2 * Copyright (c) 2004, 2005 Topspin Communications. All rights reserved.
3 * Copyright (c) 2005 Mellanox Technologies. All rights reserved.
5 * This software is available to you under a choice of one of two
6 * licenses. You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the
9 * OpenIB.org BSD license below:
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
19 * - Redistributions in binary form must reproduce the above
20 * copyright notice, this list of conditions and the following
21 * disclaimer in the documentation and/or other materials
22 * provided with the distribution.
24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33 * $Id: mthca_eq.c 1382 2004-12-24 02:21:02Z roland $
36 #include <linux/init.h>
37 #include <linux/errno.h>
38 #include <linux/interrupt.h>
39 #include <linux/pci.h>
41 #include "mthca_dev.h"
42 #include "mthca_cmd.h"
43 #include "mthca_config_reg.h"
46 MTHCA_NUM_ASYNC_EQE = 0x80,
47 MTHCA_NUM_CMD_EQE = 0x80,
48 MTHCA_NUM_SPARE_EQE = 0x80,
49 MTHCA_EQ_ENTRY_SIZE = 0x20
53 * Must be packed because start is 64 bits but only aligned to 32 bits.
55 struct mthca_eq_context {
58 __be32 logsize_usrpage;
59 __be32 tavor_pd; /* reserved for Arbel */
62 __be32 arbel_pd; /* lost_count for Tavor */
65 __be32 consumer_index;
66 __be32 producer_index;
68 } __attribute__((packed));
70 #define MTHCA_EQ_STATUS_OK ( 0 << 28)
71 #define MTHCA_EQ_STATUS_OVERFLOW ( 9 << 28)
72 #define MTHCA_EQ_STATUS_WRITE_FAIL (10 << 28)
73 #define MTHCA_EQ_OWNER_SW ( 0 << 24)
74 #define MTHCA_EQ_OWNER_HW ( 1 << 24)
75 #define MTHCA_EQ_FLAG_TR ( 1 << 18)
76 #define MTHCA_EQ_FLAG_OI ( 1 << 17)
77 #define MTHCA_EQ_STATE_ARMED ( 1 << 8)
78 #define MTHCA_EQ_STATE_FIRED ( 2 << 8)
79 #define MTHCA_EQ_STATE_ALWAYS_ARMED ( 3 << 8)
80 #define MTHCA_EQ_STATE_ARBEL ( 8 << 8)
83 MTHCA_EVENT_TYPE_COMP = 0x00,
84 MTHCA_EVENT_TYPE_PATH_MIG = 0x01,
85 MTHCA_EVENT_TYPE_COMM_EST = 0x02,
86 MTHCA_EVENT_TYPE_SQ_DRAINED = 0x03,
87 MTHCA_EVENT_TYPE_SRQ_QP_LAST_WQE = 0x13,
88 MTHCA_EVENT_TYPE_SRQ_LIMIT = 0x14,
89 MTHCA_EVENT_TYPE_CQ_ERROR = 0x04,
90 MTHCA_EVENT_TYPE_WQ_CATAS_ERROR = 0x05,
91 MTHCA_EVENT_TYPE_EEC_CATAS_ERROR = 0x06,
92 MTHCA_EVENT_TYPE_PATH_MIG_FAILED = 0x07,
93 MTHCA_EVENT_TYPE_WQ_INVAL_REQ_ERROR = 0x10,
94 MTHCA_EVENT_TYPE_WQ_ACCESS_ERROR = 0x11,
95 MTHCA_EVENT_TYPE_SRQ_CATAS_ERROR = 0x12,
