Merge branch 'linus' into timers/hrtimers
[linux-2.6] / drivers / infiniband / hw / ipath / ipath_init_chip.c
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/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 7
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  * Initialized based on number of PIO buffers if not set via module interface.
58  * The problem with this is that it's global, but we'll use different
59  * numbers for different chip types.
60  */
61 static ushort ipath_kpiobufs;
62
63 static int ipath_set_kpiobufs(const char *val, struct kernel_param *kp);
64
65 module_param_call(kpiobufs, ipath_set_kpiobufs, param_get_ushort,
66                   &ipath_kpiobufs, S_IWUSR | S_IRUGO);
67 MODULE_PARM_DESC(kpiobufs, "Set number of PIO buffers for driver");
68
69 /**
70  * create_port0_egr - allocate the eager TID buffers
71  * @dd: the infinipath device
72  *
73  * This code is now quite different for user and kernel, because
74  * the kernel uses skb's, for the accelerated network performance.
75  * This is the kernel (port0) version.
76  *
77  * Allocate the eager TID buffers and program them into infinipath.
78  * We use the network layer alloc_skb() allocator to allocate the
79  * memory, and either use the buffers as is for things like verbs
80  * packets, or pass the buffers up to the ipath layered driver and
81  * thence the network layer, replacing them as we do so (see
82  * ipath_rcv_layer()).
83  */
84 static int create_port0_egr(struct ipath_devdata *dd)
85 {
86         unsigned e, egrcnt;
87         struct ipath_skbinfo *skbinfo;
88         int ret;
89
90         egrcnt = dd->ipath_p0_rcvegrcnt;
91
92         skbinfo = vmalloc(sizeof(*dd->ipath_port0_skbinfo) * egrcnt);
93         if (skbinfo == NULL) {
94                 ipath_dev_err(dd, "allocation error for eager TID "
95                               "skb array\n");
96                 ret = -ENOMEM;
97                 goto bail;
98         }
99         for (e = 0; e < egrcnt; e++) {
100                 /*
101                  * This is a bit tricky in that we allocate extra
102                  * space for 2 bytes of the 14 byte ethernet header.
103                  * These two bytes are passed in the ipath header so
104                  * the rest of the data is word aligned.  We allocate
105                  * 4 bytes so that the data buffer stays word aligned.
106                  * See ipath_kreceive() for more details.
107                  */
108                 skbinfo[e].skb = ipath_alloc_skb(dd, GFP_KERNEL);
109                 if (!skbinfo[e].skb) {
110                         ipath_dev_err(dd, "SKB allocation error for "
111                                       "eager TID %u\n", e);
112                         while (e != 0)
113                                 dev_kfree_skb(skbinfo[--e].skb);
114                         vfree(skbinfo);
115                         ret = -ENOMEM;
116                         goto bail;
117                 }
118         }
119         /*
120          * After loop above, so we can test non-NULL to see if ready
121          * to use at receive, etc.
122          */
123         dd->ipath_port0_skbinfo = skbinfo;
124
125         for (e = 0; e < egrcnt; e++) {
126                 dd->ipath_port0_skbinfo[e].phys =
127                   ipath_map_single(dd->pcidev,
128                                    dd->ipath_port0_skbinfo[e].skb->data,
129                                    dd->ipath_ibmaxlen, PCI_DMA_FROMDEVICE);
130                 dd->ipath_f_put_tid(dd, e + (u64 __iomem *)
131                                     ((char __iomem *) dd->ipath_kregbase +
132                                      dd->ipath_rcvegrbase),
133                                     RCVHQ_RCV_TYPE_EAGER,
134                                     dd->ipath_port0_skbinfo[e].phys);
135         }
136
137         ret = 0;
138
139 bail:
140         return ret;
141 }
142
143 static int bringup_link(struct ipath_devdata *dd)
144 {
145         u64 val, ibc;
146         int ret = 0;
147
148         /* hold IBC in reset */
149         dd->ipath_control &= ~INFINIPATH_C_LINKENABLE;
150         ipath_write_kreg(dd, dd->ipath_kregs->kr_control,
151                          dd->ipath_control);
152
153         /*
154          * set initial max size pkt IBC will send, including ICRC; it's the
155          * PIO buffer size in dwords, less 1; also see ipath_set_mtu()
156          */
157         val = (dd->ipath_ibmaxlen >> 2) + 1;
158         ibc = val << dd->ibcc_mpl_shift;
159
160         /* flowcontrolwatermark is in units of KBytes */
161         ibc |= 0x5ULL << INFINIPATH_IBCC_FLOWCTRLWATERMARK_SHIFT;
162         /*
163          * How often flowctrl sent.  More or less in usecs; balance against
164          * watermark value, so that in theory senders always get a flow
165          * control update in time to not let the IB link go idle.
