2 * Copyright (c) 2006, 2007 QLogic Corporation. All rights reserved.
3 * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
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:
11 * Redistribution and use in source and binary forms, with or
12 * without modification, are permitted provided that the following
15 * - Redistributions of source code must retain the above
16 * copyright notice, this list of conditions and the following
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.
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
34 #include <linux/pci.h>
35 #include <linux/poll.h>
36 #include <linux/cdev.h>
37 #include <linux/swap.h>
38 #include <linux/vmalloc.h>
39 #include <asm/pgtable.h>
41 #include "ipath_kernel.h"
42 #include "ipath_common.h"
44 static int ipath_open(struct inode *, struct file *);
45 static int ipath_close(struct inode *, struct file *);
46 static ssize_t ipath_write(struct file *, const char __user *, size_t,
48 static unsigned int ipath_poll(struct file *, struct poll_table_struct *);
49 static int ipath_mmap(struct file *, struct vm_area_struct *);
51 static const struct file_operations ipath_file_ops = {
55 .release = ipath_close,
61 * Convert kernel virtual addresses to physical addresses so they don't
62 * potentially conflict with the chip addresses used as mmap offsets.
63 * It doesn't really matter what mmap offset we use as long as we can
64 * interpret it correctly.
66 static u64 cvt_kvaddr(void *p)
71 page = vmalloc_to_page(p);
73 paddr = page_to_pfn(page) << PAGE_SHIFT;
78 static int ipath_get_base_info(struct file *fp,
79 void __user *ubase, size_t ubase_size)
81 struct ipath_portdata *pd = port_fp(fp);
83 struct ipath_base_info *kinfo = NULL;
84 struct ipath_devdata *dd = pd->port_dd;
89 subport_cnt = pd->port_subport_cnt;
96 master = !subport_fp(fp);
100 /* If port sharing is not requested, allow the old size structure */
102 sz -= 7 * sizeof(u64);
103 if (ubase_size < sz) {
105 "Base size %zu, need %zu (version mismatch?)\n",
111 kinfo = kzalloc(sizeof(*kinfo), GFP_KERNEL);
117 ret = dd->ipath_f_get_base_info(pd, kinfo);
121 kinfo->spi_rcvhdr_cnt = dd->ipath_rcvhdrcnt;
122 kinfo->spi_rcvhdrent_size = dd->ipath_rcvhdrentsize;
123 kinfo->spi_tidegrcnt = dd->ipath_rcvegrcnt;
124 kinfo->spi_rcv_egrbufsize = dd->ipath_rcvegrbufsize;
126 * have to mmap whole thing
128 kinfo->spi_rcv_egrbuftotlen =
129 pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
130 kinfo->spi_rcv_egrperchunk = pd->port_rcvegrbufs_perchunk;
131 kinfo->spi_rcv_egrchunksize = kinfo->spi_rcv_egrbuftotlen /
132 pd->port_rcvegrbuf_chunks;
133 kinfo->spi_tidcnt = dd->ipath_rcvtidcnt / subport_cnt;
135 kinfo->spi_tidcnt += dd->ipath_rcvtidcnt % subport_cnt;
137 * for this use, may be ipath_cfgports summed over all chips that
138 * are are configured and present
140 kinfo->spi_nports = dd->ipath_cfgports;
141 /* unit (chip/board) our port is on */
142 kinfo->spi_unit = dd->ipath_unit;
143 /* for now, only a single page */
144 kinfo->spi_tid_maxsize = PAGE_SIZE;
147 * Doing this per port, and based on the skip value, etc. This has
148 * to be the actual buffer size, since the protocol code treats it
151 * These have to be set to user addresses in the user code via mmap.
152 * These values are used on return to user code for the mmap target
153 * addresses only. For 32 bit, same 44 bit address problem, so use
154 * the physical address, not virtual. Before 2.6.11, using the
155 * page_address() macro worked, but in 2.6.11, even that returns the
156 * full 64 bit address (upper bits all 1's). So far, using the
157 * physical addresses (or chip offsets, for chip mapping) works, but
158 * no doubt some future kernel release will change that, and we'll be
159 * on to yet another method of dealing with this.
161 kinfo->spi_rcvhdr_base = (u64) pd->port_rcvhdrq_phys;
162 kinfo->spi_rcvhdr_tailaddr = (u64) pd->port_rcvhdrqtailaddr_phys;
163 kinfo->spi_rcv_egrbufs = (u64) pd->port_rcvegr_phys;
164 kinfo->spi_pioavailaddr = (u64) dd->ipath_pioavailregs_phys;
165 kinfo->spi_status = (u64) kinfo->spi_pioavailaddr +
166 (void *) dd->ipath_statusp -
167 (void *) dd->ipath_pioavailregs_dma;
169 kinfo->spi_piocnt = dd->ipath_pbufsport;
170 kinfo->spi_piobufbase = (u64) pd->port_piobufs;
171 kinfo->__spi_uregbase = (u64) dd->ipath_uregbase +
172 dd->ipath_palign * pd->port_port;
174 kinfo->spi_piocnt = (dd->ipath_pbufsport / subport_cnt) +
175 (dd->ipath_pbufsport % subport_cnt);
176 /* Master's PIO buffers are after all the slave's */
177 kinfo->spi_piobufbase = (u64) pd->port_piobufs +
179 (dd->ipath_pbufsport - kinfo->spi_piocnt);
181 unsigned slave = subport_fp(fp) - 1;
183 kinfo->spi_piocnt = dd->ipath_pbufsport / subport_cnt;
184 kinfo->spi_piobufbase = (u64) pd->port_piobufs +
185 dd->ipath_palign * kinfo->spi_piocnt * slave;
188 kinfo->spi_port_uregbase = (u64) dd->ipath_uregbase +
189 dd->ipath_palign * pd->port_port;
190 kinfo->spi_port_rcvegrbuf = kinfo->spi_rcv_egrbufs;
191 kinfo->spi_port_rcvhdr_base = kinfo->spi_rcvhdr_base;
192 kinfo->spi_port_rcvhdr_tailaddr = kinfo->spi_rcvhdr_tailaddr;
194 kinfo->__spi_uregbase = cvt_kvaddr(pd->subport_uregbase +
195 PAGE_SIZE * subport_fp(fp));
197 kinfo->spi_rcvhdr_base = cvt_kvaddr(pd->subport_rcvhdr_base +
198 pd->port_rcvhdrq_size * subport_fp(fp));
199 kinfo->spi_rcvhdr_tailaddr = 0;
200 kinfo->spi_rcv_egrbufs = cvt_kvaddr(pd->subport_rcvegrbuf +
201 pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size *
204 kinfo->spi_subport_uregbase =
205 cvt_kvaddr(pd->subport_uregbase);
206 kinfo->spi_subport_rcvegrbuf =
207 cvt_kvaddr(pd->subport_rcvegrbuf);
208 kinfo->spi_subport_rcvhdr_base =
209 cvt_kvaddr(pd->subport_rcvhdr_base);
210 ipath_cdbg(PROC, "port %u flags %x %llx %llx %llx\n",
211 kinfo->spi_port, kinfo->spi_runtime_flags,
212 (unsigned long long) kinfo->spi_subport_uregbase,
213 (unsigned long long) kinfo->spi_subport_rcvegrbuf,
214 (unsigned long long) kinfo->spi_subport_rcvhdr_base);
217 kinfo->spi_pioindex = (kinfo->spi_piobufbase - dd->ipath_piobufbase) /
219 kinfo->spi_pioalign = dd->ipath_palign;
221 kinfo->spi_qpair = IPATH_KD_QP;
222 kinfo->spi_piosize = dd->ipath_ibmaxlen;
223 kinfo->spi_mtu = dd->ipath_ibmaxlen; /* maxlen, not ibmtu */
224 kinfo->spi_port = pd->port_port;
225 kinfo->spi_subport = subport_fp(fp);
226 kinfo->spi_sw_version = IPATH_KERN_SWVERSION;
227 kinfo->spi_hw_version = dd->ipath_revision;
230 kinfo->spi_runtime_flags |= IPATH_RUNTIME_MASTER;
233 sz = (ubase_size < sizeof(*kinfo)) ? ubase_size : sizeof(*kinfo);
234 if (copy_to_user(ubase, kinfo, sz))
243 * ipath_tid_update - update a port TID
245 * @fp: the ipath device file
246 * @ti: the TID information
248 * The new implementation as of Oct 2004 is that the driver assigns
249 * the tid and returns it to the caller. To make it easier to
250 * catch bugs, and to reduce search time, we keep a cursor for
251 * each port, walking the shadow tid array to find one that's not
254 * For now, if we can't allocate the full list, we fail, although
255 * in the long run, we'll allocate as many as we can, and the
256 * caller will deal with that by trying the remaining pages later.
257 * That means that when we fail, we have to mark the tids as not in
258 * use again, in our shadow copy.
260 * It's up to the caller to free the tids when they are done.
261 * We'll unlock the pages as they free them.
