2491 goto err_mr_table_free;
2493 err = mlx4_config_mad_demux(dev);
2495 mlx4_err(dev, "Failed in config_mad_demux, aborting\n");
2496 goto err_mcg_table_free;
2500 err = mlx4_init_eq_table(dev);
2502 mlx4_err(dev, "Failed to initialize event queue table, aborting\n");
2503 goto err_mcg_table_free;
2506 err = mlx4_cmd_use_events(dev);
2508 mlx4_err(dev, "Failed to switch to event-driven firmware commands, aborting\n");
2509 goto err_eq_table_free;
2512 err = mlx4_NOP(dev);
2514 if (dev->flags & MLX4_FLAG_MSI_X) {
2515 mlx4_warn(dev, "NOP command failed to generate MSI-X interrupt IRQ %d)\n",
2516 priv->eq_table.eq[MLX4_EQ_ASYNC].irq);
2517 mlx4_warn(dev, "Trying again without MSI-X\n");
2519 mlx4_err(dev, "NOP command failed to generate interrupt (IRQ %d), aborting\n",
2520 priv->eq_table.eq[MLX4_EQ_ASYNC].irq);
2521 mlx4_err(dev, "BIOS or ACPI interrupt routing problem?\n");
2527 mlx4_dbg(dev, "NOP command IRQ test passed\n");
2529 err = mlx4_init_cq_table(dev);
2531 mlx4_err(dev, "Failed to initialize completion queue table, aborting\n");
2535 err = mlx4_init_srq_table(dev);
2537 mlx4_err(dev, "Failed to initialize shared receive queue table, aborting\n");
2538 goto err_cq_table_free;
2541 err = mlx4_init_qp_table(dev);
2543 mlx4_err(dev, "Failed to initialize queue pair table, aborting\n");
2544 goto err_srq_table_free;
2547 if (!mlx4_is_slave(dev)) {
2548 err = mlx4_init_counters_table(dev);
2549 if (err && err != -ENOENT) {
2550 mlx4_err(dev, "Failed to initialize counters table, aborting\n");
2551 goto err_qp_table_free;
2555 err = mlx4_allocate_default_counters(dev);
2557 mlx4_err(dev, "Failed to allocate default counters, aborting\n");
2558 goto err_counters_table_free;
2561 if (!mlx4_is_slave(dev)) {
2562 for (port = 1; port <= dev->caps.num_ports; port++) {
2563 ib_port_default_caps = 0;
2564 err = mlx4_get_port_ib_caps(dev, port,
2565 &ib_port_default_caps);
2567 mlx4_warn(dev, "failed to get port %d default ib capabilities (%d). Continuing with caps = 0\n",
2569 dev->caps.ib_port_def_cap[port] = ib_port_default_caps;
2571 /* initialize per-slave default ib port capabilities */
2572 if (mlx4_is_master(dev)) {
2574 for (i = 0; i < dev->num_slaves; i++) {
2575 if (i == mlx4_master_func_num(dev))
2577 priv->mfunc.master.slave_state[i].ib_cap_mask[port] =
2578 ib_port_default_caps;
2582 if (mlx4_is_mfunc(dev))
2583 dev->caps.port_ib_mtu[port] = IB_MTU_2048;
2585 dev->caps.port_ib_mtu[port] = IB_MTU_4096;
2587 err = mlx4_SET_PORT(dev, port, mlx4_is_master(dev) ?
2588 dev->caps.pkey_table_len[port] : -1);
2590 mlx4_err(dev, "Failed to set port %d, aborting\n",
2592 goto err_default_countes_free;
2599 err_default_countes_free:
2600 mlx4_cleanup_default_counters(dev);
2602 err_counters_table_free:
2603 if (!mlx4_is_slave(dev))
2604 mlx4_cleanup_counters_table(dev);
2607 mlx4_cleanup_qp_table(dev);
2610 mlx4_cleanup_srq_table(dev);
2613 mlx4_cleanup_cq_table(dev);
2616 mlx4_cmd_use_polling(dev);
2619 mlx4_cleanup_eq_table(dev);
2622 if (!mlx4_is_slave(dev))
2623 mlx4_cleanup_mcg_table(dev);
2626 mlx4_cleanup_mr_table(dev);
2628 err_xrcd_table_free:
2629 mlx4_cleanup_xrcd_table(dev);
2632 mlx4_cleanup_pd_table(dev);
2638 mlx4_uar_free(dev, &priv->driver_uar);
2641 mlx4_cleanup_uar_table(dev);
2645 static int mlx4_init_affinity_hint(struct mlx4_dev *dev, int port, int eqn)
2647 int requested_cpu = 0;
2648 struct mlx4_priv *priv = mlx4_priv(dev);
2653 if (eqn > dev->caps.num_comp_vectors)
2656 for (i = 1; i < port; i++)
2657 off += mlx4_get_eqs_per_port(dev, i);
2659 requested_cpu = eqn - off - !!(eqn > MLX4_EQ_ASYNC);
2661 /* Meaning EQs are shared, and this call comes from the second port */
2662 if (requested_cpu < 0)
2665 eq = &priv->eq_table.eq[eqn];
2667 if (!zalloc_cpumask_var(&eq->affinity_mask, GFP_KERNEL))
2670 cpumask_set_cpu(requested_cpu, eq->affinity_mask);
2675 static void mlx4_enable_msi_x(struct mlx4_dev *dev)
2677 struct mlx4_priv *priv = mlx4_priv(dev);
2678 struct msix_entry *entries;
2683 int nreq = dev->caps.num_ports * num_online_cpus() + 1;
2685 nreq = min_t(int, dev->caps.num_eqs - dev->caps.reserved_eqs,
2687 if (nreq > MAX_MSIX)
2690 entries = kcalloc(nreq, sizeof *entries, GFP_KERNEL);
2694 for (i = 0; i < nreq; ++i)
2695 entries[i].entry = i;
2697 nreq = pci_enable_msix_range(dev->persist->pdev, entries, 2,
2700 if (nreq < 0 || nreq < MLX4_EQ_ASYNC) {
2704 /* 1 is reserved for events (asyncrounous EQ) */
2705 dev->caps.num_comp_vectors = nreq - 1;
2707 priv->eq_table.eq[MLX4_EQ_ASYNC].irq = entries[0].vector;
2708 bitmap_zero(priv->eq_table.eq[MLX4_EQ_ASYNC].actv_ports.ports,
2709 dev->caps.num_ports);
2711 for (i = 0; i < dev->caps.num_comp_vectors + 1; i++) {
2712 if (i == MLX4_EQ_ASYNC)
2715 priv->eq_table.eq[i].irq =
2716 entries[i + 1 - !!(i > MLX4_EQ_ASYNC)].vector;
2718 if (MLX4_IS_LEGACY_EQ_MODE(dev->caps)) {
2719 bitmap_fill(priv->eq_table.eq[i].actv_ports.ports,
2720 dev->caps.num_ports);
2721 /* We don't set affinity hint when there
2726 priv->eq_table.eq[i].actv_ports.ports);
2727 if (mlx4_init_affinity_hint(dev, port + 1, i))
2728 mlx4_warn(dev, "Couldn't init hint cpumask for EQ %d\n",
2731 /* We divide the Eqs evenly between the two ports.
