[go: nahoru, domu]

1/*
2 * Copyright (c) 2006, 2007, 2008 QLogic Corporation. All rights reserved.
3 * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
4 *
5 * This software is available to you under a choice of one of two
6 * licenses.  You may choose to be licensed under the terms of the GNU
7 * General Public License (GPL) Version 2, available from the file
8 * COPYING in the main directory of this source tree, or the
9 * OpenIB.org BSD license below:
10 *
11 *     Redistribution and use in source and binary forms, with or
12 *     without modification, are permitted provided that the following
13 *     conditions are met:
14 *
15 *      - Redistributions of source code must retain the above
16 *        copyright notice, this list of conditions and the following
17 *        disclaimer.
18 *
19 *      - Redistributions in binary form must reproduce the above
20 *        copyright notice, this list of conditions and the following
21 *        disclaimer in the documentation and/or other materials
22 *        provided with the distribution.
23 *
24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
31 * SOFTWARE.
32 */
33
34#include <linux/pci.h>
35#include <linux/poll.h>
36#include <linux/cdev.h>
37#include <linux/swap.h>
38#include <linux/export.h>
39#include <linux/vmalloc.h>
40#include <linux/slab.h>
41#include <linux/highmem.h>
42#include <linux/io.h>
43#include <linux/aio.h>
44#include <linux/jiffies.h>
45#include <linux/cpu.h>
46#include <asm/pgtable.h>
47
48#include "ipath_kernel.h"
49#include "ipath_common.h"
50#include "ipath_user_sdma.h"
51
52static int ipath_open(struct inode *, struct file *);
53static int ipath_close(struct inode *, struct file *);
54static ssize_t ipath_write(struct file *, const char __user *, size_t,
55			   loff_t *);
56static ssize_t ipath_writev(struct kiocb *, const struct iovec *,
57			    unsigned long , loff_t);
58static unsigned int ipath_poll(struct file *, struct poll_table_struct *);
59static int ipath_mmap(struct file *, struct vm_area_struct *);
60
61static const struct file_operations ipath_file_ops = {
62	.owner = THIS_MODULE,
63	.write = ipath_write,
64	.aio_write = ipath_writev,
65	.open = ipath_open,
66	.release = ipath_close,
67	.poll = ipath_poll,
68	.mmap = ipath_mmap,
69	.llseek = noop_llseek,
70};
71
72/*
73 * Convert kernel virtual addresses to physical addresses so they don't
74 * potentially conflict with the chip addresses used as mmap offsets.
75 * It doesn't really matter what mmap offset we use as long as we can
76 * interpret it correctly.
77 */
78static u64 cvt_kvaddr(void *p)
79{
80	struct page *page;
81	u64 paddr = 0;
82
83	page = vmalloc_to_page(p);
84	if (page)
85		paddr = page_to_pfn(page) << PAGE_SHIFT;
86
87	return paddr;
88}
89
90static int ipath_get_base_info(struct file *fp,
91			       void __user *ubase, size_t ubase_size)
92{
93	struct ipath_portdata *pd = port_fp(fp);
94	int ret = 0;
95	struct ipath_base_info *kinfo = NULL;
96	struct ipath_devdata *dd = pd->port_dd;
97	unsigned subport_cnt;
98	int shared, master;
99	size_t sz;
100
101	subport_cnt = pd->port_subport_cnt;
102	if (!subport_cnt) {
103		shared = 0;
104		master = 0;
105		subport_cnt = 1;
106	} else {
107		shared = 1;
108		master = !subport_fp(fp);
109	}
110
111	sz = sizeof(*kinfo);
112	/* If port sharing is not requested, allow the old size structure */
113	if (!shared)
114		sz -= 7 * sizeof(u64);
115	if (ubase_size < sz) {
116		ipath_cdbg(PROC,
117			   "Base size %zu, need %zu (version mismatch?)\n",
118			   ubase_size, sz);
119		ret = -EINVAL;
120		goto bail;
121	}
122
123	kinfo = kzalloc(sizeof(*kinfo), GFP_KERNEL);
124	if (kinfo == NULL) {
125		ret = -ENOMEM;
126		goto bail;
127	}
128
129	ret = dd->ipath_f_get_base_info(pd, kinfo);
130	if (ret < 0)
131		goto bail;
132
133	kinfo->spi_rcvhdr_cnt = dd->ipath_rcvhdrcnt;
134	kinfo->spi_rcvhdrent_size = dd->ipath_rcvhdrentsize;
135	kinfo->spi_tidegrcnt = dd->ipath_rcvegrcnt;
136	kinfo->spi_rcv_egrbufsize = dd->ipath_rcvegrbufsize;
137	/*
138	 * have to mmap whole thing
139	 */
140	kinfo->spi_rcv_egrbuftotlen =
141		pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
142	kinfo->spi_rcv_egrperchunk = pd->port_rcvegrbufs_perchunk;
143	kinfo->spi_rcv_egrchunksize = kinfo->spi_rcv_egrbuftotlen /
144		pd->port_rcvegrbuf_chunks;
145	kinfo->spi_tidcnt = dd->ipath_rcvtidcnt / subport_cnt;
146	if (master)
147		kinfo->spi_tidcnt += dd->ipath_rcvtidcnt % subport_cnt;
148	/*
149	 * for this use, may be ipath_cfgports summed over all chips that
150	 * are are configured and present
151	 */
152	kinfo->spi_nports = dd->ipath_cfgports;
153	/* unit (chip/board) our port is on */
154	kinfo->spi_unit = dd->ipath_unit;
155	/* for now, only a single page */
156	kinfo->spi_tid_maxsize = PAGE_SIZE;
157
158	/*
159	 * Doing this per port, and based on the skip value, etc.  This has
160	 * to be the actual buffer size, since the protocol code treats it
161	 * as an array.
162	 *
163	 * These have to be set to user addresses in the user code via mmap.
164	 * These values are used on return to user code for the mmap target
165	 * addresses only.  For 32 bit, same 44 bit address problem, so use
166	 * the physical address, not virtual.  Before 2.6.11, using the
167	 * page_address() macro worked, but in 2.6.11, even that returns the
168	 * full 64 bit address (upper bits all 1's).  So far, using the
169	 * physical addresses (or chip offsets, for chip mapping) works, but
170	 * no doubt some future kernel release will change that, and we'll be
171	 * on to yet another method of dealing with this.
172	 */
173	kinfo->spi_rcvhdr_base = (u64) pd->port_rcvhdrq_phys;
174	kinfo->spi_rcvhdr_tailaddr = (u64) pd->port_rcvhdrqtailaddr_phys;
175	kinfo->spi_rcv_egrbufs = (u64) pd->port_rcvegr_phys;
176	kinfo->spi_pioavailaddr = (u64) dd->ipath_pioavailregs_phys;
177	kinfo->spi_status = (u64) kinfo->spi_pioavailaddr +
178		(void *) dd->ipath_statusp -
179		(void *) dd->ipath_pioavailregs_dma;
180	if (!shared) {
181		kinfo->spi_piocnt = pd->port_piocnt;
182		kinfo->spi_piobufbase = (u64) pd->port_piobufs;
183		kinfo->__spi_uregbase = (u64) dd->ipath_uregbase +
184			dd->ipath_ureg_align * pd->port_port;
185	} else if (master) {
186		kinfo->spi_piocnt = (pd->port_piocnt / subport_cnt) +
187				    (pd->port_piocnt % subport_cnt);
188		/* Master's PIO buffers are after all the slave's */
189		kinfo->spi_piobufbase = (u64) pd->port_piobufs +
190			dd->ipath_palign *
191			(pd->port_piocnt - kinfo->spi_piocnt);
192	} else {
193		unsigned slave = subport_fp(fp) - 1;
194
195		kinfo->spi_piocnt = pd->port_piocnt / subport_cnt;
196		kinfo->spi_piobufbase = (u64) pd->port_piobufs +
197			dd->ipath_palign * kinfo->spi_piocnt * slave;
198	}
199
200	if (shared) {
201		kinfo->spi_port_uregbase = (u64) dd->ipath_uregbase +
202			dd->ipath_ureg_align * pd->port_port;
203		kinfo->spi_port_rcvegrbuf = kinfo->spi_rcv_egrbufs;
204		kinfo->spi_port_rcvhdr_base = kinfo->spi_rcvhdr_base;
205		kinfo->spi_port_rcvhdr_tailaddr = kinfo->spi_rcvhdr_tailaddr;
206
207		kinfo->__spi_uregbase = cvt_kvaddr(pd->subport_uregbase +
208			PAGE_SIZE * subport_fp(fp));
209
210		kinfo->spi_rcvhdr_base = cvt_kvaddr(pd->subport_rcvhdr_base +
211			pd->port_rcvhdrq_size * subport_fp(fp));
212		kinfo->spi_rcvhdr_tailaddr = 0;
213		kinfo->spi_rcv_egrbufs = cvt_kvaddr(pd->subport_rcvegrbuf +
214			pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size *
215			subport_fp(fp));
216
217		kinfo->spi_subport_uregbase =
218			cvt_kvaddr(pd->subport_uregbase);
219		kinfo->spi_subport_rcvegrbuf =
220			cvt_kvaddr(pd->subport_rcvegrbuf);
221		kinfo->spi_subport_rcvhdr_base =
222			cvt_kvaddr(pd->subport_rcvhdr_base);
223		ipath_cdbg(PROC, "port %u flags %x %llx %llx %llx\n",
224			kinfo->spi_port, kinfo->spi_runtime_flags,
225			(unsigned long long) kinfo->spi_subport_uregbase,
226			(unsigned long long) kinfo->spi_subport_rcvegrbuf,
227			(unsigned long long) kinfo->spi_subport_rcvhdr_base);
228	}
229
230	/*
231	 * All user buffers are 2KB buffers.  If we ever support
232	 * giving 4KB buffers to user processes, this will need some
233	 * work.
234	 */
235	kinfo->spi_pioindex = (kinfo->spi_piobufbase -
236		(dd->ipath_piobufbase & 0xffffffff)) / dd->ipath_palign;
237	kinfo->spi_pioalign = dd->ipath_palign;
238
239	kinfo->spi_qpair = IPATH_KD_QP;
240	/*
241	 * user mode PIO buffers are always 2KB, even when 4KB can
242	 * be received, and sent via the kernel; this is ibmaxlen
243	 * for 2K MTU.
244	 */
245	kinfo->spi_piosize = dd->ipath_piosize2k - 2 * sizeof(u32);
246	kinfo->spi_mtu = dd->ipath_ibmaxlen;	/* maxlen, not ibmtu */
247	kinfo->spi_port = pd->port_port;
248	kinfo->spi_subport = subport_fp(fp);
249	kinfo->spi_sw_version = IPATH_KERN_SWVERSION;
250	kinfo->spi_hw_version = dd->ipath_revision;
251
252	if (master) {
253		kinfo->spi_runtime_flags |= IPATH_RUNTIME_MASTER;
254	}
255
256	sz = (ubase_size < sizeof(*kinfo)) ? ubase_size : sizeof(*kinfo);
257	if (copy_to_user(ubase, kinfo, sz))
258		ret = -EFAULT;
259
260bail:
261	kfree(kinfo);
262	return ret;
263}
264
265/**
266 * ipath_tid_update - update a port TID
267 * @pd: the port
268 * @fp: the ipath device file
269 * @ti: the TID information
270 *
271 * The new implementation as of Oct 2004 is that the driver assigns
272 * the tid and returns it to the caller.   To make it easier to
273 * catch bugs, and to reduce search time, we keep a cursor for
274 * each port, walking the shadow tid array to find one that's not
275 * in use.
276 *
277 * For now, if we can't allocate the full list, we fail, although
278 * in the long run, we'll allocate as many as we can, and the
279 * caller will deal with that by trying the remaining pages later.
280 * That means that when we fail, we have to mark the tids as not in
281 * use again, in our shadow copy.
282 *
283 * It's up to the caller to free the tids when they are done.
284 * We'll unlock the pages as they free them.
285 *
286 * Also, right now we are locking one page at a time, but since
287 * the intended use of this routine is for a single group of
288 * virtually contiguous pages, that should change to improve
289 * performance.
