Merge branch 'next' into for-linus
[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         /*
227          * All user buffers are 2KB buffers.  If we ever support
228          * giving 4KB buffers to user processes, this will need some
229          * work.
230          */
231         kinfo->spi_pioindex = (kinfo->spi_piobufbase -
232                 (dd->ipath_piobufbase & 0xffffffff)) / dd->ipath_palign;
233         kinfo->spi_pioalign = dd->ipath_palign;
234
235         kinfo->spi_qpair = IPATH_KD_QP;
236         /*
237          * user mode PIO buffers are always 2KB, even when 4KB can
238          * be received, and sent via the kernel; this is ibmaxlen
239          * for 2K MTU.
240          */
241         kinfo->spi_piosize = dd->ipath_piosize2k - 2 * sizeof(u32);
242         kinfo->spi_mtu = dd->ipath_ibmaxlen;    /* maxlen, not ibmtu */
243         kinfo->spi_port = pd->port_port;
244         kinfo->spi_subport = subport_fp(fp);
245         kinfo->spi_sw_version = IPATH_KERN_SWVERSION;
246         kinfo->spi_hw_version = dd->ipath_revision;
247
248         if (master) {
249                 kinfo->spi_runtime_flags |= IPATH_RUNTIME_MASTER;
250         }
251
252         sz = (ubase_size < sizeof(*kinfo)) ? ubase_size : sizeof(*kinfo);
253         if (copy_to_user(ubase, kinfo, sz))
254                 ret = -EFAULT;
255
256 bail:
257         kfree(kinfo);
258         return ret;
259 }
260
261 /**
262  * ipath_tid_update - update a port TID
263  * @pd: the port
264  * @fp: the ipath device file
265  * @ti: the TID information
266  *
267  * The new implementation as of Oct 2004 is that the driver assigns
268  * the tid and returns it to the caller.   To make it easier to
269  * catch bugs, and to reduce search time, we keep a cursor for
270  * each port, walking the shadow tid array to find one that's not
271  * in use.
272  *
273  * For now, if we can't allocate the full list, we fail, although
274  * in the long run, we'll allocate as many as we can, and the
275  * caller will deal with that by trying the remaining pages later.
276  * That means that when we fail, we have to mark the tids as not in
277  * use again, in our shadow copy.
278  *
279  * It's up to the caller to free the tids when they are done.
280  * We'll unlock the pages as they free them.
281  *
282  * Also, right now we are locking one page at a time, but since
283  * the intended use of this routine is for a single group of
284  * virtually contiguous pages, that should change to improve
285  * performance.
286  */
287 static int ipath_tid_update(struct ipath_portdata *pd, struct file *fp,
288                             const struct ipath_tid_info *ti)
289 {
290         int ret = 0, ntids;
291         u32 tid, porttid, cnt, i, tidcnt, tidoff;
292         u16 *tidlist;
293         struct ipath_devdata *dd = pd->port_dd;
294         u64 physaddr;
295         unsigned long vaddr;
296         u64 __iomem *tidbase;
297         unsigned long tidmap[8];
298         struct page **pagep = NULL;
299         unsigned subport = subport_fp(fp);
300
301         if (!dd->ipath_pageshadow) {
302                 ret = -ENOMEM;
303                 goto done;
304         }
305
306         cnt = ti->tidcnt;
307         if (!cnt) {
308                 ipath_dbg("After copyin, tidcnt 0, tidlist %llx\n",
309                           (unsigned long long) ti->tidlist);
310                 /*
311                  * Should we treat as success?  likely a bug
312                  */
313                 ret = -EFAULT;
314                 goto done;
315         }
316         porttid = pd->port_port * dd->ipath_rcvtidcnt;
317         if (!pd->port_subport_cnt) {
318                 tidcnt = dd->ipath_rcvtidcnt;
319                 tid = pd->port_tidcursor;
320                 tidoff = 0;
321         } else if (!subport) {
322                 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
323                          (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
324                 tidoff = dd->ipath_rcvtidcnt - tidcnt;
325                 porttid += tidoff;
326                 tid = tidcursor_fp(fp);
327         } else {
328                 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
329                 tidoff = tidcnt * (subport - 1);
330                 porttid += tidoff;
331                 tid = tidcursor_fp(fp);
332         }
333         if (cnt > tidcnt) {
334                 /* make sure it all fits in port_tid_pg_list */
335                 dev_info(&dd->pcidev->dev, "Process tried to allocate %u "
336                          "TIDs, only trying max (%u)\n", cnt, tidcnt);
337                 cnt = tidcnt;
338         }
339         pagep = &((struct page **) pd->port_tid_pg_list)[tidoff];
340         tidlist = &((u16 *) &pagep[dd->ipath_rcvtidcnt])[tidoff];
341
342         memset(tidmap, 0, sizeof(tidmap));
343         /* before decrement; chip actual # */
344         ntids = tidcnt;
345         tidbase = (u64 __iomem *) (((char __iomem *) dd->ipath_kregbase) +
346                                    dd->ipath_rcvtidbase +
347                                    porttid * sizeof(*tidbase));
348
349         ipath_cdbg(VERBOSE, "Port%u %u tids, cursor %u, tidbase %p\n",
350                    pd->port_port, cnt, tid, tidbase);
351
352         /* virtual address of first page in transfer */
353         vaddr = ti->tidvaddr;
354         if (!access_ok(VERIFY_WRITE, (void __user *) vaddr,
355                        cnt * PAGE_SIZE)) {
356                 ipath_dbg("Fail vaddr %p, %u pages, !access_ok\n",
357                           (void *)vaddr, cnt);
358                 ret = -EFAULT;
359                 goto done;
360         }
361         ret = ipath_get_user_pages(vaddr, cnt, pagep);
362         if (ret) {
363                 if (ret == -EBUSY) {
364                         ipath_dbg("Failed to lock addr %p, %u pages "
365                                   "(already locked)\n",
366                                   (void *) vaddr, cnt);
367                         /*
368                          * for now, continue, and see what happens but with
369                          * the new implementation, this should never happen,
370                          * unless perhaps the user has mpin'ed the pages
371                          * themselves (something we need to test)
372                          */
373                         ret = 0;
374                 } else {
375                         dev_info(&dd->pcidev->dev,
376                                  "Failed to lock addr %p, %u pages: "
377                                  "errno %d\n", (void *) vaddr, cnt, -ret);
378                         goto done;
379                 }
380         }
381         for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
382                 for (; ntids--; tid++) {
383                         if (tid == tidcnt)
384                                 tid = 0;
385                         if (!dd->ipath_pageshadow[porttid + tid])
386                                 break;
387                 }
388                 if (ntids < 0) {
389                         /*
390                          * oops, wrapped all the way through their TIDs,
391                          * and didn't have enough free; see comments at
392                          * start of routine
393                          */
394                         ipath_dbg("Not enough free TIDs for %u pages "
395                                   "(index %d), failing\n", cnt, i);
396                         i--;    /* last tidlist[i] not filled in */
397                         ret = -ENOMEM;
398                         break;
399                 }
400                 tidlist[i] = tid + tidoff;
401                 ipath_cdbg(VERBOSE, "Updating idx %u to TID %u, "
402                            "vaddr %lx\n", i, tid + tidoff, vaddr);
403                 /* we "know" system pages and TID pages are same size */
404                 dd->ipath_pageshadow[porttid + tid] = pagep[i];
405                 dd->ipath_physshadow[porttid + tid] = ipath_map_page(
406                         dd->pcidev, pagep[i], 0, PAGE_SIZE,
407                         PCI_DMA_FROMDEVICE);
408                 /*
409                  * don't need atomic or it's overhead
410                  */
411                 __set_bit(tid, tidmap);
412                 physaddr = dd->ipath_physshadow[porttid + tid];
413                 ipath_stats.sps_pagelocks++;
414                 ipath_cdbg(VERBOSE,
415                            "TID %u, vaddr %lx, physaddr %llx pgp %p\n",
416                            tid, vaddr, (unsigned long long) physaddr,
417                            pagep[i]);
418                 dd->ipath_f_put_tid(dd, &tidbase[tid], RCVHQ_RCV_TYPE_EXPECTED,
419                                     physaddr);
420                 /*
421                  * don't check this tid in ipath_portshadow, since we
422                  * just filled it in; start with the next one.
423                  */
424                 tid++;
425         }
426
427         if (ret) {
428                 u32 limit;
429         cleanup:
430                 /* jump here if copy out of updated info failed... */
431                 ipath_dbg("After failure (ret=%d), undo %d of %d entries\n",
432                           -ret, i, cnt);
433                 /* same code that's in ipath_free_tid() */
434                 limit = sizeof(tidmap) * BITS_PER_BYTE;
435                 if (limit > tidcnt)
436                         /* just in case size changes in future */
437                         limit = tidcnt;
438                 tid = find_first_bit((const unsigned long *)tidmap, limit);
439                 for (; tid < limit; tid++) {
440                         if (!test_bit(tid, tidmap))
441                                 continue;
442                         if (dd->ipath_pageshadow[porttid + tid]) {
443                                 ipath_cdbg(VERBOSE, "Freeing TID %u\n",
444                                            tid);
445                                 dd->ipath_f_put_tid(dd, &tidbase[tid],
446                                                     RCVHQ_RCV_TYPE_EXPECTED,
447                                                     dd->ipath_tidinvalid);
448                                 pci_unmap_page(dd->pcidev,
449                                         dd->ipath_physshadow[porttid + tid],
450                                         PAGE_SIZE, PCI_DMA_FROMDEVICE);
451                                 dd->ipath_pageshadow[porttid + tid] = NULL;
452                                 ipath_stats.sps_pageunlocks++;
453                         }
454                 }
455                 ipath_release_user_pages(pagep, cnt);
456         } else {
457                 /*
458                  * Copy the updated array, with ipath_tid's filled in, back
459                  * to user.  Since we did the copy in already, this "should
460                  * never fail" If it does, we have to clean up...
461                  */
462                 if (copy_to_user((void __user *)
463                                  (unsigned long) ti->tidlist,
464                                  tidlist, cnt * sizeof(*tidlist))) {
465                         ret = -EFAULT;
466                         goto cleanup;
467                 }
468                 if (copy_to_user((void __user *) (unsigned long) ti->tidmap,
469                                  tidmap, sizeof tidmap)) {
470                         ret = -EFAULT;
471                         goto cleanup;
472                 }
473                 if (tid == tidcnt)
474                         tid = 0;
475                 if (!pd->port_subport_cnt)
476                         pd->port_tidcursor = tid;
477                 else
478                         tidcursor_fp(fp) = tid;
479         }
480
481 done:
482         if (ret)
483                 ipath_dbg("Failed to map %u TID pages, failing with %d\n",
484                           ti->tidcnt, -ret);
485         return ret;
486 }
487
488 /**
489  * ipath_tid_free - free a port TID
490  * @pd: the port
491  * @subport: the subport
492  * @ti: the TID info
493  *
494  * right now we are unlocking one page at a time, but since
495  * the intended use of this routine is for a single group of
496  * virtually contiguous pages, that should change to improve
497  * performance.  We check that the TID is in range for this port
498  * but otherwise don't check validity; if user has an error and
499  * frees the wrong tid, it's only their own data that can thereby
500  * be corrupted.  We do check that the TID was in use, for sanity
501  * We always use our idea of the saved address, not the address that
502  * they pass in to us.
