Merge master.kernel.org:/pub/scm/linux/kernel/git/lethal/sh-2.6
[linux-2.6] / 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                         ipath_cdbg(VERBOSE, "PID %u freeing TID %u\n",
542                                    pd->port_pid, tid);
543                         dd->ipath_f_put_tid(dd, &tidbase[tid],
544                                             RCVHQ_RCV_TYPE_EXPECTED,
545                                             dd->ipath_tidinvalid);
546                         pci_unmap_page(dd->pcidev,
547                                 dd->ipath_physshadow[porttid + tid],
548                                 PAGE_SIZE, PCI_DMA_FROMDEVICE);
549                         ipath_release_user_pages(
550                                 &dd->ipath_pageshadow[porttid + tid], 1);
551                         dd->ipath_pageshadow[porttid + tid] = NULL;
552                         ipath_stats.sps_pageunlocks++;
553                 } else
554                         ipath_dbg("Unused tid %u, ignoring\n", tid);
555         }
556         if (cnt != ti->tidcnt)
557                 ipath_dbg("passed in tidcnt %d, only %d bits set in map\n",
558                           ti->tidcnt, cnt);
559 done:
560         if (ret)
561                 ipath_dbg("Failed to unmap %u TID pages, failing with %d\n",
562                           ti->tidcnt, -ret);
563         return ret;
564 }
565
566 /**
567  * ipath_set_part_key - set a partition key
568  * @pd: the port
569  * @key: the key
570  *
571  * We can have up to 4 active at a time (other than the default, which is
572  * always allowed).  This is somewhat tricky, since multiple ports may set
573  * the same key, so we reference count them, and clean up at exit.  All 4
574  * partition keys are packed into a single infinipath register.  It's an
575  * error for a process to set the same pkey multiple times.  We provide no
576  * mechanism to de-allocate a pkey at this time, we may eventually need to
577  * do that.  I've used the atomic operations, and no locking, and only make
578  * a single pass through what's available.  This should be more than
579  * adequate for some time. I'll think about spinlocks or the like if and as
580  * it's necessary.
581  */
582 static int ipath_set_part_key(struct ipath_portdata *pd, u16 key)
583 {
584         struct ipath_devdata *dd = pd->port_dd;
585         int i, any = 0, pidx = -1;
586         u16 lkey = key & 0x7FFF;
587         int ret;
588
589         if (lkey == (IPATH_DEFAULT_P_KEY & 0x7FFF)) {
590                 /* nothing to do; this key always valid */
591                 ret = 0;
592                 goto bail;
593         }
594
595         ipath_cdbg(VERBOSE, "p%u try to set pkey %hx, current keys "
596                    "%hx:%x %hx:%x %hx:%x %hx:%x\n",
597                    pd->port_port, key, dd->ipath_pkeys[0],
598                    atomic_read(&dd->ipath_pkeyrefs[0]), dd->ipath_pkeys[1],
599                    atomic_read(&dd->ipath_pkeyrefs[1]), dd->ipath_pkeys[2],
600                    atomic_read(&dd->ipath_pkeyrefs[2]), dd->ipath_pkeys[3],
601                    atomic_read(&dd->ipath_pkeyrefs[3]));
602
603         if (!lkey) {
604                 ipath_cdbg(PROC, "p%u tries to set key 0, not allowed\n",
605                            pd->port_port);
606                 ret = -EINVAL;
607                 goto bail;
608         }
609
610         /*
611          * Set the full membership bit, because it has to be
612          * set in the register or the packet, and it seems
613          * cleaner to set in the register than to force all
614          * callers to set it. (see bug 4331)
615          */
616         key |= 0x8000;
617
618         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
619                 if (!pd->port_pkeys[i] && pidx == -1)
620                         pidx = i;
621                 if (pd->port_pkeys[i] == key) {
622                         ipath_cdbg(VERBOSE, "p%u tries to set same pkey "
623                                    "(%x) more than once\n",
624                                    pd->port_port, key);
625                         ret = -EEXIST;
626                         goto bail;
627                 }
628         }
629         if (pidx == -1) {
630                 ipath_dbg("All pkeys for port %u already in use, "
631                           "can't set %x\n", pd->port_port, key);
632                 ret = -EBUSY;
633                 goto bail;
634         }
635         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
636                 if (!dd->ipath_pkeys[i]) {
637                         any++;
638                         continue;
639                 }
640                 if (dd->ipath_pkeys[i] == key) {
641                         atomic_t *pkrefs = &dd->ipath_pkeyrefs[i];
642
643                         if (atomic_inc_return(pkrefs) > 1) {
644                                 pd->port_pkeys[pidx] = key;
645                                 ipath_cdbg(VERBOSE, "p%u set key %x "
646                                            "matches #%d, count now %d\n",
647                                            pd->port_port, key, i,
648                                            atomic_read(pkrefs));
649                                 ret = 0;
650                                 goto bail;
651                         } else {
652                                 /*
653                                  * lost race, decrement count, catch below
654                                  */
655                                 atomic_dec(pkrefs);
656                                 ipath_cdbg(VERBOSE, "Lost race, count was "
657                                            "0, after dec, it's %d\n",
658                                            atomic_read(pkrefs));
659                                 any++;
660                         }
661                 }
662                 if ((dd->ipath_pkeys[i] & 0x7FFF) == lkey) {
663                         /*
664                          * It makes no sense to have both the limited and
665                          * full membership PKEY set at the same time since
666                          * the unlimited one will disable the limited one.
667                          */
668                         ret = -EEXIST;
669                         goto bail;
670                 }
671         }
672         if (!any) {
673                 ipath_dbg("port %u, all pkeys already in use, "
674                           "can't set %x\n", pd->port_port, key);
675                 ret = -EBUSY;
676                 goto bail;
677         }
678         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
679                 if (!dd->ipath_pkeys[i] &&
680                     atomic_inc_return(&dd->ipath_pkeyrefs[i]) == 1) {
681                         u64 pkey;
682
683                         /* for ipathstats, etc. */
684                         ipath_stats.sps_pkeys[i] = lkey;
685                         pd->port_pkeys[pidx] = dd->ipath_pkeys[i] = key;
686                         pkey =
687                                 (u64) dd->ipath_pkeys[0] |
688                                 ((u64) dd->ipath_pkeys[1] << 16) |
689                                 ((u64) dd->ipath_pkeys[2] << 32) |
690                                 ((u64) dd->ipath_pkeys[3] << 48);
691                         ipath_cdbg(PROC, "p%u set key %x in #%d, "
692                                    "portidx %d, new pkey reg %llx\n",
693                                    pd->port_port, key, i, pidx,
694                                    (unsigned long long) pkey);
695                         ipath_write_kreg(
696                                 dd, dd->ipath_kregs->kr_partitionkey, pkey);
697
698                         ret = 0;
699                         goto bail;
700                 }
701         }
702         ipath_dbg("port %u, all pkeys already in use 2nd pass, "
703                   "can't set %x\n", pd->port_port, key);
704         ret = -EBUSY;
705
706 bail:
707         return ret;
708 }
709
710 /**
711  * ipath_manage_rcvq - manage a port's receive queue
712  * @pd: the port
713  * @subport: the subport
714  * @start_stop: action to carry out
715  *
716  * start_stop == 0 disables receive on the port, for use in queue
717  * overflow conditions.  start_stop==1 re-enables, to be used to
718  * re-init the software copy of the head register
719  */
720 static int ipath_manage_rcvq(struct ipath_portdata *pd, unsigned subport,
721                              int start_stop)
722 {
723         struct ipath_devdata *dd = pd->port_dd;
724
725         ipath_cdbg(PROC, "%sabling rcv for unit %u port %u:%u\n",
726                    start_stop ? "en" : "dis", dd->ipath_unit,
727                    pd->port_port, subport);
728         if (subport)
729                 goto bail;
730         /* atomically clear receive enable port. */
731         if (start_stop) {
732                 /*
733                  * On enable, force in-memory copy of the tail register to
734                  * 0, so that protocol code doesn't have to worry about
735                  * whether or not the chip has yet updated the in-memory
736                  * copy or not on return from the system call. The chip
737                  * always resets it's tail register back to 0 on a
738                  * transition from disabled to enabled.  This could cause a
739                  * problem if software was broken, and did the enable w/o
740                  * the disable, but eventually the in-memory copy will be
741                  * updated and correct itself, even in the face of software
742                  * bugs.
743                  */
744                 *(volatile u64 *)pd->port_rcvhdrtail_kvaddr = 0;
745                 set_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
746                         &dd->ipath_rcvctrl);
747         } else
748                 clear_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
749                           &dd->ipath_rcvctrl);
750         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
751                          dd->ipath_rcvctrl);
752         /* now be sure chip saw it before we return */
753         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
754         if (start_stop) {
755                 /*
756                  * And try to be sure that tail reg update has happened too.
757                  * This should in theory interlock with the RXE changes to
758                  * the tail register.  Don't assign it to the tail register
759                  * in memory copy, since we could overwrite an update by the
760                  * chip if we did.