290 */
291static int ipath_tid_update(struct ipath_portdata *pd, struct file *fp,
292			    const struct ipath_tid_info *ti)
293{
294	int ret = 0, ntids;
295	u32 tid, porttid, cnt, i, tidcnt, tidoff;
296	u16 *tidlist;
297	struct ipath_devdata *dd = pd->port_dd;
298	u64 physaddr;
299	unsigned long vaddr;
300	u64 __iomem *tidbase;
301	unsigned long tidmap[8];
302	struct page **pagep = NULL;
303	unsigned subport = subport_fp(fp);
304
305	if (!dd->ipath_pageshadow) {
306		ret = -ENOMEM;
307		goto done;
308	}
309
310	cnt = ti->tidcnt;
311	if (!cnt) {
312		ipath_dbg("After copyin, tidcnt 0, tidlist %llx\n",
313			  (unsigned long long) ti->tidlist);
314		/*
315		 * Should we treat as success?  likely a bug
316		 */
317		ret = -EFAULT;
318		goto done;
319	}
320	porttid = pd->port_port * dd->ipath_rcvtidcnt;
321	if (!pd->port_subport_cnt) {
322		tidcnt = dd->ipath_rcvtidcnt;
323		tid = pd->port_tidcursor;
324		tidoff = 0;
325	} else if (!subport) {
326		tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
327			 (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
328		tidoff = dd->ipath_rcvtidcnt - tidcnt;
329		porttid += tidoff;
330		tid = tidcursor_fp(fp);
331	} else {
332		tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
333		tidoff = tidcnt * (subport - 1);
334		porttid += tidoff;
335		tid = tidcursor_fp(fp);
336	}
337	if (cnt > tidcnt) {
338		/* make sure it all fits in port_tid_pg_list */
339		dev_info(&dd->pcidev->dev, "Process tried to allocate %u "
340			 "TIDs, only trying max (%u)\n", cnt, tidcnt);
341		cnt = tidcnt;
342	}
343	pagep = &((struct page **) pd->port_tid_pg_list)[tidoff];
344	tidlist = &((u16 *) &pagep[dd->ipath_rcvtidcnt])[tidoff];
345
346	memset(tidmap, 0, sizeof(tidmap));
347	/* before decrement; chip actual # */
348	ntids = tidcnt;
349	tidbase = (u64 __iomem *) (((char __iomem *) dd->ipath_kregbase) +
350				   dd->ipath_rcvtidbase +
351				   porttid * sizeof(*tidbase));
352
353	ipath_cdbg(VERBOSE, "Port%u %u tids, cursor %u, tidbase %p\n",
354		   pd->port_port, cnt, tid, tidbase);
355
356	/* virtual address of first page in transfer */
357	vaddr = ti->tidvaddr;
358	if (!access_ok(VERIFY_WRITE, (void __user *) vaddr,
359		       cnt * PAGE_SIZE)) {
360		ipath_dbg("Fail vaddr %p, %u pages, !access_ok\n",
361			  (void *)vaddr, cnt);
362		ret = -EFAULT;
363		goto done;
364	}
365	ret = ipath_get_user_pages(vaddr, cnt, pagep);
366	if (ret) {
367		if (ret == -EBUSY) {
368			ipath_dbg("Failed to lock addr %p, %u pages "
369				  "(already locked)\n",
370				  (void *) vaddr, cnt);
371			/*
372			 * for now, continue, and see what happens but with
373			 * the new implementation, this should never happen,
374			 * unless perhaps the user has mpin'ed the pages
375			 * themselves (something we need to test)
376			 */
377			ret = 0;
378		} else {
379			dev_info(&dd->pcidev->dev,
380				 "Failed to lock addr %p, %u pages: "
381				 "errno %d\n", (void *) vaddr, cnt, -ret);
382			goto done;
383		}
384	}
385	for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
386		for (; ntids--; tid++) {
387			if (tid == tidcnt)
388				tid = 0;
389			if (!dd->ipath_pageshadow[porttid + tid])
390				break;
391		}
392		if (ntids < 0) {
393			/*
394			 * oops, wrapped all the way through their TIDs,
395			 * and didn't have enough free; see comments at
396			 * start of routine
397			 */
398			ipath_dbg("Not enough free TIDs for %u pages "
399				  "(index %d), failing\n", cnt, i);
400			i--;	/* last tidlist[i] not filled in */
401			ret = -ENOMEM;
402			break;
403		}
404		tidlist[i] = tid + tidoff;
405		ipath_cdbg(VERBOSE, "Updating idx %u to TID %u, "
406			   "vaddr %lx\n", i, tid + tidoff, vaddr);
407		/* we "know" system pages and TID pages are same size */
408		dd->ipath_pageshadow[porttid + tid] = pagep[i];
409		dd->ipath_physshadow[porttid + tid] = ipath_map_page(
410			dd->pcidev, pagep[i], 0, PAGE_SIZE,
411			PCI_DMA_FROMDEVICE);
412		/*
413		 * don't need atomic or it's overhead
414		 */
415		__set_bit(tid, tidmap);
416		physaddr = dd->ipath_physshadow[porttid + tid];
417		ipath_stats.sps_pagelocks++;
418		ipath_cdbg(VERBOSE,
419			   "TID %u, vaddr %lx, physaddr %llx pgp %p\n",
420			   tid, vaddr, (unsigned long long) physaddr,
421			   pagep[i]);
422		dd->ipath_f_put_tid(dd, &tidbase[tid], RCVHQ_RCV_TYPE_EXPECTED,
423				    physaddr);
424		/*
425		 * don't check this tid in ipath_portshadow, since we
426		 * just filled it in; start with the next one.
427		 */
428		tid++;
429	}
430
431	if (ret) {
432		u32 limit;
433	cleanup:
434		/* jump here if copy out of updated info failed... */
435		ipath_dbg("After failure (ret=%d), undo %d of %d entries\n",
436			  -ret, i, cnt);
437		/* same code that's in ipath_free_tid() */
438		limit = sizeof(tidmap) * BITS_PER_BYTE;
439		if (limit > tidcnt)
440			/* just in case size changes in future */
441			limit = tidcnt;
442		tid = find_first_bit((const unsigned long *)tidmap, limit);
443		for (; tid < limit; tid++) {
444			if (!test_bit(tid, tidmap))
445				continue;
446			if (dd->ipath_pageshadow[porttid + tid]) {
447				ipath_cdbg(VERBOSE, "Freeing TID %u\n",
448					   tid);
449				dd->ipath_f_put_tid(dd, &tidbase[tid],
450						    RCVHQ_RCV_TYPE_EXPECTED,
451						    dd->ipath_tidinvalid);
452				pci_unmap_page(dd->pcidev,
453					dd->ipath_physshadow[porttid + tid],
454					PAGE_SIZE, PCI_DMA_FROMDEVICE);
455				dd->ipath_pageshadow[porttid + tid] = NULL;
456				ipath_stats.sps_pageunlocks++;
457			}
458		}
459		ipath_release_user_pages(pagep, cnt);
460	} else {
461		/*
462		 * Copy the updated array, with ipath_tid's filled in, back
463		 * to user.  Since we did the copy in already, this "should
464		 * never fail" If it does, we have to clean up...
465		 */
466		if (copy_to_user((void __user *)
467				 (unsigned long) ti->tidlist,
468				 tidlist, cnt * sizeof(*tidlist))) {
469			ret = -EFAULT;
470			goto cleanup;
471		}
472		if (copy_to_user((void __user *) (unsigned long) ti->tidmap,
473				 tidmap, sizeof tidmap)) {
474			ret = -EFAULT;
475			goto cleanup;
476		}
477		if (tid == tidcnt)
478			tid = 0;
479		if (!pd->port_subport_cnt)
480			pd->port_tidcursor = tid;
481		else
482			tidcursor_fp(fp) = tid;
483	}
484
485done:
486	if (ret)
487		ipath_dbg("Failed to map %u TID pages, failing with %d\n",
488			  ti->tidcnt, -ret);
489	return ret;
490}
491
492/**
493 * ipath_tid_free - free a port TID
494 * @pd: the port
495 * @subport: the subport
496 * @ti: the TID info
497 *
498 * right now we are unlocking one page at a time, but since
499 * the intended use of this routine is for a single group of
500 * virtually contiguous pages, that should change to improve
501 * performance.  We check that the TID is in range for this port
502 * but otherwise don't check validity; if user has an error and
503 * frees the wrong tid, it's only their own data that can thereby
504 * be corrupted.  We do check that the TID was in use, for sanity
505 * We always use our idea of the saved address, not the address that
506 * they pass in to us.
507 */
508
509static int ipath_tid_free(struct ipath_portdata *pd, unsigned subport,
510			  const struct ipath_tid_info *ti)
511{
512	int ret = 0;
513	u32 tid, porttid, cnt, limit, tidcnt;
514	struct ipath_devdata *dd = pd->port_dd;
515	u64 __iomem *tidbase;
516	unsigned long tidmap[8];
517
518	if (!dd->ipath_pageshadow) {
519		ret = -ENOMEM;
520		goto done;
521	}
522
523	if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap,
524			   sizeof tidmap)) {
525		ret = -EFAULT;
526		goto done;
527	}
528
529	porttid = pd->port_port * dd->ipath_rcvtidcnt;
530	if (!pd->port_subport_cnt)
531		tidcnt = dd->ipath_rcvtidcnt;
532	else if (!subport) {
533		tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) +
534			 (dd->ipath_rcvtidcnt % pd->port_subport_cnt);
535		porttid += dd->ipath_rcvtidcnt - tidcnt;
536	} else {
537		tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt;
538		porttid += tidcnt * (subport - 1);
539	}
540	tidbase = (u64 __iomem *) ((char __iomem *)(dd->ipath_kregbase) +
541				   dd->ipath_rcvtidbase +
542				   porttid * sizeof(*tidbase));
543
544	limit = sizeof(tidmap) * BITS_PER_BYTE;
545	if (limit > tidcnt)
546		/* just in case size changes in future */
547		limit = tidcnt;
548	tid = find_first_bit(tidmap, limit);
549	ipath_cdbg(VERBOSE, "Port%u free %u tids; first bit (max=%d) "
550		   "set is %d, porttid %u\n", pd->port_port, ti->tidcnt,
551		   limit, tid, porttid);
552	for (cnt = 0; tid < limit; tid++) {
553		/*
554		 * small optimization; if we detect a run of 3 or so without
555		 * any set, use find_first_bit again.  That's mainly to
556		 * accelerate the case where we wrapped, so we have some at
557		 * the beginning, and some at the end, and a big gap
558		 * in the middle.
559		 */
560		if (!test_bit(tid, tidmap))
561			continue;
562		cnt++;
563		if (dd->ipath_pageshadow[porttid + tid]) {
564			struct page *p;
565			p = dd->ipath_pageshadow[porttid + tid];
566			dd->ipath_pageshadow[porttid + tid] = NULL;
567			ipath_cdbg(VERBOSE, "PID %u freeing TID %u\n",
568				   pid_nr(pd->port_pid), tid);
569			dd->ipath_f_put_tid(dd, &tidbase[tid],
570					    RCVHQ_RCV_TYPE_EXPECTED,
571					    dd->ipath_tidinvalid);
572			pci_unmap_page(dd->pcidev,
573				dd->ipath_physshadow[porttid + tid],
574				PAGE_SIZE, PCI_DMA_FROMDEVICE);
575			ipath_release_user_pages(&p, 1);
576			ipath_stats.sps_pageunlocks++;
577		} else
578			ipath_dbg("Unused tid %u, ignoring\n", tid);
579	}
580	if (cnt != ti->tidcnt)
581		ipath_dbg("passed in tidcnt %d, only %d bits set in map\n",
582			  ti->tidcnt, cnt);
583done:
584	if (ret)
585		ipath_dbg("Failed to unmap %u TID pages, failing with %d\n",
586			  ti->tidcnt, -ret);
587	return ret;
588}
589
590/**
591 * ipath_set_part_key - set a partition key
592 * @pd: the port
593 * @key: the key
594 *
595 * We can have up to 4 active at a time (other than the default, which is
596 * always allowed).  This is somewhat tricky, since multiple ports may set
597 * the same key, so we reference count them, and clean up at exit.  All 4
598 * partition keys are packed into a single infinipath register.  It's an
599 * error for a process to set the same pkey multiple times.  We provide no
600 * mechanism to de-allocate a pkey at this time, we may eventually need to
601 * do that.  I've used the atomic operations, and no locking, and only make
602 * a single pass through what's available.  This should be more than
603 * adequate for some time. I'll think about spinlocks or the like if and as
604 * it's necessary.