503  */
504
505 static int ipath_tid_free(struct ipath_portdata *pd, unsigned subport,
506                           const struct ipath_tid_info *ti)
507 {
508         int ret = 0;
509         u32 tid, porttid, cnt, limit, tidcnt;
510         struct ipath_devdata *dd = pd->port_dd;
511         u64 __iomem *tidbase;
512         unsigned long tidmap[8];
513
514         if (!dd->ipath_pageshadow) {
515                 ret = -ENOMEM;
516                 goto done;
517         }
518
519         if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap,
520                            sizeof tidmap)) {
521                 ret = -EFAULT;
522                 goto done;
523         }
524
525         porttid = pd->port_port * dd->ipath_rcvtidcnt;
526         if (!pd->port_subport_cnt)
527                 tidcnt = dd->ipath_rcvtidcnt;
528         else if (!subport) {
529                 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
530                          (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
531                 porttid += dd->ipath_rcvtidcnt - tidcnt;
532         } else {
533                 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
534                 porttid += tidcnt * (subport - 1);
535         }
536         tidbase = (u64 __iomem *) ((char __iomem *)(dd->ipath_kregbase) +
537                                    dd->ipath_rcvtidbase +
538                                    porttid * sizeof(*tidbase));
539
540         limit = sizeof(tidmap) * BITS_PER_BYTE;
541         if (limit > tidcnt)
542                 /* just in case size changes in future */
543                 limit = tidcnt;
544         tid = find_first_bit(tidmap, limit);
545         ipath_cdbg(VERBOSE, "Port%u free %u tids; first bit (max=%d) "
546                    "set is %d, porttid %u\n", pd->port_port, ti->tidcnt,
547                    limit, tid, porttid);
548         for (cnt = 0; tid < limit; tid++) {
549                 /*
550                  * small optimization; if we detect a run of 3 or so without
551                  * any set, use find_first_bit again.  That's mainly to
552                  * accelerate the case where we wrapped, so we have some at
553                  * the beginning, and some at the end, and a big gap
554                  * in the middle.
555                  */
556                 if (!test_bit(tid, tidmap))
557                         continue;
558                 cnt++;
559                 if (dd->ipath_pageshadow[porttid + tid]) {
560                         struct page *p;
561                         p = dd->ipath_pageshadow[porttid + tid];
562                         dd->ipath_pageshadow[porttid + tid] = NULL;
563                         ipath_cdbg(VERBOSE, "PID %u freeing TID %u\n",
564                                    pid_nr(pd->port_pid), tid);
565                         dd->ipath_f_put_tid(dd, &tidbase[tid],
566                                             RCVHQ_RCV_TYPE_EXPECTED,
567                                             dd->ipath_tidinvalid);
568                         pci_unmap_page(dd->pcidev,
569                                 dd->ipath_physshadow[porttid + tid],
570                                 PAGE_SIZE, PCI_DMA_FROMDEVICE);
571                         ipath_release_user_pages(&p, 1);
572                         ipath_stats.sps_pageunlocks++;
573                 } else
574                         ipath_dbg("Unused tid %u, ignoring\n", tid);
575         }
576         if (cnt != ti->tidcnt)
577                 ipath_dbg("passed in tidcnt %d, only %d bits set in map\n",
578                           ti->tidcnt, cnt);
579 done:
580         if (ret)
581                 ipath_dbg("Failed to unmap %u TID pages, failing with %d\n",
582                           ti->tidcnt, -ret);
583         return ret;
584 }
585
586 /**
587  * ipath_set_part_key - set a partition key
588  * @pd: the port
589  * @key: the key
590  *
591  * We can have up to 4 active at a time (other than the default, which is
592  * always allowed).  This is somewhat tricky, since multiple ports may set
593  * the same key, so we reference count them, and clean up at exit.  All 4
594  * partition keys are packed into a single infinipath register.  It's an
595  * error for a process to set the same pkey multiple times.  We provide no
596  * mechanism to de-allocate a pkey at this time, we may eventually need to
597  * do that.  I've used the atomic operations, and no locking, and only make
598  * a single pass through what's available.  This should be more than
599  * adequate for some time. I'll think about spinlocks or the like if and as
600  * it's necessary.
601  */
602 static int ipath_set_part_key(struct ipath_portdata *pd, u16 key)
603 {
604         struct ipath_devdata *dd = pd->port_dd;
605         int i, any = 0, pidx = -1;
606         u16 lkey = key & 0x7FFF;
607         int ret;
608
609         if (lkey == (IPATH_DEFAULT_P_KEY & 0x7FFF)) {
610                 /* nothing to do; this key always valid */
611                 ret = 0;
612                 goto bail;
613         }
614
615         ipath_cdbg(VERBOSE, "p%u try to set pkey %hx, current keys "
616                    "%hx:%x %hx:%x %hx:%x %hx:%x\n",
617                    pd->port_port, key, dd->ipath_pkeys[0],
618                    atomic_read(&dd->ipath_pkeyrefs[0]), dd->ipath_pkeys[1],
619                    atomic_read(&dd->ipath_pkeyrefs[1]), dd->ipath_pkeys[2],
620                    atomic_read(&dd->ipath_pkeyrefs[2]), dd->ipath_pkeys[3],
621                    atomic_read(&dd->ipath_pkeyrefs[3]));
622
623         if (!lkey) {
624                 ipath_cdbg(PROC, "p%u tries to set key 0, not allowed\n",
625                            pd->port_port);
626                 ret = -EINVAL;
627                 goto bail;
628         }
629
630         /*
631          * Set the full membership bit, because it has to be
632          * set in the register or the packet, and it seems
633          * cleaner to set in the register than to force all
634          * callers to set it. (see bug 4331)
635          */
636         key |= 0x8000;
637
638         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
639                 if (!pd->port_pkeys[i] && pidx == -1)
640                         pidx = i;
641                 if (pd->port_pkeys[i] == key) {
642                         ipath_cdbg(VERBOSE, "p%u tries to set same pkey "
643                                    "(%x) more than once\n",
644                                    pd->port_port, key);
645                         ret = -EEXIST;
646                         goto bail;
647                 }
648         }
649         if (pidx == -1) {
650                 ipath_dbg("All pkeys for port %u already in use, "
651                           "can't set %x\n", pd->port_port, key);
652                 ret = -EBUSY;
653                 goto bail;
654         }
655         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
656                 if (!dd->ipath_pkeys[i]) {
657                         any++;
658                         continue;
659                 }
660                 if (dd->ipath_pkeys[i] == key) {
661                         atomic_t *pkrefs = &dd->ipath_pkeyrefs[i];
662
663                         if (atomic_inc_return(pkrefs) > 1) {
664                                 pd->port_pkeys[pidx] = key;
665                                 ipath_cdbg(VERBOSE, "p%u set key %x "
666                                            "matches #%d, count now %d\n",
667                                            pd->port_port, key, i,
668                                            atomic_read(pkrefs));
669                                 ret = 0;
670                                 goto bail;
671                         } else {
672                                 /*
673                                  * lost race, decrement count, catch below
674                                  */
675                                 atomic_dec(pkrefs);
676                                 ipath_cdbg(VERBOSE, "Lost race, count was "
677                                            "0, after dec, it's %d\n",
678                                            atomic_read(pkrefs));
679                                 any++;
680                         }
681                 }
682                 if ((dd->ipath_pkeys[i] & 0x7FFF) == lkey) {
683                         /*
684                          * It makes no sense to have both the limited and
685                          * full membership PKEY set at the same time since
686                          * the unlimited one will disable the limited one.
687                          */
688                         ret = -EEXIST;
689                         goto bail;
690                 }
691         }
692         if (!any) {
693                 ipath_dbg("port %u, all pkeys already in use, "
694                           "can't set %x\n", pd->port_port, key);
695                 ret = -EBUSY;
696                 goto bail;
697         }
698         for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
699                 if (!dd->ipath_pkeys[i] &&
700                     atomic_inc_return(&dd->ipath_pkeyrefs[i]) == 1) {
701                         u64 pkey;
702
703                         /* for ipathstats, etc. */
704                         ipath_stats.sps_pkeys[i] = lkey;
705                         pd->port_pkeys[pidx] = dd->ipath_pkeys[i] = key;
706                         pkey =
707                                 (u64) dd->ipath_pkeys[0] |
708                                 ((u64) dd->ipath_pkeys[1] << 16) |
709                                 ((u64) dd->ipath_pkeys[2] << 32) |
710                                 ((u64) dd->ipath_pkeys[3] << 48);
711                         ipath_cdbg(PROC, "p%u set key %x in #%d, "
712                                    "portidx %d, new pkey reg %llx\n",
713                                    pd->port_port, key, i, pidx,
714                                    (unsigned long long) pkey);
715                         ipath_write_kreg(
716                                 dd, dd->ipath_kregs->kr_partitionkey, pkey);
717
718                         ret = 0;
719                         goto bail;
720                 }
721         }
722         ipath_dbg("port %u, all pkeys already in use 2nd pass, "
723                   "can't set %x\n", pd->port_port, key);
724         ret = -EBUSY;
725
726 bail:
727         return ret;
728 }
729
730 /**
731  * ipath_manage_rcvq - manage a port's receive queue
732  * @pd: the port
733  * @subport: the subport
734  * @start_stop: action to carry out
735  *
736  * start_stop == 0 disables receive on the port, for use in queue
737  * overflow conditions.  start_stop==1 re-enables, to be used to
738  * re-init the software copy of the head register
739  */
740 static int ipath_manage_rcvq(struct ipath_portdata *pd, unsigned subport,
741                              int start_stop)
742 {
743         struct ipath_devdata *dd = pd->port_dd;
744
745         ipath_cdbg(PROC, "%sabling rcv for unit %u port %u:%u\n",
746                    start_stop ? "en" : "dis", dd->ipath_unit,
747                    pd->port_port, subport);
748         if (subport)
749                 goto bail;
750         /* atomically clear receive enable port. */
751         if (start_stop) {
752                 /*
753                  * On enable, force in-memory copy of the tail register to
754                  * 0, so that protocol code doesn't have to worry about
755                  * whether or not the chip has yet updated the in-memory
756                  * copy or not on return from the system call. The chip
757                  * always resets it's tail register back to 0 on a
758                  * transition from disabled to enabled.  This could cause a
759                  * problem if software was broken, and did the enable w/o
760                  * the disable, but eventually the in-memory copy will be
761                  * updated and correct itself, even in the face of software
762                  * bugs.
763                  */
764                 if (pd->port_rcvhdrtail_kvaddr)
765                         ipath_clear_rcvhdrtail(pd);
766                 set_bit(dd->ipath_r_portenable_shift + pd->port_port,
767                         &dd->ipath_rcvctrl);
768         } else
769                 clear_bit(dd->ipath_r_portenable_shift + pd->port_port,
770                           &dd->ipath_rcvctrl);
771         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
772                          dd->ipath_rcvctrl);
773         /* now be sure chip saw it before we return */
774         ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
775         if (start_stop) {
776                 /*
777                  * And try to be sure that tail reg update has happened too.
778                  * This should in theory interlock with the RXE changes to
779                  * the tail register.  Don't assign it to the tail register
780                  * in memory copy, since we could overwrite an update by the
781                  * chip if we did.
