IB/ipath: add new minor device to allow sending of diag packets
[linux-2.6] / drivers / infiniband / hw / ipath / ipath_driver.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/spinlock.h>
35 #include <linux/idr.h>
36 #include <linux/pci.h>
37 #include <linux/delay.h>
38 #include <linux/netdevice.h>
39 #include <linux/vmalloc.h>
40
41 #include "ipath_kernel.h"
42 #include "ipath_verbs.h"
43 #include "ipath_common.h"
44
45 static void ipath_update_pio_bufs(struct ipath_devdata *);
46
47 const char *ipath_get_unit_name(int unit)
48 {
49         static char iname[16];
50         snprintf(iname, sizeof iname, "infinipath%u", unit);
51         return iname;
52 }
53
54 #define DRIVER_LOAD_MSG "QLogic " IPATH_DRV_NAME " loaded: "
55 #define PFX IPATH_DRV_NAME ": "
56
57 /*
58  * The size has to be longer than this string, so we can append
59  * board/chip information to it in the init code.
60  */
61 const char ib_ipath_version[] = IPATH_IDSTR "\n";
62
63 static struct idr unit_table;
64 DEFINE_SPINLOCK(ipath_devs_lock);
65 LIST_HEAD(ipath_dev_list);
66
67 wait_queue_head_t ipath_state_wait;
68
69 unsigned ipath_debug = __IPATH_INFO;
70
71 module_param_named(debug, ipath_debug, uint, S_IWUSR | S_IRUGO);
72 MODULE_PARM_DESC(debug, "mask for debug prints");
73 EXPORT_SYMBOL_GPL(ipath_debug);
74
75 MODULE_LICENSE("GPL");
76 MODULE_AUTHOR("QLogic <support@pathscale.com>");
77 MODULE_DESCRIPTION("QLogic InfiniPath driver");
78
79 const char *ipath_ibcstatus_str[] = {
80         "Disabled",
81         "LinkUp",
82         "PollActive",
83         "PollQuiet",
84         "SleepDelay",
85         "SleepQuiet",
86         "LState6",              /* unused */
87         "LState7",              /* unused */
88         "CfgDebounce",
89         "CfgRcvfCfg",
90         "CfgWaitRmt",
91         "CfgIdle",
92         "RecovRetrain",
93         "LState0xD",            /* unused */
94         "RecovWaitRmt",
95         "RecovIdle",
96 };
97
98 /*
99  * These variables are initialized in the chip-specific files
100  * but are defined here.
101  */
102 u16 ipath_gpio_sda_num, ipath_gpio_scl_num;
103 u64 ipath_gpio_sda, ipath_gpio_scl;
104 u64 infinipath_i_bitsextant;
105 ipath_err_t infinipath_e_bitsextant, infinipath_hwe_bitsextant;
106 u32 infinipath_i_rcvavail_mask, infinipath_i_rcvurg_mask;
107
108 static void __devexit ipath_remove_one(struct pci_dev *);
109 static int __devinit ipath_init_one(struct pci_dev *,
110                                     const struct pci_device_id *);
111
112 /* Only needed for registration, nothing else needs this info */
113 #define PCI_VENDOR_ID_PATHSCALE 0x1fc1
114 #define PCI_DEVICE_ID_INFINIPATH_HT 0xd
115 #define PCI_DEVICE_ID_INFINIPATH_PE800 0x10
116
117 static const struct pci_device_id ipath_pci_tbl[] = {
118         { PCI_DEVICE(PCI_VENDOR_ID_PATHSCALE, PCI_DEVICE_ID_INFINIPATH_HT) },
119         { PCI_DEVICE(PCI_VENDOR_ID_PATHSCALE, PCI_DEVICE_ID_INFINIPATH_PE800) },
120         { 0, }
121 };
122
123 MODULE_DEVICE_TABLE(pci, ipath_pci_tbl);
124
125 static struct pci_driver ipath_driver = {
126         .name = IPATH_DRV_NAME,
127         .probe = ipath_init_one,
128         .remove = __devexit_p(ipath_remove_one),
129         .id_table = ipath_pci_tbl,
130 };
131
132
133 static inline void read_bars(struct ipath_devdata *dd, struct pci_dev *dev,
134                              u32 *bar0, u32 *bar1)
135 {
136         int ret;
137
138         ret = pci_read_config_dword(dev, PCI_BASE_ADDRESS_0, bar0);
139         if (ret)
140                 ipath_dev_err(dd, "failed to read bar0 before enable: "
141                               "error %d\n", -ret);
142
143         ret = pci_read_config_dword(dev, PCI_BASE_ADDRESS_1, bar1);
144         if (ret)
145                 ipath_dev_err(dd, "failed to read bar1 before enable: "
146                               "error %d\n", -ret);
147
148         ipath_dbg("Read bar0 %x bar1 %x\n", *bar0, *bar1);
149 }
150
151 static void ipath_free_devdata(struct pci_dev *pdev,
152                                struct ipath_devdata *dd)
153 {
154         unsigned long flags;
155
156         pci_set_drvdata(pdev, NULL);
157
158         if (dd->ipath_unit != -1) {
159                 spin_lock_irqsave(&ipath_devs_lock, flags);
160                 idr_remove(&unit_table, dd->ipath_unit);
161                 list_del(&dd->ipath_list);
162                 spin_unlock_irqrestore(&ipath_devs_lock, flags);
163         }
164         vfree(dd);
165 }
166
167 static struct ipath_devdata *ipath_alloc_devdata(struct pci_dev *pdev)
168 {
169         unsigned long flags;
170         struct ipath_devdata *dd;
171         int ret;
172
173         if (!idr_pre_get(&unit_table, GFP_KERNEL)) {
174                 dd = ERR_PTR(-ENOMEM);
175                 goto bail;
176         }
177
178         dd = vmalloc(sizeof(*dd));
179         if (!dd) {
180                 dd = ERR_PTR(-ENOMEM);
181                 goto bail;
182         }
183         memset(dd, 0, sizeof(*dd));
184         dd->ipath_unit = -1;
185
186         spin_lock_irqsave(&ipath_devs_lock, flags);
187
188         ret = idr_get_new(&unit_table, dd, &dd->ipath_unit);
189         if (ret < 0) {
190                 printk(KERN_ERR IPATH_DRV_NAME
191                        ": Could not allocate unit ID: error %d\n", -ret);
192                 ipath_free_devdata(pdev, dd);
193                 dd = ERR_PTR(ret);
194                 goto bail_unlock;
195         }
196
197         dd->pcidev = pdev;
198         pci_set_drvdata(pdev, dd);
199
200         list_add(&dd->ipath_list, &ipath_dev_list);
201
202 bail_unlock:
203         spin_unlock_irqrestore(&ipath_devs_lock, flags);
204
205 bail:
206         return dd;
207 }
208
209 static inline struct ipath_devdata *__ipath_lookup(int unit)
210 {
211         return idr_find(&unit_table, unit);
212 }
213
214 struct ipath_devdata *ipath_lookup(int unit)
215 {
216         struct ipath_devdata *dd;
217         unsigned long flags;
218
219         spin_lock_irqsave(&ipath_devs_lock, flags);
220         dd = __ipath_lookup(unit);
221         spin_unlock_irqrestore(&ipath_devs_lock, flags);
222
223         return dd;
224 }
225
226 int ipath_count_units(int *npresentp, int *nupp, u32 *maxportsp)
227 {
228         int nunits, npresent, nup;
229         struct ipath_devdata *dd;
230         unsigned long flags;
231         u32 maxports;
232
233         nunits = npresent = nup = maxports = 0;
234
235         spin_lock_irqsave(&ipath_devs_lock, flags);
236
237         list_for_each_entry(dd, &ipath_dev_list, ipath_list) {
238                 nunits++;
239                 if ((dd->ipath_flags & IPATH_PRESENT) && dd->ipath_kregbase)
240                         npresent++;
241                 if (dd->ipath_lid &&
242                     !(dd->ipath_flags & (IPATH_DISABLED | IPATH_LINKDOWN
243                                          | IPATH_LINKUNK)))
244                         nup++;
245                 if (dd->ipath_cfgports > maxports)
246                         maxports = dd->ipath_cfgports;
247         }
248
249         spin_unlock_irqrestore(&ipath_devs_lock, flags);
250
251         if (npresentp)
252                 *npresentp = npresent;
253         if (nupp)
254                 *nupp = nup;
255         if (maxportsp)
256                 *maxportsp = maxports;
257
258         return nunits;
259 }
260
261 /*
262  * These next two routines are placeholders in case we don't have per-arch
263  * code for controlling write combining.  If explicit control of write
264  * combining is not available, performance will probably be awful.
265  */
266
267 int __attribute__((weak)) ipath_enable_wc(struct ipath_devdata *dd)
268 {
269         return -EOPNOTSUPP;
270 }
271
272 void __attribute__((weak)) ipath_disable_wc(struct ipath_devdata *dd)
273 {
274 }
275
276 static int __devinit ipath_init_one(struct pci_dev *pdev,
277                                     const struct pci_device_id *ent)
278 {
279         int ret, len, j;
280         struct ipath_devdata *dd;
281         unsigned long long addr;
282         u32 bar0 = 0, bar1 = 0;
283         u8 rev;
284
285         dd = ipath_alloc_devdata(pdev);
286         if (IS_ERR(dd)) {
287                 ret = PTR_ERR(dd);
288                 printk(KERN_ERR IPATH_DRV_NAME
289                        ": Could not allocate devdata: error %d\n", -ret);
290                 goto bail;
291         }
292
293         ipath_cdbg(VERBOSE, "initializing unit #%u\n", dd->ipath_unit);
294
295         read_bars(dd, pdev, &bar0, &bar1);
296
297         ret = pci_enable_device(pdev);
298         if (ret) {
299                 /* This can happen iff:
300                  *
301                  * We did a chip reset, and then failed to reprogram the
302                  * BAR, or the chip reset due to an internal error.  We then
303                  * unloaded the driver and reloaded it.
304                  *
305                  * Both reset cases set the BAR back to initial state.  For
306                  * the latter case, the AER sticky error bit at offset 0x718
307                  * should be set, but the Linux kernel doesn't yet know
308                  * about that, it appears.  If the original BAR was retained
309                  * in the kernel data structures, this may be OK.
