Merge branch 'i7300_idle' into release
[linux-2.6] / drivers / infiniband / hw / ipath / ipath_file_ops.c
1 /*
2  * Copyright (c) 2006, 2007, 2008 QLogic Corporation. All rights reserved.
3  * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
4  *
5  * This software is available to you under a choice of one of two
6  * licenses.  You may choose to be licensed under the terms of the GNU
7  * General Public License (GPL) Version 2, available from the file
8  * COPYING in the main directory of this source tree, or the
9  * OpenIB.org BSD license below:
10  *
11  *     Redistribution and use in source and binary forms, with or
12  *     without modification, are permitted provided that the following
13  *     conditions are met:
14  *
15  *      - Redistributions of source code must retain the above
16  *        copyright notice, this list of conditions and the following
17  *        disclaimer.
18  *
19  *      - Redistributions in binary form must reproduce the above
20  *        copyright notice, this list of conditions and the following
21  *        disclaimer in the documentation and/or other materials
22  *        provided with the distribution.
23  *
24  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31  * SOFTWARE.
32  */
33
34 #include <linux/pci.h>
35 #include <linux/poll.h>
36 #include <linux/cdev.h>
37 #include <linux/swap.h>
38 #include <linux/vmalloc.h>
39 #include <linux/highmem.h>
40 #include <linux/io.h>
41 #include <linux/jiffies.h>
42 #include <linux/smp_lock.h>
43 #include <asm/pgtable.h>
44
45 #include "ipath_kernel.h"
46 #include "ipath_common.h"
47 #include "ipath_user_sdma.h"
48
49 static int ipath_open(struct inode *, struct file *);
50 static int ipath_close(struct inode *, struct file *);
51 static ssize_t ipath_write(struct file *, const char __user *, size_t,
52                            loff_t *);
53 static ssize_t ipath_writev(struct kiocb *, const struct iovec *,
54                             unsigned long , loff_t);
55 static unsigned int ipath_poll(struct file *, struct poll_table_struct *);
56 static int ipath_mmap(struct file *, struct vm_area_struct *);
57
58 static const struct file_operations ipath_file_ops = {
59         .owner = THIS_MODULE,
60         .write = ipath_write,
61         .aio_write = ipath_writev,
62         .open = ipath_open,
63         .release = ipath_close,
64         .poll = ipath_poll,
65         .mmap = ipath_mmap
66 };
67
68 /*
69  * Convert kernel virtual addresses to physical addresses so they don't
70  * potentially conflict with the chip addresses used as mmap offsets.
71  * It doesn't really matter what mmap offset we use as long as we can
72  * interpret it correctly.
73  */
74 static u64 cvt_kvaddr(void *p)
75 {
76         struct page *page;
77         u64 paddr = 0;
78
79         page = vmalloc_to_page(p);
80         if (page)
81                 paddr = page_to_pfn(page) << PAGE_SHIFT;
82
83         return paddr;
84 }
85
86 static int ipath_get_base_info(struct file *fp,
87                                void __user *ubase, size_t ubase_size)
88 {
89         struct ipath_portdata *pd = port_fp(fp);
90         int ret = 0;
91         struct ipath_base_info *kinfo = NULL;
92         struct ipath_devdata *dd = pd->port_dd;
93         unsigned subport_cnt;
94         int shared, master;
95         size_t sz;
96
97         subport_cnt = pd->port_subport_cnt;
98         if (!subport_cnt) {
99                 shared = 0;
100                 master = 0;
101                 subport_cnt = 1;
102         } else {
103                 shared = 1;
104                 master = !subport_fp(fp);
105         }
106
107         sz = sizeof(*kinfo);
108         /* If port sharing is not requested, allow the old size structure */
109         if (!shared)
110                 sz -= 7 * sizeof(u64);
111         if (ubase_size < sz) {
112                 ipath_cdbg(PROC,
113                            "Base size %zu, need %zu (version mismatch?)\n",
114                            ubase_size, sz);
115                 ret = -EINVAL;
116                 goto bail;
117         }
118
119         kinfo = kzalloc(sizeof(*kinfo), GFP_KERNEL);
120         if (kinfo == NULL) {
121                 ret = -ENOMEM;
122                 goto bail;
123         }
124
125         ret = dd->ipath_f_get_base_info(pd, kinfo);
126         if (ret < 0)
127                 goto bail;
128
129         kinfo->spi_rcvhdr_cnt = dd->ipath_rcvhdrcnt;
130         kinfo->spi_rcvhdrent_size = dd->ipath_rcvhdrentsize;
131         kinfo->spi_tidegrcnt = dd->ipath_rcvegrcnt;
132         kinfo->spi_rcv_egrbufsize = dd->ipath_rcvegrbufsize;
133         /*
134          * have to mmap whole thing
135          */
136         kinfo->spi_rcv_egrbuftotlen =
137                 pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
138         kinfo->spi_rcv_egrperchunk = pd->port_rcvegrbufs_perchunk;
139         kinfo->spi_rcv_egrchunksize = kinfo->spi_rcv_egrbuftotlen /
140                 pd->port_rcvegrbuf_chunks;
141         kinfo->spi_tidcnt = dd->ipath_rcvtidcnt / subport_cnt;
142         if (master)
143                 kinfo->spi_tidcnt += dd->ipath_rcvtidcnt % subport_cnt;
144         /*
145          * for this use, may be ipath_cfgports summed over all chips that
146          * are are configured and present
147          */
148         kinfo->spi_nports = dd->ipath_cfgports;
149         /* unit (chip/board) our port is on */
150         kinfo->spi_unit = dd->ipath_unit;
151         /* for now, only a single page */
152         kinfo->spi_tid_maxsize = PAGE_SIZE;
153
154         /*
155          * Doing this per port, and based on the skip value, etc.  This has
156          * to be the actual buffer size, since the protocol code treats it
157          * as an array.
158          *
159          * These have to be set to user addresses in the user code via mmap.
160          * These values are used on return to user code for the mmap target
161          * addresses only.  For 32 bit, same 44 bit address problem, so use
162          * the physical address, not virtual.  Before 2.6.11, using the
163          * page_address() macro worked, but in 2.6.11, even that returns the
164          * full 64 bit address (upper bits all 1's).  So far, using the
165          * physical addresses (or chip offsets, for chip mapping) works, but
166          * no doubt some future kernel release will change that, and we'll be
167          * on to yet another method of dealing with this.
168          */
169         kinfo->spi_rcvhdr_base = (u64) pd->port_rcvhdrq_phys;
170         kinfo->spi_rcvhdr_tailaddr = (u64) pd->port_rcvhdrqtailaddr_phys;
171         kinfo->spi_rcv_egrbufs = (u64) pd->port_rcvegr_phys;
172         kinfo->spi_pioavailaddr = (u64) dd->ipath_pioavailregs_phys;
173         kinfo->spi_status = (u64) kinfo->spi_pioavailaddr +
174                 (void *) dd->ipath_statusp -
175                 (void *) dd->ipath_pioavailregs_dma;
176         if (!shared) {
177                 kinfo->spi_piocnt = pd->port_piocnt;
178                 kinfo->spi_piobufbase = (u64) pd->port_piobufs;
179                 kinfo->__spi_uregbase = (u64) dd->ipath_uregbase +
180                         dd->ipath_ureg_align * pd->port_port;
181         } else if (master) {
182                 kinfo->spi_piocnt = (pd->port_piocnt / subport_cnt) +
183                                     (pd->port_piocnt % subport_cnt);
184                 /* Master's PIO buffers are after all the slave's */
185                 kinfo->spi_piobufbase = (u64) pd->port_piobufs +
186                         dd->ipath_palign *
187                         (pd->port_piocnt - kinfo->spi_piocnt);
188         } else {
189                 unsigned slave = subport_fp(fp) - 1;
190
191                 kinfo->spi_piocnt = pd->port_piocnt / subport_cnt;
192                 kinfo->spi_piobufbase = (u64) pd->port_piobufs +
193                         dd->ipath_palign * kinfo->spi_piocnt * slave;
194         }
195
196         if (shared) {
197                 kinfo->spi_port_uregbase = (u64) dd->ipath_uregbase +
198                         dd->ipath_ureg_align * pd->port_port;
199                 kinfo->spi_port_rcvegrbuf = kinfo->spi_rcv_egrbufs;
200                 kinfo->spi_port_rcvhdr_base = kinfo->spi_rcvhdr_base;
201                 kinfo->spi_port_rcvhdr_tailaddr = kinfo->spi_rcvhdr_tailaddr;
202
203                 kinfo->__spi_uregbase = cvt_kvaddr(pd->subport_uregbase +
204                         PAGE_SIZE * subport_fp(fp));
205
206                 kinfo->spi_rcvhdr_base = cvt_kvaddr(pd->subport_rcvhdr_base +
207                         pd->port_rcvhdrq_size * subport_fp(fp));
208                 kinfo->spi_rcvhdr_tailaddr = 0;
209                 kinfo->spi_rcv_egrbufs = cvt_kvaddr(pd->subport_rcvegrbuf +
210                         pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size *
211                         subport_fp(fp));
212
213                 kinfo->spi_subport_uregbase =
214                         cvt_kvaddr(pd->subport_uregbase);
215                 kinfo->spi_subport_rcvegrbuf =
216                         cvt_kvaddr(pd->subport_rcvegrbuf);
217                 kinfo->spi_subport_rcvhdr_base =
218                         cvt_kvaddr(pd->subport_rcvhdr_base);
219                 ipath_cdbg(PROC, "port %u flags %x %llx %llx %llx\n",
220                         kinfo->spi_port, kinfo->spi_runtime_flags,
221                         (unsigned long long) kinfo->spi_subport_uregbase,
222                         (unsigned long long) kinfo->spi_subport_rcvegrbuf,
223                         (unsigned long long) kinfo->spi_subport_rcvhdr_base);
224         }
225
226         kinfo->spi_pioindex = (kinfo->spi_piobufbase - dd->ipath_piobufbase) /
227                 dd->ipath_palign;
228         kinfo->spi_pioalign = dd->ipath_palign;
229
230         kinfo->spi_qpair = IPATH_KD_QP;
231         /*
232          * user mode PIO buffers are always 2KB, even when 4KB can
233          * be received, and sent via the kernel; this is ibmaxlen
234          * for 2K MTU.
235          */
236         kinfo->spi_piosize = dd->ipath_piosize2k - 2 * sizeof(u32);
237         kinfo->spi_mtu = dd->ipath_ibmaxlen;    /* maxlen, not ibmtu */
238         kinfo->spi_port = pd->port_port;
239         kinfo->spi_subport = subport_fp(fp);
240         kinfo->spi_sw_version = IPATH_KERN_SWVERSION;
241         kinfo->spi_hw_version = dd->ipath_revision;
242
243         if (master) {
244                 kinfo->spi_runtime_flags |= IPATH_RUNTIME_MASTER;
245         }
246
247         sz = (ubase_size < sizeof(*kinfo)) ? ubase_size : sizeof(*kinfo);
248         if (copy_to_user(ubase, kinfo, sz))
249                 ret = -EFAULT;
250
251 bail:
252         kfree(kinfo);
253         return ret;
254 }
255
256 /**
257  * ipath_tid_update - update a port TID
258  * @pd: the port
259  * @fp: the ipath device file
260  * @ti: the TID information
261  *
262  * The new implementation as of Oct 2004 is that the driver assigns
263  * the tid and returns it to the caller.   To make it easier to
264  * catch bugs, and to reduce search time, we keep a cursor for
265  * each port, walking the shadow tid array to find one that's not
266  * in use.
267  *
268  * For now, if we can't allocate the full list, we fail, although
269  * in the long run, we'll allocate as many as we can, and the
270  * caller will deal with that by trying the remaining pages later.
271  * That means that when we fail, we have to mark the tids as not in
272  * use again, in our shadow copy.
273  *
274  * It's up to the caller to free the tids when they are done.
275  * We'll unlock the pages as they free them.
276  *
277  * Also, right now we are locking one page at a time, but since
278  * the intended use of this routine is for a single group of
279  * virtually contiguous pages, that should change to improve
280  * performance.
