1 /*
2  * Copyright (c) 2012, 2013 Intel Corporation. All rights reserved.
3  * Copyright (c) 2006 - 2012 QLogic Corporation. All rights reserved.
4  * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved.
5  *
6  * This software is available to you under a choice of one of two
7  * licenses.  You may choose to be licensed under the terms of the GNU
8  * General Public License (GPL) Version 2, available from the file
9  * COPYING in the main directory of this source tree, or the
10  * OpenIB.org BSD license below:
11  *
12  *     Redistribution and use in source and binary forms, with or
13  *     without modification, are permitted provided that the following
14  *     conditions are met:
15  *
16  *      - Redistributions of source code must retain the above
17  *        copyright notice, this list of conditions and the following
18  *        disclaimer.
19  *
20  *      - Redistributions in binary form must reproduce the above
21  *        copyright notice, this list of conditions and the following
22  *        disclaimer in the documentation and/or other materials
23  *        provided with the distribution.
24  *
25  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
26  * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
27  * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
28  * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
29  * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
30  * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
31  * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
32  * SOFTWARE.
33  */
34 
35 #include <linux/pci.h>
36 #include <linux/poll.h>
37 #include <linux/cdev.h>
38 #include <linux/swap.h>
39 #include <linux/vmalloc.h>
40 #include <linux/highmem.h>
41 #include <linux/io.h>
42 #include <linux/jiffies.h>
43 #include <linux/delay.h>
44 #include <linux/export.h>
45 #include <linux/uio.h>
46 #include <linux/pgtable.h>
47 
48 #include <rdma/ib.h>
49 
50 #include "qib.h"
51 #include "qib_common.h"
52 #include "qib_user_sdma.h"
53 
54 #undef pr_fmt
55 #define pr_fmt(fmt) QIB_DRV_NAME ": " fmt
56 
57 static int qib_open(struct inode *, struct file *);
58 static int qib_close(struct inode *, struct file *);
59 static ssize_t qib_write(struct file *, const char __user *, size_t, loff_t *);
60 static ssize_t qib_write_iter(struct kiocb *, struct iov_iter *);
61 static __poll_t qib_poll(struct file *, struct poll_table_struct *);
62 static int qib_mmapf(struct file *, struct vm_area_struct *);
63 
64 /*
65  * This is really, really weird shit - write() and writev() here
66  * have completely unrelated semantics.  Sucky userland ABI,
67  * film at 11.
68  */
69 static const struct file_operations qib_file_ops = {
70 	.owner = THIS_MODULE,
71 	.write = qib_write,
72 	.write_iter = qib_write_iter,
73 	.open = qib_open,
74 	.release = qib_close,
75 	.poll = qib_poll,
76 	.mmap = qib_mmapf,
77 	.llseek = noop_llseek,
78 };
79 
80 /*
81  * Convert kernel virtual addresses to physical addresses so they don't
82  * potentially conflict with the chip addresses used as mmap offsets.
83  * It doesn't really matter what mmap offset we use as long as we can
84  * interpret it correctly.
85  */
86 static u64 cvt_kvaddr(void *p)
87 {
88 	struct page *page;
89 	u64 paddr = 0;
90 
91 	page = vmalloc_to_page(p);
92 	if (page)
93 		paddr = page_to_pfn(page) << PAGE_SHIFT;
94 
95 	return paddr;
96 }
97 
98 static int qib_get_base_info(struct file *fp, void __user *ubase,
99 			     size_t ubase_size)
100 {
101 	struct qib_ctxtdata *rcd = ctxt_fp(fp);
102 	int ret = 0;
103 	struct qib_base_info *kinfo = NULL;
104 	struct qib_devdata *dd = rcd->dd;
105 	struct qib_pportdata *ppd = rcd->ppd;
106 	unsigned subctxt_cnt;
107 	int shared, master;
108 	size_t sz;
109 
110 	subctxt_cnt = rcd->subctxt_cnt;
111 	if (!subctxt_cnt) {
112 		shared = 0;
113 		master = 0;
114 		subctxt_cnt = 1;
115 	} else {
116 		shared = 1;
117 		master = !subctxt_fp(fp);
118 	}
119 
120 	sz = sizeof(*kinfo);
121 	/* If context sharing is not requested, allow the old size structure */
122 	if (!shared)
123 		sz -= 7 * sizeof(u64);
124 	if (ubase_size < sz) {
125 		ret = -EINVAL;
126 		goto bail;
127 	}
128 
129 	kinfo = kzalloc(sizeof(*kinfo), GFP_KERNEL);
130 	if (kinfo == NULL) {
131 		ret = -ENOMEM;
132 		goto bail;
133 	}
134 
135 	ret = dd->f_get_base_info(rcd, kinfo);
136 	if (ret < 0)
137 		goto bail;
138 
139 	kinfo->spi_rcvhdr_cnt = dd->rcvhdrcnt;
140 	kinfo->spi_rcvhdrent_size = dd->rcvhdrentsize;
141 	kinfo->spi_tidegrcnt = rcd->rcvegrcnt;
142 	kinfo->spi_rcv_egrbufsize = dd->rcvegrbufsize;
143 	/*
144 	 * have to mmap whole thing
145 	 */
146 	kinfo->spi_rcv_egrbuftotlen =
147 		rcd->rcvegrbuf_chunks * rcd->rcvegrbuf_size;
148 	kinfo->spi_rcv_egrperchunk = rcd->rcvegrbufs_perchunk;
149 	kinfo->spi_rcv_egrchunksize = kinfo->spi_rcv_egrbuftotlen /
150 		rcd->rcvegrbuf_chunks;
151 	kinfo->spi_tidcnt = dd->rcvtidcnt / subctxt_cnt;
152 	if (master)
153 		kinfo->spi_tidcnt += dd->rcvtidcnt % subctxt_cnt;
154 	/*
155 	 * for this use, may be cfgctxts summed over all chips that
156 	 * are configured and present
157 	 */
158 	kinfo->spi_nctxts = dd->cfgctxts;
159 	/* unit (chip/board) our context is on */
160 	kinfo->spi_unit = dd->unit;
161 	kinfo->spi_port = ppd->port;
162 	/* for now, only a single page */
163 	kinfo->spi_tid_maxsize = PAGE_SIZE;
164 
165 	/*
166 	 * Doing this per context, and based on the skip value, etc.  This has
167 	 * to be the actual buffer size, since the protocol code treats it
168 	 * as an array.
