Merge git://git.infradead.org/~dwmw2/cafe-2.6
[linux-2.6] / drivers / infiniband / hw / ipath / ipath_file_ops.c
1 /*
2  * Copyright (c) 2006 QLogic, Inc. 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 /*
45  * mmap64 doesn't allow all 64 bits for 32-bit applications
46  * so only use the low 43 bits.
47  */
48 #define MMAP64_MASK     0x7FFFFFFFFFFUL
49
50 static int ipath_open(struct inode *, struct file *);
51 static int ipath_close(struct inode *, struct file *);
52 static ssize_t ipath_write(struct file *, const char __user *, size_t,
53                            loff_t *);
54 static unsigned int ipath_poll(struct file *, struct poll_table_struct *);
55 static int ipath_mmap(struct file *, struct vm_area_struct *);
56
57 static struct file_operations ipath_file_ops = {
58         .owner = THIS_MODULE,
59         .write = ipath_write,
60         .open = ipath_open,
61         .release = ipath_close,
62         .poll = ipath_poll,
63         .mmap = ipath_mmap
64 };
65
66 static int ipath_get_base_info(struct file *fp,
67                                void __user *ubase, size_t ubase_size)
68 {
69         struct ipath_portdata *pd = port_fp(fp);
70         int ret = 0;
71         struct ipath_base_info *kinfo = NULL;
72         struct ipath_devdata *dd = pd->port_dd;
73         unsigned subport_cnt;
74         int shared, master;
75         size_t sz;
76
77         subport_cnt = pd->port_subport_cnt;
78         if (!subport_cnt) {
79                 shared = 0;
80                 master = 0;
81                 subport_cnt = 1;
82         } else {
83                 shared = 1;
84                 master = !subport_fp(fp);
85         }
86
87         sz = sizeof(*kinfo);
88         /* If port sharing is not requested, allow the old size structure */
89         if (!shared)
90                 sz -= 3 * sizeof(u64);
91         if (ubase_size < sz) {
92                 ipath_cdbg(PROC,
93                            "Base size %zu, need %zu (version mismatch?)\n",
94                            ubase_size, sz);
95                 ret = -EINVAL;
96                 goto bail;
97         }
98
99         kinfo = kzalloc(sizeof(*kinfo), GFP_KERNEL);
100         if (kinfo == NULL) {
101                 ret = -ENOMEM;
102                 goto bail;
103         }
104
105         ret = dd->ipath_f_get_base_info(pd, kinfo);
106         if (ret < 0)
107                 goto bail;
108
109         kinfo->spi_rcvhdr_cnt = dd->ipath_rcvhdrcnt;
110         kinfo->spi_rcvhdrent_size = dd->ipath_rcvhdrentsize;
111         kinfo->spi_tidegrcnt = dd->ipath_rcvegrcnt;
112         kinfo->spi_rcv_egrbufsize = dd->ipath_rcvegrbufsize;
113         /*
114          * have to mmap whole thing
115          */
116         kinfo->spi_rcv_egrbuftotlen =
117                 pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
118         kinfo->spi_rcv_egrperchunk = pd->port_rcvegrbufs_perchunk;
119         kinfo->spi_rcv_egrchunksize = kinfo->spi_rcv_egrbuftotlen /
120                 pd->port_rcvegrbuf_chunks;
121         kinfo->spi_tidcnt = dd->ipath_rcvtidcnt / subport_cnt;
122         if (master)
123                 kinfo->spi_tidcnt += dd->ipath_rcvtidcnt % subport_cnt;
124         /*
125          * for this use, may be ipath_cfgports summed over all chips that
126          * are are configured and present
127          */
128         kinfo->spi_nports = dd->ipath_cfgports;
129         /* unit (chip/board) our port is on */
130         kinfo->spi_unit = dd->ipath_unit;
131         /* for now, only a single page */
132         kinfo->spi_tid_maxsize = PAGE_SIZE;
133
134         /*
135          * Doing this per port, and based on the skip value, etc.  This has
136          * to be the actual buffer size, since the protocol code treats it
137          * as an array.
138          *
139          * These have to be set to user addresses in the user code via mmap.
140          * These values are used on return to user code for the mmap target
141          * addresses only.  For 32 bit, same 44 bit address problem, so use
142          * the physical address, not virtual.  Before 2.6.11, using the
143          * page_address() macro worked, but in 2.6.11, even that returns the
144          * full 64 bit address (upper bits all 1's).  So far, using the
145          * physical addresses (or chip offsets, for chip mapping) works, but
146          * no doubt some future kernel release will change that, and we'll be
147          * on to yet another method of dealing with this.
148          */
149         kinfo->spi_rcvhdr_base = (u64) pd->port_rcvhdrq_phys;
150         kinfo->spi_rcvhdr_tailaddr = (u64) pd->port_rcvhdrqtailaddr_phys;
151         kinfo->spi_rcv_egrbufs = (u64) pd->port_rcvegr_phys;
152         kinfo->spi_pioavailaddr = (u64) dd->ipath_pioavailregs_phys;
153         kinfo->spi_status = (u64) kinfo->spi_pioavailaddr +
154                 (void *) dd->ipath_statusp -
155                 (void *) dd->ipath_pioavailregs_dma;
156         if (!shared) {
157                 kinfo->spi_piocnt = dd->ipath_pbufsport;
158                 kinfo->spi_piobufbase = (u64) pd->port_piobufs;
159                 kinfo->__spi_uregbase = (u64) dd->ipath_uregbase +
160                         dd->ipath_palign * pd->port_port;
161         } else if (master) {
162                 kinfo->spi_piocnt = (dd->ipath_pbufsport / subport_cnt) +
163                                     (dd->ipath_pbufsport % subport_cnt);
164                 /* Master's PIO buffers are after all the slave's */
165                 kinfo->spi_piobufbase = (u64) pd->port_piobufs +
166                         dd->ipath_palign *
167                         (dd->ipath_pbufsport - kinfo->spi_piocnt);
168                 kinfo->__spi_uregbase = (u64) dd->ipath_uregbase +
169                         dd->ipath_palign * pd->port_port;
170         } else {
171                 unsigned slave = subport_fp(fp) - 1;
172
173                 kinfo->spi_piocnt = dd->ipath_pbufsport / subport_cnt;
174                 kinfo->spi_piobufbase = (u64) pd->port_piobufs +
175                         dd->ipath_palign * kinfo->spi_piocnt * slave;
176                 kinfo->__spi_uregbase = ((u64) pd->subport_uregbase +
177                         PAGE_SIZE * slave) & MMAP64_MASK;
178
179                 kinfo->spi_rcvhdr_base = ((u64) pd->subport_rcvhdr_base +
180                         pd->port_rcvhdrq_size * slave) & MMAP64_MASK;
181                 kinfo->spi_rcvhdr_tailaddr =
182                         (u64) pd->port_rcvhdrqtailaddr_phys & MMAP64_MASK;
183                 kinfo->spi_rcv_egrbufs = ((u64) pd->subport_rcvegrbuf +
184                         dd->ipath_rcvegrcnt * dd->ipath_rcvegrbufsize * slave) &
185                         MMAP64_MASK;
186         }
187
188         kinfo->spi_pioindex = (kinfo->spi_piobufbase - dd->ipath_piobufbase) /
189                 dd->ipath_palign;
190         kinfo->spi_pioalign = dd->ipath_palign;
191
192         kinfo->spi_qpair = IPATH_KD_QP;
193         kinfo->spi_piosize = dd->ipath_ibmaxlen;
194         kinfo->spi_mtu = dd->ipath_ibmaxlen;    /* maxlen, not ibmtu */
195         kinfo->spi_port = pd->port_port;
196         kinfo->spi_subport = subport_fp(fp);
197         kinfo->spi_sw_version = IPATH_KERN_SWVERSION;
198         kinfo->spi_hw_version = dd->ipath_revision;
199
200         if (master) {
201                 kinfo->spi_runtime_flags |= IPATH_RUNTIME_MASTER;
202                 kinfo->spi_subport_uregbase =
203                         (u64) pd->subport_uregbase & MMAP64_MASK;
204                 kinfo->spi_subport_rcvegrbuf =
205                         (u64) pd->subport_rcvegrbuf & MMAP64_MASK;
206                 kinfo->spi_subport_rcvhdr_base =
207                         (u64) pd->subport_rcvhdr_base & MMAP64_MASK;
208                 ipath_cdbg(PROC, "port %u flags %x %llx %llx %llx\n",
209                         kinfo->spi_port, kinfo->spi_runtime_flags,
210                         (unsigned long long) kinfo->spi_subport_uregbase,
211                         (unsigned long long) kinfo->spi_subport_rcvegrbuf,
212                         (unsigned long long) kinfo->spi_subport_rcvhdr_base);
213         }
214
215         if (copy_to_user(ubase, kinfo, sizeof(*kinfo)))
216                 ret = -EFAULT;
217
218 bail:
219         kfree(kinfo);
220         return ret;
221 }
222
223 /**
224  * ipath_tid_update - update a port TID
225  * @pd: the port
226  * @fp: the ipath device file
227  * @ti: the TID information
228  *
229  * The new implementation as of Oct 2004 is that the driver assigns
230  * the tid and returns it to the caller.   To make it easier to
231  * catch bugs, and to reduce search time, we keep a cursor for
232  * each port, walking the shadow tid array to find one that's not
233  * in use.
234  *
235  * For now, if we can't allocate the full list, we fail, although
236  * in the long run, we'll allocate as many as we can, and the
237  * caller will deal with that by trying the remaining pages later.
238  * That means that when we fail, we have to mark the tids as not in
239  * use again, in our shadow copy.
240  *
241  * It's up to the caller to free the tids when they are done.
242  * We'll unlock the pages as they free them.
243  *
244  * Also, right now we are locking one page at a time, but since
245  * the intended use of this routine is for a single group of
246  * virtually contiguous pages, that should change to improve
247  * performance.
