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>
45 #include <rdma/rdma_vt.h>
48 #include "qib_common.h"
50 #ifdef CONFIG_DEBUG_FS
51 #include "qib_debugfs.h"
52 #include "qib_verbs.h"
56 #define pr_fmt(fmt) QIB_DRV_NAME ": " fmt
59 * min buffers we want to have per context, after driver
61 #define QIB_MIN_USER_CTXT_BUFCNT 7
63 #define QLOGIC_IB_R_SOFTWARE_MASK 0xFF
64 #define QLOGIC_IB_R_SOFTWARE_SHIFT 24
65 #define QLOGIC_IB_R_EMULATOR_MASK (1ULL<<62)
68 * Number of ctxts we are configured to use (to allow for more pio
69 * buffers per ctxt, etc.) Zero means use chip value.
72 module_param_named(cfgctxts, qib_cfgctxts, ushort, S_IRUGO);
73 MODULE_PARM_DESC(cfgctxts, "Set max number of contexts to use");
75 unsigned qib_numa_aware;
76 module_param_named(numa_aware, qib_numa_aware, uint, S_IRUGO);
77 MODULE_PARM_DESC(numa_aware,
78 "0 -> PSM allocation close to HCA, 1 -> PSM allocation local to process");
81 * If set, do not write to any regs if avoidable, hack to allow
82 * check for deranged default register values.
85 module_param_named(mini_init, qib_mini_init, ushort, S_IRUGO);
86 MODULE_PARM_DESC(mini_init, "If set, do minimal diag init");
88 unsigned qib_n_krcv_queues;
89 module_param_named(krcvqs, qib_n_krcv_queues, uint, S_IRUGO);
90 MODULE_PARM_DESC(krcvqs, "number of kernel receive queues per IB port");
92 unsigned qib_cc_table_size;
93 module_param_named(cc_table_size, qib_cc_table_size, uint, S_IRUGO);
94 MODULE_PARM_DESC(cc_table_size, "Congestion control table entries 0 (CCA disabled - default), min = 128, max = 1984");
96 static void verify_interrupt(struct timer_list *);
98 static struct idr qib_unit_table;
99 u32 qib_cpulist_count;
100 unsigned long *qib_cpulist;
102 /* set number of contexts we'll actually use */
103 void qib_set_ctxtcnt(struct qib_devdata *dd)
106 dd->cfgctxts = dd->first_user_ctxt + num_online_cpus();
107 if (dd->cfgctxts > dd->ctxtcnt)
108 dd->cfgctxts = dd->ctxtcnt;
109 } else if (qib_cfgctxts < dd->num_pports)
110 dd->cfgctxts = dd->ctxtcnt;
111 else if (qib_cfgctxts <= dd->ctxtcnt)
112 dd->cfgctxts = qib_cfgctxts;
114 dd->cfgctxts = dd->ctxtcnt;
115 dd->freectxts = (dd->first_user_ctxt > dd->cfgctxts) ? 0 :
116 dd->cfgctxts - dd->first_user_ctxt;
120 * Common code for creating the receive context array.
122 int qib_create_ctxts(struct qib_devdata *dd)
125 int local_node_id = pcibus_to_node(dd->pcidev->bus);
127 if (local_node_id < 0)
128 local_node_id = numa_node_id();
129 dd->assigned_node_id = local_node_id;
132 * Allocate full ctxtcnt array, rather than just cfgctxts, because
133 * cleanup iterates across all possible ctxts.
135 dd->rcd = kcalloc(dd->ctxtcnt, sizeof(*dd->rcd), GFP_KERNEL);
139 /* create (one or more) kctxt */
140 for (i = 0; i < dd->first_user_ctxt; ++i) {
141 struct qib_pportdata *ppd;
142 struct qib_ctxtdata *rcd;
144 if (dd->skip_kctxt_mask & (1 << i))
147 ppd = dd->pport + (i % dd->num_pports);
149 rcd = qib_create_ctxtdata(ppd, i, dd->assigned_node_id);
152 "Unable to allocate ctxtdata for Kernel ctxt, failing\n");
157 rcd->pkeys[0] = QIB_DEFAULT_P_KEY;
164 * Common code for user and kernel context setup.
166 struct qib_ctxtdata *qib_create_ctxtdata(struct qib_pportdata *ppd, u32 ctxt,
169 struct qib_devdata *dd = ppd->dd;
170 struct qib_ctxtdata *rcd;
172 rcd = kzalloc_node(sizeof(*rcd), GFP_KERNEL, node_id);
174 INIT_LIST_HEAD(&rcd->qp_wait_list);
175 rcd->node_id = node_id;
181 #ifdef CONFIG_DEBUG_FS
182 if (ctxt < dd->first_user_ctxt) { /* N/A for PSM contexts */
183 rcd->opstats = kzalloc_node(sizeof(*rcd->opstats),
184 GFP_KERNEL, node_id);
188 "Unable to allocate per ctxt stats buffer\n");
193 dd->f_init_ctxt(rcd);
196 * To avoid wasting a lot of memory, we allocate 32KB chunks
197 * of physically contiguous memory, advance through it until
198 * used up and then allocate more. Of course, we need
199 * memory to store those extra pointers, now. 32KB seems to
200 * be the most that is "safe" under memory pressure
201 * (creating large files and then copying them over
202 * NFS while doing lots of MPI jobs). The OOM killer can
203 * get invoked, even though we say we can sleep and this can
204 * cause significant system problems....
206 rcd->rcvegrbuf_size = 0x8000;
207 rcd->rcvegrbufs_perchunk =
208 rcd->rcvegrbuf_size / dd->rcvegrbufsize;
209 rcd->rcvegrbuf_chunks = (rcd->rcvegrcnt +
210 rcd->rcvegrbufs_perchunk - 1) /
211 rcd->rcvegrbufs_perchunk;
212 BUG_ON(!is_power_of_2(rcd->rcvegrbufs_perchunk));
213 rcd->rcvegrbufs_perchunk_shift =
214 ilog2(rcd->rcvegrbufs_perchunk);
220 * Common code for initializing the physical port structure.