263 * Also, right now we are locking one page at a time, but since
264 * the intended use of this routine is for a single group of
265 * virtually contiguous pages, that should change to improve
268 static int ipath_tid_update(struct ipath_portdata *pd, struct file *fp,
269 const struct ipath_tid_info *ti)
272 u32 tid, porttid, cnt, i, tidcnt, tidoff;
274 struct ipath_devdata *dd = pd->port_dd;
277 u64 __iomem *tidbase;
278 unsigned long tidmap[8];
279 struct page **pagep = NULL;
280 unsigned subport = subport_fp(fp);
282 if (!dd->ipath_pageshadow) {
289 ipath_dbg("After copyin, tidcnt 0, tidlist %llx\n",
290 (unsigned long long) ti->tidlist);
292 * Should we treat as success? likely a bug
297 porttid = pd->port_port * dd->ipath_rcvtidcnt;
298 if (!pd->port_subport_cnt) {
299 tidcnt = dd->ipath_rcvtidcnt;
300 tid = pd->port_tidcursor;
302 } else if (!subport) {
303 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
304 (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
305 tidoff = dd->ipath_rcvtidcnt - tidcnt;
307 tid = tidcursor_fp(fp);
309 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
310 tidoff = tidcnt * (subport - 1);
312 tid = tidcursor_fp(fp);
315 /* make sure it all fits in port_tid_pg_list */
316 dev_info(&dd->pcidev->dev, "Process tried to allocate %u "
317 "TIDs, only trying max (%u)\n", cnt, tidcnt);
320 pagep = &((struct page **) pd->port_tid_pg_list)[tidoff];
321 tidlist = &((u16 *) &pagep[dd->ipath_rcvtidcnt])[tidoff];
323 memset(tidmap, 0, sizeof(tidmap));
324 /* before decrement; chip actual # */
326 tidbase = (u64 __iomem *) (((char __iomem *) dd->ipath_kregbase) +
327 dd->ipath_rcvtidbase +
328 porttid * sizeof(*tidbase));
330 ipath_cdbg(VERBOSE, "Port%u %u tids, cursor %u, tidbase %p\n",
331 pd->port_port, cnt, tid, tidbase);
333 /* virtual address of first page in transfer */
334 vaddr = ti->tidvaddr;
335 if (!access_ok(VERIFY_WRITE, (void __user *) vaddr,
337 ipath_dbg("Fail vaddr %p, %u pages, !access_ok\n",
342 ret = ipath_get_user_pages(vaddr, cnt, pagep);
345 ipath_dbg("Failed to lock addr %p, %u pages "
346 "(already locked)\n",
347 (void *) vaddr, cnt);
349 * for now, continue, and see what happens but with
350 * the new implementation, this should never happen,
351 * unless perhaps the user has mpin'ed the pages
352 * themselves (something we need to test)
356 dev_info(&dd->pcidev->dev,
357 "Failed to lock addr %p, %u pages: "
358 "errno %d\n", (void *) vaddr, cnt, -ret);
362 for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
363 for (; ntids--; tid++) {
366 if (!dd->ipath_pageshadow[porttid + tid])
371 * oops, wrapped all the way through their TIDs,
372 * and didn't have enough free; see comments at
375 ipath_dbg("Not enough free TIDs for %u pages "
376 "(index %d), failing\n", cnt, i);
377 i--; /* last tidlist[i] not filled in */
381 tidlist[i] = tid + tidoff;
382 ipath_cdbg(VERBOSE, "Updating idx %u to TID %u, "
383 "vaddr %lx\n", i, tid + tidoff, vaddr);
384 /* we "know" system pages and TID pages are same size */
385 dd->ipath_pageshadow[porttid + tid] = pagep[i];
386 dd->ipath_physshadow[porttid + tid] = ipath_map_page(
387 dd->pcidev, pagep[i], 0, PAGE_SIZE,
390 * don't need atomic or it's overhead
392 __set_bit(tid, tidmap);
393 physaddr = dd->ipath_physshadow[porttid + tid];
394 ipath_stats.sps_pagelocks++;
396 "TID %u, vaddr %lx, physaddr %llx pgp %p\n",
397 tid, vaddr, (unsigned long long) physaddr,
399 dd->ipath_f_put_tid(dd, &tidbase[tid], RCVHQ_RCV_TYPE_EXPECTED,
402 * don't check this tid in ipath_portshadow, since we
403 * just filled it in; start with the next one.
411 /* jump here if copy out of updated info failed... */
412 ipath_dbg("After failure (ret=%d), undo %d of %d entries\n",
414 /* same code that's in ipath_free_tid() */
415 limit = sizeof(tidmap) * BITS_PER_BYTE;
417 /* just in case size changes in future */
419 tid = find_first_bit((const unsigned long *)tidmap, limit);
420 for (; tid < limit; tid++) {
421 if (!test_bit(tid, tidmap))
423 if (dd->ipath_pageshadow[porttid + tid]) {
424 ipath_cdbg(VERBOSE, "Freeing TID %u\n",
426 dd->ipath_f_put_tid(dd, &tidbase[tid],
427 RCVHQ_RCV_TYPE_EXPECTED,
428 dd->ipath_tidinvalid);
429 pci_unmap_page(dd->pcidev,
430 dd->ipath_physshadow[porttid + tid],
431 PAGE_SIZE, PCI_DMA_FROMDEVICE);
432 dd->ipath_pageshadow[porttid + tid] = NULL;
433 ipath_stats.sps_pageunlocks++;
436 ipath_release_user_pages(pagep, cnt);
439 * Copy the updated array, with ipath_tid's filled in, back
440 * to user. Since we did the copy in already, this "should
441 * never fail" If it does, we have to clean up...
443 if (copy_to_user((void __user *)
444 (unsigned long) ti->tidlist,
445 tidlist, cnt * sizeof(*tidlist))) {
449 if (copy_to_user((void __user *) (unsigned long) ti->tidmap,
450 tidmap, sizeof tidmap)) {
456 if (!pd->port_subport_cnt)
457 pd->port_tidcursor = tid;
459 tidcursor_fp(fp) = tid;
464 ipath_dbg("Failed to map %u TID pages, failing with %d\n",
470 * ipath_tid_free - free a port TID
472 * @subport: the subport
475 * right now we are unlocking one page at a time, but since
476 * the intended use of this routine is for a single group of
477 * virtually contiguous pages, that should change to improve
478 * performance. We check that the TID is in range for this port
479 * but otherwise don't check validity; if user has an error and
480 * frees the wrong tid, it's only their own data that can thereby
481 * be corrupted. We do check that the TID was in use, for sanity
482 * We always use our idea of the saved address, not the address that
483 * they pass in to us.
486 static int ipath_tid_free(struct ipath_portdata *pd, unsigned subport,
487 const struct ipath_tid_info *ti)
490 u32 tid, porttid, cnt, limit, tidcnt;
491 struct ipath_devdata *dd = pd->port_dd;
492 u64 __iomem *tidbase;
493 unsigned long tidmap[8];
495 if (!dd->ipath_pageshadow) {
500 if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap,
506 porttid = pd->port_port * dd->ipath_rcvtidcnt;
507 if (!pd->port_subport_cnt)
508 tidcnt = dd->ipath_rcvtidcnt;
510 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
511 (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
512 porttid += dd->ipath_rcvtidcnt - tidcnt;
514 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
515 porttid += tidcnt * (subport - 1);
517 tidbase = (u64 __iomem *) ((char __iomem *)(dd->ipath_kregbase) +
518 dd->ipath_rcvtidbase +
519 porttid * sizeof(*tidbase));
521 limit = sizeof(tidmap) * BITS_PER_BYTE;
523 /* just in case size changes in future */
525 tid = find_first_bit(tidmap, limit);
526 ipath_cdbg(VERBOSE, "Port%u free %u tids; first bit (max=%d) "
527 "set is %d, porttid %u\n", pd->port_port, ti->tidcnt,
528 limit, tid, porttid);
529 for (cnt = 0; tid < limit; tid++) {
531 * small optimization; if we detect a run of 3 or so without
532 * any set, use find_first_bit again. That's mainly to
533 * accelerate the case where we wrapped, so we have some at
534 * the beginning, and some at the end, and a big gap
537 if (!test_bit(tid, tidmap))
540 if (dd->ipath_pageshadow[porttid + tid]) {
541 ipath_cdbg(VERBOSE, "PID %u freeing TID %u\n",
543 dd->ipath_f_put_tid(dd, &tidbase[tid],
544 RCVHQ_RCV_TYPE_EXPECTED,
545 dd->ipath_tidinvalid);
546 pci_unmap_page(dd->pcidev,
547 dd->ipath_physshadow[porttid + tid],
548 PAGE_SIZE, PCI_DMA_FROMDEVICE);
549 ipath_release_user_pages(
550 &dd->ipath_pageshadow[porttid + tid], 1);
551 dd->ipath_pageshadow[porttid + tid] = NULL;
552 ipath_stats.sps_pageunlocks++;
554 ipath_dbg("Unused tid %u, ignoring\n", tid);
556 if (cnt != ti->tidcnt)
557 ipath_dbg("passed in tidcnt %d, only %d bits set in map\n",
561 ipath_dbg("Failed to unmap %u TID pages, failing with %d\n",
567 * ipath_set_part_key - set a partition key
571 * We can have up to 4 active at a time (other than the default, which is
572 * always allowed). This is somewhat tricky, since multiple ports may set
573 * the same key, so we reference count them, and clean up at exit. All 4
574 * partition keys are packed into a single infinipath register. It's an
575 * error for a process to set the same pkey multiple times. We provide no
576 * mechanism to de-allocate a pkey at this time, we may eventually need to