761                  */
762                 ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
763         }
764         /* always; new head should be equal to new tail; see above */
765 bail:
766         return 0;
767 }
768
769 static void ipath_clean_part_key(struct ipath_portdata *pd,
770                                  struct ipath_devdata *dd)
771 {
772         int i, j, pchanged = 0;
773         u64 oldpkey;
774
775         /* for debugging only */
776         oldpkey = (u64) dd->ipath_pkeys[0] |
777                 ((u64) dd->ipath_pkeys[1] << 16) |
778                 ((u64) dd->ipath_pkeys[2] << 32) |
779                 ((u64) dd->ipath_pkeys[3] << 48);
780
781         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
782                 if (!pd->port_pkeys[i])
783                         continue;
784                 ipath_cdbg(VERBOSE, "look for key[%d] %hx in pkeys\n", i,
785                            pd->port_pkeys[i]);
786                 for (j = 0; j < ARRAY_SIZE(dd->ipath_pkeys); j++) {
787                         /* check for match independent of the global bit */
788                         if ((dd->ipath_pkeys[j] & 0x7fff) !=
789                             (pd->port_pkeys[i] & 0x7fff))
790                                 continue;
791                         if (atomic_dec_and_test(&dd->ipath_pkeyrefs[j])) {
792                                 ipath_cdbg(VERBOSE, "p%u clear key "
793                                            "%x matches #%d\n",
794                                            pd->port_port,
795                                            pd->port_pkeys[i], j);
796                                 ipath_stats.sps_pkeys[j] =
797                                         dd->ipath_pkeys[j] = 0;
798                                 pchanged++;
799                         }
800                         else ipath_cdbg(
801                                 VERBOSE, "p%u key %x matches #%d, "
802                                 "but ref still %d\n", pd->port_port,
803                                 pd->port_pkeys[i], j,
804                                 atomic_read(&dd->ipath_pkeyrefs[j]));
805                         break;
806                 }
807                 pd->port_pkeys[i] = 0;
808         }
809         if (pchanged) {
810                 u64 pkey = (u64) dd->ipath_pkeys[0] |
811                         ((u64) dd->ipath_pkeys[1] << 16) |
812                         ((u64) dd->ipath_pkeys[2] << 32) |
813                         ((u64) dd->ipath_pkeys[3] << 48);
814                 ipath_cdbg(VERBOSE, "p%u old pkey reg %llx, "
815                            "new pkey reg %llx\n", pd->port_port,
816                            (unsigned long long) oldpkey,
817                            (unsigned long long) pkey);
818                 ipath_write_kreg(dd, dd->ipath_kregs->kr_partitionkey,
819                                  pkey);
820         }
821 }
822
823 /*
824  * Initialize the port data with the receive buffer sizes
825  * so this can be done while the master port is locked.
826  * Otherwise, there is a race with a slave opening the port
827  * and seeing these fields uninitialized.
828  */
829 static void init_user_egr_sizes(struct ipath_portdata *pd)
830 {
831         struct ipath_devdata *dd = pd->port_dd;
832         unsigned egrperchunk, egrcnt, size;
833
834         /*
835          * to avoid wasting a lot of memory, we allocate 32KB chunks of
836          * physically contiguous memory, advance through it until used up
837          * and then allocate more.  Of course, we need memory to store those
838          * extra pointers, now.  Started out with 256KB, but under heavy
839          * memory pressure (creating large files and then copying them over
840          * NFS while doing lots of MPI jobs), we hit some allocation
841          * failures, even though we can sleep...  (2.6.10) Still get
842          * failures at 64K.  32K is the lowest we can go without wasting
843          * additional memory.
844          */
845         size = 0x8000;
846         egrperchunk = size / dd->ipath_rcvegrbufsize;
847         egrcnt = dd->ipath_rcvegrcnt;
848         pd->port_rcvegrbuf_chunks = (egrcnt + egrperchunk - 1) / egrperchunk;
849         pd->port_rcvegrbufs_perchunk = egrperchunk;
850         pd->port_rcvegrbuf_size = size;
851 }
852
853 /**
854  * ipath_create_user_egr - allocate eager TID buffers
855  * @pd: the port to allocate TID buffers for
856  *
857  * This routine is now quite different for user and kernel, because
858  * the kernel uses skb's, for the accelerated network performance
859  * This is the user port version
860  *
861  * Allocate the eager TID buffers and program them into infinipath
862  * They are no longer completely contiguous, we do multiple allocation
863  * calls.
864  */
865 static int ipath_create_user_egr(struct ipath_portdata *pd)
866 {
867         struct ipath_devdata *dd = pd->port_dd;
868         unsigned e, egrcnt, egrperchunk, chunk, egrsize, egroff;
869         size_t size;
870         int ret;
871         gfp_t gfp_flags;
872
873         /*
874          * GFP_USER, but without GFP_FS, so buffer cache can be
875          * coalesced (we hope); otherwise, even at order 4,
876          * heavy filesystem activity makes these fail, and we can
877          * use compound pages.
878          */
879         gfp_flags = __GFP_WAIT | __GFP_IO | __GFP_COMP;
880
881         egrcnt = dd->ipath_rcvegrcnt;
882         /* TID number offset for this port */
883         egroff = pd->port_port * egrcnt;
884         egrsize = dd->ipath_rcvegrbufsize;
885         ipath_cdbg(VERBOSE, "Allocating %d egr buffers, at egrtid "
886                    "offset %x, egrsize %u\n", egrcnt, egroff, egrsize);
887
888         chunk = pd->port_rcvegrbuf_chunks;
889         egrperchunk = pd->port_rcvegrbufs_perchunk;
890         size = pd->port_rcvegrbuf_size;
891         pd->port_rcvegrbuf = kmalloc(chunk * sizeof(pd->port_rcvegrbuf[0]),
892                                      GFP_KERNEL);
893         if (!pd->port_rcvegrbuf) {
894                 ret = -ENOMEM;
895                 goto bail;
896         }
897         pd->port_rcvegrbuf_phys =
898                 kmalloc(chunk * sizeof(pd->port_rcvegrbuf_phys[0]),
899                         GFP_KERNEL);
900         if (!pd->port_rcvegrbuf_phys) {
901                 ret = -ENOMEM;
902                 goto bail_rcvegrbuf;
903         }
904         for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
905
906                 pd->port_rcvegrbuf[e] = dma_alloc_coherent(
907                         &dd->pcidev->dev, size, &pd->port_rcvegrbuf_phys[e],
908                         gfp_flags);
909
910                 if (!pd->port_rcvegrbuf[e]) {
911                         ret = -ENOMEM;
912                         goto bail_rcvegrbuf_phys;
913                 }
914         }
915
916         pd->port_rcvegr_phys = pd->port_rcvegrbuf_phys[0];
917
918         for (e = chunk = 0; chunk < pd->port_rcvegrbuf_chunks; chunk++) {
919                 dma_addr_t pa = pd->port_rcvegrbuf_phys[chunk];
920                 unsigned i;
921
922                 for (i = 0; e < egrcnt && i < egrperchunk; e++, i++) {
923                         dd->ipath_f_put_tid(dd, e + egroff +
924                                             (u64 __iomem *)
925                                             ((char __iomem *)
926                                              dd->ipath_kregbase +
927                                              dd->ipath_rcvegrbase),
928                                             RCVHQ_RCV_TYPE_EAGER, pa);
929                         pa += egrsize;
930                 }
931                 cond_resched(); /* don't hog the cpu */
932         }
933
934         ret = 0;
935         goto bail;
936
937 bail_rcvegrbuf_phys:
938         for (e = 0; e < pd->port_rcvegrbuf_chunks &&
939                 pd->port_rcvegrbuf[e]; e++) {
940                 dma_free_coherent(&dd->pcidev->dev, size,
941                                   pd->port_rcvegrbuf[e],
942                                   pd->port_rcvegrbuf_phys[e]);
943
944         }
945         kfree(pd->port_rcvegrbuf_phys);
946         pd->port_rcvegrbuf_phys = NULL;
947 bail_rcvegrbuf:
948         kfree(pd->port_rcvegrbuf);
949         pd->port_rcvegrbuf = NULL;
950 bail:
951         return ret;
952 }
953
954
955 /* common code for the mappings on dma_alloc_coherent mem */
956 static int ipath_mmap_mem(struct vm_area_struct *vma,
957         struct ipath_portdata *pd, unsigned len, int write_ok,
958         void *kvaddr, char *what)
959 {
960         struct ipath_devdata *dd = pd->port_dd;
961         unsigned long pfn;
962         int ret;
963
964         if ((vma->vm_end - vma->vm_start) > len) {
965                 dev_info(&dd->pcidev->dev,
966                          "FAIL on %s: len %lx > %x\n", what,
967                          vma->vm_end - vma->vm_start, len);
968                 ret = -EFAULT;
969                 goto bail;
970         }
971
972         if (!write_ok) {
973                 if (vma->vm_flags & VM_WRITE) {
974                         dev_info(&dd->pcidev->dev,
975                                  "%s must be mapped readonly\n", what);
976                         ret = -EPERM;
977                         goto bail;
978                 }
979
980                 /* don't allow them to later change with mprotect */
981                 vma->vm_flags &= ~VM_MAYWRITE;
982         }
983
984         pfn = virt_to_phys(kvaddr) >> PAGE_SHIFT;
985         ret = remap_pfn_range(vma, vma->vm_start, pfn,
986                               len, vma->vm_page_prot);
987         if (ret)
988                 dev_info(&dd->pcidev->dev, "%s port%u mmap of %lx, %x "
989                          "bytes r%c failed: %d\n", what, pd->port_port,
990                          pfn, len, write_ok?'w':'o', ret);
991         else
992                 ipath_cdbg(VERBOSE, "%s port%u mmaped %lx, %x bytes "
993                            "r%c\n", what, pd->port_port, pfn, len,
994                            write_ok?'w':'o');
995 bail:
996         return ret;
997 }
998
999 static int mmap_ureg(struct vm_area_struct *vma, struct ipath_devdata *dd,
1000                      u64 ureg)
1001 {
1002         unsigned long phys;
1003         int ret;
1004
1005         /*
1006          * This is real hardware, so use io_remap.  This is the mechanism
1007          * for the user process to update the head registers for their port
1008          * in the chip.