281  */
282 static int ipath_tid_update(struct ipath_portdata *pd, struct file *fp,
283                             const struct ipath_tid_info *ti)
284 {
285         int ret = 0, ntids;
286         u32 tid, porttid, cnt, i, tidcnt, tidoff;
287         u16 *tidlist;
288         struct ipath_devdata *dd = pd->port_dd;
289         u64 physaddr;
290         unsigned long vaddr;
291         u64 __iomem *tidbase;
292         unsigned long tidmap[8];
293         struct page **pagep = NULL;
294         unsigned subport = subport_fp(fp);
295
296         if (!dd->ipath_pageshadow) {
297                 ret = -ENOMEM;
298                 goto done;
299         }
300
301         cnt = ti->tidcnt;
302         if (!cnt) {
303                 ipath_dbg("After copyin, tidcnt 0, tidlist %llx\n",
304                           (unsigned long long) ti->tidlist);
305                 /*
306                  * Should we treat as success?  likely a bug
307                  */
308                 ret = -EFAULT;
309                 goto done;
310         }
311         porttid = pd->port_port * dd->ipath_rcvtidcnt;
312         if (!pd->port_subport_cnt) {
313                 tidcnt = dd->ipath_rcvtidcnt;
314                 tid = pd->port_tidcursor;
315                 tidoff = 0;
316         } else if (!subport) {
317                 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
318                          (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
319                 tidoff = dd->ipath_rcvtidcnt - tidcnt;
320                 porttid += tidoff;
321                 tid = tidcursor_fp(fp);
322         } else {
323                 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
324                 tidoff = tidcnt * (subport - 1);
325                 porttid += tidoff;
326                 tid = tidcursor_fp(fp);
327         }
328         if (cnt > tidcnt) {
329                 /* make sure it all fits in port_tid_pg_list */
330                 dev_info(&dd->pcidev->dev, "Process tried to allocate %u "
331                          "TIDs, only trying max (%u)\n", cnt, tidcnt);
332                 cnt = tidcnt;
333         }
334         pagep = &((struct page **) pd->port_tid_pg_list)[tidoff];
335         tidlist = &((u16 *) &pagep[dd->ipath_rcvtidcnt])[tidoff];
336
337         memset(tidmap, 0, sizeof(tidmap));
338         /* before decrement; chip actual # */
339         ntids = tidcnt;
340         tidbase = (u64 __iomem *) (((char __iomem *) dd->ipath_kregbase) +
341                                    dd->ipath_rcvtidbase +
342                                    porttid * sizeof(*tidbase));
343
344         ipath_cdbg(VERBOSE, "Port%u %u tids, cursor %u, tidbase %p\n",
345                    pd->port_port, cnt, tid, tidbase);
346
347         /* virtual address of first page in transfer */
348         vaddr = ti->tidvaddr;
349         if (!access_ok(VERIFY_WRITE, (void __user *) vaddr,
350                        cnt * PAGE_SIZE)) {
351                 ipath_dbg("Fail vaddr %p, %u pages, !access_ok\n",
352                           (void *)vaddr, cnt);
353                 ret = -EFAULT;
354                 goto done;
355         }
356         ret = ipath_get_user_pages(vaddr, cnt, pagep);
357         if (ret) {
358                 if (ret == -EBUSY) {
359                         ipath_dbg("Failed to lock addr %p, %u pages "
360                                   "(already locked)\n",
361                                   (void *) vaddr, cnt);
362                         /*
363                          * for now, continue, and see what happens but with
364                          * the new implementation, this should never happen,
365                          * unless perhaps the user has mpin'ed the pages
366                          * themselves (something we need to test)
367                          */
368                         ret = 0;
369                 } else {
370                         dev_info(&dd->pcidev->dev,
371                                  "Failed to lock addr %p, %u pages: "
372                                  "errno %d\n", (void *) vaddr, cnt, -ret);
373                         goto done;
374                 }
375         }
376         for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
377                 for (; ntids--; tid++) {
378                         if (tid == tidcnt)
379                                 tid = 0;
380                         if (!dd->ipath_pageshadow[porttid + tid])
381                                 break;
382                 }
383                 if (ntids < 0) {
384                         /*
385                          * oops, wrapped all the way through their TIDs,
386                          * and didn't have enough free; see comments at
387                          * start of routine
388                          */
389                         ipath_dbg("Not enough free TIDs for %u pages "
390                                   "(index %d), failing\n", cnt, i);
391                         i--;    /* last tidlist[i] not filled in */
392                         ret = -ENOMEM;
393                         break;
394                 }
395                 tidlist[i] = tid + tidoff;
396                 ipath_cdbg(VERBOSE, "Updating idx %u to TID %u, "
397                            "vaddr %lx\n", i, tid + tidoff, vaddr);
398                 /* we "know" system pages and TID pages are same size */
399                 dd->ipath_pageshadow[porttid + tid] = pagep[i];
400                 dd->ipath_physshadow[porttid + tid] = ipath_map_page(
401                         dd->pcidev, pagep[i], 0, PAGE_SIZE,
402                         PCI_DMA_FROMDEVICE);
403                 /*
404                  * don't need atomic or it's overhead
405                  */
406                 __set_bit(tid, tidmap);
407                 physaddr = dd->ipath_physshadow[porttid + tid];
408                 ipath_stats.sps_pagelocks++;
409                 ipath_cdbg(VERBOSE,
410                            "TID %u, vaddr %lx, physaddr %llx pgp %p\n",
411                            tid, vaddr, (unsigned long long) physaddr,
412                            pagep[i]);
413                 dd->ipath_f_put_tid(dd, &tidbase[tid], RCVHQ_RCV_TYPE_EXPECTED,
414                                     physaddr);
415                 /*
416                  * don't check this tid in ipath_portshadow, since we
417                  * just filled it in; start with the next one.
418                  */
419                 tid++;
420         }
421
422         if (ret) {
423                 u32 limit;
424         cleanup:
425                 /* jump here if copy out of updated info failed... */
426                 ipath_dbg("After failure (ret=%d), undo %d of %d entries\n",
427                           -ret, i, cnt);
428                 /* same code that's in ipath_free_tid() */
429                 limit = sizeof(tidmap) * BITS_PER_BYTE;
430                 if (limit > tidcnt)
431                         /* just in case size changes in future */
432                         limit = tidcnt;
433                 tid = find_first_bit((const unsigned long *)tidmap, limit);
434                 for (; tid < limit; tid++) {
435                         if (!test_bit(tid, tidmap))
436                                 continue;
437                         if (dd->ipath_pageshadow[porttid + tid]) {
438                                 ipath_cdbg(VERBOSE, "Freeing TID %u\n",
439                                            tid);
440                                 dd->ipath_f_put_tid(dd, &tidbase[tid],
441                                                     RCVHQ_RCV_TYPE_EXPECTED,
442                                                     dd->ipath_tidinvalid);
443                                 pci_unmap_page(dd->pcidev,
444                                         dd->ipath_physshadow[porttid + tid],
445                                         PAGE_SIZE, PCI_DMA_FROMDEVICE);
446                                 dd->ipath_pageshadow[porttid + tid] = NULL;
447                                 ipath_stats.sps_pageunlocks++;
448                         }
449                 }
450                 ipath_release_user_pages(pagep, cnt);
451         } else {
452                 /*
453                  * Copy the updated array, with ipath_tid's filled in, back
454                  * to user.  Since we did the copy in already, this "should
455                  * never fail" If it does, we have to clean up...
456                  */
457                 if (copy_to_user((void __user *)
458                                  (unsigned long) ti->tidlist,
459                                  tidlist, cnt * sizeof(*tidlist))) {
460                         ret = -EFAULT;
461                         goto cleanup;
462                 }
463                 if (copy_to_user((void __user *) (unsigned long) ti->tidmap,
464                                  tidmap, sizeof tidmap)) {
465                         ret = -EFAULT;
466                         goto cleanup;
467                 }
468                 if (tid == tidcnt)
469                         tid = 0;
470                 if (!pd->port_subport_cnt)
471                         pd->port_tidcursor = tid;
472                 else
473                         tidcursor_fp(fp) = tid;
474         }
475
476 done:
477         if (ret)
478                 ipath_dbg("Failed to map %u TID pages, failing with %d\n",
479                           ti->tidcnt, -ret);
480         return ret;
481 }
482
483 /**
484  * ipath_tid_free - free a port TID
485  * @pd: the port
486  * @subport: the subport
487  * @ti: the TID info
488  *
489  * right now we are unlocking one page at a time, but since
490  * the intended use of this routine is for a single group of
491  * virtually contiguous pages, that should change to improve
492  * performance.  We check that the TID is in range for this port
493  * but otherwise don't check validity; if user has an error and
494  * frees the wrong tid, it's only their own data that can thereby
495  * be corrupted.  We do check that the TID was in use, for sanity
496  * We always use our idea of the saved address, not the address that
497  * they pass in to us.
498  */
499
500 static int ipath_tid_free(struct ipath_portdata *pd, unsigned subport,
501                           const struct ipath_tid_info *ti)
502 {
503         int ret = 0;
504         u32 tid, porttid, cnt, limit, tidcnt;
505         struct ipath_devdata *dd = pd->port_dd;
506         u64 __iomem *tidbase;
507         unsigned long tidmap[8];
508
509         if (!dd->ipath_pageshadow) {
510                 ret = -ENOMEM;
511                 goto done;
512         }
513
514         if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap,
515                            sizeof tidmap)) {
516                 ret = -EFAULT;
517                 goto done;
518         }
519
520         porttid = pd->port_port * dd->ipath_rcvtidcnt;
521         if (!pd->port_subport_cnt)
522                 tidcnt = dd->ipath_rcvtidcnt;
523         else if (!subport) {
524                 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
525                          (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
526                 porttid += dd->ipath_rcvtidcnt - tidcnt;
527         } else {
528                 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
529                 porttid += tidcnt * (subport - 1);
530         }
531         tidbase = (u64 __iomem *) ((char __iomem *)(dd->ipath_kregbase) +
532                                    dd->ipath_rcvtidbase +
533                                    porttid * sizeof(*tidbase));
534
535         limit = sizeof(tidmap) * BITS_PER_BYTE;
536         if (limit > tidcnt)
537                 /* just in case size changes in future */
538                 limit = tidcnt;
539         tid = find_first_bit(tidmap, limit);
540         ipath_cdbg(VERBOSE, "Port%u free %u tids; first bit (max=%d) "
541                    "set is %d, porttid %u\n", pd->port_port, ti->tidcnt,
542                    limit, tid, porttid);
543         for (cnt = 0; tid < limit; tid++) {
544                 /*
545                  * small optimization; if we detect a run of 3 or so without
546                  * any set, use find_first_bit again.  That's mainly to
547                  * accelerate the case where we wrapped, so we have some at
548                  * the beginning, and some at the end, and a big gap
549                  * in the middle.
550                  */
551                 if (!test_bit(tid, tidmap))
552                         continue;
553                 cnt++;
554                 if (dd->ipath_pageshadow[porttid + tid]) {
555                         struct page *p;
556                         p = dd->ipath_pageshadow[porttid + tid];
557                         dd->ipath_pageshadow[porttid + tid] = NULL;
558                         ipath_cdbg(VERBOSE, "PID %u freeing TID %u\n",
559                                    pid_nr(pd->port_pid), tid);
560                         dd->ipath_f_put_tid(dd, &tidbase[tid],
561                                             RCVHQ_RCV_TYPE_EXPECTED,
562                                             dd->ipath_tidinvalid);
563                         pci_unmap_page(dd->pcidev,
564                                 dd->ipath_physshadow[porttid + tid],
565                                 PAGE_SIZE, PCI_DMA_FROMDEVICE);
566                         ipath_release_user_pages(&p, 1);
567                         ipath_stats.sps_pageunlocks++;
568                 } else
569                         ipath_dbg("Unused tid %u, ignoring\n", tid);
570         }
571         if (cnt != ti->tidcnt)
572                 ipath_dbg("passed in tidcnt %d, only %d bits set in map\n",
573                           ti->tidcnt, cnt);
574 done:
575         if (ret)
576                 ipath_dbg("Failed to unmap %u TID pages, failing with %d\n",
577                           ti->tidcnt, -ret);
578         return ret;
579 }
580
581 /**
582  * ipath_set_part_key - set a partition key
583  * @pd: the port
584  * @key: the key
585  *
586  * We can have up to 4 active at a time (other than the default, which is
587  * always allowed).  This is somewhat tricky, since multiple ports may set
588  * the same key, so we reference count them, and clean up at exit.  All 4
589  * partition keys are packed into a single infinipath register.  It's an
590  * error for a process to set the same pkey multiple times.  We provide no
591  * mechanism to de-allocate a pkey at this time, we may eventually need to
592  * do that.  I've used the atomic operations, and no locking, and only make
593  * a single pass through what's available.  This should be more than
594  * adequate for some time. I'll think about spinlocks or the like if and as
595  * it's necessary.