577 * do that. I've used the atomic operations, and no locking, and only make
578 * a single pass through what's available. This should be more than
579 * adequate for some time. I'll think about spinlocks or the like if and as
582 static int ipath_set_part_key(struct ipath_portdata *pd, u16 key)
584 struct ipath_devdata *dd = pd->port_dd;
585 int i, any = 0, pidx = -1;
586 u16 lkey = key & 0x7FFF;
589 if (lkey == (IPATH_DEFAULT_P_KEY & 0x7FFF)) {
590 /* nothing to do; this key always valid */
595 ipath_cdbg(VERBOSE, "p%u try to set pkey %hx, current keys "
596 "%hx:%x %hx:%x %hx:%x %hx:%x\n",
597 pd->port_port, key, dd->ipath_pkeys[0],
598 atomic_read(&dd->ipath_pkeyrefs[0]), dd->ipath_pkeys[1],
599 atomic_read(&dd->ipath_pkeyrefs[1]), dd->ipath_pkeys[2],
600 atomic_read(&dd->ipath_pkeyrefs[2]), dd->ipath_pkeys[3],
601 atomic_read(&dd->ipath_pkeyrefs[3]));
604 ipath_cdbg(PROC, "p%u tries to set key 0, not allowed\n",
611 * Set the full membership bit, because it has to be
612 * set in the register or the packet, and it seems
613 * cleaner to set in the register than to force all
614 * callers to set it. (see bug 4331)
618 for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
619 if (!pd->port_pkeys[i] && pidx == -1)
621 if (pd->port_pkeys[i] == key) {
622 ipath_cdbg(VERBOSE, "p%u tries to set same pkey "
623 "(%x) more than once\n",
630 ipath_dbg("All pkeys for port %u already in use, "
631 "can't set %x\n", pd->port_port, key);
635 for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
636 if (!dd->ipath_pkeys[i]) {
640 if (dd->ipath_pkeys[i] == key) {
641 atomic_t *pkrefs = &dd->ipath_pkeyrefs[i];
643 if (atomic_inc_return(pkrefs) > 1) {
644 pd->port_pkeys[pidx] = key;
645 ipath_cdbg(VERBOSE, "p%u set key %x "
646 "matches #%d, count now %d\n",
647 pd->port_port, key, i,
648 atomic_read(pkrefs));
653 * lost race, decrement count, catch below
656 ipath_cdbg(VERBOSE, "Lost race, count was "
657 "0, after dec, it's %d\n",
658 atomic_read(pkrefs));
662 if ((dd->ipath_pkeys[i] & 0x7FFF) == lkey) {
664 * It makes no sense to have both the limited and
665 * full membership PKEY set at the same time since
666 * the unlimited one will disable the limited one.
673 ipath_dbg("port %u, all pkeys already in use, "
674 "can't set %x\n", pd->port_port, key);
678 for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
679 if (!dd->ipath_pkeys[i] &&
680 atomic_inc_return(&dd->ipath_pkeyrefs[i]) == 1) {
683 /* for ipathstats, etc. */
684 ipath_stats.sps_pkeys[i] = lkey;
685 pd->port_pkeys[pidx] = dd->ipath_pkeys[i] = key;
687 (u64) dd->ipath_pkeys[0] |
688 ((u64) dd->ipath_pkeys[1] << 16) |
689 ((u64) dd->ipath_pkeys[2] << 32) |
690 ((u64) dd->ipath_pkeys[3] << 48);
691 ipath_cdbg(PROC, "p%u set key %x in #%d, "
692 "portidx %d, new pkey reg %llx\n",
693 pd->port_port, key, i, pidx,
694 (unsigned long long) pkey);
696 dd, dd->ipath_kregs->kr_partitionkey, pkey);
702 ipath_dbg("port %u, all pkeys already in use 2nd pass, "
703 "can't set %x\n", pd->port_port, key);
711 * ipath_manage_rcvq - manage a port's receive queue
713 * @subport: the subport
714 * @start_stop: action to carry out
716 * start_stop == 0 disables receive on the port, for use in queue
717 * overflow conditions. start_stop==1 re-enables, to be used to
718 * re-init the software copy of the head register
720 static int ipath_manage_rcvq(struct ipath_portdata *pd, unsigned subport,
723 struct ipath_devdata *dd = pd->port_dd;
725 ipath_cdbg(PROC, "%sabling rcv for unit %u port %u:%u\n",
726 start_stop ? "en" : "dis", dd->ipath_unit,
727 pd->port_port, subport);
730 /* atomically clear receive enable port. */
733 * On enable, force in-memory copy of the tail register to
734 * 0, so that protocol code doesn't have to worry about
735 * whether or not the chip has yet updated the in-memory
736 * copy or not on return from the system call. The chip
737 * always resets it's tail register back to 0 on a
738 * transition from disabled to enabled. This could cause a
739 * problem if software was broken, and did the enable w/o
740 * the disable, but eventually the in-memory copy will be
741 * updated and correct itself, even in the face of software
744 *(volatile u64 *)pd->port_rcvhdrtail_kvaddr = 0;
745 set_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
748 clear_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
750 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
752 /* now be sure chip saw it before we return */
753 ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
756 * And try to be sure that tail reg update has happened too.
757 * This should in theory interlock with the RXE changes to
758 * the tail register. Don't assign it to the tail register
759 * in memory copy, since we could overwrite an update by the
762 ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
764 /* always; new head should be equal to new tail; see above */
769 static void ipath_clean_part_key(struct ipath_portdata *pd,
770 struct ipath_devdata *dd)
772 int i, j, pchanged = 0;
775 /* for debugging only */
776 oldpkey = (u64) dd->ipath_pkeys[0] |
777 ((u64) dd->ipath_pkeys[1] << 16) |
778 ((u64) dd->ipath_pkeys[2] << 32) |
779 ((u64) dd->ipath_pkeys[3] << 48);
781 for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
782 if (!pd->port_pkeys[i])
784 ipath_cdbg(VERBOSE, "look for key[%d] %hx in pkeys\n", i,
786 for (j = 0; j < ARRAY_SIZE(dd->ipath_pkeys); j++) {
787 /* check for match independent of the global bit */
788 if ((dd->ipath_pkeys[j] & 0x7fff) !=
789 (pd->port_pkeys[i] & 0x7fff))
791 if (atomic_dec_and_test(&dd->ipath_pkeyrefs[j])) {
792 ipath_cdbg(VERBOSE, "p%u clear key "
795 pd->port_pkeys[i], j);
796 ipath_stats.sps_pkeys[j] =
797 dd->ipath_pkeys[j] = 0;
801 VERBOSE, "p%u key %x matches #%d, "
802 "but ref still %d\n", pd->port_port,
803 pd->port_pkeys[i], j,
804 atomic_read(&dd->ipath_pkeyrefs[j]));
807 pd->port_pkeys[i] = 0;
810 u64 pkey = (u64) dd->ipath_pkeys[0] |
811 ((u64) dd->ipath_pkeys[1] << 16) |
812 ((u64) dd->ipath_pkeys[2] << 32) |
813 ((u64) dd->ipath_pkeys[3] << 48);
814 ipath_cdbg(VERBOSE, "p%u old pkey reg %llx, "
815 "new pkey reg %llx\n", pd->port_port,
816 (unsigned long long) oldpkey,
817 (unsigned long long) pkey);
818 ipath_write_kreg(dd, dd->ipath_kregs->kr_partitionkey,
824 * Initialize the port data with the receive buffer sizes
825 * so this can be done while the master port is locked.
826 * Otherwise, there is a race with a slave opening the port
827 * and seeing these fields uninitialized.
829 static void init_user_egr_sizes(struct ipath_portdata *pd)
831 struct ipath_devdata *dd = pd->port_dd;
832 unsigned egrperchunk, egrcnt, size;
835 * to avoid wasting a lot of memory, we allocate 32KB chunks of
836 * physically contiguous memory, advance through it until used up
837 * and then allocate more. Of course, we need memory to store those
838 * extra pointers, now. Started out with 256KB, but under heavy
839 * memory pressure (creating large files and then copying them over
840 * NFS while doing lots of MPI jobs), we hit some allocation
841 * failures, even though we can sleep... (2.6.10) Still get
842 * failures at 64K. 32K is the lowest we can go without wasting
846 egrperchunk = size / dd->ipath_rcvegrbufsize;
847 egrcnt = dd->ipath_rcvegrcnt;
848 pd->port_rcvegrbuf_chunks = (egrcnt + egrperchunk - 1) / egrperchunk;
849 pd->port_rcvegrbufs_perchunk = egrperchunk;
850 pd->port_rcvegrbuf_size = size;
854 * ipath_create_user_egr - allocate eager TID buffers
855 * @pd: the port to allocate TID buffers for
857 * This routine is now quite different for user and kernel, because
858 * the kernel uses skb's, for the accelerated network performance
859 * This is the user port version
861 * Allocate the eager TID buffers and program them into infinipath
862 * They are no longer completely contiguous, we do multiple allocation
865 static int ipath_create_user_egr(struct ipath_portdata *pd)
867 struct ipath_devdata *dd = pd->port_dd;
868 unsigned e, egrcnt, egrperchunk, chunk, egrsize, egroff;
874 * GFP_USER, but without GFP_FS, so buffer cache can be
875 * coalesced (we hope); otherwise, even at order 4,
876 * heavy filesystem activity makes these fail, and we can
877 * use compound pages.
