2 * Copyright (c) 2006 QLogic, Inc. 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"
45 * mmap64 doesn't allow all 64 bits for 32-bit applications
46 * so only use the low 43 bits.
48 #define MMAP64_MASK 0x7FFFFFFFFFFUL
50 static int ipath_open(struct inode *, struct file *);
51 static int ipath_close(struct inode *, struct file *);
52 static ssize_t ipath_write(struct file *, const char __user *, size_t,
54 static unsigned int ipath_poll(struct file *, struct poll_table_struct *);
55 static int ipath_mmap(struct file *, struct vm_area_struct *);
57 static struct file_operations ipath_file_ops = {
61 .release = ipath_close,
66 static int ipath_get_base_info(struct file *fp,
67 void __user *ubase, size_t ubase_size)
69 struct ipath_portdata *pd = port_fp(fp);
71 struct ipath_base_info *kinfo = NULL;
72 struct ipath_devdata *dd = pd->port_dd;
77 subport_cnt = pd->port_subport_cnt;
84 master = !subport_fp(fp);
88 /* If port sharing is not requested, allow the old size structure */
90 sz -= 3 * sizeof(u64);
91 if (ubase_size < sz) {
93 "Base size %zu, need %zu (version mismatch?)\n",
99 kinfo = kzalloc(sizeof(*kinfo), GFP_KERNEL);
105 ret = dd->ipath_f_get_base_info(pd, kinfo);
109 kinfo->spi_rcvhdr_cnt = dd->ipath_rcvhdrcnt;
110 kinfo->spi_rcvhdrent_size = dd->ipath_rcvhdrentsize;
111 kinfo->spi_tidegrcnt = dd->ipath_rcvegrcnt;
112 kinfo->spi_rcv_egrbufsize = dd->ipath_rcvegrbufsize;
114 * have to mmap whole thing
116 kinfo->spi_rcv_egrbuftotlen =
117 pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
118 kinfo->spi_rcv_egrperchunk = pd->port_rcvegrbufs_perchunk;
119 kinfo->spi_rcv_egrchunksize = kinfo->spi_rcv_egrbuftotlen /
120 pd->port_rcvegrbuf_chunks;
121 kinfo->spi_tidcnt = dd->ipath_rcvtidcnt / subport_cnt;
123 kinfo->spi_tidcnt += dd->ipath_rcvtidcnt % subport_cnt;
125 * for this use, may be ipath_cfgports summed over all chips that
126 * are are configured and present
128 kinfo->spi_nports = dd->ipath_cfgports;
129 /* unit (chip/board) our port is on */
130 kinfo->spi_unit = dd->ipath_unit;
131 /* for now, only a single page */
132 kinfo->spi_tid_maxsize = PAGE_SIZE;
135 * Doing this per port, and based on the skip value, etc. This has
136 * to be the actual buffer size, since the protocol code treats it
139 * These have to be set to user addresses in the user code via mmap.
140 * These values are used on return to user code for the mmap target
141 * addresses only. For 32 bit, same 44 bit address problem, so use
142 * the physical address, not virtual. Before 2.6.11, using the
143 * page_address() macro worked, but in 2.6.11, even that returns the
144 * full 64 bit address (upper bits all 1's). So far, using the
145 * physical addresses (or chip offsets, for chip mapping) works, but
146 * no doubt some future kernel release will change that, and we'll be
147 * on to yet another method of dealing with this.
149 kinfo->spi_rcvhdr_base = (u64) pd->port_rcvhdrq_phys;
150 kinfo->spi_rcvhdr_tailaddr = (u64) pd->port_rcvhdrqtailaddr_phys;
151 kinfo->spi_rcv_egrbufs = (u64) pd->port_rcvegr_phys;
152 kinfo->spi_pioavailaddr = (u64) dd->ipath_pioavailregs_phys;
153 kinfo->spi_status = (u64) kinfo->spi_pioavailaddr +
154 (void *) dd->ipath_statusp -
155 (void *) dd->ipath_pioavailregs_dma;
157 kinfo->spi_piocnt = dd->ipath_pbufsport;
158 kinfo->spi_piobufbase = (u64) pd->port_piobufs;
159 kinfo->__spi_uregbase = (u64) dd->ipath_uregbase +
160 dd->ipath_palign * pd->port_port;
162 kinfo->spi_piocnt = (dd->ipath_pbufsport / subport_cnt) +
163 (dd->ipath_pbufsport % subport_cnt);
164 /* Master's PIO buffers are after all the slave's */
165 kinfo->spi_piobufbase = (u64) pd->port_piobufs +
167 (dd->ipath_pbufsport - kinfo->spi_piocnt);
168 kinfo->__spi_uregbase = (u64) dd->ipath_uregbase +
169 dd->ipath_palign * pd->port_port;
171 unsigned slave = subport_fp(fp) - 1;
173 kinfo->spi_piocnt = dd->ipath_pbufsport / subport_cnt;
174 kinfo->spi_piobufbase = (u64) pd->port_piobufs +
175 dd->ipath_palign * kinfo->spi_piocnt * slave;
176 kinfo->__spi_uregbase = ((u64) pd->subport_uregbase +
177 PAGE_SIZE * slave) & MMAP64_MASK;
179 kinfo->spi_rcvhdr_base = ((u64) pd->subport_rcvhdr_base +
180 pd->port_rcvhdrq_size * slave) & MMAP64_MASK;
181 kinfo->spi_rcvhdr_tailaddr =
182 (u64) pd->port_rcvhdrqtailaddr_phys & MMAP64_MASK;
183 kinfo->spi_rcv_egrbufs = ((u64) pd->subport_rcvegrbuf +
184 dd->ipath_rcvegrcnt * dd->ipath_rcvegrbufsize * slave) &
188 kinfo->spi_pioindex = (kinfo->spi_piobufbase - dd->ipath_piobufbase) /
190 kinfo->spi_pioalign = dd->ipath_palign;
192 kinfo->spi_qpair = IPATH_KD_QP;
193 kinfo->spi_piosize = dd->ipath_ibmaxlen;
194 kinfo->spi_mtu = dd->ipath_ibmaxlen; /* maxlen, not ibmtu */
195 kinfo->spi_port = pd->port_port;
196 kinfo->spi_subport = subport_fp(fp);
197 kinfo->spi_sw_version = IPATH_KERN_SWVERSION;
198 kinfo->spi_hw_version = dd->ipath_revision;
201 kinfo->spi_runtime_flags |= IPATH_RUNTIME_MASTER;
202 kinfo->spi_subport_uregbase =
203 (u64) pd->subport_uregbase & MMAP64_MASK;
204 kinfo->spi_subport_rcvegrbuf =
205 (u64) pd->subport_rcvegrbuf & MMAP64_MASK;
206 kinfo->spi_subport_rcvhdr_base =
207 (u64) pd->subport_rcvhdr_base & MMAP64_MASK;
208 ipath_cdbg(PROC, "port %u flags %x %llx %llx %llx\n",
210 kinfo->spi_runtime_flags,
211 kinfo->spi_subport_uregbase,
212 kinfo->spi_subport_rcvegrbuf,
213 kinfo->spi_subport_rcvhdr_base);
216 if (copy_to_user(ubase, kinfo, sizeof(*kinfo)))
225 * ipath_tid_update - update a port TID
227 * @fp: the ipath device file
228 * @ti: the TID information
230 * The new implementation as of Oct 2004 is that the driver assigns
231 * the tid and returns it to the caller. To make it easier to
232 * catch bugs, and to reduce search time, we keep a cursor for
233 * each port, walking the shadow tid array to find one that's not
236 * For now, if we can't allocate the full list, we fail, although
237 * in the long run, we'll allocate as many as we can, and the
238 * caller will deal with that by trying the remaining pages later.
239 * That means that when we fail, we have to mark the tids as not in
240 * use again, in our shadow copy.
242 * It's up to the caller to free the tids when they are done.
243 * We'll unlock the pages as they free them.
