IB/ipath: Log "active" time and some errors to EEPROM
[safe/jmp/linux-2.6] / drivers / infiniband / hw / ipath / ipath_init_chip.c
1 /*
2  * Copyright (c) 2006 QLogic, Inc. 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/vmalloc.h>
37
38 #include "ipath_kernel.h"
39 #include "ipath_common.h"
40
41 /*
42  * min buffers we want to have per port, after driver
43  */
44 #define IPATH_MIN_USER_PORT_BUFCNT 8
45
46 /*
47  * Number of ports we are configured to use (to allow for more pio
48  * buffers per port, etc.)  Zero means use chip value.
49  */
50 static ushort ipath_cfgports;
51
52 module_param_named(cfgports, ipath_cfgports, ushort, S_IRUGO);
53 MODULE_PARM_DESC(cfgports, "Set max number of ports to use");
54
55 /*
56  * Number of buffers reserved for driver (verbs and layered drivers.)
57  * Reserved at end of buffer list.   Initialized based on
58  * number of PIO buffers if not set via module interface.
59  * The problem with this is that it's global, but we'll use different
60  * numbers for different chip types.  So the default value is not
61  * very useful.  I've redefined it for the 1.3 release so that it's
62  * zero unless set by the user to something else, in which case we
63  * try to respect it.
64  */
65 static ushort ipath_kpiobufs;
66
67 static int ipath_set_kpiobufs(const char *val, struct kernel_param *kp);
68
69 module_param_call(kpiobufs, ipath_set_kpiobufs, param_get_ushort,
70                   &ipath_kpiobufs, S_IWUSR | S_IRUGO);
71 MODULE_PARM_DESC(kpiobufs, "Set number of PIO buffers for driver");
72
73 /**
74  * create_port0_egr - allocate the eager TID buffers
75  * @dd: the infinipath device
76  *
77  * This code is now quite different for user and kernel, because
78  * the kernel uses skb's, for the accelerated network performance.
79  * This is the kernel (port0) version.
80  *
81  * Allocate the eager TID buffers and program them into infinipath.
82  * We use the network layer alloc_skb() allocator to allocate the
83  * memory, and either use the buffers as is for things like verbs
84  * packets, or pass the buffers up to the ipath layered driver and
85  * thence the network layer, replacing them as we do so (see
86  * ipath_rcv_layer()).
87  */
88 static int create_port0_egr(struct ipath_devdata *dd)
89 {
90         unsigned e, egrcnt;
91         struct ipath_skbinfo *skbinfo;
92         int ret;
93
94         egrcnt = dd->ipath_rcvegrcnt;
95
96         skbinfo = vmalloc(sizeof(*dd->ipath_port0_skbinfo) * egrcnt);
97         if (skbinfo == NULL) {
98                 ipath_dev_err(dd, "allocation error for eager TID "
99                               "skb array\n");
100                 ret = -ENOMEM;
101                 goto bail;
102         }
103         for (e = 0; e < egrcnt; e++) {
104                 /*
105                  * This is a bit tricky in that we allocate extra
106                  * space for 2 bytes of the 14 byte ethernet header.
107                  * These two bytes are passed in the ipath header so
108                  * the rest of the data is word aligned.  We allocate
109                  * 4 bytes so that the data buffer stays word aligned.
110                  * See ipath_kreceive() for more details.
111                  */
112                 skbinfo[e].skb = ipath_alloc_skb(dd, GFP_KERNEL);
113                 if (!skbinfo[e].skb) {
114                         ipath_dev_err(dd, "SKB allocation error for "
115                                       "eager TID %u\n", e);
116                         while (e != 0)
117                                 dev_kfree_skb(skbinfo[--e].skb);
118                         vfree(skbinfo);
119                         ret = -ENOMEM;
120                         goto bail;
121                 }
122         }
123         /*
124          * After loop above, so we can test non-NULL to see if ready
125          * to use at receive, etc.
126          */
127         dd->ipath_port0_skbinfo = skbinfo;
128
129         for (e = 0; e < egrcnt; e++) {
130                 dd->ipath_port0_skbinfo[e].phys =
131                   ipath_map_single(dd->pcidev,
132                                    dd->ipath_port0_skbinfo[e].skb->data,
133                                    dd->ipath_ibmaxlen, PCI_DMA_FROMDEVICE);
134                 dd->ipath_f_put_tid(dd, e + (u64 __iomem *)
135                                     ((char __iomem *) dd->ipath_kregbase +
136                                      dd->ipath_rcvegrbase), 0,
137                                     dd->ipath_port0_skbinfo[e].phys);
138         }
139
140         ret = 0;
141
142 bail:
143         return ret;
144 }
145
146 static 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          * Note that prior to try 14 or 15 of IB, the credit scaling
158          * wasn't working, because it was swapped for writes with the
159          * 1 bit default linkstate field
160          */
161
162         /* ignore pbc and align word */
163         val = dd->ipath_piosize2k - 2 * sizeof(u32);
164         /*
165          * for ICRC, which we only send in diag test pkt mode, and we
166          * don't need to worry about that for mtu
167          */
168         val += 1;
169         /*
170          * Set the IBC maxpktlength to the size of our pio buffers the
171          * maxpktlength is in words.  This is *not* the IB data MTU.
