2 * Copyright (c) 2004, 2005 Topspin Communications. All rights reserved.
3 * Copyright (c) 2005 Sun Microsystems, Inc. All rights reserved.
4 * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
6 * This software is available to you under a choice of one of two
7 * licenses. You may choose to be licensed under the terms of the GNU
8 * General Public License (GPL) Version 2, available from the file
9 * COPYING in the main directory of this source tree, or the
10 * OpenIB.org BSD license below:
12 * Redistribution and use in source and binary forms, with or
13 * without modification, are permitted provided that the following
16 * - Redistributions of source code must retain the above
17 * copyright notice, this list of conditions and the following
20 * - Redistributions in binary form must reproduce the above
21 * copyright notice, this list of conditions and the following
22 * disclaimer in the documentation and/or other materials
23 * provided with the distribution.
25 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
35 #include <linux/module.h>
36 #include <linux/init.h>
37 #include <linux/errno.h>
38 #include <linux/pci.h>
39 #include <linux/interrupt.h>
41 #include "mthca_dev.h"
42 #include "mthca_config_reg.h"
43 #include "mthca_cmd.h"
44 #include "mthca_profile.h"
45 #include "mthca_memfree.h"
46 #include "mthca_wqe.h"
48 MODULE_AUTHOR("Roland Dreier");
49 MODULE_DESCRIPTION("Mellanox InfiniBand HCA low-level driver");
50 MODULE_LICENSE("Dual BSD/GPL");
51 MODULE_VERSION(DRV_VERSION);
53 #ifdef CONFIG_INFINIBAND_MTHCA_DEBUG
55 int mthca_debug_level = 0;
56 module_param_named(debug_level, mthca_debug_level, int, 0644);
57 MODULE_PARM_DESC(debug_level, "Enable debug tracing if > 0");
59 #endif /* CONFIG_INFINIBAND_MTHCA_DEBUG */
64 module_param(msi_x, int, 0444);
65 MODULE_PARM_DESC(msi_x, "attempt to use MSI-X if nonzero");
67 #else /* CONFIG_PCI_MSI */
71 #endif /* CONFIG_PCI_MSI */
73 static int tune_pci = 0;
74 module_param(tune_pci, int, 0444);
75 MODULE_PARM_DESC(tune_pci, "increase PCI burst from the default set by BIOS if nonzero");
77 DEFINE_MUTEX(mthca_device_mutex);
79 #define MTHCA_DEFAULT_NUM_QP (1 << 16)
80 #define MTHCA_DEFAULT_RDB_PER_QP (1 << 2)
81 #define MTHCA_DEFAULT_NUM_CQ (1 << 16)
82 #define MTHCA_DEFAULT_NUM_MCG (1 << 13)
83 #define MTHCA_DEFAULT_NUM_MPT (1 << 17)
84 #define MTHCA_DEFAULT_NUM_MTT (1 << 20)
85 #define MTHCA_DEFAULT_NUM_UDAV (1 << 15)
86 #define MTHCA_DEFAULT_NUM_RESERVED_MTTS (1 << 18)
87 #define MTHCA_DEFAULT_NUM_UARC_SIZE (1 << 18)
89 static struct mthca_profile hca_profile = {
90 .num_qp = MTHCA_DEFAULT_NUM_QP,
91 .rdb_per_qp = MTHCA_DEFAULT_RDB_PER_QP,
92 .num_cq = MTHCA_DEFAULT_NUM_CQ,
93 .num_mcg = MTHCA_DEFAULT_NUM_MCG,
94 .num_mpt = MTHCA_DEFAULT_NUM_MPT,
95 .num_mtt = MTHCA_DEFAULT_NUM_MTT,
96 .num_udav = MTHCA_DEFAULT_NUM_UDAV, /* Tavor only */
97 .fmr_reserved_mtts = MTHCA_DEFAULT_NUM_RESERVED_MTTS, /* Tavor only */
98 .uarc_size = MTHCA_DEFAULT_NUM_UARC_SIZE, /* Arbel only */
101 module_param_named(num_qp, hca_profile.num_qp, int, 0444);
102 MODULE_PARM_DESC(num_qp, "maximum number of QPs per HCA");
104 module_param_named(rdb_per_qp, hca_profile.rdb_per_qp, int, 0444);
105 MODULE_PARM_DESC(rdb_per_qp, "number of RDB buffers per QP");
107 module_param_named(num_cq, hca_profile.num_cq, int, 0444);
108 MODULE_PARM_DESC(num_cq, "maximum number of CQs per HCA");
110 module_param_named(num_mcg, hca_profile.num_mcg, int, 0444);
111 MODULE_PARM_DESC(num_mcg, "maximum number of multicast groups per HCA");
113 module_param_named(num_mpt, hca_profile.num_mpt, int, 0444);
114 MODULE_PARM_DESC(num_mpt,
115 "maximum number of memory protection table entries per HCA");
117 module_param_named(num_mtt, hca_profile.num_mtt, int, 0444);
118 MODULE_PARM_DESC(num_mtt,
119 "maximum number of memory translation table segments per HCA");
121 module_param_named(num_udav, hca_profile.num_udav, int, 0444);
122 MODULE_PARM_DESC(num_udav, "maximum number of UD address vectors per HCA");
124 module_param_named(fmr_reserved_mtts, hca_profile.fmr_reserved_mtts, int, 0444);
125 MODULE_PARM_DESC(fmr_reserved_mtts,
126 "number of memory translation table segments reserved for FMR");
