1/*
2 * Copyright (c) 2006, 2007, 2008, 2009, 2010 QLogic Corporation.
3 * All rights reserved.
4 * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
5 *
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:
11 *
12 *     Redistribution and use in source and binary forms, with or
13 *     without modification, are permitted provided that the following
14 *     conditions are met:
15 *
16 *      - Redistributions of source code must retain the above
17 *        copyright notice, this list of conditions and the following
18 *        disclaimer.
19 *
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.
24 *
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
32 * SOFTWARE.
33 */
34
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
42#include "qib.h"
43#include "qib_common.h"
44
45/*
46 * min buffers we want to have per context, after driver
47 */
48#define QIB_MIN_USER_CTXT_BUFCNT 7
49
50#define QLOGIC_IB_R_SOFTWARE_MASK 0xFF
51#define QLOGIC_IB_R_SOFTWARE_SHIFT 24
52#define QLOGIC_IB_R_EMULATOR_MASK (1ULL<<62)
53
54/*
55 * Number of ctxts we are configured to use (to allow for more pio
56 * buffers per ctxt, etc.)  Zero means use chip value.
57 */
58ushort qib_cfgctxts;
59module_param_named(cfgctxts, qib_cfgctxts, ushort, S_IRUGO);
60MODULE_PARM_DESC(cfgctxts, "Set max number of contexts to use");
61
62/*
63 * If set, do not write to any regs if avoidable, hack to allow
64 * check for deranged default register values.
65 */
66ushort qib_mini_init;
67module_param_named(mini_init, qib_mini_init, ushort, S_IRUGO);
68MODULE_PARM_DESC(mini_init, "If set, do minimal diag init");
69
70unsigned qib_n_krcv_queues;
71module_param_named(krcvqs, qib_n_krcv_queues, uint, S_IRUGO);
72MODULE_PARM_DESC(krcvqs, "number of kernel receive queues per IB port");
73
74/*
75 * qib_wc_pat parameter:
76 *      0 is WC via MTRR
77 *      1 is WC via PAT
78 *      If PAT initialization fails, code reverts back to MTRR
79 */
80unsigned qib_wc_pat = 1; /* default (1) is to use PAT, not MTRR */
81module_param_named(wc_pat, qib_wc_pat, uint, S_IRUGO);
82MODULE_PARM_DESC(wc_pat, "enable write-combining via PAT mechanism");
83
84struct workqueue_struct *qib_cq_wq;
85
86static void verify_interrupt(unsigned long);
87
88static struct idr qib_unit_table;
89u32 qib_cpulist_count;
90unsigned long *qib_cpulist;
91
92/* set number of contexts we'll actually use */
93void qib_set_ctxtcnt(struct qib_devdata *dd)
94{
95	if (!qib_cfgctxts) {
96		dd->cfgctxts = dd->first_user_ctxt + num_online_cpus();
97		if (dd->cfgctxts > dd->ctxtcnt)
98			dd->cfgctxts = dd->ctxtcnt;
99	} else if (qib_cfgctxts < dd->num_pports)
100		dd->cfgctxts = dd->ctxtcnt;
101	else if (qib_cfgctxts <= dd->ctxtcnt)
102		dd->cfgctxts = qib_cfgctxts;
103	else
104		dd->cfgctxts = dd->ctxtcnt;
105}
106
107/*
108 * Common code for creating the receive context array.
109 */
110int qib_create_ctxts(struct qib_devdata *dd)
111{
112	unsigned i;
113	int ret;
114
115	/*
116	 * Allocate full ctxtcnt array, rather than just cfgctxts, because
117	 * cleanup iterates across all possible ctxts.
118	 */
119	dd->rcd = kzalloc(sizeof(*dd->rcd) * dd->ctxtcnt, GFP_KERNEL);
120	if (!dd->rcd) {
121		qib_dev_err(dd, "Unable to allocate ctxtdata array, "
122			    "failing\n");
123		ret = -ENOMEM;
124		goto done;
125	}
126
127	/* create (one or more) kctxt */
128	for (i = 0; i < dd->first_user_ctxt; ++i) {
129		struct qib_pportdata *ppd;
130		struct qib_ctxtdata *rcd;
131
132		if (dd->skip_kctxt_mask & (1 << i))
133			continue;
134
135		ppd = dd->pport + (i % dd->num_pports);
136		rcd = qib_create_ctxtdata(ppd, i);
137		if (!rcd) {
138			qib_dev_err(dd, "Unable to allocate ctxtdata"
139				    " for Kernel ctxt, failing\n");
140			ret = -ENOMEM;
141			goto done;
142		}
143		rcd->pkeys[0] = QIB_DEFAULT_P_KEY;
144		rcd->seq_cnt = 1;
145	}
146	ret = 0;
147done:
148	return ret;
149}
150
151/*
152 * Common code for user and kernel context setup.
153 */
154struct qib_ctxtdata *qib_create_ctxtdata(struct qib_pportdata *ppd, u32 ctxt)
155{
156	struct qib_devdata *dd = ppd->dd;
157	struct qib_ctxtdata *rcd;
158
159	rcd = kzalloc(sizeof(*rcd), GFP_KERNEL);
160	if (rcd) {
161		INIT_LIST_HEAD(&rcd->qp_wait_list);
162		rcd->ppd = ppd;
163		rcd->dd = dd;
164		rcd->cnt = 1;
165		rcd->ctxt = ctxt;
166		dd->rcd[ctxt] = rcd;
167
168		dd->f_init_ctxt(rcd);
169
170		/*
171		 * To avoid wasting a lot of memory, we allocate 32KB chunks
172		 * of physically contiguous memory, advance through it until
173		 * used up and then allocate more.  Of course, we need
174		 * memory to store those extra pointers, now.  32KB seems to
175		 * be the most that is "safe" under memory pressure
176		 * (creating large files and then copying them over
177		 * NFS while doing lots of MPI jobs).  The OOM killer can
178		 * get invoked, even though we say we can sleep and this can
179		 * cause significant system problems....
180		 */
181		rcd->rcvegrbuf_size = 0x8000;
182		rcd->rcvegrbufs_perchunk =
183			rcd->rcvegrbuf_size / dd->rcvegrbufsize;
184		rcd->rcvegrbuf_chunks = (rcd->rcvegrcnt +
185			rcd->rcvegrbufs_perchunk - 1) /
186			rcd->rcvegrbufs_perchunk;
187		BUG_ON(!is_power_of_2(rcd->rcvegrbufs_perchunk));
188		rcd->rcvegrbufs_perchunk_shift =
189			ilog2(rcd->rcvegrbufs_perchunk);
190	}
191	return rcd;
192}
193
194/*
195 * Common code for initializing the physical port structure.
196 */
197void qib_init_pportdata(struct qib_pportdata *ppd, struct qib_devdata *dd,
198			u8 hw_pidx, u8 port)
199{
200	ppd->dd = dd;
201	ppd->hw_pidx = hw_pidx;
202	ppd->port = port; /* IB port number, not index */
203
204	spin_lock_init(&ppd->sdma_lock);
205	spin_lock_init(&ppd->lflags_lock);
206	init_waitqueue_head(&ppd->state_wait);
207
208	init_timer(&ppd->symerr_clear_timer);
209	ppd->symerr_clear_timer.function = qib_clear_symerror_on_linkup;
210	ppd->symerr_clear_timer.data = (unsigned long)ppd;
211}
212
213static int init_pioavailregs(struct qib_devdata *dd)
214{
215	int ret, pidx;
216	u64 *status_page;
217
218	dd->pioavailregs_dma = dma_alloc_coherent(
219		&dd->pcidev->dev, PAGE_SIZE, &dd->pioavailregs_phys,
220		GFP_KERNEL);
221	if (!dd->pioavailregs_dma) {
222		qib_dev_err(dd, "failed to allocate PIOavail reg area "
223			    "in memory\n");
224		ret = -ENOMEM;
225		goto done;
226	}
227
228	/*
229	 * We really want L2 cache aligned, but for current CPUs of
230	 * interest, they are the same.