172          */
173         ibc = (val / sizeof(u32)) << INFINIPATH_IBCC_MAXPKTLEN_SHIFT;
174         /* in KB */
175         ibc |= 0x5ULL << INFINIPATH_IBCC_FLOWCTRLWATERMARK_SHIFT;
176         /*
177          * How often flowctrl sent.  More or less in usecs; balance against
178          * watermark value, so that in theory senders always get a flow
179          * control update in time to not let the IB link go idle.
180          */
181         ibc |= 0x3ULL << INFINIPATH_IBCC_FLOWCTRLPERIOD_SHIFT;
182         /* max error tolerance */
183         ibc |= 0xfULL << INFINIPATH_IBCC_PHYERRTHRESHOLD_SHIFT;
184         /* use "real" buffer space for */
185         ibc |= 4ULL << INFINIPATH_IBCC_CREDITSCALE_SHIFT;
186         /* IB credit flow control. */
187         ibc |= 0xfULL << INFINIPATH_IBCC_OVERRUNTHRESHOLD_SHIFT;
188         /* initially come up waiting for TS1, without sending anything. */
189         dd->ipath_ibcctrl = ibc;
190         /*
191          * Want to start out with both LINKCMD and LINKINITCMD in NOP
192          * (0 and 0).  Don't put linkinitcmd in ipath_ibcctrl, want that
193          * to stay a NOP
194          */
195         ibc |= INFINIPATH_IBCC_LINKINITCMD_DISABLE <<
196                 INFINIPATH_IBCC_LINKINITCMD_SHIFT;
197         ipath_cdbg(VERBOSE, "Writing 0x%llx to ibcctrl\n",
198                    (unsigned long long) ibc);
199         ipath_write_kreg(dd, dd->ipath_kregs->kr_ibcctrl, ibc);
200
201         // be sure chip saw it
202         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
203
204         ret = dd->ipath_f_bringup_serdes(dd);
205
206         if (ret)
207                 dev_info(&dd->pcidev->dev, "Could not initialize SerDes, "
208                          "not usable\n");
209         else {
210                 /* enable IBC */
211                 dd->ipath_control |= INFINIPATH_C_LINKENABLE;
212                 ipath_write_kreg(dd, dd->ipath_kregs->kr_control,
213                                  dd->ipath_control);
214         }
215
216         return ret;
217 }
218
219 static struct ipath_portdata *create_portdata0(struct ipath_devdata *dd)
220 {
221         struct ipath_portdata *pd = NULL;
222
223         pd = kzalloc(sizeof(*pd), GFP_KERNEL);
224         if (pd) {
225                 pd->port_dd = dd;
226                 pd->port_cnt = 1;
227                 /* The port 0 pkey table is used by the layer interface. */
228                 pd->port_pkeys[0] = IPATH_DEFAULT_P_KEY;
229         }
230         return pd;
231 }
232
233 static int init_chip_first(struct ipath_devdata *dd,
234                            struct ipath_portdata **pdp)
235 {
236         struct ipath_portdata *pd = NULL;
237         int ret = 0;
238         u64 val;
239
240         /*
241          * skip cfgports stuff because we are not allocating memory,
242          * and we don't want problems if the portcnt changed due to
243          * cfgports.  We do still check and report a difference, if
244          * not same (should be impossible).
245          */
246         dd->ipath_portcnt =
247                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_portcnt);
248         if (!ipath_cfgports)
249                 dd->ipath_cfgports = dd->ipath_portcnt;
250         else if (ipath_cfgports <= dd->ipath_portcnt) {
251                 dd->ipath_cfgports = ipath_cfgports;
252                 ipath_dbg("Configured to use %u ports out of %u in chip\n",
253                           dd->ipath_cfgports, dd->ipath_portcnt);
254         } else {
255                 dd->ipath_cfgports = dd->ipath_portcnt;
256                 ipath_dbg("Tried to configured to use %u ports; chip "
257                           "only supports %u\n", ipath_cfgports,
258                           dd->ipath_portcnt);
259         }
260         /*
261          * Allocate full portcnt array, rather than just cfgports, because
262          * cleanup iterates across all possible ports.