879 gfp_flags = __GFP_WAIT | __GFP_IO | __GFP_COMP;
881 egrcnt = dd->ipath_rcvegrcnt;
882 /* TID number offset for this port */
883 egroff = pd->port_port * egrcnt;
884 egrsize = dd->ipath_rcvegrbufsize;
885 ipath_cdbg(VERBOSE, "Allocating %d egr buffers, at egrtid "
886 "offset %x, egrsize %u\n", egrcnt, egroff, egrsize);
888 chunk = pd->port_rcvegrbuf_chunks;
889 egrperchunk = pd->port_rcvegrbufs_perchunk;
890 size = pd->port_rcvegrbuf_size;
891 pd->port_rcvegrbuf = kmalloc(chunk * sizeof(pd->port_rcvegrbuf[0]),
893 if (!pd->port_rcvegrbuf) {
897 pd->port_rcvegrbuf_phys =
898 kmalloc(chunk * sizeof(pd->port_rcvegrbuf_phys[0]),
900 if (!pd->port_rcvegrbuf_phys) {
904 for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
906 pd->port_rcvegrbuf[e] = dma_alloc_coherent(
907 &dd->pcidev->dev, size, &pd->port_rcvegrbuf_phys[e],
910 if (!pd->port_rcvegrbuf[e]) {
912 goto bail_rcvegrbuf_phys;
916 pd->port_rcvegr_phys = pd->port_rcvegrbuf_phys[0];
918 for (e = chunk = 0; chunk < pd->port_rcvegrbuf_chunks; chunk++) {
919 dma_addr_t pa = pd->port_rcvegrbuf_phys[chunk];
922 for (i = 0; e < egrcnt && i < egrperchunk; e++, i++) {
923 dd->ipath_f_put_tid(dd, e + egroff +
927 dd->ipath_rcvegrbase),
928 RCVHQ_RCV_TYPE_EAGER, pa);
931 cond_resched(); /* don't hog the cpu */
938 for (e = 0; e < pd->port_rcvegrbuf_chunks &&
939 pd->port_rcvegrbuf[e]; e++) {
940 dma_free_coherent(&dd->pcidev->dev, size,
941 pd->port_rcvegrbuf[e],
942 pd->port_rcvegrbuf_phys[e]);
945 kfree(pd->port_rcvegrbuf_phys);
946 pd->port_rcvegrbuf_phys = NULL;
948 kfree(pd->port_rcvegrbuf);
949 pd->port_rcvegrbuf = NULL;
955 /* common code for the mappings on dma_alloc_coherent mem */
956 static int ipath_mmap_mem(struct vm_area_struct *vma,
957 struct ipath_portdata *pd, unsigned len, int write_ok,
958 void *kvaddr, char *what)
960 struct ipath_devdata *dd = pd->port_dd;
964 if ((vma->vm_end - vma->vm_start) > len) {
965 dev_info(&dd->pcidev->dev,
966 "FAIL on %s: len %lx > %x\n", what,
967 vma->vm_end - vma->vm_start, len);
973 if (vma->vm_flags & VM_WRITE) {
974 dev_info(&dd->pcidev->dev,
975 "%s must be mapped readonly\n", what);
980 /* don't allow them to later change with mprotect */
981 vma->vm_flags &= ~VM_MAYWRITE;
984 pfn = virt_to_phys(kvaddr) >> PAGE_SHIFT;
985 ret = remap_pfn_range(vma, vma->vm_start, pfn,
986 len, vma->vm_page_prot);
988 dev_info(&dd->pcidev->dev, "%s port%u mmap of %lx, %x "
989 "bytes r%c failed: %d\n", what, pd->port_port,
990 pfn, len, write_ok?'w':'o', ret);
992 ipath_cdbg(VERBOSE, "%s port%u mmaped %lx, %x bytes "
993 "r%c\n", what, pd->port_port, pfn, len,
999 static int mmap_ureg(struct vm_area_struct *vma, struct ipath_devdata *dd,
1006 * This is real hardware, so use io_remap. This is the mechanism
1007 * for the user process to update the head registers for their port
1010 if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) {
1011 dev_info(&dd->pcidev->dev, "FAIL mmap userreg: reqlen "
1012 "%lx > PAGE\n", vma->vm_end - vma->vm_start);
1015 phys = dd->ipath_physaddr + ureg;
1016 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1018 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1019 ret = io_remap_pfn_range(vma, vma->vm_start,
1021 vma->vm_end - vma->vm_start,
1027 static int mmap_piobufs(struct vm_area_struct *vma,
1028 struct ipath_devdata *dd,
1029 struct ipath_portdata *pd,
1030 unsigned piobufs, unsigned piocnt)
1036 * When we map the PIO buffers in the chip, we want to map them as
1037 * writeonly, no read possible. This prevents access to previous
1038 * process data, and catches users who might try to read the i/o
1039 * space due to a bug.
1041 if ((vma->vm_end - vma->vm_start) > (piocnt * dd->ipath_palign)) {
1042 dev_info(&dd->pcidev->dev, "FAIL mmap piobufs: "
1043 "reqlen %lx > PAGE\n",
1044 vma->vm_end - vma->vm_start);
1049 phys = dd->ipath_physaddr + piobufs;
1052 * Don't mark this as non-cached, or we don't get the
1053 * write combining behavior we want on the PIO buffers!
1056 #if defined(__powerpc__)
1057 /* There isn't a generic way to specify writethrough mappings */
1058 pgprot_val(vma->vm_page_prot) |= _PAGE_NO_CACHE;
1059 pgprot_val(vma->vm_page_prot) |= _PAGE_WRITETHRU;
1060 pgprot_val(vma->vm_page_prot) &= ~_PAGE_GUARDED;
1064 * don't allow them to later change to readable with mprotect (for when
1065 * not initially mapped readable, as is normally the case)
1067 vma->vm_flags &= ~VM_MAYREAD;
1068 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1070 ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
1071 vma->vm_end - vma->vm_start,
1077 static int mmap_rcvegrbufs(struct vm_area_struct *vma,
1078 struct ipath_portdata *pd)
1080 struct ipath_devdata *dd = pd->port_dd;
1081 unsigned long start, size;
1082 size_t total_size, i;
1086 size = pd->port_rcvegrbuf_size;
1087 total_size = pd->port_rcvegrbuf_chunks * size;
1088 if ((vma->vm_end - vma->vm_start) > total_size) {
1089 dev_info(&dd->pcidev->dev, "FAIL on egr bufs: "
1090 "reqlen %lx > actual %lx\n",
1091 vma->vm_end - vma->vm_start,
1092 (unsigned long) total_size);
1097 if (vma->vm_flags & VM_WRITE) {
1098 dev_info(&dd->pcidev->dev, "Can't map eager buffers as "
1099 "writable (flags=%lx)\n", vma->vm_flags);
1103 /* don't allow them to later change to writeable with mprotect */
1104 vma->vm_flags &= ~VM_MAYWRITE;
1106 start = vma->vm_start;
1108 for (i = 0; i < pd->port_rcvegrbuf_chunks; i++, start += size) {
1109 pfn = virt_to_phys(pd->port_rcvegrbuf[i]) >> PAGE_SHIFT;
1110 ret = remap_pfn_range(vma, start, pfn, size,
1122 * ipath_file_vma_nopage - handle a VMA page fault.
1124 static struct page *ipath_file_vma_nopage(struct vm_area_struct *vma,
1125 unsigned long address, int *type)
1127 unsigned long offset = address - vma->vm_start;
1128 struct page *page = NOPAGE_SIGBUS;
1132 * Convert the vmalloc address into a struct page.
1134 pageptr = (void *)(offset + (vma->vm_pgoff << PAGE_SHIFT));
1135 page = vmalloc_to_page(pageptr);
1139 /* Increment the reference count. */
1142 *type = VM_FAULT_MINOR;
1147 static struct vm_operations_struct ipath_file_vm_ops = {
1148 .nopage = ipath_file_vma_nopage,
1151 static int mmap_kvaddr(struct vm_area_struct *vma, u64 pgaddr,
1152 struct ipath_portdata *pd, unsigned subport)
1155 struct ipath_devdata *dd;
1160 /* If the port is not shared, all addresses should be physical */
1161 if (!pd->port_subport_cnt)
1165 size = pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
1168 * Each process has all the subport uregbase, rcvhdrq, and
1169 * rcvegrbufs mmapped - as an array for all the processes,
1170 * and also separately for this process.