231	 */
232	status_page = (u64 *)
233		((char *) dd->pioavailregs_dma +
234		 ((2 * L1_CACHE_BYTES +
235		   dd->pioavregs * sizeof(u64)) & ~L1_CACHE_BYTES));
236	/* device status comes first, for backwards compatibility */
237	dd->devstatusp = status_page;
238	*status_page++ = 0;
239	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
240		dd->pport[pidx].statusp = status_page;
241		*status_page++ = 0;
242	}
243
244	/*
245	 * Setup buffer to hold freeze and other messages, accessible to
246	 * apps, following statusp.  This is per-unit, not per port.
247	 */
248	dd->freezemsg = (char *) status_page;
249	*dd->freezemsg = 0;
250	/* length of msg buffer is "whatever is left" */
251	ret = (char *) status_page - (char *) dd->pioavailregs_dma;
252	dd->freezelen = PAGE_SIZE - ret;
253
254	ret = 0;
255
256done:
257	return ret;
258}
259
260/**
261 * init_shadow_tids - allocate the shadow TID array
262 * @dd: the qlogic_ib device
263 *
264 * allocate the shadow TID array, so we can qib_munlock previous
265 * entries.  It may make more sense to move the pageshadow to the
266 * ctxt data structure, so we only allocate memory for ctxts actually
267 * in use, since we at 8k per ctxt, now.
268 * We don't want failures here to prevent use of the driver/chip,
269 * so no return value.
270 */
271static void init_shadow_tids(struct qib_devdata *dd)
272{
273	struct page **pages;
274	dma_addr_t *addrs;
275
276	pages = vzalloc(dd->cfgctxts * dd->rcvtidcnt * sizeof(struct page *));
277	if (!pages) {
278		qib_dev_err(dd, "failed to allocate shadow page * "
279			    "array, no expected sends!\n");
280		goto bail;
281	}
282
283	addrs = vzalloc(dd->cfgctxts * dd->rcvtidcnt * sizeof(dma_addr_t));
284	if (!addrs) {
285		qib_dev_err(dd, "failed to allocate shadow dma handle "
286			    "array, no expected sends!\n");
287		goto bail_free;
288	}
289
290	dd->pageshadow = pages;
291	dd->physshadow = addrs;
292	return;
293
294bail_free:
295	vfree(pages);
296bail:
297	dd->pageshadow = NULL;
298}
299
300/*
301 * Do initialization for device that is only needed on
302 * first detect, not on resets.
303 */
304static int loadtime_init(struct qib_devdata *dd)
305{
306	int ret = 0;
307
308	if (((dd->revision >> QLOGIC_IB_R_SOFTWARE_SHIFT) &
309	     QLOGIC_IB_R_SOFTWARE_MASK) != QIB_CHIP_SWVERSION) {
310		qib_dev_err(dd, "Driver only handles version %d, "
311			    "chip swversion is %d (%llx), failng\n",
312			    QIB_CHIP_SWVERSION,
313			    (int)(dd->revision >>
314				QLOGIC_IB_R_SOFTWARE_SHIFT) &
315			    QLOGIC_IB_R_SOFTWARE_MASK,
316			    (unsigned long long) dd->revision);
317		ret = -ENOSYS;
318		goto done;
319	}
320
321	if (dd->revision & QLOGIC_IB_R_EMULATOR_MASK)
322		qib_devinfo(dd->pcidev, "%s", dd->boardversion);
323
324	spin_lock_init(&dd->pioavail_lock);
325	spin_lock_init(&dd->sendctrl_lock);
326	spin_lock_init(&dd->uctxt_lock);
327	spin_lock_init(&dd->qib_diag_trans_lock);
328	spin_lock_init(&dd->eep_st_lock);
329	mutex_init(&dd->eep_lock);
330
331	if (qib_mini_init)
332		goto done;
333
334	ret = init_pioavailregs(dd);
335	init_shadow_tids(dd);
336
337	qib_get_eeprom_info(dd);
338
339	/* setup time (don't start yet) to verify we got interrupt */
340	init_timer(&dd->intrchk_timer);
341	dd->intrchk_timer.function = verify_interrupt;
342	dd->intrchk_timer.data = (unsigned long) dd;
343
344done:
345	return ret;
346}
347
348/**
349 * init_after_reset - re-initialize after a reset
350 * @dd: the qlogic_ib device
351 *
352 * sanity check at least some of the values after reset, and
353 * ensure no receive or transmit (explicitly, in case reset
354 * failed
355 */
356static int init_after_reset(struct qib_devdata *dd)
357{
358	int i;
359
360	/*
361	 * Ensure chip does no sends or receives, tail updates, or
362	 * pioavail updates while we re-initialize.  This is mostly
363	 * for the driver data structures, not chip registers.
364	 */
365	for (i = 0; i < dd->num_pports; ++i) {
366		/*
367		 * ctxt == -1 means "all contexts". Only really safe for
368		 * _dis_abling things, as here.
369		 */
370		dd->f_rcvctrl(dd->pport + i, QIB_RCVCTRL_CTXT_DIS |
371				  QIB_RCVCTRL_INTRAVAIL_DIS |
372				  QIB_RCVCTRL_TAILUPD_DIS, -1);
373		/* Redundant across ports for some, but no big deal.  */
374		dd->f_sendctrl(dd->pport + i, QIB_SENDCTRL_SEND_DIS |
375			QIB_SENDCTRL_AVAIL_DIS);
376	}
377
378	return 0;
379}
380
381static void enable_chip(struct qib_devdata *dd)
382{
383	u64 rcvmask;
384	int i;
385
386	/*
387	 * Enable PIO send, and update of PIOavail regs to memory.
388	 */
389	for (i = 0; i < dd->num_pports; ++i)
390		dd->f_sendctrl(dd->pport + i, QIB_SENDCTRL_SEND_ENB |
391			QIB_SENDCTRL_AVAIL_ENB);
392	/*
393	 * Enable kernel ctxts' receive and receive interrupt.
394	 * Other ctxts done as user opens and inits them.
395	 */
396	rcvmask = QIB_RCVCTRL_CTXT_ENB | QIB_RCVCTRL_INTRAVAIL_ENB;
397	rcvmask |= (dd->flags & QIB_NODMA_RTAIL) ?
398		  QIB_RCVCTRL_TAILUPD_DIS : QIB_RCVCTRL_TAILUPD_ENB;
399	for (i = 0; dd->rcd && i < dd->first_user_ctxt; ++i) {
400		struct qib_ctxtdata *rcd = dd->rcd[i];
401
402		if (rcd)
403			dd->f_rcvctrl(rcd->ppd, rcvmask, i);
404	}
405	dd->freectxts = dd->cfgctxts - dd->first_user_ctxt;
406}
407
408static void verify_interrupt(unsigned long opaque)
409{
410	struct qib_devdata *dd = (struct qib_devdata *) opaque;
411
412	if (!dd)
413		return; /* being torn down */
414
415	/*
416	 * If we don't have a lid or any interrupts, let the user know and
417	 * don't bother checking again.
