IB/ipath: Changes to support PIO bandwidth check on IBA7220
[linux-2.6.git] / drivers / infiniband / hw / ipath / ipath_file_ops.c
1 /*
2  * Copyright (c) 2006, 2007 QLogic Corporation. All rights reserved.
3  * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33
34 #include <linux/pci.h>
35 #include <linux/poll.h>
36 #include <linux/cdev.h>
37 #include <linux/swap.h>
38 #include <linux/vmalloc.h>
39 #include <asm/pgtable.h>
40
41 #include "ipath_kernel.h"
42 #include "ipath_common.h"
43
44 static int ipath_open(struct inode *, struct file *);
45 static int ipath_close(struct inode *, struct file *);
46 static ssize_t ipath_write(struct file *, const char __user *, size_t,
47                            loff_t *);
48 static unsigned int ipath_poll(struct file *, struct poll_table_struct *);
49 static int ipath_mmap(struct file *, struct vm_area_struct *);
50
51 static const struct file_operations ipath_file_ops = {
52         .owner = THIS_MODULE,
53         .write = ipath_write,
54         .open = ipath_open,
55         .release = ipath_close,
56         .poll = ipath_poll,
57         .mmap = ipath_mmap
58 };
59
60 /*
61  * Convert kernel virtual addresses to physical addresses so they don't
62  * potentially conflict with the chip addresses used as mmap offsets.
63  * It doesn't really matter what mmap offset we use as long as we can
64  * interpret it correctly.
65  */
66 static u64 cvt_kvaddr(void *p)
67 {
68         struct page *page;
69         u64 paddr = 0;
70
71         page = vmalloc_to_page(p);
72         if (page)
73                 paddr = page_to_pfn(page) << PAGE_SHIFT;
74
75         return paddr;
76 }
77
78 static int ipath_get_base_info(struct file *fp,
79                                void __user *ubase, size_t ubase_size)
80 {
81         struct ipath_portdata *pd = port_fp(fp);
82         int ret = 0;
83         struct ipath_base_info *kinfo = NULL;
84         struct ipath_devdata *dd = pd->port_dd;
85         unsigned subport_cnt;
86         int shared, master;
87         size_t sz;
88
89         subport_cnt = pd->port_subport_cnt;
90         if (!subport_cnt) {
91                 shared = 0;
92                 master = 0;
93                 subport_cnt = 1;
94         } else {
95                 shared = 1;
96                 master = !subport_fp(fp);
97         }
98
99         sz = sizeof(*kinfo);
100         /* If port sharing is not requested, allow the old size structure */
101         if (!shared)
102                 sz -= 7 * sizeof(u64);
103         if (ubase_size < sz) {
104                 ipath_cdbg(PROC,
105                            "Base size %zu, need %zu (version mismatch?)\n",
106                            ubase_size, sz);
107                 ret = -EINVAL;
108                 goto bail;
109         }
110
111         kinfo = kzalloc(sizeof(*kinfo), GFP_KERNEL);
112         if (kinfo == NULL) {
113                 ret = -ENOMEM;
114                 goto bail;
115         }
116
117         ret = dd->ipath_f_get_base_info(pd, kinfo);
118         if (ret < 0)
119                 goto bail;
120
121         kinfo->spi_rcvhdr_cnt = dd->ipath_rcvhdrcnt;
122         kinfo->spi_rcvhdrent_size = dd->ipath_rcvhdrentsize;
123         kinfo->spi_tidegrcnt = dd->ipath_rcvegrcnt;
124         kinfo->spi_rcv_egrbufsize = dd->ipath_rcvegrbufsize;
125         /*
126          * have to mmap whole thing
127          */
128         kinfo->spi_rcv_egrbuftotlen =
129                 pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
130         kinfo->spi_rcv_egrperchunk = pd->port_rcvegrbufs_perchunk;
131         kinfo->spi_rcv_egrchunksize = kinfo->spi_rcv_egrbuftotlen /
132                 pd->port_rcvegrbuf_chunks;
133         kinfo->spi_tidcnt = dd->ipath_rcvtidcnt / subport_cnt;
134         if (master)
135                 kinfo->spi_tidcnt += dd->ipath_rcvtidcnt % subport_cnt;
136         /*
137          * for this use, may be ipath_cfgports summed over all chips that
138          * are are configured and present
139          */
140         kinfo->spi_nports = dd->ipath_cfgports;
141         /* unit (chip/board) our port is on */
142         kinfo->spi_unit = dd->ipath_unit;
143         /* for now, only a single page */
144         kinfo->spi_tid_maxsize = PAGE_SIZE;
145
146         /*
147          * Doing this per port, and based on the skip value, etc.  This has
148          * to be the actual buffer size, since the protocol code treats it
149          * as an array.
150          *
151          * These have to be set to user addresses in the user code via mmap.
152          * These values are used on return to user code for the mmap target
153          * addresses only.  For 32 bit, same 44 bit address problem, so use
154          * the physical address, not virtual.  Before 2.6.11, using the
155          * page_address() macro worked, but in 2.6.11, even that returns the
156          * full 64 bit address (upper bits all 1's).  So far, using the
157          * physical addresses (or chip offsets, for chip mapping) works, but
158          * no doubt some future kernel release will change that, and we'll be
159          * on to yet another method of dealing with this.
160          */
161         kinfo->spi_rcvhdr_base = (u64) pd->port_rcvhdrq_phys;
162         kinfo->spi_rcvhdr_tailaddr = (u64) pd->port_rcvhdrqtailaddr_phys;
163         kinfo->spi_rcv_egrbufs = (u64) pd->port_rcvegr_phys;
164         kinfo->spi_pioavailaddr = (u64) dd->ipath_pioavailregs_phys;
165         kinfo->spi_status = (u64) kinfo->spi_pioavailaddr +
166                 (void *) dd->ipath_statusp -
167                 (void *) dd->ipath_pioavailregs_dma;
168         if (!shared) {
169                 kinfo->spi_piocnt = dd->ipath_pbufsport;
170                 kinfo->spi_piobufbase = (u64) pd->port_piobufs;
171                 kinfo->__spi_uregbase = (u64) dd->ipath_uregbase +
172                         dd->ipath_ureg_align * 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_ureg_align * pd->port_port;
190                 kinfo->spi_port_rcvegrbuf = kinfo->spi_rcv_egrbufs;
191                 kinfo->spi_port_rcvhdr_base = kinfo->spi_rcvhdr_base;
192                 kinfo->spi_port_rcvhdr_tailaddr = kinfo->spi_rcvhdr_tailaddr;
193
194                 kinfo->__spi_uregbase = cvt_kvaddr(pd->subport_uregbase +
195                         PAGE_SIZE * subport_fp(fp));
196
197                 kinfo->spi_rcvhdr_base = cvt_kvaddr(pd->subport_rcvhdr_base +
198                         pd->port_rcvhdrq_size * subport_fp(fp));
199                 kinfo->spi_rcvhdr_tailaddr = 0;
200                 kinfo->spi_rcv_egrbufs = cvt_kvaddr(pd->subport_rcvegrbuf +
201                         pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size *
202                         subport_fp(fp));
203
204                 kinfo->spi_subport_uregbase =
205                         cvt_kvaddr(pd->subport_uregbase);
206                 kinfo->spi_subport_rcvegrbuf =
207                         cvt_kvaddr(pd->subport_rcvegrbuf);
208                 kinfo->spi_subport_rcvhdr_base =
209                         cvt_kvaddr(pd->subport_rcvhdr_base);
210                 ipath_cdbg(PROC, "port %u flags %x %llx %llx %llx\n",
211                         kinfo->spi_port, kinfo->spi_runtime_flags,
212                         (unsigned long long) kinfo->spi_subport_uregbase,
213                         (unsigned long long) kinfo->spi_subport_rcvegrbuf,
214                         (unsigned long long) kinfo->spi_subport_rcvhdr_base);
215         }
216
217         kinfo->spi_pioindex = (kinfo->spi_piobufbase - dd->ipath_piobufbase) /
218                 dd->ipath_palign;
219         kinfo->spi_pioalign = dd->ipath_palign;
220
221         kinfo->spi_qpair = IPATH_KD_QP;
222         kinfo->spi_piosize = dd->ipath_ibmaxlen;
223         kinfo->spi_mtu = dd->ipath_ibmaxlen;    /* maxlen, not ibmtu */
224         kinfo->spi_port = pd->port_port;
225         kinfo->spi_subport = subport_fp(fp);
226         kinfo->spi_sw_version = IPATH_KERN_SWVERSION;
227         kinfo->spi_hw_version = dd->ipath_revision;
228
229         if (master) {
230                 kinfo->spi_runtime_flags |= IPATH_RUNTIME_MASTER;
231         }
232
233         sz = (ubase_size < sizeof(*kinfo)) ? ubase_size : sizeof(*kinfo);
234         if (copy_to_user(ubase, kinfo, sz))
235                 ret = -EFAULT;
236
237 bail:
238         kfree(kinfo);
239         return ret;
240 }
241
242 /**
243  * ipath_tid_update - update a port TID
244  * @pd: the port
245  * @fp: the ipath device file
246  * @ti: the TID information
247  *
248  * The new implementation as of Oct 2004 is that the driver assigns
249  * the tid and returns it to the caller.   To make it easier to
250  * catch bugs, and to reduce search time, we keep a cursor for
251  * each port, walking the shadow tid array to find one that's not
252  * in use.
253  *
254  * For now, if we can't allocate the full list, we fail, although
255  * in the long run, we'll allocate as many as we can, and the
256  * caller will deal with that by trying the remaining pages later.
257  * That means that when we fail, we have to mark the tids as not in
258  * use again, in our shadow copy.
259  *
260  * It's up to the caller to free the tids when they are done.
261  * We'll unlock the pages as they free them.
262  *
263  * Also, right now we are locking one page at a time, but since
264  * the intended use of this routine is for a single group of
265  * virtually contiguous pages, that should change to improve
266  * performance.