1172 if (pgaddr == cvt_kvaddr(pd->subport_uregbase)) {
1173 addr = pd->subport_uregbase;
1174 size = PAGE_SIZE * pd->port_subport_cnt;
1175 } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base)) {
1176 addr = pd->subport_rcvhdr_base;
1177 size = pd->port_rcvhdrq_size * pd->port_subport_cnt;
1178 } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf)) {
1179 addr = pd->subport_rcvegrbuf;
1180 size *= pd->port_subport_cnt;
1181 } else if (pgaddr == cvt_kvaddr(pd->subport_uregbase +
1182 PAGE_SIZE * subport)) {
1183 addr = pd->subport_uregbase + PAGE_SIZE * subport;
1185 } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base +
1186 pd->port_rcvhdrq_size * subport)) {
1187 addr = pd->subport_rcvhdr_base +
1188 pd->port_rcvhdrq_size * subport;
1189 size = pd->port_rcvhdrq_size;
1190 } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf +
1192 addr = pd->subport_rcvegrbuf + size * subport;
1193 /* rcvegrbufs are read-only on the slave */
1194 if (vma->vm_flags & VM_WRITE) {
1195 dev_info(&dd->pcidev->dev,
1196 "Can't map eager buffers as "
1197 "writable (flags=%lx)\n", vma->vm_flags);
1202 * Don't allow permission to later change to writeable
1205 vma->vm_flags &= ~VM_MAYWRITE;
1209 len = vma->vm_end - vma->vm_start;
1211 ipath_cdbg(MM, "FAIL: reqlen %lx > %zx\n", len, size);
1216 vma->vm_pgoff = (unsigned long) addr >> PAGE_SHIFT;
1217 vma->vm_ops = &ipath_file_vm_ops;
1218 vma->vm_flags |= VM_RESERVED | VM_DONTEXPAND;
1226 * ipath_mmap - mmap various structures into user space
1227 * @fp: the file pointer
1230 * We use this to have a shared buffer between the kernel and the user code
1231 * for the rcvhdr queue, egr buffers, and the per-port user regs and pio
1232 * buffers in the chip. We have the open and close entries so we can bump
1233 * the ref count and keep the driver from being unloaded while still mapped.
1235 static int ipath_mmap(struct file *fp, struct vm_area_struct *vma)
1237 struct ipath_portdata *pd;
1238 struct ipath_devdata *dd;
1240 unsigned piobufs, piocnt;
1251 * This is the ipath_do_user_init() code, mapping the shared buffers
1252 * into the user process. The address referred to by vm_pgoff is the
1253 * file offset passed via mmap(). For shared ports, this is the
1254 * kernel vmalloc() address of the pages to share with the master.
1255 * For non-shared or master ports, this is a physical address.
1256 * We only do one mmap for each space mapped.
1258 pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1261 * Check for 0 in case one of the allocations failed, but user
1262 * called mmap anyway.
1269 ipath_cdbg(MM, "pgaddr %llx vm_start=%lx len %lx port %u:%u:%u\n",
1270 (unsigned long long) pgaddr, vma->vm_start,
1271 vma->vm_end - vma->vm_start, dd->ipath_unit,
1272 pd->port_port, subport_fp(fp));
1275 * Physical addresses must fit in 40 bits for our hardware.
1276 * Check for kernel virtual addresses first, anything else must
1277 * match a HW or memory address.
1279 ret = mmap_kvaddr(vma, pgaddr, pd, subport_fp(fp));
1286 ureg = dd->ipath_uregbase + dd->ipath_palign * pd->port_port;
1287 if (!pd->port_subport_cnt) {
1288 /* port is not shared */
1289 piocnt = dd->ipath_pbufsport;
1290 piobufs = pd->port_piobufs;
1291 } else if (!subport_fp(fp)) {
1292 /* caller is the master */
1293 piocnt = (dd->ipath_pbufsport / pd->port_subport_cnt) +
1294 (dd->ipath_pbufsport % pd->port_subport_cnt);
1295 piobufs = pd->port_piobufs +
1296 dd->ipath_palign * (dd->ipath_pbufsport - piocnt);
1298 unsigned slave = subport_fp(fp) - 1;
1300 /* caller is a slave */
1301 piocnt = dd->ipath_pbufsport / pd->port_subport_cnt;
1302 piobufs = pd->port_piobufs + dd->ipath_palign * piocnt * slave;
1306 ret = mmap_ureg(vma, dd, ureg);
1307 else if (pgaddr == piobufs)
1308 ret = mmap_piobufs(vma, dd, pd, piobufs, piocnt);
1309 else if (pgaddr == dd->ipath_pioavailregs_phys)
1310 /* in-memory copy of pioavail registers */
1311 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1312 (void *) dd->ipath_pioavailregs_dma,
1313 "pioavail registers");
1314 else if (pgaddr == pd->port_rcvegr_phys)
1315 ret = mmap_rcvegrbufs(vma, pd);
1316 else if (pgaddr == (u64) pd->port_rcvhdrq_phys)
1318 * The rcvhdrq itself; readonly except on HT (so have
1319 * to allow writable mapping), multiple pages, contiguous
1320 * from an i/o perspective.
1322 ret = ipath_mmap_mem(vma, pd, pd->port_rcvhdrq_size, 1,
1325 else if (pgaddr == (u64) pd->port_rcvhdrqtailaddr_phys)
1326 /* in-memory copy of rcvhdrq tail register */
1327 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1328 pd->port_rcvhdrtail_kvaddr,
1333 vma->vm_private_data = NULL;
1336 dev_info(&dd->pcidev->dev,
1337 "Failure %d on off %llx len %lx\n",
1338 -ret, (unsigned long long)pgaddr,
1339 vma->vm_end - vma->vm_start);
1344 static unsigned int ipath_poll(struct file *fp,
1345 struct poll_table_struct *pt)
1347 struct ipath_portdata *pd;
1350 unsigned pollflag = 0;
1351 struct ipath_devdata *dd;
1358 bit = pd->port_port + INFINIPATH_R_INTRAVAIL_SHIFT;
1359 set_bit(bit, &dd->ipath_rcvctrl);
1362 * Before blocking, make sure that head is still == tail,
1363 * reading from the chip, so we can be sure the interrupt
1364 * enable has made it to the chip. If not equal, disable
1365 * interrupt again and return immediately. This avoids races,
1366 * and the overhead of the chip read doesn't matter much at
1367 * this point, since we are waiting for something anyway.
1370 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1373 head = ipath_read_ureg32(dd, ur_rcvhdrhead, pd->port_port);
1374 tail = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
1377 set_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
1378 if (dd->ipath_rhdrhead_intr_off) /* arm rcv interrupt */
1379 (void)ipath_write_ureg(dd, ur_rcvhdrhead,
1380 dd->ipath_rhdrhead_intr_off
1381 | head, pd->port_port);
1382 poll_wait(fp, &pd->port_wait, pt);
1384 if (test_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag)) {
1385 /* timed out, no packets received */
1386 clear_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
1387 pd->port_rcvwait_to++;
1390 pollflag = POLLIN | POLLRDNORM;
1393 /* it's already happened; don't do wait_event overhead */
1394 pollflag = POLLIN | POLLRDNORM;
1395 pd->port_rcvnowait++;
1398 clear_bit(bit, &dd->ipath_rcvctrl);
1399 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1406 static int ipath_supports_subports(int user_swmajor, int user_swminor)
1408 /* no subport implementation prior to software version 1.3 */
1409 return (user_swmajor > 1) || (user_swminor >= 3);
1412 static int ipath_compatible_subports(int user_swmajor, int user_swminor)
1414 /* this code is written long-hand for clarity */
1415 if (IPATH_USER_SWMAJOR != user_swmajor) {
1416 /* no promise of compatibility if major mismatch */
1419 if (IPATH_USER_SWMAJOR == 1) {
1420 switch (IPATH_USER_SWMINOR) {
1424 /* no subport implementation so cannot be compatible */
1427 /* 3 is only compatible with itself */
1428 return user_swminor == 3;
1430 /* >= 4 are compatible (or are expected to be) */
1431 return user_swminor >= 4;
1434 /* make no promises yet for future major versions */
1438 static int init_subports(struct ipath_devdata *dd,
1439 struct ipath_portdata *pd,
1440 const struct ipath_user_info *uinfo)
1443 unsigned num_subports;
1447 * If the user is requesting zero or one port,
1448 * skip the subport allocation.