418	 */
419	if (dd->int_counter == 0) {
420		if (!dd->f_intr_fallback(dd))
421			dev_err(&dd->pcidev->dev, "No interrupts detected, "
422				"not usable.\n");
423		else /* re-arm the timer to see if fallback works */
424			mod_timer(&dd->intrchk_timer, jiffies + HZ/2);
425	}
426}
427
428static void init_piobuf_state(struct qib_devdata *dd)
429{
430	int i, pidx;
431	u32 uctxts;
432
433	/*
434	 * Ensure all buffers are free, and fifos empty.  Buffers
435	 * are common, so only do once for port 0.
436	 *
437	 * After enable and qib_chg_pioavailkernel so we can safely
438	 * enable pioavail updates and PIOENABLE.  After this, packets
439	 * are ready and able to go out.
440	 */
441	dd->f_sendctrl(dd->pport, QIB_SENDCTRL_DISARM_ALL);
442	for (pidx = 0; pidx < dd->num_pports; ++pidx)
443		dd->f_sendctrl(dd->pport + pidx, QIB_SENDCTRL_FLUSH);
444
445	/*
446	 * If not all sendbufs are used, add the one to each of the lower
447	 * numbered contexts.  pbufsctxt and lastctxt_piobuf are
448	 * calculated in chip-specific code because it may cause some
449	 * chip-specific adjustments to be made.
450	 */
451	uctxts = dd->cfgctxts - dd->first_user_ctxt;
452	dd->ctxts_extrabuf = dd->pbufsctxt ?
453		dd->lastctxt_piobuf - (dd->pbufsctxt * uctxts) : 0;
454
455	/*
456	 * Set up the shadow copies of the piobufavail registers,
457	 * which we compare against the chip registers for now, and
458	 * the in memory DMA'ed copies of the registers.
459	 * By now pioavail updates to memory should have occurred, so
460	 * copy them into our working/shadow registers; this is in
461	 * case something went wrong with abort, but mostly to get the
462	 * initial values of the generation bit correct.
463	 */
464	for (i = 0; i < dd->pioavregs; i++) {
465		__le64 tmp;
466
467		tmp = dd->pioavailregs_dma[i];
468		/*
469		 * Don't need to worry about pioavailkernel here
470		 * because we will call qib_chg_pioavailkernel() later
471		 * in initialization, to busy out buffers as needed.
472		 */
473		dd->pioavailshadow[i] = le64_to_cpu(tmp);
474	}
475	while (i < ARRAY_SIZE(dd->pioavailshadow))
476		dd->pioavailshadow[i++] = 0; /* for debugging sanity */
477
478	/* after pioavailshadow is setup */
479	qib_chg_pioavailkernel(dd, 0, dd->piobcnt2k + dd->piobcnt4k,
480			       TXCHK_CHG_TYPE_KERN, NULL);
481	dd->f_initvl15_bufs(dd);
482}
483
484/**
485 * qib_init - do the actual initialization sequence on the chip
486 * @dd: the qlogic_ib device
487 * @reinit: reinitializing, so don't allocate new memory
488 *
489 * Do the actual initialization sequence on the chip.  This is done
490 * both from the init routine called from the PCI infrastructure, and
491 * when we reset the chip, or detect that it was reset internally,
492 * or it's administratively re-enabled.
493 *
494 * Memory allocation here and in called routines is only done in
495 * the first case (reinit == 0).  We have to be careful, because even
496 * without memory allocation, we need to re-write all the chip registers
497 * TIDs, etc. after the reset or enable has completed.
498 */
499int qib_init(struct qib_devdata *dd, int reinit)
500{
501	int ret = 0, pidx, lastfail = 0;
502	u32 portok = 0;
503	unsigned i;
504	struct qib_ctxtdata *rcd;
505	struct qib_pportdata *ppd;
506	unsigned long flags;
507
508	/* Set linkstate to unknown, so we can watch for a transition. */
509	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
510		ppd = dd->pport + pidx;
511		spin_lock_irqsave(&ppd->lflags_lock, flags);
512		ppd->lflags &= ~(QIBL_LINKACTIVE | QIBL_LINKARMED |
513				 QIBL_LINKDOWN | QIBL_LINKINIT |
514				 QIBL_LINKV);
515		spin_unlock_irqrestore(&ppd->lflags_lock, flags);
516	}
517
518	if (reinit)
519		ret = init_after_reset(dd);
520	else
521		ret = loadtime_init(dd);
522	if (ret)
523		goto done;
524
525	/* Bypass most chip-init, to get to device creation */
526	if (qib_mini_init)
527		return 0;
528
529	ret = dd->f_late_initreg(dd);
530	if (ret)
531		goto done;
532
533	/* dd->rcd can be NULL if early init failed */
534	for (i = 0; dd->rcd && i < dd->first_user_ctxt; ++i) {
535		/*
536		 * Set up the (kernel) rcvhdr queue and egr TIDs.  If doing
537		 * re-init, the simplest way to handle this is to free
538		 * existing, and re-allocate.
539		 * Need to re-create rest of ctxt 0 ctxtdata as well.
540		 */
541		rcd = dd->rcd[i];
542		if (!rcd)
543			continue;
544
545		lastfail = qib_create_rcvhdrq(dd, rcd);
546		if (!lastfail)
547			lastfail = qib_setup_eagerbufs(rcd);
548		if (lastfail) {
549			qib_dev_err(dd, "failed to allocate kernel ctxt's "
550				    "rcvhdrq and/or egr bufs\n");
551			continue;
552		}
553	}
554
555	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
556		int mtu;
557		if (lastfail)
558			ret = lastfail;
559		ppd = dd->pport + pidx;
560		mtu = ib_mtu_enum_to_int(qib_ibmtu);
561		if (mtu == -1) {
562			mtu = QIB_DEFAULT_MTU;
563			qib_ibmtu = 0; /* don't leave invalid value */
564		}
565		/* set max we can ever have for this driver load */
566		ppd->init_ibmaxlen = min(mtu > 2048 ?
567					 dd->piosize4k : dd->piosize2k,
568					 dd->rcvegrbufsize +
569					 (dd->rcvhdrentsize << 2));
570		/*
571		 * Have to initialize ibmaxlen, but this will normally
572		 * change immediately in qib_set_mtu().
573		 */
574		ppd->ibmaxlen = ppd->init_ibmaxlen;
575		qib_set_mtu(ppd, mtu);
576
577		spin_lock_irqsave(&ppd->lflags_lock, flags);
578		ppd->lflags |= QIBL_IB_LINK_DISABLED;
579		spin_unlock_irqrestore(&ppd->lflags_lock, flags);
580
581		lastfail = dd->f_bringup_serdes(ppd);
582		if (lastfail) {
583			qib_devinfo(dd->pcidev,
584				 "Failed to bringup IB port %u\n", ppd->port);
585			lastfail = -ENETDOWN;
586			continue;
587		}
588
589		portok++;
590	}
591
592	if (!portok) {
593		/* none of the ports initialized */
594		if (!ret && lastfail)
595			ret = lastfail;
596		else if (!ret)
597			ret = -ENETDOWN;
598		/* but continue on, so we can debug cause */
599	}
600
601	enable_chip(dd);
602
603	init_piobuf_state(dd);
604
605done:
606	if (!ret) {
607		/* chip is OK for user apps; mark it as initialized */
608		for (pidx = 0; pidx < dd->num_pports; ++pidx) {
609			ppd = dd->pport + pidx;
610			/*
611			 * Set status even if port serdes is not initialized
612			 * so that diags will work.