263          */
264         dd->ipath_pd = kzalloc(sizeof(*dd->ipath_pd) * dd->ipath_portcnt,
265                                GFP_KERNEL);
266
267         if (!dd->ipath_pd) {
268                 ipath_dev_err(dd, "Unable to allocate portdata array, "
269                               "failing\n");
270                 ret = -ENOMEM;
271                 goto done;
272         }
273
274         dd->ipath_lastegrheads = kzalloc(sizeof(*dd->ipath_lastegrheads)
275                                          * dd->ipath_cfgports,
276                                          GFP_KERNEL);
277         dd->ipath_lastrcvhdrqtails =
278                 kzalloc(sizeof(*dd->ipath_lastrcvhdrqtails)
279                         * dd->ipath_cfgports, GFP_KERNEL);
280
281         if (!dd->ipath_lastegrheads || !dd->ipath_lastrcvhdrqtails) {
282                 ipath_dev_err(dd, "Unable to allocate head arrays, "
283                               "failing\n");
284                 ret = -ENOMEM;
285                 goto done;
286         }
287
288         pd = create_portdata0(dd);
289
290         if (!pd) {
291                 ipath_dev_err(dd, "Unable to allocate portdata for port "
292                               "0, failing\n");
293                 ret = -ENOMEM;
294                 goto done;
295         }
296         dd->ipath_pd[0] = pd;
297
298         dd->ipath_rcvtidcnt =
299                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvtidcnt);
300         dd->ipath_rcvtidbase =
301                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvtidbase);
302         dd->ipath_rcvegrcnt =
303                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvegrcnt);
304         dd->ipath_rcvegrbase =
305                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvegrbase);
306         dd->ipath_palign =
307                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_pagealign);
308         dd->ipath_piobufbase =
309                 ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendpiobufbase);
310         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendpiosize);
311         dd->ipath_piosize2k = val & ~0U;
312         dd->ipath_piosize4k = val >> 32;
313         dd->ipath_ibmtu = 4096; /* default to largest legal MTU */
314         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendpiobufcnt);
315         dd->ipath_piobcnt2k = val & ~0U;
316         dd->ipath_piobcnt4k = val >> 32;
317         dd->ipath_pio2kbase =
318                 (u32 __iomem *) (((char __iomem *) dd->ipath_kregbase) +
319                                  (dd->ipath_piobufbase & 0xffffffff));
320         if (dd->ipath_piobcnt4k) {
321                 dd->ipath_pio4kbase = (u32 __iomem *)
322                         (((char __iomem *) dd->ipath_kregbase) +
323                          (dd->ipath_piobufbase >> 32));
324                 /*
325                  * 4K buffers take 2 pages; we use roundup just to be
326                  * paranoid; we calculate it once here, rather than on
327                  * ever buf allocate
328                  */
329                 dd->ipath_4kalign = ALIGN(dd->ipath_piosize4k,
330                                           dd->ipath_palign);
331                 ipath_dbg("%u 2k(%x) piobufs @ %p, %u 4k(%x) @ %p "
332                           "(%x aligned)\n",
333                           dd->ipath_piobcnt2k, dd->ipath_piosize2k,
334                           dd->ipath_pio2kbase, dd->ipath_piobcnt4k,
335                           dd->ipath_piosize4k, dd->ipath_pio4kbase,
336                           dd->ipath_4kalign);
337         }
338         else ipath_dbg("%u 2k piobufs @ %p\n",
339                        dd->ipath_piobcnt2k, dd->ipath_pio2kbase);
340
341         spin_lock_init(&dd->ipath_tid_lock);
342
343         spin_lock_init(&dd->ipath_gpio_lock);
344         spin_lock_init(&dd->ipath_eep_st_lock);
345         sema_init(&dd->ipath_eep_sem, 1);
346
347 done:
348         *pdp = pd;
349         return ret;
350 }
351
352 /**
353  * init_chip_reset - re-initialize after a reset, or enable
354  * @dd: the infinipath device
355  * @pdp: output for port data
356  *
357  * sanity check at least some of the values after reset, and
358  * ensure no receive or transmit (explictly, in case reset
359  * failed
360  */
361 static int init_chip_reset(struct ipath_devdata *dd,
362                            struct ipath_portdata **pdp)
363 {
364         u32 rtmp;
365
366         *pdp = dd->ipath_pd[0];
367         /* ensure chip does no sends or receives while we re-initialize */
368         dd->ipath_control = dd->ipath_sendctrl = dd->ipath_rcvctrl = 0U;
369         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl, 0);
370         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, 0);
371         ipath_write_kreg(dd, dd->ipath_kregs->kr_control, 0);
372
373         rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_portcnt);
374         if (dd->ipath_portcnt != rtmp)
375                 dev_info(&dd->pcidev->dev, "portcnt was %u before "
376                          "reset, now %u, using original\n",
377                          dd->ipath_portcnt, rtmp);
378         rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvtidcnt);
379         if (rtmp != dd->ipath_rcvtidcnt)
380                 dev_info(&dd->pcidev->dev, "tidcnt was %u before "
381                          "reset, now %u, using original\n",
382                          dd->ipath_rcvtidcnt, rtmp);
383         rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvtidbase);
384         if (rtmp != dd->ipath_rcvtidbase)
385                 dev_info(&dd->pcidev->dev, "tidbase was %u before "
386                          "reset, now %u, using original\n",
387                          dd->ipath_rcvtidbase, rtmp);
388         rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvegrcnt);
389         if (rtmp != dd->ipath_rcvegrcnt)
390                 dev_info(&dd->pcidev->dev, "egrcnt was %u before "
391                          "reset, now %u, using original\n",
392                          dd->ipath_rcvegrcnt, rtmp);
393         rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvegrbase);
394         if (rtmp != dd->ipath_rcvegrbase)
395                 dev_info(&dd->pcidev->dev, "egrbase was %u before "
396                          "reset, now %u, using original\n",
397                          dd->ipath_rcvegrbase, rtmp);
398
399         return 0;
400 }
401
402 static int init_pioavailregs(struct ipath_devdata *dd)
403 {
404         int ret;
405
406         dd->ipath_pioavailregs_dma = dma_alloc_coherent(
407                 &dd->pcidev->dev, PAGE_SIZE, &dd->ipath_pioavailregs_phys,
408                 GFP_KERNEL);
409         if (!dd->ipath_pioavailregs_dma) {
410                 ipath_dev_err(dd, "failed to allocate PIOavail reg area "
411                               "in memory\n");
412                 ret = -ENOMEM;
413                 goto done;
414         }
415
416         /*
417          * we really want L2 cache aligned, but for current CPUs of
418          * interest, they are the same.