169 	 *
170 	 * These have to be set to user addresses in the user code via mmap.
171 	 * These values are used on return to user code for the mmap target
172 	 * addresses only.  For 32 bit, same 44 bit address problem, so use
173 	 * the physical address, not virtual.  Before 2.6.11, using the
174 	 * page_address() macro worked, but in 2.6.11, even that returns the
175 	 * full 64 bit address (upper bits all 1's).  So far, using the
176 	 * physical addresses (or chip offsets, for chip mapping) works, but
177 	 * no doubt some future kernel release will change that, and we'll be
178 	 * on to yet another method of dealing with this.
179 	 * Normally only one of rcvhdr_tailaddr or rhf_offset is useful
180 	 * since the chips with non-zero rhf_offset don't normally
181 	 * enable tail register updates to host memory, but for testing,
182 	 * both can be enabled and used.
183 	 */
184 	kinfo->spi_rcvhdr_base = (u64) rcd->rcvhdrq_phys;
185 	kinfo->spi_rcvhdr_tailaddr = (u64) rcd->rcvhdrqtailaddr_phys;
186 	kinfo->spi_rhf_offset = dd->rhf_offset;
187 	kinfo->spi_rcv_egrbufs = (u64) rcd->rcvegr_phys;
188 	kinfo->spi_pioavailaddr = (u64) dd->pioavailregs_phys;
189 	/* setup per-unit (not port) status area for user programs */
190 	kinfo->spi_status = (u64) kinfo->spi_pioavailaddr +
191 		(char *) ppd->statusp -
192 		(char *) dd->pioavailregs_dma;
193 	kinfo->spi_uregbase = (u64) dd->uregbase + dd->ureg_align * rcd->ctxt;
194 	if (!shared) {
195 		kinfo->spi_piocnt = rcd->piocnt;
196 		kinfo->spi_piobufbase = (u64) rcd->piobufs;
197 		kinfo->spi_sendbuf_status = cvt_kvaddr(rcd->user_event_mask);
198 	} else if (master) {
199 		kinfo->spi_piocnt = (rcd->piocnt / subctxt_cnt) +
200 				    (rcd->piocnt % subctxt_cnt);
201 		/* Master's PIO buffers are after all the slave's */
202 		kinfo->spi_piobufbase = (u64) rcd->piobufs +
203 			dd->palign *
204 			(rcd->piocnt - kinfo->spi_piocnt);
205 	} else {
206 		unsigned slave = subctxt_fp(fp) - 1;
207 
208 		kinfo->spi_piocnt = rcd->piocnt / subctxt_cnt;
209 		kinfo->spi_piobufbase = (u64) rcd->piobufs +
210 			dd->palign * kinfo->spi_piocnt * slave;
211 	}
212 
213 	if (shared) {
214 		kinfo->spi_sendbuf_status =
215 			cvt_kvaddr(&rcd->user_event_mask[subctxt_fp(fp)]);
216 		/* only spi_subctxt_* fields should be set in this block! */
217 		kinfo->spi_subctxt_uregbase = cvt_kvaddr(rcd->subctxt_uregbase);
218 
219 		kinfo->spi_subctxt_rcvegrbuf =
220 			cvt_kvaddr(rcd->subctxt_rcvegrbuf);
221 		kinfo->spi_subctxt_rcvhdr_base =
222 			cvt_kvaddr(rcd->subctxt_rcvhdr_base);
223 	}
224 
225 	/*
226 	 * All user buffers are 2KB buffers.  If we ever support
227 	 * giving 4KB buffers to user processes, this will need some
228 	 * work.  Can't use piobufbase directly, because it has
229 	 * both 2K and 4K buffer base values.