128 static char mthca_version[] __devinitdata =
129 DRV_NAME ": Mellanox InfiniBand HCA driver v"
130 DRV_VERSION " (" DRV_RELDATE ")\n";
132 static int mthca_tune_pci(struct mthca_dev *mdev)
137 /* First try to max out Read Byte Count */
138 if (pci_find_capability(mdev->pdev, PCI_CAP_ID_PCIX)) {
139 if (pcix_set_mmrbc(mdev->pdev, pcix_get_max_mmrbc(mdev->pdev))) {
140 mthca_err(mdev, "Couldn't set PCI-X max read count, "
144 } else if (!(mdev->mthca_flags & MTHCA_FLAG_PCIE))
145 mthca_info(mdev, "No PCI-X capability, not setting RBC.\n");
147 if (pci_find_capability(mdev->pdev, PCI_CAP_ID_EXP)) {
148 if (pcie_set_readrq(mdev->pdev, 4096)) {
149 mthca_err(mdev, "Couldn't write PCI Express read request, "
153 } else if (mdev->mthca_flags & MTHCA_FLAG_PCIE)
154 mthca_info(mdev, "No PCI Express capability, "
155 "not setting Max Read Request Size.\n");
160 static int mthca_dev_lim(struct mthca_dev *mdev, struct mthca_dev_lim *dev_lim)
165 err = mthca_QUERY_DEV_LIM(mdev, dev_lim, &status);
167 mthca_err(mdev, "QUERY_DEV_LIM command failed, aborting.\n");
171 mthca_err(mdev, "QUERY_DEV_LIM returned status 0x%02x, "
172 "aborting.\n", status);
175 if (dev_lim->min_page_sz > PAGE_SIZE) {
176 mthca_err(mdev, "HCA minimum page size of %d bigger than "
177 "kernel PAGE_SIZE of %ld, aborting.\n",
178 dev_lim->min_page_sz, PAGE_SIZE);
181 if (dev_lim->num_ports > MTHCA_MAX_PORTS) {
182 mthca_err(mdev, "HCA has %d ports, but we only support %d, "
184 dev_lim->num_ports, MTHCA_MAX_PORTS);
188 if (dev_lim->uar_size > pci_resource_len(mdev->pdev, 2)) {
189 mthca_err(mdev, "HCA reported UAR size of 0x%x bigger than "
190 "PCI resource 2 size of 0x%llx, aborting.\n",
192 (unsigned long long)pci_resource_len(mdev->pdev, 2));
196 mdev->limits.num_ports = dev_lim->num_ports;
197 mdev->limits.vl_cap = dev_lim->max_vl;
198 mdev->limits.mtu_cap = dev_lim->max_mtu;
199 mdev->limits.gid_table_len = dev_lim->max_gids;
200 mdev->limits.pkey_table_len = dev_lim->max_pkeys;
201 mdev->limits.local_ca_ack_delay = dev_lim->local_ca_ack_delay;
203 * Need to allow for worst case send WQE overhead and check
204 * whether max_desc_sz imposes a lower limit than max_sg; UD
205 * send has the biggest overhead.
207 mdev->limits.max_sg = min_t(int, dev_lim->max_sg,
208 (dev_lim->max_desc_sz -
209 sizeof (struct mthca_next_seg) -
210 (mthca_is_memfree(mdev) ?
211 sizeof (struct mthca_arbel_ud_seg) :
212 sizeof (struct mthca_tavor_ud_seg))) /
213 sizeof (struct mthca_data_seg));
214 mdev->limits.max_wqes = dev_lim->max_qp_sz;
215 mdev->limits.max_qp_init_rdma = dev_lim->max_requester_per_qp;
216 mdev->limits.reserved_qps = dev_lim->reserved_qps;
217 mdev->limits.max_srq_wqes = dev_lim->max_srq_sz;
218 mdev->limits.reserved_srqs = dev_lim->reserved_srqs;
219 mdev->limits.reserved_eecs = dev_lim->reserved_eecs;
220 mdev->limits.max_desc_sz = dev_lim->max_desc_sz;
221 mdev->limits.max_srq_sge = mthca_max_srq_sge(mdev);
223 * Subtract 1 from the limit because we need to allocate a
224 * spare CQE so the HCA HW can tell the difference between an
225 * empty CQ and a full CQ.
227 mdev->limits.max_cqes = dev_lim->max_cq_sz - 1;
228 mdev->limits.reserved_cqs = dev_lim->reserved_cqs;
229 mdev->limits.reserved_eqs = dev_lim->reserved_eqs;
230 mdev->limits.reserved_mtts = dev_lim->reserved_mtts;
231 mdev->limits.reserved_mrws = dev_lim->reserved_mrws;
232 mdev->limits.reserved_uars = dev_lim->reserved_uars;
233 mdev->limits.reserved_pds = dev_lim->reserved_pds;
234 mdev->limits.port_width_cap = dev_lim->max_port_width;
235 mdev->limits.page_size_cap = ~(u32) (dev_lim->min_page_sz - 1);
236 mdev->limits.flags = dev_lim->flags;
238 * For old FW that doesn't return static rate support, use a
239 * value of 0x3 (only static rate values of 0 or 1 are handled),
240 * except on Sinai, where even old FW can handle static rate
243 if (dev_lim->stat_rate_support)
244 mdev->limits.stat_rate_support = dev_lim->stat_rate_support;
245 else if (mdev->mthca_flags & MTHCA_FLAG_SINAI_OPT)
246 mdev->limits.stat_rate_support = 0xf;
248 mdev->limits.stat_rate_support = 0x3;
250 /* IB_DEVICE_RESIZE_MAX_WR not supported by driver.