596  */
597 static int ipath_set_part_key(struct ipath_portdata *pd, u16 key)
598 {
599         struct ipath_devdata *dd = pd->port_dd;
600         int i, any = 0, pidx = -1;
601         u16 lkey = key & 0x7FFF;
602         int ret;
603
604         if (lkey == (IPATH_DEFAULT_P_KEY & 0x7FFF)) {
605                 /* nothing to do; this key always valid */
606                 ret = 0;
607                 goto bail;
608         }
609
610         ipath_cdbg(VERBOSE, "p%u try to set pkey %hx, current keys "
611                    "%hx:%x %hx:%x %hx:%x %hx:%x\n",
612                    pd->port_port, key, dd->ipath_pkeys[0],
613                    atomic_read(&dd->ipath_pkeyrefs[0]), dd->ipath_pkeys[1],
614                    atomic_read(&dd->ipath_pkeyrefs[1]), dd->ipath_pkeys[2],
615                    atomic_read(&dd->ipath_pkeyrefs[2]), dd->ipath_pkeys[3],
616                    atomic_read(&dd->ipath_pkeyrefs[3]));
617
618         if (!lkey) {
619                 ipath_cdbg(PROC, "p%u tries to set key 0, not allowed\n",
620                            pd->port_port);
621                 ret = -EINVAL;
622                 goto bail;
623         }
624
625         /*
626          * Set the full membership bit, because it has to be
627          * set in the register or the packet, and it seems
628          * cleaner to set in the register than to force all
629          * callers to set it. (see bug 4331)
630          */
631         key |= 0x8000;
632
633         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
634                 if (!pd->port_pkeys[i] && pidx == -1)
635                         pidx = i;
636                 if (pd->port_pkeys[i] == key) {
637                         ipath_cdbg(VERBOSE, "p%u tries to set same pkey "
638                                    "(%x) more than once\n",
639                                    pd->port_port, key);
640                         ret = -EEXIST;
641                         goto bail;
642                 }
643         }
644         if (pidx == -1) {
645                 ipath_dbg("All pkeys for port %u already in use, "
646                           "can't set %x\n", pd->port_port, key);
647                 ret = -EBUSY;
648                 goto bail;
649         }
650         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
651                 if (!dd->ipath_pkeys[i]) {
652                         any++;
653                         continue;
654                 }
655                 if (dd->ipath_pkeys[i] == key) {
656                         atomic_t *pkrefs = &dd->ipath_pkeyrefs[i];
657
658                         if (atomic_inc_return(pkrefs) > 1) {
659                                 pd->port_pkeys[pidx] = key;
660                                 ipath_cdbg(VERBOSE, "p%u set key %x "
661                                            "matches #%d, count now %d\n",
662                                            pd->port_port, key, i,
663                                            atomic_read(pkrefs));
664                                 ret = 0;
665                                 goto bail;
666                         } else {
667                                 /*
668                                  * lost race, decrement count, catch below
669                                  */
670                                 atomic_dec(pkrefs);
671                                 ipath_cdbg(VERBOSE, "Lost race, count was "
672                                            "0, after dec, it's %d\n",
673                                            atomic_read(pkrefs));
674                                 any++;
675                         }
676                 }
677                 if ((dd->ipath_pkeys[i] & 0x7FFF) == lkey) {
678                         /*
679                          * It makes no sense to have both the limited and
680                          * full membership PKEY set at the same time since
681                          * the unlimited one will disable the limited one.
682                          */
683                         ret = -EEXIST;
684                         goto bail;
685                 }
686         }
687         if (!any) {
688                 ipath_dbg("port %u, all pkeys already in use, "
689                           "can't set %x\n", pd->port_port, key);
690                 ret = -EBUSY;
691                 goto bail;
692         }
693         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
694                 if (!dd->ipath_pkeys[i] &&
695                     atomic_inc_return(&dd->ipath_pkeyrefs[i]) == 1) {
696                         u64 pkey;
697
698                         /* for ipathstats, etc. */
699                         ipath_stats.sps_pkeys[i] = lkey;
700                         pd->port_pkeys[pidx] = dd->ipath_pkeys[i] = key;
701                         pkey =
702                                 (u64) dd->ipath_pkeys[0] |
703                                 ((u64) dd->ipath_pkeys[1] << 16) |
704                                 ((u64) dd->ipath_pkeys[2] << 32) |
705                                 ((u64) dd->ipath_pkeys[3] << 48);
706                         ipath_cdbg(PROC, "p%u set key %x in #%d, "
707                                    "portidx %d, new pkey reg %llx\n",
708                                    pd->port_port, key, i, pidx,
709                                    (unsigned long long) pkey);
710                         ipath_write_kreg(
711                                 dd, dd->ipath_kregs->kr_partitionkey, pkey);
712
713                         ret = 0;
714                         goto bail;
715                 }
716         }
717         ipath_dbg("port %u, all pkeys already in use 2nd pass, "
718                   "can't set %x\n", pd->port_port, key);
719         ret = -EBUSY;
720
721 bail:
722         return ret;
723 }
724
725 /**
726  * ipath_manage_rcvq - manage a port's receive queue
727  * @pd: the port
728  * @subport: the subport
729  * @start_stop: action to carry out
730  *
731  * start_stop == 0 disables receive on the port, for use in queue
732  * overflow conditions.  start_stop==1 re-enables, to be used to
733  * re-init the software copy of the head register
734  */
735 static int ipath_manage_rcvq(struct ipath_portdata *pd, unsigned subport,
736                              int start_stop)
737 {
738         struct ipath_devdata *dd = pd->port_dd;
739
740         ipath_cdbg(PROC, "%sabling rcv for unit %u port %u:%u\n",
741                    start_stop ? "en" : "dis", dd->ipath_unit,
742                    pd->port_port, subport);
743         if (subport)
744                 goto bail;
745         /* atomically clear receive enable port. */
746         if (start_stop) {
747                 /*
748                  * On enable, force in-memory copy of the tail register to
749                  * 0, so that protocol code doesn't have to worry about
750                  * whether or not the chip has yet updated the in-memory
751                  * copy or not on return from the system call. The chip
752                  * always resets it's tail register back to 0 on a
753                  * transition from disabled to enabled.  This could cause a
754                  * problem if software was broken, and did the enable w/o
755                  * the disable, but eventually the in-memory copy will be
756                  * updated and correct itself, even in the face of software
757                  * bugs.
758                  */
759                 if (pd->port_rcvhdrtail_kvaddr)
760                         ipath_clear_rcvhdrtail(pd);
761                 set_bit(dd->ipath_r_portenable_shift + pd->port_port,
762                         &dd->ipath_rcvctrl);
763         } else
764                 clear_bit(dd->ipath_r_portenable_shift + pd->port_port,
765                           &dd->ipath_rcvctrl);
766         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
767                          dd->ipath_rcvctrl);
768         /* now be sure chip saw it before we return */
769         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
770         if (start_stop) {
771                 /*
772                  * And try to be sure that tail reg update has happened too.
773                  * This should in theory interlock with the RXE changes to
774                  * the tail register.  Don't assign it to the tail register
775                  * in memory copy, since we could overwrite an update by the
776                  * chip if we did.
777                  */
778                 ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
779         }
780         /* always; new head should be equal to new tail; see above */
781 bail:
782         return 0;
783 }
784
785 static void ipath_clean_part_key(struct ipath_portdata *pd,
786                                  struct ipath_devdata *dd)
787 {
788         int i, j, pchanged = 0;
789         u64 oldpkey;
790
791         /* for debugging only */
792         oldpkey = (u64) dd->ipath_pkeys[0] |
793                 ((u64) dd->ipath_pkeys[1] << 16) |
794                 ((u64) dd->ipath_pkeys[2] << 32) |
795                 ((u64) dd->ipath_pkeys[3] << 48);
796
797         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
798                 if (!pd->port_pkeys[i])
799                         continue;
800                 ipath_cdbg(VERBOSE, "look for key[%d] %hx in pkeys\n", i,
801                            pd->port_pkeys[i]);
802                 for (j = 0; j < ARRAY_SIZE(dd->ipath_pkeys); j++) {
803                         /* check for match independent of the global bit */
804                         if ((dd->ipath_pkeys[j] & 0x7fff) !=
805                             (pd->port_pkeys[i] & 0x7fff))
806                                 continue;
807                         if (atomic_dec_and_test(&dd->ipath_pkeyrefs[j])) {
808                                 ipath_cdbg(VERBOSE, "p%u clear key "
809                                            "%x matches #%d\n",
810                                            pd->port_port,
811                                            pd->port_pkeys[i], j);
812                                 ipath_stats.sps_pkeys[j] =
813                                         dd->ipath_pkeys[j] = 0;
814                                 pchanged++;
815                         }
816                         else ipath_cdbg(
817                                 VERBOSE, "p%u key %x matches #%d, "
818                                 "but ref still %d\n", pd->port_port,
819                                 pd->port_pkeys[i], j,
820                                 atomic_read(&dd->ipath_pkeyrefs[j]));
821                         break;
822                 }
823                 pd->port_pkeys[i] = 0;
824         }
825         if (pchanged) {
826                 u64 pkey = (u64) dd->ipath_pkeys[0] |
827                         ((u64) dd->ipath_pkeys[1] << 16) |
828                         ((u64) dd->ipath_pkeys[2] << 32) |
829                         ((u64) dd->ipath_pkeys[3] << 48);
830                 ipath_cdbg(VERBOSE, "p%u old pkey reg %llx, "
831                            "new pkey reg %llx\n", pd->port_port,
832                            (unsigned long long) oldpkey,
833                            (unsigned long long) pkey);
834                 ipath_write_kreg(dd, dd->ipath_kregs->kr_partitionkey,
835                                  pkey);
836         }
837 }
838
839 /*
840  * Initialize the port data with the receive buffer sizes
841  * so this can be done while the master port is locked.
842  * Otherwise, there is a race with a slave opening the port
843  * and seeing these fields uninitialized.
844  */
845 static void init_user_egr_sizes(struct ipath_portdata *pd)
846 {
847         struct ipath_devdata *dd = pd->port_dd;
848         unsigned egrperchunk, egrcnt, size;
849
850         /*
851          * to avoid wasting a lot of memory, we allocate 32KB chunks of
852          * physically contiguous memory, advance through it until used up
853          * and then allocate more.  Of course, we need memory to store those
854          * extra pointers, now.  Started out with 256KB, but under heavy
855          * memory pressure (creating large files and then copying them over
856          * NFS while doing lots of MPI jobs), we hit some allocation
857          * failures, even though we can sleep...  (2.6.10) Still get
858          * failures at 64K.  32K is the lowest we can go without wasting
859          * additional memory.
860          */
861         size = 0x8000;
862         egrperchunk = size / dd->ipath_rcvegrbufsize;
863         egrcnt = dd->ipath_rcvegrcnt;
864         pd->port_rcvegrbuf_chunks = (egrcnt + egrperchunk - 1) / egrperchunk;
865         pd->port_rcvegrbufs_perchunk = egrperchunk;
866         pd->port_rcvegrbuf_size = size;
867 }
868
869 /**
870  * ipath_create_user_egr - allocate eager TID buffers
871  * @pd: the port to allocate TID buffers for
872  *
873  * This routine is now quite different for user and kernel, because
874  * the kernel uses skb's, for the accelerated network performance
875  * This is the user port version
876  *
877  * Allocate the eager TID buffers and program them into infinipath
878  * They are no longer completely contiguous, we do multiple allocation
879  * calls.
880  */
881 static int ipath_create_user_egr(struct ipath_portdata *pd)
882 {
883         struct ipath_devdata *dd = pd->port_dd;
884         unsigned e, egrcnt, egrperchunk, chunk, egrsize, egroff;
885         size_t size;
886         int ret;
887         gfp_t gfp_flags;
888
889         /*
890          * GFP_USER, but without GFP_FS, so buffer cache can be
891          * coalesced (we hope); otherwise, even at order 4,
892          * heavy filesystem activity makes these fail, and we can
893          * use compound pages.
