2 * Copyright (c) 2006, 2007, 2008 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 <linux/slab.h>
40 #include <linux/highmem.h>
42 #include <linux/jiffies.h>
43 #include <asm/pgtable.h>
45 #include "ipath_kernel.h"
46 #include "ipath_common.h"
47 #include "ipath_user_sdma.h"
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,
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 *);
58 static const struct file_operations ipath_file_ops = {
61 .aio_write = ipath_writev,
63 .release = ipath_close,
66 .llseek = noop_llseek,
70 * Convert kernel virtual addresses to physical addresses so they don't
71 * potentially conflict with the chip addresses used as mmap offsets.
72 * It doesn't really matter what mmap offset we use as long as we can
73 * interpret it correctly.
75 static u64 cvt_kvaddr(void *p)
80 page = vmalloc_to_page(p);
82 paddr = page_to_pfn(page) << PAGE_SHIFT;
87 static int ipath_get_base_info(struct file *fp,
88 void __user *ubase, size_t ubase_size)
90 struct ipath_portdata *pd = port_fp(fp);
92 struct ipath_base_info *kinfo = NULL;
93 struct ipath_devdata *dd = pd->port_dd;
98 subport_cnt = pd->port_subport_cnt;
105 master = !subport_fp(fp);
109 /* If port sharing is not requested, allow the old size structure */
111 sz -= 7 * sizeof(u64);
112 if (ubase_size < sz) {
114 "Base size %zu, need %zu (version mismatch?)\n",
120 kinfo = kzalloc(sizeof(*kinfo), GFP_KERNEL);
126 ret = dd->ipath_f_get_base_info(pd, kinfo);
130 kinfo->spi_rcvhdr_cnt = dd->ipath_rcvhdrcnt;
131 kinfo->spi_rcvhdrent_size = dd->ipath_rcvhdrentsize;
132 kinfo->spi_tidegrcnt = dd->ipath_rcvegrcnt;
133 kinfo->spi_rcv_egrbufsize = dd->ipath_rcvegrbufsize;
135 * have to mmap whole thing
137 kinfo->spi_rcv_egrbuftotlen =
138 pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
139 kinfo->spi_rcv_egrperchunk = pd->port_rcvegrbufs_perchunk;
140 kinfo->spi_rcv_egrchunksize = kinfo->spi_rcv_egrbuftotlen /
141 pd->port_rcvegrbuf_chunks;
142 kinfo->spi_tidcnt = dd->ipath_rcvtidcnt / subport_cnt;
144 kinfo->spi_tidcnt += dd->ipath_rcvtidcnt % subport_cnt;
146 * for this use, may be ipath_cfgports summed over all chips that
147 * are are configured and present
149 kinfo->spi_nports = dd->ipath_cfgports;
150 /* unit (chip/board) our port is on */
151 kinfo->spi_unit = dd->ipath_unit;
152 /* for now, only a single page */
153 kinfo->spi_tid_maxsize = PAGE_SIZE;
156 * Doing this per port, and based on the skip value, etc. This has
157 * to be the actual buffer size, since the protocol code treats it
160 * These have to be set to user addresses in the user code via mmap.
161 * These values are used on return to user code for the mmap target
162 * addresses only. For 32 bit, same 44 bit address problem, so use
163 * the physical address, not virtual. Before 2.6.11, using the
164 * page_address() macro worked, but in 2.6.11, even that returns the
165 * full 64 bit address (upper bits all 1's). So far, using the
166 * physical addresses (or chip offsets, for chip mapping) works, but
167 * no doubt some future kernel release will change that, and we'll be
168 * on to yet another method of dealing with this.
170 kinfo->spi_rcvhdr_base = (u64) pd->port_rcvhdrq_phys;
171 kinfo->spi_rcvhdr_tailaddr = (u64) pd->port_rcvhdrqtailaddr_phys;
172 kinfo->spi_rcv_egrbufs = (u64) pd->port_rcvegr_phys;
173 kinfo->spi_pioavailaddr = (u64) dd->ipath_pioavailregs_phys;
174 kinfo->spi_status = (u64) kinfo->spi_pioavailaddr +
175 (void *) dd->ipath_statusp -
176 (void *) dd->ipath_pioavailregs_dma;
178 kinfo->spi_piocnt = pd->port_piocnt;
179 kinfo->spi_piobufbase = (u64) pd->port_piobufs;
180 kinfo->__spi_uregbase = (u64) dd->ipath_uregbase +
181 dd->ipath_ureg_align * pd->port_port;
183 kinfo->spi_piocnt = (pd->port_piocnt / subport_cnt) +
184 (pd->port_piocnt % subport_cnt);
185 /* Master's PIO buffers are after all the slave's */
186 kinfo->spi_piobufbase = (u64) pd->port_piobufs +
188 (pd->port_piocnt - kinfo->spi_piocnt);
190 unsigned slave = subport_fp(fp) - 1;
192 kinfo->spi_piocnt = pd->port_piocnt / subport_cnt;
193 kinfo->spi_piobufbase = (u64) pd->port_piobufs +
194 dd->ipath_palign * kinfo->spi_piocnt * slave;
198 kinfo->spi_port_uregbase = (u64) dd->ipath_uregbase +
199 dd->ipath_ureg_align * pd->port_port;
200 kinfo->spi_port_rcvegrbuf = kinfo->spi_rcv_egrbufs;
201 kinfo->spi_port_rcvhdr_base = kinfo->spi_rcvhdr_base;
202 kinfo->spi_port_rcvhdr_tailaddr = kinfo->spi_rcvhdr_tailaddr;
204 kinfo->__spi_uregbase = cvt_kvaddr(pd->subport_uregbase +
205 PAGE_SIZE * subport_fp(fp));
207 kinfo->spi_rcvhdr_base = cvt_kvaddr(pd->subport_rcvhdr_base +
208 pd->port_rcvhdrq_size * subport_fp(fp));
209 kinfo->spi_rcvhdr_tailaddr = 0;
210 kinfo->spi_rcv_egrbufs = cvt_kvaddr(pd->subport_rcvegrbuf +
211 pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size *
214 kinfo->spi_subport_uregbase =
215 cvt_kvaddr(pd->subport_uregbase);
216 kinfo->spi_subport_rcvegrbuf =
217 cvt_kvaddr(pd->subport_rcvegrbuf);
218 kinfo->spi_subport_rcvhdr_base =
219 cvt_kvaddr(pd->subport_rcvhdr_base);
220 ipath_cdbg(PROC, "port %u flags %x %llx %llx %llx\n",
221 kinfo->spi_port, kinfo->spi_runtime_flags,
222 (unsigned long long) kinfo->spi_subport_uregbase,
223 (unsigned long long) kinfo->spi_subport_rcvegrbuf,
224 (unsigned long long) kinfo->spi_subport_rcvhdr_base);
228 * All user buffers are 2KB buffers. If we ever support
229 * giving 4KB buffers to user processes, this will need some
232 kinfo->spi_pioindex = (kinfo->spi_piobufbase -
233 (dd->ipath_piobufbase & 0xffffffff)) / dd->ipath_palign;
234 kinfo->spi_pioalign = dd->ipath_palign;
236 kinfo->spi_qpair = IPATH_KD_QP;
238 * user mode PIO buffers are always 2KB, even when 4KB can
239 * be received, and sent via the kernel; this is ibmaxlen
242 kinfo->spi_piosize = dd->ipath_piosize2k - 2 * sizeof(u32);
243 kinfo->spi_mtu = dd->ipath_ibmaxlen; /* maxlen, not ibmtu */
244 kinfo->spi_port = pd->port_port;
245 kinfo->spi_subport = subport_fp(fp);
246 kinfo->spi_sw_version = IPATH_KERN_SWVERSION;
247 kinfo->spi_hw_version = dd->ipath_revision;
250 kinfo->spi_runtime_flags |= IPATH_RUNTIME_MASTER;
253 sz = (ubase_size < sizeof(*kinfo)) ? ubase_size : sizeof(*kinfo);
254 if (copy_to_user(ubase, kinfo, sz))
263 * ipath_tid_update - update a port TID
265 * @fp: the ipath device file
266 * @ti: the TID information
268 * The new implementation as of Oct 2004 is that the driver assigns
269 * the tid and returns it to the caller. To make it easier to
270 * catch bugs, and to reduce search time, we keep a cursor for
271 * each port, walking the shadow tid array to find one that's not
274 * For now, if we can't allocate the full list, we fail, although
275 * in the long run, we'll allocate as many as we can, and the
276 * caller will deal with that by trying the remaining pages later.
277 * That means that when we fail, we have to mark the tids as not in
278 * use again, in our shadow copy.
280 * It's up to the caller to free the tids when they are done.
281 * We'll unlock the pages as they free them.