267  */
268 static int ipath_tid_update(struct ipath_portdata *pd, struct file *fp,
269                             const struct ipath_tid_info *ti)
270 {
271         int ret = 0, ntids;
272         u32 tid, porttid, cnt, i, tidcnt, tidoff;
273         u16 *tidlist;
274         struct ipath_devdata *dd = pd->port_dd;
275         u64 physaddr;
276         unsigned long vaddr;
277         u64 __iomem *tidbase;
278         unsigned long tidmap[8];
279         struct page **pagep = NULL;
280         unsigned subport = subport_fp(fp);
281
282         if (!dd->ipath_pageshadow) {
283                 ret = -ENOMEM;
284                 goto done;
285         }
286
287         cnt = ti->tidcnt;
288         if (!cnt) {
289                 ipath_dbg("After copyin, tidcnt 0, tidlist %llx\n",
290                           (unsigned long long) ti->tidlist);
291                 /*
292                  * Should we treat as success?  likely a bug
293                  */
294                 ret = -EFAULT;
295                 goto done;
296         }
297         porttid = pd->port_port * dd->ipath_rcvtidcnt;
298         if (!pd->port_subport_cnt) {
299                 tidcnt = dd->ipath_rcvtidcnt;
300                 tid = pd->port_tidcursor;
301                 tidoff = 0;
302         } else if (!subport) {
303                 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
304                          (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
305                 tidoff = dd->ipath_rcvtidcnt - tidcnt;
306                 porttid += tidoff;
307                 tid = tidcursor_fp(fp);
308         } else {
309                 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
310                 tidoff = tidcnt * (subport - 1);
311                 porttid += tidoff;
312                 tid = tidcursor_fp(fp);
313         }
314         if (cnt > tidcnt) {
315                 /* make sure it all fits in port_tid_pg_list */
316                 dev_info(&dd->pcidev->dev, "Process tried to allocate %u "
317                          "TIDs, only trying max (%u)\n", cnt, tidcnt);
318                 cnt = tidcnt;
319         }
320         pagep = &((struct page **) pd->port_tid_pg_list)[tidoff];
321         tidlist = &((u16 *) &pagep[dd->ipath_rcvtidcnt])[tidoff];
322
323         memset(tidmap, 0, sizeof(tidmap));
324         /* before decrement; chip actual # */
325         ntids = tidcnt;
326         tidbase = (u64 __iomem *) (((char __iomem *) dd->ipath_kregbase) +
327                                    dd->ipath_rcvtidbase +
328                                    porttid * sizeof(*tidbase));
329
330         ipath_cdbg(VERBOSE, "Port%u %u tids, cursor %u, tidbase %p\n",
331                    pd->port_port, cnt, tid, tidbase);
332
333         /* virtual address of first page in transfer */
334         vaddr = ti->tidvaddr;
335         if (!access_ok(VERIFY_WRITE, (void __user *) vaddr,
336                        cnt * PAGE_SIZE)) {
337                 ipath_dbg("Fail vaddr %p, %u pages, !access_ok\n",
338                           (void *)vaddr, cnt);
339                 ret = -EFAULT;
340                 goto done;
341         }
342         ret = ipath_get_user_pages(vaddr, cnt, pagep);
343         if (ret) {
344                 if (ret == -EBUSY) {
345                         ipath_dbg("Failed to lock addr %p, %u pages "
346                                   "(already locked)\n",
347                                   (void *) vaddr, cnt);
348                         /*
349                          * for now, continue, and see what happens but with
350                          * the new implementation, this should never happen,
351                          * unless perhaps the user has mpin'ed the pages
352                          * themselves (something we need to test)
353                          */
354                         ret = 0;
355                 } else {
356                         dev_info(&dd->pcidev->dev,
357                                  "Failed to lock addr %p, %u pages: "
358                                  "errno %d\n", (void *) vaddr, cnt, -ret);
359                         goto done;
360                 }
361         }
362         for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
363                 for (; ntids--; tid++) {
364                         if (tid == tidcnt)
365                                 tid = 0;
366                         if (!dd->ipath_pageshadow[porttid + tid])
367                                 break;
368                 }
369                 if (ntids < 0) {
370                         /*
371                          * oops, wrapped all the way through their TIDs,
372                          * and didn't have enough free; see comments at
373                          * start of routine
374                          */
375                         ipath_dbg("Not enough free TIDs for %u pages "
376                                   "(index %d), failing\n", cnt, i);
377                         i--;    /* last tidlist[i] not filled in */
378                         ret = -ENOMEM;
379                         break;
380                 }
381                 tidlist[i] = tid + tidoff;
382                 ipath_cdbg(VERBOSE, "Updating idx %u to TID %u, "
383                            "vaddr %lx\n", i, tid + tidoff, vaddr);
384                 /* we "know" system pages and TID pages are same size */
385                 dd->ipath_pageshadow[porttid + tid] = pagep[i];
386                 dd->ipath_physshadow[porttid + tid] = ipath_map_page(
387                         dd->pcidev, pagep[i], 0, PAGE_SIZE,
388                         PCI_DMA_FROMDEVICE);
389                 /*
390                  * don't need atomic or it's overhead
391                  */
392                 __set_bit(tid, tidmap);
393                 physaddr = dd->ipath_physshadow[porttid + tid];
394                 ipath_stats.sps_pagelocks++;
395                 ipath_cdbg(VERBOSE,
396                            "TID %u, vaddr %lx, physaddr %llx pgp %p\n",
397                            tid, vaddr, (unsigned long long) physaddr,
398                            pagep[i]);
399                 dd->ipath_f_put_tid(dd, &tidbase[tid], RCVHQ_RCV_TYPE_EXPECTED,
400                                     physaddr);
401                 /*
402                  * don't check this tid in ipath_portshadow, since we
403                  * just filled it in; start with the next one.
404                  */
405                 tid++;
406         }
407
408         if (ret) {
409                 u32 limit;
410         cleanup:
411                 /* jump here if copy out of updated info failed... */
412                 ipath_dbg("After failure (ret=%d), undo %d of %d entries\n",
413                           -ret, i, cnt);
414                 /* same code that's in ipath_free_tid() */
415                 limit = sizeof(tidmap) * BITS_PER_BYTE;
416                 if (limit > tidcnt)
417                         /* just in case size changes in future */
418                         limit = tidcnt;
419                 tid = find_first_bit((const unsigned long *)tidmap, limit);
420                 for (; tid < limit; tid++) {
421                         if (!test_bit(tid, tidmap))
422                                 continue;
423                         if (dd->ipath_pageshadow[porttid + tid]) {
424                                 ipath_cdbg(VERBOSE, "Freeing TID %u\n",
425                                            tid);
426                                 dd->ipath_f_put_tid(dd, &tidbase[tid],
427                                                     RCVHQ_RCV_TYPE_EXPECTED,
428                                                     dd->ipath_tidinvalid);
429                                 pci_unmap_page(dd->pcidev,
430                                         dd->ipath_physshadow[porttid + tid],
431                                         PAGE_SIZE, PCI_DMA_FROMDEVICE);
432                                 dd->ipath_pageshadow[porttid + tid] = NULL;
433                                 ipath_stats.sps_pageunlocks++;
434                         }
435                 }
436                 ipath_release_user_pages(pagep, cnt);
437         } else {
438                 /*
439                  * Copy the updated array, with ipath_tid's filled in, back
440                  * to user.  Since we did the copy in already, this "should
441                  * never fail" If it does, we have to clean up...
442                  */
443                 if (copy_to_user((void __user *)
444                                  (unsigned long) ti->tidlist,
445                                  tidlist, cnt * sizeof(*tidlist))) {
446                         ret = -EFAULT;
447                         goto cleanup;
448                 }
449                 if (copy_to_user((void __user *) (unsigned long) ti->tidmap,
450                                  tidmap, sizeof tidmap)) {
451                         ret = -EFAULT;
452                         goto cleanup;
453                 }
454                 if (tid == tidcnt)
455                         tid = 0;
456                 if (!pd->port_subport_cnt)
457                         pd->port_tidcursor = tid;
458                 else
459                         tidcursor_fp(fp) = tid;
460         }
461
462 done:
463         if (ret)
464                 ipath_dbg("Failed to map %u TID pages, failing with %d\n",
465                           ti->tidcnt, -ret);
466         return ret;
467 }
468
469 /**
470  * ipath_tid_free - free a port TID
471  * @pd: the port
472  * @subport: the subport
473  * @ti: the TID info
474  *
475  * right now we are unlocking one page at a time, but since
476  * the intended use of this routine is for a single group of
477  * virtually contiguous pages, that should change to improve
478  * performance.  We check that the TID is in range for this port
479  * but otherwise don't check validity; if user has an error and
480  * frees the wrong tid, it's only their own data that can thereby
481  * be corrupted.  We do check that the TID was in use, for sanity
482  * We always use our idea of the saved address, not the address that
483  * they pass in to us.
484  */
485
486 static int ipath_tid_free(struct ipath_portdata *pd, unsigned subport,
487                           const struct ipath_tid_info *ti)
488 {
489         int ret = 0;
490         u32 tid, porttid, cnt, limit, tidcnt;
491         struct ipath_devdata *dd = pd->port_dd;
492         u64 __iomem *tidbase;
493         unsigned long tidmap[8];
494
495         if (!dd->ipath_pageshadow) {
496                 ret = -ENOMEM;
497                 goto done;
498         }
499
500         if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap,
501                            sizeof tidmap)) {
502                 ret = -EFAULT;
503                 goto done;
504         }
505
506         porttid = pd->port_port * dd->ipath_rcvtidcnt;
507         if (!pd->port_subport_cnt)
508                 tidcnt = dd->ipath_rcvtidcnt;
509         else if (!subport) {
510                 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
511                          (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
512                 porttid += dd->ipath_rcvtidcnt - tidcnt;
513         } else {
514                 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
515                 porttid += tidcnt * (subport - 1);
516         }
517         tidbase = (u64 __iomem *) ((char __iomem *)(dd->ipath_kregbase) +
518                                    dd->ipath_rcvtidbase +
519                                    porttid * sizeof(*tidbase));
520
521         limit = sizeof(tidmap) * BITS_PER_BYTE;
522         if (limit > tidcnt)
523                 /* just in case size changes in future */
524                 limit = tidcnt;
525         tid = find_first_bit(tidmap, limit);
526         ipath_cdbg(VERBOSE, "Port%u free %u tids; first bit (max=%d) "
527                    "set is %d, porttid %u\n", pd->port_port, ti->tidcnt,
528                    limit, tid, porttid);
529         for (cnt = 0; tid < limit; tid++) {
530                 /*
531                  * small optimization; if we detect a run of 3 or so without
532                  * any set, use find_first_bit again.  That's mainly to
533                  * accelerate the case where we wrapped, so we have some at
534                  * the beginning, and some at the end, and a big gap
535                  * in the middle.
536                  */
537                 if (!test_bit(tid, tidmap))
538                         continue;
539                 cnt++;
540                 if (dd->ipath_pageshadow[porttid + tid]) {
541                         struct page *p;
542                         p = dd->ipath_pageshadow[porttid + tid];
543                         dd->ipath_pageshadow[porttid + tid] = NULL;
544                         ipath_cdbg(VERBOSE, "PID %u freeing TID %u\n",
545                                    pd->port_pid, tid);
546                         dd->ipath_f_put_tid(dd, &tidbase[tid],
547                                             RCVHQ_RCV_TYPE_EXPECTED,
548                                             dd->ipath_tidinvalid);
549                         pci_unmap_page(dd->pcidev,
550                                 dd->ipath_physshadow[porttid + tid],
551                                 PAGE_SIZE, PCI_DMA_FROMDEVICE);
552                         ipath_release_user_pages(&p, 1);
553                         ipath_stats.sps_pageunlocks++;
554                 } else
555                         ipath_dbg("Unused tid %u, ignoring\n", tid);
556         }
557         if (cnt != ti->tidcnt)
558                 ipath_dbg("passed in tidcnt %d, only %d bits set in map\n",
559                           ti->tidcnt, cnt);
560 done:
561         if (ret)
562                 ipath_dbg("Failed to unmap %u TID pages, failing with %d\n",
563                           ti->tidcnt, -ret);
564         return ret;
565 }
566
567 /**
568  * ipath_set_part_key - set a partition key
569  * @pd: the port
570  * @key: the key
571  *
572  * We can have up to 4 active at a time (other than the default, which is
573  * always allowed).  This is somewhat tricky, since multiple ports may set
574  * the same key, so we reference count them, and clean up at exit.  All 4
575  * partition keys are packed into a single infinipath register.  It's an
576  * error for a process to set the same pkey multiple times.  We provide no
577  * mechanism to de-allocate a pkey at this time, we may eventually need to
578  * do that.  I've used the atomic operations, and no locking, and only make
579  * a single pass through what's available.  This should be more than
580  * adequate for some time. I'll think about spinlocks or the like if and as
581  * it's necessary.
582  */
583 static int ipath_set_part_key(struct ipath_portdata *pd, u16 key)
584 {
585         struct ipath_devdata *dd = pd->port_dd;
586         int i, any = 0, pidx = -1;
587         u16 lkey = key & 0x7FFF;
588         int ret;
589
590         if (lkey == (IPATH_DEFAULT_P_KEY & 0x7FFF)) {
591                 /* nothing to do; this key always valid */
592                 ret = 0;
593                 goto bail;
594         }
595
596         ipath_cdbg(VERBOSE, "p%u try to set pkey %hx, current keys "
597                    "%hx:%x %hx:%x %hx:%x %hx:%x\n",
598                    pd->port_port, key, dd->ipath_pkeys[0],
599                    atomic_read(&dd->ipath_pkeyrefs[0]), dd->ipath_pkeys[1],
600                    atomic_read(&dd->ipath_pkeyrefs[1]), dd->ipath_pkeys[2],
601                    atomic_read(&dd->ipath_pkeyrefs[2]), dd->ipath_pkeys[3],
602                    atomic_read(&dd->ipath_pkeyrefs[3]));
603
604         if (!lkey) {
605                 ipath_cdbg(PROC, "p%u tries to set key 0, not allowed\n",
606                            pd->port_port);
607                 ret = -EINVAL;
608                 goto bail;
609         }
610
611         /*
612          * Set the full membership bit, because it has to be
613          * set in the register or the packet, and it seems
614          * cleaner to set in the register than to force all
615          * callers to set it. (see bug 4331)
616          */
617         key |= 0x8000;
618
619         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
620                 if (!pd->port_pkeys[i] && pidx == -1)
621                         pidx = i;
622                 if (pd->port_pkeys[i] == key) {
623                         ipath_cdbg(VERBOSE, "p%u tries to set same pkey "
624                                    "(%x) more than once\n",
625                                    pd->port_port, key);
626                         ret = -EEXIST;
627                         goto bail;
628                 }
629         }
630         if (pidx == -1) {
631                 ipath_dbg("All pkeys for port %u already in use, "
632                           "can't set %x\n", pd->port_port, key);
633                 ret = -EBUSY;
634                 goto bail;
635         }
636         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
637                 if (!dd->ipath_pkeys[i]) {
638                         any++;
639                         continue;
640                 }
641                 if (dd->ipath_pkeys[i] == key) {
642                         atomic_t *pkrefs = &dd->ipath_pkeyrefs[i];
643
644                         if (atomic_inc_return(pkrefs) > 1) {
645                                 pd->port_pkeys[pidx] = key;
646                                 ipath_cdbg(VERBOSE, "p%u set key %x "
647                                            "matches #%d, count now %d\n",
648                                            pd->port_port, key, i,
649                                            atomic_read(pkrefs));
650                                 ret = 0;
651                                 goto bail;
652                         } else {
653                                 /*
654                                  * lost race, decrement count, catch below
655                                  */
656                                 atomic_dec(pkrefs);
657                                 ipath_cdbg(VERBOSE, "Lost race, count was "
658                                            "0, after dec, it's %d\n",
659                                            atomic_read(pkrefs));
660                                 any++;
661                         }
662                 }
663                 if ((dd->ipath_pkeys[i] & 0x7FFF) == lkey) {
664                         /*
665                          * It makes no sense to have both the limited and
666                          * full membership PKEY set at the same time since
667                          * the unlimited one will disable the limited one.
