2 * Copyright (c) 2012, 2013 Intel Corporation. All rights reserved.
3 * Copyright (c) 2006 - 2012 QLogic Corporation. All rights reserved.
4 * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. 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/pci.h>
36 #include <linux/netdevice.h>
37 #include <linux/vmalloc.h>
38 #include <linux/delay.h>
39 #include <linux/idr.h>
40 #include <linux/module.h>
41 #include <linux/printk.h>
42 #ifdef CONFIG_INFINIBAND_QIB_DCA
43 #include <linux/dca.h>
47 #include "qib_common.h"
51 #define pr_fmt(fmt) QIB_DRV_NAME ": " fmt
54 * min buffers we want to have per context, after driver
56 #define QIB_MIN_USER_CTXT_BUFCNT 7
58 #define QLOGIC_IB_R_SOFTWARE_MASK 0xFF
59 #define QLOGIC_IB_R_SOFTWARE_SHIFT 24
60 #define QLOGIC_IB_R_EMULATOR_MASK (1ULL<<62)
63 * Number of ctxts we are configured to use (to allow for more pio
64 * buffers per ctxt, etc.) Zero means use chip value.
67 module_param_named(cfgctxts, qib_cfgctxts, ushort, S_IRUGO);
68 MODULE_PARM_DESC(cfgctxts, "Set max number of contexts to use");
70 unsigned qib_numa_aware;
71 module_param_named(numa_aware, qib_numa_aware, uint, S_IRUGO);
72 MODULE_PARM_DESC(numa_aware,
73 "0 -> PSM allocation close to HCA, 1 -> PSM allocation local to process");
76 * If set, do not write to any regs if avoidable, hack to allow
77 * check for deranged default register values.
80 module_param_named(mini_init, qib_mini_init, ushort, S_IRUGO);
81 MODULE_PARM_DESC(mini_init, "If set, do minimal diag init");
83 unsigned qib_n_krcv_queues;
84 module_param_named(krcvqs, qib_n_krcv_queues, uint, S_IRUGO);
85 MODULE_PARM_DESC(krcvqs, "number of kernel receive queues per IB port");
87 unsigned qib_cc_table_size;
88 module_param_named(cc_table_size, qib_cc_table_size, uint, S_IRUGO);
89 MODULE_PARM_DESC(cc_table_size, "Congestion control table entries 0 (CCA disabled - default), min = 128, max = 1984");
91 * qib_wc_pat parameter:
94 * If PAT initialization fails, code reverts back to MTRR
96 unsigned qib_wc_pat = 1; /* default (1) is to use PAT, not MTRR */
97 module_param_named(wc_pat, qib_wc_pat, uint, S_IRUGO);
98 MODULE_PARM_DESC(wc_pat, "enable write-combining via PAT mechanism");
100 struct workqueue_struct *qib_cq_wq;
102 static void verify_interrupt(unsigned long);
104 static struct idr qib_unit_table;
105 u32 qib_cpulist_count;
106 unsigned long *qib_cpulist;
108 /* set number of contexts we'll actually use */
109 void qib_set_ctxtcnt(struct qib_devdata *dd)
112 dd->cfgctxts = dd->first_user_ctxt + num_online_cpus();
113 if (dd->cfgctxts > dd->ctxtcnt)
114 dd->cfgctxts = dd->ctxtcnt;
115 } else if (qib_cfgctxts < dd->num_pports)
116 dd->cfgctxts = dd->ctxtcnt;
117 else if (qib_cfgctxts <= dd->ctxtcnt)
118 dd->cfgctxts = qib_cfgctxts;
120 dd->cfgctxts = dd->ctxtcnt;
121 dd->freectxts = (dd->first_user_ctxt > dd->cfgctxts) ? 0 :
122 dd->cfgctxts - dd->first_user_ctxt;
126 * Common code for creating the receive context array.
128 int qib_create_ctxts(struct qib_devdata *dd)
132 int local_node_id = pcibus_to_node(dd->pcidev->bus);
134 if (local_node_id < 0)
135 local_node_id = numa_node_id();
136 dd->assigned_node_id = local_node_id;
139 * Allocate full ctxtcnt array, rather than just cfgctxts, because
140 * cleanup iterates across all possible ctxts.
142 dd->rcd = kzalloc(sizeof(*dd->rcd) * dd->ctxtcnt, GFP_KERNEL);
145 "Unable to allocate ctxtdata array, failing\n");
150 /* create (one or more) kctxt */
151 for (i = 0; i < dd->first_user_ctxt; ++i) {
152 struct qib_pportdata *ppd;
153 struct qib_ctxtdata *rcd;
155 if (dd->skip_kctxt_mask & (1 << i))
158 ppd = dd->pport + (i % dd->num_pports);
160 rcd = qib_create_ctxtdata(ppd, i, dd->assigned_node_id);
163 "Unable to allocate ctxtdata for Kernel ctxt, failing\n");
167 rcd->pkeys[0] = QIB_DEFAULT_P_KEY;
176 * Common code for user and kernel context setup.
178 struct qib_ctxtdata *qib_create_ctxtdata(struct qib_pportdata *ppd, u32 ctxt,
181 struct qib_devdata *dd = ppd->dd;
182 struct qib_ctxtdata *rcd;
184 rcd = kzalloc_node(sizeof(*rcd), GFP_KERNEL, node_id);
186 INIT_LIST_HEAD(&rcd->qp_wait_list);
187 rcd->node_id = node_id;
194 dd->f_init_ctxt(rcd);
197 * To avoid wasting a lot of memory, we allocate 32KB chunks
198 * of physically contiguous memory, advance through it until
199 * used up and then allocate more. Of course, we need
200 * memory to store those extra pointers, now. 32KB seems to
201 * be the most that is "safe" under memory pressure
202 * (creating large files and then copying them over
203 * NFS while doing lots of MPI jobs). The OOM killer can
204 * get invoked, even though we say we can sleep and this can
205 * cause significant system problems....
207 rcd->rcvegrbuf_size = 0x8000;
208 rcd->rcvegrbufs_perchunk =
209 rcd->rcvegrbuf_size / dd->rcvegrbufsize;
210 rcd->rcvegrbuf_chunks = (rcd->rcvegrcnt +
211 rcd->rcvegrbufs_perchunk - 1) /
212 rcd->rcvegrbufs_perchunk;
213 BUG_ON(!is_power_of_2(rcd->rcvegrbufs_perchunk));
214 rcd->rcvegrbufs_perchunk_shift =
215 ilog2(rcd->rcvegrbufs_perchunk);
221 * Common code for initializing the physical port structure.