283 * Also, right now we are locking one page at a time, but since
284 * the intended use of this routine is for a single group of
285 * virtually contiguous pages, that should change to improve
288 static int ipath_tid_update(struct ipath_portdata *pd, struct file *fp,
289 const struct ipath_tid_info *ti)
292 u32 tid, porttid, cnt, i, tidcnt, tidoff;
294 struct ipath_devdata *dd = pd->port_dd;
297 u64 __iomem *tidbase;
298 unsigned long tidmap[8];
299 struct page **pagep = NULL;
300 unsigned subport = subport_fp(fp);
302 if (!dd->ipath_pageshadow) {
309 ipath_dbg("After copyin, tidcnt 0, tidlist %llx\n",
310 (unsigned long long) ti->tidlist);
312 * Should we treat as success? likely a bug
317 porttid = pd->port_port * dd->ipath_rcvtidcnt;
318 if (!pd->port_subport_cnt) {
319 tidcnt = dd->ipath_rcvtidcnt;
320 tid = pd->port_tidcursor;
322 } else if (!subport) {
323 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
324 (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
325 tidoff = dd->ipath_rcvtidcnt - tidcnt;
327 tid = tidcursor_fp(fp);
329 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
330 tidoff = tidcnt * (subport - 1);
332 tid = tidcursor_fp(fp);
335 /* make sure it all fits in port_tid_pg_list */
336 dev_info(&dd->pcidev->dev, "Process tried to allocate %u "
337 "TIDs, only trying max (%u)\n", cnt, tidcnt);
340 pagep = &((struct page **) pd->port_tid_pg_list)[tidoff];
341 tidlist = &((u16 *) &pagep[dd->ipath_rcvtidcnt])[tidoff];
343 memset(tidmap, 0, sizeof(tidmap));
344 /* before decrement; chip actual # */
346 tidbase = (u64 __iomem *) (((char __iomem *) dd->ipath_kregbase) +
347 dd->ipath_rcvtidbase +
348 porttid * sizeof(*tidbase));
350 ipath_cdbg(VERBOSE, "Port%u %u tids, cursor %u, tidbase %p\n",
351 pd->port_port, cnt, tid, tidbase);
353 /* virtual address of first page in transfer */
354 vaddr = ti->tidvaddr;
355 if (!access_ok(VERIFY_WRITE, (void __user *) vaddr,
357 ipath_dbg("Fail vaddr %p, %u pages, !access_ok\n",
362 ret = ipath_get_user_pages(vaddr, cnt, pagep);
365 ipath_dbg("Failed to lock addr %p, %u pages "
366 "(already locked)\n",
367 (void *) vaddr, cnt);
369 * for now, continue, and see what happens but with
370 * the new implementation, this should never happen,
371 * unless perhaps the user has mpin'ed the pages
372 * themselves (something we need to test)
376 dev_info(&dd->pcidev->dev,
377 "Failed to lock addr %p, %u pages: "
378 "errno %d\n", (void *) vaddr, cnt, -ret);
382 for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
383 for (; ntids--; tid++) {
386 if (!dd->ipath_pageshadow[porttid + tid])
391 * oops, wrapped all the way through their TIDs,
392 * and didn't have enough free; see comments at
395 ipath_dbg("Not enough free TIDs for %u pages "
396 "(index %d), failing\n", cnt, i);
397 i--; /* last tidlist[i] not filled in */
401 tidlist[i] = tid + tidoff;
402 ipath_cdbg(VERBOSE, "Updating idx %u to TID %u, "
403 "vaddr %lx\n", i, tid + tidoff, vaddr);
404 /* we "know" system pages and TID pages are same size */
405 dd->ipath_pageshadow[porttid + tid] = pagep[i];
406 dd->ipath_physshadow[porttid + tid] = ipath_map_page(
407 dd->pcidev, pagep[i], 0, PAGE_SIZE,
410 * don't need atomic or it's overhead
412 __set_bit(tid, tidmap);
413 physaddr = dd->ipath_physshadow[porttid + tid];
414 ipath_stats.sps_pagelocks++;
416 "TID %u, vaddr %lx, physaddr %llx pgp %p\n",
417 tid, vaddr, (unsigned long long) physaddr,
419 dd->ipath_f_put_tid(dd, &tidbase[tid], RCVHQ_RCV_TYPE_EXPECTED,
422 * don't check this tid in ipath_portshadow, since we
423 * just filled it in; start with the next one.
431 /* jump here if copy out of updated info failed... */
432 ipath_dbg("After failure (ret=%d), undo %d of %d entries\n",
434 /* same code that's in ipath_free_tid() */
435 limit = sizeof(tidmap) * BITS_PER_BYTE;
437 /* just in case size changes in future */
439 tid = find_first_bit((const unsigned long *)tidmap, limit);
440 for (; tid < limit; tid++) {
441 if (!test_bit(tid, tidmap))
443 if (dd->ipath_pageshadow[porttid + tid]) {
444 ipath_cdbg(VERBOSE, "Freeing TID %u\n",
446 dd->ipath_f_put_tid(dd, &tidbase[tid],
447 RCVHQ_RCV_TYPE_EXPECTED,
448 dd->ipath_tidinvalid);
449 pci_unmap_page(dd->pcidev,
450 dd->ipath_physshadow[porttid + tid],
451 PAGE_SIZE, PCI_DMA_FROMDEVICE);
452 dd->ipath_pageshadow[porttid + tid] = NULL;
453 ipath_stats.sps_pageunlocks++;
456 ipath_release_user_pages(pagep, cnt);
459 * Copy the updated array, with ipath_tid's filled in, back
460 * to user. Since we did the copy in already, this "should
461 * never fail" If it does, we have to clean up...
463 if (copy_to_user((void __user *)
464 (unsigned long) ti->tidlist,
465 tidlist, cnt * sizeof(*tidlist))) {
469 if (copy_to_user((void __user *) (unsigned long) ti->tidmap,
470 tidmap, sizeof tidmap)) {
476 if (!pd->port_subport_cnt)
477 pd->port_tidcursor = tid;
479 tidcursor_fp(fp) = tid;
484 ipath_dbg("Failed to map %u TID pages, failing with %d\n",
490 * ipath_tid_free - free a port TID
492 * @subport: the subport
495 * right now we are unlocking one page at a time, but since
496 * the intended use of this routine is for a single group of
497 * virtually contiguous pages, that should change to improve
498 * performance. We check that the TID is in range for this port
499 * but otherwise don't check validity; if user has an error and
500 * frees the wrong tid, it's only their own data that can thereby
501 * be corrupted. We do check that the TID was in use, for sanity
502 * We always use our idea of the saved address, not the address that
503 * they pass in to us.
506 static int ipath_tid_free(struct ipath_portdata *pd, unsigned subport,
507 const struct ipath_tid_info *ti)
510 u32 tid, porttid, cnt, limit, tidcnt;
511 struct ipath_devdata *dd = pd->port_dd;
512 u64 __iomem *tidbase;
513 unsigned long tidmap[8];
515 if (!dd->ipath_pageshadow) {
520 if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap,
526 porttid = pd->port_port * dd->ipath_rcvtidcnt;
527 if (!pd->port_subport_cnt)
528 tidcnt = dd->ipath_rcvtidcnt;
530 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
531 (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
532 porttid += dd->ipath_rcvtidcnt - tidcnt;
534 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
535 porttid += tidcnt * (subport - 1);
537 tidbase = (u64 __iomem *) ((char __iomem *)(dd->ipath_kregbase) +
538 dd->ipath_rcvtidbase +
539 porttid * sizeof(*tidbase));
541 limit = sizeof(tidmap) * BITS_PER_BYTE;
543 /* just in case size changes in future */
545 tid = find_first_bit(tidmap, limit);
546 ipath_cdbg(VERBOSE, "Port%u free %u tids; first bit (max=%d) "
547 "set is %d, porttid %u\n", pd->port_port, ti->tidcnt,
548 limit, tid, porttid);
549 for (cnt = 0; tid < limit; tid++) {
551 * small optimization; if we detect a run of 3 or so without
552 * any set, use find_first_bit again. That's mainly to
553 * accelerate the case where we wrapped, so we have some at
554 * the beginning, and some at the end, and a big gap
557 if (!test_bit(tid, tidmap))
560 if (dd->ipath_pageshadow[porttid + tid]) {
562 p = dd->ipath_pageshadow[porttid + tid];
563 dd->ipath_pageshadow[porttid + tid] = NULL;
564 ipath_cdbg(VERBOSE, "PID %u freeing TID %u\n",
565 pid_nr(pd->port_pid), tid);
566 dd->ipath_f_put_tid(dd, &tidbase[tid],
567 RCVHQ_RCV_TYPE_EXPECTED,
568 dd->ipath_tidinvalid);
569 pci_unmap_page(dd->pcidev,
570 dd->ipath_physshadow[porttid + tid],
571 PAGE_SIZE, PCI_DMA_FROMDEVICE);
572 ipath_release_user_pages(&p, 1);
573 ipath_stats.sps_pageunlocks++;
575 ipath_dbg("Unused tid %u, ignoring\n", tid);
577 if (cnt != ti->tidcnt)
578 ipath_dbg("passed in tidcnt %d, only %d bits set in map\n",
582 ipath_dbg("Failed to unmap %u TID pages, failing with %d\n",
588 * ipath_set_part_key - set a partition key
592 * We can have up to 4 active at a time (other than the default, which is
593 * always allowed). This is somewhat tricky, since multiple ports may set
594 * the same key, so we reference count them, and clean up at exit. All 4
595 * partition keys are packed into a single infinipath register. It's an
596 * error for a process to set the same pkey multiple times. We provide no
597 * mechanism to de-allocate a pkey at this time, we may eventually need to
598 * do that. I've used the atomic operations, and no locking, and only make
599 * a single pass through what's available. This should be more than
600 * adequate for some time. I'll think about spinlocks or the like if and as
603 static int ipath_set_part_key(struct ipath_portdata *pd, u16 key)
605 struct ipath_devdata *dd = pd->port_dd;
606 int i, any = 0, pidx = -1;
607 u16 lkey = key & 0x7FFF;
610 if (lkey == (IPATH_DEFAULT_P_KEY & 0x7FFF)) {
611 /* nothing to do; this key always valid */
616 ipath_cdbg(VERBOSE, "p%u try to set pkey %hx, current keys "
617 "%hx:%x %hx:%x %hx:%x %hx:%x\n",
618 pd->port_port, key, dd->ipath_pkeys[0],
619 atomic_read(&dd->ipath_pkeyrefs[0]), dd->ipath_pkeys[1],
620 atomic_read(&dd->ipath_pkeyrefs[1]), dd->ipath_pkeys[2],
621 atomic_read(&dd->ipath_pkeyrefs[2]), dd->ipath_pkeys[3],
622 atomic_read(&dd->ipath_pkeyrefs[3]));
625 ipath_cdbg(PROC, "p%u tries to set key 0, not allowed\n",
632 * Set the full membership bit, because it has to be
633 * set in the register or the packet, and it seems
634 * cleaner to set in the register than to force all
635 * callers to set it. (see bug 4331)
639 for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
640 if (!pd->port_pkeys[i] && pidx == -1)
642 if (pd->port_pkeys[i] == key) {
643 ipath_cdbg(VERBOSE, "p%u tries to set same pkey "
644 "(%x) more than once\n",
651 ipath_dbg("All pkeys for port %u already in use, "
652 "can't set %x\n", pd->port_port, key);
656 for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
657 if (!dd->ipath_pkeys[i]) {
661 if (dd->ipath_pkeys[i] == key) {
662 atomic_t *pkrefs = &dd->ipath_pkeyrefs[i];
664 if (atomic_inc_return(pkrefs) > 1) {
665 pd->port_pkeys[pidx] = key;
666 ipath_cdbg(VERBOSE, "p%u set key %x "
667 "matches #%d, count now %d\n",
668 pd->port_port, key, i,
669 atomic_read(pkrefs));
674 * lost race, decrement count, catch below
677 ipath_cdbg(VERBOSE, "Lost race, count was "
678 "0, after dec, it's %d\n",
679 atomic_read(pkrefs));
683 if ((dd->ipath_pkeys[i] & 0x7FFF) == lkey) {
685 * It makes no sense to have both the limited and
686 * full membership PKEY set at the same time since
687 * the unlimited one will disable the limited one.