166          */
167         ibc |= 0x3ULL << INFINIPATH_IBCC_FLOWCTRLPERIOD_SHIFT;
168         /* max error tolerance */
169         ibc |= 0xfULL << INFINIPATH_IBCC_PHYERRTHRESHOLD_SHIFT;
170         /* use "real" buffer space for */
171         ibc |= 4ULL << INFINIPATH_IBCC_CREDITSCALE_SHIFT;
172         /* IB credit flow control. */
173         ibc |= 0xfULL << INFINIPATH_IBCC_OVERRUNTHRESHOLD_SHIFT;
174         /* initially come up waiting for TS1, without sending anything. */
175         dd->ipath_ibcctrl = ibc;
176         /*
177          * Want to start out with both LINKCMD and LINKINITCMD in NOP
178          * (0 and 0).  Don't put linkinitcmd in ipath_ibcctrl, want that
179          * to stay a NOP. Flag that we are disabled, for the (unlikely)
180          * case that some recovery path is trying to bring the link up
181          * before we are ready.
182          */
183         ibc |= INFINIPATH_IBCC_LINKINITCMD_DISABLE <<
184                 INFINIPATH_IBCC_LINKINITCMD_SHIFT;
185         dd->ipath_flags |= IPATH_IB_LINK_DISABLED;
186         ipath_cdbg(VERBOSE, "Writing 0x%llx to ibcctrl\n",
187                    (unsigned long long) ibc);
188         ipath_write_kreg(dd, dd->ipath_kregs->kr_ibcctrl, ibc);
189
190         // be sure chip saw it
191         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
192
193         ret = dd->ipath_f_bringup_serdes(dd);
194
195         if (ret)
196                 dev_info(&dd->pcidev->dev, "Could not initialize SerDes, "
197                          "not usable\n");
198         else {
199                 /* enable IBC */
200                 dd->ipath_control |= INFINIPATH_C_LINKENABLE;
201                 ipath_write_kreg(dd, dd->ipath_kregs->kr_control,
202                                  dd->ipath_control);
203         }
204
205         return ret;
206 }
207
208 static struct ipath_portdata *create_portdata0(struct ipath_devdata *dd)
209 {
210         struct ipath_portdata *pd = NULL;
211
212         pd = kzalloc(sizeof(*pd), GFP_KERNEL);
213         if (pd) {
214                 pd->port_dd = dd;
215                 pd->port_cnt = 1;
216                 /* The port 0 pkey table is used by the layer interface. */
217                 pd->port_pkeys[0] = IPATH_DEFAULT_P_KEY;
218                 pd->port_seq_cnt = 1;
219         }
220         return pd;
221 }
222
223 static int init_chip_first(struct ipath_devdata *dd)
224 {
225         struct ipath_portdata *pd;
226         int ret = 0;
227         u64 val;
228
229         spin_lock_init(&dd->ipath_kernel_tid_lock);
230         spin_lock_init(&dd->ipath_user_tid_lock);
231         spin_lock_init(&dd->ipath_sendctrl_lock);
232         spin_lock_init(&dd->ipath_uctxt_lock);
233         spin_lock_init(&dd->ipath_sdma_lock);
234         spin_lock_init(&dd->ipath_gpio_lock);
235         spin_lock_init(&dd->ipath_eep_st_lock);
236         spin_lock_init(&dd->ipath_sdepb_lock);
237         mutex_init(&dd->ipath_eep_lock);
238
239         /*
240          * skip cfgports stuff because we are not allocating memory,
241          * and we don't want problems if the portcnt changed due to
242          * cfgports.  We do still check and report a difference, if
243          * not same (should be impossible).