1009          */
1010         if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) {
1011                 dev_info(&dd->pcidev->dev, "FAIL mmap userreg: reqlen "
1012                          "%lx > PAGE\n", vma->vm_end - vma->vm_start);
1013                 ret = -EFAULT;
1014         } else {
1015                 phys = dd->ipath_physaddr + ureg;
1016                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1017
1018                 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1019                 ret = io_remap_pfn_range(vma, vma->vm_start,
1020                                          phys >> PAGE_SHIFT,
1021                                          vma->vm_end - vma->vm_start,
1022                                          vma->vm_page_prot);
1023         }
1024         return ret;
1025 }
1026
1027 static int mmap_piobufs(struct vm_area_struct *vma,
1028                         struct ipath_devdata *dd,
1029                         struct ipath_portdata *pd,
1030                         unsigned piobufs, unsigned piocnt)
1031 {
1032         unsigned long phys;
1033         int ret;
1034
1035         /*
1036          * When we map the PIO buffers in the chip, we want to map them as
1037          * writeonly, no read possible.   This prevents access to previous
1038          * process data, and catches users who might try to read the i/o
1039          * space due to a bug.
1040          */
1041         if ((vma->vm_end - vma->vm_start) > (piocnt * dd->ipath_palign)) {
1042                 dev_info(&dd->pcidev->dev, "FAIL mmap piobufs: "
1043                          "reqlen %lx > PAGE\n",
1044                          vma->vm_end - vma->vm_start);
1045                 ret = -EINVAL;
1046                 goto bail;
1047         }
1048
1049         phys = dd->ipath_physaddr + piobufs;
1050
1051         /*
1052          * Don't mark this as non-cached, or we don't get the
1053          * write combining behavior we want on the PIO buffers!
1054          */
1055
1056 #if defined(__powerpc__)
1057         /* There isn't a generic way to specify writethrough mappings */
1058         pgprot_val(vma->vm_page_prot) |= _PAGE_NO_CACHE;
1059         pgprot_val(vma->vm_page_prot) |= _PAGE_WRITETHRU;
1060         pgprot_val(vma->vm_page_prot) &= ~_PAGE_GUARDED;
1061 #endif
1062
1063         /*
1064          * don't allow them to later change to readable with mprotect (for when
1065          * not initially mapped readable, as is normally the case)
1066          */
1067         vma->vm_flags &= ~VM_MAYREAD;
1068         vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1069
1070         ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
1071                                  vma->vm_end - vma->vm_start,
1072                                  vma->vm_page_prot);
1073 bail:
1074         return ret;
1075 }
1076
1077 static int mmap_rcvegrbufs(struct vm_area_struct *vma,
1078                            struct ipath_portdata *pd)
1079 {
1080         struct ipath_devdata *dd = pd->port_dd;
1081         unsigned long start, size;
1082         size_t total_size, i;
1083         unsigned long pfn;
1084         int ret;
1085
1086         size = pd->port_rcvegrbuf_size;
1087         total_size = pd->port_rcvegrbuf_chunks * size;
1088         if ((vma->vm_end - vma->vm_start) > total_size) {
1089                 dev_info(&dd->pcidev->dev, "FAIL on egr bufs: "
1090                          "reqlen %lx > actual %lx\n",
1091                          vma->vm_end - vma->vm_start,
1092                          (unsigned long) total_size);
1093                 ret = -EINVAL;
1094                 goto bail;
1095         }
1096
1097         if (vma->vm_flags & VM_WRITE) {
1098                 dev_info(&dd->pcidev->dev, "Can't map eager buffers as "
1099                          "writable (flags=%lx)\n", vma->vm_flags);
1100                 ret = -EPERM;
1101                 goto bail;
1102         }
1103         /* don't allow them to later change to writeable with mprotect */
1104         vma->vm_flags &= ~VM_MAYWRITE;
1105
1106         start = vma->vm_start;
1107
1108         for (i = 0; i < pd->port_rcvegrbuf_chunks; i++, start += size) {
1109                 pfn = virt_to_phys(pd->port_rcvegrbuf[i]) >> PAGE_SHIFT;
1110                 ret = remap_pfn_range(vma, start, pfn, size,
1111                                       vma->vm_page_prot);
1112                 if (ret < 0)
1113                         goto bail;
1114         }
1115         ret = 0;
1116
1117 bail:
1118         return ret;
1119 }
1120
1121 /*
1122  * ipath_file_vma_nopage - handle a VMA page fault.
1123  */
1124 static struct page *ipath_file_vma_nopage(struct vm_area_struct *vma,
1125                                           unsigned long address, int *type)
1126 {
1127         unsigned long offset = address - vma->vm_start;
1128         struct page *page = NOPAGE_SIGBUS;
1129         void *pageptr;
1130
1131         /*
1132          * Convert the vmalloc address into a struct page.
1133          */
1134         pageptr = (void *)(offset + (vma->vm_pgoff << PAGE_SHIFT));
1135         page = vmalloc_to_page(pageptr);
1136         if (!page)
1137                 goto out;
1138
1139         /* Increment the reference count. */
1140         get_page(page);
1141         if (type)
1142                 *type = VM_FAULT_MINOR;
1143 out:
1144         return page;
1145 }
1146
1147 static struct vm_operations_struct ipath_file_vm_ops = {
1148         .nopage = ipath_file_vma_nopage,
1149 };
1150
1151 static int mmap_kvaddr(struct vm_area_struct *vma, u64 pgaddr,
1152                        struct ipath_portdata *pd, unsigned subport)
1153 {
1154         unsigned long len;
1155         struct ipath_devdata *dd;
1156         void *addr;
1157         size_t size;
1158         int ret = 0;
1159
1160         /* If the port is not shared, all addresses should be physical */
1161         if (!pd->port_subport_cnt)
1162                 goto bail;
1163
1164         dd = pd->port_dd;
1165         size = pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
1166
1167         /*
1168          * Each process has all the subport uregbase, rcvhdrq, and
1169          * rcvegrbufs mmapped - as an array for all the processes,
1170          * and also separately for this process.
1171          */
1172         if (pgaddr == cvt_kvaddr(pd->subport_uregbase)) {
1173                 addr = pd->subport_uregbase;
1174                 size = PAGE_SIZE * pd->port_subport_cnt;
1175         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base)) {
1176                 addr = pd->subport_rcvhdr_base;
1177                 size = pd->port_rcvhdrq_size * pd->port_subport_cnt;
1178         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf)) {
1179                 addr = pd->subport_rcvegrbuf;
1180                 size *= pd->port_subport_cnt;
1181         } else if (pgaddr == cvt_kvaddr(pd->subport_uregbase +
1182                                         PAGE_SIZE * subport)) {
1183                 addr = pd->subport_uregbase + PAGE_SIZE * subport;
1184                 size = PAGE_SIZE;
1185         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base +
1186                                 pd->port_rcvhdrq_size * subport)) {
1187                 addr = pd->subport_rcvhdr_base +
1188                         pd->port_rcvhdrq_size * subport;
1189                 size = pd->port_rcvhdrq_size;
1190         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf +
1191                                size * subport)) {
1192                 addr = pd->subport_rcvegrbuf + size * subport;
1193                 /* rcvegrbufs are read-only on the slave */
1194                 if (vma->vm_flags & VM_WRITE) {
1195                         dev_info(&dd->pcidev->dev,
1196                                  "Can't map eager buffers as "
1197                                  "writable (flags=%lx)\n", vma->vm_flags);
1198                         ret = -EPERM;
1199                         goto bail;
1200                 }
1201                 /*
1202                  * Don't allow permission to later change to writeable
1203                  * with mprotect.
1204                  */
1205                 vma->vm_flags &= ~VM_MAYWRITE;
1206         } else {
1207                 goto bail;
1208         }
1209         len = vma->vm_end - vma->vm_start;
1210         if (len > size) {
1211                 ipath_cdbg(MM, "FAIL: reqlen %lx > %zx\n", len, size);
1212                 ret = -EINVAL;
1213                 goto bail;
1214         }
1215
1216         vma->vm_pgoff = (unsigned long) addr >> PAGE_SHIFT;
1217         vma->vm_ops = &ipath_file_vm_ops;
1218         vma->vm_flags |= VM_RESERVED | VM_DONTEXPAND;
1219         ret = 1;
1220
1221 bail:
1222         return ret;
1223 }
1224
1225 /**
1226  * ipath_mmap - mmap various structures into user space
1227  * @fp: the file pointer
1228  * @vma: the VM area
1229  *
1230  * We use this to have a shared buffer between the kernel and the user code
1231  * for the rcvhdr queue, egr buffers, and the per-port user regs and pio
1232  * buffers in the chip.  We have the open and close entries so we can bump
1233  * the ref count and keep the driver from being unloaded while still mapped.
1234  */
1235 static int ipath_mmap(struct file *fp, struct vm_area_struct *vma)
1236 {
1237         struct ipath_portdata *pd;
1238         struct ipath_devdata *dd;
1239         u64 pgaddr, ureg;
1240         unsigned piobufs, piocnt;
1241         int ret;
1242
1243         pd = port_fp(fp);
1244         if (!pd) {
1245                 ret = -EINVAL;
1246                 goto bail;
1247         }
1248         dd = pd->port_dd;
1249
1250         /*
1251          * This is the ipath_do_user_init() code, mapping the shared buffers
1252          * into the user process. The address referred to by vm_pgoff is the
1253          * file offset passed via mmap().  For shared ports, this is the
1254          * kernel vmalloc() address of the pages to share with the master.
1255          * For non-shared or master ports, this is a physical address.
1256          * We only do one mmap for each space mapped.
1257          */
1258         pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1259
1260         /*
1261          * Check for 0 in case one of the allocations failed, but user
1262          * called mmap anyway.