245 * Also, right now we are locking one page at a time, but since
246 * the intended use of this routine is for a single group of
247 * virtually contiguous pages, that should change to improve
250 static int ipath_tid_update(struct ipath_portdata *pd, struct file *fp,
251 const struct ipath_tid_info *ti)
254 u32 tid, porttid, cnt, i, tidcnt, tidoff;
256 struct ipath_devdata *dd = pd->port_dd;
259 u64 __iomem *tidbase;
260 unsigned long tidmap[8];
261 struct page **pagep = NULL;
262 unsigned subport = subport_fp(fp);
264 if (!dd->ipath_pageshadow) {
271 ipath_dbg("After copyin, tidcnt 0, tidlist %llx\n",
272 (unsigned long long) ti->tidlist);
274 * Should we treat as success? likely a bug
279 porttid = pd->port_port * dd->ipath_rcvtidcnt;
280 if (!pd->port_subport_cnt) {
281 tidcnt = dd->ipath_rcvtidcnt;
282 tid = pd->port_tidcursor;
284 } else if (!subport) {
285 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
286 (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
287 tidoff = dd->ipath_rcvtidcnt - tidcnt;
289 tid = tidcursor_fp(fp);
291 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
292 tidoff = tidcnt * (subport - 1);
294 tid = tidcursor_fp(fp);
297 /* make sure it all fits in port_tid_pg_list */
298 dev_info(&dd->pcidev->dev, "Process tried to allocate %u "
299 "TIDs, only trying max (%u)\n", cnt, tidcnt);
302 pagep = &((struct page **) pd->port_tid_pg_list)[tidoff];
303 tidlist = &((u16 *) &pagep[dd->ipath_rcvtidcnt])[tidoff];
305 memset(tidmap, 0, sizeof(tidmap));
306 /* before decrement; chip actual # */
308 tidbase = (u64 __iomem *) (((char __iomem *) dd->ipath_kregbase) +
309 dd->ipath_rcvtidbase +
310 porttid * sizeof(*tidbase));
312 ipath_cdbg(VERBOSE, "Port%u %u tids, cursor %u, tidbase %p\n",
313 pd->port_port, cnt, tid, tidbase);
315 /* virtual address of first page in transfer */
316 vaddr = ti->tidvaddr;
317 if (!access_ok(VERIFY_WRITE, (void __user *) vaddr,
319 ipath_dbg("Fail vaddr %p, %u pages, !access_ok\n",
324 ret = ipath_get_user_pages(vaddr, cnt, pagep);
327 ipath_dbg("Failed to lock addr %p, %u pages "
328 "(already locked)\n",
329 (void *) vaddr, cnt);
331 * for now, continue, and see what happens but with
332 * the new implementation, this should never happen,
333 * unless perhaps the user has mpin'ed the pages
334 * themselves (something we need to test)
338 dev_info(&dd->pcidev->dev,
339 "Failed to lock addr %p, %u pages: "
340 "errno %d\n", (void *) vaddr, cnt, -ret);
344 for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
345 for (; ntids--; tid++) {
348 if (!dd->ipath_pageshadow[porttid + tid])
353 * oops, wrapped all the way through their TIDs,
354 * and didn't have enough free; see comments at
357 ipath_dbg("Not enough free TIDs for %u pages "
358 "(index %d), failing\n", cnt, i);
359 i--; /* last tidlist[i] not filled in */
363 tidlist[i] = tid + tidoff;
364 ipath_cdbg(VERBOSE, "Updating idx %u to TID %u, "
365 "vaddr %lx\n", i, tid + tidoff, vaddr);
366 /* we "know" system pages and TID pages are same size */
367 dd->ipath_pageshadow[porttid + tid] = pagep[i];
369 * don't need atomic or it's overhead
371 __set_bit(tid, tidmap);
372 physaddr = page_to_phys(pagep[i]);
373 ipath_stats.sps_pagelocks++;
375 "TID %u, vaddr %lx, physaddr %llx pgp %p\n",
376 tid, vaddr, (unsigned long long) physaddr,
378 dd->ipath_f_put_tid(dd, &tidbase[tid], 1, physaddr);
380 * don't check this tid in ipath_portshadow, since we
381 * just filled it in; start with the next one.
389 /* jump here if copy out of updated info failed... */
390 ipath_dbg("After failure (ret=%d), undo %d of %d entries\n",
392 /* same code that's in ipath_free_tid() */
393 limit = sizeof(tidmap) * BITS_PER_BYTE;
395 /* just in case size changes in future */
397 tid = find_first_bit((const unsigned long *)tidmap, limit);
398 for (; tid < limit; tid++) {
399 if (!test_bit(tid, tidmap))
401 if (dd->ipath_pageshadow[porttid + tid]) {
402 ipath_cdbg(VERBOSE, "Freeing TID %u\n",
404 dd->ipath_f_put_tid(dd, &tidbase[tid], 1,
405 dd->ipath_tidinvalid);
406 dd->ipath_pageshadow[porttid + tid] = NULL;
407 ipath_stats.sps_pageunlocks++;
410 ipath_release_user_pages(pagep, cnt);
413 * Copy the updated array, with ipath_tid's filled in, back
414 * to user. Since we did the copy in already, this "should
415 * never fail" If it does, we have to clean up...
417 if (copy_to_user((void __user *)
418 (unsigned long) ti->tidlist,
419 tidlist, cnt * sizeof(*tidlist))) {
423 if (copy_to_user((void __user *) (unsigned long) ti->tidmap,
424 tidmap, sizeof tidmap)) {
430 if (!pd->port_subport_cnt)
431 pd->port_tidcursor = tid;
433 tidcursor_fp(fp) = tid;
438 ipath_dbg("Failed to map %u TID pages, failing with %d\n",
444 * ipath_tid_free - free a port TID
446 * @subport: the subport
449 * right now we are unlocking one page at a time, but since
450 * the intended use of this routine is for a single group of
451 * virtually contiguous pages, that should change to improve
452 * performance. We check that the TID is in range for this port
453 * but otherwise don't check validity; if user has an error and
454 * frees the wrong tid, it's only their own data that can thereby
455 * be corrupted. We do check that the TID was in use, for sanity
456 * We always use our idea of the saved address, not the address that
457 * they pass in to us.
460 static int ipath_tid_free(struct ipath_portdata *pd, unsigned subport,
461 const struct ipath_tid_info *ti)
464 u32 tid, porttid, cnt, limit, tidcnt;
465 struct ipath_devdata *dd = pd->port_dd;
466 u64 __iomem *tidbase;
467 unsigned long tidmap[8];
469 if (!dd->ipath_pageshadow) {
474 if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap,
480 porttid = pd->port_port * dd->ipath_rcvtidcnt;
481 if (!pd->port_subport_cnt)
482 tidcnt = dd->ipath_rcvtidcnt;
484 tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
485 (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
486 porttid += dd->ipath_rcvtidcnt - tidcnt;
488 tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
489 porttid += tidcnt * (subport - 1);
491 tidbase = (u64 __iomem *) ((char __iomem *)(dd->ipath_kregbase) +
492 dd->ipath_rcvtidbase +
493 porttid * sizeof(*tidbase));
495 limit = sizeof(tidmap) * BITS_PER_BYTE;
497 /* just in case size changes in future */
499 tid = find_first_bit(tidmap, limit);
500 ipath_cdbg(VERBOSE, "Port%u free %u tids; first bit (max=%d) "
501 "set is %d, porttid %u\n", pd->port_port, ti->tidcnt,
502 limit, tid, porttid);
503 for (cnt = 0; tid < limit; tid++) {
505 * small optimization; if we detect a run of 3 or so without
506 * any set, use find_first_bit again. That's mainly to
507 * accelerate the case where we wrapped, so we have some at
508 * the beginning, and some at the end, and a big gap
511 if (!test_bit(tid, tidmap))
514 if (dd->ipath_pageshadow[porttid + tid]) {
515 ipath_cdbg(VERBOSE, "PID %u freeing TID %u\n",
517 dd->ipath_f_put_tid(dd, &tidbase[tid], 1,
518 dd->ipath_tidinvalid);
519 ipath_release_user_pages(
520 &dd->ipath_pageshadow[porttid + tid], 1);
521 dd->ipath_pageshadow[porttid + tid] = NULL;
522 ipath_stats.sps_pageunlocks++;
524 ipath_dbg("Unused tid %u, ignoring\n", tid);
526 if (cnt != ti->tidcnt)
527 ipath_dbg("passed in tidcnt %d, only %d bits set in map\n",
531 ipath_dbg("Failed to unmap %u TID pages, failing with %d\n",
537 * ipath_set_part_key - set a partition key
541 * We can have up to 4 active at a time (other than the default, which is
542 * always allowed). This is somewhat tricky, since multiple ports may set
543 * the same key, so we reference count them, and clean up at exit. All 4
544 * partition keys are packed into a single infinipath register. It's an
545 * error for a process to set the same pkey multiple times. We provide no
546 * mechanism to de-allocate a pkey at this time, we may eventually need to
547 * do that. I've used the atomic operations, and no locking, and only make
548 * a single pass through what's available. This should be more than
549 * adequate for some time. I'll think about spinlocks or the like if and as
552 static int ipath_set_part_key(struct ipath_portdata *pd, u16 key)
554 struct ipath_devdata *dd = pd->port_dd;
555 int i, any = 0, pidx = -1;
556 u16 lkey = key & 0x7FFF;
559 if (lkey == (IPATH_DEFAULT_P_KEY & 0x7FFF)) {
560 /* nothing to do; this key always valid */
565 ipath_cdbg(VERBOSE, "p%u try to set pkey %hx, current keys "
566 "%hx:%x %hx:%x %hx:%x %hx:%x\n",
567 pd->port_port, key, dd->ipath_pkeys[0],
568 atomic_read(&dd->ipath_pkeyrefs[0]), dd->ipath_pkeys[1],
569 atomic_read(&dd->ipath_pkeyrefs[1]), dd->ipath_pkeys[2],
570 atomic_read(&dd->ipath_pkeyrefs[2]), dd->ipath_pkeys[3],
571 atomic_read(&dd->ipath_pkeyrefs[3]));
574 ipath_cdbg(PROC, "p%u tries to set key 0, not allowed\n",
581 * Set the full membership bit, because it has to be
582 * set in the register or the packet, and it seems
583 * cleaner to set in the register than to force all
584 * callers to set it. (see bug 4331)
588 for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
589 if (!pd->port_pkeys[i] && pidx == -1)
591 if (pd->port_pkeys[i] == key) {
592 ipath_cdbg(VERBOSE, "p%u tries to set same pkey "
593 "(%x) more than once\n",
600 ipath_dbg("All pkeys for port %u already in use, "
601 "can't set %x\n", pd->port_port, key);
605 for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
606 if (!dd->ipath_pkeys[i]) {
610 if (dd->ipath_pkeys[i] == key) {
611 atomic_t *pkrefs = &dd->ipath_pkeyrefs[i];
613 if (atomic_inc_return(pkrefs) > 1) {
614 pd->port_pkeys[pidx] = key;
615 ipath_cdbg(VERBOSE, "p%u set key %x "
616 "matches #%d, count now %d\n",
617 pd->port_port, key, i,
618 atomic_read(pkrefs));
623 * lost race, decrement count, catch below
626 ipath_cdbg(VERBOSE, "Lost race, count was "
627 "0, after dec, it's %d\n",
628 atomic_read(pkrefs));
632 if ((dd->ipath_pkeys[i] & 0x7FFF) == lkey) {
634 * It makes no sense to have both the limited and
635 * full membership PKEY set at the same time since
636 * the unlimited one will disable the limited one.
