IB/ipath: enable PE800 receive interrupts on user ports
[linux-2.6] / drivers / infiniband / hw / ipath / ipath_file_ops.c
1 /*
2  * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
3  *
4  * This software is available to you under a choice of one of two
5  * licenses.  You may choose to be licensed under the terms of the GNU
6  * General Public License (GPL) Version 2, available from the file
7  * COPYING in the main directory of this source tree, or the
8  * OpenIB.org BSD license below:
9  *
10  *     Redistribution and use in source and binary forms, with or
11  *     without modification, are permitted provided that the following
12  *     conditions are met:
13  *
14  *      - Redistributions of source code must retain the above
15  *        copyright notice, this list of conditions and the following
16  *        disclaimer.
17  *
18  *      - Redistributions in binary form must reproduce the above
19  *        copyright notice, this list of conditions and the following
20  *        disclaimer in the documentation and/or other materials
21  *        provided with the distribution.
22  *
23  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
30  * SOFTWARE.
31  */
32
33 #include <linux/pci.h>
34 #include <linux/poll.h>
35 #include <linux/cdev.h>
36 #include <linux/swap.h>
37 #include <linux/vmalloc.h>
38 #include <asm/pgtable.h>
39
40 #include "ipath_kernel.h"
41 #include "ips_common.h"
42 #include "ipath_layer.h"
43
44 static int ipath_open(struct inode *, struct file *);
45 static int ipath_close(struct inode *, struct file *);
46 static ssize_t ipath_write(struct file *, const char __user *, size_t,
47                            loff_t *);
48 static unsigned int ipath_poll(struct file *, struct poll_table_struct *);
49 static int ipath_mmap(struct file *, struct vm_area_struct *);
50
51 static struct file_operations ipath_file_ops = {
52         .owner = THIS_MODULE,
53         .write = ipath_write,
54         .open = ipath_open,
55         .release = ipath_close,
56         .poll = ipath_poll,
57         .mmap = ipath_mmap
58 };
59
60 static int ipath_get_base_info(struct ipath_portdata *pd,
61                                void __user *ubase, size_t ubase_size)
62 {
63         int ret = 0;
64         struct ipath_base_info *kinfo = NULL;
65         struct ipath_devdata *dd = pd->port_dd;
66
67         if (ubase_size < sizeof(*kinfo)) {
68                 ipath_cdbg(PROC,
69                            "Base size %lu, need %lu (version mismatch?)\n",
70                            (unsigned long) ubase_size,
71                            (unsigned long) sizeof(*kinfo));
72                 ret = -EINVAL;
73                 goto bail;
74         }
75
76         kinfo = kzalloc(sizeof(*kinfo), GFP_KERNEL);
77         if (kinfo == NULL) {
78                 ret = -ENOMEM;
79                 goto bail;
80         }
81
82         ret = dd->ipath_f_get_base_info(pd, kinfo);
83         if (ret < 0)
84                 goto bail;
85
86         kinfo->spi_rcvhdr_cnt = dd->ipath_rcvhdrcnt;
87         kinfo->spi_rcvhdrent_size = dd->ipath_rcvhdrentsize;
88         kinfo->spi_tidegrcnt = dd->ipath_rcvegrcnt;
89         kinfo->spi_rcv_egrbufsize = dd->ipath_rcvegrbufsize;
90         /*
91          * have to mmap whole thing
92          */
93         kinfo->spi_rcv_egrbuftotlen =
94                 pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
95         kinfo->spi_rcv_egrperchunk = pd->port_rcvegrbufs_perchunk;
96         kinfo->spi_rcv_egrchunksize = kinfo->spi_rcv_egrbuftotlen /
97                 pd->port_rcvegrbuf_chunks;
98         kinfo->spi_tidcnt = dd->ipath_rcvtidcnt;
99         /*
100          * for this use, may be ipath_cfgports summed over all chips that
101          * are are configured and present
102          */
103         kinfo->spi_nports = dd->ipath_cfgports;
104         /* unit (chip/board) our port is on */
105         kinfo->spi_unit = dd->ipath_unit;
106         /* for now, only a single page */
107         kinfo->spi_tid_maxsize = PAGE_SIZE;
108
109         /*
110          * Doing this per port, and based on the skip value, etc.  This has
111          * to be the actual buffer size, since the protocol code treats it
112          * as an array.
113          *
114          * These have to be set to user addresses in the user code via mmap.
115          * These values are used on return to user code for the mmap target
116          * addresses only.  For 32 bit, same 44 bit address problem, so use
117          * the physical address, not virtual.  Before 2.6.11, using the
118          * page_address() macro worked, but in 2.6.11, even that returns the
119          * full 64 bit address (upper bits all 1's).  So far, using the
120          * physical addresses (or chip offsets, for chip mapping) works, but
121          * no doubt some future kernel release will chang that, and we'll be
122          * on to yet another method of dealing with this
123          */
124         kinfo->spi_rcvhdr_base = (u64) pd->port_rcvhdrq_phys;
125         kinfo->spi_rcv_egrbufs = (u64) pd->port_rcvegr_phys;
126         kinfo->spi_pioavailaddr = (u64) dd->ipath_pioavailregs_phys;
127         kinfo->spi_status = (u64) kinfo->spi_pioavailaddr +
128                 (void *) dd->ipath_statusp -
129                 (void *) dd->ipath_pioavailregs_dma;
130         kinfo->spi_piobufbase = (u64) pd->port_piobufs;
131         kinfo->__spi_uregbase =
132                 dd->ipath_uregbase + dd->ipath_palign * pd->port_port;
133
134         kinfo->spi_pioindex = dd->ipath_pbufsport * (pd->port_port - 1);
135         kinfo->spi_piocnt = dd->ipath_pbufsport;
136         kinfo->spi_pioalign = dd->ipath_palign;
137
138         kinfo->spi_qpair = IPATH_KD_QP;
139         kinfo->spi_piosize = dd->ipath_ibmaxlen;
140         kinfo->spi_mtu = dd->ipath_ibmaxlen;    /* maxlen, not ibmtu */
141         kinfo->spi_port = pd->port_port;
142         kinfo->spi_sw_version = IPATH_KERN_SWVERSION;
143         kinfo->spi_hw_version = dd->ipath_revision;
144
145         if (copy_to_user(ubase, kinfo, sizeof(*kinfo)))
146                 ret = -EFAULT;
147
148 bail:
149         kfree(kinfo);
150         return ret;
151 }
152
153 /**
154  * ipath_tid_update - update a port TID
155  * @pd: the port
156  * @ti: the TID information
157  *
158  * The new implementation as of Oct 2004 is that the driver assigns
159  * the tid and returns it to the caller.   To make it easier to
160  * catch bugs, and to reduce search time, we keep a cursor for
161  * each port, walking the shadow tid array to find one that's not
162  * in use.
163  *
164  * For now, if we can't allocate the full list, we fail, although
165  * in the long run, we'll allocate as many as we can, and the
166  * caller will deal with that by trying the remaining pages later.
167  * That means that when we fail, we have to mark the tids as not in
168  * use again, in our shadow copy.
169  *
170  * It's up to the caller to free the tids when they are done.
171  * We'll unlock the pages as they free them.
172  *
173  * Also, right now we are locking one page at a time, but since
174  * the intended use of this routine is for a single group of
175  * virtually contiguous pages, that should change to improve
176  * performance.
