1/*
2 * Copyright (c) 2006, 2007, 2008 QLogic Corporation. All rights reserved.
3 * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
4 *
5 * This software is available to you under a choice of one of two
6 * licenses.  You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the
9 * OpenIB.org BSD license below:
10 *
11 *     Redistribution and use in source and binary forms, with or
12 *     without modification, are permitted provided that the following
13 *     conditions are met:
14 *
15 *      - Redistributions of source code must retain the above
16 *        copyright notice, this list of conditions and the following
17 *        disclaimer.
18 *
19 *      - Redistributions in binary form must reproduce the above
20 *        copyright notice, this list of conditions and the following
21 *        disclaimer in the documentation and/or other materials
22 *        provided with the distribution.
23 *
24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31 * SOFTWARE.
32 */
33
34#include <linux/pci.h>
35#include <linux/netdevice.h>
36#include <linux/moduleparam.h>
37#include <linux/slab.h>
38#include <linux/stat.h>
39#include <linux/vmalloc.h>
40
41#include "ipath_kernel.h"
42#include "ipath_common.h"
43
44/*
45 * min buffers we want to have per port, after driver
46 */
47#define IPATH_MIN_USER_PORT_BUFCNT 7
48
49/*
50 * Number of ports we are configured to use (to allow for more pio
51 * buffers per port, etc.)  Zero means use chip value.
52 */
53static ushort ipath_cfgports;
54
55module_param_named(cfgports, ipath_cfgports, ushort, S_IRUGO);
56MODULE_PARM_DESC(cfgports, "Set max number of ports to use");
57
58/*
59 * Number of buffers reserved for driver (verbs and layered drivers.)
60 * Initialized based on number of PIO buffers if not set via module interface.
61 * The problem with this is that it's global, but we'll use different
62 * numbers for different chip types.
63 */
64static ushort ipath_kpiobufs;
65
66static int ipath_set_kpiobufs(const char *val, struct kernel_param *kp);
67
68module_param_call(kpiobufs, ipath_set_kpiobufs, param_get_ushort,
69		  &ipath_kpiobufs, S_IWUSR | S_IRUGO);
70MODULE_PARM_DESC(kpiobufs, "Set number of PIO buffers for driver");
71
72/**
73 * create_port0_egr - allocate the eager TID buffers
74 * @dd: the infinipath device
75 *
76 * This code is now quite different for user and kernel, because
77 * the kernel uses skb's, for the accelerated network performance.
78 * This is the kernel (port0) version.
79 *
80 * Allocate the eager TID buffers and program them into infinipath.
81 * We use the network layer alloc_skb() allocator to allocate the
82 * memory, and either use the buffers as is for things like verbs
83 * packets, or pass the buffers up to the ipath layered driver and
84 * thence the network layer, replacing them as we do so (see
85 * ipath_rcv_layer()).
86 */
87static int create_port0_egr(struct ipath_devdata *dd)
88{
89	unsigned e, egrcnt;
90	struct ipath_skbinfo *skbinfo;
91	int ret;
92
93	egrcnt = dd->ipath_p0_rcvegrcnt;
94
95	skbinfo = vmalloc(sizeof(*dd->ipath_port0_skbinfo) * egrcnt);
96	if (skbinfo == NULL) {
97		ipath_dev_err(dd, "allocation error for eager TID "
98			      "skb array\n");
99		ret = -ENOMEM;
100		goto bail;
101	}
102	for (e = 0; e < egrcnt; e++) {
103		/*
104		 * This is a bit tricky in that we allocate extra
105		 * space for 2 bytes of the 14 byte ethernet header.
106		 * These two bytes are passed in the ipath header so
107		 * the rest of the data is word aligned.  We allocate
108		 * 4 bytes so that the data buffer stays word aligned.
109		 * See ipath_kreceive() for more details.
110		 */
111		skbinfo[e].skb = ipath_alloc_skb(dd, GFP_KERNEL);
112		if (!skbinfo[e].skb) {
113			ipath_dev_err(dd, "SKB allocation error for "
114				      "eager TID %u\n", e);
115			while (e != 0)
116				dev_kfree_skb(skbinfo[--e].skb);
117			vfree(skbinfo);
118			ret = -ENOMEM;
119			goto bail;
120		}
121	}
122	/*
123	 * After loop above, so we can test non-NULL to see if ready
124	 * to use at receive, etc.
125	 */
126	dd->ipath_port0_skbinfo = skbinfo;
127
128	for (e = 0; e < egrcnt; e++) {
129		dd->ipath_port0_skbinfo[e].phys =
130		  ipath_map_single(dd->pcidev,
131				   dd->ipath_port0_skbinfo[e].skb->data,
132				   dd->ipath_ibmaxlen, PCI_DMA_FROMDEVICE);
133		dd->ipath_f_put_tid(dd, e + (u64 __iomem *)
134				    ((char __iomem *) dd->ipath_kregbase +
135				     dd->ipath_rcvegrbase),
136				    RCVHQ_RCV_TYPE_EAGER,
137				    dd->ipath_port0_skbinfo[e].phys);
138	}
139
140	ret = 0;
141
142bail:
143	return ret;
144}
145
146static int bringup_link(struct ipath_devdata *dd)
147{
148	u64 val, ibc;
149	int ret = 0;
150
151	/* hold IBC in reset */
152	dd->ipath_control &= ~INFINIPATH_C_LINKENABLE;
153	ipath_write_kreg(dd, dd->ipath_kregs->kr_control,
154			 dd->ipath_control);
155
156	/*
157	 * set initial max size pkt IBC will send, including ICRC; it's the
158	 * PIO buffer size in dwords, less 1; also see ipath_set_mtu()
159	 */
160	val = (dd->ipath_ibmaxlen >> 2) + 1;
161	ibc = val << dd->ibcc_mpl_shift;
162
163	/* flowcontrolwatermark is in units of KBytes */
164	ibc |= 0x5ULL << INFINIPATH_IBCC_FLOWCTRLWATERMARK_SHIFT;
165	/*
166	 * How often flowctrl sent.  More or less in usecs; balance against
167	 * watermark value, so that in theory senders always get a flow
168	 * control update in time to not let the IB link go idle.