1263          */
1264         if (!pgaddr)  {
1265                 ret = -EINVAL;
1266                 goto bail;
1267         }
1268
1269         ipath_cdbg(MM, "pgaddr %llx vm_start=%lx len %lx port %u:%u:%u\n",
1270                    (unsigned long long) pgaddr, vma->vm_start,
1271                    vma->vm_end - vma->vm_start, dd->ipath_unit,
1272                    pd->port_port, subport_fp(fp));
1273
1274         /*
1275          * Physical addresses must fit in 40 bits for our hardware.
1276          * Check for kernel virtual addresses first, anything else must
1277          * match a HW or memory address.
1278          */
1279         ret = mmap_kvaddr(vma, pgaddr, pd, subport_fp(fp));
1280         if (ret) {
1281                 if (ret > 0)
1282                         ret = 0;
1283                 goto bail;
1284         }
1285
1286         ureg = dd->ipath_uregbase + dd->ipath_palign * pd->port_port;
1287         if (!pd->port_subport_cnt) {
1288                 /* port is not shared */
1289                 piocnt = dd->ipath_pbufsport;
1290                 piobufs = pd->port_piobufs;
1291         } else if (!subport_fp(fp)) {
1292                 /* caller is the master */
1293                 piocnt = (dd->ipath_pbufsport / pd->port_subport_cnt) +
1294                          (dd->ipath_pbufsport % pd->port_subport_cnt);
1295                 piobufs = pd->port_piobufs +
1296                         dd->ipath_palign * (dd->ipath_pbufsport - piocnt);
1297         } else {
1298                 unsigned slave = subport_fp(fp) - 1;
1299
1300                 /* caller is a slave */
1301                 piocnt = dd->ipath_pbufsport / pd->port_subport_cnt;
1302                 piobufs = pd->port_piobufs + dd->ipath_palign * piocnt * slave;
1303         }
1304
1305         if (pgaddr == ureg)
1306                 ret = mmap_ureg(vma, dd, ureg);
1307         else if (pgaddr == piobufs)
1308                 ret = mmap_piobufs(vma, dd, pd, piobufs, piocnt);
1309         else if (pgaddr == dd->ipath_pioavailregs_phys)
1310                 /* in-memory copy of pioavail registers */
1311                 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1312                                      (void *) dd->ipath_pioavailregs_dma,
1313                                      "pioavail registers");
1314         else if (pgaddr == pd->port_rcvegr_phys)
1315                 ret = mmap_rcvegrbufs(vma, pd);
1316         else if (pgaddr == (u64) pd->port_rcvhdrq_phys)
1317                 /*
1318                  * The rcvhdrq itself; readonly except on HT (so have
1319                  * to allow writable mapping), multiple pages, contiguous
1320                  * from an i/o perspective.
1321                  */
1322                 ret = ipath_mmap_mem(vma, pd, pd->port_rcvhdrq_size, 1,
1323                                      pd->port_rcvhdrq,
1324                                      "rcvhdrq");
1325         else if (pgaddr == (u64) pd->port_rcvhdrqtailaddr_phys)
1326                 /* in-memory copy of rcvhdrq tail register */
1327                 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1328                                      pd->port_rcvhdrtail_kvaddr,
1329                                      "rcvhdrq tail");
1330         else
1331                 ret = -EINVAL;
1332
1333         vma->vm_private_data = NULL;
1334
1335         if (ret < 0)
1336                 dev_info(&dd->pcidev->dev,
1337                          "Failure %d on off %llx len %lx\n",
1338                          -ret, (unsigned long long)pgaddr,
1339                          vma->vm_end - vma->vm_start);
1340 bail:
1341         return ret;
1342 }
1343
1344 static unsigned int ipath_poll_urgent(struct ipath_portdata *pd,
1345                                       struct file *fp,
1346                                       struct poll_table_struct *pt)
1347 {
1348         unsigned pollflag = 0;
1349         struct ipath_devdata *dd;
1350
1351         dd = pd->port_dd;
1352
1353         if (test_bit(IPATH_PORT_WAITING_OVERFLOW, &pd->int_flag)) {
1354                 pollflag |= POLLERR;
1355                 clear_bit(IPATH_PORT_WAITING_OVERFLOW, &pd->int_flag);
1356         }
1357
1358         if (test_bit(IPATH_PORT_WAITING_URG, &pd->int_flag)) {
1359                 pollflag |= POLLIN | POLLRDNORM;
1360                 clear_bit(IPATH_PORT_WAITING_URG, &pd->int_flag);
1361         }
1362
1363         if (!pollflag) {
1364                 set_bit(IPATH_PORT_WAITING_URG, &pd->port_flag);
1365                 if (pd->poll_type & IPATH_POLL_TYPE_OVERFLOW)
1366                         set_bit(IPATH_PORT_WAITING_OVERFLOW,
1367                                 &pd->port_flag);
1368
1369                 poll_wait(fp, &pd->port_wait, pt);
1370         }
1371
1372         return pollflag;
1373 }
1374
1375 static unsigned int ipath_poll_next(struct ipath_portdata *pd,
1376                                     struct file *fp,
1377                                     struct poll_table_struct *pt)
1378 {
1379         u32 head, tail;
1380         unsigned pollflag = 0;
1381         struct ipath_devdata *dd;
1382
1383         dd = pd->port_dd;
1384
1385         head = ipath_read_ureg32(dd, ur_rcvhdrhead, pd->port_port);
1386         tail = *(volatile u64 *)pd->port_rcvhdrtail_kvaddr;
1387
1388         if (test_bit(IPATH_PORT_WAITING_OVERFLOW, &pd->int_flag)) {
1389                 pollflag |= POLLERR;
1390                 clear_bit(IPATH_PORT_WAITING_OVERFLOW, &pd->int_flag);
1391         }
1392
1393         if (tail != head ||
1394             test_bit(IPATH_PORT_WAITING_RCV, &pd->int_flag)) {
1395                 pollflag |= POLLIN | POLLRDNORM;
1396                 clear_bit(IPATH_PORT_WAITING_RCV, &pd->int_flag);
1397         }
1398
1399         if (!pollflag) {
1400                 set_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
1401                 if (pd->poll_type & IPATH_POLL_TYPE_OVERFLOW)
1402                         set_bit(IPATH_PORT_WAITING_OVERFLOW,
1403                                 &pd->port_flag);
1404
1405                 set_bit(pd->port_port + INFINIPATH_R_INTRAVAIL_SHIFT,
1406                         &dd->ipath_rcvctrl);
1407
1408                 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1409                                  dd->ipath_rcvctrl);
1410
1411                 if (dd->ipath_rhdrhead_intr_off) /* arm rcv interrupt */
1412                         ipath_write_ureg(dd, ur_rcvhdrhead,
1413                                          dd->ipath_rhdrhead_intr_off | head,
1414                                          pd->port_port);
1415
1416                 poll_wait(fp, &pd->port_wait, pt);
1417         }
1418
1419         return pollflag;
1420 }
1421
1422 static unsigned int ipath_poll(struct file *fp,
1423                                struct poll_table_struct *pt)
1424 {
1425         struct ipath_portdata *pd;
1426         unsigned pollflag;
1427
1428         pd = port_fp(fp);
1429         if (!pd)
1430                 pollflag = 0;
1431         else if (pd->poll_type & IPATH_POLL_TYPE_URGENT)
1432                 pollflag = ipath_poll_urgent(pd, fp, pt);
1433         else
1434                 pollflag = ipath_poll_next(pd, fp, pt);
1435
1436         return pollflag;
1437 }
1438
1439 static int ipath_supports_subports(int user_swmajor, int user_swminor)
1440 {
1441         /* no subport implementation prior to software version 1.3 */
1442         return (user_swmajor > 1) || (user_swminor >= 3);
1443 }
1444
1445 static int ipath_compatible_subports(int user_swmajor, int user_swminor)
1446 {
1447         /* this code is written long-hand for clarity */
1448         if (IPATH_USER_SWMAJOR != user_swmajor) {
1449                 /* no promise of compatibility if major mismatch */
1450                 return 0;
1451         }
1452         if (IPATH_USER_SWMAJOR == 1) {
1453                 switch (IPATH_USER_SWMINOR) {
1454                 case 0:
1455                 case 1:
1456                 case 2:
1457                         /* no subport implementation so cannot be compatible */
1458                         return 0;
1459                 case 3:
1460                         /* 3 is only compatible with itself */
1461                         return user_swminor == 3;
1462                 default:
1463                         /* >= 4 are compatible (or are expected to be) */
1464                         return user_swminor >= 4;
1465                 }
1466         }
1467         /* make no promises yet for future major versions */
1468         return 0;
1469 }
1470
1471 static int init_subports(struct ipath_devdata *dd,
1472                          struct ipath_portdata *pd,
1473                          const struct ipath_user_info *uinfo)
1474 {
1475         int ret = 0;
1476         unsigned num_subports;
1477         size_t size;
1478
1479         /*
1480          * If the user is requesting zero subports,
1481          * skip the subport allocation.