222 int 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 spin_lock_init(&ppd->cc_shadow_lock);
234 init_waitqueue_head(&ppd->state_wait);
236 timer_setup(&ppd->symerr_clear_timer, qib_clear_symerror_on_linkup, 0);
239 ppd->ibport_data.pmastats =
240 alloc_percpu(struct qib_pma_counters);
241 if (!ppd->ibport_data.pmastats)
243 ppd->ibport_data.rvp.rc_acks = alloc_percpu(u64);
244 ppd->ibport_data.rvp.rc_qacks = alloc_percpu(u64);
245 ppd->ibport_data.rvp.rc_delayed_comp = alloc_percpu(u64);
246 if (!(ppd->ibport_data.rvp.rc_acks) ||
247 !(ppd->ibport_data.rvp.rc_qacks) ||
248 !(ppd->ibport_data.rvp.rc_delayed_comp))
251 if (qib_cc_table_size < IB_CCT_MIN_ENTRIES)
254 ppd->cc_supported_table_entries = min(max_t(int, qib_cc_table_size,
255 IB_CCT_MIN_ENTRIES), IB_CCT_ENTRIES*IB_CC_TABLE_CAP_DEFAULT);
257 ppd->cc_max_table_entries =
258 ppd->cc_supported_table_entries/IB_CCT_ENTRIES;
260 size = IB_CC_TABLE_CAP_DEFAULT * sizeof(struct ib_cc_table_entry)
262 ppd->ccti_entries = kzalloc(size, GFP_KERNEL);
263 if (!ppd->ccti_entries)
266 size = IB_CC_CCS_ENTRIES * sizeof(struct ib_cc_congestion_entry);
267 ppd->congestion_entries = kzalloc(size, GFP_KERNEL);
268 if (!ppd->congestion_entries)
271 size = sizeof(struct cc_table_shadow);
272 ppd->ccti_entries_shadow = kzalloc(size, GFP_KERNEL);
273 if (!ppd->ccti_entries_shadow)
276 size = sizeof(struct ib_cc_congestion_setting_attr);
277 ppd->congestion_entries_shadow = kzalloc(size, GFP_KERNEL);
278 if (!ppd->congestion_entries_shadow)
284 kfree(ppd->ccti_entries_shadow);
285 ppd->ccti_entries_shadow = NULL;
287 kfree(ppd->congestion_entries);
288 ppd->congestion_entries = NULL;
290 kfree(ppd->ccti_entries);
291 ppd->ccti_entries = NULL;
293 /* User is intentionally disabling the congestion control agent */
294 if (!qib_cc_table_size)
297 if (qib_cc_table_size < IB_CCT_MIN_ENTRIES) {
298 qib_cc_table_size = 0;
300 "Congestion Control table size %d less than minimum %d for port %d\n",
301 qib_cc_table_size, IB_CCT_MIN_ENTRIES, port);
304 qib_dev_err(dd, "Congestion Control Agent disabled for port %d\n",
309 static int init_pioavailregs(struct qib_devdata *dd)
314 dd->pioavailregs_dma = dma_alloc_coherent(
315 &dd->pcidev->dev, PAGE_SIZE, &dd->pioavailregs_phys,
317 if (!dd->pioavailregs_dma) {
319 "failed to allocate PIOavail reg area in memory\n");
325 * We really want L2 cache aligned, but for current CPUs of
326 * interest, they are the same.
328 status_page = (u64 *)
329 ((char *) dd->pioavailregs_dma +
330 ((2 * L1_CACHE_BYTES +
331 dd->pioavregs * sizeof(u64)) & ~L1_CACHE_BYTES));
332 /* device status comes first, for backwards compatibility */
333 dd->devstatusp = status_page;
335 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
336 dd->pport[pidx].statusp = status_page;
341 * Setup buffer to hold freeze and other messages, accessible to
342 * apps, following statusp. This is per-unit, not per port.
344 dd->freezemsg = (char *) status_page;
346 /* length of msg buffer is "whatever is left" */
347 ret = (char *) status_page - (char *) dd->pioavailregs_dma;
348 dd->freezelen = PAGE_SIZE - ret;
357 * init_shadow_tids - allocate the shadow TID array
358 * @dd: the qlogic_ib device
360 * allocate the shadow TID array, so we can qib_munlock previous
361 * entries. It may make more sense to move the pageshadow to the
362 * ctxt data structure, so we only allocate memory for ctxts actually
363 * in use, since we at 8k per ctxt, now.
364 * We don't want failures here to prevent use of the driver/chip,
365 * so no return value.
367 static void init_shadow_tids(struct qib_devdata *dd)
372 pages = vzalloc(dd->cfgctxts * dd->rcvtidcnt * sizeof(struct page *));
376 addrs = vzalloc(dd->cfgctxts * dd->rcvtidcnt * sizeof(dma_addr_t));
380 dd->pageshadow = pages;
381 dd->physshadow = addrs;
387 dd->pageshadow = NULL;
391 * Do initialization for device that is only needed on
392 * first detect, not on resets.
394 static int loadtime_init(struct qib_devdata *dd)
398 if (((dd->revision >> QLOGIC_IB_R_SOFTWARE_SHIFT) &
399 QLOGIC_IB_R_SOFTWARE_MASK) != QIB_CHIP_SWVERSION) {
401 "Driver only handles version %d, chip swversion is %d (%llx), failing\n",
403 (int)(dd->revision >>
404 QLOGIC_IB_R_SOFTWARE_SHIFT) &
405 QLOGIC_IB_R_SOFTWARE_MASK,
406 (unsigned long long) dd->revision);
411 if (dd->revision & QLOGIC_IB_R_EMULATOR_MASK)
412 qib_devinfo(dd->pcidev, "%s", dd->boardversion);
414 spin_lock_init(&dd->pioavail_lock);
415 spin_lock_init(&dd->sendctrl_lock);
416 spin_lock_init(&dd->uctxt_lock);
417 spin_lock_init(&dd->qib_diag_trans_lock);
418 spin_lock_init(&dd->eep_st_lock);
419 mutex_init(&dd->eep_lock);
424 ret = init_pioavailregs(dd);
425 init_shadow_tids(dd);
427 qib_get_eeprom_info(dd);
429 /* setup time (don't start yet) to verify we got interrupt */
430 timer_setup(&dd->intrchk_timer, verify_interrupt, 0);
436 * init_after_reset - re-initialize after a reset
437 * @dd: the qlogic_ib device
439 * sanity check at least some of the values after reset, and
440 * ensure no receive or transmit (explicitly, in case reset
443 static int init_after_reset(struct qib_devdata *dd)
448 * Ensure chip does no sends or receives, tail updates, or
449 * pioavail updates while we re-initialize. This is mostly
450 * for the driver data structures, not chip registers.
452 for (i = 0; i < dd->num_pports; ++i) {
454 * ctxt == -1 means "all contexts". Only really safe for
455 * _dis_abling things, as here.
457 dd->f_rcvctrl(dd->pport + i, QIB_RCVCTRL_CTXT_DIS |
458 QIB_RCVCTRL_INTRAVAIL_DIS |
459 QIB_RCVCTRL_TAILUPD_DIS, -1);
460 /* Redundant across ports for some, but no big deal. */
461 dd->f_sendctrl(dd->pport + i, QIB_SENDCTRL_SEND_DIS |
462 QIB_SENDCTRL_AVAIL_DIS);
468 static void enable_chip(struct qib_devdata *dd)
474 * Enable PIO send, and update of PIOavail regs to memory.
476 for (i = 0; i < dd->num_pports; ++i)
477 dd->f_sendctrl(dd->pport + i, QIB_SENDCTRL_SEND_ENB |
478 QIB_SENDCTRL_AVAIL_ENB);
480 * Enable kernel ctxts' receive and receive interrupt.
481 * Other ctxts done as user opens and inits them.