230 	 */
231 	kinfo->spi_pioindex = (kinfo->spi_piobufbase - dd->pio2k_bufbase) /
232 		dd->palign;
233 	kinfo->spi_pioalign = dd->palign;
234 	kinfo->spi_qpair = QIB_KD_QP;
235 	/*
236 	 * user mode PIO buffers are always 2KB, even when 4KB can
237 	 * be received, and sent via the kernel; this is ibmaxlen
238 	 * for 2K MTU.
239 	 */
240 	kinfo->spi_piosize = dd->piosize2k - 2 * sizeof(u32);
241 	kinfo->spi_mtu = ppd->ibmaxlen; /* maxlen, not ibmtu */
242 	kinfo->spi_ctxt = rcd->ctxt;
243 	kinfo->spi_subctxt = subctxt_fp(fp);
244 	kinfo->spi_sw_version = QIB_KERN_SWVERSION;
245 	kinfo->spi_sw_version |= 1U << 31; /* QLogic-built, not kernel.org */
246 	kinfo->spi_hw_version = dd->revision;
247 
248 	if (master)
249 		kinfo->spi_runtime_flags |= QIB_RUNTIME_MASTER;
250 
251 	sz = (ubase_size < sizeof(*kinfo)) ? ubase_size : sizeof(*kinfo);
252 	if (copy_to_user(ubase, kinfo, sz))
253 		ret = -EFAULT;
254 bail:
255 	kfree(kinfo);
256 	return ret;
257 }
258 
259 /**
260  * qib_tid_update - update a context TID
261  * @rcd: the context
262  * @fp: the qib device file
263  * @ti: the TID information
264  *
265  * The new implementation as of Oct 2004 is that the driver assigns
266  * the tid and returns it to the caller.   To reduce search time, we
267  * keep a cursor for each context, walking the shadow tid array to find
268  * one that's not in use.
269  *
270  * For now, if we can't allocate the full list, we fail, although
271  * in the long run, we'll allocate as many as we can, and the
272  * caller will deal with that by trying the remaining pages later.
273  * That means that when we fail, we have to mark the tids as not in
274  * use again, in our shadow copy.
275  *
276  * It's up to the caller to free the tids when they are done.
277  * We'll unlock the pages as they free them.
278  *
279  * Also, right now we are locking one page at a time, but since
280  * the intended use of this routine is for a single group of
281  * virtually contiguous pages, that should change to improve
282  * performance.
283  */
284 static int qib_tid_update(struct qib_ctxtdata *rcd, struct file *fp,
285 			  const struct qib_tid_info *ti)
286 {
287 	int ret = 0, ntids;
288 	u32 tid, ctxttid, cnt, i, tidcnt, tidoff;
289 	u16 *tidlist;
290 	struct qib_devdata *dd = rcd->dd;
291 	u64 physaddr;
292 	unsigned long vaddr;
293 	u64 __iomem *tidbase;
294 	unsigned long tidmap[8];
295 	struct page **pagep = NULL;
296 	unsigned subctxt = subctxt_fp(fp);
297 
298 	if (!dd->pageshadow) {
299 		ret = -ENOMEM;
300 		goto done;
301 	}
302 
303 	cnt = ti->tidcnt;
304 	if (!cnt) {
305 		ret = -EFAULT;
306 		goto done;
307 	}
308 	ctxttid = rcd->ctxt * dd->rcvtidcnt;
309 	if (!rcd->subctxt_cnt) {
310 		tidcnt = dd->rcvtidcnt;
311 		tid = rcd->tidcursor;
312 		tidoff = 0;
313 	} else if (!subctxt) {
314 		tidcnt = (dd->rcvtidcnt / rcd->subctxt_cnt) +
315 			 (dd->rcvtidcnt % rcd->subctxt_cnt);
316 		tidoff = dd->rcvtidcnt - tidcnt;
317 		ctxttid += tidoff;
318 		tid = tidcursor_fp(fp);
319 	} else {
320 		tidcnt = dd->rcvtidcnt / rcd->subctxt_cnt;
321 		tidoff = tidcnt * (subctxt - 1);
322 		ctxttid += tidoff;
323 		tid = tidcursor_fp(fp);
324 	}
325 	if (cnt > tidcnt) {
326 		/* make sure it all fits in tid_pg_list */
327 		qib_devinfo(dd->pcidev,
328 			"Process tried to allocate %u TIDs, only trying max (%u)\n",
329 			cnt, tidcnt);
330 		cnt = tidcnt;
331 	}
332 	pagep = (struct page **) rcd->tid_pg_list;
333 	tidlist = (u16 *) &pagep[dd->rcvtidcnt];
334 	pagep += tidoff;
335 	tidlist += tidoff;
336 
337 	memset(tidmap, 0, sizeof(tidmap));
338 	/* before decrement; chip actual # */
339 	ntids = tidcnt;
340 	tidbase = (u64 __iomem *) (((char __iomem *) dd->kregbase) +
341 				   dd->rcvtidbase +
342 				   ctxttid * sizeof(*tidbase));
343 
344 	/* virtual address of first page in transfer */
345 	vaddr = ti->tidvaddr;
346 	if (!access_ok((void __user *) vaddr,
347 		       cnt * PAGE_SIZE)) {
348 		ret = -EFAULT;
349 		goto done;
350 	}
351 	ret = qib_get_user_pages(vaddr, cnt, pagep);
352 	if (ret) {
353 		/*
354 		 * if (ret == -EBUSY)
355 		 * We can't continue because the pagep array won't be
356 		 * initialized. This should never happen,
357 		 * unless perhaps the user has mpin'ed the pages
358 		 * themselves.