1482          */
1483         if (uinfo->spu_subport_cnt <= 0)
1484                 goto bail;
1485
1486         /* Self-consistency check for ipath_compatible_subports() */
1487         if (ipath_supports_subports(IPATH_USER_SWMAJOR, IPATH_USER_SWMINOR) &&
1488             !ipath_compatible_subports(IPATH_USER_SWMAJOR,
1489                                        IPATH_USER_SWMINOR)) {
1490                 dev_info(&dd->pcidev->dev,
1491                          "Inconsistent ipath_compatible_subports()\n");
1492                 goto bail;
1493         }
1494
1495         /* Check for subport compatibility */
1496         if (!ipath_compatible_subports(uinfo->spu_userversion >> 16,
1497                                        uinfo->spu_userversion & 0xffff)) {
1498                 dev_info(&dd->pcidev->dev,
1499                          "Mismatched user version (%d.%d) and driver "
1500                          "version (%d.%d) while port sharing. Ensure "
1501                          "that driver and library are from the same "
1502                          "release.\n",
1503                          (int) (uinfo->spu_userversion >> 16),
1504                          (int) (uinfo->spu_userversion & 0xffff),
1505                          IPATH_USER_SWMAJOR,
1506                          IPATH_USER_SWMINOR);
1507                 goto bail;
1508         }
1509         if (uinfo->spu_subport_cnt > INFINIPATH_MAX_SUBPORT) {
1510                 ret = -EINVAL;
1511                 goto bail;
1512         }
1513
1514         num_subports = uinfo->spu_subport_cnt;
1515         pd->subport_uregbase = vmalloc(PAGE_SIZE * num_subports);
1516         if (!pd->subport_uregbase) {
1517                 ret = -ENOMEM;
1518                 goto bail;
1519         }
1520         /* Note: pd->port_rcvhdrq_size isn't initialized yet. */
1521         size = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
1522                      sizeof(u32), PAGE_SIZE) * num_subports;
1523         pd->subport_rcvhdr_base = vmalloc(size);
1524         if (!pd->subport_rcvhdr_base) {
1525                 ret = -ENOMEM;
1526                 goto bail_ureg;
1527         }
1528
1529         pd->subport_rcvegrbuf = vmalloc(pd->port_rcvegrbuf_chunks *
1530                                         pd->port_rcvegrbuf_size *
1531                                         num_subports);
1532         if (!pd->subport_rcvegrbuf) {
1533                 ret = -ENOMEM;
1534                 goto bail_rhdr;
1535         }
1536
1537         pd->port_subport_cnt = uinfo->spu_subport_cnt;
1538         pd->port_subport_id = uinfo->spu_subport_id;
1539         pd->active_slaves = 1;
1540         set_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
1541         memset(pd->subport_uregbase, 0, PAGE_SIZE * num_subports);
1542         memset(pd->subport_rcvhdr_base, 0, size);
1543         memset(pd->subport_rcvegrbuf, 0, pd->port_rcvegrbuf_chunks *
1544                                          pd->port_rcvegrbuf_size *
1545                                          num_subports);
1546         goto bail;
1547
1548 bail_rhdr:
1549         vfree(pd->subport_rcvhdr_base);
1550 bail_ureg:
1551         vfree(pd->subport_uregbase);
1552         pd->subport_uregbase = NULL;
1553 bail:
1554         return ret;
1555 }
1556
1557 static int try_alloc_port(struct ipath_devdata *dd, int port,
1558                           struct file *fp,
1559                           const struct ipath_user_info *uinfo)
1560 {
1561         struct ipath_portdata *pd;
1562         int ret;
1563
1564         if (!(pd = dd->ipath_pd[port])) {
1565                 void *ptmp;
1566
1567                 pd = kzalloc(sizeof(struct ipath_portdata), GFP_KERNEL);
1568
1569                 /*
1570                  * Allocate memory for use in ipath_tid_update() just once
1571                  * at open, not per call.  Reduces cost of expected send
1572                  * setup.
1573                  */
1574                 ptmp = kmalloc(dd->ipath_rcvtidcnt * sizeof(u16) +
1575                                dd->ipath_rcvtidcnt * sizeof(struct page **),
1576                                GFP_KERNEL);
1577                 if (!pd || !ptmp) {
1578                         ipath_dev_err(dd, "Unable to allocate portdata "
1579                                       "memory, failing open\n");
1580                         ret = -ENOMEM;
1581                         kfree(pd);
1582                         kfree(ptmp);
1583                         goto bail;
1584                 }
1585                 dd->ipath_pd[port] = pd;
1586                 dd->ipath_pd[port]->port_port = port;
1587                 dd->ipath_pd[port]->port_dd = dd;
1588                 dd->ipath_pd[port]->port_tid_pg_list = ptmp;
1589                 init_waitqueue_head(&dd->ipath_pd[port]->port_wait);
1590         }
1591         if (!pd->port_cnt) {
1592                 pd->userversion = uinfo->spu_userversion;
1593                 init_user_egr_sizes(pd);
1594                 if ((ret = init_subports(dd, pd, uinfo)) != 0)
1595                         goto bail;
1596                 ipath_cdbg(PROC, "%s[%u] opened unit:port %u:%u\n",
1597                            current->comm, current->pid, dd->ipath_unit,
1598                            port);
1599                 pd->port_cnt = 1;
1600                 port_fp(fp) = pd;
1601                 pd->port_pid = current->pid;
1602                 strncpy(pd->port_comm, current->comm, sizeof(pd->port_comm));
1603                 ipath_stats.sps_ports++;
1604                 ret = 0;
1605         } else
1606                 ret = -EBUSY;
1607
1608 bail:
1609         return ret;
1610 }
1611
1612 static inline int usable(struct ipath_devdata *dd)
1613 {
1614         return dd &&
1615                 (dd->ipath_flags & IPATH_PRESENT) &&
1616                 dd->ipath_kregbase &&
1617                 dd->ipath_lid &&
1618                 !(dd->ipath_flags & (IPATH_LINKDOWN | IPATH_DISABLED
1619                                      | IPATH_LINKUNK));
1620 }
1621
1622 static int find_free_port(int unit, struct file *fp,
1623                           const struct ipath_user_info *uinfo)
1624 {
1625         struct ipath_devdata *dd = ipath_lookup(unit);
1626         int ret, i;
1627
1628         if (!dd) {
1629                 ret = -ENODEV;
1630                 goto bail;
1631         }
1632
1633         if (!usable(dd)) {
1634                 ret = -ENETDOWN;
1635                 goto bail;
1636         }
1637
1638         for (i = 1; i < dd->ipath_cfgports; i++) {
1639                 ret = try_alloc_port(dd, i, fp, uinfo);
1640                 if (ret != -EBUSY)
1641                         goto bail;
1642         }
1643         ret = -EBUSY;
1644
1645 bail:
1646         return ret;
1647 }
1648
1649 static int find_best_unit(struct file *fp,
1650                           const struct ipath_user_info *uinfo)
1651 {
1652         int ret = 0, i, prefunit = -1, devmax;
1653         int maxofallports, npresent, nup;
1654         int ndev;
1655
1656         devmax = ipath_count_units(&npresent, &nup, &maxofallports);
1657
1658         /*
1659          * This code is present to allow a knowledgeable person to
1660          * specify the layout of processes to processors before opening
1661          * this driver, and then we'll assign the process to the "closest"
1662          * InfiniPath chip to that processor (we assume reasonable connectivity,
1663          * for now).  This code assumes that if affinity has been set
1664          * before this point, that at most one cpu is set; for now this
1665          * is reasonable.  I check for both cpus_empty() and cpus_full(),
1666          * in case some kernel variant sets none of the bits when no
1667          * affinity is set.  2.6.11 and 12 kernels have all present
1668          * cpus set.  Some day we'll have to fix it up further to handle
1669          * a cpu subset.  This algorithm fails for two HT chips connected
1670          * in tunnel fashion.  Eventually this needs real topology
1671          * information.  There may be some issues with dual core numbering
1672          * as well.  This needs more work prior to release.
1673          */
1674         if (!cpus_empty(current->cpus_allowed) &&
1675             !cpus_full(current->cpus_allowed)) {
1676                 int ncpus = num_online_cpus(), curcpu = -1, nset = 0;
1677                 for (i = 0; i < ncpus; i++)
1678                         if (cpu_isset(i, current->cpus_allowed)) {
1679                                 ipath_cdbg(PROC, "%s[%u] affinity set for "
1680                                            "cpu %d/%d\n", current->comm,
1681                                            current->pid, i, ncpus);
1682                                 curcpu = i;
1683                                 nset++;
1684                         }
1685                 if (curcpu != -1 && nset != ncpus) {
1686                         if (npresent) {
1687                                 prefunit = curcpu / (ncpus / npresent);
1688                                 ipath_cdbg(PROC,"%s[%u] %d chips, %d cpus, "
1689                                           "%d cpus/chip, select unit %d\n",
1690                                           current->comm, current->pid,
1691                                           npresent, ncpus, ncpus / npresent,
1692                                           prefunit);
1693                         }
1694                 }
1695         }
1696
1697         /*
1698          * user ports start at 1, kernel port is 0
1699          * For now, we do round-robin access across all chips
1700          */
1701
1702         if (prefunit != -1)
1703                 devmax = prefunit + 1;
1704 recheck:
1705         for (i = 1; i < maxofallports; i++) {
1706                 for (ndev = prefunit != -1 ? prefunit : 0; ndev < devmax;
1707                      ndev++) {
1708                         struct ipath_devdata *dd = ipath_lookup(ndev);
1709
1710                         if (!usable(dd))
1711                                 continue; /* can't use this unit */
1712                         if (i >= dd->ipath_cfgports)
1713                                 /*
1714                                  * Maxed out on users of this unit. Try
1715                                  * next.