177  */
178 static int ipath_tid_update(struct ipath_portdata *pd,
179                             const struct ipath_tid_info *ti)
180 {
181         int ret = 0, ntids;
182         u32 tid, porttid, cnt, i, tidcnt;
183         u16 *tidlist;
184         struct ipath_devdata *dd = pd->port_dd;
185         u64 physaddr;
186         unsigned long vaddr;
187         u64 __iomem *tidbase;
188         unsigned long tidmap[8];
189         struct page **pagep = NULL;
190
191         if (!dd->ipath_pageshadow) {
192                 ret = -ENOMEM;
193                 goto done;
194         }
195
196         cnt = ti->tidcnt;
197         if (!cnt) {
198                 ipath_dbg("After copyin, tidcnt 0, tidlist %llx\n",
199                           (unsigned long long) ti->tidlist);
200                 /*
201                  * Should we treat as success?  likely a bug
202                  */
203                 ret = -EFAULT;
204                 goto done;
205         }
206         tidcnt = dd->ipath_rcvtidcnt;
207         if (cnt >= tidcnt) {
208                 /* make sure it all fits in port_tid_pg_list */
209                 dev_info(&dd->pcidev->dev, "Process tried to allocate %u "
210                          "TIDs, only trying max (%u)\n", cnt, tidcnt);
211                 cnt = tidcnt;
212         }
213         pagep = (struct page **)pd->port_tid_pg_list;
214         tidlist = (u16 *) (&pagep[cnt]);
215
216         memset(tidmap, 0, sizeof(tidmap));
217         tid = pd->port_tidcursor;
218         /* before decrement; chip actual # */
219         porttid = pd->port_port * tidcnt;
220         ntids = tidcnt;
221         tidbase = (u64 __iomem *) (((char __iomem *) dd->ipath_kregbase) +
222                                    dd->ipath_rcvtidbase +
223                                    porttid * sizeof(*tidbase));
224
225         ipath_cdbg(VERBOSE, "Port%u %u tids, cursor %u, tidbase %p\n",
226                    pd->port_port, cnt, tid, tidbase);
227
228         /* virtual address of first page in transfer */
229         vaddr = ti->tidvaddr;
230         if (!access_ok(VERIFY_WRITE, (void __user *) vaddr,
231                        cnt * PAGE_SIZE)) {
232                 ipath_dbg("Fail vaddr %p, %u pages, !access_ok\n",
233                           (void *)vaddr, cnt);
234                 ret = -EFAULT;
235                 goto done;
236         }
237         ret = ipath_get_user_pages(vaddr, cnt, pagep);
238         if (ret) {
239                 if (ret == -EBUSY) {
240                         ipath_dbg("Failed to lock addr %p, %u pages "
241                                   "(already locked)\n",
242                                   (void *) vaddr, cnt);
243                         /*
244                          * for now, continue, and see what happens but with
245                          * the new implementation, this should never happen,
246                          * unless perhaps the user has mpin'ed the pages
247                          * themselves (something we need to test)
248                          */
249                         ret = 0;
250                 } else {
251                         dev_info(&dd->pcidev->dev,
252                                  "Failed to lock addr %p, %u pages: "
253                                  "errno %d\n", (void *) vaddr, cnt, -ret);
254                         goto done;
255                 }
256         }
257         for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
258                 for (; ntids--; tid++) {
259                         if (tid == tidcnt)
260                                 tid = 0;
261                         if (!dd->ipath_pageshadow[porttid + tid])
262                                 break;
263                 }
264                 if (ntids < 0) {
265                         /*
266                          * oops, wrapped all the way through their TIDs,
267                          * and didn't have enough free; see comments at
268                          * start of routine
269                          */
270                         ipath_dbg("Not enough free TIDs for %u pages "
271                                   "(index %d), failing\n", cnt, i);
272                         i--;    /* last tidlist[i] not filled in */
273                         ret = -ENOMEM;
274                         break;
275                 }
276                 tidlist[i] = tid;
277                 ipath_cdbg(VERBOSE, "Updating idx %u to TID %u, "
278                            "vaddr %lx\n", i, tid, vaddr);
279                 /* we "know" system pages and TID pages are same size */
280                 dd->ipath_pageshadow[porttid + tid] = pagep[i];
281                 /*
282                  * don't need atomic or it's overhead
283                  */
284                 __set_bit(tid, tidmap);
285                 physaddr = page_to_phys(pagep[i]);
286                 ipath_stats.sps_pagelocks++;
287                 ipath_cdbg(VERBOSE,
288                            "TID %u, vaddr %lx, physaddr %llx pgp %p\n",
289                            tid, vaddr, (unsigned long long) physaddr,
290                            pagep[i]);
291                 dd->ipath_f_put_tid(dd, &tidbase[tid], 1, physaddr);
292                 /*
293                  * don't check this tid in ipath_portshadow, since we
294                  * just filled it in; start with the next one.
295                  */
296                 tid++;
297         }
298
299         if (ret) {
300                 u32 limit;
301         cleanup:
302                 /* jump here if copy out of updated info failed... */
303                 ipath_dbg("After failure (ret=%d), undo %d of %d entries\n",
304                           -ret, i, cnt);
305                 /* same code that's in ipath_free_tid() */
306                 limit = sizeof(tidmap) * BITS_PER_BYTE;
307                 if (limit > tidcnt)
308                         /* just in case size changes in future */
309                         limit = tidcnt;
310                 tid = find_first_bit((const unsigned long *)tidmap, limit);
311                 for (; tid < limit; tid++) {
312                         if (!test_bit(tid, tidmap))
313                                 continue;
314                         if (dd->ipath_pageshadow[porttid + tid]) {
315                                 ipath_cdbg(VERBOSE, "Freeing TID %u\n",
316                                            tid);
317                                 dd->ipath_f_put_tid(dd, &tidbase[tid], 1,
318                                                     dd->ipath_tidinvalid);
319                                 dd->ipath_pageshadow[porttid + tid] = NULL;
320                                 ipath_stats.sps_pageunlocks++;
321                         }
322                 }
323                 ipath_release_user_pages(pagep, cnt);
324         } else {
325                 /*
326                  * Copy the updated array, with ipath_tid's filled in, back
327                  * to user.  Since we did the copy in already, this "should
328                  * never fail" If it does, we have to clean up...
329                  */
330                 if (copy_to_user((void __user *)
331                                  (unsigned long) ti->tidlist,
332                                  tidlist, cnt * sizeof(*tidlist))) {
333                         ret = -EFAULT;
334                         goto cleanup;
335                 }
336                 if (copy_to_user((void __user *) (unsigned long) ti->tidmap,
337                                  tidmap, sizeof tidmap)) {
338                         ret = -EFAULT;
339                         goto cleanup;
340                 }
341                 if (tid == tidcnt)
342                         tid = 0;
343                 pd->port_tidcursor = tid;
344         }
345
346 done:
347         if (ret)
348                 ipath_dbg("Failed to map %u TID pages, failing with %d\n",
349                           ti->tidcnt, -ret);
350         return ret;
351 }
352
353 /**
354  * ipath_tid_free - free a port TID
355  * @pd: the port
356  * @ti: the TID info
357  *
358  * right now we are unlocking one page at a time, but since
359  * the intended use of this routine is for a single group of
360  * virtually contiguous pages, that should change to improve
361  * performance.  We check that the TID is in range for this port
362  * but otherwise don't check validity; if user has an error and
363  * frees the wrong tid, it's only their own data that can thereby
364  * be corrupted.  We do check that the TID was in use, for sanity
365  * We always use our idea of the saved address, not the address that
366  * they pass in to us.
367  */
368
369 static int ipath_tid_free(struct ipath_portdata *pd,
370                           const struct ipath_tid_info *ti)
371 {
372         int ret = 0;
373         u32 tid, porttid, cnt, limit, tidcnt;
374         struct ipath_devdata *dd = pd->port_dd;
375         u64 __iomem *tidbase;
376         unsigned long tidmap[8];
377
378         if (!dd->ipath_pageshadow) {
379                 ret = -ENOMEM;
380                 goto done;
381         }
382
383         if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap,
384                            sizeof tidmap)) {
385                 ret = -EFAULT;
386                 goto done;
387         }
388
389         porttid = pd->port_port * dd->ipath_rcvtidcnt;
390         tidbase = (u64 __iomem *) ((char __iomem *)(dd->ipath_kregbase) +
391                                    dd->ipath_rcvtidbase +
392                                    porttid * sizeof(*tidbase));
393
394         tidcnt = dd->ipath_rcvtidcnt;
395         limit = sizeof(tidmap) * BITS_PER_BYTE;
396         if (limit > tidcnt)
397                 /* just in case size changes in future */
398                 limit = tidcnt;
399         tid = find_first_bit(tidmap, limit);
400         ipath_cdbg(VERBOSE, "Port%u free %u tids; first bit (max=%d) "
401                    "set is %d, porttid %u\n", pd->port_port, ti->tidcnt,
402                    limit, tid, porttid);
403         for (cnt = 0; tid < limit; tid++) {
404                 /*
405                  * small optimization; if we detect a run of 3 or so without
406                  * any set, use find_first_bit again.  That's mainly to
407                  * accelerate the case where we wrapped, so we have some at
408                  * the beginning, and some at the end, and a big gap
409                  * in the middle.
410                  */
411                 if (!test_bit(tid, tidmap))
412                         continue;
413                 cnt++;
414                 if (dd->ipath_pageshadow[porttid + tid]) {
415                         ipath_cdbg(VERBOSE, "PID %u freeing TID %u\n",
416                                    pd->port_pid, tid);
417                         dd->ipath_f_put_tid(dd, &tidbase[tid], 1,
418                                             dd->ipath_tidinvalid);
419                         ipath_release_user_pages(
420                                 &dd->ipath_pageshadow[porttid + tid], 1);
421                         dd->ipath_pageshadow[porttid + tid] = NULL;
422                         ipath_stats.sps_pageunlocks++;
423                 } else
424                         ipath_dbg("Unused tid %u, ignoring\n", tid);
425         }
426         if (cnt != ti->tidcnt)
427                 ipath_dbg("passed in tidcnt %d, only %d bits set in map\n",
428                           ti->tidcnt, cnt);
429 done:
430         if (ret)
431                 ipath_dbg("Failed to unmap %u TID pages, failing with %d\n",
432                           ti->tidcnt, -ret);
433         return ret;
434 }
435
436 /**
437  * ipath_set_part_key - set a partition key
438  * @pd: the port
439  * @key: the key
440  *
441  * We can have up to 4 active at a time (other than the default, which is
442  * always allowed).  This is somewhat tricky, since multiple ports may set
443  * the same key, so we reference count them, and clean up at exit.  All 4
444  * partition keys are packed into a single infinipath register.  It's an
445  * error for a process to set the same pkey multiple times.  We provide no
446  * mechanism to de-allocate a pkey at this time, we may eventually need to
447  * do that.  I've used the atomic operations, and no locking, and only make
448  * a single pass through what's available.  This should be more than
449  * adequate for some time. I'll think about spinlocks or the like if and as
450  * it's necessary.