310                  */
311                 ipath_dev_err(dd, "enable unit %d failed: error %d\n",
312                               dd->ipath_unit, -ret);
313                 goto bail_devdata;
314         }
315         addr = pci_resource_start(pdev, 0);
316         len = pci_resource_len(pdev, 0);
317         ipath_cdbg(VERBOSE, "regbase (0) %llx len %d irq %x, vend %x/%x "
318                    "driver_data %lx\n", addr, len, pdev->irq, ent->vendor,
319                    ent->device, ent->driver_data);
320
321         read_bars(dd, pdev, &bar0, &bar1);
322
323         if (!bar1 && !(bar0 & ~0xf)) {
324                 if (addr) {
325                         dev_info(&pdev->dev, "BAR is 0 (probable RESET), "
326                                  "rewriting as %llx\n", addr);
327                         ret = pci_write_config_dword(
328                                 pdev, PCI_BASE_ADDRESS_0, addr);
329                         if (ret) {
330                                 ipath_dev_err(dd, "rewrite of BAR0 "
331                                               "failed: err %d\n", -ret);
332                                 goto bail_disable;
333                         }
334                         ret = pci_write_config_dword(
335                                 pdev, PCI_BASE_ADDRESS_1, addr >> 32);
336                         if (ret) {
337                                 ipath_dev_err(dd, "rewrite of BAR1 "
338                                               "failed: err %d\n", -ret);
339                                 goto bail_disable;
340                         }
341                 } else {
342                         ipath_dev_err(dd, "BAR is 0 (probable RESET), "
343                                       "not usable until reboot\n");
344                         ret = -ENODEV;
345                         goto bail_disable;
346                 }
347         }
348
349         ret = pci_request_regions(pdev, IPATH_DRV_NAME);
350         if (ret) {
351                 dev_info(&pdev->dev, "pci_request_regions unit %u fails: "
352                          "err %d\n", dd->ipath_unit, -ret);
353                 goto bail_disable;
354         }
355
356         ret = pci_set_dma_mask(pdev, DMA_64BIT_MASK);
357         if (ret) {
358                 /*
359                  * if the 64 bit setup fails, try 32 bit.  Some systems
360                  * do not setup 64 bit maps on systems with 2GB or less
361                  * memory installed.
362                  */
363                 ret = pci_set_dma_mask(pdev, DMA_32BIT_MASK);
364                 if (ret) {
365                         dev_info(&pdev->dev,
366                                 "Unable to set DMA mask for unit %u: %d\n",
367                                 dd->ipath_unit, ret);
368                         goto bail_regions;
369                 }
370                 else {
371                         ipath_dbg("No 64bit DMA mask, used 32 bit mask\n");
372                         ret = pci_set_consistent_dma_mask(pdev, DMA_32BIT_MASK);
373                         if (ret)
374                                 dev_info(&pdev->dev,
375                                         "Unable to set DMA consistent mask "
376                                         "for unit %u: %d\n",
377                                         dd->ipath_unit, ret);
378
379                 }
380         }
381         else {
382                 ret = pci_set_consistent_dma_mask(pdev, DMA_64BIT_MASK);
383                 if (ret)
384                         dev_info(&pdev->dev,
385                                 "Unable to set DMA consistent mask "
386                                 "for unit %u: %d\n",
387                                 dd->ipath_unit, ret);
388         }
389
390         pci_set_master(pdev);
391
392         /*
393          * Save BARs to rewrite after device reset.  Save all 64 bits of
394          * BAR, just in case.
395          */
396         dd->ipath_pcibar0 = addr;
397         dd->ipath_pcibar1 = addr >> 32;
398         dd->ipath_deviceid = ent->device;       /* save for later use */
399         dd->ipath_vendorid = ent->vendor;
400
401         /* setup the chip-specific functions, as early as possible. */
402         switch (ent->device) {
403         case PCI_DEVICE_ID_INFINIPATH_HT:
404                 ipath_init_ht400_funcs(dd);
405                 break;
406         case PCI_DEVICE_ID_INFINIPATH_PE800:
407                 ipath_init_pe800_funcs(dd);
408                 break;
409         default:
410                 ipath_dev_err(dd, "Found unknown QLogic deviceid 0x%x, "
411                               "failing\n", ent->device);
412                 return -ENODEV;
413         }
414
415         for (j = 0; j < 6; j++) {
416                 if (!pdev->resource[j].start)
417                         continue;
418                 ipath_cdbg(VERBOSE, "BAR %d start %llx, end %llx, len %llx\n",
419                            j, (unsigned long long)pdev->resource[j].start,
420                            (unsigned long long)pdev->resource[j].end,
421                            (unsigned long long)pci_resource_len(pdev, j));
422         }
423
424         if (!addr) {
425                 ipath_dev_err(dd, "No valid address in BAR 0!\n");
426                 ret = -ENODEV;
427                 goto bail_regions;
428         }
429
430         dd->ipath_deviceid = ent->device;       /* save for later use */
431         dd->ipath_vendorid = ent->vendor;
432
433         ret = pci_read_config_byte(pdev, PCI_REVISION_ID, &rev);
434         if (ret) {
435                 ipath_dev_err(dd, "Failed to read PCI revision ID unit "
436                               "%u: err %d\n", dd->ipath_unit, -ret);
437                 goto bail_regions;      /* shouldn't ever happen */
438         }
439         dd->ipath_pcirev = rev;
440
441 #if defined(__powerpc__)
442         /* There isn't a generic way to specify writethrough mappings */
443         dd->ipath_kregbase = __ioremap(addr, len,
444                 (_PAGE_NO_CACHE|_PAGE_WRITETHRU));
445 #else
446         dd->ipath_kregbase = ioremap_nocache(addr, len);
447 #endif
448
449         if (!dd->ipath_kregbase) {
450                 ipath_dbg("Unable to map io addr %llx to kvirt, failing\n",
451                           addr);
452                 ret = -ENOMEM;
453                 goto bail_iounmap;
454         }
455         dd->ipath_kregend = (u64 __iomem *)
456                 ((void __iomem *)dd->ipath_kregbase + len);
457         dd->ipath_physaddr = addr;      /* used for io_remap, etc. */
458         /* for user mmap */
459         ipath_cdbg(VERBOSE, "mapped io addr %llx to kregbase %p\n",
460                    addr, dd->ipath_kregbase);
461
462         /*
463          * clear ipath_flags here instead of in ipath_init_chip as it is set
464          * by ipath_setup_htconfig.
465          */
466         dd->ipath_flags = 0;
467         dd->ipath_lli_counter = 0;
468         dd->ipath_lli_errors = 0;
469
470         if (dd->ipath_f_bus(dd, pdev))
471                 ipath_dev_err(dd, "Failed to setup config space; "
472                               "continuing anyway\n");
473
474         /*
475          * set up our interrupt handler; IRQF_SHARED probably not needed,
476          * since MSI interrupts shouldn't be shared but won't  hurt for now.
477          * check 0 irq after we return from chip-specific bus setup, since
478          * that can affect this due to setup
479          */
480         if (!pdev->irq)
481                 ipath_dev_err(dd, "irq is 0, BIOS error?  Interrupts won't "
482                               "work\n");
483         else {
484                 ret = request_irq(pdev->irq, ipath_intr, IRQF_SHARED,
485                                   IPATH_DRV_NAME, dd);
486                 if (ret) {
487                         ipath_dev_err(dd, "Couldn't setup irq handler, "
488                                       "irq=%u: %d\n", pdev->irq, ret);
489                         goto bail_iounmap;
490                 }
491         }
492
493         ret = ipath_init_chip(dd, 0);   /* do the chip-specific init */
494         if (ret)
495                 goto bail_iounmap;
496
497         ret = ipath_enable_wc(dd);
498
499         if (ret) {
500                 ipath_dev_err(dd, "Write combining not enabled "
501                               "(err %d): performance may be poor\n",
502                               -ret);
503                 ret = 0;
504         }
505
506         ipath_device_create_group(&pdev->dev, dd);
507         ipathfs_add_device(dd);
508         ipath_user_add(dd);
509         ipath_diag_add(dd);
510         ipath_register_ib_device(dd);
511
512         goto bail;
513
514 bail_iounmap:
515         iounmap((volatile void __iomem *) dd->ipath_kregbase);
516
517 bail_regions:
518         pci_release_regions(pdev);
519
520 bail_disable:
521         pci_disable_device(pdev);
522
523 bail_devdata:
524         ipath_free_devdata(pdev, dd);
525
526 bail:
527         return ret;
528 }
529
530 static void __devexit ipath_remove_one(struct pci_dev *pdev)
531 {
532         struct ipath_devdata *dd;
533
534         ipath_cdbg(VERBOSE, "removing, pdev=%p\n", pdev);
535         if (!pdev)
536                 return;
537
538         dd = pci_get_drvdata(pdev);
539         ipath_unregister_ib_device(dd->verbs_dev);
540         ipath_diag_remove(dd);
541         ipath_user_remove(dd);
542         ipathfs_remove_device(dd);
543         ipath_device_remove_group(&pdev->dev, dd);
544         ipath_cdbg(VERBOSE, "Releasing pci memory regions, dd %p, "
545                    "unit %u\n", dd, (u32) dd->ipath_unit);
546         if (dd->ipath_kregbase) {
547                 ipath_cdbg(VERBOSE, "Unmapping kregbase %p\n",
548                            dd->ipath_kregbase);
549                 iounmap((volatile void __iomem *) dd->ipath_kregbase);
550                 dd->ipath_kregbase = NULL;
551         }
552         pci_release_regions(pdev);
553         ipath_cdbg(VERBOSE, "calling pci_disable_device\n");
554         pci_disable_device(pdev);
555
556         ipath_free_devdata(pdev, dd);
557 }
558
559 /* general driver use */
560 DEFINE_MUTEX(ipath_mutex);
561
562 static DEFINE_SPINLOCK(ipath_pioavail_lock);
563
564 /**
565  * ipath_disarm_piobufs - cancel a range of PIO buffers
566  * @dd: the infinipath device
567  * @first: the first PIO buffer to cancel
568  * @cnt: the number of PIO buffers to cancel
569  *
570  * cancel a range of PIO buffers, used when they might be armed, but
571  * not triggered.  Used at init to ensure buffer state, and also user
572  * process close, in case it died while writing to a PIO buffer
573  * Also after errors.
574  */
575 void ipath_disarm_piobufs(struct ipath_devdata *dd, unsigned first,
576                           unsigned cnt)
577 {
578         unsigned i, last = first + cnt;
579         u64 sendctrl, sendorig;
580
581         ipath_cdbg(PKT, "disarm %u PIObufs first=%u\n", cnt, first);
582         sendorig = dd->ipath_sendctrl | INFINIPATH_S_DISARM;
583         for (i = first; i < last; i++) {
584                 sendctrl = sendorig |
585                         (i << INFINIPATH_S_DISARMPIOBUF_SHIFT);
586                 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
587                                  sendctrl);
588         }
589
590         /*
591          * Write it again with current value, in case ipath_sendctrl changed
592          * while we were looping; no critical bits that would require
593          * locking.
594          *
595          * Write a 0, and then the original value, reading scratch in
596          * between.  This seems to avoid a chip timing race that causes
597          * pioavail updates to memory to stop.
598          */
599         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
600                          0);
601         sendorig = ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
602         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
603                          dd->ipath_sendctrl);
604 }
605
606 /**
607  * ipath_wait_linkstate - wait for an IB link state change to occur
608  * @dd: the infinipath device
609  * @state: the state to wait for
610  * @msecs: the number of milliseconds to wait
611  *
612  * wait up to msecs milliseconds for IB link state change to occur for
613  * now, take the easy polling route.  Currently used only by
614  * ipath_set_linkstate.  Returns 0 if state reached, otherwise
615  * -ETIMEDOUT state can have multiple states set, for any of several
616  * transitions.