894          */
895         gfp_flags = __GFP_WAIT | __GFP_IO | __GFP_COMP;
896
897         egrcnt = dd->ipath_rcvegrcnt;
898         /* TID number offset for this port */
899         egroff = (pd->port_port - 1) * egrcnt + dd->ipath_p0_rcvegrcnt;
900         egrsize = dd->ipath_rcvegrbufsize;
901         ipath_cdbg(VERBOSE, "Allocating %d egr buffers, at egrtid "
902                    "offset %x, egrsize %u\n", egrcnt, egroff, egrsize);
903
904         chunk = pd->port_rcvegrbuf_chunks;
905         egrperchunk = pd->port_rcvegrbufs_perchunk;
906         size = pd->port_rcvegrbuf_size;
907         pd->port_rcvegrbuf = kmalloc(chunk * sizeof(pd->port_rcvegrbuf[0]),
908                                      GFP_KERNEL);
909         if (!pd->port_rcvegrbuf) {
910                 ret = -ENOMEM;
911                 goto bail;
912         }
913         pd->port_rcvegrbuf_phys =
914                 kmalloc(chunk * sizeof(pd->port_rcvegrbuf_phys[0]),
915                         GFP_KERNEL);
916         if (!pd->port_rcvegrbuf_phys) {
917                 ret = -ENOMEM;
918                 goto bail_rcvegrbuf;
919         }
920         for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
921
922                 pd->port_rcvegrbuf[e] = dma_alloc_coherent(
923                         &dd->pcidev->dev, size, &pd->port_rcvegrbuf_phys[e],
924                         gfp_flags);
925
926                 if (!pd->port_rcvegrbuf[e]) {
927                         ret = -ENOMEM;
928                         goto bail_rcvegrbuf_phys;
929                 }
930         }
931
932         pd->port_rcvegr_phys = pd->port_rcvegrbuf_phys[0];
933
934         for (e = chunk = 0; chunk < pd->port_rcvegrbuf_chunks; chunk++) {
935                 dma_addr_t pa = pd->port_rcvegrbuf_phys[chunk];
936                 unsigned i;
937
938                 for (i = 0; e < egrcnt && i < egrperchunk; e++, i++) {
939                         dd->ipath_f_put_tid(dd, e + egroff +
940                                             (u64 __iomem *)
941                                             ((char __iomem *)
942                                              dd->ipath_kregbase +
943                                              dd->ipath_rcvegrbase),
944                                             RCVHQ_RCV_TYPE_EAGER, pa);
945                         pa += egrsize;
946                 }
947                 cond_resched(); /* don't hog the cpu */
948         }
949
950         ret = 0;
951         goto bail;
952
953 bail_rcvegrbuf_phys:
954         for (e = 0; e < pd->port_rcvegrbuf_chunks &&
955                 pd->port_rcvegrbuf[e]; e++) {
956                 dma_free_coherent(&dd->pcidev->dev, size,
957                                   pd->port_rcvegrbuf[e],
958                                   pd->port_rcvegrbuf_phys[e]);
959
960         }
961         kfree(pd->port_rcvegrbuf_phys);
962         pd->port_rcvegrbuf_phys = NULL;
963 bail_rcvegrbuf:
964         kfree(pd->port_rcvegrbuf);
965         pd->port_rcvegrbuf = NULL;
966 bail:
967         return ret;
968 }
969
970
971 /* common code for the mappings on dma_alloc_coherent mem */
972 static int ipath_mmap_mem(struct vm_area_struct *vma,
973         struct ipath_portdata *pd, unsigned len, int write_ok,
974         void *kvaddr, char *what)
975 {
976         struct ipath_devdata *dd = pd->port_dd;
977         unsigned long pfn;
978         int ret;
979
980         if ((vma->vm_end - vma->vm_start) > len) {
981                 dev_info(&dd->pcidev->dev,
982                          "FAIL on %s: len %lx > %x\n", what,
983                          vma->vm_end - vma->vm_start, len);
984                 ret = -EFAULT;
985                 goto bail;
986         }
987
988         if (!write_ok) {
989                 if (vma->vm_flags & VM_WRITE) {
990                         dev_info(&dd->pcidev->dev,
991                                  "%s must be mapped readonly\n", what);
992                         ret = -EPERM;
993                         goto bail;
994                 }
995
996                 /* don't allow them to later change with mprotect */
997                 vma->vm_flags &= ~VM_MAYWRITE;
998         }
999
1000         pfn = virt_to_phys(kvaddr) >> PAGE_SHIFT;
1001         ret = remap_pfn_range(vma, vma->vm_start, pfn,
1002                               len, vma->vm_page_prot);
1003         if (ret)
1004                 dev_info(&dd->pcidev->dev, "%s port%u mmap of %lx, %x "
1005                          "bytes r%c failed: %d\n", what, pd->port_port,
1006                          pfn, len, write_ok?'w':'o', ret);
1007         else
1008                 ipath_cdbg(VERBOSE, "%s port%u mmaped %lx, %x bytes "
1009                            "r%c\n", what, pd->port_port, pfn, len,
1010                            write_ok?'w':'o');
1011 bail:
1012         return ret;
1013 }
1014
1015 static int mmap_ureg(struct vm_area_struct *vma, struct ipath_devdata *dd,
1016                      u64 ureg)
1017 {
1018         unsigned long phys;
1019         int ret;
1020
1021         /*
1022          * This is real hardware, so use io_remap.  This is the mechanism
1023          * for the user process to update the head registers for their port
1024          * in the chip.
1025          */
1026         if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) {
1027                 dev_info(&dd->pcidev->dev, "FAIL mmap userreg: reqlen "
1028                          "%lx > PAGE\n", vma->vm_end - vma->vm_start);
1029                 ret = -EFAULT;
1030         } else {
1031                 phys = dd->ipath_physaddr + ureg;
1032                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1033
1034                 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1035                 ret = io_remap_pfn_range(vma, vma->vm_start,
1036                                          phys >> PAGE_SHIFT,
1037                                          vma->vm_end - vma->vm_start,
1038                                          vma->vm_page_prot);
1039         }
1040         return ret;
1041 }
1042
1043 static int mmap_piobufs(struct vm_area_struct *vma,
1044                         struct ipath_devdata *dd,
1045                         struct ipath_portdata *pd,
1046                         unsigned piobufs, unsigned piocnt)
1047 {
1048         unsigned long phys;
1049         int ret;
1050
1051         /*
1052          * When we map the PIO buffers in the chip, we want to map them as
1053          * writeonly, no read possible.   This prevents access to previous
1054          * process data, and catches users who might try to read the i/o
1055          * space due to a bug.
1056          */
1057         if ((vma->vm_end - vma->vm_start) > (piocnt * dd->ipath_palign)) {
1058                 dev_info(&dd->pcidev->dev, "FAIL mmap piobufs: "
1059                          "reqlen %lx > PAGE\n",
1060                          vma->vm_end - vma->vm_start);
1061                 ret = -EINVAL;
1062                 goto bail;
1063         }
1064
1065         phys = dd->ipath_physaddr + piobufs;
1066
1067 #if defined(__powerpc__)
1068         /* There isn't a generic way to specify writethrough mappings */
1069         pgprot_val(vma->vm_page_prot) |= _PAGE_NO_CACHE;
1070         pgprot_val(vma->vm_page_prot) |= _PAGE_WRITETHRU;
1071         pgprot_val(vma->vm_page_prot) &= ~_PAGE_GUARDED;
1072 #endif
1073
1074         /*
1075          * don't allow them to later change to readable with mprotect (for when
1076          * not initially mapped readable, as is normally the case)
1077          */
1078         vma->vm_flags &= ~VM_MAYREAD;
1079         vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1080
1081         ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
1082                                  vma->vm_end - vma->vm_start,
1083                                  vma->vm_page_prot);
1084 bail:
1085         return ret;
1086 }
1087
1088 static int mmap_rcvegrbufs(struct vm_area_struct *vma,
1089                            struct ipath_portdata *pd)
1090 {
1091         struct ipath_devdata *dd = pd->port_dd;
1092         unsigned long start, size;
1093         size_t total_size, i;
1094         unsigned long pfn;
1095         int ret;
1096
1097         size = pd->port_rcvegrbuf_size;
1098         total_size = pd->port_rcvegrbuf_chunks * size;
1099         if ((vma->vm_end - vma->vm_start) > total_size) {
1100                 dev_info(&dd->pcidev->dev, "FAIL on egr bufs: "
1101                          "reqlen %lx > actual %lx\n",
1102                          vma->vm_end - vma->vm_start,
1103                          (unsigned long) total_size);
1104                 ret = -EINVAL;
1105                 goto bail;
1106         }
1107
1108         if (vma->vm_flags & VM_WRITE) {
1109                 dev_info(&dd->pcidev->dev, "Can't map eager buffers as "
1110                          "writable (flags=%lx)\n", vma->vm_flags);
1111                 ret = -EPERM;
1112                 goto bail;
1113         }
1114         /* don't allow them to later change to writeable with mprotect */
1115         vma->vm_flags &= ~VM_MAYWRITE;
1116
1117         start = vma->vm_start;
1118
1119         for (i = 0; i < pd->port_rcvegrbuf_chunks; i++, start += size) {
1120                 pfn = virt_to_phys(pd->port_rcvegrbuf[i]) >> PAGE_SHIFT;
1121                 ret = remap_pfn_range(vma, start, pfn, size,
1122                                       vma->vm_page_prot);
1123                 if (ret < 0)
1124                         goto bail;
1125         }
1126         ret = 0;
1127
1128 bail:
1129         return ret;
1130 }
1131
1132 /*
1133  * ipath_file_vma_fault - handle a VMA page fault.
1134  */
1135 static int ipath_file_vma_fault(struct vm_area_struct *vma,
1136                                         struct vm_fault *vmf)
1137 {
1138         struct page *page;
1139
1140         page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
1141         if (!page)
1142                 return VM_FAULT_SIGBUS;
1143         get_page(page);
1144         vmf->page = page;
1145
1146         return 0;
1147 }
1148
1149 static struct vm_operations_struct ipath_file_vm_ops = {
1150         .fault = ipath_file_vma_fault,
1151 };
1152
1153 static int mmap_kvaddr(struct vm_area_struct *vma, u64 pgaddr,
1154                        struct ipath_portdata *pd, unsigned subport)
1155 {
1156         unsigned long len;
1157         struct ipath_devdata *dd;
1158         void *addr;
1159         size_t size;
1160         int ret = 0;
1161
1162         /* If the port is not shared, all addresses should be physical */
1163         if (!pd->port_subport_cnt)
1164                 goto bail;
1165
1166         dd = pd->port_dd;
1167         size = pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
1168
1169         /*
1170          * Each process has all the subport uregbase, rcvhdrq, and
1171          * rcvegrbufs mmapped - as an array for all the processes,
1172          * and also separately for this process.
1173          */
1174         if (pgaddr == cvt_kvaddr(pd->subport_uregbase)) {
1175                 addr = pd->subport_uregbase;
1176                 size = PAGE_SIZE * pd->port_subport_cnt;
1177         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base)) {
1178                 addr = pd->subport_rcvhdr_base;
1179                 size = pd->port_rcvhdrq_size * pd->port_subport_cnt;
1180         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf)) {
1181                 addr = pd->subport_rcvegrbuf;
1182                 size *= pd->port_subport_cnt;
1183         } else if (pgaddr == cvt_kvaddr(pd->subport_uregbase +
1184                                         PAGE_SIZE * subport)) {
1185                 addr = pd->subport_uregbase + PAGE_SIZE * subport;
1186                 size = PAGE_SIZE;
1187         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base +
1188                                 pd->port_rcvhdrq_size * subport)) {
1189                 addr = pd->subport_rcvhdr_base +
1190                         pd->port_rcvhdrq_size * subport;
1191                 size = pd->port_rcvhdrq_size;
1192         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf +
1193                                size * subport)) {
1194                 addr = pd->subport_rcvegrbuf + size * subport;
1195                 /* rcvegrbufs are read-only on the slave */
1196                 if (vma->vm_flags & VM_WRITE) {
1197                         dev_info(&dd->pcidev->dev,
1198                                  "Can't map eager buffers as "
1199                                  "writable (flags=%lx)\n", vma->vm_flags);
1200                         ret = -EPERM;
1201                         goto bail;
1202                 }
1203                 /*
1204                  * Don't allow permission to later change to writeable
1205                  * with mprotect.
1206                  */
1207                 vma->vm_flags &= ~VM_MAYWRITE;
1208         } else {
1209                 goto bail;
1210         }
1211         len = vma->vm_end - vma->vm_start;
1212         if (len > size) {
1213                 ipath_cdbg(MM, "FAIL: reqlen %lx > %zx\n", len, size);
1214                 ret = -EINVAL;
1215                 goto bail;
1216         }
1217
1218         vma->vm_pgoff = (unsigned long) addr >> PAGE_SHIFT;
1219         vma->vm_ops = &ipath_file_vm_ops;
1220         vma->vm_flags |= VM_RESERVED | VM_DONTEXPAND;
1221         ret = 1;
1222
1223 bail:
1224         return ret;
1225 }
1226
1227 /**
1228  * ipath_mmap - mmap various structures into user space
1229  * @fp: the file pointer
1230  * @vma: the VM area
1231  *
1232  * We use this to have a shared buffer between the kernel and the user code
1233  * for the rcvhdr queue, egr buffers, and the per-port user regs and pio
1234  * buffers in the chip.  We have the open and close entries so we can bump
1235  * the ref count and keep the driver from being unloaded while still mapped.
1236  */
1237 static int ipath_mmap(struct file *fp, struct vm_area_struct *vma)
1238 {
1239         struct ipath_portdata *pd;
1240         struct ipath_devdata *dd;
1241         u64 pgaddr, ureg;
1242         unsigned piobufs, piocnt;
1243         int ret;
1244
1245         pd = port_fp(fp);
1246         if (!pd) {
1247                 ret = -EINVAL;
1248                 goto bail;
1249         }
1250         dd = pd->port_dd;
1251
1252         /*
1253          * This is the ipath_do_user_init() code, mapping the shared buffers
1254          * into the user process. The address referred to by vm_pgoff is the
1255          * file offset passed via mmap().  For shared ports, this is the
1256          * kernel vmalloc() address of the pages to share with the master.
1257          * For non-shared or master ports, this is a physical address.
1258          * We only do one mmap for each space mapped.
1259          */
1260         pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1261
1262         /*
1263          * Check for 0 in case one of the allocations failed, but user
1264          * called mmap anyway.
1265          */
1266         if (!pgaddr)  {
1267                 ret = -EINVAL;
1268                 goto bail;
1269         }
1270
1271         ipath_cdbg(MM, "pgaddr %llx vm_start=%lx len %lx port %u:%u:%u\n",
1272                    (unsigned long long) pgaddr, vma->vm_start,
1273                    vma->vm_end - vma->vm_start, dd->ipath_unit,
1274                    pd->port_port, subport_fp(fp));
1275
1276         /*
1277          * Physical addresses must fit in 40 bits for our hardware.
1278          * Check for kernel virtual addresses first, anything else must
1279          * match a HW or memory address.
