RDMA: Use vzalloc() to replace vmalloc()+memset(0)
[linux-3.10.git] / drivers / infiniband / hw / ipath / ipath_driver.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/sched.h>
35 #include <linux/spinlock.h>
36 #include <linux/idr.h>
37 #include <linux/pci.h>
38 #include <linux/io.h>
39 #include <linux/delay.h>
40 #include <linux/netdevice.h>
41 #include <linux/vmalloc.h>
42 #include <linux/bitmap.h>
43 #include <linux/slab.h>
44
45 #include "ipath_kernel.h"
46 #include "ipath_verbs.h"
47
48 static void ipath_update_pio_bufs(struct ipath_devdata *);
49
50 const char *ipath_get_unit_name(int unit)
51 {
52         static char iname[16];
53         snprintf(iname, sizeof iname, "infinipath%u", unit);
54         return iname;
55 }
56
57 #define DRIVER_LOAD_MSG "QLogic " IPATH_DRV_NAME " loaded: "
58 #define PFX IPATH_DRV_NAME ": "
59
60 /*
61  * The size has to be longer than this string, so we can append
62  * board/chip information to it in the init code.
63  */
64 const char ib_ipath_version[] = IPATH_IDSTR "\n";
65
66 static struct idr unit_table;
67 DEFINE_SPINLOCK(ipath_devs_lock);
68 LIST_HEAD(ipath_dev_list);
69
70 wait_queue_head_t ipath_state_wait;
71
72 unsigned ipath_debug = __IPATH_INFO;
73
74 module_param_named(debug, ipath_debug, uint, S_IWUSR | S_IRUGO);
75 MODULE_PARM_DESC(debug, "mask for debug prints");
76 EXPORT_SYMBOL_GPL(ipath_debug);
77
78 unsigned ipath_mtu4096 = 1; /* max 4KB IB mtu by default, if supported */
79 module_param_named(mtu4096, ipath_mtu4096, uint, S_IRUGO);
80 MODULE_PARM_DESC(mtu4096, "enable MTU of 4096 bytes, if supported");
81
82 static unsigned ipath_hol_timeout_ms = 13000;
83 module_param_named(hol_timeout_ms, ipath_hol_timeout_ms, uint, S_IRUGO);
84 MODULE_PARM_DESC(hol_timeout_ms,
85         "duration of user app suspension after link failure");
86
87 unsigned ipath_linkrecovery = 1;
88 module_param_named(linkrecovery, ipath_linkrecovery, uint, S_IWUSR | S_IRUGO);
89 MODULE_PARM_DESC(linkrecovery, "enable workaround for link recovery issue");
90
91 MODULE_LICENSE("GPL");
92 MODULE_AUTHOR("QLogic <support@qlogic.com>");
93 MODULE_DESCRIPTION("QLogic InfiniPath driver");
94
95 /*
96  * Table to translate the LINKTRAININGSTATE portion of
97  * IBCStatus to a human-readable form.
98  */
99 const char *ipath_ibcstatus_str[] = {
100         "Disabled",
101         "LinkUp",
102         "PollActive",
103         "PollQuiet",
104         "SleepDelay",
105         "SleepQuiet",
106         "LState6",              /* unused */
107         "LState7",              /* unused */
108         "CfgDebounce",
109         "CfgRcvfCfg",
110         "CfgWaitRmt",
111         "CfgIdle",
112         "RecovRetrain",
113         "CfgTxRevLane",         /* unused before IBA7220 */
114         "RecovWaitRmt",
115         "RecovIdle",
116         /* below were added for IBA7220 */
117         "CfgEnhanced",
118         "CfgTest",
119         "CfgWaitRmtTest",
120         "CfgWaitCfgEnhanced",
121         "SendTS_T",
122         "SendTstIdles",
123         "RcvTS_T",
124         "SendTst_TS1s",
125         "LTState18", "LTState19", "LTState1A", "LTState1B",
126         "LTState1C", "LTState1D", "LTState1E", "LTState1F"
127 };
128
129 static void __devexit ipath_remove_one(struct pci_dev *);
130 static int __devinit ipath_init_one(struct pci_dev *,
131                                     const struct pci_device_id *);
132
133 /* Only needed for registration, nothing else needs this info */
134 #define PCI_VENDOR_ID_PATHSCALE 0x1fc1
135 #define PCI_DEVICE_ID_INFINIPATH_HT 0xd
136
137 /* Number of seconds before our card status check...  */
138 #define STATUS_TIMEOUT 60
139
140 static const struct pci_device_id ipath_pci_tbl[] = {
141         { PCI_DEVICE(PCI_VENDOR_ID_PATHSCALE, PCI_DEVICE_ID_INFINIPATH_HT) },
142         { 0, }
143 };
144
145 MODULE_DEVICE_TABLE(pci, ipath_pci_tbl);
146
147 static struct pci_driver ipath_driver = {
148         .name = IPATH_DRV_NAME,
149         .probe = ipath_init_one,
150         .remove = __devexit_p(ipath_remove_one),
151         .id_table = ipath_pci_tbl,
152         .driver = {
153                 .groups = ipath_driver_attr_groups,
154         },
155 };
156
157 static inline void read_bars(struct ipath_devdata *dd, struct pci_dev *dev,
158                              u32 *bar0, u32 *bar1)
159 {
160         int ret;
161
162         ret = pci_read_config_dword(dev, PCI_BASE_ADDRESS_0, bar0);
163         if (ret)
164                 ipath_dev_err(dd, "failed to read bar0 before enable: "
165                               "error %d\n", -ret);
166
167         ret = pci_read_config_dword(dev, PCI_BASE_ADDRESS_1, bar1);
168         if (ret)
169                 ipath_dev_err(dd, "failed to read bar1 before enable: "
170                               "error %d\n", -ret);
171
172         ipath_dbg("Read bar0 %x bar1 %x\n", *bar0, *bar1);
173 }
174
175 static void ipath_free_devdata(struct pci_dev *pdev,
176                                struct ipath_devdata *dd)
177 {
178         unsigned long flags;
179
180         pci_set_drvdata(pdev, NULL);
181
182         if (dd->ipath_unit != -1) {
183                 spin_lock_irqsave(&ipath_devs_lock, flags);
184                 idr_remove(&unit_table, dd->ipath_unit);
185                 list_del(&dd->ipath_list);
186                 spin_unlock_irqrestore(&ipath_devs_lock, flags);
187         }
188         vfree(dd);
189 }
190
191 static struct ipath_devdata *ipath_alloc_devdata(struct pci_dev *pdev)
192 {
193         unsigned long flags;
194         struct ipath_devdata *dd;
195         int ret;
196
197         if (!idr_pre_get(&unit_table, GFP_KERNEL)) {
198                 dd = ERR_PTR(-ENOMEM);
199                 goto bail;
200         }
201
202         dd = vzalloc(sizeof(*dd));
203         if (!dd) {
204                 dd = ERR_PTR(-ENOMEM);
205                 goto bail;
206         }
207         dd->ipath_unit = -1;
208
209         spin_lock_irqsave(&ipath_devs_lock, flags);
210
211         ret = idr_get_new(&unit_table, dd, &dd->ipath_unit);
212         if (ret < 0) {
213                 printk(KERN_ERR IPATH_DRV_NAME
214                        ": Could not allocate unit ID: error %d\n", -ret);
215                 ipath_free_devdata(pdev, dd);
216                 dd = ERR_PTR(ret);
217                 goto bail_unlock;
218         }
219
220         dd->pcidev = pdev;
221         pci_set_drvdata(pdev, dd);
222
223         list_add(&dd->ipath_list, &ipath_dev_list);
224
225 bail_unlock:
226         spin_unlock_irqrestore(&ipath_devs_lock, flags);
227
228 bail:
229         return dd;
230 }
231
232 static inline struct ipath_devdata *__ipath_lookup(int unit)
233 {
234         return idr_find(&unit_table, unit);
235 }
236
237 struct ipath_devdata *ipath_lookup(int unit)
238 {
239         struct ipath_devdata *dd;
240         unsigned long flags;
241
242         spin_lock_irqsave(&ipath_devs_lock, flags);
243         dd = __ipath_lookup(unit);
244         spin_unlock_irqrestore(&ipath_devs_lock, flags);
245
246         return dd;
247 }
248
249 int ipath_count_units(int *npresentp, int *nupp, int *maxportsp)
250 {
251         int nunits, npresent, nup;
252         struct ipath_devdata *dd;
253         unsigned long flags;
254         int maxports;
255
256         nunits = npresent = nup = maxports = 0;
257
258         spin_lock_irqsave(&ipath_devs_lock, flags);
259
260         list_for_each_entry(dd, &ipath_dev_list, ipath_list) {
261                 nunits++;
262                 if ((dd->ipath_flags & IPATH_PRESENT) && dd->ipath_kregbase)
263                         npresent++;
264                 if (dd->ipath_lid &&
265                     !(dd->ipath_flags & (IPATH_DISABLED | IPATH_LINKDOWN
266                                          | IPATH_LINKUNK)))
267                         nup++;
268                 if (dd->ipath_cfgports > maxports)
269                         maxports = dd->ipath_cfgports;
270         }
271
272         spin_unlock_irqrestore(&ipath_devs_lock, flags);
273
274         if (npresentp)
275                 *npresentp = npresent;
276         if (nupp)
277                 *nupp = nup;
278         if (maxportsp)
279                 *maxportsp = maxports;
280
281         return nunits;
282 }
283
284 /*
285  * These next two routines are placeholders in case we don't have per-arch
286  * code for controlling write combining.  If explicit control of write
287  * combining is not available, performance will probably be awful.
288  */
289
290 int __attribute__((weak)) ipath_enable_wc(struct ipath_devdata *dd)
291 {
292         return -EOPNOTSUPP;
293 }
294
295 void __attribute__((weak)) ipath_disable_wc(struct ipath_devdata *dd)
296 {
297 }
298
299 /*
300  * Perform a PIO buffer bandwidth write test, to verify proper system
301  * configuration.  Even when all the setup calls work, occasionally
302  * BIOS or other issues can prevent write combining from working, or
303  * can cause other bandwidth problems to the chip.
304  *
305  * This test simply writes the same buffer over and over again, and
306  * measures close to the peak bandwidth to the chip (not testing
307  * data bandwidth to the wire).   On chips that use an address-based
308  * trigger to send packets to the wire, this is easy.  On chips that
309  * use a count to trigger, we want to make sure that the packet doesn't
310  * go out on the wire, or trigger flow control checks.
311  */
312 static void ipath_verify_pioperf(struct ipath_devdata *dd)
313 {
314         u32 pbnum, cnt, lcnt;
315         u32 __iomem *piobuf;
316         u32 *addr;
317         u64 msecs, emsecs;
318
319         piobuf = ipath_getpiobuf(dd, 0, &pbnum);
320         if (!piobuf) {
321                 dev_info(&dd->pcidev->dev,
322                         "No PIObufs for checking perf, skipping\n");
323                 return;
324         }
325
326         /*
327          * Enough to give us a reasonable test, less than piobuf size, and
328          * likely multiple of store buffer length.
329          */
330         cnt = 1024;
331
332         addr = vmalloc(cnt);
333         if (!addr) {
334                 dev_info(&dd->pcidev->dev,
335                         "Couldn't get memory for checking PIO perf,"
336                         " skipping\n");
337                 goto done;
338         }
339
340         preempt_disable();  /* we want reasonably accurate elapsed time */
341         msecs = 1 + jiffies_to_msecs(jiffies);
342         for (lcnt = 0; lcnt < 10000U; lcnt++) {
343                 /* wait until we cross msec boundary */
344                 if (jiffies_to_msecs(jiffies) >= msecs)
345                         break;
346                 udelay(1);
347         }
348
349         ipath_disable_armlaunch(dd);
350
351         /*
352          * length 0, no dwords actually sent, and mark as VL15
353          * on chips where that may matter (due to IB flowcontrol)
354          */
355         if ((dd->ipath_flags & IPATH_HAS_PBC_CNT))
356                 writeq(1UL << 63, piobuf);
357         else
358                 writeq(0, piobuf);
359         ipath_flush_wc();
360
361         /*
362          * this is only roughly accurate, since even with preempt we
363          * still take interrupts that could take a while.   Running for
364          * >= 5 msec seems to get us "close enough" to accurate values
365          */
366         msecs = jiffies_to_msecs(jiffies);
367         for (emsecs = lcnt = 0; emsecs <= 5UL; lcnt++) {
368                 __iowrite32_copy(piobuf + 64, addr, cnt >> 2);
369                 emsecs = jiffies_to_msecs(jiffies) - msecs;
370         }
371
372         /* 1 GiB/sec, slightly over IB SDR line rate */
373         if (lcnt < (emsecs * 1024U))
374                 ipath_dev_err(dd,
375                         "Performance problem: bandwidth to PIO buffers is "
376                         "only %u MiB/sec\n",
377                         lcnt / (u32) emsecs);
378         else
379                 ipath_dbg("PIO buffer bandwidth %u MiB/sec is OK\n",
380                         lcnt / (u32) emsecs);
381
382         preempt_enable();
383
384         vfree(addr);
385
386 done:
387         /* disarm piobuf, so it's available again */
388         ipath_disarm_piobufs(dd, pbnum, 1);
389         ipath_enable_armlaunch(dd);
390 }
391
392 static void cleanup_device(struct ipath_devdata *dd);
393
394 static int __devinit ipath_init_one(struct pci_dev *pdev,
395                                     const struct pci_device_id *ent)
396 {
397         int ret, len, j;
398         struct ipath_devdata *dd;
399         unsigned long long addr;
400         u32 bar0 = 0, bar1 = 0;
401         u8 rev;
402
403         dd = ipath_alloc_devdata(pdev);
404         if (IS_ERR(dd)) {
405                 ret = PTR_ERR(dd);
406                 printk(KERN_ERR IPATH_DRV_NAME
407                        ": Could not allocate devdata: error %d\n", -ret);
408                 goto bail;
409         }
410
411         ipath_cdbg(VERBOSE, "initializing unit #%u\n", dd->ipath_unit);
412
413         ret = pci_enable_device(pdev);
414         if (ret) {
415                 /* This can happen iff:
416                  *
417                  * We did a chip reset, and then failed to reprogram the
418                  * BAR, or the chip reset due to an internal error.  We then
419                  * unloaded the driver and reloaded it.
420                  *
421                  * Both reset cases set the BAR back to initial state.  For
422                  * the latter case, the AER sticky error bit at offset 0x718
423                  * should be set, but the Linux kernel doesn't yet know
424                  * about that, it appears.  If the original BAR was retained
425                  * in the kernel data structures, this may be OK.