1716                                  */
1717                                 continue;
1718                         ret = try_alloc_port(dd, i, fp, uinfo);
1719                         if (!ret)
1720                                 goto done;
1721                 }
1722         }
1723
1724         if (npresent) {
1725                 if (nup == 0) {
1726                         ret = -ENETDOWN;
1727                         ipath_dbg("No ports available (none initialized "
1728                                   "and ready)\n");
1729                 } else {
1730                         if (prefunit > 0) {
1731                                 /* if started above 0, retry from 0 */
1732                                 ipath_cdbg(PROC,
1733                                            "%s[%u] no ports on prefunit "
1734                                            "%d, clear and re-check\n",
1735                                            current->comm, current->pid,
1736                                            prefunit);
1737                                 devmax = ipath_count_units(NULL, NULL,
1738                                                            NULL);
1739                                 prefunit = -1;
1740                                 goto recheck;
1741                         }
1742                         ret = -EBUSY;
1743                         ipath_dbg("No ports available\n");
1744                 }
1745         } else {
1746                 ret = -ENXIO;
1747                 ipath_dbg("No boards found\n");
1748         }
1749
1750 done:
1751         return ret;
1752 }
1753
1754 static int find_shared_port(struct file *fp,
1755                             const struct ipath_user_info *uinfo)
1756 {
1757         int devmax, ndev, i;
1758         int ret = 0;
1759
1760         devmax = ipath_count_units(NULL, NULL, NULL);
1761
1762         for (ndev = 0; ndev < devmax; ndev++) {
1763                 struct ipath_devdata *dd = ipath_lookup(ndev);
1764
1765                 if (!dd)
1766                         continue;
1767                 for (i = 1; i < dd->ipath_cfgports; i++) {
1768                         struct ipath_portdata *pd = dd->ipath_pd[i];
1769
1770                         /* Skip ports which are not yet open */
1771                         if (!pd || !pd->port_cnt)
1772                                 continue;
1773                         /* Skip port if it doesn't match the requested one */
1774                         if (pd->port_subport_id != uinfo->spu_subport_id)
1775                                 continue;
1776                         /* Verify the sharing process matches the master */
1777                         if (pd->port_subport_cnt != uinfo->spu_subport_cnt ||
1778                             pd->userversion != uinfo->spu_userversion ||
1779                             pd->port_cnt >= pd->port_subport_cnt) {
1780                                 ret = -EINVAL;
1781                                 goto done;
1782                         }
1783                         port_fp(fp) = pd;
1784                         subport_fp(fp) = pd->port_cnt++;
1785                         tidcursor_fp(fp) = 0;
1786                         pd->active_slaves |= 1 << subport_fp(fp);
1787                         ipath_cdbg(PROC,
1788                                    "%s[%u] %u sharing %s[%u] unit:port %u:%u\n",
1789                                    current->comm, current->pid,
1790                                    subport_fp(fp),
1791                                    pd->port_comm, pd->port_pid,
1792                                    dd->ipath_unit, pd->port_port);
1793                         ret = 1;
1794                         goto done;
1795                 }
1796         }
1797
1798 done:
1799         return ret;
1800 }
1801
1802 static int ipath_open(struct inode *in, struct file *fp)
1803 {
1804         /* The real work is performed later in ipath_assign_port() */
1805         fp->private_data = kzalloc(sizeof(struct ipath_filedata), GFP_KERNEL);
1806         return fp->private_data ? 0 : -ENOMEM;
1807 }
1808
1809 /* Get port early, so can set affinity prior to memory allocation */
1810 static int ipath_assign_port(struct file *fp,
1811                               const struct ipath_user_info *uinfo)
1812 {
1813         int ret;
1814         int i_minor;
1815         unsigned swmajor, swminor;
1816
1817         /* Check to be sure we haven't already initialized this file */
1818         if (port_fp(fp)) {
1819                 ret = -EINVAL;
1820                 goto done;
1821         }
1822
1823         /* for now, if major version is different, bail */
1824         swmajor = uinfo->spu_userversion >> 16;
1825         if (swmajor != IPATH_USER_SWMAJOR) {
1826                 ipath_dbg("User major version %d not same as driver "
1827                           "major %d\n", uinfo->spu_userversion >> 16,
1828                           IPATH_USER_SWMAJOR);
1829                 ret = -ENODEV;
1830                 goto done;
1831         }
1832
1833         swminor = uinfo->spu_userversion & 0xffff;
1834         if (swminor != IPATH_USER_SWMINOR)
1835                 ipath_dbg("User minor version %d not same as driver "
1836                           "minor %d\n", swminor, IPATH_USER_SWMINOR);
1837
1838         mutex_lock(&ipath_mutex);
1839
1840         if (ipath_compatible_subports(swmajor, swminor) &&
1841             uinfo->spu_subport_cnt &&
1842             (ret = find_shared_port(fp, uinfo))) {
1843                 mutex_unlock(&ipath_mutex);
1844                 if (ret > 0)
1845                         ret = 0;
1846                 goto done;
1847         }
1848
1849         i_minor = iminor(fp->f_path.dentry->d_inode) - IPATH_USER_MINOR_BASE;
1850         ipath_cdbg(VERBOSE, "open on dev %lx (minor %d)\n",
1851                    (long)fp->f_path.dentry->d_inode->i_rdev, i_minor);
1852
1853         if (i_minor)
1854                 ret = find_free_port(i_minor - 1, fp, uinfo);
1855         else
1856                 ret = find_best_unit(fp, uinfo);
1857
1858         mutex_unlock(&ipath_mutex);
1859
1860 done:
1861         return ret;
1862 }
1863
1864
1865 static int ipath_do_user_init(struct file *fp,
1866                               const struct ipath_user_info *uinfo)
1867 {
1868         int ret;
1869         struct ipath_portdata *pd = port_fp(fp);
1870         struct ipath_devdata *dd;
1871         u32 head32;
1872
1873         /* Subports don't need to initialize anything since master did it. */
1874         if (subport_fp(fp)) {
1875                 ret = wait_event_interruptible(pd->port_wait,
1876                         !test_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag));
1877                 goto done;
1878         }
1879
1880         dd = pd->port_dd;
1881
1882         if (uinfo->spu_rcvhdrsize) {
1883                 ret = ipath_setrcvhdrsize(dd, uinfo->spu_rcvhdrsize);
1884                 if (ret)
1885                         goto done;
1886         }
1887
1888         /* for now we do nothing with rcvhdrcnt: uinfo->spu_rcvhdrcnt */
1889
1890         /* for right now, kernel piobufs are at end, so port 1 is at 0 */
1891         pd->port_piobufs = dd->ipath_piobufbase +
1892                 dd->ipath_pbufsport * (pd->port_port - 1) * dd->ipath_palign;
1893         ipath_cdbg(VERBOSE, "Set base of piobufs for port %u to 0x%x\n",
1894                    pd->port_port, pd->port_piobufs);
1895
1896         /*
1897          * Now allocate the rcvhdr Q and eager TIDs; skip the TID
1898          * array for time being.  If pd->port_port > chip-supported,
1899          * we need to do extra stuff here to handle by handling overflow
1900          * through port 0, someday
1901          */
1902         ret = ipath_create_rcvhdrq(dd, pd);
1903         if (!ret)
1904                 ret = ipath_create_user_egr(pd);
1905         if (ret)
1906                 goto done;
1907
1908         /*
1909          * set the eager head register for this port to the current values
1910          * of the tail pointers, since we don't know if they were
1911          * updated on last use of the port.
1912          */
1913         head32 = ipath_read_ureg32(dd, ur_rcvegrindextail, pd->port_port);
1914         ipath_write_ureg(dd, ur_rcvegrindexhead, head32, pd->port_port);
1915         dd->ipath_lastegrheads[pd->port_port] = -1;
1916         dd->ipath_lastrcvhdrqtails[pd->port_port] = -1;
1917         ipath_cdbg(VERBOSE, "Wrote port%d egrhead %x from tail regs\n",
1918                 pd->port_port, head32);
1919         pd->port_tidcursor = 0; /* start at beginning after open */
1920         /*
1921          * now enable the port; the tail registers will be written to memory
1922          * by the chip as soon as it sees the write to
1923          * dd->ipath_kregs->kr_rcvctrl.  The update only happens on
1924          * transition from 0 to 1, so clear it first, then set it as part of
1925          * enabling the port.  This will (very briefly) affect any other
1926          * open ports, but it shouldn't be long enough to be an issue.
1927          * We explictly set the in-memory copy to 0 beforehand, so we don't
1928          * have to wait to be sure the DMA update has happened.
1929          */
1930         *(volatile u64 *)pd->port_rcvhdrtail_kvaddr = 0ULL;
1931         set_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
1932                 &dd->ipath_rcvctrl);
1933         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1934                          dd->ipath_rcvctrl & ~INFINIPATH_R_TAILUPD);
1935         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1936                          dd->ipath_rcvctrl);
1937         /* Notify any waiting slaves */
1938         if (pd->port_subport_cnt) {
1939                 clear_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
1940                 wake_up(&pd->port_wait);
1941         }
1942 done:
1943         return ret;
1944 }
1945
1946 /**
1947  * unlock_exptid - unlock any expected TID entries port still had in use
1948  * @pd: port
1949  *
1950  * We don't actually update the chip here, because we do a bulk update
1951  * below, using ipath_f_clear_tids.