617  */
618 static int ipath_wait_linkstate(struct ipath_devdata *dd, u32 state,
619                                 int msecs)
620 {
621         dd->ipath_state_wanted = state;
622         wait_event_interruptible_timeout(ipath_state_wait,
623                                          (dd->ipath_flags & state),
624                                          msecs_to_jiffies(msecs));
625         dd->ipath_state_wanted = 0;
626
627         if (!(dd->ipath_flags & state)) {
628                 u64 val;
629                 ipath_cdbg(VERBOSE, "Didn't reach linkstate %s within %u"
630                            " ms\n",
631                            /* test INIT ahead of DOWN, both can be set */
632                            (state & IPATH_LINKINIT) ? "INIT" :
633                            ((state & IPATH_LINKDOWN) ? "DOWN" :
634                             ((state & IPATH_LINKARMED) ? "ARM" : "ACTIVE")),
635                            msecs);
636                 val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_ibcstatus);
637                 ipath_cdbg(VERBOSE, "ibcc=%llx ibcstatus=%llx (%s)\n",
638                            (unsigned long long) ipath_read_kreg64(
639                                    dd, dd->ipath_kregs->kr_ibcctrl),
640                            (unsigned long long) val,
641                            ipath_ibcstatus_str[val & 0xf]);
642         }
643         return (dd->ipath_flags & state) ? 0 : -ETIMEDOUT;
644 }
645
646 void ipath_decode_err(char *buf, size_t blen, ipath_err_t err)
647 {
648         *buf = '\0';
649         if (err & INFINIPATH_E_RHDRLEN)
650                 strlcat(buf, "rhdrlen ", blen);
651         if (err & INFINIPATH_E_RBADTID)
652                 strlcat(buf, "rbadtid ", blen);
653         if (err & INFINIPATH_E_RBADVERSION)
654                 strlcat(buf, "rbadversion ", blen);
655         if (err & INFINIPATH_E_RHDR)
656                 strlcat(buf, "rhdr ", blen);
657         if (err & INFINIPATH_E_RLONGPKTLEN)
658                 strlcat(buf, "rlongpktlen ", blen);
659         if (err & INFINIPATH_E_RSHORTPKTLEN)
660                 strlcat(buf, "rshortpktlen ", blen);
661         if (err & INFINIPATH_E_RMAXPKTLEN)
662                 strlcat(buf, "rmaxpktlen ", blen);
663         if (err & INFINIPATH_E_RMINPKTLEN)
664                 strlcat(buf, "rminpktlen ", blen);
665         if (err & INFINIPATH_E_RFORMATERR)
666                 strlcat(buf, "rformaterr ", blen);
667         if (err & INFINIPATH_E_RUNSUPVL)
668                 strlcat(buf, "runsupvl ", blen);
669         if (err & INFINIPATH_E_RUNEXPCHAR)
670                 strlcat(buf, "runexpchar ", blen);
671         if (err & INFINIPATH_E_RIBFLOW)
672                 strlcat(buf, "ribflow ", blen);
673         if (err & INFINIPATH_E_REBP)
674                 strlcat(buf, "EBP ", blen);
675         if (err & INFINIPATH_E_SUNDERRUN)
676                 strlcat(buf, "sunderrun ", blen);
677         if (err & INFINIPATH_E_SPIOARMLAUNCH)
678                 strlcat(buf, "spioarmlaunch ", blen);
679         if (err & INFINIPATH_E_SUNEXPERRPKTNUM)
680                 strlcat(buf, "sunexperrpktnum ", blen);
681         if (err & INFINIPATH_E_SDROPPEDDATAPKT)
682                 strlcat(buf, "sdroppeddatapkt ", blen);
683         if (err & INFINIPATH_E_SDROPPEDSMPPKT)
684                 strlcat(buf, "sdroppedsmppkt ", blen);
685         if (err & INFINIPATH_E_SMAXPKTLEN)
686                 strlcat(buf, "smaxpktlen ", blen);
687         if (err & INFINIPATH_E_SMINPKTLEN)
688                 strlcat(buf, "sminpktlen ", blen);
689         if (err & INFINIPATH_E_SUNSUPVL)
690                 strlcat(buf, "sunsupVL ", blen);
691         if (err & INFINIPATH_E_SPKTLEN)
692                 strlcat(buf, "spktlen ", blen);
693         if (err & INFINIPATH_E_INVALIDADDR)
694                 strlcat(buf, "invalidaddr ", blen);
695         if (err & INFINIPATH_E_RICRC)
696                 strlcat(buf, "CRC ", blen);
697         if (err & INFINIPATH_E_RVCRC)
698                 strlcat(buf, "VCRC ", blen);
699         if (err & INFINIPATH_E_RRCVEGRFULL)
700                 strlcat(buf, "rcvegrfull ", blen);
701         if (err & INFINIPATH_E_RRCVHDRFULL)
702                 strlcat(buf, "rcvhdrfull ", blen);
703         if (err & INFINIPATH_E_IBSTATUSCHANGED)
704                 strlcat(buf, "ibcstatuschg ", blen);
705         if (err & INFINIPATH_E_RIBLOSTLINK)
706                 strlcat(buf, "riblostlink ", blen);
707         if (err & INFINIPATH_E_HARDWARE)
708                 strlcat(buf, "hardware ", blen);
709         if (err & INFINIPATH_E_RESET)
710                 strlcat(buf, "reset ", blen);
711 }
712
713 /**
714  * get_rhf_errstring - decode RHF errors
715  * @err: the err number
716  * @msg: the output buffer
717  * @len: the length of the output buffer
718  *
719  * only used one place now, may want more later
720  */
721 static void get_rhf_errstring(u32 err, char *msg, size_t len)
722 {
723         /* if no errors, and so don't need to check what's first */
724         *msg = '\0';
725
726         if (err & INFINIPATH_RHF_H_ICRCERR)
727                 strlcat(msg, "icrcerr ", len);
728         if (err & INFINIPATH_RHF_H_VCRCERR)
729                 strlcat(msg, "vcrcerr ", len);
730         if (err & INFINIPATH_RHF_H_PARITYERR)
731                 strlcat(msg, "parityerr ", len);
732         if (err & INFINIPATH_RHF_H_LENERR)
733                 strlcat(msg, "lenerr ", len);
734         if (err & INFINIPATH_RHF_H_MTUERR)
735                 strlcat(msg, "mtuerr ", len);
736         if (err & INFINIPATH_RHF_H_IHDRERR)
737                 /* infinipath hdr checksum error */
738                 strlcat(msg, "ipathhdrerr ", len);
739         if (err & INFINIPATH_RHF_H_TIDERR)
740                 strlcat(msg, "tiderr ", len);
741         if (err & INFINIPATH_RHF_H_MKERR)
742                 /* bad port, offset, etc. */
743                 strlcat(msg, "invalid ipathhdr ", len);
744         if (err & INFINIPATH_RHF_H_IBERR)
745                 strlcat(msg, "iberr ", len);
746         if (err & INFINIPATH_RHF_L_SWA)
747                 strlcat(msg, "swA ", len);
748         if (err & INFINIPATH_RHF_L_SWB)
749                 strlcat(msg, "swB ", len);
750 }
751
752 /**
753  * ipath_get_egrbuf - get an eager buffer
754  * @dd: the infinipath device
755  * @bufnum: the eager buffer to get
756  * @err: unused
757  *
758  * must only be called if ipath_pd[port] is known to be allocated
759  */
760 static inline void *ipath_get_egrbuf(struct ipath_devdata *dd, u32 bufnum,
761                                      int err)
762 {
763         return dd->ipath_port0_skbs ?
764                 (void *)dd->ipath_port0_skbs[bufnum]->data : NULL;
765 }
766
767 /**
768  * ipath_alloc_skb - allocate an skb and buffer with possible constraints
769  * @dd: the infinipath device
770  * @gfp_mask: the sk_buff SFP mask
771  */
772 struct sk_buff *ipath_alloc_skb(struct ipath_devdata *dd,
773                                 gfp_t gfp_mask)
774 {
775         struct sk_buff *skb;
776         u32 len;
777
778         /*
779          * Only fully supported way to handle this is to allocate lots
780          * extra, align as needed, and then do skb_reserve().  That wastes
781          * a lot of memory...  I'll have to hack this into infinipath_copy
782          * also.
783          */
784
785         /*
786          * We need 4 extra bytes for unaligned transfer copying
787          */
788         if (dd->ipath_flags & IPATH_4BYTE_TID) {
789                 /* we need a 4KB multiple alignment, and there is no way
790                  * to do it except to allocate extra and then skb_reserve
791                  * enough to bring it up to the right alignment.
792                  */
793                 len = dd->ipath_ibmaxlen + 4 + (1 << 11) - 1;
794         }
795         else
796                 len = dd->ipath_ibmaxlen + 4;
797         skb = __dev_alloc_skb(len, gfp_mask);
798         if (!skb) {
799                 ipath_dev_err(dd, "Failed to allocate skbuff, length %u\n",
800                               len);
801                 goto bail;
802         }
803         if (dd->ipath_flags & IPATH_4BYTE_TID) {
804                 u32 una = ((1 << 11) - 1) & (unsigned long)(skb->data + 4);
805                 if (una)
806                         skb_reserve(skb, 4 + (1 << 11) - una);
807                 else
808                         skb_reserve(skb, 4);
809         } else
810                 skb_reserve(skb, 4);
811
812 bail:
813         return skb;
814 }
815
816 static void ipath_rcv_hdrerr(struct ipath_devdata *dd,
817                              u32 eflags,
818                              u32 l,
819                              u32 etail,
820                              u64 *rc)
821 {
822         char emsg[128];
823         struct ipath_message_header *hdr;
824
825         get_rhf_errstring(eflags, emsg, sizeof emsg);
826         hdr = (struct ipath_message_header *)&rc[1];
827         ipath_cdbg(PKT, "RHFerrs %x hdrqtail=%x typ=%u "
828                    "tlen=%x opcode=%x egridx=%x: %s\n",
829                    eflags, l,
830                    ipath_hdrget_rcv_type((__le32 *) rc),
831                    ipath_hdrget_length_in_bytes((__le32 *) rc),
832                    be32_to_cpu(hdr->bth[0]) >> 24,
833                    etail, emsg);
834
835         /* Count local link integrity errors. */
836         if (eflags & (INFINIPATH_RHF_H_ICRCERR | INFINIPATH_RHF_H_VCRCERR)) {
837                 u8 n = (dd->ipath_ibcctrl >>
838                         INFINIPATH_IBCC_PHYERRTHRESHOLD_SHIFT) &
839                         INFINIPATH_IBCC_PHYERRTHRESHOLD_MASK;
840
841                 if (++dd->ipath_lli_counter > n) {
842                         dd->ipath_lli_counter = 0;
843                         dd->ipath_lli_errors++;
844                 }
845         }
846 }
847
848 /*
849  * ipath_kreceive - receive a packet
850  * @dd: the infinipath device
851  *
852  * called from interrupt handler for errors or receive interrupt
853  */
854 void ipath_kreceive(struct ipath_devdata *dd)
855 {
856         u64 *rc;
857         void *ebuf;
858         const u32 rsize = dd->ipath_rcvhdrentsize;      /* words */
859         const u32 maxcnt = dd->ipath_rcvhdrcnt * rsize; /* words */
860         u32 etail = -1, l, hdrqtail;
861         struct ipath_message_header *hdr;
862         u32 eflags, i, etype, tlen, pkttot = 0, updegr=0, reloop=0;
863         static u64 totcalls;    /* stats, may eventually remove */
864
865         if (!dd->ipath_hdrqtailptr) {
866                 ipath_dev_err(dd,
867                               "hdrqtailptr not set, can't do receives\n");
868                 goto bail;
869         }
870
871         /* There is already a thread processing this queue. */
872         if (test_and_set_bit(0, &dd->ipath_rcv_pending))
873                 goto bail;
874
875         l = dd->ipath_port0head;
876         hdrqtail = (u32) le64_to_cpu(*dd->ipath_hdrqtailptr);
877         if (l == hdrqtail)
878                 goto done;
879
880 reloop:
881         for (i = 0; l != hdrqtail; i++) {
882                 u32 qp;
883                 u8 *bthbytes;
884
885                 rc = (u64 *) (dd->ipath_pd[0]->port_rcvhdrq + (l << 2));
886                 hdr = (struct ipath_message_header *)&rc[1];
887                 /*
888                  * could make a network order version of IPATH_KD_QP, and
889                  * do the obvious shift before masking to speed this up.
