IB/ipath: Convert from .nopage to .fault
[linux-2.6.git] / drivers / infiniband / hw / ipath / ipath_file_ops.c
1 /*
2  * Copyright (c) 2006, 2007 QLogic Corporation. All rights reserved.
3  * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33
34 #include <linux/pci.h>
35 #include <linux/poll.h>
36 #include <linux/cdev.h>
37 #include <linux/swap.h>
38 #include <linux/vmalloc.h>
39 #include <asm/pgtable.h>
40
41 #include "ipath_kernel.h"
42 #include "ipath_common.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 const 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 /*
61  * Convert kernel virtual addresses to physical addresses so they don't
62  * potentially conflict with the chip addresses used as mmap offsets.
63  * It doesn't really matter what mmap offset we use as long as we can
64  * interpret it correctly.
65  */
66 static u64 cvt_kvaddr(void *p)
67 {
68         struct page *page;
69         u64 paddr = 0;
70
71         page = vmalloc_to_page(p);
72         if (page)
73                 paddr = page_to_pfn(page) << PAGE_SHIFT;
74
75         return paddr;
76 }
77
78 static int ipath_get_base_info(struct file *fp,
79                                void __user *ubase, size_t ubase_size)
80 {
81         struct ipath_portdata *pd = port_fp(fp);
82         int ret = 0;
83         struct ipath_base_info *kinfo = NULL;
84         struct ipath_devdata *dd = pd->port_dd;
85         unsigned subport_cnt;
86         int shared, master;
87         size_t sz;
88
89         subport_cnt = pd->port_subport_cnt;
90         if (!subport_cnt) {
91                 shared = 0;
92                 master = 0;
93                 subport_cnt = 1;
94         } else {
95                 shared = 1;
96                 master = !subport_fp(fp);
97         }
98
99         sz = sizeof(*kinfo);
100         /* If port sharing is not requested, allow the old size structure */
101         if (!shared)
102                 sz -= 7 * sizeof(u64);
103         if (ubase_size < sz) {
104                 ipath_cdbg(PROC,
105                            "Base size %zu, need %zu (version mismatch?)\n",
106                            ubase_size, sz);
107                 ret = -EINVAL;
108                 goto bail;
109         }
110
111         kinfo = kzalloc(sizeof(*kinfo), GFP_KERNEL);
112         if (kinfo == NULL) {
113                 ret = -ENOMEM;
114                 goto bail;
115         }
116
117         ret = dd->ipath_f_get_base_info(pd, kinfo);
118         if (ret < 0)
119                 goto bail;
120
121         kinfo->spi_rcvhdr_cnt = dd->ipath_rcvhdrcnt;
122         kinfo->spi_rcvhdrent_size = dd->ipath_rcvhdrentsize;
123         kinfo->spi_tidegrcnt = dd->ipath_rcvegrcnt;
124         kinfo->spi_rcv_egrbufsize = dd->ipath_rcvegrbufsize;
125         /*
126          * have to mmap whole thing
127          */
128         kinfo->spi_rcv_egrbuftotlen =
129                 pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
130         kinfo->spi_rcv_egrperchunk = pd->port_rcvegrbufs_perchunk;
131         kinfo->spi_rcv_egrchunksize = kinfo->spi_rcv_egrbuftotlen /
132                 pd->port_rcvegrbuf_chunks;
133         kinfo->spi_tidcnt = dd->ipath_rcvtidcnt / subport_cnt;
134         if (master)
135                 kinfo->spi_tidcnt += dd->ipath_rcvtidcnt % subport_cnt;
136         /*
137          * for this use, may be ipath_cfgports summed over all chips that
138          * are are configured and present
139          */
140         kinfo->spi_nports = dd->ipath_cfgports;
141         /* unit (chip/board) our port is on */
142         kinfo->spi_unit = dd->ipath_unit;
143         /* for now, only a single page */
144         kinfo->spi_tid_maxsize = PAGE_SIZE;
145
146         /*
147          * Doing this per port, and based on the skip value, etc.  This has
148          * to be the actual buffer size, since the protocol code treats it
149          * as an array.
150          *
151          * These have to be set to user addresses in the user code via mmap.
152          * These values are used on return to user code for the mmap target
153          * addresses only.  For 32 bit, same 44 bit address problem, so use
154          * the physical address, not virtual.  Before 2.6.11, using the
155          * page_address() macro worked, but in 2.6.11, even that returns the
156          * full 64 bit address (upper bits all 1's).  So far, using the
157          * physical addresses (or chip offsets, for chip mapping) works, but
158          * no doubt some future kernel release will change that, and we'll be
159          * on to yet another method of dealing with this.
160          */
161         kinfo->spi_rcvhdr_base = (u64) pd->port_rcvhdrq_phys;
162         kinfo->spi_rcvhdr_tailaddr = (u64) pd->port_rcvhdrqtailaddr_phys;
163         kinfo->spi_rcv_egrbufs = (u64) pd->port_rcvegr_phys;
164         kinfo->spi_pioavailaddr = (u64) dd->ipath_pioavailregs_phys;
165         kinfo->spi_status = (u64) kinfo->spi_pioavailaddr +
166                 (void *) dd->ipath_statusp -
167                 (void *) dd->ipath_pioavailregs_dma;
168         if (!shared) {
169                 kinfo->spi_piocnt = dd->ipath_pbufsport;
170                 kinfo->spi_piobufbase = (u64) pd->port_piobufs;
171                 kinfo->__spi_uregbase = (u64) dd->ipath_uregbase +
172                         dd->ipath_palign * pd->port_port;
173         } else if (master) {
174                 kinfo->spi_piocnt = (dd->ipath_pbufsport / subport_cnt) +
175                                     (dd->ipath_pbufsport % subport_cnt);
176                 /* Master's PIO buffers are after all the slave's */
177                 kinfo->spi_piobufbase = (u64) pd->port_piobufs +
178                         dd->ipath_palign *
179                         (dd->ipath_pbufsport - kinfo->spi_piocnt);
180         } else {
181                 unsigned slave = subport_fp(fp) - 1;
182
183                 kinfo->spi_piocnt = dd->ipath_pbufsport / subport_cnt;
184                 kinfo->spi_piobufbase = (u64) pd->port_piobufs +
185                         dd->ipath_palign * kinfo->spi_piocnt * slave;
186         }
187         if (shared) {
188                 kinfo->spi_port_uregbase = (u64) dd->ipath_uregbase +
189                         dd->ipath_palign * pd->port_port;
190                 kinfo->spi_port_rcvegrbuf = kinfo->spi_rcv_egrbufs;
191                 kinfo->spi_port_rcvhdr_base = kinfo->spi_rcvhdr_base;
192                 kinfo->spi_port_rcvhdr_tailaddr = kinfo->spi_rcvhdr_tailaddr;
193
194                 kinfo->__spi_uregbase = cvt_kvaddr(pd->subport_uregbase +
195                         PAGE_SIZE * subport_fp(fp));
196
197                 kinfo->spi_rcvhdr_base = cvt_kvaddr(pd->subport_rcvhdr_base +
198                         pd->port_rcvhdrq_size * subport_fp(fp));
199                 kinfo->spi_rcvhdr_tailaddr = 0;
200                 kinfo->spi_rcv_egrbufs = cvt_kvaddr(pd->subport_rcvegrbuf +
201                         pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size *
202                         subport_fp(fp));
203
204                 kinfo->spi_subport_uregbase =
205                         cvt_kvaddr(pd->subport_uregbase);
206                 kinfo->spi_subport_rcvegrbuf =
207                         cvt_kvaddr(pd->subport_rcvegrbuf);
208                 kinfo->spi_subport_rcvhdr_base =
209                         cvt_kvaddr(pd->subport_rcvhdr_base);
210                 ipath_cdbg(PROC, "port %u flags %x %llx %llx %llx\n",
211                         kinfo->spi_port, kinfo->spi_runtime_flags,
212                         (unsigned long long) kinfo->spi_subport_uregbase,
213                         (unsigned long long) kinfo->spi_subport_rcvegrbuf,
214                         (unsigned long long) kinfo->spi_subport_rcvhdr_base);
215         }
216
217         kinfo->spi_pioindex = (kinfo->spi_piobufbase - dd->ipath_piobufbase) /
218                 dd->ipath_palign;
219         kinfo->spi_pioalign = dd->ipath_palign;
220
221         kinfo->spi_qpair = IPATH_KD_QP;
222         kinfo->spi_piosize = dd->ipath_ibmaxlen;
223         kinfo->spi_mtu = dd->ipath_ibmaxlen;    /* maxlen, not ibmtu */
224         kinfo->spi_port = pd->port_port;
225         kinfo->spi_subport = subport_fp(fp);
226         kinfo->spi_sw_version = IPATH_KERN_SWVERSION;
227         kinfo->spi_hw_version = dd->ipath_revision;
228
229         if (master) {
230                 kinfo->spi_runtime_flags |= IPATH_RUNTIME_MASTER;
231         }
232
233         sz = (ubase_size < sizeof(*kinfo)) ? ubase_size : sizeof(*kinfo);
234         if (copy_to_user(ubase, kinfo, sz))
235                 ret = -EFAULT;
236
237 bail:
238         kfree(kinfo);
239         return ret;
240 }
241
242 /**
243  * ipath_tid_update - update a port TID
244  * @pd: the port
245  * @fp: the ipath device file
246  * @ti: the TID information
247  *
248  * The new implementation as of Oct 2004 is that the driver assigns
249  * the tid and returns it to the caller.   To make it easier to
250  * catch bugs, and to reduce search time, we keep a cursor for
251  * each port, walking the shadow tid array to find one that's not
252  * in use.
253  *
254  * For now, if we can't allocate the full list, we fail, although
255  * in the long run, we'll allocate as many as we can, and the
256  * caller will deal with that by trying the remaining pages later.
257  * That means that when we fail, we have to mark the tids as not in
258  * use again, in our shadow copy.
259  *
260  * It's up to the caller to free the tids when they are done.
261  * We'll unlock the pages as they free them.
262  *
263  * Also, right now we are locking one page at a time, but since
264  * the intended use of this routine is for a single group of
265  * virtually contiguous pages, that should change to improve
266  * performance.