643 ipath_dbg("port %u, all pkeys already in use, "
644 "can't set %x\n", pd->port_port, key);
648 for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
649 if (!dd->ipath_pkeys[i] &&
650 atomic_inc_return(&dd->ipath_pkeyrefs[i]) == 1) {
653 /* for ipathstats, etc. */
654 ipath_stats.sps_pkeys[i] = lkey;
655 pd->port_pkeys[pidx] = dd->ipath_pkeys[i] = key;
657 (u64) dd->ipath_pkeys[0] |
658 ((u64) dd->ipath_pkeys[1] << 16) |
659 ((u64) dd->ipath_pkeys[2] << 32) |
660 ((u64) dd->ipath_pkeys[3] << 48);
661 ipath_cdbg(PROC, "p%u set key %x in #%d, "
662 "portidx %d, new pkey reg %llx\n",
663 pd->port_port, key, i, pidx,
664 (unsigned long long) pkey);
666 dd, dd->ipath_kregs->kr_partitionkey, pkey);
672 ipath_dbg("port %u, all pkeys already in use 2nd pass, "
673 "can't set %x\n", pd->port_port, key);
681 * ipath_manage_rcvq - manage a port's receive queue
683 * @subport: the subport
684 * @start_stop: action to carry out
686 * start_stop == 0 disables receive on the port, for use in queue
687 * overflow conditions. start_stop==1 re-enables, to be used to
688 * re-init the software copy of the head register
690 static int ipath_manage_rcvq(struct ipath_portdata *pd, unsigned subport,
693 struct ipath_devdata *dd = pd->port_dd;
696 ipath_cdbg(PROC, "%sabling rcv for unit %u port %u:%u\n",
697 start_stop ? "en" : "dis", dd->ipath_unit,
698 pd->port_port, subport);
701 /* atomically clear receive enable port. */
704 * On enable, force in-memory copy of the tail register to
705 * 0, so that protocol code doesn't have to worry about
706 * whether or not the chip has yet updated the in-memory
707 * copy or not on return from the system call. The chip
708 * always resets it's tail register back to 0 on a
709 * transition from disabled to enabled. This could cause a
710 * problem if software was broken, and did the enable w/o
711 * the disable, but eventually the in-memory copy will be
712 * updated and correct itself, even in the face of software
715 *pd->port_rcvhdrtail_kvaddr = 0;
716 set_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
719 clear_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
721 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
723 /* now be sure chip saw it before we return */
724 tval = ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
727 * And try to be sure that tail reg update has happened too.
728 * This should in theory interlock with the RXE changes to
729 * the tail register. Don't assign it to the tail register
730 * in memory copy, since we could overwrite an update by the
733 tval = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
735 /* always; new head should be equal to new tail; see above */
740 static void ipath_clean_part_key(struct ipath_portdata *pd,
741 struct ipath_devdata *dd)
743 int i, j, pchanged = 0;
746 /* for debugging only */
747 oldpkey = (u64) dd->ipath_pkeys[0] |
748 ((u64) dd->ipath_pkeys[1] << 16) |
749 ((u64) dd->ipath_pkeys[2] << 32) |
750 ((u64) dd->ipath_pkeys[3] << 48);
752 for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
753 if (!pd->port_pkeys[i])
755 ipath_cdbg(VERBOSE, "look for key[%d] %hx in pkeys\n", i,
757 for (j = 0; j < ARRAY_SIZE(dd->ipath_pkeys); j++) {
758 /* check for match independent of the global bit */
759 if ((dd->ipath_pkeys[j] & 0x7fff) !=
760 (pd->port_pkeys[i] & 0x7fff))
762 if (atomic_dec_and_test(&dd->ipath_pkeyrefs[j])) {
763 ipath_cdbg(VERBOSE, "p%u clear key "
766 pd->port_pkeys[i], j);
767 ipath_stats.sps_pkeys[j] =
768 dd->ipath_pkeys[j] = 0;
772 VERBOSE, "p%u key %x matches #%d, "
773 "but ref still %d\n", pd->port_port,
774 pd->port_pkeys[i], j,
775 atomic_read(&dd->ipath_pkeyrefs[j]));
778 pd->port_pkeys[i] = 0;
781 u64 pkey = (u64) dd->ipath_pkeys[0] |
782 ((u64) dd->ipath_pkeys[1] << 16) |
783 ((u64) dd->ipath_pkeys[2] << 32) |
784 ((u64) dd->ipath_pkeys[3] << 48);
785 ipath_cdbg(VERBOSE, "p%u old pkey reg %llx, "
786 "new pkey reg %llx\n", pd->port_port,
787 (unsigned long long) oldpkey,
788 (unsigned long long) pkey);
789 ipath_write_kreg(dd, dd->ipath_kregs->kr_partitionkey,
795 * Initialize the port data with the receive buffer sizes
796 * so this can be done while the master port is locked.
797 * Otherwise, there is a race with a slave opening the port
798 * and seeing these fields uninitialized.
800 static void init_user_egr_sizes(struct ipath_portdata *pd)
802 struct ipath_devdata *dd = pd->port_dd;
803 unsigned egrperchunk, egrcnt, size;
806 * to avoid wasting a lot of memory, we allocate 32KB chunks of
807 * physically contiguous memory, advance through it until used up
808 * and then allocate more. Of course, we need memory to store those
809 * extra pointers, now. Started out with 256KB, but under heavy
810 * memory pressure (creating large files and then copying them over
811 * NFS while doing lots of MPI jobs), we hit some allocation
812 * failures, even though we can sleep... (2.6.10) Still get
813 * failures at 64K. 32K is the lowest we can go without wasting
817 egrperchunk = size / dd->ipath_rcvegrbufsize;
818 egrcnt = dd->ipath_rcvegrcnt;
819 pd->port_rcvegrbuf_chunks = (egrcnt + egrperchunk - 1) / egrperchunk;
820 pd->port_rcvegrbufs_perchunk = egrperchunk;
821 pd->port_rcvegrbuf_size = size;
825 * ipath_create_user_egr - allocate eager TID buffers
826 * @pd: the port to allocate TID buffers for
828 * This routine is now quite different for user and kernel, because
829 * the kernel uses skb's, for the accelerated network performance
830 * This is the user port version
832 * Allocate the eager TID buffers and program them into infinipath
833 * They are no longer completely contiguous, we do multiple allocation
836 static int ipath_create_user_egr(struct ipath_portdata *pd)
838 struct ipath_devdata *dd = pd->port_dd;
839 unsigned e, egrcnt, egrperchunk, chunk, egrsize, egroff;
845 * GFP_USER, but without GFP_FS, so buffer cache can be
846 * coalesced (we hope); otherwise, even at order 4,
847 * heavy filesystem activity makes these fail, and we can
848 * use compound pages.