890                  */
891                 qp = ntohl(hdr->bth[1]) & 0xffffff;
892                 bthbytes = (u8 *) hdr->bth;
893
894                 eflags = ipath_hdrget_err_flags((__le32 *) rc);
895                 etype = ipath_hdrget_rcv_type((__le32 *) rc);
896                 /* total length */
897                 tlen = ipath_hdrget_length_in_bytes((__le32 *) rc);
898                 ebuf = NULL;
899                 if (etype != RCVHQ_RCV_TYPE_EXPECTED) {
900                         /*
901                          * it turns out that the chips uses an eager buffer
902                          * for all non-expected packets, whether it "needs"
903                          * one or not.  So always get the index, but don't
904                          * set ebuf (so we try to copy data) unless the
905                          * length requires it.
906                          */
907                         etail = ipath_hdrget_index((__le32 *) rc);
908                         if (tlen > sizeof(*hdr) ||
909                             etype == RCVHQ_RCV_TYPE_NON_KD)
910                                 ebuf = ipath_get_egrbuf(dd, etail, 0);
911                 }
912
913                 /*
914                  * both tiderr and ipathhdrerr are set for all plain IB
915                  * packets; only ipathhdrerr should be set.
916                  */
917
918                 if (etype != RCVHQ_RCV_TYPE_NON_KD && etype !=
919                     RCVHQ_RCV_TYPE_ERROR && ipath_hdrget_ipath_ver(
920                             hdr->iph.ver_port_tid_offset) !=
921                     IPS_PROTO_VERSION) {
922                         ipath_cdbg(PKT, "Bad InfiniPath protocol version "
923                                    "%x\n", etype);
924                 }
925
926                 if (unlikely(eflags))
927                         ipath_rcv_hdrerr(dd, eflags, l, etail, rc);
928                 else if (etype == RCVHQ_RCV_TYPE_NON_KD) {
929                         ipath_ib_rcv(dd->verbs_dev, rc + 1, ebuf, tlen);
930                         if (dd->ipath_lli_counter)
931                                 dd->ipath_lli_counter--;
932                         ipath_cdbg(PKT, "typ %x, opcode %x (eager, "
933                                    "qp=%x), len %x; ignored\n",
934                                    etype, bthbytes[0], qp, tlen);
935                 }
936                 else if (etype == RCVHQ_RCV_TYPE_EAGER)
937                         ipath_cdbg(PKT, "typ %x, opcode %x (eager, "
938                                    "qp=%x), len %x; ignored\n",
939                                    etype, bthbytes[0], qp, tlen);
940                 else if (etype == RCVHQ_RCV_TYPE_EXPECTED)
941                         ipath_dbg("Bug: Expected TID, opcode %x; ignored\n",
942                                   be32_to_cpu(hdr->bth[0]) & 0xff);
943                 else {
944                         /*
945                          * error packet, type of error  unknown.
946                          * Probably type 3, but we don't know, so don't
947                          * even try to print the opcode, etc.
948                          */
949                         ipath_dbg("Error Pkt, but no eflags! egrbuf %x, "
950                                   "len %x\nhdrq@%lx;hdrq+%x rhf: %llx; "
951                                   "hdr %llx %llx %llx %llx %llx\n",
952                                   etail, tlen, (unsigned long) rc, l,
953                                   (unsigned long long) rc[0],
954                                   (unsigned long long) rc[1],
955                                   (unsigned long long) rc[2],
956                                   (unsigned long long) rc[3],
957                                   (unsigned long long) rc[4],
958                                   (unsigned long long) rc[5]);
959                 }
960                 l += rsize;
961                 if (l >= maxcnt)
962                         l = 0;
963                 if (etype != RCVHQ_RCV_TYPE_EXPECTED)
964                     updegr = 1;
965                 /*
966                  * update head regs on last packet, and every 16 packets.
967                  * Reduce bus traffic, while still trying to prevent
968                  * rcvhdrq overflows, for when the queue is nearly full
969                  */
970                 if (l == hdrqtail || (i && !(i&0xf))) {
971                         u64 lval;
972                         if (l == hdrqtail) /* PE-800 interrupt only on last */
973                                 lval = dd->ipath_rhdrhead_intr_off | l;
974                         else
975                                 lval = l;
976                         (void)ipath_write_ureg(dd, ur_rcvhdrhead, lval, 0);
977                         if (updegr) {
978                                 (void)ipath_write_ureg(dd, ur_rcvegrindexhead,
979                                                        etail, 0);
980                                 updegr = 0;
981                         }
982                 }
983         }
984
985         if (!dd->ipath_rhdrhead_intr_off && !reloop) {
986                 /* HT-400 workaround; we can have a race clearing chip
987                  * interrupt with another interrupt about to be delivered,
988                  * and can clear it before it is delivered on the GPIO
989                  * workaround.  By doing the extra check here for the
990                  * in-memory tail register updating while we were doing
991                  * earlier packets, we "almost" guarantee we have covered
992                  * that case.
993                  */
994                 u32 hqtail = (u32)le64_to_cpu(*dd->ipath_hdrqtailptr);
995                 if (hqtail != hdrqtail) {
996                         hdrqtail = hqtail;
997                         reloop = 1; /* loop 1 extra time at most */
998                         goto reloop;
999                 }
1000         }
1001
1002         pkttot += i;
1003
1004         dd->ipath_port0head = l;
1005
1006         if (pkttot > ipath_stats.sps_maxpkts_call)
1007                 ipath_stats.sps_maxpkts_call = pkttot;
1008         ipath_stats.sps_port0pkts += pkttot;
1009         ipath_stats.sps_avgpkts_call =
1010                 ipath_stats.sps_port0pkts / ++totcalls;
1011
1012 done:
1013         clear_bit(0, &dd->ipath_rcv_pending);
1014         smp_mb__after_clear_bit();
1015
1016 bail:;
1017 }
1018
1019 /**
1020  * ipath_update_pio_bufs - update shadow copy of the PIO availability map
1021  * @dd: the infinipath device
1022  *
1023  * called whenever our local copy indicates we have run out of send buffers
1024  * NOTE: This can be called from interrupt context by some code
1025  * and from non-interrupt context by ipath_getpiobuf().
1026  */
1027
1028 static void ipath_update_pio_bufs(struct ipath_devdata *dd)
1029 {
1030         unsigned long flags;
1031         int i;
1032         const unsigned piobregs = (unsigned)dd->ipath_pioavregs;
1033
1034         /* If the generation (check) bits have changed, then we update the
1035          * busy bit for the corresponding PIO buffer.  This algorithm will
1036          * modify positions to the value they already have in some cases
1037          * (i.e., no change), but it's faster than changing only the bits
1038          * that have changed.
1039          *
1040          * We would like to do this atomicly, to avoid spinlocks in the
1041          * critical send path, but that's not really possible, given the
1042          * type of changes, and that this routine could be called on
1043          * multiple cpu's simultaneously, so we lock in this routine only,
1044          * to avoid conflicting updates; all we change is the shadow, and
1045          * it's a single 64 bit memory location, so by definition the update
1046          * is atomic in terms of what other cpu's can see in testing the
1047          * bits.  The spin_lock overhead isn't too bad, since it only
1048          * happens when all buffers are in use, so only cpu overhead, not
1049          * latency or bandwidth is affected.