96 MTHCA_EVENT_TYPE_LOCAL_CATAS_ERROR = 0x08,
97 MTHCA_EVENT_TYPE_PORT_CHANGE = 0x09,
98 MTHCA_EVENT_TYPE_EQ_OVERFLOW = 0x0f,
99 MTHCA_EVENT_TYPE_ECC_DETECT = 0x0e,
100 MTHCA_EVENT_TYPE_CMD = 0x0a
103 #define MTHCA_ASYNC_EVENT_MASK ((1ULL << MTHCA_EVENT_TYPE_PATH_MIG) | \
104 (1ULL << MTHCA_EVENT_TYPE_COMM_EST) | \
105 (1ULL << MTHCA_EVENT_TYPE_SQ_DRAINED) | \
106 (1ULL << MTHCA_EVENT_TYPE_CQ_ERROR) | \
107 (1ULL << MTHCA_EVENT_TYPE_WQ_CATAS_ERROR) | \
108 (1ULL << MTHCA_EVENT_TYPE_EEC_CATAS_ERROR) | \
109 (1ULL << MTHCA_EVENT_TYPE_PATH_MIG_FAILED) | \
110 (1ULL << MTHCA_EVENT_TYPE_WQ_INVAL_REQ_ERROR) | \
111 (1ULL << MTHCA_EVENT_TYPE_WQ_ACCESS_ERROR) | \
112 (1ULL << MTHCA_EVENT_TYPE_LOCAL_CATAS_ERROR) | \
113 (1ULL << MTHCA_EVENT_TYPE_PORT_CHANGE) | \
114 (1ULL << MTHCA_EVENT_TYPE_ECC_DETECT))
115 #define MTHCA_SRQ_EVENT_MASK ((1ULL << MTHCA_EVENT_TYPE_SRQ_CATAS_ERROR) | \
116 (1ULL << MTHCA_EVENT_TYPE_SRQ_QP_LAST_WQE) | \
117 (1ULL << MTHCA_EVENT_TYPE_SRQ_LIMIT))
118 #define MTHCA_CMD_EVENT_MASK (1ULL << MTHCA_EVENT_TYPE_CMD)
120 #define MTHCA_EQ_DB_INC_CI (1 << 24)
121 #define MTHCA_EQ_DB_REQ_NOT (2 << 24)
122 #define MTHCA_EQ_DB_DISARM_CQ (3 << 24)
123 #define MTHCA_EQ_DB_SET_CI (4 << 24)
124 #define MTHCA_EQ_DB_ALWAYS_ARM (5 << 24)
135 } __attribute__((packed)) comp;
143 } __attribute__((packed)) cmd;
146 } __attribute__((packed)) qp;
149 } __attribute__((packed)) srq;
155 } __attribute__((packed)) cq_err;
159 } __attribute__((packed)) port_change;
163 } __attribute__((packed));
165 #define MTHCA_EQ_ENTRY_OWNER_SW (0 << 7)
166 #define MTHCA_EQ_ENTRY_OWNER_HW (1 << 7)
168 static inline u64 async_mask(struct mthca_dev *dev)
170 return dev->mthca_flags & MTHCA_FLAG_SRQ ?
171 MTHCA_ASYNC_EVENT_MASK | MTHCA_SRQ_EVENT_MASK :
172 MTHCA_ASYNC_EVENT_MASK;
175 static inline void tavor_set_eq_ci(struct mthca_dev *dev, struct mthca_eq *eq, u32 ci)
179 doorbell[0] = cpu_to_be32(MTHCA_EQ_DB_SET_CI | eq->eqn);
180 doorbell[1] = cpu_to_be32(ci & (eq->nent - 1));
183 * This barrier makes sure that all updates to ownership bits
184 * done by set_eqe_hw() hit memory before the consumer index
185 * is updated. set_eq_ci() allows the HCA to possibly write
186 * more EQ entries, and we want to avoid the exceedingly
187 * unlikely possibility of the HCA writing an entry and then
188 * having set_eqe_hw() overwrite the owner field.