1952  */
1953 static void unlock_expected_tids(struct ipath_portdata *pd)
1954 {
1955         struct ipath_devdata *dd = pd->port_dd;
1956         int port_tidbase = pd->port_port * dd->ipath_rcvtidcnt;
1957         int i, cnt = 0, maxtid = port_tidbase + dd->ipath_rcvtidcnt;
1958
1959         ipath_cdbg(VERBOSE, "Port %u unlocking any locked expTID pages\n",
1960                    pd->port_port);
1961         for (i = port_tidbase; i < maxtid; i++) {
1962                 if (!dd->ipath_pageshadow[i])
1963                         continue;
1964
1965                 pci_unmap_page(dd->pcidev, dd->ipath_physshadow[i],
1966                         PAGE_SIZE, PCI_DMA_FROMDEVICE);
1967                 ipath_release_user_pages_on_close(&dd->ipath_pageshadow[i],
1968                                                   1);
1969                 dd->ipath_pageshadow[i] = NULL;
1970                 cnt++;
1971                 ipath_stats.sps_pageunlocks++;
1972         }
1973         if (cnt)
1974                 ipath_cdbg(VERBOSE, "Port %u locked %u expTID entries\n",
1975                            pd->port_port, cnt);
1976
1977         if (ipath_stats.sps_pagelocks || ipath_stats.sps_pageunlocks)
1978                 ipath_cdbg(VERBOSE, "%llu pages locked, %llu unlocked\n",
1979                            (unsigned long long) ipath_stats.sps_pagelocks,
1980                            (unsigned long long)
1981                            ipath_stats.sps_pageunlocks);
1982 }
1983
1984 static int ipath_close(struct inode *in, struct file *fp)
1985 {
1986         int ret = 0;
1987         struct ipath_filedata *fd;
1988         struct ipath_portdata *pd;
1989         struct ipath_devdata *dd;
1990         unsigned port;
1991
1992         ipath_cdbg(VERBOSE, "close on dev %lx, private data %p\n",
1993                    (long)in->i_rdev, fp->private_data);
1994
1995         mutex_lock(&ipath_mutex);
1996
1997         fd = (struct ipath_filedata *) fp->private_data;
1998         fp->private_data = NULL;
1999         pd = fd->pd;
2000         if (!pd) {
2001                 mutex_unlock(&ipath_mutex);
2002                 goto bail;
2003         }
2004         if (--pd->port_cnt) {
2005                 /*
2006                  * XXX If the master closes the port before the slave(s),
2007                  * revoke the mmap for the eager receive queue so
2008                  * the slave(s) don't wait for receive data forever.
2009                  */
2010                 pd->active_slaves &= ~(1 << fd->subport);
2011                 mutex_unlock(&ipath_mutex);
2012                 goto bail;
2013         }
2014         port = pd->port_port;
2015         dd = pd->port_dd;
2016
2017         if (pd->port_hdrqfull) {
2018                 ipath_cdbg(PROC, "%s[%u] had %u rcvhdrqfull errors "
2019                            "during run\n", pd->port_comm, pd->port_pid,
2020                            pd->port_hdrqfull);
2021                 pd->port_hdrqfull = 0;
2022         }
2023
2024         if (pd->port_rcvwait_to || pd->port_piowait_to
2025             || pd->port_rcvnowait || pd->port_pionowait) {
2026                 ipath_cdbg(VERBOSE, "port%u, %u rcv, %u pio wait timeo; "
2027                            "%u rcv %u, pio already\n",
2028                            pd->port_port, pd->port_rcvwait_to,
2029                            pd->port_piowait_to, pd->port_rcvnowait,
2030                            pd->port_pionowait);
2031                 pd->port_rcvwait_to = pd->port_piowait_to =
2032                         pd->port_rcvnowait = pd->port_pionowait = 0;
2033         }
2034         if (pd->port_flag) {
2035                 ipath_dbg("port %u port_flag still set to 0x%lx\n",
2036                           pd->port_port, pd->port_flag);
2037                 pd->port_flag = 0;
2038         }
2039
2040         if (dd->ipath_kregbase) {
2041                 int i;
2042                 /* atomically clear receive enable port. */
2043                 clear_bit(INFINIPATH_R_PORTENABLE_SHIFT + port,
2044                           &dd->ipath_rcvctrl);
2045                 ipath_write_kreg( dd, dd->ipath_kregs->kr_rcvctrl,
2046                         dd->ipath_rcvctrl);
2047                 /* and read back from chip to be sure that nothing
2048                  * else is in flight when we do the rest */
2049                 (void)ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
2050
2051                 /* clean up the pkeys for this port user */
2052                 ipath_clean_part_key(pd, dd);
2053                 /*
2054                  * be paranoid, and never write 0's to these, just use an
2055                  * unused part of the port 0 tail page.  Of course,
2056                  * rcvhdraddr points to a large chunk of memory, so this
2057                  * could still trash things, but at least it won't trash
2058                  * page 0, and by disabling the port, it should stop "soon",
2059                  * even if a packet or two is in already in flight after we
2060                  * disabled the port.
2061                  */
2062                 ipath_write_kreg_port(dd,
2063                         dd->ipath_kregs->kr_rcvhdrtailaddr, port,
2064                         dd->ipath_dummy_hdrq_phys);
2065                 ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdraddr,
2066                         pd->port_port, dd->ipath_dummy_hdrq_phys);
2067
2068                 i = dd->ipath_pbufsport * (port - 1);
2069                 ipath_disarm_piobufs(dd, i, dd->ipath_pbufsport);
2070
2071                 dd->ipath_f_clear_tids(dd, pd->port_port);
2072
2073                 if (dd->ipath_pageshadow)
2074                         unlock_expected_tids(pd);
2075                 ipath_stats.sps_ports--;
2076                 ipath_cdbg(PROC, "%s[%u] closed port %u:%u\n",
2077                            pd->port_comm, pd->port_pid,
2078                            dd->ipath_unit, port);
2079         }
2080
2081         pd->port_pid = 0;
2082         dd->ipath_pd[pd->port_port] = NULL; /* before releasing mutex */
2083         mutex_unlock(&ipath_mutex);
2084         ipath_free_pddata(dd, pd); /* after releasing the mutex */
2085
2086 bail:
2087         kfree(fd);
2088         return ret;
2089 }
2090
2091 static int ipath_port_info(struct ipath_portdata *pd, u16 subport,
2092                            struct ipath_port_info __user *uinfo)
2093 {
2094         struct ipath_port_info info;
2095         int nup;
2096         int ret;
2097         size_t sz;
2098
2099         (void) ipath_count_units(NULL, &nup, NULL);
2100         info.num_active = nup;
2101         info.unit = pd->port_dd->ipath_unit;
2102         info.port = pd->port_port;
2103         info.subport = subport;
2104         /* Don't return new fields if old library opened the port. */
2105         if (ipath_supports_subports(pd->userversion >> 16,
2106                                     pd->userversion & 0xffff)) {
2107                 /* Number of user ports available for this device. */
2108                 info.num_ports = pd->port_dd->ipath_cfgports - 1;
2109                 info.num_subports = pd->port_subport_cnt;
2110                 sz = sizeof(info);
2111         } else
2112                 sz = sizeof(info) - 2 * sizeof(u16);
2113
2114         if (copy_to_user(uinfo, &info, sz)) {
2115                 ret = -EFAULT;
2116                 goto bail;
2117         }
2118         ret = 0;
2119
2120 bail:
2121         return ret;
2122 }
2123
2124 static int ipath_get_slave_info(struct ipath_portdata *pd,
2125                                 void __user *slave_mask_addr)
2126 {
2127         int ret = 0;
2128
2129         if (copy_to_user(slave_mask_addr, &pd->active_slaves, sizeof(u32)))
2130                 ret = -EFAULT;
2131         return ret;
2132 }
2133
2134 static int ipath_force_pio_avail_update(struct ipath_devdata *dd)
2135 {
2136         u64 reg = dd->ipath_sendctrl;
2137
2138         clear_bit(IPATH_S_PIOBUFAVAILUPD, &reg);
2139         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, reg);
2140         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, dd->ipath_sendctrl);
2141
2142         return 0;
2143 }
2144
2145 static ssize_t ipath_write(struct file *fp, const char __user *data,
2146                            size_t count, loff_t *off)
2147 {
2148         const struct ipath_cmd __user *ucmd;
2149         struct ipath_portdata *pd;
2150         const void __user *src;
2151         size_t consumed, copy;
2152         struct ipath_cmd cmd;
2153         ssize_t ret = 0;
2154         void *dest;
2155
2156         if (count < sizeof(cmd.type)) {
2157                 ret = -EINVAL;
2158                 goto bail;
2159         }
2160
2161         ucmd = (const struct ipath_cmd __user *) data;
2162
2163         if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) {
2164                 ret = -EFAULT;
2165                 goto bail;
2166         }
2167
2168         consumed = sizeof(cmd.type);
2169
2170         switch (cmd.type) {
2171         case IPATH_CMD_ASSIGN_PORT:
2172         case __IPATH_CMD_USER_INIT:
2173         case IPATH_CMD_USER_INIT:
2174                 copy = sizeof(cmd.cmd.user_info);
2175                 dest = &cmd.cmd.user_info;
2176                 src = &ucmd->cmd.user_info;
2177                 break;
2178         case IPATH_CMD_RECV_CTRL:
2179                 copy = sizeof(cmd.cmd.recv_ctrl);
2180                 dest = &cmd.cmd.recv_ctrl;
2181                 src = &ucmd->cmd.recv_ctrl;
2182                 break;
2183         case IPATH_CMD_PORT_INFO:
2184                 copy = sizeof(cmd.cmd.port_info);
2185                 dest = &cmd.cmd.port_info;
2186                 src = &ucmd->cmd.port_info;
2187                 break;
2188         case IPATH_CMD_TID_UPDATE:
2189         case IPATH_CMD_TID_FREE:
2190                 copy = sizeof(cmd.cmd.tid_info);
2191                 dest = &cmd.cmd.tid_info;
2192                 src = &ucmd->cmd.tid_info;
2193                 break;
2194         case IPATH_CMD_SET_PART_KEY:
2195                 copy = sizeof(cmd.cmd.part_key);
2196                 dest = &cmd.cmd.part_key;