1050          */
1051 #define _IPATH_ALL_CHECKBITS 0x5555555555555555ULL
1052         if (!dd->ipath_pioavailregs_dma) {
1053                 ipath_dbg("Update shadow pioavail, but regs_dma NULL!\n");
1054                 return;
1055         }
1056         if (ipath_debug & __IPATH_VERBDBG) {
1057                 /* only if packet debug and verbose */
1058                 volatile __le64 *dma = dd->ipath_pioavailregs_dma;
1059                 unsigned long *shadow = dd->ipath_pioavailshadow;
1060
1061                 ipath_cdbg(PKT, "Refill avail, dma0=%llx shad0=%lx, "
1062                            "d1=%llx s1=%lx, d2=%llx s2=%lx, d3=%llx "
1063                            "s3=%lx\n",
1064                            (unsigned long long) le64_to_cpu(dma[0]),
1065                            shadow[0],
1066                            (unsigned long long) le64_to_cpu(dma[1]),
1067                            shadow[1],
1068                            (unsigned long long) le64_to_cpu(dma[2]),
1069                            shadow[2],
1070                            (unsigned long long) le64_to_cpu(dma[3]),
1071                            shadow[3]);
1072                 if (piobregs > 4)
1073                         ipath_cdbg(
1074                                 PKT, "2nd group, dma4=%llx shad4=%lx, "
1075                                 "d5=%llx s5=%lx, d6=%llx s6=%lx, "
1076                                 "d7=%llx s7=%lx\n",
1077                                 (unsigned long long) le64_to_cpu(dma[4]),
1078                                 shadow[4],
1079                                 (unsigned long long) le64_to_cpu(dma[5]),
1080                                 shadow[5],
1081                                 (unsigned long long) le64_to_cpu(dma[6]),
1082                                 shadow[6],
1083                                 (unsigned long long) le64_to_cpu(dma[7]),
1084                                 shadow[7]);
1085         }
1086         spin_lock_irqsave(&ipath_pioavail_lock, flags);
1087         for (i = 0; i < piobregs; i++) {
1088                 u64 pchbusy, pchg, piov, pnew;
1089                 /*
1090                  * Chip Errata: bug 6641; even and odd qwords>3 are swapped
1091                  */
1092                 if (i > 3) {
1093                         if (i & 1)
1094                                 piov = le64_to_cpu(
1095                                         dd->ipath_pioavailregs_dma[i - 1]);
1096                         else
1097                                 piov = le64_to_cpu(
1098                                         dd->ipath_pioavailregs_dma[i + 1]);
1099                 } else
1100                         piov = le64_to_cpu(dd->ipath_pioavailregs_dma[i]);
1101                 pchg = _IPATH_ALL_CHECKBITS &
1102                         ~(dd->ipath_pioavailshadow[i] ^ piov);
1103                 pchbusy = pchg << INFINIPATH_SENDPIOAVAIL_BUSY_SHIFT;
1104                 if (pchg && (pchbusy & dd->ipath_pioavailshadow[i])) {
1105                         pnew = dd->ipath_pioavailshadow[i] & ~pchbusy;
1106                         pnew |= piov & pchbusy;
1107                         dd->ipath_pioavailshadow[i] = pnew;
1108                 }
1109         }
1110         spin_unlock_irqrestore(&ipath_pioavail_lock, flags);
1111 }
1112
1113 /**
1114  * ipath_setrcvhdrsize - set the receive header size
1115  * @dd: the infinipath device
1116  * @rhdrsize: the receive header size
1117  *
1118  * called from user init code, and also layered driver init
1119  */
1120 int ipath_setrcvhdrsize(struct ipath_devdata *dd, unsigned rhdrsize)
1121 {
1122         int ret = 0;
1123
1124         if (dd->ipath_flags & IPATH_RCVHDRSZ_SET) {
1125                 if (dd->ipath_rcvhdrsize != rhdrsize) {
1126                         dev_info(&dd->pcidev->dev,
1127                                  "Error: can't set protocol header "
1128                                  "size %u, already %u\n",
1129                                  rhdrsize, dd->ipath_rcvhdrsize);
1130                         ret = -EAGAIN;
1131                 } else
1132                         ipath_cdbg(VERBOSE, "Reuse same protocol header "
1133                                    "size %u\n", dd->ipath_rcvhdrsize);
1134         } else if (rhdrsize > (dd->ipath_rcvhdrentsize -
1135                                (sizeof(u64) / sizeof(u32)))) {
1136                 ipath_dbg("Error: can't set protocol header size %u "
1137                           "(> max %u)\n", rhdrsize,
1138                           dd->ipath_rcvhdrentsize -
1139                           (u32) (sizeof(u64) / sizeof(u32)));
1140                 ret = -EOVERFLOW;
1141         } else {
1142                 dd->ipath_flags |= IPATH_RCVHDRSZ_SET;
1143                 dd->ipath_rcvhdrsize = rhdrsize;
1144                 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvhdrsize,
1145                                  dd->ipath_rcvhdrsize);
1146                 ipath_cdbg(VERBOSE, "Set protocol header size to %u\n",
1147                            dd->ipath_rcvhdrsize);
1148         }
1149         return ret;
1150 }
1151
1152 /**
1153  * ipath_getpiobuf - find an available pio buffer
1154  * @dd: the infinipath device
1155  * @pbufnum: the buffer number is placed here
1156  *
1157  * do appropriate marking as busy, etc.
1158  * returns buffer number if one found (>=0), negative number is error.
1159  * Used by ipath_layer_send
1160  */
1161 u32 __iomem *ipath_getpiobuf(struct ipath_devdata *dd, u32 * pbufnum)
1162 {
1163         int i, j, starti, updated = 0;
1164         unsigned piobcnt, iter;
1165         unsigned long flags;
1166         unsigned long *shadow = dd->ipath_pioavailshadow;
1167         u32 __iomem *buf;
1168
1169         piobcnt = (unsigned)(dd->ipath_piobcnt2k
1170                              + dd->ipath_piobcnt4k);
1171         starti = dd->ipath_lastport_piobuf;
1172         iter = piobcnt - starti;
1173         if (dd->ipath_upd_pio_shadow) {
1174                 /*
1175                  * Minor optimization.  If we had no buffers on last call,
1176                  * start out by doing the update; continue and do scan even
1177                  * if no buffers were updated, to be paranoid
1178                  */
1179                 ipath_update_pio_bufs(dd);
1180                 /* we scanned here, don't do it at end of scan */
1181                 updated = 1;
1182                 i = starti;
1183         } else
1184                 i = dd->ipath_lastpioindex;
1185
1186 rescan:
1187         /*
1188          * while test_and_set_bit() is atomic, we do that and then the
1189          * change_bit(), and the pair is not.  See if this is the cause
1190          * of the remaining armlaunch errors.
1191          */
1192         spin_lock_irqsave(&ipath_pioavail_lock, flags);
1193         for (j = 0; j < iter; j++, i++) {
1194                 if (i >= piobcnt)
1195                         i = starti;
1196                 /*
1197                  * To avoid bus lock overhead, we first find a candidate
1198                  * buffer, then do the test and set, and continue if that
1199                  * fails.
1200                  */
1201                 if (test_bit((2 * i) + 1, shadow) ||
1202                     test_and_set_bit((2 * i) + 1, shadow))
1203                         continue;
1204                 /* flip generation bit */
1205                 change_bit(2 * i, shadow);
1206                 break;
1207         }
1208         spin_unlock_irqrestore(&ipath_pioavail_lock, flags);
1209
1210         if (j == iter) {
1211                 volatile __le64 *dma = dd->ipath_pioavailregs_dma;
1212
1213                 /*
1214                  * first time through; shadow exhausted, but may be real
1215                  * buffers available, so go see; if any updated, rescan
1216                  * (once)
1217                  */
1218                 if (!updated) {
1219                         ipath_update_pio_bufs(dd);
1220                         updated = 1;
1221                         i = starti;
1222                         goto rescan;
1223                 }
1224                 dd->ipath_upd_pio_shadow = 1;
1225                 /*
1226                  * not atomic, but if we lose one once in a while, that's OK
1227                  */
1228                 ipath_stats.sps_nopiobufs++;
1229                 if (!(++dd->ipath_consec_nopiobuf % 100000)) {
1230                         ipath_dbg(
1231                                 "%u pio sends with no bufavail; dmacopy: "
1232                                 "%llx %llx %llx %llx; shadow:  "
1233                                 "%lx %lx %lx %lx\n",
1234                                 dd->ipath_consec_nopiobuf,
1235                                 (unsigned long long) le64_to_cpu(dma[0]),
1236                                 (unsigned long long) le64_to_cpu(dma[1]),
1237                                 (unsigned long long) le64_to_cpu(dma[2]),
1238                                 (unsigned long long) le64_to_cpu(dma[3]),
1239                                 shadow[0], shadow[1], shadow[2],
1240                                 shadow[3]);
1241                         /*
1242                          * 4 buffers per byte, 4 registers above, cover rest
1243                          * below
1244                          */
1245                         if ((dd->ipath_piobcnt2k + dd->ipath_piobcnt4k) >
1246                             (sizeof(shadow[0]) * 4 * 4))
1247                                 ipath_dbg("2nd group: dmacopy: %llx %llx "
1248                                           "%llx %llx; shadow: %lx %lx "
1249                                           "%lx %lx\n",
1250                                           (unsigned long long)
1251                                           le64_to_cpu(dma[4]),
1252                                           (unsigned long long)
1253                                           le64_to_cpu(dma[5]),
1254                                           (unsigned long long)
1255                                           le64_to_cpu(dma[6]),
1256                                           (unsigned long long)
1257                                           le64_to_cpu(dma[7]),
1258                                           shadow[4], shadow[5],
1259                                           shadow[6], shadow[7]);
1260                 }
1261                 buf = NULL;
1262                 goto bail;
1263         }
1264
1265         /*
1266          * set next starting place.  Since it's just an optimization,
1267          * it doesn't matter who wins on this, so no locking
1268          */
1269         dd->ipath_lastpioindex = i + 1;
1270         if (dd->ipath_upd_pio_shadow)
1271                 dd->ipath_upd_pio_shadow = 0;
1272         if (dd->ipath_consec_nopiobuf)
1273                 dd->ipath_consec_nopiobuf = 0;
1274         if (i < dd->ipath_piobcnt2k)
1275                 buf = (u32 __iomem *) (dd->ipath_pio2kbase +
1276                                        i * dd->ipath_palign);
1277         else
1278                 buf = (u32 __iomem *)
1279                         (dd->ipath_pio4kbase +
1280                          (i - dd->ipath_piobcnt2k) * dd->ipath_4kalign);
1281         ipath_cdbg(VERBOSE, "Return piobuf%u %uk @ %p\n",
1282                    i, (i < dd->ipath_piobcnt2k) ? 2 : 4, buf);
1283         if (pbufnum)
1284                 *pbufnum = i;
1285
1286 bail:
1287         return buf;
1288 }
1289
1290 /**
1291  * ipath_create_rcvhdrq - create a receive header queue
1292  * @dd: the infinipath device
1293  * @pd: the port data
1294  *
1295  * this must be contiguous memory (from an i/o perspective), and must be
1296  * DMA'able (which means for some systems, it will go through an IOMMU,
1297  * or be forced into a low address range).