244          */
245         dd->ipath_f_config_ports(dd, ipath_cfgports);
246         if (!ipath_cfgports)
247                 dd->ipath_cfgports = dd->ipath_portcnt;
248         else if (ipath_cfgports <= dd->ipath_portcnt) {
249                 dd->ipath_cfgports = ipath_cfgports;
250                 ipath_dbg("Configured to use %u ports out of %u in chip\n",
251                           dd->ipath_cfgports, ipath_read_kreg32(dd,
252                           dd->ipath_kregs->kr_portcnt));
253         } else {
254                 dd->ipath_cfgports = dd->ipath_portcnt;
255                 ipath_dbg("Tried to configured to use %u ports; chip "
256                           "only supports %u\n", ipath_cfgports,
257                           ipath_read_kreg32(dd,
258                                   dd->ipath_kregs->kr_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         pd = create_portdata0(dd);
275         if (!pd) {
276                 ipath_dev_err(dd, "Unable to allocate portdata for port "
277                               "0, failing\n");
278                 ret = -ENOMEM;
279                 goto done;
280         }
281         dd->ipath_pd[0] = pd;
282
283         dd->ipath_rcvtidcnt =
284                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvtidcnt);
285         dd->ipath_rcvtidbase =
286                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvtidbase);
287         dd->ipath_rcvegrcnt =
288                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvegrcnt);
289         dd->ipath_rcvegrbase =
290                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvegrbase);
291         dd->ipath_palign =
292                 ipath_read_kreg32(dd, dd->ipath_kregs->kr_pagealign);
293         dd->ipath_piobufbase =
294                 ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendpiobufbase);
295         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendpiosize);
296         dd->ipath_piosize2k = val & ~0U;
297         dd->ipath_piosize4k = val >> 32;
298         if (dd->ipath_piosize4k == 0 && ipath_mtu4096)
299                 ipath_mtu4096 = 0; /* 4KB not supported by this chip */
300         dd->ipath_ibmtu = ipath_mtu4096 ? 4096 : 2048;
301         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendpiobufcnt);
302         dd->ipath_piobcnt2k = val & ~0U;
303         dd->ipath_piobcnt4k = val >> 32;
304         dd->ipath_pio2kbase =
305                 (u32 __iomem *) (((char __iomem *) dd->ipath_kregbase) +
306                                  (dd->ipath_piobufbase & 0xffffffff));
307         if (dd->ipath_piobcnt4k) {
308                 dd->ipath_pio4kbase = (u32 __iomem *)
309                         (((char __iomem *) dd->ipath_kregbase) +
310                          (dd->ipath_piobufbase >> 32));
311                 /*
312                  * 4K buffers take 2 pages; we use roundup just to be
313                  * paranoid; we calculate it once here, rather than on
314                  * ever buf allocate
315                  */
316                 dd->ipath_4kalign = ALIGN(dd->ipath_piosize4k,
317                                           dd->ipath_palign);
318                 ipath_dbg("%u 2k(%x) piobufs @ %p, %u 4k(%x) @ %p "
319                           "(%x aligned)\n",
320                           dd->ipath_piobcnt2k, dd->ipath_piosize2k,
321                           dd->ipath_pio2kbase, dd->ipath_piobcnt4k,
322                           dd->ipath_piosize4k, dd->ipath_pio4kbase,
323                           dd->ipath_4kalign);
324         }
325         else ipath_dbg("%u 2k piobufs @ %p\n",
326                        dd->ipath_piobcnt2k, dd->ipath_pio2kbase);
327
328 done:
329         return ret;
330 }
331
332 /**
333  * init_chip_reset - re-initialize after a reset, or enable
334  * @dd: the infinipath device
335  *
336  * sanity check at least some of the values after reset, and
337  * ensure no receive or transmit (explictly, in case reset
338  * failed
339  */
340 static int init_chip_reset(struct ipath_devdata *dd)
341 {
342         u32 rtmp;
343         int i;