1280          */
1281         ret = mmap_kvaddr(vma, pgaddr, pd, subport_fp(fp));
1282         if (ret) {
1283                 if (ret > 0)
1284                         ret = 0;
1285                 goto bail;
1286         }
1287
1288         ureg = dd->ipath_uregbase + dd->ipath_ureg_align * pd->port_port;
1289         if (!pd->port_subport_cnt) {
1290                 /* port is not shared */
1291                 piocnt = pd->port_piocnt;
1292                 piobufs = pd->port_piobufs;
1293         } else if (!subport_fp(fp)) {
1294                 /* caller is the master */
1295                 piocnt = (pd->port_piocnt / pd->port_subport_cnt) +
1296                          (pd->port_piocnt % pd->port_subport_cnt);
1297                 piobufs = pd->port_piobufs +
1298                         dd->ipath_palign * (pd->port_piocnt - piocnt);
1299         } else {
1300                 unsigned slave = subport_fp(fp) - 1;
1301
1302                 /* caller is a slave */
1303                 piocnt = pd->port_piocnt / pd->port_subport_cnt;
1304                 piobufs = pd->port_piobufs + dd->ipath_palign * piocnt * slave;
1305         }
1306
1307         if (pgaddr == ureg)
1308                 ret = mmap_ureg(vma, dd, ureg);
1309         else if (pgaddr == piobufs)
1310                 ret = mmap_piobufs(vma, dd, pd, piobufs, piocnt);
1311         else if (pgaddr == dd->ipath_pioavailregs_phys)
1312                 /* in-memory copy of pioavail registers */
1313                 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1314                                      (void *) dd->ipath_pioavailregs_dma,
1315                                      "pioavail registers");
1316         else if (pgaddr == pd->port_rcvegr_phys)
1317                 ret = mmap_rcvegrbufs(vma, pd);
1318         else if (pgaddr == (u64) pd->port_rcvhdrq_phys)
1319                 /*
1320                  * The rcvhdrq itself; readonly except on HT (so have
1321                  * to allow writable mapping), multiple pages, contiguous
1322                  * from an i/o perspective.
1323                  */
1324                 ret = ipath_mmap_mem(vma, pd, pd->port_rcvhdrq_size, 1,
1325                                      pd->port_rcvhdrq,
1326                                      "rcvhdrq");
1327         else if (pgaddr == (u64) pd->port_rcvhdrqtailaddr_phys)
1328                 /* in-memory copy of rcvhdrq tail register */
1329                 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1330                                      pd->port_rcvhdrtail_kvaddr,
1331                                      "rcvhdrq tail");
1332         else
1333                 ret = -EINVAL;
1334
1335         vma->vm_private_data = NULL;
1336
1337         if (ret < 0)
1338                 dev_info(&dd->pcidev->dev,
1339                          "Failure %d on off %llx len %lx\n",
1340                          -ret, (unsigned long long)pgaddr,
1341                          vma->vm_end - vma->vm_start);
1342 bail:
1343         return ret;
1344 }
1345
1346 static unsigned ipath_poll_hdrqfull(struct ipath_portdata *pd)
1347 {
1348         unsigned pollflag = 0;
1349
1350         if ((pd->poll_type & IPATH_POLL_TYPE_OVERFLOW) &&
1351             pd->port_hdrqfull != pd->port_hdrqfull_poll) {
1352                 pollflag |= POLLIN | POLLRDNORM;
1353                 pd->port_hdrqfull_poll = pd->port_hdrqfull;
1354         }
1355
1356         return pollflag;
1357 }
1358
1359 static unsigned int ipath_poll_urgent(struct ipath_portdata *pd,
1360                                       struct file *fp,
1361                                       struct poll_table_struct *pt)
1362 {
1363         unsigned pollflag = 0;
1364         struct ipath_devdata *dd;
1365
1366         dd = pd->port_dd;
1367
1368         /* variable access in ipath_poll_hdrqfull() needs this */
1369         rmb();
1370         pollflag = ipath_poll_hdrqfull(pd);
1371
1372         if (pd->port_urgent != pd->port_urgent_poll) {
1373                 pollflag |= POLLIN | POLLRDNORM;
1374                 pd->port_urgent_poll = pd->port_urgent;
1375         }
1376
1377         if (!pollflag) {
1378                 /* this saves a spin_lock/unlock in interrupt handler... */
1379                 set_bit(IPATH_PORT_WAITING_URG, &pd->port_flag);
1380                 /* flush waiting flag so don't miss an event... */
1381                 wmb();
1382                 poll_wait(fp, &pd->port_wait, pt);
1383         }
1384
1385         return pollflag;
1386 }
1387
1388 static unsigned int ipath_poll_next(struct ipath_portdata *pd,
1389                                     struct file *fp,
1390                                     struct poll_table_struct *pt)
1391 {
1392         u32 head;
1393         u32 tail;
1394         unsigned pollflag = 0;
1395         struct ipath_devdata *dd;
1396
1397         dd = pd->port_dd;
1398
1399         /* variable access in ipath_poll_hdrqfull() needs this */
1400         rmb();
1401         pollflag = ipath_poll_hdrqfull(pd);
1402
1403         head = ipath_read_ureg32(dd, ur_rcvhdrhead, pd->port_port);
1404         if (pd->port_rcvhdrtail_kvaddr)
1405                 tail = ipath_get_rcvhdrtail(pd);
1406         else
1407                 tail = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
1408
1409         if (head != tail)
1410                 pollflag |= POLLIN | POLLRDNORM;
1411         else {
1412                 /* this saves a spin_lock/unlock in interrupt handler */
1413                 set_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
1414                 /* flush waiting flag so we don't miss an event */
1415                 wmb();
1416
1417                 set_bit(pd->port_port + dd->ipath_r_intravail_shift,
1418                         &dd->ipath_rcvctrl);
1419
1420                 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1421                                  dd->ipath_rcvctrl);
1422
1423                 if (dd->ipath_rhdrhead_intr_off) /* arm rcv interrupt */
1424                         ipath_write_ureg(dd, ur_rcvhdrhead,
1425                                          dd->ipath_rhdrhead_intr_off | head,
1426                                          pd->port_port);
1427
1428                 poll_wait(fp, &pd->port_wait, pt);
1429         }
1430
1431         return pollflag;
1432 }
1433
1434 static unsigned int ipath_poll(struct file *fp,
1435                                struct poll_table_struct *pt)
1436 {
1437         struct ipath_portdata *pd;
1438         unsigned pollflag;
1439
1440         pd = port_fp(fp);
1441         if (!pd)
1442                 pollflag = 0;
1443         else if (pd->poll_type & IPATH_POLL_TYPE_URGENT)
1444                 pollflag = ipath_poll_urgent(pd, fp, pt);
1445         else
1446                 pollflag = ipath_poll_next(pd, fp, pt);
1447
1448         return pollflag;
1449 }
1450
1451 static int ipath_supports_subports(int user_swmajor, int user_swminor)
1452 {
1453         /* no subport implementation prior to software version 1.3 */
1454         return (user_swmajor > 1) || (user_swminor >= 3);
1455 }
1456
1457 static int ipath_compatible_subports(int user_swmajor, int user_swminor)
1458 {
1459         /* this code is written long-hand for clarity */
1460         if (IPATH_USER_SWMAJOR != user_swmajor) {
1461                 /* no promise of compatibility if major mismatch */
1462                 return 0;
1463         }
1464         if (IPATH_USER_SWMAJOR == 1) {
1465                 switch (IPATH_USER_SWMINOR) {
1466                 case 0:
1467                 case 1:
1468                 case 2:
1469                         /* no subport implementation so cannot be compatible */
1470                         return 0;
1471                 case 3:
1472                         /* 3 is only compatible with itself */
1473                         return user_swminor == 3;
1474                 default:
1475                         /* >= 4 are compatible (or are expected to be) */
1476                         return user_swminor >= 4;
1477                 }
1478         }
1479         /* make no promises yet for future major versions */
1480         return 0;
1481 }
1482
1483 static int init_subports(struct ipath_devdata *dd,
1484                          struct ipath_portdata *pd,
1485                          const struct ipath_user_info *uinfo)
1486 {
1487         int ret = 0;
1488         unsigned num_subports;
1489         size_t size;
1490
1491         /*
1492          * If the user is requesting zero subports,
1493          * skip the subport allocation.
1494          */
1495         if (uinfo->spu_subport_cnt <= 0)
1496                 goto bail;
1497
1498         /* Self-consistency check for ipath_compatible_subports() */
1499         if (ipath_supports_subports(IPATH_USER_SWMAJOR, IPATH_USER_SWMINOR) &&
1500             !ipath_compatible_subports(IPATH_USER_SWMAJOR,
1501                                        IPATH_USER_SWMINOR)) {
1502                 dev_info(&dd->pcidev->dev,
1503                          "Inconsistent ipath_compatible_subports()\n");
1504                 goto bail;
1505         }
1506
1507         /* Check for subport compatibility */
1508         if (!ipath_compatible_subports(uinfo->spu_userversion >> 16,
1509                                        uinfo->spu_userversion & 0xffff)) {
1510                 dev_info(&dd->pcidev->dev,
1511                          "Mismatched user version (%d.%d) and driver "
1512                          "version (%d.%d) while port sharing. Ensure "
1513                          "that driver and library are from the same "
1514                          "release.\n",
1515                          (int) (uinfo->spu_userversion >> 16),
1516                          (int) (uinfo->spu_userversion & 0xffff),
1517                          IPATH_USER_SWMAJOR,
1518                          IPATH_USER_SWMINOR);
1519                 goto bail;
1520         }
1521         if (uinfo->spu_subport_cnt > INFINIPATH_MAX_SUBPORT) {
1522                 ret = -EINVAL;
1523                 goto bail;
1524         }
1525
1526         num_subports = uinfo->spu_subport_cnt;
1527         pd->subport_uregbase = vmalloc(PAGE_SIZE * num_subports);
1528         if (!pd->subport_uregbase) {
1529                 ret = -ENOMEM;
1530                 goto bail;
1531         }
1532         /* Note: pd->port_rcvhdrq_size isn't initialized yet. */
1533         size = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
1534                      sizeof(u32), PAGE_SIZE) * num_subports;
1535         pd->subport_rcvhdr_base = vmalloc(size);
1536         if (!pd->subport_rcvhdr_base) {
1537                 ret = -ENOMEM;
1538                 goto bail_ureg;
1539         }
1540
1541         pd->subport_rcvegrbuf = vmalloc(pd->port_rcvegrbuf_chunks *
1542                                         pd->port_rcvegrbuf_size *
1543                                         num_subports);
1544         if (!pd->subport_rcvegrbuf) {
1545                 ret = -ENOMEM;
1546                 goto bail_rhdr;
1547         }
1548
1549         pd->port_subport_cnt = uinfo->spu_subport_cnt;
1550         pd->port_subport_id = uinfo->spu_subport_id;
1551         pd->active_slaves = 1;
1552         set_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
1553         memset(pd->subport_uregbase, 0, PAGE_SIZE * num_subports);
1554         memset(pd->subport_rcvhdr_base, 0, size);
1555         memset(pd->subport_rcvegrbuf, 0, pd->port_rcvegrbuf_chunks *
1556                                          pd->port_rcvegrbuf_size *
1557                                          num_subports);
1558         goto bail;
1559
1560 bail_rhdr:
1561         vfree(pd->subport_rcvhdr_base);
1562 bail_ureg:
1563         vfree(pd->subport_uregbase);
1564         pd->subport_uregbase = NULL;
1565 bail:
1566         return ret;
1567 }
1568
1569 static int try_alloc_port(struct ipath_devdata *dd, int port,
1570                           struct file *fp,
1571                           const struct ipath_user_info *uinfo)
1572 {
1573         struct ipath_portdata *pd;
1574         int ret;
1575
1576         if (!(pd = dd->ipath_pd[port])) {
1577                 void *ptmp;
1578
1579                 pd = kzalloc(sizeof(struct ipath_portdata), GFP_KERNEL);
1580
1581                 /*
1582                  * Allocate memory for use in ipath_tid_update() just once
1583                  * at open, not per call.  Reduces cost of expected send
1584                  * setup.