267  */
268 static int ipath_tid_update(struct ipath_portdata *pd, struct file *fp,
269                             const struct ipath_tid_info *ti)
270 {
271         int ret = 0, ntids;
272         u32 tid, porttid, cnt, i, tidcnt, tidoff;
273         u16 *tidlist;
274         struct ipath_devdata *dd = pd->port_dd;
275         u64 physaddr;
276         unsigned long vaddr;
277         u64 __iomem *tidbase;
278         unsigned long tidmap[8];
279         struct page **pagep = NULL;
280         unsigned subport = subport_fp(fp);
281
282         if (!dd->ipath_pageshadow) {
283                 ret = -ENOMEM;
284                 goto done;
285         }
286
287         cnt = ti->tidcnt;
288         if (!cnt) {
289                 ipath_dbg("After copyin, tidcnt 0, tidlist %llx\n",
290                           (unsigned long long) ti->tidlist);
291                 /*
292                  * Should we treat as success?  likely a bug
293                  */
294                 ret = -EFAULT;
295                 goto done;
296         }
297         porttid = pd->port_port * dd->ipath_rcvtidcnt;
298         if (!pd->port_subport_cnt) {
299                 tidcnt = dd->ipath_rcvtidcnt;
300                 tid = pd->port_tidcursor;
301                 tidoff = 0;
302         } else if (!subport) {
303                 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
304                          (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
305                 tidoff = dd->ipath_rcvtidcnt - tidcnt;
306                 porttid += tidoff;
307                 tid = tidcursor_fp(fp);
308         } else {
309                 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
310                 tidoff = tidcnt * (subport - 1);
311                 porttid += tidoff;
312                 tid = tidcursor_fp(fp);
313         }
314         if (cnt > tidcnt) {
315                 /* make sure it all fits in port_tid_pg_list */
316                 dev_info(&dd->pcidev->dev, "Process tried to allocate %u "
317                          "TIDs, only trying max (%u)\n", cnt, tidcnt);
318                 cnt = tidcnt;
319         }
320         pagep = &((struct page **) pd->port_tid_pg_list)[tidoff];
321         tidlist = &((u16 *) &pagep[dd->ipath_rcvtidcnt])[tidoff];
322
323         memset(tidmap, 0, sizeof(tidmap));
324         /* before decrement; chip actual # */
325         ntids = tidcnt;
326         tidbase = (u64 __iomem *) (((char __iomem *) dd->ipath_kregbase) +
327                                    dd->ipath_rcvtidbase +
328                                    porttid * sizeof(*tidbase));
329
330         ipath_cdbg(VERBOSE, "Port%u %u tids, cursor %u, tidbase %p\n",
331                    pd->port_port, cnt, tid, tidbase);
332
333         /* virtual address of first page in transfer */
334         vaddr = ti->tidvaddr;
335         if (!access_ok(VERIFY_WRITE, (void __user *) vaddr,
336                        cnt * PAGE_SIZE)) {
337                 ipath_dbg("Fail vaddr %p, %u pages, !access_ok\n",
338                           (void *)vaddr, cnt);
339                 ret = -EFAULT;
340                 goto done;
341         }
342         ret = ipath_get_user_pages(vaddr, cnt, pagep);
343         if (ret) {
344                 if (ret == -EBUSY) {
345                         ipath_dbg("Failed to lock addr %p, %u pages "
346                                   "(already locked)\n",
347                                   (void *) vaddr, cnt);
348                         /*
349                          * for now, continue, and see what happens but with
350                          * the new implementation, this should never happen,
351                          * unless perhaps the user has mpin'ed the pages
352                          * themselves (something we need to test)
353                          */
354                         ret = 0;
355                 } else {
356                         dev_info(&dd->pcidev->dev,
357                                  "Failed to lock addr %p, %u pages: "
358                                  "errno %d\n", (void *) vaddr, cnt, -ret);
359                         goto done;
360                 }
361         }
362         for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
363                 for (; ntids--; tid++) {
364                         if (tid == tidcnt)
365                                 tid = 0;
366                         if (!dd->ipath_pageshadow[porttid + tid])
367                                 break;
368                 }
369                 if (ntids < 0) {
370                         /*
371                          * oops, wrapped all the way through their TIDs,
372                          * and didn't have enough free; see comments at
373                          * start of routine
374                          */
375                         ipath_dbg("Not enough free TIDs for %u pages "
376                                   "(index %d), failing\n", cnt, i);
377                         i--;    /* last tidlist[i] not filled in */
378                         ret = -ENOMEM;
379                         break;
380                 }
381                 tidlist[i] = tid + tidoff;
382                 ipath_cdbg(VERBOSE, "Updating idx %u to TID %u, "
383                            "vaddr %lx\n", i, tid + tidoff, vaddr);
384                 /* we "know" system pages and TID pages are same size */
385                 dd->ipath_pageshadow[porttid + tid] = pagep[i];
386                 dd->ipath_physshadow[porttid + tid] = ipath_map_page(
387                         dd->pcidev, pagep[i], 0, PAGE_SIZE,
388                         PCI_DMA_FROMDEVICE);
389                 /*
390                  * don't need atomic or it's overhead
391                  */
392                 __set_bit(tid, tidmap);
393                 physaddr = dd->ipath_physshadow[porttid + tid];
394                 ipath_stats.sps_pagelocks++;
395                 ipath_cdbg(VERBOSE,
396                            "TID %u, vaddr %lx, physaddr %llx pgp %p\n",
397                            tid, vaddr, (unsigned long long) physaddr,
398                            pagep[i]);
399                 dd->ipath_f_put_tid(dd, &tidbase[tid], RCVHQ_RCV_TYPE_EXPECTED,
400                                     physaddr);
401                 /*
402                  * don't check this tid in ipath_portshadow, since we
403                  * just filled it in; start with the next one.
404                  */
405                 tid++;
406         }
407
408         if (ret) {
409                 u32 limit;
410         cleanup:
411                 /* jump here if copy out of updated info failed... */
412                 ipath_dbg("After failure (ret=%d), undo %d of %d entries\n",
413                           -ret, i, cnt);
414                 /* same code that's in ipath_free_tid() */
415                 limit = sizeof(tidmap) * BITS_PER_BYTE;
416                 if (limit > tidcnt)
417                         /* just in case size changes in future */
418                         limit = tidcnt;
419                 tid = find_first_bit((const unsigned long *)tidmap, limit);
420                 for (; tid < limit; tid++) {
421                         if (!test_bit(tid, tidmap))
422                                 continue;
423                         if (dd->ipath_pageshadow[porttid + tid]) {
424                                 ipath_cdbg(VERBOSE, "Freeing TID %u\n",
425                                            tid);
426                                 dd->ipath_f_put_tid(dd, &tidbase[tid],
427                                                     RCVHQ_RCV_TYPE_EXPECTED,
428                                                     dd->ipath_tidinvalid);
429                                 pci_unmap_page(dd->pcidev,
430                                         dd->ipath_physshadow[porttid + tid],
431                                         PAGE_SIZE, PCI_DMA_FROMDEVICE);
432                                 dd->ipath_pageshadow[porttid + tid] = NULL;
433                                 ipath_stats.sps_pageunlocks++;
434                         }
435                 }
436                 ipath_release_user_pages(pagep, cnt);
437         } else {
438                 /*
439                  * Copy the updated array, with ipath_tid's filled in, back
440                  * to user.  Since we did the copy in already, this "should
441                  * never fail" If it does, we have to clean up...
442                  */
443                 if (copy_to_user((void __user *)
444                                  (unsigned long) ti->tidlist,
445                                  tidlist, cnt * sizeof(*tidlist))) {
446                         ret = -EFAULT;
447                         goto cleanup;
448                 }
449                 if (copy_to_user((void __user *) (unsigned long) ti->tidmap,
450                                  tidmap, sizeof tidmap)) {
451                         ret = -EFAULT;
452                         goto cleanup;
453                 }
454                 if (tid == tidcnt)
455                         tid = 0;
456                 if (!pd->port_subport_cnt)
457                         pd->port_tidcursor = tid;
458                 else
459                         tidcursor_fp(fp) = tid;
460         }
461
462 done:
463         if (ret)
464                 ipath_dbg("Failed to map %u TID pages, failing with %d\n",
465                           ti->tidcnt, -ret);
466         return ret;
467 }
468
469 /**
470  * ipath_tid_free - free a port TID
471  * @pd: the port
472  * @subport: the subport
473  * @ti: the TID info
474  *
475  * right now we are unlocking one page at a time, but since
476  * the intended use of this routine is for a single group of
477  * virtually contiguous pages, that should change to improve
478  * performance.  We check that the TID is in range for this port
479  * but otherwise don't check validity; if user has an error and
480  * frees the wrong tid, it's only their own data that can thereby
481  * be corrupted.  We do check that the TID was in use, for sanity
482  * We always use our idea of the saved address, not the address that
483  * they pass in to us.
484  */
485
486 static int ipath_tid_free(struct ipath_portdata *pd, unsigned subport,
487                           const struct ipath_tid_info *ti)
488 {
489         int ret = 0;
490         u32 tid, porttid, cnt, limit, tidcnt;
491         struct ipath_devdata *dd = pd->port_dd;
492         u64 __iomem *tidbase;
493         unsigned long tidmap[8];
494
495         if (!dd->ipath_pageshadow) {
496                 ret = -ENOMEM;
497                 goto done;
498         }
499
500         if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap,
501                            sizeof tidmap)) {
502                 ret = -EFAULT;
503                 goto done;
504         }
505
506         porttid = pd->port_port * dd->ipath_rcvtidcnt;
507         if (!pd->port_subport_cnt)
508                 tidcnt = dd->ipath_rcvtidcnt;
509         else if (!subport) {
510                 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
511                          (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
512                 porttid += dd->ipath_rcvtidcnt - tidcnt;
513         } else {
514                 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
515                 porttid += tidcnt * (subport - 1);
516         }
517         tidbase = (u64 __iomem *) ((char __iomem *)(dd->ipath_kregbase) +
518                                    dd->ipath_rcvtidbase +
519                                    porttid * sizeof(*tidbase));
520
521         limit = sizeof(tidmap) * BITS_PER_BYTE;
522         if (limit > tidcnt)
523                 /* just in case size changes in future */
524                 limit = tidcnt;
525         tid = find_first_bit(tidmap, limit);
526         ipath_cdbg(VERBOSE, "Port%u free %u tids; first bit (max=%d) "
527                    "set is %d, porttid %u\n", pd->port_port, ti->tidcnt,
528                    limit, tid, porttid);
529         for (cnt = 0; tid < limit; tid++) {
530                 /*
531                  * small optimization; if we detect a run of 3 or so without
532                  * any set, use find_first_bit again.  That's mainly to
533                  * accelerate the case where we wrapped, so we have some at
534                  * the beginning, and some at the end, and a big gap
535                  * in the middle.
536                  */
537                 if (!test_bit(tid, tidmap))
538                         continue;
539                 cnt++;
540                 if (dd->ipath_pageshadow[porttid + tid]) {
541                         struct page *p;
542                         p = dd->ipath_pageshadow[porttid + tid];
543                         dd->ipath_pageshadow[porttid + tid] = NULL;
544                         ipath_cdbg(VERBOSE, "PID %u freeing TID %u\n",
545                                    pd->port_pid, tid);
546                         dd->ipath_f_put_tid(dd, &tidbase[tid],
547                                             RCVHQ_RCV_TYPE_EXPECTED,
548                                             dd->ipath_tidinvalid);
549                         pci_unmap_page(dd->pcidev,
550                                 dd->ipath_physshadow[porttid + tid],
551                                 PAGE_SIZE, PCI_DMA_FROMDEVICE);
552                         ipath_release_user_pages(&p, 1);
553                         ipath_stats.sps_pageunlocks++;
554                 } else
555                         ipath_dbg("Unused tid %u, ignoring\n", tid);
556         }
557         if (cnt != ti->tidcnt)
558                 ipath_dbg("passed in tidcnt %d, only %d bits set in map\n",
559                           ti->tidcnt, cnt);
560 done:
561         if (ret)
562                 ipath_dbg("Failed to unmap %u TID pages, failing with %d\n",
563                           ti->tidcnt, -ret);
564         return ret;
565 }
566
567 /**
568  * ipath_set_part_key - set a partition key
569  * @pd: the port
570  * @key: the key
571  *
572  * We can have up to 4 active at a time (other than the default, which is
573  * always allowed).  This is somewhat tricky, since multiple ports may set
574  * the same key, so we reference count them, and clean up at exit.  All 4
575  * partition keys are packed into a single infinipath register.  It's an
576  * error for a process to set the same pkey multiple times.  We provide no
577  * mechanism to de-allocate a pkey at this time, we may eventually need to
578  * do that.  I've used the atomic operations, and no locking, and only make
579  * a single pass through what's available.  This should be more than
580  * adequate for some time. I'll think about spinlocks or the like if and as
581  * it's necessary.