191 mthca_write64(doorbell,
192 dev->kar + MTHCA_EQ_DOORBELL,
193 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
196 static inline void arbel_set_eq_ci(struct mthca_dev *dev, struct mthca_eq *eq, u32 ci)
198 /* See comment in tavor_set_eq_ci() above. */
200 __raw_writel((__force u32) cpu_to_be32(ci),
201 dev->eq_regs.arbel.eq_set_ci_base + eq->eqn * 8);
202 /* We still want ordering, just not swabbing, so add a barrier */
206 static inline void set_eq_ci(struct mthca_dev *dev, struct mthca_eq *eq, u32 ci)
208 if (mthca_is_memfree(dev))
209 arbel_set_eq_ci(dev, eq, ci);
211 tavor_set_eq_ci(dev, eq, ci);
214 static inline void tavor_eq_req_not(struct mthca_dev *dev, int eqn)
218 doorbell[0] = cpu_to_be32(MTHCA_EQ_DB_REQ_NOT | eqn);
221 mthca_write64(doorbell,
222 dev->kar + MTHCA_EQ_DOORBELL,
223 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
226 static inline void arbel_eq_req_not(struct mthca_dev *dev, u32 eqn_mask)
228 writel(eqn_mask, dev->eq_regs.arbel.eq_arm);
231 static inline void disarm_cq(struct mthca_dev *dev, int eqn, int cqn)
233 if (!mthca_is_memfree(dev)) {
236 doorbell[0] = cpu_to_be32(MTHCA_EQ_DB_DISARM_CQ | eqn);
237 doorbell[1] = cpu_to_be32(cqn);
239 mthca_write64(doorbell,
240 dev->kar + MTHCA_EQ_DOORBELL,
241 MTHCA_GET_DOORBELL_LOCK(&dev->doorbell_lock));
245 static inline struct mthca_eqe *get_eqe(struct mthca_eq *eq, u32 entry)
247 unsigned long off = (entry & (eq->nent - 1)) * MTHCA_EQ_ENTRY_SIZE;
248 return eq->page_list[off / PAGE_SIZE].buf + off % PAGE_SIZE;
251 static inline struct mthca_eqe* next_eqe_sw(struct mthca_eq *eq)
253 struct mthca_eqe* eqe;
254 eqe = get_eqe(eq, eq->cons_index);
255 return (MTHCA_EQ_ENTRY_OWNER_HW & eqe->owner) ? NULL : eqe;
258 static inline void set_eqe_hw(struct mthca_eqe *eqe)
260 eqe->owner = MTHCA_EQ_ENTRY_OWNER_HW;
263 static void port_change(struct mthca_dev *dev, int port, int active)
265 struct ib_event record;
267 mthca_dbg(dev, "Port change to %s for port %d\n",
268 active ? "active" : "down", port);
270 record.device = &dev->ib_dev;
271 record.event = active ? IB_EVENT_PORT_ACTIVE : IB_EVENT_PORT_ERR;
272 record.element.port_num = port;
274 ib_dispatch_event(&record);
277 static int mthca_eq_int(struct mthca_dev *dev, struct mthca_eq *eq)
279 struct mthca_eqe *eqe;
284 while ((eqe = next_eqe_sw(eq))) {
286 * Make sure we read EQ entry contents after we've
287 * checked the ownership bit.
292 case MTHCA_EVENT_TYPE_COMP:
293 disarm_cqn = be32_to_cpu(eqe->event.comp.cqn) & 0xffffff;
294 disarm_cq(dev, eq->eqn, disarm_cqn);
295 mthca_cq_completion(dev, disarm_cqn);
298 case MTHCA_EVENT_TYPE_PATH_MIG:
299 mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
303 case MTHCA_EVENT_TYPE_COMM_EST:
304 mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
308 case MTHCA_EVENT_TYPE_SQ_DRAINED:
309 mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
310 IB_EVENT_SQ_DRAINED);
313 case MTHCA_EVENT_TYPE_SRQ_QP_LAST_WQE:
314 mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
315 IB_EVENT_QP_LAST_WQE_REACHED);
318 case MTHCA_EVENT_TYPE_SRQ_LIMIT:
319 mthca_srq_event(dev, be32_to_cpu(eqe->event.srq.srqn) & 0xffffff,
320 IB_EVENT_SRQ_LIMIT_REACHED);
323 case MTHCA_EVENT_TYPE_WQ_CATAS_ERROR:
324 mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
328 case MTHCA_EVENT_TYPE_PATH_MIG_FAILED:
329 mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
330 IB_EVENT_PATH_MIG_ERR);
333 case MTHCA_EVENT_TYPE_WQ_INVAL_REQ_ERROR:
334 mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
335 IB_EVENT_QP_REQ_ERR);
338 case MTHCA_EVENT_TYPE_WQ_ACCESS_ERROR:
339 mthca_qp_event(dev, be32_to_cpu(eqe->event.qp.qpn) & 0xffffff,
340 IB_EVENT_QP_ACCESS_ERR);
343 case MTHCA_EVENT_TYPE_CMD:
345 be16_to_cpu(eqe->event.cmd.token),
346 eqe->event.cmd.status,
347 be64_to_cpu(eqe->event.cmd.out_param));
350 case MTHCA_EVENT_TYPE_PORT_CHANGE:
352 (be32_to_cpu(eqe->event.port_change.port) >> 28) & 3,
353 eqe->subtype == 0x4);
356 case MTHCA_EVENT_TYPE_CQ_ERROR:
357 mthca_warn(dev, "CQ %s on CQN %06x\n",
358 eqe->event.cq_err.syndrome == 1 ?