850 gfp_flags = __GFP_WAIT | __GFP_IO | __GFP_COMP;
852 egrcnt = dd->ipath_rcvegrcnt;
853 /* TID number offset for this port */
854 egroff = pd->port_port * egrcnt;
855 egrsize = dd->ipath_rcvegrbufsize;
856 ipath_cdbg(VERBOSE, "Allocating %d egr buffers, at egrtid "
857 "offset %x, egrsize %u\n", egrcnt, egroff, egrsize);
859 chunk = pd->port_rcvegrbuf_chunks;
860 egrperchunk = pd->port_rcvegrbufs_perchunk;
861 size = pd->port_rcvegrbuf_size;
862 pd->port_rcvegrbuf = kmalloc(chunk * sizeof(pd->port_rcvegrbuf[0]),
864 if (!pd->port_rcvegrbuf) {
868 pd->port_rcvegrbuf_phys =
869 kmalloc(chunk * sizeof(pd->port_rcvegrbuf_phys[0]),
871 if (!pd->port_rcvegrbuf_phys) {
875 for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
877 pd->port_rcvegrbuf[e] = dma_alloc_coherent(
878 &dd->pcidev->dev, size, &pd->port_rcvegrbuf_phys[e],
881 if (!pd->port_rcvegrbuf[e]) {
883 goto bail_rcvegrbuf_phys;
887 pd->port_rcvegr_phys = pd->port_rcvegrbuf_phys[0];
889 for (e = chunk = 0; chunk < pd->port_rcvegrbuf_chunks; chunk++) {
890 dma_addr_t pa = pd->port_rcvegrbuf_phys[chunk];
893 for (i = 0; e < egrcnt && i < egrperchunk; e++, i++) {
894 dd->ipath_f_put_tid(dd, e + egroff +
898 dd->ipath_rcvegrbase), 0, pa);
901 cond_resched(); /* don't hog the cpu */
908 for (e = 0; e < pd->port_rcvegrbuf_chunks &&
909 pd->port_rcvegrbuf[e]; e++) {
910 dma_free_coherent(&dd->pcidev->dev, size,
911 pd->port_rcvegrbuf[e],
912 pd->port_rcvegrbuf_phys[e]);
915 kfree(pd->port_rcvegrbuf_phys);
916 pd->port_rcvegrbuf_phys = NULL;
918 kfree(pd->port_rcvegrbuf);
919 pd->port_rcvegrbuf = NULL;
925 /* common code for the mappings on dma_alloc_coherent mem */
926 static int ipath_mmap_mem(struct vm_area_struct *vma,
927 struct ipath_portdata *pd, unsigned len,
928 int write_ok, dma_addr_t addr, char *what)
930 struct ipath_devdata *dd = pd->port_dd;
931 unsigned pfn = (unsigned long)addr >> PAGE_SHIFT;
934 if ((vma->vm_end - vma->vm_start) > len) {
935 dev_info(&dd->pcidev->dev,
936 "FAIL on %s: len %lx > %x\n", what,
937 vma->vm_end - vma->vm_start, len);
943 if (vma->vm_flags & VM_WRITE) {
944 dev_info(&dd->pcidev->dev,
945 "%s must be mapped readonly\n", what);
950 /* don't allow them to later change with mprotect */
951 vma->vm_flags &= ~VM_MAYWRITE;
954 ret = remap_pfn_range(vma, vma->vm_start, pfn,
955 len, vma->vm_page_prot);
957 dev_info(&dd->pcidev->dev,
958 "%s port%u mmap of %lx, %x bytes r%c failed: %d\n",
959 what, pd->port_port, (unsigned long)addr, len,
960 write_ok?'w':'o', ret);
962 ipath_cdbg(VERBOSE, "%s port%u mmaped %lx, %x bytes r%c\n",
963 what, pd->port_port, (unsigned long)addr, len,
969 static int mmap_ureg(struct vm_area_struct *vma, struct ipath_devdata *dd,
976 * This is real hardware, so use io_remap. This is the mechanism
977 * for the user process to update the head registers for their port
980 if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) {
981 dev_info(&dd->pcidev->dev, "FAIL mmap userreg: reqlen "
982 "%lx > PAGE\n", vma->vm_end - vma->vm_start);
985 phys = dd->ipath_physaddr + ureg;
986 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
988 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
989 ret = io_remap_pfn_range(vma, vma->vm_start,
991 vma->vm_end - vma->vm_start,
997 static int mmap_piobufs(struct vm_area_struct *vma,
998 struct ipath_devdata *dd,
999 struct ipath_portdata *pd,
1000 unsigned piobufs, unsigned piocnt)
1006 * When we map the PIO buffers in the chip, we want to map them as
1007 * writeonly, no read possible. This prevents access to previous
1008 * process data, and catches users who might try to read the i/o
1009 * space due to a bug.
1011 if ((vma->vm_end - vma->vm_start) > (piocnt * dd->ipath_palign)) {
1012 dev_info(&dd->pcidev->dev, "FAIL mmap piobufs: "
1013 "reqlen %lx > PAGE\n",
1014 vma->vm_end - vma->vm_start);
1019 phys = dd->ipath_physaddr + piobufs;
1022 * Don't mark this as non-cached, or we don't get the
1023 * write combining behavior we want on the PIO buffers!
1026 #if defined(__powerpc__)
1027 /* There isn't a generic way to specify writethrough mappings */
1028 pgprot_val(vma->vm_page_prot) |= _PAGE_NO_CACHE;
1029 pgprot_val(vma->vm_page_prot) |= _PAGE_WRITETHRU;
1030 pgprot_val(vma->vm_page_prot) &= ~_PAGE_GUARDED;
1034 * don't allow them to later change to readable with mprotect (for when
1035 * not initially mapped readable, as is normally the case)
1037 vma->vm_flags &= ~VM_MAYREAD;
1038 vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1040 ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
1041 vma->vm_end - vma->vm_start,
1047 static int mmap_rcvegrbufs(struct vm_area_struct *vma,
1048 struct ipath_portdata *pd)
1050 struct ipath_devdata *dd = pd->port_dd;
1051 unsigned long start, size;
1052 size_t total_size, i;
1056 size = pd->port_rcvegrbuf_size;
1057 total_size = pd->port_rcvegrbuf_chunks * size;
1058 if ((vma->vm_end - vma->vm_start) > total_size) {
1059 dev_info(&dd->pcidev->dev, "FAIL on egr bufs: "
1060 "reqlen %lx > actual %lx\n",
1061 vma->vm_end - vma->vm_start,
1062 (unsigned long) total_size);
1067 if (vma->vm_flags & VM_WRITE) {
1068 dev_info(&dd->pcidev->dev, "Can't map eager buffers as "
1069 "writable (flags=%lx)\n", vma->vm_flags);
1073 /* don't allow them to later change to writeable with mprotect */
1074 vma->vm_flags &= ~VM_MAYWRITE;
1076 start = vma->vm_start;
1077 phys = pd->port_rcvegrbuf_phys;
1079 for (i = 0; i < pd->port_rcvegrbuf_chunks; i++, start += size) {
1080 ret = remap_pfn_range(vma, start, phys[i] >> PAGE_SHIFT,
1081 size, vma->vm_page_prot);
1092 * ipath_file_vma_nopage - handle a VMA page fault.
1094 static struct page *ipath_file_vma_nopage(struct vm_area_struct *vma,
1095 unsigned long address, int *type)
1097 unsigned long offset = address - vma->vm_start;
1098 struct page *page = NOPAGE_SIGBUS;
1102 * Convert the vmalloc address into a struct page.
1104 pageptr = (void *)(offset + (vma->vm_pgoff << PAGE_SHIFT));
1105 page = vmalloc_to_page(pageptr);
1109 /* Increment the reference count. */
1112 *type = VM_FAULT_MINOR;
1117 static struct vm_operations_struct ipath_file_vm_ops = {
1118 .nopage = ipath_file_vma_nopage,
1121 static int mmap_kvaddr(struct vm_area_struct *vma, u64 pgaddr,
1122 struct ipath_portdata *pd, unsigned subport)
1125 struct ipath_devdata *dd;
1130 /* If the port is not shared, all addresses should be physical */
1131 if (!pd->port_subport_cnt) {
1137 size = pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
1140 * Master has all the slave uregbase, rcvhdrq, and
1141 * rcvegrbufs mmapped.