426                  */
427                 ipath_dev_err(dd, "enable unit %d failed: error %d\n",
428                               dd->ipath_unit, -ret);
429                 goto bail_devdata;
430         }
431         addr = pci_resource_start(pdev, 0);
432         len = pci_resource_len(pdev, 0);
433         ipath_cdbg(VERBOSE, "regbase (0) %llx len %d irq %d, vend %x/%x "
434                    "driver_data %lx\n", addr, len, pdev->irq, ent->vendor,
435                    ent->device, ent->driver_data);
436
437         read_bars(dd, pdev, &bar0, &bar1);
438
439         if (!bar1 && !(bar0 & ~0xf)) {
440                 if (addr) {
441                         dev_info(&pdev->dev, "BAR is 0 (probable RESET), "
442                                  "rewriting as %llx\n", addr);
443                         ret = pci_write_config_dword(
444                                 pdev, PCI_BASE_ADDRESS_0, addr);
445                         if (ret) {
446                                 ipath_dev_err(dd, "rewrite of BAR0 "
447                                               "failed: err %d\n", -ret);
448                                 goto bail_disable;
449                         }
450                         ret = pci_write_config_dword(
451                                 pdev, PCI_BASE_ADDRESS_1, addr >> 32);
452                         if (ret) {
453                                 ipath_dev_err(dd, "rewrite of BAR1 "
454                                               "failed: err %d\n", -ret);
455                                 goto bail_disable;
456                         }
457                 } else {
458                         ipath_dev_err(dd, "BAR is 0 (probable RESET), "
459                                       "not usable until reboot\n");
460                         ret = -ENODEV;
461                         goto bail_disable;
462                 }
463         }
464
465         ret = pci_request_regions(pdev, IPATH_DRV_NAME);
466         if (ret) {
467                 dev_info(&pdev->dev, "pci_request_regions unit %u fails: "
468                          "err %d\n", dd->ipath_unit, -ret);
469                 goto bail_disable;
470         }
471
472         ret = pci_set_dma_mask(pdev, DMA_BIT_MASK(64));
473         if (ret) {
474                 /*
475                  * if the 64 bit setup fails, try 32 bit.  Some systems
476                  * do not setup 64 bit maps on systems with 2GB or less
477                  * memory installed.
478                  */
479                 ret = pci_set_dma_mask(pdev, DMA_BIT_MASK(32));
480                 if (ret) {
481                         dev_info(&pdev->dev,
482                                 "Unable to set DMA mask for unit %u: %d\n",
483                                 dd->ipath_unit, ret);
484                         goto bail_regions;
485                 }
486                 else {
487                         ipath_dbg("No 64bit DMA mask, used 32 bit mask\n");
488                         ret = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(32));
489                         if (ret)
490                                 dev_info(&pdev->dev,
491                                         "Unable to set DMA consistent mask "
492                                         "for unit %u: %d\n",
493                                         dd->ipath_unit, ret);
494
495                 }
496         }
497         else {
498                 ret = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(64));
499                 if (ret)
500                         dev_info(&pdev->dev,
501                                 "Unable to set DMA consistent mask "
502                                 "for unit %u: %d\n",
503                                 dd->ipath_unit, ret);
504         }
505
506         pci_set_master(pdev);
507
508         /*
509          * Save BARs to rewrite after device reset.  Save all 64 bits of
510          * BAR, just in case.
511          */
512         dd->ipath_pcibar0 = addr;
513         dd->ipath_pcibar1 = addr >> 32;
514         dd->ipath_deviceid = ent->device;       /* save for later use */
515         dd->ipath_vendorid = ent->vendor;
516
517         /* setup the chip-specific functions, as early as possible. */
518         switch (ent->device) {
519         case PCI_DEVICE_ID_INFINIPATH_HT:
520                 ipath_init_iba6110_funcs(dd);
521                 break;
522
523         default:
524                 ipath_dev_err(dd, "Found unknown QLogic deviceid 0x%x, "
525                               "failing\n", ent->device);
526                 return -ENODEV;
527         }
528
529         for (j = 0; j < 6; j++) {
530                 if (!pdev->resource[j].start)
531                         continue;
532                 ipath_cdbg(VERBOSE, "BAR %d %pR, len %llx\n",
533                            j, &pdev->resource[j],
534                            (unsigned long long)pci_resource_len(pdev, j));
535         }
536
537         if (!addr) {
538                 ipath_dev_err(dd, "No valid address in BAR 0!\n");
539                 ret = -ENODEV;
540                 goto bail_regions;
541         }
542
543         ret = pci_read_config_byte(pdev, PCI_REVISION_ID, &rev);
544         if (ret) {
545                 ipath_dev_err(dd, "Failed to read PCI revision ID unit "
546                               "%u: err %d\n", dd->ipath_unit, -ret);
547                 goto bail_regions;      /* shouldn't ever happen */
548         }
549         dd->ipath_pcirev = rev;
550
551 #if defined(__powerpc__)
552         /* There isn't a generic way to specify writethrough mappings */
553         dd->ipath_kregbase = __ioremap(addr, len,
554                 (_PAGE_NO_CACHE|_PAGE_WRITETHRU));
555 #else
556         dd->ipath_kregbase = ioremap_nocache(addr, len);
557 #endif
558
559         if (!dd->ipath_kregbase) {
560                 ipath_dbg("Unable to map io addr %llx to kvirt, failing\n",
561                           addr);
562                 ret = -ENOMEM;
563                 goto bail_iounmap;
564         }
565         dd->ipath_kregend = (u64 __iomem *)
566                 ((void __iomem *)dd->ipath_kregbase + len);
567         dd->ipath_physaddr = addr;      /* used for io_remap, etc. */
568         /* for user mmap */
569         ipath_cdbg(VERBOSE, "mapped io addr %llx to kregbase %p\n",
570                    addr, dd->ipath_kregbase);
571
572         if (dd->ipath_f_bus(dd, pdev))
573                 ipath_dev_err(dd, "Failed to setup config space; "
574                               "continuing anyway\n");
575
576         /*
577          * set up our interrupt handler; IRQF_SHARED probably not needed,
578          * since MSI interrupts shouldn't be shared but won't  hurt for now.
579          * check 0 irq after we return from chip-specific bus setup, since
580          * that can affect this due to setup
581          */
582         if (!dd->ipath_irq)
583                 ipath_dev_err(dd, "irq is 0, BIOS error?  Interrupts won't "
584                               "work\n");
585         else {
586                 ret = request_irq(dd->ipath_irq, ipath_intr, IRQF_SHARED,
587                                   IPATH_DRV_NAME, dd);
588                 if (ret) {
589                         ipath_dev_err(dd, "Couldn't setup irq handler, "
590                                       "irq=%d: %d\n", dd->ipath_irq, ret);
591                         goto bail_iounmap;
592                 }
593         }
594
595         ret = ipath_init_chip(dd, 0);   /* do the chip-specific init */
596         if (ret)
597                 goto bail_irqsetup;
598
599         ret = ipath_enable_wc(dd);
600
601         if (ret) {
602                 ipath_dev_err(dd, "Write combining not enabled "
603                               "(err %d): performance may be poor\n",
604                               -ret);
605                 ret = 0;
606         }
607
608         ipath_verify_pioperf(dd);
609
610         ipath_device_create_group(&pdev->dev, dd);
611         ipathfs_add_device(dd);
612         ipath_user_add(dd);
613         ipath_diag_add(dd);
614         ipath_register_ib_device(dd);
615
616         goto bail;
617
618 bail_irqsetup:
619         cleanup_device(dd);
620
621         if (dd->ipath_irq)
622                 dd->ipath_f_free_irq(dd);
623
624         if (dd->ipath_f_cleanup)
625                 dd->ipath_f_cleanup(dd);
626
627 bail_iounmap:
628         iounmap((volatile void __iomem *) dd->ipath_kregbase);
629
630 bail_regions:
631         pci_release_regions(pdev);
632
633 bail_disable:
634         pci_disable_device(pdev);
635
636 bail_devdata:
637         ipath_free_devdata(pdev, dd);
638
639 bail:
640         return ret;
641 }
642
643 static void cleanup_device(struct ipath_devdata *dd)
644 {
645         int port;
646         struct ipath_portdata **tmp;
647         unsigned long flags;
648
649         if (*dd->ipath_statusp & IPATH_STATUS_CHIP_PRESENT) {
650                 /* can't do anything more with chip; needs re-init */
651                 *dd->ipath_statusp &= ~IPATH_STATUS_CHIP_PRESENT;
652                 if (dd->ipath_kregbase) {
653                         /*
654                          * if we haven't already cleaned up before these are
655                          * to ensure any register reads/writes "fail" until
656                          * re-init
657                          */
658                         dd->ipath_kregbase = NULL;
659                         dd->ipath_uregbase = 0;
660                         dd->ipath_sregbase = 0;
661                         dd->ipath_cregbase = 0;
662                         dd->ipath_kregsize = 0;
663                 }
664                 ipath_disable_wc(dd);
665         }
666
667         if (dd->ipath_spectriggerhit)
668                 dev_info(&dd->pcidev->dev, "%lu special trigger hits\n",
669                          dd->ipath_spectriggerhit);
670
671         if (dd->ipath_pioavailregs_dma) {
672                 dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE,
673                                   (void *) dd->ipath_pioavailregs_dma,
674                                   dd->ipath_pioavailregs_phys);
675                 dd->ipath_pioavailregs_dma = NULL;
676         }
677         if (dd->ipath_dummy_hdrq) {
678                 dma_free_coherent(&dd->pcidev->dev,
679                         dd->ipath_pd[0]->port_rcvhdrq_size,
680                         dd->ipath_dummy_hdrq, dd->ipath_dummy_hdrq_phys);
681                 dd->ipath_dummy_hdrq = NULL;
682         }
683
684         if (dd->ipath_pageshadow) {
685                 struct page **tmpp = dd->ipath_pageshadow;
686                 dma_addr_t *tmpd = dd->ipath_physshadow;
687                 int i, cnt = 0;
688
689                 ipath_cdbg(VERBOSE, "Unlocking any expTID pages still "
690                            "locked\n");
691                 for (port = 0; port < dd->ipath_cfgports; port++) {
692                         int port_tidbase = port * dd->ipath_rcvtidcnt;
693                         int maxtid = port_tidbase + dd->ipath_rcvtidcnt;
694                         for (i = port_tidbase; i < maxtid; i++) {
695                                 if (!tmpp[i])
696                                         continue;
697                                 pci_unmap_page(dd->pcidev, tmpd[i],
698                                         PAGE_SIZE, PCI_DMA_FROMDEVICE);
699                                 ipath_release_user_pages(&tmpp[i], 1);
700                                 tmpp[i] = NULL;
701                                 cnt++;
702                         }
703                 }
704                 if (cnt) {
705                         ipath_stats.sps_pageunlocks += cnt;
706                         ipath_cdbg(VERBOSE, "There were still %u expTID "
707                                    "entries locked\n", cnt);
708                 }
709                 if (ipath_stats.sps_pagelocks ||
710                     ipath_stats.sps_pageunlocks)
711                         ipath_cdbg(VERBOSE, "%llu pages locked, %llu "
712                                    "unlocked via ipath_m{un}lock\n",
713                                    (unsigned long long)
714                                    ipath_stats.sps_pagelocks,
715                                    (unsigned long long)
716                                    ipath_stats.sps_pageunlocks);
717
718                 ipath_cdbg(VERBOSE, "Free shadow page tid array at %p\n",
719                            dd->ipath_pageshadow);
720                 tmpp = dd->ipath_pageshadow;
721                 dd->ipath_pageshadow = NULL;
722                 vfree(tmpp);
723
724                 dd->ipath_egrtidbase = NULL;
725         }
726
727         /*
728          * free any resources still in use (usually just kernel ports)
729          * at unload; we do for portcnt, because that's what we allocate.
730          * We acquire lock to be really paranoid that ipath_pd isn't being
731          * accessed from some interrupt-related code (that should not happen,
732          * but best to be sure).
733          */
734         spin_lock_irqsave(&dd->ipath_uctxt_lock, flags);
735         tmp = dd->ipath_pd;
736         dd->ipath_pd = NULL;
737         spin_unlock_irqrestore(&dd->ipath_uctxt_lock, flags);
738         for (port = 0; port < dd->ipath_portcnt; port++) {
739                 struct ipath_portdata *pd = tmp[port];
740                 tmp[port] = NULL; /* debugging paranoia */
741                 ipath_free_pddata(dd, pd);
742         }
743         kfree(tmp);
744 }
745
746 static void __devexit ipath_remove_one(struct pci_dev *pdev)
747 {
748         struct ipath_devdata *dd = pci_get_drvdata(pdev);
749
750         ipath_cdbg(VERBOSE, "removing, pdev=%p, dd=%p\n", pdev, dd);
751
752         /*
753          * disable the IB link early, to be sure no new packets arrive, which
754          * complicates the shutdown process
755          */
756         ipath_shutdown_device(dd);
757
758         flush_scheduled_work();
759
760         if (dd->verbs_dev)
761                 ipath_unregister_ib_device(dd->verbs_dev);
762
763         ipath_diag_remove(dd);
764         ipath_user_remove(dd);
765         ipathfs_remove_device(dd);
766         ipath_device_remove_group(&pdev->dev, dd);
767
768         ipath_cdbg(VERBOSE, "Releasing pci memory regions, dd %p, "
769                    "unit %u\n", dd, (u32) dd->ipath_unit);
770
771         cleanup_device(dd);
772
773         /*
774          * turn off rcv, send, and interrupts for all ports, all drivers
775          * should also hard reset the chip here?
776          * free up port 0 (kernel) rcvhdr, egr bufs, and eventually tid bufs
777          * for all versions of the driver, if they were allocated
778          */
779         if (dd->ipath_irq) {
780                 ipath_cdbg(VERBOSE, "unit %u free irq %d\n",
781                            dd->ipath_unit, dd->ipath_irq);
782                 dd->ipath_f_free_irq(dd);
783         } else
784                 ipath_dbg("irq is 0, not doing free_irq "
785                           "for unit %u\n", dd->ipath_unit);
786         /*
787          * we check for NULL here, because it's outside
788          * the kregbase check, and we need to call it
789          * after the free_irq.  Thus it's possible that
790          * the function pointers were never initialized.
791          */
792         if (dd->ipath_f_cleanup)
793                 /* clean up chip-specific stuff */
794                 dd->ipath_f_cleanup(dd);
795
796         ipath_cdbg(VERBOSE, "Unmapping kregbase %p\n", dd->ipath_kregbase);
797         iounmap((volatile void __iomem *) dd->ipath_kregbase);
798         pci_release_regions(pdev);
799         ipath_cdbg(VERBOSE, "calling pci_disable_device\n");
800         pci_disable_device(pdev);
801
802         ipath_free_devdata(pdev, dd);
803 }
804
805 /* general driver use */
806 DEFINE_MUTEX(ipath_mutex);
807
808 static DEFINE_SPINLOCK(ipath_pioavail_lock);
809
810 /**
811  * ipath_disarm_piobufs - cancel a range of PIO buffers
812  * @dd: the infinipath device
813  * @first: the first PIO buffer to cancel
814  * @cnt: the number of PIO buffers to cancel
815  *
816  * cancel a range of PIO buffers, used when they might be armed, but
817  * not triggered.  Used at init to ensure buffer state, and also user
818  * process close, in case it died while writing to a PIO buffer
819  * Also after errors.