169	 */
170	ibc |= 0x3ULL << INFINIPATH_IBCC_FLOWCTRLPERIOD_SHIFT;
171	/* max error tolerance */
172	ibc |= 0xfULL << INFINIPATH_IBCC_PHYERRTHRESHOLD_SHIFT;
173	/* use "real" buffer space for */
174	ibc |= 4ULL << INFINIPATH_IBCC_CREDITSCALE_SHIFT;
175	/* IB credit flow control. */
176	ibc |= 0xfULL << INFINIPATH_IBCC_OVERRUNTHRESHOLD_SHIFT;
177	/* initially come up waiting for TS1, without sending anything. */
178	dd->ipath_ibcctrl = ibc;
179	/*
180	 * Want to start out with both LINKCMD and LINKINITCMD in NOP
181	 * (0 and 0).  Don't put linkinitcmd in ipath_ibcctrl, want that
182	 * to stay a NOP. Flag that we are disabled, for the (unlikely)
183	 * case that some recovery path is trying to bring the link up
184	 * before we are ready.
185	 */
186	ibc |= INFINIPATH_IBCC_LINKINITCMD_DISABLE <<
187		INFINIPATH_IBCC_LINKINITCMD_SHIFT;
188	dd->ipath_flags |= IPATH_IB_LINK_DISABLED;
189	ipath_cdbg(VERBOSE, "Writing 0x%llx to ibcctrl\n",
190		   (unsigned long long) ibc);
191	ipath_write_kreg(dd, dd->ipath_kregs->kr_ibcctrl, ibc);
192
193	// be sure chip saw it
194	val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
195
196	ret = dd->ipath_f_bringup_serdes(dd);
197
198	if (ret)
199		dev_info(&dd->pcidev->dev, "Could not initialize SerDes, "
200			 "not usable\n");
201	else {
202		/* enable IBC */
203		dd->ipath_control |= INFINIPATH_C_LINKENABLE;
204		ipath_write_kreg(dd, dd->ipath_kregs->kr_control,
205				 dd->ipath_control);
206	}
207
208	return ret;
209}
210
211static struct ipath_portdata *create_portdata0(struct ipath_devdata *dd)
212{
213	struct ipath_portdata *pd = NULL;
214
215	pd = kzalloc(sizeof(*pd), GFP_KERNEL);
216	if (pd) {
217		pd->port_dd = dd;
218		pd->port_cnt = 1;
219		/* The port 0 pkey table is used by the layer interface. */
220		pd->port_pkeys[0] = IPATH_DEFAULT_P_KEY;
221		pd->port_seq_cnt = 1;
222	}
223	return pd;
224}
225
226static int init_chip_first(struct ipath_devdata *dd)
227{
228	struct ipath_portdata *pd;
229	int ret = 0;
230	u64 val;
231
232	spin_lock_init(&dd->ipath_kernel_tid_lock);
233	spin_lock_init(&dd->ipath_user_tid_lock);
234	spin_lock_init(&dd->ipath_sendctrl_lock);
235	spin_lock_init(&dd->ipath_uctxt_lock);
236	spin_lock_init(&dd->ipath_sdma_lock);
237	spin_lock_init(&dd->ipath_gpio_lock);
238	spin_lock_init(&dd->ipath_eep_st_lock);
239	spin_lock_init(&dd->ipath_sdepb_lock);
240	mutex_init(&dd->ipath_eep_lock);
241
242	/*
243	 * skip cfgports stuff because we are not allocating memory,
244	 * and we don't want problems if the portcnt changed due to
245	 * cfgports.  We do still check and report a difference, if
246	 * not same (should be impossible).
247	 */
248	dd->ipath_f_config_ports(dd, ipath_cfgports);
249	if (!ipath_cfgports)
250		dd->ipath_cfgports = dd->ipath_portcnt;
251	else if (ipath_cfgports <= dd->ipath_portcnt) {
252		dd->ipath_cfgports = ipath_cfgports;
253		ipath_dbg("Configured to use %u ports out of %u in chip\n",
254			  dd->ipath_cfgports, ipath_read_kreg32(dd,
255			  dd->ipath_kregs->kr_portcnt));
256	} else {
257		dd->ipath_cfgports = dd->ipath_portcnt;
258		ipath_dbg("Tried to configured to use %u ports; chip "
259			  "only supports %u\n", ipath_cfgports,
260			  ipath_read_kreg32(dd,
261				  dd->ipath_kregs->kr_portcnt));
262	}
263	/*
264	 * Allocate full portcnt array, rather than just cfgports, because
265	 * cleanup iterates across all possible ports.