2197                 src = &ucmd->cmd.part_key;
2198                 break;
2199         case __IPATH_CMD_SLAVE_INFO:
2200                 copy = sizeof(cmd.cmd.slave_mask_addr);
2201                 dest = &cmd.cmd.slave_mask_addr;
2202                 src = &ucmd->cmd.slave_mask_addr;
2203                 break;
2204         case IPATH_CMD_PIOAVAILUPD:     // force an update of PIOAvail reg
2205                 copy = 0;
2206                 src = NULL;
2207                 dest = NULL;
2208                 break;
2209         case IPATH_CMD_POLL_TYPE:
2210                 copy = sizeof(cmd.cmd.poll_type);
2211                 dest = &cmd.cmd.poll_type;
2212                 src = &ucmd->cmd.poll_type;
2213                 break;
2214         default:
2215                 ret = -EINVAL;
2216                 goto bail;
2217         }
2218
2219         if (copy) {
2220                 if ((count - consumed) < copy) {
2221                         ret = -EINVAL;
2222                         goto bail;
2223                 }
2224
2225                 if (copy_from_user(dest, src, copy)) {
2226                         ret = -EFAULT;
2227                         goto bail;
2228                 }
2229
2230                 consumed += copy;
2231         }
2232
2233         pd = port_fp(fp);
2234         if (!pd && cmd.type != __IPATH_CMD_USER_INIT &&
2235                 cmd.type != IPATH_CMD_ASSIGN_PORT) {
2236                 ret = -EINVAL;
2237                 goto bail;
2238         }
2239
2240         switch (cmd.type) {
2241         case IPATH_CMD_ASSIGN_PORT:
2242                 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2243                 if (ret)
2244                         goto bail;
2245                 break;
2246         case __IPATH_CMD_USER_INIT:
2247                 /* backwards compatibility, get port first */
2248                 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2249                 if (ret)
2250                         goto bail;
2251                 /* and fall through to current version. */
2252         case IPATH_CMD_USER_INIT:
2253                 ret = ipath_do_user_init(fp, &cmd.cmd.user_info);
2254                 if (ret)
2255                         goto bail;
2256                 ret = ipath_get_base_info(
2257                         fp, (void __user *) (unsigned long)
2258                         cmd.cmd.user_info.spu_base_info,
2259                         cmd.cmd.user_info.spu_base_info_size);
2260                 break;
2261         case IPATH_CMD_RECV_CTRL:
2262                 ret = ipath_manage_rcvq(pd, subport_fp(fp), cmd.cmd.recv_ctrl);
2263                 break;
2264         case IPATH_CMD_PORT_INFO:
2265                 ret = ipath_port_info(pd, subport_fp(fp),
2266                                       (struct ipath_port_info __user *)
2267                                       (unsigned long) cmd.cmd.port_info);
2268                 break;
2269         case IPATH_CMD_TID_UPDATE:
2270                 ret = ipath_tid_update(pd, fp, &cmd.cmd.tid_info);
2271                 break;
2272         case IPATH_CMD_TID_FREE:
2273                 ret = ipath_tid_free(pd, subport_fp(fp), &cmd.cmd.tid_info);
2274                 break;
2275         case IPATH_CMD_SET_PART_KEY:
2276                 ret = ipath_set_part_key(pd, cmd.cmd.part_key);
2277                 break;
2278         case __IPATH_CMD_SLAVE_INFO:
2279                 ret = ipath_get_slave_info(pd,
2280                                            (void __user *) (unsigned long)
2281                                            cmd.cmd.slave_mask_addr);
2282                 break;
2283         case IPATH_CMD_PIOAVAILUPD:
2284                 ret = ipath_force_pio_avail_update(pd->port_dd);
2285                 break;
2286         case IPATH_CMD_POLL_TYPE:
2287                 pd->poll_type = cmd.cmd.poll_type;
2288                 break;
2289         }
2290
2291         if (ret >= 0)
2292                 ret = consumed;
2293
2294 bail:
2295         return ret;
2296 }
2297
2298 static struct class *ipath_class;
2299
2300 static int init_cdev(int minor, char *name, const struct file_operations *fops,
2301                      struct cdev **cdevp, struct class_device **class_devp)
2302 {
2303         const dev_t dev = MKDEV(IPATH_MAJOR, minor);
2304         struct cdev *cdev = NULL;
2305         struct class_device *class_dev = NULL;
2306         int ret;
2307
2308         cdev = cdev_alloc();
2309         if (!cdev) {
2310                 printk(KERN_ERR IPATH_DRV_NAME
2311                        ": Could not allocate cdev for minor %d, %s\n",
2312                        minor, name);
2313                 ret = -ENOMEM;
2314                 goto done;
2315         }
2316
2317         cdev->owner = THIS_MODULE;
2318         cdev->ops = fops;
2319         kobject_set_name(&cdev->kobj, name);
2320
2321         ret = cdev_add(cdev, dev, 1);
2322         if (ret < 0) {
2323                 printk(KERN_ERR IPATH_DRV_NAME
2324                        ": Could not add cdev for minor %d, %s (err %d)\n",
2325                        minor, name, -ret);
2326                 goto err_cdev;
2327         }
2328
2329         class_dev = class_device_create(ipath_class, NULL, dev, NULL, name);
2330
2331         if (IS_ERR(class_dev)) {
2332                 ret = PTR_ERR(class_dev);
2333                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2334                        "class_dev for minor %d, %s (err %d)\n",
2335                        minor, name, -ret);
2336                 goto err_cdev;
2337         }
2338
2339         goto done;
2340
2341 err_cdev:
2342         cdev_del(cdev);
2343         cdev = NULL;
2344
2345 done:
2346         if (ret >= 0) {
2347                 *cdevp = cdev;
2348                 *class_devp = class_dev;
2349         } else {
2350                 *cdevp = NULL;
2351                 *class_devp = NULL;
2352         }
2353
2354         return ret;
2355 }
2356
2357 int ipath_cdev_init(int minor, char *name, const struct file_operations *fops,
2358                     struct cdev **cdevp, struct class_device **class_devp)
2359 {
2360         return init_cdev(minor, name, fops, cdevp, class_devp);
2361 }
2362
2363 static void cleanup_cdev(struct cdev **cdevp,
2364                          struct class_device **class_devp)
2365 {
2366         struct class_device *class_dev = *class_devp;
2367
2368         if (class_dev) {
2369                 class_device_unregister(class_dev);
2370                 *class_devp = NULL;
2371         }
2372
2373         if (*cdevp) {
2374                 cdev_del(*cdevp);
2375                 *cdevp = NULL;
2376         }
2377 }
2378
2379 void ipath_cdev_cleanup(struct cdev **cdevp,
2380                         struct class_device **class_devp)
2381 {
2382         cleanup_cdev(cdevp, class_devp);
2383 }
2384
2385 static struct cdev *wildcard_cdev;
2386 static struct class_device *wildcard_class_dev;
2387
2388 static const dev_t dev = MKDEV(IPATH_MAJOR, 0);
2389
2390 static int user_init(void)
2391 {
2392         int ret;
2393
2394         ret = register_chrdev_region(dev, IPATH_NMINORS, IPATH_DRV_NAME);
2395         if (ret < 0) {
2396                 printk(KERN_ERR IPATH_DRV_NAME ": Could not register "
2397                        "chrdev region (err %d)\n", -ret);
2398                 goto done;
2399         }
2400
2401         ipath_class = class_create(THIS_MODULE, IPATH_DRV_NAME);
2402
2403         if (IS_ERR(ipath_class)) {
2404                 ret = PTR_ERR(ipath_class);
2405                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2406                        "device class (err %d)\n", -ret);
2407                 goto bail;
2408         }
2409
2410         goto done;
2411 bail:
2412         unregister_chrdev_region(dev, IPATH_NMINORS);
2413 done:
2414         return ret;
2415 }
2416
2417 static void user_cleanup(void)
2418 {
2419         if (ipath_class) {
2420                 class_destroy(ipath_class);
2421                 ipath_class = NULL;
2422         }
2423
2424         unregister_chrdev_region(dev, IPATH_NMINORS);
2425 }
2426
2427 static atomic_t user_count = ATOMIC_INIT(0);
2428 static atomic_t user_setup = ATOMIC_INIT(0);
2429
2430 int ipath_user_add(struct ipath_devdata *dd)
2431 {
2432         char name[10];
2433         int ret;
2434
2435         if (atomic_inc_return(&user_count) == 1) {
2436                 ret = user_init();
2437                 if (ret < 0) {
2438                         ipath_dev_err(dd, "Unable to set up user support: "
2439                                       "error %d\n", -ret);
2440                         goto bail;
2441                 }
2442                 ret = init_cdev(0, "ipath", &ipath_file_ops, &wildcard_cdev,
2443                                 &wildcard_class_dev);
2444                 if (ret < 0) {
2445                         ipath_dev_err(dd, "Could not create wildcard "
2446                                       "minor: error %d\n", -ret);
2447                         goto bail_user;
2448                 }
2449
2450                 atomic_set(&user_setup, 1);
2451         }
2452
2453         snprintf(name, sizeof(name), "ipath%d", dd->ipath_unit);
2454
2455         ret = init_cdev(dd->ipath_unit + 1, name, &ipath_file_ops,
2456                         &dd->user_cdev, &dd->user_class_dev);
2457         if (ret < 0)
2458                 ipath_dev_err(dd, "Could not create user minor %d, %s\n",
2459                               dd->ipath_unit + 1, name);
2460
2461         goto bail;
2462
2463 bail_user:
2464         user_cleanup();
2465 bail:
2466         return ret;
2467 }
2468
2469 void ipath_user_remove(struct ipath_devdata *dd)
2470 {
2471         cleanup_cdev(&dd->user_cdev, &dd->user_class_dev);
2472
2473         if (atomic_dec_return(&user_count) == 0) {
2474                 if (atomic_read(&user_setup) == 0)
2475                         goto bail;
2476
2477                 cleanup_cdev(&wildcard_cdev, &wildcard_class_dev);
2478                 user_cleanup();
2479
2480                 atomic_set(&user_setup, 0);
2481         }
2482 bail:
2483         return;
2484 }