2732 * (dev->caps.num_comp_vectors / dev->caps.num_ports)
2733 * refers to the number of Eqs per port
2734 * (i.e eqs_per_port). Theoretically, we would like to
2735 * write something like (i + 1) % eqs_per_port == 0.
2736 * However, since there's an asynchronous Eq, we have
2737 * to skip over it by comparing this condition to
2738 * !!((i + 1) > MLX4_EQ_ASYNC).
2740 if ((dev->caps.num_comp_vectors > dev->caps.num_ports) &&
2742 (dev->caps.num_comp_vectors / dev->caps.num_ports)) ==
2743 !!((i + 1) > MLX4_EQ_ASYNC))
2744 /* If dev->caps.num_comp_vectors < dev->caps.num_ports,
2745 * everything is shared anyway.
2750 dev->flags |= MLX4_FLAG_MSI_X;
2757 dev->caps.num_comp_vectors = 1;
2759 BUG_ON(MLX4_EQ_ASYNC >= 2);
2760 for (i = 0; i < 2; ++i) {
2761 priv->eq_table.eq[i].irq = dev->persist->pdev->irq;
2762 if (i != MLX4_EQ_ASYNC) {
2763 bitmap_fill(priv->eq_table.eq[i].actv_ports.ports,
2764 dev->caps.num_ports);
2769 static int mlx4_init_port_info(struct mlx4_dev *dev, int port)
2771 struct mlx4_port_info *info = &mlx4_priv(dev)->port[port];
2776 if (!mlx4_is_slave(dev)) {
2777 mlx4_init_mac_table(dev, &info->mac_table);
2778 mlx4_init_vlan_table(dev, &info->vlan_table);
2779 mlx4_init_roce_gid_table(dev, &info->gid_table);
2780 info->base_qpn = mlx4_get_base_qpn(dev, port);
2783 sprintf(info->dev_name, "mlx4_port%d", port);
2784 info->port_attr.attr.name = info->dev_name;
2785 if (mlx4_is_mfunc(dev))
2786 info->port_attr.attr.mode = S_IRUGO;
2788 info->port_attr.attr.mode = S_IRUGO | S_IWUSR;
2789 info->port_attr.store = set_port_type;
2791 info->port_attr.show = show_port_type;
2792 sysfs_attr_init(&info->port_attr.attr);
2794 err = device_create_file(&dev->persist->pdev->dev, &info->port_attr);
2796 mlx4_err(dev, "Failed to create file for port %d\n", port);
2800 sprintf(info->dev_mtu_name, "mlx4_port%d_mtu", port);
2801 info->port_mtu_attr.attr.name = info->dev_mtu_name;
2802 if (mlx4_is_mfunc(dev))
2803 info->port_mtu_attr.attr.mode = S_IRUGO;
2805 info->port_mtu_attr.attr.mode = S_IRUGO | S_IWUSR;
2806 info->port_mtu_attr.store = set_port_ib_mtu;
2808 info->port_mtu_attr.show = show_port_ib_mtu;
2809 sysfs_attr_init(&info->port_mtu_attr.attr);
2811 err = device_create_file(&dev->persist->pdev->dev,
2812 &info->port_mtu_attr);
2814 mlx4_err(dev, "Failed to create mtu file for port %d\n", port);
2815 device_remove_file(&info->dev->persist->pdev->dev,
2823 static void mlx4_cleanup_port_info(struct mlx4_port_info *info)
2828 device_remove_file(&info->dev->persist->pdev->dev, &info->port_attr);
2829 device_remove_file(&info->dev->persist->pdev->dev,
2830 &info->port_mtu_attr);
2831 #ifdef CONFIG_RFS_ACCEL
2832 free_irq_cpu_rmap(info->rmap);
2837 static int mlx4_init_steering(struct mlx4_dev *dev)
2839 struct mlx4_priv *priv = mlx4_priv(dev);
2840 int num_entries = dev->caps.num_ports;
2843 priv->steer = kzalloc(sizeof(struct mlx4_steer) * num_entries, GFP_KERNEL);
2847 for (i = 0; i < num_entries; i++)
2848 for (j = 0; j < MLX4_NUM_STEERS; j++) {
2849 INIT_LIST_HEAD(&priv->steer[i].promisc_qps[j]);
2850 INIT_LIST_HEAD(&priv->steer[i].steer_entries[j]);
2855 static void mlx4_clear_steering(struct mlx4_dev *dev)
2857 struct mlx4_priv *priv = mlx4_priv(dev);
2858 struct mlx4_steer_index *entry, *tmp_entry;
2859 struct mlx4_promisc_qp *pqp, *tmp_pqp;
2860 int num_entries = dev->caps.num_ports;
2863 for (i = 0; i < num_entries; i++) {
2864 for (j = 0; j < MLX4_NUM_STEERS; j++) {
2865 list_for_each_entry_safe(pqp, tmp_pqp,
2866 &priv->steer[i].promisc_qps[j],
2868 list_del(&pqp->list);
2871 list_for_each_entry_safe(entry, tmp_entry,
2872 &priv->steer[i].steer_entries[j],
2874 list_del(&entry->list);
2875 list_for_each_entry_safe(pqp, tmp_pqp,
2878 list_del(&pqp->list);
2888 static int extended_func_num(struct pci_dev *pdev)
2890 return PCI_SLOT(pdev->devfn) * 8 + PCI_FUNC(pdev->devfn);
2893 #define MLX4_OWNER_BASE 0x8069c
2894 #define MLX4_OWNER_SIZE 4