451  */
452 static int ipath_set_part_key(struct ipath_portdata *pd, u16 key)
453 {
454         struct ipath_devdata *dd = pd->port_dd;
455         int i, any = 0, pidx = -1;
456         u16 lkey = key & 0x7FFF;
457         int ret;
458
459         if (lkey == (IPS_DEFAULT_P_KEY & 0x7FFF)) {
460                 /* nothing to do; this key always valid */
461                 ret = 0;
462                 goto bail;
463         }
464
465         ipath_cdbg(VERBOSE, "p%u try to set pkey %hx, current keys "
466                    "%hx:%x %hx:%x %hx:%x %hx:%x\n",
467                    pd->port_port, key, dd->ipath_pkeys[0],
468                    atomic_read(&dd->ipath_pkeyrefs[0]), dd->ipath_pkeys[1],
469                    atomic_read(&dd->ipath_pkeyrefs[1]), dd->ipath_pkeys[2],
470                    atomic_read(&dd->ipath_pkeyrefs[2]), dd->ipath_pkeys[3],
471                    atomic_read(&dd->ipath_pkeyrefs[3]));
472
473         if (!lkey) {
474                 ipath_cdbg(PROC, "p%u tries to set key 0, not allowed\n",
475                            pd->port_port);
476                 ret = -EINVAL;
477                 goto bail;
478         }
479
480         /*
481          * Set the full membership bit, because it has to be
482          * set in the register or the packet, and it seems
483          * cleaner to set in the register than to force all
484          * callers to set it. (see bug 4331)
485          */
486         key |= 0x8000;
487
488         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
489                 if (!pd->port_pkeys[i] && pidx == -1)
490                         pidx = i;
491                 if (pd->port_pkeys[i] == key) {
492                         ipath_cdbg(VERBOSE, "p%u tries to set same pkey "
493                                    "(%x) more than once\n",
494                                    pd->port_port, key);
495                         ret = -EEXIST;
496                         goto bail;
497                 }
498         }
499         if (pidx == -1) {
500                 ipath_dbg("All pkeys for port %u already in use, "
501                           "can't set %x\n", pd->port_port, key);
502                 ret = -EBUSY;
503                 goto bail;
504         }
505         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
506                 if (!dd->ipath_pkeys[i]) {
507                         any++;
508                         continue;
509                 }
510                 if (dd->ipath_pkeys[i] == key) {
511                         atomic_t *pkrefs = &dd->ipath_pkeyrefs[i];
512
513                         if (atomic_inc_return(pkrefs) > 1) {
514                                 pd->port_pkeys[pidx] = key;
515                                 ipath_cdbg(VERBOSE, "p%u set key %x "
516                                            "matches #%d, count now %d\n",
517                                            pd->port_port, key, i,
518                                            atomic_read(pkrefs));
519                                 ret = 0;
520                                 goto bail;
521                         } else {
522                                 /*
523                                  * lost race, decrement count, catch below
524                                  */
525                                 atomic_dec(pkrefs);
526                                 ipath_cdbg(VERBOSE, "Lost race, count was "
527                                            "0, after dec, it's %d\n",
528                                            atomic_read(pkrefs));
529                                 any++;
530                         }
531                 }
532                 if ((dd->ipath_pkeys[i] & 0x7FFF) == lkey) {
533                         /*
534                          * It makes no sense to have both the limited and
535                          * full membership PKEY set at the same time since
536                          * the unlimited one will disable the limited one.
537                          */
538                         ret = -EEXIST;
539                         goto bail;
540                 }
541         }
542         if (!any) {
543                 ipath_dbg("port %u, all pkeys already in use, "
544                           "can't set %x\n", pd->port_port, key);
545                 ret = -EBUSY;
546                 goto bail;
547         }
548         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
549                 if (!dd->ipath_pkeys[i] &&
550                     atomic_inc_return(&dd->ipath_pkeyrefs[i]) == 1) {
551                         u64 pkey;
552
553                         /* for ipathstats, etc. */
554                         ipath_stats.sps_pkeys[i] = lkey;
555                         pd->port_pkeys[pidx] = dd->ipath_pkeys[i] = key;
556                         pkey =
557                                 (u64) dd->ipath_pkeys[0] |
558                                 ((u64) dd->ipath_pkeys[1] << 16) |
559                                 ((u64) dd->ipath_pkeys[2] << 32) |
560                                 ((u64) dd->ipath_pkeys[3] << 48);
561                         ipath_cdbg(PROC, "p%u set key %x in #%d, "
562                                    "portidx %d, new pkey reg %llx\n",
563                                    pd->port_port, key, i, pidx,
564                                    (unsigned long long) pkey);
565                         ipath_write_kreg(
566                                 dd, dd->ipath_kregs->kr_partitionkey, pkey);
567
568                         ret = 0;
569                         goto bail;
570                 }
571         }
572         ipath_dbg("port %u, all pkeys already in use 2nd pass, "
573                   "can't set %x\n", pd->port_port, key);
574         ret = -EBUSY;
575
576 bail:
577         return ret;
578 }
579
580 /**
581  * ipath_manage_rcvq - manage a port's receive queue
582  * @pd: the port
583  * @start_stop: action to carry out
584  *
585  * start_stop == 0 disables receive on the port, for use in queue
586  * overflow conditions.  start_stop==1 re-enables, to be used to
587  * re-init the software copy of the head register
588  */
589 static int ipath_manage_rcvq(struct ipath_portdata *pd, int start_stop)
590 {
591         struct ipath_devdata *dd = pd->port_dd;
592         u64 tval;
593
594         ipath_cdbg(PROC, "%sabling rcv for unit %u port %u\n",
595                    start_stop ? "en" : "dis", dd->ipath_unit,
596                    pd->port_port);
597         /* atomically clear receive enable port. */
598         if (start_stop) {
599                 /*
600                  * On enable, force in-memory copy of the tail register to
601                  * 0, so that protocol code doesn't have to worry about
602                  * whether or not the chip has yet updated the in-memory
603                  * copy or not on return from the system call. The chip
604                  * always resets it's tail register back to 0 on a
605                  * transition from disabled to enabled.  This could cause a
606                  * problem if software was broken, and did the enable w/o
607                  * the disable, but eventually the in-memory copy will be
608                  * updated and correct itself, even in the face of software
609                  * bugs.
610                  */
611                 *pd->port_rcvhdrtail_kvaddr = 0;
612                 set_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
613                         &dd->ipath_rcvctrl);
614         } else
615                 clear_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
616                           &dd->ipath_rcvctrl);
617         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
618                          dd->ipath_rcvctrl);
619         /* now be sure chip saw it before we return */
620         tval = ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
621         if (start_stop) {
622                 /*
623                  * And try to be sure that tail reg update has happened too.
624                  * This should in theory interlock with the RXE changes to
625                  * the tail register.  Don't assign it to the tail register
626                  * in memory copy, since we could overwrite an update by the
627                  * chip if we did.
628                  */
629                 tval = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
630         }
631         /* always; new head should be equal to new tail; see above */
632         return 0;
633 }
634
635 static void ipath_clean_part_key(struct ipath_portdata *pd,
636                                  struct ipath_devdata *dd)
637 {
638         int i, j, pchanged = 0;
639         u64 oldpkey;
640
641         /* for debugging only */
642         oldpkey = (u64) dd->ipath_pkeys[0] |
643                 ((u64) dd->ipath_pkeys[1] << 16) |
644                 ((u64) dd->ipath_pkeys[2] << 32) |
645                 ((u64) dd->ipath_pkeys[3] << 48);
646
647         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
648                 if (!pd->port_pkeys[i])
649                         continue;
650                 ipath_cdbg(VERBOSE, "look for key[%d] %hx in pkeys\n", i,
651                            pd->port_pkeys[i]);
652                 for (j = 0; j < ARRAY_SIZE(dd->ipath_pkeys); j++) {
653                         /* check for match independent of the global bit */
654                         if ((dd->ipath_pkeys[j] & 0x7fff) !=
655                             (pd->port_pkeys[i] & 0x7fff))
656                                 continue;
657                         if (atomic_dec_and_test(&dd->ipath_pkeyrefs[j])) {
658                                 ipath_cdbg(VERBOSE, "p%u clear key "
659                                            "%x matches #%d\n",
660                                            pd->port_port,
661                                            pd->port_pkeys[i], j);
662                                 ipath_stats.sps_pkeys[j] =
663                                         dd->ipath_pkeys[j] = 0;
664                                 pchanged++;
665                         }
666                         else ipath_cdbg(
667                                 VERBOSE, "p%u key %x matches #%d, "
668                                 "but ref still %d\n", pd->port_port,
669                                 pd->port_pkeys[i], j,
670                                 atomic_read(&dd->ipath_pkeyrefs[j]));
671                         break;
672                 }
673                 pd->port_pkeys[i] = 0;
674         }
675         if (pchanged) {
676                 u64 pkey = (u64) dd->ipath_pkeys[0] |
677                         ((u64) dd->ipath_pkeys[1] << 16) |
678                         ((u64) dd->ipath_pkeys[2] << 32) |
679                         ((u64) dd->ipath_pkeys[3] << 48);
680                 ipath_cdbg(VERBOSE, "p%u old pkey reg %llx, "
681                            "new pkey reg %llx\n", pd->port_port,
682                            (unsigned long long) oldpkey,
683                            (unsigned long long) pkey);
684                 ipath_write_kreg(dd, dd->ipath_kregs->kr_partitionkey,
685                                  pkey);
686         }
687 }
688
689 /**
690  * ipath_create_user_egr - allocate eager TID buffers
691  * @pd: the port to allocate TID buffers for
692  *
693  * This routine is now quite different for user and kernel, because
694  * the kernel uses skb's, for the accelerated network performance
695  * This is the user port version
696  *
697  * Allocate the eager TID buffers and program them into infinipath
698  * They are no longer completely contiguous, we do multiple allocation
699  * calls.