359 "overrun" : "access violation",
360 be32_to_cpu(eqe->event.cq_err.cqn) & 0xffffff);
361 mthca_cq_event(dev, be32_to_cpu(eqe->event.cq_err.cqn),
365 case MTHCA_EVENT_TYPE_EQ_OVERFLOW:
366 mthca_warn(dev, "EQ overrun on EQN %d\n", eq->eqn);
369 case MTHCA_EVENT_TYPE_EEC_CATAS_ERROR:
370 case MTHCA_EVENT_TYPE_SRQ_CATAS_ERROR:
371 case MTHCA_EVENT_TYPE_LOCAL_CATAS_ERROR:
372 case MTHCA_EVENT_TYPE_ECC_DETECT:
374 mthca_warn(dev, "Unhandled event %02x(%02x) on EQ %d\n",
375 eqe->type, eqe->subtype, eq->eqn);
385 * The HCA will think the queue has overflowed if we
386 * don't tell it we've been processing events. We
387 * create our EQs with MTHCA_NUM_SPARE_EQE extra
388 * entries, so we must update our consumer index at
391 if (unlikely(set_ci >= MTHCA_NUM_SPARE_EQE)) {
393 * Conditional on hca_type is OK here because
394 * this is a rare case, not the fast path.
396 set_eq_ci(dev, eq, eq->cons_index);
402 * Rely on caller to set consumer index so that we don't have
403 * to test hca_type in our interrupt handling fast path.
408 static irqreturn_t mthca_tavor_interrupt(int irq, void *dev_ptr, struct pt_regs *regs)
410 struct mthca_dev *dev = dev_ptr;
414 if (dev->eq_table.clr_mask)
415 writel(dev->eq_table.clr_mask, dev->eq_table.clr_int);
417 ecr = readl(dev->eq_regs.tavor.ecr_base + 4);
421 writel(ecr, dev->eq_regs.tavor.ecr_base +
422 MTHCA_ECR_CLR_BASE - MTHCA_ECR_BASE + 4);
424 for (i = 0; i < MTHCA_NUM_EQ; ++i)
425 if (ecr & dev->eq_table.eq[i].eqn_mask) {
426 if (mthca_eq_int(dev, &dev->eq_table.eq[i]))
427 tavor_set_eq_ci(dev, &dev->eq_table.eq[i],
428 dev->eq_table.eq[i].cons_index);
429 tavor_eq_req_not(dev, dev->eq_table.eq[i].eqn);
435 static irqreturn_t mthca_tavor_msi_x_interrupt(int irq, void *eq_ptr,
436 struct pt_regs *regs)
438 struct mthca_eq *eq = eq_ptr;
439 struct mthca_dev *dev = eq->dev;
441 mthca_eq_int(dev, eq);
442 tavor_set_eq_ci(dev, eq, eq->cons_index);
443 tavor_eq_req_not(dev, eq->eqn);
445 /* MSI-X vectors always belong to us */
449 static irqreturn_t mthca_arbel_interrupt(int irq, void *dev_ptr, struct pt_regs *regs)
451 struct mthca_dev *dev = dev_ptr;
455 if (dev->eq_table.clr_mask)
456 writel(dev->eq_table.clr_mask, dev->eq_table.clr_int);
458 for (i = 0; i < MTHCA_NUM_EQ; ++i)
459 if (mthca_eq_int(dev, &dev->eq_table.eq[i])) {
461 arbel_set_eq_ci(dev, &dev->eq_table.eq[i],
462 dev->eq_table.eq[i].cons_index);
465 arbel_eq_req_not(dev, dev->eq_table.arm_mask);
467 return IRQ_RETVAL(work);
470 static irqreturn_t mthca_arbel_msi_x_interrupt(int irq, void *eq_ptr,
471 struct pt_regs *regs)
473 struct mthca_eq *eq = eq_ptr;
474 struct mthca_dev *dev = eq->dev;
476 mthca_eq_int(dev, eq);
477 arbel_set_eq_ci(dev, eq, eq->cons_index);
478 arbel_eq_req_not(dev, eq->eqn_mask);
480 /* MSI-X vectors always belong to us */
484 static int __devinit mthca_create_eq(struct mthca_dev *dev,