2896 static int mlx4_get_ownership(struct mlx4_dev *dev)
2898 void __iomem *owner;
2901 if (pci_channel_offline(dev->persist->pdev))
2904 owner = ioremap(pci_resource_start(dev->persist->pdev, 0) +
2908 mlx4_err(dev, "Failed to obtain ownership bit\n");
2917 static void mlx4_free_ownership(struct mlx4_dev *dev)
2919 void __iomem *owner;
2921 if (pci_channel_offline(dev->persist->pdev))
2924 owner = ioremap(pci_resource_start(dev->persist->pdev, 0) +
2928 mlx4_err(dev, "Failed to obtain ownership bit\n");
2936 #define SRIOV_VALID_STATE(flags) (!!((flags) & MLX4_FLAG_SRIOV) ==\
2937 !!((flags) & MLX4_FLAG_MASTER))
2939 static u64 mlx4_enable_sriov(struct mlx4_dev *dev, struct pci_dev *pdev,
2940 u8 total_vfs, int existing_vfs, int reset_flow)
2942 u64 dev_flags = dev->flags;
2944 int fw_enabled_sriov_vfs = min(pci_sriov_get_totalvfs(pdev),
2948 dev->dev_vfs = kcalloc(total_vfs, sizeof(*dev->dev_vfs),
2955 atomic_inc(&pf_loading);
2956 if (dev->flags & MLX4_FLAG_SRIOV) {
2957 if (existing_vfs != total_vfs) {
2958 mlx4_err(dev, "SR-IOV was already enabled, but with num_vfs (%d) different than requested (%d)\n",
2959 existing_vfs, total_vfs);
2960 total_vfs = existing_vfs;
2964 dev->dev_vfs = kzalloc(total_vfs * sizeof(*dev->dev_vfs), GFP_KERNEL);
2965 if (NULL == dev->dev_vfs) {
2966 mlx4_err(dev, "Failed to allocate memory for VFs\n");
2970 if (!(dev->flags & MLX4_FLAG_SRIOV)) {
2971 if (total_vfs > fw_enabled_sriov_vfs) {
2972 mlx4_err(dev, "requested vfs (%d) > available vfs (%d). Continuing without SR_IOV\n",
2973 total_vfs, fw_enabled_sriov_vfs);
2977 mlx4_warn(dev, "Enabling SR-IOV with %d VFs\n", total_vfs);
2978 err = pci_enable_sriov(pdev, total_vfs);
2981 mlx4_err(dev, "Failed to enable SR-IOV, continuing without SR-IOV (err = %d)\n",
2985 mlx4_warn(dev, "Running in master mode\n");
2986 dev_flags |= MLX4_FLAG_SRIOV |
2988 dev_flags &= ~MLX4_FLAG_SLAVE;
2989 dev->persist->num_vfs = total_vfs;
2994 atomic_dec(&pf_loading);
2996 dev->persist->num_vfs = 0;
2997 kfree(dev->dev_vfs);
2998 dev->dev_vfs = NULL;
2999 return dev_flags & ~MLX4_FLAG_MASTER;
3003 MLX4_DEV_CAP_CHECK_NUM_VFS_ABOVE_64 = -1,
3006 static int mlx4_check_dev_cap(struct mlx4_dev *dev, struct mlx4_dev_cap *dev_cap,
3009 int requested_vfs = nvfs[0] + nvfs[1] + nvfs[2];
3010 /* Checking for 64 VFs as a limitation of CX2 */
3011 if (!(dev_cap->flags2 & MLX4_DEV_CAP_FLAG2_80_VFS) &&
3012 requested_vfs >= 64) {
3013 mlx4_err(dev, "Requested %d VFs, but FW does not support more than 64\n",
3015 return MLX4_DEV_CAP_CHECK_NUM_VFS_ABOVE_64;
3020 static int mlx4_load_one(struct pci_dev *pdev, int pci_dev_data,
3021 int total_vfs, int *nvfs, struct mlx4_priv *priv,
3024 struct mlx4_dev *dev;
3029 struct mlx4_dev_cap *dev_cap = NULL;
3030 int existing_vfs = 0;
3034 INIT_LIST_HEAD(&priv->ctx_list);
3035 spin_lock_init(&priv->ctx_lock);
3037 mutex_init(&priv->port_mutex);
3038 mutex_init(&priv->bond_mutex);
3040 INIT_LIST_HEAD(&priv->pgdir_list);
3041 mutex_init(&priv->pgdir_mutex);
3043 INIT_LIST_HEAD(&priv->bf_list);
3044 mutex_init(&priv->bf_mutex);
3046 dev->rev_id = pdev->revision;
3047 dev->numa_node = dev_to_node(&pdev->dev);
3049 /* Detect if this device is a virtual function */
3050 if (pci_dev_data & MLX4_PCI_DEV_IS_VF) {
3051 mlx4_warn(dev, "Detected virtual function - running in slave mode\n");
3052 dev->flags |= MLX4_FLAG_SLAVE;
3054 /* We reset the device and enable SRIOV only for physical
3055 * devices. Try to claim ownership on the device;
3056 * if already taken, skip -- do not allow multiple PFs */
3057 err = mlx4_get_ownership(dev);
3062 mlx4_warn(dev, "Multiple PFs not yet supported - Skipping PF\n");
3067 atomic_set(&priv->opreq_count, 0);
3068 INIT_WORK(&priv->opreq_task, mlx4_opreq_action);
3071 * Now reset the HCA before we touch the PCI capabilities or
3072 * attempt a firmware command, since a boot ROM may have left
3073 * the HCA in an undefined state.