605 */
606static int ipath_set_part_key(struct ipath_portdata *pd, u16 key)
607{
608	struct ipath_devdata *dd = pd->port_dd;
609	int i, any = 0, pidx = -1;
610	u16 lkey = key & 0x7FFF;
611	int ret;
612
613	if (lkey == (IPATH_DEFAULT_P_KEY & 0x7FFF)) {
614		/* nothing to do; this key always valid */
615		ret = 0;
616		goto bail;
617	}
618
619	ipath_cdbg(VERBOSE, "p%u try to set pkey %hx, current keys "
620		   "%hx:%x %hx:%x %hx:%x %hx:%x\n",
621		   pd->port_port, key, dd->ipath_pkeys[0],
622		   atomic_read(&dd->ipath_pkeyrefs[0]), dd->ipath_pkeys[1],
623		   atomic_read(&dd->ipath_pkeyrefs[1]), dd->ipath_pkeys[2],
624		   atomic_read(&dd->ipath_pkeyrefs[2]), dd->ipath_pkeys[3],
625		   atomic_read(&dd->ipath_pkeyrefs[3]));
626
627	if (!lkey) {
628		ipath_cdbg(PROC, "p%u tries to set key 0, not allowed\n",
629			   pd->port_port);
630		ret = -EINVAL;
631		goto bail;
632	}
633
634	/*
635	 * Set the full membership bit, because it has to be
636	 * set in the register or the packet, and it seems
637	 * cleaner to set in the register than to force all
638	 * callers to set it. (see bug 4331)
639	 */
640	key |= 0x8000;
641
642	for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
643		if (!pd->port_pkeys[i] && pidx == -1)
644			pidx = i;
645		if (pd->port_pkeys[i] == key) {
646			ipath_cdbg(VERBOSE, "p%u tries to set same pkey "
647				   "(%x) more than once\n",
648				   pd->port_port, key);
649			ret = -EEXIST;
650			goto bail;
651		}
652	}
653	if (pidx == -1) {
654		ipath_dbg("All pkeys for port %u already in use, "
655			  "can't set %x\n", pd->port_port, key);
656		ret = -EBUSY;
657		goto bail;
658	}
659	for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
660		if (!dd->ipath_pkeys[i]) {
661			any++;
662			continue;
663		}
664		if (dd->ipath_pkeys[i] == key) {
665			atomic_t *pkrefs = &dd->ipath_pkeyrefs[i];
666
667			if (atomic_inc_return(pkrefs) > 1) {
668				pd->port_pkeys[pidx] = key;
669				ipath_cdbg(VERBOSE, "p%u set key %x "
670					   "matches #%d, count now %d\n",
671					   pd->port_port, key, i,
672					   atomic_read(pkrefs));
673				ret = 0;
674				goto bail;
675			} else {
676				/*
677				 * lost race, decrement count, catch below
678				 */
679				atomic_dec(pkrefs);
680				ipath_cdbg(VERBOSE, "Lost race, count was "
681					   "0, after dec, it's %d\n",
682					   atomic_read(pkrefs));
683				any++;
684			}
685		}
686		if ((dd->ipath_pkeys[i] & 0x7FFF) == lkey) {
687			/*
688			 * It makes no sense to have both the limited and
689			 * full membership PKEY set at the same time since
690			 * the unlimited one will disable the limited one.
691			 */
692			ret = -EEXIST;
693			goto bail;
694		}
695	}
696	if (!any) {
697		ipath_dbg("port %u, all pkeys already in use, "
698			  "can't set %x\n", pd->port_port, key);
699		ret = -EBUSY;
700		goto bail;
701	}
702	for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) {
703		if (!dd->ipath_pkeys[i] &&
704		    atomic_inc_return(&dd->ipath_pkeyrefs[i]) == 1) {
705			u64 pkey;
706
707			/* for ipathstats, etc. */
708			ipath_stats.sps_pkeys[i] = lkey;
709			pd->port_pkeys[pidx] = dd->ipath_pkeys[i] = key;
710			pkey =
711				(u64) dd->ipath_pkeys[0] |
712				((u64) dd->ipath_pkeys[1] << 16) |
713				((u64) dd->ipath_pkeys[2] << 32) |
714				((u64) dd->ipath_pkeys[3] << 48);
715			ipath_cdbg(PROC, "p%u set key %x in #%d, "
716				   "portidx %d, new pkey reg %llx\n",
717				   pd->port_port, key, i, pidx,
718				   (unsigned long long) pkey);
719			ipath_write_kreg(
720				dd, dd->ipath_kregs->kr_partitionkey, pkey);
721
722			ret = 0;
723			goto bail;
724		}
725	}
726	ipath_dbg("port %u, all pkeys already in use 2nd pass, "
727		  "can't set %x\n", pd->port_port, key);
728	ret = -EBUSY;
729
730bail:
731	return ret;
732}
733
734/**
735 * ipath_manage_rcvq - manage a port's receive queue
736 * @pd: the port
737 * @subport: the subport
738 * @start_stop: action to carry out
739 *
740 * start_stop == 0 disables receive on the port, for use in queue
741 * overflow conditions.  start_stop==1 re-enables, to be used to
742 * re-init the software copy of the head register
743 */
744static int ipath_manage_rcvq(struct ipath_portdata *pd, unsigned subport,
745			     int start_stop)
746{
747	struct ipath_devdata *dd = pd->port_dd;
748
749	ipath_cdbg(PROC, "%sabling rcv for unit %u port %u:%u\n",
750		   start_stop ? "en" : "dis", dd->ipath_unit,
751		   pd->port_port, subport);
752	if (subport)
753		goto bail;
754	/* atomically clear receive enable port. */
755	if (start_stop) {
756		/*
757		 * On enable, force in-memory copy of the tail register to
758		 * 0, so that protocol code doesn't have to worry about
759		 * whether or not the chip has yet updated the in-memory
760		 * copy or not on return from the system call. The chip
761		 * always resets it's tail register back to 0 on a
762		 * transition from disabled to enabled.  This could cause a
763		 * problem if software was broken, and did the enable w/o
764		 * the disable, but eventually the in-memory copy will be
765		 * updated and correct itself, even in the face of software
766		 * bugs.
767		 */
768		if (pd->port_rcvhdrtail_kvaddr)
769			ipath_clear_rcvhdrtail(pd);
770		set_bit(dd->ipath_r_portenable_shift + pd->port_port,
771			&dd->ipath_rcvctrl);
772	} else
773		clear_bit(dd->ipath_r_portenable_shift + pd->port_port,
774			  &dd->ipath_rcvctrl);
775	ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
776			 dd->ipath_rcvctrl);
777	/* now be sure chip saw it before we return */
778	ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
779	if (start_stop) {
780		/*
781		 * And try to be sure that tail reg update has happened too.
782		 * This should in theory interlock with the RXE changes to
783		 * the tail register.  Don't assign it to the tail register
784		 * in memory copy, since we could overwrite an update by the
785		 * chip if we did.
786		 */
787		ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
788	}
789	/* always; new head should be equal to new tail; see above */
790bail:
791	return 0;
792}
793
794static void ipath_clean_part_key(struct ipath_portdata *pd,
795				 struct ipath_devdata *dd)
796{
797	int i, j, pchanged = 0;
798	u64 oldpkey;
799
800	/* for debugging only */
801	oldpkey = (u64) dd->ipath_pkeys[0] |
802		((u64) dd->ipath_pkeys[1] << 16) |
803		((u64) dd->ipath_pkeys[2] << 32) |
804		((u64) dd->ipath_pkeys[3] << 48);
805
806	for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) {
807		if (!pd->port_pkeys[i])
808			continue;
809		ipath_cdbg(VERBOSE, "look for key[%d] %hx in pkeys\n", i,
810			   pd->port_pkeys[i]);
811		for (j = 0; j < ARRAY_SIZE(dd->ipath_pkeys); j++) {
812			/* check for match independent of the global bit */
813			if ((dd->ipath_pkeys[j] & 0x7fff) !=
814			    (pd->port_pkeys[i] & 0x7fff))
815				continue;
816			if (atomic_dec_and_test(&dd->ipath_pkeyrefs[j])) {
817				ipath_cdbg(VERBOSE, "p%u clear key "
818					   "%x matches #%d\n",
819					   pd->port_port,
820					   pd->port_pkeys[i], j);
821				ipath_stats.sps_pkeys[j] =
822					dd->ipath_pkeys[j] = 0;
823				pchanged++;
824			}
825			else ipath_cdbg(
826				VERBOSE, "p%u key %x matches #%d, "
827				"but ref still %d\n", pd->port_port,
828				pd->port_pkeys[i], j,
829				atomic_read(&dd->ipath_pkeyrefs[j]));
830			break;
831		}
832		pd->port_pkeys[i] = 0;
833	}
834	if (pchanged) {
835		u64 pkey = (u64) dd->ipath_pkeys[0] |
836			((u64) dd->ipath_pkeys[1] << 16) |
837			((u64) dd->ipath_pkeys[2] << 32) |
838			((u64) dd->ipath_pkeys[3] << 48);
839		ipath_cdbg(VERBOSE, "p%u old pkey reg %llx, "
840			   "new pkey reg %llx\n", pd->port_port,
841			   (unsigned long long) oldpkey,
842			   (unsigned long long) pkey);
843		ipath_write_kreg(dd, dd->ipath_kregs->kr_partitionkey,
844				 pkey);
845	}
846}
847
848/*
849 * Initialize the port data with the receive buffer sizes
850 * so this can be done while the master port is locked.
851 * Otherwise, there is a race with a slave opening the port
852 * and seeing these fields uninitialized.
853 */
854static void init_user_egr_sizes(struct ipath_portdata *pd)
855{
856	struct ipath_devdata *dd = pd->port_dd;
857	unsigned egrperchunk, egrcnt, size;
858
859	/*
860	 * to avoid wasting a lot of memory, we allocate 32KB chunks of
861	 * physically contiguous memory, advance through it until used up
862	 * and then allocate more.  Of course, we need memory to store those
863	 * extra pointers, now.  Started out with 256KB, but under heavy
864	 * memory pressure (creating large files and then copying them over
865	 * NFS while doing lots of MPI jobs), we hit some allocation
866	 * failures, even though we can sleep...  (2.6.10) Still get
867	 * failures at 64K.  32K is the lowest we can go without wasting
868	 * additional memory.
869	 */
870	size = 0x8000;
871	egrperchunk = size / dd->ipath_rcvegrbufsize;
872	egrcnt = dd->ipath_rcvegrcnt;
873	pd->port_rcvegrbuf_chunks = (egrcnt + egrperchunk - 1) / egrperchunk;
874	pd->port_rcvegrbufs_perchunk = egrperchunk;
875	pd->port_rcvegrbuf_size = size;
876}
877
878/**
879 * ipath_create_user_egr - allocate eager TID buffers
880 * @pd: the port to allocate TID buffers for
881 *
882 * This routine is now quite different for user and kernel, because
883 * the kernel uses skb's, for the accelerated network performance
884 * This is the user port version
885 *
886 * Allocate the eager TID buffers and program them into infinipath
887 * They are no longer completely contiguous, we do multiple allocation
888 * calls.
889 */
890static int ipath_create_user_egr(struct ipath_portdata *pd)
891{
892	struct ipath_devdata *dd = pd->port_dd;
893	unsigned e, egrcnt, egrperchunk, chunk, egrsize, egroff;
894	size_t size;
895	int ret;
896	gfp_t gfp_flags;
897
898	/*
899	 * GFP_USER, but without GFP_FS, so buffer cache can be
900	 * coalesced (we hope); otherwise, even at order 4,
901	 * heavy filesystem activity makes these fail, and we can
902	 * use compound pages.