266	 */
267	dd->ipath_pd = kzalloc(sizeof(*dd->ipath_pd) * dd->ipath_portcnt,
268			       GFP_KERNEL);
269
270	if (!dd->ipath_pd) {
271		ipath_dev_err(dd, "Unable to allocate portdata array, "
272			      "failing\n");
273		ret = -ENOMEM;
274		goto done;
275	}
276
277	pd = create_portdata0(dd);
278	if (!pd) {
279		ipath_dev_err(dd, "Unable to allocate portdata for port "
280			      "0, failing\n");
281		ret = -ENOMEM;
282		goto done;
283	}
284	dd->ipath_pd[0] = pd;
285
286	dd->ipath_rcvtidcnt =
287		ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvtidcnt);
288	dd->ipath_rcvtidbase =
289		ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvtidbase);
290	dd->ipath_rcvegrcnt =
291		ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvegrcnt);
292	dd->ipath_rcvegrbase =
293		ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvegrbase);
294	dd->ipath_palign =
295		ipath_read_kreg32(dd, dd->ipath_kregs->kr_pagealign);
296	dd->ipath_piobufbase =
297		ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendpiobufbase);
298	val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendpiosize);
299	dd->ipath_piosize2k = val & ~0U;
300	dd->ipath_piosize4k = val >> 32;
301	if (dd->ipath_piosize4k == 0 && ipath_mtu4096)
302		ipath_mtu4096 = 0; /* 4KB not supported by this chip */
303	dd->ipath_ibmtu = ipath_mtu4096 ? 4096 : 2048;
304	val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendpiobufcnt);
305	dd->ipath_piobcnt2k = val & ~0U;
306	dd->ipath_piobcnt4k = val >> 32;
307	dd->ipath_pio2kbase =
308		(u32 __iomem *) (((char __iomem *) dd->ipath_kregbase) +
309				 (dd->ipath_piobufbase & 0xffffffff));
310	if (dd->ipath_piobcnt4k) {
311		dd->ipath_pio4kbase = (u32 __iomem *)
312			(((char __iomem *) dd->ipath_kregbase) +
313			 (dd->ipath_piobufbase >> 32));
314		/*
315		 * 4K buffers take 2 pages; we use roundup just to be
316		 * paranoid; we calculate it once here, rather than on
317		 * ever buf allocate
318		 */
319		dd->ipath_4kalign = ALIGN(dd->ipath_piosize4k,
320					  dd->ipath_palign);
321		ipath_dbg("%u 2k(%x) piobufs @ %p, %u 4k(%x) @ %p "
322			  "(%x aligned)\n",
323			  dd->ipath_piobcnt2k, dd->ipath_piosize2k,
324			  dd->ipath_pio2kbase, dd->ipath_piobcnt4k,
325			  dd->ipath_piosize4k, dd->ipath_pio4kbase,
326			  dd->ipath_4kalign);
327	}
328	else ipath_dbg("%u 2k piobufs @ %p\n",
329		       dd->ipath_piobcnt2k, dd->ipath_pio2kbase);
330
331done:
332	return ret;
333}
334
335/**
336 * init_chip_reset - re-initialize after a reset, or enable
337 * @dd: the infinipath device
338 *
339 * sanity check at least some of the values after reset, and
340 * ensure no receive or transmit (explicitly, in case reset
341 * failed
342 */
343static int init_chip_reset(struct ipath_devdata *dd)
344{
345	u32 rtmp;
346	int i;
347	unsigned long flags;
348
349	/*
350	 * ensure chip does no sends or receives, tail updates, or
351	 * pioavail updates while we re-initialize
352	 */
353	dd->ipath_rcvctrl &= ~(1ULL << dd->ipath_r_tailupd_shift);
354	for (i = 0; i < dd->ipath_portcnt; i++) {
355		clear_bit(dd->ipath_r_portenable_shift + i,
356			  &dd->ipath_rcvctrl);
357		clear_bit(dd->ipath_r_intravail_shift + i,
358			  &dd->ipath_rcvctrl);
359	}
360	ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
361		dd->ipath_rcvctrl);
362
363	spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
364	dd->ipath_sendctrl = 0U; /* no sdma, etc */
365	ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, dd->ipath_sendctrl);
366	ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
367	spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
368
369	ipath_write_kreg(dd, dd->ipath_kregs->kr_control, 0ULL);
370
371	rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvtidcnt);
372	if (rtmp != dd->ipath_rcvtidcnt)
373		dev_info(&dd->pcidev->dev, "tidcnt was %u before "
374			 "reset, now %u, using original\n",
375			 dd->ipath_rcvtidcnt, rtmp);
376	rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvtidbase);
377	if (rtmp != dd->ipath_rcvtidbase)
378		dev_info(&dd->pcidev->dev, "tidbase was %u before "
379			 "reset, now %u, using original\n",
380			 dd->ipath_rcvtidbase, rtmp);
381	rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvegrcnt);
382	if (rtmp != dd->ipath_rcvegrcnt)
383		dev_info(&dd->pcidev->dev, "egrcnt was %u before "
384			 "reset, now %u, using original\n",
385			 dd->ipath_rcvegrcnt, rtmp);
386	rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvegrbase);
387	if (rtmp != dd->ipath_rcvegrbase)
388		dev_info(&dd->pcidev->dev, "egrbase was %u before "
389			 "reset, now %u, using original\n",
390			 dd->ipath_rcvegrbase, rtmp);
391
392	return 0;
393}
394
395static int init_pioavailregs(struct ipath_devdata *dd)
396{
397	int ret;
398
399	dd->ipath_pioavailregs_dma = dma_alloc_coherent(
400		&dd->pcidev->dev, PAGE_SIZE, &dd->ipath_pioavailregs_phys,
401		GFP_KERNEL);
402	if (!dd->ipath_pioavailregs_dma) {
403		ipath_dev_err(dd, "failed to allocate PIOavail reg area "
404			      "in memory\n");
405		ret = -ENOMEM;
406		goto done;
407	}
408
409	/*
410	 * we really want L2 cache aligned, but for current CPUs of
411	 * interest, they are the same.