1298  */
1299 int ipath_create_rcvhdrq(struct ipath_devdata *dd,
1300                          struct ipath_portdata *pd)
1301 {
1302         int ret = 0;
1303
1304         if (!pd->port_rcvhdrq) {
1305                 dma_addr_t phys_hdrqtail;
1306                 gfp_t gfp_flags = GFP_USER | __GFP_COMP;
1307                 int amt = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
1308                                 sizeof(u32), PAGE_SIZE);
1309
1310                 pd->port_rcvhdrq = dma_alloc_coherent(
1311                         &dd->pcidev->dev, amt, &pd->port_rcvhdrq_phys,
1312                         gfp_flags);
1313
1314                 if (!pd->port_rcvhdrq) {
1315                         ipath_dev_err(dd, "attempt to allocate %d bytes "
1316                                       "for port %u rcvhdrq failed\n",
1317                                       amt, pd->port_port);
1318                         ret = -ENOMEM;
1319                         goto bail;
1320                 }
1321                 pd->port_rcvhdrtail_kvaddr = dma_alloc_coherent(
1322                         &dd->pcidev->dev, PAGE_SIZE, &phys_hdrqtail, GFP_KERNEL);
1323                 if (!pd->port_rcvhdrtail_kvaddr) {
1324                         ipath_dev_err(dd, "attempt to allocate 1 page "
1325                                       "for port %u rcvhdrqtailaddr failed\n",
1326                                       pd->port_port);
1327                         ret = -ENOMEM;
1328                         goto bail;
1329                 }
1330                 pd->port_rcvhdrqtailaddr_phys = phys_hdrqtail;
1331
1332                 pd->port_rcvhdrq_size = amt;
1333
1334                 ipath_cdbg(VERBOSE, "%d pages at %p (phys %lx) size=%lu "
1335                            "for port %u rcvhdr Q\n",
1336                            amt >> PAGE_SHIFT, pd->port_rcvhdrq,
1337                            (unsigned long) pd->port_rcvhdrq_phys,
1338                            (unsigned long) pd->port_rcvhdrq_size,
1339                            pd->port_port);
1340
1341                 ipath_cdbg(VERBOSE, "port %d hdrtailaddr, %llx physical\n",
1342                            pd->port_port,
1343                            (unsigned long long) phys_hdrqtail);
1344         }
1345         else
1346                 ipath_cdbg(VERBOSE, "reuse port %d rcvhdrq @%p %llx phys; "
1347                            "hdrtailaddr@%p %llx physical\n",
1348                            pd->port_port, pd->port_rcvhdrq,
1349                            pd->port_rcvhdrq_phys, pd->port_rcvhdrtail_kvaddr,
1350                            (unsigned long long)pd->port_rcvhdrqtailaddr_phys);
1351
1352         /* clear for security and sanity on each use */
1353         memset(pd->port_rcvhdrq, 0, pd->port_rcvhdrq_size);
1354         memset((void *)pd->port_rcvhdrtail_kvaddr, 0, PAGE_SIZE);
1355
1356         /*
1357          * tell chip each time we init it, even if we are re-using previous
1358          * memory (we zero the register at process close)
1359          */
1360         ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdrtailaddr,
1361                               pd->port_port, pd->port_rcvhdrqtailaddr_phys);
1362         ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdraddr,
1363                               pd->port_port, pd->port_rcvhdrq_phys);
1364
1365         ret = 0;
1366 bail:
1367         return ret;
1368 }
1369
1370 int ipath_waitfor_complete(struct ipath_devdata *dd, ipath_kreg reg_id,
1371                            u64 bits_to_wait_for, u64 * valp)
1372 {
1373         unsigned long timeout;
1374         u64 lastval, val;
1375         int ret;
1376
1377         lastval = ipath_read_kreg64(dd, reg_id);
1378         /* wait a ridiculously long time */
1379         timeout = jiffies + msecs_to_jiffies(5);
1380         do {
1381                 val = ipath_read_kreg64(dd, reg_id);
1382                 /* set so they have something, even on failures. */
1383                 *valp = val;
1384                 if ((val & bits_to_wait_for) == bits_to_wait_for) {
1385                         ret = 0;
1386                         break;
1387                 }
1388                 if (val != lastval)
1389                         ipath_cdbg(VERBOSE, "Changed from %llx to %llx, "
1390                                    "waiting for %llx bits\n",
1391                                    (unsigned long long) lastval,
1392                                    (unsigned long long) val,
1393                                    (unsigned long long) bits_to_wait_for);
1394                 cond_resched();
1395                 if (time_after(jiffies, timeout)) {
1396                         ipath_dbg("Didn't get bits %llx in register 0x%x, "
1397                                   "got %llx\n",
1398                                   (unsigned long long) bits_to_wait_for,
1399                                   reg_id, (unsigned long long) *valp);
1400                         ret = -ENODEV;
1401                         break;
1402                 }
1403         } while (1);
1404
1405         return ret;
1406 }
1407
1408 /**
1409  * ipath_waitfor_mdio_cmdready - wait for last command to complete
1410  * @dd: the infinipath device
1411  *
1412  * Like ipath_waitfor_complete(), but we wait for the CMDVALID bit to go
1413  * away indicating the last command has completed.  It doesn't return data
1414  */
1415 int ipath_waitfor_mdio_cmdready(struct ipath_devdata *dd)
1416 {
1417         unsigned long timeout;
1418         u64 val;
1419         int ret;
1420
1421         /* wait a ridiculously long time */
1422         timeout = jiffies + msecs_to_jiffies(5);
1423         do {
1424                 val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_mdio);
1425                 if (!(val & IPATH_MDIO_CMDVALID)) {
1426                         ret = 0;
1427                         break;
1428                 }
1429                 cond_resched();
1430                 if (time_after(jiffies, timeout)) {
1431                         ipath_dbg("CMDVALID stuck in mdio reg? (%llx)\n",
1432                                   (unsigned long long) val);
1433                         ret = -ENODEV;
1434                         break;
1435                 }
1436         } while (1);
1437
1438         return ret;
1439 }
1440
1441 static void ipath_set_ib_lstate(struct ipath_devdata *dd, int which)
1442 {
1443         static const char *what[4] = {
1444                 [0] = "DOWN",
1445                 [INFINIPATH_IBCC_LINKCMD_INIT] = "INIT",
1446                 [INFINIPATH_IBCC_LINKCMD_ARMED] = "ARMED",
1447                 [INFINIPATH_IBCC_LINKCMD_ACTIVE] = "ACTIVE"
1448         };
1449         int linkcmd = (which >> INFINIPATH_IBCC_LINKCMD_SHIFT) &
1450                         INFINIPATH_IBCC_LINKCMD_MASK;
1451
1452         ipath_cdbg(VERBOSE, "Trying to move unit %u to %s, current ltstate "
1453                    "is %s\n", dd->ipath_unit,
1454                    what[linkcmd],
1455                    ipath_ibcstatus_str[
1456                            (ipath_read_kreg64
1457                             (dd, dd->ipath_kregs->kr_ibcstatus) >>
1458                             INFINIPATH_IBCS_LINKTRAININGSTATE_SHIFT) &
1459                            INFINIPATH_IBCS_LINKTRAININGSTATE_MASK]);
1460         /* flush all queued sends when going to DOWN or INIT, to be sure that
1461          * they don't block MAD packets */
1462         if (!linkcmd || linkcmd == INFINIPATH_IBCC_LINKCMD_INIT) {
1463                 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
1464                                  INFINIPATH_S_ABORT);
1465                 ipath_disarm_piobufs(dd, dd->ipath_lastport_piobuf,
1466                                     (unsigned)(dd->ipath_piobcnt2k +
1467                                     dd->ipath_piobcnt4k) -
1468                                     dd->ipath_lastport_piobuf);
1469         }
1470
1471         ipath_write_kreg(dd, dd->ipath_kregs->kr_ibcctrl,
1472                          dd->ipath_ibcctrl | which);
1473 }
1474
1475 int ipath_set_linkstate(struct ipath_devdata *dd, u8 newstate)
1476 {
1477         u32 lstate;
1478         int ret;
1479
1480         switch (newstate) {
1481         case IPATH_IB_LINKDOWN:
1482                 ipath_set_ib_lstate(dd, INFINIPATH_IBCC_LINKINITCMD_POLL <<
1483                                     INFINIPATH_IBCC_LINKINITCMD_SHIFT);
1484                 /* don't wait */
1485                 ret = 0;
1486                 goto bail;
1487
1488         case IPATH_IB_LINKDOWN_SLEEP:
1489                 ipath_set_ib_lstate(dd, INFINIPATH_IBCC_LINKINITCMD_SLEEP <<
1490                                     INFINIPATH_IBCC_LINKINITCMD_SHIFT);
1491                 /* don't wait */
1492                 ret = 0;
1493                 goto bail;
1494
1495         case IPATH_IB_LINKDOWN_DISABLE:
1496                 ipath_set_ib_lstate(dd,
1497                                     INFINIPATH_IBCC_LINKINITCMD_DISABLE <<
1498                                     INFINIPATH_IBCC_LINKINITCMD_SHIFT);
1499                 /* don't wait */
1500                 ret = 0;
1501                 goto bail;
1502
1503         case IPATH_IB_LINKINIT:
1504                 if (dd->ipath_flags & IPATH_LINKINIT) {
1505                         ret = 0;
1506                         goto bail;
1507                 }
1508                 ipath_set_ib_lstate(dd, INFINIPATH_IBCC_LINKCMD_INIT <<
1509                                     INFINIPATH_IBCC_LINKCMD_SHIFT);
1510                 lstate = IPATH_LINKINIT;
1511                 break;
1512
1513         case IPATH_IB_LINKARM:
1514                 if (dd->ipath_flags & IPATH_LINKARMED) {
1515                         ret = 0;
1516                         goto bail;
1517                 }
1518                 if (!(dd->ipath_flags &
1519                       (IPATH_LINKINIT | IPATH_LINKACTIVE))) {
1520                         ret = -EINVAL;
1521                         goto bail;
1522                 }
1523                 ipath_set_ib_lstate(dd, INFINIPATH_IBCC_LINKCMD_ARMED <<
1524                                     INFINIPATH_IBCC_LINKCMD_SHIFT);
1525                 /*
1526                  * Since the port can transition to ACTIVE by receiving
1527                  * a non VL 15 packet, wait for either state.
1528                  */
1529                 lstate = IPATH_LINKARMED | IPATH_LINKACTIVE;
1530                 break;
1531
1532         case IPATH_IB_LINKACTIVE:
1533                 if (dd->ipath_flags & IPATH_LINKACTIVE) {
1534                         ret = 0;
1535                         goto bail;
1536                 }
1537                 if (!(dd->ipath_flags & IPATH_LINKARMED)) {
1538                         ret = -EINVAL;
1539                         goto bail;
1540                 }
1541                 ipath_set_ib_lstate(dd, INFINIPATH_IBCC_LINKCMD_ACTIVE <<
1542                                     INFINIPATH_IBCC_LINKCMD_SHIFT);
1543                 lstate = IPATH_LINKACTIVE;
1544                 break;
1545
1546         default:
1547                 ipath_dbg("Invalid linkstate 0x%x requested\n", newstate);
1548                 ret = -EINVAL;
1549                 goto bail;
1550         }
1551         ret = ipath_wait_linkstate(dd, lstate, 2000);
1552
1553 bail:
1554         return ret;
1555 }
1556
1557 /**
1558  * ipath_set_mtu - set the MTU
1559  * @dd: the infinipath device
1560  * @arg: the new MTU
1561  *
1562  * we can handle "any" incoming size, the issue here is whether we
1563  * need to restrict our outgoing size.   For now, we don't do any
1564  * sanity checking on this, and we don't deal with what happens to
1565  * programs that are already running when the size changes.
1566  * NOTE: changing the MTU will usually cause the IBC to go back to
1567  * link initialize (IPATH_IBSTATE_INIT) state...
1568  */
1569 int ipath_set_mtu(struct ipath_devdata *dd, u16 arg)
1570 {
1571         u32 piosize;
1572         int changed = 0;
1573         int ret;
1574
1575         /*
1576          * mtu is IB data payload max.  It's the largest power of 2 less
1577          * than piosize (or even larger, since it only really controls the
1578          * largest we can receive; we can send the max of the mtu and
1579          * piosize).  We check that it's one of the valid IB sizes.