1450 if (uinfo->spu_subport_cnt <= 1)
1453 /* Self-consistency check for ipath_compatible_subports() */
1454 if (ipath_supports_subports(IPATH_USER_SWMAJOR, IPATH_USER_SWMINOR) &&
1455 !ipath_compatible_subports(IPATH_USER_SWMAJOR,
1456 IPATH_USER_SWMINOR)) {
1457 dev_info(&dd->pcidev->dev,
1458 "Inconsistent ipath_compatible_subports()\n");
1462 /* Check for subport compatibility */
1463 if (!ipath_compatible_subports(uinfo->spu_userversion >> 16,
1464 uinfo->spu_userversion & 0xffff)) {
1465 dev_info(&dd->pcidev->dev,
1466 "Mismatched user version (%d.%d) and driver "
1467 "version (%d.%d) while port sharing. Ensure "
1468 "that driver and library are from the same "
1470 (int) (uinfo->spu_userversion >> 16),
1471 (int) (uinfo->spu_userversion & 0xffff),
1473 IPATH_USER_SWMINOR);
1476 if (uinfo->spu_subport_cnt > INFINIPATH_MAX_SUBPORT) {
1481 num_subports = uinfo->spu_subport_cnt;
1482 pd->subport_uregbase = vmalloc(PAGE_SIZE * num_subports);
1483 if (!pd->subport_uregbase) {
1487 /* Note: pd->port_rcvhdrq_size isn't initialized yet. */
1488 size = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
1489 sizeof(u32), PAGE_SIZE) * num_subports;
1490 pd->subport_rcvhdr_base = vmalloc(size);
1491 if (!pd->subport_rcvhdr_base) {
1496 pd->subport_rcvegrbuf = vmalloc(pd->port_rcvegrbuf_chunks *
1497 pd->port_rcvegrbuf_size *
1499 if (!pd->subport_rcvegrbuf) {
1504 pd->port_subport_cnt = uinfo->spu_subport_cnt;
1505 pd->port_subport_id = uinfo->spu_subport_id;
1506 pd->active_slaves = 1;
1507 set_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
1508 memset(pd->subport_uregbase, 0, PAGE_SIZE * num_subports);
1509 memset(pd->subport_rcvhdr_base, 0, size);
1510 memset(pd->subport_rcvegrbuf, 0, pd->port_rcvegrbuf_chunks *
1511 pd->port_rcvegrbuf_size *
1516 vfree(pd->subport_rcvhdr_base);
1518 vfree(pd->subport_uregbase);
1519 pd->subport_uregbase = NULL;
1524 static int try_alloc_port(struct ipath_devdata *dd, int port,
1526 const struct ipath_user_info *uinfo)
1528 struct ipath_portdata *pd;
1531 if (!(pd = dd->ipath_pd[port])) {
1534 pd = kzalloc(sizeof(struct ipath_portdata), GFP_KERNEL);
1537 * Allocate memory for use in ipath_tid_update() just once
1538 * at open, not per call. Reduces cost of expected send
1541 ptmp = kmalloc(dd->ipath_rcvtidcnt * sizeof(u16) +
1542 dd->ipath_rcvtidcnt * sizeof(struct page **),
1545 ipath_dev_err(dd, "Unable to allocate portdata "
1546 "memory, failing open\n");
1552 dd->ipath_pd[port] = pd;
1553 dd->ipath_pd[port]->port_port = port;
1554 dd->ipath_pd[port]->port_dd = dd;
1555 dd->ipath_pd[port]->port_tid_pg_list = ptmp;
1556 init_waitqueue_head(&dd->ipath_pd[port]->port_wait);
1558 if (!pd->port_cnt) {
1559 pd->userversion = uinfo->spu_userversion;
1560 init_user_egr_sizes(pd);
1561 if ((ret = init_subports(dd, pd, uinfo)) != 0)
1563 ipath_cdbg(PROC, "%s[%u] opened unit:port %u:%u\n",
1564 current->comm, current->pid, dd->ipath_unit,
1568 pd->port_pid = current->pid;
1569 strncpy(pd->port_comm, current->comm, sizeof(pd->port_comm));
1570 ipath_stats.sps_ports++;
1579 static inline int usable(struct ipath_devdata *dd)
1582 (dd->ipath_flags & IPATH_PRESENT) &&
1583 dd->ipath_kregbase &&
1585 !(dd->ipath_flags & (IPATH_LINKDOWN | IPATH_DISABLED
1589 static int find_free_port(int unit, struct file *fp,
1590 const struct ipath_user_info *uinfo)
1592 struct ipath_devdata *dd = ipath_lookup(unit);
1605 for (i = 1; i < dd->ipath_cfgports; i++) {
1606 ret = try_alloc_port(dd, i, fp, uinfo);
1616 static int find_best_unit(struct file *fp,
1617 const struct ipath_user_info *uinfo)
1619 int ret = 0, i, prefunit = -1, devmax;
1620 int maxofallports, npresent, nup;
1623 devmax = ipath_count_units(&npresent, &nup, &maxofallports);
1626 * This code is present to allow a knowledgeable person to
1627 * specify the layout of processes to processors before opening
1628 * this driver, and then we'll assign the process to the "closest"
1629 * InfiniPath chip to that processor (we assume reasonable connectivity,
1630 * for now). This code assumes that if affinity has been set
1631 * before this point, that at most one cpu is set; for now this
1632 * is reasonable. I check for both cpus_empty() and cpus_full(),
1633 * in case some kernel variant sets none of the bits when no
1634 * affinity is set. 2.6.11 and 12 kernels have all present
1635 * cpus set. Some day we'll have to fix it up further to handle
1636 * a cpu subset. This algorithm fails for two HT chips connected
1637 * in tunnel fashion. Eventually this needs real topology
1638 * information. There may be some issues with dual core numbering
1639 * as well. This needs more work prior to release.
1641 if (!cpus_empty(current->cpus_allowed) &&
1642 !cpus_full(current->cpus_allowed)) {
1643 int ncpus = num_online_cpus(), curcpu = -1, nset = 0;
1644 for (i = 0; i < ncpus; i++)
1645 if (cpu_isset(i, current->cpus_allowed)) {
1646 ipath_cdbg(PROC, "%s[%u] affinity set for "
1647 "cpu %d/%d\n", current->comm,
1648 current->pid, i, ncpus);
1652 if (curcpu != -1 && nset != ncpus) {
1654 prefunit = curcpu / (ncpus / npresent);
1655 ipath_cdbg(PROC,"%s[%u] %d chips, %d cpus, "
1656 "%d cpus/chip, select unit %d\n",
1657 current->comm, current->pid,
1658 npresent, ncpus, ncpus / npresent,
1665 * user ports start at 1, kernel port is 0
1666 * For now, we do round-robin access across all chips
1670 devmax = prefunit + 1;
1672 for (i = 1; i < maxofallports; i++) {
1673 for (ndev = prefunit != -1 ? prefunit : 0; ndev < devmax;
1675 struct ipath_devdata *dd = ipath_lookup(ndev);
1678 continue; /* can't use this unit */
1679 if (i >= dd->ipath_cfgports)
1681 * Maxed out on users of this unit. Try
1685 ret = try_alloc_port(dd, i, fp, uinfo);
1694 ipath_dbg("No ports available (none initialized "
1698 /* if started above 0, retry from 0 */
1700 "%s[%u] no ports on prefunit "
1701 "%d, clear and re-check\n",
1702 current->comm, current->pid,
1704 devmax = ipath_count_units(NULL, NULL,
1710 ipath_dbg("No ports available\n");
1714 ipath_dbg("No boards found\n");
1721 static int find_shared_port(struct file *fp,
1722 const struct ipath_user_info *uinfo)
1724 int devmax, ndev, i;
1727 devmax = ipath_count_units(NULL, NULL, NULL);
1729 for (ndev = 0; ndev < devmax; ndev++) {
1730 struct ipath_devdata *dd = ipath_lookup(ndev);
1734 for (i = 1; i < dd->ipath_cfgports; i++) {
1735 struct ipath_portdata *pd = dd->ipath_pd[i];
1737 /* Skip ports which are not yet open */
1738 if (!pd || !pd->port_cnt)
1740 /* Skip port if it doesn't match the requested one */
1741 if (pd->port_subport_id != uinfo->spu_subport_id)
1743 /* Verify the sharing process matches the master */
1744 if (pd->port_subport_cnt != uinfo->spu_subport_cnt ||
1745 pd->userversion != uinfo->spu_userversion ||
1746 pd->port_cnt >= pd->port_subport_cnt) {
1751 subport_fp(fp) = pd->port_cnt++;
1752 tidcursor_fp(fp) = 0;
1753 pd->active_slaves |= 1 << subport_fp(fp);
1755 "%s[%u] %u sharing %s[%u] unit:port %u:%u\n",
1756 current->comm, current->pid,
1758 pd->port_comm, pd->port_pid,
1759 dd->ipath_unit, pd->port_port);
1769 static int ipath_open(struct inode *in, struct file *fp)
1771 /* The real work is performed later in ipath_assign_port() */
1772 fp->private_data = kzalloc(sizeof(struct ipath_filedata), GFP_KERNEL);
1773 return fp->private_data ? 0 : -ENOMEM;
1776 /* Get port early, so can set affinity prior to memory allocation */
1777 static int ipath_assign_port(struct file *fp,
1778 const struct ipath_user_info *uinfo)
1782 unsigned swmajor, swminor;
1784 /* Check to be sure we haven't already initialized this file */
1790 /* for now, if major version is different, bail */
1791 swmajor = uinfo->spu_userversion >> 16;
1792 if (swmajor != IPATH_USER_SWMAJOR) {
1793 ipath_dbg("User major version %d not same as driver "
1794 "major %d\n", uinfo->spu_userversion >> 16,
1795 IPATH_USER_SWMAJOR);
1800 swminor = uinfo->spu_userversion & 0xffff;
1801 if (swminor != IPATH_USER_SWMINOR)
1802 ipath_dbg("User minor version %d not same as driver "
1803 "minor %d\n", swminor, IPATH_USER_SWMINOR);
1805 mutex_lock(&ipath_mutex);
1807 if (ipath_compatible_subports(swmajor, swminor) &&
1808 uinfo->spu_subport_cnt &&
1809 (ret = find_shared_port(fp, uinfo))) {
1810 mutex_unlock(&ipath_mutex);
1816 i_minor = iminor(fp->f_path.dentry->d_inode) - IPATH_USER_MINOR_BASE;
1817 ipath_cdbg(VERBOSE, "open on dev %lx (minor %d)\n",
1818 (long)fp->f_path.dentry->d_inode->i_rdev, i_minor);
1821 ret = find_free_port(i_minor - 1, fp, uinfo);
1823 ret = find_best_unit(fp, uinfo);
1825 mutex_unlock(&ipath_mutex);
1832 static int ipath_do_user_init(struct file *fp,
1833 const struct ipath_user_info *uinfo)
1836 struct ipath_portdata *pd = port_fp(fp);
1837 struct ipath_devdata *dd;
1840 /* Subports don't need to initialize anything since master did it. */
1841 if (subport_fp(fp)) {
1842 ret = wait_event_interruptible(pd->port_wait,
1843 !test_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag));
1849 if (uinfo->spu_rcvhdrsize) {
1850 ret = ipath_setrcvhdrsize(dd, uinfo->spu_rcvhdrsize);
1855 /* for now we do nothing with rcvhdrcnt: uinfo->spu_rcvhdrcnt */
1857 /* for right now, kernel piobufs are at end, so port 1 is at 0 */
1858 pd->port_piobufs = dd->ipath_piobufbase +
1859 dd->ipath_pbufsport * (pd->port_port - 1) * dd->ipath_palign;
1860 ipath_cdbg(VERBOSE, "Set base of piobufs for port %u to 0x%x\n",
1861 pd->port_port, pd->port_piobufs);
1864 * Now allocate the rcvhdr Q and eager TIDs; skip the TID
1865 * array for time being. If pd->port_port > chip-supported,
1866 * we need to do extra stuff here to handle by handling overflow
1867 * through port 0, someday
1869 ret = ipath_create_rcvhdrq(dd, pd);
1871 ret = ipath_create_user_egr(pd);
1876 * set the eager head register for this port to the current values
1877 * of the tail pointers, since we don't know if they were
1878 * updated on last use of the port.