582  */
583 static int ipath_set_part_key(struct ipath_portdata *pd, u16 key)
584 {
585         struct ipath_devdata *dd = pd->port_dd;
586         int i, any = 0, pidx = -1;
587         u16 lkey = key & 0x7FFF;
588         int ret;
589
590         if (lkey == (IPATH_DEFAULT_P_KEY & 0x7FFF)) {
591                 /* nothing to do; this key always valid */
592                 ret = 0;
593                 goto bail;
594         }
595
596         ipath_cdbg(VERBOSE, "p%u try to set pkey %hx, current keys "
597                    "%hx:%x %hx:%x %hx:%x %hx:%x\n",
598                    pd->port_port, key, dd->ipath_pkeys[0],
599                    atomic_read(&dd->ipath_pkeyrefs[0]), dd->ipath_pkeys[1],
600                    atomic_read(&dd->ipath_pkeyrefs[1]), dd->ipath_pkeys[2],
601                    atomic_read(&dd->ipath_pkeyrefs[2]), dd->ipath_pkeys[3],
602                    atomic_read(&dd->ipath_pkeyrefs[3]));
603
604         if (!lkey) {
605                 ipath_cdbg(PROC, "p%u tries to set key 0, not allowed\n",
606                            pd->port_port);
607                 ret = -EINVAL;
608                 goto bail;
609         }
610
611         /*
612          * Set the full membership bit, because it has to be
613          * set in the register or the packet, and it seems
614          * cleaner to set in the register than to force all
615          * callers to set it. (see bug 4331)
616          */
617         key |= 0x8000;
618
619         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
620                 if (!pd->port_pkeys[i] && pidx == -1)
621                         pidx = i;
622                 if (pd->port_pkeys[i] == key) {
623                         ipath_cdbg(VERBOSE, "p%u tries to set same pkey "
624                                    "(%x) more than once\n",
625                                    pd->port_port, key);
626                         ret = -EEXIST;
627                         goto bail;
628                 }
629         }
630         if (pidx == -1) {
631                 ipath_dbg("All pkeys for port %u already in use, "
632                           "can't set %x\n", pd->port_port, key);
633                 ret = -EBUSY;
634                 goto bail;
635         }
636         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
637                 if (!dd->ipath_pkeys[i]) {
638                         any++;
639                         continue;
640                 }
641                 if (dd->ipath_pkeys[i] == key) {
642                         atomic_t *pkrefs = &dd->ipath_pkeyrefs[i];
643
644                         if (atomic_inc_return(pkrefs) > 1) {
645                                 pd->port_pkeys[pidx] = key;
646                                 ipath_cdbg(VERBOSE, "p%u set key %x "
647                                            "matches #%d, count now %d\n",
648                                            pd->port_port, key, i,
649                                            atomic_read(pkrefs));
650                                 ret = 0;
651                                 goto bail;
652                         } else {
653                                 /*
654                                  * lost race, decrement count, catch below
655                                  */
656                                 atomic_dec(pkrefs);
657                                 ipath_cdbg(VERBOSE, "Lost race, count was "
658                                            "0, after dec, it's %d\n",
659                                            atomic_read(pkrefs));
660                                 any++;
661                         }
662                 }
663                 if ((dd->ipath_pkeys[i] & 0x7FFF) == lkey) {
664                         /*
665                          * It makes no sense to have both the limited and
666                          * full membership PKEY set at the same time since
667                          * the unlimited one will disable the limited one.
668                          */
669                         ret = -EEXIST;
670                         goto bail;
671                 }
672         }
673         if (!any) {
674                 ipath_dbg("port %u, all pkeys already in use, "
675                           "can't set %x\n", pd->port_port, key);
676                 ret = -EBUSY;
677                 goto bail;
678         }
679         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
680                 if (!dd->ipath_pkeys[i] &&
681                     atomic_inc_return(&dd->ipath_pkeyrefs[i]) == 1) {
682                         u64 pkey;
683
684                         /* for ipathstats, etc. */
685                         ipath_stats.sps_pkeys[i] = lkey;
686                         pd->port_pkeys[pidx] = dd->ipath_pkeys[i] = key;
687                         pkey =
688                                 (u64) dd->ipath_pkeys[0] |
689                                 ((u64) dd->ipath_pkeys[1] << 16) |
690                                 ((u64) dd->ipath_pkeys[2] << 32) |
691                                 ((u64) dd->ipath_pkeys[3] << 48);
692                         ipath_cdbg(PROC, "p%u set key %x in #%d, "
693                                    "portidx %d, new pkey reg %llx\n",
694                                    pd->port_port, key, i, pidx,
695                                    (unsigned long long) pkey);
696                         ipath_write_kreg(
697                                 dd, dd->ipath_kregs->kr_partitionkey, pkey);
698
699                         ret = 0;
700                         goto bail;
701                 }
702         }
703         ipath_dbg("port %u, all pkeys already in use 2nd pass, "
704                   "can't set %x\n", pd->port_port, key);
705         ret = -EBUSY;
706
707 bail:
708         return ret;
709 }
710
711 /**
712  * ipath_manage_rcvq - manage a port's receive queue
713  * @pd: the port
714  * @subport: the subport
715  * @start_stop: action to carry out
716  *
717  * start_stop == 0 disables receive on the port, for use in queue
718  * overflow conditions.  start_stop==1 re-enables, to be used to
719  * re-init the software copy of the head register
720  */
721 static int ipath_manage_rcvq(struct ipath_portdata *pd, unsigned subport,
722                              int start_stop)
723 {
724         struct ipath_devdata *dd = pd->port_dd;
725
726         ipath_cdbg(PROC, "%sabling rcv for unit %u port %u:%u\n",
727                    start_stop ? "en" : "dis", dd->ipath_unit,
728                    pd->port_port, subport);
729         if (subport)
730                 goto bail;
731         /* atomically clear receive enable port. */
732         if (start_stop) {
733                 /*
734                  * On enable, force in-memory copy of the tail register to
735                  * 0, so that protocol code doesn't have to worry about
736                  * whether or not the chip has yet updated the in-memory
737                  * copy or not on return from the system call. The chip
738                  * always resets it's tail register back to 0 on a
739                  * transition from disabled to enabled.  This could cause a
740                  * problem if software was broken, and did the enable w/o
741                  * the disable, but eventually the in-memory copy will be
742                  * updated and correct itself, even in the face of software
743                  * bugs.
744                  */
745                 *(volatile u64 *)pd->port_rcvhdrtail_kvaddr = 0;
746                 set_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
747                         &dd->ipath_rcvctrl);
748         } else
749                 clear_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
750                           &dd->ipath_rcvctrl);
751         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
752                          dd->ipath_rcvctrl);
753         /* now be sure chip saw it before we return */
754         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
755         if (start_stop) {
756                 /*
757                  * And try to be sure that tail reg update has happened too.
758                  * This should in theory interlock with the RXE changes to
759                  * the tail register.  Don't assign it to the tail register
760                  * in memory copy, since we could overwrite an update by the
761                  * chip if we did.
762                  */
763                 ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
764         }
765         /* always; new head should be equal to new tail; see above */
766 bail:
767         return 0;
768 }
769
770 static void ipath_clean_part_key(struct ipath_portdata *pd,
771                                  struct ipath_devdata *dd)
772 {
773         int i, j, pchanged = 0;
774         u64 oldpkey;
775
776         /* for debugging only */
777         oldpkey = (u64) dd->ipath_pkeys[0] |
778                 ((u64) dd->ipath_pkeys[1] << 16) |
779                 ((u64) dd->ipath_pkeys[2] << 32) |
780                 ((u64) dd->ipath_pkeys[3] << 48);
781
782         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
783                 if (!pd->port_pkeys[i])
784                         continue;
785                 ipath_cdbg(VERBOSE, "look for key[%d] %hx in pkeys\n", i,
786                            pd->port_pkeys[i]);
787                 for (j = 0; j < ARRAY_SIZE(dd->ipath_pkeys); j++) {
788                         /* check for match independent of the global bit */
789                         if ((dd->ipath_pkeys[j] & 0x7fff) !=
790                             (pd->port_pkeys[i] & 0x7fff))
791                                 continue;
792                         if (atomic_dec_and_test(&dd->ipath_pkeyrefs[j])) {
793                                 ipath_cdbg(VERBOSE, "p%u clear key "
794                                            "%x matches #%d\n",
795                                            pd->port_port,
796                                            pd->port_pkeys[i], j);
797                                 ipath_stats.sps_pkeys[j] =
798                                         dd->ipath_pkeys[j] = 0;
799                                 pchanged++;
800                         }
801                         else ipath_cdbg(
802                                 VERBOSE, "p%u key %x matches #%d, "
803                                 "but ref still %d\n", pd->port_port,
804                                 pd->port_pkeys[i], j,
805                                 atomic_read(&dd->ipath_pkeyrefs[j]));
806                         break;
807                 }
808                 pd->port_pkeys[i] = 0;
809         }
810         if (pchanged) {
811                 u64 pkey = (u64) dd->ipath_pkeys[0] |
812                         ((u64) dd->ipath_pkeys[1] << 16) |
813                         ((u64) dd->ipath_pkeys[2] << 32) |
814                         ((u64) dd->ipath_pkeys[3] << 48);
815                 ipath_cdbg(VERBOSE, "p%u old pkey reg %llx, "
816                            "new pkey reg %llx\n", pd->port_port,
817                            (unsigned long long) oldpkey,
818                            (unsigned long long) pkey);
819                 ipath_write_kreg(dd, dd->ipath_kregs->kr_partitionkey,
820                                  pkey);
821         }
822 }
823
824 /*
825  * Initialize the port data with the receive buffer sizes
826  * so this can be done while the master port is locked.
827  * Otherwise, there is a race with a slave opening the port
828  * and seeing these fields uninitialized.
829  */
830 static void init_user_egr_sizes(struct ipath_portdata *pd)
831 {
832         struct ipath_devdata *dd = pd->port_dd;
833         unsigned egrperchunk, egrcnt, size;
834
835         /*
836          * to avoid wasting a lot of memory, we allocate 32KB chunks of
837          * physically contiguous memory, advance through it until used up
838          * and then allocate more.  Of course, we need memory to store those
839          * extra pointers, now.  Started out with 256KB, but under heavy
840          * memory pressure (creating large files and then copying them over
841          * NFS while doing lots of MPI jobs), we hit some allocation
842          * failures, even though we can sleep...  (2.6.10) Still get
843          * failures at 64K.  32K is the lowest we can go without wasting
844          * additional memory.
845          */
846         size = 0x8000;
847         egrperchunk = size / dd->ipath_rcvegrbufsize;
848         egrcnt = dd->ipath_rcvegrcnt;
849         pd->port_rcvegrbuf_chunks = (egrcnt + egrperchunk - 1) / egrperchunk;
850         pd->port_rcvegrbufs_perchunk = egrperchunk;
851         pd->port_rcvegrbuf_size = size;
852 }
853
854 /**
855  * ipath_create_user_egr - allocate eager TID buffers
856  * @pd: the port to allocate TID buffers for
857  *
858  * This routine is now quite different for user and kernel, because
859  * the kernel uses skb's, for the accelerated network performance
860  * This is the user port version
861  *
862  * Allocate the eager TID buffers and program them into infinipath
863  * They are no longer completely contiguous, we do multiple allocation
864  * calls.
865  */
866 static int ipath_create_user_egr(struct ipath_portdata *pd)
867 {
868         struct ipath_devdata *dd = pd->port_dd;
869         unsigned e, egrcnt, egrperchunk, chunk, egrsize, egroff;
870         size_t size;
871         int ret;
872         gfp_t gfp_flags;
873
874         /*
875          * GFP_USER, but without GFP_FS, so buffer cache can be
876          * coalesced (we hope); otherwise, even at order 4,
877          * heavy filesystem activity makes these fail, and we can
878          * use compound pages.