251 May be doable since hardware supports it for SRQ.
253 IB_DEVICE_N_NOTIFY_CQ is supported by hardware but not by driver.
255 IB_DEVICE_SRQ_RESIZE is supported by hardware but SRQ is not
256 supported by driver. */
257 mdev->device_cap_flags = IB_DEVICE_CHANGE_PHY_PORT |
258 IB_DEVICE_PORT_ACTIVE_EVENT |
259 IB_DEVICE_SYS_IMAGE_GUID |
260 IB_DEVICE_RC_RNR_NAK_GEN;
262 if (dev_lim->flags & DEV_LIM_FLAG_BAD_PKEY_CNTR)
263 mdev->device_cap_flags |= IB_DEVICE_BAD_PKEY_CNTR;
265 if (dev_lim->flags & DEV_LIM_FLAG_BAD_QKEY_CNTR)
266 mdev->device_cap_flags |= IB_DEVICE_BAD_QKEY_CNTR;
268 if (dev_lim->flags & DEV_LIM_FLAG_RAW_MULTI)
269 mdev->device_cap_flags |= IB_DEVICE_RAW_MULTI;
271 if (dev_lim->flags & DEV_LIM_FLAG_AUTO_PATH_MIG)
272 mdev->device_cap_flags |= IB_DEVICE_AUTO_PATH_MIG;
274 if (dev_lim->flags & DEV_LIM_FLAG_UD_AV_PORT_ENFORCE)
275 mdev->device_cap_flags |= IB_DEVICE_UD_AV_PORT_ENFORCE;
277 if (dev_lim->flags & DEV_LIM_FLAG_SRQ)
278 mdev->mthca_flags |= MTHCA_FLAG_SRQ;
280 if (mthca_is_memfree(mdev))
281 if (dev_lim->flags & DEV_LIM_FLAG_IPOIB_CSUM)
282 mdev->device_cap_flags |= IB_DEVICE_UD_IP_CSUM;
287 static int mthca_init_tavor(struct mthca_dev *mdev)
292 struct mthca_dev_lim dev_lim;
293 struct mthca_profile profile;
294 struct mthca_init_hca_param init_hca;
296 err = mthca_SYS_EN(mdev, &status);
298 mthca_err(mdev, "SYS_EN command failed, aborting.\n");
302 mthca_err(mdev, "SYS_EN returned status 0x%02x, "
303 "aborting.\n", status);
307 err = mthca_QUERY_FW(mdev, &status);
309 mthca_err(mdev, "QUERY_FW command failed, aborting.\n");
313 mthca_err(mdev, "QUERY_FW returned status 0x%02x, "
314 "aborting.\n", status);
318 err = mthca_QUERY_DDR(mdev, &status);
320 mthca_err(mdev, "QUERY_DDR command failed, aborting.\n");
324 mthca_err(mdev, "QUERY_DDR returned status 0x%02x, "
325 "aborting.\n", status);
330 err = mthca_dev_lim(mdev, &dev_lim);
332 mthca_err(mdev, "QUERY_DEV_LIM command failed, aborting.\n");
336 profile = hca_profile;
337 profile.num_uar = dev_lim.uar_size / PAGE_SIZE;
338 profile.uarc_size = 0;
339 if (mdev->mthca_flags & MTHCA_FLAG_SRQ)
340 profile.num_srq = dev_lim.max_srqs;
342 size = mthca_make_profile(mdev, &profile, &dev_lim, &init_hca);
348 err = mthca_INIT_HCA(mdev, &init_hca, &status);
350 mthca_err(mdev, "INIT_HCA command failed, aborting.\n");
354 mthca_err(mdev, "INIT_HCA returned status 0x%02x, "
355 "aborting.\n", status);
363 mthca_SYS_DIS(mdev, &status);
368 static int mthca_load_fw(struct mthca_dev *mdev)
373 /* FIXME: use HCA-attached memory for FW if present */
375 mdev->fw.arbel.fw_icm =
376 mthca_alloc_icm(mdev, mdev->fw.arbel.fw_pages,
377 GFP_HIGHUSER | __GFP_NOWARN, 0);
378 if (!mdev->fw.arbel.fw_icm) {
379 mthca_err(mdev, "Couldn't allocate FW area, aborting.\n");
383 err = mthca_MAP_FA(mdev, mdev->fw.arbel.fw_icm, &status);
385 mthca_err(mdev, "MAP_FA command failed, aborting.\n");
389 mthca_err(mdev, "MAP_FA returned status 0x%02x, aborting.\n", status);
393 err = mthca_RUN_FW(mdev, &status);
395 mthca_err(mdev, "RUN_FW command failed, aborting.\n");
399 mthca_err(mdev, "RUN_FW returned status 0x%02x, aborting.\n", status);
407 mthca_UNMAP_FA(mdev, &status);
410 mthca_free_icm(mdev, mdev->fw.arbel.fw_icm, 0);
414 static int mthca_init_icm(struct mthca_dev *mdev,
415 struct mthca_dev_lim *dev_lim,
416 struct mthca_init_hca_param *init_hca,
423 err = mthca_SET_ICM_SIZE(mdev, icm_size, &aux_pages, &status);
425 mthca_err(mdev, "SET_ICM_SIZE command failed, aborting.\n");
429 mthca_err(mdev, "SET_ICM_SIZE returned status 0x%02x, "
430 "aborting.\n", status);
434 mthca_dbg(mdev, "%lld KB of HCA context requires %lld KB aux memory.\n",
435 (unsigned long long) icm_size >> 10,
436 (unsigned long long) aux_pages << 2);
438 mdev->fw.arbel.aux_icm = mthca_alloc_icm(mdev, aux_pages,
439 GFP_HIGHUSER | __GFP_NOWARN, 0);
440 if (!mdev->fw.arbel.aux_icm) {
441 mthca_err(mdev, "Couldn't allocate aux memory, aborting.\n");
445 err = mthca_MAP_ICM_AUX(mdev, mdev->fw.arbel.aux_icm, &status);
447 mthca_err(mdev, "MAP_ICM_AUX command failed, aborting.\n");
451 mthca_err(mdev, "MAP_ICM_AUX returned status 0x%02x, aborting.\n", status);
456 err = mthca_map_eq_icm(mdev, init_hca->eqc_base);
458 mthca_err(mdev, "Failed to map EQ context memory, aborting.\n");
462 /* CPU writes to non-reserved MTTs, while HCA might DMA to reserved mtts */