668                          */
669                         ret = -EEXIST;
670                         goto bail;
671                 }
672         }
673         if (!any) {
674                 ipath_dbg("port %u, all pkeys already in use, "
675                           "can't set %x\n", pd->port_port, key);
676                 ret = -EBUSY;
677                 goto bail;
678         }
679         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
680                 if (!dd->ipath_pkeys[i] &&
681                     atomic_inc_return(&dd->ipath_pkeyrefs[i]) == 1) {
682                         u64 pkey;
683
684                         /* for ipathstats, etc. */
685                         ipath_stats.sps_pkeys[i] = lkey;
686                         pd->port_pkeys[pidx] = dd->ipath_pkeys[i] = key;
687                         pkey =
688                                 (u64) dd->ipath_pkeys[0] |
689                                 ((u64) dd->ipath_pkeys[1] << 16) |
690                                 ((u64) dd->ipath_pkeys[2] << 32) |
691                                 ((u64) dd->ipath_pkeys[3] << 48);
692                         ipath_cdbg(PROC, "p%u set key %x in #%d, "
693                                    "portidx %d, new pkey reg %llx\n",
694                                    pd->port_port, key, i, pidx,
695                                    (unsigned long long) pkey);
696                         ipath_write_kreg(
697                                 dd, dd->ipath_kregs->kr_partitionkey, pkey);
698
699                         ret = 0;
700                         goto bail;
701                 }
702         }
703         ipath_dbg("port %u, all pkeys already in use 2nd pass, "
704                   "can't set %x\n", pd->port_port, key);
705         ret = -EBUSY;
706
707 bail:
708         return ret;
709 }
710
711 /**
712  * ipath_manage_rcvq - manage a port's receive queue
713  * @pd: the port
714  * @subport: the subport
715  * @start_stop: action to carry out
716  *
717  * start_stop == 0 disables receive on the port, for use in queue
718  * overflow conditions.  start_stop==1 re-enables, to be used to
719  * re-init the software copy of the head register
720  */
721 static int ipath_manage_rcvq(struct ipath_portdata *pd, unsigned subport,
722                              int start_stop)
723 {
724         struct ipath_devdata *dd = pd->port_dd;
725
726         ipath_cdbg(PROC, "%sabling rcv for unit %u port %u:%u\n",
727                    start_stop ? "en" : "dis", dd->ipath_unit,
728                    pd->port_port, subport);
729         if (subport)
730                 goto bail;
731         /* atomically clear receive enable port. */
732         if (start_stop) {
733                 /*
734                  * On enable, force in-memory copy of the tail register to
735                  * 0, so that protocol code doesn't have to worry about
736                  * whether or not the chip has yet updated the in-memory
737                  * copy or not on return from the system call. The chip
738                  * always resets it's tail register back to 0 on a
739                  * transition from disabled to enabled.  This could cause a
740                  * problem if software was broken, and did the enable w/o
741                  * the disable, but eventually the in-memory copy will be
742                  * updated and correct itself, even in the face of software
743                  * bugs.
744                  */
745                 if (pd->port_rcvhdrtail_kvaddr)
746                         ipath_clear_rcvhdrtail(pd);
747                 set_bit(dd->ipath_r_portenable_shift + pd->port_port,
748                         &dd->ipath_rcvctrl);
749         } else
750                 clear_bit(dd->ipath_r_portenable_shift + pd->port_port,
751                           &dd->ipath_rcvctrl);
752         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
753                          dd->ipath_rcvctrl);
754         /* now be sure chip saw it before we return */
755         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
756         if (start_stop) {
757                 /*
758                  * And try to be sure that tail reg update has happened too.
759                  * This should in theory interlock with the RXE changes to
760                  * the tail register.  Don't assign it to the tail register
761                  * in memory copy, since we could overwrite an update by the
762                  * chip if we did.
763                  */
764                 ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
765         }
766         /* always; new head should be equal to new tail; see above */
767 bail:
768         return 0;
769 }
770
771 static void ipath_clean_part_key(struct ipath_portdata *pd,
772                                  struct ipath_devdata *dd)
773 {
774         int i, j, pchanged = 0;
775         u64 oldpkey;
776
777         /* for debugging only */
778         oldpkey = (u64) dd->ipath_pkeys[0] |
779                 ((u64) dd->ipath_pkeys[1] << 16) |
780                 ((u64) dd->ipath_pkeys[2] << 32) |
781                 ((u64) dd->ipath_pkeys[3] << 48);
782
783         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
784                 if (!pd->port_pkeys[i])
785                         continue;
786                 ipath_cdbg(VERBOSE, "look for key[%d] %hx in pkeys\n", i,
787                            pd->port_pkeys[i]);
788                 for (j = 0; j < ARRAY_SIZE(dd->ipath_pkeys); j++) {
789                         /* check for match independent of the global bit */
790                         if ((dd->ipath_pkeys[j] & 0x7fff) !=
791                             (pd->port_pkeys[i] & 0x7fff))
792                                 continue;
793                         if (atomic_dec_and_test(&dd->ipath_pkeyrefs[j])) {
794                                 ipath_cdbg(VERBOSE, "p%u clear key "
795                                            "%x matches #%d\n",
796                                            pd->port_port,
797                                            pd->port_pkeys[i], j);
798                                 ipath_stats.sps_pkeys[j] =
799                                         dd->ipath_pkeys[j] = 0;
800                                 pchanged++;
801                         }
802                         else ipath_cdbg(
803                                 VERBOSE, "p%u key %x matches #%d, "
804                                 "but ref still %d\n", pd->port_port,
805                                 pd->port_pkeys[i], j,
806                                 atomic_read(&dd->ipath_pkeyrefs[j]));
807                         break;
808                 }
809                 pd->port_pkeys[i] = 0;
810         }
811         if (pchanged) {
812                 u64 pkey = (u64) dd->ipath_pkeys[0] |
813                         ((u64) dd->ipath_pkeys[1] << 16) |
814                         ((u64) dd->ipath_pkeys[2] << 32) |
815                         ((u64) dd->ipath_pkeys[3] << 48);
816                 ipath_cdbg(VERBOSE, "p%u old pkey reg %llx, "
817                            "new pkey reg %llx\n", pd->port_port,
818                            (unsigned long long) oldpkey,
819                            (unsigned long long) pkey);
820                 ipath_write_kreg(dd, dd->ipath_kregs->kr_partitionkey,
821                                  pkey);
822         }
823 }
824
825 /*
826  * Initialize the port data with the receive buffer sizes
827  * so this can be done while the master port is locked.
828  * Otherwise, there is a race with a slave opening the port
829  * and seeing these fields uninitialized.
830  */
831 static void init_user_egr_sizes(struct ipath_portdata *pd)
832 {
833         struct ipath_devdata *dd = pd->port_dd;
834         unsigned egrperchunk, egrcnt, size;
835
836         /*
837          * to avoid wasting a lot of memory, we allocate 32KB chunks of
838          * physically contiguous memory, advance through it until used up
839          * and then allocate more.  Of course, we need memory to store those
840          * extra pointers, now.  Started out with 256KB, but under heavy
841          * memory pressure (creating large files and then copying them over
842          * NFS while doing lots of MPI jobs), we hit some allocation
843          * failures, even though we can sleep...  (2.6.10) Still get
844          * failures at 64K.  32K is the lowest we can go without wasting
845          * additional memory.
846          */
847         size = 0x8000;
848         egrperchunk = size / dd->ipath_rcvegrbufsize;
849         egrcnt = dd->ipath_rcvegrcnt;
850         pd->port_rcvegrbuf_chunks = (egrcnt + egrperchunk - 1) / egrperchunk;
851         pd->port_rcvegrbufs_perchunk = egrperchunk;
852         pd->port_rcvegrbuf_size = size;
853 }
854
855 /**
856  * ipath_create_user_egr - allocate eager TID buffers
857  * @pd: the port to allocate TID buffers for
858  *
859  * This routine is now quite different for user and kernel, because
860  * the kernel uses skb's, for the accelerated network performance
861  * This is the user port version
862  *
863  * Allocate the eager TID buffers and program them into infinipath
864  * They are no longer completely contiguous, we do multiple allocation
865  * calls.
866  */
867 static int ipath_create_user_egr(struct ipath_portdata *pd)
868 {
869         struct ipath_devdata *dd = pd->port_dd;
870         unsigned e, egrcnt, egrperchunk, chunk, egrsize, egroff;
871         size_t size;
872         int ret;
873         gfp_t gfp_flags;
874
875         /*
876          * GFP_USER, but without GFP_FS, so buffer cache can be
877          * coalesced (we hope); otherwise, even at order 4,
878          * heavy filesystem activity makes these fail, and we can
879          * use compound pages.