1144 unsigned num_slaves = pd->port_subport_cnt - 1;
1146 if (pgaddr == ((u64) pd->subport_uregbase & MMAP64_MASK)) {
1147 addr = pd->subport_uregbase;
1148 size = PAGE_SIZE * num_slaves;
1149 } else if (pgaddr == ((u64) pd->subport_rcvhdr_base &
1151 addr = pd->subport_rcvhdr_base;
1152 size = pd->port_rcvhdrq_size * num_slaves;
1153 } else if (pgaddr == ((u64) pd->subport_rcvegrbuf &
1155 addr = pd->subport_rcvegrbuf;
1161 } else if (pgaddr == (((u64) pd->subport_uregbase +
1162 PAGE_SIZE * (subport - 1)) & MMAP64_MASK)) {
1163 addr = pd->subport_uregbase + PAGE_SIZE * (subport - 1);
1165 } else if (pgaddr == (((u64) pd->subport_rcvhdr_base +
1166 pd->port_rcvhdrq_size * (subport - 1)) &
1168 addr = pd->subport_rcvhdr_base +
1169 pd->port_rcvhdrq_size * (subport - 1);
1170 size = pd->port_rcvhdrq_size;
1171 } else if (pgaddr == (((u64) pd->subport_rcvegrbuf +
1172 size * (subport - 1)) & MMAP64_MASK)) {
1173 addr = pd->subport_rcvegrbuf + size * (subport - 1);
1174 /* rcvegrbufs are read-only on the slave */
1175 if (vma->vm_flags & VM_WRITE) {
1176 dev_info(&dd->pcidev->dev,
1177 "Can't map eager buffers as "
1178 "writable (flags=%lx)\n", vma->vm_flags);
1183 * Don't allow permission to later change to writeable
1186 vma->vm_flags &= ~VM_MAYWRITE;
1191 len = vma->vm_end - vma->vm_start;
1193 ipath_cdbg(MM, "FAIL: reqlen %lx > %zx\n", len, size);
1198 vma->vm_pgoff = (unsigned long) addr >> PAGE_SHIFT;
1199 vma->vm_ops = &ipath_file_vm_ops;
1200 vma->vm_flags |= VM_RESERVED | VM_DONTEXPAND;
1208 * ipath_mmap - mmap various structures into user space
1209 * @fp: the file pointer
1212 * We use this to have a shared buffer between the kernel and the user code
1213 * for the rcvhdr queue, egr buffers, and the per-port user regs and pio
1214 * buffers in the chip. We have the open and close entries so we can bump
1215 * the ref count and keep the driver from being unloaded while still mapped.
1217 static int ipath_mmap(struct file *fp, struct vm_area_struct *vma)
1219 struct ipath_portdata *pd;
1220 struct ipath_devdata *dd;
1222 unsigned piobufs, piocnt;
1233 * This is the ipath_do_user_init() code, mapping the shared buffers
1234 * into the user process. The address referred to by vm_pgoff is the
1235 * file offset passed via mmap(). For shared ports, this is the
1236 * kernel vmalloc() address of the pages to share with the master.
1237 * For non-shared or master ports, this is a physical address.
1238 * We only do one mmap for each space mapped.
1240 pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1243 * Check for 0 in case one of the allocations failed, but user
1244 * called mmap anyway.
1251 ipath_cdbg(MM, "pgaddr %llx vm_start=%lx len %lx port %u:%u:%u\n",
1252 (unsigned long long) pgaddr, vma->vm_start,
1253 vma->vm_end - vma->vm_start, dd->ipath_unit,
1254 pd->port_port, subport_fp(fp));
1257 * Physical addresses must fit in 40 bits for our hardware.
1258 * Check for kernel virtual addresses first, anything else must
1259 * match a HW or memory address.
1261 if (pgaddr >= (1ULL<<40)) {
1262 ret = mmap_kvaddr(vma, pgaddr, pd, subport_fp(fp));
1266 if (!pd->port_subport_cnt) {
1267 /* port is not shared */
1268 ureg = dd->ipath_uregbase + dd->ipath_palign * pd->port_port;
1269 piocnt = dd->ipath_pbufsport;
1270 piobufs = pd->port_piobufs;
1271 } else if (!subport_fp(fp)) {
1272 /* caller is the master */
1273 ureg = dd->ipath_uregbase + dd->ipath_palign * pd->port_port;
1274 piocnt = (dd->ipath_pbufsport / pd->port_subport_cnt) +
1275 (dd->ipath_pbufsport % pd->port_subport_cnt);
1276 piobufs = pd->port_piobufs +
1277 dd->ipath_palign * (dd->ipath_pbufsport - piocnt);
1279 unsigned slave = subport_fp(fp) - 1;
1281 /* caller is a slave */
1283 piocnt = dd->ipath_pbufsport / pd->port_subport_cnt;
1284 piobufs = pd->port_piobufs + dd->ipath_palign * piocnt * slave;
1288 ret = mmap_ureg(vma, dd, ureg);
1289 else if (pgaddr == piobufs)
1290 ret = mmap_piobufs(vma, dd, pd, piobufs, piocnt);
1291 else if (pgaddr == dd->ipath_pioavailregs_phys)
1292 /* in-memory copy of pioavail registers */
1293 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1294 dd->ipath_pioavailregs_phys,
1295 "pioavail registers");
1296 else if (subport_fp(fp))
1297 /* Subports don't mmap the physical receive buffers */
1299 else if (pgaddr == pd->port_rcvegr_phys)
1300 ret = mmap_rcvegrbufs(vma, pd);
1301 else if (pgaddr == (u64) pd->port_rcvhdrq_phys)
1303 * The rcvhdrq itself; readonly except on HT (so have
1304 * to allow writable mapping), multiple pages, contiguous
1305 * from an i/o perspective.
1307 ret = ipath_mmap_mem(vma, pd, pd->port_rcvhdrq_size, 1,
1308 pd->port_rcvhdrq_phys,
1310 else if (pgaddr == (u64) pd->port_rcvhdrqtailaddr_phys)
1311 /* in-memory copy of rcvhdrq tail register */
1312 ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1313 pd->port_rcvhdrqtailaddr_phys,
1318 vma->vm_private_data = NULL;
1321 dev_info(&dd->pcidev->dev,
1322 "Failure %d on off %llx len %lx\n",
1323 -ret, (unsigned long long)pgaddr,
1324 vma->vm_end - vma->vm_start);
1329 static unsigned int ipath_poll(struct file *fp,
1330 struct poll_table_struct *pt)
1332 struct ipath_portdata *pd;
1335 unsigned pollflag = 0;
1336 struct ipath_devdata *dd;
1343 bit = pd->port_port + INFINIPATH_R_INTRAVAIL_SHIFT;
1344 set_bit(bit, &dd->ipath_rcvctrl);
1347 * Before blocking, make sure that head is still == tail,
1348 * reading from the chip, so we can be sure the interrupt
1349 * enable has made it to the chip. If not equal, disable
1350 * interrupt again and return immediately. This avoids races,
1351 * and the overhead of the chip read doesn't matter much at
1352 * this point, since we are waiting for something anyway.
1355 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1358 head = ipath_read_ureg32(dd, ur_rcvhdrhead, pd->port_port);
1359 tail = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
1362 set_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
1363 if (dd->ipath_rhdrhead_intr_off) /* arm rcv interrupt */
1364 (void)ipath_write_ureg(dd, ur_rcvhdrhead,
1365 dd->ipath_rhdrhead_intr_off
1366 | head, pd->port_port);
1367 poll_wait(fp, &pd->port_wait, pt);
1369 if (test_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag)) {
1370 /* timed out, no packets received */
1371 clear_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
1372 pd->port_rcvwait_to++;
1375 pollflag = POLLIN | POLLRDNORM;
1378 /* it's already happened; don't do wait_event overhead */
1379 pollflag = POLLIN | POLLRDNORM;
1380 pd->port_rcvnowait++;
1383 clear_bit(bit, &dd->ipath_rcvctrl);
1384 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1391 static int init_subports(struct ipath_devdata *dd,
1392 struct ipath_portdata *pd,
1393 const struct ipath_user_info *uinfo)
1396 unsigned num_slaves;
1399 /* Old user binaries don't know about subports */
1400 if ((uinfo->spu_userversion & 0xffff) != IPATH_USER_SWMINOR)
1403 * If the user is requesting zero or one port,
1404 * skip the subport allocation.