359 		 */
360 		qib_devinfo(
361 			dd->pcidev,
362 			"Failed to lock addr %p, %u pages: errno %d\n",
363 			(void *) vaddr, cnt, -ret);
364 		goto done;
365 	}
366 	for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) {
367 		dma_addr_t daddr;
368 
369 		for (; ntids--; tid++) {
370 			if (tid == tidcnt)
371 				tid = 0;
372 			if (!dd->pageshadow[ctxttid + tid])
373 				break;
374 		}
375 		if (ntids < 0) {
376 			/*
377 			 * Oops, wrapped all the way through their TIDs,
378 			 * and didn't have enough free; see comments at
379 			 * start of routine
380 			 */
381 			i--;    /* last tidlist[i] not filled in */
382 			ret = -ENOMEM;
383 			break;
384 		}
385 		ret = qib_map_page(dd->pcidev, pagep[i], &daddr);
386 		if (ret)
387 			break;
388 
389 		tidlist[i] = tid + tidoff;
390 		/* we "know" system pages and TID pages are same size */
391 		dd->pageshadow[ctxttid + tid] = pagep[i];
392 		dd->physshadow[ctxttid + tid] = daddr;
393 		/*
394 		 * don't need atomic or it's overhead
395 		 */
396 		__set_bit(tid, tidmap);
397 		physaddr = dd->physshadow[ctxttid + tid];
398 		/* PERFORMANCE: below should almost certainly be cached */
399 		dd->f_put_tid(dd, &tidbase[tid],
400 				  RCVHQ_RCV_TYPE_EXPECTED, physaddr);
401 		/*
402 		 * don't check this tid in qib_ctxtshadow, since we
403 		 * just filled it in; start with the next one.
404 		 */
405 		tid++;
406 	}
407 
408 	if (ret) {
409 		u32 limit;
410 cleanup:
411 		/* jump here if copy out of updated info failed... */
412 		/* same code that's in qib_free_tid() */
413 		limit = sizeof(tidmap) * BITS_PER_BYTE;
414 		if (limit > tidcnt)
415 			/* just in case size changes in future */
416 			limit = tidcnt;
417 		tid = find_first_bit((const unsigned long *)tidmap, limit);
418 		for (; tid < limit; tid++) {
419 			if (!test_bit(tid, tidmap))
420 				continue;
421 			if (dd->pageshadow[ctxttid + tid]) {
422 				dma_addr_t phys;
423 
424 				phys = dd->physshadow[ctxttid + tid];
425 				dd->physshadow[ctxttid + tid] = dd->tidinvalid;
426 				/* PERFORMANCE: below should almost certainly
427 				 * be cached
428 				 */
429 				dd->f_put_tid(dd, &tidbase[tid],
430 					      RCVHQ_RCV_TYPE_EXPECTED,
431 					      dd->tidinvalid);
432 				dma_unmap_page(&dd->pcidev->dev, phys,
433 					       PAGE_SIZE, DMA_FROM_DEVICE);
434 				dd->pageshadow[ctxttid + tid] = NULL;
435 			}
436 		}
437 		qib_release_user_pages(pagep, cnt);
438 	} else {
439 		/*
440 		 * Copy the updated array, with qib_tid's filled in, back
441 		 * to user.  Since we did the copy in already, this "should
442 		 * never fail" If it does, we have to clean up...
443 		 */
444 		if (copy_to_user((void __user *)
445 				 (unsigned long) ti->tidlist,
446 				 tidlist, cnt * sizeof(*tidlist))) {
447 			ret = -EFAULT;
448 			goto cleanup;
449 		}
450 		if (copy_to_user(u64_to_user_ptr(ti->tidmap),
451 				 tidmap, sizeof(tidmap))) {
452 			ret = -EFAULT;
453 			goto cleanup;
454 		}
455 		if (tid == tidcnt)
456 			tid = 0;
457 		if (!rcd->subctxt_cnt)
458 			rcd->tidcursor = tid;
459 		else
460 			tidcursor_fp(fp) = tid;
461 	}
462 
463 done:
464 	return ret;
465 }
466 
467 /**
468  * qib_tid_free - free a context TID
469  * @rcd: the context
470  * @subctxt: the subcontext
471  * @ti: the TID info
472  *
473  * right now we are unlocking one page at a time, but since
474  * the intended use of this routine is for a single group of
475  * virtually contiguous pages, that should change to improve
476  * performance.  We check that the TID is in range for this context
477  * but otherwise don't check validity; if user has an error and
478  * frees the wrong tid, it's only their own data that can thereby
479  * be corrupted.  We do check that the TID was in use, for sanity
480  * We always use our idea of the saved address, not the address that
481  * they pass in to us.