1880 head32 = ipath_read_ureg32(dd, ur_rcvegrindextail, pd->port_port);
1881 ipath_write_ureg(dd, ur_rcvegrindexhead, head32, pd->port_port);
1882 dd->ipath_lastegrheads[pd->port_port] = -1;
1883 dd->ipath_lastrcvhdrqtails[pd->port_port] = -1;
1884 ipath_cdbg(VERBOSE, "Wrote port%d egrhead %x from tail regs\n",
1885 pd->port_port, head32);
1886 pd->port_tidcursor = 0; /* start at beginning after open */
1888 * now enable the port; the tail registers will be written to memory
1889 * by the chip as soon as it sees the write to
1890 * dd->ipath_kregs->kr_rcvctrl. The update only happens on
1891 * transition from 0 to 1, so clear it first, then set it as part of
1892 * enabling the port. This will (very briefly) affect any other
1893 * open ports, but it shouldn't be long enough to be an issue.
1894 * We explictly set the in-memory copy to 0 beforehand, so we don't
1895 * have to wait to be sure the DMA update has happened.
1897 *(volatile u64 *)pd->port_rcvhdrtail_kvaddr = 0ULL;
1898 set_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
1899 &dd->ipath_rcvctrl);
1900 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1901 dd->ipath_rcvctrl & ~INFINIPATH_R_TAILUPD);
1902 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1904 /* Notify any waiting slaves */
1905 if (pd->port_subport_cnt) {
1906 clear_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
1907 wake_up(&pd->port_wait);
1914 * unlock_exptid - unlock any expected TID entries port still had in use
1917 * We don't actually update the chip here, because we do a bulk update
1918 * below, using ipath_f_clear_tids.
1920 static void unlock_expected_tids(struct ipath_portdata *pd)
1922 struct ipath_devdata *dd = pd->port_dd;
1923 int port_tidbase = pd->port_port * dd->ipath_rcvtidcnt;
1924 int i, cnt = 0, maxtid = port_tidbase + dd->ipath_rcvtidcnt;
1926 ipath_cdbg(VERBOSE, "Port %u unlocking any locked expTID pages\n",
1928 for (i = port_tidbase; i < maxtid; i++) {
1929 if (!dd->ipath_pageshadow[i])
1932 pci_unmap_page(dd->pcidev, dd->ipath_physshadow[i],
1933 PAGE_SIZE, PCI_DMA_FROMDEVICE);
1934 ipath_release_user_pages_on_close(&dd->ipath_pageshadow[i],
1936 dd->ipath_pageshadow[i] = NULL;
1938 ipath_stats.sps_pageunlocks++;
1941 ipath_cdbg(VERBOSE, "Port %u locked %u expTID entries\n",
1942 pd->port_port, cnt);
1944 if (ipath_stats.sps_pagelocks || ipath_stats.sps_pageunlocks)
1945 ipath_cdbg(VERBOSE, "%llu pages locked, %llu unlocked\n",
1946 (unsigned long long) ipath_stats.sps_pagelocks,
1947 (unsigned long long)
1948 ipath_stats.sps_pageunlocks);
1951 static int ipath_close(struct inode *in, struct file *fp)
1954 struct ipath_filedata *fd;
1955 struct ipath_portdata *pd;
1956 struct ipath_devdata *dd;
1959 ipath_cdbg(VERBOSE, "close on dev %lx, private data %p\n",
1960 (long)in->i_rdev, fp->private_data);
1962 mutex_lock(&ipath_mutex);
1964 fd = (struct ipath_filedata *) fp->private_data;
1965 fp->private_data = NULL;
1968 mutex_unlock(&ipath_mutex);
1971 if (--pd->port_cnt) {
1973 * XXX If the master closes the port before the slave(s),
1974 * revoke the mmap for the eager receive queue so
1975 * the slave(s) don't wait for receive data forever.
1977 pd->active_slaves &= ~(1 << fd->subport);
1978 mutex_unlock(&ipath_mutex);
1981 port = pd->port_port;
1984 if (pd->port_hdrqfull) {
1985 ipath_cdbg(PROC, "%s[%u] had %u rcvhdrqfull errors "
1986 "during run\n", pd->port_comm, pd->port_pid,
1988 pd->port_hdrqfull = 0;
1991 if (pd->port_rcvwait_to || pd->port_piowait_to
1992 || pd->port_rcvnowait || pd->port_pionowait) {
1993 ipath_cdbg(VERBOSE, "port%u, %u rcv, %u pio wait timeo; "
1994 "%u rcv %u, pio already\n",
1995 pd->port_port, pd->port_rcvwait_to,
1996 pd->port_piowait_to, pd->port_rcvnowait,
1997 pd->port_pionowait);
1998 pd->port_rcvwait_to = pd->port_piowait_to =
1999 pd->port_rcvnowait = pd->port_pionowait = 0;
2001 if (pd->port_flag) {
2002 ipath_dbg("port %u port_flag still set to 0x%lx\n",
2003 pd->port_port, pd->port_flag);
2007 if (dd->ipath_kregbase) {
2009 /* atomically clear receive enable port. */
2010 clear_bit(INFINIPATH_R_PORTENABLE_SHIFT + port,
2011 &dd->ipath_rcvctrl);
2012 ipath_write_kreg( dd, dd->ipath_kregs->kr_rcvctrl,
2014 /* and read back from chip to be sure that nothing
2015 * else is in flight when we do the rest */
2016 (void)ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
2018 /* clean up the pkeys for this port user */
2019 ipath_clean_part_key(pd, dd);
2021 * be paranoid, and never write 0's to these, just use an
2022 * unused part of the port 0 tail page. Of course,
2023 * rcvhdraddr points to a large chunk of memory, so this
2024 * could still trash things, but at least it won't trash
2025 * page 0, and by disabling the port, it should stop "soon",
2026 * even if a packet or two is in already in flight after we
2027 * disabled the port.