879          */
880         gfp_flags = __GFP_WAIT | __GFP_IO | __GFP_COMP;
881
882         egrcnt = dd->ipath_rcvegrcnt;
883         /* TID number offset for this port */
884         egroff = pd->port_port * egrcnt;
885         egrsize = dd->ipath_rcvegrbufsize;
886         ipath_cdbg(VERBOSE, "Allocating %d egr buffers, at egrtid "
887                    "offset %x, egrsize %u\n", egrcnt, egroff, egrsize);
888
889         chunk = pd->port_rcvegrbuf_chunks;
890         egrperchunk = pd->port_rcvegrbufs_perchunk;
891         size = pd->port_rcvegrbuf_size;
892         pd->port_rcvegrbuf = kmalloc(chunk * sizeof(pd->port_rcvegrbuf[0]),
893                                      GFP_KERNEL);
894         if (!pd->port_rcvegrbuf) {
895                 ret = -ENOMEM;
896                 goto bail;
897         }
898         pd->port_rcvegrbuf_phys =
899                 kmalloc(chunk * sizeof(pd->port_rcvegrbuf_phys[0]),
900                         GFP_KERNEL);
901         if (!pd->port_rcvegrbuf_phys) {
902                 ret = -ENOMEM;
903                 goto bail_rcvegrbuf;
904         }
905         for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
906
907                 pd->port_rcvegrbuf[e] = dma_alloc_coherent(
908                         &dd->pcidev->dev, size, &pd->port_rcvegrbuf_phys[e],
909                         gfp_flags);
910
911                 if (!pd->port_rcvegrbuf[e]) {
912                         ret = -ENOMEM;
913                         goto bail_rcvegrbuf_phys;
914                 }
915         }
916
917         pd->port_rcvegr_phys = pd->port_rcvegrbuf_phys[0];
918
919         for (e = chunk = 0; chunk < pd->port_rcvegrbuf_chunks; chunk++) {
920                 dma_addr_t pa = pd->port_rcvegrbuf_phys[chunk];
921                 unsigned i;
922
923                 for (i = 0; e < egrcnt && i < egrperchunk; e++, i++) {
924                         dd->ipath_f_put_tid(dd, e + egroff +
925                                             (u64 __iomem *)
926                                             ((char __iomem *)
927                                              dd->ipath_kregbase +
928                                              dd->ipath_rcvegrbase),
929                                             RCVHQ_RCV_TYPE_EAGER, pa);
930                         pa += egrsize;
931                 }
932                 cond_resched(); /* don't hog the cpu */
933         }
934
935         ret = 0;
936         goto bail;
937
938 bail_rcvegrbuf_phys:
939         for (e = 0; e < pd->port_rcvegrbuf_chunks &&
940                 pd->port_rcvegrbuf[e]; e++) {
941                 dma_free_coherent(&dd->pcidev->dev, size,
942                                   pd->port_rcvegrbuf[e],
943                                   pd->port_rcvegrbuf_phys[e]);
944
945         }
946         kfree(pd->port_rcvegrbuf_phys);
947         pd->port_rcvegrbuf_phys = NULL;
948 bail_rcvegrbuf:
949         kfree(pd->port_rcvegrbuf);
950         pd->port_rcvegrbuf = NULL;
951 bail:
952         return ret;
953 }
954
955
956 /* common code for the mappings on dma_alloc_coherent mem */
957 static int ipath_mmap_mem(struct vm_area_struct *vma,
958         struct ipath_portdata *pd, unsigned len, int write_ok,
959         void *kvaddr, char *what)
960 {
961         struct ipath_devdata *dd = pd->port_dd;
962         unsigned long pfn;
963         int ret;
964
965         if ((vma->vm_end - vma->vm_start) > len) {
966                 dev_info(&dd->pcidev->dev,
967                          "FAIL on %s: len %lx > %x\n", what,
968                          vma->vm_end - vma->vm_start, len);
969                 ret = -EFAULT;
970                 goto bail;
971         }
972
973         if (!write_ok) {
974                 if (vma->vm_flags & VM_WRITE) {
975                         dev_info(&dd->pcidev->dev,
976                                  "%s must be mapped readonly\n", what);
977                         ret = -EPERM;
978                         goto bail;
979                 }
980
981                 /* don't allow them to later change with mprotect */
982                 vma->vm_flags &= ~VM_MAYWRITE;
983         }
984
985         pfn = virt_to_phys(kvaddr) >> PAGE_SHIFT;
986         ret = remap_pfn_range(vma, vma->vm_start, pfn,
987                               len, vma->vm_page_prot);
988         if (ret)
989                 dev_info(&dd->pcidev->dev, "%s port%u mmap of %lx, %x "
990                          "bytes r%c failed: %d\n", what, pd->port_port,
991                          pfn, len, write_ok?'w':'o', ret);
992         else
993                 ipath_cdbg(VERBOSE, "%s port%u mmaped %lx, %x bytes "
994                            "r%c\n", what, pd->port_port, pfn, len,
995                            write_ok?'w':'o');
996 bail:
997         return ret;
998 }
999
1000 static int mmap_ureg(struct vm_area_struct *vma, struct ipath_devdata *dd,
1001                      u64 ureg)
1002 {
1003         unsigned long phys;
1004         int ret;
1005
1006         /*
1007          * This is real hardware, so use io_remap.  This is the mechanism
1008          * for the user process to update the head registers for their port
1009          * in the chip.
1010          */
1011         if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) {
1012                 dev_info(&dd->pcidev->dev, "FAIL mmap userreg: reqlen "
1013                          "%lx > PAGE\n", vma->vm_end - vma->vm_start);
1014                 ret = -EFAULT;
1015         } else {
1016                 phys = dd->ipath_physaddr + ureg;
1017                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1018
1019                 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1020                 ret = io_remap_pfn_range(vma, vma->vm_start,
1021                                          phys >> PAGE_SHIFT,
1022                                          vma->vm_end - vma->vm_start,
1023                                          vma->vm_page_prot);
1024         }
1025         return ret;
1026 }
1027
1028 static int mmap_piobufs(struct vm_area_struct *vma,
1029                         struct ipath_devdata *dd,
1030                         struct ipath_portdata *pd,
1031                         unsigned piobufs, unsigned piocnt)
1032 {
1033         unsigned long phys;
1034         int ret;
1035
1036         /*
1037          * When we map the PIO buffers in the chip, we want to map them as
1038          * writeonly, no read possible.   This prevents access to previous
1039          * process data, and catches users who might try to read the i/o
1040          * space due to a bug.
1041          */
1042         if ((vma->vm_end - vma->vm_start) > (piocnt * dd->ipath_palign)) {
1043                 dev_info(&dd->pcidev->dev, "FAIL mmap piobufs: "
1044                          "reqlen %lx > PAGE\n",
1045                          vma->vm_end - vma->vm_start);
1046                 ret = -EINVAL;
1047                 goto bail;
1048         }
1049
1050         phys = dd->ipath_physaddr + piobufs;
1051
1052         /*
1053          * Don't mark this as non-cached, or we don't get the
1054          * write combining behavior we want on the PIO buffers!
1055          */
1056
1057 #if defined(__powerpc__)
1058         /* There isn't a generic way to specify writethrough mappings */
1059         pgprot_val(vma->vm_page_prot) |= _PAGE_NO_CACHE;
1060         pgprot_val(vma->vm_page_prot) |= _PAGE_WRITETHRU;
1061         pgprot_val(vma->vm_page_prot) &= ~_PAGE_GUARDED;
1062 #endif
1063
1064         /*
1065          * don't allow them to later change to readable with mprotect (for when
1066          * not initially mapped readable, as is normally the case)
1067          */
1068         vma->vm_flags &= ~VM_MAYREAD;
1069         vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1070
1071         ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
1072                                  vma->vm_end - vma->vm_start,
1073                                  vma->vm_page_prot);
1074 bail:
1075         return ret;
1076 }
1077
1078 static int mmap_rcvegrbufs(struct vm_area_struct *vma,
1079                            struct ipath_portdata *pd)
1080 {
1081         struct ipath_devdata *dd = pd->port_dd;
1082         unsigned long start, size;
1083         size_t total_size, i;
1084         unsigned long pfn;
1085         int ret;
1086
1087         size = pd->port_rcvegrbuf_size;
1088         total_size = pd->port_rcvegrbuf_chunks * size;
1089         if ((vma->vm_end - vma->vm_start) > total_size) {
1090                 dev_info(&dd->pcidev->dev, "FAIL on egr bufs: "
1091                          "reqlen %lx > actual %lx\n",
1092                          vma->vm_end - vma->vm_start,
1093                          (unsigned long) total_size);
1094                 ret = -EINVAL;
1095                 goto bail;
1096         }
1097
1098         if (vma->vm_flags & VM_WRITE) {
1099                 dev_info(&dd->pcidev->dev, "Can't map eager buffers as "
1100                          "writable (flags=%lx)\n", vma->vm_flags);
1101                 ret = -EPERM;
1102                 goto bail;
1103         }
1104         /* don't allow them to later change to writeable with mprotect */
1105         vma->vm_flags &= ~VM_MAYWRITE;
1106
1107         start = vma->vm_start;
1108
1109         for (i = 0; i < pd->port_rcvegrbuf_chunks; i++, start += size) {
1110                 pfn = virt_to_phys(pd->port_rcvegrbuf[i]) >> PAGE_SHIFT;
1111                 ret = remap_pfn_range(vma, start, pfn, size,
1112                                       vma->vm_page_prot);
1113                 if (ret < 0)
1114                         goto bail;
1115         }
1116         ret = 0;
1117
1118 bail:
1119         return ret;
1120 }
1121
1122 /*
1123  * ipath_file_vma_fault - handle a VMA page fault.
1124  */
1125 static int ipath_file_vma_fault(struct vm_area_struct *vma,
1126                                         struct vm_fault *vmf)
1127 {
1128         struct page *page;
1129
1130         page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
1131         if (!page)
1132                 return VM_FAULT_SIGBUS;
1133         get_page(page);
1134         vmf->page = page;
1135
1136         return 0;
1137 }
1138
1139 static struct vm_operations_struct ipath_file_vm_ops = {
1140         .fault = ipath_file_vma_fault,
1141 };
1142
1143 static int mmap_kvaddr(struct vm_area_struct *vma, u64 pgaddr,
1144                        struct ipath_portdata *pd, unsigned subport)
1145 {
1146         unsigned long len;
1147         struct ipath_devdata *dd;
1148         void *addr;
1149         size_t size;
1150         int ret = 0;
1151
1152         /* If the port is not shared, all addresses should be physical */
1153         if (!pd->port_subport_cnt)
1154                 goto bail;
1155
1156         dd = pd->port_dd;
1157         size = pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
1158
1159         /*
1160          * Each process has all the subport uregbase, rcvhdrq, and
1161          * rcvegrbufs mmapped - as an array for all the processes,
1162          * and also separately for this process.
1163          */
1164         if (pgaddr == cvt_kvaddr(pd->subport_uregbase)) {
1165                 addr = pd->subport_uregbase;
1166                 size = PAGE_SIZE * pd->port_subport_cnt;
1167         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base)) {
1168                 addr = pd->subport_rcvhdr_base;
1169                 size = pd->port_rcvhdrq_size * pd->port_subport_cnt;
1170         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf)) {
1171                 addr = pd->subport_rcvegrbuf;
1172                 size *= pd->port_subport_cnt;
1173         } else if (pgaddr == cvt_kvaddr(pd->subport_uregbase +
1174                                         PAGE_SIZE * subport)) {
1175                 addr = pd->subport_uregbase + PAGE_SIZE * subport;
1176                 size = PAGE_SIZE;
1177         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base +
1178                                 pd->port_rcvhdrq_size * subport)) {
1179                 addr = pd->subport_rcvhdr_base +
1180                         pd->port_rcvhdrq_size * subport;
1181                 size = pd->port_rcvhdrq_size;
1182         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf +
1183                                size * subport)) {
1184                 addr = pd->subport_rcvegrbuf + size * subport;
1185                 /* rcvegrbufs are read-only on the slave */
1186                 if (vma->vm_flags & VM_WRITE) {
1187                         dev_info(&dd->pcidev->dev,
1188                                  "Can't map eager buffers as "
1189                                  "writable (flags=%lx)\n", vma->vm_flags);
1190                         ret = -EPERM;
1191                         goto bail;
1192                 }
1193                 /*
1194                  * Don't allow permission to later change to writeable
1195                  * with mprotect.
1196                  */
1197                 vma->vm_flags &= ~VM_MAYWRITE;
1198         } else {
1199                 goto bail;
1200         }
1201         len = vma->vm_end - vma->vm_start;
1202         if (len > size) {
1203                 ipath_cdbg(MM, "FAIL: reqlen %lx > %zx\n", len, size);
1204                 ret = -EINVAL;
1205                 goto bail;
1206         }
1207
1208         vma->vm_pgoff = (unsigned long) addr >> PAGE_SHIFT;
1209         vma->vm_ops = &ipath_file_vm_ops;
1210         vma->vm_flags |= VM_RESERVED | VM_DONTEXPAND;
1211         ret = 1;
1212
1213 bail:
1214         return ret;
1215 }
1216
1217 /**
1218  * ipath_mmap - mmap various structures into user space
1219  * @fp: the file pointer
1220  * @vma: the VM area
1221  *
1222  * We use this to have a shared buffer between the kernel and the user code
1223  * for the rcvhdr queue, egr buffers, and the per-port user regs and pio
1224  * buffers in the chip.  We have the open and close entries so we can bump
1225  * the ref count and keep the driver from being unloaded while still mapped.
1226  */
1227 static int ipath_mmap(struct file *fp, struct vm_area_struct *vma)
1228 {
1229         struct ipath_portdata *pd;
1230         struct ipath_devdata *dd;
1231         u64 pgaddr, ureg;
1232         unsigned piobufs, piocnt;
1233         int ret;
1234
1235         pd = port_fp(fp);
1236         if (!pd) {
1237                 ret = -EINVAL;
1238                 goto bail;
1239         }
1240         dd = pd->port_dd;
1241
1242         /*
1243          * This is the ipath_do_user_init() code, mapping the shared buffers
1244          * into the user process. The address referred to by vm_pgoff is the
1245          * file offset passed via mmap().  For shared ports, this is the
1246          * kernel vmalloc() address of the pages to share with the master.