482  */
483 static int qib_tid_free(struct qib_ctxtdata *rcd, unsigned subctxt,
484 			const struct qib_tid_info *ti)
485 {
486 	int ret = 0;
487 	u32 tid, ctxttid, cnt, limit, tidcnt;
488 	struct qib_devdata *dd = rcd->dd;
489 	u64 __iomem *tidbase;
490 	unsigned long tidmap[8];
491 
492 	if (!dd->pageshadow) {
493 		ret = -ENOMEM;
494 		goto done;
495 	}
496 
497 	if (copy_from_user(tidmap, u64_to_user_ptr(ti->tidmap),
498 			   sizeof(tidmap))) {
499 		ret = -EFAULT;
500 		goto done;
501 	}
502 
503 	ctxttid = rcd->ctxt * dd->rcvtidcnt;
504 	if (!rcd->subctxt_cnt)
505 		tidcnt = dd->rcvtidcnt;
506 	else if (!subctxt) {
507 		tidcnt = (dd->rcvtidcnt / rcd->subctxt_cnt) +
508 			 (dd->rcvtidcnt % rcd->subctxt_cnt);
509 		ctxttid += dd->rcvtidcnt - tidcnt;
510 	} else {
511 		tidcnt = dd->rcvtidcnt / rcd->subctxt_cnt;
512 		ctxttid += tidcnt * (subctxt - 1);
513 	}
514 	tidbase = (u64 __iomem *) ((char __iomem *)(dd->kregbase) +
515 				   dd->rcvtidbase +
516 				   ctxttid * sizeof(*tidbase));
517 
518 	limit = sizeof(tidmap) * BITS_PER_BYTE;
519 	if (limit > tidcnt)
520 		/* just in case size changes in future */
521 		limit = tidcnt;
522 	tid = find_first_bit(tidmap, limit);
523 	for (cnt = 0; tid < limit; tid++) {
524 		/*
525 		 * small optimization; if we detect a run of 3 or so without
526 		 * any set, use find_first_bit again.  That's mainly to
527 		 * accelerate the case where we wrapped, so we have some at
528 		 * the beginning, and some at the end, and a big gap
529 		 * in the middle.
530 		 */
531 		if (!test_bit(tid, tidmap))
532 			continue;
533 		cnt++;
534 		if (dd->pageshadow[ctxttid + tid]) {
535 			struct page *p;
536 			dma_addr_t phys;
537 
538 			p = dd->pageshadow[ctxttid + tid];
539 			dd->pageshadow[ctxttid + tid] = NULL;
540 			phys = dd->physshadow[ctxttid + tid];
541 			dd->physshadow[ctxttid + tid] = dd->tidinvalid;
542 			/* PERFORMANCE: below should almost certainly be
543 			 * cached
544 			 */
545 			dd->f_put_tid(dd, &tidbase[tid],
546 				      RCVHQ_RCV_TYPE_EXPECTED, dd->tidinvalid);
547 			dma_unmap_page(&dd->pcidev->dev, phys, PAGE_SIZE,
548 				       DMA_FROM_DEVICE);
549 			qib_release_user_pages(&p, 1);
550 		}
551 	}
552 done:
553 	return ret;
554 }
555 
556 /**
557  * qib_set_part_key - set a partition key
558  * @rcd: the context
559  * @key: the key
560  *
561  * We can have up to 4 active at a time (other than the default, which is
562  * always allowed).  This is somewhat tricky, since multiple contexts may set
563  * the same key, so we reference count them, and clean up at exit.  All 4
564  * partition keys are packed into a single qlogic_ib register.  It's an
565  * error for a process to set the same pkey multiple times.  We provide no
566  * mechanism to de-allocate a pkey at this time, we may eventually need to
567  * do that.  I've used the atomic operations, and no locking, and only make
568  * a single pass through what's available.  This should be more than
569  * adequate for some time. I'll think about spinlocks or the like if and as
570  * it's necessary.
571  */
572 static int qib_set_part_key(struct qib_ctxtdata *rcd, u16 key)
573 {
574 	struct qib_pportdata *ppd = rcd->ppd;
575 	int i, pidx = -1;
576 	bool any = false;
577 	u16 lkey = key & 0x7FFF;
578 
579 	if (lkey == (QIB_DEFAULT_P_KEY & 0x7FFF))
580 		/* nothing to do; this key always valid */
581 		return 0;
582 
583 	if (!lkey)
584 		return -EINVAL;
585 
586 	/*
587 	 * Set the full membership bit, because it has to be
588 	 * set in the register or the packet, and it seems
589 	 * cleaner to set in the register than to force all
590 	 * callers to set it.