463 mdev->limits.reserved_mtts = ALIGN(mdev->limits.reserved_mtts * MTHCA_MTT_SEG_SIZE,
464 dma_get_cache_alignment()) / MTHCA_MTT_SEG_SIZE;
466 mdev->mr_table.mtt_table = mthca_alloc_icm_table(mdev, init_hca->mtt_base,
468 mdev->limits.num_mtt_segs,
469 mdev->limits.reserved_mtts,
471 if (!mdev->mr_table.mtt_table) {
472 mthca_err(mdev, "Failed to map MTT context memory, aborting.\n");
477 mdev->mr_table.mpt_table = mthca_alloc_icm_table(mdev, init_hca->mpt_base,
478 dev_lim->mpt_entry_sz,
479 mdev->limits.num_mpts,
480 mdev->limits.reserved_mrws,
482 if (!mdev->mr_table.mpt_table) {
483 mthca_err(mdev, "Failed to map MPT context memory, aborting.\n");
488 mdev->qp_table.qp_table = mthca_alloc_icm_table(mdev, init_hca->qpc_base,
489 dev_lim->qpc_entry_sz,
490 mdev->limits.num_qps,
491 mdev->limits.reserved_qps,
493 if (!mdev->qp_table.qp_table) {
494 mthca_err(mdev, "Failed to map QP context memory, aborting.\n");
499 mdev->qp_table.eqp_table = mthca_alloc_icm_table(mdev, init_hca->eqpc_base,
500 dev_lim->eqpc_entry_sz,
501 mdev->limits.num_qps,
502 mdev->limits.reserved_qps,
504 if (!mdev->qp_table.eqp_table) {
505 mthca_err(mdev, "Failed to map EQP context memory, aborting.\n");
510 mdev->qp_table.rdb_table = mthca_alloc_icm_table(mdev, init_hca->rdb_base,
511 MTHCA_RDB_ENTRY_SIZE,
512 mdev->limits.num_qps <<
513 mdev->qp_table.rdb_shift, 0,
515 if (!mdev->qp_table.rdb_table) {
516 mthca_err(mdev, "Failed to map RDB context memory, aborting\n");
521 mdev->cq_table.table = mthca_alloc_icm_table(mdev, init_hca->cqc_base,
522 dev_lim->cqc_entry_sz,
523 mdev->limits.num_cqs,
524 mdev->limits.reserved_cqs,
526 if (!mdev->cq_table.table) {
527 mthca_err(mdev, "Failed to map CQ context memory, aborting.\n");
532 if (mdev->mthca_flags & MTHCA_FLAG_SRQ) {
533 mdev->srq_table.table =
534 mthca_alloc_icm_table(mdev, init_hca->srqc_base,
535 dev_lim->srq_entry_sz,
536 mdev->limits.num_srqs,
537 mdev->limits.reserved_srqs,
539 if (!mdev->srq_table.table) {
540 mthca_err(mdev, "Failed to map SRQ context memory, "
548 * It's not strictly required, but for simplicity just map the
549 * whole multicast group table now. The table isn't very big
550 * and it's a lot easier than trying to track ref counts.
552 mdev->mcg_table.table = mthca_alloc_icm_table(mdev, init_hca->mc_base,
553 MTHCA_MGM_ENTRY_SIZE,
554 mdev->limits.num_mgms +
555 mdev->limits.num_amgms,
556 mdev->limits.num_mgms +
557 mdev->limits.num_amgms,
559 if (!mdev->mcg_table.table) {
560 mthca_err(mdev, "Failed to map MCG context memory, aborting.\n");
568 if (mdev->mthca_flags & MTHCA_FLAG_SRQ)
569 mthca_free_icm_table(mdev, mdev->srq_table.table);
572 mthca_free_icm_table(mdev, mdev->cq_table.table);
575 mthca_free_icm_table(mdev, mdev->qp_table.rdb_table);
578 mthca_free_icm_table(mdev, mdev->qp_table.eqp_table);
581 mthca_free_icm_table(mdev, mdev->qp_table.qp_table);
584 mthca_free_icm_table(mdev, mdev->mr_table.mpt_table);
587 mthca_free_icm_table(mdev, mdev->mr_table.mtt_table);
590 mthca_unmap_eq_icm(mdev);
593 mthca_UNMAP_ICM_AUX(mdev, &status);
596 mthca_free_icm(mdev, mdev->fw.arbel.aux_icm, 0);
601 static void mthca_free_icms(struct mthca_dev *mdev)
605 mthca_free_icm_table(mdev, mdev->mcg_table.table);
606 if (mdev->mthca_flags & MTHCA_FLAG_SRQ)
607 mthca_free_icm_table(mdev, mdev->srq_table.table);
608 mthca_free_icm_table(mdev, mdev->cq_table.table);
609 mthca_free_icm_table(mdev, mdev->qp_table.rdb_table);
610 mthca_free_icm_table(mdev, mdev->qp_table.eqp_table);
611 mthca_free_icm_table(mdev, mdev->qp_table.qp_table);
612 mthca_free_icm_table(mdev, mdev->mr_table.mpt_table);
613 mthca_free_icm_table(mdev, mdev->mr_table.mtt_table);
614 mthca_unmap_eq_icm(mdev);
616 mthca_UNMAP_ICM_AUX(mdev, &status);
617 mthca_free_icm(mdev, mdev->fw.arbel.aux_icm, 0);
620 static int mthca_init_arbel(struct mthca_dev *mdev)
622 struct mthca_dev_lim dev_lim;
623 struct mthca_profile profile;
624 struct mthca_init_hca_param init_hca;
629 err = mthca_QUERY_FW(mdev, &status);
631 mthca_err(mdev, "QUERY_FW command failed, aborting.\n");
635 mthca_err(mdev, "QUERY_FW returned status 0x%02x, "
636 "aborting.\n", status);
640 err = mthca_ENABLE_LAM(mdev, &status);
642 mthca_err(mdev, "ENABLE_LAM command failed, aborting.\n");
645 if (status == MTHCA_CMD_STAT_LAM_NOT_PRE) {
646 mthca_dbg(mdev, "No HCA-attached memory (running in MemFree mode)\n");
647 mdev->mthca_flags |= MTHCA_FLAG_NO_LAM;