694 ipath_dbg("port %u, all pkeys already in use, "
695 "can't set %x\n", pd->port_port, key);
699 for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
700 if (!dd->ipath_pkeys[i] &&
701 atomic_inc_return(&dd->ipath_pkeyrefs[i]) == 1) {
704 /* for ipathstats, etc. */
705 ipath_stats.sps_pkeys[i] = lkey;
706 pd->port_pkeys[pidx] = dd->ipath_pkeys[i] = key;
708 (u64) dd->ipath_pkeys[0] |
709 ((u64) dd->ipath_pkeys[1] << 16) |
710 ((u64) dd->ipath_pkeys[2] << 32) |
711 ((u64) dd->ipath_pkeys[3] << 48);
712 ipath_cdbg(PROC, "p%u set key %x in #%d, "
713 "portidx %d, new pkey reg %llx\n",
714 pd->port_port, key, i, pidx,
715 (unsigned long long) pkey);
717 dd, dd->ipath_kregs->kr_partitionkey, pkey);
723 ipath_dbg("port %u, all pkeys already in use 2nd pass, "
724 "can't set %x\n", pd->port_port, key);
732 * ipath_manage_rcvq - manage a port's receive queue
734 * @subport: the subport
735 * @start_stop: action to carry out
737 * start_stop == 0 disables receive on the port, for use in queue
738 * overflow conditions. start_stop==1 re-enables, to be used to
739 * re-init the software copy of the head register
741 static int ipath_manage_rcvq(struct ipath_portdata *pd, unsigned subport,
744 struct ipath_devdata *dd = pd->port_dd;
746 ipath_cdbg(PROC, "%sabling rcv for unit %u port %u:%u\n",
747 start_stop ? "en" : "dis", dd->ipath_unit,
748 pd->port_port, subport);
751 /* atomically clear receive enable port. */
754 * On enable, force in-memory copy of the tail register to
755 * 0, so that protocol code doesn't have to worry about
756 * whether or not the chip has yet updated the in-memory
757 * copy or not on return from the system call. The chip
758 * always resets it's tail register back to 0 on a
759 * transition from disabled to enabled. This could cause a
760 * problem if software was broken, and did the enable w/o
761 * the disable, but eventually the in-memory copy will be
762 * updated and correct itself, even in the face of software
765 if (pd->port_rcvhdrtail_kvaddr)
766 ipath_clear_rcvhdrtail(pd);
767 set_bit(dd->ipath_r_portenable_shift + pd->port_port,
770 clear_bit(dd->ipath_r_portenable_shift + pd->port_port,
772 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
774 /* now be sure chip saw it before we return */
775 ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
778 * And try to be sure that tail reg update has happened too.
779 * This should in theory interlock with the RXE changes to
780 * the tail register. Don't assign it to the tail register
781 * in memory copy, since we could overwrite an update by the
784 ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
786 /* always; new head should be equal to new tail; see above */
791 static void ipath_clean_part_key(struct ipath_portdata *pd,
792 struct ipath_devdata *dd)
794 int i, j, pchanged = 0;
797 /* for debugging only */
798 oldpkey = (u64) dd->ipath_pkeys[0] |
799 ((u64) dd->ipath_pkeys[1] << 16) |
800 ((u64) dd->ipath_pkeys[2] << 32) |
801 ((u64) dd->ipath_pkeys[3] << 48);
803 for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
804 if (!pd->port_pkeys[i])
806 ipath_cdbg(VERBOSE, "look for key[%d] %hx in pkeys\n", i,
808 for (j = 0; j < ARRAY_SIZE(dd->ipath_pkeys); j++) {
809 /* check for match independent of the global bit */
810 if ((dd->ipath_pkeys[j] & 0x7fff) !=
811 (pd->port_pkeys[i] & 0x7fff))
813 if (atomic_dec_and_test(&dd->ipath_pkeyrefs[j])) {
814 ipath_cdbg(VERBOSE, "p%u clear key "
817 pd->port_pkeys[i], j);
818 ipath_stats.sps_pkeys[j] =
819 dd->ipath_pkeys[j] = 0;
823 VERBOSE, "p%u key %x matches #%d, "
824 "but ref still %d\n", pd->port_port,
825 pd->port_pkeys[i], j,
826 atomic_read(&dd->ipath_pkeyrefs[j]));
829 pd->port_pkeys[i] = 0;
832 u64 pkey = (u64) dd->ipath_pkeys[0] |
833 ((u64) dd->ipath_pkeys[1] << 16) |
834 ((u64) dd->ipath_pkeys[2] << 32) |
835 ((u64) dd->ipath_pkeys[3] << 48);
836 ipath_cdbg(VERBOSE, "p%u old pkey reg %llx, "
837 "new pkey reg %llx\n", pd->port_port,
838 (unsigned long long) oldpkey,
839 (unsigned long long) pkey);
840 ipath_write_kreg(dd, dd->ipath_kregs->kr_partitionkey,
846 * Initialize the port data with the receive buffer sizes
847 * so this can be done while the master port is locked.
848 * Otherwise, there is a race with a slave opening the port
849 * and seeing these fields uninitialized.
851 static void init_user_egr_sizes(struct ipath_portdata *pd)
853 struct ipath_devdata *dd = pd->port_dd;
854 unsigned egrperchunk, egrcnt, size;
857 * to avoid wasting a lot of memory, we allocate 32KB chunks of
858 * physically contiguous memory, advance through it until used up
859 * and then allocate more. Of course, we need memory to store those
860 * extra pointers, now. Started out with 256KB, but under heavy
861 * memory pressure (creating large files and then copying them over
862 * NFS while doing lots of MPI jobs), we hit some allocation
863 * failures, even though we can sleep... (2.6.10) Still get
864 * failures at 64K. 32K is the lowest we can go without wasting
868 egrperchunk = size / dd->ipath_rcvegrbufsize;
869 egrcnt = dd->ipath_rcvegrcnt;
870 pd->port_rcvegrbuf_chunks = (egrcnt + egrperchunk - 1) / egrperchunk;
871 pd->port_rcvegrbufs_perchunk = egrperchunk;
872 pd->port_rcvegrbuf_size = size;
876 * ipath_create_user_egr - allocate eager TID buffers
877 * @pd: the port to allocate TID buffers for
879 * This routine is now quite different for user and kernel, because
880 * the kernel uses skb's, for the accelerated network performance
881 * This is the user port version
883 * Allocate the eager TID buffers and program them into infinipath
884 * They are no longer completely contiguous, we do multiple allocation
887 static int ipath_create_user_egr(struct ipath_portdata *pd)
889 struct ipath_devdata *dd = pd->port_dd;
890 unsigned e, egrcnt, egrperchunk, chunk, egrsize, egroff;
896 * GFP_USER, but without GFP_FS, so buffer cache can be
897 * coalesced (we hope); otherwise, even at order 4,
898 * heavy filesystem activity makes these fail, and we can
899 * use compound pages.