591 	 */
592 	key |= 0x8000;
593 
594 	for (i = 0; i < ARRAY_SIZE(rcd->pkeys); i++) {
595 		if (!rcd->pkeys[i] && pidx == -1)
596 			pidx = i;
597 		if (rcd->pkeys[i] == key)
598 			return -EEXIST;
599 	}
600 	if (pidx == -1)
601 		return -EBUSY;
602 	for (i = 0; i < ARRAY_SIZE(ppd->pkeys); i++) {
603 		if (!ppd->pkeys[i]) {
604 			any = true;
605 			continue;
606 		}
607 		if (ppd->pkeys[i] == key) {
608 			atomic_t *pkrefs = &ppd->pkeyrefs[i];
609 
610 			if (atomic_inc_return(pkrefs) > 1) {
611 				rcd->pkeys[pidx] = key;
612 				return 0;
613 			}
614 			/*
615 			 * lost race, decrement count, catch below
616 			 */
617 			atomic_dec(pkrefs);
618 			any = true;
619 		}
620 		if ((ppd->pkeys[i] & 0x7FFF) == lkey)
621 			/*
622 			 * It makes no sense to have both the limited and
623 			 * full membership PKEY set at the same time since
624 			 * the unlimited one will disable the limited one.
625 			 */
626 			return -EEXIST;
627 	}
628 	if (!any)
629 		return -EBUSY;
630 	for (i = 0; i < ARRAY_SIZE(ppd->pkeys); i++) {
631 		if (!ppd->pkeys[i] &&
632 		    atomic_inc_return(&ppd->pkeyrefs[i]) == 1) {
633 			rcd->pkeys[pidx] = key;
634 			ppd->pkeys[i] = key;
635 			(void) ppd->dd->f_set_ib_cfg(ppd, QIB_IB_CFG_PKEYS, 0);
636 			return 0;
637 		}
638 	}
639 	return -EBUSY;
640 }
641 
642 /**
643  * qib_manage_rcvq - manage a context's receive queue
644  * @rcd: the context
645  * @subctxt: the subcontext
646  * @start_stop: action to carry out
647  *
648  * start_stop == 0 disables receive on the context, for use in queue
649  * overflow conditions.  start_stop==1 re-enables, to be used to
650  * re-init the software copy of the head register
651  */
652 static int qib_manage_rcvq(struct qib_ctxtdata *rcd, unsigned subctxt,
653 			   int start_stop)
654 {
655 	struct qib_devdata *dd = rcd->dd;
656 	unsigned int rcvctrl_op;
657 
658 	if (subctxt)
659 		goto bail;
660 	/* atomically clear receive enable ctxt. */
661 	if (start_stop) {
662 		/*
663 		 * On enable, force in-memory copy of the tail register to
664 		 * 0, so that protocol code doesn't have to worry about
665 		 * whether or not the chip has yet updated the in-memory
666 		 * copy or not on return from the system call. The chip
667 		 * always resets it's tail register back to 0 on a
668 		 * transition from disabled to enabled.
669 		 */
670 		if (rcd->rcvhdrtail_kvaddr)
671 			qib_clear_rcvhdrtail(rcd);
672 		rcvctrl_op = QIB_RCVCTRL_CTXT_ENB;
673 	} else
674 		rcvctrl_op = QIB_RCVCTRL_CTXT_DIS;
675 	dd->f_rcvctrl(rcd->ppd, rcvctrl_op, rcd->ctxt);
676 	/* always; new head should be equal to new tail; see above */
677 bail:
678 	return 0;
679 }
680 
681 static void qib_clean_part_key(struct qib_ctxtdata *rcd,
682 			       struct qib_devdata *dd)
683 {
684 	int i, j, pchanged = 0;
685 	struct qib_pportdata *ppd = rcd->ppd;
686 
687 	for (i = 0; i < ARRAY_SIZE(rcd->pkeys); i++) {
688 		if (!rcd->pkeys[i])
689 			continue;
690 		for (j = 0; j < ARRAY_SIZE(ppd->pkeys); j++) {
691 			/* check for match independent of the global bit */
692 			if ((ppd->pkeys[j] & 0x7fff) !=
693 			    (rcd->pkeys[i] & 0x7fff))
694 				continue;
695 			if (atomic_dec_and_test(&ppd->pkeyrefs[j])) {
696 				ppd->pkeys[j] = 0;
697 				pchanged++;
698 			}
699 			break;
700 		}
701 		rcd->pkeys[i] = 0;
702 	}
703 	if (pchanged)
704 		(void) ppd->dd->f_set_ib_cfg(ppd, QIB_IB_CFG_PKEYS, 0);
705 }
706 
707 /* common code for the mappings on dma_alloc_coherent mem */
708 static int qib_mmap_mem(struct vm_area_struct *vma, struct qib_ctxtdata *rcd,
709 			unsigned len, void *kvaddr, u32 write_ok, char *what)
710 {
711 	struct qib_devdata *dd = rcd->dd;
712 	unsigned long pfn;
713 	int ret;
714 
715 	if ((vma->vm_end - vma->vm_start) > len) {
716 		qib_devinfo(dd->pcidev,
717 			 "FAIL on %s: len %lx > %x\n", what,
718 			 vma->vm_end - vma->vm_start, len);
719 		ret = -EFAULT;
720 		goto bail;
721 	}
722 
723 	/*
724 	 * shared context user code requires rcvhdrq mapped r/w, others
725 	 * only allowed readonly mapping.