2029 ipath_write_kreg_port(dd,
2030 dd->ipath_kregs->kr_rcvhdrtailaddr, port,
2031 dd->ipath_dummy_hdrq_phys);
2032 ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdraddr,
2033 pd->port_port, dd->ipath_dummy_hdrq_phys);
2035 i = dd->ipath_pbufsport * (port - 1);
2036 ipath_disarm_piobufs(dd, i, dd->ipath_pbufsport);
2038 dd->ipath_f_clear_tids(dd, pd->port_port);
2040 if (dd->ipath_pageshadow)
2041 unlock_expected_tids(pd);
2042 ipath_stats.sps_ports--;
2043 ipath_cdbg(PROC, "%s[%u] closed port %u:%u\n",
2044 pd->port_comm, pd->port_pid,
2045 dd->ipath_unit, port);
2049 dd->ipath_pd[pd->port_port] = NULL; /* before releasing mutex */
2050 mutex_unlock(&ipath_mutex);
2051 ipath_free_pddata(dd, pd); /* after releasing the mutex */
2058 static int ipath_port_info(struct ipath_portdata *pd, u16 subport,
2059 struct ipath_port_info __user *uinfo)
2061 struct ipath_port_info info;
2066 (void) ipath_count_units(NULL, &nup, NULL);
2067 info.num_active = nup;
2068 info.unit = pd->port_dd->ipath_unit;
2069 info.port = pd->port_port;
2070 info.subport = subport;
2071 /* Don't return new fields if old library opened the port. */
2072 if (ipath_supports_subports(pd->userversion >> 16,
2073 pd->userversion & 0xffff)) {
2074 /* Number of user ports available for this device. */
2075 info.num_ports = pd->port_dd->ipath_cfgports - 1;
2076 info.num_subports = pd->port_subport_cnt;
2079 sz = sizeof(info) - 2 * sizeof(u16);
2081 if (copy_to_user(uinfo, &info, sz)) {
2091 static int ipath_get_slave_info(struct ipath_portdata *pd,
2092 void __user *slave_mask_addr)
2096 if (copy_to_user(slave_mask_addr, &pd->active_slaves, sizeof(u32)))
2101 static int ipath_force_pio_avail_update(struct ipath_devdata *dd)
2103 u64 reg = dd->ipath_sendctrl;
2105 clear_bit(IPATH_S_PIOBUFAVAILUPD, ®);
2106 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, reg);
2107 ipath_write_kreg(dd, dd->ipath_kregs->kr_sendctrl, dd->ipath_sendctrl);
2112 static ssize_t ipath_write(struct file *fp, const char __user *data,
2113 size_t count, loff_t *off)
2115 const struct ipath_cmd __user *ucmd;
2116 struct ipath_portdata *pd;
2117 const void __user *src;
2118 size_t consumed, copy;
2119 struct ipath_cmd cmd;
2123 if (count < sizeof(cmd.type)) {
2128 ucmd = (const struct ipath_cmd __user *) data;
2130 if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) {
2135 consumed = sizeof(cmd.type);
2138 case IPATH_CMD_ASSIGN_PORT:
2139 case __IPATH_CMD_USER_INIT:
2140 case IPATH_CMD_USER_INIT:
2141 copy = sizeof(cmd.cmd.user_info);
2142 dest = &cmd.cmd.user_info;
2143 src = &ucmd->cmd.user_info;
2145 case IPATH_CMD_RECV_CTRL:
2146 copy = sizeof(cmd.cmd.recv_ctrl);
2147 dest = &cmd.cmd.recv_ctrl;
2148 src = &ucmd->cmd.recv_ctrl;
2150 case IPATH_CMD_PORT_INFO:
2151 copy = sizeof(cmd.cmd.port_info);
2152 dest = &cmd.cmd.port_info;
2153 src = &ucmd->cmd.port_info;
2155 case IPATH_CMD_TID_UPDATE:
2156 case IPATH_CMD_TID_FREE:
2157 copy = sizeof(cmd.cmd.tid_info);
2158 dest = &cmd.cmd.tid_info;
2159 src = &ucmd->cmd.tid_info;
2161 case IPATH_CMD_SET_PART_KEY:
2162 copy = sizeof(cmd.cmd.part_key);
2163 dest = &cmd.cmd.part_key;
2164 src = &ucmd->cmd.part_key;
2166 case __IPATH_CMD_SLAVE_INFO:
2167 copy = sizeof(cmd.cmd.slave_mask_addr);
2168 dest = &cmd.cmd.slave_mask_addr;
2169 src = &ucmd->cmd.slave_mask_addr;
2171 case IPATH_CMD_PIOAVAILUPD: // force an update of PIOAvail reg
2182 if ((count - consumed) < copy) {
2187 if (copy_from_user(dest, src, copy)) {
2196 if (!pd && cmd.type != __IPATH_CMD_USER_INIT &&
2197 cmd.type != IPATH_CMD_ASSIGN_PORT) {
2203 case IPATH_CMD_ASSIGN_PORT:
2204 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2208 case __IPATH_CMD_USER_INIT:
2209 /* backwards compatibility, get port first */
2210 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2213 /* and fall through to current version. */
2214 case IPATH_CMD_USER_INIT:
2215 ret = ipath_do_user_init(fp, &cmd.cmd.user_info);
2218 ret = ipath_get_base_info(
2219 fp, (void __user *) (unsigned long)
2220 cmd.cmd.user_info.spu_base_info,
2221 cmd.cmd.user_info.spu_base_info_size);
2223 case IPATH_CMD_RECV_CTRL:
2224 ret = ipath_manage_rcvq(pd, subport_fp(fp), cmd.cmd.recv_ctrl);
2226 case IPATH_CMD_PORT_INFO:
2227 ret = ipath_port_info(pd, subport_fp(fp),
2228 (struct ipath_port_info __user *)
2229 (unsigned long) cmd.cmd.port_info);
2231 case IPATH_CMD_TID_UPDATE:
2232 ret = ipath_tid_update(pd, fp, &cmd.cmd.tid_info);
2234 case IPATH_CMD_TID_FREE:
2235 ret = ipath_tid_free(pd, subport_fp(fp), &cmd.cmd.tid_info);
2237 case IPATH_CMD_SET_PART_KEY:
2238 ret = ipath_set_part_key(pd, cmd.cmd.part_key);
2240 case __IPATH_CMD_SLAVE_INFO:
2241 ret = ipath_get_slave_info(pd,
2242 (void __user *) (unsigned long)
2243 cmd.cmd.slave_mask_addr);
2245 case IPATH_CMD_PIOAVAILUPD:
2246 ret = ipath_force_pio_avail_update(pd->port_dd);
2257 static struct class *ipath_class;
2259 static int init_cdev(int minor, char *name, const struct file_operations *fops,
2260 struct cdev **cdevp, struct class_device **class_devp)
2262 const dev_t dev = MKDEV(IPATH_MAJOR, minor);
2263 struct cdev *cdev = NULL;
2264 struct class_device *class_dev = NULL;
2267 cdev = cdev_alloc();
2269 printk(KERN_ERR IPATH_DRV_NAME
2270 ": Could not allocate cdev for minor %d, %s\n",
2276 cdev->owner = THIS_MODULE;
2278 kobject_set_name(&cdev->kobj, name);
2280 ret = cdev_add(cdev, dev, 1);
2282 printk(KERN_ERR IPATH_DRV_NAME
2283 ": Could not add cdev for minor %d, %s (err %d)\n",
2288 class_dev = class_device_create(ipath_class, NULL, dev, NULL, name);
2290 if (IS_ERR(class_dev)) {
2291 ret = PTR_ERR(class_dev);
2292 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2293 "class_dev for minor %d, %s (err %d)\n",
2307 *class_devp = class_dev;
2316 int ipath_cdev_init(int minor, char *name, const struct file_operations *fops,
2317 struct cdev **cdevp, struct class_device **class_devp)
2319 return init_cdev(minor, name, fops, cdevp, class_devp);
2322 static void cleanup_cdev(struct cdev **cdevp,
2323 struct class_device **class_devp)
2325 struct class_device *class_dev = *class_devp;
2328 class_device_unregister(class_dev);
2338 void ipath_cdev_cleanup(struct cdev **cdevp,
2339 struct class_device **class_devp)
2341 cleanup_cdev(cdevp, class_devp);
2344 static struct cdev *wildcard_cdev;
2345 static struct class_device *wildcard_class_dev;
2347 static const dev_t dev = MKDEV(IPATH_MAJOR, 0);
2349 static int user_init(void)
2353 ret = register_chrdev_region(dev, IPATH_NMINORS, IPATH_DRV_NAME);
2355 printk(KERN_ERR IPATH_DRV_NAME ": Could not register "
2356 "chrdev region (err %d)\n", -ret);
2360 ipath_class = class_create(THIS_MODULE, IPATH_DRV_NAME);
2362 if (IS_ERR(ipath_class)) {
2363 ret = PTR_ERR(ipath_class);
2364 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2365 "device class (err %d)\n", -ret);
2371 unregister_chrdev_region(dev, IPATH_NMINORS);
2376 static void user_cleanup(void)
2379 class_destroy(ipath_class);
2383 unregister_chrdev_region(dev, IPATH_NMINORS);
2386 static atomic_t user_count = ATOMIC_INIT(0);
2387 static atomic_t user_setup = ATOMIC_INIT(0);
2389 int ipath_user_add(struct ipath_devdata *dd)
2394 if (atomic_inc_return(&user_count) == 1) {
2397 ipath_dev_err(dd, "Unable to set up user support: "
2398 "error %d\n", -ret);
2401 ret = init_cdev(0, "ipath", &ipath_file_ops, &wildcard_cdev,
2402 &wildcard_class_dev);
2404 ipath_dev_err(dd, "Could not create wildcard "
2405 "minor: error %d\n", -ret);
2409 atomic_set(&user_setup, 1);
2412 snprintf(name, sizeof(name), "ipath%d", dd->ipath_unit);
2414 ret = init_cdev(dd->ipath_unit + 1, name, &ipath_file_ops,
2415 &dd->user_cdev, &dd->user_class_dev);
2417 ipath_dev_err(dd, "Could not create user minor %d, %s\n",
2418 dd->ipath_unit + 1, name);
2428 void ipath_user_remove(struct ipath_devdata *dd)
2430 cleanup_cdev(&dd->user_cdev, &dd->user_class_dev);
2432 if (atomic_dec_return(&user_count) == 0) {
2433 if (atomic_read(&user_setup) == 0)
2436 cleanup_cdev(&wildcard_cdev, &wildcard_class_dev);
2439 atomic_set(&user_setup, 0);