901 gfp_flags = __GFP_WAIT | __GFP_IO | __GFP_COMP;
903 egrcnt = dd->ipath_rcvegrcnt;
904 /* TID number offset for this port */
905 egroff = (pd->port_port - 1) * egrcnt + dd->ipath_p0_rcvegrcnt;
906 egrsize = dd->ipath_rcvegrbufsize;
907 ipath_cdbg(VERBOSE, "Allocating %d egr buffers, at egrtid "
908 "offset %x, egrsize %u\n", egrcnt, egroff, egrsize);
910 chunk = pd->port_rcvegrbuf_chunks;
911 egrperchunk = pd->port_rcvegrbufs_perchunk;
912 size = pd->port_rcvegrbuf_size;
913 pd->port_rcvegrbuf = kmalloc(chunk * sizeof(pd->port_rcvegrbuf[0]),
915 if (!pd->port_rcvegrbuf) {
919 pd->port_rcvegrbuf_phys =
920 kmalloc(chunk * sizeof(pd->port_rcvegrbuf_phys[0]),
922 if (!pd->port_rcvegrbuf_phys) {
926 for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
928 pd->port_rcvegrbuf[e] = dma_alloc_coherent(
929 &dd->pcidev->dev, size, &pd->port_rcvegrbuf_phys[e],
932 if (!pd->port_rcvegrbuf[e]) {
934 goto bail_rcvegrbuf_phys;
938 pd->port_rcvegr_phys = pd->port_rcvegrbuf_phys[0];
940 for (e = chunk = 0; chunk < pd->port_rcvegrbuf_chunks; chunk++) {
941 dma_addr_t pa = pd->port_rcvegrbuf_phys[chunk];
944 for (i = 0; e < egrcnt && i < egrperchunk; e++, i++) {
945 dd->ipath_f_put_tid(dd, e + egroff +
949 dd->ipath_rcvegrbase),
950 RCVHQ_RCV_TYPE_EAGER, pa);
953 cond_resched(); /* don't hog the cpu */
960 for (e = 0; e < pd->port_rcvegrbuf_chunks &&
961 pd->port_rcvegrbuf[e]; e++) {
962 dma_free_coherent(&dd->pcidev->dev, size,
963 pd->port_rcvegrbuf[e],
964 pd->port_rcvegrbuf_phys[e]);
967 kfree(pd->port_rcvegrbuf_phys);
968 pd->port_rcvegrbuf_phys = NULL;
970 kfree(pd->port_rcvegrbuf);
971 pd->port_rcvegrbuf = NULL;
977 /* common code for the mappings on dma_alloc_coherent mem */
978 static int ipath_mmap_mem(struct vm_area_struct *vma,
979 struct ipath_portdata *pd, unsigned len, int write_ok,
980 void *kvaddr, char *what)
982 struct ipath_devdata *dd = pd->port_dd;
986 if ((vma->vm_end - vma->vm_start) > len) {
987 dev_info(&dd->pcidev->dev,
988 "FAIL on %s: len %lx > %x\n", what,
989 vma->vm_end - vma->vm_start, len);
995 if (vma->vm_flags & VM_WRITE) {
996 dev_info(&dd->pcidev->dev,
997 "%s must be mapped readonly\n", what);
1002 /* don't allow them to later change with mprotect */
1003 vma->vm_flags &= ~VM_MAYWRITE;
1006 pfn = virt_to_phys(kvaddr) >> PAGE_SHIFT;
1007 ret = remap_pfn_range(vma, vma->vm_start, pfn,
1008 len, vma->vm_page_prot);
1010 dev_info(&dd->pcidev->dev, "%s port%u mmap of %lx, %x "
1011 "bytes r%c failed: %d\n", what, pd->port_port,
1012 pfn, len, write_ok?'w':'o', ret);
1014 ipath_cdbg(VERBOSE, "%s port%u mmaped %lx, %x bytes "
1015 "r%c\n", what, pd->port_port, pfn, len,
1021 static int mmap_ureg(struct vm_area_struct *vma, struct ipath_devdata *dd,
1028 * This is real hardware, so use io_remap. This is the mechanism
1029 * for the user process to update the head registers for their port
1032 if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) {
1033 dev_info(&dd->pcidev->dev, "FAIL mmap userreg: reqlen "
1034 "%lx > PAGE\n", vma->vm_end - vma->vm_start);
1037 phys = dd->ipath_physaddr + ureg;
1038 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1040 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1041 ret = io_remap_pfn_range(vma, vma->vm_start,
1043 vma->vm_end - vma->vm_start,
1049 static int mmap_piobufs(struct vm_area_struct *vma,
1050 struct ipath_devdata *dd,
1051 struct ipath_portdata *pd,
1052 unsigned piobufs, unsigned piocnt)
1058 * When we map the PIO buffers in the chip, we want to map them as
1059 * writeonly, no read possible. This prevents access to previous
1060 * process data, and catches users who might try to read the i/o
1061 * space due to a bug.
1063 if ((vma->vm_end - vma->vm_start) > (piocnt * dd->ipath_palign)) {
1064 dev_info(&dd->pcidev->dev, "FAIL mmap piobufs: "
1065 "reqlen %lx > PAGE\n",
1066 vma->vm_end - vma->vm_start);
1071 phys = dd->ipath_physaddr + piobufs;
1073 #if defined(__powerpc__)
1074 /* There isn't a generic way to specify writethrough mappings */
1075 pgprot_val(vma->vm_page_prot) |= _PAGE_NO_CACHE;
1076 pgprot_val(vma->vm_page_prot) |= _PAGE_WRITETHRU;
1077 pgprot_val(vma->vm_page_prot) &= ~_PAGE_GUARDED;
1081 * don't allow them to later change to readable with mprotect (for when
1082 * not initially mapped readable, as is normally the case)
1084 vma->vm_flags &= ~VM_MAYREAD;
1085 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1087 ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
1088 vma->vm_end - vma->vm_start,
1094 static int mmap_rcvegrbufs(struct vm_area_struct *vma,
1095 struct ipath_portdata *pd)
1097 struct ipath_devdata *dd = pd->port_dd;
1098 unsigned long start, size;
1099 size_t total_size, i;
1103 size = pd->port_rcvegrbuf_size;
1104 total_size = pd->port_rcvegrbuf_chunks * size;
1105 if ((vma->vm_end - vma->vm_start) > total_size) {
1106 dev_info(&dd->pcidev->dev, "FAIL on egr bufs: "
1107 "reqlen %lx > actual %lx\n",
1108 vma->vm_end - vma->vm_start,
1109 (unsigned long) total_size);
1114 if (vma->vm_flags & VM_WRITE) {
1115 dev_info(&dd->pcidev->dev, "Can't map eager buffers as "
1116 "writable (flags=%lx)\n", vma->vm_flags);
1120 /* don't allow them to later change to writeable with mprotect */
1121 vma->vm_flags &= ~VM_MAYWRITE;
1123 start = vma->vm_start;
1125 for (i = 0; i < pd->port_rcvegrbuf_chunks; i++, start += size) {
1126 pfn = virt_to_phys(pd->port_rcvegrbuf[i]) >> PAGE_SHIFT;
1127 ret = remap_pfn_range(vma, start, pfn, size,
1139 * ipath_file_vma_fault - handle a VMA page fault.
1141 static int ipath_file_vma_fault(struct vm_area_struct *vma,
1142 struct vm_fault *vmf)
1146 page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
1148 return VM_FAULT_SIGBUS;
1155 static const struct vm_operations_struct ipath_file_vm_ops = {
1156 .fault = ipath_file_vma_fault,
1159 static int mmap_kvaddr(struct vm_area_struct *vma, u64 pgaddr,
1160 struct ipath_portdata *pd, unsigned subport)
1163 struct ipath_devdata *dd;
1168 /* If the port is not shared, all addresses should be physical */
1169 if (!pd->port_subport_cnt)
1173 size = pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
1176 * Each process has all the subport uregbase, rcvhdrq, and
1177 * rcvegrbufs mmapped - as an array for all the processes,
1178 * and also separately for this process.
1180 if (pgaddr == cvt_kvaddr(pd->subport_uregbase)) {
1181 addr = pd->subport_uregbase;
1182 size = PAGE_SIZE * pd->port_subport_cnt;
1183 } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base)) {
1184 addr = pd->subport_rcvhdr_base;
1185 size = pd->port_rcvhdrq_size * pd->port_subport_cnt;
1186 } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf)) {
1187 addr = pd->subport_rcvegrbuf;
1188 size *= pd->port_subport_cnt;
1189 } else if (pgaddr == cvt_kvaddr(pd->subport_uregbase +
1190 PAGE_SIZE * subport)) {
1191 addr = pd->subport_uregbase + PAGE_SIZE * subport;
1193 } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base +
1194 pd->port_rcvhdrq_size * subport)) {
1195 addr = pd->subport_rcvhdr_base +
1196 pd->port_rcvhdrq_size * subport;
1197 size = pd->port_rcvhdrq_size;
1198 } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf +
1200 addr = pd->subport_rcvegrbuf + size * subport;
1201 /* rcvegrbufs are read-only on the slave */
1202 if (vma->vm_flags & VM_WRITE) {
1203 dev_info(&dd->pcidev->dev,
1204 "Can't map eager buffers as "
1205 "writable (flags=%lx)\n", vma->vm_flags);
1210 * Don't allow permission to later change to writeable
1213 vma->vm_flags &= ~VM_MAYWRITE;
1217 len = vma->vm_end - vma->vm_start;
1219 ipath_cdbg(MM, "FAIL: reqlen %lx > %zx\n", len, size);
1224 vma->vm_pgoff = (unsigned long) addr >> PAGE_SHIFT;
1225 vma->vm_ops = &ipath_file_vm_ops;
1226 vma->vm_flags |= VM_RESERVED | VM_DONTEXPAND;
1234 * ipath_mmap - mmap various structures into user space
1235 * @fp: the file pointer
1238 * We use this to have a shared buffer between the kernel and the user code
1239 * for the rcvhdr queue, egr buffers, and the per-port user regs and pio
1240 * buffers in the chip. We have the open and close entries so we can bump
1241 * the ref count and keep the driver from being unloaded while still mapped.
1243 static int ipath_mmap(struct file *fp, struct vm_area_struct *vma)
1245 struct ipath_portdata *pd;
1246 struct ipath_devdata *dd;
1248 unsigned piobufs, piocnt;
1259 * This is the ipath_do_user_init() code, mapping the shared buffers
1260 * into the user process. The address referred to by vm_pgoff is the
1261 * file offset passed via mmap(). For shared ports, this is the
1262 * kernel vmalloc() address of the pages to share with the master.
1263 * For non-shared or master ports, this is a physical address.
1264 * We only do one mmap for each space mapped.
1266 pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1269 * Check for 0 in case one of the allocations failed, but user
1270 * called mmap anyway.
1277 ipath_cdbg(MM, "pgaddr %llx vm_start=%lx len %lx port %u:%u:%u\n",
1278 (unsigned long long) pgaddr, vma->vm_start,
1279 vma->vm_end - vma->vm_start, dd->ipath_unit,
1280 pd->port_port, subport_fp(fp));
1283 * Physical addresses must fit in 40 bits for our hardware.