1247          * For non-shared or master ports, this is a physical address.
1248          * We only do one mmap for each space mapped.
1249          */
1250         pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1251
1252         /*
1253          * Check for 0 in case one of the allocations failed, but user
1254          * called mmap anyway.
1255          */
1256         if (!pgaddr)  {
1257                 ret = -EINVAL;
1258                 goto bail;
1259         }
1260
1261         ipath_cdbg(MM, "pgaddr %llx vm_start=%lx len %lx port %u:%u:%u\n",
1262                    (unsigned long long) pgaddr, vma->vm_start,
1263                    vma->vm_end - vma->vm_start, dd->ipath_unit,
1264                    pd->port_port, subport_fp(fp));
1265
1266         /*
1267          * Physical addresses must fit in 40 bits for our hardware.
1268          * Check for kernel virtual addresses first, anything else must
1269          * match a HW or memory address.
1270          */
1271         ret = mmap_kvaddr(vma, pgaddr, pd, subport_fp(fp));
1272         if (ret) {
1273                 if (ret > 0)
1274                         ret = 0;
1275                 goto bail;
1276         }
1277
1278         ureg = dd->ipath_uregbase + dd->ipath_palign * pd->port_port;
1279         if (!pd->port_subport_cnt) {
1280                 /* port is not shared */
1281                 piocnt = dd->ipath_pbufsport;
1282                 piobufs = pd->port_piobufs;
1283         } else if (!subport_fp(fp)) {
1284                 /* caller is the master */
1285                 piocnt = (dd->ipath_pbufsport / pd->port_subport_cnt) +
1286                          (dd->ipath_pbufsport % pd->port_subport_cnt);
1287                 piobufs = pd->port_piobufs +
1288                         dd->ipath_palign * (dd->ipath_pbufsport - piocnt);
1289         } else {
1290                 unsigned slave = subport_fp(fp) - 1;
1291
1292                 /* caller is a slave */
1293                 piocnt = dd->ipath_pbufsport / pd->port_subport_cnt;
1294                 piobufs = pd->port_piobufs + dd->ipath_palign * piocnt * slave;
1295         }
1296
1297         if (pgaddr == ureg)
1298                 ret = mmap_ureg(vma, dd, ureg);
1299         else if (pgaddr == piobufs)
1300                 ret = mmap_piobufs(vma, dd, pd, piobufs, piocnt);
1301         else if (pgaddr == dd->ipath_pioavailregs_phys)
1302                 /* in-memory copy of pioavail registers */
1303                 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1304                                      (void *) dd->ipath_pioavailregs_dma,
1305                                      "pioavail registers");
1306         else if (pgaddr == pd->port_rcvegr_phys)
1307                 ret = mmap_rcvegrbufs(vma, pd);
1308         else if (pgaddr == (u64) pd->port_rcvhdrq_phys)
1309                 /*
1310                  * The rcvhdrq itself; readonly except on HT (so have
1311                  * to allow writable mapping), multiple pages, contiguous
1312                  * from an i/o perspective.
1313                  */
1314                 ret = ipath_mmap_mem(vma, pd, pd->port_rcvhdrq_size, 1,
1315                                      pd->port_rcvhdrq,
1316                                      "rcvhdrq");
1317         else if (pgaddr == (u64) pd->port_rcvhdrqtailaddr_phys)
1318                 /* in-memory copy of rcvhdrq tail register */
1319                 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1320                                      pd->port_rcvhdrtail_kvaddr,
1321                                      "rcvhdrq tail");
1322         else
1323                 ret = -EINVAL;
1324
1325         vma->vm_private_data = NULL;
1326
1327         if (ret < 0)
1328                 dev_info(&dd->pcidev->dev,
1329                          "Failure %d on off %llx len %lx\n",
1330                          -ret, (unsigned long long)pgaddr,
1331                          vma->vm_end - vma->vm_start);
1332 bail:
1333         return ret;
1334 }
1335
1336 static unsigned ipath_poll_hdrqfull(struct ipath_portdata *pd)
1337 {
1338         unsigned pollflag = 0;
1339
1340         if ((pd->poll_type & IPATH_POLL_TYPE_OVERFLOW) &&
1341             pd->port_hdrqfull != pd->port_hdrqfull_poll) {
1342                 pollflag |= POLLIN | POLLRDNORM;
1343                 pd->port_hdrqfull_poll = pd->port_hdrqfull;
1344         }
1345
1346         return pollflag;
1347 }
1348
1349 static unsigned int ipath_poll_urgent(struct ipath_portdata *pd,
1350                                       struct file *fp,
1351                                       struct poll_table_struct *pt)
1352 {
1353         unsigned pollflag = 0;
1354         struct ipath_devdata *dd;
1355
1356         dd = pd->port_dd;
1357
1358         /* variable access in ipath_poll_hdrqfull() needs this */
1359         rmb();
1360         pollflag = ipath_poll_hdrqfull(pd);
1361
1362         if (pd->port_urgent != pd->port_urgent_poll) {
1363                 pollflag |= POLLIN | POLLRDNORM;
1364                 pd->port_urgent_poll = pd->port_urgent;
1365         }
1366
1367         if (!pollflag) {
1368                 /* this saves a spin_lock/unlock in interrupt handler... */
1369                 set_bit(IPATH_PORT_WAITING_URG, &pd->port_flag);
1370                 /* flush waiting flag so don't miss an event... */
1371                 wmb();
1372                 poll_wait(fp, &pd->port_wait, pt);
1373         }
1374
1375         return pollflag;
1376 }
1377
1378 static unsigned int ipath_poll_next(struct ipath_portdata *pd,
1379                                     struct file *fp,
1380                                     struct poll_table_struct *pt)
1381 {
1382         u32 head;
1383         u32 tail;
1384         unsigned pollflag = 0;
1385         struct ipath_devdata *dd;
1386
1387         dd = pd->port_dd;
1388
1389         /* variable access in ipath_poll_hdrqfull() needs this */
1390         rmb();
1391         pollflag = ipath_poll_hdrqfull(pd);
1392
1393         head = ipath_read_ureg32(dd, ur_rcvhdrhead, pd->port_port);
1394         tail = *(volatile u64 *)pd->port_rcvhdrtail_kvaddr;
1395
1396         if (head != tail)
1397                 pollflag |= POLLIN | POLLRDNORM;
1398         else {
1399                 /* this saves a spin_lock/unlock in interrupt handler */
1400                 set_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
1401                 /* flush waiting flag so we don't miss an event */
1402                 wmb();
1403
1404                 set_bit(pd->port_port + INFINIPATH_R_INTRAVAIL_SHIFT,
1405                         &dd->ipath_rcvctrl);
1406
1407                 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1408                                  dd->ipath_rcvctrl);
1409
1410                 if (dd->ipath_rhdrhead_intr_off) /* arm rcv interrupt */
1411                         ipath_write_ureg(dd, ur_rcvhdrhead,
1412                                          dd->ipath_rhdrhead_intr_off | head,
1413                                          pd->port_port);
1414
1415                 poll_wait(fp, &pd->port_wait, pt);
1416         }
1417
1418         return pollflag;
1419 }
1420
1421 static unsigned int ipath_poll(struct file *fp,
1422                                struct poll_table_struct *pt)
1423 {
1424         struct ipath_portdata *pd;
1425         unsigned pollflag;
1426
1427         pd = port_fp(fp);
1428         if (!pd)
1429                 pollflag = 0;
1430         else if (pd->poll_type & IPATH_POLL_TYPE_URGENT)
1431                 pollflag = ipath_poll_urgent(pd, fp, pt);
1432         else
1433                 pollflag = ipath_poll_next(pd, fp, pt);
1434
1435         return pollflag;
1436 }
1437
1438 static int ipath_supports_subports(int user_swmajor, int user_swminor)
1439 {
1440         /* no subport implementation prior to software version 1.3 */
1441         return (user_swmajor > 1) || (user_swminor >= 3);
1442 }
1443
1444 static int ipath_compatible_subports(int user_swmajor, int user_swminor)
1445 {
1446         /* this code is written long-hand for clarity */
1447         if (IPATH_USER_SWMAJOR != user_swmajor) {
1448                 /* no promise of compatibility if major mismatch */
1449                 return 0;
1450         }
1451         if (IPATH_USER_SWMAJOR == 1) {
1452                 switch (IPATH_USER_SWMINOR) {
1453                 case 0:
1454                 case 1:
1455                 case 2:
1456                         /* no subport implementation so cannot be compatible */
1457                         return 0;
1458                 case 3:
1459                         /* 3 is only compatible with itself */
1460                         return user_swminor == 3;
1461                 default:
1462                         /* >= 4 are compatible (or are expected to be) */
1463                         return user_swminor >= 4;
1464                 }
1465         }
1466         /* make no promises yet for future major versions */
1467         return 0;
1468 }
1469
1470 static int init_subports(struct ipath_devdata *dd,
1471                          struct ipath_portdata *pd,
1472                          const struct ipath_user_info *uinfo)
1473 {
1474         int ret = 0;
1475         unsigned num_subports;
1476         size_t size;
1477
1478         /*
1479          * If the user is requesting zero subports,
1480          * skip the subport allocation.
1481          */
1482         if (uinfo->spu_subport_cnt <= 0)
1483                 goto bail;
1484
1485         /* Self-consistency check for ipath_compatible_subports() */
1486         if (ipath_supports_subports(IPATH_USER_SWMAJOR, IPATH_USER_SWMINOR) &&
1487             !ipath_compatible_subports(IPATH_USER_SWMAJOR,
1488                                        IPATH_USER_SWMINOR)) {
1489                 dev_info(&dd->pcidev->dev,
1490                          "Inconsistent ipath_compatible_subports()\n");
1491                 goto bail;
1492         }
1493
1494         /* Check for subport compatibility */
1495         if (!ipath_compatible_subports(uinfo->spu_userversion >> 16,
1496                                        uinfo->spu_userversion & 0xffff)) {
1497                 dev_info(&dd->pcidev->dev,
1498                          "Mismatched user version (%d.%d) and driver "
1499                          "version (%d.%d) while port sharing. Ensure "
1500                          "that driver and library are from the same "
1501                          "release.\n",
1502                          (int) (uinfo->spu_userversion >> 16),
1503                          (int) (uinfo->spu_userversion & 0xffff),
1504                          IPATH_USER_SWMAJOR,
1505                          IPATH_USER_SWMINOR);
1506                 goto bail;
1507         }
1508         if (uinfo->spu_subport_cnt > INFINIPATH_MAX_SUBPORT) {
1509                 ret = -EINVAL;
1510                 goto bail;
1511         }
1512
1513         num_subports = uinfo->spu_subport_cnt;
1514         pd->subport_uregbase = vmalloc(PAGE_SIZE * num_subports);
1515         if (!pd->subport_uregbase) {
1516                 ret = -ENOMEM;
1517                 goto bail;
1518         }
1519         /* Note: pd->port_rcvhdrq_size isn't initialized yet. */
1520         size = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
1521                      sizeof(u32), PAGE_SIZE) * num_subports;
1522         pd->subport_rcvhdr_base = vmalloc(size);
1523         if (!pd->subport_rcvhdr_base) {
1524                 ret = -ENOMEM;
1525                 goto bail_ureg;
1526         }
1527
1528         pd->subport_rcvegrbuf = vmalloc(pd->port_rcvegrbuf_chunks *
1529                                         pd->port_rcvegrbuf_size *
1530                                         num_subports);
1531         if (!pd->subport_rcvegrbuf) {
1532                 ret = -ENOMEM;
1533                 goto bail_rhdr;
1534         }
1535
1536         pd->port_subport_cnt = uinfo->spu_subport_cnt;
1537         pd->port_subport_id = uinfo->spu_subport_id;
1538         pd->active_slaves = 1;
1539         set_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
1540         memset(pd->subport_uregbase, 0, PAGE_SIZE * num_subports);
1541         memset(pd->subport_rcvhdr_base, 0, size);
1542         memset(pd->subport_rcvegrbuf, 0, pd->port_rcvegrbuf_chunks *
1543                                          pd->port_rcvegrbuf_size *
1544                                          num_subports);
1545         goto bail;
1546
1547 bail_rhdr:
1548         vfree(pd->subport_rcvhdr_base);
1549 bail_ureg:
1550         vfree(pd->subport_uregbase);
1551         pd->subport_uregbase = NULL;
1552 bail:
1553         return ret;
1554 }
1555
1556 static int try_alloc_port(struct ipath_devdata *dd, int port,
1557                           struct file *fp,
1558                           const struct ipath_user_info *uinfo)
1559 {
1560         struct ipath_portdata *pd;
1561         int ret;
1562
1563         if (!(pd = dd->ipath_pd[port])) {
1564                 void *ptmp;
1565
1566                 pd = kzalloc(sizeof(struct ipath_portdata), GFP_KERNEL);
1567
1568                 /*
1569                  * Allocate memory for use in ipath_tid_update() just once
1570                  * at open, not per call.  Reduces cost of expected send
1571                  * setup.