419          */
420         dd->ipath_statusp = (u64 *)
421                 ((char *)dd->ipath_pioavailregs_dma +
422                  ((2 * L1_CACHE_BYTES +
423                    dd->ipath_pioavregs * sizeof(u64)) & ~L1_CACHE_BYTES));
424         /* copy the current value now that it's really allocated */
425         *dd->ipath_statusp = dd->_ipath_status;
426         /*
427          * setup buffer to hold freeze msg, accessible to apps,
428          * following statusp
429          */
430         dd->ipath_freezemsg = (char *)&dd->ipath_statusp[1];
431         /* and its length */
432         dd->ipath_freezelen = L1_CACHE_BYTES - sizeof(dd->ipath_statusp[0]);
433
434         ret = 0;
435
436 done:
437         return ret;
438 }
439
440 /**
441  * init_shadow_tids - allocate the shadow TID array
442  * @dd: the infinipath device
443  *
444  * allocate the shadow TID array, so we can ipath_munlock previous
445  * entries.  It may make more sense to move the pageshadow to the
446  * port data structure, so we only allocate memory for ports actually
447  * in use, since we at 8k per port, now.
448  */
449 static void init_shadow_tids(struct ipath_devdata *dd)
450 {
451         struct page **pages;
452         dma_addr_t *addrs;
453
454         pages = vmalloc(dd->ipath_cfgports * dd->ipath_rcvtidcnt *
455                         sizeof(struct page *));
456         if (!pages) {
457                 ipath_dev_err(dd, "failed to allocate shadow page * "
458                               "array, no expected sends!\n");
459                 dd->ipath_pageshadow = NULL;
460                 return;
461         }
462
463         addrs = vmalloc(dd->ipath_cfgports * dd->ipath_rcvtidcnt *
464                         sizeof(dma_addr_t));
465         if (!addrs) {
466                 ipath_dev_err(dd, "failed to allocate shadow dma handle "
467                               "array, no expected sends!\n");
468                 vfree(dd->ipath_pageshadow);
469                 dd->ipath_pageshadow = NULL;
470                 return;
471         }
472
473         memset(pages, 0, dd->ipath_cfgports * dd->ipath_rcvtidcnt *
474                sizeof(struct page *));
475
476         dd->ipath_pageshadow = pages;
477         dd->ipath_physshadow = addrs;
478 }
479
480 static void enable_chip(struct ipath_devdata *dd,
481                         struct ipath_portdata *pd, int reinit)
482 {
483         u32 val;
484         int i;
485
486         if (!reinit)
487                 init_waitqueue_head(&ipath_state_wait);
488
489         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
490                          dd->ipath_rcvctrl);
491
492         /* Enable PIO send, and update of PIOavail regs to memory. */
493         dd->ipath_sendctrl = INFINIPATH_S_PIOENABLE |
494                 INFINIPATH_S_PIOBUFAVAILUPD;
495         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
496                          dd->ipath_sendctrl);
497
498         /*
499          * enable port 0 receive, and receive interrupt.  other ports
500          * done as user opens and inits them.
501          */
502         dd->ipath_rcvctrl = INFINIPATH_R_TAILUPD |
503                 (1ULL << INFINIPATH_R_PORTENABLE_SHIFT) |
504                 (1ULL << INFINIPATH_R_INTRAVAIL_SHIFT);
505         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
506                          dd->ipath_rcvctrl);
507
508         /*
509          * now ready for use.  this should be cleared whenever we
510          * detect a reset, or initiate one.
511          */
512         dd->ipath_flags |= IPATH_INITTED;
513
514         /*
515          * init our shadow copies of head from tail values, and write
516          * head values to match.
517          */
518         val = ipath_read_ureg32(dd, ur_rcvegrindextail, 0);
519         (void)ipath_write_ureg(dd, ur_rcvegrindexhead, val, 0);
520         dd->ipath_port0head = ipath_read_ureg32(dd, ur_rcvhdrtail, 0);
521
522         /* Initialize so we interrupt on next packet received */
523         (void)ipath_write_ureg(dd, ur_rcvhdrhead,
524                                dd->ipath_rhdrhead_intr_off |
525                                dd->ipath_port0head, 0);
526
527         /*
528          * by now pioavail updates to memory should have occurred, so
529          * copy them into our working/shadow registers; this is in
530          * case something went wrong with abort, but mostly to get the
531          * initial values of the generation bit correct.