1585                  */
1586                 ptmp = kmalloc(dd->ipath_rcvtidcnt * sizeof(u16) +
1587                                dd->ipath_rcvtidcnt * sizeof(struct page **),
1588                                GFP_KERNEL);
1589                 if (!pd || !ptmp) {
1590                         ipath_dev_err(dd, "Unable to allocate portdata "
1591                                       "memory, failing open\n");
1592                         ret = -ENOMEM;
1593                         kfree(pd);
1594                         kfree(ptmp);
1595                         goto bail;
1596                 }
1597                 dd->ipath_pd[port] = pd;
1598                 dd->ipath_pd[port]->port_port = port;
1599                 dd->ipath_pd[port]->port_dd = dd;
1600                 dd->ipath_pd[port]->port_tid_pg_list = ptmp;
1601                 init_waitqueue_head(&dd->ipath_pd[port]->port_wait);
1602         }
1603         if (!pd->port_cnt) {
1604                 pd->userversion = uinfo->spu_userversion;
1605                 init_user_egr_sizes(pd);
1606                 if ((ret = init_subports(dd, pd, uinfo)) != 0)
1607                         goto bail;
1608                 ipath_cdbg(PROC, "%s[%u] opened unit:port %u:%u\n",
1609                            current->comm, current->pid, dd->ipath_unit,
1610                            port);
1611                 pd->port_cnt = 1;
1612                 port_fp(fp) = pd;
1613                 pd->port_pid = get_pid(task_pid(current));
1614                 strncpy(pd->port_comm, current->comm, sizeof(pd->port_comm));
1615                 ipath_stats.sps_ports++;
1616                 ret = 0;
1617         } else
1618                 ret = -EBUSY;
1619
1620 bail:
1621         return ret;
1622 }
1623
1624 static inline int usable(struct ipath_devdata *dd)
1625 {
1626         return dd &&
1627                 (dd->ipath_flags & IPATH_PRESENT) &&
1628                 dd->ipath_kregbase &&
1629                 dd->ipath_lid &&
1630                 !(dd->ipath_flags & (IPATH_LINKDOWN | IPATH_DISABLED
1631                                      | IPATH_LINKUNK));
1632 }
1633
1634 static int find_free_port(int unit, struct file *fp,
1635                           const struct ipath_user_info *uinfo)
1636 {
1637         struct ipath_devdata *dd = ipath_lookup(unit);
1638         int ret, i;
1639
1640         if (!dd) {
1641                 ret = -ENODEV;
1642                 goto bail;
1643         }
1644
1645         if (!usable(dd)) {
1646                 ret = -ENETDOWN;
1647                 goto bail;
1648         }
1649
1650         for (i = 1; i < dd->ipath_cfgports; i++) {
1651                 ret = try_alloc_port(dd, i, fp, uinfo);
1652                 if (ret != -EBUSY)
1653                         goto bail;
1654         }
1655         ret = -EBUSY;
1656
1657 bail:
1658         return ret;
1659 }
1660
1661 static int find_best_unit(struct file *fp,
1662                           const struct ipath_user_info *uinfo)
1663 {
1664         int ret = 0, i, prefunit = -1, devmax;
1665         int maxofallports, npresent, nup;
1666         int ndev;
1667
1668         devmax = ipath_count_units(&npresent, &nup, &maxofallports);
1669
1670         /*
1671          * This code is present to allow a knowledgeable person to
1672          * specify the layout of processes to processors before opening
1673          * this driver, and then we'll assign the process to the "closest"
1674          * InfiniPath chip to that processor (we assume reasonable connectivity,
1675          * for now).  This code assumes that if affinity has been set
1676          * before this point, that at most one cpu is set; for now this
1677          * is reasonable.  I check for both cpus_empty() and cpus_full(),
1678          * in case some kernel variant sets none of the bits when no
1679          * affinity is set.  2.6.11 and 12 kernels have all present
1680          * cpus set.  Some day we'll have to fix it up further to handle
1681          * a cpu subset.  This algorithm fails for two HT chips connected
1682          * in tunnel fashion.  Eventually this needs real topology
1683          * information.  There may be some issues with dual core numbering
1684          * as well.  This needs more work prior to release.
1685          */
1686         if (!cpus_empty(current->cpus_allowed) &&
1687             !cpus_full(current->cpus_allowed)) {
1688                 int ncpus = num_online_cpus(), curcpu = -1, nset = 0;
1689                 for (i = 0; i < ncpus; i++)
1690                         if (cpu_isset(i, current->cpus_allowed)) {
1691                                 ipath_cdbg(PROC, "%s[%u] affinity set for "
1692                                            "cpu %d/%d\n", current->comm,
1693                                            current->pid, i, ncpus);
1694                                 curcpu = i;
1695                                 nset++;
1696                         }
1697                 if (curcpu != -1 && nset != ncpus) {
1698                         if (npresent) {
1699                                 prefunit = curcpu / (ncpus / npresent);
1700                                 ipath_cdbg(PROC,"%s[%u] %d chips, %d cpus, "
1701                                           "%d cpus/chip, select unit %d\n",
1702                                           current->comm, current->pid,
1703                                           npresent, ncpus, ncpus / npresent,
1704                                           prefunit);
1705                         }
1706                 }
1707         }
1708
1709         /*
1710          * user ports start at 1, kernel port is 0
1711          * For now, we do round-robin access across all chips
1712          */
1713
1714         if (prefunit != -1)
1715                 devmax = prefunit + 1;
1716 recheck:
1717         for (i = 1; i < maxofallports; i++) {
1718                 for (ndev = prefunit != -1 ? prefunit : 0; ndev < devmax;
1719                      ndev++) {
1720                         struct ipath_devdata *dd = ipath_lookup(ndev);
1721
1722                         if (!usable(dd))
1723                                 continue; /* can't use this unit */
1724                         if (i >= dd->ipath_cfgports)
1725                                 /*
1726                                  * Maxed out on users of this unit. Try
1727                                  * next.
1728                                  */
1729                                 continue;
1730                         ret = try_alloc_port(dd, i, fp, uinfo);
1731                         if (!ret)
1732                                 goto done;
1733                 }
1734         }
1735
1736         if (npresent) {
1737                 if (nup == 0) {
1738                         ret = -ENETDOWN;
1739                         ipath_dbg("No ports available (none initialized "
1740                                   "and ready)\n");
1741                 } else {
1742                         if (prefunit > 0) {
1743                                 /* if started above 0, retry from 0 */
1744                                 ipath_cdbg(PROC,
1745                                            "%s[%u] no ports on prefunit "
1746                                            "%d, clear and re-check\n",
1747                                            current->comm, current->pid,
1748                                            prefunit);
1749                                 devmax = ipath_count_units(NULL, NULL,
1750                                                            NULL);
1751                                 prefunit = -1;
1752                                 goto recheck;
1753                         }
1754                         ret = -EBUSY;
1755                         ipath_dbg("No ports available\n");
1756                 }
1757         } else {
1758                 ret = -ENXIO;
1759                 ipath_dbg("No boards found\n");
1760         }
1761
1762 done:
1763         return ret;
1764 }
1765
1766 static int find_shared_port(struct file *fp,
1767                             const struct ipath_user_info *uinfo)
1768 {
1769         int devmax, ndev, i;
1770         int ret = 0;
1771
1772         devmax = ipath_count_units(NULL, NULL, NULL);
1773
1774         for (ndev = 0; ndev < devmax; ndev++) {
1775                 struct ipath_devdata *dd = ipath_lookup(ndev);
1776
1777                 if (!usable(dd))
1778                         continue;
1779                 for (i = 1; i < dd->ipath_cfgports; i++) {
1780                         struct ipath_portdata *pd = dd->ipath_pd[i];
1781
1782                         /* Skip ports which are not yet open */
1783                         if (!pd || !pd->port_cnt)
1784                                 continue;
1785                         /* Skip port if it doesn't match the requested one */
1786                         if (pd->port_subport_id != uinfo->spu_subport_id)
1787                                 continue;
1788                         /* Verify the sharing process matches the master */
1789                         if (pd->port_subport_cnt != uinfo->spu_subport_cnt ||
1790                             pd->userversion != uinfo->spu_userversion ||
1791                             pd->port_cnt >= pd->port_subport_cnt) {
1792                                 ret = -EINVAL;
1793                                 goto done;
1794                         }
1795                         port_fp(fp) = pd;
1796                         subport_fp(fp) = pd->port_cnt++;
1797                         pd->port_subpid[subport_fp(fp)] =
1798                                 get_pid(task_pid(current));
1799                         tidcursor_fp(fp) = 0;
1800                         pd->active_slaves |= 1 << subport_fp(fp);
1801                         ipath_cdbg(PROC,
1802                                    "%s[%u] %u sharing %s[%u] unit:port %u:%u\n",
1803                                    current->comm, current->pid,
1804                                    subport_fp(fp),
1805                                    pd->port_comm, pid_nr(pd->port_pid),
1806                                    dd->ipath_unit, pd->port_port);
1807                         ret = 1;
1808                         goto done;
1809                 }
1810         }
1811
1812 done:
1813         return ret;
1814 }
1815
1816 static int ipath_open(struct inode *in, struct file *fp)
1817 {
1818         /* The real work is performed later in ipath_assign_port() */
1819         cycle_kernel_lock();
1820         fp->private_data = kzalloc(sizeof(struct ipath_filedata), GFP_KERNEL);
1821         return fp->private_data ? 0 : -ENOMEM;
1822 }
1823
1824 /* Get port early, so can set affinity prior to memory allocation */
1825 static int ipath_assign_port(struct file *fp,
1826                               const struct ipath_user_info *uinfo)
1827 {
1828         int ret;
1829         int i_minor;
1830         unsigned swmajor, swminor;
1831
1832         /* Check to be sure we haven't already initialized this file */
1833         if (port_fp(fp)) {
1834                 ret = -EINVAL;
1835                 goto done;
1836         }
1837
1838         /* for now, if major version is different, bail */
1839         swmajor = uinfo->spu_userversion >> 16;
1840         if (swmajor != IPATH_USER_SWMAJOR) {
1841                 ipath_dbg("User major version %d not same as driver "
1842                           "major %d\n", uinfo->spu_userversion >> 16,
1843                           IPATH_USER_SWMAJOR);
1844                 ret = -ENODEV;
1845                 goto done;
1846         }
1847
1848         swminor = uinfo->spu_userversion & 0xffff;
1849         if (swminor != IPATH_USER_SWMINOR)
1850                 ipath_dbg("User minor version %d not same as driver "
1851                           "minor %d\n", swminor, IPATH_USER_SWMINOR);
1852
1853         mutex_lock(&ipath_mutex);
1854
1855         if (ipath_compatible_subports(swmajor, swminor) &&
1856             uinfo->spu_subport_cnt &&
1857             (ret = find_shared_port(fp, uinfo))) {
1858                 if (ret > 0)
1859                         ret = 0;
1860                 goto done_chk_sdma;
1861         }
1862
1863         i_minor = iminor(fp->f_path.dentry->d_inode) - IPATH_USER_MINOR_BASE;
1864         ipath_cdbg(VERBOSE, "open on dev %lx (minor %d)\n",
1865                    (long)fp->f_path.dentry->d_inode->i_rdev, i_minor);
1866
1867         if (i_minor)
1868                 ret = find_free_port(i_minor - 1, fp, uinfo);
1869         else
1870                 ret = find_best_unit(fp, uinfo);
1871
1872 done_chk_sdma:
1873         if (!ret) {
1874                 struct ipath_filedata *fd = fp->private_data;
1875                 const struct ipath_portdata *pd = fd->pd;
1876                 const struct ipath_devdata *dd = pd->port_dd;
1877
1878                 fd->pq = ipath_user_sdma_queue_create(&dd->pcidev->dev,
1879                                                       dd->ipath_unit,
1880                                                       pd->port_port,
1881                                                       fd->subport);
1882
1883                 if (!fd->pq)
1884                         ret = -ENOMEM;
1885         }
1886
1887         mutex_unlock(&ipath_mutex);
1888
1889 done:
1890         return ret;
1891 }
1892
1893
1894 static int ipath_do_user_init(struct file *fp,
1895                               const struct ipath_user_info *uinfo)
1896 {
1897         int ret;
1898         struct ipath_portdata *pd = port_fp(fp);
1899         struct ipath_devdata *dd;
1900         u32 head32;
1901
1902         /* Subports don't need to initialize anything since master did it. */
1903         if (subport_fp(fp)) {
1904                 ret = wait_event_interruptible(pd->port_wait,
1905                         !test_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag));
1906                 goto done;
1907         }
1908
1909         dd = pd->port_dd;
1910
1911         if (uinfo->spu_rcvhdrsize) {
1912                 ret = ipath_setrcvhdrsize(dd, uinfo->spu_rcvhdrsize);
1913                 if (ret)
1914                         goto done;
1915         }
1916
1917         /* for now we do nothing with rcvhdrcnt: uinfo->spu_rcvhdrcnt */
1918
1919         /* some ports may get extra buffers, calculate that here */
1920         if (pd->port_port <= dd->ipath_ports_extrabuf)
1921                 pd->port_piocnt = dd->ipath_pbufsport + 1;
1922         else
1923                 pd->port_piocnt = dd->ipath_pbufsport;
1924
1925         /* for right now, kernel piobufs are at end, so port 1 is at 0 */
1926         if (pd->port_port <= dd->ipath_ports_extrabuf)
1927                 pd->port_pio_base = (dd->ipath_pbufsport + 1)
1928                         * (pd->port_port - 1);
1929         else
1930                 pd->port_pio_base = dd->ipath_ports_extrabuf +
1931                         dd->ipath_pbufsport * (pd->port_port - 1);
1932         pd->port_piobufs = dd->ipath_piobufbase +
1933                 pd->port_pio_base * dd->ipath_palign;
1934         ipath_cdbg(VERBOSE, "piobuf base for port %u is 0x%x, piocnt %u,"
1935                 " first pio %u\n", pd->port_port, pd->port_piobufs,
1936                 pd->port_piocnt, pd->port_pio_base);
1937         ipath_chg_pioavailkernel(dd, pd->port_pio_base, pd->port_piocnt, 0);
1938
1939         /*
1940          * Now allocate the rcvhdr Q and eager TIDs; skip the TID
1941          * array for time being.  If pd->port_port > chip-supported,
1942          * we need to do extra stuff here to handle by handling overflow
1943          * through port 0, someday
1944          */
1945         ret = ipath_create_rcvhdrq(dd, pd);
1946         if (!ret)
1947                 ret = ipath_create_user_egr(pd);
1948         if (ret)
1949                 goto done;
1950
1951         /*
1952          * set the eager head register for this port to the current values
1953          * of the tail pointers, since we don't know if they were
1954          * updated on last use of the port.