782                  */
783                 ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
784         }
785         /* always; new head should be equal to new tail; see above */
786 bail:
787         return 0;
788 }
789
790 static void ipath_clean_part_key(struct ipath_portdata *pd,
791                                  struct ipath_devdata *dd)
792 {
793         int i, j, pchanged = 0;
794         u64 oldpkey;
795
796         /* for debugging only */
797         oldpkey = (u64) dd->ipath_pkeys[0] |
798                 ((u64) dd->ipath_pkeys[1] << 16) |
799                 ((u64) dd->ipath_pkeys[2] << 32) |
800                 ((u64) dd->ipath_pkeys[3] << 48);
801
802         for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
803                 if (!pd->port_pkeys[i])
804                         continue;
805                 ipath_cdbg(VERBOSE, "look for key[%d] %hx in pkeys\n", i,
806                            pd->port_pkeys[i]);
807                 for (j = 0; j < ARRAY_SIZE(dd->ipath_pkeys); j++) {
808                         /* check for match independent of the global bit */
809                         if ((dd->ipath_pkeys[j] & 0x7fff) !=
810                             (pd->port_pkeys[i] & 0x7fff))
811                                 continue;
812                         if (atomic_dec_and_test(&dd->ipath_pkeyrefs[j])) {
813                                 ipath_cdbg(VERBOSE, "p%u clear key "
814                                            "%x matches #%d\n",
815                                            pd->port_port,
816                                            pd->port_pkeys[i], j);
817                                 ipath_stats.sps_pkeys[j] =
818                                         dd->ipath_pkeys[j] = 0;
819                                 pchanged++;
820                         }
821                         else ipath_cdbg(
822                                 VERBOSE, "p%u key %x matches #%d, "
823                                 "but ref still %d\n", pd->port_port,
824                                 pd->port_pkeys[i], j,
825                                 atomic_read(&dd->ipath_pkeyrefs[j]));
826                         break;
827                 }
828                 pd->port_pkeys[i] = 0;
829         }
830         if (pchanged) {
831                 u64 pkey = (u64) dd->ipath_pkeys[0] |
832                         ((u64) dd->ipath_pkeys[1] << 16) |
833                         ((u64) dd->ipath_pkeys[2] << 32) |
834                         ((u64) dd->ipath_pkeys[3] << 48);
835                 ipath_cdbg(VERBOSE, "p%u old pkey reg %llx, "
836                            "new pkey reg %llx\n", pd->port_port,
837                            (unsigned long long) oldpkey,
838                            (unsigned long long) pkey);
839                 ipath_write_kreg(dd, dd->ipath_kregs->kr_partitionkey,
840                                  pkey);
841         }
842 }
843
844 /*
845  * Initialize the port data with the receive buffer sizes
846  * so this can be done while the master port is locked.
847  * Otherwise, there is a race with a slave opening the port
848  * and seeing these fields uninitialized.
849  */
850 static void init_user_egr_sizes(struct ipath_portdata *pd)
851 {
852         struct ipath_devdata *dd = pd->port_dd;
853         unsigned egrperchunk, egrcnt, size;
854
855         /*
856          * to avoid wasting a lot of memory, we allocate 32KB chunks of
857          * physically contiguous memory, advance through it until used up
858          * and then allocate more.  Of course, we need memory to store those
859          * extra pointers, now.  Started out with 256KB, but under heavy
860          * memory pressure (creating large files and then copying them over
861          * NFS while doing lots of MPI jobs), we hit some allocation
862          * failures, even though we can sleep...  (2.6.10) Still get
863          * failures at 64K.  32K is the lowest we can go without wasting
864          * additional memory.
865          */
866         size = 0x8000;
867         egrperchunk = size / dd->ipath_rcvegrbufsize;
868         egrcnt = dd->ipath_rcvegrcnt;
869         pd->port_rcvegrbuf_chunks = (egrcnt + egrperchunk - 1) / egrperchunk;
870         pd->port_rcvegrbufs_perchunk = egrperchunk;
871         pd->port_rcvegrbuf_size = size;
872 }
873
874 /**
875  * ipath_create_user_egr - allocate eager TID buffers
876  * @pd: the port to allocate TID buffers for
877  *
878  * This routine is now quite different for user and kernel, because
879  * the kernel uses skb's, for the accelerated network performance
880  * This is the user port version
881  *
882  * Allocate the eager TID buffers and program them into infinipath
883  * They are no longer completely contiguous, we do multiple allocation
884  * calls.
885  */
886 static int ipath_create_user_egr(struct ipath_portdata *pd)
887 {
888         struct ipath_devdata *dd = pd->port_dd;
889         unsigned e, egrcnt, egrperchunk, chunk, egrsize, egroff;
890         size_t size;
891         int ret;
892         gfp_t gfp_flags;
893
894         /*
895          * GFP_USER, but without GFP_FS, so buffer cache can be
896          * coalesced (we hope); otherwise, even at order 4,
897          * heavy filesystem activity makes these fail, and we can
898          * use compound pages.
899          */
900         gfp_flags = __GFP_WAIT | __GFP_IO | __GFP_COMP;
901
902         egrcnt = dd->ipath_rcvegrcnt;
903         /* TID number offset for this port */
904         egroff = (pd->port_port - 1) * egrcnt + dd->ipath_p0_rcvegrcnt;
905         egrsize = dd->ipath_rcvegrbufsize;
906         ipath_cdbg(VERBOSE, "Allocating %d egr buffers, at egrtid "
907                    "offset %x, egrsize %u\n", egrcnt, egroff, egrsize);
908
909         chunk = pd->port_rcvegrbuf_chunks;
910         egrperchunk = pd->port_rcvegrbufs_perchunk;
911         size = pd->port_rcvegrbuf_size;
912         pd->port_rcvegrbuf = kmalloc(chunk * sizeof(pd->port_rcvegrbuf[0]),
913                                      GFP_KERNEL);
914         if (!pd->port_rcvegrbuf) {
915                 ret = -ENOMEM;
916                 goto bail;
917         }
918         pd->port_rcvegrbuf_phys =
919                 kmalloc(chunk * sizeof(pd->port_rcvegrbuf_phys[0]),
920                         GFP_KERNEL);
921         if (!pd->port_rcvegrbuf_phys) {
922                 ret = -ENOMEM;
923                 goto bail_rcvegrbuf;
924         }
925         for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
926
927                 pd->port_rcvegrbuf[e] = dma_alloc_coherent(
928                         &dd->pcidev->dev, size, &pd->port_rcvegrbuf_phys[e],
929                         gfp_flags);
930
931                 if (!pd->port_rcvegrbuf[e]) {
932                         ret = -ENOMEM;
933                         goto bail_rcvegrbuf_phys;
934                 }
935         }
936
937         pd->port_rcvegr_phys = pd->port_rcvegrbuf_phys[0];
938
939         for (e = chunk = 0; chunk < pd->port_rcvegrbuf_chunks; chunk++) {
940                 dma_addr_t pa = pd->port_rcvegrbuf_phys[chunk];
941                 unsigned i;
942
943                 for (i = 0; e < egrcnt && i < egrperchunk; e++, i++) {
944                         dd->ipath_f_put_tid(dd, e + egroff +
945                                             (u64 __iomem *)
946                                             ((char __iomem *)
947                                              dd->ipath_kregbase +
948                                              dd->ipath_rcvegrbase),
949                                             RCVHQ_RCV_TYPE_EAGER, pa);
950                         pa += egrsize;
951                 }
952                 cond_resched(); /* don't hog the cpu */
953         }
954
955         ret = 0;
956         goto bail;
957
958 bail_rcvegrbuf_phys:
959         for (e = 0; e < pd->port_rcvegrbuf_chunks &&
960                 pd->port_rcvegrbuf[e]; e++) {
961                 dma_free_coherent(&dd->pcidev->dev, size,
962                                   pd->port_rcvegrbuf[e],
963                                   pd->port_rcvegrbuf_phys[e]);
964
965         }
966         kfree(pd->port_rcvegrbuf_phys);
967         pd->port_rcvegrbuf_phys = NULL;
968 bail_rcvegrbuf:
969         kfree(pd->port_rcvegrbuf);
970         pd->port_rcvegrbuf = NULL;
971 bail:
972         return ret;
973 }
974
975
976 /* common code for the mappings on dma_alloc_coherent mem */
977 static int ipath_mmap_mem(struct vm_area_struct *vma,
978         struct ipath_portdata *pd, unsigned len, int write_ok,
979         void *kvaddr, char *what)
980 {
981         struct ipath_devdata *dd = pd->port_dd;
982         unsigned long pfn;
983         int ret;
984
985         if ((vma->vm_end - vma->vm_start) > len) {
986                 dev_info(&dd->pcidev->dev,
987                          "FAIL on %s: len %lx > %x\n", what,
988                          vma->vm_end - vma->vm_start, len);
989                 ret = -EFAULT;
990                 goto bail;
991         }
992
993         if (!write_ok) {
994                 if (vma->vm_flags & VM_WRITE) {
995                         dev_info(&dd->pcidev->dev,
996                                  "%s must be mapped readonly\n", what);
997                         ret = -EPERM;
998                         goto bail;
999                 }
1000
1001                 /* don't allow them to later change with mprotect */
1002                 vma->vm_flags &= ~VM_MAYWRITE;
1003         }
1004
1005         pfn = virt_to_phys(kvaddr) >> PAGE_SHIFT;
1006         ret = remap_pfn_range(vma, vma->vm_start, pfn,
1007                               len, vma->vm_page_prot);
1008         if (ret)
1009                 dev_info(&dd->pcidev->dev, "%s port%u mmap of %lx, %x "
1010                          "bytes r%c failed: %d\n", what, pd->port_port,
1011                          pfn, len, write_ok?'w':'o', ret);
1012         else
1013                 ipath_cdbg(VERBOSE, "%s port%u mmaped %lx, %x bytes "
1014                            "r%c\n", what, pd->port_port, pfn, len,
1015                            write_ok?'w':'o');
1016 bail:
1017         return ret;
1018 }
1019
1020 static int mmap_ureg(struct vm_area_struct *vma, struct ipath_devdata *dd,
1021                      u64 ureg)
1022 {
1023         unsigned long phys;
1024         int ret;
1025
1026         /*
1027          * This is real hardware, so use io_remap.  This is the mechanism
1028          * for the user process to update the head registers for their port
1029          * in the chip.
1030          */
1031         if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) {
1032                 dev_info(&dd->pcidev->dev, "FAIL mmap userreg: reqlen "
1033                          "%lx > PAGE\n", vma->vm_end - vma->vm_start);
1034                 ret = -EFAULT;
1035         } else {
1036                 phys = dd->ipath_physaddr + ureg;
1037                 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1038
1039                 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1040                 ret = io_remap_pfn_range(vma, vma->vm_start,
1041                                          phys >> PAGE_SHIFT,
1042                                          vma->vm_end - vma->vm_start,
1043                                          vma->vm_page_prot);
1044         }
1045         return ret;
1046 }
1047
1048 static int mmap_piobufs(struct vm_area_struct *vma,
1049                         struct ipath_devdata *dd,
1050                         struct ipath_portdata *pd,
1051                         unsigned piobufs, unsigned piocnt)
1052 {
1053         unsigned long phys;
1054         int ret;
1055
1056         /*
1057          * When we map the PIO buffers in the chip, we want to map them as
1058          * writeonly, no read possible.   This prevents access to previous
1059          * process data, and catches users who might try to read the i/o
1060          * space due to a bug.
1061          */
1062         if ((vma->vm_end - vma->vm_start) > (piocnt * dd->ipath_palign)) {
1063                 dev_info(&dd->pcidev->dev, "FAIL mmap piobufs: "
1064                          "reqlen %lx > PAGE\n",
1065                          vma->vm_end - vma->vm_start);
1066                 ret = -EINVAL;
1067                 goto bail;
1068         }
1069
1070         phys = dd->ipath_physaddr + piobufs;
1071
1072 #if defined(__powerpc__)
1073         /* There isn't a generic way to specify writethrough mappings */
1074         pgprot_val(vma->vm_page_prot) |= _PAGE_NO_CACHE;
1075         pgprot_val(vma->vm_page_prot) |= _PAGE_WRITETHRU;
1076         pgprot_val(vma->vm_page_prot) &= ~_PAGE_GUARDED;
1077 #endif
1078
1079         /*
1080          * don't allow them to later change to readable with mprotect (for when
1081          * not initially mapped readable, as is normally the case)
1082          */
1083         vma->vm_flags &= ~VM_MAYREAD;
1084         vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1085
1086         ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
1087                                  vma->vm_end - vma->vm_start,
1088                                  vma->vm_page_prot);
1089 bail:
1090         return ret;
1091 }
1092
1093 static int mmap_rcvegrbufs(struct vm_area_struct *vma,
1094                            struct ipath_portdata *pd)
1095 {
1096         struct ipath_devdata *dd = pd->port_dd;
1097         unsigned long start, size;
1098         size_t total_size, i;
1099         unsigned long pfn;
1100         int ret;
1101
1102         size = pd->port_rcvegrbuf_size;
1103         total_size = pd->port_rcvegrbuf_chunks * size;
1104         if ((vma->vm_end - vma->vm_start) > total_size) {
1105                 dev_info(&dd->pcidev->dev, "FAIL on egr bufs: "
1106                          "reqlen %lx > actual %lx\n",
1107                          vma->vm_end - vma->vm_start,
1108                          (unsigned long) total_size);
1109                 ret = -EINVAL;
1110                 goto bail;
1111         }
1112
1113         if (vma->vm_flags & VM_WRITE) {
1114                 dev_info(&dd->pcidev->dev, "Can't map eager buffers as "
1115                          "writable (flags=%lx)\n", vma->vm_flags);
1116                 ret = -EPERM;
1117                 goto bail;
1118         }
1119         /* don't allow them to later change to writeable with mprotect */
1120         vma->vm_flags &= ~VM_MAYWRITE;
1121
1122         start = vma->vm_start;
1123
1124         for (i = 0; i < pd->port_rcvegrbuf_chunks; i++, start += size) {
1125                 pfn = virt_to_phys(pd->port_rcvegrbuf[i]) >> PAGE_SHIFT;
1126                 ret = remap_pfn_range(vma, start, pfn, size,
1127                                       vma->vm_page_prot);
1128                 if (ret < 0)
1129                         goto bail;
1130         }
1131         ret = 0;
1132
1133 bail:
1134         return ret;
1135 }
1136
1137 /*
1138  * ipath_file_vma_fault - handle a VMA page fault.