880          */
881         gfp_flags = __GFP_WAIT | __GFP_IO | __GFP_COMP;
882
883         egrcnt = dd->ipath_rcvegrcnt;
884         /* TID number offset for this port */
885         egroff = (pd->port_port - 1) * egrcnt + dd->ipath_p0_rcvegrcnt;
886         egrsize = dd->ipath_rcvegrbufsize;
887         ipath_cdbg(VERBOSE, "Allocating %d egr buffers, at egrtid "
888                    "offset %x, egrsize %u\n", egrcnt, egroff, egrsize);
889
890         chunk = pd->port_rcvegrbuf_chunks;
891         egrperchunk = pd->port_rcvegrbufs_perchunk;
892         size = pd->port_rcvegrbuf_size;
893         pd->port_rcvegrbuf = kmalloc(chunk * sizeof(pd->port_rcvegrbuf[0]),
894                                      GFP_KERNEL);
895         if (!pd->port_rcvegrbuf) {
896                 ret = -ENOMEM;
897                 goto bail;
898         }
899         pd->port_rcvegrbuf_phys =
900                 kmalloc(chunk * sizeof(pd->port_rcvegrbuf_phys[0]),
901                         GFP_KERNEL);
902         if (!pd->port_rcvegrbuf_phys) {
903                 ret = -ENOMEM;
904                 goto bail_rcvegrbuf;
905         }
906         for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
907
908                 pd->port_rcvegrbuf[e] = dma_alloc_coherent(
909                         &dd->pcidev->dev, size, &pd->port_rcvegrbuf_phys[e],
910                         gfp_flags);
911
912                 if (!pd->port_rcvegrbuf[e]) {
913                         ret = -ENOMEM;
914                         goto bail_rcvegrbuf_phys;
915                 }
916         }
917
918         pd->port_rcvegr_phys = pd->port_rcvegrbuf_phys[0];
919
920         for (e = chunk = 0; chunk < pd->port_rcvegrbuf_chunks; chunk++) {
921                 dma_addr_t pa = pd->port_rcvegrbuf_phys[chunk];
922                 unsigned i;
923
924                 for (i = 0; e < egrcnt && i < egrperchunk; e++, i++) {
925                         dd->ipath_f_put_tid(dd, e + egroff +
926                                             (u64 __iomem *)
927                                             ((char __iomem *)
928                                              dd->ipath_kregbase +
929                                              dd->ipath_rcvegrbase),
930                                             RCVHQ_RCV_TYPE_EAGER, pa);
931                         pa += egrsize;
932                 }
933                 cond_resched(); /* don't hog the cpu */
934         }
935
936         ret = 0;
937         goto bail;
938
939 bail_rcvegrbuf_phys:
940         for (e = 0; e < pd->port_rcvegrbuf_chunks &&
941                 pd->port_rcvegrbuf[e]; e++) {
942                 dma_free_coherent(&dd->pcidev->dev, size,
943                                   pd->port_rcvegrbuf[e],
944                                   pd->port_rcvegrbuf_phys[e]);
945
946         }
947         kfree(pd->port_rcvegrbuf_phys);
948         pd->port_rcvegrbuf_phys = NULL;
949 bail_rcvegrbuf:
950         kfree(pd->port_rcvegrbuf);
951         pd->port_rcvegrbuf = NULL;
952 bail:
953         return ret;
954 }
955
956
957 /* common code for the mappings on dma_alloc_coherent mem */
958 static int ipath_mmap_mem(struct vm_area_struct *vma,
959         struct ipath_portdata *pd, unsigned len, int write_ok,
960         void *kvaddr, char *what)
961 {
962         struct ipath_devdata *dd = pd->port_dd;
963         unsigned long pfn;
964         int ret;
965
966         if ((vma->vm_end - vma->vm_start) > len) {
967                 dev_info(&dd->pcidev->dev,
968                          "FAIL on %s: len %lx > %x\n", what,
969                          vma->vm_end - vma->vm_start, len);
970                 ret = -EFAULT;
971                 goto bail;
972         }
973
974         if (!write_ok) {
975                 if (vma->vm_flags & VM_WRITE) {
976                         dev_info(&dd->pcidev->dev,
977                                  "%s must be mapped readonly\n", what);
978                         ret = -EPERM;
979                         goto bail;
980                 }
981
982                 /* don't allow them to later change with mprotect */
983                 vma->vm_flags &= ~VM_MAYWRITE;
984         }
985
986         pfn = virt_to_phys(kvaddr) >> PAGE_SHIFT;
987         ret = remap_pfn_range(vma, vma->vm_start, pfn,
988                               len, vma->vm_page_prot);
989         if (ret)
990                 dev_info(&dd->pcidev->dev, "%s port%u mmap of %lx, %x "
991                          "bytes r%c failed: %d\n", what, pd->port_port,
992                          pfn, len, write_ok?'w':'o', ret);
993         else
994                 ipath_cdbg(VERBOSE, "%s port%u mmaped %lx, %x bytes "
995                            "r%c\n", what, pd->port_port, pfn, len,
996                            write_ok?'w':'o');
997 bail:
998         return ret;
999 }
1000
1001 static int mmap_ureg(struct vm_area_struct *vma, struct ipath_devdata *dd,
1002                      u64 ureg)
1003 {
1004         unsigned long phys;
1005         int ret;
1006
1007         /*
1008          * This is real hardware, so use io_remap.  This is the mechanism
1009          * for the user process to update the head registers for their port
1010          * in the chip.
1011          */
1012         if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) {
1013                 dev_info(&dd->pcidev->dev, "FAIL mmap userreg: reqlen "
1014                          "%lx > PAGE\n", vma->vm_end - vma->vm_start);
1015                 ret = -EFAULT;
1016         } else {
1017                 phys = dd->ipath_physaddr + ureg;
1018                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1019
1020                 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1021                 ret = io_remap_pfn_range(vma, vma->vm_start,
1022                                          phys >> PAGE_SHIFT,
1023                                          vma->vm_end - vma->vm_start,
1024                                          vma->vm_page_prot);
1025         }
1026         return ret;
1027 }
1028
1029 static int mmap_piobufs(struct vm_area_struct *vma,
1030                         struct ipath_devdata *dd,
1031                         struct ipath_portdata *pd,
1032                         unsigned piobufs, unsigned piocnt)
1033 {
1034         unsigned long phys;
1035         int ret;
1036
1037         /*
1038          * When we map the PIO buffers in the chip, we want to map them as
1039          * writeonly, no read possible.   This prevents access to previous
1040          * process data, and catches users who might try to read the i/o
1041          * space due to a bug.
1042          */
1043         if ((vma->vm_end - vma->vm_start) > (piocnt * dd->ipath_palign)) {
1044                 dev_info(&dd->pcidev->dev, "FAIL mmap piobufs: "
1045                          "reqlen %lx > PAGE\n",
1046                          vma->vm_end - vma->vm_start);
1047                 ret = -EINVAL;
1048                 goto bail;
1049         }
1050
1051         phys = dd->ipath_physaddr + piobufs;
1052
1053 #if defined(__powerpc__)
1054         /* There isn't a generic way to specify writethrough mappings */
1055         pgprot_val(vma->vm_page_prot) |= _PAGE_NO_CACHE;
1056         pgprot_val(vma->vm_page_prot) |= _PAGE_WRITETHRU;
1057         pgprot_val(vma->vm_page_prot) &= ~_PAGE_GUARDED;
1058 #endif
1059
1060         /*
1061          * don't allow them to later change to readable with mprotect (for when
1062          * not initially mapped readable, as is normally the case)
1063          */
1064         vma->vm_flags &= ~VM_MAYREAD;
1065         vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1066
1067         ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
1068                                  vma->vm_end - vma->vm_start,
1069                                  vma->vm_page_prot);
1070 bail:
1071         return ret;
1072 }
1073
1074 static int mmap_rcvegrbufs(struct vm_area_struct *vma,
1075                            struct ipath_portdata *pd)
1076 {
1077         struct ipath_devdata *dd = pd->port_dd;
1078         unsigned long start, size;
1079         size_t total_size, i;
1080         unsigned long pfn;
1081         int ret;
1082
1083         size = pd->port_rcvegrbuf_size;
1084         total_size = pd->port_rcvegrbuf_chunks * size;
1085         if ((vma->vm_end - vma->vm_start) > total_size) {
1086                 dev_info(&dd->pcidev->dev, "FAIL on egr bufs: "
1087                          "reqlen %lx > actual %lx\n",
1088                          vma->vm_end - vma->vm_start,
1089                          (unsigned long) total_size);
1090                 ret = -EINVAL;
1091                 goto bail;
1092         }
1093
1094         if (vma->vm_flags & VM_WRITE) {
1095                 dev_info(&dd->pcidev->dev, "Can't map eager buffers as "
1096                          "writable (flags=%lx)\n", vma->vm_flags);
1097                 ret = -EPERM;
1098                 goto bail;
1099         }
1100         /* don't allow them to later change to writeable with mprotect */
1101         vma->vm_flags &= ~VM_MAYWRITE;
1102
1103         start = vma->vm_start;
1104
1105         for (i = 0; i < pd->port_rcvegrbuf_chunks; i++, start += size) {
1106                 pfn = virt_to_phys(pd->port_rcvegrbuf[i]) >> PAGE_SHIFT;
1107                 ret = remap_pfn_range(vma, start, pfn, size,
1108                                       vma->vm_page_prot);
1109                 if (ret < 0)
1110                         goto bail;
1111         }
1112         ret = 0;
1113
1114 bail:
1115         return ret;
1116 }
1117
1118 /*
1119  * ipath_file_vma_fault - handle a VMA page fault.
1120  */
1121 static int ipath_file_vma_fault(struct vm_area_struct *vma,
1122                                         struct vm_fault *vmf)
1123 {
1124         struct page *page;
1125
1126         page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
1127         if (!page)
1128                 return VM_FAULT_SIGBUS;
1129         get_page(page);
1130         vmf->page = page;
1131
1132         return 0;
1133 }
1134
1135 static struct vm_operations_struct ipath_file_vm_ops = {
1136         .fault = ipath_file_vma_fault,
1137 };
1138
1139 static int mmap_kvaddr(struct vm_area_struct *vma, u64 pgaddr,
1140                        struct ipath_portdata *pd, unsigned subport)
1141 {
1142         unsigned long len;
1143         struct ipath_devdata *dd;
1144         void *addr;
1145         size_t size;
1146         int ret = 0;
1147
1148         /* If the port is not shared, all addresses should be physical */
1149         if (!pd->port_subport_cnt)
1150                 goto bail;
1151
1152         dd = pd->port_dd;
1153         size = pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
1154
1155         /*
1156          * Each process has all the subport uregbase, rcvhdrq, and
1157          * rcvegrbufs mmapped - as an array for all the processes,
1158          * and also separately for this process.
1159          */
1160         if (pgaddr == cvt_kvaddr(pd->subport_uregbase)) {
1161                 addr = pd->subport_uregbase;
1162                 size = PAGE_SIZE * pd->port_subport_cnt;
1163         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base)) {
1164                 addr = pd->subport_rcvhdr_base;
1165                 size = pd->port_rcvhdrq_size * pd->port_subport_cnt;
1166         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf)) {
1167                 addr = pd->subport_rcvegrbuf;
1168                 size *= pd->port_subport_cnt;
1169         } else if (pgaddr == cvt_kvaddr(pd->subport_uregbase +
1170                                         PAGE_SIZE * subport)) {
1171                 addr = pd->subport_uregbase + PAGE_SIZE * subport;
1172                 size = PAGE_SIZE;
1173         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base +
1174                                 pd->port_rcvhdrq_size * subport)) {
1175                 addr = pd->subport_rcvhdr_base +
1176                         pd->port_rcvhdrq_size * subport;
1177                 size = pd->port_rcvhdrq_size;
1178         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf +
1179                                size * subport)) {
1180                 addr = pd->subport_rcvegrbuf + size * subport;
1181                 /* rcvegrbufs are read-only on the slave */
1182                 if (vma->vm_flags & VM_WRITE) {
1183                         dev_info(&dd->pcidev->dev,
1184                                  "Can't map eager buffers as "
1185                                  "writable (flags=%lx)\n", vma->vm_flags);
1186                         ret = -EPERM;
1187                         goto bail;
1188                 }
1189                 /*
1190                  * Don't allow permission to later change to writeable
1191                  * with mprotect.
1192                  */
1193                 vma->vm_flags &= ~VM_MAYWRITE;
1194         } else {
1195                 goto bail;
1196         }
1197         len = vma->vm_end - vma->vm_start;
1198         if (len > size) {
1199                 ipath_cdbg(MM, "FAIL: reqlen %lx > %zx\n", len, size);
1200                 ret = -EINVAL;
1201                 goto bail;
1202         }
1203
1204         vma->vm_pgoff = (unsigned long) addr >> PAGE_SHIFT;
1205         vma->vm_ops = &ipath_file_vm_ops;
1206         vma->vm_flags |= VM_RESERVED | VM_DONTEXPAND;
1207         ret = 1;
1208
1209 bail:
1210         return ret;
1211 }
1212
1213 /**
1214  * ipath_mmap - mmap various structures into user space
1215  * @fp: the file pointer
1216  * @vma: the VM area
1217  *
1218  * We use this to have a shared buffer between the kernel and the user code
1219  * for the rcvhdr queue, egr buffers, and the per-port user regs and pio
1220  * buffers in the chip.  We have the open and close entries so we can bump
1221  * the ref count and keep the driver from being unloaded while still mapped.
1222  */
1223 static int ipath_mmap(struct file *fp, struct vm_area_struct *vma)
1224 {
1225         struct ipath_portdata *pd;
1226         struct ipath_devdata *dd;
1227         u64 pgaddr, ureg;
1228         unsigned piobufs, piocnt;
1229         int ret;
1230
1231         pd = port_fp(fp);
1232         if (!pd) {
1233                 ret = -EINVAL;
1234                 goto bail;
1235         }
1236         dd = pd->port_dd;
1237
1238         /*
1239          * This is the ipath_do_user_init() code, mapping the shared buffers
1240          * into the user process. The address referred to by vm_pgoff is the
1241          * file offset passed via mmap().  For shared ports, this is the
1242          * kernel vmalloc() address of the pages to share with the master.
1243          * For non-shared or master ports, this is a physical address.
1244          * We only do one mmap for each space mapped.
1245          */
1246         pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1247
1248         /*
1249          * Check for 0 in case one of the allocations failed, but user
1250          * called mmap anyway.