649 mthca_err(mdev, "ENABLE_LAM returned status 0x%02x, "
650 "aborting.\n", status);
654 err = mthca_load_fw(mdev);
656 mthca_err(mdev, "Failed to start FW, aborting.\n");
660 err = mthca_dev_lim(mdev, &dev_lim);
662 mthca_err(mdev, "QUERY_DEV_LIM command failed, aborting.\n");
666 profile = hca_profile;
667 profile.num_uar = dev_lim.uar_size / PAGE_SIZE;
668 profile.num_udav = 0;
669 if (mdev->mthca_flags & MTHCA_FLAG_SRQ)
670 profile.num_srq = dev_lim.max_srqs;
672 icm_size = mthca_make_profile(mdev, &profile, &dev_lim, &init_hca);
678 err = mthca_init_icm(mdev, &dev_lim, &init_hca, icm_size);
682 err = mthca_INIT_HCA(mdev, &init_hca, &status);
684 mthca_err(mdev, "INIT_HCA command failed, aborting.\n");
688 mthca_err(mdev, "INIT_HCA returned status 0x%02x, "
689 "aborting.\n", status);
697 mthca_free_icms(mdev);
700 mthca_UNMAP_FA(mdev, &status);
701 mthca_free_icm(mdev, mdev->fw.arbel.fw_icm, 0);
704 if (!(mdev->mthca_flags & MTHCA_FLAG_NO_LAM))
705 mthca_DISABLE_LAM(mdev, &status);
710 static void mthca_close_hca(struct mthca_dev *mdev)
714 mthca_CLOSE_HCA(mdev, 0, &status);
716 if (mthca_is_memfree(mdev)) {
717 mthca_free_icms(mdev);
719 mthca_UNMAP_FA(mdev, &status);
720 mthca_free_icm(mdev, mdev->fw.arbel.fw_icm, 0);
722 if (!(mdev->mthca_flags & MTHCA_FLAG_NO_LAM))
723 mthca_DISABLE_LAM(mdev, &status);
725 mthca_SYS_DIS(mdev, &status);
728 static int mthca_init_hca(struct mthca_dev *mdev)
732 struct mthca_adapter adapter;
734 if (mthca_is_memfree(mdev))
735 err = mthca_init_arbel(mdev);
737 err = mthca_init_tavor(mdev);
742 err = mthca_QUERY_ADAPTER(mdev, &adapter, &status);
744 mthca_err(mdev, "QUERY_ADAPTER command failed, aborting.\n");
748 mthca_err(mdev, "QUERY_ADAPTER returned status 0x%02x, "
749 "aborting.\n", status);
754 mdev->eq_table.inta_pin = adapter.inta_pin;
755 if (!mthca_is_memfree(mdev))
756 mdev->rev_id = adapter.revision_id;
757 memcpy(mdev->board_id, adapter.board_id, sizeof mdev->board_id);
762 mthca_close_hca(mdev);
766 static int mthca_setup_hca(struct mthca_dev *dev)
771 MTHCA_INIT_DOORBELL_LOCK(&dev->doorbell_lock);
773 err = mthca_init_uar_table(dev);
775 mthca_err(dev, "Failed to initialize "
776 "user access region table, aborting.\n");
780 err = mthca_uar_alloc(dev, &dev->driver_uar);
782 mthca_err(dev, "Failed to allocate driver access region, "
784 goto err_uar_table_free;
787 dev->kar = ioremap(dev->driver_uar.pfn << PAGE_SHIFT, PAGE_SIZE);
789 mthca_err(dev, "Couldn't map kernel access region, "
795 err = mthca_init_pd_table(dev);
797 mthca_err(dev, "Failed to initialize "
798 "protection domain table, aborting.\n");
802 err = mthca_init_mr_table(dev);
804 mthca_err(dev, "Failed to initialize "
805 "memory region table, aborting.\n");
806 goto err_pd_table_free;
809 err = mthca_pd_alloc(dev, 1, &dev->driver_pd);
811 mthca_err(dev, "Failed to create driver PD, "
813 goto err_mr_table_free;
816 err = mthca_init_eq_table(dev);
818 mthca_err(dev, "Failed to initialize "
819 "event queue table, aborting.\n");
823 err = mthca_cmd_use_events(dev);
825 mthca_err(dev, "Failed to switch to event-driven "
826 "firmware commands, aborting.\n");
827 goto err_eq_table_free;
830 err = mthca_NOP(dev, &status);
832 if (dev->mthca_flags & MTHCA_FLAG_MSI_X) {
833 mthca_warn(dev, "NOP command failed to generate interrupt "
835 dev->eq_table.eq[MTHCA_EQ_CMD].msi_x_vector);
836 mthca_warn(dev, "Trying again with MSI-X disabled.\n");
838 mthca_err(dev, "NOP command failed to generate interrupt "
839 "(IRQ %d), aborting.\n",
841 mthca_err(dev, "BIOS or ACPI interrupt routing problem?\n");
847 mthca_dbg(dev, "NOP command IRQ test passed\n");
849 err = mthca_init_cq_table(dev);
851 mthca_err(dev, "Failed to initialize "
852 "completion queue table, aborting.\n");
856 err = mthca_init_srq_table(dev);
858 mthca_err(dev, "Failed to initialize "
859 "shared receive queue table, aborting.\n");
860 goto err_cq_table_free;
863 err = mthca_init_qp_table(dev);
865 mthca_err(dev, "Failed to initialize "
866 "queue pair table, aborting.\n");
867 goto err_srq_table_free;
870 err = mthca_init_av_table(dev);
872 mthca_err(dev, "Failed to initialize "
873 "address vector table, aborting.\n");
874 goto err_qp_table_free;
877 err = mthca_init_mcg_table(dev);
879 mthca_err(dev, "Failed to initialize "
880 "multicast group table, aborting.\n");
881 goto err_av_table_free;
887 mthca_cleanup_av_table(dev);