1406 if (uinfo->spu_subport_cnt <= 1)
1408 if (uinfo->spu_subport_cnt > 4) {
1413 num_slaves = uinfo->spu_subport_cnt - 1;
1414 pd->subport_uregbase = vmalloc(PAGE_SIZE * num_slaves);
1415 if (!pd->subport_uregbase) {
1419 /* Note: pd->port_rcvhdrq_size isn't initialized yet. */
1420 size = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
1421 sizeof(u32), PAGE_SIZE) * num_slaves;
1422 pd->subport_rcvhdr_base = vmalloc(size);
1423 if (!pd->subport_rcvhdr_base) {
1428 pd->subport_rcvegrbuf = vmalloc(pd->port_rcvegrbuf_chunks *
1429 pd->port_rcvegrbuf_size *
1431 if (!pd->subport_rcvegrbuf) {
1436 pd->port_subport_cnt = uinfo->spu_subport_cnt;
1437 pd->port_subport_id = uinfo->spu_subport_id;
1438 pd->active_slaves = 1;
1442 vfree(pd->subport_rcvhdr_base);
1444 vfree(pd->subport_uregbase);
1445 pd->subport_uregbase = NULL;
1450 static int try_alloc_port(struct ipath_devdata *dd, int port,
1452 const struct ipath_user_info *uinfo)
1454 struct ipath_portdata *pd;
1457 if (!(pd = dd->ipath_pd[port])) {
1460 pd = kzalloc(sizeof(struct ipath_portdata), GFP_KERNEL);
1463 * Allocate memory for use in ipath_tid_update() just once
1464 * at open, not per call. Reduces cost of expected send
1467 ptmp = kmalloc(dd->ipath_rcvtidcnt * sizeof(u16) +
1468 dd->ipath_rcvtidcnt * sizeof(struct page **),
1471 ipath_dev_err(dd, "Unable to allocate portdata "
1472 "memory, failing open\n");
1478 dd->ipath_pd[port] = pd;
1479 dd->ipath_pd[port]->port_port = port;
1480 dd->ipath_pd[port]->port_dd = dd;
1481 dd->ipath_pd[port]->port_tid_pg_list = ptmp;
1482 init_waitqueue_head(&dd->ipath_pd[port]->port_wait);
1484 if (!pd->port_cnt) {
1485 pd->userversion = uinfo->spu_userversion;
1486 init_user_egr_sizes(pd);
1487 if ((ret = init_subports(dd, pd, uinfo)) != 0)
1489 ipath_cdbg(PROC, "%s[%u] opened unit:port %u:%u\n",
1490 current->comm, current->pid, dd->ipath_unit,
1494 pd->port_pid = current->pid;
1495 strncpy(pd->port_comm, current->comm, sizeof(pd->port_comm));
1496 ipath_stats.sps_ports++;
1505 static inline int usable(struct ipath_devdata *dd)
1508 (dd->ipath_flags & IPATH_PRESENT) &&
1509 dd->ipath_kregbase &&
1511 !(dd->ipath_flags & (IPATH_LINKDOWN | IPATH_DISABLED
1515 static int find_free_port(int unit, struct file *fp,
1516 const struct ipath_user_info *uinfo)
1518 struct ipath_devdata *dd = ipath_lookup(unit);
1531 for (i = 1; i < dd->ipath_cfgports; i++) {
1532 ret = try_alloc_port(dd, i, fp, uinfo);
1542 static int find_best_unit(struct file *fp,
1543 const struct ipath_user_info *uinfo)
1545 int ret = 0, i, prefunit = -1, devmax;
1546 int maxofallports, npresent, nup;
1549 devmax = ipath_count_units(&npresent, &nup, &maxofallports);
1552 * This code is present to allow a knowledgeable person to
1553 * specify the layout of processes to processors before opening
1554 * this driver, and then we'll assign the process to the "closest"
1555 * InfiniPath chip to that processor (we assume reasonable connectivity,
1556 * for now). This code assumes that if affinity has been set
1557 * before this point, that at most one cpu is set; for now this
1558 * is reasonable. I check for both cpus_empty() and cpus_full(),
1559 * in case some kernel variant sets none of the bits when no
1560 * affinity is set. 2.6.11 and 12 kernels have all present
1561 * cpus set. Some day we'll have to fix it up further to handle
1562 * a cpu subset. This algorithm fails for two HT chips connected
1563 * in tunnel fashion. Eventually this needs real topology
1564 * information. There may be some issues with dual core numbering
1565 * as well. This needs more work prior to release.
1567 if (!cpus_empty(current->cpus_allowed) &&
1568 !cpus_full(current->cpus_allowed)) {
1569 int ncpus = num_online_cpus(), curcpu = -1;
1570 for (i = 0; i < ncpus; i++)
1571 if (cpu_isset(i, current->cpus_allowed)) {
1572 ipath_cdbg(PROC, "%s[%u] affinity set for "
1573 "cpu %d\n", current->comm,
1579 prefunit = curcpu / (ncpus / npresent);
1580 ipath_dbg("%s[%u] %d chips, %d cpus, "
1581 "%d cpus/chip, select unit %d\n",
1582 current->comm, current->pid,
1583 npresent, ncpus, ncpus / npresent,
1590 * user ports start at 1, kernel port is 0
1591 * For now, we do round-robin access across all chips
1595 devmax = prefunit + 1;
1597 for (i = 1; i < maxofallports; i++) {
1598 for (ndev = prefunit != -1 ? prefunit : 0; ndev < devmax;
1600 struct ipath_devdata *dd = ipath_lookup(ndev);
1603 continue; /* can't use this unit */
1604 if (i >= dd->ipath_cfgports)
1606 * Maxed out on users of this unit. Try
1610 ret = try_alloc_port(dd, i, fp, uinfo);
1619 ipath_dbg("No ports available (none initialized "
1623 /* if started above 0, retry from 0 */
1625 "%s[%u] no ports on prefunit "
1626 "%d, clear and re-check\n",
1627 current->comm, current->pid,
1629 devmax = ipath_count_units(NULL, NULL,
1635 ipath_dbg("No ports available\n");
1639 ipath_dbg("No boards found\n");
1646 static int find_shared_port(struct file *fp,
1647 const struct ipath_user_info *uinfo)
1649 int devmax, ndev, i;
1652 devmax = ipath_count_units(NULL, NULL, NULL);
1654 for (ndev = 0; ndev < devmax; ndev++) {
1655 struct ipath_devdata *dd = ipath_lookup(ndev);
1659 for (i = 1; i < dd->ipath_cfgports; i++) {
1660 struct ipath_portdata *pd = dd->ipath_pd[i];
1662 /* Skip ports which are not yet open */
1663 if (!pd || !pd->port_cnt)
1665 /* Skip port if it doesn't match the requested one */
1666 if (pd->port_subport_id != uinfo->spu_subport_id)
1668 /* Verify the sharing process matches the master */
1669 if (pd->port_subport_cnt != uinfo->spu_subport_cnt ||
1670 pd->userversion != uinfo->spu_userversion ||
1671 pd->port_cnt >= pd->port_subport_cnt) {
1676 subport_fp(fp) = pd->port_cnt++;
1677 tidcursor_fp(fp) = 0;
1678 pd->active_slaves |= 1 << subport_fp(fp);
1680 "%s[%u] %u sharing %s[%u] unit:port %u:%u\n",
1681 current->comm, current->pid,
1683 pd->port_comm, pd->port_pid,
1684 dd->ipath_unit, pd->port_port);
1694 static int ipath_open(struct inode *in, struct file *fp)
1696 /* The real work is performed later in ipath_do_user_init() */
1697 fp->private_data = kzalloc(sizeof(struct ipath_filedata), GFP_KERNEL);
1698 return fp->private_data ? 0 : -ENOMEM;
1701 static int ipath_do_user_init(struct file *fp,
1702 const struct ipath_user_info *uinfo)
1705 struct ipath_portdata *pd;
1706 struct ipath_devdata *dd;
1711 /* Check to be sure we haven't already initialized this file */
1717 /* for now, if major version is different, bail */
1718 if ((uinfo->spu_userversion >> 16) != IPATH_USER_SWMAJOR) {
1719 ipath_dbg("User major version %d not same as driver "
1720 "major %d\n", uinfo->spu_userversion >> 16,
1721 IPATH_USER_SWMAJOR);
1726 swminor = uinfo->spu_userversion & 0xffff;
1727 if (swminor != IPATH_USER_SWMINOR)
1728 ipath_dbg("User minor version %d not same as driver "
1729 "minor %d\n", swminor, IPATH_USER_SWMINOR);
1731 mutex_lock(&ipath_mutex);
1733 if (swminor == IPATH_USER_SWMINOR && uinfo->spu_subport_cnt &&
1734 (ret = find_shared_port(fp, uinfo))) {
1735 mutex_unlock(&ipath_mutex);
1741 i_minor = iminor(fp->f_dentry->d_inode) - IPATH_USER_MINOR_BASE;
1742 ipath_cdbg(VERBOSE, "open on dev %lx (minor %d)\n",
1743 (long)fp->f_dentry->d_inode->i_rdev, i_minor);
1746 ret = find_free_port(i_minor - 1, fp, uinfo);
1748 ret = find_best_unit(fp, uinfo);
1750 mutex_unlock(&ipath_mutex);
1758 if (uinfo->spu_rcvhdrsize) {
1759 ret = ipath_setrcvhdrsize(dd, uinfo->spu_rcvhdrsize);
1764 /* for now we do nothing with rcvhdrcnt: uinfo->spu_rcvhdrcnt */
1766 /* for right now, kernel piobufs are at end, so port 1 is at 0 */
1767 pd->port_piobufs = dd->ipath_piobufbase +
1768 dd->ipath_pbufsport * (pd->port_port - 1) * dd->ipath_palign;
1769 ipath_cdbg(VERBOSE, "Set base of piobufs for port %u to 0x%x\n",
1770 pd->port_port, pd->port_piobufs);
1773 * Now allocate the rcvhdr Q and eager TIDs; skip the TID
1774 * array for time being. If pd->port_port > chip-supported,
1775 * we need to do extra stuff here to handle by handling overflow
1776 * through port 0, someday
1778 ret = ipath_create_rcvhdrq(dd, pd);
1780 ret = ipath_create_user_egr(pd);
1785 * set the eager head register for this port to the current values
1786 * of the tail pointers, since we don't know if they were
1787 * updated on last use of the port.