1251          */
1252         if (!pgaddr)  {
1253                 ret = -EINVAL;
1254                 goto bail;
1255         }
1256
1257         ipath_cdbg(MM, "pgaddr %llx vm_start=%lx len %lx port %u:%u:%u\n",
1258                    (unsigned long long) pgaddr, vma->vm_start,
1259                    vma->vm_end - vma->vm_start, dd->ipath_unit,
1260                    pd->port_port, subport_fp(fp));
1261
1262         /*
1263          * Physical addresses must fit in 40 bits for our hardware.
1264          * Check for kernel virtual addresses first, anything else must
1265          * match a HW or memory address.
1266          */
1267         ret = mmap_kvaddr(vma, pgaddr, pd, subport_fp(fp));
1268         if (ret) {
1269                 if (ret > 0)
1270                         ret = 0;
1271                 goto bail;
1272         }
1273
1274         ureg = dd->ipath_uregbase + dd->ipath_ureg_align * pd->port_port;
1275         if (!pd->port_subport_cnt) {
1276                 /* port is not shared */
1277                 piocnt = dd->ipath_pbufsport;
1278                 piobufs = pd->port_piobufs;
1279         } else if (!subport_fp(fp)) {
1280                 /* caller is the master */
1281                 piocnt = (dd->ipath_pbufsport / pd->port_subport_cnt) +
1282                          (dd->ipath_pbufsport % pd->port_subport_cnt);
1283                 piobufs = pd->port_piobufs +
1284                         dd->ipath_palign * (dd->ipath_pbufsport - piocnt);
1285         } else {
1286                 unsigned slave = subport_fp(fp) - 1;
1287
1288                 /* caller is a slave */
1289                 piocnt = dd->ipath_pbufsport / pd->port_subport_cnt;
1290                 piobufs = pd->port_piobufs + dd->ipath_palign * piocnt * slave;
1291         }
1292
1293         if (pgaddr == ureg)
1294                 ret = mmap_ureg(vma, dd, ureg);
1295         else if (pgaddr == piobufs)
1296                 ret = mmap_piobufs(vma, dd, pd, piobufs, piocnt);
1297         else if (pgaddr == dd->ipath_pioavailregs_phys)
1298                 /* in-memory copy of pioavail registers */
1299                 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1300                                      (void *) dd->ipath_pioavailregs_dma,
1301                                      "pioavail registers");
1302         else if (pgaddr == pd->port_rcvegr_phys)
1303                 ret = mmap_rcvegrbufs(vma, pd);
1304         else if (pgaddr == (u64) pd->port_rcvhdrq_phys)
1305                 /*
1306                  * The rcvhdrq itself; readonly except on HT (so have
1307                  * to allow writable mapping), multiple pages, contiguous
1308                  * from an i/o perspective.
1309                  */
1310                 ret = ipath_mmap_mem(vma, pd, pd->port_rcvhdrq_size, 1,
1311                                      pd->port_rcvhdrq,
1312                                      "rcvhdrq");
1313         else if (pgaddr == (u64) pd->port_rcvhdrqtailaddr_phys)
1314                 /* in-memory copy of rcvhdrq tail register */
1315                 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1316                                      pd->port_rcvhdrtail_kvaddr,
1317                                      "rcvhdrq tail");
1318         else
1319                 ret = -EINVAL;
1320
1321         vma->vm_private_data = NULL;
1322
1323         if (ret < 0)
1324                 dev_info(&dd->pcidev->dev,
1325                          "Failure %d on off %llx len %lx\n",
1326                          -ret, (unsigned long long)pgaddr,
1327                          vma->vm_end - vma->vm_start);
1328 bail:
1329         return ret;
1330 }
1331
1332 static unsigned ipath_poll_hdrqfull(struct ipath_portdata *pd)
1333 {
1334         unsigned pollflag = 0;
1335
1336         if ((pd->poll_type & IPATH_POLL_TYPE_OVERFLOW) &&
1337             pd->port_hdrqfull != pd->port_hdrqfull_poll) {
1338                 pollflag |= POLLIN | POLLRDNORM;
1339                 pd->port_hdrqfull_poll = pd->port_hdrqfull;
1340         }
1341
1342         return pollflag;
1343 }
1344
1345 static unsigned int ipath_poll_urgent(struct ipath_portdata *pd,
1346                                       struct file *fp,
1347                                       struct poll_table_struct *pt)
1348 {
1349         unsigned pollflag = 0;
1350         struct ipath_devdata *dd;
1351
1352         dd = pd->port_dd;
1353
1354         /* variable access in ipath_poll_hdrqfull() needs this */
1355         rmb();
1356         pollflag = ipath_poll_hdrqfull(pd);
1357
1358         if (pd->port_urgent != pd->port_urgent_poll) {
1359                 pollflag |= POLLIN | POLLRDNORM;
1360                 pd->port_urgent_poll = pd->port_urgent;
1361         }
1362
1363         if (!pollflag) {
1364                 /* this saves a spin_lock/unlock in interrupt handler... */
1365                 set_bit(IPATH_PORT_WAITING_URG, &pd->port_flag);
1366                 /* flush waiting flag so don't miss an event... */
1367                 wmb();
1368                 poll_wait(fp, &pd->port_wait, pt);
1369         }
1370
1371         return pollflag;
1372 }
1373
1374 static unsigned int ipath_poll_next(struct ipath_portdata *pd,
1375                                     struct file *fp,
1376                                     struct poll_table_struct *pt)
1377 {
1378         u32 head;
1379         u32 tail;
1380         unsigned pollflag = 0;
1381         struct ipath_devdata *dd;
1382
1383         dd = pd->port_dd;
1384
1385         /* variable access in ipath_poll_hdrqfull() needs this */
1386         rmb();
1387         pollflag = ipath_poll_hdrqfull(pd);
1388
1389         head = ipath_read_ureg32(dd, ur_rcvhdrhead, pd->port_port);
1390         if (pd->port_rcvhdrtail_kvaddr)
1391                 tail = ipath_get_rcvhdrtail(pd);
1392         else
1393                 tail = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
1394
1395         if (head != tail)
1396                 pollflag |= POLLIN | POLLRDNORM;
1397         else {
1398                 /* this saves a spin_lock/unlock in interrupt handler */
1399                 set_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
1400                 /* flush waiting flag so we don't miss an event */
1401                 wmb();
1402
1403                 set_bit(pd->port_port + dd->ipath_r_intravail_shift,
1404                         &dd->ipath_rcvctrl);
1405
1406                 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1407                                  dd->ipath_rcvctrl);
1408
1409                 if (dd->ipath_rhdrhead_intr_off) /* arm rcv interrupt */
1410                         ipath_write_ureg(dd, ur_rcvhdrhead,
1411                                          dd->ipath_rhdrhead_intr_off | head,
1412                                          pd->port_port);
1413
1414                 poll_wait(fp, &pd->port_wait, pt);
1415         }
1416
1417         return pollflag;
1418 }
1419
1420 static unsigned int ipath_poll(struct file *fp,
1421                                struct poll_table_struct *pt)
1422 {
1423         struct ipath_portdata *pd;
1424         unsigned pollflag;
1425
1426         pd = port_fp(fp);
1427         if (!pd)
1428                 pollflag = 0;
1429         else if (pd->poll_type & IPATH_POLL_TYPE_URGENT)
1430                 pollflag = ipath_poll_urgent(pd, fp, pt);
1431         else
1432                 pollflag = ipath_poll_next(pd, fp, pt);
1433
1434         return pollflag;
1435 }
1436
1437 static int ipath_supports_subports(int user_swmajor, int user_swminor)
1438 {
1439         /* no subport implementation prior to software version 1.3 */
1440         return (user_swmajor > 1) || (user_swminor >= 3);
1441 }
1442
1443 static int ipath_compatible_subports(int user_swmajor, int user_swminor)
1444 {
1445         /* this code is written long-hand for clarity */
1446         if (IPATH_USER_SWMAJOR != user_swmajor) {
1447                 /* no promise of compatibility if major mismatch */
1448                 return 0;
1449         }
1450         if (IPATH_USER_SWMAJOR == 1) {
1451                 switch (IPATH_USER_SWMINOR) {
1452                 case 0:
1453                 case 1:
1454                 case 2:
1455                         /* no subport implementation so cannot be compatible */
1456                         return 0;
1457                 case 3:
1458                         /* 3 is only compatible with itself */
1459                         return user_swminor == 3;
1460                 default:
1461                         /* >= 4 are compatible (or are expected to be) */
1462                         return user_swminor >= 4;
1463                 }
1464         }
1465         /* make no promises yet for future major versions */
1466         return 0;
1467 }
1468
1469 static int init_subports(struct ipath_devdata *dd,
1470                          struct ipath_portdata *pd,
1471                          const struct ipath_user_info *uinfo)
1472 {
1473         int ret = 0;
1474         unsigned num_subports;
1475         size_t size;
1476
1477         /*
1478          * If the user is requesting zero subports,
1479          * skip the subport allocation.
1480          */
1481         if (uinfo->spu_subport_cnt <= 0)
1482                 goto bail;
1483
1484         /* Self-consistency check for ipath_compatible_subports() */
1485         if (ipath_supports_subports(IPATH_USER_SWMAJOR, IPATH_USER_SWMINOR) &&
1486             !ipath_compatible_subports(IPATH_USER_SWMAJOR,
1487                                        IPATH_USER_SWMINOR)) {
1488                 dev_info(&dd->pcidev->dev,
1489                          "Inconsistent ipath_compatible_subports()\n");
1490                 goto bail;
1491         }
1492
1493         /* Check for subport compatibility */
1494         if (!ipath_compatible_subports(uinfo->spu_userversion >> 16,
1495                                        uinfo->spu_userversion & 0xffff)) {
1496                 dev_info(&dd->pcidev->dev,
1497                          "Mismatched user version (%d.%d) and driver "
1498                          "version (%d.%d) while port sharing. Ensure "
1499                          "that driver and library are from the same "
1500                          "release.\n",
1501                          (int) (uinfo->spu_userversion >> 16),
1502                          (int) (uinfo->spu_userversion & 0xffff),
1503                          IPATH_USER_SWMAJOR,
1504                          IPATH_USER_SWMINOR);
1505                 goto bail;
1506         }
1507         if (uinfo->spu_subport_cnt > INFINIPATH_MAX_SUBPORT) {
1508                 ret = -EINVAL;
1509                 goto bail;
1510         }
1511
1512         num_subports = uinfo->spu_subport_cnt;
1513         pd->subport_uregbase = vmalloc(PAGE_SIZE * num_subports);
1514         if (!pd->subport_uregbase) {
1515                 ret = -ENOMEM;
1516                 goto bail;
1517         }
1518         /* Note: pd->port_rcvhdrq_size isn't initialized yet. */
1519         size = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
1520                      sizeof(u32), PAGE_SIZE) * num_subports;
1521         pd->subport_rcvhdr_base = vmalloc(size);
1522         if (!pd->subport_rcvhdr_base) {
1523                 ret = -ENOMEM;
1524                 goto bail_ureg;
1525         }
1526
1527         pd->subport_rcvegrbuf = vmalloc(pd->port_rcvegrbuf_chunks *
1528                                         pd->port_rcvegrbuf_size *
1529                                         num_subports);
1530         if (!pd->subport_rcvegrbuf) {
1531                 ret = -ENOMEM;
1532                 goto bail_rhdr;
1533         }
1534
1535         pd->port_subport_cnt = uinfo->spu_subport_cnt;
1536         pd->port_subport_id = uinfo->spu_subport_id;
1537         pd->active_slaves = 1;
1538         set_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
1539         memset(pd->subport_uregbase, 0, PAGE_SIZE * num_subports);
1540         memset(pd->subport_rcvhdr_base, 0, size);
1541         memset(pd->subport_rcvegrbuf, 0, pd->port_rcvegrbuf_chunks *
1542                                          pd->port_rcvegrbuf_size *
1543                                          num_subports);
1544         goto bail;
1545
1546 bail_rhdr:
1547         vfree(pd->subport_rcvhdr_base);
1548 bail_ureg:
1549         vfree(pd->subport_uregbase);
1550         pd->subport_uregbase = NULL;
1551 bail:
1552         return ret;
1553 }
1554
1555 static int try_alloc_port(struct ipath_devdata *dd, int port,
1556                           struct file *fp,
1557                           const struct ipath_user_info *uinfo)
1558 {
1559         struct ipath_portdata *pd;
1560         int ret;
1561
1562         if (!(pd = dd->ipath_pd[port])) {
1563                 void *ptmp;
1564
1565                 pd = kzalloc(sizeof(struct ipath_portdata), GFP_KERNEL);
1566
1567                 /*
1568                  * Allocate memory for use in ipath_tid_update() just once
1569                  * at open, not per call.  Reduces cost of expected send
1570                  * setup.