248  */
249 static int ipath_tid_update(struct ipath_portdata *pd, struct file *fp,
250                             const struct ipath_tid_info *ti)
251 {
252         int ret = 0, ntids;
253         u32 tid, porttid, cnt, i, tidcnt, tidoff;
254         u16 *tidlist;
255         struct ipath_devdata *dd = pd->port_dd;
256         u64 physaddr;
257         unsigned long vaddr;
258         u64 __iomem *tidbase;
259         unsigned long tidmap[8];
260         struct page **pagep = NULL;
261         unsigned subport = subport_fp(fp);
262
263         if (!dd->ipath_pageshadow) {
264                 ret = -ENOMEM;
265                 goto done;
266         }
267
268         cnt = ti->tidcnt;
269         if (!cnt) {
270                 ipath_dbg("After copyin, tidcnt 0, tidlist %llx\n",
271                           (unsigned long long) ti->tidlist);
272                 /*
273                  * Should we treat as success?  likely a bug
274                  */
275                 ret = -EFAULT;
276                 goto done;
277         }
278         porttid = pd->port_port * dd->ipath_rcvtidcnt;
279         if (!pd->port_subport_cnt) {
280                 tidcnt = dd->ipath_rcvtidcnt;
281                 tid = pd->port_tidcursor;
282                 tidoff = 0;
283         } else if (!subport) {
284                 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
285                          (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
286                 tidoff = dd->ipath_rcvtidcnt - tidcnt;
287                 porttid += tidoff;
288                 tid = tidcursor_fp(fp);
289         } else {
290                 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
291                 tidoff = tidcnt * (subport - 1);
292                 porttid += tidoff;
293                 tid = tidcursor_fp(fp);
294         }
295         if (cnt > tidcnt) {
296                 /* make sure it all fits in port_tid_pg_list */
297                 dev_info(&dd->pcidev->dev, "Process tried to allocate %u "
298                          "TIDs, only trying max (%u)\n", cnt, tidcnt);
299                 cnt = tidcnt;
300         }
301         pagep = &((struct page **) pd->port_tid_pg_list)[tidoff];
302         tidlist = &((u16 *) &pagep[dd->ipath_rcvtidcnt])[tidoff];
303
304         memset(tidmap, 0, sizeof(tidmap));
305         /* before decrement; chip actual # */
306         ntids = tidcnt;
307         tidbase = (u64 __iomem *) (((char __iomem *) dd->ipath_kregbase) +
308                                    dd->ipath_rcvtidbase +
309                                    porttid * sizeof(*tidbase));
310
311         ipath_cdbg(VERBOSE, "Port%u %u tids, cursor %u, tidbase %p\n",
312                    pd->port_port, cnt, tid, tidbase);
313
314         /* virtual address of first page in transfer */
315         vaddr = ti->tidvaddr;
316         if (!access_ok(VERIFY_WRITE, (void __user *) vaddr,
317                        cnt * PAGE_SIZE)) {
318                 ipath_dbg("Fail vaddr %p, %u pages, !access_ok\n",
319                           (void *)vaddr, cnt);
320                 ret = -EFAULT;
321                 goto done;
322         }
323         ret = ipath_get_user_pages(vaddr, cnt, pagep);
324         if (ret) {
325                 if (ret == -EBUSY) {
326                         ipath_dbg("Failed to lock addr %p, %u pages "
327                                   "(already locked)\n",
328                                   (void *) vaddr, cnt);
329                         /*
330                          * for now, continue, and see what happens but with
331                          * the new implementation, this should never happen,
332                          * unless perhaps the user has mpin'ed the pages
333                          * themselves (something we need to test)
334                          */
335                         ret = 0;
336                 } else {
337                         dev_info(&dd->pcidev->dev,
338                                  "Failed to lock addr %p, %u pages: "
339                                  "errno %d\n", (void *) vaddr, cnt, -ret);
340                         goto done;
341                 }
342         }
343         for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
344                 for (; ntids--; tid++) {
345                         if (tid == tidcnt)
346                                 tid = 0;
347                         if (!dd->ipath_pageshadow[porttid + tid])
348                                 break;
349                 }
350                 if (ntids < 0) {
351                         /*
352                          * oops, wrapped all the way through their TIDs,
353                          * and didn't have enough free; see comments at
354                          * start of routine
355                          */
356                         ipath_dbg("Not enough free TIDs for %u pages "
357                                   "(index %d), failing\n", cnt, i);
358                         i--;    /* last tidlist[i] not filled in */
359                         ret = -ENOMEM;
360                         break;
361                 }
362                 tidlist[i] = tid + tidoff;
363                 ipath_cdbg(VERBOSE, "Updating idx %u to TID %u, "
364                            "vaddr %lx\n", i, tid + tidoff, vaddr);
365                 /* we "know" system pages and TID pages are same size */
366                 dd->ipath_pageshadow[porttid + tid] = pagep[i];
367                 dd->ipath_physshadow[porttid + tid] = ipath_map_page(
368                         dd->pcidev, pagep[i], 0, PAGE_SIZE,
369                         PCI_DMA_FROMDEVICE);
370                 /*
371                  * don't need atomic or it's overhead
372                  */
373                 __set_bit(tid, tidmap);
374                 physaddr = dd->ipath_physshadow[porttid + tid];
375                 ipath_stats.sps_pagelocks++;
376                 ipath_cdbg(VERBOSE,
377                            "TID %u, vaddr %lx, physaddr %llx pgp %p\n",
378                            tid, vaddr, (unsigned long long) physaddr,
379                            pagep[i]);
380                 dd->ipath_f_put_tid(dd, &tidbase[tid], 1, physaddr);
381                 /*
382                  * don't check this tid in ipath_portshadow, since we
383                  * just filled it in; start with the next one.
384                  */
385                 tid++;
386         }
387
388         if (ret) {
389                 u32 limit;
390         cleanup:
391                 /* jump here if copy out of updated info failed... */
392                 ipath_dbg("After failure (ret=%d), undo %d of %d entries\n",
393                           -ret, i, cnt);
394                 /* same code that's in ipath_free_tid() */
395                 limit = sizeof(tidmap) * BITS_PER_BYTE;
396                 if (limit > tidcnt)
397                         /* just in case size changes in future */
398                         limit = tidcnt;
399                 tid = find_first_bit((const unsigned long *)tidmap, limit);
400                 for (; tid < limit; tid++) {
401                         if (!test_bit(tid, tidmap))
402                                 continue;
403                         if (dd->ipath_pageshadow[porttid + tid]) {
404                                 ipath_cdbg(VERBOSE, "Freeing TID %u\n",
405                                            tid);
406                                 dd->ipath_f_put_tid(dd, &tidbase[tid], 1,
407                                                     dd->ipath_tidinvalid);
408                                 pci_unmap_page(dd->pcidev,
409                                         dd->ipath_physshadow[porttid + tid],
410                                         PAGE_SIZE, PCI_DMA_FROMDEVICE);
411                                 dd->ipath_pageshadow[porttid + tid] = NULL;
412                                 ipath_stats.sps_pageunlocks++;
413                         }
414                 }
415                 ipath_release_user_pages(pagep, cnt);
416         } else {
417                 /*
418                  * Copy the updated array, with ipath_tid's filled in, back
419                  * to user.  Since we did the copy in already, this "should
420                  * never fail" If it does, we have to clean up...
421                  */
422                 if (copy_to_user((void __user *)
423                                  (unsigned long) ti->tidlist,
424                                  tidlist, cnt * sizeof(*tidlist))) {
425                         ret = -EFAULT;
426                         goto cleanup;
427                 }
428                 if (copy_to_user((void __user *) (unsigned long) ti->tidmap,
429                                  tidmap, sizeof tidmap)) {
430                         ret = -EFAULT;
431                         goto cleanup;
432                 }
433                 if (tid == tidcnt)
434                         tid = 0;
435                 if (!pd->port_subport_cnt)
436                         pd->port_tidcursor = tid;
437                 else
438                         tidcursor_fp(fp) = tid;
439         }
440
441 done:
442         if (ret)
443                 ipath_dbg("Failed to map %u TID pages, failing with %d\n",
444                           ti->tidcnt, -ret);
445         return ret;
446 }
447
448 /**
449  * ipath_tid_free - free a port TID
450  * @pd: the port
451  * @subport: the subport
452  * @ti: the TID info
453  *
454  * right now we are unlocking one page at a time, but since
455  * the intended use of this routine is for a single group of
456  * virtually contiguous pages, that should change to improve
457  * performance.  We check that the TID is in range for this port
458  * but otherwise don't check validity; if user has an error and
459  * frees the wrong tid, it's only their own data that can thereby
460  * be corrupted.  We do check that the TID was in use, for sanity
461  * We always use our idea of the saved address, not the address that
462  * they pass in to us.
463  */
464
465 static int ipath_tid_free(struct ipath_portdata *pd, unsigned subport,
466                           const struct ipath_tid_info *ti)
467 {
468         int ret = 0;
469         u32 tid, porttid, cnt, limit, tidcnt;
470         struct ipath_devdata *dd = pd->port_dd;
471         u64 __iomem *tidbase;
472         unsigned long tidmap[8];
473
474         if (!dd->ipath_pageshadow) {
475                 ret = -ENOMEM;
476                 goto done;
477         }
478
479         if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap,
480                            sizeof tidmap)) {
481                 ret = -EFAULT;
482                 goto done;
483         }
484
485         porttid = pd->port_port * dd->ipath_rcvtidcnt;
486         if (!pd->port_subport_cnt)
487                 tidcnt = dd->ipath_rcvtidcnt;
488         else if (!subport) {
489                 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
490                          (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
491                 porttid += dd->ipath_rcvtidcnt - tidcnt;
492         } else {
493                 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
494                 porttid += tidcnt * (subport - 1);
495         }
496         tidbase = (u64 __iomem *) ((char __iomem *)(dd->ipath_kregbase) +
497                                    dd->ipath_rcvtidbase +
498                                    porttid * sizeof(*tidbase));
499
500         limit = sizeof(tidmap) * BITS_PER_BYTE;
501         if (limit > tidcnt)
502                 /* just in case size changes in future */
503                 limit = tidcnt;
504         tid = find_first_bit(tidmap, limit);
505         ipath_cdbg(VERBOSE, "Port%u free %u tids; first bit (max=%d) "
506                    "set is %d, porttid %u\n", pd->port_port, ti->tidcnt,
507                    limit, tid, porttid);
508         for (cnt = 0; tid < limit; tid++) {
509                 /*
510                  * small optimization; if we detect a run of 3 or so without
511                  * any set, use find_first_bit again.  That's mainly to
512                  * accelerate the case where we wrapped, so we have some at
513                  * the beginning, and some at the end, and a big gap
514                  * in the middle.
515                  */
516                 if (!test_bit(tid, tidmap))
517                         continue;
518                 cnt++;
519                 if (dd->ipath_pageshadow[porttid + tid]) {
520                         ipath_cdbg(VERBOSE, "PID %u freeing TID %u\n",
521                                    pd->port_pid, tid);
522                         dd->ipath_f_put_tid(dd, &tidbase[tid], 1,
523                                             dd->ipath_tidinvalid);
524                         pci_unmap_page(dd->pcidev,
525                                 dd->ipath_physshadow[porttid + tid],
526                                 PAGE_SIZE, PCI_DMA_FROMDEVICE);
527                         ipath_release_user_pages(
528                                 &dd->ipath_pageshadow[porttid + tid], 1);
529                         dd->ipath_pageshadow[porttid + tid] = NULL;
530                         ipath_stats.sps_pageunlocks++;
531                 } else
532                         ipath_dbg("Unused tid %u, ignoring\n", tid);
533         }
534         if (cnt != ti->tidcnt)
535                 ipath_dbg("passed in tidcnt %d, only %d bits set in map\n",
536                           ti->tidcnt, cnt);
537 done:
538         if (ret)
539                 ipath_dbg("Failed to unmap %u TID pages, failing with %d\n",
540                           ti->tidcnt, -ret);
541         return ret;
542 }
543
544 /**
545  * ipath_set_part_key - set a partition key
546  * @pd: the port
547  * @key: the key
548  *
549  * We can have up to 4 active at a time (other than the default, which is
550  * always allowed).  This is somewhat tricky, since multiple ports may set
551  * the same key, so we reference count them, and clean up at exit.  All 4
552  * partition keys are packed into a single infinipath register.  It's an
553  * error for a process to set the same pkey multiple times.  We provide no
554  * mechanism to de-allocate a pkey at this time, we may eventually need to
555  * do that.  I've used the atomic operations, and no locking, and only make
556  * a single pass through what's available.  This should be more than
557  * adequate for some time. I'll think about spinlocks or the like if and as
558  * it's necessary.