412	 */
413	dd->ipath_statusp = (u64 *)
414		((char *)dd->ipath_pioavailregs_dma +
415		 ((2 * L1_CACHE_BYTES +
416		   dd->ipath_pioavregs * sizeof(u64)) & ~L1_CACHE_BYTES));
417	/* copy the current value now that it's really allocated */
418	*dd->ipath_statusp = dd->_ipath_status;
419	/*
420	 * setup buffer to hold freeze msg, accessible to apps,
421	 * following statusp
422	 */
423	dd->ipath_freezemsg = (char *)&dd->ipath_statusp[1];
424	/* and its length */
425	dd->ipath_freezelen = L1_CACHE_BYTES - sizeof(dd->ipath_statusp[0]);
426
427	ret = 0;
428
429done:
430	return ret;
431}
432
433/**
434 * init_shadow_tids - allocate the shadow TID array
435 * @dd: the infinipath device
436 *
437 * allocate the shadow TID array, so we can ipath_munlock previous
438 * entries.  It may make more sense to move the pageshadow to the
439 * port data structure, so we only allocate memory for ports actually
440 * in use, since we at 8k per port, now.
441 */
442static void init_shadow_tids(struct ipath_devdata *dd)
443{
444	struct page **pages;
445	dma_addr_t *addrs;
446
447	pages = vzalloc(dd->ipath_cfgports * dd->ipath_rcvtidcnt *
448			sizeof(struct page *));
449	if (!pages) {
450		ipath_dev_err(dd, "failed to allocate shadow page * "
451			      "array, no expected sends!\n");
452		dd->ipath_pageshadow = NULL;
453		return;
454	}
455
456	addrs = vmalloc(dd->ipath_cfgports * dd->ipath_rcvtidcnt *
457			sizeof(dma_addr_t));
458	if (!addrs) {
459		ipath_dev_err(dd, "failed to allocate shadow dma handle "
460			      "array, no expected sends!\n");
461		vfree(pages);
462		dd->ipath_pageshadow = NULL;
463		return;
464	}
465
466	dd->ipath_pageshadow = pages;
467	dd->ipath_physshadow = addrs;
468}
469
470static void enable_chip(struct ipath_devdata *dd, int reinit)
471{
472	u32 val;
473	u64 rcvmask;
474	unsigned long flags;
475	int i;
476
477	if (!reinit)
478		init_waitqueue_head(&ipath_state_wait);
479
480	ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
481			 dd->ipath_rcvctrl);
482
483	spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
484	/* Enable PIO send, and update of PIOavail regs to memory. */
485	dd->ipath_sendctrl = INFINIPATH_S_PIOENABLE |
486		INFINIPATH_S_PIOBUFAVAILUPD;
487
488	/*
489	 * Set the PIO avail update threshold to host memory
490	 * on chips that support it.
491	 */
492	if (dd->ipath_pioupd_thresh)
493		dd->ipath_sendctrl |= dd->ipath_pioupd_thresh
494			<< INFINIPATH_S_UPDTHRESH_SHIFT;
495	ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, dd->ipath_sendctrl);
496	ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
497	spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
498
499	/*
500	 * Enable kernel ports' receive and receive interrupt.
501	 * Other ports done as user opens and inits them.
502	 */
503	rcvmask = 1ULL;
504	dd->ipath_rcvctrl |= (rcvmask << dd->ipath_r_portenable_shift) |
505		(rcvmask << dd->ipath_r_intravail_shift);
506	if (!(dd->ipath_flags & IPATH_NODMA_RTAIL))
507		dd->ipath_rcvctrl |= (1ULL << dd->ipath_r_tailupd_shift);
508
509	ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
510			 dd->ipath_rcvctrl);
511
512	/*
513	 * now ready for use.  this should be cleared whenever we
514	 * detect a reset, or initiate one.
515	 */
516	dd->ipath_flags |= IPATH_INITTED;
517
518	/*
519	 * Init our shadow copies of head from tail values,
520	 * and write head values to match.
521	 */
522	val = ipath_read_ureg32(dd, ur_rcvegrindextail, 0);
523	ipath_write_ureg(dd, ur_rcvegrindexhead, val, 0);
524
525	/* Initialize so we interrupt on next packet received */
526	ipath_write_ureg(dd, ur_rcvhdrhead,
527			 dd->ipath_rhdrhead_intr_off |
528			 dd->ipath_pd[0]->port_head, 0);
529
530	/*
531	 * by now pioavail updates to memory should have occurred, so
532	 * copy them into our working/shadow registers; this is in
533	 * case something went wrong with abort, but mostly to get the
534	 * initial values of the generation bit correct.
535	 */
536	for (i = 0; i < dd->ipath_pioavregs; i++) {
537		__le64 pioavail;
538
539		/*
540		 * Chip Errata bug 6641; even and odd qwords>3 are swapped.