344         unsigned long flags;
345
346         /*
347          * ensure chip does no sends or receives, tail updates, or
348          * pioavail updates while we re-initialize
349          */
350         dd->ipath_rcvctrl &= ~(1ULL << dd->ipath_r_tailupd_shift);
351         for (i = 0; i < dd->ipath_portcnt; i++) {
352                 clear_bit(dd->ipath_r_portenable_shift + i,
353                           &dd->ipath_rcvctrl);
354                 clear_bit(dd->ipath_r_intravail_shift + i,
355                           &dd->ipath_rcvctrl);
356         }
357         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
358                 dd->ipath_rcvctrl);
359
360         spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
361         dd->ipath_sendctrl = 0U; /* no sdma, etc */
362         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, dd->ipath_sendctrl);
363         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
364         spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
365
366         ipath_write_kreg(dd, dd->ipath_kregs->kr_control, 0ULL);
367
368         rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvtidcnt);
369         if (rtmp != dd->ipath_rcvtidcnt)
370                 dev_info(&dd->pcidev->dev, "tidcnt was %u before "
371                          "reset, now %u, using original\n",
372                          dd->ipath_rcvtidcnt, rtmp);
373         rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvtidbase);
374         if (rtmp != dd->ipath_rcvtidbase)
375                 dev_info(&dd->pcidev->dev, "tidbase was %u before "
376                          "reset, now %u, using original\n",
377                          dd->ipath_rcvtidbase, rtmp);
378         rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvegrcnt);
379         if (rtmp != dd->ipath_rcvegrcnt)
380                 dev_info(&dd->pcidev->dev, "egrcnt was %u before "
381                          "reset, now %u, using original\n",
382                          dd->ipath_rcvegrcnt, rtmp);
383         rtmp = ipath_read_kreg32(dd, dd->ipath_kregs->kr_rcvegrbase);
384         if (rtmp != dd->ipath_rcvegrbase)
385                 dev_info(&dd->pcidev->dev, "egrbase was %u before "
386                          "reset, now %u, using original\n",
387                          dd->ipath_rcvegrbase, rtmp);
388
389         return 0;
390 }
391
392 static int init_pioavailregs(struct ipath_devdata *dd)
393 {
394         int ret;
395
396         dd->ipath_pioavailregs_dma = dma_alloc_coherent(
397                 &dd->pcidev->dev, PAGE_SIZE, &dd->ipath_pioavailregs_phys,
398                 GFP_KERNEL);
399         if (!dd->ipath_pioavailregs_dma) {
400                 ipath_dev_err(dd, "failed to allocate PIOavail reg area "
401                               "in memory\n");
402                 ret = -ENOMEM;
403                 goto done;
404         }
405
406         /*
407          * we really want L2 cache aligned, but for current CPUs of
408          * interest, they are the same.
409          */
410         dd->ipath_statusp = (u64 *)
411                 ((char *)dd->ipath_pioavailregs_dma +
412                  ((2 * L1_CACHE_BYTES +
413                    dd->ipath_pioavregs * sizeof(u64)) & ~L1_CACHE_BYTES));
414         /* copy the current value now that it's really allocated */
415         *dd->ipath_statusp = dd->_ipath_status;
416         /*
417          * setup buffer to hold freeze msg, accessible to apps,
418          * following statusp
419          */
420         dd->ipath_freezemsg = (char *)&dd->ipath_statusp[1];
421         /* and its length */
422         dd->ipath_freezelen = L1_CACHE_BYTES - sizeof(dd->ipath_statusp[0]);
423
424         ret = 0;
425
426 done:
427         return ret;
428 }
429
430 /**
431  * init_shadow_tids - allocate the shadow TID array
432  * @dd: the infinipath device
433  *
434  * allocate the shadow TID array, so we can ipath_munlock previous
435  * entries.  It may make more sense to move the pageshadow to the
436  * port data structure, so we only allocate memory for ports actually
437  * in use, since we at 8k per port, now.