223 void qib_init_pportdata(struct qib_pportdata *ppd, struct qib_devdata *dd,
228 ppd->hw_pidx = hw_pidx;
229 ppd->port = port; /* IB port number, not index */
231 spin_lock_init(&ppd->sdma_lock);
232 spin_lock_init(&ppd->lflags_lock);
233 init_waitqueue_head(&ppd->state_wait);
235 init_timer(&ppd->symerr_clear_timer);
236 ppd->symerr_clear_timer.function = qib_clear_symerror_on_linkup;
237 ppd->symerr_clear_timer.data = (unsigned long)ppd;
241 spin_lock_init(&ppd->cc_shadow_lock);
243 if (qib_cc_table_size < IB_CCT_MIN_ENTRIES)
246 ppd->cc_supported_table_entries = min(max_t(int, qib_cc_table_size,
247 IB_CCT_MIN_ENTRIES), IB_CCT_ENTRIES*IB_CC_TABLE_CAP_DEFAULT);
249 ppd->cc_max_table_entries =
250 ppd->cc_supported_table_entries/IB_CCT_ENTRIES;
252 size = IB_CC_TABLE_CAP_DEFAULT * sizeof(struct ib_cc_table_entry)
254 ppd->ccti_entries = kzalloc(size, GFP_KERNEL);
255 if (!ppd->ccti_entries) {
257 "failed to allocate congestion control table for port %d!\n",
262 size = IB_CC_CCS_ENTRIES * sizeof(struct ib_cc_congestion_entry);
263 ppd->congestion_entries = kzalloc(size, GFP_KERNEL);
264 if (!ppd->congestion_entries) {
266 "failed to allocate congestion setting list for port %d!\n",
271 size = sizeof(struct cc_table_shadow);
272 ppd->ccti_entries_shadow = kzalloc(size, GFP_KERNEL);
273 if (!ppd->ccti_entries_shadow) {
275 "failed to allocate shadow ccti list for port %d!\n",
280 size = sizeof(struct ib_cc_congestion_setting_attr);
281 ppd->congestion_entries_shadow = kzalloc(size, GFP_KERNEL);
282 if (!ppd->congestion_entries_shadow) {
284 "failed to allocate shadow congestion setting list for port %d!\n",
292 kfree(ppd->ccti_entries_shadow);
293 ppd->ccti_entries_shadow = NULL;
295 kfree(ppd->congestion_entries);
296 ppd->congestion_entries = NULL;
298 kfree(ppd->ccti_entries);
299 ppd->ccti_entries = NULL;
301 /* User is intentionally disabling the congestion control agent */
302 if (!qib_cc_table_size)
305 if (qib_cc_table_size < IB_CCT_MIN_ENTRIES) {
306 qib_cc_table_size = 0;
308 "Congestion Control table size %d less than minimum %d for port %d\n",
309 qib_cc_table_size, IB_CCT_MIN_ENTRIES, port);
312 qib_dev_err(dd, "Congestion Control Agent disabled for port %d\n",
317 static int init_pioavailregs(struct qib_devdata *dd)
322 dd->pioavailregs_dma = dma_alloc_coherent(
323 &dd->pcidev->dev, PAGE_SIZE, &dd->pioavailregs_phys,
325 if (!dd->pioavailregs_dma) {
327 "failed to allocate PIOavail reg area in memory\n");
333 * We really want L2 cache aligned, but for current CPUs of
334 * interest, they are the same.
336 status_page = (u64 *)
337 ((char *) dd->pioavailregs_dma +
338 ((2 * L1_CACHE_BYTES +
339 dd->pioavregs * sizeof(u64)) & ~L1_CACHE_BYTES));
340 /* device status comes first, for backwards compatibility */
341 dd->devstatusp = status_page;
343 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
344 dd->pport[pidx].statusp = status_page;
349 * Setup buffer to hold freeze and other messages, accessible to
350 * apps, following statusp. This is per-unit, not per port.
352 dd->freezemsg = (char *) status_page;
354 /* length of msg buffer is "whatever is left" */
355 ret = (char *) status_page - (char *) dd->pioavailregs_dma;
356 dd->freezelen = PAGE_SIZE - ret;
365 * init_shadow_tids - allocate the shadow TID array
366 * @dd: the qlogic_ib device
368 * allocate the shadow TID array, so we can qib_munlock previous
369 * entries. It may make more sense to move the pageshadow to the
370 * ctxt data structure, so we only allocate memory for ctxts actually
371 * in use, since we at 8k per ctxt, now.
372 * We don't want failures here to prevent use of the driver/chip,
373 * so no return value.
375 static void init_shadow_tids(struct qib_devdata *dd)
380 pages = vzalloc(dd->cfgctxts * dd->rcvtidcnt * sizeof(struct page *));
383 "failed to allocate shadow page * array, no expected sends!\n");
387 addrs = vzalloc(dd->cfgctxts * dd->rcvtidcnt * sizeof(dma_addr_t));
390 "failed to allocate shadow dma handle array, no expected sends!\n");
394 dd->pageshadow = pages;
395 dd->physshadow = addrs;
401 dd->pageshadow = NULL;
405 * Do initialization for device that is only needed on
406 * first detect, not on resets.
408 static int loadtime_init(struct qib_devdata *dd)
412 if (((dd->revision >> QLOGIC_IB_R_SOFTWARE_SHIFT) &
413 QLOGIC_IB_R_SOFTWARE_MASK) != QIB_CHIP_SWVERSION) {
415 "Driver only handles version %d, chip swversion is %d (%llx), failng\n",
417 (int)(dd->revision >>
418 QLOGIC_IB_R_SOFTWARE_SHIFT) &
419 QLOGIC_IB_R_SOFTWARE_MASK,
420 (unsigned long long) dd->revision);
425 if (dd->revision & QLOGIC_IB_R_EMULATOR_MASK)
426 qib_devinfo(dd->pcidev, "%s", dd->boardversion);
428 spin_lock_init(&dd->pioavail_lock);
429 spin_lock_init(&dd->sendctrl_lock);
430 spin_lock_init(&dd->uctxt_lock);
431 spin_lock_init(&dd->qib_diag_trans_lock);
432 spin_lock_init(&dd->eep_st_lock);
433 mutex_init(&dd->eep_lock);
438 ret = init_pioavailregs(dd);
439 init_shadow_tids(dd);
441 qib_get_eeprom_info(dd);
443 /* setup time (don't start yet) to verify we got interrupt */
444 init_timer(&dd->intrchk_timer);
445 dd->intrchk_timer.function = verify_interrupt;
446 dd->intrchk_timer.data = (unsigned long) dd;
453 * init_after_reset - re-initialize after a reset
454 * @dd: the qlogic_ib device
456 * sanity check at least some of the values after reset, and
457 * ensure no receive or transmit (explicitly, in case reset
460 static int init_after_reset(struct qib_devdata *dd)
465 * Ensure chip does no sends or receives, tail updates, or
466 * pioavail updates while we re-initialize. This is mostly
467 * for the driver data structures, not chip registers.