559  */
560 static int ipath_set_part_key(struct ipath_portdata *pd, u16 key)
561 {
562         struct ipath_devdata *dd = pd->port_dd;
563         int i, any = 0, pidx = -1;
564         u16 lkey = key & 0x7FFF;
565         int ret;
566
567         if (lkey == (IPATH_DEFAULT_P_KEY & 0x7FFF)) {
568                 /* nothing to do; this key always valid */
569                 ret = 0;
570                 goto bail;
571         }
572
573         ipath_cdbg(VERBOSE, "p%u try to set pkey %hx, current keys "
574                    "%hx:%x %hx:%x %hx:%x %hx:%x\n",
575                    pd->port_port, key, dd->ipath_pkeys[0],
576                    atomic_read(&dd->ipath_pkeyrefs[0]), dd->ipath_pkeys[1],
577                    atomic_read(&dd->ipath_pkeyrefs[1]), dd->ipath_pkeys[2],
578                    atomic_read(&dd->ipath_pkeyrefs[2]), dd->ipath_pkeys[3],
579                    atomic_read(&dd->ipath_pkeyrefs[3]));
580
581         if (!lkey) {
582                 ipath_cdbg(PROC, "p%u tries to set key 0, not allowed\n",
583                            pd->port_port);
584                 ret = -EINVAL;
585                 goto bail;
586         }
587
588         /*
589          * Set the full membership bit, because it has to be
590          * set in the register or the packet, and it seems
591          * cleaner to set in the register than to force all
592          * callers to set it. (see bug 4331)
593          */
594         key |= 0x8000;
595
596         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
597                 if (!pd->port_pkeys[i] && pidx == -1)
598                         pidx = i;
599                 if (pd->port_pkeys[i] == key) {
600                         ipath_cdbg(VERBOSE, "p%u tries to set same pkey "
601                                    "(%x) more than once\n",
602                                    pd->port_port, key);
603                         ret = -EEXIST;
604                         goto bail;
605                 }
606         }
607         if (pidx == -1) {
608                 ipath_dbg("All pkeys for port %u already in use, "
609                           "can't set %x\n", pd->port_port, key);
610                 ret = -EBUSY;
611                 goto bail;
612         }
613         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
614                 if (!dd->ipath_pkeys[i]) {
615                         any++;
616                         continue;
617                 }
618                 if (dd->ipath_pkeys[i] == key) {
619                         atomic_t *pkrefs = &dd->ipath_pkeyrefs[i];
620
621                         if (atomic_inc_return(pkrefs) > 1) {
622                                 pd->port_pkeys[pidx] = key;
623                                 ipath_cdbg(VERBOSE, "p%u set key %x "
624                                            "matches #%d, count now %d\n",
625                                            pd->port_port, key, i,
626                                            atomic_read(pkrefs));
627                                 ret = 0;
628                                 goto bail;
629                         } else {
630                                 /*
631                                  * lost race, decrement count, catch below
632                                  */
633                                 atomic_dec(pkrefs);
634                                 ipath_cdbg(VERBOSE, "Lost race, count was "
635                                            "0, after dec, it's %d\n",
636                                            atomic_read(pkrefs));
637                                 any++;
638                         }
639                 }
640                 if ((dd->ipath_pkeys[i] & 0x7FFF) == lkey) {
641                         /*
642                          * It makes no sense to have both the limited and
643                          * full membership PKEY set at the same time since
644                          * the unlimited one will disable the limited one.
645                          */
646                         ret = -EEXIST;
647                         goto bail;
648                 }
649         }
650         if (!any) {
651                 ipath_dbg("port %u, all pkeys already in use, "
652                           "can't set %x\n", pd->port_port, key);
653                 ret = -EBUSY;
654                 goto bail;
655         }
656         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
657                 if (!dd->ipath_pkeys[i] &&
658                     atomic_inc_return(&dd->ipath_pkeyrefs[i]) == 1) {
659                         u64 pkey;
660
661                         /* for ipathstats, etc. */
662                         ipath_stats.sps_pkeys[i] = lkey;
663                         pd->port_pkeys[pidx] = dd->ipath_pkeys[i] = key;
664                         pkey =
665                                 (u64) dd->ipath_pkeys[0] |
666                                 ((u64) dd->ipath_pkeys[1] << 16) |
667                                 ((u64) dd->ipath_pkeys[2] << 32) |
668                                 ((u64) dd->ipath_pkeys[3] << 48);
669                         ipath_cdbg(PROC, "p%u set key %x in #%d, "
670                                    "portidx %d, new pkey reg %llx\n",
671                                    pd->port_port, key, i, pidx,
672                                    (unsigned long long) pkey);
673                         ipath_write_kreg(
674                                 dd, dd->ipath_kregs->kr_partitionkey, pkey);
675
676                         ret = 0;
677                         goto bail;
678                 }
679         }
680         ipath_dbg("port %u, all pkeys already in use 2nd pass, "
681                   "can't set %x\n", pd->port_port, key);
682         ret = -EBUSY;
683
684 bail:
685         return ret;
686 }
687
688 /**
689  * ipath_manage_rcvq - manage a port's receive queue
690  * @pd: the port
691  * @subport: the subport
692  * @start_stop: action to carry out
693  *
694  * start_stop == 0 disables receive on the port, for use in queue
695  * overflow conditions.  start_stop==1 re-enables, to be used to
696  * re-init the software copy of the head register
697  */
698 static int ipath_manage_rcvq(struct ipath_portdata *pd, unsigned subport,
699                              int start_stop)
700 {
701         struct ipath_devdata *dd = pd->port_dd;
702         u64 tval;
703
704         ipath_cdbg(PROC, "%sabling rcv for unit %u port %u:%u\n",
705                    start_stop ? "en" : "dis", dd->ipath_unit,
706                    pd->port_port, subport);
707         if (subport)
708                 goto bail;
709         /* atomically clear receive enable port. */
710         if (start_stop) {
711                 /*
712                  * On enable, force in-memory copy of the tail register to
713                  * 0, so that protocol code doesn't have to worry about
714                  * whether or not the chip has yet updated the in-memory
715                  * copy or not on return from the system call. The chip
716                  * always resets it's tail register back to 0 on a
717                  * transition from disabled to enabled.  This could cause a
718                  * problem if software was broken, and did the enable w/o
719                  * the disable, but eventually the in-memory copy will be
720                  * updated and correct itself, even in the face of software
721                  * bugs.
722                  */
723                 *(volatile u64 *)pd->port_rcvhdrtail_kvaddr = 0;
724                 set_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
725                         &dd->ipath_rcvctrl);
726         } else
727                 clear_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
728                           &dd->ipath_rcvctrl);
729         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
730                          dd->ipath_rcvctrl);
731         /* now be sure chip saw it before we return */
732         tval = ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
733         if (start_stop) {
734                 /*
735                  * And try to be sure that tail reg update has happened too.
736                  * This should in theory interlock with the RXE changes to
737                  * the tail register.  Don't assign it to the tail register
738                  * in memory copy, since we could overwrite an update by the
739                  * chip if we did.
740                  */
741                 tval = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
742         }
743         /* always; new head should be equal to new tail; see above */
744 bail:
745         return 0;
746 }
747
748 static void ipath_clean_part_key(struct ipath_portdata *pd,
749                                  struct ipath_devdata *dd)
750 {
751         int i, j, pchanged = 0;
752         u64 oldpkey;
753
754         /* for debugging only */
755         oldpkey = (u64) dd->ipath_pkeys[0] |
756                 ((u64) dd->ipath_pkeys[1] << 16) |
757                 ((u64) dd->ipath_pkeys[2] << 32) |
758                 ((u64) dd->ipath_pkeys[3] << 48);
759
760         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
761                 if (!pd->port_pkeys[i])
762                         continue;
763                 ipath_cdbg(VERBOSE, "look for key[%d] %hx in pkeys\n", i,
764                            pd->port_pkeys[i]);
765                 for (j = 0; j < ARRAY_SIZE(dd->ipath_pkeys); j++) {
766                         /* check for match independent of the global bit */
767                         if ((dd->ipath_pkeys[j] & 0x7fff) !=
768                             (pd->port_pkeys[i] & 0x7fff))
769                                 continue;
770                         if (atomic_dec_and_test(&dd->ipath_pkeyrefs[j])) {
771                                 ipath_cdbg(VERBOSE, "p%u clear key "
772                                            "%x matches #%d\n",
773                                            pd->port_port,
774                                            pd->port_pkeys[i], j);
775                                 ipath_stats.sps_pkeys[j] =
776                                         dd->ipath_pkeys[j] = 0;
777                                 pchanged++;
778                         }
779                         else ipath_cdbg(
780                                 VERBOSE, "p%u key %x matches #%d, "
781                                 "but ref still %d\n", pd->port_port,
782                                 pd->port_pkeys[i], j,
783                                 atomic_read(&dd->ipath_pkeyrefs[j]));
784                         break;
785                 }
786                 pd->port_pkeys[i] = 0;
787         }
788         if (pchanged) {
789                 u64 pkey = (u64) dd->ipath_pkeys[0] |
790                         ((u64) dd->ipath_pkeys[1] << 16) |
791                         ((u64) dd->ipath_pkeys[2] << 32) |
792                         ((u64) dd->ipath_pkeys[3] << 48);
793                 ipath_cdbg(VERBOSE, "p%u old pkey reg %llx, "
794                            "new pkey reg %llx\n", pd->port_port,
795                            (unsigned long long) oldpkey,
796                            (unsigned long long) pkey);
797                 ipath_write_kreg(dd, dd->ipath_kregs->kr_partitionkey,
798                                  pkey);
799         }
800 }
801
802 /*
803  * Initialize the port data with the receive buffer sizes
804  * so this can be done while the master port is locked.
805  * Otherwise, there is a race with a slave opening the port
806  * and seeing these fields uninitialized.
807  */
808 static void init_user_egr_sizes(struct ipath_portdata *pd)
809 {
810         struct ipath_devdata *dd = pd->port_dd;
811         unsigned egrperchunk, egrcnt, size;
812
813         /*
814          * to avoid wasting a lot of memory, we allocate 32KB chunks of
815          * physically contiguous memory, advance through it until used up
816          * and then allocate more.  Of course, we need memory to store those
817          * extra pointers, now.  Started out with 256KB, but under heavy
818          * memory pressure (creating large files and then copying them over
819          * NFS while doing lots of MPI jobs), we hit some allocation
820          * failures, even though we can sleep...  (2.6.10) Still get
821          * failures at 64K.  32K is the lowest we can go without wasting
822          * additional memory.
823          */
824         size = 0x8000;
825         egrperchunk = size / dd->ipath_rcvegrbufsize;
826         egrcnt = dd->ipath_rcvegrcnt;
827         pd->port_rcvegrbuf_chunks = (egrcnt + egrperchunk - 1) / egrperchunk;
828         pd->port_rcvegrbufs_perchunk = egrperchunk;
829         pd->port_rcvegrbuf_size = size;
830 }
831
832 /**
833  * ipath_create_user_egr - allocate eager TID buffers
834  * @pd: the port to allocate TID buffers for
835  *
836  * This routine is now quite different for user and kernel, because
837  * the kernel uses skb's, for the accelerated network performance
838  * This is the user port version
839  *
840  * Allocate the eager TID buffers and program them into infinipath
841  * They are no longer completely contiguous, we do multiple allocation
842  * calls.
843  */
844 static int ipath_create_user_egr(struct ipath_portdata *pd)
845 {
846         struct ipath_devdata *dd = pd->port_dd;
847         unsigned e, egrcnt, egrperchunk, chunk, egrsize, egroff;
848         size_t size;
849         int ret;
850         gfp_t gfp_flags;
851
852         /*
853          * GFP_USER, but without GFP_FS, so buffer cache can be
854          * coalesced (we hope); otherwise, even at order 4,
855          * heavy filesystem activity makes these fail, and we can
856          * use compound pages.