532          */
533         for (i = 0; i < dd->ipath_pioavregs; i++) {
534                 __le64 val;
535
536                 /*
537                  * Chip Errata bug 6641; even and odd qwords>3 are swapped.
538                  */
539                 if (i > 3) {
540                         if (i & 1)
541                                 val = dd->ipath_pioavailregs_dma[i - 1];
542                         else
543                                 val = dd->ipath_pioavailregs_dma[i + 1];
544                 }
545                 else
546                         val = dd->ipath_pioavailregs_dma[i];
547                 dd->ipath_pioavailshadow[i] = le64_to_cpu(val);
548         }
549         /* can get counters, stats, etc. */
550         dd->ipath_flags |= IPATH_PRESENT;
551 }
552
553 static int init_housekeeping(struct ipath_devdata *dd,
554                              struct ipath_portdata **pdp, int reinit)
555 {
556         char boardn[32];
557         int ret = 0;
558
559         /*
560          * have to clear shadow copies of registers at init that are
561          * not otherwise set here, or all kinds of bizarre things
562          * happen with driver on chip reset
563          */
564         dd->ipath_rcvhdrsize = 0;
565
566         /*
567          * Don't clear ipath_flags as 8bit mode was set before
568          * entering this func. However, we do set the linkstate to
569          * unknown, so we can watch for a transition.
570          * PRESENT is set because we want register reads to work,
571          * and the kernel infrastructure saw it in config space;
572          * We clear it if we have failures.
573          */
574         dd->ipath_flags |= IPATH_LINKUNK | IPATH_PRESENT;
575         dd->ipath_flags &= ~(IPATH_LINKACTIVE | IPATH_LINKARMED |
576                              IPATH_LINKDOWN | IPATH_LINKINIT);
577
578         ipath_cdbg(VERBOSE, "Try to read spc chip revision\n");
579         dd->ipath_revision =
580                 ipath_read_kreg64(dd, dd->ipath_kregs->kr_revision);
581
582         /*
583          * set up fundamental info we need to use the chip; we assume
584          * if the revision reg and these regs are OK, we don't need to
585          * special case the rest
586          */
587         dd->ipath_sregbase =
588                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_sendregbase);
589         dd->ipath_cregbase =
590                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_counterregbase);
591         dd->ipath_uregbase =
592                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_userregbase);
593         ipath_cdbg(VERBOSE, "ipath_kregbase %p, sendbase %x usrbase %x, "
594                    "cntrbase %x\n", dd->ipath_kregbase, dd->ipath_sregbase,
595                    dd->ipath_uregbase, dd->ipath_cregbase);
596         if ((dd->ipath_revision & 0xffffffff) == 0xffffffff
597             || (dd->ipath_sregbase & 0xffffffff) == 0xffffffff
598             || (dd->ipath_cregbase & 0xffffffff) == 0xffffffff
599             || (dd->ipath_uregbase & 0xffffffff) == 0xffffffff) {
600                 ipath_dev_err(dd, "Register read failures from chip, "
601                               "giving up initialization\n");
602                 dd->ipath_flags &= ~IPATH_PRESENT;
603                 ret = -ENODEV;
604                 goto done;
605         }
606
607
608         /* clear diagctrl register, in case diags were running and crashed */
609         ipath_write_kreg (dd, dd->ipath_kregs->kr_hwdiagctrl, 0);
610
611         /* clear the initial reset flag, in case first driver load */
612         ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear,
613                          INFINIPATH_E_RESET);
614
615         if (reinit)
616                 ret = init_chip_reset(dd, pdp);
617         else
618                 ret = init_chip_first(dd, pdp);
619
620         if (ret)
621                 goto done;
622
623         ipath_cdbg(VERBOSE, "Revision %llx (PCI %x), %u ports, %u tids, "
624                    "%u egrtids\n", (unsigned long long) dd->ipath_revision,
625                    dd->ipath_pcirev, dd->ipath_portcnt, dd->ipath_rcvtidcnt,
626                    dd->ipath_rcvegrcnt);
627
628         if (((dd->ipath_revision >> INFINIPATH_R_SOFTWARE_SHIFT) &
629              INFINIPATH_R_SOFTWARE_MASK) != IPATH_CHIP_SWVERSION) {
630                 ipath_dev_err(dd, "Driver only handles version %d, "
631                               "chip swversion is %d (%llx), failng\n",
632                               IPATH_CHIP_SWVERSION,
633                               (int)(dd->ipath_revision >>
634                                     INFINIPATH_R_SOFTWARE_SHIFT) &
635                               INFINIPATH_R_SOFTWARE_MASK,
636                               (unsigned long long) dd->ipath_revision);
637                 ret = -ENOSYS;
638                 goto done;
639         }
640         dd->ipath_majrev = (u8) ((dd->ipath_revision >>
641                                   INFINIPATH_R_CHIPREVMAJOR_SHIFT) &