890 mthca_cleanup_qp_table(dev);
893 mthca_cleanup_srq_table(dev);
896 mthca_cleanup_cq_table(dev);
899 mthca_cmd_use_polling(dev);
902 mthca_cleanup_eq_table(dev);
905 mthca_pd_free(dev, &dev->driver_pd);
908 mthca_cleanup_mr_table(dev);
911 mthca_cleanup_pd_table(dev);
917 mthca_uar_free(dev, &dev->driver_uar);
920 mthca_cleanup_uar_table(dev);
924 static int mthca_request_regions(struct pci_dev *pdev, int ddr_hidden)
929 * We can't just use pci_request_regions() because the MSI-X
930 * table is right in the middle of the first BAR. If we did
931 * pci_request_region and grab all of the first BAR, then
932 * setting up MSI-X would fail, since the PCI core wants to do
933 * request_mem_region on the MSI-X vector table.
935 * So just request what we need right now, and request any
936 * other regions we need when setting up EQs.
938 if (!request_mem_region(pci_resource_start(pdev, 0) + MTHCA_HCR_BASE,
939 MTHCA_HCR_SIZE, DRV_NAME))
942 err = pci_request_region(pdev, 2, DRV_NAME);
944 goto err_bar2_failed;
947 err = pci_request_region(pdev, 4, DRV_NAME);
949 goto err_bar4_failed;
955 pci_release_region(pdev, 2);
958 release_mem_region(pci_resource_start(pdev, 0) + MTHCA_HCR_BASE,
964 static void mthca_release_regions(struct pci_dev *pdev,
968 pci_release_region(pdev, 4);
970 pci_release_region(pdev, 2);
972 release_mem_region(pci_resource_start(pdev, 0) + MTHCA_HCR_BASE,
976 static int mthca_enable_msi_x(struct mthca_dev *mdev)
978 struct msix_entry entries[3];
981 entries[0].entry = 0;
982 entries[1].entry = 1;
983 entries[2].entry = 2;
985 err = pci_enable_msix(mdev->pdev, entries, ARRAY_SIZE(entries));
988 mthca_info(mdev, "Only %d MSI-X vectors available, "
989 "not using MSI-X\n", err);
993 mdev->eq_table.eq[MTHCA_EQ_COMP ].msi_x_vector = entries[0].vector;
994 mdev->eq_table.eq[MTHCA_EQ_ASYNC].msi_x_vector = entries[1].vector;
995 mdev->eq_table.eq[MTHCA_EQ_CMD ].msi_x_vector = entries[2].vector;
1000 /* Types of supported HCA */
1002 TAVOR, /* MT23108 */
1003 ARBEL_COMPAT, /* MT25208 in Tavor compat mode */
1004 ARBEL_NATIVE, /* MT25208 with extended features */
1008 #define MTHCA_FW_VER(major, minor, subminor) \
1009 (((u64) (major) << 32) | ((u64) (minor) << 16) | (u64) (subminor))
1014 } mthca_hca_table[] = {
1015 [TAVOR] = { .latest_fw = MTHCA_FW_VER(3, 5, 0),
1017 [ARBEL_COMPAT] = { .latest_fw = MTHCA_FW_VER(4, 8, 200),
1018 .flags = MTHCA_FLAG_PCIE },
1019 [ARBEL_NATIVE] = { .latest_fw = MTHCA_FW_VER(5, 3, 0),
1020 .flags = MTHCA_FLAG_MEMFREE |
1022 [SINAI] = { .latest_fw = MTHCA_FW_VER(1, 2, 0),
1023 .flags = MTHCA_FLAG_MEMFREE |
1025 MTHCA_FLAG_SINAI_OPT }
1028 static int __mthca_init_one(struct pci_dev *pdev, int hca_type)
1032 struct mthca_dev *mdev;
1034 printk(KERN_INFO PFX "Initializing %s\n",
1037 err = pci_enable_device(pdev);
1039 dev_err(&pdev->dev, "Cannot enable PCI device, "
1045 * Check for BARs. We expect 0: 1MB, 2: 8MB, 4: DDR (may not
1048 if (!(pci_resource_flags(pdev, 0) & IORESOURCE_MEM) ||
1049 pci_resource_len(pdev, 0) != 1 << 20) {
1050 dev_err(&pdev->dev, "Missing DCS, aborting.\n");
1052 goto err_disable_pdev;
1054 if (!(pci_resource_flags(pdev, 2) & IORESOURCE_MEM)) {
1055 dev_err(&pdev->dev, "Missing UAR, aborting.\n");
1057 goto err_disable_pdev;
1059 if (!(pci_resource_flags(pdev, 4) & IORESOURCE_MEM))
1062 err = mthca_request_regions(pdev, ddr_hidden);
1064 dev_err(&pdev->dev, "Cannot obtain PCI resources, "
1066 goto err_disable_pdev;
1069 pci_set_master(pdev);
1071 err = pci_set_dma_mask(pdev, DMA_64BIT_MASK);
1073 dev_warn(&pdev->dev, "Warning: couldn't set 64-bit PCI DMA mask.\n");
1074 err = pci_set_dma_mask(pdev, DMA_32BIT_MASK);
1076 dev_err(&pdev->dev, "Can't set PCI DMA mask, aborting.\n");
1080 err = pci_set_consistent_dma_mask(pdev, DMA_64BIT_MASK);
1082 dev_warn(&pdev->dev, "Warning: couldn't set 64-bit "
1083 "consistent PCI DMA mask.\n");
1084 err = pci_set_consistent_dma_mask(pdev, DMA_32BIT_MASK);
1086 dev_err(&pdev->dev, "Can't set consistent PCI DMA mask, "
1092 mdev = (struct mthca_dev *) ib_alloc_device(sizeof *mdev);
1094 dev_err(&pdev->dev, "Device struct alloc failed, "
1102 mdev->mthca_flags = mthca_hca_table[hca_type].flags;
1104 mdev->mthca_flags |= MTHCA_FLAG_DDR_HIDDEN;
1107 * Now reset the HCA before we touch the PCI capabilities or
1108 * attempt a firmware command, since a boot ROM may have left
1109 * the HCA in an undefined state.