469 for (i = 0; i < dd->num_pports; ++i) {
471 * ctxt == -1 means "all contexts". Only really safe for
472 * _dis_abling things, as here.
474 dd->f_rcvctrl(dd->pport + i, QIB_RCVCTRL_CTXT_DIS |
475 QIB_RCVCTRL_INTRAVAIL_DIS |
476 QIB_RCVCTRL_TAILUPD_DIS, -1);
477 /* Redundant across ports for some, but no big deal. */
478 dd->f_sendctrl(dd->pport + i, QIB_SENDCTRL_SEND_DIS |
479 QIB_SENDCTRL_AVAIL_DIS);
485 static void enable_chip(struct qib_devdata *dd)
491 * Enable PIO send, and update of PIOavail regs to memory.
493 for (i = 0; i < dd->num_pports; ++i)
494 dd->f_sendctrl(dd->pport + i, QIB_SENDCTRL_SEND_ENB |
495 QIB_SENDCTRL_AVAIL_ENB);
497 * Enable kernel ctxts' receive and receive interrupt.
498 * Other ctxts done as user opens and inits them.
500 rcvmask = QIB_RCVCTRL_CTXT_ENB | QIB_RCVCTRL_INTRAVAIL_ENB;
501 rcvmask |= (dd->flags & QIB_NODMA_RTAIL) ?
502 QIB_RCVCTRL_TAILUPD_DIS : QIB_RCVCTRL_TAILUPD_ENB;
503 for (i = 0; dd->rcd && i < dd->first_user_ctxt; ++i) {
504 struct qib_ctxtdata *rcd = dd->rcd[i];
507 dd->f_rcvctrl(rcd->ppd, rcvmask, i);
511 static void verify_interrupt(unsigned long opaque)
513 struct qib_devdata *dd = (struct qib_devdata *) opaque;
516 return; /* being torn down */
519 * If we don't have a lid or any interrupts, let the user know and
520 * don't bother checking again.
522 if (dd->int_counter == 0) {
523 if (!dd->f_intr_fallback(dd))
524 dev_err(&dd->pcidev->dev,
525 "No interrupts detected, not usable.\n");
526 else /* re-arm the timer to see if fallback works */
527 mod_timer(&dd->intrchk_timer, jiffies + HZ/2);
531 static void init_piobuf_state(struct qib_devdata *dd)
537 * Ensure all buffers are free, and fifos empty. Buffers
538 * are common, so only do once for port 0.
540 * After enable and qib_chg_pioavailkernel so we can safely
541 * enable pioavail updates and PIOENABLE. After this, packets
542 * are ready and able to go out.
544 dd->f_sendctrl(dd->pport, QIB_SENDCTRL_DISARM_ALL);
545 for (pidx = 0; pidx < dd->num_pports; ++pidx)
546 dd->f_sendctrl(dd->pport + pidx, QIB_SENDCTRL_FLUSH);
549 * If not all sendbufs are used, add the one to each of the lower
550 * numbered contexts. pbufsctxt and lastctxt_piobuf are
551 * calculated in chip-specific code because it may cause some
552 * chip-specific adjustments to be made.
554 uctxts = dd->cfgctxts - dd->first_user_ctxt;
555 dd->ctxts_extrabuf = dd->pbufsctxt ?
556 dd->lastctxt_piobuf - (dd->pbufsctxt * uctxts) : 0;
559 * Set up the shadow copies of the piobufavail registers,
560 * which we compare against the chip registers for now, and
561 * the in memory DMA'ed copies of the registers.
562 * By now pioavail updates to memory should have occurred, so
563 * copy them into our working/shadow registers; this is in
564 * case something went wrong with abort, but mostly to get the
565 * initial values of the generation bit correct.
567 for (i = 0; i < dd->pioavregs; i++) {
570 tmp = dd->pioavailregs_dma[i];
572 * Don't need to worry about pioavailkernel here
573 * because we will call qib_chg_pioavailkernel() later
574 * in initialization, to busy out buffers as needed.
576 dd->pioavailshadow[i] = le64_to_cpu(tmp);
578 while (i < ARRAY_SIZE(dd->pioavailshadow))
579 dd->pioavailshadow[i++] = 0; /* for debugging sanity */
581 /* after pioavailshadow is setup */
582 qib_chg_pioavailkernel(dd, 0, dd->piobcnt2k + dd->piobcnt4k,
583 TXCHK_CHG_TYPE_KERN, NULL);
584 dd->f_initvl15_bufs(dd);
588 * qib_create_workqueues - create per port workqueues
589 * @dd: the qlogic_ib device
591 static int qib_create_workqueues(struct qib_devdata *dd)
594 struct qib_pportdata *ppd;
596 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
597 ppd = dd->pport + pidx;
599 char wq_name[8]; /* 3 + 2 + 1 + 1 + 1 */
600 snprintf(wq_name, sizeof(wq_name), "qib%d_%d",
603 create_singlethread_workqueue(wq_name);
610 pr_err("create_singlethread_workqueue failed for port %d\n",
612 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
613 ppd = dd->pport + pidx;
615 destroy_workqueue(ppd->qib_wq);
623 * qib_init - do the actual initialization sequence on the chip
624 * @dd: the qlogic_ib device
625 * @reinit: reinitializing, so don't allocate new memory
627 * Do the actual initialization sequence on the chip. This is done
628 * both from the init routine called from the PCI infrastructure, and
629 * when we reset the chip, or detect that it was reset internally,
630 * or it's administratively re-enabled.
632 * Memory allocation here and in called routines is only done in
633 * the first case (reinit == 0). We have to be careful, because even
634 * without memory allocation, we need to re-write all the chip registers
635 * TIDs, etc. after the reset or enable has completed.
637 int qib_init(struct qib_devdata *dd, int reinit)
639 int ret = 0, pidx, lastfail = 0;
642 struct qib_ctxtdata *rcd;
643 struct qib_pportdata *ppd;
646 /* Set linkstate to unknown, so we can watch for a transition. */
647 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
648 ppd = dd->pport + pidx;
649 spin_lock_irqsave(&ppd->lflags_lock, flags);
650 ppd->lflags &= ~(QIBL_LINKACTIVE | QIBL_LINKARMED |
651 QIBL_LINKDOWN | QIBL_LINKINIT |
653 spin_unlock_irqrestore(&ppd->lflags_lock, flags);
657 ret = init_after_reset(dd);
659 ret = loadtime_init(dd);
663 /* Bypass most chip-init, to get to device creation */
667 ret = dd->f_late_initreg(dd);
671 /* dd->rcd can be NULL if early init failed */
672 for (i = 0; dd->rcd && i < dd->first_user_ctxt; ++i) {
674 * Set up the (kernel) rcvhdr queue and egr TIDs. If doing
675 * re-init, the simplest way to handle this is to free
676 * existing, and re-allocate.