1139  */
1140 static int ipath_file_vma_fault(struct vm_area_struct *vma,
1141                                         struct vm_fault *vmf)
1142 {
1143         struct page *page;
1144
1145         page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
1146         if (!page)
1147                 return VM_FAULT_SIGBUS;
1148         get_page(page);
1149         vmf->page = page;
1150
1151         return 0;
1152 }
1153
1154 static struct vm_operations_struct ipath_file_vm_ops = {
1155         .fault = ipath_file_vma_fault,
1156 };
1157
1158 static int mmap_kvaddr(struct vm_area_struct *vma, u64 pgaddr,
1159                        struct ipath_portdata *pd, unsigned subport)
1160 {
1161         unsigned long len;
1162         struct ipath_devdata *dd;
1163         void *addr;
1164         size_t size;
1165         int ret = 0;
1166
1167         /* If the port is not shared, all addresses should be physical */
1168         if (!pd->port_subport_cnt)
1169                 goto bail;
1170
1171         dd = pd->port_dd;
1172         size = pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
1173
1174         /*
1175          * Each process has all the subport uregbase, rcvhdrq, and
1176          * rcvegrbufs mmapped - as an array for all the processes,
1177          * and also separately for this process.
1178          */
1179         if (pgaddr == cvt_kvaddr(pd->subport_uregbase)) {
1180                 addr = pd->subport_uregbase;
1181                 size = PAGE_SIZE * pd->port_subport_cnt;
1182         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base)) {
1183                 addr = pd->subport_rcvhdr_base;
1184                 size = pd->port_rcvhdrq_size * pd->port_subport_cnt;
1185         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf)) {
1186                 addr = pd->subport_rcvegrbuf;
1187                 size *= pd->port_subport_cnt;
1188         } else if (pgaddr == cvt_kvaddr(pd->subport_uregbase +
1189                                         PAGE_SIZE * subport)) {
1190                 addr = pd->subport_uregbase + PAGE_SIZE * subport;
1191                 size = PAGE_SIZE;
1192         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base +
1193                                 pd->port_rcvhdrq_size * subport)) {
1194                 addr = pd->subport_rcvhdr_base +
1195                         pd->port_rcvhdrq_size * subport;
1196                 size = pd->port_rcvhdrq_size;
1197         } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf +
1198                                size * subport)) {
1199                 addr = pd->subport_rcvegrbuf + size * subport;
1200                 /* rcvegrbufs are read-only on the slave */
1201                 if (vma->vm_flags & VM_WRITE) {
1202                         dev_info(&dd->pcidev->dev,
1203                                  "Can't map eager buffers as "
1204                                  "writable (flags=%lx)\n", vma->vm_flags);
1205                         ret = -EPERM;
1206                         goto bail;
1207                 }
1208                 /*
1209                  * Don't allow permission to later change to writeable
1210                  * with mprotect.
1211                  */
1212                 vma->vm_flags &= ~VM_MAYWRITE;
1213         } else {
1214                 goto bail;
1215         }
1216         len = vma->vm_end - vma->vm_start;
1217         if (len > size) {
1218                 ipath_cdbg(MM, "FAIL: reqlen %lx > %zx\n", len, size);
1219                 ret = -EINVAL;
1220                 goto bail;
1221         }
1222
1223         vma->vm_pgoff = (unsigned long) addr >> PAGE_SHIFT;
1224         vma->vm_ops = &ipath_file_vm_ops;
1225         vma->vm_flags |= VM_RESERVED | VM_DONTEXPAND;
1226         ret = 1;
1227
1228 bail:
1229         return ret;
1230 }
1231
1232 /**
1233  * ipath_mmap - mmap various structures into user space
1234  * @fp: the file pointer
1235  * @vma: the VM area
1236  *
1237  * We use this to have a shared buffer between the kernel and the user code
1238  * for the rcvhdr queue, egr buffers, and the per-port user regs and pio
1239  * buffers in the chip.  We have the open and close entries so we can bump
1240  * the ref count and keep the driver from being unloaded while still mapped.
1241  */
1242 static int ipath_mmap(struct file *fp, struct vm_area_struct *vma)
1243 {
1244         struct ipath_portdata *pd;
1245         struct ipath_devdata *dd;
1246         u64 pgaddr, ureg;
1247         unsigned piobufs, piocnt;
1248         int ret;
1249
1250         pd = port_fp(fp);
1251         if (!pd) {
1252                 ret = -EINVAL;
1253                 goto bail;
1254         }
1255         dd = pd->port_dd;
1256
1257         /*
1258          * This is the ipath_do_user_init() code, mapping the shared buffers
1259          * into the user process. The address referred to by vm_pgoff is the
1260          * file offset passed via mmap().  For shared ports, this is the
1261          * kernel vmalloc() address of the pages to share with the master.
1262          * For non-shared or master ports, this is a physical address.
1263          * We only do one mmap for each space mapped.
1264          */
1265         pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1266
1267         /*
1268          * Check for 0 in case one of the allocations failed, but user
1269          * called mmap anyway.
1270          */
1271         if (!pgaddr)  {
1272                 ret = -EINVAL;
1273                 goto bail;
1274         }
1275
1276         ipath_cdbg(MM, "pgaddr %llx vm_start=%lx len %lx port %u:%u:%u\n",
1277                    (unsigned long long) pgaddr, vma->vm_start,
1278                    vma->vm_end - vma->vm_start, dd->ipath_unit,
1279                    pd->port_port, subport_fp(fp));
1280
1281         /*
1282          * Physical addresses must fit in 40 bits for our hardware.
1283          * Check for kernel virtual addresses first, anything else must
1284          * match a HW or memory address.
1285          */
1286         ret = mmap_kvaddr(vma, pgaddr, pd, subport_fp(fp));
1287         if (ret) {
1288                 if (ret > 0)
1289                         ret = 0;
1290                 goto bail;
1291         }
1292
1293         ureg = dd->ipath_uregbase + dd->ipath_ureg_align * pd->port_port;
1294         if (!pd->port_subport_cnt) {
1295                 /* port is not shared */
1296                 piocnt = pd->port_piocnt;
1297                 piobufs = pd->port_piobufs;
1298         } else if (!subport_fp(fp)) {
1299                 /* caller is the master */
1300                 piocnt = (pd->port_piocnt / pd->port_subport_cnt) +
1301                          (pd->port_piocnt % pd->port_subport_cnt);
1302                 piobufs = pd->port_piobufs +
1303                         dd->ipath_palign * (pd->port_piocnt - piocnt);
1304         } else {
1305                 unsigned slave = subport_fp(fp) - 1;
1306
1307                 /* caller is a slave */
1308                 piocnt = pd->port_piocnt / pd->port_subport_cnt;
1309                 piobufs = pd->port_piobufs + dd->ipath_palign * piocnt * slave;
1310         }
1311
1312         if (pgaddr == ureg)
1313                 ret = mmap_ureg(vma, dd, ureg);
1314         else if (pgaddr == piobufs)
1315                 ret = mmap_piobufs(vma, dd, pd, piobufs, piocnt);
1316         else if (pgaddr == dd->ipath_pioavailregs_phys)
1317                 /* in-memory copy of pioavail registers */
1318                 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1319                                      (void *) dd->ipath_pioavailregs_dma,
1320                                      "pioavail registers");
1321         else if (pgaddr == pd->port_rcvegr_phys)
1322                 ret = mmap_rcvegrbufs(vma, pd);
1323         else if (pgaddr == (u64) pd->port_rcvhdrq_phys)
1324                 /*
1325                  * The rcvhdrq itself; readonly except on HT (so have
1326                  * to allow writable mapping), multiple pages, contiguous
1327                  * from an i/o perspective.