1111 err = mthca_reset(mdev);
1113 mthca_err(mdev, "Failed to reset HCA, aborting.\n");
1117 if (mthca_cmd_init(mdev)) {
1118 mthca_err(mdev, "Failed to init command interface, aborting.\n");
1122 err = mthca_tune_pci(mdev);
1126 err = mthca_init_hca(mdev);
1130 if (mdev->fw_ver < mthca_hca_table[hca_type].latest_fw) {
1131 mthca_warn(mdev, "HCA FW version %d.%d.%03d is old (%d.%d.%03d is current).\n",
1132 (int) (mdev->fw_ver >> 32), (int) (mdev->fw_ver >> 16) & 0xffff,
1133 (int) (mdev->fw_ver & 0xffff),
1134 (int) (mthca_hca_table[hca_type].latest_fw >> 32),
1135 (int) (mthca_hca_table[hca_type].latest_fw >> 16) & 0xffff,
1136 (int) (mthca_hca_table[hca_type].latest_fw & 0xffff));
1137 mthca_warn(mdev, "If you have problems, try updating your HCA FW.\n");
1140 if (msi_x && !mthca_enable_msi_x(mdev))
1141 mdev->mthca_flags |= MTHCA_FLAG_MSI_X;
1143 err = mthca_setup_hca(mdev);
1144 if (err == -EBUSY && (mdev->mthca_flags & MTHCA_FLAG_MSI_X)) {
1145 if (mdev->mthca_flags & MTHCA_FLAG_MSI_X)
1146 pci_disable_msix(pdev);
1147 mdev->mthca_flags &= ~MTHCA_FLAG_MSI_X;
1149 err = mthca_setup_hca(mdev);
1155 err = mthca_register_device(mdev);
1159 err = mthca_create_agents(mdev);
1161 goto err_unregister;
1163 pci_set_drvdata(pdev, mdev);
1164 mdev->hca_type = hca_type;
1169 mthca_unregister_device(mdev);
1172 mthca_cleanup_mcg_table(mdev);
1173 mthca_cleanup_av_table(mdev);
1174 mthca_cleanup_qp_table(mdev);
1175 mthca_cleanup_srq_table(mdev);
1176 mthca_cleanup_cq_table(mdev);
1177 mthca_cmd_use_polling(mdev);
1178 mthca_cleanup_eq_table(mdev);
1180 mthca_pd_free(mdev, &mdev->driver_pd);
1182 mthca_cleanup_mr_table(mdev);
1183 mthca_cleanup_pd_table(mdev);
1184 mthca_cleanup_uar_table(mdev);
1187 if (mdev->mthca_flags & MTHCA_FLAG_MSI_X)
1188 pci_disable_msix(pdev);
1190 mthca_close_hca(mdev);
1193 mthca_cmd_cleanup(mdev);
1196 ib_dealloc_device(&mdev->ib_dev);
1199 mthca_release_regions(pdev, ddr_hidden);
1202 pci_disable_device(pdev);
1203 pci_set_drvdata(pdev, NULL);
1207 static void __mthca_remove_one(struct pci_dev *pdev)
1209 struct mthca_dev *mdev = pci_get_drvdata(pdev);
1214 mthca_free_agents(mdev);
1215 mthca_unregister_device(mdev);
1217 for (p = 1; p <= mdev->limits.num_ports; ++p)
1218 mthca_CLOSE_IB(mdev, p, &status);
1220 mthca_cleanup_mcg_table(mdev);
1221 mthca_cleanup_av_table(mdev);
1222 mthca_cleanup_qp_table(mdev);
1223 mthca_cleanup_srq_table(mdev);
1224 mthca_cleanup_cq_table(mdev);
1225 mthca_cmd_use_polling(mdev);
1226 mthca_cleanup_eq_table(mdev);
1228 mthca_pd_free(mdev, &mdev->driver_pd);
1230 mthca_cleanup_mr_table(mdev);
1231 mthca_cleanup_pd_table(mdev);
1234 mthca_uar_free(mdev, &mdev->driver_uar);
1235 mthca_cleanup_uar_table(mdev);
1236 mthca_close_hca(mdev);
1237 mthca_cmd_cleanup(mdev);
1239 if (mdev->mthca_flags & MTHCA_FLAG_MSI_X)
1240 pci_disable_msix(pdev);
1242 ib_dealloc_device(&mdev->ib_dev);
1243 mthca_release_regions(pdev, mdev->mthca_flags &
1244 MTHCA_FLAG_DDR_HIDDEN);
1245 pci_disable_device(pdev);
1246 pci_set_drvdata(pdev, NULL);
1250 int __mthca_restart_one(struct pci_dev *pdev)
1252 struct mthca_dev *mdev;
1255 mdev = pci_get_drvdata(pdev);
1258 hca_type = mdev->hca_type;
1259 __mthca_remove_one(pdev);
1260 return __mthca_init_one(pdev, hca_type);
1263 static int __devinit mthca_init_one(struct pci_dev *pdev,
1264 const struct pci_device_id *id)
1266 static int mthca_version_printed = 0;
1269 mutex_lock(&mthca_device_mutex);
1271 if (!mthca_version_printed) {
1272 printk(KERN_INFO "%s", mthca_version);
1273 ++mthca_version_printed;
1276 if (id->driver_data >= ARRAY_SIZE(mthca_hca_table)) {