857          */
858         gfp_flags = __GFP_WAIT | __GFP_IO | __GFP_COMP;
859
860         egrcnt = dd->ipath_rcvegrcnt;
861         /* TID number offset for this port */
862         egroff = pd->port_port * egrcnt;
863         egrsize = dd->ipath_rcvegrbufsize;
864         ipath_cdbg(VERBOSE, "Allocating %d egr buffers, at egrtid "
865                    "offset %x, egrsize %u\n", egrcnt, egroff, egrsize);
866
867         chunk = pd->port_rcvegrbuf_chunks;
868         egrperchunk = pd->port_rcvegrbufs_perchunk;
869         size = pd->port_rcvegrbuf_size;
870         pd->port_rcvegrbuf = kmalloc(chunk * sizeof(pd->port_rcvegrbuf[0]),
871                                      GFP_KERNEL);
872         if (!pd->port_rcvegrbuf) {
873                 ret = -ENOMEM;
874                 goto bail;
875         }
876         pd->port_rcvegrbuf_phys =
877                 kmalloc(chunk * sizeof(pd->port_rcvegrbuf_phys[0]),
878                         GFP_KERNEL);
879         if (!pd->port_rcvegrbuf_phys) {
880                 ret = -ENOMEM;
881                 goto bail_rcvegrbuf;
882         }
883         for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
884
885                 pd->port_rcvegrbuf[e] = dma_alloc_coherent(
886                         &dd->pcidev->dev, size, &pd->port_rcvegrbuf_phys[e],
887                         gfp_flags);
888
889                 if (!pd->port_rcvegrbuf[e]) {
890                         ret = -ENOMEM;
891                         goto bail_rcvegrbuf_phys;
892                 }
893         }
894
895         pd->port_rcvegr_phys = pd->port_rcvegrbuf_phys[0];
896
897         for (e = chunk = 0; chunk < pd->port_rcvegrbuf_chunks; chunk++) {
898                 dma_addr_t pa = pd->port_rcvegrbuf_phys[chunk];
899                 unsigned i;
900
901                 for (i = 0; e < egrcnt && i < egrperchunk; e++, i++) {
902                         dd->ipath_f_put_tid(dd, e + egroff +
903                                             (u64 __iomem *)
904                                             ((char __iomem *)
905                                              dd->ipath_kregbase +
906                                              dd->ipath_rcvegrbase), 0, pa);
907                         pa += egrsize;
908                 }
909                 cond_resched(); /* don't hog the cpu */
910         }
911
912         ret = 0;
913         goto bail;
914
915 bail_rcvegrbuf_phys:
916         for (e = 0; e < pd->port_rcvegrbuf_chunks &&
917                 pd->port_rcvegrbuf[e]; e++) {
918                 dma_free_coherent(&dd->pcidev->dev, size,
919                                   pd->port_rcvegrbuf[e],
920                                   pd->port_rcvegrbuf_phys[e]);
921
922         }
923         kfree(pd->port_rcvegrbuf_phys);
924         pd->port_rcvegrbuf_phys = NULL;
925 bail_rcvegrbuf:
926         kfree(pd->port_rcvegrbuf);
927         pd->port_rcvegrbuf = NULL;
928 bail:
929         return ret;
930 }
931
932
933 /* common code for the mappings on dma_alloc_coherent mem */
934 static int ipath_mmap_mem(struct vm_area_struct *vma,
935         struct ipath_portdata *pd, unsigned len, int write_ok,
936         void *kvaddr, char *what)
937 {
938         struct ipath_devdata *dd = pd->port_dd;
939         unsigned long pfn;
940         int ret;
941
942         if ((vma->vm_end - vma->vm_start) > len) {
943                 dev_info(&dd->pcidev->dev,
944                          "FAIL on %s: len %lx > %x\n", what,
945                          vma->vm_end - vma->vm_start, len);
946                 ret = -EFAULT;
947                 goto bail;
948         }
949
950         if (!write_ok) {
951                 if (vma->vm_flags & VM_WRITE) {
952                         dev_info(&dd->pcidev->dev,
953                                  "%s must be mapped readonly\n", what);
954                         ret = -EPERM;
955                         goto bail;
956                 }
957
958                 /* don't allow them to later change with mprotect */
959                 vma->vm_flags &= ~VM_MAYWRITE;
960         }
961
962         pfn = virt_to_phys(kvaddr) >> PAGE_SHIFT;
963         ret = remap_pfn_range(vma, vma->vm_start, pfn,
964                               len, vma->vm_page_prot);
965         if (ret)
966                 dev_info(&dd->pcidev->dev, "%s port%u mmap of %lx, %x "
967                          "bytes r%c failed: %d\n", what, pd->port_port,
968                          pfn, len, write_ok?'w':'o', ret);
969         else
970                 ipath_cdbg(VERBOSE, "%s port%u mmaped %lx, %x bytes "
971                            "r%c\n", what, pd->port_port, pfn, len,
972                            write_ok?'w':'o');
973 bail:
974         return ret;
975 }
976
977 static int mmap_ureg(struct vm_area_struct *vma, struct ipath_devdata *dd,
978                      u64 ureg)
979 {
980         unsigned long phys;
981         int ret;
982
983         /*
984          * This is real hardware, so use io_remap.  This is the mechanism
985          * for the user process to update the head registers for their port
986          * in the chip.
987          */
988         if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) {
989                 dev_info(&dd->pcidev->dev, "FAIL mmap userreg: reqlen "
990                          "%lx > PAGE\n", vma->vm_end - vma->vm_start);
991                 ret = -EFAULT;
992         } else {
993                 phys = dd->ipath_physaddr + ureg;
994                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
995
996                 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
997                 ret = io_remap_pfn_range(vma, vma->vm_start,
998                                          phys >> PAGE_SHIFT,
999                                          vma->vm_end - vma->vm_start,
1000                                          vma->vm_page_prot);
1001         }
1002         return ret;
1003 }
1004
1005 static int mmap_piobufs(struct vm_area_struct *vma,
1006                         struct ipath_devdata *dd,
1007                         struct ipath_portdata *pd,
1008                         unsigned piobufs, unsigned piocnt)
1009 {
1010         unsigned long phys;
1011         int ret;
1012
1013         /*
1014          * When we map the PIO buffers in the chip, we want to map them as
1015          * writeonly, no read possible.   This prevents access to previous
1016          * process data, and catches users who might try to read the i/o
1017          * space due to a bug.
1018          */
1019         if ((vma->vm_end - vma->vm_start) > (piocnt * dd->ipath_palign)) {
1020                 dev_info(&dd->pcidev->dev, "FAIL mmap piobufs: "
1021                          "reqlen %lx > PAGE\n",
1022                          vma->vm_end - vma->vm_start);
1023                 ret = -EINVAL;
1024                 goto bail;
1025         }
1026
1027         phys = dd->ipath_physaddr + piobufs;
1028
1029         /*
1030          * Don't mark this as non-cached, or we don't get the
1031          * write combining behavior we want on the PIO buffers!
1032          */
1033
1034 #if defined(__powerpc__)
1035         /* There isn't a generic way to specify writethrough mappings */
1036         pgprot_val(vma->vm_page_prot) |= _PAGE_NO_CACHE;
1037         pgprot_val(vma->vm_page_prot) |= _PAGE_WRITETHRU;
1038         pgprot_val(vma->vm_page_prot) &= ~_PAGE_GUARDED;
1039 #endif
1040
1041         /*
1042          * don't allow them to later change to readable with mprotect (for when
1043          * not initially mapped readable, as is normally the case)
1044          */
1045         vma->vm_flags &= ~VM_MAYREAD;
1046         vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1047
1048         ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
1049                                  vma->vm_end - vma->vm_start,
1050                                  vma->vm_page_prot);
1051 bail:
1052         return ret;
1053 }
1054
1055 static int mmap_rcvegrbufs(struct vm_area_struct *vma,
1056                            struct ipath_portdata *pd)
1057 {
1058         struct ipath_devdata *dd = pd->port_dd;
1059         unsigned long start, size;
1060         size_t total_size, i;
1061         unsigned long pfn;
1062         int ret;
1063
1064         size = pd->port_rcvegrbuf_size;
1065         total_size = pd->port_rcvegrbuf_chunks * size;
1066         if ((vma->vm_end - vma->vm_start) > total_size) {
1067                 dev_info(&dd->pcidev->dev, "FAIL on egr bufs: "
1068                          "reqlen %lx > actual %lx\n",
1069                          vma->vm_end - vma->vm_start,
1070                          (unsigned long) total_size);
1071                 ret = -EINVAL;
1072                 goto bail;
1073         }
1074
1075         if (vma->vm_flags & VM_WRITE) {
1076                 dev_info(&dd->pcidev->dev, "Can't map eager buffers as "
1077                          "writable (flags=%lx)\n", vma->vm_flags);
1078                 ret = -EPERM;
1079                 goto bail;
1080         }
1081         /* don't allow them to later change to writeable with mprotect */
1082         vma->vm_flags &= ~VM_MAYWRITE;
1083
1084         start = vma->vm_start;
1085
1086         for (i = 0; i < pd->port_rcvegrbuf_chunks; i++, start += size) {
1087                 pfn = virt_to_phys(pd->port_rcvegrbuf[i]) >> PAGE_SHIFT;
1088                 ret = remap_pfn_range(vma, start, pfn, size,
1089                                       vma->vm_page_prot);
1090                 if (ret < 0)
1091                         goto bail;
1092         }
1093         ret = 0;
1094
1095 bail:
1096         return ret;
1097 }
1098
1099 /*
1100  * ipath_file_vma_nopage - handle a VMA page fault.
1101  */
1102 static struct page *ipath_file_vma_nopage(struct vm_area_struct *vma,
1103                                           unsigned long address, int *type)
1104 {
1105         unsigned long offset = address - vma->vm_start;
1106         struct page *page = NOPAGE_SIGBUS;
1107         void *pageptr;
1108
1109         /*
1110          * Convert the vmalloc address into a struct page.
1111          */
1112         pageptr = (void *)(offset + (vma->vm_pgoff << PAGE_SHIFT));
1113         page = vmalloc_to_page(pageptr);
1114         if (!page)
1115                 goto out;
1116
1117         /* Increment the reference count. */
1118         get_page(page);
1119         if (type)
1120                 *type = VM_FAULT_MINOR;
1121 out:
1122         return page;
1123 }
1124
1125 static struct vm_operations_struct ipath_file_vm_ops = {
1126         .nopage = ipath_file_vma_nopage,
1127 };
1128
1129 static int mmap_kvaddr(struct vm_area_struct *vma, u64 pgaddr,
1130                        struct ipath_portdata *pd, unsigned subport)
1131 {
1132         unsigned long len;
1133         struct ipath_devdata *dd;
1134         void *addr;
1135         size_t size;
1136         int ret;
1137
1138         /* If the port is not shared, all addresses should be physical */
1139         if (!pd->port_subport_cnt) {
1140                 ret = -EINVAL;
1141                 goto bail;
1142         }
1143
1144         dd = pd->port_dd;
1145         size = pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
1146
1147         /*
1148          * Master has all the slave uregbase, rcvhdrq, and
1149          * rcvegrbufs mmapped.