1284 * Check for kernel virtual addresses first, anything else must
1285 * match a HW or memory address.
1287 ret = mmap_kvaddr(vma, pgaddr, pd, subport_fp(fp));
1294 ureg = dd->ipath_uregbase + dd->ipath_ureg_align * pd->port_port;
1295 if (!pd->port_subport_cnt) {
1296 /* port is not shared */
1297 piocnt = pd->port_piocnt;
1298 piobufs = pd->port_piobufs;
1299 } else if (!subport_fp(fp)) {
1300 /* caller is the master */
1301 piocnt = (pd->port_piocnt / pd->port_subport_cnt) +
1302 (pd->port_piocnt % pd->port_subport_cnt);
1303 piobufs = pd->port_piobufs +
1304 dd->ipath_palign * (pd->port_piocnt - piocnt);
1306 unsigned slave = subport_fp(fp) - 1;
1308 /* caller is a slave */
1309 piocnt = pd->port_piocnt / pd->port_subport_cnt;
1310 piobufs = pd->port_piobufs + dd->ipath_palign * piocnt * slave;
1314 ret = mmap_ureg(vma, dd, ureg);
1315 else if (pgaddr == piobufs)
1316 ret = mmap_piobufs(vma, dd, pd, piobufs, piocnt);
1317 else if (pgaddr == dd->ipath_pioavailregs_phys)
1318 /* in-memory copy of pioavail registers */
1319 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1320 (void *) dd->ipath_pioavailregs_dma,
1321 "pioavail registers");
1322 else if (pgaddr == pd->port_rcvegr_phys)
1323 ret = mmap_rcvegrbufs(vma, pd);
1324 else if (pgaddr == (u64) pd->port_rcvhdrq_phys)
1326 * The rcvhdrq itself; readonly except on HT (so have
1327 * to allow writable mapping), multiple pages, contiguous
1328 * from an i/o perspective.
1330 ret = ipath_mmap_mem(vma, pd, pd->port_rcvhdrq_size, 1,
1333 else if (pgaddr == (u64) pd->port_rcvhdrqtailaddr_phys)
1334 /* in-memory copy of rcvhdrq tail register */
1335 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1336 pd->port_rcvhdrtail_kvaddr,
1341 vma->vm_private_data = NULL;
1344 dev_info(&dd->pcidev->dev,
1345 "Failure %d on off %llx len %lx\n",
1346 -ret, (unsigned long long)pgaddr,
1347 vma->vm_end - vma->vm_start);
1352 static unsigned ipath_poll_hdrqfull(struct ipath_portdata *pd)
1354 unsigned pollflag = 0;
1356 if ((pd->poll_type & IPATH_POLL_TYPE_OVERFLOW) &&
1357 pd->port_hdrqfull != pd->port_hdrqfull_poll) {
1358 pollflag |= POLLIN | POLLRDNORM;
1359 pd->port_hdrqfull_poll = pd->port_hdrqfull;
1365 static unsigned int ipath_poll_urgent(struct ipath_portdata *pd,
1367 struct poll_table_struct *pt)
1369 unsigned pollflag = 0;
1370 struct ipath_devdata *dd;
1374 /* variable access in ipath_poll_hdrqfull() needs this */
1376 pollflag = ipath_poll_hdrqfull(pd);
1378 if (pd->port_urgent != pd->port_urgent_poll) {
1379 pollflag |= POLLIN | POLLRDNORM;
1380 pd->port_urgent_poll = pd->port_urgent;
1384 /* this saves a spin_lock/unlock in interrupt handler... */
1385 set_bit(IPATH_PORT_WAITING_URG, &pd->port_flag);
1386 /* flush waiting flag so don't miss an event... */
1388 poll_wait(fp, &pd->port_wait, pt);
1394 static unsigned int ipath_poll_next(struct ipath_portdata *pd,
1396 struct poll_table_struct *pt)
1400 unsigned pollflag = 0;
1401 struct ipath_devdata *dd;
1405 /* variable access in ipath_poll_hdrqfull() needs this */
1407 pollflag = ipath_poll_hdrqfull(pd);
1409 head = ipath_read_ureg32(dd, ur_rcvhdrhead, pd->port_port);
1410 if (pd->port_rcvhdrtail_kvaddr)
1411 tail = ipath_get_rcvhdrtail(pd);
1413 tail = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
1416 pollflag |= POLLIN | POLLRDNORM;
1418 /* this saves a spin_lock/unlock in interrupt handler */
1419 set_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
1420 /* flush waiting flag so we don't miss an event */
1423 set_bit(pd->port_port + dd->ipath_r_intravail_shift,
1424 &dd->ipath_rcvctrl);
1426 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1429 if (dd->ipath_rhdrhead_intr_off) /* arm rcv interrupt */
1430 ipath_write_ureg(dd, ur_rcvhdrhead,
1431 dd->ipath_rhdrhead_intr_off | head,
1434 poll_wait(fp, &pd->port_wait, pt);
1440 static unsigned int ipath_poll(struct file *fp,
1441 struct poll_table_struct *pt)
1443 struct ipath_portdata *pd;
1449 else if (pd->poll_type & IPATH_POLL_TYPE_URGENT)
1450 pollflag = ipath_poll_urgent(pd, fp, pt);
1452 pollflag = ipath_poll_next(pd, fp, pt);
1457 static int ipath_supports_subports(int user_swmajor, int user_swminor)
1459 /* no subport implementation prior to software version 1.3 */
1460 return (user_swmajor > 1) || (user_swminor >= 3);
1463 static int ipath_compatible_subports(int user_swmajor, int user_swminor)
1465 /* this code is written long-hand for clarity */
1466 if (IPATH_USER_SWMAJOR != user_swmajor) {
1467 /* no promise of compatibility if major mismatch */
1470 if (IPATH_USER_SWMAJOR == 1) {
1471 switch (IPATH_USER_SWMINOR) {
1475 /* no subport implementation so cannot be compatible */
1478 /* 3 is only compatible with itself */
1479 return user_swminor == 3;
1481 /* >= 4 are compatible (or are expected to be) */
1482 return user_swminor >= 4;
1485 /* make no promises yet for future major versions */
1489 static int init_subports(struct ipath_devdata *dd,
1490 struct ipath_portdata *pd,
1491 const struct ipath_user_info *uinfo)
1494 unsigned num_subports;
1498 * If the user is requesting zero subports,
1499 * skip the subport allocation.
1501 if (uinfo->spu_subport_cnt <= 0)
1504 /* Self-consistency check for ipath_compatible_subports() */
1505 if (ipath_supports_subports(IPATH_USER_SWMAJOR, IPATH_USER_SWMINOR) &&
1506 !ipath_compatible_subports(IPATH_USER_SWMAJOR,
1507 IPATH_USER_SWMINOR)) {
1508 dev_info(&dd->pcidev->dev,
1509 "Inconsistent ipath_compatible_subports()\n");
1513 /* Check for subport compatibility */
1514 if (!ipath_compatible_subports(uinfo->spu_userversion >> 16,
1515 uinfo->spu_userversion & 0xffff)) {
1516 dev_info(&dd->pcidev->dev,
1517 "Mismatched user version (%d.%d) and driver "
1518 "version (%d.%d) while port sharing. Ensure "
1519 "that driver and library are from the same "
1521 (int) (uinfo->spu_userversion >> 16),
1522 (int) (uinfo->spu_userversion & 0xffff),
1524 IPATH_USER_SWMINOR);
1527 if (uinfo->spu_subport_cnt > INFINIPATH_MAX_SUBPORT) {
1532 num_subports = uinfo->spu_subport_cnt;
1533 pd->subport_uregbase = vzalloc(PAGE_SIZE * num_subports);
1534 if (!pd->subport_uregbase) {
1538 /* Note: pd->port_rcvhdrq_size isn't initialized yet. */
1539 size = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
1540 sizeof(u32), PAGE_SIZE) * num_subports;
1541 pd->subport_rcvhdr_base = vzalloc(size);
1542 if (!pd->subport_rcvhdr_base) {
1547 pd->subport_rcvegrbuf = vzalloc(pd->port_rcvegrbuf_chunks *
1548 pd->port_rcvegrbuf_size *
1550 if (!pd->subport_rcvegrbuf) {
1555 pd->port_subport_cnt = uinfo->spu_subport_cnt;
1556 pd->port_subport_id = uinfo->spu_subport_id;
1557 pd->active_slaves = 1;
1558 set_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
1562 vfree(pd->subport_rcvhdr_base);
1564 vfree(pd->subport_uregbase);
1565 pd->subport_uregbase = NULL;
1570 static int try_alloc_port(struct ipath_devdata *dd, int port,
1572 const struct ipath_user_info *uinfo)
1574 struct ipath_portdata *pd;
1577 if (!(pd = dd->ipath_pd[port])) {
1580 pd = kzalloc(sizeof(struct ipath_portdata), GFP_KERNEL);
1583 * Allocate memory for use in ipath_tid_update() just once
1584 * at open, not per call. Reduces cost of expected send
1587 ptmp = kmalloc(dd->ipath_rcvtidcnt * sizeof(u16) +
1588 dd->ipath_rcvtidcnt * sizeof(struct page **),
1591 ipath_dev_err(dd, "Unable to allocate portdata "
1592 "memory, failing open\n");
1598 dd->ipath_pd[port] = pd;
1599 dd->ipath_pd[port]->port_port = port;
1600 dd->ipath_pd[port]->port_dd = dd;
1601 dd->ipath_pd[port]->port_tid_pg_list = ptmp;
1602 init_waitqueue_head(&dd->ipath_pd[port]->port_wait);
1604 if (!pd->port_cnt) {
1605 pd->userversion = uinfo->spu_userversion;
1606 init_user_egr_sizes(pd);
1607 if ((ret = init_subports(dd, pd, uinfo)) != 0)
1609 ipath_cdbg(PROC, "%s[%u] opened unit:port %u:%u\n",
1610 current->comm, current->pid, dd->ipath_unit,
1614 pd->port_pid = get_pid(task_pid(current));
1615 strlcpy(pd->port_comm, current->comm, sizeof(pd->port_comm));
1616 ipath_stats.sps_ports++;
1625 static inline int usable(struct ipath_devdata *dd)
1628 (dd->ipath_flags & IPATH_PRESENT) &&
1629 dd->ipath_kregbase &&
1631 !(dd->ipath_flags & (IPATH_LINKDOWN | IPATH_DISABLED
1635 static int find_free_port(int unit, struct file *fp,
1636 const struct ipath_user_info *uinfo)
1638 struct ipath_devdata *dd = ipath_lookup(unit);
1651 for (i = 1; i < dd->ipath_cfgports; i++) {
1652 ret = try_alloc_port(dd, i, fp, uinfo);
1662 static int find_best_unit(struct file *fp,
1663 const struct ipath_user_info *uinfo)
1665 int ret = 0, i, prefunit = -1, devmax;
1666 int maxofallports, npresent, nup;
1669 devmax = ipath_count_units(&npresent, &nup, &maxofallports);
1672 * This code is present to allow a knowledgeable person to
1673 * specify the layout of processes to processors before opening
1674 * this driver, and then we'll assign the process to the "closest"
1675 * InfiniPath chip to that processor (we assume reasonable connectivity,
1676 * for now). This code assumes that if affinity has been set
1677 * before this point, that at most one cpu is set; for now this
1678 * is reasonable. I check for both cpumask_empty() and cpumask_full(),
1679 * in case some kernel variant sets none of the bits when no
1680 * affinity is set. 2.6.11 and 12 kernels have all present
1681 * cpus set. Some day we'll have to fix it up further to handle
1682 * a cpu subset. This algorithm fails for two HT chips connected
1683 * in tunnel fashion. Eventually this needs real topology
1684 * information. There may be some issues with dual core numbering
1685 * as well. This needs more work prior to release.