541		 */
542		if (i > 3 && (dd->ipath_flags & IPATH_SWAP_PIOBUFS))
543			pioavail = dd->ipath_pioavailregs_dma[i ^ 1];
544		else
545			pioavail = dd->ipath_pioavailregs_dma[i];
546		/*
547		 * don't need to worry about ipath_pioavailkernel here
548		 * because we will call ipath_chg_pioavailkernel() later
549		 * in initialization, to busy out buffers as needed
550		 */
551		dd->ipath_pioavailshadow[i] = le64_to_cpu(pioavail);
552	}
553	/* can get counters, stats, etc. */
554	dd->ipath_flags |= IPATH_PRESENT;
555}
556
557static int init_housekeeping(struct ipath_devdata *dd, int reinit)
558{
559	char boardn[40];
560	int ret = 0;
561
562	/*
563	 * have to clear shadow copies of registers at init that are
564	 * not otherwise set here, or all kinds of bizarre things
565	 * happen with driver on chip reset
566	 */
567	dd->ipath_rcvhdrsize = 0;
568
569	/*
570	 * Don't clear ipath_flags as 8bit mode was set before
571	 * entering this func. However, we do set the linkstate to
572	 * unknown, so we can watch for a transition.
573	 * PRESENT is set because we want register reads to work,
574	 * and the kernel infrastructure saw it in config space;
575	 * We clear it if we have failures.
576	 */
577	dd->ipath_flags |= IPATH_LINKUNK | IPATH_PRESENT;
578	dd->ipath_flags &= ~(IPATH_LINKACTIVE | IPATH_LINKARMED |
579			     IPATH_LINKDOWN | IPATH_LINKINIT);
580
581	ipath_cdbg(VERBOSE, "Try to read spc chip revision\n");
582	dd->ipath_revision =
583		ipath_read_kreg64(dd, dd->ipath_kregs->kr_revision);
584
585	/*
586	 * set up fundamental info we need to use the chip; we assume
587	 * if the revision reg and these regs are OK, we don't need to
588	 * special case the rest
589	 */
590	dd->ipath_sregbase =
591		ipath_read_kreg32(dd, dd->ipath_kregs->kr_sendregbase);
592	dd->ipath_cregbase =
593		ipath_read_kreg32(dd, dd->ipath_kregs->kr_counterregbase);
594	dd->ipath_uregbase =
595		ipath_read_kreg32(dd, dd->ipath_kregs->kr_userregbase);
596	ipath_cdbg(VERBOSE, "ipath_kregbase %p, sendbase %x usrbase %x, "
597		   "cntrbase %x\n", dd->ipath_kregbase, dd->ipath_sregbase,
598		   dd->ipath_uregbase, dd->ipath_cregbase);
599	if ((dd->ipath_revision & 0xffffffff) == 0xffffffff
600	    || (dd->ipath_sregbase & 0xffffffff) == 0xffffffff
601	    || (dd->ipath_cregbase & 0xffffffff) == 0xffffffff
602	    || (dd->ipath_uregbase & 0xffffffff) == 0xffffffff) {
603		ipath_dev_err(dd, "Register read failures from chip, "
604			      "giving up initialization\n");
605		dd->ipath_flags &= ~IPATH_PRESENT;
606		ret = -ENODEV;
607		goto done;
608	}
609
610
611	/* clear diagctrl register, in case diags were running and crashed */
612	ipath_write_kreg (dd, dd->ipath_kregs->kr_hwdiagctrl, 0);
613
614	/* clear the initial reset flag, in case first driver load */
615	ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear,
616			 INFINIPATH_E_RESET);
617
618	ipath_cdbg(VERBOSE, "Revision %llx (PCI %x)\n",
619		   (unsigned long long) dd->ipath_revision,
620		   dd->ipath_pcirev);
621
622	if (((dd->ipath_revision >> INFINIPATH_R_SOFTWARE_SHIFT) &
623	     INFINIPATH_R_SOFTWARE_MASK) != IPATH_CHIP_SWVERSION) {
624		ipath_dev_err(dd, "Driver only handles version %d, "
625			      "chip swversion is %d (%llx), failng\n",
626			      IPATH_CHIP_SWVERSION,
627			      (int)(dd->ipath_revision >>
628				    INFINIPATH_R_SOFTWARE_SHIFT) &
629			      INFINIPATH_R_SOFTWARE_MASK,
630			      (unsigned long long) dd->ipath_revision);
631		ret = -ENOSYS;
632		goto done;
633	}
634	dd->ipath_majrev = (u8) ((dd->ipath_revision >>
635				  INFINIPATH_R_CHIPREVMAJOR_SHIFT) &
636				 INFINIPATH_R_CHIPREVMAJOR_MASK);
637	dd->ipath_minrev = (u8) ((dd->ipath_revision >>
638				  INFINIPATH_R_CHIPREVMINOR_SHIFT) &
639				 INFINIPATH_R_CHIPREVMINOR_MASK);
640	dd->ipath_boardrev = (u8) ((dd->ipath_revision >>
641				    INFINIPATH_R_BOARDID_SHIFT) &
642				   INFINIPATH_R_BOARDID_MASK);
643
644	ret = dd->ipath_f_get_boardname(dd, boardn, sizeof boardn);
645
646	snprintf(dd->ipath_boardversion, sizeof(dd->ipath_boardversion),
647		 "ChipABI %u.%u, %s, InfiniPath%u %u.%u, PCI %u, "
648		 "SW Compat %u\n",
649		 IPATH_CHIP_VERS_MAJ, IPATH_CHIP_VERS_MIN, boardn,
650		 (unsigned)(dd->ipath_revision >> INFINIPATH_R_ARCH_SHIFT) &
651		 INFINIPATH_R_ARCH_MASK,
652		 dd->ipath_majrev, dd->ipath_minrev, dd->ipath_pcirev,
653		 (unsigned)(dd->ipath_revision >>
654			    INFINIPATH_R_SOFTWARE_SHIFT) &
655		 INFINIPATH_R_SOFTWARE_MASK);
656
657	ipath_dbg("%s", dd->ipath_boardversion);
658
659	if (ret)
660		goto done;
661
662	if (reinit)
663		ret = init_chip_reset(dd);
664	else
665		ret = init_chip_first(dd);
666
667done:
668	return ret;
669}
670
671static void verify_interrupt(unsigned long opaque)
672{
673	struct ipath_devdata *dd = (struct ipath_devdata *) opaque;
674
675	if (!dd)
676		return; /* being torn down */
677
678	/*
679	 * If we don't have any interrupts, let the user know and
680	 * don't bother checking again.