1571                  */
1572                 ptmp = kmalloc(dd->ipath_rcvtidcnt * sizeof(u16) +
1573                                dd->ipath_rcvtidcnt * sizeof(struct page **),
1574                                GFP_KERNEL);
1575                 if (!pd || !ptmp) {
1576                         ipath_dev_err(dd, "Unable to allocate portdata "
1577                                       "memory, failing open\n");
1578                         ret = -ENOMEM;
1579                         kfree(pd);
1580                         kfree(ptmp);
1581                         goto bail;
1582                 }
1583                 dd->ipath_pd[port] = pd;
1584                 dd->ipath_pd[port]->port_port = port;
1585                 dd->ipath_pd[port]->port_dd = dd;
1586                 dd->ipath_pd[port]->port_tid_pg_list = ptmp;
1587                 init_waitqueue_head(&dd->ipath_pd[port]->port_wait);
1588         }
1589         if (!pd->port_cnt) {
1590                 pd->userversion = uinfo->spu_userversion;
1591                 init_user_egr_sizes(pd);
1592                 if ((ret = init_subports(dd, pd, uinfo)) != 0)
1593                         goto bail;
1594                 ipath_cdbg(PROC, "%s[%u] opened unit:port %u:%u\n",
1595                            current->comm, current->pid, dd->ipath_unit,
1596                            port);
1597                 pd->port_cnt = 1;
1598                 port_fp(fp) = pd;
1599                 pd->port_pid = current->pid;
1600                 strncpy(pd->port_comm, current->comm, sizeof(pd->port_comm));
1601                 ipath_stats.sps_ports++;
1602                 ret = 0;
1603         } else
1604                 ret = -EBUSY;
1605
1606 bail:
1607         return ret;
1608 }
1609
1610 static inline int usable(struct ipath_devdata *dd)
1611 {
1612         return dd &&
1613                 (dd->ipath_flags & IPATH_PRESENT) &&
1614                 dd->ipath_kregbase &&
1615                 dd->ipath_lid &&
1616                 !(dd->ipath_flags & (IPATH_LINKDOWN | IPATH_DISABLED
1617                                      | IPATH_LINKUNK));
1618 }
1619
1620 static int find_free_port(int unit, struct file *fp,
1621                           const struct ipath_user_info *uinfo)
1622 {
1623         struct ipath_devdata *dd = ipath_lookup(unit);
1624         int ret, i;
1625
1626         if (!dd) {
1627                 ret = -ENODEV;
1628                 goto bail;
1629         }
1630
1631         if (!usable(dd)) {
1632                 ret = -ENETDOWN;
1633                 goto bail;
1634         }
1635
1636         for (i = 1; i < dd->ipath_cfgports; i++) {
1637                 ret = try_alloc_port(dd, i, fp, uinfo);
1638                 if (ret != -EBUSY)
1639                         goto bail;
1640         }
1641         ret = -EBUSY;
1642
1643 bail:
1644         return ret;
1645 }
1646
1647 static int find_best_unit(struct file *fp,
1648                           const struct ipath_user_info *uinfo)
1649 {
1650         int ret = 0, i, prefunit = -1, devmax;
1651         int maxofallports, npresent, nup;
1652         int ndev;
1653
1654         devmax = ipath_count_units(&npresent, &nup, &maxofallports);
1655
1656         /*
1657          * This code is present to allow a knowledgeable person to
1658          * specify the layout of processes to processors before opening
1659          * this driver, and then we'll assign the process to the "closest"
1660          * InfiniPath chip to that processor (we assume reasonable connectivity,
1661          * for now).  This code assumes that if affinity has been set
1662          * before this point, that at most one cpu is set; for now this
1663          * is reasonable.  I check for both cpus_empty() and cpus_full(),
1664          * in case some kernel variant sets none of the bits when no
1665          * affinity is set.  2.6.11 and 12 kernels have all present
1666          * cpus set.  Some day we'll have to fix it up further to handle
1667          * a cpu subset.  This algorithm fails for two HT chips connected
1668          * in tunnel fashion.  Eventually this needs real topology
1669          * information.  There may be some issues with dual core numbering
1670          * as well.  This needs more work prior to release.
1671          */
1672         if (!cpus_empty(current->cpus_allowed) &&
1673             !cpus_full(current->cpus_allowed)) {
1674                 int ncpus = num_online_cpus(), curcpu = -1, nset = 0;
1675                 for (i = 0; i < ncpus; i++)
1676                         if (cpu_isset(i, current->cpus_allowed)) {
1677                                 ipath_cdbg(PROC, "%s[%u] affinity set for "
1678                                            "cpu %d/%d\n", current->comm,
1679                                            current->pid, i, ncpus);
1680                                 curcpu = i;
1681                                 nset++;
1682                         }
1683                 if (curcpu != -1 && nset != ncpus) {
1684                         if (npresent) {
1685                                 prefunit = curcpu / (ncpus / npresent);
1686                                 ipath_cdbg(PROC,"%s[%u] %d chips, %d cpus, "
1687                                           "%d cpus/chip, select unit %d\n",
1688                                           current->comm, current->pid,
1689                                           npresent, ncpus, ncpus / npresent,
1690                                           prefunit);
1691                         }
1692                 }
1693         }
1694
1695         /*
1696          * user ports start at 1, kernel port is 0
1697          * For now, we do round-robin access across all chips
1698          */
1699
1700         if (prefunit != -1)
1701                 devmax = prefunit + 1;
1702 recheck:
1703         for (i = 1; i < maxofallports; i++) {
1704                 for (ndev = prefunit != -1 ? prefunit : 0; ndev < devmax;
1705                      ndev++) {
1706                         struct ipath_devdata *dd = ipath_lookup(ndev);
1707
1708                         if (!usable(dd))
1709                                 continue; /* can't use this unit */
1710                         if (i >= dd->ipath_cfgports)
1711                                 /*
1712                                  * Maxed out on users of this unit. Try
1713                                  * next.
1714                                  */
1715                                 continue;
1716                         ret = try_alloc_port(dd, i, fp, uinfo);
1717                         if (!ret)
1718                                 goto done;
1719                 }
1720         }
1721
1722         if (npresent) {
1723                 if (nup == 0) {
1724                         ret = -ENETDOWN;
1725                         ipath_dbg("No ports available (none initialized "
1726                                   "and ready)\n");
1727                 } else {
1728                         if (prefunit > 0) {
1729                                 /* if started above 0, retry from 0 */
1730                                 ipath_cdbg(PROC,
1731                                            "%s[%u] no ports on prefunit "
1732                                            "%d, clear and re-check\n",
1733                                            current->comm, current->pid,
1734                                            prefunit);
1735                                 devmax = ipath_count_units(NULL, NULL,
1736                                                            NULL);
1737                                 prefunit = -1;
1738                                 goto recheck;
1739                         }
1740                         ret = -EBUSY;
1741                         ipath_dbg("No ports available\n");
1742                 }
1743         } else {
1744                 ret = -ENXIO;
1745                 ipath_dbg("No boards found\n");
1746         }
1747
1748 done:
1749         return ret;
1750 }
1751
1752 static int find_shared_port(struct file *fp,
1753                             const struct ipath_user_info *uinfo)
1754 {
1755         int devmax, ndev, i;
1756         int ret = 0;
1757
1758         devmax = ipath_count_units(NULL, NULL, NULL);
1759
1760         for (ndev = 0; ndev < devmax; ndev++) {
1761                 struct ipath_devdata *dd = ipath_lookup(ndev);
1762
1763                 if (!dd)
1764                         continue;
1765                 for (i = 1; i < dd->ipath_cfgports; i++) {
1766                         struct ipath_portdata *pd = dd->ipath_pd[i];
1767
1768                         /* Skip ports which are not yet open */
1769                         if (!pd || !pd->port_cnt)
1770                                 continue;
1771                         /* Skip port if it doesn't match the requested one */
1772                         if (pd->port_subport_id != uinfo->spu_subport_id)
1773                                 continue;
1774                         /* Verify the sharing process matches the master */
1775                         if (pd->port_subport_cnt != uinfo->spu_subport_cnt ||
1776                             pd->userversion != uinfo->spu_userversion ||
1777                             pd->port_cnt >= pd->port_subport_cnt) {
1778                                 ret = -EINVAL;
1779                                 goto done;
1780                         }
1781                         port_fp(fp) = pd;
1782                         subport_fp(fp) = pd->port_cnt++;
1783                         pd->port_subpid[subport_fp(fp)] = current->pid;
1784                         tidcursor_fp(fp) = 0;
1785                         pd->active_slaves |= 1 << subport_fp(fp);
1786                         ipath_cdbg(PROC,
1787                                    "%s[%u] %u sharing %s[%u] unit:port %u:%u\n",
1788                                    current->comm, current->pid,
1789                                    subport_fp(fp),
1790                                    pd->port_comm, pd->port_pid,
1791                                    dd->ipath_unit, pd->port_port);
1792                         ret = 1;
1793                         goto done;
1794                 }
1795         }
1796
1797 done:
1798         return ret;
1799 }
1800
1801 static int ipath_open(struct inode *in, struct file *fp)
1802 {
1803         /* The real work is performed later in ipath_assign_port() */
1804         fp->private_data = kzalloc(sizeof(struct ipath_filedata), GFP_KERNEL);
1805         return fp->private_data ? 0 : -ENOMEM;
1806 }
1807
1808 /* Get port early, so can set affinity prior to memory allocation */
1809 static int ipath_assign_port(struct file *fp,
1810                               const struct ipath_user_info *uinfo)
1811 {
1812         int ret;
1813         int i_minor;
1814         unsigned swmajor, swminor;
1815
1816         /* Check to be sure we haven't already initialized this file */
1817         if (port_fp(fp)) {
1818                 ret = -EINVAL;
1819                 goto done;
1820         }
1821
1822         /* for now, if major version is different, bail */
1823         swmajor = uinfo->spu_userversion >> 16;
1824         if (swmajor != IPATH_USER_SWMAJOR) {
1825                 ipath_dbg("User major version %d not same as driver "
1826                           "major %d\n", uinfo->spu_userversion >> 16,
1827                           IPATH_USER_SWMAJOR);
1828                 ret = -ENODEV;
1829                 goto done;
1830         }
1831
1832         swminor = uinfo->spu_userversion & 0xffff;
1833         if (swminor != IPATH_USER_SWMINOR)
1834                 ipath_dbg("User minor version %d not same as driver "
1835                           "minor %d\n", swminor, IPATH_USER_SWMINOR);
1836
1837         mutex_lock(&ipath_mutex);
1838
1839         if (ipath_compatible_subports(swmajor, swminor) &&
1840             uinfo->spu_subport_cnt &&
1841             (ret = find_shared_port(fp, uinfo))) {
1842                 mutex_unlock(&ipath_mutex);
1843                 if (ret > 0)
1844                         ret = 0;
1845                 goto done;
1846         }
1847
1848         i_minor = iminor(fp->f_path.dentry->d_inode) - IPATH_USER_MINOR_BASE;
1849         ipath_cdbg(VERBOSE, "open on dev %lx (minor %d)\n",
1850                    (long)fp->f_path.dentry->d_inode->i_rdev, i_minor);
1851
1852         if (i_minor)
1853                 ret = find_free_port(i_minor - 1, fp, uinfo);
1854         else
1855                 ret = find_best_unit(fp, uinfo);
1856
1857         mutex_unlock(&ipath_mutex);
1858
1859 done:
1860         return ret;
1861 }
1862
1863
1864 static int ipath_do_user_init(struct file *fp,
1865                               const struct ipath_user_info *uinfo)
1866 {
1867         int ret;
1868         struct ipath_portdata *pd = port_fp(fp);
1869         struct ipath_devdata *dd;
1870         u32 head32;
1871
1872         /* Subports don't need to initialize anything since master did it. */
1873         if (subport_fp(fp)) {
1874                 ret = wait_event_interruptible(pd->port_wait,
1875                         !test_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag));
1876                 goto done;
1877         }
1878
1879         dd = pd->port_dd;
1880
1881         if (uinfo->spu_rcvhdrsize) {
1882                 ret = ipath_setrcvhdrsize(dd, uinfo->spu_rcvhdrsize);
1883                 if (ret)
1884                         goto done;
1885         }
1886
1887         /* for now we do nothing with rcvhdrcnt: uinfo->spu_rcvhdrcnt */
1888
1889         /* for right now, kernel piobufs are at end, so port 1 is at 0 */
1890         pd->port_piobufs = dd->ipath_piobufbase +
1891                 dd->ipath_pbufsport * (pd->port_port - 1) * dd->ipath_palign;
1892         ipath_cdbg(VERBOSE, "Set base of piobufs for port %u to 0x%x\n",
1893                    pd->port_port, pd->port_piobufs);
1894
1895         /*
1896          * Now allocate the rcvhdr Q and eager TIDs; skip the TID
1897          * array for time being.  If pd->port_port > chip-supported,
1898          * we need to do extra stuff here to handle by handling overflow
1899          * through port 0, someday
1900          */
1901         ret = ipath_create_rcvhdrq(dd, pd);
1902         if (!ret)
1903                 ret = ipath_create_user_egr(pd);
1904         if (ret)
1905                 goto done;
1906
1907         /*
1908          * set the eager head register for this port to the current values
1909          * of the tail pointers, since we don't know if they were
1910          * updated on last use of the port.