1150          */
1151         if (subport == 0) {
1152                 unsigned num_slaves = pd->port_subport_cnt - 1;
1153
1154                 if (pgaddr == ((u64) pd->subport_uregbase & MMAP64_MASK)) {
1155                         addr = pd->subport_uregbase;
1156                         size = PAGE_SIZE * num_slaves;
1157                 } else if (pgaddr == ((u64) pd->subport_rcvhdr_base &
1158                                       MMAP64_MASK)) {
1159                         addr = pd->subport_rcvhdr_base;
1160                         size = pd->port_rcvhdrq_size * num_slaves;
1161                 } else if (pgaddr == ((u64) pd->subport_rcvegrbuf &
1162                                       MMAP64_MASK)) {
1163                         addr = pd->subport_rcvegrbuf;
1164                         size *= num_slaves;
1165                 } else {
1166                         ret = -EINVAL;
1167                         goto bail;
1168                 }
1169         } else if (pgaddr == (((u64) pd->subport_uregbase +
1170                                PAGE_SIZE * (subport - 1)) & MMAP64_MASK)) {
1171                 addr = pd->subport_uregbase + PAGE_SIZE * (subport - 1);
1172                 size = PAGE_SIZE;
1173         } else if (pgaddr == (((u64) pd->subport_rcvhdr_base +
1174                                pd->port_rcvhdrq_size * (subport - 1)) &
1175                               MMAP64_MASK)) {
1176                 addr = pd->subport_rcvhdr_base +
1177                         pd->port_rcvhdrq_size * (subport - 1);
1178                 size = pd->port_rcvhdrq_size;
1179         } else if (pgaddr == (((u64) pd->subport_rcvegrbuf +
1180                                size * (subport - 1)) & MMAP64_MASK)) {
1181                 addr = pd->subport_rcvegrbuf + size * (subport - 1);
1182                 /* rcvegrbufs are read-only on the slave */
1183                 if (vma->vm_flags & VM_WRITE) {
1184                         dev_info(&dd->pcidev->dev,
1185                                  "Can't map eager buffers as "
1186                                  "writable (flags=%lx)\n", vma->vm_flags);
1187                         ret = -EPERM;
1188                         goto bail;
1189                 }
1190                 /*
1191                  * Don't allow permission to later change to writeable
1192                  * with mprotect.
1193                  */
1194                 vma->vm_flags &= ~VM_MAYWRITE;
1195         } else {
1196                 ret = -EINVAL;
1197                 goto bail;
1198         }
1199         len = vma->vm_end - vma->vm_start;
1200         if (len > size) {
1201                 ipath_cdbg(MM, "FAIL: reqlen %lx > %zx\n", len, size);
1202                 ret = -EINVAL;
1203                 goto bail;
1204         }
1205
1206         vma->vm_pgoff = (unsigned long) addr >> PAGE_SHIFT;
1207         vma->vm_ops = &ipath_file_vm_ops;
1208         vma->vm_flags |= VM_RESERVED | VM_DONTEXPAND;
1209         ret = 0;
1210
1211 bail:
1212         return ret;
1213 }
1214
1215 /**
1216  * ipath_mmap - mmap various structures into user space
1217  * @fp: the file pointer
1218  * @vma: the VM area
1219  *
1220  * We use this to have a shared buffer between the kernel and the user code
1221  * for the rcvhdr queue, egr buffers, and the per-port user regs and pio
1222  * buffers in the chip.  We have the open and close entries so we can bump
1223  * the ref count and keep the driver from being unloaded while still mapped.
1224  */
1225 static int ipath_mmap(struct file *fp, struct vm_area_struct *vma)
1226 {
1227         struct ipath_portdata *pd;
1228         struct ipath_devdata *dd;
1229         u64 pgaddr, ureg;
1230         unsigned piobufs, piocnt;
1231         int ret;
1232
1233         pd = port_fp(fp);
1234         if (!pd) {
1235                 ret = -EINVAL;
1236                 goto bail;
1237         }
1238         dd = pd->port_dd;
1239
1240         /*
1241          * This is the ipath_do_user_init() code, mapping the shared buffers
1242          * into the user process. The address referred to by vm_pgoff is the
1243          * file offset passed via mmap().  For shared ports, this is the
1244          * kernel vmalloc() address of the pages to share with the master.
1245          * For non-shared or master ports, this is a physical address.
1246          * We only do one mmap for each space mapped.
1247          */
1248         pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1249
1250         /*
1251          * Check for 0 in case one of the allocations failed, but user
1252          * called mmap anyway.
1253          */
1254         if (!pgaddr)  {
1255                 ret = -EINVAL;
1256                 goto bail;
1257         }
1258
1259         ipath_cdbg(MM, "pgaddr %llx vm_start=%lx len %lx port %u:%u:%u\n",
1260                    (unsigned long long) pgaddr, vma->vm_start,
1261                    vma->vm_end - vma->vm_start, dd->ipath_unit,
1262                    pd->port_port, subport_fp(fp));
1263
1264         /*
1265          * Physical addresses must fit in 40 bits for our hardware.
1266          * Check for kernel virtual addresses first, anything else must
1267          * match a HW or memory address.
1268          */
1269         if (pgaddr >= (1ULL<<40)) {
1270                 ret = mmap_kvaddr(vma, pgaddr, pd, subport_fp(fp));
1271                 goto bail;
1272         }
1273
1274         if (!pd->port_subport_cnt) {
1275                 /* port is not shared */
1276                 ureg = dd->ipath_uregbase + dd->ipath_palign * pd->port_port;
1277                 piocnt = dd->ipath_pbufsport;
1278                 piobufs = pd->port_piobufs;
1279         } else if (!subport_fp(fp)) {
1280                 /* caller is the master */
1281                 ureg = dd->ipath_uregbase + dd->ipath_palign * pd->port_port;
1282                 piocnt = (dd->ipath_pbufsport / pd->port_subport_cnt) +
1283                          (dd->ipath_pbufsport % pd->port_subport_cnt);
1284                 piobufs = pd->port_piobufs +
1285                         dd->ipath_palign * (dd->ipath_pbufsport - piocnt);
1286         } else {
1287                 unsigned slave = subport_fp(fp) - 1;
1288
1289                 /* caller is a slave */
1290                 ureg = 0;
1291                 piocnt = dd->ipath_pbufsport / pd->port_subport_cnt;
1292                 piobufs = pd->port_piobufs + dd->ipath_palign * piocnt * slave;
1293         }
1294
1295         if (pgaddr == ureg)
1296                 ret = mmap_ureg(vma, dd, ureg);
1297         else if (pgaddr == piobufs)
1298                 ret = mmap_piobufs(vma, dd, pd, piobufs, piocnt);
1299         else if (pgaddr == dd->ipath_pioavailregs_phys)
1300                 /* in-memory copy of pioavail registers */
1301                 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1302                                      (void *) dd->ipath_pioavailregs_dma,
1303                                      "pioavail registers");
1304         else if (subport_fp(fp))
1305                 /* Subports don't mmap the physical receive buffers */
1306                 ret = -EINVAL;
1307         else if (pgaddr == pd->port_rcvegr_phys)
1308                 ret = mmap_rcvegrbufs(vma, pd);
1309         else if (pgaddr == (u64) pd->port_rcvhdrq_phys)
1310                 /*
1311                  * The rcvhdrq itself; readonly except on HT (so have
1312                  * to allow writable mapping), multiple pages, contiguous
1313                  * from an i/o perspective.
1314                  */
1315                 ret = ipath_mmap_mem(vma, pd, pd->port_rcvhdrq_size, 1,
1316                                      pd->port_rcvhdrq,
1317                                      "rcvhdrq");
1318         else if (pgaddr == (u64) pd->port_rcvhdrqtailaddr_phys)
1319                 /* in-memory copy of rcvhdrq tail register */
1320                 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1321                                      pd->port_rcvhdrtail_kvaddr,
1322                                      "rcvhdrq tail");
1323         else
1324                 ret = -EINVAL;
1325
1326         vma->vm_private_data = NULL;
1327
1328         if (ret < 0)
1329                 dev_info(&dd->pcidev->dev,
1330                          "Failure %d on off %llx len %lx\n",
1331                          -ret, (unsigned long long)pgaddr,
1332                          vma->vm_end - vma->vm_start);
1333 bail:
1334         return ret;
1335 }
1336
1337 static unsigned int ipath_poll(struct file *fp,
1338                                struct poll_table_struct *pt)
1339 {
1340         struct ipath_portdata *pd;
1341         u32 head, tail;
1342         int bit;
1343         unsigned pollflag = 0;
1344         struct ipath_devdata *dd;
1345
1346         pd = port_fp(fp);
1347         if (!pd)
1348                 goto bail;
1349         dd = pd->port_dd;
1350
1351         bit = pd->port_port + INFINIPATH_R_INTRAVAIL_SHIFT;
1352         set_bit(bit, &dd->ipath_rcvctrl);
1353
1354         /*
1355          * Before blocking, make sure that head is still == tail,
1356          * reading from the chip, so we can be sure the interrupt
1357          * enable has made it to the chip.  If not equal, disable
1358          * interrupt again and return immediately.  This avoids races,
1359          * and the overhead of the chip read doesn't matter much at
1360          * this point, since we are waiting for something anyway.
1361          */
1362
1363         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1364                          dd->ipath_rcvctrl);
1365
1366         head = ipath_read_ureg32(dd, ur_rcvhdrhead, pd->port_port);
1367         tail = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
1368
1369         if (tail == head) {
1370                 set_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
1371                 if (dd->ipath_rhdrhead_intr_off) /* arm rcv interrupt */
1372                         (void)ipath_write_ureg(dd, ur_rcvhdrhead,
1373                                                dd->ipath_rhdrhead_intr_off
1374                                                | head, pd->port_port);
1375                 poll_wait(fp, &pd->port_wait, pt);
1376
1377                 if (test_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag)) {
1378                         /* timed out, no packets received */
1379                         clear_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
1380                         pd->port_rcvwait_to++;
1381                 }
1382                 else
1383                         pollflag = POLLIN | POLLRDNORM;
1384         }
1385         else {
1386                 /* it's already happened; don't do wait_event overhead */
1387                 pollflag = POLLIN | POLLRDNORM;
1388                 pd->port_rcvnowait++;
1389         }
1390
1391         clear_bit(bit, &dd->ipath_rcvctrl);
1392         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1393                          dd->ipath_rcvctrl);
1394
1395 bail:
1396         return pollflag;
1397 }
1398
1399 static int init_subports(struct ipath_devdata *dd,
1400                          struct ipath_portdata *pd,
1401                          const struct ipath_user_info *uinfo)
1402 {
1403         int ret = 0;
1404         unsigned num_slaves;
1405         size_t size;
1406
1407         /* Old user binaries don't know about subports */
1408         if ((uinfo->spu_userversion & 0xffff) != IPATH_USER_SWMINOR)
1409                 goto bail;
1410         /*
1411          * If the user is requesting zero or one port,
1412          * skip the subport allocation.
1413          */
1414         if (uinfo->spu_subport_cnt <= 1)
1415                 goto bail;
1416         if (uinfo->spu_subport_cnt > 4) {
1417                 ret = -EINVAL;
1418                 goto bail;
1419         }
1420
1421         num_slaves = uinfo->spu_subport_cnt - 1;
1422         pd->subport_uregbase = vmalloc(PAGE_SIZE * num_slaves);
1423         if (!pd->subport_uregbase) {
1424                 ret = -ENOMEM;
1425                 goto bail;
1426         }
1427         /* Note: pd->port_rcvhdrq_size isn't initialized yet. */
1428         size = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
1429                      sizeof(u32), PAGE_SIZE) * num_slaves;
1430         pd->subport_rcvhdr_base = vmalloc(size);
1431         if (!pd->subport_rcvhdr_base) {
1432                 ret = -ENOMEM;
1433                 goto bail_ureg;
1434         }
1435
1436         pd->subport_rcvegrbuf = vmalloc(pd->port_rcvegrbuf_chunks *
1437                                         pd->port_rcvegrbuf_size *
1438                                         num_slaves);
1439         if (!pd->subport_rcvegrbuf) {
1440                 ret = -ENOMEM;
1441                 goto bail_rhdr;
1442         }
1443
1444         pd->port_subport_cnt = uinfo->spu_subport_cnt;
1445         pd->port_subport_id = uinfo->spu_subport_id;
1446         pd->active_slaves = 1;
1447         goto bail;
1448
1449 bail_rhdr:
1450         vfree(pd->subport_rcvhdr_base);
1451 bail_ureg:
1452         vfree(pd->subport_uregbase);
1453         pd->subport_uregbase = NULL;
1454 bail:
1455         return ret;
1456 }
1457
1458 static int try_alloc_port(struct ipath_devdata *dd, int port,
1459                           struct file *fp,
1460                           const struct ipath_user_info *uinfo)
1461 {
1462         struct ipath_portdata *pd;
1463         int ret;
1464
1465         if (!(pd = dd->ipath_pd[port])) {
1466                 void *ptmp;
1467
1468                 pd = kzalloc(sizeof(struct ipath_portdata), GFP_KERNEL);
1469
1470                 /*
1471                  * Allocate memory for use in ipath_tid_update() just once
1472                  * at open, not per call.  Reduces cost of expected send
1473                  * setup.