820  */
821 void ipath_disarm_piobufs(struct ipath_devdata *dd, unsigned first,
822                           unsigned cnt)
823 {
824         unsigned i, last = first + cnt;
825         unsigned long flags;
826
827         ipath_cdbg(PKT, "disarm %u PIObufs first=%u\n", cnt, first);
828         for (i = first; i < last; i++) {
829                 spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
830                 /*
831                  * The disarm-related bits are write-only, so it
832                  * is ok to OR them in with our copy of sendctrl
833                  * while we hold the lock.
834                  */
835                 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
836                         dd->ipath_sendctrl | INFINIPATH_S_DISARM |
837                         (i << INFINIPATH_S_DISARMPIOBUF_SHIFT));
838                 /* can't disarm bufs back-to-back per iba7220 spec */
839                 ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
840                 spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
841         }
842         /* on some older chips, update may not happen after cancel */
843         ipath_force_pio_avail_update(dd);
844 }
845
846 /**
847  * ipath_wait_linkstate - wait for an IB link state change to occur
848  * @dd: the infinipath device
849  * @state: the state to wait for
850  * @msecs: the number of milliseconds to wait
851  *
852  * wait up to msecs milliseconds for IB link state change to occur for
853  * now, take the easy polling route.  Currently used only by
854  * ipath_set_linkstate.  Returns 0 if state reached, otherwise
855  * -ETIMEDOUT state can have multiple states set, for any of several
856  * transitions.
857  */
858 int ipath_wait_linkstate(struct ipath_devdata *dd, u32 state, int msecs)
859 {
860         dd->ipath_state_wanted = state;
861         wait_event_interruptible_timeout(ipath_state_wait,
862                                          (dd->ipath_flags & state),
863                                          msecs_to_jiffies(msecs));
864         dd->ipath_state_wanted = 0;
865
866         if (!(dd->ipath_flags & state)) {
867                 u64 val;
868                 ipath_cdbg(VERBOSE, "Didn't reach linkstate %s within %u"
869                            " ms\n",
870                            /* test INIT ahead of DOWN, both can be set */
871                            (state & IPATH_LINKINIT) ? "INIT" :
872                            ((state & IPATH_LINKDOWN) ? "DOWN" :
873                             ((state & IPATH_LINKARMED) ? "ARM" : "ACTIVE")),
874                            msecs);
875                 val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_ibcstatus);
876                 ipath_cdbg(VERBOSE, "ibcc=%llx ibcstatus=%llx (%s)\n",
877                            (unsigned long long) ipath_read_kreg64(
878                                    dd, dd->ipath_kregs->kr_ibcctrl),
879                            (unsigned long long) val,
880                            ipath_ibcstatus_str[val & dd->ibcs_lts_mask]);
881         }
882         return (dd->ipath_flags & state) ? 0 : -ETIMEDOUT;
883 }
884
885 static void decode_sdma_errs(struct ipath_devdata *dd, ipath_err_t err,
886         char *buf, size_t blen)
887 {
888         static const struct {
889                 ipath_err_t err;
890                 const char *msg;
891         } errs[] = {
892                 { INFINIPATH_E_SDMAGENMISMATCH, "SDmaGenMismatch" },
893                 { INFINIPATH_E_SDMAOUTOFBOUND, "SDmaOutOfBound" },
894                 { INFINIPATH_E_SDMATAILOUTOFBOUND, "SDmaTailOutOfBound" },
895                 { INFINIPATH_E_SDMABASE, "SDmaBase" },
896                 { INFINIPATH_E_SDMA1STDESC, "SDma1stDesc" },
897                 { INFINIPATH_E_SDMARPYTAG, "SDmaRpyTag" },
898                 { INFINIPATH_E_SDMADWEN, "SDmaDwEn" },
899                 { INFINIPATH_E_SDMAMISSINGDW, "SDmaMissingDw" },
900                 { INFINIPATH_E_SDMAUNEXPDATA, "SDmaUnexpData" },
901                 { INFINIPATH_E_SDMADESCADDRMISALIGN, "SDmaDescAddrMisalign" },
902                 { INFINIPATH_E_SENDBUFMISUSE, "SendBufMisuse" },
903                 { INFINIPATH_E_SDMADISABLED, "SDmaDisabled" },
904         };
905         int i;
906         int expected;
907         size_t bidx = 0;
908
909         for (i = 0; i < ARRAY_SIZE(errs); i++) {
910                 expected = (errs[i].err != INFINIPATH_E_SDMADISABLED) ? 0 :
911                         test_bit(IPATH_SDMA_ABORTING, &dd->ipath_sdma_status);
912                 if ((err & errs[i].err) && !expected)
913                         bidx += snprintf(buf + bidx, blen - bidx,
914                                          "%s ", errs[i].msg);
915         }
916 }
917
918 /*
919  * Decode the error status into strings, deciding whether to always
920  * print * it or not depending on "normal packet errors" vs everything
921  * else.   Return 1 if "real" errors, otherwise 0 if only packet
922  * errors, so caller can decide what to print with the string.
923  */
924 int ipath_decode_err(struct ipath_devdata *dd, char *buf, size_t blen,
925         ipath_err_t err)
926 {
927         int iserr = 1;
928         *buf = '\0';
929         if (err & INFINIPATH_E_PKTERRS) {
930                 if (!(err & ~INFINIPATH_E_PKTERRS))
931                         iserr = 0; // if only packet errors.
932                 if (ipath_debug & __IPATH_ERRPKTDBG) {
933                         if (err & INFINIPATH_E_REBP)
934                                 strlcat(buf, "EBP ", blen);
935                         if (err & INFINIPATH_E_RVCRC)
936                                 strlcat(buf, "VCRC ", blen);
937                         if (err & INFINIPATH_E_RICRC) {
938                                 strlcat(buf, "CRC ", blen);
939                                 // clear for check below, so only once
940                                 err &= INFINIPATH_E_RICRC;
941                         }
942                         if (err & INFINIPATH_E_RSHORTPKTLEN)
943                                 strlcat(buf, "rshortpktlen ", blen);
944                         if (err & INFINIPATH_E_SDROPPEDDATAPKT)
945                                 strlcat(buf, "sdroppeddatapkt ", blen);
946                         if (err & INFINIPATH_E_SPKTLEN)
947                                 strlcat(buf, "spktlen ", blen);
948                 }
949                 if ((err & INFINIPATH_E_RICRC) &&
950                         !(err&(INFINIPATH_E_RVCRC|INFINIPATH_E_REBP)))
951                         strlcat(buf, "CRC ", blen);
952                 if (!iserr)
953                         goto done;
954         }
955         if (err & INFINIPATH_E_RHDRLEN)
956                 strlcat(buf, "rhdrlen ", blen);
957         if (err & INFINIPATH_E_RBADTID)
958                 strlcat(buf, "rbadtid ", blen);
959         if (err & INFINIPATH_E_RBADVERSION)
960                 strlcat(buf, "rbadversion ", blen);
961         if (err & INFINIPATH_E_RHDR)
962                 strlcat(buf, "rhdr ", blen);
963         if (err & INFINIPATH_E_SENDSPECIALTRIGGER)
964                 strlcat(buf, "sendspecialtrigger ", blen);
965         if (err & INFINIPATH_E_RLONGPKTLEN)
966                 strlcat(buf, "rlongpktlen ", blen);
967         if (err & INFINIPATH_E_RMAXPKTLEN)
968                 strlcat(buf, "rmaxpktlen ", blen);
969         if (err & INFINIPATH_E_RMINPKTLEN)
970                 strlcat(buf, "rminpktlen ", blen);
971         if (err & INFINIPATH_E_SMINPKTLEN)
972                 strlcat(buf, "sminpktlen ", blen);
973         if (err & INFINIPATH_E_RFORMATERR)
974                 strlcat(buf, "rformaterr ", blen);
975         if (err & INFINIPATH_E_RUNSUPVL)
976                 strlcat(buf, "runsupvl ", blen);
977         if (err & INFINIPATH_E_RUNEXPCHAR)
978                 strlcat(buf, "runexpchar ", blen);
979         if (err & INFINIPATH_E_RIBFLOW)
980                 strlcat(buf, "ribflow ", blen);
981         if (err & INFINIPATH_E_SUNDERRUN)
982                 strlcat(buf, "sunderrun ", blen);
983         if (err & INFINIPATH_E_SPIOARMLAUNCH)
984                 strlcat(buf, "spioarmlaunch ", blen);
985         if (err & INFINIPATH_E_SUNEXPERRPKTNUM)
986                 strlcat(buf, "sunexperrpktnum ", blen);
987         if (err & INFINIPATH_E_SDROPPEDSMPPKT)
988                 strlcat(buf, "sdroppedsmppkt ", blen);
989         if (err & INFINIPATH_E_SMAXPKTLEN)
990                 strlcat(buf, "smaxpktlen ", blen);
991         if (err & INFINIPATH_E_SUNSUPVL)
992                 strlcat(buf, "sunsupVL ", blen);
993         if (err & INFINIPATH_E_INVALIDADDR)
994                 strlcat(buf, "invalidaddr ", blen);
995         if (err & INFINIPATH_E_RRCVEGRFULL)
996                 strlcat(buf, "rcvegrfull ", blen);
997         if (err & INFINIPATH_E_RRCVHDRFULL)
998                 strlcat(buf, "rcvhdrfull ", blen);
999         if (err & INFINIPATH_E_IBSTATUSCHANGED)
1000                 strlcat(buf, "ibcstatuschg ", blen);
1001         if (err & INFINIPATH_E_RIBLOSTLINK)
1002                 strlcat(buf, "riblostlink ", blen);
1003         if (err & INFINIPATH_E_HARDWARE)
1004                 strlcat(buf, "hardware ", blen);
1005         if (err & INFINIPATH_E_RESET)
1006                 strlcat(buf, "reset ", blen);
1007         if (err & INFINIPATH_E_SDMAERRS)
1008                 decode_sdma_errs(dd, err, buf, blen);
1009         if (err & INFINIPATH_E_INVALIDEEPCMD)
1010                 strlcat(buf, "invalideepromcmd ", blen);
1011 done:
1012         return iserr;
1013 }
1014
1015 /**
1016  * get_rhf_errstring - decode RHF errors
1017  * @err: the err number
1018  * @msg: the output buffer
1019  * @len: the length of the output buffer
1020  *
1021  * only used one place now, may want more later
1022  */
1023 static void get_rhf_errstring(u32 err, char *msg, size_t len)
1024 {
1025         /* if no errors, and so don't need to check what's first */
1026         *msg = '\0';
1027
1028         if (err & INFINIPATH_RHF_H_ICRCERR)
1029                 strlcat(msg, "icrcerr ", len);
1030         if (err & INFINIPATH_RHF_H_VCRCERR)
1031                 strlcat(msg, "vcrcerr ", len);
1032         if (err & INFINIPATH_RHF_H_PARITYERR)
1033                 strlcat(msg, "parityerr ", len);
1034         if (err & INFINIPATH_RHF_H_LENERR)
1035                 strlcat(msg, "lenerr ", len);
1036         if (err & INFINIPATH_RHF_H_MTUERR)
1037                 strlcat(msg, "mtuerr ", len);
1038         if (err & INFINIPATH_RHF_H_IHDRERR)
1039                 /* infinipath hdr checksum error */
1040                 strlcat(msg, "ipathhdrerr ", len);
1041         if (err & INFINIPATH_RHF_H_TIDERR)
1042                 strlcat(msg, "tiderr ", len);
1043         if (err & INFINIPATH_RHF_H_MKERR)
1044                 /* bad port, offset, etc. */
1045                 strlcat(msg, "invalid ipathhdr ", len);
1046         if (err & INFINIPATH_RHF_H_IBERR)
1047                 strlcat(msg, "iberr ", len);
1048         if (err & INFINIPATH_RHF_L_SWA)
1049                 strlcat(msg, "swA ", len);
1050         if (err & INFINIPATH_RHF_L_SWB)
1051                 strlcat(msg, "swB ", len);
1052 }
1053
1054 /**
1055  * ipath_get_egrbuf - get an eager buffer
1056  * @dd: the infinipath device
1057  * @bufnum: the eager buffer to get
1058  *
1059  * must only be called if ipath_pd[port] is known to be allocated
1060  */
1061 static inline void *ipath_get_egrbuf(struct ipath_devdata *dd, u32 bufnum)
1062 {
1063         return dd->ipath_port0_skbinfo ?
1064                 (void *) dd->ipath_port0_skbinfo[bufnum].skb->data : NULL;
1065 }
1066
1067 /**
1068  * ipath_alloc_skb - allocate an skb and buffer with possible constraints
1069  * @dd: the infinipath device
1070  * @gfp_mask: the sk_buff SFP mask
1071  */
1072 struct sk_buff *ipath_alloc_skb(struct ipath_devdata *dd,
1073                                 gfp_t gfp_mask)
1074 {
1075         struct sk_buff *skb;
1076         u32 len;
1077
1078         /*
1079          * Only fully supported way to handle this is to allocate lots
1080          * extra, align as needed, and then do skb_reserve().  That wastes
1081          * a lot of memory...  I'll have to hack this into infinipath_copy
1082          * also.
1083          */
1084
1085         /*
1086          * We need 2 extra bytes for ipath_ether data sent in the
1087          * key header.  In order to keep everything dword aligned,
1088          * we'll reserve 4 bytes.
1089          */
1090         len = dd->ipath_ibmaxlen + 4;
1091
1092         if (dd->ipath_flags & IPATH_4BYTE_TID) {
1093                 /* We need a 2KB multiple alignment, and there is no way
1094                  * to do it except to allocate extra and then skb_reserve
1095                  * enough to bring it up to the right alignment.