3075 err = mlx4_reset(dev);
3077 mlx4_err(dev, "Failed to reset HCA, aborting\n");
3082 dev->flags = MLX4_FLAG_MASTER;
3083 existing_vfs = pci_num_vf(pdev);
3085 dev->flags |= MLX4_FLAG_SRIOV;
3086 dev->persist->num_vfs = total_vfs;
3090 /* on load remove any previous indication of internal error,
3093 dev->persist->state = MLX4_DEVICE_STATE_UP;
3096 err = mlx4_cmd_init(dev);
3098 mlx4_err(dev, "Failed to init command interface, aborting\n");
3102 /* In slave functions, the communication channel must be initialized
3103 * before posting commands. Also, init num_slaves before calling
3105 if (mlx4_is_mfunc(dev)) {
3106 if (mlx4_is_master(dev)) {
3107 dev->num_slaves = MLX4_MAX_NUM_SLAVES;
3110 dev->num_slaves = 0;
3111 err = mlx4_multi_func_init(dev);
3113 mlx4_err(dev, "Failed to init slave mfunc interface, aborting\n");
3119 err = mlx4_init_fw(dev);
3121 mlx4_err(dev, "Failed to init fw, aborting.\n");
3125 if (mlx4_is_master(dev)) {
3126 /* when we hit the goto slave_start below, dev_cap already initialized */
3128 dev_cap = kzalloc(sizeof(*dev_cap), GFP_KERNEL);
3135 err = mlx4_QUERY_DEV_CAP(dev, dev_cap);
3137 mlx4_err(dev, "QUERY_DEV_CAP command failed, aborting.\n");
3141 if (mlx4_check_dev_cap(dev, dev_cap, nvfs))
3144 if (!(dev_cap->flags2 & MLX4_DEV_CAP_FLAG2_SYS_EQS)) {
3145 u64 dev_flags = mlx4_enable_sriov(dev, pdev,
3151 mlx4_cmd_cleanup(dev, MLX4_CMD_CLEANUP_ALL);
3152 dev->flags = dev_flags;
3153 if (!SRIOV_VALID_STATE(dev->flags)) {
3154 mlx4_err(dev, "Invalid SRIOV state\n");
3157 err = mlx4_reset(dev);
3159 mlx4_err(dev, "Failed to reset HCA, aborting.\n");
3165 /* Legacy mode FW requires SRIOV to be enabled before
3166 * doing QUERY_DEV_CAP, since max_eq's value is different if
3169 memset(dev_cap, 0, sizeof(*dev_cap));
3170 err = mlx4_QUERY_DEV_CAP(dev, dev_cap);
3172 mlx4_err(dev, "QUERY_DEV_CAP command failed, aborting.\n");
3176 if (mlx4_check_dev_cap(dev, dev_cap, nvfs))
3181 err = mlx4_init_hca(dev);
3183 if (err == -EACCES) {
3184 /* Not primary Physical function
3185 * Running in slave mode */
3186 mlx4_cmd_cleanup(dev, MLX4_CMD_CLEANUP_ALL);
3187 /* We're not a PF */
3188 if (dev->flags & MLX4_FLAG_SRIOV) {
3190 pci_disable_sriov(pdev);
3191 if (mlx4_is_master(dev) && !reset_flow)
3192 atomic_dec(&pf_loading);
3193 dev->flags &= ~MLX4_FLAG_SRIOV;
3195 if (!mlx4_is_slave(dev))
3196 mlx4_free_ownership(dev);
3197 dev->flags |= MLX4_FLAG_SLAVE;
3198 dev->flags &= ~MLX4_FLAG_MASTER;
3204 if (mlx4_is_master(dev) && (dev_cap->flags2 & MLX4_DEV_CAP_FLAG2_SYS_EQS)) {
3205 u64 dev_flags = mlx4_enable_sriov(dev, pdev, total_vfs,
3206 existing_vfs, reset_flow);
3208 if ((dev->flags ^ dev_flags) & (MLX4_FLAG_MASTER | MLX4_FLAG_SLAVE)) {
3209 mlx4_cmd_cleanup(dev, MLX4_CMD_CLEANUP_VHCR);
3210 dev->flags = dev_flags;
3211 err = mlx4_cmd_init(dev);
3213 /* Only VHCR is cleaned up, so could still
3216 mlx4_err(dev, "Failed to init VHCR command interface, aborting\n");
3220 dev->flags = dev_flags;
3223 if (!SRIOV_VALID_STATE(dev->flags)) {
3224 mlx4_err(dev, "Invalid SRIOV state\n");
3230 /* check if the device is functioning at its maximum possible speed.
3231 * No return code for this call, just warn the user in case of PCI
3232 * express device capabilities are under-satisfied by the bus.
3234 if (!mlx4_is_slave(dev))
3235 mlx4_check_pcie_caps(dev);
3237 /* In master functions, the communication channel must be initialized
3238 * after obtaining its address from fw */
3239 if (mlx4_is_master(dev)) {
3240 if (dev->caps.num_ports < 2 &&
3244 "Error: Trying to configure VFs on port 2, but HCA has only %d physical ports\n",
3245 dev->caps.num_ports);
3248 memcpy(dev->persist->nvfs, nvfs, sizeof(dev->persist->nvfs));
3251 i < sizeof(dev->persist->nvfs)/
3252 sizeof(dev->persist->nvfs[0]); i++) {
3255 for (j = 0; j < dev->persist->nvfs[i]; ++sum, ++j) {
3256 dev->dev_vfs[sum].min_port = i < 2 ? i + 1 : 1;
3257 dev->dev_vfs[sum].n_ports = i < 2 ? 1 :
3258 dev->caps.num_ports;
3262 /* In master functions, the communication channel
3263 * must be initialized after obtaining its address from fw
3265 err = mlx4_multi_func_init(dev);
3267 mlx4_err(dev, "Failed to init master mfunc interface, aborting.\n");
3272 err = mlx4_alloc_eq_table(dev);
3274 goto err_master_mfunc;
3276 bitmap_zero(priv->msix_ctl.pool_bm, MAX_MSIX);
3277 mutex_init(&priv->msix_ctl.pool_lock);
3279 mlx4_enable_msi_x(dev);
3280 if ((mlx4_is_mfunc(dev)) &&
3281 !(dev->flags & MLX4_FLAG_MSI_X)) {
3283 mlx4_err(dev, "INTx is not supported in multi-function mode, aborting\n");
3287 if (!mlx4_is_slave(dev)) {
3288 err = mlx4_init_steering(dev);
3290 goto err_disable_msix;
3293 err = mlx4_setup_hca(dev);
3294 if (err == -EBUSY && (dev->flags & MLX4_FLAG_MSI_X) &&
3295 !mlx4_is_mfunc(dev)) {
3296 dev->flags &= ~MLX4_FLAG_MSI_X;
3297 dev->caps.num_comp_vectors = 1;
3298 pci_disable_msix(pdev);
3299 err = mlx4_setup_hca(dev);
3305 mlx4_init_quotas(dev);
3306 /* When PF resources are ready arm its comm channel to enable
3309 if (mlx4_is_master(dev)) {
3310 err = mlx4_ARM_COMM_CHANNEL(dev);
3312 mlx4_err(dev, " Failed to arm comm channel eq: %x\n",