642                                  INFINIPATH_R_CHIPREVMAJOR_MASK);
643         dd->ipath_minrev = (u8) ((dd->ipath_revision >>
644                                   INFINIPATH_R_CHIPREVMINOR_SHIFT) &
645                                  INFINIPATH_R_CHIPREVMINOR_MASK);
646         dd->ipath_boardrev = (u8) ((dd->ipath_revision >>
647                                     INFINIPATH_R_BOARDID_SHIFT) &
648                                    INFINIPATH_R_BOARDID_MASK);
649
650         ret = dd->ipath_f_get_boardname(dd, boardn, sizeof boardn);
651
652         snprintf(dd->ipath_boardversion, sizeof(dd->ipath_boardversion),
653                  "Driver %u.%u, %s, InfiniPath%u %u.%u, PCI %u, "
654                  "SW Compat %u\n",
655                  IPATH_CHIP_VERS_MAJ, IPATH_CHIP_VERS_MIN, boardn,
656                  (unsigned)(dd->ipath_revision >> INFINIPATH_R_ARCH_SHIFT) &
657                  INFINIPATH_R_ARCH_MASK,
658                  dd->ipath_majrev, dd->ipath_minrev, dd->ipath_pcirev,
659                  (unsigned)(dd->ipath_revision >>
660                             INFINIPATH_R_SOFTWARE_SHIFT) &
661                  INFINIPATH_R_SOFTWARE_MASK);
662
663         ipath_dbg("%s", dd->ipath_boardversion);
664
665 done:
666         return ret;
667 }
668
669
670 /**
671  * ipath_init_chip - do the actual initialization sequence on the chip
672  * @dd: the infinipath device
673  * @reinit: reinitializing, so don't allocate new memory
674  *
675  * Do the actual initialization sequence on the chip.  This is done
676  * both from the init routine called from the PCI infrastructure, and
677  * when we reset the chip, or detect that it was reset internally,
678  * or it's administratively re-enabled.
679  *
680  * Memory allocation here and in called routines is only done in
681  * the first case (reinit == 0).  We have to be careful, because even
682  * without memory allocation, we need to re-write all the chip registers
683  * TIDs, etc. after the reset or enable has completed.
684  */
685 int ipath_init_chip(struct ipath_devdata *dd, int reinit)
686 {
687         int ret = 0, i;
688         u32 val32, kpiobufs;
689         u32 piobufs, uports;
690         u64 val;
691         struct ipath_portdata *pd = NULL; /* keep gcc4 happy */
692         gfp_t gfp_flags = GFP_USER | __GFP_COMP;
693
694         ret = init_housekeeping(dd, &pd, reinit);
695         if (ret)
696                 goto done;
697
698         /*
699          * we ignore most issues after reporting them, but have to specially
700          * handle hardware-disabled chips.
701          */
702         if (ret == 2) {
703                 /* unique error, known to ipath_init_one */
704                 ret = -EPERM;
705                 goto done;
706         }
707
708         /*
709          * We could bump this to allow for full rcvegrcnt + rcvtidcnt,
710          * but then it no longer nicely fits power of two, and since
711          * we now use routines that backend onto __get_free_pages, the
712          * rest would be wasted.
713          */
714         dd->ipath_rcvhdrcnt = dd->ipath_rcvegrcnt;
715         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvhdrcnt,
716                          dd->ipath_rcvhdrcnt);
717
718         /*
719          * Set up the shadow copies of the piobufavail registers,
720          * which we compare against the chip registers for now, and
721          * the in memory DMA'ed copies of the registers.  This has to
722          * be done early, before we calculate lastport, etc.
723          */
724         piobufs = dd->ipath_piobcnt2k + dd->ipath_piobcnt4k;
725         /*
726          * calc number of pioavail registers, and save it; we have 2
727          * bits per buffer.
728          */
729         dd->ipath_pioavregs = ALIGN(piobufs, sizeof(u64) * BITS_PER_BYTE / 2)
730                 / (sizeof(u64) * BITS_PER_BYTE / 2);
731         uports = dd->ipath_cfgports ? dd->ipath_cfgports - 1 : 0;
732         if (ipath_kpiobufs == 0) {
733                 /* not set by user (this is default) */
734                 if (piobufs >= (uports * IPATH_MIN_USER_PORT_BUFCNT) + 32)
735                         kpiobufs = 32;
736                 else
737                         kpiobufs = 16;
738         }
739         else
740                 kpiobufs = ipath_kpiobufs;
741
742         if (kpiobufs + (uports * IPATH_MIN_USER_PORT_BUFCNT) > piobufs) {
743                 i = (int) piobufs -
744                         (int) (uports * IPATH_MIN_USER_PORT_BUFCNT);
745                 if (i < 0)
746                         i = 0;
747                 dev_info(&dd->pcidev->dev, "Allocating %d PIO bufs of "
748                          "%d for kernel leaves too few for %d user ports "
749                          "(%d each); using %u\n", kpiobufs,
750                          piobufs, uports, IPATH_MIN_USER_PORT_BUFCNT, i);
751                 /*
752                  * shouldn't change ipath_kpiobufs, because could be
753                  * different for different devices...