1096                  */
1097                 len += 2047;
1098         }
1099
1100         skb = __dev_alloc_skb(len, gfp_mask);
1101         if (!skb) {
1102                 ipath_dev_err(dd, "Failed to allocate skbuff, length %u\n",
1103                               len);
1104                 goto bail;
1105         }
1106
1107         skb_reserve(skb, 4);
1108
1109         if (dd->ipath_flags & IPATH_4BYTE_TID) {
1110                 u32 una = (unsigned long)skb->data & 2047;
1111                 if (una)
1112                         skb_reserve(skb, 2048 - una);
1113         }
1114
1115 bail:
1116         return skb;
1117 }
1118
1119 static void ipath_rcv_hdrerr(struct ipath_devdata *dd,
1120                              u32 eflags,
1121                              u32 l,
1122                              u32 etail,
1123                              __le32 *rhf_addr,
1124                              struct ipath_message_header *hdr)
1125 {
1126         char emsg[128];
1127
1128         get_rhf_errstring(eflags, emsg, sizeof emsg);
1129         ipath_cdbg(PKT, "RHFerrs %x hdrqtail=%x typ=%u "
1130                    "tlen=%x opcode=%x egridx=%x: %s\n",
1131                    eflags, l,
1132                    ipath_hdrget_rcv_type(rhf_addr),
1133                    ipath_hdrget_length_in_bytes(rhf_addr),
1134                    be32_to_cpu(hdr->bth[0]) >> 24,
1135                    etail, emsg);
1136
1137         /* Count local link integrity errors. */
1138         if (eflags & (INFINIPATH_RHF_H_ICRCERR | INFINIPATH_RHF_H_VCRCERR)) {
1139                 u8 n = (dd->ipath_ibcctrl >>
1140                         INFINIPATH_IBCC_PHYERRTHRESHOLD_SHIFT) &
1141                         INFINIPATH_IBCC_PHYERRTHRESHOLD_MASK;
1142
1143                 if (++dd->ipath_lli_counter > n) {
1144                         dd->ipath_lli_counter = 0;
1145                         dd->ipath_lli_errors++;
1146                 }
1147         }
1148 }
1149
1150 /*
1151  * ipath_kreceive - receive a packet
1152  * @pd: the infinipath port
1153  *
1154  * called from interrupt handler for errors or receive interrupt
1155  */
1156 void ipath_kreceive(struct ipath_portdata *pd)
1157 {
1158         struct ipath_devdata *dd = pd->port_dd;
1159         __le32 *rhf_addr;
1160         void *ebuf;
1161         const u32 rsize = dd->ipath_rcvhdrentsize;      /* words */
1162         const u32 maxcnt = dd->ipath_rcvhdrcnt * rsize; /* words */
1163         u32 etail = -1, l, hdrqtail;
1164         struct ipath_message_header *hdr;
1165         u32 eflags, i, etype, tlen, pkttot = 0, updegr = 0, reloop = 0;
1166         static u64 totcalls;    /* stats, may eventually remove */
1167         int last;
1168
1169         l = pd->port_head;
1170         rhf_addr = (__le32 *) pd->port_rcvhdrq + l + dd->ipath_rhf_offset;
1171         if (dd->ipath_flags & IPATH_NODMA_RTAIL) {
1172                 u32 seq = ipath_hdrget_seq(rhf_addr);
1173
1174                 if (seq != pd->port_seq_cnt)
1175                         goto bail;
1176                 hdrqtail = 0;
1177         } else {
1178                 hdrqtail = ipath_get_rcvhdrtail(pd);
1179                 if (l == hdrqtail)
1180                         goto bail;
1181                 smp_rmb();
1182         }
1183
1184 reloop:
1185         for (last = 0, i = 1; !last; i += !last) {
1186                 hdr = dd->ipath_f_get_msgheader(dd, rhf_addr);
1187                 eflags = ipath_hdrget_err_flags(rhf_addr);
1188                 etype = ipath_hdrget_rcv_type(rhf_addr);
1189                 /* total length */
1190                 tlen = ipath_hdrget_length_in_bytes(rhf_addr);
1191                 ebuf = NULL;
1192                 if ((dd->ipath_flags & IPATH_NODMA_RTAIL) ?
1193                     ipath_hdrget_use_egr_buf(rhf_addr) :
1194                     (etype != RCVHQ_RCV_TYPE_EXPECTED)) {
1195                         /*
1196                          * It turns out that the chip uses an eager buffer
1197                          * for all non-expected packets, whether it "needs"
1198                          * one or not.  So always get the index, but don't
1199                          * set ebuf (so we try to copy data) unless the
1200                          * length requires it.
1201                          */
1202                         etail = ipath_hdrget_index(rhf_addr);
1203                         updegr = 1;
1204                         if (tlen > sizeof(*hdr) ||
1205                             etype == RCVHQ_RCV_TYPE_NON_KD)
1206                                 ebuf = ipath_get_egrbuf(dd, etail);
1207                 }
1208
1209                 /*
1210                  * both tiderr and ipathhdrerr are set for all plain IB
1211                  * packets; only ipathhdrerr should be set.
1212                  */
1213
1214                 if (etype != RCVHQ_RCV_TYPE_NON_KD &&
1215                     etype != RCVHQ_RCV_TYPE_ERROR &&
1216                     ipath_hdrget_ipath_ver(hdr->iph.ver_port_tid_offset) !=
1217                     IPS_PROTO_VERSION)
1218                         ipath_cdbg(PKT, "Bad InfiniPath protocol version "
1219                                    "%x\n", etype);
1220
1221                 if (unlikely(eflags))
1222                         ipath_rcv_hdrerr(dd, eflags, l, etail, rhf_addr, hdr);
1223                 else if (etype == RCVHQ_RCV_TYPE_NON_KD) {
1224                         ipath_ib_rcv(dd->verbs_dev, (u32 *)hdr, ebuf, tlen);
1225                         if (dd->ipath_lli_counter)
1226                                 dd->ipath_lli_counter--;
1227                 } else if (etype == RCVHQ_RCV_TYPE_EAGER) {
1228                         u8 opcode = be32_to_cpu(hdr->bth[0]) >> 24;
1229                         u32 qp = be32_to_cpu(hdr->bth[1]) & 0xffffff;
1230                         ipath_cdbg(PKT, "typ %x, opcode %x (eager, "
1231                                    "qp=%x), len %x; ignored\n",
1232                                    etype, opcode, qp, tlen);
1233                 }
1234                 else if (etype == RCVHQ_RCV_TYPE_EXPECTED)
1235                         ipath_dbg("Bug: Expected TID, opcode %x; ignored\n",
1236                                   be32_to_cpu(hdr->bth[0]) >> 24);
1237                 else {
1238                         /*
1239                          * error packet, type of error unknown.
1240                          * Probably type 3, but we don't know, so don't
1241                          * even try to print the opcode, etc.
1242                          * Usually caused by a "bad packet", that has no
1243                          * BTH, when the LRH says it should.
1244                          */
1245                         ipath_cdbg(ERRPKT, "Error Pkt, but no eflags! egrbuf"
1246                                   " %x, len %x hdrq+%x rhf: %Lx\n",
1247                                   etail, tlen, l, (unsigned long long)
1248                                   le64_to_cpu(*(__le64 *) rhf_addr));
1249                         if (ipath_debug & __IPATH_ERRPKTDBG) {
1250                                 u32 j, *d, dw = rsize-2;
1251                                 if (rsize > (tlen>>2))
1252                                         dw = tlen>>2;
1253                                 d = (u32 *)hdr;
1254                                 printk(KERN_DEBUG "EPkt rcvhdr(%x dw):\n",
1255                                         dw);
1256                                 for (j = 0; j < dw; j++)
1257                                         printk(KERN_DEBUG "%8x%s", d[j],
1258                                                 (j%8) == 7 ? "\n" : " ");
1259                                 printk(KERN_DEBUG ".\n");
1260                         }
1261                 }
1262                 l += rsize;
1263                 if (l >= maxcnt)
1264                         l = 0;
1265                 rhf_addr = (__le32 *) pd->port_rcvhdrq +
1266                         l + dd->ipath_rhf_offset;
1267                 if (dd->ipath_flags & IPATH_NODMA_RTAIL) {
1268                         u32 seq = ipath_hdrget_seq(rhf_addr);
1269
1270                         if (++pd->port_seq_cnt > 13)
1271                                 pd->port_seq_cnt = 1;
1272                         if (seq != pd->port_seq_cnt)
1273                                 last = 1;
1274                 } else if (l == hdrqtail)
1275                         last = 1;
1276                 /*
1277                  * update head regs on last packet, and every 16 packets.
1278                  * Reduce bus traffic, while still trying to prevent
1279                  * rcvhdrq overflows, for when the queue is nearly full
1280                  */
1281                 if (last || !(i & 0xf)) {
1282                         u64 lval = l;
1283
1284                         /* request IBA6120 and 7220 interrupt only on last */
1285                         if (last)
1286                                 lval |= dd->ipath_rhdrhead_intr_off;
1287                         ipath_write_ureg(dd, ur_rcvhdrhead, lval,
1288                                 pd->port_port);
1289                         if (updegr) {
1290                                 ipath_write_ureg(dd, ur_rcvegrindexhead,
1291                                                  etail, pd->port_port);
1292                                 updegr = 0;
1293                         }
1294                 }
1295         }
1296
1297         if (!dd->ipath_rhdrhead_intr_off && !reloop &&
1298             !(dd->ipath_flags & IPATH_NODMA_RTAIL)) {
1299                 /* IBA6110 workaround; we can have a race clearing chip
1300                  * interrupt with another interrupt about to be delivered,
1301                  * and can clear it before it is delivered on the GPIO
1302                  * workaround.  By doing the extra check here for the
1303                  * in-memory tail register updating while we were doing
1304                  * earlier packets, we "almost" guarantee we have covered
1305                  * that case.
1306                  */
1307                 u32 hqtail = ipath_get_rcvhdrtail(pd);
1308                 if (hqtail != hdrqtail) {
1309                         hdrqtail = hqtail;
1310                         reloop = 1; /* loop 1 extra time at most */
1311                         goto reloop;
1312                 }
1313         }
1314
1315         pkttot += i;
1316
1317         pd->port_head = l;
1318
1319         if (pkttot > ipath_stats.sps_maxpkts_call)
1320                 ipath_stats.sps_maxpkts_call = pkttot;
1321         ipath_stats.sps_port0pkts += pkttot;
1322         ipath_stats.sps_avgpkts_call =
1323                 ipath_stats.sps_port0pkts / ++totcalls;
1324
1325 bail:;
1326 }
1327
1328 /**
1329  * ipath_update_pio_bufs - update shadow copy of the PIO availability map
1330  * @dd: the infinipath device
1331  *
1332  * called whenever our local copy indicates we have run out of send buffers
1333  * NOTE: This can be called from interrupt context by some code
1334  * and from non-interrupt context by ipath_getpiobuf().
1335  */
1336
1337 static void ipath_update_pio_bufs(struct ipath_devdata *dd)
1338 {
1339         unsigned long flags;
1340         int i;
1341         const unsigned piobregs = (unsigned)dd->ipath_pioavregs;
1342
1343         /* If the generation (check) bits have changed, then we update the
1344          * busy bit for the corresponding PIO buffer.  This algorithm will
1345          * modify positions to the value they already have in some cases
1346          * (i.e., no change), but it's faster than changing only the bits
1347          * that have changed.
1348          *
1349          * We would like to do this atomicly, to avoid spinlocks in the
1350          * critical send path, but that's not really possible, given the
1351          * type of changes, and that this routine could be called on
1352          * multiple cpu's simultaneously, so we lock in this routine only,
1353          * to avoid conflicting updates; all we change is the shadow, and
1354          * it's a single 64 bit memory location, so by definition the update
1355          * is atomic in terms of what other cpu's can see in testing the
1356          * bits.  The spin_lock overhead isn't too bad, since it only
1357          * happens when all buffers are in use, so only cpu overhead, not
1358          * latency or bandwidth is affected.
1359          */
1360         if (!dd->ipath_pioavailregs_dma) {
1361                 ipath_dbg("Update shadow pioavail, but regs_dma NULL!\n");
1362                 return;
1363         }
1364         if (ipath_debug & __IPATH_VERBDBG) {
1365                 /* only if packet debug and verbose */
1366                 volatile __le64 *dma = dd->ipath_pioavailregs_dma;
1367                 unsigned long *shadow = dd->ipath_pioavailshadow;
1368
1369                 ipath_cdbg(PKT, "Refill avail, dma0=%llx shad0=%lx, "
1370                            "d1=%llx s1=%lx, d2=%llx s2=%lx, d3=%llx "
1371                            "s3=%lx\n",
1372                            (unsigned long long) le64_to_cpu(dma[0]),
1373                            shadow[0],
1374                            (unsigned long long) le64_to_cpu(dma[1]),
1375                            shadow[1],
1376                            (unsigned long long) le64_to_cpu(dma[2]),
1377                            shadow[2],
1378                            (unsigned long long) le64_to_cpu(dma[3]),
1379                            shadow[3]);
1380                 if (piobregs > 4)
1381                         ipath_cdbg(
1382                                 PKT, "2nd group, dma4=%llx shad4=%lx, "
1383                                 "d5=%llx s5=%lx, d6=%llx s6=%lx, "
1384                                 "d7=%llx s7=%lx\n",
1385                                 (unsigned long long) le64_to_cpu(dma[4]),
1386                                 shadow[4],
1387                                 (unsigned long long) le64_to_cpu(dma[5]),
1388                                 shadow[5],
1389                                 (unsigned long long) le64_to_cpu(dma[6]),
1390                                 shadow[6],
1391                                 (unsigned long long) le64_to_cpu(dma[7]),
1392                                 shadow[7]);
1393         }
1394         spin_lock_irqsave(&ipath_pioavail_lock, flags);
1395         for (i = 0; i < piobregs; i++) {
1396                 u64 pchbusy, pchg, piov, pnew;
1397                 /*
1398                  * Chip Errata: bug 6641; even and odd qwords>3 are swapped
1399                  */
1400                 if (i > 3 && (dd->ipath_flags & IPATH_SWAP_PIOBUFS))
1401                         piov = le64_to_cpu(dd->ipath_pioavailregs_dma[i ^ 1]);
1402                 else
1403                         piov = le64_to_cpu(dd->ipath_pioavailregs_dma[i]);
1404                 pchg = dd->ipath_pioavailkernel[i] &
1405                         ~(dd->ipath_pioavailshadow[i] ^ piov);
1406                 pchbusy = pchg << INFINIPATH_SENDPIOAVAIL_BUSY_SHIFT;
1407                 if (pchg && (pchbusy & dd->ipath_pioavailshadow[i])) {
1408                         pnew = dd->ipath_pioavailshadow[i] & ~pchbusy;
1409                         pnew |= piov & pchbusy;
1410                         dd->ipath_pioavailshadow[i] = pnew;
1411                 }
1412         }
1413         spin_unlock_irqrestore(&ipath_pioavail_lock, flags);
1414 }
1415
1416 /*
1417  * used to force update of pioavailshadow if we can't get a pio buffer.
1418  * Needed primarily due to exitting freeze mode after recovering
1419  * from errors.  Done lazily, because it's safer (known to not
1420  * be writing pio buffers).
1421  */
1422 static void ipath_reset_availshadow(struct ipath_devdata *dd)
1423 {
1424         int i, im;
1425         unsigned long flags;
1426
1427         spin_lock_irqsave(&ipath_pioavail_lock, flags);
1428         for (i = 0; i < dd->ipath_pioavregs; i++) {
1429                 u64 val, oldval;
1430                 /* deal with 6110 chip bug on high register #s */
1431                 im = (i > 3 && (dd->ipath_flags & IPATH_SWAP_PIOBUFS)) ?
1432                         i ^ 1 : i;
1433                 val = le64_to_cpu(dd->ipath_pioavailregs_dma[im]);
1434                 /*
1435                  * busy out the buffers not in the kernel avail list,
1436                  * without changing the generation bits.