1789 head32 = ipath_read_ureg32(dd, ur_rcvegrindextail, pd->port_port);
1790 ipath_write_ureg(dd, ur_rcvegrindexhead, head32, pd->port_port);
1791 dd->ipath_lastegrheads[pd->port_port] = -1;
1792 dd->ipath_lastrcvhdrqtails[pd->port_port] = -1;
1793 ipath_cdbg(VERBOSE, "Wrote port%d egrhead %x from tail regs\n",
1794 pd->port_port, head32);
1795 pd->port_tidcursor = 0; /* start at beginning after open */
1797 * now enable the port; the tail registers will be written to memory
1798 * by the chip as soon as it sees the write to
1799 * dd->ipath_kregs->kr_rcvctrl. The update only happens on
1800 * transition from 0 to 1, so clear it first, then set it as part of
1801 * enabling the port. This will (very briefly) affect any other
1802 * open ports, but it shouldn't be long enough to be an issue.
1803 * We explictly set the in-memory copy to 0 beforehand, so we don't
1804 * have to wait to be sure the DMA update has happened.
1806 *pd->port_rcvhdrtail_kvaddr = 0ULL;
1807 set_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port,
1808 &dd->ipath_rcvctrl);
1809 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1810 dd->ipath_rcvctrl & ~INFINIPATH_R_TAILUPD);
1811 ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1818 * unlock_exptid - unlock any expected TID entries port still had in use
1821 * We don't actually update the chip here, because we do a bulk update
1822 * below, using ipath_f_clear_tids.
1824 static void unlock_expected_tids(struct ipath_portdata *pd)
1826 struct ipath_devdata *dd = pd->port_dd;
1827 int port_tidbase = pd->port_port * dd->ipath_rcvtidcnt;
1828 int i, cnt = 0, maxtid = port_tidbase + dd->ipath_rcvtidcnt;
1830 ipath_cdbg(VERBOSE, "Port %u unlocking any locked expTID pages\n",
1832 for (i = port_tidbase; i < maxtid; i++) {
1833 if (!dd->ipath_pageshadow[i])
1836 ipath_release_user_pages_on_close(&dd->ipath_pageshadow[i],
1838 dd->ipath_pageshadow[i] = NULL;
1840 ipath_stats.sps_pageunlocks++;
1843 ipath_cdbg(VERBOSE, "Port %u locked %u expTID entries\n",
1844 pd->port_port, cnt);
1846 if (ipath_stats.sps_pagelocks || ipath_stats.sps_pageunlocks)
1847 ipath_cdbg(VERBOSE, "%llu pages locked, %llu unlocked\n",
1848 (unsigned long long) ipath_stats.sps_pagelocks,
1849 (unsigned long long)
1850 ipath_stats.sps_pageunlocks);
1853 static int ipath_close(struct inode *in, struct file *fp)
1856 struct ipath_filedata *fd;
1857 struct ipath_portdata *pd;
1858 struct ipath_devdata *dd;
1861 ipath_cdbg(VERBOSE, "close on dev %lx, private data %p\n",
1862 (long)in->i_rdev, fp->private_data);
1864 mutex_lock(&ipath_mutex);
1866 fd = (struct ipath_filedata *) fp->private_data;
1867 fp->private_data = NULL;
1870 mutex_unlock(&ipath_mutex);
1873 if (--pd->port_cnt) {
1875 * XXX If the master closes the port before the slave(s),
1876 * revoke the mmap for the eager receive queue so
1877 * the slave(s) don't wait for receive data forever.
1879 pd->active_slaves &= ~(1 << fd->subport);
1880 mutex_unlock(&ipath_mutex);
1883 port = pd->port_port;
1886 if (pd->port_hdrqfull) {
1887 ipath_cdbg(PROC, "%s[%u] had %u rcvhdrqfull errors "
1888 "during run\n", pd->port_comm, pd->port_pid,
1890 pd->port_hdrqfull = 0;
1893 if (pd->port_rcvwait_to || pd->port_piowait_to
1894 || pd->port_rcvnowait || pd->port_pionowait) {
1895 ipath_cdbg(VERBOSE, "port%u, %u rcv, %u pio wait timeo; "
1896 "%u rcv %u, pio already\n",
1897 pd->port_port, pd->port_rcvwait_to,
1898 pd->port_piowait_to, pd->port_rcvnowait,
1899 pd->port_pionowait);
1900 pd->port_rcvwait_to = pd->port_piowait_to =
1901 pd->port_rcvnowait = pd->port_pionowait = 0;
1903 if (pd->port_flag) {
1904 ipath_dbg("port %u port_flag still set to 0x%lx\n",
1905 pd->port_port, pd->port_flag);
1909 if (dd->ipath_kregbase) {
1911 /* atomically clear receive enable port. */
1912 clear_bit(INFINIPATH_R_PORTENABLE_SHIFT + port,
1913 &dd->ipath_rcvctrl);
1914 ipath_write_kreg( dd, dd->ipath_kregs->kr_rcvctrl,
1916 /* and read back from chip to be sure that nothing
1917 * else is in flight when we do the rest */
1918 (void)ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
1920 /* clean up the pkeys for this port user */
1921 ipath_clean_part_key(pd, dd);
1923 * be paranoid, and never write 0's to these, just use an
1924 * unused part of the port 0 tail page. Of course,
1925 * rcvhdraddr points to a large chunk of memory, so this
1926 * could still trash things, but at least it won't trash
1927 * page 0, and by disabling the port, it should stop "soon",
1928 * even if a packet or two is in already in flight after we
1929 * disabled the port.