483 rcvmask = QIB_RCVCTRL_CTXT_ENB | QIB_RCVCTRL_INTRAVAIL_ENB;
484 rcvmask |= (dd->flags & QIB_NODMA_RTAIL) ?
485 QIB_RCVCTRL_TAILUPD_DIS : QIB_RCVCTRL_TAILUPD_ENB;
486 for (i = 0; dd->rcd && i < dd->first_user_ctxt; ++i) {
487 struct qib_ctxtdata *rcd = dd->rcd[i];
490 dd->f_rcvctrl(rcd->ppd, rcvmask, i);
494 static void verify_interrupt(struct timer_list *t)
496 struct qib_devdata *dd = from_timer(dd, t, intrchk_timer);
500 return; /* being torn down */
503 * If we don't have a lid or any interrupts, let the user know and
504 * don't bother checking again.
506 int_counter = qib_int_counter(dd) - dd->z_int_counter;
507 if (int_counter == 0) {
508 if (!dd->f_intr_fallback(dd))
509 dev_err(&dd->pcidev->dev,
510 "No interrupts detected, not usable.\n");
511 else /* re-arm the timer to see if fallback works */
512 mod_timer(&dd->intrchk_timer, jiffies + HZ/2);
516 static void init_piobuf_state(struct qib_devdata *dd)
522 * Ensure all buffers are free, and fifos empty. Buffers
523 * are common, so only do once for port 0.
525 * After enable and qib_chg_pioavailkernel so we can safely
526 * enable pioavail updates and PIOENABLE. After this, packets
527 * are ready and able to go out.
529 dd->f_sendctrl(dd->pport, QIB_SENDCTRL_DISARM_ALL);
530 for (pidx = 0; pidx < dd->num_pports; ++pidx)
531 dd->f_sendctrl(dd->pport + pidx, QIB_SENDCTRL_FLUSH);
534 * If not all sendbufs are used, add the one to each of the lower
535 * numbered contexts. pbufsctxt and lastctxt_piobuf are
536 * calculated in chip-specific code because it may cause some
537 * chip-specific adjustments to be made.
539 uctxts = dd->cfgctxts - dd->first_user_ctxt;
540 dd->ctxts_extrabuf = dd->pbufsctxt ?
541 dd->lastctxt_piobuf - (dd->pbufsctxt * uctxts) : 0;
544 * Set up the shadow copies of the piobufavail registers,
545 * which we compare against the chip registers for now, and
546 * the in memory DMA'ed copies of the registers.
547 * By now pioavail updates to memory should have occurred, so
548 * copy them into our working/shadow registers; this is in
549 * case something went wrong with abort, but mostly to get the
550 * initial values of the generation bit correct.
552 for (i = 0; i < dd->pioavregs; i++) {
555 tmp = dd->pioavailregs_dma[i];
557 * Don't need to worry about pioavailkernel here
558 * because we will call qib_chg_pioavailkernel() later
559 * in initialization, to busy out buffers as needed.
561 dd->pioavailshadow[i] = le64_to_cpu(tmp);
563 while (i < ARRAY_SIZE(dd->pioavailshadow))
564 dd->pioavailshadow[i++] = 0; /* for debugging sanity */
566 /* after pioavailshadow is setup */
567 qib_chg_pioavailkernel(dd, 0, dd->piobcnt2k + dd->piobcnt4k,
568 TXCHK_CHG_TYPE_KERN, NULL);
569 dd->f_initvl15_bufs(dd);
573 * qib_create_workqueues - create per port workqueues
574 * @dd: the qlogic_ib device
576 static int qib_create_workqueues(struct qib_devdata *dd)
579 struct qib_pportdata *ppd;
581 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
582 ppd = dd->pport + pidx;
584 char wq_name[8]; /* 3 + 2 + 1 + 1 + 1 */
586 snprintf(wq_name, sizeof(wq_name), "qib%d_%d",
588 ppd->qib_wq = alloc_ordered_workqueue(wq_name,
596 pr_err("create_singlethread_workqueue failed for port %d\n",
598 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
599 ppd = dd->pport + pidx;
601 destroy_workqueue(ppd->qib_wq);
608 static void qib_free_pportdata(struct qib_pportdata *ppd)
610 free_percpu(ppd->ibport_data.pmastats);
611 free_percpu(ppd->ibport_data.rvp.rc_acks);
612 free_percpu(ppd->ibport_data.rvp.rc_qacks);
613 free_percpu(ppd->ibport_data.rvp.rc_delayed_comp);
614 ppd->ibport_data.pmastats = NULL;
618 * qib_init - do the actual initialization sequence on the chip
619 * @dd: the qlogic_ib device
620 * @reinit: reinitializing, so don't allocate new memory
622 * Do the actual initialization sequence on the chip. This is done
623 * both from the init routine called from the PCI infrastructure, and
624 * when we reset the chip, or detect that it was reset internally,
625 * or it's administratively re-enabled.
627 * Memory allocation here and in called routines is only done in
628 * the first case (reinit == 0). We have to be careful, because even
629 * without memory allocation, we need to re-write all the chip registers
630 * TIDs, etc. after the reset or enable has completed.
632 int qib_init(struct qib_devdata *dd, int reinit)
634 int ret = 0, pidx, lastfail = 0;
637 struct qib_ctxtdata *rcd;
638 struct qib_pportdata *ppd;
641 /* Set linkstate to unknown, so we can watch for a transition. */
642 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
643 ppd = dd->pport + pidx;
644 spin_lock_irqsave(&ppd->lflags_lock, flags);
645 ppd->lflags &= ~(QIBL_LINKACTIVE | QIBL_LINKARMED |
646 QIBL_LINKDOWN | QIBL_LINKINIT |
648 spin_unlock_irqrestore(&ppd->lflags_lock, flags);
652 ret = init_after_reset(dd);
654 ret = loadtime_init(dd);
658 /* Bypass most chip-init, to get to device creation */
662 ret = dd->f_late_initreg(dd);
666 /* dd->rcd can be NULL if early init failed */
667 for (i = 0; dd->rcd && i < dd->first_user_ctxt; ++i) {
669 * Set up the (kernel) rcvhdr queue and egr TIDs. If doing
670 * re-init, the simplest way to handle this is to free
671 * existing, and re-allocate.
672 * Need to re-create rest of ctxt 0 ctxtdata as well.
678 lastfail = qib_create_rcvhdrq(dd, rcd);
680 lastfail = qib_setup_eagerbufs(rcd);
683 "failed to allocate kernel ctxt's rcvhdrq and/or egr bufs\n");
686 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
691 ppd = dd->pport + pidx;
692 mtu = ib_mtu_enum_to_int(qib_ibmtu);
694 mtu = QIB_DEFAULT_MTU;
695 qib_ibmtu = 0; /* don't leave invalid value */
697 /* set max we can ever have for this driver load */
698 ppd->init_ibmaxlen = min(mtu > 2048 ?