1328                  */
1329                 ret = ipath_mmap_mem(vma, pd, pd->port_rcvhdrq_size, 1,
1330                                      pd->port_rcvhdrq,
1331                                      "rcvhdrq");
1332         else if (pgaddr == (u64) pd->port_rcvhdrqtailaddr_phys)
1333                 /* in-memory copy of rcvhdrq tail register */
1334                 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1335                                      pd->port_rcvhdrtail_kvaddr,
1336                                      "rcvhdrq tail");
1337         else
1338                 ret = -EINVAL;
1339
1340         vma->vm_private_data = NULL;
1341
1342         if (ret < 0)
1343                 dev_info(&dd->pcidev->dev,
1344                          "Failure %d on off %llx len %lx\n",
1345                          -ret, (unsigned long long)pgaddr,
1346                          vma->vm_end - vma->vm_start);
1347 bail:
1348         return ret;
1349 }
1350
1351 static unsigned ipath_poll_hdrqfull(struct ipath_portdata *pd)
1352 {
1353         unsigned pollflag = 0;
1354
1355         if ((pd->poll_type & IPATH_POLL_TYPE_OVERFLOW) &&
1356             pd->port_hdrqfull != pd->port_hdrqfull_poll) {
1357                 pollflag |= POLLIN | POLLRDNORM;
1358                 pd->port_hdrqfull_poll = pd->port_hdrqfull;
1359         }
1360
1361         return pollflag;
1362 }
1363
1364 static unsigned int ipath_poll_urgent(struct ipath_portdata *pd,
1365                                       struct file *fp,
1366                                       struct poll_table_struct *pt)
1367 {
1368         unsigned pollflag = 0;
1369         struct ipath_devdata *dd;
1370
1371         dd = pd->port_dd;
1372
1373         /* variable access in ipath_poll_hdrqfull() needs this */
1374         rmb();
1375         pollflag = ipath_poll_hdrqfull(pd);
1376
1377         if (pd->port_urgent != pd->port_urgent_poll) {
1378                 pollflag |= POLLIN | POLLRDNORM;
1379                 pd->port_urgent_poll = pd->port_urgent;
1380         }
1381
1382         if (!pollflag) {
1383                 /* this saves a spin_lock/unlock in interrupt handler... */
1384                 set_bit(IPATH_PORT_WAITING_URG, &pd->port_flag);
1385                 /* flush waiting flag so don't miss an event... */
1386                 wmb();
1387                 poll_wait(fp, &pd->port_wait, pt);
1388         }
1389
1390         return pollflag;
1391 }
1392
1393 static unsigned int ipath_poll_next(struct ipath_portdata *pd,
1394                                     struct file *fp,
1395                                     struct poll_table_struct *pt)
1396 {
1397         u32 head;
1398         u32 tail;
1399         unsigned pollflag = 0;
1400         struct ipath_devdata *dd;
1401
1402         dd = pd->port_dd;
1403
1404         /* variable access in ipath_poll_hdrqfull() needs this */
1405         rmb();
1406         pollflag = ipath_poll_hdrqfull(pd);
1407
1408         head = ipath_read_ureg32(dd, ur_rcvhdrhead, pd->port_port);
1409         if (pd->port_rcvhdrtail_kvaddr)
1410                 tail = ipath_get_rcvhdrtail(pd);
1411         else
1412                 tail = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
1413
1414         if (head != tail)
1415                 pollflag |= POLLIN | POLLRDNORM;
1416         else {
1417                 /* this saves a spin_lock/unlock in interrupt handler */
1418                 set_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
1419                 /* flush waiting flag so we don't miss an event */
1420                 wmb();
1421
1422                 set_bit(pd->port_port + dd->ipath_r_intravail_shift,
1423                         &dd->ipath_rcvctrl);
1424
1425                 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1426                                  dd->ipath_rcvctrl);
1427
1428                 if (dd->ipath_rhdrhead_intr_off) /* arm rcv interrupt */
1429                         ipath_write_ureg(dd, ur_rcvhdrhead,
1430                                          dd->ipath_rhdrhead_intr_off | head,
1431                                          pd->port_port);
1432
1433                 poll_wait(fp, &pd->port_wait, pt);
1434         }
1435
1436         return pollflag;
1437 }
1438
1439 static unsigned int ipath_poll(struct file *fp,
1440                                struct poll_table_struct *pt)
1441 {
1442         struct ipath_portdata *pd;
1443         unsigned pollflag;
1444
1445         pd = port_fp(fp);
1446         if (!pd)
1447                 pollflag = 0;
1448         else if (pd->poll_type & IPATH_POLL_TYPE_URGENT)
1449                 pollflag = ipath_poll_urgent(pd, fp, pt);
1450         else
1451                 pollflag = ipath_poll_next(pd, fp, pt);
1452
1453         return pollflag;
1454 }
1455
1456 static int ipath_supports_subports(int user_swmajor, int user_swminor)
1457 {
1458         /* no subport implementation prior to software version 1.3 */
1459         return (user_swmajor > 1) || (user_swminor >= 3);
1460 }
1461
1462 static int ipath_compatible_subports(int user_swmajor, int user_swminor)
1463 {
1464         /* this code is written long-hand for clarity */
1465         if (IPATH_USER_SWMAJOR != user_swmajor) {
1466                 /* no promise of compatibility if major mismatch */
1467                 return 0;
1468         }
1469         if (IPATH_USER_SWMAJOR == 1) {
1470                 switch (IPATH_USER_SWMINOR) {
1471                 case 0:
1472                 case 1:
1473                 case 2:
1474                         /* no subport implementation so cannot be compatible */
1475                         return 0;
1476                 case 3:
1477                         /* 3 is only compatible with itself */
1478                         return user_swminor == 3;
1479                 default:
1480                         /* >= 4 are compatible (or are expected to be) */
1481                         return user_swminor >= 4;
1482                 }
1483         }
1484         /* make no promises yet for future major versions */
1485         return 0;
1486 }
1487
1488 static int init_subports(struct ipath_devdata *dd,
1489                          struct ipath_portdata *pd,
1490                          const struct ipath_user_info *uinfo)
1491 {
1492         int ret = 0;
1493         unsigned num_subports;
1494         size_t size;
1495
1496         /*
1497          * If the user is requesting zero subports,
1498          * skip the subport allocation.
1499          */
1500         if (uinfo->spu_subport_cnt <= 0)
1501                 goto bail;
1502
1503         /* Self-consistency check for ipath_compatible_subports() */
1504         if (ipath_supports_subports(IPATH_USER_SWMAJOR, IPATH_USER_SWMINOR) &&
1505             !ipath_compatible_subports(IPATH_USER_SWMAJOR,
1506                                        IPATH_USER_SWMINOR)) {
1507                 dev_info(&dd->pcidev->dev,
1508                          "Inconsistent ipath_compatible_subports()\n");
1509                 goto bail;
1510         }
1511
1512         /* Check for subport compatibility */
1513         if (!ipath_compatible_subports(uinfo->spu_userversion >> 16,
1514                                        uinfo->spu_userversion & 0xffff)) {
1515                 dev_info(&dd->pcidev->dev,
1516                          "Mismatched user version (%d.%d) and driver "
1517                          "version (%d.%d) while port sharing. Ensure "
1518                          "that driver and library are from the same "
1519                          "release.\n",
1520                          (int) (uinfo->spu_userversion >> 16),
1521                          (int) (uinfo->spu_userversion & 0xffff),
1522                          IPATH_USER_SWMAJOR,
1523                          IPATH_USER_SWMINOR);
1524                 goto bail;
1525         }
1526         if (uinfo->spu_subport_cnt > INFINIPATH_MAX_SUBPORT) {
1527                 ret = -EINVAL;
1528                 goto bail;
1529         }
1530
1531         num_subports = uinfo->spu_subport_cnt;
1532         pd->subport_uregbase = vmalloc(PAGE_SIZE * num_subports);
1533         if (!pd->subport_uregbase) {
1534                 ret = -ENOMEM;
1535                 goto bail;
1536         }
1537         /* Note: pd->port_rcvhdrq_size isn't initialized yet. */
1538         size = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
1539                      sizeof(u32), PAGE_SIZE) * num_subports;
1540         pd->subport_rcvhdr_base = vmalloc(size);
1541         if (!pd->subport_rcvhdr_base) {
1542                 ret = -ENOMEM;
1543                 goto bail_ureg;
1544         }
1545
1546         pd->subport_rcvegrbuf = vmalloc(pd->port_rcvegrbuf_chunks *
1547                                         pd->port_rcvegrbuf_size *
1548                                         num_subports);
1549         if (!pd->subport_rcvegrbuf) {
1550                 ret = -ENOMEM;
1551                 goto bail_rhdr;
1552         }
1553
1554         pd->port_subport_cnt = uinfo->spu_subport_cnt;
1555         pd->port_subport_id = uinfo->spu_subport_id;
1556         pd->active_slaves = 1;
1557         set_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
1558         memset(pd->subport_uregbase, 0, PAGE_SIZE * num_subports);
1559         memset(pd->subport_rcvhdr_base, 0, size);
1560         memset(pd->subport_rcvegrbuf, 0, pd->port_rcvegrbuf_chunks *
1561                                          pd->port_rcvegrbuf_size *
1562                                          num_subports);
1563         goto bail;
1564
1565 bail_rhdr:
1566         vfree(pd->subport_rcvhdr_base);
1567 bail_ureg:
1568         vfree(pd->subport_uregbase);
1569         pd->subport_uregbase = NULL;
1570 bail:
1571         return ret;
1572 }
1573
1574 static int try_alloc_port(struct ipath_devdata *dd, int port,
1575                           struct file *fp,
1576                           const struct ipath_user_info *uinfo)
1577 {
1578         struct ipath_portdata *pd;
1579         int ret;
1580
1581         if (!(pd = dd->ipath_pd[port])) {
1582                 void *ptmp;
1583
1584                 pd = kzalloc(sizeof(struct ipath_portdata), GFP_KERNEL);
1585
1586                 /*
1587                  * Allocate memory for use in ipath_tid_update() just once
1588                  * at open, not per call.  Reduces cost of expected send
1589                  * setup.
1590                  */
1591                 ptmp = kmalloc(dd->ipath_rcvtidcnt * sizeof(u16) +
1592                                dd->ipath_rcvtidcnt * sizeof(struct page **),
1593                                GFP_KERNEL);
1594                 if (!pd || !ptmp) {
1595                         ipath_dev_err(dd, "Unable to allocate portdata "
1596                                       "memory, failing open\n");
1597                         ret = -ENOMEM;
1598                         kfree(pd);
1599                         kfree(ptmp);
1600                         goto bail;
1601                 }
1602                 dd->ipath_pd[port] = pd;
1603                 dd->ipath_pd[port]->port_port = port;
1604                 dd->ipath_pd[port]->port_dd = dd;
1605                 dd->ipath_pd[port]->port_tid_pg_list = ptmp;
1606                 init_waitqueue_head(&dd->ipath_pd[port]->port_wait);
1607         }
1608         if (!pd->port_cnt) {
1609                 pd->userversion = uinfo->spu_userversion;
1610                 init_user_egr_sizes(pd);
1611                 if ((ret = init_subports(dd, pd, uinfo)) != 0)
1612                         goto bail;
1613                 ipath_cdbg(PROC, "%s[%u] opened unit:port %u:%u\n",
1614                            current->comm, current->pid, dd->ipath_unit,
1615                            port);
1616                 pd->port_cnt = 1;
1617                 port_fp(fp) = pd;
1618                 pd->port_pid = get_pid(task_pid(current));
1619                 strncpy(pd->port_comm, current->comm, sizeof(pd->port_comm));
1620                 ipath_stats.sps_ports++;
1621                 ret = 0;
1622         } else
1623                 ret = -EBUSY;
1624
1625 bail:
1626         return ret;
1627 }
1628
1629 static inline int usable(struct ipath_devdata *dd)
1630 {
1631         return dd &&
1632                 (dd->ipath_flags & IPATH_PRESENT) &&
1633                 dd->ipath_kregbase &&
1634                 dd->ipath_lid &&
1635                 !(dd->ipath_flags & (IPATH_LINKDOWN | IPATH_DISABLED
1636                                      | IPATH_LINKUNK));
1637 }
1638
1639 static int find_free_port(int unit, struct file *fp,
1640                           const struct ipath_user_info *uinfo)
1641 {
1642         struct ipath_devdata *dd = ipath_lookup(unit);
1643         int ret, i;
1644
1645         if (!dd) {
1646                 ret = -ENODEV;
1647                 goto bail;
1648         }
1649
1650         if (!usable(dd)) {
1651                 ret = -ENETDOWN;
1652                 goto bail;
1653         }
1654
1655         for (i = 1; i < dd->ipath_cfgports; i++) {
1656                 ret = try_alloc_port(dd, i, fp, uinfo);
1657                 if (ret != -EBUSY)
1658                         goto bail;
1659         }
1660         ret = -EBUSY;
1661
1662 bail:
1663         return ret;
1664 }
1665
1666 static int find_best_unit(struct file *fp,
1667                           const struct ipath_user_info *uinfo)
1668 {
1669         int ret = 0, i, prefunit = -1, devmax;
1670         int maxofallports, npresent, nup;
1671         int ndev;
1672
1673         devmax = ipath_count_units(&npresent, &nup, &maxofallports);
1674
1675         /*
1676          * This code is present to allow a knowledgeable person to
1677          * specify the layout of processes to processors before opening
1678          * this driver, and then we'll assign the process to the "closest"
1679          * InfiniPath chip to that processor (we assume reasonable connectivity,
1680          * for now).  This code assumes that if affinity has been set
1681          * before this point, that at most one cpu is set; for now this
1682          * is reasonable.  I check for both cpumask_empty() and cpumask_full(),
1683          * in case some kernel variant sets none of the bits when no
1684          * affinity is set.  2.6.11 and 12 kernels have all present
1685          * cpus set.  Some day we'll have to fix it up further to handle
1686          * a cpu subset.  This algorithm fails for two HT chips connected
1687          * in tunnel fashion.  Eventually this needs real topology
1688          * information.  There may be some issues with dual core numbering
1689          * as well.  This needs more work prior to release.