3318 for (port = 1; port <= dev->caps.num_ports; port++) {
3319 err = mlx4_init_port_info(dev, port);
3324 priv->v2p.port1 = 1;
3325 priv->v2p.port2 = 2;
3327 err = mlx4_register_device(dev);
3331 mlx4_request_modules(dev);
3333 mlx4_sense_init(dev);
3334 mlx4_start_sense(dev);
3338 if (mlx4_is_master(dev) && dev->persist->num_vfs && !reset_flow)
3339 atomic_dec(&pf_loading);
3345 for (--port; port >= 1; --port)
3346 mlx4_cleanup_port_info(&priv->port[port]);
3348 mlx4_cleanup_default_counters(dev);
3349 if (!mlx4_is_slave(dev))
3350 mlx4_cleanup_counters_table(dev);
3351 mlx4_cleanup_qp_table(dev);
3352 mlx4_cleanup_srq_table(dev);
3353 mlx4_cleanup_cq_table(dev);
3354 mlx4_cmd_use_polling(dev);
3355 mlx4_cleanup_eq_table(dev);
3356 mlx4_cleanup_mcg_table(dev);
3357 mlx4_cleanup_mr_table(dev);
3358 mlx4_cleanup_xrcd_table(dev);
3359 mlx4_cleanup_pd_table(dev);
3360 mlx4_cleanup_uar_table(dev);
3363 if (!mlx4_is_slave(dev))
3364 mlx4_clear_steering(dev);
3367 if (dev->flags & MLX4_FLAG_MSI_X)
3368 pci_disable_msix(pdev);
3371 mlx4_free_eq_table(dev);
3374 if (mlx4_is_master(dev)) {
3375 mlx4_free_resource_tracker(dev, RES_TR_FREE_STRUCTS_ONLY);
3376 mlx4_multi_func_cleanup(dev);
3379 if (mlx4_is_slave(dev)) {
3380 kfree(dev->caps.qp0_qkey);
3381 kfree(dev->caps.qp0_tunnel);
3382 kfree(dev->caps.qp0_proxy);
3383 kfree(dev->caps.qp1_tunnel);
3384 kfree(dev->caps.qp1_proxy);
3388 mlx4_close_hca(dev);
3394 if (mlx4_is_slave(dev))
3395 mlx4_multi_func_cleanup(dev);
3398 mlx4_cmd_cleanup(dev, MLX4_CMD_CLEANUP_ALL);
3401 if (dev->flags & MLX4_FLAG_SRIOV && !existing_vfs) {
3402 pci_disable_sriov(pdev);
3403 dev->flags &= ~MLX4_FLAG_SRIOV;
3406 if (mlx4_is_master(dev) && dev->persist->num_vfs && !reset_flow)
3407 atomic_dec(&pf_loading);
3409 kfree(priv->dev.dev_vfs);
3411 if (!mlx4_is_slave(dev))
3412 mlx4_free_ownership(dev);
3418 static int __mlx4_init_one(struct pci_dev *pdev, int pci_dev_data,
3419 struct mlx4_priv *priv)
3422 int nvfs[MLX4_MAX_PORTS + 1] = {0, 0, 0};
3423 int prb_vf[MLX4_MAX_PORTS + 1] = {0, 0, 0};
3424 const int param_map[MLX4_MAX_PORTS + 1][MLX4_MAX_PORTS + 1] = {
3425 {2, 0, 0}, {0, 1, 2}, {0, 1, 2} };
3426 unsigned total_vfs = 0;
3429 pr_info(DRV_NAME ": Initializing %s\n", pci_name(pdev));
3431 err = pci_enable_device(pdev);
3433 dev_err(&pdev->dev, "Cannot enable PCI device, aborting\n");
3437 /* Due to requirement that all VFs and the PF are *guaranteed* 2 MACS
3438 * per port, we must limit the number of VFs to 63 (since their are
3441 for (i = 0; i < sizeof(nvfs)/sizeof(nvfs[0]) && i < num_vfs_argc;
3442 total_vfs += nvfs[param_map[num_vfs_argc - 1][i]], i++) {
3443 nvfs[param_map[num_vfs_argc - 1][i]] = num_vfs[i];
3445 dev_err(&pdev->dev, "num_vfs module parameter cannot be negative\n");
3447 goto err_disable_pdev;
3450 for (i = 0; i < sizeof(prb_vf)/sizeof(prb_vf[0]) && i < probe_vfs_argc;
3452 prb_vf[param_map[probe_vfs_argc - 1][i]] = probe_vf[i];
3453 if (prb_vf[i] < 0 || prb_vf[i] > nvfs[i]) {
3454 dev_err(&pdev->dev, "probe_vf module parameter cannot be negative or greater than num_vfs\n");
3456 goto err_disable_pdev;
3459 if (total_vfs > MLX4_MAX_NUM_VF) {
3461 "Requested more VF's (%d) than allowed by hw (%d)\n",
3462 total_vfs, MLX4_MAX_NUM_VF);
3464 goto err_disable_pdev;
3467 for (i = 0; i < MLX4_MAX_PORTS; i++) {
3468 if (nvfs[i] + nvfs[2] > MLX4_MAX_NUM_VF_P_PORT) {
3470 "Requested more VF's (%d) for port (%d) than allowed by driver (%d)\n",
3471 nvfs[i] + nvfs[2], i + 1,
3472 MLX4_MAX_NUM_VF_P_PORT);
3474 goto err_disable_pdev;
3478 /* Check for BARs. */
3479 if (!(pci_dev_data & MLX4_PCI_DEV_IS_VF) &&
3480 !(pci_resource_flags(pdev, 0) & IORESOURCE_MEM)) {
3481 dev_err(&pdev->dev, "Missing DCS, aborting (driver_data: 0x%x, pci_resource_flags(pdev, 0):0x%lx)\n",
3482 pci_dev_data, pci_resource_flags(pdev, 0));
3484 goto err_disable_pdev;
3486 if (!(pci_resource_flags(pdev, 2) & IORESOURCE_MEM)) {
3487 dev_err(&pdev->dev, "Missing UAR, aborting\n");
3489 goto err_disable_pdev;
3492 err = pci_request_regions(pdev, DRV_NAME);
3494 dev_err(&pdev->dev, "Couldn't get PCI resources, aborting\n");
3495 goto err_disable_pdev;
3498 pci_set_master(pdev);
3500 err = pci_set_dma_mask(pdev, DMA_BIT_MASK(64));
3502 dev_warn(&pdev->dev, "Warning: couldn't set 64-bit PCI DMA mask\n");
3503 err = pci_set_dma_mask(pdev, DMA_BIT_MASK(32));
3505 dev_err(&pdev->dev, "Can't set PCI DMA mask, aborting\n");
3506 goto err_release_regions;
3509 err = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64));
3511 dev_warn(&pdev->dev, "Warning: couldn't set 64-bit consistent PCI DMA mask\n");
3512 err = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(32));
3514 dev_err(&pdev->dev, "Can't set consistent PCI DMA mask, aborting\n");
3515 goto err_release_regions;
3519 /* Allow large DMA segments, up to the firmware limit of 1 GB */
3520 dma_set_max_seg_size(&pdev->dev, 1024 * 1024 * 1024);
3521 /* Detect if this device is a virtual function */
3522 if (pci_dev_data & MLX4_PCI_DEV_IS_VF) {
3523 /* When acting as pf, we normally skip vfs unless explicitly
3524 * requested to probe them.