699 dd->piosize4k : dd->piosize2k,
701 (dd->rcvhdrentsize << 2));
703 * Have to initialize ibmaxlen, but this will normally
704 * change immediately in qib_set_mtu().
706 ppd->ibmaxlen = ppd->init_ibmaxlen;
707 qib_set_mtu(ppd, mtu);
709 spin_lock_irqsave(&ppd->lflags_lock, flags);
710 ppd->lflags |= QIBL_IB_LINK_DISABLED;
711 spin_unlock_irqrestore(&ppd->lflags_lock, flags);
713 lastfail = dd->f_bringup_serdes(ppd);
715 qib_devinfo(dd->pcidev,
716 "Failed to bringup IB port %u\n", ppd->port);
717 lastfail = -ENETDOWN;
725 /* none of the ports initialized */
726 if (!ret && lastfail)
730 /* but continue on, so we can debug cause */
735 init_piobuf_state(dd);
739 /* chip is OK for user apps; mark it as initialized */
740 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
741 ppd = dd->pport + pidx;
743 * Set status even if port serdes is not initialized
744 * so that diags will work.
746 *ppd->statusp |= QIB_STATUS_CHIP_PRESENT |
748 if (!ppd->link_speed_enabled)
750 if (dd->flags & QIB_HAS_SEND_DMA)
751 ret = qib_setup_sdma(ppd);
752 timer_setup(&ppd->hol_timer, qib_hol_event, 0);
753 ppd->hol_state = QIB_HOL_UP;
756 /* now we can enable all interrupts from the chip */
757 dd->f_set_intr_state(dd, 1);
760 * Setup to verify we get an interrupt, and fallback
761 * to an alternate if necessary and possible.
763 mod_timer(&dd->intrchk_timer, jiffies + HZ/2);
764 /* start stats retrieval timer */
765 mod_timer(&dd->stats_timer, jiffies + HZ * ACTIVITY_TIMER);
768 /* if ret is non-zero, we probably should do some cleanup here... */
773 * These next two routines are placeholders in case we don't have per-arch
774 * code for controlling write combining. If explicit control of write
775 * combining is not available, performance will probably be awful.
778 int __attribute__((weak)) qib_enable_wc(struct qib_devdata *dd)
783 void __attribute__((weak)) qib_disable_wc(struct qib_devdata *dd)
787 static inline struct qib_devdata *__qib_lookup(int unit)
789 return idr_find(&qib_unit_table, unit);
792 struct qib_devdata *qib_lookup(int unit)
794 struct qib_devdata *dd;
797 spin_lock_irqsave(&qib_devs_lock, flags);
798 dd = __qib_lookup(unit);
799 spin_unlock_irqrestore(&qib_devs_lock, flags);
805 * Stop the timers during unit shutdown, or after an error late
808 static void qib_stop_timers(struct qib_devdata *dd)
810 struct qib_pportdata *ppd;
813 if (dd->stats_timer.function)
814 del_timer_sync(&dd->stats_timer);
815 if (dd->intrchk_timer.function)
816 del_timer_sync(&dd->intrchk_timer);
817 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
818 ppd = dd->pport + pidx;
819 if (ppd->hol_timer.function)
820 del_timer_sync(&ppd->hol_timer);
821 if (ppd->led_override_timer.function) {
822 del_timer_sync(&ppd->led_override_timer);
823 atomic_set(&ppd->led_override_timer_active, 0);
825 if (ppd->symerr_clear_timer.function)
826 del_timer_sync(&ppd->symerr_clear_timer);
831 * qib_shutdown_device - shut down a device
832 * @dd: the qlogic_ib device
834 * This is called to make the device quiet when we are about to
835 * unload the driver, and also when the device is administratively
836 * disabled. It does not free any data structures.
837 * Everything it does has to be setup again by qib_init(dd, 1)
839 static void qib_shutdown_device(struct qib_devdata *dd)
841 struct qib_pportdata *ppd;
844 if (dd->flags & QIB_SHUTDOWN)
846 dd->flags |= QIB_SHUTDOWN;
848 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
849 ppd = dd->pport + pidx;
851 spin_lock_irq(&ppd->lflags_lock);
852 ppd->lflags &= ~(QIBL_LINKDOWN | QIBL_LINKINIT |
853 QIBL_LINKARMED | QIBL_LINKACTIVE |
855 spin_unlock_irq(&ppd->lflags_lock);
856 *ppd->statusp &= ~(QIB_STATUS_IB_CONF | QIB_STATUS_IB_READY);
858 dd->flags &= ~QIB_INITTED;
860 /* mask interrupts, but not errors */
861 dd->f_set_intr_state(dd, 0);
863 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
864 ppd = dd->pport + pidx;
865 dd->f_rcvctrl(ppd, QIB_RCVCTRL_TAILUPD_DIS |
866 QIB_RCVCTRL_CTXT_DIS |
867 QIB_RCVCTRL_INTRAVAIL_DIS |
868 QIB_RCVCTRL_PKEY_ENB, -1);
870 * Gracefully stop all sends allowing any in progress to
873 dd->f_sendctrl(ppd, QIB_SENDCTRL_CLEAR);
877 * Enough for anything that's going to trickle out to have actually
882 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
883 ppd = dd->pport + pidx;
884 dd->f_setextled(ppd, 0); /* make sure LEDs are off */
886 if (dd->flags & QIB_HAS_SEND_DMA)
887 qib_teardown_sdma(ppd);
889 dd->f_sendctrl(ppd, QIB_SENDCTRL_AVAIL_DIS |
890 QIB_SENDCTRL_SEND_DIS);
892 * Clear SerdesEnable.
893 * We can't count on interrupts since we are stopping.
895 dd->f_quiet_serdes(ppd);
898 destroy_workqueue(ppd->qib_wq);
901 qib_free_pportdata(ppd);
907 * qib_free_ctxtdata - free a context's allocated data
908 * @dd: the qlogic_ib device
909 * @rcd: the ctxtdata structure
911 * free up any allocated data for a context
912 * This should not touch anything that would affect a simultaneous
913 * re-allocation of context data, because it is called after qib_mutex
914 * is released (and can be called from reinit as well).
915 * It should never change any chip state, or global driver state.