677 * Need to re-create rest of ctxt 0 ctxtdata as well.
683 lastfail = qib_create_rcvhdrq(dd, rcd);
685 lastfail = qib_setup_eagerbufs(rcd);
688 "failed to allocate kernel ctxt's rcvhdrq and/or egr bufs\n");
693 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
697 ppd = dd->pport + pidx;
698 mtu = ib_mtu_enum_to_int(qib_ibmtu);
700 mtu = QIB_DEFAULT_MTU;
701 qib_ibmtu = 0; /* don't leave invalid value */
703 /* set max we can ever have for this driver load */
704 ppd->init_ibmaxlen = min(mtu > 2048 ?
705 dd->piosize4k : dd->piosize2k,
707 (dd->rcvhdrentsize << 2));
709 * Have to initialize ibmaxlen, but this will normally
710 * change immediately in qib_set_mtu().
712 ppd->ibmaxlen = ppd->init_ibmaxlen;
713 qib_set_mtu(ppd, mtu);
715 spin_lock_irqsave(&ppd->lflags_lock, flags);
716 ppd->lflags |= QIBL_IB_LINK_DISABLED;
717 spin_unlock_irqrestore(&ppd->lflags_lock, flags);
719 lastfail = dd->f_bringup_serdes(ppd);
721 qib_devinfo(dd->pcidev,
722 "Failed to bringup IB port %u\n", ppd->port);
723 lastfail = -ENETDOWN;
731 /* none of the ports initialized */
732 if (!ret && lastfail)
736 /* but continue on, so we can debug cause */
741 init_piobuf_state(dd);
745 /* chip is OK for user apps; mark it as initialized */
746 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
747 ppd = dd->pport + pidx;
749 * Set status even if port serdes is not initialized
750 * so that diags will work.
752 *ppd->statusp |= QIB_STATUS_CHIP_PRESENT |
754 if (!ppd->link_speed_enabled)
756 if (dd->flags & QIB_HAS_SEND_DMA)
757 ret = qib_setup_sdma(ppd);
758 init_timer(&ppd->hol_timer);
759 ppd->hol_timer.function = qib_hol_event;
760 ppd->hol_timer.data = (unsigned long)ppd;
761 ppd->hol_state = QIB_HOL_UP;
764 /* now we can enable all interrupts from the chip */
765 dd->f_set_intr_state(dd, 1);
768 * Setup to verify we get an interrupt, and fallback
769 * to an alternate if necessary and possible.
771 mod_timer(&dd->intrchk_timer, jiffies + HZ/2);
772 /* start stats retrieval timer */
773 mod_timer(&dd->stats_timer, jiffies + HZ * ACTIVITY_TIMER);
776 /* if ret is non-zero, we probably should do some cleanup here... */
781 * These next two routines are placeholders in case we don't have per-arch
782 * code for controlling write combining. If explicit control of write
783 * combining is not available, performance will probably be awful.
786 int __attribute__((weak)) qib_enable_wc(struct qib_devdata *dd)
791 void __attribute__((weak)) qib_disable_wc(struct qib_devdata *dd)
795 static inline struct qib_devdata *__qib_lookup(int unit)
797 return idr_find(&qib_unit_table, unit);
800 struct qib_devdata *qib_lookup(int unit)
802 struct qib_devdata *dd;
805 spin_lock_irqsave(&qib_devs_lock, flags);
806 dd = __qib_lookup(unit);
807 spin_unlock_irqrestore(&qib_devs_lock, flags);
813 * Stop the timers during unit shutdown, or after an error late
816 static void qib_stop_timers(struct qib_devdata *dd)
818 struct qib_pportdata *ppd;
821 if (dd->stats_timer.data) {
822 del_timer_sync(&dd->stats_timer);
823 dd->stats_timer.data = 0;
825 if (dd->intrchk_timer.data) {
826 del_timer_sync(&dd->intrchk_timer);
827 dd->intrchk_timer.data = 0;
829 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
830 ppd = dd->pport + pidx;
831 if (ppd->hol_timer.data)
832 del_timer_sync(&ppd->hol_timer);
833 if (ppd->led_override_timer.data) {
834 del_timer_sync(&ppd->led_override_timer);
835 atomic_set(&ppd->led_override_timer_active, 0);
837 if (ppd->symerr_clear_timer.data)
838 del_timer_sync(&ppd->symerr_clear_timer);
843 * qib_shutdown_device - shut down a device
844 * @dd: the qlogic_ib device
846 * This is called to make the device quiet when we are about to
847 * unload the driver, and also when the device is administratively
848 * disabled. It does not free any data structures.
849 * Everything it does has to be setup again by qib_init(dd, 1)
851 static void qib_shutdown_device(struct qib_devdata *dd)
853 struct qib_pportdata *ppd;
856 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
857 ppd = dd->pport + pidx;
859 spin_lock_irq(&ppd->lflags_lock);
860 ppd->lflags &= ~(QIBL_LINKDOWN | QIBL_LINKINIT |
861 QIBL_LINKARMED | QIBL_LINKACTIVE |
863 spin_unlock_irq(&ppd->lflags_lock);
864 *ppd->statusp &= ~(QIB_STATUS_IB_CONF | QIB_STATUS_IB_READY);
866 dd->flags &= ~QIB_INITTED;
868 /* mask interrupts, but not errors */
869 dd->f_set_intr_state(dd, 0);
871 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
872 ppd = dd->pport + pidx;
873 dd->f_rcvctrl(ppd, QIB_RCVCTRL_TAILUPD_DIS |
874 QIB_RCVCTRL_CTXT_DIS |
875 QIB_RCVCTRL_INTRAVAIL_DIS |
876 QIB_RCVCTRL_PKEY_ENB, -1);
878 * Gracefully stop all sends allowing any in progress to
881 dd->f_sendctrl(ppd, QIB_SENDCTRL_CLEAR);
885 * Enough for anything that's going to trickle out to have actually
890 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
891 ppd = dd->pport + pidx;
892 dd->f_setextled(ppd, 0); /* make sure LEDs are off */
894 if (dd->flags & QIB_HAS_SEND_DMA)
895 qib_teardown_sdma(ppd);
897 dd->f_sendctrl(ppd, QIB_SENDCTRL_AVAIL_DIS |
898 QIB_SENDCTRL_SEND_DIS);
900 * Clear SerdesEnable.
901 * We can't count on interrupts since we are stopping.
903 dd->f_quiet_serdes(ppd);
906 destroy_workqueue(ppd->qib_wq);
911 qib_update_eeprom_log(dd);
915 * qib_free_ctxtdata - free a context's allocated data
916 * @dd: the qlogic_ib device
917 * @rcd: the ctxtdata structure
919 * free up any allocated data for a context
920 * This should not touch anything that would affect a simultaneous
921 * re-allocation of context data, because it is called after qib_mutex
922 * is released (and can be called from reinit as well).