3527 unsigned vfs_offset = 0;
3529 for (i = 0; i < sizeof(nvfs)/sizeof(nvfs[0]) &&
3530 vfs_offset + nvfs[i] < extended_func_num(pdev);
3531 vfs_offset += nvfs[i], i++)
3533 if (i == sizeof(nvfs)/sizeof(nvfs[0])) {
3535 goto err_release_regions;
3537 if ((extended_func_num(pdev) - vfs_offset)
3539 dev_warn(&pdev->dev, "Skipping virtual function:%d\n",
3540 extended_func_num(pdev));
3542 goto err_release_regions;
3547 err = mlx4_catas_init(&priv->dev);
3549 goto err_release_regions;
3551 err = mlx4_load_one(pdev, pci_dev_data, total_vfs, nvfs, priv, 0);
3558 mlx4_catas_end(&priv->dev);
3560 err_release_regions:
3561 pci_release_regions(pdev);
3564 pci_disable_device(pdev);
3565 pci_set_drvdata(pdev, NULL);
3569 static int mlx4_init_one(struct pci_dev *pdev, const struct pci_device_id *id)
3571 struct mlx4_priv *priv;
3572 struct mlx4_dev *dev;
3575 printk_once(KERN_INFO "%s", mlx4_version);
3577 priv = kzalloc(sizeof(*priv), GFP_KERNEL);
3582 dev->persist = kzalloc(sizeof(*dev->persist), GFP_KERNEL);
3583 if (!dev->persist) {
3587 dev->persist->pdev = pdev;
3588 dev->persist->dev = dev;
3589 pci_set_drvdata(pdev, dev->persist);
3590 priv->pci_dev_data = id->driver_data;
3591 mutex_init(&dev->persist->device_state_mutex);
3592 mutex_init(&dev->persist->interface_state_mutex);
3594 ret = __mlx4_init_one(pdev, id->driver_data, priv);
3596 kfree(dev->persist);
3599 pci_save_state(pdev);
3605 static void mlx4_clean_dev(struct mlx4_dev *dev)
3607 struct mlx4_dev_persistent *persist = dev->persist;
3608 struct mlx4_priv *priv = mlx4_priv(dev);
3609 unsigned long flags = (dev->flags & RESET_PERSIST_MASK_FLAGS);
3611 memset(priv, 0, sizeof(*priv));
3612 priv->dev.persist = persist;
3613 priv->dev.flags = flags;
3616 static void mlx4_unload_one(struct pci_dev *pdev)
3618 struct mlx4_dev_persistent *persist = pci_get_drvdata(pdev);
3619 struct mlx4_dev *dev = persist->dev;
3620 struct mlx4_priv *priv = mlx4_priv(dev);
3627 /* saving current ports type for further use */
3628 for (i = 0; i < dev->caps.num_ports; i++) {
3629 dev->persist->curr_port_type[i] = dev->caps.port_type[i + 1];
3630 dev->persist->curr_port_poss_type[i] = dev->caps.
3631 possible_type[i + 1];
3634 pci_dev_data = priv->pci_dev_data;
3636 mlx4_stop_sense(dev);
3637 mlx4_unregister_device(dev);
3639 for (p = 1; p <= dev->caps.num_ports; p++) {
3640 mlx4_cleanup_port_info(&priv->port[p]);
3641 mlx4_CLOSE_PORT(dev, p);
3644 if (mlx4_is_master(dev))
3645 mlx4_free_resource_tracker(dev,
3646 RES_TR_FREE_SLAVES_ONLY);
3648 mlx4_cleanup_default_counters(dev);
3649 if (!mlx4_is_slave(dev))
3650 mlx4_cleanup_counters_table(dev);
3651 mlx4_cleanup_qp_table(dev);
3652 mlx4_cleanup_srq_table(dev);
3653 mlx4_cleanup_cq_table(dev);
3654 mlx4_cmd_use_polling(dev);
3655 mlx4_cleanup_eq_table(dev);
3656 mlx4_cleanup_mcg_table(dev);
3657 mlx4_cleanup_mr_table(dev);
3658 mlx4_cleanup_xrcd_table(dev);
3659 mlx4_cleanup_pd_table(dev);
3661 if (mlx4_is_master(dev))
3662 mlx4_free_resource_tracker(dev,
3663 RES_TR_FREE_STRUCTS_ONLY);
3666 mlx4_uar_free(dev, &priv->driver_uar);
3667 mlx4_cleanup_uar_table(dev);
3668 if (!mlx4_is_slave(dev))
3669 mlx4_clear_steering(dev);
3670 mlx4_free_eq_table(dev);
3671 if (mlx4_is_master(dev))
3672 mlx4_multi_func_cleanup(dev);
3673 mlx4_close_hca(dev);
3675 if (mlx4_is_slave(dev))
3676 mlx4_multi_func_cleanup(dev);
3677 mlx4_cmd_cleanup(dev, MLX4_CMD_CLEANUP_ALL);
3679 if (dev->flags & MLX4_FLAG_MSI_X)
3680 pci_disable_msix(pdev);
3682 if (!mlx4_is_slave(dev))
3683 mlx4_free_ownership(dev);
3685 kfree(dev->caps.qp0_qkey);
3686 kfree(dev->caps.qp0_tunnel);
3687 kfree(dev->caps.qp0_proxy);
3688 kfree(dev->caps.qp1_tunnel);
3689 kfree(dev->caps.qp1_proxy);
3690 kfree(dev->dev_vfs);
3692 mlx4_clean_dev(dev);
3693 priv->pci_dev_data = pci_dev_data;
3697 static void mlx4_remove_one(struct pci_dev *pdev)
3699 struct mlx4_dev_persistent *persist = pci_get_drvdata(pdev);
3700 struct mlx4_dev *dev = persist->dev;
3701 struct mlx4_priv *priv = mlx4_priv(dev);
3704 if (mlx4_is_slave(dev))
3705 persist->interface_state |= MLX4_INTERFACE_STATE_NOWAIT;
3707 mutex_lock(&persist->interface_state_mutex);
3708 persist->interface_state |= MLX4_INTERFACE_STATE_DELETION;
3709 mutex_unlock(&persist->interface_state_mutex);
3711 /* Disabling SR-IOV is not allowed while there are active vf's */
3712 if (mlx4_is_master(dev) && dev->flags & MLX4_FLAG_SRIOV) {
3713 active_vfs = mlx4_how_many_lives_vf(dev);