917 void qib_free_ctxtdata(struct qib_devdata *dd, struct qib_ctxtdata *rcd)
923 dma_free_coherent(&dd->pcidev->dev, rcd->rcvhdrq_size,
924 rcd->rcvhdrq, rcd->rcvhdrq_phys);
926 if (rcd->rcvhdrtail_kvaddr) {
927 dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE,
928 rcd->rcvhdrtail_kvaddr,
929 rcd->rcvhdrqtailaddr_phys);
930 rcd->rcvhdrtail_kvaddr = NULL;
933 if (rcd->rcvegrbuf) {
936 for (e = 0; e < rcd->rcvegrbuf_chunks; e++) {
937 void *base = rcd->rcvegrbuf[e];
938 size_t size = rcd->rcvegrbuf_size;
940 dma_free_coherent(&dd->pcidev->dev, size,
941 base, rcd->rcvegrbuf_phys[e]);
943 kfree(rcd->rcvegrbuf);
944 rcd->rcvegrbuf = NULL;
945 kfree(rcd->rcvegrbuf_phys);
946 rcd->rcvegrbuf_phys = NULL;
947 rcd->rcvegrbuf_chunks = 0;
950 kfree(rcd->tid_pg_list);
951 vfree(rcd->user_event_mask);
952 vfree(rcd->subctxt_uregbase);
953 vfree(rcd->subctxt_rcvegrbuf);
954 vfree(rcd->subctxt_rcvhdr_base);
955 #ifdef CONFIG_DEBUG_FS
963 * Perform a PIO buffer bandwidth write test, to verify proper system
964 * configuration. Even when all the setup calls work, occasionally
965 * BIOS or other issues can prevent write combining from working, or
966 * can cause other bandwidth problems to the chip.
968 * This test simply writes the same buffer over and over again, and
969 * measures close to the peak bandwidth to the chip (not testing
970 * data bandwidth to the wire). On chips that use an address-based
971 * trigger to send packets to the wire, this is easy. On chips that
972 * use a count to trigger, we want to make sure that the packet doesn't
973 * go out on the wire, or trigger flow control checks.
975 static void qib_verify_pioperf(struct qib_devdata *dd)
977 u32 pbnum, cnt, lcnt;
982 piobuf = dd->f_getsendbuf(dd->pport, 0ULL, &pbnum);
984 qib_devinfo(dd->pcidev,
985 "No PIObufs for checking perf, skipping\n");
990 * Enough to give us a reasonable test, less than piobuf size, and
991 * likely multiple of store buffer length.
999 preempt_disable(); /* we want reasonably accurate elapsed time */
1000 msecs = 1 + jiffies_to_msecs(jiffies);
1001 for (lcnt = 0; lcnt < 10000U; lcnt++) {
1002 /* wait until we cross msec boundary */
1003 if (jiffies_to_msecs(jiffies) >= msecs)
1008 dd->f_set_armlaunch(dd, 0);
1011 * length 0, no dwords actually sent
1017 * This is only roughly accurate, since even with preempt we
1018 * still take interrupts that could take a while. Running for
1019 * >= 5 msec seems to get us "close enough" to accurate values.
1021 msecs = jiffies_to_msecs(jiffies);
1022 for (emsecs = lcnt = 0; emsecs <= 5UL; lcnt++) {
1023 qib_pio_copy(piobuf + 64, addr, cnt >> 2);
1024 emsecs = jiffies_to_msecs(jiffies) - msecs;
1027 /* 1 GiB/sec, slightly over IB SDR line rate */
1028 if (lcnt < (emsecs * 1024U))
1030 "Performance problem: bandwidth to PIO buffers is only %u MiB/sec\n",
1031 lcnt / (u32) emsecs);
1038 /* disarm piobuf, so it's available again */
1039 dd->f_sendctrl(dd->pport, QIB_SENDCTRL_DISARM_BUF(pbnum));
1040 qib_sendbuf_done(dd, pbnum);
1041 dd->f_set_armlaunch(dd, 1);
1044 void qib_free_devdata(struct qib_devdata *dd)
1046 unsigned long flags;
1048 spin_lock_irqsave(&qib_devs_lock, flags);
1049 idr_remove(&qib_unit_table, dd->unit);
1050 list_del(&dd->list);
1051 spin_unlock_irqrestore(&qib_devs_lock, flags);
1053 #ifdef CONFIG_DEBUG_FS
1054 qib_dbg_ibdev_exit(&dd->verbs_dev);
1056 free_percpu(dd->int_counter);
1057 rvt_dealloc_device(&dd->verbs_dev.rdi);
1060 u64 qib_int_counter(struct qib_devdata *dd)
1063 u64 int_counter = 0;
1065 for_each_possible_cpu(cpu)
1066 int_counter += *per_cpu_ptr(dd->int_counter, cpu);
1070 u64 qib_sps_ints(void)
1072 unsigned long flags;
1073 struct qib_devdata *dd;
1076 spin_lock_irqsave(&qib_devs_lock, flags);
1077 list_for_each_entry(dd, &qib_dev_list, list) {
1078 sps_ints += qib_int_counter(dd);
1080 spin_unlock_irqrestore(&qib_devs_lock, flags);
1085 * Allocate our primary per-unit data structure. Must be done via verbs
1086 * allocator, because the verbs cleanup process both does cleanup and
1087 * free of the data structure.
1088 * "extra" is for chip-specific data.
1090 * Use the idr mechanism to get a unit number for this unit.
1092 struct qib_devdata *qib_alloc_devdata(struct pci_dev *pdev, size_t extra)
1094 unsigned long flags;
1095 struct qib_devdata *dd;
1098 /* extra is * number of ports */
1099 nports = extra / sizeof(struct qib_pportdata);
1100 dd = (struct qib_devdata *)rvt_alloc_device(sizeof(*dd) + extra,
1103 return ERR_PTR(-ENOMEM);
1105 INIT_LIST_HEAD(&dd->list);
1107 idr_preload(GFP_KERNEL);
1108 spin_lock_irqsave(&qib_devs_lock, flags);
1110 ret = idr_alloc(&qib_unit_table, dd, 0, 0, GFP_NOWAIT);
1113 list_add(&dd->list, &qib_dev_list);
1116 spin_unlock_irqrestore(&qib_devs_lock, flags);
1120 qib_early_err(&pdev->dev,
1121 "Could not allocate unit ID: error %d\n", -ret);
1124 rvt_set_ibdev_name(&dd->verbs_dev.rdi, "%s%d", "qib", dd->unit);
1126 dd->int_counter = alloc_percpu(u64);
1127 if (!dd->int_counter) {
1129 qib_early_err(&pdev->dev,
1130 "Could not allocate per-cpu int_counter\n");
1134 if (!qib_cpulist_count) {
1135 u32 count = num_online_cpus();
1137 qib_cpulist = kcalloc(BITS_TO_LONGS(count), sizeof(long),
1140 qib_cpulist_count = count;
1142 #ifdef CONFIG_DEBUG_FS
1143 qib_dbg_ibdev_init(&dd->verbs_dev);
1147 if (!list_empty(&dd->list))
1148 list_del_init(&dd->list);
1149 rvt_dealloc_device(&dd->verbs_dev.rdi);
1150 return ERR_PTR(ret);
1154 * Called from freeze mode handlers, and from PCI error
1155 * reporting code. Should be paranoid about state of
1156 * system and data structures.
1158 void qib_disable_after_error(struct qib_devdata *dd)
1160 if (dd->flags & QIB_INITTED) {
1163 dd->flags &= ~QIB_INITTED;
1165 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
1166 struct qib_pportdata *ppd;
1168 ppd = dd->pport + pidx;
1169 if (dd->flags & QIB_PRESENT) {
1170 qib_set_linkstate(ppd,
1171 QIB_IB_LINKDOWN_DISABLE);
1172 dd->f_setextled(ppd, 0);
1174 *ppd->statusp &= ~QIB_STATUS_IB_READY;
1179 * Mark as having had an error for driver, and also
1180 * for /sys and status word mapped to user programs.