923 * It should never change any chip state, or global driver state.
925 void qib_free_ctxtdata(struct qib_devdata *dd, struct qib_ctxtdata *rcd)
931 dma_free_coherent(&dd->pcidev->dev, rcd->rcvhdrq_size,
932 rcd->rcvhdrq, rcd->rcvhdrq_phys);
934 if (rcd->rcvhdrtail_kvaddr) {
935 dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE,
936 rcd->rcvhdrtail_kvaddr,
937 rcd->rcvhdrqtailaddr_phys);
938 rcd->rcvhdrtail_kvaddr = NULL;
941 if (rcd->rcvegrbuf) {
944 for (e = 0; e < rcd->rcvegrbuf_chunks; e++) {
945 void *base = rcd->rcvegrbuf[e];
946 size_t size = rcd->rcvegrbuf_size;
948 dma_free_coherent(&dd->pcidev->dev, size,
949 base, rcd->rcvegrbuf_phys[e]);
951 kfree(rcd->rcvegrbuf);
952 rcd->rcvegrbuf = NULL;
953 kfree(rcd->rcvegrbuf_phys);
954 rcd->rcvegrbuf_phys = NULL;
955 rcd->rcvegrbuf_chunks = 0;
958 kfree(rcd->tid_pg_list);
959 vfree(rcd->user_event_mask);
960 vfree(rcd->subctxt_uregbase);
961 vfree(rcd->subctxt_rcvegrbuf);
962 vfree(rcd->subctxt_rcvhdr_base);
967 * Perform a PIO buffer bandwidth write test, to verify proper system
968 * configuration. Even when all the setup calls work, occasionally
969 * BIOS or other issues can prevent write combining from working, or
970 * can cause other bandwidth problems to the chip.
972 * This test simply writes the same buffer over and over again, and
973 * measures close to the peak bandwidth to the chip (not testing
974 * data bandwidth to the wire). On chips that use an address-based
975 * trigger to send packets to the wire, this is easy. On chips that
976 * use a count to trigger, we want to make sure that the packet doesn't
977 * go out on the wire, or trigger flow control checks.
979 static void qib_verify_pioperf(struct qib_devdata *dd)
981 u32 pbnum, cnt, lcnt;
986 piobuf = dd->f_getsendbuf(dd->pport, 0ULL, &pbnum);
988 qib_devinfo(dd->pcidev,
989 "No PIObufs for checking perf, skipping\n");
994 * Enough to give us a reasonable test, less than piobuf size, and
995 * likely multiple of store buffer length.
1001 qib_devinfo(dd->pcidev,
1002 "Couldn't get memory for checking PIO perf,"
1007 preempt_disable(); /* we want reasonably accurate elapsed time */
1008 msecs = 1 + jiffies_to_msecs(jiffies);
1009 for (lcnt = 0; lcnt < 10000U; lcnt++) {
1010 /* wait until we cross msec boundary */
1011 if (jiffies_to_msecs(jiffies) >= msecs)
1016 dd->f_set_armlaunch(dd, 0);
1019 * length 0, no dwords actually sent
1025 * This is only roughly accurate, since even with preempt we
1026 * still take interrupts that could take a while. Running for
1027 * >= 5 msec seems to get us "close enough" to accurate values.
1029 msecs = jiffies_to_msecs(jiffies);
1030 for (emsecs = lcnt = 0; emsecs <= 5UL; lcnt++) {
1031 qib_pio_copy(piobuf + 64, addr, cnt >> 2);
1032 emsecs = jiffies_to_msecs(jiffies) - msecs;
1035 /* 1 GiB/sec, slightly over IB SDR line rate */
1036 if (lcnt < (emsecs * 1024U))
1038 "Performance problem: bandwidth to PIO buffers is only %u MiB/sec\n",
1039 lcnt / (u32) emsecs);
1046 /* disarm piobuf, so it's available again */
1047 dd->f_sendctrl(dd->pport, QIB_SENDCTRL_DISARM_BUF(pbnum));
1048 qib_sendbuf_done(dd, pbnum);
1049 dd->f_set_armlaunch(dd, 1);
1053 void qib_free_devdata(struct qib_devdata *dd)
1055 unsigned long flags;
1057 spin_lock_irqsave(&qib_devs_lock, flags);
1058 idr_remove(&qib_unit_table, dd->unit);
1059 list_del(&dd->list);
1060 spin_unlock_irqrestore(&qib_devs_lock, flags);
1062 ib_dealloc_device(&dd->verbs_dev.ibdev);
1066 * Allocate our primary per-unit data structure. Must be done via verbs
1067 * allocator, because the verbs cleanup process both does cleanup and
1068 * free of the data structure.
1069 * "extra" is for chip-specific data.
1071 * Use the idr mechanism to get a unit number for this unit.
1073 struct qib_devdata *qib_alloc_devdata(struct pci_dev *pdev, size_t extra)
1075 unsigned long flags;
1076 struct qib_devdata *dd;
1079 dd = (struct qib_devdata *) ib_alloc_device(sizeof(*dd) + extra);
1081 dd = ERR_PTR(-ENOMEM);
1085 idr_preload(GFP_KERNEL);
1086 spin_lock_irqsave(&qib_devs_lock, flags);
1088 ret = idr_alloc(&qib_unit_table, dd, 0, 0, GFP_NOWAIT);
1091 list_add(&dd->list, &qib_dev_list);
1094 spin_unlock_irqrestore(&qib_devs_lock, flags);
1098 qib_early_err(&pdev->dev,
1099 "Could not allocate unit ID: error %d\n", -ret);
1100 ib_dealloc_device(&dd->verbs_dev.ibdev);
1105 if (!qib_cpulist_count) {
1106 u32 count = num_online_cpus();
1107 qib_cpulist = kzalloc(BITS_TO_LONGS(count) *
1108 sizeof(long), GFP_KERNEL);
1110 qib_cpulist_count = count;
1112 qib_early_err(&pdev->dev,
1113 "Could not alloc cpulist info, cpu affinity might be wrong\n");
1121 * Called from freeze mode handlers, and from PCI error
1122 * reporting code. Should be paranoid about state of
1123 * system and data structures.
1125 void qib_disable_after_error(struct qib_devdata *dd)
1127 if (dd->flags & QIB_INITTED) {
1130 dd->flags &= ~QIB_INITTED;
1132 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
1133 struct qib_pportdata *ppd;
1135 ppd = dd->pport + pidx;
1136 if (dd->flags & QIB_PRESENT) {
1137 qib_set_linkstate(ppd,
1138 QIB_IB_LINKDOWN_DISABLE);
1139 dd->f_setextled(ppd, 0);
1141 *ppd->statusp &= ~QIB_STATUS_IB_READY;
1146 * Mark as having had an error for driver, and also
1147 * for /sys and status word mapped to user programs.