1955          */
1956         head32 = ipath_read_ureg32(dd, ur_rcvegrindextail, pd->port_port);
1957         ipath_write_ureg(dd, ur_rcvegrindexhead, head32, pd->port_port);
1958         pd->port_lastrcvhdrqtail = -1;
1959         ipath_cdbg(VERBOSE, "Wrote port%d egrhead %x from tail regs\n",
1960                 pd->port_port, head32);
1961         pd->port_tidcursor = 0; /* start at beginning after open */
1962
1963         /* initialize poll variables... */
1964         pd->port_urgent = 0;
1965         pd->port_urgent_poll = 0;
1966         pd->port_hdrqfull_poll = pd->port_hdrqfull;
1967
1968         /*
1969          * Now enable the port for receive.
1970          * For chips that are set to DMA the tail register to memory
1971          * when they change (and when the update bit transitions from
1972          * 0 to 1.  So for those chips, we turn it off and then back on.
1973          * This will (very briefly) affect any other open ports, but the
1974          * duration is very short, and therefore isn't an issue.  We
1975          * explictly set the in-memory tail copy to 0 beforehand, so we
1976          * don't have to wait to be sure the DMA update has happened
1977          * (chip resets head/tail to 0 on transition to enable).
1978          */
1979         set_bit(dd->ipath_r_portenable_shift + pd->port_port,
1980                 &dd->ipath_rcvctrl);
1981         if (!(dd->ipath_flags & IPATH_NODMA_RTAIL)) {
1982                 if (pd->port_rcvhdrtail_kvaddr)
1983                         ipath_clear_rcvhdrtail(pd);
1984                 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1985                         dd->ipath_rcvctrl &
1986                         ~(1ULL << dd->ipath_r_tailupd_shift));
1987         }
1988         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1989                          dd->ipath_rcvctrl);
1990         /* Notify any waiting slaves */
1991         if (pd->port_subport_cnt) {
1992                 clear_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
1993                 wake_up(&pd->port_wait);
1994         }
1995 done:
1996         return ret;
1997 }
1998
1999 /**
2000  * unlock_exptid - unlock any expected TID entries port still had in use
2001  * @pd: port
2002  *
2003  * We don't actually update the chip here, because we do a bulk update
2004  * below, using ipath_f_clear_tids.
2005  */
2006 static void unlock_expected_tids(struct ipath_portdata *pd)
2007 {
2008         struct ipath_devdata *dd = pd->port_dd;
2009         int port_tidbase = pd->port_port * dd->ipath_rcvtidcnt;
2010         int i, cnt = 0, maxtid = port_tidbase + dd->ipath_rcvtidcnt;
2011
2012         ipath_cdbg(VERBOSE, "Port %u unlocking any locked expTID pages\n",
2013                    pd->port_port);
2014         for (i = port_tidbase; i < maxtid; i++) {
2015                 struct page *ps = dd->ipath_pageshadow[i];
2016
2017                 if (!ps)
2018                         continue;
2019
2020                 dd->ipath_pageshadow[i] = NULL;
2021                 pci_unmap_page(dd->pcidev, dd->ipath_physshadow[i],
2022                         PAGE_SIZE, PCI_DMA_FROMDEVICE);
2023                 ipath_release_user_pages_on_close(&ps, 1);
2024                 cnt++;
2025                 ipath_stats.sps_pageunlocks++;
2026         }
2027         if (cnt)
2028                 ipath_cdbg(VERBOSE, "Port %u locked %u expTID entries\n",
2029                            pd->port_port, cnt);
2030
2031         if (ipath_stats.sps_pagelocks || ipath_stats.sps_pageunlocks)
2032                 ipath_cdbg(VERBOSE, "%llu pages locked, %llu unlocked\n",
2033                            (unsigned long long) ipath_stats.sps_pagelocks,
2034                            (unsigned long long)
2035                            ipath_stats.sps_pageunlocks);
2036 }
2037
2038 static int ipath_close(struct inode *in, struct file *fp)
2039 {
2040         int ret = 0;
2041         struct ipath_filedata *fd;
2042         struct ipath_portdata *pd;
2043         struct ipath_devdata *dd;
2044         unsigned port;
2045
2046         ipath_cdbg(VERBOSE, "close on dev %lx, private data %p\n",
2047                    (long)in->i_rdev, fp->private_data);
2048
2049         mutex_lock(&ipath_mutex);
2050
2051         fd = (struct ipath_filedata *) fp->private_data;
2052         fp->private_data = NULL;
2053         pd = fd->pd;
2054         if (!pd) {
2055                 mutex_unlock(&ipath_mutex);
2056                 goto bail;
2057         }
2058
2059         dd = pd->port_dd;
2060
2061         /* drain user sdma queue */
2062         ipath_user_sdma_queue_drain(dd, fd->pq);
2063         ipath_user_sdma_queue_destroy(fd->pq);
2064
2065         if (--pd->port_cnt) {
2066                 /*
2067                  * XXX If the master closes the port before the slave(s),
2068                  * revoke the mmap for the eager receive queue so
2069                  * the slave(s) don't wait for receive data forever.
2070                  */
2071                 pd->active_slaves &= ~(1 << fd->subport);
2072                 put_pid(pd->port_subpid[fd->subport]);
2073                 pd->port_subpid[fd->subport] = NULL;
2074                 mutex_unlock(&ipath_mutex);
2075                 goto bail;
2076         }
2077         port = pd->port_port;
2078
2079         if (pd->port_hdrqfull) {
2080                 ipath_cdbg(PROC, "%s[%u] had %u rcvhdrqfull errors "
2081                            "during run\n", pd->port_comm, pid_nr(pd->port_pid),
2082                            pd->port_hdrqfull);
2083                 pd->port_hdrqfull = 0;
2084         }
2085
2086         if (pd->port_rcvwait_to || pd->port_piowait_to
2087             || pd->port_rcvnowait || pd->port_pionowait) {
2088                 ipath_cdbg(VERBOSE, "port%u, %u rcv, %u pio wait timeo; "
2089                            "%u rcv %u, pio already\n",
2090                            pd->port_port, pd->port_rcvwait_to,
2091                            pd->port_piowait_to, pd->port_rcvnowait,
2092                            pd->port_pionowait);
2093                 pd->port_rcvwait_to = pd->port_piowait_to =
2094                         pd->port_rcvnowait = pd->port_pionowait = 0;
2095         }
2096         if (pd->port_flag) {
2097                 ipath_cdbg(PROC, "port %u port_flag set: 0x%lx\n",
2098                           pd->port_port, pd->port_flag);
2099                 pd->port_flag = 0;
2100         }
2101
2102         if (dd->ipath_kregbase) {
2103                 /* atomically clear receive enable port and intr avail. */
2104                 clear_bit(dd->ipath_r_portenable_shift + port,
2105                           &dd->ipath_rcvctrl);
2106                 clear_bit(pd->port_port + dd->ipath_r_intravail_shift,
2107                           &dd->ipath_rcvctrl);
2108                 ipath_write_kreg( dd, dd->ipath_kregs->kr_rcvctrl,
2109                         dd->ipath_rcvctrl);
2110                 /* and read back from chip to be sure that nothing
2111                  * else is in flight when we do the rest */
2112                 (void)ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
2113
2114                 /* clean up the pkeys for this port user */
2115                 ipath_clean_part_key(pd, dd);
2116                 /*
2117                  * be paranoid, and never write 0's to these, just use an
2118                  * unused part of the port 0 tail page.  Of course,
2119                  * rcvhdraddr points to a large chunk of memory, so this
2120                  * could still trash things, but at least it won't trash
2121                  * page 0, and by disabling the port, it should stop "soon",
2122                  * even if a packet or two is in already in flight after we
2123                  * disabled the port.
2124                  */
2125                 ipath_write_kreg_port(dd,
2126                         dd->ipath_kregs->kr_rcvhdrtailaddr, port,
2127                         dd->ipath_dummy_hdrq_phys);
2128                 ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdraddr,
2129                         pd->port_port, dd->ipath_dummy_hdrq_phys);
2130
2131                 ipath_disarm_piobufs(dd, pd->port_pio_base, pd->port_piocnt);
2132                 ipath_chg_pioavailkernel(dd, pd->port_pio_base,
2133                         pd->port_piocnt, 1);
2134
2135                 dd->ipath_f_clear_tids(dd, pd->port_port);
2136
2137                 if (dd->ipath_pageshadow)
2138                         unlock_expected_tids(pd);
2139                 ipath_stats.sps_ports--;
2140                 ipath_cdbg(PROC, "%s[%u] closed port %u:%u\n",
2141                            pd->port_comm, pid_nr(pd->port_pid),
2142                            dd->ipath_unit, port);
2143         }
2144
2145         put_pid(pd->port_pid);
2146         pd->port_pid = NULL;
2147         dd->ipath_pd[pd->port_port] = NULL; /* before releasing mutex */
2148         mutex_unlock(&ipath_mutex);
2149         ipath_free_pddata(dd, pd); /* after releasing the mutex */
2150
2151 bail:
2152         kfree(fd);
2153         return ret;
2154 }
2155
2156 static int ipath_port_info(struct ipath_portdata *pd, u16 subport,
2157                            struct ipath_port_info __user *uinfo)
2158 {
2159         struct ipath_port_info info;
2160         int nup;
2161         int ret;
2162         size_t sz;
2163
2164         (void) ipath_count_units(NULL, &nup, NULL);
2165         info.num_active = nup;
2166         info.unit = pd->port_dd->ipath_unit;
2167         info.port = pd->port_port;
2168         info.subport = subport;
2169         /* Don't return new fields if old library opened the port. */
2170         if (ipath_supports_subports(pd->userversion >> 16,
2171                                     pd->userversion & 0xffff)) {
2172                 /* Number of user ports available for this device. */
2173                 info.num_ports = pd->port_dd->ipath_cfgports - 1;
2174                 info.num_subports = pd->port_subport_cnt;
2175                 sz = sizeof(info);
2176         } else
2177                 sz = sizeof(info) - 2 * sizeof(u16);
2178
2179         if (copy_to_user(uinfo, &info, sz)) {
2180                 ret = -EFAULT;
2181                 goto bail;
2182         }
2183         ret = 0;
2184
2185 bail:
2186         return ret;
2187 }
2188
2189 static int ipath_get_slave_info(struct ipath_portdata *pd,
2190                                 void __user *slave_mask_addr)
2191 {
2192         int ret = 0;
2193
2194         if (copy_to_user(slave_mask_addr, &pd->active_slaves, sizeof(u32)))
2195                 ret = -EFAULT;
2196         return ret;
2197 }
2198
2199 static int ipath_sdma_get_inflight(struct ipath_user_sdma_queue *pq,
2200                                    u32 __user *inflightp)
2201 {
2202         const u32 val = ipath_user_sdma_inflight_counter(pq);
2203
2204         if (put_user(val, inflightp))
2205                 return -EFAULT;
2206
2207         return 0;
2208 }
2209
2210 static int ipath_sdma_get_complete(struct ipath_devdata *dd,
2211                                    struct ipath_user_sdma_queue *pq,
2212                                    u32 __user *completep)
2213 {
2214         u32 val;
2215         int err;
2216
2217         err = ipath_user_sdma_make_progress(dd, pq);
2218         if (err < 0)
2219                 return err;
2220
2221         val = ipath_user_sdma_complete_counter(pq);
2222         if (put_user(val, completep))
2223                 return -EFAULT;
2224
2225         return 0;
2226 }
2227
2228 static ssize_t ipath_write(struct file *fp, const char __user *data,
2229                            size_t count, loff_t *off)
2230 {
2231         const struct ipath_cmd __user *ucmd;
2232         struct ipath_portdata *pd;
2233         const void __user *src;
2234         size_t consumed, copy;
2235         struct ipath_cmd cmd;
2236         ssize_t ret = 0;
2237         void *dest;
2238
2239         if (count < sizeof(cmd.type)) {
2240                 ret = -EINVAL;
2241                 goto bail;
2242         }
2243
2244         ucmd = (const struct ipath_cmd __user *) data;
2245
2246         if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) {
2247                 ret = -EFAULT;
2248                 goto bail;
2249         }
2250
2251         consumed = sizeof(cmd.type);
2252
2253         switch (cmd.type) {
2254         case IPATH_CMD_ASSIGN_PORT:
2255         case __IPATH_CMD_USER_INIT:
2256         case IPATH_CMD_USER_INIT:
2257                 copy = sizeof(cmd.cmd.user_info);
2258                 dest = &cmd.cmd.user_info;
2259                 src = &ucmd->cmd.user_info;
2260                 break;
2261         case IPATH_CMD_RECV_CTRL:
2262                 copy = sizeof(cmd.cmd.recv_ctrl);
2263                 dest = &cmd.cmd.recv_ctrl;