490 u64 *dma_list = NULL;
492 struct mthca_mailbox *mailbox;
493 struct mthca_eq_context *eq_context;
499 eq->nent = roundup_pow_of_two(max(nent, 2));
500 npages = ALIGN(eq->nent * MTHCA_EQ_ENTRY_SIZE, PAGE_SIZE) / PAGE_SIZE;
502 eq->page_list = kmalloc(npages * sizeof *eq->page_list,
507 for (i = 0; i < npages; ++i)
508 eq->page_list[i].buf = NULL;
510 dma_list = kmalloc(npages * sizeof *dma_list, GFP_KERNEL);
514 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
517 eq_context = mailbox->buf;
519 for (i = 0; i < npages; ++i) {
520 eq->page_list[i].buf = dma_alloc_coherent(&dev->pdev->dev,
521 PAGE_SIZE, &t, GFP_KERNEL);
522 if (!eq->page_list[i].buf)
523 goto err_out_free_pages;
526 pci_unmap_addr_set(&eq->page_list[i], mapping, t);
528 memset(eq->page_list[i].buf, 0, PAGE_SIZE);
531 for (i = 0; i < eq->nent; ++i)
532 set_eqe_hw(get_eqe(eq, i));
534 eq->eqn = mthca_alloc(&dev->eq_table.alloc);
536 goto err_out_free_pages;
538 err = mthca_mr_alloc_phys(dev, dev->driver_pd.pd_num,
539 dma_list, PAGE_SHIFT, npages,
540 0, npages * PAGE_SIZE,
541 MTHCA_MPT_FLAG_LOCAL_WRITE |
542 MTHCA_MPT_FLAG_LOCAL_READ,
545 goto err_out_free_eq;
547 memset(eq_context, 0, sizeof *eq_context);
548 eq_context->flags = cpu_to_be32(MTHCA_EQ_STATUS_OK |
550 MTHCA_EQ_STATE_ARMED |
552 if (mthca_is_memfree(dev))
553 eq_context->flags |= cpu_to_be32(MTHCA_EQ_STATE_ARBEL);
555 eq_context->logsize_usrpage = cpu_to_be32((ffs(eq->nent) - 1) << 24);
556 if (mthca_is_memfree(dev)) {
557 eq_context->arbel_pd = cpu_to_be32(dev->driver_pd.pd_num);
559 eq_context->logsize_usrpage |= cpu_to_be32(dev->driver_uar.index);
560 eq_context->tavor_pd = cpu_to_be32(dev->driver_pd.pd_num);
562 eq_context->intr = intr;
563 eq_context->lkey = cpu_to_be32(eq->mr.ibmr.lkey);
565 err = mthca_SW2HW_EQ(dev, mailbox, eq->eqn, &status);
567 mthca_warn(dev, "SW2HW_EQ failed (%d)\n", err);
568 goto err_out_free_mr;
571 mthca_warn(dev, "SW2HW_EQ returned status 0x%02x\n",
574 goto err_out_free_mr;
578 mthca_free_mailbox(dev, mailbox);
580 eq->eqn_mask = swab32(1 << eq->eqn);
583 dev->eq_table.arm_mask |= eq->eqn_mask;
585 mthca_dbg(dev, "Allocated EQ %d with %d entries\n",
591 mthca_free_mr(dev, &eq->mr);
594 mthca_free(&dev->eq_table.alloc, eq->eqn);
597 for (i = 0; i < npages; ++i)
598 if (eq->page_list[i].buf)
599 dma_free_coherent(&dev->pdev->dev, PAGE_SIZE,
600 eq->page_list[i].buf,
601 pci_unmap_addr(&eq->page_list[i],
604 mthca_free_mailbox(dev, mailbox);
607 kfree(eq->page_list);
614 static void mthca_free_eq(struct mthca_dev *dev,
617 struct mthca_mailbox *mailbox;
620 int npages = (eq->nent * MTHCA_EQ_ENTRY_SIZE + PAGE_SIZE - 1) /
624 mailbox = mthca_alloc_mailbox(dev, GFP_KERNEL);
628 err = mthca_HW2SW_EQ(dev, mailbox, eq->eqn, &status);
630 mthca_warn(dev, "HW2SW_EQ failed (%d)\n", err);