1148 * This marks unit as not usable, until reset.
1151 *dd->devstatusp |= QIB_STATUS_HWERROR;
1154 static void qib_remove_one(struct pci_dev *);
1155 static int qib_init_one(struct pci_dev *, const struct pci_device_id *);
1157 #define DRIVER_LOAD_MSG "Intel " QIB_DRV_NAME " loaded: "
1158 #define PFX QIB_DRV_NAME ": "
1160 static DEFINE_PCI_DEVICE_TABLE(qib_pci_tbl) = {
1161 { PCI_DEVICE(PCI_VENDOR_ID_PATHSCALE, PCI_DEVICE_ID_QLOGIC_IB_6120) },
1162 { PCI_DEVICE(PCI_VENDOR_ID_QLOGIC, PCI_DEVICE_ID_QLOGIC_IB_7220) },
1163 { PCI_DEVICE(PCI_VENDOR_ID_QLOGIC, PCI_DEVICE_ID_QLOGIC_IB_7322) },
1167 MODULE_DEVICE_TABLE(pci, qib_pci_tbl);
1169 struct pci_driver qib_driver = {
1170 .name = QIB_DRV_NAME,
1171 .probe = qib_init_one,
1172 .remove = qib_remove_one,
1173 .id_table = qib_pci_tbl,
1174 .err_handler = &qib_pci_err_handler,
1177 #ifdef CONFIG_INFINIBAND_QIB_DCA
1179 static int qib_notify_dca(struct notifier_block *, unsigned long, void *);
1180 static struct notifier_block dca_notifier = {
1181 .notifier_call = qib_notify_dca,
1186 static int qib_notify_dca_device(struct device *device, void *data)
1188 struct qib_devdata *dd = dev_get_drvdata(device);
1189 unsigned long event = *(unsigned long *)data;
1191 return dd->f_notify_dca(dd, event);
1194 static int qib_notify_dca(struct notifier_block *nb, unsigned long event,
1199 rval = driver_for_each_device(&qib_driver.driver, NULL,
1200 &event, qib_notify_dca_device);
1201 return rval ? NOTIFY_BAD : NOTIFY_DONE;
1207 * Do all the generic driver unit- and chip-independent memory
1208 * allocation and initialization.
1210 static int __init qlogic_ib_init(void)
1214 ret = qib_dev_init();
1218 qib_cq_wq = create_singlethread_workqueue("qib_cq");
1225 * These must be called before the driver is registered with
1226 * the PCI subsystem.
1228 idr_init(&qib_unit_table);
1230 #ifdef CONFIG_INFINIBAND_QIB_DCA
1231 dca_register_notify(&dca_notifier);
1233 ret = pci_register_driver(&qib_driver);
1235 pr_err("Unable to register driver: error %d\n", -ret);
1239 /* not fatal if it doesn't work */
1240 if (qib_init_qibfs())
1241 pr_err("Unable to register ipathfs\n");
1242 goto bail; /* all OK */
1245 #ifdef CONFIG_INFINIBAND_QIB_DCA
1246 dca_unregister_notify(&dca_notifier);
1248 idr_destroy(&qib_unit_table);
1249 destroy_workqueue(qib_cq_wq);
1256 module_init(qlogic_ib_init);
1259 * Do the non-unit driver cleanup, memory free, etc. at unload.
1261 static void __exit qlogic_ib_cleanup(void)
1265 ret = qib_exit_qibfs();
1268 "Unable to cleanup counter filesystem: error %d\n",
1271 #ifdef CONFIG_INFINIBAND_QIB_DCA
1272 dca_unregister_notify(&dca_notifier);
1274 pci_unregister_driver(&qib_driver);
1276 destroy_workqueue(qib_cq_wq);
1278 qib_cpulist_count = 0;
1281 idr_destroy(&qib_unit_table);
1285 module_exit(qlogic_ib_cleanup);
1287 /* this can only be called after a successful initialization */
1288 static void cleanup_device_data(struct qib_devdata *dd)
1292 struct qib_ctxtdata **tmp;
1293 unsigned long flags;
1295 /* users can't do anything more with chip */
1296 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
1297 if (dd->pport[pidx].statusp)
1298 *dd->pport[pidx].statusp &= ~QIB_STATUS_CHIP_PRESENT;
1300 spin_lock(&dd->pport[pidx].cc_shadow_lock);
1302 kfree(dd->pport[pidx].congestion_entries);
1303 dd->pport[pidx].congestion_entries = NULL;
1304 kfree(dd->pport[pidx].ccti_entries);
1305 dd->pport[pidx].ccti_entries = NULL;
1306 kfree(dd->pport[pidx].ccti_entries_shadow);
1307 dd->pport[pidx].ccti_entries_shadow = NULL;
1308 kfree(dd->pport[pidx].congestion_entries_shadow);
1309 dd->pport[pidx].congestion_entries_shadow = NULL;
1311 spin_unlock(&dd->pport[pidx].cc_shadow_lock);
1317 if (dd->pioavailregs_dma) {
1318 dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE,
1319 (void *) dd->pioavailregs_dma,
1320 dd->pioavailregs_phys);
1321 dd->pioavailregs_dma = NULL;
1324 if (dd->pageshadow) {
1325 struct page **tmpp = dd->pageshadow;
1326 dma_addr_t *tmpd = dd->physshadow;
1329 for (ctxt = 0; ctxt < dd->cfgctxts; ctxt++) {
1330 int ctxt_tidbase = ctxt * dd->rcvtidcnt;
1331 int maxtid = ctxt_tidbase + dd->rcvtidcnt;
1333 for (i = ctxt_tidbase; i < maxtid; i++) {
1336 pci_unmap_page(dd->pcidev, tmpd[i],
1337 PAGE_SIZE, PCI_DMA_FROMDEVICE);
1338 qib_release_user_pages(&tmpp[i], 1);
1344 tmpp = dd->pageshadow;
1345 dd->pageshadow = NULL;
1350 * Free any resources still in use (usually just kernel contexts)
1351 * at unload; we do for ctxtcnt, because that's what we allocate.
1352 * We acquire lock to be really paranoid that rcd isn't being
1353 * accessed from some interrupt-related code (that should not happen,
1354 * but best to be sure).
1356 spin_lock_irqsave(&dd->uctxt_lock, flags);
1359 spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1360 for (ctxt = 0; tmp && ctxt < dd->ctxtcnt; ctxt++) {
1361 struct qib_ctxtdata *rcd = tmp[ctxt];
1363 tmp[ctxt] = NULL; /* debugging paranoia */
1364 qib_free_ctxtdata(dd, rcd);
1367 kfree(dd->boardname);
1371 * Clean up on unit shutdown, or error during unit load after
1372 * successful initialization.