1277 printk(KERN_ERR PFX "%s has invalid driver data %lx\n",
1278 pci_name(pdev), id->driver_data);
1279 mutex_unlock(&mthca_device_mutex);
1283 ret = __mthca_init_one(pdev, id->driver_data);
1285 mutex_unlock(&mthca_device_mutex);
1290 static void __devexit mthca_remove_one(struct pci_dev *pdev)
1292 mutex_lock(&mthca_device_mutex);
1293 __mthca_remove_one(pdev);
1294 mutex_unlock(&mthca_device_mutex);
1297 static struct pci_device_id mthca_pci_table[] = {
1298 { PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_TAVOR),
1299 .driver_data = TAVOR },
1300 { PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_TAVOR),
1301 .driver_data = TAVOR },
1302 { PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT),
1303 .driver_data = ARBEL_COMPAT },
1304 { PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_ARBEL_COMPAT),
1305 .driver_data = ARBEL_COMPAT },
1306 { PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_ARBEL),
1307 .driver_data = ARBEL_NATIVE },
1308 { PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_ARBEL),
1309 .driver_data = ARBEL_NATIVE },
1310 { PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_SINAI),
1311 .driver_data = SINAI },
1312 { PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_SINAI),
1313 .driver_data = SINAI },
1314 { PCI_DEVICE(PCI_VENDOR_ID_MELLANOX, PCI_DEVICE_ID_MELLANOX_SINAI_OLD),
1315 .driver_data = SINAI },
1316 { PCI_DEVICE(PCI_VENDOR_ID_TOPSPIN, PCI_DEVICE_ID_MELLANOX_SINAI_OLD),
1317 .driver_data = SINAI },
1321 MODULE_DEVICE_TABLE(pci, mthca_pci_table);
1323 static struct pci_driver mthca_driver = {
1325 .id_table = mthca_pci_table,
1326 .probe = mthca_init_one,
1327 .remove = __devexit_p(mthca_remove_one)
1330 static void __init __mthca_check_profile_val(const char *name, int *pval,
1333 /* value must be positive and power of 2 */
1334 int old_pval = *pval;
1337 *pval = pval_default;
1339 *pval = roundup_pow_of_two(old_pval);
1341 if (old_pval != *pval) {
1342 printk(KERN_WARNING PFX "Invalid value %d for %s in module parameter.\n",
1344 printk(KERN_WARNING PFX "Corrected %s to %d.\n", name, *pval);
1348 #define mthca_check_profile_val(name, default) \
1349 __mthca_check_profile_val(#name, &hca_profile.name, default)
1351 static void __init mthca_validate_profile(void)
1353 mthca_check_profile_val(num_qp, MTHCA_DEFAULT_NUM_QP);
1354 mthca_check_profile_val(rdb_per_qp, MTHCA_DEFAULT_RDB_PER_QP);
1355 mthca_check_profile_val(num_cq, MTHCA_DEFAULT_NUM_CQ);
1356 mthca_check_profile_val(num_mcg, MTHCA_DEFAULT_NUM_MCG);
1357 mthca_check_profile_val(num_mpt, MTHCA_DEFAULT_NUM_MPT);
1358 mthca_check_profile_val(num_mtt, MTHCA_DEFAULT_NUM_MTT);
1359 mthca_check_profile_val(num_udav, MTHCA_DEFAULT_NUM_UDAV);
1360 mthca_check_profile_val(fmr_reserved_mtts, MTHCA_DEFAULT_NUM_RESERVED_MTTS);
1362 if (hca_profile.fmr_reserved_mtts >= hca_profile.num_mtt) {
1363 printk(KERN_WARNING PFX "Invalid fmr_reserved_mtts module parameter %d.\n",
1364 hca_profile.fmr_reserved_mtts);
1365 printk(KERN_WARNING PFX "(Must be smaller than num_mtt %d)\n",
1366 hca_profile.num_mtt);
1367 hca_profile.fmr_reserved_mtts = hca_profile.num_mtt / 2;
1368 printk(KERN_WARNING PFX "Corrected fmr_reserved_mtts to %d.\n",
1369 hca_profile.fmr_reserved_mtts);
1373 static int __init mthca_init(void)
1377 mthca_validate_profile();
1379 ret = mthca_catas_init();
1383 ret = pci_register_driver(&mthca_driver);
1385 mthca_catas_cleanup();
1392 static void __exit mthca_cleanup(void)
1394 pci_unregister_driver(&mthca_driver);
1395 mthca_catas_cleanup();
1398 module_init(mthca_init);
1399 module_exit(mthca_cleanup);