700  */
701 static int ipath_create_user_egr(struct ipath_portdata *pd)
702 {
703         struct ipath_devdata *dd = pd->port_dd;
704         unsigned e, egrcnt, alloced, egrperchunk, chunk, egrsize, egroff;
705         size_t size;
706         int ret;
707
708         egrcnt = dd->ipath_rcvegrcnt;
709         /* TID number offset for this port */
710         egroff = pd->port_port * egrcnt;
711         egrsize = dd->ipath_rcvegrbufsize;
712         ipath_cdbg(VERBOSE, "Allocating %d egr buffers, at egrtid "
713                    "offset %x, egrsize %u\n", egrcnt, egroff, egrsize);
714
715         /*
716          * to avoid wasting a lot of memory, we allocate 32KB chunks of
717          * physically contiguous memory, advance through it until used up
718          * and then allocate more.  Of course, we need memory to store those
719          * extra pointers, now.  Started out with 256KB, but under heavy
720          * memory pressure (creating large files and then copying them over
721          * NFS while doing lots of MPI jobs), we hit some allocation
722          * failures, even though we can sleep...  (2.6.10) Still get
723          * failures at 64K.  32K is the lowest we can go without waiting
724          * more memory again.  It seems likely that the coalescing in
725          * free_pages, etc. still has issues (as it has had previously
726          * during 2.6.x development).
727          */
728         size = 0x8000;
729         alloced = ALIGN(egrsize * egrcnt, size);
730         egrperchunk = size / egrsize;
731         chunk = (egrcnt + egrperchunk - 1) / egrperchunk;
732         pd->port_rcvegrbuf_chunks = chunk;
733         pd->port_rcvegrbufs_perchunk = egrperchunk;
734         pd->port_rcvegrbuf_size = size;
735         pd->port_rcvegrbuf = vmalloc(chunk * sizeof(pd->port_rcvegrbuf[0]));
736         if (!pd->port_rcvegrbuf) {
737                 ret = -ENOMEM;
738                 goto bail;
739         }
740         pd->port_rcvegrbuf_phys =
741                 vmalloc(chunk * sizeof(pd->port_rcvegrbuf_phys[0]));
742         if (!pd->port_rcvegrbuf_phys) {
743                 ret = -ENOMEM;
744                 goto bail_rcvegrbuf;
745         }
746         for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
747                 /*
748                  * GFP_USER, but without GFP_FS, so buffer cache can be
749                  * coalesced (we hope); otherwise, even at order 4,
750                  * heavy filesystem activity makes these fail
751                  */
752                 gfp_t gfp_flags = __GFP_WAIT | __GFP_IO | __GFP_COMP;
753
754                 pd->port_rcvegrbuf[e] = dma_alloc_coherent(
755                         &dd->pcidev->dev, size, &pd->port_rcvegrbuf_phys[e],
756                         gfp_flags);
757
758                 if (!pd->port_rcvegrbuf[e]) {
759                         ret = -ENOMEM;
760                         goto bail_rcvegrbuf_phys;
761                 }
762         }
763
764         pd->port_rcvegr_phys = pd->port_rcvegrbuf_phys[0];
765
766         for (e = chunk = 0; chunk < pd->port_rcvegrbuf_chunks; chunk++) {
767                 dma_addr_t pa = pd->port_rcvegrbuf_phys[chunk];
768                 unsigned i;
769
770                 for (i = 0; e < egrcnt && i < egrperchunk; e++, i++) {
771                         dd->ipath_f_put_tid(dd, e + egroff +
772                                             (u64 __iomem *)
773                                             ((char __iomem *)
774                                              dd->ipath_kregbase +
775                                              dd->ipath_rcvegrbase), 0, pa);
776                         pa += egrsize;
777                 }
778                 cond_resched(); /* don't hog the cpu */
779         }
780
781         ret = 0;
782         goto bail;
783
784 bail_rcvegrbuf_phys:
785         for (e = 0; e < pd->port_rcvegrbuf_chunks &&
786                      pd->port_rcvegrbuf[e]; e++)
787                 dma_free_coherent(&dd->pcidev->dev, size,
788                                   pd->port_rcvegrbuf[e],
789                                   pd->port_rcvegrbuf_phys[e]);
790
791         vfree(pd->port_rcvegrbuf_phys);
792         pd->port_rcvegrbuf_phys = NULL;
793 bail_rcvegrbuf:
794         vfree(pd->port_rcvegrbuf);
795         pd->port_rcvegrbuf = NULL;
796 bail:
797         return ret;
798 }
799
800 static int ipath_do_user_init(struct ipath_portdata *pd,
801                               const struct ipath_user_info *uinfo)
802 {
803         int ret = 0;
804         struct ipath_devdata *dd = pd->port_dd;
805         u64 physaddr, uaddr, off, atmp;
806         struct page *pagep;
807         u32 head32;
808         u64 head;
809
810         /* for now, if major version is different, bail */
811         if ((uinfo->spu_userversion >> 16) != IPATH_USER_SWMAJOR) {
812                 dev_info(&dd->pcidev->dev,
813                          "User major version %d not same as driver "
814                          "major %d\n", uinfo->spu_userversion >> 16,
815                          IPATH_USER_SWMAJOR);
816                 ret = -ENODEV;
817                 goto done;
818         }
819
820         if ((uinfo->spu_userversion & 0xffff) != IPATH_USER_SWMINOR)
821                 ipath_dbg("User minor version %d not same as driver "
822                           "minor %d\n", uinfo->spu_userversion & 0xffff,
823                           IPATH_USER_SWMINOR);
824
825         if (uinfo->spu_rcvhdrsize) {
826                 ret = ipath_setrcvhdrsize(dd, uinfo->spu_rcvhdrsize);
827                 if (ret)
828                         goto done;
829         }
830
831         /* for now we do nothing with rcvhdrcnt: uinfo->spu_rcvhdrcnt */
832
833         /* set up for the rcvhdr Q tail register writeback to user memory */
834         if (!uinfo->spu_rcvhdraddr ||
835             !access_ok(VERIFY_WRITE, (u64 __user *) (unsigned long)
836                        uinfo->spu_rcvhdraddr, sizeof(u64))) {
837                 ipath_dbg("Port %d rcvhdrtail addr %llx not valid\n",
838                           pd->port_port,
839                           (unsigned long long) uinfo->spu_rcvhdraddr);
840                 ret = -EINVAL;
841                 goto done;
842         }
843
844         off = offset_in_page(uinfo->spu_rcvhdraddr);
845         uaddr = PAGE_MASK & (unsigned long) uinfo->spu_rcvhdraddr;
846         ret = ipath_get_user_pages_nocopy(uaddr, &pagep);
847         if (ret) {
848                 dev_info(&dd->pcidev->dev, "Failed to lookup and lock "
849                          "address %llx for rcvhdrtail: errno %d\n",
850                          (unsigned long long) uinfo->spu_rcvhdraddr, -ret);
851                 goto done;
852         }
853         ipath_stats.sps_pagelocks++;
854         pd->port_rcvhdrtail_uaddr = uaddr;
855         pd->port_rcvhdrtail_pagep = pagep;
856         pd->port_rcvhdrtail_kvaddr =
857                 page_address(pagep);
858         pd->port_rcvhdrtail_kvaddr += off;
859         physaddr = page_to_phys(pagep) + off;
860         ipath_cdbg(VERBOSE, "port %d user addr %llx hdrtailaddr, %llx "
861                    "physical (off=%llx)\n",
862                    pd->port_port,
863                    (unsigned long long) uinfo->spu_rcvhdraddr,
864                    (unsigned long long) physaddr, (unsigned long long) off);
865         ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdrtailaddr,
866                               pd->port_port, physaddr);
867         atmp = ipath_read_kreg64_port(dd,
868                                       dd->ipath_kregs->kr_rcvhdrtailaddr,
869                                       pd->port_port);
870         if (physaddr != atmp) {
871                 ipath_dev_err(dd,
872                               "Catastrophic software error, "
873                               "RcvHdrTailAddr%u written as %llx, "
874                               "read back as %llx\n", pd->port_port,
875                               (unsigned long long) physaddr,
876                               (unsigned long long) atmp);
877                 ret = -EINVAL;
878                 goto done;
879         }
880
881         /* for right now, kernel piobufs are at end, so port 1 is at 0 */
882         pd->port_piobufs = dd->ipath_piobufbase +
883                 dd->ipath_pbufsport * (pd->port_port -
884                                        1) * dd->ipath_palign;
885         ipath_cdbg(VERBOSE, "Set base of piobufs for port %u to 0x%x\n",
886                    pd->port_port, pd->port_piobufs);
887
888         /*
889          * Now allocate the rcvhdr Q and eager TIDs; skip the TID
890          * array for time being.  If pd->port_port > chip-supported,
891          * we need to do extra stuff here to handle by handling overflow
892          * through port 0, someday
893          */
894         ret = ipath_create_rcvhdrq(dd, pd);
895         if (!ret)
896                 ret = ipath_create_user_egr(pd);
897         if (ret)
898                 goto done;
899         /* enable receives now */
900         /* atomically set enable bit for this port */
901         set_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
902                 &dd->ipath_rcvctrl);
903
904         /*
905          * set the head registers for this port to the current values
906          * of the tail pointers, since we don't know if they were
907          * updated on last use of the port.