1374 static void qib_postinit_cleanup(struct qib_devdata *dd)
1377 * Clean up chip-specific stuff.
1378 * We check for NULL here, because it's outside
1379 * the kregbase check, and we need to call it
1380 * after the free_irq. Thus it's possible that
1381 * the function pointers were never initialized.
1386 qib_pcie_ddcleanup(dd);
1388 cleanup_device_data(dd);
1390 qib_free_devdata(dd);
1393 static int qib_init_one(struct pci_dev *pdev, const struct pci_device_id *ent)
1395 int ret, j, pidx, initfail;
1396 struct qib_devdata *dd = NULL;
1398 ret = qib_pcie_init(pdev, ent);
1403 * Do device-specific initialiation, function table setup, dd
1406 switch (ent->device) {
1407 case PCI_DEVICE_ID_QLOGIC_IB_6120:
1408 #ifdef CONFIG_PCI_MSI
1409 dd = qib_init_iba6120_funcs(pdev, ent);
1411 qib_early_err(&pdev->dev,
1412 "Intel PCIE device 0x%x cannot work if CONFIG_PCI_MSI is not enabled\n",
1414 dd = ERR_PTR(-ENODEV);
1418 case PCI_DEVICE_ID_QLOGIC_IB_7220:
1419 dd = qib_init_iba7220_funcs(pdev, ent);
1422 case PCI_DEVICE_ID_QLOGIC_IB_7322:
1423 dd = qib_init_iba7322_funcs(pdev, ent);
1427 qib_early_err(&pdev->dev,
1428 "Failing on unknown Intel deviceid 0x%x\n",
1436 goto bail; /* error already printed */
1438 ret = qib_create_workqueues(dd);
1442 /* do the generic initialization */
1443 initfail = qib_init(dd, 0);
1445 ret = qib_register_ib_device(dd);
1448 * Now ready for use. this should be cleared whenever we
1449 * detect a reset, or initiate one. If earlier failure,
1450 * we still create devices, so diags, etc. can be used
1451 * to determine cause of problem.
1453 if (!qib_mini_init && !initfail && !ret)
1454 dd->flags |= QIB_INITTED;
1456 j = qib_device_create(dd);
1458 qib_dev_err(dd, "Failed to create /dev devices: %d\n", -j);
1461 qib_dev_err(dd, "Failed filesystem setup for counters: %d\n",
1464 if (qib_mini_init || initfail || ret) {
1465 qib_stop_timers(dd);
1466 flush_workqueue(ib_wq);
1467 for (pidx = 0; pidx < dd->num_pports; ++pidx)
1468 dd->f_quiet_serdes(dd->pport + pidx);
1472 (void) qibfs_remove(dd);
1473 qib_device_remove(dd);
1476 qib_unregister_ib_device(dd);
1477 qib_postinit_cleanup(dd);
1484 ret = qib_enable_wc(dd);
1487 "Write combining not enabled (err %d): performance may be poor\n",
1493 qib_verify_pioperf(dd);
1498 static void qib_remove_one(struct pci_dev *pdev)
1500 struct qib_devdata *dd = pci_get_drvdata(pdev);
1503 /* unregister from IB core */
1504 qib_unregister_ib_device(dd);
1507 * Disable the IB link, disable interrupts on the device,
1508 * clear dma engines, etc.
1511 qib_shutdown_device(dd);
1513 qib_stop_timers(dd);
1515 /* wait until all of our (qsfp) queue_work() calls complete */
1516 flush_workqueue(ib_wq);
1518 ret = qibfs_remove(dd);
1520 qib_dev_err(dd, "Failed counters filesystem cleanup: %d\n",
1523 qib_device_remove(dd);
1525 qib_postinit_cleanup(dd);
1529 * qib_create_rcvhdrq - create a receive header queue
1530 * @dd: the qlogic_ib device
1531 * @rcd: the context data
1533 * This must be contiguous memory (from an i/o perspective), and must be
1534 * DMA'able (which means for some systems, it will go through an IOMMU,
1535 * or be forced into a low address range).
1537 int qib_create_rcvhdrq(struct qib_devdata *dd, struct qib_ctxtdata *rcd)
1542 if (!rcd->rcvhdrq) {
1543 dma_addr_t phys_hdrqtail;
1546 amt = ALIGN(dd->rcvhdrcnt * dd->rcvhdrentsize *
1547 sizeof(u32), PAGE_SIZE);
1548 gfp_flags = (rcd->ctxt >= dd->first_user_ctxt) ?
1549 GFP_USER : GFP_KERNEL;
1551 old_node_id = dev_to_node(&dd->pcidev->dev);
1552 set_dev_node(&dd->pcidev->dev, rcd->node_id);
1553 rcd->rcvhdrq = dma_alloc_coherent(
1554 &dd->pcidev->dev, amt, &rcd->rcvhdrq_phys,
1555 gfp_flags | __GFP_COMP);
1556 set_dev_node(&dd->pcidev->dev, old_node_id);
1558 if (!rcd->rcvhdrq) {
1560 "attempt to allocate %d bytes for ctxt %u rcvhdrq failed\n",
1565 if (rcd->ctxt >= dd->first_user_ctxt) {
1566 rcd->user_event_mask = vmalloc_user(PAGE_SIZE);
1567 if (!rcd->user_event_mask)
1568 goto bail_free_hdrq;
1571 if (!(dd->flags & QIB_NODMA_RTAIL)) {
1572 set_dev_node(&dd->pcidev->dev, rcd->node_id);
1573 rcd->rcvhdrtail_kvaddr = dma_alloc_coherent(
1574 &dd->pcidev->dev, PAGE_SIZE, &phys_hdrqtail,
1576 set_dev_node(&dd->pcidev->dev, old_node_id);
1577 if (!rcd->rcvhdrtail_kvaddr)
1579 rcd->rcvhdrqtailaddr_phys = phys_hdrqtail;
1582 rcd->rcvhdrq_size = amt;
1585 /* clear for security and sanity on each use */
1586 memset(rcd->rcvhdrq, 0, rcd->rcvhdrq_size);
1587 if (rcd->rcvhdrtail_kvaddr)
1588 memset(rcd->rcvhdrtail_kvaddr, 0, PAGE_SIZE);
1593 "attempt to allocate 1 page for ctxt %u rcvhdrqtailaddr failed\n",
1595 vfree(rcd->user_event_mask);
1596 rcd->user_event_mask = NULL;
1598 dma_free_coherent(&dd->pcidev->dev, amt, rcd->rcvhdrq,
1600 rcd->rcvhdrq = NULL;
1606 * allocate eager buffers, both kernel and user contexts.
1607 * @rcd: the context we are setting up.
1609 * Allocate the eager TID buffers and program them into hip.