2264                 src = &ucmd->cmd.recv_ctrl;
2265                 break;
2266         case IPATH_CMD_PORT_INFO:
2267                 copy = sizeof(cmd.cmd.port_info);
2268                 dest = &cmd.cmd.port_info;
2269                 src = &ucmd->cmd.port_info;
2270                 break;
2271         case IPATH_CMD_TID_UPDATE:
2272         case IPATH_CMD_TID_FREE:
2273                 copy = sizeof(cmd.cmd.tid_info);
2274                 dest = &cmd.cmd.tid_info;
2275                 src = &ucmd->cmd.tid_info;
2276                 break;
2277         case IPATH_CMD_SET_PART_KEY:
2278                 copy = sizeof(cmd.cmd.part_key);
2279                 dest = &cmd.cmd.part_key;
2280                 src = &ucmd->cmd.part_key;
2281                 break;
2282         case __IPATH_CMD_SLAVE_INFO:
2283                 copy = sizeof(cmd.cmd.slave_mask_addr);
2284                 dest = &cmd.cmd.slave_mask_addr;
2285                 src = &ucmd->cmd.slave_mask_addr;
2286                 break;
2287         case IPATH_CMD_PIOAVAILUPD:     // force an update of PIOAvail reg
2288                 copy = 0;
2289                 src = NULL;
2290                 dest = NULL;
2291                 break;
2292         case IPATH_CMD_POLL_TYPE:
2293                 copy = sizeof(cmd.cmd.poll_type);
2294                 dest = &cmd.cmd.poll_type;
2295                 src = &ucmd->cmd.poll_type;
2296                 break;
2297         case IPATH_CMD_ARMLAUNCH_CTRL:
2298                 copy = sizeof(cmd.cmd.armlaunch_ctrl);
2299                 dest = &cmd.cmd.armlaunch_ctrl;
2300                 src = &ucmd->cmd.armlaunch_ctrl;
2301                 break;
2302         case IPATH_CMD_SDMA_INFLIGHT:
2303                 copy = sizeof(cmd.cmd.sdma_inflight);
2304                 dest = &cmd.cmd.sdma_inflight;
2305                 src = &ucmd->cmd.sdma_inflight;
2306                 break;
2307         case IPATH_CMD_SDMA_COMPLETE:
2308                 copy = sizeof(cmd.cmd.sdma_complete);
2309                 dest = &cmd.cmd.sdma_complete;
2310                 src = &ucmd->cmd.sdma_complete;
2311                 break;
2312         default:
2313                 ret = -EINVAL;
2314                 goto bail;
2315         }
2316
2317         if (copy) {
2318                 if ((count - consumed) < copy) {
2319                         ret = -EINVAL;
2320                         goto bail;
2321                 }
2322
2323                 if (copy_from_user(dest, src, copy)) {
2324                         ret = -EFAULT;
2325                         goto bail;
2326                 }
2327
2328                 consumed += copy;
2329         }
2330
2331         pd = port_fp(fp);
2332         if (!pd && cmd.type != __IPATH_CMD_USER_INIT &&
2333                 cmd.type != IPATH_CMD_ASSIGN_PORT) {
2334                 ret = -EINVAL;
2335                 goto bail;
2336         }
2337
2338         switch (cmd.type) {
2339         case IPATH_CMD_ASSIGN_PORT:
2340                 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2341                 if (ret)
2342                         goto bail;
2343                 break;
2344         case __IPATH_CMD_USER_INIT:
2345                 /* backwards compatibility, get port first */
2346                 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2347                 if (ret)
2348                         goto bail;
2349                 /* and fall through to current version. */
2350         case IPATH_CMD_USER_INIT:
2351                 ret = ipath_do_user_init(fp, &cmd.cmd.user_info);
2352                 if (ret)
2353                         goto bail;
2354                 ret = ipath_get_base_info(
2355                         fp, (void __user *) (unsigned long)
2356                         cmd.cmd.user_info.spu_base_info,
2357                         cmd.cmd.user_info.spu_base_info_size);
2358                 break;
2359         case IPATH_CMD_RECV_CTRL:
2360                 ret = ipath_manage_rcvq(pd, subport_fp(fp), cmd.cmd.recv_ctrl);
2361                 break;
2362         case IPATH_CMD_PORT_INFO:
2363                 ret = ipath_port_info(pd, subport_fp(fp),
2364                                       (struct ipath_port_info __user *)
2365                                       (unsigned long) cmd.cmd.port_info);
2366                 break;
2367         case IPATH_CMD_TID_UPDATE:
2368                 ret = ipath_tid_update(pd, fp, &cmd.cmd.tid_info);
2369                 break;
2370         case IPATH_CMD_TID_FREE:
2371                 ret = ipath_tid_free(pd, subport_fp(fp), &cmd.cmd.tid_info);
2372                 break;
2373         case IPATH_CMD_SET_PART_KEY:
2374                 ret = ipath_set_part_key(pd, cmd.cmd.part_key);
2375                 break;
2376         case __IPATH_CMD_SLAVE_INFO:
2377                 ret = ipath_get_slave_info(pd,
2378                                            (void __user *) (unsigned long)
2379                                            cmd.cmd.slave_mask_addr);
2380                 break;
2381         case IPATH_CMD_PIOAVAILUPD:
2382                 ipath_force_pio_avail_update(pd->port_dd);
2383                 break;
2384         case IPATH_CMD_POLL_TYPE:
2385                 pd->poll_type = cmd.cmd.poll_type;
2386                 break;
2387         case IPATH_CMD_ARMLAUNCH_CTRL:
2388                 if (cmd.cmd.armlaunch_ctrl)
2389                         ipath_enable_armlaunch(pd->port_dd);
2390                 else
2391                         ipath_disable_armlaunch(pd->port_dd);
2392                 break;
2393         case IPATH_CMD_SDMA_INFLIGHT:
2394                 ret = ipath_sdma_get_inflight(user_sdma_queue_fp(fp),
2395                                               (u32 __user *) (unsigned long)
2396                                               cmd.cmd.sdma_inflight);
2397                 break;
2398         case IPATH_CMD_SDMA_COMPLETE:
2399                 ret = ipath_sdma_get_complete(pd->port_dd,
2400                                               user_sdma_queue_fp(fp),
2401                                               (u32 __user *) (unsigned long)
2402                                               cmd.cmd.sdma_complete);
2403                 break;
2404         }
2405
2406         if (ret >= 0)
2407                 ret = consumed;
2408
2409 bail:
2410         return ret;
2411 }
2412
2413 static ssize_t ipath_writev(struct kiocb *iocb, const struct iovec *iov,
2414                             unsigned long dim, loff_t off)
2415 {
2416         struct file *filp = iocb->ki_filp;
2417         struct ipath_filedata *fp = filp->private_data;
2418         struct ipath_portdata *pd = port_fp(filp);
2419         struct ipath_user_sdma_queue *pq = fp->pq;
2420
2421         if (!dim)
2422                 return -EINVAL;
2423
2424         return ipath_user_sdma_writev(pd->port_dd, pq, iov, dim);
2425 }
2426
2427 static struct class *ipath_class;
2428
2429 static int init_cdev(int minor, char *name, const struct file_operations *fops,
2430                      struct cdev **cdevp, struct device **devp)
2431 {
2432         const dev_t dev = MKDEV(IPATH_MAJOR, minor);
2433         struct cdev *cdev = NULL;
2434         struct device *device = NULL;
2435         int ret;
2436
2437         cdev = cdev_alloc();
2438         if (!cdev) {
2439                 printk(KERN_ERR IPATH_DRV_NAME
2440                        ": Could not allocate cdev for minor %d, %s\n",
2441                        minor, name);
2442                 ret = -ENOMEM;
2443                 goto done;
2444         }
2445
2446         cdev->owner = THIS_MODULE;
2447         cdev->ops = fops;
2448         kobject_set_name(&cdev->kobj, name);
2449
2450         ret = cdev_add(cdev, dev, 1);
2451         if (ret < 0) {
2452                 printk(KERN_ERR IPATH_DRV_NAME
2453                        ": Could not add cdev for minor %d, %s (err %d)\n",
2454                        minor, name, -ret);
2455                 goto err_cdev;
2456         }
2457
2458         device = device_create(ipath_class, NULL, dev, NULL, name);
2459
2460         if (IS_ERR(device)) {
2461                 ret = PTR_ERR(device);
2462                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2463                        "device for minor %d, %s (err %d)\n",
2464                        minor, name, -ret);
2465                 goto err_cdev;
2466         }
2467
2468         goto done;
2469
2470 err_cdev:
2471         cdev_del(cdev);
2472         cdev = NULL;
2473
2474 done:
2475         if (ret >= 0) {
2476                 *cdevp = cdev;
2477                 *devp = device;
2478         } else {
2479                 *cdevp = NULL;
2480                 *devp = NULL;
2481         }
2482
2483         return ret;
2484 }
2485
2486 int ipath_cdev_init(int minor, char *name, const struct file_operations *fops,
2487                     struct cdev **cdevp, struct device **devp)
2488 {
2489         return init_cdev(minor, name, fops, cdevp, devp);
2490 }
2491
2492 static void cleanup_cdev(struct cdev **cdevp,
2493                          struct device **devp)
2494 {
2495         struct device *dev = *devp;
2496
2497         if (dev) {
2498                 device_unregister(dev);
2499                 *devp = NULL;
2500         }
2501
2502         if (*cdevp) {
2503                 cdev_del(*cdevp);
2504                 *cdevp = NULL;
2505         }
2506 }
2507
2508 void ipath_cdev_cleanup(struct cdev **cdevp,
2509                         struct device **devp)
2510 {
2511         cleanup_cdev(cdevp, devp);
2512 }
2513
2514 static struct cdev *wildcard_cdev;
2515 static struct device *wildcard_dev;
2516
2517 static const dev_t dev = MKDEV(IPATH_MAJOR, 0);
2518
2519 static int user_init(void)
2520 {
2521         int ret;
2522
2523         ret = register_chrdev_region(dev, IPATH_NMINORS, IPATH_DRV_NAME);
2524         if (ret < 0) {
2525                 printk(KERN_ERR IPATH_DRV_NAME ": Could not register "
2526                        "chrdev region (err %d)\n", -ret);
2527                 goto done;
2528         }
2529
2530         ipath_class = class_create(THIS_MODULE, IPATH_DRV_NAME);
2531
2532         if (IS_ERR(ipath_class)) {
2533                 ret = PTR_ERR(ipath_class);
2534                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2535                        "device class (err %d)\n", -ret);
2536                 goto bail;
2537         }
2538
2539         goto done;
2540 bail:
2541         unregister_chrdev_region(dev, IPATH_NMINORS);
2542 done:
2543         return ret;
2544 }
2545
2546 static void user_cleanup(void)
2547 {
2548         if (ipath_class) {
2549                 class_destroy(ipath_class);
2550                 ipath_class = NULL;
2551         }
2552
2553         unregister_chrdev_region(dev, IPATH_NMINORS);
2554 }
2555
2556 static atomic_t user_count = ATOMIC_INIT(0);
2557 static atomic_t user_setup = ATOMIC_INIT(0);
2558
2559 int ipath_user_add(struct ipath_devdata *dd)
2560 {
2561         char name[10];
2562         int ret;
2563
2564         if (atomic_inc_return(&user_count) == 1) {
2565                 ret = user_init();
2566                 if (ret < 0) {
2567                         ipath_dev_err(dd, "Unable to set up user support: "
2568                                       "error %d\n", -ret);
2569                         goto bail;
2570                 }
2571                 ret = init_cdev(0, "ipath", &ipath_file_ops, &wildcard_cdev,
2572                                 &wildcard_dev);
2573                 if (ret < 0) {
2574                         ipath_dev_err(dd, "Could not create wildcard "
2575                                       "minor: error %d\n", -ret);
2576                         goto bail_user;
2577                 }
2578
2579                 atomic_set(&user_setup, 1);
2580         }
2581
2582         snprintf(name, sizeof(name), "ipath%d", dd->ipath_unit);
2583
2584         ret = init_cdev(dd->ipath_unit + 1, name, &ipath_file_ops,
2585                         &dd->user_cdev, &dd->user_dev);
2586         if (ret < 0)
2587                 ipath_dev_err(dd, "Could not create user minor %d, %s\n",
2588                               dd->ipath_unit + 1, name);
2589
2590         goto bail;
2591
2592 bail_user:
2593         user_cleanup();
2594 bail:
2595         return ret;
2596 }
2597
2598 void ipath_user_remove(struct ipath_devdata *dd)
2599 {
2600         cleanup_cdev(&dd->user_cdev, &dd->user_dev);
2601
2602         if (atomic_dec_return(&user_count) == 0) {
2603                 if (atomic_read(&user_setup) == 0)
2604                         goto bail;
2605
2606                 cleanup_cdev(&wildcard_cdev, &wildcard_dev);
2607                 user_cleanup();
2608
2609                 atomic_set(&user_setup, 0);
2610         }
2611 bail:
2612         return;
2613 }