908          */
909         head32 = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
910         head = (u64) head32;
911         ipath_write_ureg(dd, ur_rcvhdrhead, head, pd->port_port);
912         head32 = ipath_read_ureg32(dd, ur_rcvegrindextail, pd->port_port);
913         ipath_write_ureg(dd, ur_rcvegrindexhead, head32, pd->port_port);
914         dd->ipath_lastegrheads[pd->port_port] = -1;
915         dd->ipath_lastrcvhdrqtails[pd->port_port] = -1;
916         ipath_cdbg(VERBOSE, "Wrote port%d head %llx, egrhead %x from "
917                    "tail regs\n", pd->port_port,
918                    (unsigned long long) head, head32);
919         pd->port_tidcursor = 0; /* start at beginning after open */
920         /*
921          * now enable the port; the tail registers will be written to memory
922          * by the chip as soon as it sees the write to
923          * dd->ipath_kregs->kr_rcvctrl.  The update only happens on
924          * transition from 0 to 1, so clear it first, then set it as part of
925          * enabling the port.  This will (very briefly) affect any other
926          * open ports, but it shouldn't be long enough to be an issue.
927          */
928         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
929                          dd->ipath_rcvctrl & ~INFINIPATH_R_TAILUPD);
930         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
931                          dd->ipath_rcvctrl);
932
933 done:
934         return ret;
935 }
936
937 static int mmap_ureg(struct vm_area_struct *vma, struct ipath_devdata *dd,
938                      u64 ureg)
939 {
940         unsigned long phys;
941         int ret;
942
943         /* it's the real hardware, so io_remap works */
944
945         if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) {
946                 dev_info(&dd->pcidev->dev, "FAIL mmap userreg: reqlen "
947                          "%lx > PAGE\n", vma->vm_end - vma->vm_start);
948                 ret = -EFAULT;
949         } else {
950                 phys = dd->ipath_physaddr + ureg;
951                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
952
953                 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
954                 ret = io_remap_pfn_range(vma, vma->vm_start,
955                                          phys >> PAGE_SHIFT,
956                                          vma->vm_end - vma->vm_start,
957                                          vma->vm_page_prot);
958         }
959         return ret;
960 }
961
962 static int mmap_piobufs(struct vm_area_struct *vma,
963                         struct ipath_devdata *dd,
964                         struct ipath_portdata *pd)
965 {
966         unsigned long phys;
967         int ret;
968
969         /*
970          * When we map the PIO buffers, we want to map them as writeonly, no
971          * read possible.
972          */
973
974         if ((vma->vm_end - vma->vm_start) >
975             (dd->ipath_pbufsport * dd->ipath_palign)) {
976                 dev_info(&dd->pcidev->dev, "FAIL mmap piobufs: "
977                          "reqlen %lx > PAGE\n",
978                          vma->vm_end - vma->vm_start);
979                 ret = -EFAULT;
980                 goto bail;
981         }
982
983         phys = dd->ipath_physaddr + pd->port_piobufs;
984         /*
985          * Do *NOT* mark this as non-cached (PWT bit), or we don't get the
986          * write combining behavior we want on the PIO buffers!
987          * vma->vm_page_prot =
988          *        pgprot_noncached(vma->vm_page_prot);
989          */
990
991         if (vma->vm_flags & VM_READ) {
992                 dev_info(&dd->pcidev->dev,
993                          "Can't map piobufs as readable (flags=%lx)\n",
994                          vma->vm_flags);
995                 ret = -EPERM;
996                 goto bail;
997         }
998
999         /* don't allow them to later change to readable with mprotect */
1000
1001         vma->vm_flags &= ~VM_MAYWRITE;
1002         vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1003
1004         ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
1005                                  vma->vm_end - vma->vm_start,
1006                                  vma->vm_page_prot);
1007 bail:
1008         return ret;
1009 }
1010
1011 static int mmap_rcvegrbufs(struct vm_area_struct *vma,
1012                            struct ipath_portdata *pd)
1013 {
1014         struct ipath_devdata *dd = pd->port_dd;
1015         unsigned long start, size;
1016         size_t total_size, i;
1017         dma_addr_t *phys;
1018         int ret;
1019
1020         if (!pd->port_rcvegrbuf) {
1021                 ret = -EFAULT;
1022                 goto bail;
1023         }
1024
1025         size = pd->port_rcvegrbuf_size;
1026         total_size = pd->port_rcvegrbuf_chunks * size;
1027         if ((vma->vm_end - vma->vm_start) > total_size) {
1028                 dev_info(&dd->pcidev->dev, "FAIL on egr bufs: "
1029                          "reqlen %lx > actual %lx\n",
1030                          vma->vm_end - vma->vm_start,
1031                          (unsigned long) total_size);
1032                 ret = -EFAULT;
1033                 goto bail;
1034         }
1035
1036         if (vma->vm_flags & VM_WRITE) {
1037                 dev_info(&dd->pcidev->dev, "Can't map eager buffers as "
1038                          "writable (flags=%lx)\n", vma->vm_flags);
1039                 ret = -EPERM;
1040                 goto bail;
1041         }
1042
1043         start = vma->vm_start;
1044         phys = pd->port_rcvegrbuf_phys;
1045
1046         /* don't allow them to later change to writeable with mprotect */
1047         vma->vm_flags &= ~VM_MAYWRITE;
1048
1049         for (i = 0; i < pd->port_rcvegrbuf_chunks; i++, start += size) {
1050                 ret = remap_pfn_range(vma, start, phys[i] >> PAGE_SHIFT,
1051                                       size, vma->vm_page_prot);
1052                 if (ret < 0)
1053                         goto bail;
1054         }
1055         ret = 0;
1056
1057 bail:
1058         return ret;
1059 }
1060
1061 static int mmap_rcvhdrq(struct vm_area_struct *vma,
1062                         struct ipath_portdata *pd)
1063 {
1064         struct ipath_devdata *dd = pd->port_dd;
1065         size_t total_size;
1066         int ret;
1067
1068         /*
1069          * kmalloc'ed memory, physically contiguous; this is from
1070          * spi_rcvhdr_base; we allow user to map read-write so they can
1071          * write hdrq entries to allow protocol code to directly poll
1072          * whether a hdrq entry has been written.
1073          */
1074         total_size = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
1075                            sizeof(u32), PAGE_SIZE);
1076         if ((vma->vm_end - vma->vm_start) > total_size) {
1077                 dev_info(&dd->pcidev->dev,
1078                          "FAIL on rcvhdrq: reqlen %lx > actual %lx\n",
1079                          vma->vm_end - vma->vm_start,
1080                          (unsigned long) total_size);
1081                 ret = -EFAULT;
1082                 goto bail;
1083         }
1084
1085         ret = remap_pfn_range(vma, vma->vm_start,
1086                               pd->port_rcvhdrq_phys >> PAGE_SHIFT,
1087                               vma->vm_end - vma->vm_start,
1088                               vma->vm_page_prot);
1089 bail:
1090         return ret;
1091 }
1092
1093 static int mmap_pioavailregs(struct vm_area_struct *vma,
1094                              struct ipath_portdata *pd)
1095 {
1096         struct ipath_devdata *dd = pd->port_dd;
1097         int ret;
1098
1099         /*
1100          * when we map the PIO bufferavail registers, we want to map them as
1101          * readonly, no write possible.
1102          *
1103          * kmalloc'ed memory, physically contiguous, one page only, readonly
1104          */
1105
1106         if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) {
1107                 dev_info(&dd->pcidev->dev, "FAIL on pioavailregs_dma: "
1108                          "reqlen %lx > actual %lx\n",
1109                          vma->vm_end - vma->vm_start,
1110                          (unsigned long) PAGE_SIZE);
1111                 ret = -EFAULT;
1112                 goto bail;
1113         }
1114
1115         if (vma->vm_flags & VM_WRITE) {
1116                 dev_info(&dd->pcidev->dev,
1117                          "Can't map pioavailregs as writable (flags=%lx)\n",
1118                          vma->vm_flags);
1119                 ret = -EPERM;
1120                 goto bail;
1121         }
1122
1123         /* don't allow them to later change with mprotect */
1124         vma->vm_flags &= ~VM_MAYWRITE;
1125
1126         ret = remap_pfn_range(vma, vma->vm_start,
1127                               dd->ipath_pioavailregs_phys >> PAGE_SHIFT,
1128                               PAGE_SIZE, vma->vm_page_prot);
1129 bail:
1130         return ret;
1131 }
1132
1133 /**
1134  * ipath_mmap - mmap various structures into user space
1135  * @fp: the file pointer
1136  * @vma: the VM area
1137  *
1138  * We use this to have a shared buffer between the kernel and the user code
1139  * for the rcvhdr queue, egr buffers, and the per-port user regs and pio
1140  * buffers in the chip.  We have the open and close entries so we can bump
1141  * the ref count and keep the driver from being unloaded while still mapped.
1142  */
1143 static int ipath_mmap(struct file *fp, struct vm_area_struct *vma)
1144 {
1145         struct ipath_portdata *pd;
1146         struct ipath_devdata *dd;
1147         u64 pgaddr, ureg;
1148         int ret;
1149
1150         pd = port_fp(fp);
1151         dd = pd->port_dd;
1152         /*
1153          * This is the ipath_do_user_init() code, mapping the shared buffers
1154          * into the user process. The address referred to by vm_pgoff is the
1155          * virtual, not physical, address; we only do one mmap for each
1156          * space mapped.