1572                  */
1573                 ptmp = kmalloc(dd->ipath_rcvtidcnt * sizeof(u16) +
1574                                dd->ipath_rcvtidcnt * sizeof(struct page **),
1575                                GFP_KERNEL);
1576                 if (!pd || !ptmp) {
1577                         ipath_dev_err(dd, "Unable to allocate portdata "
1578                                       "memory, failing open\n");
1579                         ret = -ENOMEM;
1580                         kfree(pd);
1581                         kfree(ptmp);
1582                         goto bail;
1583                 }
1584                 dd->ipath_pd[port] = pd;
1585                 dd->ipath_pd[port]->port_port = port;
1586                 dd->ipath_pd[port]->port_dd = dd;
1587                 dd->ipath_pd[port]->port_tid_pg_list = ptmp;
1588                 init_waitqueue_head(&dd->ipath_pd[port]->port_wait);
1589         }
1590         if (!pd->port_cnt) {
1591                 pd->userversion = uinfo->spu_userversion;
1592                 init_user_egr_sizes(pd);
1593                 if ((ret = init_subports(dd, pd, uinfo)) != 0)
1594                         goto bail;
1595                 ipath_cdbg(PROC, "%s[%u] opened unit:port %u:%u\n",
1596                            current->comm, current->pid, dd->ipath_unit,
1597                            port);
1598                 pd->port_cnt = 1;
1599                 port_fp(fp) = pd;
1600                 pd->port_pid = current->pid;
1601                 strncpy(pd->port_comm, current->comm, sizeof(pd->port_comm));
1602                 ipath_stats.sps_ports++;
1603                 ret = 0;
1604         } else
1605                 ret = -EBUSY;
1606
1607 bail:
1608         return ret;
1609 }
1610
1611 static inline int usable(struct ipath_devdata *dd)
1612 {
1613         return dd &&
1614                 (dd->ipath_flags & IPATH_PRESENT) &&
1615                 dd->ipath_kregbase &&
1616                 dd->ipath_lid &&
1617                 !(dd->ipath_flags & (IPATH_LINKDOWN | IPATH_DISABLED
1618                                      | IPATH_LINKUNK));
1619 }
1620
1621 static int find_free_port(int unit, struct file *fp,
1622                           const struct ipath_user_info *uinfo)
1623 {
1624         struct ipath_devdata *dd = ipath_lookup(unit);
1625         int ret, i;
1626
1627         if (!dd) {
1628                 ret = -ENODEV;
1629                 goto bail;
1630         }
1631
1632         if (!usable(dd)) {
1633                 ret = -ENETDOWN;
1634                 goto bail;
1635         }
1636
1637         for (i = 1; i < dd->ipath_cfgports; i++) {
1638                 ret = try_alloc_port(dd, i, fp, uinfo);
1639                 if (ret != -EBUSY)
1640                         goto bail;
1641         }
1642         ret = -EBUSY;
1643
1644 bail:
1645         return ret;
1646 }
1647
1648 static int find_best_unit(struct file *fp,
1649                           const struct ipath_user_info *uinfo)
1650 {
1651         int ret = 0, i, prefunit = -1, devmax;
1652         int maxofallports, npresent, nup;
1653         int ndev;
1654
1655         devmax = ipath_count_units(&npresent, &nup, &maxofallports);
1656
1657         /*
1658          * This code is present to allow a knowledgeable person to
1659          * specify the layout of processes to processors before opening
1660          * this driver, and then we'll assign the process to the "closest"
1661          * InfiniPath chip to that processor (we assume reasonable connectivity,
1662          * for now).  This code assumes that if affinity has been set
1663          * before this point, that at most one cpu is set; for now this
1664          * is reasonable.  I check for both cpus_empty() and cpus_full(),
1665          * in case some kernel variant sets none of the bits when no
1666          * affinity is set.  2.6.11 and 12 kernels have all present
1667          * cpus set.  Some day we'll have to fix it up further to handle
1668          * a cpu subset.  This algorithm fails for two HT chips connected
1669          * in tunnel fashion.  Eventually this needs real topology
1670          * information.  There may be some issues with dual core numbering
1671          * as well.  This needs more work prior to release.
1672          */
1673         if (!cpus_empty(current->cpus_allowed) &&
1674             !cpus_full(current->cpus_allowed)) {
1675                 int ncpus = num_online_cpus(), curcpu = -1, nset = 0;
1676                 for (i = 0; i < ncpus; i++)
1677                         if (cpu_isset(i, current->cpus_allowed)) {
1678                                 ipath_cdbg(PROC, "%s[%u] affinity set for "
1679                                            "cpu %d/%d\n", current->comm,
1680                                            current->pid, i, ncpus);
1681                                 curcpu = i;
1682                                 nset++;
1683                         }
1684                 if (curcpu != -1 && nset != ncpus) {
1685                         if (npresent) {
1686                                 prefunit = curcpu / (ncpus / npresent);
1687                                 ipath_cdbg(PROC,"%s[%u] %d chips, %d cpus, "
1688                                           "%d cpus/chip, select unit %d\n",
1689                                           current->comm, current->pid,
1690                                           npresent, ncpus, ncpus / npresent,
1691                                           prefunit);
1692                         }
1693                 }
1694         }
1695
1696         /*
1697          * user ports start at 1, kernel port is 0
1698          * For now, we do round-robin access across all chips
1699          */
1700
1701         if (prefunit != -1)
1702                 devmax = prefunit + 1;
1703 recheck:
1704         for (i = 1; i < maxofallports; i++) {
1705                 for (ndev = prefunit != -1 ? prefunit : 0; ndev < devmax;
1706                      ndev++) {
1707                         struct ipath_devdata *dd = ipath_lookup(ndev);
1708
1709                         if (!usable(dd))
1710                                 continue; /* can't use this unit */
1711                         if (i >= dd->ipath_cfgports)
1712                                 /*
1713                                  * Maxed out on users of this unit. Try
1714                                  * next.
1715                                  */
1716                                 continue;
1717                         ret = try_alloc_port(dd, i, fp, uinfo);
1718                         if (!ret)
1719                                 goto done;
1720                 }
1721         }
1722
1723         if (npresent) {
1724                 if (nup == 0) {
1725                         ret = -ENETDOWN;
1726                         ipath_dbg("No ports available (none initialized "
1727                                   "and ready)\n");
1728                 } else {
1729                         if (prefunit > 0) {
1730                                 /* if started above 0, retry from 0 */
1731                                 ipath_cdbg(PROC,
1732                                            "%s[%u] no ports on prefunit "
1733                                            "%d, clear and re-check\n",
1734                                            current->comm, current->pid,
1735                                            prefunit);
1736                                 devmax = ipath_count_units(NULL, NULL,
1737                                                            NULL);
1738                                 prefunit = -1;
1739                                 goto recheck;
1740                         }
1741                         ret = -EBUSY;
1742                         ipath_dbg("No ports available\n");
1743                 }
1744         } else {
1745                 ret = -ENXIO;
1746                 ipath_dbg("No boards found\n");
1747         }
1748
1749 done:
1750         return ret;
1751 }
1752
1753 static int find_shared_port(struct file *fp,
1754                             const struct ipath_user_info *uinfo)
1755 {
1756         int devmax, ndev, i;
1757         int ret = 0;
1758
1759         devmax = ipath_count_units(NULL, NULL, NULL);
1760
1761         for (ndev = 0; ndev < devmax; ndev++) {
1762                 struct ipath_devdata *dd = ipath_lookup(ndev);
1763
1764                 if (!dd)
1765                         continue;
1766                 for (i = 1; i < dd->ipath_cfgports; i++) {
1767                         struct ipath_portdata *pd = dd->ipath_pd[i];
1768
1769                         /* Skip ports which are not yet open */
1770                         if (!pd || !pd->port_cnt)
1771                                 continue;
1772                         /* Skip port if it doesn't match the requested one */
1773                         if (pd->port_subport_id != uinfo->spu_subport_id)
1774                                 continue;
1775                         /* Verify the sharing process matches the master */
1776                         if (pd->port_subport_cnt != uinfo->spu_subport_cnt ||
1777                             pd->userversion != uinfo->spu_userversion ||
1778                             pd->port_cnt >= pd->port_subport_cnt) {
1779                                 ret = -EINVAL;
1780                                 goto done;
1781                         }
1782                         port_fp(fp) = pd;
1783                         subport_fp(fp) = pd->port_cnt++;
1784                         tidcursor_fp(fp) = 0;
1785                         pd->active_slaves |= 1 << subport_fp(fp);
1786                         ipath_cdbg(PROC,
1787                                    "%s[%u] %u sharing %s[%u] unit:port %u:%u\n",
1788                                    current->comm, current->pid,
1789                                    subport_fp(fp),
1790                                    pd->port_comm, pd->port_pid,
1791                                    dd->ipath_unit, pd->port_port);
1792                         ret = 1;
1793                         goto done;
1794                 }
1795         }
1796
1797 done:
1798         return ret;
1799 }
1800
1801 static int ipath_open(struct inode *in, struct file *fp)
1802 {
1803         /* The real work is performed later in ipath_assign_port() */
1804         fp->private_data = kzalloc(sizeof(struct ipath_filedata), GFP_KERNEL);
1805         return fp->private_data ? 0 : -ENOMEM;
1806 }
1807
1808 /* Get port early, so can set affinity prior to memory allocation */
1809 static int ipath_assign_port(struct file *fp,
1810                               const struct ipath_user_info *uinfo)
1811 {
1812         int ret;
1813         int i_minor;
1814         unsigned swmajor, swminor;
1815
1816         /* Check to be sure we haven't already initialized this file */
1817         if (port_fp(fp)) {
1818                 ret = -EINVAL;
1819                 goto done;
1820         }
1821
1822         /* for now, if major version is different, bail */
1823         swmajor = uinfo->spu_userversion >> 16;
1824         if (swmajor != IPATH_USER_SWMAJOR) {
1825                 ipath_dbg("User major version %d not same as driver "
1826                           "major %d\n", uinfo->spu_userversion >> 16,
1827                           IPATH_USER_SWMAJOR);
1828                 ret = -ENODEV;
1829                 goto done;
1830         }
1831
1832         swminor = uinfo->spu_userversion & 0xffff;
1833         if (swminor != IPATH_USER_SWMINOR)
1834                 ipath_dbg("User minor version %d not same as driver "
1835                           "minor %d\n", swminor, IPATH_USER_SWMINOR);
1836
1837         mutex_lock(&ipath_mutex);
1838
1839         if (ipath_compatible_subports(swmajor, swminor) &&
1840             uinfo->spu_subport_cnt &&
1841             (ret = find_shared_port(fp, uinfo))) {
1842                 mutex_unlock(&ipath_mutex);
1843                 if (ret > 0)
1844                         ret = 0;
1845                 goto done;
1846         }
1847
1848         i_minor = iminor(fp->f_path.dentry->d_inode) - IPATH_USER_MINOR_BASE;
1849         ipath_cdbg(VERBOSE, "open on dev %lx (minor %d)\n",
1850                    (long)fp->f_path.dentry->d_inode->i_rdev, i_minor);
1851
1852         if (i_minor)
1853                 ret = find_free_port(i_minor - 1, fp, uinfo);
1854         else
1855                 ret = find_best_unit(fp, uinfo);
1856
1857         mutex_unlock(&ipath_mutex);
1858
1859 done:
1860         return ret;
1861 }
1862
1863
1864 static int ipath_do_user_init(struct file *fp,
1865                               const struct ipath_user_info *uinfo)
1866 {
1867         int ret;
1868         struct ipath_portdata *pd = port_fp(fp);
1869         struct ipath_devdata *dd;
1870         u32 head32;
1871
1872         /* Subports don't need to initialize anything since master did it. */
1873         if (subport_fp(fp)) {
1874                 ret = wait_event_interruptible(pd->port_wait,
1875                         !test_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag));
1876                 goto done;
1877         }
1878
1879         dd = pd->port_dd;
1880
1881         if (uinfo->spu_rcvhdrsize) {
1882                 ret = ipath_setrcvhdrsize(dd, uinfo->spu_rcvhdrsize);
1883                 if (ret)
1884                         goto done;
1885         }
1886
1887         /* for now we do nothing with rcvhdrcnt: uinfo->spu_rcvhdrcnt */
1888
1889         /* for right now, kernel piobufs are at end, so port 1 is at 0 */
1890         pd->port_piobufs = dd->ipath_piobufbase +
1891                 dd->ipath_pbufsport * (pd->port_port - 1) * dd->ipath_palign;
1892         ipath_cdbg(VERBOSE, "Set base of piobufs for port %u to 0x%x\n",
1893                    pd->port_port, pd->port_piobufs);
1894
1895         /*
1896          * Now allocate the rcvhdr Q and eager TIDs; skip the TID
1897          * array for time being.  If pd->port_port > chip-supported,
1898          * we need to do extra stuff here to handle by handling overflow
1899          * through port 0, someday
1900          */
1901         ret = ipath_create_rcvhdrq(dd, pd);
1902         if (!ret)
1903                 ret = ipath_create_user_egr(pd);
1904         if (ret)
1905                 goto done;
1906
1907         /*
1908          * set the eager head register for this port to the current values
1909          * of the tail pointers, since we don't know if they were
1910          * updated on last use of the port.