726 	 */
727 	if (!write_ok) {
728 		if (vma->vm_flags & VM_WRITE) {
729 			qib_devinfo(dd->pcidev,
730 				 "%s must be mapped readonly\n", what);
731 			ret = -EPERM;
732 			goto bail;
733 		}
734 
735 		/* don't allow them to later change with mprotect */
736 		vma->vm_flags &= ~VM_MAYWRITE;
737 	}
738 
739 	pfn = virt_to_phys(kvaddr) >> PAGE_SHIFT;
740 	ret = remap_pfn_range(vma, vma->vm_start, pfn,
741 			      len, vma->vm_page_prot);
742 	if (ret)
743 		qib_devinfo(dd->pcidev,
744 			"%s ctxt%u mmap of %lx, %x bytes failed: %d\n",
745 			what, rcd->ctxt, pfn, len, ret);
746 bail:
747 	return ret;
748 }
749 
750 static int mmap_ureg(struct vm_area_struct *vma, struct qib_devdata *dd,
751 		     u64 ureg)
752 {
753 	unsigned long phys;
754 	unsigned long sz;
755 	int ret;
756 
757 	/*
758 	 * This is real hardware, so use io_remap.  This is the mechanism
759 	 * for the user process to update the head registers for their ctxt
760 	 * in the chip.
761 	 */
762 	sz = dd->flags & QIB_HAS_HDRSUPP ? 2 * PAGE_SIZE : PAGE_SIZE;
763 	if ((vma->vm_end - vma->vm_start) > sz) {
764 		qib_devinfo(dd->pcidev,
765 			"FAIL mmap userreg: reqlen %lx > PAGE\n",
766 			vma->vm_end - vma->vm_start);
767 		ret = -EFAULT;
768 	} else {
769 		phys = dd->physaddr + ureg;
770 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
771 
772 		vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
773 		ret = io_remap_pfn_range(vma, vma->vm_start,
774 					 phys >> PAGE_SHIFT,
775 					 vma->vm_end - vma->vm_start,
776 					 vma->vm_page_prot);
777 	}
778 	return ret;
779 }
780 
781 static int mmap_piobufs(struct vm_area_struct *vma,
782 			struct qib_devdata *dd,
783 			struct qib_ctxtdata *rcd,
784 			unsigned piobufs, unsigned piocnt)
785 {
786 	unsigned long phys;
787 	int ret;
788 
789 	/*
790 	 * When we map the PIO buffers in the chip, we want to map them as
791 	 * writeonly, no read possible; unfortunately, x86 doesn't allow
792 	 * for this in hardware, but we still prevent users from asking
793 	 * for it.
794 	 */
795 	if ((vma->vm_end - vma->vm_start) > (piocnt * dd->palign)) {
796 		qib_devinfo(dd->pcidev,
797 			"FAIL mmap piobufs: reqlen %lx > PAGE\n",
798 			 vma->vm_end - vma->vm_start);
799 		ret = -EINVAL;
800 		goto bail;
801 	}
802 
803 	phys = dd->physaddr + piobufs;
804 
805 #if defined(__powerpc__)
806 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
807 #endif
808 
809 	/*
810 	 * don't allow them to later change to readable with mprotect (for when
811 	 * not initially mapped readable, as is normally the case)
812 	 */
813 	vma->vm_flags &= ~VM_MAYREAD;
814 	vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND;
815 
816 	/* We used PAT if wc_cookie == 0 */
817 	if (!dd->wc_cookie)
818 		vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
819 
820 	ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT,
821 				 vma->vm_end - vma->vm_start,
822 				 vma->vm_page_prot);
823 bail:
824 	return ret;
825 }
826 
827 static int mmap_rcvegrbufs(struct vm_area_struct *vma,
828 			   struct qib_ctxtdata *rcd)
829 {
830 	struct qib_devdata *dd = rcd->dd;
831 	unsigned long start, size;
832 	size_t total_size, i;
833 	unsigned long pfn;
834 	int ret;
835 
836 	size = rcd->rcvegrbuf_size;
837 	total_size = rcd->rcvegrbuf_chunks * size;
838 	if ((vma->vm_end - vma->vm_start) > total_size) {
839 		qib_devinfo(dd->pcidev,
840 			"FAIL on egr bufs: reqlen %lx > actual %lx\n",
841 			 vma->vm_end - vma->vm_start,
842 			 (unsigned long) total_size);
843 		ret = -EINVAL;
844 		goto bail;
845 	}
846 
847 	if (vma->vm_flags & VM_WRITE) {
848 		qib_devinfo(dd->pcidev,
849 			"Can't map eager buffers as writable (flags=%lx)\n",
850 			vma->vm_flags);
851 		ret = -EPERM;
852 		goto bail;
853 	}
854 	/* don't allow them to later change to writable with mprotect */
855 	vma->vm_flags &= ~VM_MAYWRITE;
856 
857 	start = vma->vm_start;
858 
859 	for (i = 0; i < rcd->rcvegrbuf_chunks; i++, start += size) {
860 		pfn = virt_to_phys(rcd->rcvegrbuf[i]) >> PAGE_SHIFT;
861 		ret = remap_pfn_range(vma, start, pfn, size,
862 				      vma->vm_page_prot);
863 		if (ret < 0)
864 			goto bail;
865 	}
866 	ret = 0;
867 
868 bail:
869 	return ret;
870 }
871 
872 /*
873  * qib_file_vma_fault - handle a VMA page fault.