1580          */
1581         if (arg != 256 && arg != 512 && arg != 1024 && arg != 2048 &&
1582             arg != 4096) {
1583                 ipath_dbg("Trying to set invalid mtu %u, failing\n", arg);
1584                 ret = -EINVAL;
1585                 goto bail;
1586         }
1587         if (dd->ipath_ibmtu == arg) {
1588                 ret = 0;        /* same as current */
1589                 goto bail;
1590         }
1591
1592         piosize = dd->ipath_ibmaxlen;
1593         dd->ipath_ibmtu = arg;
1594
1595         if (arg >= (piosize - IPATH_PIO_MAXIBHDR)) {
1596                 /* Only if it's not the initial value (or reset to it) */
1597                 if (piosize != dd->ipath_init_ibmaxlen) {
1598                         dd->ipath_ibmaxlen = piosize;
1599                         changed = 1;
1600                 }
1601         } else if ((arg + IPATH_PIO_MAXIBHDR) != dd->ipath_ibmaxlen) {
1602                 piosize = arg + IPATH_PIO_MAXIBHDR;
1603                 ipath_cdbg(VERBOSE, "ibmaxlen was 0x%x, setting to 0x%x "
1604                            "(mtu 0x%x)\n", dd->ipath_ibmaxlen, piosize,
1605                            arg);
1606                 dd->ipath_ibmaxlen = piosize;
1607                 changed = 1;
1608         }
1609
1610         if (changed) {
1611                 /*
1612                  * set the IBC maxpktlength to the size of our pio
1613                  * buffers in words
1614                  */
1615                 u64 ibc = dd->ipath_ibcctrl;
1616                 ibc &= ~(INFINIPATH_IBCC_MAXPKTLEN_MASK <<
1617                          INFINIPATH_IBCC_MAXPKTLEN_SHIFT);
1618
1619                 piosize = piosize - 2 * sizeof(u32);    /* ignore pbc */
1620                 dd->ipath_ibmaxlen = piosize;
1621                 piosize /= sizeof(u32); /* in words */
1622                 /*
1623                  * for ICRC, which we only send in diag test pkt mode, and
1624                  * we don't need to worry about that for mtu
1625                  */
1626                 piosize += 1;
1627
1628                 ibc |= piosize << INFINIPATH_IBCC_MAXPKTLEN_SHIFT;
1629                 dd->ipath_ibcctrl = ibc;
1630                 ipath_write_kreg(dd, dd->ipath_kregs->kr_ibcctrl,
1631                                  dd->ipath_ibcctrl);
1632                 dd->ipath_f_tidtemplate(dd);
1633         }
1634
1635         ret = 0;
1636
1637 bail:
1638         return ret;
1639 }
1640
1641 int ipath_set_lid(struct ipath_devdata *dd, u32 arg, u8 lmc)
1642 {
1643         dd->ipath_lid = arg;
1644         dd->ipath_lmc = lmc;
1645
1646         return 0;
1647 }
1648
1649 /**
1650  * ipath_read_kreg64_port - read a device's per-port 64-bit kernel register
1651  * @dd: the infinipath device
1652  * @regno: the register number to read
1653  * @port: the port containing the register
1654  *
1655  * Registers that vary with the chip implementation constants (port)
1656  * use this routine.
1657  */
1658 u64 ipath_read_kreg64_port(const struct ipath_devdata *dd, ipath_kreg regno,
1659                            unsigned port)
1660 {
1661         u16 where;
1662
1663         if (port < dd->ipath_portcnt &&
1664             (regno == dd->ipath_kregs->kr_rcvhdraddr ||
1665              regno == dd->ipath_kregs->kr_rcvhdrtailaddr))
1666                 where = regno + port;
1667         else
1668                 where = -1;
1669
1670         return ipath_read_kreg64(dd, where);
1671 }
1672
1673 /**
1674  * ipath_write_kreg_port - write a device's per-port 64-bit kernel register
1675  * @dd: the infinipath device
1676  * @regno: the register number to write
1677  * @port: the port containing the register
1678  * @value: the value to write
1679  *
1680  * Registers that vary with the chip implementation constants (port)
1681  * use this routine.
1682  */
1683 void ipath_write_kreg_port(const struct ipath_devdata *dd, ipath_kreg regno,
1684                           unsigned port, u64 value)
1685 {
1686         u16 where;
1687
1688         if (port < dd->ipath_portcnt &&
1689             (regno == dd->ipath_kregs->kr_rcvhdraddr ||
1690              regno == dd->ipath_kregs->kr_rcvhdrtailaddr))
1691                 where = regno + port;
1692         else
1693                 where = -1;
1694
1695         ipath_write_kreg(dd, where, value);
1696 }
1697
1698 /**
1699  * ipath_shutdown_device - shut down a device
1700  * @dd: the infinipath device
1701  *
1702  * This is called to make the device quiet when we are about to
1703  * unload the driver, and also when the device is administratively
1704  * disabled.   It does not free any data structures.
1705  * Everything it does has to be setup again by ipath_init_chip(dd,1)
1706  */
1707 void ipath_shutdown_device(struct ipath_devdata *dd)
1708 {
1709         u64 val;
1710
1711         ipath_dbg("Shutting down the device\n");
1712
1713         dd->ipath_flags |= IPATH_LINKUNK;
1714         dd->ipath_flags &= ~(IPATH_INITTED | IPATH_LINKDOWN |
1715                              IPATH_LINKINIT | IPATH_LINKARMED |
1716                              IPATH_LINKACTIVE);
1717         *dd->ipath_statusp &= ~(IPATH_STATUS_IB_CONF |
1718                                 IPATH_STATUS_IB_READY);
1719
1720         /* mask interrupts, but not errors */
1721         ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask, 0ULL);
1722
1723         dd->ipath_rcvctrl = 0;
1724         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1725                          dd->ipath_rcvctrl);
1726
1727         /*
1728          * gracefully stop all sends allowing any in progress to trickle out
1729          * first.
1730          */
1731         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, 0ULL);
1732         /* flush it */
1733         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
1734         /*
1735          * enough for anything that's going to trickle out to have actually
1736          * done so.
1737          */
1738         udelay(5);
1739
1740         /*
1741          * abort any armed or launched PIO buffers that didn't go. (self
1742          * clearing).  Will cause any packet currently being transmitted to
1743          * go out with an EBP, and may also cause a short packet error on
1744          * the receiver.
1745          */
1746         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
1747                          INFINIPATH_S_ABORT);
1748
1749         ipath_set_ib_lstate(dd, INFINIPATH_IBCC_LINKINITCMD_DISABLE <<
1750                             INFINIPATH_IBCC_LINKINITCMD_SHIFT);
1751
1752         /* disable IBC */
1753         dd->ipath_control &= ~INFINIPATH_C_LINKENABLE;
1754         ipath_write_kreg(dd, dd->ipath_kregs->kr_control,
1755                          dd->ipath_control | INFINIPATH_C_FREEZEMODE);
1756
1757         /*
1758          * clear SerdesEnable and turn the leds off; do this here because
1759          * we are unloading, so don't count on interrupts to move along
1760          * Turn the LEDs off explictly for the same reason.
1761          */
1762         dd->ipath_f_quiet_serdes(dd);
1763         dd->ipath_f_setextled(dd, 0, 0);
1764
1765         if (dd->ipath_stats_timer_active) {
1766                 del_timer_sync(&dd->ipath_stats_timer);
1767                 dd->ipath_stats_timer_active = 0;
1768         }
1769
1770         /*
1771          * clear all interrupts and errors, so that the next time the driver
1772          * is loaded or device is enabled, we know that whatever is set
1773          * happened while we were unloaded
1774          */
1775         ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear,
1776                          ~0ULL & ~INFINIPATH_HWE_MEMBISTFAILED);
1777         ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear, -1LL);
1778         ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, -1LL);
1779 }
1780
1781 /**
1782  * ipath_free_pddata - free a port's allocated data
1783  * @dd: the infinipath device
1784  * @pd: the portdata structure
1785  *
1786  * free up any allocated data for a port
1787  * This should not touch anything that would affect a simultaneous
1788  * re-allocation of port data, because it is called after ipath_mutex
1789  * is released (and can be called from reinit as well).
1790  * It should never change any chip state, or global driver state.
1791  * (The only exception to global state is freeing the port0 port0_skbs.)
1792  */
1793 void ipath_free_pddata(struct ipath_devdata *dd, struct ipath_portdata *pd)
1794 {
1795         if (!pd)
1796                 return;
1797
1798         if (pd->port_rcvhdrq) {
1799                 ipath_cdbg(VERBOSE, "free closed port %d rcvhdrq @ %p "
1800                            "(size=%lu)\n", pd->port_port, pd->port_rcvhdrq,
1801                            (unsigned long) pd->port_rcvhdrq_size);
1802                 dma_free_coherent(&dd->pcidev->dev, pd->port_rcvhdrq_size,
1803                                   pd->port_rcvhdrq, pd->port_rcvhdrq_phys);
1804                 pd->port_rcvhdrq = NULL;
1805                 if (pd->port_rcvhdrtail_kvaddr) {
1806                         dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE,
1807                                          (void *)pd->port_rcvhdrtail_kvaddr,
1808                                          pd->port_rcvhdrqtailaddr_phys);
1809                         pd->port_rcvhdrtail_kvaddr = NULL;
1810                 }
1811         }
1812         if (pd->port_port && pd->port_rcvegrbuf) {
1813                 unsigned e;
1814
1815                 for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
1816                         void *base = pd->port_rcvegrbuf[e];
1817                         size_t size = pd->port_rcvegrbuf_size;
1818
1819                         ipath_cdbg(VERBOSE, "egrbuf free(%p, %lu), "
1820                                    "chunk %u/%u\n", base,
1821                                    (unsigned long) size,
1822                                    e, pd->port_rcvegrbuf_chunks);
1823                         dma_free_coherent(&dd->pcidev->dev, size,
1824                                 base, pd->port_rcvegrbuf_phys[e]);
1825                 }
1826                 vfree(pd->port_rcvegrbuf);
1827                 pd->port_rcvegrbuf = NULL;
1828                 vfree(pd->port_rcvegrbuf_phys);
1829                 pd->port_rcvegrbuf_phys = NULL;
1830                 pd->port_rcvegrbuf_chunks = 0;
1831         } else if (pd->port_port == 0 && dd->ipath_port0_skbs) {
1832                 unsigned e;
1833                 struct sk_buff **skbs = dd->ipath_port0_skbs;
1834
1835                 dd->ipath_port0_skbs = NULL;
1836                 ipath_cdbg(VERBOSE, "free closed port %d ipath_port0_skbs "
1837                            "@ %p\n", pd->port_port, skbs);
1838                 for (e = 0; e < dd->ipath_rcvegrcnt; e++)
1839                         if (skbs[e])
1840                                 dev_kfree_skb(skbs[e]);
1841                 vfree(skbs);
1842         }
1843         kfree(pd->port_tid_pg_list);
1844         kfree(pd);
1845 }
1846
1847 static int __init infinipath_init(void)
1848 {
1849         int ret;
1850
1851         ipath_dbg(KERN_INFO DRIVER_LOAD_MSG "%s", ib_ipath_version);
1852
1853         /*
1854          * These must be called before the driver is registered with
1855          * the PCI subsystem.