1911          */
1912         head32 = ipath_read_ureg32(dd, ur_rcvegrindextail, pd->port_port);
1913         ipath_write_ureg(dd, ur_rcvegrindexhead, head32, pd->port_port);
1914         pd->port_lastrcvhdrqtail = -1;
1915         ipath_cdbg(VERBOSE, "Wrote port%d egrhead %x from tail regs\n",
1916                 pd->port_port, head32);
1917         pd->port_tidcursor = 0; /* start at beginning after open */
1918
1919         /* initialize poll variables... */
1920         pd->port_urgent = 0;
1921         pd->port_urgent_poll = 0;
1922         pd->port_hdrqfull_poll = pd->port_hdrqfull;
1923
1924         /*
1925          * now enable the port; the tail registers will be written to memory
1926          * by the chip as soon as it sees the write to
1927          * dd->ipath_kregs->kr_rcvctrl.  The update only happens on
1928          * transition from 0 to 1, so clear it first, then set it as part of
1929          * enabling the port.  This will (very briefly) affect any other
1930          * open ports, but it shouldn't be long enough to be an issue.
1931          * We explictly set the in-memory copy to 0 beforehand, so we don't
1932          * have to wait to be sure the DMA update has happened.
1933          */
1934         if (pd->port_rcvhdrtail_kvaddr)
1935                 ipath_clear_rcvhdrtail(pd);
1936         set_bit(dd->ipath_r_portenable_shift + pd->port_port,
1937                 &dd->ipath_rcvctrl);
1938         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1939                         dd->ipath_rcvctrl &
1940                         ~(1ULL << dd->ipath_r_tailupd_shift));
1941         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1942                          dd->ipath_rcvctrl);
1943         /* Notify any waiting slaves */
1944         if (pd->port_subport_cnt) {
1945                 clear_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
1946                 wake_up(&pd->port_wait);
1947         }
1948 done:
1949         return ret;
1950 }
1951
1952 /**
1953  * unlock_exptid - unlock any expected TID entries port still had in use
1954  * @pd: port
1955  *
1956  * We don't actually update the chip here, because we do a bulk update
1957  * below, using ipath_f_clear_tids.
1958  */
1959 static void unlock_expected_tids(struct ipath_portdata *pd)
1960 {
1961         struct ipath_devdata *dd = pd->port_dd;
1962         int port_tidbase = pd->port_port * dd->ipath_rcvtidcnt;
1963         int i, cnt = 0, maxtid = port_tidbase + dd->ipath_rcvtidcnt;
1964
1965         ipath_cdbg(VERBOSE, "Port %u unlocking any locked expTID pages\n",
1966                    pd->port_port);
1967         for (i = port_tidbase; i < maxtid; i++) {
1968                 if (!dd->ipath_pageshadow[i])
1969                         continue;
1970
1971                 pci_unmap_page(dd->pcidev, dd->ipath_physshadow[i],
1972                         PAGE_SIZE, PCI_DMA_FROMDEVICE);
1973                 ipath_release_user_pages_on_close(&dd->ipath_pageshadow[i],
1974                                                   1);
1975                 dd->ipath_pageshadow[i] = NULL;
1976                 cnt++;
1977                 ipath_stats.sps_pageunlocks++;
1978         }
1979         if (cnt)
1980                 ipath_cdbg(VERBOSE, "Port %u locked %u expTID entries\n",
1981                            pd->port_port, cnt);
1982
1983         if (ipath_stats.sps_pagelocks || ipath_stats.sps_pageunlocks)
1984                 ipath_cdbg(VERBOSE, "%llu pages locked, %llu unlocked\n",
1985                            (unsigned long long) ipath_stats.sps_pagelocks,
1986                            (unsigned long long)
1987                            ipath_stats.sps_pageunlocks);
1988 }
1989
1990 static int ipath_close(struct inode *in, struct file *fp)
1991 {
1992         int ret = 0;
1993         struct ipath_filedata *fd;
1994         struct ipath_portdata *pd;
1995         struct ipath_devdata *dd;
1996         unsigned port;
1997
1998         ipath_cdbg(VERBOSE, "close on dev %lx, private data %p\n",
1999                    (long)in->i_rdev, fp->private_data);
2000
2001         mutex_lock(&ipath_mutex);
2002
2003         fd = (struct ipath_filedata *) fp->private_data;
2004         fp->private_data = NULL;
2005         pd = fd->pd;
2006         if (!pd) {
2007                 mutex_unlock(&ipath_mutex);
2008                 goto bail;
2009         }
2010         if (--pd->port_cnt) {
2011                 /*
2012                  * XXX If the master closes the port before the slave(s),
2013                  * revoke the mmap for the eager receive queue so
2014                  * the slave(s) don't wait for receive data forever.
2015                  */
2016                 pd->active_slaves &= ~(1 << fd->subport);
2017                 pd->port_subpid[fd->subport] = 0;
2018                 mutex_unlock(&ipath_mutex);
2019                 goto bail;
2020         }
2021         port = pd->port_port;
2022         dd = pd->port_dd;
2023
2024         if (pd->port_hdrqfull) {
2025                 ipath_cdbg(PROC, "%s[%u] had %u rcvhdrqfull errors "
2026                            "during run\n", pd->port_comm, pd->port_pid,
2027                            pd->port_hdrqfull);
2028                 pd->port_hdrqfull = 0;
2029         }
2030
2031         if (pd->port_rcvwait_to || pd->port_piowait_to
2032             || pd->port_rcvnowait || pd->port_pionowait) {
2033                 ipath_cdbg(VERBOSE, "port%u, %u rcv, %u pio wait timeo; "
2034                            "%u rcv %u, pio already\n",
2035                            pd->port_port, pd->port_rcvwait_to,
2036                            pd->port_piowait_to, pd->port_rcvnowait,
2037                            pd->port_pionowait);
2038                 pd->port_rcvwait_to = pd->port_piowait_to =
2039                         pd->port_rcvnowait = pd->port_pionowait = 0;
2040         }
2041         if (pd->port_flag) {
2042                 ipath_dbg("port %u port_flag still set to 0x%lx\n",
2043                           pd->port_port, pd->port_flag);
2044                 pd->port_flag = 0;
2045         }
2046
2047         if (dd->ipath_kregbase) {
2048                 int i;
2049                 /* atomically clear receive enable port and intr avail. */
2050                 clear_bit(dd->ipath_r_portenable_shift + port,
2051                           &dd->ipath_rcvctrl);
2052                 clear_bit(pd->port_port + dd->ipath_r_intravail_shift,
2053                           &dd->ipath_rcvctrl);
2054                 ipath_write_kreg( dd, dd->ipath_kregs->kr_rcvctrl,
2055                         dd->ipath_rcvctrl);
2056                 /* and read back from chip to be sure that nothing
2057                  * else is in flight when we do the rest */
2058                 (void)ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
2059
2060                 /* clean up the pkeys for this port user */
2061                 ipath_clean_part_key(pd, dd);
2062                 /*
2063                  * be paranoid, and never write 0's to these, just use an
2064                  * unused part of the port 0 tail page.  Of course,
2065                  * rcvhdraddr points to a large chunk of memory, so this
2066                  * could still trash things, but at least it won't trash
2067                  * page 0, and by disabling the port, it should stop "soon",
2068                  * even if a packet or two is in already in flight after we
2069                  * disabled the port.
2070                  */
2071                 ipath_write_kreg_port(dd,
2072                         dd->ipath_kregs->kr_rcvhdrtailaddr, port,
2073                         dd->ipath_dummy_hdrq_phys);
2074                 ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdraddr,
2075                         pd->port_port, dd->ipath_dummy_hdrq_phys);
2076
2077                 i = dd->ipath_pbufsport * (port - 1);
2078                 ipath_disarm_piobufs(dd, i, dd->ipath_pbufsport);
2079
2080                 dd->ipath_f_clear_tids(dd, pd->port_port);
2081
2082                 if (dd->ipath_pageshadow)
2083                         unlock_expected_tids(pd);
2084                 ipath_stats.sps_ports--;
2085                 ipath_cdbg(PROC, "%s[%u] closed port %u:%u\n",
2086                            pd->port_comm, pd->port_pid,
2087                            dd->ipath_unit, port);
2088         }
2089
2090         pd->port_pid = 0;
2091         dd->ipath_pd[pd->port_port] = NULL; /* before releasing mutex */
2092         mutex_unlock(&ipath_mutex);
2093         ipath_free_pddata(dd, pd); /* after releasing the mutex */
2094
2095 bail:
2096         kfree(fd);
2097         return ret;
2098 }
2099
2100 static int ipath_port_info(struct ipath_portdata *pd, u16 subport,
2101                            struct ipath_port_info __user *uinfo)
2102 {
2103         struct ipath_port_info info;
2104         int nup;
2105         int ret;
2106         size_t sz;
2107
2108         (void) ipath_count_units(NULL, &nup, NULL);
2109         info.num_active = nup;
2110         info.unit = pd->port_dd->ipath_unit;
2111         info.port = pd->port_port;
2112         info.subport = subport;
2113         /* Don't return new fields if old library opened the port. */
2114         if (ipath_supports_subports(pd->userversion >> 16,
2115                                     pd->userversion & 0xffff)) {
2116                 /* Number of user ports available for this device. */
2117                 info.num_ports = pd->port_dd->ipath_cfgports - 1;
2118                 info.num_subports = pd->port_subport_cnt;
2119                 sz = sizeof(info);
2120         } else
2121                 sz = sizeof(info) - 2 * sizeof(u16);
2122
2123         if (copy_to_user(uinfo, &info, sz)) {
2124                 ret = -EFAULT;
2125                 goto bail;
2126         }
2127         ret = 0;
2128
2129 bail:
2130         return ret;
2131 }
2132
2133 static int ipath_get_slave_info(struct ipath_portdata *pd,
2134                                 void __user *slave_mask_addr)
2135 {
2136         int ret = 0;
2137
2138         if (copy_to_user(slave_mask_addr, &pd->active_slaves, sizeof(u32)))
2139                 ret = -EFAULT;
2140         return ret;
2141 }
2142
2143 static int ipath_force_pio_avail_update(struct ipath_devdata *dd)
2144 {
2145         unsigned long flags;
2146
2147         spin_lock_irqsave(&dd->ipath_sendctrl_lock, flags);
2148         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl,
2149                 dd->ipath_sendctrl & ~INFINIPATH_S_PIOBUFAVAILUPD);
2150         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
2151         ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, dd->ipath_sendctrl);
2152         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
2153         spin_unlock_irqrestore(&dd->ipath_sendctrl_lock, flags);
2154
2155         return 0;
2156 }
2157
2158 static ssize_t ipath_write(struct file *fp, const char __user *data,
2159                            size_t count, loff_t *off)
2160 {
2161         const struct ipath_cmd __user *ucmd;
2162         struct ipath_portdata *pd;
2163         const void __user *src;
2164         size_t consumed, copy;
2165         struct ipath_cmd cmd;
2166         ssize_t ret = 0;
2167         void *dest;
2168
2169         if (count < sizeof(cmd.type)) {
2170                 ret = -EINVAL;
2171                 goto bail;
2172         }
2173
2174         ucmd = (const struct ipath_cmd __user *) data;
2175
2176         if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) {
2177                 ret = -EFAULT;
2178                 goto bail;
2179         }
2180
2181         consumed = sizeof(cmd.type);
2182
2183         switch (cmd.type) {
2184         case IPATH_CMD_ASSIGN_PORT:
2185         case __IPATH_CMD_USER_INIT:
2186         case IPATH_CMD_USER_INIT:
2187                 copy = sizeof(cmd.cmd.user_info);
2188                 dest = &cmd.cmd.user_info;
2189                 src = &ucmd->cmd.user_info;
2190                 break;
2191         case IPATH_CMD_RECV_CTRL:
2192                 copy = sizeof(cmd.cmd.recv_ctrl);
2193                 dest = &cmd.cmd.recv_ctrl;
2194                 src = &ucmd->cmd.recv_ctrl;
2195                 break;
2196         case IPATH_CMD_PORT_INFO:
2197                 copy = sizeof(cmd.cmd.port_info);
2198                 dest = &cmd.cmd.port_info;
2199                 src = &ucmd->cmd.port_info;
2200                 break;
2201         case IPATH_CMD_TID_UPDATE:
2202         case IPATH_CMD_TID_FREE:
2203                 copy = sizeof(cmd.cmd.tid_info);
2204                 dest = &cmd.cmd.tid_info;
2205                 src = &ucmd->cmd.tid_info;
2206                 break;
2207         case IPATH_CMD_SET_PART_KEY:
2208                 copy = sizeof(cmd.cmd.part_key);
2209                 dest = &cmd.cmd.part_key;
2210                 src = &ucmd->cmd.part_key;
2211                 break;
2212         case __IPATH_CMD_SLAVE_INFO:
2213                 copy = sizeof(cmd.cmd.slave_mask_addr);
2214                 dest = &cmd.cmd.slave_mask_addr;
2215                 src = &ucmd->cmd.slave_mask_addr;
2216                 break;
2217         case IPATH_CMD_PIOAVAILUPD:     // force an update of PIOAvail reg
2218                 copy = 0;
2219                 src = NULL;
2220                 dest = NULL;
2221                 break;
2222         case IPATH_CMD_POLL_TYPE:
2223                 copy = sizeof(cmd.cmd.poll_type);
2224                 dest = &cmd.cmd.poll_type;
2225                 src = &ucmd->cmd.poll_type;
2226                 break;
2227         case IPATH_CMD_ARMLAUNCH_CTRL:
2228                 copy = sizeof(cmd.cmd.armlaunch_ctrl);
2229                 dest = &cmd.cmd.armlaunch_ctrl;
2230                 src = &ucmd->cmd.armlaunch_ctrl;
2231                 break;
2232         default:
2233                 ret = -EINVAL;
2234                 goto bail;
2235         }
2236
2237         if (copy) {
2238                 if ((count - consumed) < copy) {
2239                         ret = -EINVAL;
2240                         goto bail;
2241                 }
2242
2243                 if (copy_from_user(dest, src, copy)) {
2244                         ret = -EFAULT;
2245                         goto bail;
2246                 }
2247
2248                 consumed += copy;
2249         }
2250
2251         pd = port_fp(fp);
2252         if (!pd && cmd.type != __IPATH_CMD_USER_INIT &&
2253                 cmd.type != IPATH_CMD_ASSIGN_PORT) {
2254                 ret = -EINVAL;
2255                 goto bail;
2256         }
2257
2258         switch (cmd.type) {
2259         case IPATH_CMD_ASSIGN_PORT:
2260                 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2261                 if (ret)
2262                         goto bail;
2263                 break;
2264         case __IPATH_CMD_USER_INIT:
2265                 /* backwards compatibility, get port first */
2266                 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2267                 if (ret)
2268                         goto bail;
2269                 /* and fall through to current version. */
2270         case IPATH_CMD_USER_INIT:
2271                 ret = ipath_do_user_init(fp, &cmd.cmd.user_info);
2272                 if (ret)
2273                         goto bail;
2274                 ret = ipath_get_base_info(
2275                         fp, (void __user *) (unsigned long)
2276                         cmd.cmd.user_info.spu_base_info,
2277                         cmd.cmd.user_info.spu_base_info_size);
2278                 break;
2279         case IPATH_CMD_RECV_CTRL:
2280                 ret = ipath_manage_rcvq(pd, subport_fp(fp), cmd.cmd.recv_ctrl);
2281                 break;
2282         case IPATH_CMD_PORT_INFO:
2283                 ret = ipath_port_info(pd, subport_fp(fp),
2284                                       (struct ipath_port_info __user *)
2285                                       (unsigned long) cmd.cmd.port_info);
2286                 break;
2287         case IPATH_CMD_TID_UPDATE:
2288                 ret = ipath_tid_update(pd, fp, &cmd.cmd.tid_info);
2289                 break;
2290         case IPATH_CMD_TID_FREE:
2291                 ret = ipath_tid_free(pd, subport_fp(fp), &cmd.cmd.tid_info);
2292                 break;
2293         case IPATH_CMD_SET_PART_KEY:
2294                 ret = ipath_set_part_key(pd, cmd.cmd.part_key);
2295                 break;
2296         case __IPATH_CMD_SLAVE_INFO:
2297                 ret = ipath_get_slave_info(pd,
2298                                            (void __user *) (unsigned long)
2299                                            cmd.cmd.slave_mask_addr);
2300                 break;
2301         case IPATH_CMD_PIOAVAILUPD:
2302                 ret = ipath_force_pio_avail_update(pd->port_dd);
2303                 break;
2304         case IPATH_CMD_POLL_TYPE:
2305                 pd->poll_type = cmd.cmd.poll_type;
2306                 break;
2307         case IPATH_CMD_ARMLAUNCH_CTRL:
2308                 if (cmd.cmd.armlaunch_ctrl)
2309                         ipath_enable_armlaunch(pd->port_dd);
2310                 else
2311                         ipath_disable_armlaunch(pd->port_dd);
2312                 break;
2313         }
2314
2315         if (ret >= 0)
2316                 ret = consumed;
2317
2318 bail:
2319         return ret;
2320 }
2321
2322 static struct class *ipath_class;
2323
2324 static int init_cdev(int minor, char *name, const struct file_operations *fops,
2325                      struct cdev **cdevp, struct class_device **class_devp)
2326 {
2327         const dev_t dev = MKDEV(IPATH_MAJOR, minor);
2328         struct cdev *cdev = NULL;
2329         struct class_device *class_dev = NULL;
2330         int ret;
2331
2332         cdev = cdev_alloc();
2333         if (!cdev) {
2334                 printk(KERN_ERR IPATH_DRV_NAME
2335                        ": Could not allocate cdev for minor %d, %s\n",
2336                        minor, name);
2337                 ret = -ENOMEM;
2338                 goto done;
2339         }
2340
2341         cdev->owner = THIS_MODULE;
2342         cdev->ops = fops;
2343         kobject_set_name(&cdev->kobj, name);
2344
2345         ret = cdev_add(cdev, dev, 1);
2346         if (ret < 0) {
2347                 printk(KERN_ERR IPATH_DRV_NAME
2348                        ": Could not add cdev for minor %d, %s (err %d)\n",
2349                        minor, name, -ret);
2350                 goto err_cdev;
2351         }
2352
2353         class_dev = class_device_create(ipath_class, NULL, dev, NULL, name);
2354
2355         if (IS_ERR(class_dev)) {
2356                 ret = PTR_ERR(class_dev);
2357                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2358                        "class_dev for minor %d, %s (err %d)\n",
2359                        minor, name, -ret);
2360                 goto err_cdev;
2361         }
2362
2363         goto done;
2364
2365 err_cdev:
2366         cdev_del(cdev);
2367         cdev = NULL;
2368
2369 done:
2370         if (ret >= 0) {
2371                 *cdevp = cdev;
2372                 *class_devp = class_dev;
2373         } else {
2374                 *cdevp = NULL;
2375                 *class_devp = NULL;
2376         }
2377
2378         return ret;
2379 }
2380
2381 int ipath_cdev_init(int minor, char *name, const struct file_operations *fops,
2382                     struct cdev **cdevp, struct class_device **class_devp)
2383 {
2384         return init_cdev(minor, name, fops, cdevp, class_devp);
2385 }
2386
2387 static void cleanup_cdev(struct cdev **cdevp,
2388                          struct class_device **class_devp)
2389 {
2390         struct class_device *class_dev = *class_devp;
2391
2392         if (class_dev) {
2393                 class_device_unregister(class_dev);
2394                 *class_devp = NULL;
2395         }
2396
2397         if (*cdevp) {
2398                 cdev_del(*cdevp);
2399                 *cdevp = NULL;
2400         }
2401 }
2402
2403 void ipath_cdev_cleanup(struct cdev **cdevp,
2404                         struct class_device **class_devp)
2405 {
2406         cleanup_cdev(cdevp, class_devp);
2407 }
2408
2409 static struct cdev *wildcard_cdev;
2410 static struct class_device *wildcard_class_dev;
2411
2412 static const dev_t dev = MKDEV(IPATH_MAJOR, 0);
2413
2414 static int user_init(void)
2415 {
2416         int ret;
2417
2418         ret = register_chrdev_region(dev, IPATH_NMINORS, IPATH_DRV_NAME);
2419         if (ret < 0) {
2420                 printk(KERN_ERR IPATH_DRV_NAME ": Could not register "
2421                        "chrdev region (err %d)\n", -ret);
2422                 goto done;
2423         }
2424
2425         ipath_class = class_create(THIS_MODULE, IPATH_DRV_NAME);
2426
2427         if (IS_ERR(ipath_class)) {
2428                 ret = PTR_ERR(ipath_class);
2429                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2430                        "device class (err %d)\n", -ret);
2431                 goto bail;
2432         }
2433
2434         goto done;
2435 bail:
2436         unregister_chrdev_region(dev, IPATH_NMINORS);
2437 done:
2438         return ret;
2439 }
2440
2441 static void user_cleanup(void)
2442 {
2443         if (ipath_class) {
2444                 class_destroy(ipath_class);
2445                 ipath_class = NULL;
2446         }
2447
2448         unregister_chrdev_region(dev, IPATH_NMINORS);
2449 }
2450
2451 static atomic_t user_count = ATOMIC_INIT(0);
2452 static atomic_t user_setup = ATOMIC_INIT(0);
2453
2454 int ipath_user_add(struct ipath_devdata *dd)
2455 {
2456         char name[10];
2457         int ret;
2458
2459         if (atomic_inc_return(&user_count) == 1) {
2460                 ret = user_init();
2461                 if (ret < 0) {
2462                         ipath_dev_err(dd, "Unable to set up user support: "
2463                                       "error %d\n", -ret);
2464                         goto bail;
2465                 }
2466                 ret = init_cdev(0, "ipath", &ipath_file_ops, &wildcard_cdev,
2467                                 &wildcard_class_dev);
2468                 if (ret < 0) {
2469                         ipath_dev_err(dd, "Could not create wildcard "
2470                                       "minor: error %d\n", -ret);
2471                         goto bail_user;
2472                 }
2473
2474                 atomic_set(&user_setup, 1);
2475         }
2476
2477         snprintf(name, sizeof(name), "ipath%d", dd->ipath_unit);
2478
2479         ret = init_cdev(dd->ipath_unit + 1, name, &ipath_file_ops,
2480                         &dd->user_cdev, &dd->user_class_dev);
2481         if (ret < 0)
2482                 ipath_dev_err(dd, "Could not create user minor %d, %s\n",
2483                               dd->ipath_unit + 1, name);
2484
2485         goto bail;
2486
2487 bail_user:
2488         user_cleanup();
2489 bail:
2490         return ret;
2491 }
2492
2493 void ipath_user_remove(struct ipath_devdata *dd)
2494 {
2495         cleanup_cdev(&dd->user_cdev, &dd->user_class_dev);
2496
2497         if (atomic_dec_return(&user_count) == 0) {
2498                 if (atomic_read(&user_setup) == 0)
2499                         goto bail;
2500
2501                 cleanup_cdev(&wildcard_cdev, &wildcard_class_dev);
2502                 user_cleanup();
2503
2504                 atomic_set(&user_setup, 0);
2505         }
2506 bail:
2507         return;
2508 }