1690          */
1691         if (!cpumask_empty(&current->cpus_allowed) &&
1692             !cpumask_full(&current->cpus_allowed)) {
1693                 int ncpus = num_online_cpus(), curcpu = -1, nset = 0;
1694                 for (i = 0; i < ncpus; i++)
1695                         if (cpumask_test_cpu(i, &current->cpus_allowed)) {
1696                                 ipath_cdbg(PROC, "%s[%u] affinity set for "
1697                                            "cpu %d/%d\n", current->comm,
1698                                            current->pid, i, ncpus);
1699                                 curcpu = i;
1700                                 nset++;
1701                         }
1702                 if (curcpu != -1 && nset != ncpus) {
1703                         if (npresent) {
1704                                 prefunit = curcpu / (ncpus / npresent);
1705                                 ipath_cdbg(PROC,"%s[%u] %d chips, %d cpus, "
1706                                           "%d cpus/chip, select unit %d\n",
1707                                           current->comm, current->pid,
1708                                           npresent, ncpus, ncpus / npresent,
1709                                           prefunit);
1710                         }
1711                 }
1712         }
1713
1714         /*
1715          * user ports start at 1, kernel port is 0
1716          * For now, we do round-robin access across all chips
1717          */
1718
1719         if (prefunit != -1)
1720                 devmax = prefunit + 1;
1721 recheck:
1722         for (i = 1; i < maxofallports; i++) {
1723                 for (ndev = prefunit != -1 ? prefunit : 0; ndev < devmax;
1724                      ndev++) {
1725                         struct ipath_devdata *dd = ipath_lookup(ndev);
1726
1727                         if (!usable(dd))
1728                                 continue; /* can't use this unit */
1729                         if (i >= dd->ipath_cfgports)
1730                                 /*
1731                                  * Maxed out on users of this unit. Try
1732                                  * next.
1733                                  */
1734                                 continue;
1735                         ret = try_alloc_port(dd, i, fp, uinfo);
1736                         if (!ret)
1737                                 goto done;
1738                 }
1739         }
1740
1741         if (npresent) {
1742                 if (nup == 0) {
1743                         ret = -ENETDOWN;
1744                         ipath_dbg("No ports available (none initialized "
1745                                   "and ready)\n");
1746                 } else {
1747                         if (prefunit > 0) {
1748                                 /* if started above 0, retry from 0 */
1749                                 ipath_cdbg(PROC,
1750                                            "%s[%u] no ports on prefunit "
1751                                            "%d, clear and re-check\n",
1752                                            current->comm, current->pid,
1753                                            prefunit);
1754                                 devmax = ipath_count_units(NULL, NULL,
1755                                                            NULL);
1756                                 prefunit = -1;
1757                                 goto recheck;
1758                         }
1759                         ret = -EBUSY;
1760                         ipath_dbg("No ports available\n");
1761                 }
1762         } else {
1763                 ret = -ENXIO;
1764                 ipath_dbg("No boards found\n");
1765         }
1766
1767 done:
1768         return ret;
1769 }
1770
1771 static int find_shared_port(struct file *fp,
1772                             const struct ipath_user_info *uinfo)
1773 {
1774         int devmax, ndev, i;
1775         int ret = 0;
1776
1777         devmax = ipath_count_units(NULL, NULL, NULL);
1778
1779         for (ndev = 0; ndev < devmax; ndev++) {
1780                 struct ipath_devdata *dd = ipath_lookup(ndev);
1781
1782                 if (!usable(dd))
1783                         continue;
1784                 for (i = 1; i < dd->ipath_cfgports; i++) {
1785                         struct ipath_portdata *pd = dd->ipath_pd[i];
1786
1787                         /* Skip ports which are not yet open */
1788                         if (!pd || !pd->port_cnt)
1789                                 continue;
1790                         /* Skip port if it doesn't match the requested one */
1791                         if (pd->port_subport_id != uinfo->spu_subport_id)
1792                                 continue;
1793                         /* Verify the sharing process matches the master */
1794                         if (pd->port_subport_cnt != uinfo->spu_subport_cnt ||
1795                             pd->userversion != uinfo->spu_userversion ||
1796                             pd->port_cnt >= pd->port_subport_cnt) {
1797                                 ret = -EINVAL;
1798                                 goto done;
1799                         }
1800                         port_fp(fp) = pd;
1801                         subport_fp(fp) = pd->port_cnt++;
1802                         pd->port_subpid[subport_fp(fp)] =
1803                                 get_pid(task_pid(current));
1804                         tidcursor_fp(fp) = 0;
1805                         pd->active_slaves |= 1 << subport_fp(fp);
1806                         ipath_cdbg(PROC,
1807                                    "%s[%u] %u sharing %s[%u] unit:port %u:%u\n",
1808                                    current->comm, current->pid,
1809                                    subport_fp(fp),
1810                                    pd->port_comm, pid_nr(pd->port_pid),
1811                                    dd->ipath_unit, pd->port_port);
1812                         ret = 1;
1813                         goto done;
1814                 }
1815         }
1816
1817 done:
1818         return ret;
1819 }
1820
1821 static int ipath_open(struct inode *in, struct file *fp)
1822 {
1823         /* The real work is performed later in ipath_assign_port() */
1824         cycle_kernel_lock();
1825         fp->private_data = kzalloc(sizeof(struct ipath_filedata), GFP_KERNEL);
1826         return fp->private_data ? 0 : -ENOMEM;
1827 }
1828
1829 /* Get port early, so can set affinity prior to memory allocation */
1830 static int ipath_assign_port(struct file *fp,
1831                               const struct ipath_user_info *uinfo)
1832 {
1833         int ret;
1834         int i_minor;
1835         unsigned swmajor, swminor;
1836
1837         /* Check to be sure we haven't already initialized this file */
1838         if (port_fp(fp)) {
1839                 ret = -EINVAL;
1840                 goto done;
1841         }
1842
1843         /* for now, if major version is different, bail */
1844         swmajor = uinfo->spu_userversion >> 16;
1845         if (swmajor != IPATH_USER_SWMAJOR) {
1846                 ipath_dbg("User major version %d not same as driver "
1847                           "major %d\n", uinfo->spu_userversion >> 16,
1848                           IPATH_USER_SWMAJOR);
1849                 ret = -ENODEV;
1850                 goto done;
1851         }
1852
1853         swminor = uinfo->spu_userversion & 0xffff;
1854         if (swminor != IPATH_USER_SWMINOR)
1855                 ipath_dbg("User minor version %d not same as driver "
1856                           "minor %d\n", swminor, IPATH_USER_SWMINOR);
1857
1858         mutex_lock(&ipath_mutex);
1859
1860         if (ipath_compatible_subports(swmajor, swminor) &&
1861             uinfo->spu_subport_cnt &&
1862             (ret = find_shared_port(fp, uinfo))) {
1863                 if (ret > 0)
1864                         ret = 0;
1865                 goto done_chk_sdma;
1866         }
1867
1868         i_minor = iminor(fp->f_path.dentry->d_inode) - IPATH_USER_MINOR_BASE;
1869         ipath_cdbg(VERBOSE, "open on dev %lx (minor %d)\n",
1870                    (long)fp->f_path.dentry->d_inode->i_rdev, i_minor);
1871
1872         if (i_minor)
1873                 ret = find_free_port(i_minor - 1, fp, uinfo);
1874         else
1875                 ret = find_best_unit(fp, uinfo);
1876
1877 done_chk_sdma:
1878         if (!ret) {
1879                 struct ipath_filedata *fd = fp->private_data;
1880                 const struct ipath_portdata *pd = fd->pd;
1881                 const struct ipath_devdata *dd = pd->port_dd;
1882
1883                 fd->pq = ipath_user_sdma_queue_create(&dd->pcidev->dev,
1884                                                       dd->ipath_unit,
1885                                                       pd->port_port,
1886                                                       fd->subport);
1887
1888                 if (!fd->pq)
1889                         ret = -ENOMEM;
1890         }
1891
1892         mutex_unlock(&ipath_mutex);
1893
1894 done:
1895         return ret;
1896 }
1897
1898
1899 static int ipath_do_user_init(struct file *fp,
1900                               const struct ipath_user_info *uinfo)
1901 {
1902         int ret;
1903         struct ipath_portdata *pd = port_fp(fp);
1904         struct ipath_devdata *dd;
1905         u32 head32;
1906
1907         /* Subports don't need to initialize anything since master did it. */
1908         if (subport_fp(fp)) {
1909                 ret = wait_event_interruptible(pd->port_wait,
1910                         !test_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag));
1911                 goto done;
1912         }
1913
1914         dd = pd->port_dd;
1915
1916         if (uinfo->spu_rcvhdrsize) {
1917                 ret = ipath_setrcvhdrsize(dd, uinfo->spu_rcvhdrsize);
1918                 if (ret)
1919                         goto done;
1920         }
1921
1922         /* for now we do nothing with rcvhdrcnt: uinfo->spu_rcvhdrcnt */
1923
1924         /* some ports may get extra buffers, calculate that here */
1925         if (pd->port_port <= dd->ipath_ports_extrabuf)
1926                 pd->port_piocnt = dd->ipath_pbufsport + 1;
1927         else
1928                 pd->port_piocnt = dd->ipath_pbufsport;
1929
1930         /* for right now, kernel piobufs are at end, so port 1 is at 0 */
1931         if (pd->port_port <= dd->ipath_ports_extrabuf)
1932                 pd->port_pio_base = (dd->ipath_pbufsport + 1)
1933                         * (pd->port_port - 1);
1934         else
1935                 pd->port_pio_base = dd->ipath_ports_extrabuf +
1936                         dd->ipath_pbufsport * (pd->port_port - 1);
1937         pd->port_piobufs = dd->ipath_piobufbase +
1938                 pd->port_pio_base * dd->ipath_palign;
1939         ipath_cdbg(VERBOSE, "piobuf base for port %u is 0x%x, piocnt %u,"
1940                 " first pio %u\n", pd->port_port, pd->port_piobufs,
1941                 pd->port_piocnt, pd->port_pio_base);
1942         ipath_chg_pioavailkernel(dd, pd->port_pio_base, pd->port_piocnt, 0);
1943
1944         /*
1945          * Now allocate the rcvhdr Q and eager TIDs; skip the TID
1946          * array for time being.  If pd->port_port > chip-supported,
1947          * we need to do extra stuff here to handle by handling overflow
1948          * through port 0, someday
1949          */
1950         ret = ipath_create_rcvhdrq(dd, pd);
1951         if (!ret)
1952                 ret = ipath_create_user_egr(pd);
1953         if (ret)
1954                 goto done;
1955
1956         /*
1957          * set the eager head register for this port to the current values
1958          * of the tail pointers, since we don't know if they were
1959          * updated on last use of the port.
1960          */
1961         head32 = ipath_read_ureg32(dd, ur_rcvegrindextail, pd->port_port);
1962         ipath_write_ureg(dd, ur_rcvegrindexhead, head32, pd->port_port);
1963         pd->port_lastrcvhdrqtail = -1;
1964         ipath_cdbg(VERBOSE, "Wrote port%d egrhead %x from tail regs\n",
1965                 pd->port_port, head32);
1966         pd->port_tidcursor = 0; /* start at beginning after open */
1967
1968         /* initialize poll variables... */
1969         pd->port_urgent = 0;
1970         pd->port_urgent_poll = 0;
1971         pd->port_hdrqfull_poll = pd->port_hdrqfull;
1972
1973         /*
1974          * Now enable the port for receive.
1975          * For chips that are set to DMA the tail register to memory
1976          * when they change (and when the update bit transitions from
1977          * 0 to 1.  So for those chips, we turn it off and then back on.
1978          * This will (very briefly) affect any other open ports, but the
1979          * duration is very short, and therefore isn't an issue.  We
1980          * explictly set the in-memory tail copy to 0 beforehand, so we
1981          * don't have to wait to be sure the DMA update has happened
1982          * (chip resets head/tail to 0 on transition to enable).
1983          */
1984         set_bit(dd->ipath_r_portenable_shift + pd->port_port,
1985                 &dd->ipath_rcvctrl);
1986         if (!(dd->ipath_flags & IPATH_NODMA_RTAIL)) {
1987                 if (pd->port_rcvhdrtail_kvaddr)
1988                         ipath_clear_rcvhdrtail(pd);
1989                 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1990                         dd->ipath_rcvctrl &
1991                         ~(1ULL << dd->ipath_r_tailupd_shift));
1992         }
1993         ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1994                          dd->ipath_rcvctrl);
1995         /* Notify any waiting slaves */
1996         if (pd->port_subport_cnt) {
1997                 clear_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
1998                 wake_up(&pd->port_wait);
1999         }
2000 done:
2001         return ret;
2002 }
2003
2004 /**
2005  * unlock_exptid - unlock any expected TID entries port still had in use
2006  * @pd: port
2007  *
2008  * We don't actually update the chip here, because we do a bulk update
2009  * below, using ipath_f_clear_tids.