613			 */
614			*ppd->statusp |= QIB_STATUS_CHIP_PRESENT |
615				QIB_STATUS_INITTED;
616			if (!ppd->link_speed_enabled)
617				continue;
618			if (dd->flags & QIB_HAS_SEND_DMA)
619				ret = qib_setup_sdma(ppd);
620			init_timer(&ppd->hol_timer);
621			ppd->hol_timer.function = qib_hol_event;
622			ppd->hol_timer.data = (unsigned long)ppd;
623			ppd->hol_state = QIB_HOL_UP;
624		}
625
626		/* now we can enable all interrupts from the chip */
627		dd->f_set_intr_state(dd, 1);
628
629		/*
630		 * Setup to verify we get an interrupt, and fallback
631		 * to an alternate if necessary and possible.
632		 */
633		mod_timer(&dd->intrchk_timer, jiffies + HZ/2);
634		/* start stats retrieval timer */
635		mod_timer(&dd->stats_timer, jiffies + HZ * ACTIVITY_TIMER);
636	}
637
638	/* if ret is non-zero, we probably should do some cleanup here... */
639	return ret;
640}
641
642/*
643 * These next two routines are placeholders in case we don't have per-arch
644 * code for controlling write combining.  If explicit control of write
645 * combining is not available, performance will probably be awful.
646 */
647
648int __attribute__((weak)) qib_enable_wc(struct qib_devdata *dd)
649{
650	return -EOPNOTSUPP;
651}
652
653void __attribute__((weak)) qib_disable_wc(struct qib_devdata *dd)
654{
655}
656
657static inline struct qib_devdata *__qib_lookup(int unit)
658{
659	return idr_find(&qib_unit_table, unit);
660}
661
662struct qib_devdata *qib_lookup(int unit)
663{
664	struct qib_devdata *dd;
665	unsigned long flags;
666
667	spin_lock_irqsave(&qib_devs_lock, flags);
668	dd = __qib_lookup(unit);
669	spin_unlock_irqrestore(&qib_devs_lock, flags);
670
671	return dd;
672}
673
674/*
675 * Stop the timers during unit shutdown, or after an error late
676 * in initialization.
677 */
678static void qib_stop_timers(struct qib_devdata *dd)
679{
680	struct qib_pportdata *ppd;
681	int pidx;
682
683	if (dd->stats_timer.data) {
684		del_timer_sync(&dd->stats_timer);
685		dd->stats_timer.data = 0;
686	}
687	if (dd->intrchk_timer.data) {
688		del_timer_sync(&dd->intrchk_timer);
689		dd->intrchk_timer.data = 0;
690	}
691	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
692		ppd = dd->pport + pidx;
693		if (ppd->hol_timer.data)
694			del_timer_sync(&ppd->hol_timer);
695		if (ppd->led_override_timer.data) {
696			del_timer_sync(&ppd->led_override_timer);
697			atomic_set(&ppd->led_override_timer_active, 0);
698		}
699		if (ppd->symerr_clear_timer.data)
700			del_timer_sync(&ppd->symerr_clear_timer);
701	}
702}
703
704/**
705 * qib_shutdown_device - shut down a device
706 * @dd: the qlogic_ib device
707 *
708 * This is called to make the device quiet when we are about to
709 * unload the driver, and also when the device is administratively
710 * disabled.   It does not free any data structures.
711 * Everything it does has to be setup again by qib_init(dd, 1)
712 */
713static void qib_shutdown_device(struct qib_devdata *dd)
714{
715	struct qib_pportdata *ppd;
716	unsigned pidx;
717
718	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
719		ppd = dd->pport + pidx;
720
721		spin_lock_irq(&ppd->lflags_lock);
722		ppd->lflags &= ~(QIBL_LINKDOWN | QIBL_LINKINIT |
723				 QIBL_LINKARMED | QIBL_LINKACTIVE |
724				 QIBL_LINKV);
725		spin_unlock_irq(&ppd->lflags_lock);
726		*ppd->statusp &= ~(QIB_STATUS_IB_CONF | QIB_STATUS_IB_READY);
727	}
728	dd->flags &= ~QIB_INITTED;
729
730	/* mask interrupts, but not errors */
731	dd->f_set_intr_state(dd, 0);
732
733	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
734		ppd = dd->pport + pidx;
735		dd->f_rcvctrl(ppd, QIB_RCVCTRL_TAILUPD_DIS |
736				   QIB_RCVCTRL_CTXT_DIS |
737				   QIB_RCVCTRL_INTRAVAIL_DIS |
738				   QIB_RCVCTRL_PKEY_ENB, -1);
739		/*
740		 * Gracefully stop all sends allowing any in progress to
741		 * trickle out first.
742		 */
743		dd->f_sendctrl(ppd, QIB_SENDCTRL_CLEAR);
744	}
745
746	/*
747	 * Enough for anything that's going to trickle out to have actually
748	 * done so.
749	 */
750	udelay(20);
751
752	for (pidx = 0; pidx < dd->num_pports; ++pidx) {
753		ppd = dd->pport + pidx;
754		dd->f_setextled(ppd, 0); /* make sure LEDs are off */
755
756		if (dd->flags & QIB_HAS_SEND_DMA)
757			qib_teardown_sdma(ppd);
758
759		dd->f_sendctrl(ppd, QIB_SENDCTRL_AVAIL_DIS |
760				    QIB_SENDCTRL_SEND_DIS);
761		/*
762		 * Clear SerdesEnable.
763		 * We can't count on interrupts since we are stopping.
764		 */
765		dd->f_quiet_serdes(ppd);
766	}
767
768	qib_update_eeprom_log(dd);
769}
770
771/**
772 * qib_free_ctxtdata - free a context's allocated data
773 * @dd: the qlogic_ib device
774 * @rcd: the ctxtdata structure
775 *
776 * free up any allocated data for a context
777 * This should not touch anything that would affect a simultaneous
778 * re-allocation of context data, because it is called after qib_mutex
779 * is released (and can be called from reinit as well).
780 * It should never change any chip state, or global driver state.
781 */
782void qib_free_ctxtdata(struct qib_devdata *dd, struct qib_ctxtdata *rcd)
783{
784	if (!rcd)
785		return;
786
787	if (rcd->rcvhdrq) {
788		dma_free_coherent(&dd->pcidev->dev, rcd->rcvhdrq_size,
789				  rcd->rcvhdrq, rcd->rcvhdrq_phys);
790		rcd->rcvhdrq = NULL;
791		if (rcd->rcvhdrtail_kvaddr) {
792			dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE,
793					  rcd->rcvhdrtail_kvaddr,
794					  rcd->rcvhdrqtailaddr_phys);
795			rcd->rcvhdrtail_kvaddr = NULL;
796		}
797	}
798	if (rcd->rcvegrbuf) {
799		unsigned e;
800
801		for (e = 0; e < rcd->rcvegrbuf_chunks; e++) {
802			void *base = rcd->rcvegrbuf[e];
803			size_t size = rcd->rcvegrbuf_size;
804
805			dma_free_coherent(&dd->pcidev->dev, size,
806					  base, rcd->rcvegrbuf_phys[e]);
807		}
808		kfree(rcd->rcvegrbuf);
809		rcd->rcvegrbuf = NULL;
810		kfree(rcd->rcvegrbuf_phys);
811		rcd->rcvegrbuf_phys = NULL;
812		rcd->rcvegrbuf_chunks = 0;
813	}
814
815	kfree(rcd->tid_pg_list);
816	vfree(rcd->user_event_mask);
817	vfree(rcd->subctxt_uregbase);
818	vfree(rcd->subctxt_rcvegrbuf);
819	vfree(rcd->subctxt_rcvhdr_base);
820	kfree(rcd);
821}
822
823/*
824 * Perform a PIO buffer bandwidth write test, to verify proper system
825 * configuration.  Even when all the setup calls work, occasionally
826 * BIOS or other issues can prevent write combining from working, or
827 * can cause other bandwidth problems to the chip.