1437                  */
1438                 oldval = dd->ipath_pioavailshadow[i];
1439                 dd->ipath_pioavailshadow[i] = val |
1440                         ((~dd->ipath_pioavailkernel[i] <<
1441                         INFINIPATH_SENDPIOAVAIL_BUSY_SHIFT) &
1442                         0xaaaaaaaaaaaaaaaaULL); /* All BUSY bits in qword */
1443                 if (oldval != dd->ipath_pioavailshadow[i])
1444                         ipath_dbg("shadow[%d] was %Lx, now %lx\n",
1445                                 i, (unsigned long long) oldval,
1446                                 dd->ipath_pioavailshadow[i]);
1447         }
1448         spin_unlock_irqrestore(&ipath_pioavail_lock, flags);
1449 }
1450
1451 /**
1452  * ipath_setrcvhdrsize - set the receive header size
1453  * @dd: the infinipath device
1454  * @rhdrsize: the receive header size
1455  *
1456  * called from user init code, and also layered driver init
1457  */
1458 int ipath_setrcvhdrsize(struct ipath_devdata *dd, unsigned rhdrsize)
1459 {
1460         int ret = 0;
1461
1462         if (dd->ipath_flags & IPATH_RCVHDRSZ_SET) {
1463                 if (dd->ipath_rcvhdrsize != rhdrsize) {
1464                         dev_info(&dd->pcidev->dev,
1465                                  "Error: can't set protocol header "
1466                                  "size %u, already %u\n",
1467                                  rhdrsize, dd->ipath_rcvhdrsize);
1468                         ret = -EAGAIN;
1469                 } else
1470                         ipath_cdbg(VERBOSE, "Reuse same protocol header "
1471                                    "size %u\n", dd->ipath_rcvhdrsize);
1472         } else if (rhdrsize > (dd->ipath_rcvhdrentsize -
1473                                (sizeof(u64) / sizeof(u32)))) {
1474                 ipath_dbg("Error: can't set protocol header size %u "
1475                           "(> max %u)\n", rhdrsize,
1476                           dd->ipath_rcvhdrentsize -
1477                           (u32) (sizeof(u64) / sizeof(u32)));
1478                 ret = -EOVERFLOW;
1479         } else {
1480                 dd->ipath_flags |= IPATH_RCVHDRSZ_SET;
1481                 dd->ipath_rcvhdrsize = rhdrsize;
1482                 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvhdrsize,
1483                                  dd->ipath_rcvhdrsize);
1484                 ipath_cdbg(VERBOSE, "Set protocol header size to %u\n",
1485                            dd->ipath_rcvhdrsize);
1486         }
1487         return ret;
1488 }
1489
1490 /*
1491  * debugging code and stats updates if no pio buffers available.
1492  */
1493 static noinline void no_pio_bufs(struct ipath_devdata *dd)
1494 {
1495         unsigned long *shadow = dd->ipath_pioavailshadow;
1496         __le64 *dma = (__le64 *)dd->ipath_pioavailregs_dma;
1497
1498         dd->ipath_upd_pio_shadow = 1;
1499
1500         /*
1501          * not atomic, but if we lose a stat count in a while, that's OK
1502          */
1503         ipath_stats.sps_nopiobufs++;
1504         if (!(++dd->ipath_consec_nopiobuf % 100000)) {
1505                 ipath_force_pio_avail_update(dd); /* at start */
1506                 ipath_dbg("%u tries no piobufavail ts%lx; dmacopy: "
1507                         "%llx %llx %llx %llx\n"
1508                         "ipath  shadow:  %lx %lx %lx %lx\n",
1509                         dd->ipath_consec_nopiobuf,
1510                         (unsigned long)get_cycles(),
1511                         (unsigned long long) le64_to_cpu(dma[0]),
1512                         (unsigned long long) le64_to_cpu(dma[1]),
1513                         (unsigned long long) le64_to_cpu(dma[2]),
1514                         (unsigned long long) le64_to_cpu(dma[3]),
1515                         shadow[0], shadow[1], shadow[2], shadow[3]);
1516                 /*
1517                  * 4 buffers per byte, 4 registers above, cover rest
1518                  * below
1519                  */
1520                 if ((dd->ipath_piobcnt2k + dd->ipath_piobcnt4k) >
1521                     (sizeof(shadow[0]) * 4 * 4))
1522                         ipath_dbg("2nd group: dmacopy: "
1523                                   "%llx %llx %llx %llx\n"
1524                                   "ipath  shadow:  %lx %lx %lx %lx\n",
1525                                   (unsigned long long)le64_to_cpu(dma[4]),
1526                                   (unsigned long long)le64_to_cpu(dma[5]),
1527                                   (unsigned long long)le64_to_cpu(dma[6]),
1528                                   (unsigned long long)le64_to_cpu(dma[7]),
1529                                   shadow[4], shadow[5], shadow[6], shadow[7]);
1530
1531                 /* at end, so update likely happened */
1532                 ipath_reset_availshadow(dd);
1533         }
1534 }
1535
1536 /*
1537  * common code for normal driver pio buffer allocation, and reserved
1538  * allocation.
1539  *
1540  * do appropriate marking as busy, etc.
1541  * returns buffer number if one found (>=0), negative number is error.
1542  */
1543 static u32 __iomem *ipath_getpiobuf_range(struct ipath_devdata *dd,
1544         u32 *pbufnum, u32 first, u32 last, u32 firsti)
1545 {
1546         int i, j, updated = 0;
1547         unsigned piobcnt;
1548         unsigned long flags;
1549         unsigned long *shadow = dd->ipath_pioavailshadow;
1550         u32 __iomem *buf;
1551
1552         piobcnt = last - first;
1553         if (dd->ipath_upd_pio_shadow) {
1554                 /*
1555                  * Minor optimization.  If we had no buffers on last call,
1556                  * start out by doing the update; continue and do scan even
1557                  * if no buffers were updated, to be paranoid
1558                  */
1559                 ipath_update_pio_bufs(dd);
1560                 updated++;
1561                 i = first;
1562         } else
1563                 i = firsti;
1564 rescan:
1565         /*
1566          * while test_and_set_bit() is atomic, we do that and then the
1567          * change_bit(), and the pair is not.  See if this is the cause
1568          * of the remaining armlaunch errors.
1569          */
1570         spin_lock_irqsave(&ipath_pioavail_lock, flags);
1571         for (j = 0; j < piobcnt; j++, i++) {
1572                 if (i >= last)
1573                         i = first;
1574                 if (__test_and_set_bit((2 * i) + 1, shadow))
1575                         continue;
1576                 /* flip generation bit */
1577                 __change_bit(2 * i, shadow);
1578                 break;
1579         }
1580         spin_unlock_irqrestore(&ipath_pioavail_lock, flags);
1581
1582         if (j == piobcnt) {
1583                 if (!updated) {
1584                         /*
1585                          * first time through; shadow exhausted, but may be
1586                          * buffers available, try an update and then rescan.
1587                          */
1588                         ipath_update_pio_bufs(dd);
1589                         updated++;
1590                         i = first;
1591                         goto rescan;
1592                 } else if (updated == 1 && piobcnt <=
1593                         ((dd->ipath_sendctrl
1594                         >> INFINIPATH_S_UPDTHRESH_SHIFT) &
1595                         INFINIPATH_S_UPDTHRESH_MASK)) {
1596                         /*
1597                          * for chips supporting and using the update
1598                          * threshold we need to force an update of the
1599                          * in-memory copy if the count is less than the
1600                          * thershold, then check one more time.
1601                          */
1602                         ipath_force_pio_avail_update(dd);
1603                         ipath_update_pio_bufs(dd);
1604                         updated++;
1605                         i = first;
1606                         goto rescan;
1607                 }
1608
1609                 no_pio_bufs(dd);
1610                 buf = NULL;
1611         } else {
1612                 if (i < dd->ipath_piobcnt2k)
1613                         buf = (u32 __iomem *) (dd->ipath_pio2kbase +
1614                                                i * dd->ipath_palign);
1615                 else
1616                         buf = (u32 __iomem *)
1617                                 (dd->ipath_pio4kbase +
1618                                  (i - dd->ipath_piobcnt2k) * dd->ipath_4kalign);
1619                 if (pbufnum)
1620                         *pbufnum = i;
1621         }
1622
1623         return buf;
1624 }
1625
1626 /**
1627  * ipath_getpiobuf - find an available pio buffer
1628  * @dd: the infinipath device
1629  * @plen: the size of the PIO buffer needed in 32-bit words
1630  * @pbufnum: the buffer number is placed here
1631  */
1632 u32 __iomem *ipath_getpiobuf(struct ipath_devdata *dd, u32 plen, u32 *pbufnum)
1633 {
1634         u32 __iomem *buf;
1635         u32 pnum, nbufs;
1636         u32 first, lasti;
1637
1638         if (plen + 1 >= IPATH_SMALLBUF_DWORDS) {
1639                 first = dd->ipath_piobcnt2k;
1640                 lasti = dd->ipath_lastpioindexl;
1641         } else {
1642                 first = 0;
1643                 lasti = dd->ipath_lastpioindex;
1644         }
1645         nbufs = dd->ipath_piobcnt2k + dd->ipath_piobcnt4k;
1646         buf = ipath_getpiobuf_range(dd, &pnum, first, nbufs, lasti);
1647
1648         if (buf) {
1649                 /*
1650                  * Set next starting place.  It's just an optimization,
1651                  * it doesn't matter who wins on this, so no locking
1652                  */
1653                 if (plen + 1 >= IPATH_SMALLBUF_DWORDS)
1654                         dd->ipath_lastpioindexl = pnum + 1;
1655                 else
1656                         dd->ipath_lastpioindex = pnum + 1;
1657                 if (dd->ipath_upd_pio_shadow)
1658                         dd->ipath_upd_pio_shadow = 0;
1659                 if (dd->ipath_consec_nopiobuf)
1660                         dd->ipath_consec_nopiobuf = 0;
1661                 ipath_cdbg(VERBOSE, "Return piobuf%u %uk @ %p\n",
1662                            pnum, (pnum < dd->ipath_piobcnt2k) ? 2 : 4, buf);
1663                 if (pbufnum)
1664                         *pbufnum = pnum;
1665
1666         }
1667         return buf;
1668 }
1669
1670 /**
1671  * ipath_chg_pioavailkernel - change which send buffers are available for kernel
1672  * @dd: the infinipath device
1673  * @start: the starting send buffer number
1674  * @len: the number of send buffers
1675  * @avail: true if the buffers are available for kernel use, false otherwise
1676  */
1677 void ipath_chg_pioavailkernel(struct ipath_devdata *dd, unsigned start,
1678                               unsigned len, int avail)
1679 {
1680         unsigned long flags;
1681         unsigned end, cnt = 0;
1682
1683         /* There are two bits per send buffer (busy and generation) */
1684         start *= 2;
1685         end = start + len * 2;
1686
1687         spin_lock_irqsave(&ipath_pioavail_lock, flags);
1688         /* Set or clear the busy bit in the shadow. */
1689         while (start < end) {
1690                 if (avail) {
1691                         unsigned long dma;
1692                         int i, im;
1693                         /*
1694                          * the BUSY bit will never be set, because we disarm
1695                          * the user buffers before we hand them back to the
1696                          * kernel.  We do have to make sure the generation
1697                          * bit is set correctly in shadow, since it could
1698                          * have changed many times while allocated to user.
1699                          * We can't use the bitmap functions on the full
1700                          * dma array because it is always little-endian, so
1701                          * we have to flip to host-order first.
1702                          * BITS_PER_LONG is slightly wrong, since it's
1703                          * always 64 bits per register in chip...
1704                          * We only work on 64 bit kernels, so that's OK.
1705                          */
1706                         /* deal with 6110 chip bug on high register #s */
1707                         i = start / BITS_PER_LONG;
1708                         im = (i > 3 && (dd->ipath_flags & IPATH_SWAP_PIOBUFS)) ?
1709                                 i ^ 1 : i;
1710                         __clear_bit(INFINIPATH_SENDPIOAVAIL_BUSY_SHIFT
1711                                 + start, dd->ipath_pioavailshadow);
1712                         dma = (unsigned long) le64_to_cpu(
1713                                 dd->ipath_pioavailregs_dma[im]);
1714                         if (test_bit((INFINIPATH_SENDPIOAVAIL_CHECK_SHIFT
1715                                 + start) % BITS_PER_LONG, &dma))
1716                                 __set_bit(INFINIPATH_SENDPIOAVAIL_CHECK_SHIFT
1717                                         + start, dd->ipath_pioavailshadow);
1718                         else
1719                                 __clear_bit(INFINIPATH_SENDPIOAVAIL_CHECK_SHIFT
1720                                         + start, dd->ipath_pioavailshadow);
1721                         __set_bit(start, dd->ipath_pioavailkernel);
1722                 } else {
1723                         __set_bit(start + INFINIPATH_SENDPIOAVAIL_BUSY_SHIFT,
1724                                 dd->ipath_pioavailshadow);
1725                         __clear_bit(start, dd->ipath_pioavailkernel);
1726                 }
1727                 start += 2;
1728         }
1729
1730         if (dd->ipath_pioupd_thresh) {
1731                 end = 2 * (dd->ipath_piobcnt2k + dd->ipath_piobcnt4k);
1732                 cnt = bitmap_weight(dd->ipath_pioavailkernel, end);
1733         }
1734         spin_unlock_irqrestore(&ipath_pioavail_lock, flags);
1735
1736         /*
1737          * When moving buffers from kernel to user, if number assigned to
1738          * the user is less than the pio update threshold, and threshold
1739          * is supported (cnt was computed > 0), drop the update threshold
1740          * so we update at least once per allocated number of buffers.
1741          * In any case, if the kernel buffers are less than the threshold,
1742          * drop the threshold.  We don't bother increasing it, having once
1743          * decreased it, since it would typically just cycle back and forth.
1744          * If we don't decrease below buffers in use, we can wait a long
1745          * time for an update, until some other context uses PIO buffers.
1746          */
1747         if (!avail && len < cnt)
1748                 cnt = len;
1749         if (cnt < dd->ipath_pioupd_thresh) {
1750                 dd->ipath_pioupd_thresh = cnt;
1751                 ipath_dbg("Decreased pio update threshold to %u\n",
1752                         dd->ipath_pioupd_thresh);
1753                 spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
1754                 dd->ipath_sendctrl &= ~(INFINIPATH_S_UPDTHRESH_MASK
1755                         << INFINIPATH_S_UPDTHRESH_SHIFT);
1756                 dd->ipath_sendctrl |= dd->ipath_pioupd_thresh
1757                         << INFINIPATH_S_UPDTHRESH_SHIFT;
1758                 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
1759                         dd->ipath_sendctrl);
1760                 spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
1761         }
1762 }
1763
1764 /**
1765  * ipath_create_rcvhdrq - create a receive header queue
1766  * @dd: the infinipath device
1767  * @pd: the port data
1768  *
1769  * this must be contiguous memory (from an i/o perspective), and must be
1770  * DMA'able (which means for some systems, it will go through an IOMMU,
1771  * or be forced into a low address range).