903	 */
904	gfp_flags = __GFP_WAIT | __GFP_IO | __GFP_COMP;
905
906	egrcnt = dd->ipath_rcvegrcnt;
907	/* TID number offset for this port */
908	egroff = (pd->port_port - 1) * egrcnt + dd->ipath_p0_rcvegrcnt;
909	egrsize = dd->ipath_rcvegrbufsize;
910	ipath_cdbg(VERBOSE, "Allocating %d egr buffers, at egrtid "
911		   "offset %x, egrsize %u\n", egrcnt, egroff, egrsize);
912
913	chunk = pd->port_rcvegrbuf_chunks;
914	egrperchunk = pd->port_rcvegrbufs_perchunk;
915	size = pd->port_rcvegrbuf_size;
916	pd->port_rcvegrbuf = kmalloc(chunk * sizeof(pd->port_rcvegrbuf[0]),
917				     GFP_KERNEL);
918	if (!pd->port_rcvegrbuf) {
919		ret = -ENOMEM;
920		goto bail;
921	}
922	pd->port_rcvegrbuf_phys =
923		kmalloc(chunk * sizeof(pd->port_rcvegrbuf_phys[0]),
924			GFP_KERNEL);
925	if (!pd->port_rcvegrbuf_phys) {
926		ret = -ENOMEM;
927		goto bail_rcvegrbuf;
928	}
929	for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) {
930
931		pd->port_rcvegrbuf[e] = dma_alloc_coherent(
932			&dd->pcidev->dev, size, &pd->port_rcvegrbuf_phys[e],
933			gfp_flags);
934
935		if (!pd->port_rcvegrbuf[e]) {
936			ret = -ENOMEM;
937			goto bail_rcvegrbuf_phys;
938		}
939	}
940
941	pd->port_rcvegr_phys = pd->port_rcvegrbuf_phys[0];
942
943	for (e = chunk = 0; chunk < pd->port_rcvegrbuf_chunks; chunk++) {
944		dma_addr_t pa = pd->port_rcvegrbuf_phys[chunk];
945		unsigned i;
946
947		for (i = 0; e < egrcnt && i < egrperchunk; e++, i++) {
948			dd->ipath_f_put_tid(dd, e + egroff +
949					    (u64 __iomem *)
950					    ((char __iomem *)
951					     dd->ipath_kregbase +
952					     dd->ipath_rcvegrbase),
953					    RCVHQ_RCV_TYPE_EAGER, pa);
954			pa += egrsize;
955		}
956		cond_resched();	/* don't hog the cpu */
957	}
958
959	ret = 0;
960	goto bail;
961
962bail_rcvegrbuf_phys:
963	for (e = 0; e < pd->port_rcvegrbuf_chunks &&
964		pd->port_rcvegrbuf[e]; e++) {
965		dma_free_coherent(&dd->pcidev->dev, size,
966				  pd->port_rcvegrbuf[e],
967				  pd->port_rcvegrbuf_phys[e]);
968
969	}
970	kfree(pd->port_rcvegrbuf_phys);
971	pd->port_rcvegrbuf_phys = NULL;
972bail_rcvegrbuf:
973	kfree(pd->port_rcvegrbuf);
974	pd->port_rcvegrbuf = NULL;
975bail:
976	return ret;
977}
978
979
980/* common code for the mappings on dma_alloc_coherent mem */
981static int ipath_mmap_mem(struct vm_area_struct *vma,
982	struct ipath_portdata *pd, unsigned len, int write_ok,
983	void *kvaddr, char *what)
984{
985	struct ipath_devdata *dd = pd->port_dd;
986	unsigned long pfn;
987	int ret;
988
989	if ((vma->vm_end - vma->vm_start) > len) {
990		dev_info(&dd->pcidev->dev,
991		         "FAIL on %s: len %lx > %x\n", what,
992			 vma->vm_end - vma->vm_start, len);
993		ret = -EFAULT;
994		goto bail;
995	}
996
997	if (!write_ok) {
998		if (vma->vm_flags & VM_WRITE) {
999			dev_info(&dd->pcidev->dev,
1000				 "%s must be mapped readonly\n", what);
1001			ret = -EPERM;
1002			goto bail;
1003		}
1004
1005		/* don't allow them to later change with mprotect */
1006		vma->vm_flags &= ~VM_MAYWRITE;
1007	}
1008
1009	pfn = virt_to_phys(kvaddr) >> PAGE_SHIFT;
1010	ret = remap_pfn_range(vma, vma->vm_start, pfn,
1011			      len, vma->vm_page_prot);
1012	if (ret)
1013		dev_info(&dd->pcidev->dev, "%s port%u mmap of %lx, %x "
1014			 "bytes r%c failed: %d\n", what, pd->port_port,
1015			 pfn, len, write_ok?'w':'o', ret);
1016	else
1017		ipath_cdbg(VERBOSE, "%s port%u mmaped %lx, %x bytes "
1018			   "r%c\n", what, pd->port_port, pfn, len,
1019			   write_ok?'w':'o');
1020bail:
1021	return ret;
1022}
1023
1024static int mmap_ureg(struct vm_area_struct *vma, struct ipath_devdata *dd,
1025		     u64 ureg)
1026{
1027	unsigned long phys;
1028	int ret;
1029
1030	/*
1031	 * This is real hardware, so use io_remap.  This is the mechanism
1032	 * for the user process to update the head registers for their port
1033	 * in the chip.
1034	 */
1035	if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) {
1036		dev_info(&dd->pcidev->dev, "FAIL mmap userreg: reqlen "
1037			 "%lx > PAGE\n", vma->vm_end - vma->vm_start);
1038		ret = -EFAULT;
1039	} else {
1040		phys = dd->ipath_physaddr + ureg;
1041		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
1042
1043		vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1044		ret = io_remap_pfn_range(vma, vma->vm_start,
1045					 phys >> PAGE_SHIFT,
1046					 vma->vm_end - vma->vm_start,
1047					 vma->vm_page_prot);
1048	}
1049	return ret;
1050}
1051
1052static int mmap_piobufs(struct vm_area_struct *vma,
1053			struct ipath_devdata *dd,
1054			struct ipath_portdata *pd,
1055			unsigned piobufs, unsigned piocnt)
1056{
1057	unsigned long phys;
1058	int ret;
1059
1060	/*
1061	 * When we map the PIO buffers in the chip, we want to map them as
1062	 * writeonly, no read possible.   This prevents access to previous
1063	 * process data, and catches users who might try to read the i/o
1064	 * space due to a bug.
1065	 */
1066	if ((vma->vm_end - vma->vm_start) > (piocnt * dd->ipath_palign)) {
1067		dev_info(&dd->pcidev->dev, "FAIL mmap piobufs: "
1068			 "reqlen %lx > PAGE\n",
1069			 vma->vm_end - vma->vm_start);
1070		ret = -EINVAL;
1071		goto bail;
1072	}
1073
1074	phys = dd->ipath_physaddr + piobufs;
1075
1076#if defined(__powerpc__)
1077	/* There isn't a generic way to specify writethrough mappings */
1078	pgprot_val(vma->vm_page_prot) |= _PAGE_NO_CACHE;
1079	pgprot_val(vma->vm_page_prot) |= _PAGE_WRITETHRU;
1080	pgprot_val(vma->vm_page_prot) &= ~_PAGE_GUARDED;
1081#endif
1082
1083	/*
1084	 * don't allow them to later change to readable with mprotect (for when
1085	 * not initially mapped readable, as is normally the case)
1086	 */
1087	vma->vm_flags &= ~VM_MAYREAD;
1088	vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
1089
1090	ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
1091				 vma->vm_end - vma->vm_start,
1092				 vma->vm_page_prot);
1093bail:
1094	return ret;
1095}
1096
1097static int mmap_rcvegrbufs(struct vm_area_struct *vma,
1098			   struct ipath_portdata *pd)
1099{
1100	struct ipath_devdata *dd = pd->port_dd;
1101	unsigned long start, size;
1102	size_t total_size, i;
1103	unsigned long pfn;
1104	int ret;
1105
1106	size = pd->port_rcvegrbuf_size;
1107	total_size = pd->port_rcvegrbuf_chunks * size;
1108	if ((vma->vm_end - vma->vm_start) > total_size) {
1109		dev_info(&dd->pcidev->dev, "FAIL on egr bufs: "
1110			 "reqlen %lx > actual %lx\n",
1111			 vma->vm_end - vma->vm_start,
1112			 (unsigned long) total_size);
1113		ret = -EINVAL;
1114		goto bail;
1115	}
1116
1117	if (vma->vm_flags & VM_WRITE) {
1118		dev_info(&dd->pcidev->dev, "Can't map eager buffers as "
1119			 "writable (flags=%lx)\n", vma->vm_flags);
1120		ret = -EPERM;
1121		goto bail;
1122	}
1123	/* don't allow them to later change to writeable with mprotect */
1124	vma->vm_flags &= ~VM_MAYWRITE;
1125
1126	start = vma->vm_start;
1127
1128	for (i = 0; i < pd->port_rcvegrbuf_chunks; i++, start += size) {
1129		pfn = virt_to_phys(pd->port_rcvegrbuf[i]) >> PAGE_SHIFT;
1130		ret = remap_pfn_range(vma, start, pfn, size,
1131				      vma->vm_page_prot);
1132		if (ret < 0)
1133			goto bail;
1134	}
1135	ret = 0;
1136
1137bail:
1138	return ret;
1139}
1140
1141/*
1142 * ipath_file_vma_fault - handle a VMA page fault.
1143 */
1144static int ipath_file_vma_fault(struct vm_area_struct *vma,
1145					struct vm_fault *vmf)
1146{
1147	struct page *page;
1148
1149	page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
1150	if (!page)
1151		return VM_FAULT_SIGBUS;
1152	get_page(page);
1153	vmf->page = page;
1154
1155	return 0;
1156}
1157
1158static const struct vm_operations_struct ipath_file_vm_ops = {
1159	.fault = ipath_file_vma_fault,
1160};
1161
1162static int mmap_kvaddr(struct vm_area_struct *vma, u64 pgaddr,
1163		       struct ipath_portdata *pd, unsigned subport)
1164{
1165	unsigned long len;
1166	struct ipath_devdata *dd;
1167	void *addr;
1168	size_t size;
1169	int ret = 0;
1170
1171	/* If the port is not shared, all addresses should be physical */
1172	if (!pd->port_subport_cnt)
1173		goto bail;
1174
1175	dd = pd->port_dd;
1176	size = pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size;
1177
1178	/*
1179	 * Each process has all the subport uregbase, rcvhdrq, and
1180	 * rcvegrbufs mmapped - as an array for all the processes,
1181	 * and also separately for this process.
1182	 */
1183	if (pgaddr == cvt_kvaddr(pd->subport_uregbase)) {
1184		addr = pd->subport_uregbase;
1185		size = PAGE_SIZE * pd->port_subport_cnt;
1186	} else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base)) {
1187		addr = pd->subport_rcvhdr_base;
1188		size = pd->port_rcvhdrq_size * pd->port_subport_cnt;
1189	} else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf)) {
1190		addr = pd->subport_rcvegrbuf;
1191		size *= pd->port_subport_cnt;
1192        } else if (pgaddr == cvt_kvaddr(pd->subport_uregbase +
1193                                        PAGE_SIZE * subport)) {
1194                addr = pd->subport_uregbase + PAGE_SIZE * subport;
1195                size = PAGE_SIZE;
1196        } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base +
1197                                pd->port_rcvhdrq_size * subport)) {
1198                addr = pd->subport_rcvhdr_base +
1199                        pd->port_rcvhdrq_size * subport;
1200                size = pd->port_rcvhdrq_size;
1201        } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf +
1202                               size * subport)) {
1203                addr = pd->subport_rcvegrbuf + size * subport;
1204                /* rcvegrbufs are read-only on the slave */
1205                if (vma->vm_flags & VM_WRITE) {
1206                        dev_info(&dd->pcidev->dev,
1207                                 "Can't map eager buffers as "
1208                                 "writable (flags=%lx)\n", vma->vm_flags);
1209                        ret = -EPERM;
1210                        goto bail;
1211                }
1212                /*
1213                 * Don't allow permission to later change to writeable
1214                 * with mprotect.
1215                 */
1216                vma->vm_flags &= ~VM_MAYWRITE;
1217	} else {
1218		goto bail;
1219	}
1220	len = vma->vm_end - vma->vm_start;
1221	if (len > size) {
1222		ipath_cdbg(MM, "FAIL: reqlen %lx > %zx\n", len, size);
1223		ret = -EINVAL;
1224		goto bail;
1225	}
1226
1227	vma->vm_pgoff = (unsigned long) addr >> PAGE_SHIFT;
1228	vma->vm_ops = &ipath_file_vm_ops;
1229	vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
1230	ret = 1;
1231
1232bail:
1233	return ret;
1234}
1235
1236/**
1237 * ipath_mmap - mmap various structures into user space
1238 * @fp: the file pointer
1239 * @vma: the VM area
1240 *
1241 * We use this to have a shared buffer between the kernel and the user code
1242 * for the rcvhdr queue, egr buffers, and the per-port user regs and pio
1243 * buffers in the chip.  We have the open and close entries so we can bump
1244 * the ref count and keep the driver from being unloaded while still mapped.
1245 */
1246static int ipath_mmap(struct file *fp, struct vm_area_struct *vma)
1247{
1248	struct ipath_portdata *pd;
1249	struct ipath_devdata *dd;
1250	u64 pgaddr, ureg;
1251	unsigned piobufs, piocnt;
1252	int ret;
1253
1254	pd = port_fp(fp);
1255	if (!pd) {
1256		ret = -EINVAL;
1257		goto bail;
1258	}
1259	dd = pd->port_dd;
1260
1261	/*
1262	 * This is the ipath_do_user_init() code, mapping the shared buffers
1263	 * into the user process. The address referred to by vm_pgoff is the
1264	 * file offset passed via mmap().  For shared ports, this is the
1265	 * kernel vmalloc() address of the pages to share with the master.
1266	 * For non-shared or master ports, this is a physical address.
1267	 * We only do one mmap for each space mapped.
1268	 */
1269	pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1270
1271	/*
1272	 * Check for 0 in case one of the allocations failed, but user
1273	 * called mmap anyway.
1274	 */
1275	if (!pgaddr)  {
1276		ret = -EINVAL;
1277		goto bail;
1278	}
1279
1280	ipath_cdbg(MM, "pgaddr %llx vm_start=%lx len %lx port %u:%u:%u\n",
1281		   (unsigned long long) pgaddr, vma->vm_start,
1282		   vma->vm_end - vma->vm_start, dd->ipath_unit,
1283		   pd->port_port, subport_fp(fp));
1284
1285	/*
1286	 * Physical addresses must fit in 40 bits for our hardware.