1610 * They are no longer completely contiguous, we do multiple allocation
1611 * calls. Otherwise we get the OOM code involved, by asking for too
1612 * much per call, with disastrous results on some kernels.
1614 int qib_setup_eagerbufs(struct qib_ctxtdata *rcd)
1616 struct qib_devdata *dd = rcd->dd;
1617 unsigned e, egrcnt, egrperchunk, chunk, egrsize, egroff;
1623 * GFP_USER, but without GFP_FS, so buffer cache can be
1624 * coalesced (we hope); otherwise, even at order 4,
1625 * heavy filesystem activity makes these fail, and we can
1626 * use compound pages.
1628 gfp_flags = __GFP_WAIT | __GFP_IO | __GFP_COMP;
1630 egrcnt = rcd->rcvegrcnt;
1631 egroff = rcd->rcvegr_tid_base;
1632 egrsize = dd->rcvegrbufsize;
1634 chunk = rcd->rcvegrbuf_chunks;
1635 egrperchunk = rcd->rcvegrbufs_perchunk;
1636 size = rcd->rcvegrbuf_size;
1637 if (!rcd->rcvegrbuf) {
1639 kzalloc_node(chunk * sizeof(rcd->rcvegrbuf[0]),
1640 GFP_KERNEL, rcd->node_id);
1641 if (!rcd->rcvegrbuf)
1644 if (!rcd->rcvegrbuf_phys) {
1645 rcd->rcvegrbuf_phys =
1646 kmalloc_node(chunk * sizeof(rcd->rcvegrbuf_phys[0]),
1647 GFP_KERNEL, rcd->node_id);
1648 if (!rcd->rcvegrbuf_phys)
1649 goto bail_rcvegrbuf;
1651 for (e = 0; e < rcd->rcvegrbuf_chunks; e++) {
1652 if (rcd->rcvegrbuf[e])
1655 old_node_id = dev_to_node(&dd->pcidev->dev);
1656 set_dev_node(&dd->pcidev->dev, rcd->node_id);
1658 dma_alloc_coherent(&dd->pcidev->dev, size,
1659 &rcd->rcvegrbuf_phys[e],
1661 set_dev_node(&dd->pcidev->dev, old_node_id);
1662 if (!rcd->rcvegrbuf[e])
1663 goto bail_rcvegrbuf_phys;
1666 rcd->rcvegr_phys = rcd->rcvegrbuf_phys[0];
1668 for (e = chunk = 0; chunk < rcd->rcvegrbuf_chunks; chunk++) {
1669 dma_addr_t pa = rcd->rcvegrbuf_phys[chunk];
1672 /* clear for security and sanity on each use */
1673 memset(rcd->rcvegrbuf[chunk], 0, size);
1675 for (i = 0; e < egrcnt && i < egrperchunk; e++, i++) {
1676 dd->f_put_tid(dd, e + egroff +
1681 RCVHQ_RCV_TYPE_EAGER, pa);
1684 cond_resched(); /* don't hog the cpu */
1689 bail_rcvegrbuf_phys:
1690 for (e = 0; e < rcd->rcvegrbuf_chunks && rcd->rcvegrbuf[e]; e++)
1691 dma_free_coherent(&dd->pcidev->dev, size,
1692 rcd->rcvegrbuf[e], rcd->rcvegrbuf_phys[e]);
1693 kfree(rcd->rcvegrbuf_phys);
1694 rcd->rcvegrbuf_phys = NULL;
1696 kfree(rcd->rcvegrbuf);
1697 rcd->rcvegrbuf = NULL;
1703 * Note: Changes to this routine should be mirrored
1704 * for the diagnostics routine qib_remap_ioaddr32().
1705 * There is also related code for VL15 buffers in qib_init_7322_variables().
1706 * The teardown code that unmaps is in qib_pcie_ddcleanup()
1708 int init_chip_wc_pat(struct qib_devdata *dd, u32 vl15buflen)
1710 u64 __iomem *qib_kregbase = NULL;
1711 void __iomem *qib_piobase = NULL;
1712 u64 __iomem *qib_userbase = NULL;
1714 u64 qib_pio2koffset = dd->piobufbase & 0xffffffff;
1715 u64 qib_pio4koffset = dd->piobufbase >> 32;
1716 u64 qib_pio2klen = dd->piobcnt2k * dd->palign;
1717 u64 qib_pio4klen = dd->piobcnt4k * dd->align4k;
1718 u64 qib_physaddr = dd->physaddr;
1720 u64 qib_userlen = 0;
1723 * Free the old mapping because the kernel will try to reuse the
1724 * old mapping and not create a new mapping with the
1725 * write combining attribute.
1727 iounmap(dd->kregbase);
1728 dd->kregbase = NULL;
1731 * Assumes chip address space looks like:
1732 * - kregs + sregs + cregs + uregs (in any order)
1733 * - piobufs (2K and 4K bufs in either order)
1735 * - kregs + sregs + cregs (in any order)
1736 * - piobufs (2K and 4K bufs in either order)
1739 if (dd->piobcnt4k == 0) {
1740 qib_kreglen = qib_pio2koffset;
1741 qib_piolen = qib_pio2klen;
1742 } else if (qib_pio2koffset < qib_pio4koffset) {
1743 qib_kreglen = qib_pio2koffset;
1744 qib_piolen = qib_pio4koffset + qib_pio4klen - qib_kreglen;
1746 qib_kreglen = qib_pio4koffset;
1747 qib_piolen = qib_pio2koffset + qib_pio2klen - qib_kreglen;
1749 qib_piolen += vl15buflen;
1750 /* Map just the configured ports (not all hw ports) */
1751 if (dd->uregbase > qib_kreglen)
1752 qib_userlen = dd->ureg_align * dd->cfgctxts;
1754 /* Sanity checks passed, now create the new mappings */
1755 qib_kregbase = ioremap_nocache(qib_physaddr, qib_kreglen);
1759 qib_piobase = ioremap_wc(qib_physaddr + qib_kreglen, qib_piolen);
1764 qib_userbase = ioremap_nocache(qib_physaddr + dd->uregbase,
1770 dd->kregbase = qib_kregbase;
1771 dd->kregend = (u64 __iomem *)
1772 ((char __iomem *) qib_kregbase + qib_kreglen);
1773 dd->piobase = qib_piobase;
1774 dd->pio2kbase = (void __iomem *)
1775 (((char __iomem *) dd->piobase) +
1776 qib_pio2koffset - qib_kreglen);
1778 dd->pio4kbase = (void __iomem *)
1779 (((char __iomem *) dd->piobase) +
1780 qib_pio4koffset - qib_kreglen);
1782 /* ureg will now be accessed relative to dd->userbase */
1783 dd->userbase = qib_userbase;
1787 iounmap(qib_piobase);
1789 iounmap(qib_kregbase);