754                  */
755                 kpiobufs = i;
756         }
757         dd->ipath_lastport_piobuf = piobufs - kpiobufs;
758         dd->ipath_pbufsport =
759                 uports ? dd->ipath_lastport_piobuf / uports : 0;
760         val32 = dd->ipath_lastport_piobuf - (dd->ipath_pbufsport * uports);
761         if (val32 > 0) {
762                 ipath_dbg("allocating %u pbufs/port leaves %u unused, "
763                           "add to kernel\n", dd->ipath_pbufsport, val32);
764                 dd->ipath_lastport_piobuf -= val32;
765                 ipath_dbg("%u pbufs/port leaves %u unused, add to kernel\n",
766                           dd->ipath_pbufsport, val32);
767         }
768         dd->ipath_lastpioindex = dd->ipath_lastport_piobuf;
769         ipath_cdbg(VERBOSE, "%d PIO bufs for kernel out of %d total %u "
770                    "each for %u user ports\n", kpiobufs,
771                    piobufs, dd->ipath_pbufsport, uports);
772
773         dd->ipath_f_early_init(dd);
774
775         /* early_init sets rcvhdrentsize and rcvhdrsize, so this must be
776          * done after early_init */
777         dd->ipath_hdrqlast =
778                 dd->ipath_rcvhdrentsize * (dd->ipath_rcvhdrcnt - 1);
779         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvhdrentsize,
780                          dd->ipath_rcvhdrentsize);
781         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvhdrsize,
782                          dd->ipath_rcvhdrsize);
783
784         if (!reinit) {
785                 ret = init_pioavailregs(dd);
786                 init_shadow_tids(dd);
787                 if (ret)
788                         goto done;
789         }
790
791         (void)ipath_write_kreg(dd, dd->ipath_kregs->kr_sendpioavailaddr,
792                                dd->ipath_pioavailregs_phys);
793         /*
794          * this is to detect s/w errors, which the h/w works around by
795          * ignoring the low 6 bits of address, if it wasn't aligned.
796          */
797         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendpioavailaddr);
798         if (val != dd->ipath_pioavailregs_phys) {
799                 ipath_dev_err(dd, "Catastrophic software error, "
800                               "SendPIOAvailAddr written as %lx, "
801                               "read back as %llx\n",
802                               (unsigned long) dd->ipath_pioavailregs_phys,
803                               (unsigned long long) val);
804                 ret = -EINVAL;
805                 goto done;
806         }
807
808         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvbthqp, IPATH_KD_QP);
809
810         /*
811          * make sure we are not in freeze, and PIO send enabled, so
812          * writes to pbc happen
813          */
814         ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrmask, 0ULL);
815         ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear,
816                          ~0ULL&~INFINIPATH_HWE_MEMBISTFAILED);
817         ipath_write_kreg(dd, dd->ipath_kregs->kr_control, 0ULL);
818         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
819                          INFINIPATH_S_PIOENABLE);
820
821         /*
822          * before error clears, since we expect serdes pll errors during
823          * this, the first time after reset
824          */
825         if (bringup_link(dd)) {
826                 dev_info(&dd->pcidev->dev, "Failed to bringup IB link\n");
827                 ret = -ENETDOWN;
828                 goto done;
829         }
830
831         /*
832          * clear any "expected" hwerrs from reset and/or initialization
833          * clear any that aren't enabled (at least this once), and then
834          * set the enable mask
835          */
836         dd->ipath_f_init_hwerrors(dd);
837         ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear,
838                          ~0ULL&~INFINIPATH_HWE_MEMBISTFAILED);
839         ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrmask,
840                          dd->ipath_hwerrmask);
841
842         dd->ipath_maskederrs = dd->ipath_ignorederrs;
843         /* clear all */
844         ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear, -1LL);
845         /* enable errors that are masked, at least this first time. */
846         ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
847                          ~dd->ipath_maskederrs);
848         /* clear any interrups up to this point (ints still not enabled) */
849         ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, -1LL);
850
851         /*
852          * Set up the port 0 (kernel) rcvhdr q and egr TIDs.  If doing
853          * re-init, the simplest way to handle this is to free
854          * existing, and re-allocate.
855          * Need to re-create rest of port 0 portdata as well.
856          */
857         if (reinit) {
858                 /* Alloc and init new ipath_portdata for port0,
859                  * Then free old pd. Could lead to fragmentation, but also
860                  * makes later support for hot-swap easier.