681	 */
682	if (dd->ipath_int_counter == 0) {
683		if (!dd->ipath_f_intr_fallback(dd))
684			dev_err(&dd->pcidev->dev, "No interrupts detected, "
685				"not usable.\n");
686		else /* re-arm the timer to see if fallback works */
687			mod_timer(&dd->ipath_intrchk_timer, jiffies + HZ/2);
688	} else
689		ipath_cdbg(VERBOSE, "%u interrupts at timer check\n",
690			dd->ipath_int_counter);
691}
692
693/**
694 * ipath_init_chip - do the actual initialization sequence on the chip
695 * @dd: the infinipath device
696 * @reinit: reinitializing, so don't allocate new memory
697 *
698 * Do the actual initialization sequence on the chip.  This is done
699 * both from the init routine called from the PCI infrastructure, and
700 * when we reset the chip, or detect that it was reset internally,
701 * or it's administratively re-enabled.
702 *
703 * Memory allocation here and in called routines is only done in
704 * the first case (reinit == 0).  We have to be careful, because even
705 * without memory allocation, we need to re-write all the chip registers
706 * TIDs, etc. after the reset or enable has completed.
707 */
708int ipath_init_chip(struct ipath_devdata *dd, int reinit)
709{
710	int ret = 0;
711	u32 kpiobufs, defkbufs;
712	u32 piobufs, uports;
713	u64 val;
714	struct ipath_portdata *pd;
715	gfp_t gfp_flags = GFP_USER | __GFP_COMP;
716
717	ret = init_housekeeping(dd, reinit);
718	if (ret)
719		goto done;
720
721	/*
722	 * We could bump this to allow for full rcvegrcnt + rcvtidcnt,
723	 * but then it no longer nicely fits power of two, and since
724	 * we now use routines that backend onto __get_free_pages, the
725	 * rest would be wasted.
726	 */
727	dd->ipath_rcvhdrcnt = max(dd->ipath_p0_rcvegrcnt, dd->ipath_rcvegrcnt);
728	ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvhdrcnt,
729			 dd->ipath_rcvhdrcnt);
730
731	/*
732	 * Set up the shadow copies of the piobufavail registers,
733	 * which we compare against the chip registers for now, and
734	 * the in memory DMA'ed copies of the registers.  This has to
735	 * be done early, before we calculate lastport, etc.
736	 */
737	piobufs = dd->ipath_piobcnt2k + dd->ipath_piobcnt4k;
738	/*
739	 * calc number of pioavail registers, and save it; we have 2
740	 * bits per buffer.
741	 */
742	dd->ipath_pioavregs = ALIGN(piobufs, sizeof(u64) * BITS_PER_BYTE / 2)
743		/ (sizeof(u64) * BITS_PER_BYTE / 2);
744	uports = dd->ipath_cfgports ? dd->ipath_cfgports - 1 : 0;
745	if (piobufs > 144)
746		defkbufs = 32 + dd->ipath_pioreserved;
747	else
748		defkbufs = 16 + dd->ipath_pioreserved;
749
750	if (ipath_kpiobufs && (ipath_kpiobufs +
751		(uports * IPATH_MIN_USER_PORT_BUFCNT)) > piobufs) {
752		int i = (int) piobufs -
753			(int) (uports * IPATH_MIN_USER_PORT_BUFCNT);
754		if (i < 1)
755			i = 1;
756		dev_info(&dd->pcidev->dev, "Allocating %d PIO bufs of "
757			 "%d for kernel leaves too few for %d user ports "
758			 "(%d each); using %u\n", ipath_kpiobufs,
759			 piobufs, uports, IPATH_MIN_USER_PORT_BUFCNT, i);
760		/*
761		 * shouldn't change ipath_kpiobufs, because could be
762		 * different for different devices...
763		 */
764		kpiobufs = i;
765	} else if (ipath_kpiobufs)
766		kpiobufs = ipath_kpiobufs;
767	else
768		kpiobufs = defkbufs;
769	dd->ipath_lastport_piobuf = piobufs - kpiobufs;
770	dd->ipath_pbufsport =
771		uports ? dd->ipath_lastport_piobuf / uports : 0;
772	/* if not an even divisor, some user ports get extra buffers */
773	dd->ipath_ports_extrabuf = dd->ipath_lastport_piobuf -
774		(dd->ipath_pbufsport * uports);
775	if (dd->ipath_ports_extrabuf)
776		ipath_dbg("%u pbufs/port leaves some unused, add 1 buffer to "
777			"ports <= %u\n", dd->ipath_pbufsport,
778			dd->ipath_ports_extrabuf);
779	dd->ipath_lastpioindex = 0;
780	dd->ipath_lastpioindexl = dd->ipath_piobcnt2k;
781	/* ipath_pioavailshadow initialized earlier */
782	ipath_cdbg(VERBOSE, "%d PIO bufs for kernel out of %d total %u "
783		   "each for %u user ports\n", kpiobufs,
784		   piobufs, dd->ipath_pbufsport, uports);
785	ret = dd->ipath_f_early_init(dd);
786	if (ret) {
787		ipath_dev_err(dd, "Early initialization failure\n");
788		goto done;
789	}
790
791	/*
792	 * Early_init sets rcvhdrentsize and rcvhdrsize, so this must be
793	 * done after early_init.