1687 if (!cpumask_empty(¤t->cpus_allowed) &&
1688 !cpumask_full(¤t->cpus_allowed)) {
1689 int ncpus = num_online_cpus(), curcpu = -1, nset = 0;
1690 for (i = 0; i < ncpus; i++)
1691 if (cpumask_test_cpu(i, ¤t->cpus_allowed)) {
1692 ipath_cdbg(PROC, "%s[%u] affinity set for "
1693 "cpu %d/%d\n", current->comm,
1694 current->pid, i, ncpus);
1698 if (curcpu != -1 && nset != ncpus) {
1700 prefunit = curcpu / (ncpus / npresent);
1701 ipath_cdbg(PROC,"%s[%u] %d chips, %d cpus, "
1702 "%d cpus/chip, select unit %d\n",
1703 current->comm, current->pid,
1704 npresent, ncpus, ncpus / npresent,
1711 * user ports start at 1, kernel port is 0
1712 * For now, we do round-robin access across all chips
1716 devmax = prefunit + 1;
1718 for (i = 1; i < maxofallports; i++) {
1719 for (ndev = prefunit != -1 ? prefunit : 0; ndev < devmax;
1721 struct ipath_devdata *dd = ipath_lookup(ndev);
1724 continue; /* can't use this unit */
1725 if (i >= dd->ipath_cfgports)
1727 * Maxed out on users of this unit. Try
1731 ret = try_alloc_port(dd, i, fp, uinfo);
1740 ipath_dbg("No ports available (none initialized "
1744 /* if started above 0, retry from 0 */
1746 "%s[%u] no ports on prefunit "
1747 "%d, clear and re-check\n",
1748 current->comm, current->pid,
1750 devmax = ipath_count_units(NULL, NULL,
1756 ipath_dbg("No ports available\n");
1760 ipath_dbg("No boards found\n");
1767 static int find_shared_port(struct file *fp,
1768 const struct ipath_user_info *uinfo)
1770 int devmax, ndev, i;
1773 devmax = ipath_count_units(NULL, NULL, NULL);
1775 for (ndev = 0; ndev < devmax; ndev++) {
1776 struct ipath_devdata *dd = ipath_lookup(ndev);
1780 for (i = 1; i < dd->ipath_cfgports; i++) {
1781 struct ipath_portdata *pd = dd->ipath_pd[i];
1783 /* Skip ports which are not yet open */
1784 if (!pd || !pd->port_cnt)
1786 /* Skip port if it doesn't match the requested one */
1787 if (pd->port_subport_id != uinfo->spu_subport_id)
1789 /* Verify the sharing process matches the master */
1790 if (pd->port_subport_cnt != uinfo->spu_subport_cnt ||
1791 pd->userversion != uinfo->spu_userversion ||
1792 pd->port_cnt >= pd->port_subport_cnt) {
1797 subport_fp(fp) = pd->port_cnt++;
1798 pd->port_subpid[subport_fp(fp)] =
1799 get_pid(task_pid(current));
1800 tidcursor_fp(fp) = 0;
1801 pd->active_slaves |= 1 << subport_fp(fp);
1803 "%s[%u] %u sharing %s[%u] unit:port %u:%u\n",
1804 current->comm, current->pid,
1806 pd->port_comm, pid_nr(pd->port_pid),
1807 dd->ipath_unit, pd->port_port);
1817 static int ipath_open(struct inode *in, struct file *fp)
1819 /* The real work is performed later in ipath_assign_port() */
1820 fp->private_data = kzalloc(sizeof(struct ipath_filedata), GFP_KERNEL);
1821 return fp->private_data ? 0 : -ENOMEM;
1824 /* Get port early, so can set affinity prior to memory allocation */
1825 static int ipath_assign_port(struct file *fp,
1826 const struct ipath_user_info *uinfo)
1830 unsigned swmajor, swminor;
1832 /* Check to be sure we haven't already initialized this file */
1838 /* for now, if major version is different, bail */
1839 swmajor = uinfo->spu_userversion >> 16;
1840 if (swmajor != IPATH_USER_SWMAJOR) {
1841 ipath_dbg("User major version %d not same as driver "
1842 "major %d\n", uinfo->spu_userversion >> 16,
1843 IPATH_USER_SWMAJOR);
1848 swminor = uinfo->spu_userversion & 0xffff;
1849 if (swminor != IPATH_USER_SWMINOR)
1850 ipath_dbg("User minor version %d not same as driver "
1851 "minor %d\n", swminor, IPATH_USER_SWMINOR);
1853 mutex_lock(&ipath_mutex);
1855 if (ipath_compatible_subports(swmajor, swminor) &&
1856 uinfo->spu_subport_cnt &&
1857 (ret = find_shared_port(fp, uinfo))) {
1863 i_minor = iminor(fp->f_path.dentry->d_inode) - IPATH_USER_MINOR_BASE;
1864 ipath_cdbg(VERBOSE, "open on dev %lx (minor %d)\n",
1865 (long)fp->f_path.dentry->d_inode->i_rdev, i_minor);
1868 ret = find_free_port(i_minor - 1, fp, uinfo);
1870 ret = find_best_unit(fp, uinfo);
1874 struct ipath_filedata *fd = fp->private_data;
1875 const struct ipath_portdata *pd = fd->pd;
1876 const struct ipath_devdata *dd = pd->port_dd;
1878 fd->pq = ipath_user_sdma_queue_create(&dd->pcidev->dev,
1887 mutex_unlock(&ipath_mutex);
1894 static int ipath_do_user_init(struct file *fp,
1895 const struct ipath_user_info *uinfo)
1898 struct ipath_portdata *pd = port_fp(fp);
1899 struct ipath_devdata *dd;
1902 /* Subports don't need to initialize anything since master did it. */
1903 if (subport_fp(fp)) {
1904 ret = wait_event_interruptible(pd->port_wait,
1905 !test_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag));
1911 if (uinfo->spu_rcvhdrsize) {
1912 ret = ipath_setrcvhdrsize(dd, uinfo->spu_rcvhdrsize);
1917 /* for now we do nothing with rcvhdrcnt: uinfo->spu_rcvhdrcnt */
1919 /* some ports may get extra buffers, calculate that here */
1920 if (pd->port_port <= dd->ipath_ports_extrabuf)
1921 pd->port_piocnt = dd->ipath_pbufsport + 1;
1923 pd->port_piocnt = dd->ipath_pbufsport;
1925 /* for right now, kernel piobufs are at end, so port 1 is at 0 */
1926 if (pd->port_port <= dd->ipath_ports_extrabuf)
1927 pd->port_pio_base = (dd->ipath_pbufsport + 1)
1928 * (pd->port_port - 1);
1930 pd->port_pio_base = dd->ipath_ports_extrabuf +
1931 dd->ipath_pbufsport * (pd->port_port - 1);
1932 pd->port_piobufs = dd->ipath_piobufbase +
1933 pd->port_pio_base * dd->ipath_palign;
1934 ipath_cdbg(VERBOSE, "piobuf base for port %u is 0x%x, piocnt %u,"
1935 " first pio %u\n", pd->port_port, pd->port_piobufs,
1936 pd->port_piocnt, pd->port_pio_base);
1937 ipath_chg_pioavailkernel(dd, pd->port_pio_base, pd->port_piocnt, 0);
1940 * Now allocate the rcvhdr Q and eager TIDs; skip the TID
1941 * array for time being. If pd->port_port > chip-supported,
1942 * we need to do extra stuff here to handle by handling overflow
1943 * through port 0, someday
1945 ret = ipath_create_rcvhdrq(dd, pd);
1947 ret = ipath_create_user_egr(pd);
1952 * set the eager head register for this port to the current values
1953 * of the tail pointers, since we don't know if they were
1954 * updated on last use of the port.