1287	 * Check for kernel virtual addresses first, anything else must
1288	 * match a HW or memory address.
1289	 */
1290	ret = mmap_kvaddr(vma, pgaddr, pd, subport_fp(fp));
1291	if (ret) {
1292		if (ret > 0)
1293			ret = 0;
1294		goto bail;
1295	}
1296
1297	ureg = dd->ipath_uregbase + dd->ipath_ureg_align * pd->port_port;
1298	if (!pd->port_subport_cnt) {
1299		/* port is not shared */
1300		piocnt = pd->port_piocnt;
1301		piobufs = pd->port_piobufs;
1302	} else if (!subport_fp(fp)) {
1303		/* caller is the master */
1304		piocnt = (pd->port_piocnt / pd->port_subport_cnt) +
1305			 (pd->port_piocnt % pd->port_subport_cnt);
1306		piobufs = pd->port_piobufs +
1307			dd->ipath_palign * (pd->port_piocnt - piocnt);
1308	} else {
1309		unsigned slave = subport_fp(fp) - 1;
1310
1311		/* caller is a slave */
1312		piocnt = pd->port_piocnt / pd->port_subport_cnt;
1313		piobufs = pd->port_piobufs + dd->ipath_palign * piocnt * slave;
1314	}
1315
1316	if (pgaddr == ureg)
1317		ret = mmap_ureg(vma, dd, ureg);
1318	else if (pgaddr == piobufs)
1319		ret = mmap_piobufs(vma, dd, pd, piobufs, piocnt);
1320	else if (pgaddr == dd->ipath_pioavailregs_phys)
1321		/* in-memory copy of pioavail registers */
1322		ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1323			      	     (void *) dd->ipath_pioavailregs_dma,
1324				     "pioavail registers");
1325	else if (pgaddr == pd->port_rcvegr_phys)
1326		ret = mmap_rcvegrbufs(vma, pd);
1327	else if (pgaddr == (u64) pd->port_rcvhdrq_phys)
1328		/*
1329		 * The rcvhdrq itself; readonly except on HT (so have
1330		 * to allow writable mapping), multiple pages, contiguous
1331		 * from an i/o perspective.
1332		 */
1333		ret = ipath_mmap_mem(vma, pd, pd->port_rcvhdrq_size, 1,
1334				     pd->port_rcvhdrq,
1335				     "rcvhdrq");
1336	else if (pgaddr == (u64) pd->port_rcvhdrqtailaddr_phys)
1337		/* in-memory copy of rcvhdrq tail register */
1338		ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0,
1339				     pd->port_rcvhdrtail_kvaddr,
1340				     "rcvhdrq tail");
1341	else
1342		ret = -EINVAL;
1343
1344	vma->vm_private_data = NULL;
1345
1346	if (ret < 0)
1347		dev_info(&dd->pcidev->dev,
1348			 "Failure %d on off %llx len %lx\n",
1349			 -ret, (unsigned long long)pgaddr,
1350			 vma->vm_end - vma->vm_start);
1351bail:
1352	return ret;
1353}
1354
1355static unsigned ipath_poll_hdrqfull(struct ipath_portdata *pd)
1356{
1357	unsigned pollflag = 0;
1358
1359	if ((pd->poll_type & IPATH_POLL_TYPE_OVERFLOW) &&
1360	    pd->port_hdrqfull != pd->port_hdrqfull_poll) {
1361		pollflag |= POLLIN | POLLRDNORM;
1362		pd->port_hdrqfull_poll = pd->port_hdrqfull;
1363	}
1364
1365	return pollflag;
1366}
1367
1368static unsigned int ipath_poll_urgent(struct ipath_portdata *pd,
1369				      struct file *fp,
1370				      struct poll_table_struct *pt)
1371{
1372	unsigned pollflag = 0;
1373	struct ipath_devdata *dd;
1374
1375	dd = pd->port_dd;
1376
1377	/* variable access in ipath_poll_hdrqfull() needs this */
1378	rmb();
1379	pollflag = ipath_poll_hdrqfull(pd);
1380
1381	if (pd->port_urgent != pd->port_urgent_poll) {
1382		pollflag |= POLLIN | POLLRDNORM;
1383		pd->port_urgent_poll = pd->port_urgent;
1384	}
1385
1386	if (!pollflag) {
1387		/* this saves a spin_lock/unlock in interrupt handler... */
1388		set_bit(IPATH_PORT_WAITING_URG, &pd->port_flag);
1389		/* flush waiting flag so don't miss an event... */
1390		wmb();
1391		poll_wait(fp, &pd->port_wait, pt);
1392	}
1393
1394	return pollflag;
1395}
1396
1397static unsigned int ipath_poll_next(struct ipath_portdata *pd,
1398				    struct file *fp,
1399				    struct poll_table_struct *pt)
1400{
1401	u32 head;
1402	u32 tail;
1403	unsigned pollflag = 0;
1404	struct ipath_devdata *dd;
1405
1406	dd = pd->port_dd;
1407
1408	/* variable access in ipath_poll_hdrqfull() needs this */
1409	rmb();
1410	pollflag = ipath_poll_hdrqfull(pd);
1411
1412	head = ipath_read_ureg32(dd, ur_rcvhdrhead, pd->port_port);
1413	if (pd->port_rcvhdrtail_kvaddr)
1414		tail = ipath_get_rcvhdrtail(pd);
1415	else
1416		tail = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port);
1417
1418	if (head != tail)
1419		pollflag |= POLLIN | POLLRDNORM;
1420	else {
1421		/* this saves a spin_lock/unlock in interrupt handler */
1422		set_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag);
1423		/* flush waiting flag so we don't miss an event */
1424		wmb();
1425
1426		set_bit(pd->port_port + dd->ipath_r_intravail_shift,
1427			&dd->ipath_rcvctrl);
1428
1429		ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1430				 dd->ipath_rcvctrl);
1431
1432		if (dd->ipath_rhdrhead_intr_off) /* arm rcv interrupt */
1433			ipath_write_ureg(dd, ur_rcvhdrhead,
1434					 dd->ipath_rhdrhead_intr_off | head,
1435					 pd->port_port);
1436
1437		poll_wait(fp, &pd->port_wait, pt);
1438	}
1439
1440	return pollflag;
1441}
1442
1443static unsigned int ipath_poll(struct file *fp,
1444			       struct poll_table_struct *pt)
1445{
1446	struct ipath_portdata *pd;
1447	unsigned pollflag;
1448
1449	pd = port_fp(fp);
1450	if (!pd)
1451		pollflag = 0;
1452	else if (pd->poll_type & IPATH_POLL_TYPE_URGENT)
1453		pollflag = ipath_poll_urgent(pd, fp, pt);
1454	else
1455		pollflag = ipath_poll_next(pd, fp, pt);
1456
1457	return pollflag;
1458}
1459
1460static int ipath_supports_subports(int user_swmajor, int user_swminor)
1461{
1462	/* no subport implementation prior to software version 1.3 */
1463	return (user_swmajor > 1) || (user_swminor >= 3);
1464}
1465
1466static int ipath_compatible_subports(int user_swmajor, int user_swminor)
1467{
1468	/* this code is written long-hand for clarity */
1469	if (IPATH_USER_SWMAJOR != user_swmajor) {
1470		/* no promise of compatibility if major mismatch */
1471		return 0;
1472	}
1473	if (IPATH_USER_SWMAJOR == 1) {
1474		switch (IPATH_USER_SWMINOR) {
1475		case 0:
1476		case 1:
1477		case 2:
1478			/* no subport implementation so cannot be compatible */
1479			return 0;
1480		case 3:
1481			/* 3 is only compatible with itself */
1482			return user_swminor == 3;
1483		default:
1484			/* >= 4 are compatible (or are expected to be) */
1485			return user_swminor >= 4;
1486		}
1487	}
1488	/* make no promises yet for future major versions */
1489	return 0;
1490}
1491
1492static int init_subports(struct ipath_devdata *dd,
1493			 struct ipath_portdata *pd,
1494			 const struct ipath_user_info *uinfo)
1495{
1496	int ret = 0;
1497	unsigned num_subports;
1498	size_t size;
1499
1500	/*
1501	 * If the user is requesting zero subports,
1502	 * skip the subport allocation.
1503	 */
1504	if (uinfo->spu_subport_cnt <= 0)
1505		goto bail;
1506
1507	/* Self-consistency check for ipath_compatible_subports() */
1508	if (ipath_supports_subports(IPATH_USER_SWMAJOR, IPATH_USER_SWMINOR) &&
1509	    !ipath_compatible_subports(IPATH_USER_SWMAJOR,
1510				       IPATH_USER_SWMINOR)) {
1511		dev_info(&dd->pcidev->dev,
1512			 "Inconsistent ipath_compatible_subports()\n");
1513		goto bail;
1514	}
1515
1516	/* Check for subport compatibility */
1517	if (!ipath_compatible_subports(uinfo->spu_userversion >> 16,
1518				       uinfo->spu_userversion & 0xffff)) {
1519		dev_info(&dd->pcidev->dev,
1520			 "Mismatched user version (%d.%d) and driver "
1521			 "version (%d.%d) while port sharing. Ensure "
1522                         "that driver and library are from the same "
1523                         "release.\n",
1524			 (int) (uinfo->spu_userversion >> 16),
1525                         (int) (uinfo->spu_userversion & 0xffff),
1526			 IPATH_USER_SWMAJOR,
1527	                 IPATH_USER_SWMINOR);
1528		goto bail;
1529	}
1530	if (uinfo->spu_subport_cnt > INFINIPATH_MAX_SUBPORT) {
1531		ret = -EINVAL;
1532		goto bail;
1533	}
1534
1535	num_subports = uinfo->spu_subport_cnt;
1536	pd->subport_uregbase = vzalloc(PAGE_SIZE * num_subports);
1537	if (!pd->subport_uregbase) {
1538		ret = -ENOMEM;
1539		goto bail;
1540	}
1541	/* Note: pd->port_rcvhdrq_size isn't initialized yet. */
1542	size = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize *
1543		     sizeof(u32), PAGE_SIZE) * num_subports;
1544	pd->subport_rcvhdr_base = vzalloc(size);
1545	if (!pd->subport_rcvhdr_base) {
1546		ret = -ENOMEM;
1547		goto bail_ureg;
1548	}
1549
1550	pd->subport_rcvegrbuf = vzalloc(pd->port_rcvegrbuf_chunks *
1551					pd->port_rcvegrbuf_size *
1552					num_subports);
1553	if (!pd->subport_rcvegrbuf) {
1554		ret = -ENOMEM;
1555		goto bail_rhdr;
1556	}
1557
1558	pd->port_subport_cnt = uinfo->spu_subport_cnt;
1559	pd->port_subport_id = uinfo->spu_subport_id;
1560	pd->active_slaves = 1;
1561	set_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
1562	goto bail;
1563
1564bail_rhdr:
1565	vfree(pd->subport_rcvhdr_base);
1566bail_ureg:
1567	vfree(pd->subport_uregbase);
1568	pd->subport_uregbase = NULL;
1569bail:
1570	return ret;
1571}
1572
1573static int try_alloc_port(struct ipath_devdata *dd, int port,
1574			  struct file *fp,
1575			  const struct ipath_user_info *uinfo)
1576{
1577	struct ipath_portdata *pd;
1578	int ret;
1579
1580	if (!(pd = dd->ipath_pd[port])) {
1581		void *ptmp;
1582
1583		pd = kzalloc(sizeof(struct ipath_portdata), GFP_KERNEL);
1584
1585		/*
1586		 * Allocate memory for use in ipath_tid_update() just once
1587		 * at open, not per call.  Reduces cost of expected send
1588		 * setup.