794	 */
795	dd->ipath_hdrqlast =
796		dd->ipath_rcvhdrentsize * (dd->ipath_rcvhdrcnt - 1);
797	ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvhdrentsize,
798			 dd->ipath_rcvhdrentsize);
799	ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvhdrsize,
800			 dd->ipath_rcvhdrsize);
801
802	if (!reinit) {
803		ret = init_pioavailregs(dd);
804		init_shadow_tids(dd);
805		if (ret)
806			goto done;
807	}
808
809	ipath_write_kreg(dd, dd->ipath_kregs->kr_sendpioavailaddr,
810			 dd->ipath_pioavailregs_phys);
811
812	/*
813	 * this is to detect s/w errors, which the h/w works around by
814	 * ignoring the low 6 bits of address, if it wasn't aligned.
815	 */
816	val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendpioavailaddr);
817	if (val != dd->ipath_pioavailregs_phys) {
818		ipath_dev_err(dd, "Catastrophic software error, "
819			      "SendPIOAvailAddr written as %lx, "
820			      "read back as %llx\n",
821			      (unsigned long) dd->ipath_pioavailregs_phys,
822			      (unsigned long long) val);
823		ret = -EINVAL;
824		goto done;
825	}
826
827	ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvbthqp, IPATH_KD_QP);
828
829	/*
830	 * make sure we are not in freeze, and PIO send enabled, so
831	 * writes to pbc happen
832	 */
833	ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrmask, 0ULL);
834	ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear,
835			 ~0ULL&~INFINIPATH_HWE_MEMBISTFAILED);
836	ipath_write_kreg(dd, dd->ipath_kregs->kr_control, 0ULL);
837
838	/*
839	 * before error clears, since we expect serdes pll errors during
840	 * this, the first time after reset
841	 */
842	if (bringup_link(dd)) {
843		dev_info(&dd->pcidev->dev, "Failed to bringup IB link\n");
844		ret = -ENETDOWN;
845		goto done;
846	}
847
848	/*
849	 * clear any "expected" hwerrs from reset and/or initialization
850	 * clear any that aren't enabled (at least this once), and then
851	 * set the enable mask
852	 */
853	dd->ipath_f_init_hwerrors(dd);
854	ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear,
855			 ~0ULL&~INFINIPATH_HWE_MEMBISTFAILED);
856	ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrmask,
857			 dd->ipath_hwerrmask);
858
859	/* clear all */
860	ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear, -1LL);
861	/* enable errors that are masked, at least this first time. */
862	ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
863			 ~dd->ipath_maskederrs);
864	dd->ipath_maskederrs = 0; /* don't re-enable ignored in timer */
865	dd->ipath_errormask =
866		ipath_read_kreg64(dd, dd->ipath_kregs->kr_errormask);
867	/* clear any interrupts up to this point (ints still not enabled) */
868	ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, -1LL);
869
870	dd->ipath_f_tidtemplate(dd);
871
872	/*
873	 * Set up the port 0 (kernel) rcvhdr q and egr TIDs.  If doing
874	 * re-init, the simplest way to handle this is to free
875	 * existing, and re-allocate.
876	 * Need to re-create rest of port 0 portdata as well.
877	 */
878	pd = dd->ipath_pd[0];
879	if (reinit) {
880		struct ipath_portdata *npd;
881
882		/*
883		 * Alloc and init new ipath_portdata for port0,
884		 * Then free old pd. Could lead to fragmentation, but also
885		 * makes later support for hot-swap easier.
886		 */
887		npd = create_portdata0(dd);
888		if (npd) {
889			ipath_free_pddata(dd, pd);
890			dd->ipath_pd[0] = npd;
891			pd = npd;
892		} else {
893			ipath_dev_err(dd, "Unable to allocate portdata"
894				      " for port 0, failing\n");
895			ret = -ENOMEM;
896			goto done;
897		}
898	}
899	ret = ipath_create_rcvhdrq(dd, pd);
900	if (!ret)
901		ret = create_port0_egr(dd);
902	if (ret) {
903		ipath_dev_err(dd, "failed to allocate kernel port's "
904			      "rcvhdrq and/or egr bufs\n");
905		goto done;
906	}
907	else
908		enable_chip(dd, reinit);
909
910	/* after enable_chip, so pioavailshadow setup */
911	ipath_chg_pioavailkernel(dd, 0, piobufs, 1);
912
913	/*
914	 * Cancel any possible active sends from early driver load.
915	 * Follows early_init because some chips have to initialize
916	 * PIO buffers in early_init to avoid false parity errors.
917	 * After enable and ipath_chg_pioavailkernel so we can safely
918	 * enable pioavail updates and PIOENABLE; packets are now
919	 * ready to go out.
920	 */
921	ipath_cancel_sends(dd, 1);
922
923	if (!reinit) {
924		/*
925		 * Used when we close a port, for DMA already in flight
926		 * at close.