1157          */
1158         pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1159
1160         /*
1161          * note that ureg does *NOT* have the kregvirt as part of it, to be
1162          * sure that for 32 bit programs, we don't end up trying to map a >
1163          * 44 address.  Has to match ipath_get_base_info() code that sets
1164          * __spi_uregbase
1165          */
1166
1167         ureg = dd->ipath_uregbase + dd->ipath_palign * pd->port_port;
1168
1169         ipath_cdbg(MM, "ushare: pgaddr %llx vm_start=%lx, vmlen %lx\n",
1170                    (unsigned long long) pgaddr, vma->vm_start,
1171                    vma->vm_end - vma->vm_start);
1172
1173         if (pgaddr == ureg)
1174                 ret = mmap_ureg(vma, dd, ureg);
1175         else if (pgaddr == pd->port_piobufs)
1176                 ret = mmap_piobufs(vma, dd, pd);
1177         else if (pgaddr == (u64) pd->port_rcvegr_phys)
1178                 ret = mmap_rcvegrbufs(vma, pd);
1179         else if (pgaddr == (u64) pd->port_rcvhdrq_phys)
1180                 ret = mmap_rcvhdrq(vma, pd);
1181         else if (pgaddr == dd->ipath_pioavailregs_phys)
1182                 ret = mmap_pioavailregs(vma, pd);
1183         else
1184                 ret = -EINVAL;
1185
1186         vma->vm_private_data = NULL;
1187
1188         if (ret < 0)
1189                 dev_info(&dd->pcidev->dev,
1190                          "Failure %d on addr %lx, off %lx\n",
1191                          -ret, vma->vm_start, vma->vm_pgoff);
1192
1193         return ret;
1194 }
1195
1196 static unsigned int ipath_poll(struct file *fp,
1197                                struct poll_table_struct *pt)
1198 {
1199         struct ipath_portdata *pd;
1200         u32 head, tail;
1201         int bit;
1202         struct ipath_devdata *dd;
1203
1204         pd = port_fp(fp);
1205         dd = pd->port_dd;
1206
1207         bit = pd->port_port + INFINIPATH_R_INTRAVAIL_SHIFT;
1208         set_bit(bit, &dd->ipath_rcvctrl);
1209
1210         /*
1211          * Before blocking, make sure that head is still == tail,
1212          * reading from the chip, so we can be sure the interrupt
1213          * enable has made it to the chip.  If not equal, disable
1214          * interrupt again and return immediately.  This avoids races,
1215          * and the overhead of the chip read doesn't matter much at
1216          * this point, since we are waiting for something anyway.
1217          */
1218
1219         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1220                          dd->ipath_rcvctrl);
1221
1222         head = ipath_read_ureg32(dd, ur_rcvhdrhead, pd->port_port);
1223         tail = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
1224
1225         if (tail == head) {
1226                 set_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
1227                 if(dd->ipath_rhdrhead_intr_off) /* arm rcv interrupt */
1228                         (void)ipath_write_ureg(dd, ur_rcvhdrhead,
1229                                                dd->ipath_rhdrhead_intr_off
1230                                                | head, pd->port_port);
1231                 poll_wait(fp, &pd->port_wait, pt);
1232
1233                 if (test_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag)) {
1234                         /* timed out, no packets received */
1235                         clear_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
1236                         pd->port_rcvwait_to++;
1237                 }
1238         }
1239         else {
1240                 /* it's already happened; don't do wait_event overhead */
1241                 pd->port_rcvnowait++;
1242         }
1243
1244         clear_bit(bit, &dd->ipath_rcvctrl);
1245         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1246                          dd->ipath_rcvctrl);
1247
1248         return 0;
1249 }
1250
1251 static int try_alloc_port(struct ipath_devdata *dd, int port,
1252                           struct file *fp)
1253 {
1254         int ret;
1255
1256         if (!dd->ipath_pd[port]) {
1257                 void *p, *ptmp;
1258
1259                 p = kzalloc(sizeof(struct ipath_portdata), GFP_KERNEL);
1260
1261                 /*
1262                  * Allocate memory for use in ipath_tid_update() just once
1263                  * at open, not per call.  Reduces cost of expected send
1264                  * setup.
1265                  */
1266                 ptmp = kmalloc(dd->ipath_rcvtidcnt * sizeof(u16) +
1267                                dd->ipath_rcvtidcnt * sizeof(struct page **),
1268                                GFP_KERNEL);
1269                 if (!p || !ptmp) {
1270                         ipath_dev_err(dd, "Unable to allocate portdata "
1271                                       "memory, failing open\n");
1272                         ret = -ENOMEM;
1273                         kfree(p);
1274                         kfree(ptmp);
1275                         goto bail;
1276                 }
1277                 dd->ipath_pd[port] = p;
1278                 dd->ipath_pd[port]->port_port = port;
1279                 dd->ipath_pd[port]->port_dd = dd;
1280                 dd->ipath_pd[port]->port_tid_pg_list = ptmp;
1281                 init_waitqueue_head(&dd->ipath_pd[port]->port_wait);
1282         }
1283         if (!dd->ipath_pd[port]->port_cnt) {
1284                 dd->ipath_pd[port]->port_cnt = 1;
1285                 fp->private_data = (void *) dd->ipath_pd[port];
1286                 ipath_cdbg(PROC, "%s[%u] opened unit:port %u:%u\n",
1287                            current->comm, current->pid, dd->ipath_unit,
1288                            port);
1289                 dd->ipath_pd[port]->port_pid = current->pid;
1290                 strncpy(dd->ipath_pd[port]->port_comm, current->comm,
1291                         sizeof(dd->ipath_pd[port]->port_comm));
1292                 ipath_stats.sps_ports++;
1293                 ret = 0;
1294                 goto bail;
1295         }
1296         ret = -EBUSY;
1297
1298 bail:
1299         return ret;
1300 }
1301
1302 static inline int usable(struct ipath_devdata *dd)
1303 {
1304         return dd &&
1305                 (dd->ipath_flags & IPATH_PRESENT) &&
1306                 dd->ipath_kregbase &&
1307                 dd->ipath_lid &&
1308                 !(dd->ipath_flags & (IPATH_LINKDOWN | IPATH_DISABLED
1309                                      | IPATH_LINKUNK));
1310 }
1311
1312 static int find_free_port(int unit, struct file *fp)
1313 {
1314         struct ipath_devdata *dd = ipath_lookup(unit);
1315         int ret, i;
1316
1317         if (!dd) {
1318                 ret = -ENODEV;
1319                 goto bail;
1320         }
1321
1322         if (!usable(dd)) {
1323                 ret = -ENETDOWN;
1324                 goto bail;
1325         }
1326
1327         for (i = 0; i < dd->ipath_cfgports; i++) {
1328                 ret = try_alloc_port(dd, i, fp);
1329                 if (ret != -EBUSY)
1330                         goto bail;
1331         }
1332         ret = -EBUSY;
1333
1334 bail:
1335         return ret;
1336 }
1337
1338 static int find_best_unit(struct file *fp)
1339 {
1340         int ret = 0, i, prefunit = -1, devmax;
1341         int maxofallports, npresent, nup;
1342         int ndev;
1343
1344         (void) ipath_count_units(&npresent, &nup, &maxofallports);
1345
1346         /*
1347          * This code is present to allow a knowledgeable person to
1348          * specify the layout of processes to processors before opening
1349          * this driver, and then we'll assign the process to the "closest"
1350          * HT-400 to that processor (we assume reasonable connectivity,
1351          * for now).  This code assumes that if affinity has been set
1352          * before this point, that at most one cpu is set; for now this
1353          * is reasonable.  I check for both cpus_empty() and cpus_full(),
1354          * in case some kernel variant sets none of the bits when no
1355          * affinity is set.  2.6.11 and 12 kernels have all present
1356          * cpus set.  Some day we'll have to fix it up further to handle
1357          * a cpu subset.  This algorithm fails for two HT-400's connected
1358          * in tunnel fashion.  Eventually this needs real topology
1359          * information.  There may be some issues with dual core numbering
1360          * as well.  This needs more work prior to release.
1361          */
1362         if (!cpus_empty(current->cpus_allowed) &&
1363             !cpus_full(current->cpus_allowed)) {
1364                 int ncpus = num_online_cpus(), curcpu = -1;
1365                 for (i = 0; i < ncpus; i++)
1366                         if (cpu_isset(i, current->cpus_allowed)) {
1367                                 ipath_cdbg(PROC, "%s[%u] affinity set for "
1368                                            "cpu %d\n", current->comm,
1369                                            current->pid, i);
1370                                 curcpu = i;
1371                         }
1372                 if (curcpu != -1) {
1373                         if (npresent) {
1374                                 prefunit = curcpu / (ncpus / npresent);
1375                                 ipath_dbg("%s[%u] %d chips, %d cpus, "
1376                                           "%d cpus/chip, select unit %d\n",
1377                                           current->comm, current->pid,
1378                                           npresent, ncpus, ncpus / npresent,
1379                                           prefunit);
1380                         }
1381                 }
1382         }
1383
1384         /*
1385          * user ports start at 1, kernel port is 0
1386          * For now, we do round-robin access across all chips
1387          */
1388
1389         if (prefunit != -1)
1390                 devmax = prefunit + 1;
1391         else
1392                 devmax = ipath_count_units(NULL, NULL, NULL);
1393 recheck:
1394         for (i = 1; i < maxofallports; i++) {
1395                 for (ndev = prefunit != -1 ? prefunit : 0; ndev < devmax;
1396                      ndev++) {
1397                         struct ipath_devdata *dd = ipath_lookup(ndev);
1398
1399                         if (!usable(dd))
1400                                 continue; /* can't use this unit */
1401                         if (i >= dd->ipath_cfgports)
1402                                 /*
1403                                  * Maxed out on users of this unit. Try
1404                                  * next.