1772  */
1773 int ipath_create_rcvhdrq(struct ipath_devdata *dd,
1774                          struct ipath_portdata *pd)
1775 {
1776         int ret = 0;
1777
1778         if (!pd->port_rcvhdrq) {
1779                 dma_addr_t phys_hdrqtail;
1780                 gfp_t gfp_flags = GFP_USER | __GFP_COMP;
1781                 int amt = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
1782                                 sizeof(u32), PAGE_SIZE);
1783
1784                 pd->port_rcvhdrq = dma_alloc_coherent(
1785                         &dd->pcidev->dev, amt, &pd->port_rcvhdrq_phys,
1786                         gfp_flags);
1787
1788                 if (!pd->port_rcvhdrq) {
1789                         ipath_dev_err(dd, "attempt to allocate %d bytes "
1790                                       "for port %u rcvhdrq failed\n",
1791                                       amt, pd->port_port);
1792                         ret = -ENOMEM;
1793                         goto bail;
1794                 }
1795
1796                 if (!(dd->ipath_flags & IPATH_NODMA_RTAIL)) {
1797                         pd->port_rcvhdrtail_kvaddr = dma_alloc_coherent(
1798                                 &dd->pcidev->dev, PAGE_SIZE, &phys_hdrqtail,
1799                                 GFP_KERNEL);
1800                         if (!pd->port_rcvhdrtail_kvaddr) {
1801                                 ipath_dev_err(dd, "attempt to allocate 1 page "
1802                                         "for port %u rcvhdrqtailaddr "
1803                                         "failed\n", pd->port_port);
1804                                 ret = -ENOMEM;
1805                                 dma_free_coherent(&dd->pcidev->dev, amt,
1806                                         pd->port_rcvhdrq,
1807                                         pd->port_rcvhdrq_phys);
1808                                 pd->port_rcvhdrq = NULL;
1809                                 goto bail;
1810                         }
1811                         pd->port_rcvhdrqtailaddr_phys = phys_hdrqtail;
1812                         ipath_cdbg(VERBOSE, "port %d hdrtailaddr, %llx "
1813                                    "physical\n", pd->port_port,
1814                                    (unsigned long long) phys_hdrqtail);
1815                 }
1816
1817                 pd->port_rcvhdrq_size = amt;
1818
1819                 ipath_cdbg(VERBOSE, "%d pages at %p (phys %lx) size=%lu "
1820                            "for port %u rcvhdr Q\n",
1821                            amt >> PAGE_SHIFT, pd->port_rcvhdrq,
1822                            (unsigned long) pd->port_rcvhdrq_phys,
1823                            (unsigned long) pd->port_rcvhdrq_size,
1824                            pd->port_port);
1825         }
1826         else
1827                 ipath_cdbg(VERBOSE, "reuse port %d rcvhdrq @%p %llx phys; "
1828                            "hdrtailaddr@%p %llx physical\n",
1829                            pd->port_port, pd->port_rcvhdrq,
1830                            (unsigned long long) pd->port_rcvhdrq_phys,
1831                            pd->port_rcvhdrtail_kvaddr, (unsigned long long)
1832                            pd->port_rcvhdrqtailaddr_phys);
1833
1834         /* clear for security and sanity on each use */
1835         memset(pd->port_rcvhdrq, 0, pd->port_rcvhdrq_size);
1836         if (pd->port_rcvhdrtail_kvaddr)
1837                 memset(pd->port_rcvhdrtail_kvaddr, 0, PAGE_SIZE);
1838
1839         /*
1840          * tell chip each time we init it, even if we are re-using previous
1841          * memory (we zero the register at process close)
1842          */
1843         ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdrtailaddr,
1844                               pd->port_port, pd->port_rcvhdrqtailaddr_phys);
1845         ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdraddr,
1846                               pd->port_port, pd->port_rcvhdrq_phys);
1847
1848 bail:
1849         return ret;
1850 }
1851
1852
1853 /*
1854  * Flush all sends that might be in the ready to send state, as well as any
1855  * that are in the process of being sent.   Used whenever we need to be
1856  * sure the send side is idle.  Cleans up all buffer state by canceling
1857  * all pio buffers, and issuing an abort, which cleans up anything in the
1858  * launch fifo.  The cancel is superfluous on some chip versions, but
1859  * it's safer to always do it.
1860  * PIOAvail bits are updated by the chip as if normal send had happened.
1861  */
1862 void ipath_cancel_sends(struct ipath_devdata *dd, int restore_sendctrl)
1863 {
1864         unsigned long flags;
1865
1866         if (dd->ipath_flags & IPATH_IB_AUTONEG_INPROG) {
1867                 ipath_cdbg(VERBOSE, "Ignore while in autonegotiation\n");
1868                 goto bail;
1869         }
1870         /*
1871          * If we have SDMA, and it's not disabled, we have to kick off the
1872          * abort state machine, provided we aren't already aborting.
1873          * If we are in the process of aborting SDMA (!DISABLED, but ABORTING),
1874          * we skip the rest of this routine. It is already "in progress"
1875          */
1876         if (dd->ipath_flags & IPATH_HAS_SEND_DMA) {
1877                 int skip_cancel;
1878                 unsigned long *statp = &dd->ipath_sdma_status;
1879
1880                 spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
1881                 skip_cancel =
1882                         test_and_set_bit(IPATH_SDMA_ABORTING, statp)
1883                         && !test_bit(IPATH_SDMA_DISABLED, statp);
1884                 spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
1885                 if (skip_cancel)
1886                         goto bail;
1887         }
1888
1889         ipath_dbg("Cancelling all in-progress send buffers\n");
1890
1891         /* skip armlaunch errs for a while */
1892         dd->ipath_lastcancel = jiffies + HZ / 2;
1893
1894         /*
1895          * The abort bit is auto-clearing.  We also don't want pioavail
1896          * update happening during this, and we don't want any other
1897          * sends going out, so turn those off for the duration.  We read
1898          * the scratch register to be sure that cancels and the abort
1899          * have taken effect in the chip.  Otherwise two parts are same
1900          * as ipath_force_pio_avail_update()
1901          */
1902         spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
1903         dd->ipath_sendctrl &= ~(INFINIPATH_S_PIOBUFAVAILUPD
1904                 | INFINIPATH_S_PIOENABLE);
1905         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
1906                 dd->ipath_sendctrl | INFINIPATH_S_ABORT);
1907         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
1908         spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
1909
1910         /* disarm all send buffers */
1911         ipath_disarm_piobufs(dd, 0,
1912                 dd->ipath_piobcnt2k + dd->ipath_piobcnt4k);
1913
1914         if (dd->ipath_flags & IPATH_HAS_SEND_DMA)
1915                 set_bit(IPATH_SDMA_DISARMED, &dd->ipath_sdma_status);
1916
1917         if (restore_sendctrl) {
1918                 /* else done by caller later if needed */
1919                 spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
1920                 dd->ipath_sendctrl |= INFINIPATH_S_PIOBUFAVAILUPD |
1921                         INFINIPATH_S_PIOENABLE;
1922                 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
1923                         dd->ipath_sendctrl);
1924                 /* and again, be sure all have hit the chip */
1925                 ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
1926                 spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
1927         }
1928
1929         if ((dd->ipath_flags & IPATH_HAS_SEND_DMA) &&
1930             !test_bit(IPATH_SDMA_DISABLED, &dd->ipath_sdma_status) &&
1931             test_bit(IPATH_SDMA_RUNNING, &dd->ipath_sdma_status)) {
1932                 spin_lock_irqsave(&dd->ipath_sdma_lock, flags);
1933                 /* only wait so long for intr */
1934                 dd->ipath_sdma_abort_intr_timeout = jiffies + HZ;
1935                 dd->ipath_sdma_reset_wait = 200;
1936                 if (!test_bit(IPATH_SDMA_SHUTDOWN, &dd->ipath_sdma_status))
1937                         tasklet_hi_schedule(&dd->ipath_sdma_abort_task);
1938                 spin_unlock_irqrestore(&dd->ipath_sdma_lock, flags);
1939         }
1940 bail:;
1941 }
1942
1943 /*
1944  * Force an update of in-memory copy of the pioavail registers, when
1945  * needed for any of a variety of reasons.  We read the scratch register
1946  * to make it highly likely that the update will have happened by the
1947  * time we return.  If already off (as in cancel_sends above), this
1948  * routine is a nop, on the assumption that the caller will "do the
1949  * right thing".
1950  */
1951 void ipath_force_pio_avail_update(struct ipath_devdata *dd)
1952 {
1953         unsigned long flags;
1954
1955         spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
1956         if (dd->ipath_sendctrl & INFINIPATH_S_PIOBUFAVAILUPD) {
1957                 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
1958                         dd->ipath_sendctrl & ~INFINIPATH_S_PIOBUFAVAILUPD);
1959                 ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
1960                 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
1961                         dd->ipath_sendctrl);
1962                 ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
1963         }
1964         spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
1965 }
1966
1967 static void ipath_set_ib_lstate(struct ipath_devdata *dd, int linkcmd,
1968                                 int linitcmd)
1969 {
1970         u64 mod_wd;
1971         static const char *what[4] = {
1972                 [0] = "NOP",
1973                 [INFINIPATH_IBCC_LINKCMD_DOWN] = "DOWN",
1974                 [INFINIPATH_IBCC_LINKCMD_ARMED] = "ARMED",
1975                 [INFINIPATH_IBCC_LINKCMD_ACTIVE] = "ACTIVE"
1976         };
1977
1978         if (linitcmd == INFINIPATH_IBCC_LINKINITCMD_DISABLE) {
1979                 /*
1980                  * If we are told to disable, note that so link-recovery
1981                  * code does not attempt to bring us back up.
1982                  */
1983                 preempt_disable();
1984                 dd->ipath_flags |= IPATH_IB_LINK_DISABLED;
1985                 preempt_enable();
1986         } else if (linitcmd) {
1987                 /*
1988                  * Any other linkinitcmd will lead to LINKDOWN and then
1989                  * to INIT (if all is well), so clear flag to let
1990                  * link-recovery code attempt to bring us back up.
1991                  */
1992                 preempt_disable();
1993                 dd->ipath_flags &= ~IPATH_IB_LINK_DISABLED;
1994                 preempt_enable();
1995         }
1996
1997         mod_wd = (linkcmd << dd->ibcc_lc_shift) |
1998                 (linitcmd << INFINIPATH_IBCC_LINKINITCMD_SHIFT);
1999         ipath_cdbg(VERBOSE,
2000                 "Moving unit %u to %s (initcmd=0x%x), current ltstate is %s\n",
2001                 dd->ipath_unit, what[linkcmd], linitcmd,
2002                 ipath_ibcstatus_str[ipath_ib_linktrstate(dd,
2003                         ipath_read_kreg64(dd, dd->ipath_kregs->kr_ibcstatus))]);
2004
2005         ipath_write_kreg(dd, dd->ipath_kregs->kr_ibcctrl,
2006                          dd->ipath_ibcctrl | mod_wd);
2007         /* read from chip so write is flushed */
2008         (void) ipath_read_kreg64(dd, dd->ipath_kregs->kr_ibcstatus);
2009 }
2010
2011 int ipath_set_linkstate(struct ipath_devdata *dd, u8 newstate)
2012 {
2013         u32 lstate;
2014         int ret;
2015
2016         switch (newstate) {
2017         case IPATH_IB_LINKDOWN_ONLY:
2018                 ipath_set_ib_lstate(dd, INFINIPATH_IBCC_LINKCMD_DOWN, 0);
2019                 /* don't wait */
2020                 ret = 0;
2021                 goto bail;
2022
2023         case IPATH_IB_LINKDOWN:
2024                 ipath_set_ib_lstate(dd, INFINIPATH_IBCC_LINKCMD_DOWN,
2025                                         INFINIPATH_IBCC_LINKINITCMD_POLL);
2026                 /* don't wait */
2027                 ret = 0;
2028                 goto bail;
2029
2030         case IPATH_IB_LINKDOWN_SLEEP:
2031                 ipath_set_ib_lstate(dd, INFINIPATH_IBCC_LINKCMD_DOWN,
2032                                         INFINIPATH_IBCC_LINKINITCMD_SLEEP);
2033                 /* don't wait */
2034                 ret = 0;
2035                 goto bail;
2036
2037         case IPATH_IB_LINKDOWN_DISABLE:
2038                 ipath_set_ib_lstate(dd, INFINIPATH_IBCC_LINKCMD_DOWN,
2039                                         INFINIPATH_IBCC_LINKINITCMD_DISABLE);
2040                 /* don't wait */
2041                 ret = 0;
2042                 goto bail;
2043
2044         case IPATH_IB_LINKARM:
2045                 if (dd->ipath_flags & IPATH_LINKARMED) {
2046                         ret = 0;
2047                         goto bail;
2048                 }
2049                 if (!(dd->ipath_flags &
2050                       (IPATH_LINKINIT | IPATH_LINKACTIVE))) {
2051                         ret = -EINVAL;
2052                         goto bail;
2053                 }
2054                 ipath_set_ib_lstate(dd, INFINIPATH_IBCC_LINKCMD_ARMED, 0);
2055
2056                 /*
2057                  * Since the port can transition to ACTIVE by receiving
2058                  * a non VL 15 packet, wait for either state.
2059                  */
2060                 lstate = IPATH_LINKARMED | IPATH_LINKACTIVE;
2061                 break;
2062
2063         case IPATH_IB_LINKACTIVE:
2064                 if (dd->ipath_flags & IPATH_LINKACTIVE) {
2065                         ret = 0;
2066                         goto bail;
2067                 }
2068                 if (!(dd->ipath_flags & IPATH_LINKARMED)) {
2069                         ret = -EINVAL;
2070                         goto bail;
2071                 }
2072                 ipath_set_ib_lstate(dd, INFINIPATH_IBCC_LINKCMD_ACTIVE, 0);
2073                 lstate = IPATH_LINKACTIVE;
2074                 break;
2075
2076         case IPATH_IB_LINK_LOOPBACK:
2077                 dev_info(&dd->pcidev->dev, "Enabling IB local loopback\n");
2078                 dd->ipath_ibcctrl |= INFINIPATH_IBCC_LOOPBACK;
2079                 ipath_write_kreg(dd, dd->ipath_kregs->kr_ibcctrl,
2080                                  dd->ipath_ibcctrl);
2081
2082                 /* turn heartbeat off, as it causes loopback to fail */
2083                 dd->ipath_f_set_ib_cfg(dd, IPATH_IB_CFG_HRTBT,
2084                                        IPATH_IB_HRTBT_OFF);
2085                 /* don't wait */
2086                 ret = 0;
2087                 goto bail;
2088
2089         case IPATH_IB_LINK_EXTERNAL:
2090                 dev_info(&dd->pcidev->dev,
2091                         "Disabling IB local loopback (normal)\n");
2092                 dd->ipath_f_set_ib_cfg(dd, IPATH_IB_CFG_HRTBT,
2093                                        IPATH_IB_HRTBT_ON);
2094                 dd->ipath_ibcctrl &= ~INFINIPATH_IBCC_LOOPBACK;
2095                 ipath_write_kreg(dd, dd->ipath_kregs->kr_ibcctrl,
2096                                  dd->ipath_ibcctrl);
2097                 /* don't wait */
2098                 ret = 0;
2099                 goto bail;
2100
2101         /*
2102          * Heartbeat can be explicitly enabled by the user via
2103          * "hrtbt_enable" "file", and if disabled, trying to enable here
2104          * will have no effect.  Implicit changes (heartbeat off when
2105          * loopback on, and vice versa) are included to ease testing.