1474                  */
1475                 ptmp = kmalloc(dd->ipath_rcvtidcnt * sizeof(u16) +
1476                                dd->ipath_rcvtidcnt * sizeof(struct page **),
1477                                GFP_KERNEL);
1478                 if (!pd || !ptmp) {
1479                         ipath_dev_err(dd, "Unable to allocate portdata "
1480                                       "memory, failing open\n");
1481                         ret = -ENOMEM;
1482                         kfree(pd);
1483                         kfree(ptmp);
1484                         goto bail;
1485                 }
1486                 dd->ipath_pd[port] = pd;
1487                 dd->ipath_pd[port]->port_port = port;
1488                 dd->ipath_pd[port]->port_dd = dd;
1489                 dd->ipath_pd[port]->port_tid_pg_list = ptmp;
1490                 init_waitqueue_head(&dd->ipath_pd[port]->port_wait);
1491         }
1492         if (!pd->port_cnt) {
1493                 pd->userversion = uinfo->spu_userversion;
1494                 init_user_egr_sizes(pd);
1495                 if ((ret = init_subports(dd, pd, uinfo)) != 0)
1496                         goto bail;
1497                 ipath_cdbg(PROC, "%s[%u] opened unit:port %u:%u\n",
1498                            current->comm, current->pid, dd->ipath_unit,
1499                            port);
1500                 pd->port_cnt = 1;
1501                 port_fp(fp) = pd;
1502                 pd->port_pid = current->pid;
1503                 strncpy(pd->port_comm, current->comm, sizeof(pd->port_comm));
1504                 ipath_stats.sps_ports++;
1505                 ret = 0;
1506         } else
1507                 ret = -EBUSY;
1508
1509 bail:
1510         return ret;
1511 }
1512
1513 static inline int usable(struct ipath_devdata *dd)
1514 {
1515         return dd &&
1516                 (dd->ipath_flags & IPATH_PRESENT) &&
1517                 dd->ipath_kregbase &&
1518                 dd->ipath_lid &&
1519                 !(dd->ipath_flags & (IPATH_LINKDOWN | IPATH_DISABLED
1520                                      | IPATH_LINKUNK));
1521 }
1522
1523 static int find_free_port(int unit, struct file *fp,
1524                           const struct ipath_user_info *uinfo)
1525 {
1526         struct ipath_devdata *dd = ipath_lookup(unit);
1527         int ret, i;
1528
1529         if (!dd) {
1530                 ret = -ENODEV;
1531                 goto bail;
1532         }
1533
1534         if (!usable(dd)) {
1535                 ret = -ENETDOWN;
1536                 goto bail;
1537         }
1538
1539         for (i = 1; i < dd->ipath_cfgports; i++) {
1540                 ret = try_alloc_port(dd, i, fp, uinfo);
1541                 if (ret != -EBUSY)
1542                         goto bail;
1543         }
1544         ret = -EBUSY;
1545
1546 bail:
1547         return ret;
1548 }
1549
1550 static int find_best_unit(struct file *fp,
1551                           const struct ipath_user_info *uinfo)
1552 {
1553         int ret = 0, i, prefunit = -1, devmax;
1554         int maxofallports, npresent, nup;
1555         int ndev;
1556
1557         devmax = ipath_count_units(&npresent, &nup, &maxofallports);
1558
1559         /*
1560          * This code is present to allow a knowledgeable person to
1561          * specify the layout of processes to processors before opening
1562          * this driver, and then we'll assign the process to the "closest"
1563          * InfiniPath chip to that processor (we assume reasonable connectivity,
1564          * for now).  This code assumes that if affinity has been set
1565          * before this point, that at most one cpu is set; for now this
1566          * is reasonable.  I check for both cpus_empty() and cpus_full(),
1567          * in case some kernel variant sets none of the bits when no
1568          * affinity is set.  2.6.11 and 12 kernels have all present
1569          * cpus set.  Some day we'll have to fix it up further to handle
1570          * a cpu subset.  This algorithm fails for two HT chips connected
1571          * in tunnel fashion.  Eventually this needs real topology
1572          * information.  There may be some issues with dual core numbering
1573          * as well.  This needs more work prior to release.
1574          */
1575         if (!cpus_empty(current->cpus_allowed) &&
1576             !cpus_full(current->cpus_allowed)) {
1577                 int ncpus = num_online_cpus(), curcpu = -1;
1578                 for (i = 0; i < ncpus; i++)
1579                         if (cpu_isset(i, current->cpus_allowed)) {
1580                                 ipath_cdbg(PROC, "%s[%u] affinity set for "
1581                                            "cpu %d\n", current->comm,
1582                                            current->pid, i);
1583                                 curcpu = i;
1584                         }
1585                 if (curcpu != -1) {
1586                         if (npresent) {
1587                                 prefunit = curcpu / (ncpus / npresent);
1588                                 ipath_dbg("%s[%u] %d chips, %d cpus, "
1589                                           "%d cpus/chip, select unit %d\n",
1590                                           current->comm, current->pid,
1591                                           npresent, ncpus, ncpus / npresent,
1592                                           prefunit);
1593                         }
1594                 }
1595         }
1596
1597         /*
1598          * user ports start at 1, kernel port is 0
1599          * For now, we do round-robin access across all chips
1600          */
1601
1602         if (prefunit != -1)
1603                 devmax = prefunit + 1;
1604 recheck:
1605         for (i = 1; i < maxofallports; i++) {
1606                 for (ndev = prefunit != -1 ? prefunit : 0; ndev < devmax;
1607                      ndev++) {
1608                         struct ipath_devdata *dd = ipath_lookup(ndev);
1609
1610                         if (!usable(dd))
1611                                 continue; /* can't use this unit */
1612                         if (i >= dd->ipath_cfgports)
1613                                 /*
1614                                  * Maxed out on users of this unit. Try
1615                                  * next.
1616                                  */
1617                                 continue;
1618                         ret = try_alloc_port(dd, i, fp, uinfo);
1619                         if (!ret)
1620                                 goto done;
1621                 }
1622         }
1623
1624         if (npresent) {
1625                 if (nup == 0) {
1626                         ret = -ENETDOWN;
1627                         ipath_dbg("No ports available (none initialized "
1628                                   "and ready)\n");
1629                 } else {
1630                         if (prefunit > 0) {
1631                                 /* if started above 0, retry from 0 */
1632                                 ipath_cdbg(PROC,
1633                                            "%s[%u] no ports on prefunit "
1634                                            "%d, clear and re-check\n",
1635                                            current->comm, current->pid,
1636                                            prefunit);
1637                                 devmax = ipath_count_units(NULL, NULL,
1638                                                            NULL);
1639                                 prefunit = -1;
1640                                 goto recheck;
1641                         }
1642                         ret = -EBUSY;
1643                         ipath_dbg("No ports available\n");
1644                 }
1645         } else {
1646                 ret = -ENXIO;
1647                 ipath_dbg("No boards found\n");
1648         }
1649
1650 done:
1651         return ret;
1652 }
1653
1654 static int find_shared_port(struct file *fp,
1655                             const struct ipath_user_info *uinfo)
1656 {
1657         int devmax, ndev, i;
1658         int ret = 0;
1659
1660         devmax = ipath_count_units(NULL, NULL, NULL);
1661
1662         for (ndev = 0; ndev < devmax; ndev++) {
1663                 struct ipath_devdata *dd = ipath_lookup(ndev);
1664
1665                 if (!dd)
1666                         continue;
1667                 for (i = 1; i < dd->ipath_cfgports; i++) {
1668                         struct ipath_portdata *pd = dd->ipath_pd[i];
1669
1670                         /* Skip ports which are not yet open */
1671                         if (!pd || !pd->port_cnt)
1672                                 continue;
1673                         /* Skip port if it doesn't match the requested one */
1674                         if (pd->port_subport_id != uinfo->spu_subport_id)
1675                                 continue;
1676                         /* Verify the sharing process matches the master */
1677                         if (pd->port_subport_cnt != uinfo->spu_subport_cnt ||
1678                             pd->userversion != uinfo->spu_userversion ||
1679                             pd->port_cnt >= pd->port_subport_cnt) {
1680                                 ret = -EINVAL;
1681                                 goto done;
1682                         }
1683                         port_fp(fp) = pd;
1684                         subport_fp(fp) = pd->port_cnt++;
1685                         tidcursor_fp(fp) = 0;
1686                         pd->active_slaves |= 1 << subport_fp(fp);
1687                         ipath_cdbg(PROC,
1688                                    "%s[%u] %u sharing %s[%u] unit:port %u:%u\n",
1689                                    current->comm, current->pid,
1690                                    subport_fp(fp),
1691                                    pd->port_comm, pd->port_pid,
1692                                    dd->ipath_unit, pd->port_port);
1693                         ret = 1;
1694                         goto done;
1695                 }
1696         }
1697
1698 done:
1699         return ret;
1700 }
1701
1702 static int ipath_open(struct inode *in, struct file *fp)
1703 {
1704         /* The real work is performed later in ipath_assign_port() */
1705         fp->private_data = kzalloc(sizeof(struct ipath_filedata), GFP_KERNEL);
1706         return fp->private_data ? 0 : -ENOMEM;
1707 }
1708
1709
1710 /* Get port early, so can set affinity prior to memory allocation */
1711 static int ipath_assign_port(struct file *fp,
1712                               const struct ipath_user_info *uinfo)
1713 {
1714         int ret;
1715         int i_minor;
1716         unsigned swminor;
1717
1718         /* Check to be sure we haven't already initialized this file */
1719         if (port_fp(fp)) {
1720                 ret = -EINVAL;
1721                 goto done;
1722         }
1723
1724         /* for now, if major version is different, bail */
1725         if ((uinfo->spu_userversion >> 16) != IPATH_USER_SWMAJOR) {
1726                 ipath_dbg("User major version %d not same as driver "
1727                           "major %d\n", uinfo->spu_userversion >> 16,
1728                           IPATH_USER_SWMAJOR);
1729                 ret = -ENODEV;
1730                 goto done;
1731         }
1732
1733         swminor = uinfo->spu_userversion & 0xffff;
1734         if (swminor != IPATH_USER_SWMINOR)
1735                 ipath_dbg("User minor version %d not same as driver "
1736                           "minor %d\n", swminor, IPATH_USER_SWMINOR);
1737
1738         mutex_lock(&ipath_mutex);
1739
1740         if (swminor == IPATH_USER_SWMINOR && uinfo->spu_subport_cnt &&
1741             (ret = find_shared_port(fp, uinfo))) {
1742                 mutex_unlock(&ipath_mutex);
1743                 if (ret > 0)
1744                         ret = 0;
1745                 goto done;
1746         }
1747
1748         i_minor = iminor(fp->f_dentry->d_inode) - IPATH_USER_MINOR_BASE;
1749         ipath_cdbg(VERBOSE, "open on dev %lx (minor %d)\n",
1750                    (long)fp->f_dentry->d_inode->i_rdev, i_minor);
1751
1752         if (i_minor)
1753                 ret = find_free_port(i_minor - 1, fp, uinfo);
1754         else
1755                 ret = find_best_unit(fp, uinfo);
1756
1757         mutex_unlock(&ipath_mutex);
1758
1759 done:
1760         return ret;
1761 }
1762
1763
1764 static int ipath_do_user_init(struct file *fp,
1765                               const struct ipath_user_info *uinfo)
1766 {
1767         int ret;
1768         struct ipath_portdata *pd;
1769         struct ipath_devdata *dd;
1770         u32 head32;
1771
1772         pd = port_fp(fp);
1773         dd = pd->port_dd;
1774
1775         if (uinfo->spu_rcvhdrsize) {
1776                 ret = ipath_setrcvhdrsize(dd, uinfo->spu_rcvhdrsize);
1777                 if (ret)
1778                         goto done;
1779         }
1780
1781         /* for now we do nothing with rcvhdrcnt: uinfo->spu_rcvhdrcnt */
1782
1783         /* for right now, kernel piobufs are at end, so port 1 is at 0 */
1784         pd->port_piobufs = dd->ipath_piobufbase +
1785                 dd->ipath_pbufsport * (pd->port_port - 1) * dd->ipath_palign;
1786         ipath_cdbg(VERBOSE, "Set base of piobufs for port %u to 0x%x\n",
1787                    pd->port_port, pd->port_piobufs);
1788
1789         /*
1790          * Now allocate the rcvhdr Q and eager TIDs; skip the TID
1791          * array for time being.  If pd->port_port > chip-supported,
1792          * we need to do extra stuff here to handle by handling overflow
1793          * through port 0, someday
1794          */
1795         ret = ipath_create_rcvhdrq(dd, pd);
1796         if (!ret)
1797                 ret = ipath_create_user_egr(pd);
1798         if (ret)
1799                 goto done;
1800
1801         /*
1802          * set the eager head register for this port to the current values
1803          * of the tail pointers, since we don't know if they were
1804          * updated on last use of the port.