1405                                  */
1406                                 continue;
1407                         ret = try_alloc_port(dd, i, fp);
1408                         if (!ret)
1409                                 goto done;
1410                 }
1411         }
1412
1413         if (npresent) {
1414                 if (nup == 0) {
1415                         ret = -ENETDOWN;
1416                         ipath_dbg("No ports available (none initialized "
1417                                   "and ready)\n");
1418                 } else {
1419                         if (prefunit > 0) {
1420                                 /* if started above 0, retry from 0 */
1421                                 ipath_cdbg(PROC,
1422                                            "%s[%u] no ports on prefunit "
1423                                            "%d, clear and re-check\n",
1424                                            current->comm, current->pid,
1425                                            prefunit);
1426                                 devmax = ipath_count_units(NULL, NULL,
1427                                                            NULL);
1428                                 prefunit = -1;
1429                                 goto recheck;
1430                         }
1431                         ret = -EBUSY;
1432                         ipath_dbg("No ports available\n");
1433                 }
1434         } else {
1435                 ret = -ENXIO;
1436                 ipath_dbg("No boards found\n");
1437         }
1438
1439 done:
1440         return ret;
1441 }
1442
1443 static int ipath_open(struct inode *in, struct file *fp)
1444 {
1445         int ret, minor;
1446
1447         mutex_lock(&ipath_mutex);
1448
1449         minor = iminor(in);
1450         ipath_cdbg(VERBOSE, "open on dev %lx (minor %d)\n",
1451                    (long)in->i_rdev, minor);
1452
1453         if (minor)
1454                 ret = find_free_port(minor - 1, fp);
1455         else
1456                 ret = find_best_unit(fp);
1457
1458         mutex_unlock(&ipath_mutex);
1459         return ret;
1460 }
1461
1462 /**
1463  * unlock_exptid - unlock any expected TID entries port still had in use
1464  * @pd: port
1465  *
1466  * We don't actually update the chip here, because we do a bulk update
1467  * below, using ipath_f_clear_tids.
1468  */
1469 static void unlock_expected_tids(struct ipath_portdata *pd)
1470 {
1471         struct ipath_devdata *dd = pd->port_dd;
1472         int port_tidbase = pd->port_port * dd->ipath_rcvtidcnt;
1473         int i, cnt = 0, maxtid = port_tidbase + dd->ipath_rcvtidcnt;
1474
1475         ipath_cdbg(VERBOSE, "Port %u unlocking any locked expTID pages\n",
1476                    pd->port_port);
1477         for (i = port_tidbase; i < maxtid; i++) {
1478                 if (!dd->ipath_pageshadow[i])
1479                         continue;
1480
1481                 ipath_release_user_pages_on_close(&dd->ipath_pageshadow[i],
1482                                                   1);
1483                 dd->ipath_pageshadow[i] = NULL;
1484                 cnt++;
1485                 ipath_stats.sps_pageunlocks++;
1486         }
1487         if (cnt)
1488                 ipath_cdbg(VERBOSE, "Port %u locked %u expTID entries\n",
1489                            pd->port_port, cnt);
1490
1491         if (ipath_stats.sps_pagelocks || ipath_stats.sps_pageunlocks)
1492                 ipath_cdbg(VERBOSE, "%llu pages locked, %llu unlocked\n",
1493                            (unsigned long long) ipath_stats.sps_pagelocks,
1494                            (unsigned long long)
1495                            ipath_stats.sps_pageunlocks);
1496 }
1497
1498 static int ipath_close(struct inode *in, struct file *fp)
1499 {
1500         int ret = 0;
1501         struct ipath_portdata *pd;
1502         struct ipath_devdata *dd;
1503         unsigned port;
1504
1505         ipath_cdbg(VERBOSE, "close on dev %lx, private data %p\n",
1506                    (long)in->i_rdev, fp->private_data);
1507
1508         mutex_lock(&ipath_mutex);
1509
1510         pd = port_fp(fp);
1511         port = pd->port_port;
1512         fp->private_data = NULL;
1513         dd = pd->port_dd;
1514
1515         if (pd->port_hdrqfull) {
1516                 ipath_cdbg(PROC, "%s[%u] had %u rcvhdrqfull errors "
1517                            "during run\n", pd->port_comm, pd->port_pid,
1518                            pd->port_hdrqfull);
1519                 pd->port_hdrqfull = 0;
1520         }
1521
1522         if (pd->port_rcvwait_to || pd->port_piowait_to
1523             || pd->port_rcvnowait || pd->port_pionowait) {
1524                 ipath_cdbg(VERBOSE, "port%u, %u rcv, %u pio wait timeo; "
1525                            "%u rcv %u, pio already\n",
1526                            pd->port_port, pd->port_rcvwait_to,
1527                            pd->port_piowait_to, pd->port_rcvnowait,
1528                            pd->port_pionowait);
1529                 pd->port_rcvwait_to = pd->port_piowait_to =
1530                         pd->port_rcvnowait = pd->port_pionowait = 0;
1531         }
1532         if (pd->port_flag) {
1533                 ipath_dbg("port %u port_flag still set to 0x%lx\n",
1534                           pd->port_port, pd->port_flag);
1535                 pd->port_flag = 0;
1536         }
1537
1538         if (dd->ipath_kregbase) {
1539                 if (pd->port_rcvhdrtail_uaddr) {
1540                         pd->port_rcvhdrtail_uaddr = 0;
1541                         pd->port_rcvhdrtail_kvaddr = NULL;
1542                         ipath_release_user_pages_on_close(
1543                                 &pd->port_rcvhdrtail_pagep, 1);
1544                         pd->port_rcvhdrtail_pagep = NULL;
1545                         ipath_stats.sps_pageunlocks++;
1546                 }
1547                 ipath_write_kreg_port(
1548                         dd, dd->ipath_kregs->kr_rcvhdrtailaddr,
1549                         port, 0ULL);
1550                 ipath_write_kreg_port(
1551                         dd, dd->ipath_kregs->kr_rcvhdraddr,
1552                         pd->port_port, 0);
1553
1554                 /* clean up the pkeys for this port user */
1555                 ipath_clean_part_key(pd, dd);
1556
1557                 if (port < dd->ipath_cfgports) {
1558                         int i = dd->ipath_pbufsport * (port - 1);
1559                         ipath_disarm_piobufs(dd, i, dd->ipath_pbufsport);
1560
1561                         /* atomically clear receive enable port. */
1562                         clear_bit(INFINIPATH_R_PORTENABLE_SHIFT + port,
1563                                   &dd->ipath_rcvctrl);
1564                         ipath_write_kreg(
1565                                 dd,
1566                                 dd->ipath_kregs->kr_rcvctrl,
1567                                 dd->ipath_rcvctrl);
1568
1569                         if (dd->ipath_pageshadow)
1570                                 unlock_expected_tids(pd);
1571                         ipath_stats.sps_ports--;
1572                         ipath_cdbg(PROC, "%s[%u] closed port %u:%u\n",
1573                                    pd->port_comm, pd->port_pid,
1574                                    dd->ipath_unit, port);
1575                 }
1576         }
1577
1578         pd->port_cnt = 0;
1579         pd->port_pid = 0;
1580
1581         dd->ipath_f_clear_tids(dd, pd->port_port);
1582
1583         ipath_free_pddata(dd, pd->port_port, 0);
1584
1585         mutex_unlock(&ipath_mutex);
1586
1587         return ret;
1588 }
1589
1590 static int ipath_port_info(struct ipath_portdata *pd,
1591                            struct ipath_port_info __user *uinfo)
1592 {
1593         struct ipath_port_info info;
1594         int nup;
1595         int ret;
1596
1597         (void) ipath_count_units(NULL, &nup, NULL);
1598         info.num_active = nup;
1599         info.unit = pd->port_dd->ipath_unit;
1600         info.port = pd->port_port;
1601
1602         if (copy_to_user(uinfo, &info, sizeof(info))) {
1603                 ret = -EFAULT;
1604                 goto bail;
1605         }
1606         ret = 0;
1607
1608 bail:
1609         return ret;
1610 }
1611
1612 static ssize_t ipath_write(struct file *fp, const char __user *data,
1613                            size_t count, loff_t *off)
1614 {
1615         const struct ipath_cmd __user *ucmd;
1616         struct ipath_portdata *pd;
1617         const void __user *src;
1618         size_t consumed, copy;
1619         struct ipath_cmd cmd;
1620         ssize_t ret = 0;
1621         void *dest;
1622
1623         if (count < sizeof(cmd.type)) {
1624                 ret = -EINVAL;
1625                 goto bail;
1626         }
1627
1628         ucmd = (const struct ipath_cmd __user *) data;
1629
1630         if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) {
1631                 ret = -EFAULT;
1632                 goto bail;
1633         }
1634
1635         consumed = sizeof(cmd.type);
1636
1637         switch (cmd.type) {
1638         case IPATH_CMD_USER_INIT:
1639                 copy = sizeof(cmd.cmd.user_info);
1640                 dest = &cmd.cmd.user_info;
1641                 src = &ucmd->cmd.user_info;