1956 head32 = ipath_read_ureg32(dd, ur_rcvegrindextail, pd->port_port);
1957 ipath_write_ureg(dd, ur_rcvegrindexhead, head32, pd->port_port);
1958 pd->port_lastrcvhdrqtail = -1;
1959 ipath_cdbg(VERBOSE, "Wrote port%d egrhead %x from tail regs\n",
1960 pd->port_port, head32);
1961 pd->port_tidcursor = 0; /* start at beginning after open */
1963 /* initialize poll variables... */
1964 pd->port_urgent = 0;
1965 pd->port_urgent_poll = 0;
1966 pd->port_hdrqfull_poll = pd->port_hdrqfull;
1969 * Now enable the port for receive.
1970 * For chips that are set to DMA the tail register to memory
1971 * when they change (and when the update bit transitions from
1972 * 0 to 1. So for those chips, we turn it off and then back on.
1973 * This will (very briefly) affect any other open ports, but the
1974 * duration is very short, and therefore isn't an issue. We
1975 * explicitly set the in-memory tail copy to 0 beforehand, so we
1976 * don't have to wait to be sure the DMA update has happened
1977 * (chip resets head/tail to 0 on transition to enable).
1979 set_bit(dd->ipath_r_portenable_shift + pd->port_port,
1980 &dd->ipath_rcvctrl);
1981 if (!(dd->ipath_flags & IPATH_NODMA_RTAIL)) {
1982 if (pd->port_rcvhdrtail_kvaddr)
1983 ipath_clear_rcvhdrtail(pd);
1984 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1986 ~(1ULL << dd->ipath_r_tailupd_shift));
1988 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1990 /* Notify any waiting slaves */
1991 if (pd->port_subport_cnt) {
1992 clear_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
1993 wake_up(&pd->port_wait);
2000 * unlock_exptid - unlock any expected TID entries port still had in use
2003 * We don't actually update the chip here, because we do a bulk update
2004 * below, using ipath_f_clear_tids.
2006 static void unlock_expected_tids(struct ipath_portdata *pd)
2008 struct ipath_devdata *dd = pd->port_dd;
2009 int port_tidbase = pd->port_port * dd->ipath_rcvtidcnt;
2010 int i, cnt = 0, maxtid = port_tidbase + dd->ipath_rcvtidcnt;
2012 ipath_cdbg(VERBOSE, "Port %u unlocking any locked expTID pages\n",
2014 for (i = port_tidbase; i < maxtid; i++) {
2015 struct page *ps = dd->ipath_pageshadow[i];
2020 dd->ipath_pageshadow[i] = NULL;
2021 pci_unmap_page(dd->pcidev, dd->ipath_physshadow[i],
2022 PAGE_SIZE, PCI_DMA_FROMDEVICE);
2023 ipath_release_user_pages_on_close(&ps, 1);
2025 ipath_stats.sps_pageunlocks++;
2028 ipath_cdbg(VERBOSE, "Port %u locked %u expTID entries\n",
2029 pd->port_port, cnt);
2031 if (ipath_stats.sps_pagelocks || ipath_stats.sps_pageunlocks)
2032 ipath_cdbg(VERBOSE, "%llu pages locked, %llu unlocked\n",
2033 (unsigned long long) ipath_stats.sps_pagelocks,
2034 (unsigned long long)
2035 ipath_stats.sps_pageunlocks);
2038 static int ipath_close(struct inode *in, struct file *fp)
2041 struct ipath_filedata *fd;
2042 struct ipath_portdata *pd;
2043 struct ipath_devdata *dd;
2044 unsigned long flags;
2048 ipath_cdbg(VERBOSE, "close on dev %lx, private data %p\n",
2049 (long)in->i_rdev, fp->private_data);
2051 mutex_lock(&ipath_mutex);
2053 fd = fp->private_data;
2054 fp->private_data = NULL;
2057 mutex_unlock(&ipath_mutex);
2063 /* drain user sdma queue */
2064 ipath_user_sdma_queue_drain(dd, fd->pq);
2065 ipath_user_sdma_queue_destroy(fd->pq);
2067 if (--pd->port_cnt) {
2069 * XXX If the master closes the port before the slave(s),
2070 * revoke the mmap for the eager receive queue so
2071 * the slave(s) don't wait for receive data forever.
2073 pd->active_slaves &= ~(1 << fd->subport);
2074 put_pid(pd->port_subpid[fd->subport]);
2075 pd->port_subpid[fd->subport] = NULL;
2076 mutex_unlock(&ipath_mutex);
2079 /* early; no interrupt users after this */
2080 spin_lock_irqsave(&dd->ipath_uctxt_lock, flags);
2081 port = pd->port_port;
2082 dd->ipath_pd[port] = NULL;
2084 pd->port_pid = NULL;
2085 spin_unlock_irqrestore(&dd->ipath_uctxt_lock, flags);
2087 if (pd->port_rcvwait_to || pd->port_piowait_to
2088 || pd->port_rcvnowait || pd->port_pionowait) {
2089 ipath_cdbg(VERBOSE, "port%u, %u rcv, %u pio wait timeo; "
2090 "%u rcv %u, pio already\n",
2091 pd->port_port, pd->port_rcvwait_to,
2092 pd->port_piowait_to, pd->port_rcvnowait,
2093 pd->port_pionowait);
2094 pd->port_rcvwait_to = pd->port_piowait_to =
2095 pd->port_rcvnowait = pd->port_pionowait = 0;
2097 if (pd->port_flag) {
2098 ipath_cdbg(PROC, "port %u port_flag set: 0x%lx\n",
2099 pd->port_port, pd->port_flag);
2103 if (dd->ipath_kregbase) {
2104 /* atomically clear receive enable port and intr avail. */
2105 clear_bit(dd->ipath_r_portenable_shift + port,
2106 &dd->ipath_rcvctrl);
2107 clear_bit(pd->port_port + dd->ipath_r_intravail_shift,
2108 &dd->ipath_rcvctrl);
2109 ipath_write_kreg( dd, dd->ipath_kregs->kr_rcvctrl,
2111 /* and read back from chip to be sure that nothing
2112 * else is in flight when we do the rest */
2113 (void)ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
2115 /* clean up the pkeys for this port user */
2116 ipath_clean_part_key(pd, dd);
2118 * be paranoid, and never write 0's to these, just use an
2119 * unused part of the port 0 tail page. Of course,
2120 * rcvhdraddr points to a large chunk of memory, so this
2121 * could still trash things, but at least it won't trash
2122 * page 0, and by disabling the port, it should stop "soon",
2123 * even if a packet or two is in already in flight after we
2124 * disabled the port.
2126 ipath_write_kreg_port(dd,
2127 dd->ipath_kregs->kr_rcvhdrtailaddr, port,
2128 dd->ipath_dummy_hdrq_phys);
2129 ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdraddr,
2130 pd->port_port, dd->ipath_dummy_hdrq_phys);
2132 ipath_disarm_piobufs(dd, pd->port_pio_base, pd->port_piocnt);
2133 ipath_chg_pioavailkernel(dd, pd->port_pio_base,
2134 pd->port_piocnt, 1);
2136 dd->ipath_f_clear_tids(dd, pd->port_port);
2138 if (dd->ipath_pageshadow)
2139 unlock_expected_tids(pd);
2140 ipath_stats.sps_ports--;
2141 ipath_cdbg(PROC, "%s[%u] closed port %u:%u\n",
2142 pd->port_comm, pid_nr(pid),
2143 dd->ipath_unit, port);
2147 mutex_unlock(&ipath_mutex);
2148 ipath_free_pddata(dd, pd); /* after releasing the mutex */
2155 static int ipath_port_info(struct ipath_portdata *pd, u16 subport,
2156 struct ipath_port_info __user *uinfo)
2158 struct ipath_port_info info;
2163 (void) ipath_count_units(NULL, &nup, NULL);
2164 info.num_active = nup;
2165 info.unit = pd->port_dd->ipath_unit;
2166 info.port = pd->port_port;
2167 info.subport = subport;
2168 /* Don't return new fields if old library opened the port. */
2169 if (ipath_supports_subports(pd->userversion >> 16,
2170 pd->userversion & 0xffff)) {
2171 /* Number of user ports available for this device. */
2172 info.num_ports = pd->port_dd->ipath_cfgports - 1;
2173 info.num_subports = pd->port_subport_cnt;
2176 sz = sizeof(info) - 2 * sizeof(u16);
2178 if (copy_to_user(uinfo, &info, sz)) {
2188 static int ipath_get_slave_info(struct ipath_portdata *pd,
2189 void __user *slave_mask_addr)
2193 if (copy_to_user(slave_mask_addr, &pd->active_slaves, sizeof(u32)))
2198 static int ipath_sdma_get_inflight(struct ipath_user_sdma_queue *pq,
2199 u32 __user *inflightp)
2201 const u32 val = ipath_user_sdma_inflight_counter(pq);
2203 if (put_user(val, inflightp))
2209 static int ipath_sdma_get_complete(struct ipath_devdata *dd,
2210 struct ipath_user_sdma_queue *pq,
2211 u32 __user *completep)
2216 err = ipath_user_sdma_make_progress(dd, pq);
2220 val = ipath_user_sdma_complete_counter(pq);