1856          */
1857         idr_init(&unit_table);
1858         if (!idr_pre_get(&unit_table, GFP_KERNEL)) {
1859                 ret = -ENOMEM;
1860                 goto bail;
1861         }
1862
1863         ret = pci_register_driver(&ipath_driver);
1864         if (ret < 0) {
1865                 printk(KERN_ERR IPATH_DRV_NAME
1866                        ": Unable to register driver: error %d\n", -ret);
1867                 goto bail_unit;
1868         }
1869
1870         ret = ipath_driver_create_group(&ipath_driver.driver);
1871         if (ret < 0) {
1872                 printk(KERN_ERR IPATH_DRV_NAME ": Unable to create driver "
1873                        "sysfs entries: error %d\n", -ret);
1874                 goto bail_pci;
1875         }
1876
1877         ret = ipath_init_ipathfs();
1878         if (ret < 0) {
1879                 printk(KERN_ERR IPATH_DRV_NAME ": Unable to create "
1880                        "ipathfs: error %d\n", -ret);
1881                 goto bail_group;
1882         }
1883
1884         ret = ipath_diagpkt_add();
1885         if (ret < 0) {
1886                 printk(KERN_ERR IPATH_DRV_NAME ": Unable to create "
1887                        "diag data device: error %d\n", -ret);
1888                 goto bail_ipathfs;
1889         }
1890
1891         goto bail;
1892
1893 bail_ipathfs:
1894         ipath_exit_ipathfs();
1895
1896 bail_group:
1897         ipath_driver_remove_group(&ipath_driver.driver);
1898
1899 bail_pci:
1900         pci_unregister_driver(&ipath_driver);
1901
1902 bail_unit:
1903         idr_destroy(&unit_table);
1904
1905 bail:
1906         return ret;
1907 }
1908
1909 static void cleanup_device(struct ipath_devdata *dd)
1910 {
1911         int port;
1912
1913         ipath_shutdown_device(dd);
1914
1915         if (*dd->ipath_statusp & IPATH_STATUS_CHIP_PRESENT) {
1916                 /* can't do anything more with chip; needs re-init */
1917                 *dd->ipath_statusp &= ~IPATH_STATUS_CHIP_PRESENT;
1918                 if (dd->ipath_kregbase) {
1919                         /*
1920                          * if we haven't already cleaned up before these are
1921                          * to ensure any register reads/writes "fail" until
1922                          * re-init
1923                          */
1924                         dd->ipath_kregbase = NULL;
1925                         dd->ipath_uregbase = 0;
1926                         dd->ipath_sregbase = 0;
1927                         dd->ipath_cregbase = 0;
1928                         dd->ipath_kregsize = 0;
1929                 }
1930                 ipath_disable_wc(dd);
1931         }
1932
1933         if (dd->ipath_pioavailregs_dma) {
1934                 dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE,
1935                                   (void *) dd->ipath_pioavailregs_dma,
1936                                   dd->ipath_pioavailregs_phys);
1937                 dd->ipath_pioavailregs_dma = NULL;
1938         }
1939         if (dd->ipath_dummy_hdrq) {
1940                 dma_free_coherent(&dd->pcidev->dev,
1941                         dd->ipath_pd[0]->port_rcvhdrq_size,
1942                         dd->ipath_dummy_hdrq, dd->ipath_dummy_hdrq_phys);
1943                 dd->ipath_dummy_hdrq = NULL;
1944         }
1945
1946         if (dd->ipath_pageshadow) {
1947                 struct page **tmpp = dd->ipath_pageshadow;
1948                 int i, cnt = 0;
1949
1950                 ipath_cdbg(VERBOSE, "Unlocking any expTID pages still "
1951                            "locked\n");
1952                 for (port = 0; port < dd->ipath_cfgports; port++) {
1953                         int port_tidbase = port * dd->ipath_rcvtidcnt;
1954                         int maxtid = port_tidbase + dd->ipath_rcvtidcnt;
1955                         for (i = port_tidbase; i < maxtid; i++) {
1956                                 if (!tmpp[i])
1957                                         continue;
1958                                 ipath_release_user_pages(&tmpp[i], 1);
1959                                 tmpp[i] = NULL;
1960                                 cnt++;
1961                         }
1962                 }
1963                 if (cnt) {
1964                         ipath_stats.sps_pageunlocks += cnt;
1965                         ipath_cdbg(VERBOSE, "There were still %u expTID "
1966                                    "entries locked\n", cnt);
1967                 }
1968                 if (ipath_stats.sps_pagelocks ||
1969                     ipath_stats.sps_pageunlocks)
1970                         ipath_cdbg(VERBOSE, "%llu pages locked, %llu "
1971                                    "unlocked via ipath_m{un}lock\n",
1972                                    (unsigned long long)
1973                                    ipath_stats.sps_pagelocks,
1974                                    (unsigned long long)
1975                                    ipath_stats.sps_pageunlocks);
1976
1977                 ipath_cdbg(VERBOSE, "Free shadow page tid array at %p\n",
1978                            dd->ipath_pageshadow);
1979                 vfree(dd->ipath_pageshadow);
1980                 dd->ipath_pageshadow = NULL;
1981         }
1982
1983         /*
1984          * free any resources still in use (usually just kernel ports)
1985          * at unload; we do for portcnt, not cfgports, because cfgports
1986          * could have changed while we were loaded.
1987          */
1988         for (port = 0; port < dd->ipath_portcnt; port++) {
1989                 struct ipath_portdata *pd = dd->ipath_pd[port];
1990                 dd->ipath_pd[port] = NULL;
1991                 ipath_free_pddata(dd, pd);
1992         }
1993         kfree(dd->ipath_pd);
1994         /*
1995          * debuggability, in case some cleanup path tries to use it
1996          * after this
1997          */
1998         dd->ipath_pd = NULL;
1999 }
2000
2001 static void __exit infinipath_cleanup(void)
2002 {
2003         struct ipath_devdata *dd, *tmp;
2004         unsigned long flags;
2005
2006         ipath_diagpkt_remove();
2007
2008         ipath_exit_ipathfs();
2009
2010         ipath_driver_remove_group(&ipath_driver.driver);
2011
2012         spin_lock_irqsave(&ipath_devs_lock, flags);
2013
2014         /*
2015          * turn off rcv, send, and interrupts for all ports, all drivers
2016          * should also hard reset the chip here?
2017          * free up port 0 (kernel) rcvhdr, egr bufs, and eventually tid bufs
2018          * for all versions of the driver, if they were allocated
2019          */
2020         list_for_each_entry_safe(dd, tmp, &ipath_dev_list, ipath_list) {
2021                 spin_unlock_irqrestore(&ipath_devs_lock, flags);
2022
2023                 if (dd->ipath_kregbase)
2024                         cleanup_device(dd);
2025
2026                 if (dd->pcidev) {
2027                         if (dd->pcidev->irq) {
2028                                 ipath_cdbg(VERBOSE,
2029                                            "unit %u free_irq of irq %x\n",
2030                                            dd->ipath_unit, dd->pcidev->irq);
2031                                 free_irq(dd->pcidev->irq, dd);
2032                         } else
2033                                 ipath_dbg("irq is 0, not doing free_irq "
2034                                           "for unit %u\n", dd->ipath_unit);
2035
2036                         /*
2037                          * we check for NULL here, because it's outside
2038                          * the kregbase check, and we need to call it
2039                          * after the free_irq.  Thus it's possible that
2040                          * the function pointers were never initialized.
2041                          */
2042                         if (dd->ipath_f_cleanup)
2043                                 /* clean up chip-specific stuff */
2044                                 dd->ipath_f_cleanup(dd);
2045
2046                         dd->pcidev = NULL;
2047                 }
2048                 spin_lock_irqsave(&ipath_devs_lock, flags);
2049         }
2050
2051         spin_unlock_irqrestore(&ipath_devs_lock, flags);
2052
2053         ipath_cdbg(VERBOSE, "Unregistering pci driver\n");
2054         pci_unregister_driver(&ipath_driver);
2055
2056         idr_destroy(&unit_table);
2057 }
2058
2059 /**
2060  * ipath_reset_device - reset the chip if possible
2061  * @unit: the device to reset
2062  *
2063  * Whether or not reset is successful, we attempt to re-initialize the chip
2064  * (that is, much like a driver unload/reload).  We clear the INITTED flag
2065  * so that the various entry points will fail until we reinitialize.  For
2066  * now, we only allow this if no user ports are open that use chip resources
2067  */
2068 int ipath_reset_device(int unit)
2069 {
2070         int ret, i;
2071         struct ipath_devdata *dd = ipath_lookup(unit);
2072
2073         if (!dd) {
2074                 ret = -ENODEV;
2075                 goto bail;
2076         }
2077
2078         dev_info(&dd->pcidev->dev, "Reset on unit %u requested\n", unit);
2079
2080         if (!dd->ipath_kregbase || !(dd->ipath_flags & IPATH_PRESENT)) {
2081                 dev_info(&dd->pcidev->dev, "Invalid unit number %u or "
2082                          "not initialized or not present\n", unit);
2083                 ret = -ENXIO;
2084                 goto bail;
2085         }
2086
2087         if (dd->ipath_pd)
2088                 for (i = 1; i < dd->ipath_cfgports; i++) {
2089                         if (dd->ipath_pd[i] && dd->ipath_pd[i]->port_cnt) {
2090                                 ipath_dbg("unit %u port %d is in use "
2091                                           "(PID %u cmd %s), can't reset\n",
2092                                           unit, i,
2093                                           dd->ipath_pd[i]->port_pid,
2094                                           dd->ipath_pd[i]->port_comm);
2095                                 ret = -EBUSY;
2096                                 goto bail;
2097                         }
2098                 }
2099
2100         dd->ipath_flags &= ~IPATH_INITTED;
2101         ret = dd->ipath_f_reset(dd);
2102         if (ret != 1)
2103                 ipath_dbg("reset was not successful\n");
2104         ipath_dbg("Trying to reinitialize unit %u after reset attempt\n",
2105                   unit);
2106         ret = ipath_init_chip(dd, 1);
2107         if (ret)
2108                 ipath_dev_err(dd, "Reinitialize unit %u after "
2109                               "reset failed with %d\n", unit, ret);
2110         else
2111                 dev_info(&dd->pcidev->dev, "Reinitialized unit %u after "
2112                          "resetting\n", unit);
2113
2114 bail:
2115         return ret;
2116 }
2117
2118 module_init(infinipath_init);
2119 module_exit(infinipath_cleanup);