1181 * This marks unit as not usable, until reset.
1184 *dd->devstatusp |= QIB_STATUS_HWERROR;
1187 static void qib_remove_one(struct pci_dev *);
1188 static int qib_init_one(struct pci_dev *, const struct pci_device_id *);
1189 static void qib_shutdown_one(struct pci_dev *);
1191 #define DRIVER_LOAD_MSG "Intel " QIB_DRV_NAME " loaded: "
1192 #define PFX QIB_DRV_NAME ": "
1194 static const struct pci_device_id qib_pci_tbl[] = {
1195 { PCI_DEVICE(PCI_VENDOR_ID_PATHSCALE, PCI_DEVICE_ID_QLOGIC_IB_6120) },
1196 { PCI_DEVICE(PCI_VENDOR_ID_QLOGIC, PCI_DEVICE_ID_QLOGIC_IB_7220) },
1197 { PCI_DEVICE(PCI_VENDOR_ID_QLOGIC, PCI_DEVICE_ID_QLOGIC_IB_7322) },
1201 MODULE_DEVICE_TABLE(pci, qib_pci_tbl);
1203 static struct pci_driver qib_driver = {
1204 .name = QIB_DRV_NAME,
1205 .probe = qib_init_one,
1206 .remove = qib_remove_one,
1207 .shutdown = qib_shutdown_one,
1208 .id_table = qib_pci_tbl,
1209 .err_handler = &qib_pci_err_handler,
1212 #ifdef CONFIG_INFINIBAND_QIB_DCA
1214 static int qib_notify_dca(struct notifier_block *, unsigned long, void *);
1215 static struct notifier_block dca_notifier = {
1216 .notifier_call = qib_notify_dca,
1221 static int qib_notify_dca_device(struct device *device, void *data)
1223 struct qib_devdata *dd = dev_get_drvdata(device);
1224 unsigned long event = *(unsigned long *)data;
1226 return dd->f_notify_dca(dd, event);
1229 static int qib_notify_dca(struct notifier_block *nb, unsigned long event,
1234 rval = driver_for_each_device(&qib_driver.driver, NULL,
1235 &event, qib_notify_dca_device);
1236 return rval ? NOTIFY_BAD : NOTIFY_DONE;
1242 * Do all the generic driver unit- and chip-independent memory
1243 * allocation and initialization.
1245 static int __init qib_ib_init(void)
1249 ret = qib_dev_init();
1254 * These must be called before the driver is registered with
1255 * the PCI subsystem.
1257 idr_init(&qib_unit_table);
1259 #ifdef CONFIG_INFINIBAND_QIB_DCA
1260 dca_register_notify(&dca_notifier);
1262 #ifdef CONFIG_DEBUG_FS
1265 ret = pci_register_driver(&qib_driver);
1267 pr_err("Unable to register driver: error %d\n", -ret);
1271 /* not fatal if it doesn't work */
1272 if (qib_init_qibfs())
1273 pr_err("Unable to register ipathfs\n");
1274 goto bail; /* all OK */
1277 #ifdef CONFIG_INFINIBAND_QIB_DCA
1278 dca_unregister_notify(&dca_notifier);
1280 #ifdef CONFIG_DEBUG_FS
1283 idr_destroy(&qib_unit_table);
1289 module_init(qib_ib_init);
1292 * Do the non-unit driver cleanup, memory free, etc. at unload.
1294 static void __exit qib_ib_cleanup(void)
1298 ret = qib_exit_qibfs();
1301 "Unable to cleanup counter filesystem: error %d\n",
1304 #ifdef CONFIG_INFINIBAND_QIB_DCA
1305 dca_unregister_notify(&dca_notifier);
1307 pci_unregister_driver(&qib_driver);
1308 #ifdef CONFIG_DEBUG_FS
1312 qib_cpulist_count = 0;
1315 idr_destroy(&qib_unit_table);
1319 module_exit(qib_ib_cleanup);
1321 /* this can only be called after a successful initialization */
1322 static void cleanup_device_data(struct qib_devdata *dd)
1326 struct qib_ctxtdata **tmp;
1327 unsigned long flags;
1329 /* users can't do anything more with chip */
1330 for (pidx = 0; pidx < dd->num_pports; ++pidx) {
1331 if (dd->pport[pidx].statusp)
1332 *dd->pport[pidx].statusp &= ~QIB_STATUS_CHIP_PRESENT;
1334 spin_lock(&dd->pport[pidx].cc_shadow_lock);
1336 kfree(dd->pport[pidx].congestion_entries);
1337 dd->pport[pidx].congestion_entries = NULL;
1338 kfree(dd->pport[pidx].ccti_entries);
1339 dd->pport[pidx].ccti_entries = NULL;
1340 kfree(dd->pport[pidx].ccti_entries_shadow);
1341 dd->pport[pidx].ccti_entries_shadow = NULL;
1342 kfree(dd->pport[pidx].congestion_entries_shadow);
1343 dd->pport[pidx].congestion_entries_shadow = NULL;
1345 spin_unlock(&dd->pport[pidx].cc_shadow_lock);
1350 if (dd->pioavailregs_dma) {
1351 dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE,
1352 (void *) dd->pioavailregs_dma,
1353 dd->pioavailregs_phys);
1354 dd->pioavailregs_dma = NULL;
1357 if (dd->pageshadow) {
1358 struct page **tmpp = dd->pageshadow;
1359 dma_addr_t *tmpd = dd->physshadow;
1362 for (ctxt = 0; ctxt < dd->cfgctxts; ctxt++) {
1363 int ctxt_tidbase = ctxt * dd->rcvtidcnt;
1364 int maxtid = ctxt_tidbase + dd->rcvtidcnt;
1366 for (i = ctxt_tidbase; i < maxtid; i++) {
1369 pci_unmap_page(dd->pcidev, tmpd[i],
1370 PAGE_SIZE, PCI_DMA_FROMDEVICE);
1371 qib_release_user_pages(&tmpp[i], 1);
1376 dd->pageshadow = NULL;
1378 dd->physshadow = NULL;
1383 * Free any resources still in use (usually just kernel contexts)
1384 * at unload; we do for ctxtcnt, because that's what we allocate.
1385 * We acquire lock to be really paranoid that rcd isn't being
1386 * accessed from some interrupt-related code (that should not happen,
1387 * but best to be sure).
1389 spin_lock_irqsave(&dd->uctxt_lock, flags);
1392 spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1393 for (ctxt = 0; tmp && ctxt < dd->ctxtcnt; ctxt++) {
1394 struct qib_ctxtdata *rcd = tmp[ctxt];
1396 tmp[ctxt] = NULL; /* debugging paranoia */
1397 qib_free_ctxtdata(dd, rcd);
1403 * Clean up on unit shutdown, or error during unit load after
1404 * successful initialization.