2106          */
2107         case IPATH_IB_LINK_HRTBT:
2108                 ret = dd->ipath_f_set_ib_cfg(dd, IPATH_IB_CFG_HRTBT,
2109                         IPATH_IB_HRTBT_ON);
2110                 goto bail;
2111
2112         case IPATH_IB_LINK_NO_HRTBT:
2113                 ret = dd->ipath_f_set_ib_cfg(dd, IPATH_IB_CFG_HRTBT,
2114                         IPATH_IB_HRTBT_OFF);
2115                 goto bail;
2116
2117         default:
2118                 ipath_dbg("Invalid linkstate 0x%x requested\n", newstate);
2119                 ret = -EINVAL;
2120                 goto bail;
2121         }
2122         ret = ipath_wait_linkstate(dd, lstate, 2000);
2123
2124 bail:
2125         return ret;
2126 }
2127
2128 /**
2129  * ipath_set_mtu - set the MTU
2130  * @dd: the infinipath device
2131  * @arg: the new MTU
2132  *
2133  * we can handle "any" incoming size, the issue here is whether we
2134  * need to restrict our outgoing size.   For now, we don't do any
2135  * sanity checking on this, and we don't deal with what happens to
2136  * programs that are already running when the size changes.
2137  * NOTE: changing the MTU will usually cause the IBC to go back to
2138  * link INIT state...
2139  */
2140 int ipath_set_mtu(struct ipath_devdata *dd, u16 arg)
2141 {
2142         u32 piosize;
2143         int changed = 0;
2144         int ret;
2145
2146         /*
2147          * mtu is IB data payload max.  It's the largest power of 2 less
2148          * than piosize (or even larger, since it only really controls the
2149          * largest we can receive; we can send the max of the mtu and
2150          * piosize).  We check that it's one of the valid IB sizes.
2151          */
2152         if (arg != 256 && arg != 512 && arg != 1024 && arg != 2048 &&
2153             (arg != 4096 || !ipath_mtu4096)) {
2154                 ipath_dbg("Trying to set invalid mtu %u, failing\n", arg);
2155                 ret = -EINVAL;
2156                 goto bail;
2157         }
2158         if (dd->ipath_ibmtu == arg) {
2159                 ret = 0;        /* same as current */
2160                 goto bail;
2161         }
2162
2163         piosize = dd->ipath_ibmaxlen;
2164         dd->ipath_ibmtu = arg;
2165
2166         if (arg >= (piosize - IPATH_PIO_MAXIBHDR)) {
2167                 /* Only if it's not the initial value (or reset to it) */
2168                 if (piosize != dd->ipath_init_ibmaxlen) {
2169                         if (arg > piosize && arg <= dd->ipath_init_ibmaxlen)
2170                                 piosize = dd->ipath_init_ibmaxlen;
2171                         dd->ipath_ibmaxlen = piosize;
2172                         changed = 1;
2173                 }
2174         } else if ((arg + IPATH_PIO_MAXIBHDR) != dd->ipath_ibmaxlen) {
2175                 piosize = arg + IPATH_PIO_MAXIBHDR;
2176                 ipath_cdbg(VERBOSE, "ibmaxlen was 0x%x, setting to 0x%x "
2177                            "(mtu 0x%x)\n", dd->ipath_ibmaxlen, piosize,
2178                            arg);
2179                 dd->ipath_ibmaxlen = piosize;
2180                 changed = 1;
2181         }
2182
2183         if (changed) {
2184                 u64 ibc = dd->ipath_ibcctrl, ibdw;
2185                 /*
2186                  * update our housekeeping variables, and set IBC max
2187                  * size, same as init code; max IBC is max we allow in
2188                  * buffer, less the qword pbc, plus 1 for ICRC, in dwords
2189                  */
2190                 dd->ipath_ibmaxlen = piosize - 2 * sizeof(u32);
2191                 ibdw = (dd->ipath_ibmaxlen >> 2) + 1;
2192                 ibc &= ~(INFINIPATH_IBCC_MAXPKTLEN_MASK <<
2193                          dd->ibcc_mpl_shift);
2194                 ibc |= ibdw << dd->ibcc_mpl_shift;
2195                 dd->ipath_ibcctrl = ibc;
2196                 ipath_write_kreg(dd, dd->ipath_kregs->kr_ibcctrl,
2197                                  dd->ipath_ibcctrl);
2198                 dd->ipath_f_tidtemplate(dd);
2199         }
2200
2201         ret = 0;
2202
2203 bail:
2204         return ret;
2205 }
2206
2207 int ipath_set_lid(struct ipath_devdata *dd, u32 lid, u8 lmc)
2208 {
2209         dd->ipath_lid = lid;
2210         dd->ipath_lmc = lmc;
2211
2212         dd->ipath_f_set_ib_cfg(dd, IPATH_IB_CFG_LIDLMC, lid |
2213                 (~((1U << lmc) - 1)) << 16);
2214
2215         dev_info(&dd->pcidev->dev, "We got a lid: 0x%x\n", lid);
2216
2217         return 0;
2218 }
2219
2220
2221 /**
2222  * ipath_write_kreg_port - write a device's per-port 64-bit kernel register
2223  * @dd: the infinipath device
2224  * @regno: the register number to write
2225  * @port: the port containing the register
2226  * @value: the value to write
2227  *
2228  * Registers that vary with the chip implementation constants (port)
2229  * use this routine.
2230  */
2231 void ipath_write_kreg_port(const struct ipath_devdata *dd, ipath_kreg regno,
2232                           unsigned port, u64 value)
2233 {
2234         u16 where;
2235
2236         if (port < dd->ipath_portcnt &&
2237             (regno == dd->ipath_kregs->kr_rcvhdraddr ||
2238              regno == dd->ipath_kregs->kr_rcvhdrtailaddr))
2239                 where = regno + port;
2240         else
2241                 where = -1;
2242
2243         ipath_write_kreg(dd, where, value);
2244 }
2245
2246 /*
2247  * Following deal with the "obviously simple" task of overriding the state
2248  * of the LEDS, which normally indicate link physical and logical status.
2249  * The complications arise in dealing with different hardware mappings
2250  * and the board-dependent routine being called from interrupts.
2251  * and then there's the requirement to _flash_ them.
2252  */
2253 #define LED_OVER_FREQ_SHIFT 8
2254 #define LED_OVER_FREQ_MASK (0xFF<<LED_OVER_FREQ_SHIFT)
2255 /* Below is "non-zero" to force override, but both actual LEDs are off */
2256 #define LED_OVER_BOTH_OFF (8)
2257
2258 static void ipath_run_led_override(unsigned long opaque)
2259 {
2260         struct ipath_devdata *dd = (struct ipath_devdata *)opaque;
2261         int timeoff;
2262         int pidx;
2263         u64 lstate, ltstate, val;
2264
2265         if (!(dd->ipath_flags & IPATH_INITTED))
2266                 return;
2267
2268         pidx = dd->ipath_led_override_phase++ & 1;
2269         dd->ipath_led_override = dd->ipath_led_override_vals[pidx];
2270         timeoff = dd->ipath_led_override_timeoff;
2271
2272         /*
2273          * below potentially restores the LED values per current status,
2274          * should also possibly setup the traffic-blink register,
2275          * but leave that to per-chip functions.
2276          */
2277         val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_ibcstatus);
2278         ltstate = ipath_ib_linktrstate(dd, val);
2279         lstate = ipath_ib_linkstate(dd, val);
2280
2281         dd->ipath_f_setextled(dd, lstate, ltstate);
2282         mod_timer(&dd->ipath_led_override_timer, jiffies + timeoff);
2283 }
2284
2285 void ipath_set_led_override(struct ipath_devdata *dd, unsigned int val)
2286 {
2287         int timeoff, freq;
2288
2289         if (!(dd->ipath_flags & IPATH_INITTED))
2290                 return;
2291
2292         /* First check if we are blinking. If not, use 1HZ polling */
2293         timeoff = HZ;
2294         freq = (val & LED_OVER_FREQ_MASK) >> LED_OVER_FREQ_SHIFT;
2295
2296         if (freq) {
2297                 /* For blink, set each phase from one nybble of val */
2298                 dd->ipath_led_override_vals[0] = val & 0xF;
2299                 dd->ipath_led_override_vals[1] = (val >> 4) & 0xF;
2300                 timeoff = (HZ << 4)/freq;
2301         } else {
2302                 /* Non-blink set both phases the same. */
2303                 dd->ipath_led_override_vals[0] = val & 0xF;
2304                 dd->ipath_led_override_vals[1] = val & 0xF;
2305         }
2306         dd->ipath_led_override_timeoff = timeoff;
2307
2308         /*
2309          * If the timer has not already been started, do so. Use a "quick"
2310          * timeout so the function will be called soon, to look at our request.
2311          */
2312         if (atomic_inc_return(&dd->ipath_led_override_timer_active) == 1) {
2313                 /* Need to start timer */
2314                 init_timer(&dd->ipath_led_override_timer);
2315                 dd->ipath_led_override_timer.function =
2316                                                  ipath_run_led_override;
2317                 dd->ipath_led_override_timer.data = (unsigned long) dd;
2318                 dd->ipath_led_override_timer.expires = jiffies + 1;
2319                 add_timer(&dd->ipath_led_override_timer);
2320         } else
2321                 atomic_dec(&dd->ipath_led_override_timer_active);
2322 }
2323
2324 /**
2325  * ipath_shutdown_device - shut down a device
2326  * @dd: the infinipath device
2327  *
2328  * This is called to make the device quiet when we are about to
2329  * unload the driver, and also when the device is administratively
2330  * disabled.   It does not free any data structures.
2331  * Everything it does has to be setup again by ipath_init_chip(dd,1)
2332  */
2333 void ipath_shutdown_device(struct ipath_devdata *dd)
2334 {
2335         unsigned long flags;
2336
2337         ipath_dbg("Shutting down the device\n");
2338
2339         ipath_hol_up(dd); /* make sure user processes aren't suspended */
2340
2341         dd->ipath_flags |= IPATH_LINKUNK;
2342         dd->ipath_flags &= ~(IPATH_INITTED | IPATH_LINKDOWN |
2343                              IPATH_LINKINIT | IPATH_LINKARMED |
2344                              IPATH_LINKACTIVE);
2345         *dd->ipath_statusp &= ~(IPATH_STATUS_IB_CONF |
2346                                 IPATH_STATUS_IB_READY);
2347
2348         /* mask interrupts, but not errors */
2349         ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask, 0ULL);
2350
2351         dd->ipath_rcvctrl = 0;
2352         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
2353                          dd->ipath_rcvctrl);
2354
2355         if (dd->ipath_flags & IPATH_HAS_SEND_DMA)
2356                 teardown_sdma(dd);
2357
2358         /*
2359          * gracefully stop all sends allowing any in progress to trickle out
2360          * first.
2361          */
2362         spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
2363         dd->ipath_sendctrl = 0;
2364         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, dd->ipath_sendctrl);
2365         /* flush it */
2366         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
2367         spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
2368
2369         /*
2370          * enough for anything that's going to trickle out to have actually
2371          * done so.
2372          */
2373         udelay(5);
2374
2375         dd->ipath_f_setextled(dd, 0, 0); /* make sure LEDs are off */
2376
2377         ipath_set_ib_lstate(dd, 0, INFINIPATH_IBCC_LINKINITCMD_DISABLE);
2378         ipath_cancel_sends(dd, 0);
2379
2380         /*
2381          * we are shutting down, so tell components that care.  We don't do
2382          * this on just a link state change, much like ethernet, a cable
2383          * unplug, etc. doesn't change driver state
2384          */
2385         signal_ib_event(dd, IB_EVENT_PORT_ERR);
2386
2387         /* disable IBC */
2388         dd->ipath_control &= ~INFINIPATH_C_LINKENABLE;
2389         ipath_write_kreg(dd, dd->ipath_kregs->kr_control,
2390                          dd->ipath_control | INFINIPATH_C_FREEZEMODE);
2391
2392         /*
2393          * clear SerdesEnable and turn the leds off; do this here because
2394          * we are unloading, so don't count on interrupts to move along
2395          * Turn the LEDs off explictly for the same reason.
2396          */
2397         dd->ipath_f_quiet_serdes(dd);
2398
2399         /* stop all the timers that might still be running */
2400         del_timer_sync(&dd->ipath_hol_timer);
2401         if (dd->ipath_stats_timer_active) {
2402                 del_timer_sync(&dd->ipath_stats_timer);
2403                 dd->ipath_stats_timer_active = 0;
2404         }
2405         if (dd->ipath_intrchk_timer.data) {
2406                 del_timer_sync(&dd->ipath_intrchk_timer);
2407                 dd->ipath_intrchk_timer.data = 0;
2408         }
2409         if (atomic_read(&dd->ipath_led_override_timer_active)) {
2410                 del_timer_sync(&dd->ipath_led_override_timer);
2411                 atomic_set(&dd->ipath_led_override_timer_active, 0);
2412         }
2413
2414         /*
2415          * clear all interrupts and errors, so that the next time the driver
2416          * is loaded or device is enabled, we know that whatever is set
2417          * happened while we were unloaded
2418          */
2419         ipath_write_kreg(dd, dd->ipath_kregs->kr_hwerrclear,
2420                          ~0ULL & ~INFINIPATH_HWE_MEMBISTFAILED);
2421         ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear, -1LL);
2422         ipath_write_kreg(dd, dd->ipath_kregs->kr_intclear, -1LL);
2423
2424         ipath_cdbg(VERBOSE, "Flush time and errors to EEPROM\n");
2425         ipath_update_eeprom_log(dd);
2426 }
2427
2428 /**
2429  * ipath_free_pddata - free a port's allocated data
2430  * @dd: the infinipath device
2431  * @pd: the portdata structure
2432  *
2433  * free up any allocated data for a port
2434  * This should not touch anything that would affect a simultaneous
2435  * re-allocation of port data, because it is called after ipath_mutex
2436  * is released (and can be called from reinit as well).
2437  * It should never change any chip state, or global driver state.
2438  * (The only exception to global state is freeing the port0 port0_skbs.)