1911          */
1912         head32 = ipath_read_ureg32(dd, ur_rcvegrindextail, pd->port_port);
1913         ipath_write_ureg(dd, ur_rcvegrindexhead, head32, pd->port_port);
1914         dd->ipath_lastegrheads[pd->port_port] = -1;
1915         dd->ipath_lastrcvhdrqtails[pd->port_port] = -1;
1916         ipath_cdbg(VERBOSE, "Wrote port%d egrhead %x from tail regs\n",
1917                 pd->port_port, head32);
1918         pd->port_tidcursor = 0; /* start at beginning after open */
1919
1920         /* initialize poll variables... */
1921         pd->port_urgent = 0;
1922         pd->port_urgent_poll = 0;
1923         pd->port_hdrqfull_poll = pd->port_hdrqfull;
1924
1925         /*
1926          * now enable the port; the tail registers will be written to memory
1927          * by the chip as soon as it sees the write to
1928          * dd->ipath_kregs->kr_rcvctrl.  The update only happens on
1929          * transition from 0 to 1, so clear it first, then set it as part of
1930          * enabling the port.  This will (very briefly) affect any other
1931          * open ports, but it shouldn't be long enough to be an issue.
1932          * We explictly set the in-memory copy to 0 beforehand, so we don't
1933          * have to wait to be sure the DMA update has happened.
1934          */
1935         *(volatile u64 *)pd->port_rcvhdrtail_kvaddr = 0ULL;
1936         set_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
1937                 &dd->ipath_rcvctrl);
1938         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1939                          dd->ipath_rcvctrl & ~INFINIPATH_R_TAILUPD);
1940         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1941                          dd->ipath_rcvctrl);
1942         /* Notify any waiting slaves */
1943         if (pd->port_subport_cnt) {
1944                 clear_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
1945                 wake_up(&pd->port_wait);
1946         }
1947 done:
1948         return ret;
1949 }
1950
1951 /**
1952  * unlock_exptid - unlock any expected TID entries port still had in use
1953  * @pd: port
1954  *
1955  * We don't actually update the chip here, because we do a bulk update
1956  * below, using ipath_f_clear_tids.
1957  */
1958 static void unlock_expected_tids(struct ipath_portdata *pd)
1959 {
1960         struct ipath_devdata *dd = pd->port_dd;
1961         int port_tidbase = pd->port_port * dd->ipath_rcvtidcnt;
1962         int i, cnt = 0, maxtid = port_tidbase + dd->ipath_rcvtidcnt;
1963
1964         ipath_cdbg(VERBOSE, "Port %u unlocking any locked expTID pages\n",
1965                    pd->port_port);
1966         for (i = port_tidbase; i < maxtid; i++) {
1967                 if (!dd->ipath_pageshadow[i])
1968                         continue;
1969
1970                 pci_unmap_page(dd->pcidev, dd->ipath_physshadow[i],
1971                         PAGE_SIZE, PCI_DMA_FROMDEVICE);
1972                 ipath_release_user_pages_on_close(&dd->ipath_pageshadow[i],
1973                                                   1);
1974                 dd->ipath_pageshadow[i] = NULL;
1975                 cnt++;
1976                 ipath_stats.sps_pageunlocks++;
1977         }
1978         if (cnt)
1979                 ipath_cdbg(VERBOSE, "Port %u locked %u expTID entries\n",
1980                            pd->port_port, cnt);
1981
1982         if (ipath_stats.sps_pagelocks || ipath_stats.sps_pageunlocks)
1983                 ipath_cdbg(VERBOSE, "%llu pages locked, %llu unlocked\n",
1984                            (unsigned long long) ipath_stats.sps_pagelocks,
1985                            (unsigned long long)
1986                            ipath_stats.sps_pageunlocks);
1987 }
1988
1989 static int ipath_close(struct inode *in, struct file *fp)
1990 {
1991         int ret = 0;
1992         struct ipath_filedata *fd;
1993         struct ipath_portdata *pd;
1994         struct ipath_devdata *dd;
1995         unsigned port;
1996
1997         ipath_cdbg(VERBOSE, "close on dev %lx, private data %p\n",
1998                    (long)in->i_rdev, fp->private_data);
1999
2000         mutex_lock(&ipath_mutex);
2001
2002         fd = (struct ipath_filedata *) fp->private_data;
2003         fp->private_data = NULL;
2004         pd = fd->pd;
2005         if (!pd) {
2006                 mutex_unlock(&ipath_mutex);
2007                 goto bail;
2008         }
2009         if (--pd->port_cnt) {
2010                 /*
2011                  * XXX If the master closes the port before the slave(s),
2012                  * revoke the mmap for the eager receive queue so
2013                  * the slave(s) don't wait for receive data forever.
2014                  */
2015                 pd->active_slaves &= ~(1 << fd->subport);
2016                 mutex_unlock(&ipath_mutex);
2017                 goto bail;
2018         }
2019         port = pd->port_port;
2020         dd = pd->port_dd;
2021
2022         if (pd->port_hdrqfull) {
2023                 ipath_cdbg(PROC, "%s[%u] had %u rcvhdrqfull errors "
2024                            "during run\n", pd->port_comm, pd->port_pid,
2025                            pd->port_hdrqfull);
2026                 pd->port_hdrqfull = 0;
2027         }
2028
2029         if (pd->port_rcvwait_to || pd->port_piowait_to
2030             || pd->port_rcvnowait || pd->port_pionowait) {
2031                 ipath_cdbg(VERBOSE, "port%u, %u rcv, %u pio wait timeo; "
2032                            "%u rcv %u, pio already\n",
2033                            pd->port_port, pd->port_rcvwait_to,
2034                            pd->port_piowait_to, pd->port_rcvnowait,
2035                            pd->port_pionowait);
2036                 pd->port_rcvwait_to = pd->port_piowait_to =
2037                         pd->port_rcvnowait = pd->port_pionowait = 0;
2038         }
2039         if (pd->port_flag) {
2040                 ipath_dbg("port %u port_flag still set to 0x%lx\n",
2041                           pd->port_port, pd->port_flag);
2042                 pd->port_flag = 0;
2043         }
2044
2045         if (dd->ipath_kregbase) {
2046                 int i;
2047                 /* atomically clear receive enable port and intr avail. */
2048                 clear_bit(INFINIPATH_R_PORTENABLE_SHIFT + port,
2049                           &dd->ipath_rcvctrl);
2050                 clear_bit(pd->port_port + INFINIPATH_R_INTRAVAIL_SHIFT,
2051                           &dd->ipath_rcvctrl);
2052                 ipath_write_kreg( dd, dd->ipath_kregs->kr_rcvctrl,
2053                         dd->ipath_rcvctrl);
2054                 /* and read back from chip to be sure that nothing
2055                  * else is in flight when we do the rest */
2056                 (void)ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
2057
2058                 /* clean up the pkeys for this port user */
2059                 ipath_clean_part_key(pd, dd);
2060                 /*
2061                  * be paranoid, and never write 0's to these, just use an
2062                  * unused part of the port 0 tail page.  Of course,
2063                  * rcvhdraddr points to a large chunk of memory, so this
2064                  * could still trash things, but at least it won't trash
2065                  * page 0, and by disabling the port, it should stop "soon",
2066                  * even if a packet or two is in already in flight after we
2067                  * disabled the port.