632 mthca_warn(dev, "HW2SW_EQ returned status 0x%02x\n", status);
634 dev->eq_table.arm_mask &= ~eq->eqn_mask;
637 mthca_dbg(dev, "Dumping EQ context %02x:\n", eq->eqn);
638 for (i = 0; i < sizeof (struct mthca_eq_context) / 4; ++i) {
640 printk("[%02x] ", i * 4);
641 printk(" %08x", be32_to_cpup(mailbox->buf + i * 4));
642 if ((i + 1) % 4 == 0)
647 mthca_free_mr(dev, &eq->mr);
648 for (i = 0; i < npages; ++i)
649 pci_free_consistent(dev->pdev, PAGE_SIZE,
650 eq->page_list[i].buf,
651 pci_unmap_addr(&eq->page_list[i], mapping));
653 kfree(eq->page_list);
654 mthca_free_mailbox(dev, mailbox);
657 static void mthca_free_irqs(struct mthca_dev *dev)
661 if (dev->eq_table.have_irq)
662 free_irq(dev->pdev->irq, dev);
663 for (i = 0; i < MTHCA_NUM_EQ; ++i)
664 if (dev->eq_table.eq[i].have_irq)
665 free_irq(dev->eq_table.eq[i].msi_x_vector,
666 dev->eq_table.eq + i);
669 static int __devinit mthca_map_reg(struct mthca_dev *dev,
670 unsigned long offset, unsigned long size,
673 unsigned long base = pci_resource_start(dev->pdev, 0);
675 if (!request_mem_region(base + offset, size, DRV_NAME))
678 *map = ioremap(base + offset, size);
680 release_mem_region(base + offset, size);
687 static void mthca_unmap_reg(struct mthca_dev *dev, unsigned long offset,
688 unsigned long size, void __iomem *map)
690 unsigned long base = pci_resource_start(dev->pdev, 0);
692 release_mem_region(base + offset, size);
696 static int __devinit mthca_map_eq_regs(struct mthca_dev *dev)
698 unsigned long mthca_base;
700 mthca_base = pci_resource_start(dev->pdev, 0);
702 if (mthca_is_memfree(dev)) {
704 * We assume that the EQ arm and EQ set CI registers
705 * fall within the first BAR. We can't trust the
706 * values firmware gives us, since those addresses are
707 * valid on the HCA's side of the PCI bus but not
708 * necessarily the host side.
710 if (mthca_map_reg(dev, (pci_resource_len(dev->pdev, 0) - 1) &
711 dev->fw.arbel.clr_int_base, MTHCA_CLR_INT_SIZE,
713 mthca_err(dev, "Couldn't map interrupt clear register, "
719 * Add 4 because we limit ourselves to EQs 0 ... 31,
720 * so we only need the low word of the register.
722 if (mthca_map_reg(dev, ((pci_resource_len(dev->pdev, 0) - 1) &
723 dev->fw.arbel.eq_arm_base) + 4, 4,
724 &dev->eq_regs.arbel.eq_arm)) {
725 mthca_err(dev, "Couldn't map EQ arm register, aborting.\n");
726 mthca_unmap_reg(dev, (pci_resource_len(dev->pdev, 0) - 1) &
727 dev->fw.arbel.clr_int_base, MTHCA_CLR_INT_SIZE,
732 if (mthca_map_reg(dev, (pci_resource_len(dev->pdev, 0) - 1) &
733 dev->fw.arbel.eq_set_ci_base,
734 MTHCA_EQ_SET_CI_SIZE,
735 &dev->eq_regs.arbel.eq_set_ci_base)) {
736 mthca_err(dev, "Couldn't map EQ CI register, aborting.\n");
737 mthca_unmap_reg(dev, ((pci_resource_len(dev->pdev, 0) - 1) &
738 dev->fw.arbel.eq_arm_base) + 4, 4,
739 dev->eq_regs.arbel.eq_arm);