2221 if (put_user(val, completep))
2227 static ssize_t ipath_write(struct file *fp, const char __user *data,
2228 size_t count, loff_t *off)
2230 const struct ipath_cmd __user *ucmd;
2231 struct ipath_portdata *pd;
2232 const void __user *src;
2233 size_t consumed, copy;
2234 struct ipath_cmd cmd;
2238 if (count < sizeof(cmd.type)) {
2243 ucmd = (const struct ipath_cmd __user *) data;
2245 if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) {
2250 consumed = sizeof(cmd.type);
2253 case IPATH_CMD_ASSIGN_PORT:
2254 case __IPATH_CMD_USER_INIT:
2255 case IPATH_CMD_USER_INIT:
2256 copy = sizeof(cmd.cmd.user_info);
2257 dest = &cmd.cmd.user_info;
2258 src = &ucmd->cmd.user_info;
2260 case IPATH_CMD_RECV_CTRL:
2261 copy = sizeof(cmd.cmd.recv_ctrl);
2262 dest = &cmd.cmd.recv_ctrl;
2263 src = &ucmd->cmd.recv_ctrl;
2265 case IPATH_CMD_PORT_INFO:
2266 copy = sizeof(cmd.cmd.port_info);
2267 dest = &cmd.cmd.port_info;
2268 src = &ucmd->cmd.port_info;
2270 case IPATH_CMD_TID_UPDATE:
2271 case IPATH_CMD_TID_FREE:
2272 copy = sizeof(cmd.cmd.tid_info);
2273 dest = &cmd.cmd.tid_info;
2274 src = &ucmd->cmd.tid_info;
2276 case IPATH_CMD_SET_PART_KEY:
2277 copy = sizeof(cmd.cmd.part_key);
2278 dest = &cmd.cmd.part_key;
2279 src = &ucmd->cmd.part_key;
2281 case __IPATH_CMD_SLAVE_INFO:
2282 copy = sizeof(cmd.cmd.slave_mask_addr);
2283 dest = &cmd.cmd.slave_mask_addr;
2284 src = &ucmd->cmd.slave_mask_addr;
2286 case IPATH_CMD_PIOAVAILUPD: // force an update of PIOAvail reg
2291 case IPATH_CMD_POLL_TYPE:
2292 copy = sizeof(cmd.cmd.poll_type);
2293 dest = &cmd.cmd.poll_type;
2294 src = &ucmd->cmd.poll_type;
2296 case IPATH_CMD_ARMLAUNCH_CTRL:
2297 copy = sizeof(cmd.cmd.armlaunch_ctrl);
2298 dest = &cmd.cmd.armlaunch_ctrl;
2299 src = &ucmd->cmd.armlaunch_ctrl;
2301 case IPATH_CMD_SDMA_INFLIGHT:
2302 copy = sizeof(cmd.cmd.sdma_inflight);
2303 dest = &cmd.cmd.sdma_inflight;
2304 src = &ucmd->cmd.sdma_inflight;
2306 case IPATH_CMD_SDMA_COMPLETE:
2307 copy = sizeof(cmd.cmd.sdma_complete);
2308 dest = &cmd.cmd.sdma_complete;
2309 src = &ucmd->cmd.sdma_complete;
2317 if ((count - consumed) < copy) {
2322 if (copy_from_user(dest, src, copy)) {
2331 if (!pd && cmd.type != __IPATH_CMD_USER_INIT &&
2332 cmd.type != IPATH_CMD_ASSIGN_PORT) {
2338 case IPATH_CMD_ASSIGN_PORT:
2339 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2343 case __IPATH_CMD_USER_INIT:
2344 /* backwards compatibility, get port first */
2345 ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2348 /* and fall through to current version. */
2349 case IPATH_CMD_USER_INIT:
2350 ret = ipath_do_user_init(fp, &cmd.cmd.user_info);
2353 ret = ipath_get_base_info(
2354 fp, (void __user *) (unsigned long)
2355 cmd.cmd.user_info.spu_base_info,
2356 cmd.cmd.user_info.spu_base_info_size);
2358 case IPATH_CMD_RECV_CTRL:
2359 ret = ipath_manage_rcvq(pd, subport_fp(fp), cmd.cmd.recv_ctrl);
2361 case IPATH_CMD_PORT_INFO:
2362 ret = ipath_port_info(pd, subport_fp(fp),
2363 (struct ipath_port_info __user *)
2364 (unsigned long) cmd.cmd.port_info);
2366 case IPATH_CMD_TID_UPDATE:
2367 ret = ipath_tid_update(pd, fp, &cmd.cmd.tid_info);
2369 case IPATH_CMD_TID_FREE:
2370 ret = ipath_tid_free(pd, subport_fp(fp), &cmd.cmd.tid_info);
2372 case IPATH_CMD_SET_PART_KEY:
2373 ret = ipath_set_part_key(pd, cmd.cmd.part_key);
2375 case __IPATH_CMD_SLAVE_INFO:
2376 ret = ipath_get_slave_info(pd,
2377 (void __user *) (unsigned long)
2378 cmd.cmd.slave_mask_addr);
2380 case IPATH_CMD_PIOAVAILUPD:
2381 ipath_force_pio_avail_update(pd->port_dd);
2383 case IPATH_CMD_POLL_TYPE:
2384 pd->poll_type = cmd.cmd.poll_type;
2386 case IPATH_CMD_ARMLAUNCH_CTRL:
2387 if (cmd.cmd.armlaunch_ctrl)
2388 ipath_enable_armlaunch(pd->port_dd);
2390 ipath_disable_armlaunch(pd->port_dd);
2392 case IPATH_CMD_SDMA_INFLIGHT:
2393 ret = ipath_sdma_get_inflight(user_sdma_queue_fp(fp),
2394 (u32 __user *) (unsigned long)
2395 cmd.cmd.sdma_inflight);
2397 case IPATH_CMD_SDMA_COMPLETE:
2398 ret = ipath_sdma_get_complete(pd->port_dd,
2399 user_sdma_queue_fp(fp),
2400 (u32 __user *) (unsigned long)
2401 cmd.cmd.sdma_complete);
2412 static ssize_t ipath_writev(struct kiocb *iocb, const struct iovec *iov,
2413 unsigned long dim, loff_t off)
2415 struct file *filp = iocb->ki_filp;
2416 struct ipath_filedata *fp = filp->private_data;
2417 struct ipath_portdata *pd = port_fp(filp);
2418 struct ipath_user_sdma_queue *pq = fp->pq;
2423 return ipath_user_sdma_writev(pd->port_dd, pq, iov, dim);
2426 static struct class *ipath_class;
2428 static int init_cdev(int minor, char *name, const struct file_operations *fops,
2429 struct cdev **cdevp, struct device **devp)
2431 const dev_t dev = MKDEV(IPATH_MAJOR, minor);
2432 struct cdev *cdev = NULL;
2433 struct device *device = NULL;
2436 cdev = cdev_alloc();
2438 printk(KERN_ERR IPATH_DRV_NAME
2439 ": Could not allocate cdev for minor %d, %s\n",
2445 cdev->owner = THIS_MODULE;
2447 kobject_set_name(&cdev->kobj, name);
2449 ret = cdev_add(cdev, dev, 1);
2451 printk(KERN_ERR IPATH_DRV_NAME
2452 ": Could not add cdev for minor %d, %s (err %d)\n",
2457 device = device_create(ipath_class, NULL, dev, NULL, name);
2459 if (IS_ERR(device)) {
2460 ret = PTR_ERR(device);
2461 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2462 "device for minor %d, %s (err %d)\n",
2485 int ipath_cdev_init(int minor, char *name, const struct file_operations *fops,
2486 struct cdev **cdevp, struct device **devp)
2488 return init_cdev(minor, name, fops, cdevp, devp);
2491 static void cleanup_cdev(struct cdev **cdevp,
2492 struct device **devp)
2494 struct device *dev = *devp;
2497 device_unregister(dev);
2507 void ipath_cdev_cleanup(struct cdev **cdevp,
2508 struct device **devp)
2510 cleanup_cdev(cdevp, devp);
2513 static struct cdev *wildcard_cdev;
2514 static struct device *wildcard_dev;
2516 static const dev_t dev = MKDEV(IPATH_MAJOR, 0);
2518 static int user_init(void)
2522 ret = register_chrdev_region(dev, IPATH_NMINORS, IPATH_DRV_NAME);
2524 printk(KERN_ERR IPATH_DRV_NAME ": Could not register "
2525 "chrdev region (err %d)\n", -ret);
2529 ipath_class = class_create(THIS_MODULE, IPATH_DRV_NAME);
2531 if (IS_ERR(ipath_class)) {
2532 ret = PTR_ERR(ipath_class);
2533 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2534 "device class (err %d)\n", -ret);
2540 unregister_chrdev_region(dev, IPATH_NMINORS);
2545 static void user_cleanup(void)
2548 class_destroy(ipath_class);
2552 unregister_chrdev_region(dev, IPATH_NMINORS);
2555 static atomic_t user_count = ATOMIC_INIT(0);
2556 static atomic_t user_setup = ATOMIC_INIT(0);
2558 int ipath_user_add(struct ipath_devdata *dd)
2563 if (atomic_inc_return(&user_count) == 1) {
2566 ipath_dev_err(dd, "Unable to set up user support: "
2567 "error %d\n", -ret);
2570 ret = init_cdev(0, "ipath", &ipath_file_ops, &wildcard_cdev,
2573 ipath_dev_err(dd, "Could not create wildcard "
2574 "minor: error %d\n", -ret);
2578 atomic_set(&user_setup, 1);
2581 snprintf(name, sizeof(name), "ipath%d", dd->ipath_unit);
2583 ret = init_cdev(dd->ipath_unit + 1, name, &ipath_file_ops,
2584 &dd->user_cdev, &dd->user_dev);
2586 ipath_dev_err(dd, "Could not create user minor %d, %s\n",
2587 dd->ipath_unit + 1, name);
2597 void ipath_user_remove(struct ipath_devdata *dd)
2599 cleanup_cdev(&dd->user_cdev, &dd->user_dev);
2601 if (atomic_dec_return(&user_count) == 0) {
2602 if (atomic_read(&user_setup) == 0)
2605 cleanup_cdev(&wildcard_cdev, &wildcard_dev);
2608 atomic_set(&user_setup, 0);