438  */
439 static void init_shadow_tids(struct ipath_devdata *dd)
440 {
441         struct page **pages;
442         dma_addr_t *addrs;
443
444         pages = vmalloc(dd->ipath_cfgports * dd->ipath_rcvtidcnt *
445                         sizeof(struct page *));
446         if (!pages) {
447                 ipath_dev_err(dd, "failed to allocate shadow page * "
448                               "array, no expected sends!\n");
449                 dd->ipath_pageshadow = NULL;
450                 return;
451         }
452
453         addrs = vmalloc(dd->ipath_cfgports * dd->ipath_rcvtidcnt *
454                         sizeof(dma_addr_t));
455         if (!addrs) {
456                 ipath_dev_err(dd, "failed to allocate shadow dma handle "
457                               "array, no expected sends!\n");
458                 vfree(dd->ipath_pageshadow);
459                 dd->ipath_pageshadow = NULL;
460                 return;
461         }
462
463         memset(pages, 0, dd->ipath_cfgports * dd->ipath_rcvtidcnt *
464                sizeof(struct page *));
465
466         dd->ipath_pageshadow = pages;
467         dd->ipath_physshadow = addrs;
468 }
469
470 static 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
557 static 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
667 done:
668         return ret;
669 }
670
671 static 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  */
708 int 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 ignore most issues after reporting them, but have to specially
723          * handle hardware-disabled chips.
724          */
725         if (ret == 2) {
726                 /* unique error, known to ipath_init_one */
727                 ret = -EPERM;
728                 goto done;
729         }
730
731         /*
732          * We could bump this to allow for full rcvegrcnt + rcvtidcnt,
733          * but then it no longer nicely fits power of two, and since
734          * we now use routines that backend onto __get_free_pages, the
735          * rest would be wasted.
736          */
737         dd->ipath_rcvhdrcnt = max(dd->ipath_p0_rcvegrcnt, dd->ipath_rcvegrcnt);
738         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvhdrcnt,
739                          dd->ipath_rcvhdrcnt);
740
741         /*
742          * Set up the shadow copies of the piobufavail registers,
743          * which we compare against the chip registers for now, and
744          * the in memory DMA'ed copies of the registers.  This has to
745          * be done early, before we calculate lastport, etc.
746          */
747         piobufs = dd->ipath_piobcnt2k + dd->ipath_piobcnt4k;
748         /*
749          * calc number of pioavail registers, and save it; we have 2
750          * bits per buffer.
751          */
752         dd->ipath_pioavregs = ALIGN(piobufs, sizeof(u64) * BITS_PER_BYTE / 2)
753                 / (sizeof(u64) * BITS_PER_BYTE / 2);
754         uports = dd->ipath_cfgports ? dd->ipath_cfgports - 1 : 0;
755         if (piobufs > 144)
756                 defkbufs = 32 + dd->ipath_pioreserved;
757         else
758                 defkbufs = 16 + dd->ipath_pioreserved;
759
760         if (ipath_kpiobufs && (ipath_kpiobufs +
761                 (uports * IPATH_MIN_USER_PORT_BUFCNT)) > piobufs) {
762                 int i = (int) piobufs -
763                         (int) (uports * IPATH_MIN_USER_PORT_BUFCNT);
764                 if (i < 1)
765                         i = 1;
766                 dev_info(&dd->pcidev->dev, "Allocating %d PIO bufs of "
767                          "%d for kernel leaves too few for %d user ports "
768                          "(%d each); using %u\n", ipath_kpiobufs,
769                          piobufs, uports, IPATH_MIN_USER_PORT_BUFCNT, i);
770                 /*
771                  * shouldn't change ipath_kpiobufs, because could be
772                  * different for different devices...