1589		 */
1590		ptmp = kmalloc(dd->ipath_rcvtidcnt * sizeof(u16) +
1591			       dd->ipath_rcvtidcnt * sizeof(struct page **),
1592			       GFP_KERNEL);
1593		if (!pd || !ptmp) {
1594			ipath_dev_err(dd, "Unable to allocate portdata "
1595				      "memory, failing open\n");
1596			ret = -ENOMEM;
1597			kfree(pd);
1598			kfree(ptmp);
1599			goto bail;
1600		}
1601		dd->ipath_pd[port] = pd;
1602		dd->ipath_pd[port]->port_port = port;
1603		dd->ipath_pd[port]->port_dd = dd;
1604		dd->ipath_pd[port]->port_tid_pg_list = ptmp;
1605		init_waitqueue_head(&dd->ipath_pd[port]->port_wait);
1606	}
1607	if (!pd->port_cnt) {
1608		pd->userversion = uinfo->spu_userversion;
1609		init_user_egr_sizes(pd);
1610		if ((ret = init_subports(dd, pd, uinfo)) != 0)
1611			goto bail;
1612		ipath_cdbg(PROC, "%s[%u] opened unit:port %u:%u\n",
1613			   current->comm, current->pid, dd->ipath_unit,
1614			   port);
1615		pd->port_cnt = 1;
1616		port_fp(fp) = pd;
1617		pd->port_pid = get_pid(task_pid(current));
1618		strlcpy(pd->port_comm, current->comm, sizeof(pd->port_comm));
1619		ipath_stats.sps_ports++;
1620		ret = 0;
1621	} else
1622		ret = -EBUSY;
1623
1624bail:
1625	return ret;
1626}
1627
1628static inline int usable(struct ipath_devdata *dd)
1629{
1630	return dd &&
1631		(dd->ipath_flags & IPATH_PRESENT) &&
1632		dd->ipath_kregbase &&
1633		dd->ipath_lid &&
1634		!(dd->ipath_flags & (IPATH_LINKDOWN | IPATH_DISABLED
1635				     | IPATH_LINKUNK));
1636}
1637
1638static int find_free_port(int unit, struct file *fp,
1639			  const struct ipath_user_info *uinfo)
1640{
1641	struct ipath_devdata *dd = ipath_lookup(unit);
1642	int ret, i;
1643
1644	if (!dd) {
1645		ret = -ENODEV;
1646		goto bail;
1647	}
1648
1649	if (!usable(dd)) {
1650		ret = -ENETDOWN;
1651		goto bail;
1652	}
1653
1654	for (i = 1; i < dd->ipath_cfgports; i++) {
1655		ret = try_alloc_port(dd, i, fp, uinfo);
1656		if (ret != -EBUSY)
1657			goto bail;
1658	}
1659	ret = -EBUSY;
1660
1661bail:
1662	return ret;
1663}
1664
1665static int find_best_unit(struct file *fp,
1666			  const struct ipath_user_info *uinfo)
1667{
1668	int ret = 0, i, prefunit = -1, devmax;
1669	int maxofallports, npresent, nup;
1670	int ndev;
1671
1672	devmax = ipath_count_units(&npresent, &nup, &maxofallports);
1673
1674	/*
1675	 * This code is present to allow a knowledgeable person to
1676	 * specify the layout of processes to processors before opening
1677	 * this driver, and then we'll assign the process to the "closest"
1678	 * InfiniPath chip to that processor (we assume reasonable connectivity,
1679	 * for now).  This code assumes that if affinity has been set
1680	 * before this point, that at most one cpu is set; for now this
1681	 * is reasonable.  I check for both cpumask_empty() and cpumask_full(),
1682	 * in case some kernel variant sets none of the bits when no
1683	 * affinity is set.  2.6.11 and 12 kernels have all present
1684	 * cpus set.  Some day we'll have to fix it up further to handle
1685	 * a cpu subset.  This algorithm fails for two HT chips connected
1686	 * in tunnel fashion.  Eventually this needs real topology
1687	 * information.  There may be some issues with dual core numbering
1688	 * as well.  This needs more work prior to release.
1689	 */
1690	if (!cpumask_empty(tsk_cpus_allowed(current)) &&
1691	    !cpumask_full(tsk_cpus_allowed(current))) {
1692		int ncpus = num_online_cpus(), curcpu = -1, nset = 0;
1693		get_online_cpus();
1694		for_each_online_cpu(i)
1695			if (cpumask_test_cpu(i, tsk_cpus_allowed(current))) {
1696				ipath_cdbg(PROC, "%s[%u] affinity set for "
1697					   "cpu %d/%d\n", current->comm,
1698					   current->pid, i, ncpus);
1699				curcpu = i;
1700				nset++;
1701			}
1702		put_online_cpus();
1703		if (curcpu != -1 && nset != ncpus) {
1704			if (npresent) {
1705				prefunit = curcpu / (ncpus / npresent);
1706				ipath_cdbg(PROC,"%s[%u] %d chips, %d cpus, "
1707					  "%d cpus/chip, select unit %d\n",
1708					  current->comm, current->pid,
1709					  npresent, ncpus, ncpus / npresent,
1710					  prefunit);
1711			}
1712		}
1713	}
1714
1715	/*
1716	 * user ports start at 1, kernel port is 0
1717	 * For now, we do round-robin access across all chips
1718	 */
1719
1720	if (prefunit != -1)
1721		devmax = prefunit + 1;
1722recheck:
1723	for (i = 1; i < maxofallports; i++) {
1724		for (ndev = prefunit != -1 ? prefunit : 0; ndev < devmax;
1725		     ndev++) {
1726			struct ipath_devdata *dd = ipath_lookup(ndev);
1727
1728			if (!usable(dd))
1729				continue; /* can't use this unit */
1730			if (i >= dd->ipath_cfgports)
1731				/*
1732				 * Maxed out on users of this unit. Try
1733				 * next.
1734				 */
1735				continue;
1736			ret = try_alloc_port(dd, i, fp, uinfo);
1737			if (!ret)
1738				goto done;
1739		}
1740	}
1741
1742	if (npresent) {
1743		if (nup == 0) {
1744			ret = -ENETDOWN;
1745			ipath_dbg("No ports available (none initialized "
1746				  "and ready)\n");
1747		} else {
1748			if (prefunit > 0) {
1749				/* if started above 0, retry from 0 */
1750				ipath_cdbg(PROC,
1751					   "%s[%u] no ports on prefunit "
1752					   "%d, clear and re-check\n",
1753					   current->comm, current->pid,
1754					   prefunit);
1755				devmax = ipath_count_units(NULL, NULL,
1756							   NULL);
1757				prefunit = -1;
1758				goto recheck;
1759			}
1760			ret = -EBUSY;
1761			ipath_dbg("No ports available\n");
1762		}
1763	} else {
1764		ret = -ENXIO;
1765		ipath_dbg("No boards found\n");
1766	}
1767
1768done:
1769	return ret;
1770}
1771
1772static int find_shared_port(struct file *fp,
1773			    const struct ipath_user_info *uinfo)
1774{
1775	int devmax, ndev, i;
1776	int ret = 0;
1777
1778	devmax = ipath_count_units(NULL, NULL, NULL);
1779
1780	for (ndev = 0; ndev < devmax; ndev++) {
1781		struct ipath_devdata *dd = ipath_lookup(ndev);
1782
1783		if (!usable(dd))
1784			continue;
1785		for (i = 1; i < dd->ipath_cfgports; i++) {
1786			struct ipath_portdata *pd = dd->ipath_pd[i];
1787
1788			/* Skip ports which are not yet open */
1789			if (!pd || !pd->port_cnt)
1790				continue;
1791			/* Skip port if it doesn't match the requested one */
1792			if (pd->port_subport_id != uinfo->spu_subport_id)
1793				continue;
1794			/* Verify the sharing process matches the master */
1795			if (pd->port_subport_cnt != uinfo->spu_subport_cnt ||
1796			    pd->userversion != uinfo->spu_userversion ||
1797			    pd->port_cnt >= pd->port_subport_cnt) {
1798				ret = -EINVAL;
1799				goto done;
1800			}
1801			port_fp(fp) = pd;
1802			subport_fp(fp) = pd->port_cnt++;
1803			pd->port_subpid[subport_fp(fp)] =
1804				get_pid(task_pid(current));
1805			tidcursor_fp(fp) = 0;
1806			pd->active_slaves |= 1 << subport_fp(fp);
1807			ipath_cdbg(PROC,
1808				   "%s[%u] %u sharing %s[%u] unit:port %u:%u\n",
1809				   current->comm, current->pid,
1810				   subport_fp(fp),
1811				   pd->port_comm, pid_nr(pd->port_pid),
1812				   dd->ipath_unit, pd->port_port);
1813			ret = 1;
1814			goto done;
1815		}
1816	}
1817
1818done:
1819	return ret;
1820}
1821
1822static int ipath_open(struct inode *in, struct file *fp)
1823{
1824	/* The real work is performed later in ipath_assign_port() */
1825	fp->private_data = kzalloc(sizeof(struct ipath_filedata), GFP_KERNEL);
1826	return fp->private_data ? 0 : -ENOMEM;
1827}
1828
1829/* Get port early, so can set affinity prior to memory allocation */
1830static int ipath_assign_port(struct file *fp,
1831			      const struct ipath_user_info *uinfo)
1832{
1833	int ret;
1834	int i_minor;
1835	unsigned swmajor, swminor;
1836
1837	/* Check to be sure we haven't already initialized this file */
1838	if (port_fp(fp)) {
1839		ret = -EINVAL;
1840		goto done;
1841	}
1842
1843	/* for now, if major version is different, bail */
1844	swmajor = uinfo->spu_userversion >> 16;
1845	if (swmajor != IPATH_USER_SWMAJOR) {
1846		ipath_dbg("User major version %d not same as driver "
1847			  "major %d\n", uinfo->spu_userversion >> 16,
1848			  IPATH_USER_SWMAJOR);
1849		ret = -ENODEV;
1850		goto done;
1851	}
1852
1853	swminor = uinfo->spu_userversion & 0xffff;
1854	if (swminor != IPATH_USER_SWMINOR)
1855		ipath_dbg("User minor version %d not same as driver "
1856			  "minor %d\n", swminor, IPATH_USER_SWMINOR);
1857
1858	mutex_lock(&ipath_mutex);
1859
1860	if (ipath_compatible_subports(swmajor, swminor) &&
1861	    uinfo->spu_subport_cnt &&
1862	    (ret = find_shared_port(fp, uinfo))) {
1863		if (ret > 0)
1864			ret = 0;
1865		goto done_chk_sdma;
1866	}
1867
1868	i_minor = iminor(file_inode(fp)) - IPATH_USER_MINOR_BASE;
1869	ipath_cdbg(VERBOSE, "open on dev %lx (minor %d)\n",
1870		   (long)file_inode(fp)->i_rdev, i_minor);
1871
1872	if (i_minor)
1873		ret = find_free_port(i_minor - 1, fp, uinfo);
1874	else
1875		ret = find_best_unit(fp, uinfo);
1876
1877done_chk_sdma:
1878	if (!ret) {
1879		struct ipath_filedata *fd = fp->private_data;
1880		const struct ipath_portdata *pd = fd->pd;
1881		const struct ipath_devdata *dd = pd->port_dd;
1882
1883		fd->pq = ipath_user_sdma_queue_create(&dd->pcidev->dev,
1884						      dd->ipath_unit,
1885						      pd->port_port,
1886						      fd->subport);
1887
1888		if (!fd->pq)
1889			ret = -ENOMEM;
1890	}
1891
1892	mutex_unlock(&ipath_mutex);
1893
1894done:
1895	return ret;
1896}
1897
1898
1899static int ipath_do_user_init(struct file *fp,
1900			      const struct ipath_user_info *uinfo)
1901{
1902	int ret;
1903	struct ipath_portdata *pd = port_fp(fp);
1904	struct ipath_devdata *dd;
1905	u32 head32;
1906
1907	/* Subports don't need to initialize anything since master did it. */
1908	if (subport_fp(fp)) {
1909		ret = wait_event_interruptible(pd->port_wait,
1910			!test_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag));
1911		goto done;
1912	}
1913
1914	dd = pd->port_dd;
1915
1916	if (uinfo->spu_rcvhdrsize) {
1917		ret = ipath_setrcvhdrsize(dd, uinfo->spu_rcvhdrsize);
1918		if (ret)
1919			goto done;
1920	}
1921
1922	/* for now we do nothing with rcvhdrcnt: uinfo->spu_rcvhdrcnt */
1923
1924	/* some ports may get extra buffers, calculate that here */
1925	if (pd->port_port <= dd->ipath_ports_extrabuf)
1926		pd->port_piocnt = dd->ipath_pbufsport + 1;
1927	else
1928		pd->port_piocnt = dd->ipath_pbufsport;
1929
1930	/* for right now, kernel piobufs are at end, so port 1 is at 0 */
1931	if (pd->port_port <= dd->ipath_ports_extrabuf)
1932		pd->port_pio_base = (dd->ipath_pbufsport + 1)
1933			* (pd->port_port - 1);
1934	else
1935		pd->port_pio_base = dd->ipath_ports_extrabuf +
1936			dd->ipath_pbufsport * (pd->port_port - 1);
1937	pd->port_piobufs = dd->ipath_piobufbase +
1938		pd->port_pio_base * dd->ipath_palign;
1939	ipath_cdbg(VERBOSE, "piobuf base for port %u is 0x%x, piocnt %u,"
1940		" first pio %u\n", pd->port_port, pd->port_piobufs,
1941		pd->port_piocnt, pd->port_pio_base);
1942	ipath_chg_pioavailkernel(dd, pd->port_pio_base, pd->port_piocnt, 0);
1943
1944	/*
1945	 * Now allocate the rcvhdr Q and eager TIDs; skip the TID
1946	 * array for time being.  If pd->port_port > chip-supported,
1947	 * we need to do extra stuff here to handle by handling overflow
1948	 * through port 0, someday
1949	 */
1950	ret = ipath_create_rcvhdrq(dd, pd);
1951	if (!ret)
1952		ret = ipath_create_user_egr(pd);
1953	if (ret)
1954		goto done;
1955
1956	/*
1957	 * set the eager head register for this port to the current values
1958	 * of the tail pointers, since we don't know if they were
1959	 * updated on last use of the port.