828 *
829 * This test simply writes the same buffer over and over again, and
830 * measures close to the peak bandwidth to the chip (not testing
831 * data bandwidth to the wire).   On chips that use an address-based
832 * trigger to send packets to the wire, this is easy.  On chips that
833 * use a count to trigger, we want to make sure that the packet doesn't
834 * go out on the wire, or trigger flow control checks.
835 */
836static void qib_verify_pioperf(struct qib_devdata *dd)
837{
838	u32 pbnum, cnt, lcnt;
839	u32 __iomem *piobuf;
840	u32 *addr;
841	u64 msecs, emsecs;
842
843	piobuf = dd->f_getsendbuf(dd->pport, 0ULL, &pbnum);
844	if (!piobuf) {
845		qib_devinfo(dd->pcidev,
846			 "No PIObufs for checking perf, skipping\n");
847		return;
848	}
849
850	/*
851	 * Enough to give us a reasonable test, less than piobuf size, and
852	 * likely multiple of store buffer length.
853	 */
854	cnt = 1024;
855
856	addr = vmalloc(cnt);
857	if (!addr) {
858		qib_devinfo(dd->pcidev,
859			 "Couldn't get memory for checking PIO perf,"
860			 " skipping\n");
861		goto done;
862	}
863
864	preempt_disable();  /* we want reasonably accurate elapsed time */
865	msecs = 1 + jiffies_to_msecs(jiffies);
866	for (lcnt = 0; lcnt < 10000U; lcnt++) {
867		/* wait until we cross msec boundary */
868		if (jiffies_to_msecs(jiffies) >= msecs)
869			break;
870		udelay(1);
871	}
872
873	dd->f_set_armlaunch(dd, 0);
874
875	/*
876	 * length 0, no dwords actually sent
877	 */
878	writeq(0, piobuf);
879	qib_flush_wc();
880
881	/*
882	 * This is only roughly accurate, since even with preempt we
883	 * still take interrupts that could take a while.   Running for
884	 * >= 5 msec seems to get us "close enough" to accurate values.
885	 */
886	msecs = jiffies_to_msecs(jiffies);
887	for (emsecs = lcnt = 0; emsecs <= 5UL; lcnt++) {
888		qib_pio_copy(piobuf + 64, addr, cnt >> 2);
889		emsecs = jiffies_to_msecs(jiffies) - msecs;
890	}
891
892	/* 1 GiB/sec, slightly over IB SDR line rate */
893	if (lcnt < (emsecs * 1024U))
894		qib_dev_err(dd,
895			    "Performance problem: bandwidth to PIO buffers is "
896			    "only %u MiB/sec\n",
897			    lcnt / (u32) emsecs);
898
899	preempt_enable();
900
901	vfree(addr);
902
903done:
904	/* disarm piobuf, so it's available again */
905	dd->f_sendctrl(dd->pport, QIB_SENDCTRL_DISARM_BUF(pbnum));
906	qib_sendbuf_done(dd, pbnum);
907	dd->f_set_armlaunch(dd, 1);
908}
909
910
911void qib_free_devdata(struct qib_devdata *dd)
912{
913	unsigned long flags;
914
915	spin_lock_irqsave(&qib_devs_lock, flags);
916	idr_remove(&qib_unit_table, dd->unit);
917	list_del(&dd->list);
918	spin_unlock_irqrestore(&qib_devs_lock, flags);
919
920	ib_dealloc_device(&dd->verbs_dev.ibdev);
921}
922
923/*
924 * Allocate our primary per-unit data structure.  Must be done via verbs
925 * allocator, because the verbs cleanup process both does cleanup and
926 * free of the data structure.
927 * "extra" is for chip-specific data.
928 *
929 * Use the idr mechanism to get a unit number for this unit.
930 */
931struct qib_devdata *qib_alloc_devdata(struct pci_dev *pdev, size_t extra)
932{
933	unsigned long flags;
934	struct qib_devdata *dd;
935	int ret;
936
937	if (!idr_pre_get(&qib_unit_table, GFP_KERNEL)) {
938		dd = ERR_PTR(-ENOMEM);
939		goto bail;
940	}
941
942	dd = (struct qib_devdata *) ib_alloc_device(sizeof(*dd) + extra);
943	if (!dd) {
944		dd = ERR_PTR(-ENOMEM);
945		goto bail;
946	}
947
948	spin_lock_irqsave(&qib_devs_lock, flags);
949	ret = idr_get_new(&qib_unit_table, dd, &dd->unit);
950	if (ret >= 0)
951		list_add(&dd->list, &qib_dev_list);
952	spin_unlock_irqrestore(&qib_devs_lock, flags);
953
954	if (ret < 0) {
955		qib_early_err(&pdev->dev,
956			      "Could not allocate unit ID: error %d\n", -ret);
957		ib_dealloc_device(&dd->verbs_dev.ibdev);
958		dd = ERR_PTR(ret);
959		goto bail;
960	}
961
962	if (!qib_cpulist_count) {
963		u32 count = num_online_cpus();
964		qib_cpulist = kzalloc(BITS_TO_LONGS(count) *
965				      sizeof(long), GFP_KERNEL);
966		if (qib_cpulist)
967			qib_cpulist_count = count;
968		else
969			qib_early_err(&pdev->dev, "Could not alloc cpulist "
970				      "info, cpu affinity might be wrong\n");
971	}
972
973bail:
974	return dd;
975}
976
977/*
978 * Called from freeze mode handlers, and from PCI error
979 * reporting code.  Should be paranoid about state of
980 * system and data structures.
981 */
982void qib_disable_after_error(struct qib_devdata *dd)
983{
984	if (dd->flags & QIB_INITTED) {
985		u32 pidx;
986
987		dd->flags &= ~QIB_INITTED;
988		if (dd->pport)
989			for (pidx = 0; pidx < dd->num_pports; ++pidx) {
990				struct qib_pportdata *ppd;
991
992				ppd = dd->pport + pidx;
993				if (dd->flags & QIB_PRESENT) {
994					qib_set_linkstate(ppd,
995						QIB_IB_LINKDOWN_DISABLE);
996					dd->f_setextled(ppd, 0);
997				}
998				*ppd->statusp &= ~QIB_STATUS_IB_READY;
999			}
1000	}
1001
1002	/*
1003	 * Mark as having had an error for driver, and also
1004	 * for /sys and status word mapped to user programs.
1005	 * This marks unit as not usable, until reset.
1006	 */
1007	if (dd->devstatusp)
1008		*dd->devstatusp |= QIB_STATUS_HWERROR;
1009}
1010
1011static void __devexit qib_remove_one(struct pci_dev *);
1012static int __devinit qib_init_one(struct pci_dev *,
1013				  const struct pci_device_id *);
1014
1015#define DRIVER_LOAD_MSG "QLogic " QIB_DRV_NAME " loaded: "
1016#define PFX QIB_DRV_NAME ": "
1017
1018static DEFINE_PCI_DEVICE_TABLE(qib_pci_tbl) = {
1019	{ PCI_DEVICE(PCI_VENDOR_ID_PATHSCALE, PCI_DEVICE_ID_QLOGIC_IB_6120) },
1020	{ PCI_DEVICE(PCI_VENDOR_ID_QLOGIC, PCI_DEVICE_ID_QLOGIC_IB_7220) },
1021	{ PCI_DEVICE(PCI_VENDOR_ID_QLOGIC, PCI_DEVICE_ID_QLOGIC_IB_7322) },
1022	{ 0, }
1023};
1024
1025MODULE_DEVICE_TABLE(pci, qib_pci_tbl);
1026
1027struct pci_driver qib_driver = {
1028	.name = QIB_DRV_NAME,
1029	.probe = qib_init_one,
1030	.remove = __devexit_p(qib_remove_one),
1031	.id_table = qib_pci_tbl,
1032	.err_handler = &qib_pci_err_handler,
1033};
1034
1035/*
1036 * Do all the generic driver unit- and chip-independent memory
1037 * allocation and initialization.