773                  */
774                 kpiobufs = i;
775         } else if (ipath_kpiobufs)
776                 kpiobufs = ipath_kpiobufs;
777         else
778                 kpiobufs = defkbufs;
779         dd->ipath_lastport_piobuf = piobufs - kpiobufs;
780         dd->ipath_pbufsport =
781                 uports ? dd->ipath_lastport_piobuf / uports : 0;
782         /* if not an even divisor, some user ports get extra buffers */
783         dd->ipath_ports_extrabuf = dd->ipath_lastport_piobuf -
784                 (dd->ipath_pbufsport * uports);
785         if (dd->ipath_ports_extrabuf)
786                 ipath_dbg("%u pbufs/port leaves some unused, add 1 buffer to "
787                         "ports <= %u\n", dd->ipath_pbufsport,
788                         dd->ipath_ports_extrabuf);
789         dd->ipath_lastpioindex = 0;
790         dd->ipath_lastpioindexl = dd->ipath_piobcnt2k;
791         /* ipath_pioavailshadow initialized earlier */
792         ipath_cdbg(VERBOSE, "%d PIO bufs for kernel out of %d total %u "
793                    "each for %u user ports\n", kpiobufs,
794                    piobufs, dd->ipath_pbufsport, uports);
795         ret = dd->ipath_f_early_init(dd);
796         if (ret) {
797                 ipath_dev_err(dd, "Early initialization failure\n");
798                 goto done;
799         }
800
801         /*
802          * Early_init sets rcvhdrentsize and rcvhdrsize, so this must be
803          * done after early_init.
804          */
805         dd->ipath_hdrqlast =
806                 dd->ipath_rcvhdrentsize * (dd->ipath_rcvhdrcnt - 1);
807         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvhdrentsize,
808                          dd->ipath_rcvhdrentsize);
809         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvhdrsize,
810                          dd->ipath_rcvhdrsize);
811
812         if (!reinit) {
813                 ret = init_pioavailregs(dd);
814                 init_shadow_tids(dd);
815                 if (ret)
816                         goto done;
817         }
818
819         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendpioavailaddr,
820                          dd->ipath_pioavailregs_phys);
821
822         /*
823          * this is to detect s/w errors, which the h/w works around by
824          * ignoring the low 6 bits of address, if it wasn't aligned.
825          */
826         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_sendpioavailaddr);
827         if (val != dd->ipath_pioavailregs_phys) {
828                 ipath_dev_err(dd, "Catastrophic software error, "
829                               "SendPIOAvailAddr written as %lx, "
830                               "read back as %llx\n",
831                               (unsigned long) dd->ipath_pioavailregs_phys,
832                               (unsigned long long) val);
833                 ret = -EINVAL;
834                 goto done;
835         }
836
837         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvbthqp, IPATH_KD_QP);
838
839         /*
840          * make sure we are not in freeze, and PIO send enabled, so
841          * writes to pbc happen
842          */
843         ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrmask, 0ULL);
844         ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear,
845                          ~0ULL&~INFINIPATH_HWE_MEMBISTFAILED);
846         ipath_write_kreg(dd, dd->ipath_kregs->kr_control, 0ULL);
847
848         /*
849          * before error clears, since we expect serdes pll errors during
850          * this, the first time after reset
851          */
852         if (bringup_link(dd)) {
853                 dev_info(&dd->pcidev->dev, "Failed to bringup IB link\n");
854                 ret = -ENETDOWN;
855                 goto done;
856         }
857
858         /*
859          * clear any "expected" hwerrs from reset and/or initialization
860          * clear any that aren't enabled (at least this once), and then
861          * set the enable mask
862          */
863         dd->ipath_f_init_hwerrors(dd);
864         ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear,
865                          ~0ULL&~INFINIPATH_HWE_MEMBISTFAILED);
866         ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrmask,
867                          dd->ipath_hwerrmask);
868
869         /* clear all */
870         ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear, -1LL);
871         /* enable errors that are masked, at least this first time. */
872         ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
873                          ~dd->ipath_maskederrs);
874         dd->ipath_maskederrs = 0; /* don't re-enable ignored in timer */
875         dd->ipath_errormask =
876                 ipath_read_kreg64(dd, dd->ipath_kregs->kr_errormask);
877         /* clear any interrupts up to this point (ints still not enabled) */
878         ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, -1LL);
879
880         dd->ipath_f_tidtemplate(dd);
881
882         /*
883          * Set up the port 0 (kernel) rcvhdr q and egr TIDs.  If doing
884          * re-init, the simplest way to handle this is to free
885          * existing, and re-allocate.