740 mthca_unmap_reg(dev, (pci_resource_len(dev->pdev, 0) - 1) &
741 dev->fw.arbel.clr_int_base, MTHCA_CLR_INT_SIZE,
746 if (mthca_map_reg(dev, MTHCA_CLR_INT_BASE, MTHCA_CLR_INT_SIZE,
748 mthca_err(dev, "Couldn't map interrupt clear register, "
753 if (mthca_map_reg(dev, MTHCA_ECR_BASE,
754 MTHCA_ECR_SIZE + MTHCA_ECR_CLR_SIZE,
755 &dev->eq_regs.tavor.ecr_base)) {
756 mthca_err(dev, "Couldn't map ecr register, "
758 mthca_unmap_reg(dev, MTHCA_CLR_INT_BASE, MTHCA_CLR_INT_SIZE,
768 static void mthca_unmap_eq_regs(struct mthca_dev *dev)
770 if (mthca_is_memfree(dev)) {
771 mthca_unmap_reg(dev, (pci_resource_len(dev->pdev, 0) - 1) &
772 dev->fw.arbel.eq_set_ci_base,
773 MTHCA_EQ_SET_CI_SIZE,
774 dev->eq_regs.arbel.eq_set_ci_base);
775 mthca_unmap_reg(dev, ((pci_resource_len(dev->pdev, 0) - 1) &
776 dev->fw.arbel.eq_arm_base) + 4, 4,
777 dev->eq_regs.arbel.eq_arm);
778 mthca_unmap_reg(dev, (pci_resource_len(dev->pdev, 0) - 1) &
779 dev->fw.arbel.clr_int_base, MTHCA_CLR_INT_SIZE,
782 mthca_unmap_reg(dev, MTHCA_ECR_BASE,
783 MTHCA_ECR_SIZE + MTHCA_ECR_CLR_SIZE,
784 dev->eq_regs.tavor.ecr_base);
785 mthca_unmap_reg(dev, MTHCA_CLR_INT_BASE, MTHCA_CLR_INT_SIZE,
790 int __devinit mthca_map_eq_icm(struct mthca_dev *dev, u64 icm_virt)
796 * We assume that mapping one page is enough for the whole EQ
797 * context table. This is fine with all current HCAs, because
798 * we only use 32 EQs and each EQ uses 32 bytes of context
799 * memory, or 1 KB total.
801 dev->eq_table.icm_virt = icm_virt;
802 dev->eq_table.icm_page = alloc_page(GFP_HIGHUSER);
803 if (!dev->eq_table.icm_page)
805 dev->eq_table.icm_dma = pci_map_page(dev->pdev, dev->eq_table.icm_page, 0,
806 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
807 if (pci_dma_mapping_error(dev->eq_table.icm_dma)) {
808 __free_page(dev->eq_table.icm_page);
812 ret = mthca_MAP_ICM_page(dev, dev->eq_table.icm_dma, icm_virt, &status);
816 pci_unmap_page(dev->pdev, dev->eq_table.icm_dma, PAGE_SIZE,
817 PCI_DMA_BIDIRECTIONAL);
818 __free_page(dev->eq_table.icm_page);
824 void mthca_unmap_eq_icm(struct mthca_dev *dev)
828 mthca_UNMAP_ICM(dev, dev->eq_table.icm_virt, 1, &status);
829 pci_unmap_page(dev->pdev, dev->eq_table.icm_dma, PAGE_SIZE,
830 PCI_DMA_BIDIRECTIONAL);
831 __free_page(dev->eq_table.icm_page);
834 int __devinit mthca_init_eq_table(struct mthca_dev *dev)
841 err = mthca_alloc_init(&dev->eq_table.alloc,
843 dev->limits.num_eqs - 1,
844 dev->limits.reserved_eqs);
848 err = mthca_map_eq_regs(dev);
852 if (dev->mthca_flags & MTHCA_FLAG_MSI ||
853 dev->mthca_flags & MTHCA_FLAG_MSI_X) {
854 dev->eq_table.clr_mask = 0;
856 dev->eq_table.clr_mask =
857 swab32(1 << (dev->eq_table.inta_pin & 31));
858 dev->eq_table.clr_int = dev->clr_base +
859 (dev->eq_table.inta_pin < 32 ? 4 : 0);
862 dev->eq_table.arm_mask = 0;
864 intr = (dev->mthca_flags & MTHCA_FLAG_MSI) ?