1038 */
1039static int __init qlogic_ib_init(void)
1040{
1041	int ret;
1042
1043	ret = qib_dev_init();
1044	if (ret)
1045		goto bail;
1046
1047	qib_cq_wq = create_singlethread_workqueue("qib_cq");
1048	if (!qib_cq_wq) {
1049		ret = -ENOMEM;
1050		goto bail_dev;
1051	}
1052
1053	/*
1054	 * These must be called before the driver is registered with
1055	 * the PCI subsystem.
1056	 */
1057	idr_init(&qib_unit_table);
1058	if (!idr_pre_get(&qib_unit_table, GFP_KERNEL)) {
1059		printk(KERN_ERR QIB_DRV_NAME ": idr_pre_get() failed\n");
1060		ret = -ENOMEM;
1061		goto bail_cq_wq;
1062	}
1063
1064	ret = pci_register_driver(&qib_driver);
1065	if (ret < 0) {
1066		printk(KERN_ERR QIB_DRV_NAME
1067		       ": Unable to register driver: error %d\n", -ret);
1068		goto bail_unit;
1069	}
1070
1071	/* not fatal if it doesn't work */
1072	if (qib_init_qibfs())
1073		printk(KERN_ERR QIB_DRV_NAME ": Unable to register ipathfs\n");
1074	goto bail; /* all OK */
1075
1076bail_unit:
1077	idr_destroy(&qib_unit_table);
1078bail_cq_wq:
1079	destroy_workqueue(qib_cq_wq);
1080bail_dev:
1081	qib_dev_cleanup();
1082bail:
1083	return ret;
1084}
1085
1086module_init(qlogic_ib_init);
1087
1088/*
1089 * Do the non-unit driver cleanup, memory free, etc. at unload.
1090 */
1091static void __exit qlogic_ib_cleanup(void)
1092{
1093	int ret;
1094
1095	ret = qib_exit_qibfs();
1096	if (ret)
1097		printk(KERN_ERR QIB_DRV_NAME ": "
1098			"Unable to cleanup counter filesystem: "
1099			"error %d\n", -ret);
1100
1101	pci_unregister_driver(&qib_driver);
1102
1103	destroy_workqueue(qib_cq_wq);
1104
1105	qib_cpulist_count = 0;
1106	kfree(qib_cpulist);
1107
1108	idr_destroy(&qib_unit_table);
1109	qib_dev_cleanup();
1110}
1111
1112module_exit(qlogic_ib_cleanup);
1113
1114/* this can only be called after a successful initialization */
1115static void cleanup_device_data(struct qib_devdata *dd)
1116{
1117	int ctxt;
1118	int pidx;
1119	struct qib_ctxtdata **tmp;
1120	unsigned long flags;
1121
1122	/* users can't do anything more with chip */
1123	for (pidx = 0; pidx < dd->num_pports; ++pidx)
1124		if (dd->pport[pidx].statusp)
1125			*dd->pport[pidx].statusp &= ~QIB_STATUS_CHIP_PRESENT;
1126
1127	if (!qib_wc_pat)
1128		qib_disable_wc(dd);
1129
1130	if (dd->pioavailregs_dma) {
1131		dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE,
1132				  (void *) dd->pioavailregs_dma,
1133				  dd->pioavailregs_phys);
1134		dd->pioavailregs_dma = NULL;
1135	}
1136
1137	if (dd->pageshadow) {
1138		struct page **tmpp = dd->pageshadow;
1139		dma_addr_t *tmpd = dd->physshadow;
1140		int i, cnt = 0;
1141
1142		for (ctxt = 0; ctxt < dd->cfgctxts; ctxt++) {
1143			int ctxt_tidbase = ctxt * dd->rcvtidcnt;
1144			int maxtid = ctxt_tidbase + dd->rcvtidcnt;
1145
1146			for (i = ctxt_tidbase; i < maxtid; i++) {
1147				if (!tmpp[i])
1148					continue;
1149				pci_unmap_page(dd->pcidev, tmpd[i],
1150					       PAGE_SIZE, PCI_DMA_FROMDEVICE);
1151				qib_release_user_pages(&tmpp[i], 1);
1152				tmpp[i] = NULL;
1153				cnt++;
1154			}
1155		}
1156
1157		tmpp = dd->pageshadow;
1158		dd->pageshadow = NULL;
1159		vfree(tmpp);
1160	}
1161
1162	/*
1163	 * Free any resources still in use (usually just kernel contexts)
1164	 * at unload; we do for ctxtcnt, because that's what we allocate.
1165	 * We acquire lock to be really paranoid that rcd isn't being
1166	 * accessed from some interrupt-related code (that should not happen,
1167	 * but best to be sure).
1168	 */
1169	spin_lock_irqsave(&dd->uctxt_lock, flags);
1170	tmp = dd->rcd;
1171	dd->rcd = NULL;
1172	spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1173	for (ctxt = 0; tmp && ctxt < dd->ctxtcnt; ctxt++) {
1174		struct qib_ctxtdata *rcd = tmp[ctxt];
1175
1176		tmp[ctxt] = NULL; /* debugging paranoia */
1177		qib_free_ctxtdata(dd, rcd);
1178	}
1179	kfree(tmp);
1180	kfree(dd->boardname);
1181}
1182
1183/*
1184 * Clean up on unit shutdown, or error during unit load after
1185 * successful initialization.
1186 */
1187static void qib_postinit_cleanup(struct qib_devdata *dd)
1188{
1189	/*
1190	 * Clean up chip-specific stuff.
1191	 * We check for NULL here, because it's outside
1192	 * the kregbase check, and we need to call it
1193	 * after the free_irq.  Thus it's possible that
1194	 * the function pointers were never initialized.
1195	 */
1196	if (dd->f_cleanup)
1197		dd->f_cleanup(dd);
1198
1199	qib_pcie_ddcleanup(dd);
1200
1201	cleanup_device_data(dd);
1202
1203	qib_free_devdata(dd);
1204}
1205
1206static int __devinit qib_init_one(struct pci_dev *pdev,
1207				  const struct pci_device_id *ent)
1208{
1209	int ret, j, pidx, initfail;
1210	struct qib_devdata *dd = NULL;
1211
1212	ret = qib_pcie_init(pdev, ent);
1213	if (ret)
1214		goto bail;
1215
1216	/*
1217	 * Do device-specific initialiation, function table setup, dd
1218	 * allocation, etc.