1805          */
1806         head32 = ipath_read_ureg32(dd, ur_rcvegrindextail, pd->port_port);
1807         ipath_write_ureg(dd, ur_rcvegrindexhead, head32, pd->port_port);
1808         dd->ipath_lastegrheads[pd->port_port] = -1;
1809         dd->ipath_lastrcvhdrqtails[pd->port_port] = -1;
1810         ipath_cdbg(VERBOSE, "Wrote port%d egrhead %x from tail regs\n",
1811                 pd->port_port, head32);
1812         pd->port_tidcursor = 0; /* start at beginning after open */
1813         /*
1814          * now enable the port; the tail registers will be written to memory
1815          * by the chip as soon as it sees the write to
1816          * dd->ipath_kregs->kr_rcvctrl.  The update only happens on
1817          * transition from 0 to 1, so clear it first, then set it as part of
1818          * enabling the port.  This will (very briefly) affect any other
1819          * open ports, but it shouldn't be long enough to be an issue.
1820          * We explictly set the in-memory copy to 0 beforehand, so we don't
1821          * have to wait to be sure the DMA update has happened.
1822          */
1823         *(volatile u64 *)pd->port_rcvhdrtail_kvaddr = 0ULL;
1824         set_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
1825                 &dd->ipath_rcvctrl);
1826         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1827                          dd->ipath_rcvctrl & ~INFINIPATH_R_TAILUPD);
1828         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1829                          dd->ipath_rcvctrl);
1830 done:
1831         return ret;
1832 }
1833
1834 /**
1835  * unlock_exptid - unlock any expected TID entries port still had in use
1836  * @pd: port
1837  *
1838  * We don't actually update the chip here, because we do a bulk update
1839  * below, using ipath_f_clear_tids.
1840  */
1841 static void unlock_expected_tids(struct ipath_portdata *pd)
1842 {
1843         struct ipath_devdata *dd = pd->port_dd;
1844         int port_tidbase = pd->port_port * dd->ipath_rcvtidcnt;
1845         int i, cnt = 0, maxtid = port_tidbase + dd->ipath_rcvtidcnt;
1846
1847         ipath_cdbg(VERBOSE, "Port %u unlocking any locked expTID pages\n",
1848                    pd->port_port);
1849         for (i = port_tidbase; i < maxtid; i++) {
1850                 if (!dd->ipath_pageshadow[i])
1851                         continue;
1852
1853                 pci_unmap_page(dd->pcidev, dd->ipath_physshadow[i],
1854                         PAGE_SIZE, PCI_DMA_FROMDEVICE);
1855                 ipath_release_user_pages_on_close(&dd->ipath_pageshadow[i],
1856                                                   1);
1857                 dd->ipath_pageshadow[i] = NULL;
1858                 cnt++;
1859                 ipath_stats.sps_pageunlocks++;
1860         }
1861         if (cnt)
1862                 ipath_cdbg(VERBOSE, "Port %u locked %u expTID entries\n",
1863                            pd->port_port, cnt);
1864
1865         if (ipath_stats.sps_pagelocks || ipath_stats.sps_pageunlocks)
1866                 ipath_cdbg(VERBOSE, "%llu pages locked, %llu unlocked\n",
1867                            (unsigned long long) ipath_stats.sps_pagelocks,
1868                            (unsigned long long)
1869                            ipath_stats.sps_pageunlocks);
1870 }
1871
1872 static int ipath_close(struct inode *in, struct file *fp)
1873 {
1874         int ret = 0;
1875         struct ipath_filedata *fd;
1876         struct ipath_portdata *pd;
1877         struct ipath_devdata *dd;
1878         unsigned port;
1879
1880         ipath_cdbg(VERBOSE, "close on dev %lx, private data %p\n",
1881                    (long)in->i_rdev, fp->private_data);
1882
1883         mutex_lock(&ipath_mutex);
1884
1885         fd = (struct ipath_filedata *) fp->private_data;
1886         fp->private_data = NULL;
1887         pd = fd->pd;
1888         if (!pd) {
1889                 mutex_unlock(&ipath_mutex);
1890                 goto bail;
1891         }
1892         if (--pd->port_cnt) {
1893                 /*
1894                  * XXX If the master closes the port before the slave(s),
1895                  * revoke the mmap for the eager receive queue so
1896                  * the slave(s) don't wait for receive data forever.
1897                  */
1898                 pd->active_slaves &= ~(1 << fd->subport);
1899                 mutex_unlock(&ipath_mutex);
1900                 goto bail;
1901         }
1902         port = pd->port_port;
1903         dd = pd->port_dd;
1904
1905         if (pd->port_hdrqfull) {
1906                 ipath_cdbg(PROC, "%s[%u] had %u rcvhdrqfull errors "
1907                            "during run\n", pd->port_comm, pd->port_pid,
1908                            pd->port_hdrqfull);
1909                 pd->port_hdrqfull = 0;
1910         }
1911
1912         if (pd->port_rcvwait_to || pd->port_piowait_to
1913             || pd->port_rcvnowait || pd->port_pionowait) {
1914                 ipath_cdbg(VERBOSE, "port%u, %u rcv, %u pio wait timeo; "
1915                            "%u rcv %u, pio already\n",
1916                            pd->port_port, pd->port_rcvwait_to,
1917                            pd->port_piowait_to, pd->port_rcvnowait,
1918                            pd->port_pionowait);
1919                 pd->port_rcvwait_to = pd->port_piowait_to =
1920                         pd->port_rcvnowait = pd->port_pionowait = 0;
1921         }
1922         if (pd->port_flag) {
1923                 ipath_dbg("port %u port_flag still set to 0x%lx\n",
1924                           pd->port_port, pd->port_flag);
1925                 pd->port_flag = 0;
1926         }
1927
1928         if (dd->ipath_kregbase) {
1929                 int i;
1930                 /* atomically clear receive enable port. */
1931                 clear_bit(INFINIPATH_R_PORTENABLE_SHIFT + port,
1932                           &dd->ipath_rcvctrl);
1933                 ipath_write_kreg( dd, dd->ipath_kregs->kr_rcvctrl,
1934                         dd->ipath_rcvctrl);
1935                 /* and read back from chip to be sure that nothing
1936                  * else is in flight when we do the rest */
1937                 (void)ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
1938
1939                 /* clean up the pkeys for this port user */
1940                 ipath_clean_part_key(pd, dd);
1941                 /*
1942                  * be paranoid, and never write 0's to these, just use an
1943                  * unused part of the port 0 tail page.  Of course,
1944                  * rcvhdraddr points to a large chunk of memory, so this
1945                  * could still trash things, but at least it won't trash
1946                  * page 0, and by disabling the port, it should stop "soon",
1947                  * even if a packet or two is in already in flight after we
1948                  * disabled the port.