1931 ipath_write_kreg_port(dd,
1932 dd->ipath_kregs->kr_rcvhdrtailaddr, port,
1933 dd->ipath_dummy_hdrq_phys);
1934 ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdraddr,
1935 pd->port_port, dd->ipath_dummy_hdrq_phys);
1937 i = dd->ipath_pbufsport * (port - 1);
1938 ipath_disarm_piobufs(dd, i, dd->ipath_pbufsport);
1940 if (dd->ipath_pageshadow)
1941 unlock_expected_tids(pd);
1942 ipath_stats.sps_ports--;
1943 ipath_cdbg(PROC, "%s[%u] closed port %u:%u\n",
1944 pd->port_comm, pd->port_pid,
1945 dd->ipath_unit, port);
1947 dd->ipath_f_clear_tids(dd, pd->port_port);
1951 dd->ipath_pd[pd->port_port] = NULL; /* before releasing mutex */
1952 mutex_unlock(&ipath_mutex);
1953 ipath_free_pddata(dd, pd); /* after releasing the mutex */
1960 static int ipath_port_info(struct ipath_portdata *pd, u16 subport,
1961 struct ipath_port_info __user *uinfo)
1963 struct ipath_port_info info;
1968 (void) ipath_count_units(NULL, &nup, NULL);
1969 info.num_active = nup;
1970 info.unit = pd->port_dd->ipath_unit;
1971 info.port = pd->port_port;
1972 info.subport = subport;
1973 /* Don't return new fields if old library opened the port. */
1974 if ((pd->userversion & 0xffff) == IPATH_USER_SWMINOR) {
1975 /* Number of user ports available for this device. */
1976 info.num_ports = pd->port_dd->ipath_cfgports - 1;
1977 info.num_subports = pd->port_subport_cnt;
1980 sz = sizeof(info) - 2 * sizeof(u16);
1982 if (copy_to_user(uinfo, &info, sz)) {
1992 static int ipath_get_slave_info(struct ipath_portdata *pd,
1993 void __user *slave_mask_addr)
1997 if (copy_to_user(slave_mask_addr, &pd->active_slaves, sizeof(u32)))
2002 static ssize_t ipath_write(struct file *fp, const char __user *data,
2003 size_t count, loff_t *off)
2005 const struct ipath_cmd __user *ucmd;
2006 struct ipath_portdata *pd;
2007 const void __user *src;
2008 size_t consumed, copy;
2009 struct ipath_cmd cmd;
2013 if (count < sizeof(cmd.type)) {
2018 ucmd = (const struct ipath_cmd __user *) data;
2020 if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) {
2025 consumed = sizeof(cmd.type);
2028 case IPATH_CMD_USER_INIT:
2029 copy = sizeof(cmd.cmd.user_info);
2030 dest = &cmd.cmd.user_info;
2031 src = &ucmd->cmd.user_info;
2033 case IPATH_CMD_RECV_CTRL:
2034 copy = sizeof(cmd.cmd.recv_ctrl);
2035 dest = &cmd.cmd.recv_ctrl;
2036 src = &ucmd->cmd.recv_ctrl;
2038 case IPATH_CMD_PORT_INFO:
2039 copy = sizeof(cmd.cmd.port_info);
2040 dest = &cmd.cmd.port_info;
2041 src = &ucmd->cmd.port_info;
2043 case IPATH_CMD_TID_UPDATE:
2044 case IPATH_CMD_TID_FREE:
2045 copy = sizeof(cmd.cmd.tid_info);
2046 dest = &cmd.cmd.tid_info;
2047 src = &ucmd->cmd.tid_info;
2049 case IPATH_CMD_SET_PART_KEY:
2050 copy = sizeof(cmd.cmd.part_key);
2051 dest = &cmd.cmd.part_key;
2052 src = &ucmd->cmd.part_key;
2054 case IPATH_CMD_SLAVE_INFO:
2055 copy = sizeof(cmd.cmd.slave_mask_addr);
2056 dest = &cmd.cmd.slave_mask_addr;
2057 src = &ucmd->cmd.slave_mask_addr;
2064 if ((count - consumed) < copy) {
2069 if (copy_from_user(dest, src, copy)) {
2076 if (!pd && cmd.type != IPATH_CMD_USER_INIT) {
2082 case IPATH_CMD_USER_INIT:
2083 ret = ipath_do_user_init(fp, &cmd.cmd.user_info);
2086 ret = ipath_get_base_info(
2087 fp, (void __user *) (unsigned long)
2088 cmd.cmd.user_info.spu_base_info,
2089 cmd.cmd.user_info.spu_base_info_size);
2091 case IPATH_CMD_RECV_CTRL:
2092 ret = ipath_manage_rcvq(pd, subport_fp(fp), cmd.cmd.recv_ctrl);
2094 case IPATH_CMD_PORT_INFO:
2095 ret = ipath_port_info(pd, subport_fp(fp),
2096 (struct ipath_port_info __user *)
2097 (unsigned long) cmd.cmd.port_info);
2099 case IPATH_CMD_TID_UPDATE:
2100 ret = ipath_tid_update(pd, fp, &cmd.cmd.tid_info);
2102 case IPATH_CMD_TID_FREE:
2103 ret = ipath_tid_free(pd, subport_fp(fp), &cmd.cmd.tid_info);
2105 case IPATH_CMD_SET_PART_KEY:
2106 ret = ipath_set_part_key(pd, cmd.cmd.part_key);
2108 case IPATH_CMD_SLAVE_INFO:
2109 ret = ipath_get_slave_info(pd,
2110 (void __user *) (unsigned long)
2111 cmd.cmd.slave_mask_addr);
2122 static struct class *ipath_class;
2124 static int init_cdev(int minor, char *name, struct file_operations *fops,
2125 struct cdev **cdevp, struct class_device **class_devp)
2127 const dev_t dev = MKDEV(IPATH_MAJOR, minor);
2128 struct cdev *cdev = NULL;
2129 struct class_device *class_dev = NULL;
2132 cdev = cdev_alloc();
2134 printk(KERN_ERR IPATH_DRV_NAME
2135 ": Could not allocate cdev for minor %d, %s\n",
2141 cdev->owner = THIS_MODULE;
2143 kobject_set_name(&cdev->kobj, name);
2145 ret = cdev_add(cdev, dev, 1);
2147 printk(KERN_ERR IPATH_DRV_NAME
2148 ": Could not add cdev for minor %d, %s (err %d)\n",
2153 class_dev = class_device_create(ipath_class, NULL, dev, NULL, name);
2155 if (IS_ERR(class_dev)) {
2156 ret = PTR_ERR(class_dev);
2157 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2158 "class_dev for minor %d, %s (err %d)\n",
2172 *class_devp = class_dev;
2181 int ipath_cdev_init(int minor, char *name, struct file_operations *fops,
2182 struct cdev **cdevp, struct class_device **class_devp)
2184 return init_cdev(minor, name, fops, cdevp, class_devp);
2187 static void cleanup_cdev(struct cdev **cdevp,
2188 struct class_device **class_devp)
2190 struct class_device *class_dev = *class_devp;
2193 class_device_unregister(class_dev);
2203 void ipath_cdev_cleanup(struct cdev **cdevp,
2204 struct class_device **class_devp)
2206 cleanup_cdev(cdevp, class_devp);
2209 static struct cdev *wildcard_cdev;
2210 static struct class_device *wildcard_class_dev;
2212 static const dev_t dev = MKDEV(IPATH_MAJOR, 0);
2214 static int user_init(void)
2218 ret = register_chrdev_region(dev, IPATH_NMINORS, IPATH_DRV_NAME);
2220 printk(KERN_ERR IPATH_DRV_NAME ": Could not register "
2221 "chrdev region (err %d)\n", -ret);
2225 ipath_class = class_create(THIS_MODULE, IPATH_DRV_NAME);
2227 if (IS_ERR(ipath_class)) {
2228 ret = PTR_ERR(ipath_class);
2229 printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2230 "device class (err %d)\n", -ret);
2236 unregister_chrdev_region(dev, IPATH_NMINORS);
2241 static void user_cleanup(void)
2244 class_destroy(ipath_class);
2248 unregister_chrdev_region(dev, IPATH_NMINORS);
2251 static atomic_t user_count = ATOMIC_INIT(0);
2252 static atomic_t user_setup = ATOMIC_INIT(0);
2254 int ipath_user_add(struct ipath_devdata *dd)
2259 if (atomic_inc_return(&user_count) == 1) {
2262 ipath_dev_err(dd, "Unable to set up user support: "
2263 "error %d\n", -ret);
2266 ret = init_cdev(0, "ipath", &ipath_file_ops, &wildcard_cdev,
2267 &wildcard_class_dev);
2269 ipath_dev_err(dd, "Could not create wildcard "
2270 "minor: error %d\n", -ret);
2274 atomic_set(&user_setup, 1);
2277 snprintf(name, sizeof(name), "ipath%d", dd->ipath_unit);
2279 ret = init_cdev(dd->ipath_unit + 1, name, &ipath_file_ops,
2280 &dd->user_cdev, &dd->user_class_dev);
2282 ipath_dev_err(dd, "Could not create user minor %d, %s\n",
2283 dd->ipath_unit + 1, name);
2293 void ipath_user_remove(struct ipath_devdata *dd)
2295 cleanup_cdev(&dd->user_cdev, &dd->user_class_dev);
2297 if (atomic_dec_return(&user_count) == 0) {
2298 if (atomic_read(&user_setup) == 0)
2301 cleanup_cdev(&wildcard_cdev, &wildcard_class_dev);
2304 atomic_set(&user_setup, 0);