1219	 */
1220	switch (ent->device) {
1221	case PCI_DEVICE_ID_QLOGIC_IB_6120:
1222#ifdef CONFIG_PCI_MSI
1223		dd = qib_init_iba6120_funcs(pdev, ent);
1224#else
1225		qib_early_err(&pdev->dev, "QLogic PCIE device 0x%x cannot "
1226		      "work if CONFIG_PCI_MSI is not enabled\n",
1227		      ent->device);
1228		dd = ERR_PTR(-ENODEV);
1229#endif
1230		break;
1231
1232	case PCI_DEVICE_ID_QLOGIC_IB_7220:
1233		dd = qib_init_iba7220_funcs(pdev, ent);
1234		break;
1235
1236	case PCI_DEVICE_ID_QLOGIC_IB_7322:
1237		dd = qib_init_iba7322_funcs(pdev, ent);
1238		break;
1239
1240	default:
1241		qib_early_err(&pdev->dev, "Failing on unknown QLogic "
1242			      "deviceid 0x%x\n", ent->device);
1243		ret = -ENODEV;
1244	}
1245
1246	if (IS_ERR(dd))
1247		ret = PTR_ERR(dd);
1248	if (ret)
1249		goto bail; /* error already printed */
1250
1251	/* do the generic initialization */
1252	initfail = qib_init(dd, 0);
1253
1254	ret = qib_register_ib_device(dd);
1255
1256	/*
1257	 * Now ready for use.  this should be cleared whenever we
1258	 * detect a reset, or initiate one.  If earlier failure,
1259	 * we still create devices, so diags, etc. can be used
1260	 * to determine cause of problem.
1261	 */
1262	if (!qib_mini_init && !initfail && !ret)
1263		dd->flags |= QIB_INITTED;
1264
1265	j = qib_device_create(dd);
1266	if (j)
1267		qib_dev_err(dd, "Failed to create /dev devices: %d\n", -j);
1268	j = qibfs_add(dd);
1269	if (j)
1270		qib_dev_err(dd, "Failed filesystem setup for counters: %d\n",
1271			    -j);
1272
1273	if (qib_mini_init || initfail || ret) {
1274		qib_stop_timers(dd);
1275		flush_workqueue(ib_wq);
1276		for (pidx = 0; pidx < dd->num_pports; ++pidx)
1277			dd->f_quiet_serdes(dd->pport + pidx);
1278		if (qib_mini_init)
1279			goto bail;
1280		if (!j) {
1281			(void) qibfs_remove(dd);
1282			qib_device_remove(dd);
1283		}
1284		if (!ret)
1285			qib_unregister_ib_device(dd);
1286		qib_postinit_cleanup(dd);
1287		if (initfail)
1288			ret = initfail;
1289		goto bail;
1290	}
1291
1292	if (!qib_wc_pat) {
1293		ret = qib_enable_wc(dd);
1294		if (ret) {
1295			qib_dev_err(dd, "Write combining not enabled "
1296				    "(err %d): performance may be poor\n",
1297				    -ret);
1298			ret = 0;
1299		}
1300	}
1301
1302	qib_verify_pioperf(dd);
1303bail:
1304	return ret;
1305}
1306
1307static void __devexit qib_remove_one(struct pci_dev *pdev)
1308{
1309	struct qib_devdata *dd = pci_get_drvdata(pdev);
1310	int ret;
1311
1312	/* unregister from IB core */
1313	qib_unregister_ib_device(dd);
1314
1315	/*
1316	 * Disable the IB link, disable interrupts on the device,
1317	 * clear dma engines, etc.
1318	 */
1319	if (!qib_mini_init)
1320		qib_shutdown_device(dd);
1321
1322	qib_stop_timers(dd);
1323
1324	/* wait until all of our (qsfp) queue_work() calls complete */
1325	flush_workqueue(ib_wq);
1326
1327	ret = qibfs_remove(dd);
1328	if (ret)
1329		qib_dev_err(dd, "Failed counters filesystem cleanup: %d\n",
1330			    -ret);
1331
1332	qib_device_remove(dd);
1333
1334	qib_postinit_cleanup(dd);
1335}
1336
1337/**
1338 * qib_create_rcvhdrq - create a receive header queue
1339 * @dd: the qlogic_ib device
1340 * @rcd: the context data
1341 *
1342 * This must be contiguous memory (from an i/o perspective), and must be
1343 * DMA'able (which means for some systems, it will go through an IOMMU,
1344 * or be forced into a low address range).
1345 */
1346int qib_create_rcvhdrq(struct qib_devdata *dd, struct qib_ctxtdata *rcd)
1347{
1348	unsigned amt;
1349
1350	if (!rcd->rcvhdrq) {
1351		dma_addr_t phys_hdrqtail;
1352		gfp_t gfp_flags;
1353
1354		amt = ALIGN(dd->rcvhdrcnt * dd->rcvhdrentsize *
1355			    sizeof(u32), PAGE_SIZE);
1356		gfp_flags = (rcd->ctxt >= dd->first_user_ctxt) ?
1357			GFP_USER : GFP_KERNEL;
1358		rcd->rcvhdrq = dma_alloc_coherent(
1359			&dd->pcidev->dev, amt, &rcd->rcvhdrq_phys,
1360			gfp_flags | __GFP_COMP);
1361
1362		if (!rcd->rcvhdrq) {
1363			qib_dev_err(dd, "attempt to allocate %d bytes "
1364				    "for ctxt %u rcvhdrq failed\n",
1365				    amt, rcd->ctxt);
1366			goto bail;
1367		}
1368
1369		if (rcd->ctxt >= dd->first_user_ctxt) {
1370			rcd->user_event_mask = vmalloc_user(PAGE_SIZE);
1371			if (!rcd->user_event_mask)
1372				goto bail_free_hdrq;
1373		}
1374
1375		if (!(dd->flags & QIB_NODMA_RTAIL)) {
1376			rcd->rcvhdrtail_kvaddr = dma_alloc_coherent(
1377				&dd->pcidev->dev, PAGE_SIZE, &phys_hdrqtail,
1378				gfp_flags);
1379			if (!rcd->rcvhdrtail_kvaddr)
1380				goto bail_free;
1381			rcd->rcvhdrqtailaddr_phys = phys_hdrqtail;
1382		}
1383
1384		rcd->rcvhdrq_size = amt;
1385	}
1386
1387	/* clear for security and sanity on each use */
1388	memset(rcd->rcvhdrq, 0, rcd->rcvhdrq_size);
1389	if (rcd->rcvhdrtail_kvaddr)
1390		memset(rcd->rcvhdrtail_kvaddr, 0, PAGE_SIZE);
1391	return 0;
1392
1393bail_free:
1394	qib_dev_err(dd, "attempt to allocate 1 page for ctxt %u "
1395		    "rcvhdrqtailaddr failed\n", rcd->ctxt);
1396	vfree(rcd->user_event_mask);
1397	rcd->user_event_mask = NULL;
1398bail_free_hdrq:
1399	dma_free_coherent(&dd->pcidev->dev, amt, rcd->rcvhdrq,
1400			  rcd->rcvhdrq_phys);
1401	rcd->rcvhdrq = NULL;
1402bail:
1403	return -ENOMEM;
1404}
1405
1406/**
1407 * allocate eager buffers, both kernel and user contexts.
1408 * @rcd: the context we are setting up.
1409 *
1410 * Allocate the eager TID buffers and program them into hip.
1411 * They are no longer completely contiguous, we do multiple allocation
1412 * calls.  Otherwise we get the OOM code involved, by asking for too
1413 * much per call, with disastrous results on some kernels.
1414 */
1415int qib_setup_eagerbufs(struct qib_ctxtdata *rcd)
1416{
1417	struct qib_devdata *dd = rcd->dd;
1418	unsigned e, egrcnt, egrperchunk, chunk, egrsize, egroff;
1419	size_t size;
1420	gfp_t gfp_flags;
1421
1422	/*
1423	 * GFP_USER, but without GFP_FS, so buffer cache can be
1424	 * coalesced (we hope); otherwise, even at order 4,
1425	 * heavy filesystem activity makes these fail, and we can
1426	 * use compound pages.