1949                  */
1950                 ipath_write_kreg_port(dd,
1951                         dd->ipath_kregs->kr_rcvhdrtailaddr, port,
1952                         dd->ipath_dummy_hdrq_phys);
1953                 ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdraddr,
1954                         pd->port_port, dd->ipath_dummy_hdrq_phys);
1955
1956                 i = dd->ipath_pbufsport * (port - 1);
1957                 ipath_disarm_piobufs(dd, i, dd->ipath_pbufsport);
1958
1959                 dd->ipath_f_clear_tids(dd, pd->port_port);
1960
1961                 if (dd->ipath_pageshadow)
1962                         unlock_expected_tids(pd);
1963                 ipath_stats.sps_ports--;
1964                 ipath_cdbg(PROC, "%s[%u] closed port %u:%u\n",
1965                            pd->port_comm, pd->port_pid,
1966                            dd->ipath_unit, port);
1967         }
1968
1969         pd->port_pid = 0;
1970         dd->ipath_pd[pd->port_port] = NULL; /* before releasing mutex */
1971         mutex_unlock(&ipath_mutex);
1972         ipath_free_pddata(dd, pd); /* after releasing the mutex */
1973
1974 bail:
1975         kfree(fd);
1976         return ret;
1977 }
1978
1979 static int ipath_port_info(struct ipath_portdata *pd, u16 subport,
1980                            struct ipath_port_info __user *uinfo)
1981 {
1982         struct ipath_port_info info;
1983         int nup;
1984         int ret;
1985         size_t sz;
1986
1987         (void) ipath_count_units(NULL, &nup, NULL);
1988         info.num_active = nup;
1989         info.unit = pd->port_dd->ipath_unit;
1990         info.port = pd->port_port;
1991         info.subport = subport;
1992         /* Don't return new fields if old library opened the port. */
1993         if ((pd->userversion & 0xffff) == IPATH_USER_SWMINOR) {
1994                 /* Number of user ports available for this device. */
1995                 info.num_ports = pd->port_dd->ipath_cfgports - 1;
1996                 info.num_subports = pd->port_subport_cnt;
1997                 sz = sizeof(info);
1998         } else
1999                 sz = sizeof(info) - 2 * sizeof(u16);
2000
2001         if (copy_to_user(uinfo, &info, sz)) {
2002                 ret = -EFAULT;
2003                 goto bail;
2004         }
2005         ret = 0;
2006
2007 bail:
2008         return ret;
2009 }
2010
2011 static int ipath_get_slave_info(struct ipath_portdata *pd,
2012                                 void __user *slave_mask_addr)
2013 {
2014         int ret = 0;
2015
2016         if (copy_to_user(slave_mask_addr, &pd->active_slaves, sizeof(u32)))
2017                 ret = -EFAULT;
2018         return ret;
2019 }
2020
2021 static ssize_t ipath_write(struct file *fp, const char __user *data,
2022                            size_t count, loff_t *off)
2023 {
2024         const struct ipath_cmd __user *ucmd;
2025         struct ipath_portdata *pd;
2026         const void __user *src;
2027         size_t consumed, copy;
2028         struct ipath_cmd cmd;
2029         ssize_t ret = 0;
2030         void *dest;
2031
2032         if (count < sizeof(cmd.type)) {
2033                 ret = -EINVAL;
2034                 goto bail;
2035         }
2036
2037         ucmd = (const struct ipath_cmd __user *) data;
2038
2039         if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) {
2040                 ret = -EFAULT;
2041                 goto bail;
2042         }
2043
2044         consumed = sizeof(cmd.type);
2045
2046         switch (cmd.type) {
2047         case IPATH_CMD_ASSIGN_PORT:
2048         case __IPATH_CMD_USER_INIT:
2049         case IPATH_CMD_USER_INIT:
2050                 copy = sizeof(cmd.cmd.user_info);
2051                 dest = &cmd.cmd.user_info;
2052                 src = &ucmd->cmd.user_info;
2053                 break;
2054         case IPATH_CMD_RECV_CTRL:
2055                 copy = sizeof(cmd.cmd.recv_ctrl);
2056                 dest = &cmd.cmd.recv_ctrl;
2057                 src = &ucmd->cmd.recv_ctrl;
2058                 break;
2059         case IPATH_CMD_PORT_INFO:
2060                 copy = sizeof(cmd.cmd.port_info);
2061                 dest = &cmd.cmd.port_info;
2062                 src = &ucmd->cmd.port_info;
2063                 break;
2064         case IPATH_CMD_TID_UPDATE:
2065         case IPATH_CMD_TID_FREE:
2066                 copy = sizeof(cmd.cmd.tid_info);
2067                 dest = &cmd.cmd.tid_info;
2068                 src = &ucmd->cmd.tid_info;
2069                 break;
2070         case IPATH_CMD_SET_PART_KEY:
2071                 copy = sizeof(cmd.cmd.part_key);
2072                 dest = &cmd.cmd.part_key;
2073                 src = &ucmd->cmd.part_key;
2074                 break;
2075         case IPATH_CMD_SLAVE_INFO:
2076                 copy = sizeof(cmd.cmd.slave_mask_addr);
2077                 dest = &cmd.cmd.slave_mask_addr;
2078                 src = &ucmd->cmd.slave_mask_addr;
2079                 break;
2080         default:
2081                 ret = -EINVAL;
2082                 goto bail;
2083         }
2084
2085         if ((count - consumed) < copy) {
2086                 ret = -EINVAL;
2087                 goto bail;
2088         }
2089
2090         if (copy_from_user(dest, src, copy)) {
2091                 ret = -EFAULT;
2092                 goto bail;
2093         }
2094
2095         consumed += copy;
2096         pd = port_fp(fp);
2097         if (!pd && cmd.type != __IPATH_CMD_USER_INIT &&
2098                 cmd.type != IPATH_CMD_ASSIGN_PORT) {
2099                 ret = -EINVAL;
2100                 goto bail;
2101         }
2102
2103         switch (cmd.type) {
2104         case IPATH_CMD_ASSIGN_PORT:
2105                 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2106                 if (ret)
2107                         goto bail;
2108                 break;
2109         case __IPATH_CMD_USER_INIT:
2110                 /* backwards compatibility, get port first */
2111                 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2112                 if (ret)
2113                         goto bail;
2114                 /* and fall through to current version. */
2115         case IPATH_CMD_USER_INIT:
2116                 ret = ipath_do_user_init(fp, &cmd.cmd.user_info);
2117                 if (ret)
2118                         goto bail;
2119                 ret = ipath_get_base_info(
2120                         fp, (void __user *) (unsigned long)
2121                         cmd.cmd.user_info.spu_base_info,
2122                         cmd.cmd.user_info.spu_base_info_size);
2123                 break;
2124         case IPATH_CMD_RECV_CTRL:
2125                 ret = ipath_manage_rcvq(pd, subport_fp(fp), cmd.cmd.recv_ctrl);
2126                 break;
2127         case IPATH_CMD_PORT_INFO:
2128                 ret = ipath_port_info(pd, subport_fp(fp),
2129                                       (struct ipath_port_info __user *)
2130                                       (unsigned long) cmd.cmd.port_info);
2131                 break;
2132         case IPATH_CMD_TID_UPDATE:
2133                 ret = ipath_tid_update(pd, fp, &cmd.cmd.tid_info);
2134                 break;
2135         case IPATH_CMD_TID_FREE:
2136                 ret = ipath_tid_free(pd, subport_fp(fp), &cmd.cmd.tid_info);
2137                 break;
2138         case IPATH_CMD_SET_PART_KEY:
2139                 ret = ipath_set_part_key(pd, cmd.cmd.part_key);
2140                 break;
2141         case IPATH_CMD_SLAVE_INFO:
2142                 ret = ipath_get_slave_info(pd,
2143                                            (void __user *) (unsigned long)
2144                                            cmd.cmd.slave_mask_addr);
2145                 break;
2146         }
2147
2148         if (ret >= 0)
2149                 ret = consumed;
2150
2151 bail:
2152         return ret;
2153 }
2154
2155 static struct class *ipath_class;
2156
2157 static int init_cdev(int minor, char *name, struct file_operations *fops,
2158                      struct cdev **cdevp, struct class_device **class_devp)
2159 {
2160         const dev_t dev = MKDEV(IPATH_MAJOR, minor);
2161         struct cdev *cdev = NULL;
2162         struct class_device *class_dev = NULL;
2163         int ret;
2164
2165         cdev = cdev_alloc();
2166         if (!cdev) {
2167                 printk(KERN_ERR IPATH_DRV_NAME
2168                        ": Could not allocate cdev for minor %d, %s\n",
2169                        minor, name);
2170                 ret = -ENOMEM;
2171                 goto done;
2172         }
2173
2174         cdev->owner = THIS_MODULE;
2175         cdev->ops = fops;
2176         kobject_set_name(&cdev->kobj, name);
2177
2178         ret = cdev_add(cdev, dev, 1);
2179         if (ret < 0) {
2180                 printk(KERN_ERR IPATH_DRV_NAME
2181                        ": Could not add cdev for minor %d, %s (err %d)\n",
2182                        minor, name, -ret);
2183                 goto err_cdev;
2184         }
2185
2186         class_dev = class_device_create(ipath_class, NULL, dev, NULL, name);
2187
2188         if (IS_ERR(class_dev)) {
2189                 ret = PTR_ERR(class_dev);
2190                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2191                        "class_dev for minor %d, %s (err %d)\n",
2192                        minor, name, -ret);
2193                 goto err_cdev;
2194         }
2195
2196         goto done;
2197
2198 err_cdev:
2199         cdev_del(cdev);
2200         cdev = NULL;
2201
2202 done:
2203         if (ret >= 0) {
2204                 *cdevp = cdev;
2205                 *class_devp = class_dev;
2206         } else {
2207                 *cdevp = NULL;
2208                 *class_devp = NULL;
2209         }
2210
2211         return ret;
2212 }
2213
2214 int ipath_cdev_init(int minor, char *name, struct file_operations *fops,
2215                     struct cdev **cdevp, struct class_device **class_devp)
2216 {
2217         return init_cdev(minor, name, fops, cdevp, class_devp);
2218 }
2219
2220 static void cleanup_cdev(struct cdev **cdevp,
2221                          struct class_device **class_devp)
2222 {
2223         struct class_device *class_dev = *class_devp;
2224
2225         if (class_dev) {
2226                 class_device_unregister(class_dev);
2227                 *class_devp = NULL;
2228         }
2229
2230         if (*cdevp) {
2231                 cdev_del(*cdevp);
2232                 *cdevp = NULL;
2233         }
2234 }
2235
2236 void ipath_cdev_cleanup(struct cdev **cdevp,
2237                         struct class_device **class_devp)
2238 {
2239         cleanup_cdev(cdevp, class_devp);
2240 }
2241
2242 static struct cdev *wildcard_cdev;
2243 static struct class_device *wildcard_class_dev;
2244
2245 static const dev_t dev = MKDEV(IPATH_MAJOR, 0);
2246
2247 static int user_init(void)
2248 {
2249         int ret;
2250
2251         ret = register_chrdev_region(dev, IPATH_NMINORS, IPATH_DRV_NAME);
2252         if (ret < 0) {
2253                 printk(KERN_ERR IPATH_DRV_NAME ": Could not register "
2254                        "chrdev region (err %d)\n", -ret);
2255                 goto done;
2256         }
2257
2258         ipath_class = class_create(THIS_MODULE, IPATH_DRV_NAME);
2259
2260         if (IS_ERR(ipath_class)) {
2261                 ret = PTR_ERR(ipath_class);
2262                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2263                        "device class (err %d)\n", -ret);
2264                 goto bail;
2265         }
2266
2267         goto done;
2268 bail:
2269         unregister_chrdev_region(dev, IPATH_NMINORS);
2270 done:
2271         return ret;
2272 }
2273
2274 static void user_cleanup(void)
2275 {
2276         if (ipath_class) {
2277                 class_destroy(ipath_class);
2278                 ipath_class = NULL;
2279         }
2280
2281         unregister_chrdev_region(dev, IPATH_NMINORS);
2282 }
2283
2284 static atomic_t user_count = ATOMIC_INIT(0);
2285 static atomic_t user_setup = ATOMIC_INIT(0);
2286
2287 int ipath_user_add(struct ipath_devdata *dd)
2288 {
2289         char name[10];
2290         int ret;
2291
2292         if (atomic_inc_return(&user_count) == 1) {
2293                 ret = user_init();
2294                 if (ret < 0) {
2295                         ipath_dev_err(dd, "Unable to set up user support: "
2296                                       "error %d\n", -ret);
2297                         goto bail;
2298                 }
2299                 ret = init_cdev(0, "ipath", &ipath_file_ops, &wildcard_cdev,
2300                                 &wildcard_class_dev);
2301                 if (ret < 0) {
2302                         ipath_dev_err(dd, "Could not create wildcard "
2303                                       "minor: error %d\n", -ret);
2304                         goto bail_user;
2305                 }
2306
2307                 atomic_set(&user_setup, 1);
2308         }
2309
2310         snprintf(name, sizeof(name), "ipath%d", dd->ipath_unit);
2311
2312         ret = init_cdev(dd->ipath_unit + 1, name, &ipath_file_ops,
2313                         &dd->user_cdev, &dd->user_class_dev);
2314         if (ret < 0)
2315                 ipath_dev_err(dd, "Could not create user minor %d, %s\n",
2316                               dd->ipath_unit + 1, name);
2317
2318         goto bail;
2319
2320 bail_user:
2321         user_cleanup();
2322 bail:
2323         return ret;
2324 }
2325
2326 void ipath_user_remove(struct ipath_devdata *dd)
2327 {
2328         cleanup_cdev(&dd->user_cdev, &dd->user_class_dev);
2329
2330         if (atomic_dec_return(&user_count) == 0) {
2331                 if (atomic_read(&user_setup) == 0)
2332                         goto bail;
2333
2334                 cleanup_cdev(&wildcard_cdev, &wildcard_class_dev);
2335                 user_cleanup();
2336
2337                 atomic_set(&user_setup, 0);
2338         }
2339 bail:
2340         return;
2341 }