2010  */
2011 static void unlock_expected_tids(struct ipath_portdata *pd)
2012 {
2013         struct ipath_devdata *dd = pd->port_dd;
2014         int port_tidbase = pd->port_port * dd->ipath_rcvtidcnt;
2015         int i, cnt = 0, maxtid = port_tidbase + dd->ipath_rcvtidcnt;
2016
2017         ipath_cdbg(VERBOSE, "Port %u unlocking any locked expTID pages\n",
2018                    pd->port_port);
2019         for (i = port_tidbase; i < maxtid; i++) {
2020                 struct page *ps = dd->ipath_pageshadow[i];
2021
2022                 if (!ps)
2023                         continue;
2024
2025                 dd->ipath_pageshadow[i] = NULL;
2026                 pci_unmap_page(dd->pcidev, dd->ipath_physshadow[i],
2027                         PAGE_SIZE, PCI_DMA_FROMDEVICE);
2028                 ipath_release_user_pages_on_close(&ps, 1);
2029                 cnt++;
2030                 ipath_stats.sps_pageunlocks++;
2031         }
2032         if (cnt)
2033                 ipath_cdbg(VERBOSE, "Port %u locked %u expTID entries\n",
2034                            pd->port_port, cnt);
2035
2036         if (ipath_stats.sps_pagelocks || ipath_stats.sps_pageunlocks)
2037                 ipath_cdbg(VERBOSE, "%llu pages locked, %llu unlocked\n",
2038                            (unsigned long long) ipath_stats.sps_pagelocks,
2039                            (unsigned long long)
2040                            ipath_stats.sps_pageunlocks);
2041 }
2042
2043 static int ipath_close(struct inode *in, struct file *fp)
2044 {
2045         int ret = 0;
2046         struct ipath_filedata *fd;
2047         struct ipath_portdata *pd;
2048         struct ipath_devdata *dd;
2049         unsigned long flags;
2050         unsigned port;
2051         struct pid *pid;
2052
2053         ipath_cdbg(VERBOSE, "close on dev %lx, private data %p\n",
2054                    (long)in->i_rdev, fp->private_data);
2055
2056         mutex_lock(&ipath_mutex);
2057
2058         fd = (struct ipath_filedata *) fp->private_data;
2059         fp->private_data = NULL;
2060         pd = fd->pd;
2061         if (!pd) {
2062                 mutex_unlock(&ipath_mutex);
2063                 goto bail;
2064         }
2065
2066         dd = pd->port_dd;
2067
2068         /* drain user sdma queue */
2069         ipath_user_sdma_queue_drain(dd, fd->pq);
2070         ipath_user_sdma_queue_destroy(fd->pq);
2071
2072         if (--pd->port_cnt) {
2073                 /*
2074                  * XXX If the master closes the port before the slave(s),
2075                  * revoke the mmap for the eager receive queue so
2076                  * the slave(s) don't wait for receive data forever.
2077                  */
2078                 pd->active_slaves &= ~(1 << fd->subport);
2079                 put_pid(pd->port_subpid[fd->subport]);
2080                 pd->port_subpid[fd->subport] = NULL;
2081                 mutex_unlock(&ipath_mutex);
2082                 goto bail;
2083         }
2084         /* early; no interrupt users after this */
2085         spin_lock_irqsave(&dd->ipath_uctxt_lock, flags);
2086         port = pd->port_port;
2087         dd->ipath_pd[port] = NULL;
2088         pid = pd->port_pid;
2089         pd->port_pid = NULL;
2090         spin_unlock_irqrestore(&dd->ipath_uctxt_lock, flags);
2091
2092         if (pd->port_rcvwait_to || pd->port_piowait_to
2093             || pd->port_rcvnowait || pd->port_pionowait) {
2094                 ipath_cdbg(VERBOSE, "port%u, %u rcv, %u pio wait timeo; "
2095                            "%u rcv %u, pio already\n",
2096                            pd->port_port, pd->port_rcvwait_to,
2097                            pd->port_piowait_to, pd->port_rcvnowait,
2098                            pd->port_pionowait);
2099                 pd->port_rcvwait_to = pd->port_piowait_to =
2100                         pd->port_rcvnowait = pd->port_pionowait = 0;
2101         }
2102         if (pd->port_flag) {
2103                 ipath_cdbg(PROC, "port %u port_flag set: 0x%lx\n",
2104                           pd->port_port, pd->port_flag);
2105                 pd->port_flag = 0;
2106         }
2107
2108         if (dd->ipath_kregbase) {
2109                 /* atomically clear receive enable port and intr avail. */
2110                 clear_bit(dd->ipath_r_portenable_shift + port,
2111                           &dd->ipath_rcvctrl);
2112                 clear_bit(pd->port_port + dd->ipath_r_intravail_shift,
2113                           &dd->ipath_rcvctrl);
2114                 ipath_write_kreg( dd, dd->ipath_kregs->kr_rcvctrl,
2115                         dd->ipath_rcvctrl);
2116                 /* and read back from chip to be sure that nothing
2117                  * else is in flight when we do the rest */
2118                 (void)ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
2119
2120                 /* clean up the pkeys for this port user */
2121                 ipath_clean_part_key(pd, dd);
2122                 /*
2123                  * be paranoid, and never write 0's to these, just use an
2124                  * unused part of the port 0 tail page.  Of course,
2125                  * rcvhdraddr points to a large chunk of memory, so this
2126                  * could still trash things, but at least it won't trash
2127                  * page 0, and by disabling the port, it should stop "soon",
2128                  * even if a packet or two is in already in flight after we
2129                  * disabled the port.
2130                  */
2131                 ipath_write_kreg_port(dd,
2132                         dd->ipath_kregs->kr_rcvhdrtailaddr, port,
2133                         dd->ipath_dummy_hdrq_phys);
2134                 ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdraddr,
2135                         pd->port_port, dd->ipath_dummy_hdrq_phys);
2136
2137                 ipath_disarm_piobufs(dd, pd->port_pio_base, pd->port_piocnt);
2138                 ipath_chg_pioavailkernel(dd, pd->port_pio_base,
2139                         pd->port_piocnt, 1);
2140
2141                 dd->ipath_f_clear_tids(dd, pd->port_port);
2142
2143                 if (dd->ipath_pageshadow)
2144                         unlock_expected_tids(pd);
2145                 ipath_stats.sps_ports--;
2146                 ipath_cdbg(PROC, "%s[%u] closed port %u:%u\n",
2147                            pd->port_comm, pid_nr(pid),
2148                            dd->ipath_unit, port);
2149         }
2150
2151         put_pid(pid);
2152         mutex_unlock(&ipath_mutex);
2153         ipath_free_pddata(dd, pd); /* after releasing the mutex */
2154
2155 bail:
2156         kfree(fd);
2157         return ret;
2158 }
2159
2160 static int ipath_port_info(struct ipath_portdata *pd, u16 subport,
2161                            struct ipath_port_info __user *uinfo)
2162 {
2163         struct ipath_port_info info;
2164         int nup;
2165         int ret;
2166         size_t sz;
2167
2168         (void) ipath_count_units(NULL, &nup, NULL);
2169         info.num_active = nup;
2170         info.unit = pd->port_dd->ipath_unit;
2171         info.port = pd->port_port;
2172         info.subport = subport;
2173         /* Don't return new fields if old library opened the port. */
2174         if (ipath_supports_subports(pd->userversion >> 16,
2175                                     pd->userversion & 0xffff)) {
2176                 /* Number of user ports available for this device. */
2177                 info.num_ports = pd->port_dd->ipath_cfgports - 1;
2178                 info.num_subports = pd->port_subport_cnt;
2179                 sz = sizeof(info);
2180         } else
2181                 sz = sizeof(info) - 2 * sizeof(u16);
2182
2183         if (copy_to_user(uinfo, &info, sz)) {
2184                 ret = -EFAULT;
2185                 goto bail;
2186         }
2187         ret = 0;
2188
2189 bail:
2190         return ret;
2191 }
2192
2193 static int ipath_get_slave_info(struct ipath_portdata *pd,
2194                                 void __user *slave_mask_addr)
2195 {
2196         int ret = 0;
2197
2198         if (copy_to_user(slave_mask_addr, &pd->active_slaves, sizeof(u32)))
2199                 ret = -EFAULT;
2200         return ret;
2201 }
2202
2203 static int ipath_sdma_get_inflight(struct ipath_user_sdma_queue *pq,
2204                                    u32 __user *inflightp)
2205 {
2206         const u32 val = ipath_user_sdma_inflight_counter(pq);
2207
2208         if (put_user(val, inflightp))
2209                 return -EFAULT;
2210
2211         return 0;
2212 }
2213
2214 static int ipath_sdma_get_complete(struct ipath_devdata *dd,
2215                                    struct ipath_user_sdma_queue *pq,
2216                                    u32 __user *completep)
2217 {
2218         u32 val;
2219         int err;
2220
2221         err = ipath_user_sdma_make_progress(dd, pq);
2222         if (err < 0)
2223                 return err;
2224
2225         val = ipath_user_sdma_complete_counter(pq);
2226         if (put_user(val, completep))
2227                 return -EFAULT;
2228
2229         return 0;
2230 }
2231
2232 static ssize_t ipath_write(struct file *fp, const char __user *data,
2233                            size_t count, loff_t *off)
2234 {
2235         const struct ipath_cmd __user *ucmd;
2236         struct ipath_portdata *pd;
2237         const void __user *src;
2238         size_t consumed, copy;
2239         struct ipath_cmd cmd;
2240         ssize_t ret = 0;
2241         void *dest;
2242
2243         if (count < sizeof(cmd.type)) {
2244                 ret = -EINVAL;
2245                 goto bail;
2246         }
2247
2248         ucmd = (const struct ipath_cmd __user *) data;
2249
2250         if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) {
2251                 ret = -EFAULT;
2252                 goto bail;
2253         }
2254
2255         consumed = sizeof(cmd.type);
2256
2257         switch (cmd.type) {
2258         case IPATH_CMD_ASSIGN_PORT:
2259         case __IPATH_CMD_USER_INIT:
2260         case IPATH_CMD_USER_INIT:
2261                 copy = sizeof(cmd.cmd.user_info);
2262                 dest = &cmd.cmd.user_info;
2263                 src = &ucmd->cmd.user_info;
2264                 break;
2265         case IPATH_CMD_RECV_CTRL:
2266                 copy = sizeof(cmd.cmd.recv_ctrl);
2267                 dest = &cmd.cmd.recv_ctrl;
2268                 src = &ucmd->cmd.recv_ctrl;
2269                 break;
2270         case IPATH_CMD_PORT_INFO:
2271                 copy = sizeof(cmd.cmd.port_info);
2272                 dest = &cmd.cmd.port_info;
2273                 src = &ucmd->cmd.port_info;
2274                 break;
2275         case IPATH_CMD_TID_UPDATE:
2276         case IPATH_CMD_TID_FREE:
2277                 copy = sizeof(cmd.cmd.tid_info);
2278                 dest = &cmd.cmd.tid_info;
2279                 src = &ucmd->cmd.tid_info;
2280                 break;
2281         case IPATH_CMD_SET_PART_KEY:
2282                 copy = sizeof(cmd.cmd.part_key);
2283                 dest = &cmd.cmd.part_key;
2284                 src = &ucmd->cmd.part_key;
2285                 break;
2286         case __IPATH_CMD_SLAVE_INFO:
2287                 copy = sizeof(cmd.cmd.slave_mask_addr);
2288                 dest = &cmd.cmd.slave_mask_addr;
2289                 src = &ucmd->cmd.slave_mask_addr;
2290                 break;
2291         case IPATH_CMD_PIOAVAILUPD:     // force an update of PIOAvail reg
2292                 copy = 0;
2293                 src = NULL;
2294                 dest = NULL;
2295                 break;