1960	 */
1961	head32 = ipath_read_ureg32(dd, ur_rcvegrindextail, pd->port_port);
1962	ipath_write_ureg(dd, ur_rcvegrindexhead, head32, pd->port_port);
1963	pd->port_lastrcvhdrqtail = -1;
1964	ipath_cdbg(VERBOSE, "Wrote port%d egrhead %x from tail regs\n",
1965		pd->port_port, head32);
1966	pd->port_tidcursor = 0;	/* start at beginning after open */
1967
1968	/* initialize poll variables... */
1969	pd->port_urgent = 0;
1970	pd->port_urgent_poll = 0;
1971	pd->port_hdrqfull_poll = pd->port_hdrqfull;
1972
1973	/*
1974	 * Now enable the port for receive.
1975	 * For chips that are set to DMA the tail register to memory
1976	 * when they change (and when the update bit transitions from
1977	 * 0 to 1.  So for those chips, we turn it off and then back on.
1978	 * This will (very briefly) affect any other open ports, but the
1979	 * duration is very short, and therefore isn't an issue.  We
1980	 * explicitly set the in-memory tail copy to 0 beforehand, so we
1981	 * don't have to wait to be sure the DMA update has happened
1982	 * (chip resets head/tail to 0 on transition to enable).
1983	 */
1984	set_bit(dd->ipath_r_portenable_shift + pd->port_port,
1985		&dd->ipath_rcvctrl);
1986	if (!(dd->ipath_flags & IPATH_NODMA_RTAIL)) {
1987		if (pd->port_rcvhdrtail_kvaddr)
1988			ipath_clear_rcvhdrtail(pd);
1989		ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1990			dd->ipath_rcvctrl &
1991			~(1ULL << dd->ipath_r_tailupd_shift));
1992	}
1993	ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl,
1994			 dd->ipath_rcvctrl);
1995	/* Notify any waiting slaves */
1996	if (pd->port_subport_cnt) {
1997		clear_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag);
1998		wake_up(&pd->port_wait);
1999	}
2000done:
2001	return ret;
2002}
2003
2004/**
2005 * unlock_exptid - unlock any expected TID entries port still had in use
2006 * @pd: port
2007 *
2008 * We don't actually update the chip here, because we do a bulk update
2009 * below, using ipath_f_clear_tids.
2010 */
2011static void unlock_expected_tids(struct ipath_portdata *pd)
2012{
2013	struct ipath_devdata *dd = pd->port_dd;
2014	int port_tidbase = pd->port_port * dd->ipath_rcvtidcnt;
2015	int i, cnt = 0, maxtid = port_tidbase + dd->ipath_rcvtidcnt;
2016
2017	ipath_cdbg(VERBOSE, "Port %u unlocking any locked expTID pages\n",
2018		   pd->port_port);
2019	for (i = port_tidbase; i < maxtid; i++) {
2020		struct page *ps = dd->ipath_pageshadow[i];
2021
2022		if (!ps)
2023			continue;
2024
2025		dd->ipath_pageshadow[i] = NULL;
2026		pci_unmap_page(dd->pcidev, dd->ipath_physshadow[i],
2027			PAGE_SIZE, PCI_DMA_FROMDEVICE);
2028		ipath_release_user_pages_on_close(&ps, 1);
2029		cnt++;
2030		ipath_stats.sps_pageunlocks++;
2031	}
2032	if (cnt)
2033		ipath_cdbg(VERBOSE, "Port %u locked %u expTID entries\n",
2034			   pd->port_port, cnt);
2035
2036	if (ipath_stats.sps_pagelocks || ipath_stats.sps_pageunlocks)
2037		ipath_cdbg(VERBOSE, "%llu pages locked, %llu unlocked\n",
2038			   (unsigned long long) ipath_stats.sps_pagelocks,
2039			   (unsigned long long)
2040			   ipath_stats.sps_pageunlocks);
2041}
2042
2043static int ipath_close(struct inode *in, struct file *fp)
2044{
2045	int ret = 0;
2046	struct ipath_filedata *fd;
2047	struct ipath_portdata *pd;
2048	struct ipath_devdata *dd;
2049	unsigned long flags;
2050	unsigned port;
2051	struct pid *pid;
2052
2053	ipath_cdbg(VERBOSE, "close on dev %lx, private data %p\n",
2054		   (long)in->i_rdev, fp->private_data);
2055
2056	mutex_lock(&ipath_mutex);
2057
2058	fd = fp->private_data;
2059	fp->private_data = NULL;
2060	pd = fd->pd;
2061	if (!pd) {
2062		mutex_unlock(&ipath_mutex);
2063		goto bail;
2064	}
2065
2066	dd = pd->port_dd;
2067
2068	/* drain user sdma queue */
2069	ipath_user_sdma_queue_drain(dd, fd->pq);
2070	ipath_user_sdma_queue_destroy(fd->pq);
2071
2072	if (--pd->port_cnt) {
2073		/*
2074		 * XXX If the master closes the port before the slave(s),
2075		 * revoke the mmap for the eager receive queue so
2076		 * the slave(s) don't wait for receive data forever.
2077		 */
2078		pd->active_slaves &= ~(1 << fd->subport);
2079		put_pid(pd->port_subpid[fd->subport]);
2080		pd->port_subpid[fd->subport] = NULL;
2081		mutex_unlock(&ipath_mutex);
2082		goto bail;
2083	}
2084	/* early; no interrupt users after this */
2085	spin_lock_irqsave(&dd->ipath_uctxt_lock, flags);
2086	port = pd->port_port;
2087	dd->ipath_pd[port] = NULL;
2088	pid = pd->port_pid;
2089	pd->port_pid = NULL;
2090	spin_unlock_irqrestore(&dd->ipath_uctxt_lock, flags);
2091
2092	if (pd->port_rcvwait_to || pd->port_piowait_to
2093	    || pd->port_rcvnowait || pd->port_pionowait) {
2094		ipath_cdbg(VERBOSE, "port%u, %u rcv, %u pio wait timeo; "
2095			   "%u rcv %u, pio already\n",
2096			   pd->port_port, pd->port_rcvwait_to,
2097			   pd->port_piowait_to, pd->port_rcvnowait,
2098			   pd->port_pionowait);
2099		pd->port_rcvwait_to = pd->port_piowait_to =
2100			pd->port_rcvnowait = pd->port_pionowait = 0;
2101	}
2102	if (pd->port_flag) {
2103		ipath_cdbg(PROC, "port %u port_flag set: 0x%lx\n",
2104			  pd->port_port, pd->port_flag);
2105		pd->port_flag = 0;
2106	}
2107
2108	if (dd->ipath_kregbase) {
2109		/* atomically clear receive enable port and intr avail. */
2110		clear_bit(dd->ipath_r_portenable_shift + port,
2111			  &dd->ipath_rcvctrl);
2112		clear_bit(pd->port_port + dd->ipath_r_intravail_shift,
2113			  &dd->ipath_rcvctrl);
2114		ipath_write_kreg( dd, dd->ipath_kregs->kr_rcvctrl,
2115			dd->ipath_rcvctrl);
2116		/* and read back from chip to be sure that nothing
2117		 * else is in flight when we do the rest */
2118		(void)ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch);
2119
2120		/* clean up the pkeys for this port user */
2121		ipath_clean_part_key(pd, dd);
2122		/*
2123		 * be paranoid, and never write 0's to these, just use an
2124		 * unused part of the port 0 tail page.  Of course,
2125		 * rcvhdraddr points to a large chunk of memory, so this
2126		 * could still trash things, but at least it won't trash
2127		 * page 0, and by disabling the port, it should stop "soon",
2128		 * even if a packet or two is in already in flight after we
2129		 * disabled the port.
2130		 */
2131		ipath_write_kreg_port(dd,
2132		        dd->ipath_kregs->kr_rcvhdrtailaddr, port,
2133			dd->ipath_dummy_hdrq_phys);
2134		ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdraddr,
2135			pd->port_port, dd->ipath_dummy_hdrq_phys);
2136
2137		ipath_disarm_piobufs(dd, pd->port_pio_base, pd->port_piocnt);
2138		ipath_chg_pioavailkernel(dd, pd->port_pio_base,
2139			pd->port_piocnt, 1);
2140
2141		dd->ipath_f_clear_tids(dd, pd->port_port);
2142
2143		if (dd->ipath_pageshadow)
2144			unlock_expected_tids(pd);
2145		ipath_stats.sps_ports--;
2146		ipath_cdbg(PROC, "%s[%u] closed port %u:%u\n",
2147			   pd->port_comm, pid_nr(pid),
2148			   dd->ipath_unit, port);
2149	}
2150
2151	put_pid(pid);
2152	mutex_unlock(&ipath_mutex);
2153	ipath_free_pddata(dd, pd); /* after releasing the mutex */
2154
2155bail:
2156	kfree(fd);
2157	return ret;
2158}
2159
2160static int ipath_port_info(struct ipath_portdata *pd, u16 subport,
2161			   struct ipath_port_info __user *uinfo)
2162{
2163	struct ipath_port_info info;
2164	int nup;
2165	int ret;
2166	size_t sz;
2167
2168	(void) ipath_count_units(NULL, &nup, NULL);
2169	info.num_active = nup;
2170	info.unit = pd->port_dd->ipath_unit;
2171	info.port = pd->port_port;
2172	info.subport = subport;
2173	/* Don't return new fields if old library opened the port. */
2174	if (ipath_supports_subports(pd->userversion >> 16,
2175				    pd->userversion & 0xffff)) {
2176		/* Number of user ports available for this device. */
2177		info.num_ports = pd->port_dd->ipath_cfgports - 1;
2178		info.num_subports = pd->port_subport_cnt;
2179		sz = sizeof(info);
2180	} else
2181		sz = sizeof(info) - 2 * sizeof(u16);
2182
2183	if (copy_to_user(uinfo, &info, sz)) {
2184		ret = -EFAULT;
2185		goto bail;
2186	}
2187	ret = 0;
2188
2189bail:
2190	return ret;
2191}
2192
2193static int ipath_get_slave_info(struct ipath_portdata *pd,
2194				void __user *slave_mask_addr)
2195{
2196	int ret = 0;
2197
2198	if (copy_to_user(slave_mask_addr, &pd->active_slaves, sizeof(u32)))
2199		ret = -EFAULT;
2200	return ret;
2201}
2202
2203static int ipath_sdma_get_inflight(struct ipath_user_sdma_queue *pq,
2204				   u32 __user *inflightp)
2205{
2206	const u32 val = ipath_user_sdma_inflight_counter(pq);
2207
2208	if (put_user(val, inflightp))
2209		return -EFAULT;
2210
2211	return 0;
2212}
2213
2214static int ipath_sdma_get_complete(struct ipath_devdata *dd,
2215				   struct ipath_user_sdma_queue *pq,
2216				   u32 __user *completep)
2217{
2218	u32 val;
2219	int err;
2220
2221	err = ipath_user_sdma_make_progress(dd, pq);
2222	if (err < 0)
2223		return err;
2224
2225	val = ipath_user_sdma_complete_counter(pq);
2226	if (put_user(val, completep))
2227		return -EFAULT;
2228
2229	return 0;
2230}
2231
2232static ssize_t ipath_write(struct file *fp, const char __user *data,
2233			   size_t count, loff_t *off)
2234{
2235	const struct ipath_cmd __user *ucmd;
2236	struct ipath_portdata *pd;
2237	const void __user *src;
2238	size_t consumed, copy;
2239	struct ipath_cmd cmd;
2240	ssize_t ret = 0;
2241	void *dest;
2242
2243	if (count < sizeof(cmd.type)) {
2244		ret = -EINVAL;
2245		goto bail;
2246	}
2247
2248	ucmd = (const struct ipath_cmd __user *) data;
2249
2250	if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) {
2251		ret = -EFAULT;
2252		goto bail;
2253	}
2254
2255	consumed = sizeof(cmd.type);
2256
2257	switch (cmd.type) {
2258	case IPATH_CMD_ASSIGN_PORT:
2259	case __IPATH_CMD_USER_INIT:
2260	case IPATH_CMD_USER_INIT:
2261		copy = sizeof(cmd.cmd.user_info);
2262		dest = &cmd.cmd.user_info;
2263		src = &ucmd->cmd.user_info;
2264		break;
2265	case IPATH_CMD_RECV_CTRL:
2266		copy = sizeof(cmd.cmd.recv_ctrl);
2267		dest = &cmd.cmd.recv_ctrl;
2268		src = &ucmd->cmd.recv_ctrl;
2269		break;
2270	case IPATH_CMD_PORT_INFO:
2271		copy = sizeof(cmd.cmd.port_info);
2272		dest = &cmd.cmd.port_info;
2273		src = &ucmd->cmd.port_info;
2274		break;
2275	case IPATH_CMD_TID_UPDATE:
2276	case IPATH_CMD_TID_FREE:
2277		copy = sizeof(cmd.cmd.tid_info);
2278		dest = &cmd.cmd.tid_info;
2279		src = &ucmd->cmd.tid_info;
2280		break;
2281	case IPATH_CMD_SET_PART_KEY:
2282		copy = sizeof(cmd.cmd.part_key);
2283		dest = &cmd.cmd.part_key;
2284		src = &ucmd->cmd.part_key;
2285		break;
2286	case __IPATH_CMD_SLAVE_INFO:
2287		copy = sizeof(cmd.cmd.slave_mask_addr);
2288		dest = &cmd.cmd.slave_mask_addr;
2289		src = &ucmd->cmd.slave_mask_addr;
2290		break;
2291	case IPATH_CMD_PIOAVAILUPD:	// force an update of PIOAvail reg
2292		copy = 0;
2293		src = NULL;
2294		dest = NULL;
2295		break;
2296	case IPATH_CMD_POLL_TYPE:
2297		copy = sizeof(cmd.cmd.poll_type);
2298		dest = &cmd.cmd.poll_type;
2299		src = &ucmd->cmd.poll_type;
2300		break;
2301	case IPATH_CMD_ARMLAUNCH_CTRL:
2302		copy = sizeof(cmd.cmd.armlaunch_ctrl);
2303		dest = &cmd.cmd.armlaunch_ctrl;
2304		src = &ucmd->cmd.armlaunch_ctrl;
2305		break;
2306	case IPATH_CMD_SDMA_INFLIGHT:
2307		copy = sizeof(cmd.cmd.sdma_inflight);
2308		dest = &cmd.cmd.sdma_inflight;
2309		src = &ucmd->cmd.sdma_inflight;
2310		break;
2311	case IPATH_CMD_SDMA_COMPLETE:
2312		copy = sizeof(cmd.cmd.sdma_complete);
2313		dest = &cmd.cmd.sdma_complete;
2314		src = &ucmd->cmd.sdma_complete;
2315		break;
2316	default:
2317		ret = -EINVAL;
2318		goto bail;
2319	}
2320
2321	if (copy) {
2322		if ((count - consumed) < copy) {
2323			ret = -EINVAL;
2324			goto bail;
2325		}
2326
2327		if (copy_from_user(dest, src, copy)) {
2328			ret = -EFAULT;
2329			goto bail;
2330		}
2331
2332		consumed += copy;
2333	}
2334
2335	pd = port_fp(fp);
2336	if (!pd && cmd.type != __IPATH_CMD_USER_INIT &&
2337		cmd.type != IPATH_CMD_ASSIGN_PORT) {
2338		ret = -EINVAL;
2339		goto bail;
2340	}
2341
2342	switch (cmd.type) {
2343	case IPATH_CMD_ASSIGN_PORT:
2344		ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2345		if (ret)
2346			goto bail;
2347		break;
2348	case __IPATH_CMD_USER_INIT:
2349		/* backwards compatibility, get port first */
2350		ret = ipath_assign_port(fp, &cmd.cmd.user_info);
2351		if (ret)
2352			goto bail;
2353		/* and fall through to current version. */
2354	case IPATH_CMD_USER_INIT:
2355		ret = ipath_do_user_init(fp, &cmd.cmd.user_info);
2356		if (ret)
2357			goto bail;
2358		ret = ipath_get_base_info(
2359			fp, (void __user *) (unsigned long)
2360			cmd.cmd.user_info.spu_base_info,
2361			cmd.cmd.user_info.spu_base_info_size);
2362		break;
2363	case IPATH_CMD_RECV_CTRL:
2364		ret = ipath_manage_rcvq(pd, subport_fp(fp), cmd.cmd.recv_ctrl);
2365		break;
2366	case IPATH_CMD_PORT_INFO:
2367		ret = ipath_port_info(pd, subport_fp(fp),
2368				      (struct ipath_port_info __user *)
2369				      (unsigned long) cmd.cmd.port_info);
2370		break;
2371	case IPATH_CMD_TID_UPDATE:
2372		ret = ipath_tid_update(pd, fp, &cmd.cmd.tid_info);
2373		break;
2374	case IPATH_CMD_TID_FREE:
2375		ret = ipath_tid_free(pd, subport_fp(fp), &cmd.cmd.tid_info);
2376		break;
2377	case IPATH_CMD_SET_PART_KEY:
2378		ret = ipath_set_part_key(pd, cmd.cmd.part_key);
2379		break;
2380	case __IPATH_CMD_SLAVE_INFO:
2381		ret = ipath_get_slave_info(pd,
2382					   (void __user *) (unsigned long)
2383					   cmd.cmd.slave_mask_addr);
2384		break;
2385	case IPATH_CMD_PIOAVAILUPD:
2386		ipath_force_pio_avail_update(pd->port_dd);
2387		break;
2388	case IPATH_CMD_POLL_TYPE:
2389		pd->poll_type = cmd.cmd.poll_type;
2390		break;
2391	case IPATH_CMD_ARMLAUNCH_CTRL:
2392		if (cmd.cmd.armlaunch_ctrl)
2393			ipath_enable_armlaunch(pd->port_dd);
2394		else
2395			ipath_disable_armlaunch(pd->port_dd);
2396		break;
2397	case IPATH_CMD_SDMA_INFLIGHT:
2398		ret = ipath_sdma_get_inflight(user_sdma_queue_fp(fp),
2399					      (u32 __user *) (unsigned long)
2400					      cmd.cmd.sdma_inflight);
2401		break;
2402	case IPATH_CMD_SDMA_COMPLETE:
2403		ret = ipath_sdma_get_complete(pd->port_dd,
2404					      user_sdma_queue_fp(fp),
2405					      (u32 __user *) (unsigned long)
2406					      cmd.cmd.sdma_complete);
2407		break;
2408	}
2409
2410	if (ret >= 0)
2411		ret = consumed;
2412
2413bail:
2414	return ret;
2415}
2416
2417static ssize_t ipath_writev(struct kiocb *iocb, const struct iovec *iov,
2418			    unsigned long dim, loff_t off)
2419{
2420	struct file *filp = iocb->ki_filp;
2421	struct ipath_filedata *fp = filp->private_data;
2422	struct ipath_portdata *pd = port_fp(filp);
2423	struct ipath_user_sdma_queue *pq = fp->pq;
2424
2425	if (!dim)
2426		return -EINVAL;
2427
2428	return ipath_user_sdma_writev(pd->port_dd, pq, iov, dim);
2429}
2430
2431static struct class *ipath_class;
2432
2433static int init_cdev(int minor, char *name, const struct file_operations *fops,
2434		     struct cdev **cdevp, struct device **devp)
2435{
2436	const dev_t dev = MKDEV(IPATH_MAJOR, minor);
2437	struct cdev *cdev = NULL;
2438	struct device *device = NULL;
2439	int ret;
2440
2441	cdev = cdev_alloc();
2442	if (!cdev) {
2443		printk(KERN_ERR IPATH_DRV_NAME
2444		       ": Could not allocate cdev for minor %d, %s\n",
2445		       minor, name);
2446		ret = -ENOMEM;
2447		goto done;
2448	}
2449
2450	cdev->owner = THIS_MODULE;
2451	cdev->ops = fops;
2452	kobject_set_name(&cdev->kobj, name);
2453
2454	ret = cdev_add(cdev, dev, 1);
2455	if (ret < 0) {
2456		printk(KERN_ERR IPATH_DRV_NAME
2457		       ": Could not add cdev for minor %d, %s (err %d)\n",
2458		       minor, name, -ret);
2459		goto err_cdev;
2460	}
2461
2462	device = device_create(ipath_class, NULL, dev, NULL, name);
2463
2464	if (IS_ERR(device)) {
2465		ret = PTR_ERR(device);
2466		printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2467		       "device for minor %d, %s (err %d)\n",
2468		       minor, name, -ret);
2469		goto err_cdev;
2470	}
2471
2472	goto done;
2473
2474err_cdev:
2475	cdev_del(cdev);
2476	cdev = NULL;
2477
2478done:
2479	if (ret >= 0) {
2480		*cdevp = cdev;
2481		*devp = device;
2482	} else {
2483		*cdevp = NULL;
2484		*devp = NULL;
2485	}
2486
2487	return ret;
2488}
2489
2490int ipath_cdev_init(int minor, char *name, const struct file_operations *fops,
2491		    struct cdev **cdevp, struct device **devp)
2492{
2493	return init_cdev(minor, name, fops, cdevp, devp);
2494}
2495
2496static void cleanup_cdev(struct cdev **cdevp,
2497			 struct device **devp)
2498{
2499	struct device *dev = *devp;
2500
2501	if (dev) {
2502		device_unregister(dev);
2503		*devp = NULL;
2504	}
2505
2506	if (*cdevp) {
2507		cdev_del(*cdevp);
2508		*cdevp = NULL;
2509	}
2510}
2511
2512void ipath_cdev_cleanup(struct cdev **cdevp,
2513			struct device **devp)
2514{
2515	cleanup_cdev(cdevp, devp);
2516}
2517
2518static struct cdev *wildcard_cdev;
2519static struct device *wildcard_dev;
2520
2521static const dev_t dev = MKDEV(IPATH_MAJOR, 0);
2522
2523static int user_init(void)
2524{
2525	int ret;
2526
2527	ret = register_chrdev_region(dev, IPATH_NMINORS, IPATH_DRV_NAME);
2528	if (ret < 0) {
2529		printk(KERN_ERR IPATH_DRV_NAME ": Could not register "
2530		       "chrdev region (err %d)\n", -ret);
2531		goto done;
2532	}
2533
2534	ipath_class = class_create(THIS_MODULE, IPATH_DRV_NAME);
2535
2536	if (IS_ERR(ipath_class)) {
2537		ret = PTR_ERR(ipath_class);
2538		printk(KERN_ERR IPATH_DRV_NAME ": Could not create "
2539		       "device class (err %d)\n", -ret);
2540		goto bail;
2541	}
2542
2543	goto done;
2544bail:
2545	unregister_chrdev_region(dev, IPATH_NMINORS);
2546done:
2547	return ret;
2548}
2549
2550static void user_cleanup(void)
2551{
2552	if (ipath_class) {
2553		class_destroy(ipath_class);
2554		ipath_class = NULL;
2555	}
2556
2557	unregister_chrdev_region(dev, IPATH_NMINORS);
2558}
2559
2560static atomic_t user_count = ATOMIC_INIT(0);
2561static atomic_t user_setup = ATOMIC_INIT(0);
2562
2563int ipath_user_add(struct ipath_devdata *dd)
2564{
2565	char name[10];
2566	int ret;
2567
2568	if (atomic_inc_return(&user_count) == 1) {
2569		ret = user_init();
2570		if (ret < 0) {
2571			ipath_dev_err(dd, "Unable to set up user support: "
2572				      "error %d\n", -ret);
2573			goto bail;
2574		}
2575		ret = init_cdev(0, "ipath", &ipath_file_ops, &wildcard_cdev,
2576				&wildcard_dev);
2577		if (ret < 0) {
2578			ipath_dev_err(dd, "Could not create wildcard "
2579				      "minor: error %d\n", -ret);
2580			goto bail_user;
2581		}
2582
2583		atomic_set(&user_setup, 1);
2584	}
2585
2586	snprintf(name, sizeof(name), "ipath%d", dd->ipath_unit);
2587
2588	ret = init_cdev(dd->ipath_unit + 1, name, &ipath_file_ops,
2589			&dd->user_cdev, &dd->user_dev);
2590	if (ret < 0)
2591		ipath_dev_err(dd, "Could not create user minor %d, %s\n",
2592			      dd->ipath_unit + 1, name);
2593
2594	goto bail;
2595
2596bail_user:
2597	user_cleanup();
2598bail:
2599	return ret;
2600}
2601
2602void ipath_user_remove(struct ipath_devdata *dd)
2603{
2604	cleanup_cdev(&dd->user_cdev, &dd->user_dev);
2605
2606	if (atomic_dec_return(&user_count) == 0) {
2607		if (atomic_read(&user_setup) == 0)
2608			goto bail;
2609
2610		cleanup_cdev(&wildcard_cdev, &wildcard_dev);
2611		user_cleanup();
2612
2613		atomic_set(&user_setup, 0);
2614	}
2615bail:
2616	return;
2617}
2618