1427	 */
1428	gfp_flags = __GFP_WAIT | __GFP_IO | __GFP_COMP;
1429
1430	egrcnt = rcd->rcvegrcnt;
1431	egroff = rcd->rcvegr_tid_base;
1432	egrsize = dd->rcvegrbufsize;
1433
1434	chunk = rcd->rcvegrbuf_chunks;
1435	egrperchunk = rcd->rcvegrbufs_perchunk;
1436	size = rcd->rcvegrbuf_size;
1437	if (!rcd->rcvegrbuf) {
1438		rcd->rcvegrbuf =
1439			kzalloc(chunk * sizeof(rcd->rcvegrbuf[0]),
1440				GFP_KERNEL);
1441		if (!rcd->rcvegrbuf)
1442			goto bail;
1443	}
1444	if (!rcd->rcvegrbuf_phys) {
1445		rcd->rcvegrbuf_phys =
1446			kmalloc(chunk * sizeof(rcd->rcvegrbuf_phys[0]),
1447				GFP_KERNEL);
1448		if (!rcd->rcvegrbuf_phys)
1449			goto bail_rcvegrbuf;
1450	}
1451	for (e = 0; e < rcd->rcvegrbuf_chunks; e++) {
1452		if (rcd->rcvegrbuf[e])
1453			continue;
1454		rcd->rcvegrbuf[e] =
1455			dma_alloc_coherent(&dd->pcidev->dev, size,
1456					   &rcd->rcvegrbuf_phys[e],
1457					   gfp_flags);
1458		if (!rcd->rcvegrbuf[e])
1459			goto bail_rcvegrbuf_phys;
1460	}
1461
1462	rcd->rcvegr_phys = rcd->rcvegrbuf_phys[0];
1463
1464	for (e = chunk = 0; chunk < rcd->rcvegrbuf_chunks; chunk++) {
1465		dma_addr_t pa = rcd->rcvegrbuf_phys[chunk];
1466		unsigned i;
1467
1468		/* clear for security and sanity on each use */
1469		memset(rcd->rcvegrbuf[chunk], 0, size);
1470
1471		for (i = 0; e < egrcnt && i < egrperchunk; e++, i++) {
1472			dd->f_put_tid(dd, e + egroff +
1473					  (u64 __iomem *)
1474					  ((char __iomem *)
1475					   dd->kregbase +
1476					   dd->rcvegrbase),
1477					  RCVHQ_RCV_TYPE_EAGER, pa);
1478			pa += egrsize;
1479		}
1480		cond_resched(); /* don't hog the cpu */
1481	}
1482
1483	return 0;
1484
1485bail_rcvegrbuf_phys:
1486	for (e = 0; e < rcd->rcvegrbuf_chunks && rcd->rcvegrbuf[e]; e++)
1487		dma_free_coherent(&dd->pcidev->dev, size,
1488				  rcd->rcvegrbuf[e], rcd->rcvegrbuf_phys[e]);
1489	kfree(rcd->rcvegrbuf_phys);
1490	rcd->rcvegrbuf_phys = NULL;
1491bail_rcvegrbuf:
1492	kfree(rcd->rcvegrbuf);
1493	rcd->rcvegrbuf = NULL;
1494bail:
1495	return -ENOMEM;
1496}
1497
1498/*
1499 * Note: Changes to this routine should be mirrored
1500 * for the diagnostics routine qib_remap_ioaddr32().
1501 * There is also related code for VL15 buffers in qib_init_7322_variables().
1502 * The teardown code that unmaps is in qib_pcie_ddcleanup()
1503 */
1504int init_chip_wc_pat(struct qib_devdata *dd, u32 vl15buflen)
1505{
1506	u64 __iomem *qib_kregbase = NULL;
1507	void __iomem *qib_piobase = NULL;
1508	u64 __iomem *qib_userbase = NULL;
1509	u64 qib_kreglen;
1510	u64 qib_pio2koffset = dd->piobufbase & 0xffffffff;
1511	u64 qib_pio4koffset = dd->piobufbase >> 32;
1512	u64 qib_pio2klen = dd->piobcnt2k * dd->palign;
1513	u64 qib_pio4klen = dd->piobcnt4k * dd->align4k;
1514	u64 qib_physaddr = dd->physaddr;
1515	u64 qib_piolen;
1516	u64 qib_userlen = 0;
1517
1518	/*
1519	 * Free the old mapping because the kernel will try to reuse the
1520	 * old mapping and not create a new mapping with the
1521	 * write combining attribute.
1522	 */
1523	iounmap(dd->kregbase);
1524	dd->kregbase = NULL;
1525
1526	/*
1527	 * Assumes chip address space looks like:
1528	 *	- kregs + sregs + cregs + uregs (in any order)
1529	 *	- piobufs (2K and 4K bufs in either order)
1530	 * or:
1531	 *	- kregs + sregs + cregs (in any order)
1532	 *	- piobufs (2K and 4K bufs in either order)
1533	 *	- uregs
1534	 */
1535	if (dd->piobcnt4k == 0) {
1536		qib_kreglen = qib_pio2koffset;
1537		qib_piolen = qib_pio2klen;
1538	} else if (qib_pio2koffset < qib_pio4koffset) {
1539		qib_kreglen = qib_pio2koffset;
1540		qib_piolen = qib_pio4koffset + qib_pio4klen - qib_kreglen;
1541	} else {
1542		qib_kreglen = qib_pio4koffset;
1543		qib_piolen = qib_pio2koffset + qib_pio2klen - qib_kreglen;
1544	}
1545	qib_piolen += vl15buflen;
1546	/* Map just the configured ports (not all hw ports) */
1547	if (dd->uregbase > qib_kreglen)
1548		qib_userlen = dd->ureg_align * dd->cfgctxts;
1549
1550	/* Sanity checks passed, now create the new mappings */
1551	qib_kregbase = ioremap_nocache(qib_physaddr, qib_kreglen);
1552	if (!qib_kregbase)
1553		goto bail;
1554
1555	qib_piobase = ioremap_wc(qib_physaddr + qib_kreglen, qib_piolen);
1556	if (!qib_piobase)
1557		goto bail_kregbase;
1558
1559	if (qib_userlen) {
1560		qib_userbase = ioremap_nocache(qib_physaddr + dd->uregbase,
1561					       qib_userlen);
1562		if (!qib_userbase)
1563			goto bail_piobase;
1564	}
1565
1566	dd->kregbase = qib_kregbase;
1567	dd->kregend = (u64 __iomem *)
1568		((char __iomem *) qib_kregbase + qib_kreglen);
1569	dd->piobase = qib_piobase;
1570	dd->pio2kbase = (void __iomem *)
1571		(((char __iomem *) dd->piobase) +
1572		 qib_pio2koffset - qib_kreglen);
1573	if (dd->piobcnt4k)
1574		dd->pio4kbase = (void __iomem *)
1575			(((char __iomem *) dd->piobase) +
1576			 qib_pio4koffset - qib_kreglen);
1577	if (qib_userlen)
1578		/* ureg will now be accessed relative to dd->userbase */
1579		dd->userbase = qib_userbase;
1580	return 0;
1581
1582bail_piobase:
1583	iounmap(qib_piobase);
1584bail_kregbase:
1585	iounmap(qib_kregbase);
1586bail:
1587	return -ENOMEM;
1588}
1589