2296         case IPATH_CMD_POLL_TYPE:
2297                 copy = sizeof(cmd.cmd.poll_type);
2298                 dest = &cmd.cmd.poll_type;
2299                 src = &ucmd->cmd.poll_type;
2300                 break;
2301         case IPATH_CMD_ARMLAUNCH_CTRL:
2302                 copy = sizeof(cmd.cmd.armlaunch_ctrl);
2303                 dest = &cmd.cmd.armlaunch_ctrl;
2304                 src = &ucmd->cmd.armlaunch_ctrl;
2305                 break;
2306         case IPATH_CMD_SDMA_INFLIGHT:
2307                 copy = sizeof(cmd.cmd.sdma_inflight);
2308                 dest = &cmd.cmd.sdma_inflight;
2309                 src = &ucmd->cmd.sdma_inflight;
2310                 break;
2311         case IPATH_CMD_SDMA_COMPLETE:
2312                 copy = sizeof(cmd.cmd.sdma_complete);
2313                 dest = &cmd.cmd.sdma_complete;
2314                 src = &ucmd->cmd.sdma_complete;
2315                 break;
2316         default:
2317                 ret = -EINVAL;
2318                 goto bail;
2319         }
2320
2321         if (copy) {
2322                 if ((count - consumed) < copy) {
2323                         ret = -EINVAL;
2324                         goto bail;
2325                 }
2326
2327                 if (copy_from_user(dest, src, copy)) {
2328                         ret = -EFAULT;
2329                         goto bail;
2330                 }
2331
2332                 consumed += copy;
2333         }
2334
2335         pd = port_fp(fp);
2336         if (!pd && cmd.type != __IPATH_CMD_USER_INIT &&
2337                 cmd.type != IPATH_CMD_ASSIGN_PORT) {
2338                 ret = -EINVAL;
2339                 goto bail;
2340         }
2341
2342         switch (cmd.type) {
2343         case IPATH_CMD_ASSIGN_PORT:
2344                 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2345                 if (ret)
2346                         goto bail;
2347                 break;
2348         case __IPATH_CMD_USER_INIT:
2349                 /* backwards compatibility, get port first */
2350                 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2351                 if (ret)
2352                         goto bail;
2353                 /* and fall through to current version. */
2354         case IPATH_CMD_USER_INIT:
2355                 ret = ipath_do_user_init(fp, &cmd.cmd.user_info);
2356                 if (ret)
2357                         goto bail;
2358                 ret = ipath_get_base_info(
2359                         fp, (void __user *) (unsigned long)
2360                         cmd.cmd.user_info.spu_base_info,
2361                         cmd.cmd.user_info.spu_base_info_size);
2362                 break;
2363         case IPATH_CMD_RECV_CTRL:
2364                 ret = ipath_manage_rcvq(pd, subport_fp(fp), cmd.cmd.recv_ctrl);
2365                 break;
2366         case IPATH_CMD_PORT_INFO:
2367                 ret = ipath_port_info(pd, subport_fp(fp),
2368                                       (struct ipath_port_info __user *)
2369                                       (unsigned long) cmd.cmd.port_info);
2370                 break;
2371         case IPATH_CMD_TID_UPDATE:
2372                 ret = ipath_tid_update(pd, fp, &cmd.cmd.tid_info);
2373                 break;
2374         case IPATH_CMD_TID_FREE:
2375                 ret = ipath_tid_free(pd, subport_fp(fp), &cmd.cmd.tid_info);
2376                 break;
2377         case IPATH_CMD_SET_PART_KEY:
2378                 ret = ipath_set_part_key(pd, cmd.cmd.part_key);
2379                 break;
2380         case __IPATH_CMD_SLAVE_INFO:
2381                 ret = ipath_get_slave_info(pd,
2382                                            (void __user *) (unsigned long)
2383                                            cmd.cmd.slave_mask_addr);
2384                 break;
2385         case IPATH_CMD_PIOAVAILUPD:
2386                 ipath_force_pio_avail_update(pd->port_dd);
2387                 break;
2388         case IPATH_CMD_POLL_TYPE:
2389                 pd->poll_type = cmd.cmd.poll_type;
2390                 break;
2391         case IPATH_CMD_ARMLAUNCH_CTRL:
2392                 if (cmd.cmd.armlaunch_ctrl)
2393                         ipath_enable_armlaunch(pd->port_dd);
2394                 else
2395                         ipath_disable_armlaunch(pd->port_dd);
2396                 break;
2397         case IPATH_CMD_SDMA_INFLIGHT:
2398                 ret = ipath_sdma_get_inflight(user_sdma_queue_fp(fp),
2399                                               (u32 __user *) (unsigned long)
2400                                               cmd.cmd.sdma_inflight);
2401                 break;
2402         case IPATH_CMD_SDMA_COMPLETE:
2403                 ret = ipath_sdma_get_complete(pd->port_dd,
2404                                               user_sdma_queue_fp(fp),
2405                                               (u32 __user *) (unsigned long)
2406                                               cmd.cmd.sdma_complete);
2407                 break;
2408         }
2409
2410         if (ret >= 0)
2411                 ret = consumed;
2412
2413 bail:
2414         return ret;
2415 }
2416
2417 static ssize_t ipath_writev(struct kiocb *iocb, const struct iovec *iov,
2418                             unsigned long dim, loff_t off)
2419 {
2420         struct file *filp = iocb->ki_filp;
2421         struct ipath_filedata *fp = filp->private_data;
2422         struct ipath_portdata *pd = port_fp(filp);
2423         struct ipath_user_sdma_queue *pq = fp->pq;
2424
2425         if (!dim)
2426                 return -EINVAL;
2427
2428         return ipath_user_sdma_writev(pd->port_dd, pq, iov, dim);
2429 }
2430
2431 static struct class *ipath_class;
2432
2433 static int init_cdev(int minor, char *name, const struct file_operations *fops,
2434                      struct cdev **cdevp, struct device **devp)
2435 {
2436         const dev_t dev = MKDEV(IPATH_MAJOR, minor);
2437         struct cdev *cdev = NULL;
2438         struct device *device = NULL;
2439         int ret;
2440
2441         cdev = cdev_alloc();
2442         if (!cdev) {
2443                 printk(KERN_ERR IPATH_DRV_NAME
2444                        ": Could not allocate cdev for minor %d, %s\n",
2445                        minor, name);
2446                 ret = -ENOMEM;
2447                 goto done;
2448         }
2449
2450         cdev->owner = THIS_MODULE;
2451         cdev->ops = fops;
2452         kobject_set_name(&cdev->kobj, name);
2453
2454         ret = cdev_add(cdev, dev, 1);
2455         if (ret < 0) {
2456                 printk(KERN_ERR IPATH_DRV_NAME
2457                        ": Could not add cdev for minor %d, %s (err %d)\n",
2458                        minor, name, -ret);
2459                 goto err_cdev;
2460         }
2461
2462         device = device_create(ipath_class, NULL, dev, NULL, name);
2463
2464         if (IS_ERR(device)) {
2465                 ret = PTR_ERR(device);
2466                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2467                        "device for minor %d, %s (err %d)\n",
2468                        minor, name, -ret);
2469                 goto err_cdev;
2470         }
2471
2472         goto done;
2473
2474 err_cdev:
2475         cdev_del(cdev);
2476         cdev = NULL;
2477
2478 done:
2479         if (ret >= 0) {
2480                 *cdevp = cdev;
2481                 *devp = device;
2482         } else {
2483                 *cdevp = NULL;
2484                 *devp = NULL;
2485         }
2486
2487         return ret;
2488 }
2489
2490 int ipath_cdev_init(int minor, char *name, const struct file_operations *fops,
2491                     struct cdev **cdevp, struct device **devp)
2492 {
2493         return init_cdev(minor, name, fops, cdevp, devp);
2494 }
2495
2496 static void cleanup_cdev(struct cdev **cdevp,
2497                          struct device **devp)
2498 {
2499         struct device *dev = *devp;
2500
2501         if (dev) {
2502                 device_unregister(dev);
2503                 *devp = NULL;
2504         }
2505
2506         if (*cdevp) {
2507                 cdev_del(*cdevp);
2508                 *cdevp = NULL;
2509         }
2510 }
2511
2512 void ipath_cdev_cleanup(struct cdev **cdevp,
2513                         struct device **devp)
2514 {
2515         cleanup_cdev(cdevp, devp);
2516 }
2517
2518 static struct cdev *wildcard_cdev;
2519 static struct device *wildcard_dev;
2520
2521 static const dev_t dev = MKDEV(IPATH_MAJOR, 0);
2522
2523 static int user_init(void)
2524 {
2525         int ret;
2526
2527         ret = register_chrdev_region(dev, IPATH_NMINORS, IPATH_DRV_NAME);
2528         if (ret < 0) {
2529                 printk(KERN_ERR IPATH_DRV_NAME ": Could not register "
2530                        "chrdev region (err %d)\n", -ret);
2531                 goto done;
2532         }
2533
2534         ipath_class = class_create(THIS_MODULE, IPATH_DRV_NAME);
2535
2536         if (IS_ERR(ipath_class)) {
2537                 ret = PTR_ERR(ipath_class);
2538                 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2539                        "device class (err %d)\n", -ret);
2540                 goto bail;
2541         }
2542
2543         goto done;
2544 bail:
2545         unregister_chrdev_region(dev, IPATH_NMINORS);
2546 done:
2547         return ret;
2548 }
2549
2550 static void user_cleanup(void)
2551 {
2552         if (ipath_class) {
2553                 class_destroy(ipath_class);
2554                 ipath_class = NULL;
2555         }
2556
2557         unregister_chrdev_region(dev, IPATH_NMINORS);
2558 }
2559
2560 static atomic_t user_count = ATOMIC_INIT(0);
2561 static atomic_t user_setup = ATOMIC_INIT(0);
2562
2563 int ipath_user_add(struct ipath_devdata *dd)
2564 {
2565         char name[10];
2566         int ret;
2567
2568         if (atomic_inc_return(&user_count) == 1) {
2569                 ret = user_init();
2570                 if (ret < 0) {
2571                         ipath_dev_err(dd, "Unable to set up user support: "
2572                                       "error %d\n", -ret);
2573                         goto bail;
2574                 }
2575                 ret = init_cdev(0, "ipath", &ipath_file_ops, &wildcard_cdev,
2576                                 &wildcard_dev);
2577                 if (ret < 0) {
2578                         ipath_dev_err(dd, "Could not create wildcard "
2579                                       "minor: error %d\n", -ret);
2580                         goto bail_user;
2581                 }
2582
2583                 atomic_set(&user_setup, 1);
2584         }
2585
2586         snprintf(name, sizeof(name), "ipath%d", dd->ipath_unit);
2587
2588         ret = init_cdev(dd->ipath_unit + 1, name, &ipath_file_ops,
2589                         &dd->user_cdev, &dd->user_dev);
2590         if (ret < 0)
2591                 ipath_dev_err(dd, "Could not create user minor %d, %s\n",
2592                               dd->ipath_unit + 1, name);
2593
2594         goto bail;
2595
2596 bail_user:
2597         user_cleanup();
2598 bail:
2599         return ret;
2600 }
2601
2602 void ipath_user_remove(struct ipath_devdata *dd)
2603 {
2604         cleanup_cdev(&dd->user_cdev, &dd->user_dev);
2605
2606         if (atomic_dec_return(&user_count) == 0) {
2607                 if (atomic_read(&user_setup) == 0)
2608                         goto bail;
2609
2610                 cleanup_cdev(&wildcard_cdev, &wildcard_dev);
2611                 user_cleanup();
2612
2613                 atomic_set(&user_setup, 0);
2614         }
2615 bail:
2616         return;
2617 }