865 128 : dev->eq_table.inta_pin;
867 err = mthca_create_eq(dev, dev->limits.num_cqs + MTHCA_NUM_SPARE_EQE,
868 (dev->mthca_flags & MTHCA_FLAG_MSI_X) ? 128 : intr,
869 &dev->eq_table.eq[MTHCA_EQ_COMP]);
873 err = mthca_create_eq(dev, MTHCA_NUM_ASYNC_EQE + MTHCA_NUM_SPARE_EQE,
874 (dev->mthca_flags & MTHCA_FLAG_MSI_X) ? 129 : intr,
875 &dev->eq_table.eq[MTHCA_EQ_ASYNC]);
879 err = mthca_create_eq(dev, MTHCA_NUM_CMD_EQE + MTHCA_NUM_SPARE_EQE,
880 (dev->mthca_flags & MTHCA_FLAG_MSI_X) ? 130 : intr,
881 &dev->eq_table.eq[MTHCA_EQ_CMD]);
885 if (dev->mthca_flags & MTHCA_FLAG_MSI_X) {
886 static const char *eq_name[] = {
887 [MTHCA_EQ_COMP] = DRV_NAME " (comp)",
888 [MTHCA_EQ_ASYNC] = DRV_NAME " (async)",
889 [MTHCA_EQ_CMD] = DRV_NAME " (cmd)"
892 for (i = 0; i < MTHCA_NUM_EQ; ++i) {
893 err = request_irq(dev->eq_table.eq[i].msi_x_vector,
894 mthca_is_memfree(dev) ?
895 mthca_arbel_msi_x_interrupt :
896 mthca_tavor_msi_x_interrupt,
897 0, eq_name[i], dev->eq_table.eq + i);
900 dev->eq_table.eq[i].have_irq = 1;
903 err = request_irq(dev->pdev->irq,
904 mthca_is_memfree(dev) ?
905 mthca_arbel_interrupt :
906 mthca_tavor_interrupt,
907 SA_SHIRQ, DRV_NAME, dev);
910 dev->eq_table.have_irq = 1;
913 err = mthca_MAP_EQ(dev, async_mask(dev),
914 0, dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn, &status);
916 mthca_warn(dev, "MAP_EQ for async EQ %d failed (%d)\n",
917 dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn, err);
919 mthca_warn(dev, "MAP_EQ for async EQ %d returned status 0x%02x\n",
920 dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn, status);
922 err = mthca_MAP_EQ(dev, MTHCA_CMD_EVENT_MASK,
923 0, dev->eq_table.eq[MTHCA_EQ_CMD].eqn, &status);
925 mthca_warn(dev, "MAP_EQ for cmd EQ %d failed (%d)\n",
926 dev->eq_table.eq[MTHCA_EQ_CMD].eqn, err);
928 mthca_warn(dev, "MAP_EQ for cmd EQ %d returned status 0x%02x\n",
929 dev->eq_table.eq[MTHCA_EQ_CMD].eqn, status);
931 for (i = 0; i < MTHCA_NUM_EQ; ++i)
932 if (mthca_is_memfree(dev))
933 arbel_eq_req_not(dev, dev->eq_table.eq[i].eqn_mask);
935 tavor_eq_req_not(dev, dev->eq_table.eq[i].eqn);
940 mthca_free_irqs(dev);
941 mthca_free_eq(dev, &dev->eq_table.eq[MTHCA_EQ_CMD]);
944 mthca_free_eq(dev, &dev->eq_table.eq[MTHCA_EQ_ASYNC]);
947 mthca_free_eq(dev, &dev->eq_table.eq[MTHCA_EQ_COMP]);
950 mthca_unmap_eq_regs(dev);
953 mthca_alloc_cleanup(&dev->eq_table.alloc);
957 void mthca_cleanup_eq_table(struct mthca_dev *dev)
962 mthca_free_irqs(dev);
964 mthca_MAP_EQ(dev, async_mask(dev),
965 1, dev->eq_table.eq[MTHCA_EQ_ASYNC].eqn, &status);
966 mthca_MAP_EQ(dev, MTHCA_CMD_EVENT_MASK,
967 1, dev->eq_table.eq[MTHCA_EQ_CMD].eqn, &status);
969 for (i = 0; i < MTHCA_NUM_EQ; ++i)
970 mthca_free_eq(dev, &dev->eq_table.eq[i]);
972 mthca_unmap_eq_regs(dev);
974 mthca_alloc_cleanup(&dev->eq_table.alloc);