1406 static void qib_postinit_cleanup(struct qib_devdata *dd)
1409 * Clean up chip-specific stuff.
1410 * We check for NULL here, because it's outside
1411 * the kregbase check, and we need to call it
1412 * after the free_irq. Thus it's possible that
1413 * the function pointers were never initialized.
1418 qib_pcie_ddcleanup(dd);
1420 cleanup_device_data(dd);
1422 qib_free_devdata(dd);
1425 static int qib_init_one(struct pci_dev *pdev, const struct pci_device_id *ent)
1427 int ret, j, pidx, initfail;
1428 struct qib_devdata *dd = NULL;
1430 ret = qib_pcie_init(pdev, ent);
1435 * Do device-specific initialiation, function table setup, dd
1438 switch (ent->device) {
1439 case PCI_DEVICE_ID_QLOGIC_IB_6120:
1440 #ifdef CONFIG_PCI_MSI
1441 dd = qib_init_iba6120_funcs(pdev, ent);
1443 qib_early_err(&pdev->dev,
1444 "Intel PCIE device 0x%x cannot work if CONFIG_PCI_MSI is not enabled\n",
1446 dd = ERR_PTR(-ENODEV);
1450 case PCI_DEVICE_ID_QLOGIC_IB_7220:
1451 dd = qib_init_iba7220_funcs(pdev, ent);
1454 case PCI_DEVICE_ID_QLOGIC_IB_7322:
1455 dd = qib_init_iba7322_funcs(pdev, ent);
1459 qib_early_err(&pdev->dev,
1460 "Failing on unknown Intel deviceid 0x%x\n",
1468 goto bail; /* error already printed */
1470 ret = qib_create_workqueues(dd);
1474 /* do the generic initialization */
1475 initfail = qib_init(dd, 0);
1477 ret = qib_register_ib_device(dd);
1480 * Now ready for use. this should be cleared whenever we
1481 * detect a reset, or initiate one. If earlier failure,
1482 * we still create devices, so diags, etc. can be used
1483 * to determine cause of problem.
1485 if (!qib_mini_init && !initfail && !ret)
1486 dd->flags |= QIB_INITTED;
1488 j = qib_device_create(dd);
1490 qib_dev_err(dd, "Failed to create /dev devices: %d\n", -j);
1493 qib_dev_err(dd, "Failed filesystem setup for counters: %d\n",
1496 if (qib_mini_init || initfail || ret) {
1497 qib_stop_timers(dd);
1498 flush_workqueue(ib_wq);
1499 for (pidx = 0; pidx < dd->num_pports; ++pidx)
1500 dd->f_quiet_serdes(dd->pport + pidx);
1504 (void) qibfs_remove(dd);
1505 qib_device_remove(dd);
1508 qib_unregister_ib_device(dd);
1509 qib_postinit_cleanup(dd);
1515 ret = qib_enable_wc(dd);
1518 "Write combining not enabled (err %d): performance may be poor\n",
1523 qib_verify_pioperf(dd);
1528 static void qib_remove_one(struct pci_dev *pdev)
1530 struct qib_devdata *dd = pci_get_drvdata(pdev);
1533 /* unregister from IB core */
1534 qib_unregister_ib_device(dd);
1537 * Disable the IB link, disable interrupts on the device,
1538 * clear dma engines, etc.
1541 qib_shutdown_device(dd);
1543 qib_stop_timers(dd);
1545 /* wait until all of our (qsfp) queue_work() calls complete */
1546 flush_workqueue(ib_wq);
1548 ret = qibfs_remove(dd);
1550 qib_dev_err(dd, "Failed counters filesystem cleanup: %d\n",
1553 qib_device_remove(dd);
1555 qib_postinit_cleanup(dd);
1558 static void qib_shutdown_one(struct pci_dev *pdev)
1560 struct qib_devdata *dd = pci_get_drvdata(pdev);
1562 qib_shutdown_device(dd);
1566 * qib_create_rcvhdrq - create a receive header queue
1567 * @dd: the qlogic_ib device
1568 * @rcd: the context data
1570 * This must be contiguous memory (from an i/o perspective), and must be
1571 * DMA'able (which means for some systems, it will go through an IOMMU,
1572 * or be forced into a low address range).
1574 int qib_create_rcvhdrq(struct qib_devdata *dd, struct qib_ctxtdata *rcd)
1579 if (!rcd->rcvhdrq) {
1580 dma_addr_t phys_hdrqtail;
1583 amt = ALIGN(dd->rcvhdrcnt * dd->rcvhdrentsize *
1584 sizeof(u32), PAGE_SIZE);
1585 gfp_flags = (rcd->ctxt >= dd->first_user_ctxt) ?
1586 GFP_USER : GFP_KERNEL;
1588 old_node_id = dev_to_node(&dd->pcidev->dev);
1589 set_dev_node(&dd->pcidev->dev, rcd->node_id);
1590 rcd->rcvhdrq = dma_alloc_coherent(
1591 &dd->pcidev->dev, amt, &rcd->rcvhdrq_phys,
1592 gfp_flags | __GFP_COMP);
1593 set_dev_node(&dd->pcidev->dev, old_node_id);
1595 if (!rcd->rcvhdrq) {
1597 "attempt to allocate %d bytes for ctxt %u rcvhdrq failed\n",
1602 if (rcd->ctxt >= dd->first_user_ctxt) {
1603 rcd->user_event_mask = vmalloc_user(PAGE_SIZE);
1604 if (!rcd->user_event_mask)
1605 goto bail_free_hdrq;
1608 if (!(dd->flags & QIB_NODMA_RTAIL)) {
1609 set_dev_node(&dd->pcidev->dev, rcd->node_id);
1610 rcd->rcvhdrtail_kvaddr = dma_alloc_coherent(
1611 &dd->pcidev->dev, PAGE_SIZE, &phys_hdrqtail,
1613 set_dev_node(&dd->pcidev->dev, old_node_id);
1614 if (!rcd->rcvhdrtail_kvaddr)
1616 rcd->rcvhdrqtailaddr_phys = phys_hdrqtail;
1619 rcd->rcvhdrq_size = amt;
1622 /* clear for security and sanity on each use */
1623 memset(rcd->rcvhdrq, 0, rcd->rcvhdrq_size);
1624 if (rcd->rcvhdrtail_kvaddr)
1625 memset(rcd->rcvhdrtail_kvaddr, 0, PAGE_SIZE);
1630 "attempt to allocate 1 page for ctxt %u rcvhdrqtailaddr failed\n",
1632 vfree(rcd->user_event_mask);
1633 rcd->user_event_mask = NULL;
1635 dma_free_coherent(&dd->pcidev->dev, amt, rcd->rcvhdrq,
1637 rcd->rcvhdrq = NULL;
1643 * allocate eager buffers, both kernel and user contexts.