2068                  */
2069                 ipath_write_kreg_port(dd,
2070                         dd->ipath_kregs->kr_rcvhdrtailaddr, port,
2071                         dd->ipath_dummy_hdrq_phys);
2072                 ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdraddr,
2073                         pd->port_port, dd->ipath_dummy_hdrq_phys);
2074
2075                 i = dd->ipath_pbufsport * (port - 1);
2076                 ipath_disarm_piobufs(dd, i, dd->ipath_pbufsport);
2077
2078                 dd->ipath_f_clear_tids(dd, pd->port_port);
2079
2080                 if (dd->ipath_pageshadow)
2081                         unlock_expected_tids(pd);
2082                 ipath_stats.sps_ports--;
2083                 ipath_cdbg(PROC, "%s[%u] closed port %u:%u\n",
2084                            pd->port_comm, pd->port_pid,
2085                            dd->ipath_unit, port);
2086         }
2087
2088         pd->port_pid = 0;
2089         dd->ipath_pd[pd->port_port] = NULL; /* before releasing mutex */
2090         mutex_unlock(&ipath_mutex);
2091         ipath_free_pddata(dd, pd); /* after releasing the mutex */
2092
2093 bail:
2094         kfree(fd);
2095         return ret;
2096 }
2097
2098 static int ipath_port_info(struct ipath_portdata *pd, u16 subport,
2099                            struct ipath_port_info __user *uinfo)
2100 {
2101         struct ipath_port_info info;
2102         int nup;
2103         int ret;
2104         size_t sz;
2105
2106         (void) ipath_count_units(NULL, &nup, NULL);
2107         info.num_active = nup;
2108         info.unit = pd->port_dd->ipath_unit;
2109         info.port = pd->port_port;
2110         info.subport = subport;
2111         /* Don't return new fields if old library opened the port. */
2112         if (ipath_supports_subports(pd->userversion >> 16,
2113                                     pd->userversion & 0xffff)) {
2114                 /* Number of user ports available for this device. */
2115                 info.num_ports = pd->port_dd->ipath_cfgports - 1;
2116                 info.num_subports = pd->port_subport_cnt;
2117                 sz = sizeof(info);
2118         } else
2119                 sz = sizeof(info) - 2 * sizeof(u16);
2120
2121         if (copy_to_user(uinfo, &info, sz)) {
2122                 ret = -EFAULT;
2123                 goto bail;
2124         }
2125         ret = 0;
2126
2127 bail:
2128         return ret;
2129 }
2130
2131 static int ipath_get_slave_info(struct ipath_portdata *pd,
2132                                 void __user *slave_mask_addr)
2133 {
2134         int ret = 0;
2135
2136         if (copy_to_user(slave_mask_addr, &pd->active_slaves, sizeof(u32)))
2137                 ret = -EFAULT;
2138         return ret;
2139 }
2140
2141 static int ipath_force_pio_avail_update(struct ipath_devdata *dd)
2142 {
2143         unsigned long flags;
2144
2145         spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
2146         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
2147                 dd->ipath_sendctrl & ~INFINIPATH_S_PIOBUFAVAILUPD);
2148         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
2149         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, dd->ipath_sendctrl);
2150         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
2151         spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
2152
2153         return 0;
2154 }
2155
2156 static ssize_t ipath_write(struct file *fp, const char __user *data,
2157                            size_t count, loff_t *off)
2158 {
2159         const struct ipath_cmd __user *ucmd;
2160         struct ipath_portdata *pd;
2161         const void __user *src;
2162         size_t consumed, copy;
2163         struct ipath_cmd cmd;
2164         ssize_t ret = 0;
2165         void *dest;
2166
2167         if (count < sizeof(cmd.type)) {
2168                 ret = -EINVAL;
2169                 goto bail;
2170         }
2171
2172         ucmd = (const struct ipath_cmd __user *) data;
2173
2174         if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) {
2175                 ret = -EFAULT;
2176                 goto bail;
2177         }
2178
2179         consumed = sizeof(cmd.type);
2180
2181         switch (cmd.type) {
2182         case IPATH_CMD_ASSIGN_PORT:
2183         case __IPATH_CMD_USER_INIT:
2184         case IPATH_CMD_USER_INIT:
2185                 copy = sizeof(cmd.cmd.user_info);
2186                 dest = &cmd.cmd.user_info;
2187                 src = &ucmd->cmd.user_info;
2188                 break;
2189         case IPATH_CMD_RECV_CTRL:
2190                 copy = sizeof(cmd.cmd.recv_ctrl);
2191                 dest = &cmd.cmd.recv_ctrl;
2192                 src = &ucmd->cmd.recv_ctrl;
2193                 break;
2194         case IPATH_CMD_PORT_INFO:
2195                 copy = sizeof(cmd.cmd.port_info);
2196                 dest = &cmd.cmd.port_info;
2197                 src = &ucmd->cmd.port_info;
2198                 break;
2199         case IPATH_CMD_TID_UPDATE:
2200         case IPATH_CMD_TID_FREE:
2201                 copy = sizeof(cmd.cmd.tid_info);
2202                 dest = &cmd.cmd.tid_info;
2203                 src = &ucmd->cmd.tid_info;
2204                 break;
2205         case IPATH_CMD_SET_PART_KEY:
2206                 copy = sizeof(cmd.cmd.part_key);
2207                 dest = &cmd.cmd.part_key;
2208                 src = &ucmd->cmd.part_key;
2209                 break;
2210         case __IPATH_CMD_SLAVE_INFO:
2211                 copy = sizeof(cmd.cmd.slave_mask_addr);
2212                 dest = &cmd.cmd.slave_mask_addr;
2213                 src = &ucmd->cmd.slave_mask_addr;
2214                 break;
2215         case IPATH_CMD_PIOAVAILUPD:     // force an update of PIOAvail reg
2216                 copy = 0;
2217                 src = NULL;
2218                 dest = NULL;
2219                 break;
2220         case IPATH_CMD_POLL_TYPE:
2221                 copy = sizeof(cmd.cmd.poll_type);
2222                 dest = &cmd.cmd.poll_type;
2223                 src = &ucmd->cmd.poll_type;
2224                 break;
2225         default:
2226                 ret = -EINVAL;
2227                 goto bail;
2228         }
2229
2230         if (copy) {
2231                 if ((count - consumed) < copy) {
2232                         ret = -EINVAL;
2233                         goto bail;
2234                 }
2235
2236                 if (copy_from_user(dest, src, copy)) {
2237                         ret = -EFAULT;
2238                         goto bail;
2239                 }
2240
2241                 consumed += copy;
2242         }
2243
2244         pd = port_fp(fp);
2245         if (!pd && cmd.type != __IPATH_CMD_USER_INIT &&
2246                 cmd.type != IPATH_CMD_ASSIGN_PORT) {
2247                 ret = -EINVAL;
2248                 goto bail;
2249         }
2250
2251         switch (cmd.type) {
2252         case IPATH_CMD_ASSIGN_PORT:
2253                 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2254                 if (ret)
2255                         goto bail;
2256                 break;
2257         case __IPATH_CMD_USER_INIT:
2258                 /* backwards compatibility, get port first */
2259                 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2260                 if (ret)
2261                         goto bail;
2262                 /* and fall through to current version. */
2263         case IPATH_CMD_USER_INIT:
2264                 ret = ipath_do_user_init(fp, &cmd.cmd.user_info);
2265                 if (ret)
2266                         goto bail;
2267                 ret = ipath_get_base_info(
2268                         fp, (void __user *) (unsigned long)
2269                         cmd.cmd.user_info.spu_base_info,
2270                         cmd.cmd.user_info.spu_base_info_size);
2271                 break;
2272         case IPATH_CMD_RECV_CTRL:
2273                 ret = ipath_manage_rcvq(pd, subport_fp(fp), cmd.cmd.recv_ctrl);
2274                 break;
2275         case IPATH_CMD_PORT_INFO:
2276                 ret = ipath_port_info(pd, subport_fp(fp),
2277                                       (struct ipath_port_info __user *)
2278                                       (unsigned long) cmd.cmd.port_info);
2279                 break;
2280         case IPATH_CMD_TID_UPDATE:
2281                 ret = ipath_tid_update(pd, fp, &cmd.cmd.tid_info);
2282                 break;
2283         case IPATH_CMD_TID_FREE:
2284                 ret = ipath_tid_free(pd, subport_fp(fp), &cmd.cmd.tid_info);
2285                 break;
2286         case IPATH_CMD_SET_PART_KEY:
2287                 ret = ipath_set_part_key(pd, cmd.cmd.part_key);
2288                 break;
2289         case __IPATH_CMD_SLAVE_INFO:
2290                 ret = ipath_get_slave_info(pd,
2291                                            (void __user *) (unsigned long)
2292                                            cmd.cmd.slave_mask_addr);
2293                 break;
2294         case IPATH_CMD_PIOAVAILUPD:
2295                 ret = ipath_force_pio_avail_update(pd->port_dd);
2296                 break;
2297         case IPATH_CMD_POLL_TYPE:
2298                 pd->poll_type = cmd.cmd.poll_type;
2299                 break;
2300         }
2301
2302         if (ret >= 0)
2303                 ret = consumed;
2304
2305 bail:
2306         return ret;
2307 }
2308
2309 static struct class *ipath_class;
2310
2311 static int init_cdev(int minor, char *name, const struct file_operations *fops,
2312                      struct cdev **cdevp, struct class_device **class_devp)
2313 {
2314         const dev_t dev = MKDEV(IPATH_MAJOR, minor);
2315         struct cdev *cdev = NULL;
2316         struct class_device *class_dev = NULL;
2317         int ret;
2318
2319         cdev = cdev_alloc();
2320         if (!cdev) {
2321                 printk(KERN_ERR IPATH_DRV_NAME
2322                        ": Could not allocate cdev for minor %d, %s\n",
2323                        minor, name);
2324                 ret = -ENOMEM;
2325                 goto done;
2326         }
2327
2328         cdev->owner = THIS_MODULE;
2329         cdev->ops = fops;
2330         kobject_set_name(&cdev->kobj, name);
2331
2332         ret = cdev_add(cdev, dev, 1);
2333         if (ret < 0) {
2334                 printk(KERN_ERR IPATH_DRV_NAME
2335                        ": Could not add cdev for minor %d, %s (err %d)\n",
2336                        minor, name, -ret);
2337                 goto err_cdev;
2338         }
2339
2340         class_dev = class_device_create(ipath_class, NULL, dev, NULL, name);
2341
2342         if (IS_ERR(class_dev)) {
2343                 ret = PTR_ERR(class_dev);
2344                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2345                        "class_dev for minor %d, %s (err %d)\n",
2346                        minor, name, -ret);
2347                 goto err_cdev;
2348         }
2349
2350         goto done;
2351
2352 err_cdev:
2353         cdev_del(cdev);
2354         cdev = NULL;
2355
2356 done:
2357         if (ret >= 0) {
2358                 *cdevp = cdev;
2359                 *class_devp = class_dev;
2360         } else {
2361                 *cdevp = NULL;
2362                 *class_devp = NULL;
2363         }
2364
2365         return ret;
2366 }
2367
2368 int ipath_cdev_init(int minor, char *name, const struct file_operations *fops,
2369                     struct cdev **cdevp, struct class_device **class_devp)
2370 {
2371         return init_cdev(minor, name, fops, cdevp, class_devp);
2372 }
2373
2374 static void cleanup_cdev(struct cdev **cdevp,
2375                          struct class_device **class_devp)
2376 {
2377         struct class_device *class_dev = *class_devp;
2378
2379         if (class_dev) {
2380                 class_device_unregister(class_dev);
2381                 *class_devp = NULL;
2382         }
2383
2384         if (*cdevp) {
2385                 cdev_del(*cdevp);
2386                 *cdevp = NULL;
2387         }
2388 }
2389
2390 void ipath_cdev_cleanup(struct cdev **cdevp,
2391                         struct class_device **class_devp)
2392 {
2393         cleanup_cdev(cdevp, class_devp);
2394 }
2395
2396 static struct cdev *wildcard_cdev;
2397 static struct class_device *wildcard_class_dev;
2398
2399 static const dev_t dev = MKDEV(IPATH_MAJOR, 0);
2400
2401 static int user_init(void)
2402 {
2403         int ret;
2404
2405         ret = register_chrdev_region(dev, IPATH_NMINORS, IPATH_DRV_NAME);
2406         if (ret < 0) {
2407                 printk(KERN_ERR IPATH_DRV_NAME ": Could not register "
2408                        "chrdev region (err %d)\n", -ret);
2409                 goto done;
2410         }
2411
2412         ipath_class = class_create(THIS_MODULE, IPATH_DRV_NAME);
2413
2414         if (IS_ERR(ipath_class)) {
2415                 ret = PTR_ERR(ipath_class);
2416                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2417                        "device class (err %d)\n", -ret);
2418                 goto bail;
2419         }
2420
2421         goto done;
2422 bail:
2423         unregister_chrdev_region(dev, IPATH_NMINORS);
2424 done:
2425         return ret;
2426 }
2427
2428 static void user_cleanup(void)
2429 {
2430         if (ipath_class) {
2431                 class_destroy(ipath_class);
2432                 ipath_class = NULL;
2433         }
2434
2435         unregister_chrdev_region(dev, IPATH_NMINORS);
2436 }
2437
2438 static atomic_t user_count = ATOMIC_INIT(0);
2439 static atomic_t user_setup = ATOMIC_INIT(0);
2440
2441 int ipath_user_add(struct ipath_devdata *dd)
2442 {
2443         char name[10];
2444         int ret;
2445
2446         if (atomic_inc_return(&user_count) == 1) {
2447                 ret = user_init();
2448                 if (ret < 0) {
2449                         ipath_dev_err(dd, "Unable to set up user support: "
2450                                       "error %d\n", -ret);
2451                         goto bail;
2452                 }
2453                 ret = init_cdev(0, "ipath", &ipath_file_ops, &wildcard_cdev,
2454                                 &wildcard_class_dev);
2455                 if (ret < 0) {
2456                         ipath_dev_err(dd, "Could not create wildcard "
2457                                       "minor: error %d\n", -ret);
2458                         goto bail_user;
2459                 }
2460
2461                 atomic_set(&user_setup, 1);
2462         }
2463
2464         snprintf(name, sizeof(name), "ipath%d", dd->ipath_unit);
2465
2466         ret = init_cdev(dd->ipath_unit + 1, name, &ipath_file_ops,
2467                         &dd->user_cdev, &dd->user_class_dev);
2468         if (ret < 0)
2469                 ipath_dev_err(dd, "Could not create user minor %d, %s\n",
2470                               dd->ipath_unit + 1, name);
2471
2472         goto bail;
2473
2474 bail_user:
2475         user_cleanup();
2476 bail:
2477         return ret;
2478 }
2479
2480 void ipath_user_remove(struct ipath_devdata *dd)
2481 {
2482         cleanup_cdev(&dd->user_cdev, &dd->user_class_dev);
2483
2484         if (atomic_dec_return(&user_count) == 0) {
2485                 if (atomic_read(&user_setup) == 0)
2486                         goto bail;
2487
2488                 cleanup_cdev(&wildcard_cdev, &wildcard_class_dev);
2489                 user_cleanup();
2490
2491                 atomic_set(&user_setup, 0);
2492         }
2493 bail:
2494         return;
2495 }