861                  */
862                 struct ipath_portdata *npd;
863                 npd = create_portdata0(dd);
864                 if (npd) {
865                         ipath_free_pddata(dd, pd);
866                         dd->ipath_pd[0] = pd = npd;
867                 } else {
868                         ipath_dev_err(dd, "Unable to allocate portdata for"
869                                       "  port 0, failing\n");
870                         ret = -ENOMEM;
871                         goto done;
872                 }
873         }
874         dd->ipath_f_tidtemplate(dd);
875         ret = ipath_create_rcvhdrq(dd, pd);
876         if (!ret) {
877                 dd->ipath_hdrqtailptr =
878                         (volatile __le64 *)pd->port_rcvhdrtail_kvaddr;
879                 ret = create_port0_egr(dd);
880         }
881         if (ret)
882                 ipath_dev_err(dd, "failed to allocate port 0 (kernel) "
883                               "rcvhdrq and/or egr bufs\n");
884         else
885                 enable_chip(dd, pd, reinit);
886
887
888         if (!ret && !reinit) {
889             /* used when we close a port, for DMA already in flight at close */
890                 dd->ipath_dummy_hdrq = dma_alloc_coherent(
891                         &dd->pcidev->dev, pd->port_rcvhdrq_size,
892                         &dd->ipath_dummy_hdrq_phys,
893                         gfp_flags);
894                 if (!dd->ipath_dummy_hdrq ) {
895                         dev_info(&dd->pcidev->dev,
896                                 "Couldn't allocate 0x%lx bytes for dummy hdrq\n",
897                                 pd->port_rcvhdrq_size);
898                         /* fallback to just 0'ing */
899                         dd->ipath_dummy_hdrq_phys = 0UL;
900                 }
901         }
902
903         /*
904          * cause retrigger of pending interrupts ignored during init,
905          * even if we had errors
906          */
907         ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, 0ULL);
908
909         if(!dd->ipath_stats_timer_active) {
910                 /*
911                  * first init, or after an admin disable/enable
912                  * set up stats retrieval timer, even if we had errors
913                  * in last portion of setup
914                  */
915                 init_timer(&dd->ipath_stats_timer);
916                 dd->ipath_stats_timer.function = ipath_get_faststats;
917                 dd->ipath_stats_timer.data = (unsigned long) dd;
918                 /* every 5 seconds; */
919                 dd->ipath_stats_timer.expires = jiffies + 5 * HZ;
920                 /* takes ~16 seconds to overflow at full IB 4x bandwdith */
921                 add_timer(&dd->ipath_stats_timer);
922                 dd->ipath_stats_timer_active = 1;
923         }
924
925 done:
926         if (!ret) {
927                 *dd->ipath_statusp |= IPATH_STATUS_CHIP_PRESENT;
928                 if (!dd->ipath_f_intrsetup(dd)) {
929                         /* now we can enable all interrupts from the chip */
930                         ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask,
931                                          -1LL);
932                         /* force re-interrupt of any pending interrupts. */
933                         ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear,
934                                          0ULL);
935                         /* chip is usable; mark it as initialized */
936                         *dd->ipath_statusp |= IPATH_STATUS_INITTED;
937                 } else
938                         ipath_dev_err(dd, "No interrupts enabled, couldn't "
939                                       "setup interrupt address\n");
940
941                 if (dd->ipath_cfgports > ipath_stats.sps_nports)
942                         /*
943                          * sps_nports is a global, so, we set it to
944                          * the highest number of ports of any of the
945                          * chips we find; we never decrement it, at
946                          * least for now.  Since this might have changed
947                          * over disable/enable or prior to reset, always
948                          * do the check and potentially adjust.
949                          */
950                         ipath_stats.sps_nports = dd->ipath_cfgports;
951         } else
952                 ipath_dbg("Failed (%d) to initialize chip\n", ret);
953
954         /* if ret is non-zero, we probably should do some cleanup
955            here... */
956         return ret;
957 }
958
959 static int ipath_set_kpiobufs(const char *str, struct kernel_param *kp)
960 {
961         struct ipath_devdata *dd;
962         unsigned long flags;
963         unsigned short val;
964         int ret;
965
966         ret = ipath_parse_ushort(str, &val);
967
968         spin_lock_irqsave(&ipath_devs_lock, flags);
969
970         if (ret < 0)
971                 goto bail;
972
973         if (val == 0) {
974                 ret = -EINVAL;
975                 goto bail;
976         }
977
978         list_for_each_entry(dd, &ipath_dev_list, ipath_list) {
979                 if (dd->ipath_kregbase)
980                         continue;
981                 if (val > (dd->ipath_piobcnt2k + dd->ipath_piobcnt4k -
982                            (dd->ipath_cfgports *
983                             IPATH_MIN_USER_PORT_BUFCNT)))
984                 {
985                         ipath_dev_err(
986                                 dd,
987                                 "Allocating %d PIO bufs for kernel leaves "
988                                 "too few for %d user ports (%d each)\n",
989                                 val, dd->ipath_cfgports - 1,
990                                 IPATH_MIN_USER_PORT_BUFCNT);
991                         ret = -EINVAL;
992                         goto bail;
993                 }
994                 dd->ipath_lastport_piobuf =
995                         dd->ipath_piobcnt2k + dd->ipath_piobcnt4k - val;
996         }
997
998         ipath_kpiobufs = val;
999         ret = 0;
1000 bail:
1001         spin_unlock_irqrestore(&ipath_devs_lock, flags);
1002
1003         return ret;
1004 }