1644 * @rcd: the context we are setting up.
1646 * Allocate the eager TID buffers and program them into hip.
1647 * They are no longer completely contiguous, we do multiple allocation
1648 * calls. Otherwise we get the OOM code involved, by asking for too
1649 * much per call, with disastrous results on some kernels.
1651 int qib_setup_eagerbufs(struct qib_ctxtdata *rcd)
1653 struct qib_devdata *dd = rcd->dd;
1654 unsigned e, egrcnt, egrperchunk, chunk, egrsize, egroff;
1660 * GFP_USER, but without GFP_FS, so buffer cache can be
1661 * coalesced (we hope); otherwise, even at order 4,
1662 * heavy filesystem activity makes these fail, and we can
1663 * use compound pages.
1665 gfp_flags = __GFP_RECLAIM | __GFP_IO | __GFP_COMP;
1667 egrcnt = rcd->rcvegrcnt;
1668 egroff = rcd->rcvegr_tid_base;
1669 egrsize = dd->rcvegrbufsize;
1671 chunk = rcd->rcvegrbuf_chunks;
1672 egrperchunk = rcd->rcvegrbufs_perchunk;
1673 size = rcd->rcvegrbuf_size;
1674 if (!rcd->rcvegrbuf) {
1676 kzalloc_node(chunk * sizeof(rcd->rcvegrbuf[0]),
1677 GFP_KERNEL, rcd->node_id);
1678 if (!rcd->rcvegrbuf)
1681 if (!rcd->rcvegrbuf_phys) {
1682 rcd->rcvegrbuf_phys =
1683 kmalloc_array_node(chunk,
1684 sizeof(rcd->rcvegrbuf_phys[0]),
1685 GFP_KERNEL, rcd->node_id);
1686 if (!rcd->rcvegrbuf_phys)
1687 goto bail_rcvegrbuf;
1689 for (e = 0; e < rcd->rcvegrbuf_chunks; e++) {
1690 if (rcd->rcvegrbuf[e])
1693 old_node_id = dev_to_node(&dd->pcidev->dev);
1694 set_dev_node(&dd->pcidev->dev, rcd->node_id);
1696 dma_alloc_coherent(&dd->pcidev->dev, size,
1697 &rcd->rcvegrbuf_phys[e],
1699 set_dev_node(&dd->pcidev->dev, old_node_id);
1700 if (!rcd->rcvegrbuf[e])
1701 goto bail_rcvegrbuf_phys;
1704 rcd->rcvegr_phys = rcd->rcvegrbuf_phys[0];
1706 for (e = chunk = 0; chunk < rcd->rcvegrbuf_chunks; chunk++) {
1707 dma_addr_t pa = rcd->rcvegrbuf_phys[chunk];
1710 /* clear for security and sanity on each use */
1711 memset(rcd->rcvegrbuf[chunk], 0, size);
1713 for (i = 0; e < egrcnt && i < egrperchunk; e++, i++) {
1714 dd->f_put_tid(dd, e + egroff +
1719 RCVHQ_RCV_TYPE_EAGER, pa);
1722 cond_resched(); /* don't hog the cpu */
1727 bail_rcvegrbuf_phys:
1728 for (e = 0; e < rcd->rcvegrbuf_chunks && rcd->rcvegrbuf[e]; e++)
1729 dma_free_coherent(&dd->pcidev->dev, size,
1730 rcd->rcvegrbuf[e], rcd->rcvegrbuf_phys[e]);
1731 kfree(rcd->rcvegrbuf_phys);
1732 rcd->rcvegrbuf_phys = NULL;
1734 kfree(rcd->rcvegrbuf);
1735 rcd->rcvegrbuf = NULL;
1741 * Note: Changes to this routine should be mirrored
1742 * for the diagnostics routine qib_remap_ioaddr32().
1743 * There is also related code for VL15 buffers in qib_init_7322_variables().
1744 * The teardown code that unmaps is in qib_pcie_ddcleanup()
1746 int init_chip_wc_pat(struct qib_devdata *dd, u32 vl15buflen)
1748 u64 __iomem *qib_kregbase = NULL;
1749 void __iomem *qib_piobase = NULL;
1750 u64 __iomem *qib_userbase = NULL;
1752 u64 qib_pio2koffset = dd->piobufbase & 0xffffffff;
1753 u64 qib_pio4koffset = dd->piobufbase >> 32;
1754 u64 qib_pio2klen = dd->piobcnt2k * dd->palign;
1755 u64 qib_pio4klen = dd->piobcnt4k * dd->align4k;
1756 u64 qib_physaddr = dd->physaddr;
1758 u64 qib_userlen = 0;
1761 * Free the old mapping because the kernel will try to reuse the
1762 * old mapping and not create a new mapping with the
1763 * write combining attribute.
1765 iounmap(dd->kregbase);
1766 dd->kregbase = NULL;
1769 * Assumes chip address space looks like:
1770 * - kregs + sregs + cregs + uregs (in any order)
1771 * - piobufs (2K and 4K bufs in either order)
1773 * - kregs + sregs + cregs (in any order)
1774 * - piobufs (2K and 4K bufs in either order)
1777 if (dd->piobcnt4k == 0) {
1778 qib_kreglen = qib_pio2koffset;
1779 qib_piolen = qib_pio2klen;
1780 } else if (qib_pio2koffset < qib_pio4koffset) {
1781 qib_kreglen = qib_pio2koffset;
1782 qib_piolen = qib_pio4koffset + qib_pio4klen - qib_kreglen;
1784 qib_kreglen = qib_pio4koffset;
1785 qib_piolen = qib_pio2koffset + qib_pio2klen - qib_kreglen;
1787 qib_piolen += vl15buflen;
1788 /* Map just the configured ports (not all hw ports) */
1789 if (dd->uregbase > qib_kreglen)
1790 qib_userlen = dd->ureg_align * dd->cfgctxts;
1792 /* Sanity checks passed, now create the new mappings */
1793 qib_kregbase = ioremap_nocache(qib_physaddr, qib_kreglen);
1797 qib_piobase = ioremap_wc(qib_physaddr + qib_kreglen, qib_piolen);
1802 qib_userbase = ioremap_nocache(qib_physaddr + dd->uregbase,
1808 dd->kregbase = qib_kregbase;
1809 dd->kregend = (u64 __iomem *)
1810 ((char __iomem *) qib_kregbase + qib_kreglen);
1811 dd->piobase = qib_piobase;
1812 dd->pio2kbase = (void __iomem *)
1813 (((char __iomem *) dd->piobase) +
1814 qib_pio2koffset - qib_kreglen);
1816 dd->pio4kbase = (void __iomem *)
1817 (((char __iomem *) dd->piobase) +
1818 qib_pio4koffset - qib_kreglen);
1820 /* ureg will now be accessed relative to dd->userbase */
1821 dd->userbase = qib_userbase;
1825 iounmap(qib_piobase);
1827 iounmap(qib_kregbase);