886          * Need to re-create rest of port 0 portdata as well.
887          */
888         pd = dd->ipath_pd[0];
889         if (reinit) {
890                 struct ipath_portdata *npd;
891
892                 /*
893                  * Alloc and init new ipath_portdata for port0,
894                  * Then free old pd. Could lead to fragmentation, but also
895                  * makes later support for hot-swap easier.
896                  */
897                 npd = create_portdata0(dd);
898                 if (npd) {
899                         ipath_free_pddata(dd, pd);
900                         dd->ipath_pd[0] = npd;
901                         pd = npd;
902                 } else {
903                         ipath_dev_err(dd, "Unable to allocate portdata"
904                                       " for port 0, failing\n");
905                         ret = -ENOMEM;
906                         goto done;
907                 }
908         }
909         ret = ipath_create_rcvhdrq(dd, pd);
910         if (!ret)
911                 ret = create_port0_egr(dd);
912         if (ret) {
913                 ipath_dev_err(dd, "failed to allocate kernel port's "
914                               "rcvhdrq and/or egr bufs\n");
915                 goto done;
916         }
917         else
918                 enable_chip(dd, reinit);
919
920         /* after enable_chip, so pioavailshadow setup */
921         ipath_chg_pioavailkernel(dd, 0, piobufs, 1);
922
923         /*
924          * Cancel any possible active sends from early driver load.
925          * Follows early_init because some chips have to initialize
926          * PIO buffers in early_init to avoid false parity errors.
927          * After enable and ipath_chg_pioavailkernel so we can safely
928          * enable pioavail updates and PIOENABLE; packets are now
929          * ready to go out.
930          */
931         ipath_cancel_sends(dd, 1);
932
933         if (!reinit) {
934                 /*
935                  * Used when we close a port, for DMA already in flight
936                  * at close.
937                  */
938                 dd->ipath_dummy_hdrq = dma_alloc_coherent(
939                         &dd->pcidev->dev, dd->ipath_pd[0]->port_rcvhdrq_size,
940                         &dd->ipath_dummy_hdrq_phys,
941                         gfp_flags);
942                 if (!dd->ipath_dummy_hdrq) {
943                         dev_info(&dd->pcidev->dev,
944                                 "Couldn't allocate 0x%lx bytes for dummy hdrq\n",
945                                 dd->ipath_pd[0]->port_rcvhdrq_size);
946                         /* fallback to just 0'ing */
947                         dd->ipath_dummy_hdrq_phys = 0UL;
948                 }
949         }
950
951         /*
952          * cause retrigger of pending interrupts ignored during init,
953          * even if we had errors
954          */
955         ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, 0ULL);
956
957         if (!dd->ipath_stats_timer_active) {
958                 /*
959                  * first init, or after an admin disable/enable
960                  * set up stats retrieval timer, even if we had errors
961                  * in last portion of setup
962                  */
963                 init_timer(&dd->ipath_stats_timer);
964                 dd->ipath_stats_timer.function = ipath_get_faststats;
965                 dd->ipath_stats_timer.data = (unsigned long) dd;
966                 /* every 5 seconds; */
967                 dd->ipath_stats_timer.expires = jiffies + 5 * HZ;
968                 /* takes ~16 seconds to overflow at full IB 4x bandwdith */
969                 add_timer(&dd->ipath_stats_timer);
970                 dd->ipath_stats_timer_active = 1;
971         }
972
973         /* Set up SendDMA if chip supports it */
974         if (dd->ipath_flags & IPATH_HAS_SEND_DMA)
975                 ret = setup_sdma(dd);
976
977         /* Set up HoL state */
978         init_timer(&dd->ipath_hol_timer);
979         dd->ipath_hol_timer.function = ipath_hol_event;
980         dd->ipath_hol_timer.data = (unsigned long)dd;
981         dd->ipath_hol_state = IPATH_HOL_UP;
982
983 done:
984         if (!ret) {