927		 */
928		dd->ipath_dummy_hdrq = dma_alloc_coherent(
929			&dd->pcidev->dev, dd->ipath_pd[0]->port_rcvhdrq_size,
930			&dd->ipath_dummy_hdrq_phys,
931			gfp_flags);
932		if (!dd->ipath_dummy_hdrq) {
933			dev_info(&dd->pcidev->dev,
934				"Couldn't allocate 0x%lx bytes for dummy hdrq\n",
935				dd->ipath_pd[0]->port_rcvhdrq_size);
936			/* fallback to just 0'ing */
937			dd->ipath_dummy_hdrq_phys = 0UL;
938		}
939	}
940
941	/*
942	 * cause retrigger of pending interrupts ignored during init,
943	 * even if we had errors
944	 */
945	ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, 0ULL);
946
947	if (!dd->ipath_stats_timer_active) {
948		/*
949		 * first init, or after an admin disable/enable
950		 * set up stats retrieval timer, even if we had errors
951		 * in last portion of setup
952		 */
953		init_timer(&dd->ipath_stats_timer);
954		dd->ipath_stats_timer.function = ipath_get_faststats;
955		dd->ipath_stats_timer.data = (unsigned long) dd;
956		/* every 5 seconds; */
957		dd->ipath_stats_timer.expires = jiffies + 5 * HZ;
958		/* takes ~16 seconds to overflow at full IB 4x bandwdith */
959		add_timer(&dd->ipath_stats_timer);
960		dd->ipath_stats_timer_active = 1;
961	}
962
963	/* Set up SendDMA if chip supports it */
964	if (dd->ipath_flags & IPATH_HAS_SEND_DMA)
965		ret = setup_sdma(dd);
966
967	/* Set up HoL state */
968	init_timer(&dd->ipath_hol_timer);
969	dd->ipath_hol_timer.function = ipath_hol_event;
970	dd->ipath_hol_timer.data = (unsigned long)dd;
971	dd->ipath_hol_state = IPATH_HOL_UP;
972
973done:
974	if (!ret) {
975		*dd->ipath_statusp |= IPATH_STATUS_CHIP_PRESENT;
976		if (!dd->ipath_f_intrsetup(dd)) {
977			/* now we can enable all interrupts from the chip */
978			ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask,
979					 -1LL);
980			/* force re-interrupt of any pending interrupts. */
981			ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear,
982					 0ULL);
983			/* chip is usable; mark it as initialized */
984			*dd->ipath_statusp |= IPATH_STATUS_INITTED;
985
986			/*
987			 * setup to verify we get an interrupt, and fallback
988			 * to an alternate if necessary and possible
989			 */
990			if (!reinit) {
991				init_timer(&dd->ipath_intrchk_timer);
992				dd->ipath_intrchk_timer.function =
993					verify_interrupt;
994				dd->ipath_intrchk_timer.data =
995					(unsigned long) dd;
996			}
997			dd->ipath_intrchk_timer.expires = jiffies + HZ/2;
998			add_timer(&dd->ipath_intrchk_timer);
999		} else
1000			ipath_dev_err(dd, "No interrupts enabled, couldn't "
1001				      "setup interrupt address\n");
1002
1003		if (dd->ipath_cfgports > ipath_stats.sps_nports)
1004			/*
1005			 * sps_nports is a global, so, we set it to
1006			 * the highest number of ports of any of the
1007			 * chips we find; we never decrement it, at
1008			 * least for now.  Since this might have changed
1009			 * over disable/enable or prior to reset, always
1010			 * do the check and potentially adjust.
1011			 */
1012			ipath_stats.sps_nports = dd->ipath_cfgports;
1013	} else
1014		ipath_dbg("Failed (%d) to initialize chip\n", ret);
1015
1016	/* if ret is non-zero, we probably should do some cleanup
1017	   here... */
1018	return ret;
1019}
1020
1021static int ipath_set_kpiobufs(const char *str, struct kernel_param *kp)
1022{
1023	struct ipath_devdata *dd;
1024	unsigned long flags;
1025	unsigned short val;
1026	int ret;
1027
1028	ret = ipath_parse_ushort(str, &val);
1029
1030	spin_lock_irqsave(&ipath_devs_lock, flags);
1031
1032	if (ret < 0)
1033		goto bail;
1034
1035	if (val == 0) {
1036		ret = -EINVAL;
1037		goto bail;
1038	}
1039
1040	list_for_each_entry(dd, &ipath_dev_list, ipath_list) {
1041		if (dd->ipath_kregbase)
1042			continue;
1043		if (val > (dd->ipath_piobcnt2k + dd->ipath_piobcnt4k -
1044			   (dd->ipath_cfgports *
1045			    IPATH_MIN_USER_PORT_BUFCNT)))
1046		{
1047			ipath_dev_err(
1048				dd,
1049				"Allocating %d PIO bufs for kernel leaves "
1050				"too few for %d user ports (%d each)\n",
1051				val, dd->ipath_cfgports - 1,
1052				IPATH_MIN_USER_PORT_BUFCNT);
1053			ret = -EINVAL;
1054			goto bail;
1055		}
1056		dd->ipath_lastport_piobuf =
1057			dd->ipath_piobcnt2k + dd->ipath_piobcnt4k - val;
1058	}
1059
1060	ipath_kpiobufs = val;
1061	ret = 0;
1062bail:
1063	spin_unlock_irqrestore(&ipath_devs_lock, flags);
1064
1065	return ret;
1066}
1067