3715 pr_warn("Removing PF when there are active VF's !!\n");
3716 pr_warn("Will not disable SR-IOV.\n");
3720 /* device marked to be under deletion running now without the lock
3721 * letting other tasks to be terminated
3723 if (persist->interface_state & MLX4_INTERFACE_STATE_UP)
3724 mlx4_unload_one(pdev);
3726 mlx4_info(dev, "%s: interface is down\n", __func__);
3727 mlx4_catas_end(dev);
3728 if (dev->flags & MLX4_FLAG_SRIOV && !active_vfs) {
3729 mlx4_warn(dev, "Disabling SR-IOV\n");
3730 pci_disable_sriov(pdev);
3733 pci_release_regions(pdev);
3734 pci_disable_device(pdev);
3735 kfree(dev->persist);
3737 pci_set_drvdata(pdev, NULL);
3740 static int restore_current_port_types(struct mlx4_dev *dev,
3741 enum mlx4_port_type *types,
3742 enum mlx4_port_type *poss_types)
3744 struct mlx4_priv *priv = mlx4_priv(dev);
3747 mlx4_stop_sense(dev);
3749 mutex_lock(&priv->port_mutex);
3750 for (i = 0; i < dev->caps.num_ports; i++)
3751 dev->caps.possible_type[i + 1] = poss_types[i];
3752 err = mlx4_change_port_types(dev, types);
3753 mlx4_start_sense(dev);
3754 mutex_unlock(&priv->port_mutex);
3759 int mlx4_restart_one(struct pci_dev *pdev)
3761 struct mlx4_dev_persistent *persist = pci_get_drvdata(pdev);
3762 struct mlx4_dev *dev = persist->dev;
3763 struct mlx4_priv *priv = mlx4_priv(dev);
3764 int nvfs[MLX4_MAX_PORTS + 1] = {0, 0, 0};
3765 int pci_dev_data, err, total_vfs;
3767 pci_dev_data = priv->pci_dev_data;
3768 total_vfs = dev->persist->num_vfs;
3769 memcpy(nvfs, dev->persist->nvfs, sizeof(dev->persist->nvfs));
3771 mlx4_unload_one(pdev);
3772 err = mlx4_load_one(pdev, pci_dev_data, total_vfs, nvfs, priv, 1);
3774 mlx4_err(dev, "%s: ERROR: mlx4_load_one failed, pci_name=%s, err=%d\n",
3775 __func__, pci_name(pdev), err);
3779 err = restore_current_port_types(dev, dev->persist->curr_port_type,
3780 dev->persist->curr_port_poss_type);
3782 mlx4_err(dev, "could not restore original port types (%d)\n",
3788 static const struct pci_device_id mlx4_pci_table[] = {
3789 /* MT25408 "Hermon" SDR */
3790 { PCI_VDEVICE(MELLANOX, 0x6340), MLX4_PCI_DEV_FORCE_SENSE_PORT },
3791 /* MT25408 "Hermon" DDR */
3792 { PCI_VDEVICE(MELLANOX, 0x634a), MLX4_PCI_DEV_FORCE_SENSE_PORT },
3793 /* MT25408 "Hermon" QDR */
3794 { PCI_VDEVICE(MELLANOX, 0x6354), MLX4_PCI_DEV_FORCE_SENSE_PORT },
3795 /* MT25408 "Hermon" DDR PCIe gen2 */
3796 { PCI_VDEVICE(MELLANOX, 0x6732), MLX4_PCI_DEV_FORCE_SENSE_PORT },
3797 /* MT25408 "Hermon" QDR PCIe gen2 */
3798 { PCI_VDEVICE(MELLANOX, 0x673c), MLX4_PCI_DEV_FORCE_SENSE_PORT },
3799 /* MT25408 "Hermon" EN 10GigE */
3800 { PCI_VDEVICE(MELLANOX, 0x6368), MLX4_PCI_DEV_FORCE_SENSE_PORT },
3801 /* MT25408 "Hermon" EN 10GigE PCIe gen2 */
3802 { PCI_VDEVICE(MELLANOX, 0x6750), MLX4_PCI_DEV_FORCE_SENSE_PORT },
3803 /* MT25458 ConnectX EN 10GBASE-T 10GigE */
3804 { PCI_VDEVICE(MELLANOX, 0x6372), MLX4_PCI_DEV_FORCE_SENSE_PORT },
3805 /* MT25458 ConnectX EN 10GBASE-T+Gen2 10GigE */
3806 { PCI_VDEVICE(MELLANOX, 0x675a), MLX4_PCI_DEV_FORCE_SENSE_PORT },
3807 /* MT26468 ConnectX EN 10GigE PCIe gen2*/
3808 { PCI_VDEVICE(MELLANOX, 0x6764), MLX4_PCI_DEV_FORCE_SENSE_PORT },
3809 /* MT26438 ConnectX EN 40GigE PCIe gen2 5GT/s */
3810 { PCI_VDEVICE(MELLANOX, 0x6746), MLX4_PCI_DEV_FORCE_SENSE_PORT },
3811 /* MT26478 ConnectX2 40GigE PCIe gen2 */
3812 { PCI_VDEVICE(MELLANOX, 0x676e), MLX4_PCI_DEV_FORCE_SENSE_PORT },
3813 /* MT25400 Family [ConnectX-2 Virtual Function] */
3814 { PCI_VDEVICE(MELLANOX, 0x1002), MLX4_PCI_DEV_IS_VF },
3815 /* MT27500 Family [ConnectX-3] */
3816 { PCI_VDEVICE(MELLANOX, 0x1003), 0 },
3817 /* MT27500 Family [ConnectX-3 Virtual Function] */
3818 { PCI_VDEVICE(MELLANOX, 0x1004), MLX4_PCI_DEV_IS_VF },
3819 { PCI_VDEVICE(MELLANOX, 0x1005), 0 }, /* MT27510 Family */
3820 { PCI_VDEVICE(MELLANOX, 0x1006), 0 }, /* MT27511 Family */
3821 { PCI_VDEVICE(MELLANOX, 0x1007), 0 }, /* MT27520 Family */
3822 { PCI_VDEVICE(MELLANOX, 0x1008), 0 }, /* MT27521 Family */
3823 { PCI_VDEVICE(MELLANOX, 0x1009), 0 }, /* MT27530 Family */
3824 { PCI_VDEVICE(MELLANOX, 0x100a), 0 }, /* MT27531 Family */
3825 { PCI_VDEVICE(MELLANOX, 0x100b), 0 }, /* MT27540 Family */
3826 { PCI_VDEVICE(MELLANOX, 0x100c), 0 }, /* MT27541 Family */
3827 { PCI_VDEVICE(MELLANOX, 0x100d), 0 }, /* MT27550 Family */
3828 { PCI_VDEVICE(MELLANOX, 0x100e), 0 }, /* MT27551 Family */
3829 { PCI_VDEVICE(MELLANOX, 0x100f), 0 }, /* MT27560 Family */
3830 { PCI_VDEVICE(MELLANOX, 0x1010), 0 }, /* MT27561 Family */