1642                 break;
1643         case IPATH_CMD_RECV_CTRL:
1644                 copy = sizeof(cmd.cmd.recv_ctrl);
1645                 dest = &cmd.cmd.recv_ctrl;
1646                 src = &ucmd->cmd.recv_ctrl;
1647                 break;
1648         case IPATH_CMD_PORT_INFO:
1649                 copy = sizeof(cmd.cmd.port_info);
1650                 dest = &cmd.cmd.port_info;
1651                 src = &ucmd->cmd.port_info;
1652                 break;
1653         case IPATH_CMD_TID_UPDATE:
1654         case IPATH_CMD_TID_FREE:
1655                 copy = sizeof(cmd.cmd.tid_info);
1656                 dest = &cmd.cmd.tid_info;
1657                 src = &ucmd->cmd.tid_info;
1658                 break;
1659         case IPATH_CMD_SET_PART_KEY:
1660                 copy = sizeof(cmd.cmd.part_key);
1661                 dest = &cmd.cmd.part_key;
1662                 src = &ucmd->cmd.part_key;
1663                 break;
1664         default:
1665                 ret = -EINVAL;
1666                 goto bail;
1667         }
1668
1669         if ((count - consumed) < copy) {
1670                 ret = -EINVAL;
1671                 goto bail;
1672         }
1673
1674         if (copy_from_user(dest, src, copy)) {
1675                 ret = -EFAULT;
1676                 goto bail;
1677         }
1678
1679         consumed += copy;
1680         pd = port_fp(fp);
1681
1682         switch (cmd.type) {
1683         case IPATH_CMD_USER_INIT:
1684                 ret = ipath_do_user_init(pd, &cmd.cmd.user_info);
1685                 if (ret < 0)
1686                         goto bail;
1687                 ret = ipath_get_base_info(
1688                         pd, (void __user *) (unsigned long)
1689                         cmd.cmd.user_info.spu_base_info,
1690                         cmd.cmd.user_info.spu_base_info_size);
1691                 break;
1692         case IPATH_CMD_RECV_CTRL:
1693                 ret = ipath_manage_rcvq(pd, cmd.cmd.recv_ctrl);
1694                 break;
1695         case IPATH_CMD_PORT_INFO:
1696                 ret = ipath_port_info(pd,
1697                                       (struct ipath_port_info __user *)
1698                                       (unsigned long) cmd.cmd.port_info);
1699                 break;
1700         case IPATH_CMD_TID_UPDATE:
1701                 ret = ipath_tid_update(pd, &cmd.cmd.tid_info);
1702                 break;
1703         case IPATH_CMD_TID_FREE:
1704                 ret = ipath_tid_free(pd, &cmd.cmd.tid_info);
1705                 break;
1706         case IPATH_CMD_SET_PART_KEY:
1707                 ret = ipath_set_part_key(pd, cmd.cmd.part_key);
1708                 break;
1709         }
1710
1711         if (ret >= 0)
1712                 ret = consumed;
1713
1714 bail:
1715         return ret;
1716 }
1717
1718 static struct class *ipath_class;
1719
1720 static int init_cdev(int minor, char *name, struct file_operations *fops,
1721                      struct cdev **cdevp, struct class_device **class_devp)
1722 {
1723         const dev_t dev = MKDEV(IPATH_MAJOR, minor);
1724         struct cdev *cdev = NULL;
1725         struct class_device *class_dev = NULL;
1726         int ret;
1727
1728         cdev = cdev_alloc();
1729         if (!cdev) {
1730                 printk(KERN_ERR IPATH_DRV_NAME
1731                        ": Could not allocate cdev for minor %d, %s\n",
1732                        minor, name);
1733                 ret = -ENOMEM;
1734                 goto done;
1735         }
1736
1737         cdev->owner = THIS_MODULE;
1738         cdev->ops = fops;
1739         kobject_set_name(&cdev->kobj, name);
1740
1741         ret = cdev_add(cdev, dev, 1);
1742         if (ret < 0) {
1743                 printk(KERN_ERR IPATH_DRV_NAME
1744                        ": Could not add cdev for minor %d, %s (err %d)\n",
1745                        minor, name, -ret);
1746                 goto err_cdev;
1747         }
1748
1749         class_dev = class_device_create(ipath_class, NULL, dev, NULL, name);
1750
1751         if (IS_ERR(class_dev)) {
1752                 ret = PTR_ERR(class_dev);
1753                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
1754                        "class_dev for minor %d, %s (err %d)\n",
1755                        minor, name, -ret);
1756                 goto err_cdev;
1757         }
1758
1759         goto done;
1760
1761 err_cdev:
1762         cdev_del(cdev);
1763         cdev = NULL;
1764
1765 done:
1766         if (ret >= 0) {
1767                 *cdevp = cdev;
1768                 *class_devp = class_dev;
1769         } else {
1770                 *cdevp = NULL;
1771                 *class_devp = NULL;
1772         }
1773
1774         return ret;
1775 }
1776
1777 int ipath_cdev_init(int minor, char *name, struct file_operations *fops,
1778                     struct cdev **cdevp, struct class_device **class_devp)
1779 {
1780         return init_cdev(minor, name, fops, cdevp, class_devp);
1781 }
1782
1783 static void cleanup_cdev(struct cdev **cdevp,
1784                          struct class_device **class_devp)
1785 {
1786         struct class_device *class_dev = *class_devp;
1787
1788         if (class_dev) {
1789                 class_device_unregister(class_dev);
1790                 *class_devp = NULL;
1791         }
1792
1793         if (*cdevp) {
1794                 cdev_del(*cdevp);
1795                 *cdevp = NULL;
1796         }
1797 }
1798
1799 void ipath_cdev_cleanup(struct cdev **cdevp,
1800                         struct class_device **class_devp)
1801 {
1802         cleanup_cdev(cdevp, class_devp);
1803 }
1804
1805 static struct cdev *wildcard_cdev;
1806 static struct class_device *wildcard_class_dev;
1807
1808 static const dev_t dev = MKDEV(IPATH_MAJOR, 0);
1809
1810 static int user_init(void)
1811 {
1812         int ret;
1813
1814         ret = register_chrdev_region(dev, IPATH_NMINORS, IPATH_DRV_NAME);
1815         if (ret < 0) {
1816                 printk(KERN_ERR IPATH_DRV_NAME ": Could not register "
1817                        "chrdev region (err %d)\n", -ret);
1818                 goto done;
1819         }
1820
1821         ipath_class = class_create(THIS_MODULE, IPATH_DRV_NAME);
1822
1823         if (IS_ERR(ipath_class)) {
1824                 ret = PTR_ERR(ipath_class);
1825                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
1826                        "device class (err %d)\n", -ret);
1827                 goto bail;
1828         }
1829
1830         goto done;
1831 bail:
1832         unregister_chrdev_region(dev, IPATH_NMINORS);
1833 done:
1834         return ret;
1835 }
1836
1837 static void user_cleanup(void)
1838 {
1839         if (ipath_class) {
1840                 class_destroy(ipath_class);
1841                 ipath_class = NULL;
1842         }
1843
1844         unregister_chrdev_region(dev, IPATH_NMINORS);
1845 }
1846
1847 static atomic_t user_count = ATOMIC_INIT(0);
1848 static atomic_t user_setup = ATOMIC_INIT(0);
1849
1850 int ipath_user_add(struct ipath_devdata *dd)
1851 {
1852         char name[10];
1853         int ret;
1854
1855         if (atomic_inc_return(&user_count) == 1) {
1856                 ret = user_init();
1857                 if (ret < 0) {
1858                         ipath_dev_err(dd, "Unable to set up user support: "
1859                                       "error %d\n", -ret);
1860                         goto bail;
1861                 }
1862                 ret = ipath_diag_init();
1863                 if (ret < 0) {
1864                         ipath_dev_err(dd, "Unable to set up diag support: "
1865                                       "error %d\n", -ret);
1866                         goto bail_sma;
1867                 }
1868
1869                 ret = init_cdev(0, "ipath", &ipath_file_ops, &wildcard_cdev,
1870                                 &wildcard_class_dev);
1871                 if (ret < 0) {
1872                         ipath_dev_err(dd, "Could not create wildcard "
1873                                       "minor: error %d\n", -ret);
1874                         goto bail_diag;
1875                 }
1876
1877                 atomic_set(&user_setup, 1);
1878         }
1879
1880         snprintf(name, sizeof(name), "ipath%d", dd->ipath_unit);
1881
1882         ret = init_cdev(dd->ipath_unit + 1, name, &ipath_file_ops,
1883                         &dd->cdev, &dd->class_dev);
1884         if (ret < 0)
1885                 ipath_dev_err(dd, "Could not create user minor %d, %s\n",
1886                               dd->ipath_unit + 1, name);
1887
1888         goto bail;
1889
1890 bail_diag:
1891         ipath_diag_cleanup();
1892 bail_sma:
1893         user_cleanup();
1894 bail:
1895         return ret;
1896 }
1897
1898 void ipath_user_del(struct ipath_devdata *dd)
1899 {
1900         cleanup_cdev(&dd->cdev, &dd->class_dev);
1901
1902         if (atomic_dec_return(&user_count) == 0) {
1903                 if (atomic_read(&user_setup) == 0)
1904                         goto bail;
1905
1906                 cleanup_cdev(&wildcard_cdev, &wildcard_class_dev);
1907                 ipath_diag_cleanup();
1908                 user_cleanup();
1909
1910                 atomic_set(&user_setup, 0);
1911         }
1912 bail:
1913         return;
1914 }