985                 *dd->ipath_statusp |= IPATH_STATUS_CHIP_PRESENT;
986                 if (!dd->ipath_f_intrsetup(dd)) {
987                         /* now we can enable all interrupts from the chip */
988                         ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask,
989                                          -1LL);
990                         /* force re-interrupt of any pending interrupts. */
991                         ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear,
992                                          0ULL);
993                         /* chip is usable; mark it as initialized */
994                         *dd->ipath_statusp |= IPATH_STATUS_INITTED;
995
996                         /*
997                          * setup to verify we get an interrupt, and fallback
998                          * to an alternate if necessary and possible
999                          */
1000                         if (!reinit) {
1001                                 init_timer(&dd->ipath_intrchk_timer);
1002                                 dd->ipath_intrchk_timer.function =
1003                                         verify_interrupt;
1004                                 dd->ipath_intrchk_timer.data =
1005                                         (unsigned long) dd;
1006                         }
1007                         dd->ipath_intrchk_timer.expires = jiffies + HZ/2;
1008                         add_timer(&dd->ipath_intrchk_timer);
1009                 } else
1010                         ipath_dev_err(dd, "No interrupts enabled, couldn't "
1011                                       "setup interrupt address\n");
1012
1013                 if (dd->ipath_cfgports > ipath_stats.sps_nports)
1014                         /*
1015                          * sps_nports is a global, so, we set it to
1016                          * the highest number of ports of any of the
1017                          * chips we find; we never decrement it, at
1018                          * least for now.  Since this might have changed
1019                          * over disable/enable or prior to reset, always
1020                          * do the check and potentially adjust.
1021                          */
1022                         ipath_stats.sps_nports = dd->ipath_cfgports;
1023         } else
1024                 ipath_dbg("Failed (%d) to initialize chip\n", ret);
1025
1026         /* if ret is non-zero, we probably should do some cleanup
1027            here... */
1028         return ret;
1029 }
1030
1031 static int ipath_set_kpiobufs(const char *str, struct kernel_param *kp)
1032 {
1033         struct ipath_devdata *dd;
1034         unsigned long flags;
1035         unsigned short val;
1036         int ret;
1037
1038         ret = ipath_parse_ushort(str, &val);
1039
1040         spin_lock_irqsave(&ipath_devs_lock, flags);
1041
1042         if (ret < 0)
1043                 goto bail;
1044
1045         if (val == 0) {
1046                 ret = -EINVAL;
1047                 goto bail;
1048         }
1049
1050         list_for_each_entry(dd, &ipath_dev_list, ipath_list) {
1051                 if (dd->ipath_kregbase)
1052                         continue;
1053                 if (val > (dd->ipath_piobcnt2k + dd->ipath_piobcnt4k -
1054                            (dd->ipath_cfgports *
1055                             IPATH_MIN_USER_PORT_BUFCNT)))
1056                 {
1057                         ipath_dev_err(
1058                                 dd,
1059                                 "Allocating %d PIO bufs for kernel leaves "
1060                                 "too few for %d user ports (%d each)\n",
1061                                 val, dd->ipath_cfgports - 1,
1062                                 IPATH_MIN_USER_PORT_BUFCNT);
1063                         ret = -EINVAL;
1064                         goto bail;
1065                 }
1066                 dd->ipath_lastport_piobuf =
1067                         dd->ipath_piobcnt2k + dd->ipath_piobcnt4k - val;
1068         }
1069
1070         ipath_kpiobufs = val;
1071         ret = 0;
1072 bail:
1073         spin_unlock_irqrestore(&ipath_devs_lock, flags);
1074
1075         return ret;
1076 }