3834 MODULE_DEVICE_TABLE(pci, mlx4_pci_table);
3836 static pci_ers_result_t mlx4_pci_err_detected(struct pci_dev *pdev,
3837 pci_channel_state_t state)
3839 struct mlx4_dev_persistent *persist = pci_get_drvdata(pdev);
3841 mlx4_err(persist->dev, "mlx4_pci_err_detected was called\n");
3842 mlx4_enter_error_state(persist);
3844 mutex_lock(&persist->interface_state_mutex);
3845 if (persist->interface_state & MLX4_INTERFACE_STATE_UP)
3846 mlx4_unload_one(pdev);
3848 mutex_unlock(&persist->interface_state_mutex);
3849 if (state == pci_channel_io_perm_failure)
3850 return PCI_ERS_RESULT_DISCONNECT;
3852 pci_disable_device(pdev);
3853 return PCI_ERS_RESULT_NEED_RESET;
3856 static pci_ers_result_t mlx4_pci_slot_reset(struct pci_dev *pdev)
3858 struct mlx4_dev_persistent *persist = pci_get_drvdata(pdev);
3859 struct mlx4_dev *dev = persist->dev;
3862 mlx4_err(dev, "mlx4_pci_slot_reset was called\n");
3863 err = pci_enable_device(pdev);
3865 mlx4_err(dev, "Can not re-enable device, err=%d\n", err);
3866 return PCI_ERS_RESULT_DISCONNECT;
3869 pci_set_master(pdev);
3870 pci_restore_state(pdev);
3871 pci_save_state(pdev);
3872 return PCI_ERS_RESULT_RECOVERED;
3875 static void mlx4_pci_resume(struct pci_dev *pdev)
3877 struct mlx4_dev_persistent *persist = pci_get_drvdata(pdev);
3878 struct mlx4_dev *dev = persist->dev;
3879 struct mlx4_priv *priv = mlx4_priv(dev);
3880 int nvfs[MLX4_MAX_PORTS + 1] = {0, 0, 0};
3884 mlx4_err(dev, "%s was called\n", __func__);
3885 total_vfs = dev->persist->num_vfs;
3886 memcpy(nvfs, dev->persist->nvfs, sizeof(dev->persist->nvfs));
3888 mutex_lock(&persist->interface_state_mutex);
3889 if (!(persist->interface_state & MLX4_INTERFACE_STATE_UP)) {
3890 err = mlx4_load_one(pdev, priv->pci_dev_data, total_vfs, nvfs,
3893 mlx4_err(dev, "%s: mlx4_load_one failed, err=%d\n",
3898 err = restore_current_port_types(dev, dev->persist->
3899 curr_port_type, dev->persist->
3900 curr_port_poss_type);
3902 mlx4_err(dev, "could not restore original port types (%d)\n", err);
3905 mutex_unlock(&persist->interface_state_mutex);
3909 static void mlx4_shutdown(struct pci_dev *pdev)
3911 struct mlx4_dev_persistent *persist = pci_get_drvdata(pdev);
3913 mlx4_info(persist->dev, "mlx4_shutdown was called\n");
3914 mutex_lock(&persist->interface_state_mutex);
3915 if (persist->interface_state & MLX4_INTERFACE_STATE_UP)
3916 mlx4_unload_one(pdev);
3917 mutex_unlock(&persist->interface_state_mutex);
3920 static const struct pci_error_handlers mlx4_err_handler = {
3921 .error_detected = mlx4_pci_err_detected,
3922 .slot_reset = mlx4_pci_slot_reset,
3923 .resume = mlx4_pci_resume,
3926 static struct pci_driver mlx4_driver = {
3928 .id_table = mlx4_pci_table,
3929 .probe = mlx4_init_one,
3930 .shutdown = mlx4_shutdown,
3931 .remove = mlx4_remove_one,
3932 .err_handler = &mlx4_err_handler,
3935 static int __init mlx4_verify_params(void)
3937 if ((log_num_mac < 0) || (log_num_mac > 7)) {
3938 pr_warn("mlx4_core: bad num_mac: %d\n", log_num_mac);
3942 if (log_num_vlan != 0)
3943 pr_warn("mlx4_core: log_num_vlan - obsolete module param, using %d\n",
3944 MLX4_LOG_NUM_VLANS);
3947 pr_warn("mlx4_core: use_prio - obsolete module param, ignored\n");
3949 if ((log_mtts_per_seg < 1) || (log_mtts_per_seg > 7)) {
3950 pr_warn("mlx4_core: bad log_mtts_per_seg: %d\n",
3955 /* Check if module param for ports type has legal combination */
3956 if (port_type_array[0] == false && port_type_array[1] == true) {
3957 pr_warn("Module parameter configuration ETH/IB is not supported. Switching to default configuration IB/IB\n");
3958 port_type_array[0] = true;
3961 if (mlx4_log_num_mgm_entry_size < -7 ||
3962 (mlx4_log_num_mgm_entry_size > 0 &&
3963 (mlx4_log_num_mgm_entry_size < MLX4_MIN_MGM_LOG_ENTRY_SIZE ||
3964 mlx4_log_num_mgm_entry_size > MLX4_MAX_MGM_LOG_ENTRY_SIZE))) {
3965 pr_warn("mlx4_core: mlx4_log_num_mgm_entry_size (%d) not in legal range (-7..0 or %d..%d)\n",
3966 mlx4_log_num_mgm_entry_size,
3967 MLX4_MIN_MGM_LOG_ENTRY_SIZE,
3968 MLX4_MAX_MGM_LOG_ENTRY_SIZE);
3975 static int __init mlx4_init(void)
3979 if (mlx4_verify_params())
3983 mlx4_wq = create_singlethread_workqueue("mlx4");
3987 ret = pci_register_driver(&mlx4_driver);
3989 destroy_workqueue(mlx4_wq);
3990 return ret < 0 ? ret : 0;
3993 static void __exit mlx4_cleanup(void)
3995 pci_unregister_driver(&mlx4_driver);
3996 destroy_workqueue(mlx4_wq);
3999 module_init(mlx4_init);
4000 module_exit(mlx4_cleanup);