2439  */
2440 void ipath_free_pddata(struct ipath_devdata *dd, struct ipath_portdata *pd)
2441 {
2442         if (!pd)
2443                 return;
2444
2445         if (pd->port_rcvhdrq) {
2446                 ipath_cdbg(VERBOSE, "free closed port %d rcvhdrq @ %p "
2447                            "(size=%lu)\n", pd->port_port, pd->port_rcvhdrq,
2448                            (unsigned long) pd->port_rcvhdrq_size);
2449                 dma_free_coherent(&dd->pcidev->dev, pd->port_rcvhdrq_size,
2450                                   pd->port_rcvhdrq, pd->port_rcvhdrq_phys);
2451                 pd->port_rcvhdrq = NULL;
2452                 if (pd->port_rcvhdrtail_kvaddr) {
2453                         dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE,
2454                                          pd->port_rcvhdrtail_kvaddr,
2455                                          pd->port_rcvhdrqtailaddr_phys);
2456                         pd->port_rcvhdrtail_kvaddr = NULL;
2457                 }
2458         }
2459         if (pd->port_port && pd->port_rcvegrbuf) {
2460                 unsigned e;
2461
2462                 for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
2463                         void *base = pd->port_rcvegrbuf[e];
2464                         size_t size = pd->port_rcvegrbuf_size;
2465
2466                         ipath_cdbg(VERBOSE, "egrbuf free(%p, %lu), "
2467                                    "chunk %u/%u\n", base,
2468                                    (unsigned long) size,
2469                                    e, pd->port_rcvegrbuf_chunks);
2470                         dma_free_coherent(&dd->pcidev->dev, size,
2471                                 base, pd->port_rcvegrbuf_phys[e]);
2472                 }
2473                 kfree(pd->port_rcvegrbuf);
2474                 pd->port_rcvegrbuf = NULL;
2475                 kfree(pd->port_rcvegrbuf_phys);
2476                 pd->port_rcvegrbuf_phys = NULL;
2477                 pd->port_rcvegrbuf_chunks = 0;
2478         } else if (pd->port_port == 0 && dd->ipath_port0_skbinfo) {
2479                 unsigned e;
2480                 struct ipath_skbinfo *skbinfo = dd->ipath_port0_skbinfo;
2481
2482                 dd->ipath_port0_skbinfo = NULL;
2483                 ipath_cdbg(VERBOSE, "free closed port %d "
2484                            "ipath_port0_skbinfo @ %p\n", pd->port_port,
2485                            skbinfo);
2486                 for (e = 0; e < dd->ipath_p0_rcvegrcnt; e++)
2487                         if (skbinfo[e].skb) {
2488                                 pci_unmap_single(dd->pcidev, skbinfo[e].phys,
2489                                                  dd->ipath_ibmaxlen,
2490                                                  PCI_DMA_FROMDEVICE);
2491                                 dev_kfree_skb(skbinfo[e].skb);
2492                         }
2493                 vfree(skbinfo);
2494         }
2495         kfree(pd->port_tid_pg_list);
2496         vfree(pd->subport_uregbase);
2497         vfree(pd->subport_rcvegrbuf);
2498         vfree(pd->subport_rcvhdr_base);
2499         kfree(pd);
2500 }
2501
2502 static int __init infinipath_init(void)
2503 {
2504         int ret;
2505
2506         if (ipath_debug & __IPATH_DBG)
2507                 printk(KERN_INFO DRIVER_LOAD_MSG "%s", ib_ipath_version);
2508
2509         /*
2510          * These must be called before the driver is registered with
2511          * the PCI subsystem.
2512          */
2513         idr_init(&unit_table);
2514         if (!idr_pre_get(&unit_table, GFP_KERNEL)) {
2515                 printk(KERN_ERR IPATH_DRV_NAME ": idr_pre_get() failed\n");
2516                 ret = -ENOMEM;
2517                 goto bail;
2518         }
2519
2520         ret = pci_register_driver(&ipath_driver);
2521         if (ret < 0) {
2522                 printk(KERN_ERR IPATH_DRV_NAME
2523                        ": Unable to register driver: error %d\n", -ret);
2524                 goto bail_unit;
2525         }
2526
2527         ret = ipath_init_ipathfs();
2528         if (ret < 0) {
2529                 printk(KERN_ERR IPATH_DRV_NAME ": Unable to create "
2530                        "ipathfs: error %d\n", -ret);
2531                 goto bail_pci;
2532         }
2533
2534         goto bail;
2535
2536 bail_pci:
2537         pci_unregister_driver(&ipath_driver);
2538
2539 bail_unit:
2540         idr_destroy(&unit_table);
2541
2542 bail:
2543         return ret;
2544 }
2545
2546 static void __exit infinipath_cleanup(void)
2547 {
2548         ipath_exit_ipathfs();
2549
2550         ipath_cdbg(VERBOSE, "Unregistering pci driver\n");
2551         pci_unregister_driver(&ipath_driver);
2552
2553         idr_destroy(&unit_table);
2554 }
2555
2556 /**
2557  * ipath_reset_device - reset the chip if possible
2558  * @unit: the device to reset
2559  *
2560  * Whether or not reset is successful, we attempt to re-initialize the chip
2561  * (that is, much like a driver unload/reload).  We clear the INITTED flag
2562  * so that the various entry points will fail until we reinitialize.  For
2563  * now, we only allow this if no user ports are open that use chip resources
2564  */
2565 int ipath_reset_device(int unit)
2566 {
2567         int ret, i;
2568         struct ipath_devdata *dd = ipath_lookup(unit);
2569         unsigned long flags;
2570
2571         if (!dd) {
2572                 ret = -ENODEV;
2573                 goto bail;
2574         }
2575
2576         if (atomic_read(&dd->ipath_led_override_timer_active)) {
2577                 /* Need to stop LED timer, _then_ shut off LEDs */
2578                 del_timer_sync(&dd->ipath_led_override_timer);
2579                 atomic_set(&dd->ipath_led_override_timer_active, 0);
2580         }
2581
2582         /* Shut off LEDs after we are sure timer is not running */
2583         dd->ipath_led_override = LED_OVER_BOTH_OFF;
2584         dd->ipath_f_setextled(dd, 0, 0);
2585
2586         dev_info(&dd->pcidev->dev, "Reset on unit %u requested\n", unit);
2587
2588         if (!dd->ipath_kregbase || !(dd->ipath_flags & IPATH_PRESENT)) {
2589                 dev_info(&dd->pcidev->dev, "Invalid unit number %u or "
2590                          "not initialized or not present\n", unit);
2591                 ret = -ENXIO;
2592                 goto bail;
2593         }
2594
2595         spin_lock_irqsave(&dd->ipath_uctxt_lock, flags);
2596         if (dd->ipath_pd)
2597                 for (i = 1; i < dd->ipath_cfgports; i++) {
2598                         if (!dd->ipath_pd[i] || !dd->ipath_pd[i]->port_cnt)
2599                                 continue;
2600                         spin_unlock_irqrestore(&dd->ipath_uctxt_lock, flags);
2601                         ipath_dbg("unit %u port %d is in use "
2602                                   "(PID %u cmd %s), can't reset\n",
2603                                   unit, i,
2604                                   pid_nr(dd->ipath_pd[i]->port_pid),
2605                                   dd->ipath_pd[i]->port_comm);
2606                         ret = -EBUSY;
2607                         goto bail;
2608                 }
2609         spin_unlock_irqrestore(&dd->ipath_uctxt_lock, flags);
2610
2611         if (dd->ipath_flags & IPATH_HAS_SEND_DMA)
2612                 teardown_sdma(dd);
2613
2614         dd->ipath_flags &= ~IPATH_INITTED;
2615         ipath_write_kreg(dd, dd->ipath_kregs->kr_intmask, 0ULL);
2616         ret = dd->ipath_f_reset(dd);
2617         if (ret == 1) {
2618                 ipath_dbg("Reinitializing unit %u after reset attempt\n",
2619                           unit);
2620                 ret = ipath_init_chip(dd, 1);
2621         } else
2622                 ret = -EAGAIN;
2623         if (ret)
2624                 ipath_dev_err(dd, "Reinitialize unit %u after "
2625                               "reset failed with %d\n", unit, ret);
2626         else
2627                 dev_info(&dd->pcidev->dev, "Reinitialized unit %u after "
2628                          "resetting\n", unit);
2629
2630 bail:
2631         return ret;
2632 }
2633
2634 /*
2635  * send a signal to all the processes that have the driver open
2636  * through the normal interfaces (i.e., everything other than diags
2637  * interface).  Returns number of signalled processes.
2638  */
2639 static int ipath_signal_procs(struct ipath_devdata *dd, int sig)
2640 {
2641         int i, sub, any = 0;
2642         struct pid *pid;
2643         unsigned long flags;
2644
2645         if (!dd->ipath_pd)
2646                 return 0;
2647
2648         spin_lock_irqsave(&dd->ipath_uctxt_lock, flags);
2649         for (i = 1; i < dd->ipath_cfgports; i++) {
2650                 if (!dd->ipath_pd[i] || !dd->ipath_pd[i]->port_cnt)
2651                         continue;
2652                 pid = dd->ipath_pd[i]->port_pid;
2653                 if (!pid)
2654                         continue;
2655
2656                 dev_info(&dd->pcidev->dev, "context %d in use "
2657                           "(PID %u), sending signal %d\n",
2658                           i, pid_nr(pid), sig);
2659                 kill_pid(pid, sig, 1);
2660                 any++;
2661                 for (sub = 0; sub < INFINIPATH_MAX_SUBPORT; sub++) {
2662                         pid = dd->ipath_pd[i]->port_subpid[sub];
2663                         if (!pid)
2664                                 continue;
2665                         dev_info(&dd->pcidev->dev, "sub-context "
2666                                 "%d:%d in use (PID %u), sending "
2667                                 "signal %d\n", i, sub, pid_nr(pid), sig);
2668                         kill_pid(pid, sig, 1);
2669                         any++;
2670                 }
2671         }
2672         spin_unlock_irqrestore(&dd->ipath_uctxt_lock, flags);
2673         return any;
2674 }
2675
2676 static void ipath_hol_signal_down(struct ipath_devdata *dd)
2677 {
2678         if (ipath_signal_procs(dd, SIGSTOP))
2679                 ipath_dbg("Stopped some processes\n");
2680         ipath_cancel_sends(dd, 1);
2681 }
2682
2683
2684 static void ipath_hol_signal_up(struct ipath_devdata *dd)
2685 {
2686         if (ipath_signal_procs(dd, SIGCONT))
2687                 ipath_dbg("Continued some processes\n");
2688 }
2689
2690 /*
2691  * link is down, stop any users processes, and flush pending sends
2692  * to prevent HoL blocking, then start the HoL timer that
2693  * periodically continues, then stop procs, so they can detect
2694  * link down if they want, and do something about it.
2695  * Timer may already be running, so use mod_timer, not add_timer.
2696  */
2697 void ipath_hol_down(struct ipath_devdata *dd)
2698 {
2699         dd->ipath_hol_state = IPATH_HOL_DOWN;
2700         ipath_hol_signal_down(dd);
2701         dd->ipath_hol_next = IPATH_HOL_DOWNCONT;
2702         dd->ipath_hol_timer.expires = jiffies +
2703                 msecs_to_jiffies(ipath_hol_timeout_ms);
2704         mod_timer(&dd->ipath_hol_timer, dd->ipath_hol_timer.expires);
2705 }
2706
2707 /*
2708  * link is up, continue any user processes, and ensure timer
2709  * is a nop, if running.  Let timer keep running, if set; it
2710  * will nop when it sees the link is up
2711  */
2712 void ipath_hol_up(struct ipath_devdata *dd)
2713 {
2714         ipath_hol_signal_up(dd);
2715         dd->ipath_hol_state = IPATH_HOL_UP;
2716 }
2717
2718 /*
2719  * toggle the running/not running state of user proceses
2720  * to prevent HoL blocking on chip resources, but still allow
2721  * user processes to do link down special case handling.
2722  * Should only be called via the timer
2723  */
2724 void ipath_hol_event(unsigned long opaque)
2725 {
2726         struct ipath_devdata *dd = (struct ipath_devdata *)opaque;
2727
2728         if (dd->ipath_hol_next == IPATH_HOL_DOWNSTOP
2729                 && dd->ipath_hol_state != IPATH_HOL_UP) {
2730                 dd->ipath_hol_next = IPATH_HOL_DOWNCONT;
2731                 ipath_dbg("Stopping processes\n");
2732                 ipath_hol_signal_down(dd);
2733         } else { /* may do "extra" if also in ipath_hol_up() */
2734                 dd->ipath_hol_next = IPATH_HOL_DOWNSTOP;
2735                 ipath_dbg("Continuing processes\n");
2736                 ipath_hol_signal_up(dd);
2737         }
2738         if (dd->ipath_hol_state == IPATH_HOL_UP)
2739                 ipath_dbg("link's up, don't resched timer\n");
2740         else {
2741                 dd->ipath_hol_timer.expires = jiffies +
2742                         msecs_to_jiffies(ipath_hol_timeout_ms);
2743                 mod_timer(&dd->ipath_hol_timer,
2744                         dd->ipath_hol_timer.expires);
2745         }
2746 }
2747
2748 int ipath_set_rx_pol_inv(struct ipath_devdata *dd, u8 new_pol_inv)
2749 {
2750         u64 val;
2751
2752         if (new_pol_inv > INFINIPATH_XGXS_RX_POL_MASK)
2753                 return -1;
2754         if (dd->ipath_rx_pol_inv != new_pol_inv) {
2755                 dd->ipath_rx_pol_inv = new_pol_inv;
2756                 val = ipath_read_kreg64(dd, dd->ipath_kregs->kr_xgxsconfig);
2757                 val &= ~(INFINIPATH_XGXS_RX_POL_MASK <<
2758                          INFINIPATH_XGXS_RX_POL_SHIFT);
2759                 val |= ((u64)dd->ipath_rx_pol_inv) <<
2760                         INFINIPATH_XGXS_RX_POL_SHIFT;
2761                 ipath_write_kreg(dd, dd->ipath_kregs->kr_xgxsconfig, val);
2762         }
2763         return 0;
2764 }
2765
2766 /*
2767  * Disable and enable the armlaunch error.  Used for PIO bandwidth testing on
2768  * the 7220, which is count-based, rather than trigger-based.  Safe for the
2769  * driver check, since it's at init.   Not completely safe when used for
2770  * user-mode checking, since some error checking can be lost, but not
2771  * particularly risky, and only has problematic side-effects in the face of
2772  * very buggy user code.  There is no reference counting, but that's also
2773  * fine, given the intended use.
2774  */
2775 void ipath_enable_armlaunch(struct ipath_devdata *dd)
2776 {
2777         dd->ipath_lasterror &= ~INFINIPATH_E_SPIOARMLAUNCH;
2778         ipath_write_kreg(dd, dd->ipath_kregs->kr_errorclear,
2779                 INFINIPATH_E_SPIOARMLAUNCH);
2780         dd->ipath_errormask |= INFINIPATH_E_SPIOARMLAUNCH;
2781         ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
2782                 dd->ipath_errormask);
2783 }
2784
2785 void ipath_disable_armlaunch(struct ipath_devdata *dd)
2786 {
2787         /* so don't re-enable if already set */
2788         dd->ipath_maskederrs &= ~INFINIPATH_E_SPIOARMLAUNCH;
2789         dd->ipath_errormask &= ~INFINIPATH_E_SPIOARMLAUNCH;
2790         ipath_write_kreg(dd, dd->ipath_kregs->kr_errormask,
2791                 dd->ipath_errormask);
2792 }
2793
2794 module_init(infinipath_init);
2795 module_exit(infinipath_cleanup);