874  */
875 static vm_fault_t qib_file_vma_fault(struct vm_fault *vmf)
876 {
877 	struct page *page;
878 
879 	page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
880 	if (!page)
881 		return VM_FAULT_SIGBUS;
882 
883 	get_page(page);
884 	vmf->page = page;
885 
886 	return 0;
887 }
888 
889 static const struct vm_operations_struct qib_file_vm_ops = {
890 	.fault = qib_file_vma_fault,
891 };
892 
893 static int mmap_kvaddr(struct vm_area_struct *vma, u64 pgaddr,
894 		       struct qib_ctxtdata *rcd, unsigned subctxt)
895 {
896 	struct qib_devdata *dd = rcd->dd;
897 	unsigned subctxt_cnt;
898 	unsigned long len;
899 	void *addr;
900 	size_t size;
901 	int ret = 0;
902 
903 	subctxt_cnt = rcd->subctxt_cnt;
904 	size = rcd->rcvegrbuf_chunks * rcd->rcvegrbuf_size;
905 
906 	/*
907 	 * Each process has all the subctxt uregbase, rcvhdrq, and
908 	 * rcvegrbufs mmapped - as an array for all the processes,
909 	 * and also separately for this process.
910 	 */
911 	if (pgaddr == cvt_kvaddr(rcd->subctxt_uregbase)) {
912 		addr = rcd->subctxt_uregbase;
913 		size = PAGE_SIZE * subctxt_cnt;
914 	} else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvhdr_base)) {
915 		addr = rcd->subctxt_rcvhdr_base;
916 		size = rcd->rcvhdrq_size * subctxt_cnt;
917 	} else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvegrbuf)) {
918 		addr = rcd->subctxt_rcvegrbuf;
919 		size *= subctxt_cnt;
920 	} else if (pgaddr == cvt_kvaddr(rcd->subctxt_uregbase +
921 					PAGE_SIZE * subctxt)) {
922 		addr = rcd->subctxt_uregbase + PAGE_SIZE * subctxt;
923 		size = PAGE_SIZE;
924 	} else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvhdr_base +
925 					rcd->rcvhdrq_size * subctxt)) {
926 		addr = rcd->subctxt_rcvhdr_base +
927 			rcd->rcvhdrq_size * subctxt;
928 		size = rcd->rcvhdrq_size;
929 	} else if (pgaddr == cvt_kvaddr(&rcd->user_event_mask[subctxt])) {
930 		addr = rcd->user_event_mask;
931 		size = PAGE_SIZE;
932 	} else if (pgaddr == cvt_kvaddr(rcd->subctxt_rcvegrbuf +
933 					size * subctxt)) {
934 		addr = rcd->subctxt_rcvegrbuf + size * subctxt;
935 		/* rcvegrbufs are read-only on the slave */
936 		if (vma->vm_flags & VM_WRITE) {
937 			qib_devinfo(dd->pcidev,
938 				 "Can't map eager buffers as writable (flags=%lx)\n",
939 				 vma->vm_flags);
940 			ret = -EPERM;
941 			goto bail;
942 		}
943 		/*
944 		 * Don't allow permission to later change to writable
945 		 * with mprotect.
946 		 */
947 		vma->vm_flags &= ~VM_MAYWRITE;
948 	} else
949 		goto bail;
950 	len = vma->vm_end - vma->vm_start;
951 	if (len > size) {
952 		ret = -EINVAL;
953 		goto bail;
954 	}
955 
956 	vma->vm_pgoff = (unsigned long) addr >> PAGE_SHIFT;
957 	vma->vm_ops = &qib_file_vm_ops;
958 	vma->vm_flags |= VM_DONTEXPAND | VM_DONTDUMP;
959 	ret = 1;
960 
961 bail:
962 	return ret;
963 }
964 
965 /**
966  * qib_mmapf - mmap various structures into user space
967  * @fp: the file pointer
968  * @vma: the VM area
969  *
970  * We use this to have a shared buffer between the kernel and the user code
971  * for the rcvhdr queue, egr buffers, and the per-context user regs and pio
972  * buffers in the chip.  We have the open and close entries so we can bump
973  * the ref count and keep the driver from being unloaded while still mapped.
974  */
975 static int qib_mmapf(struct file *fp, struct vm_area_struct *vma)
976 {
977 	struct qib_ctxtdata *rcd;
978 	struct qib_devdata *dd;
979 	u64 pgaddr, ureg;
980 	unsigned piobufs, piocnt;
981 	int ret, match = 1;
982 
983 	rcd = ctxt_fp(fp);
984 	if (!rcd || !(vma->vm_flags & VM_SHARED)) {
985 		ret = -EINVAL;
986 		goto bail;
987 	}
988 	dd = rcd->dd;
989 
990 	/*
991 	 * This is the qib_do_user_init() code, mapping the shared buffers
992 	 * and per-context user registers into the user process. The address
993 	 * referred to by vm_pgoff is the file offset passed via mmap().
994 	 * For shared contexts, this is the kernel vmalloc() address of the
995 	 * pages to share with the master.
996 	 * For non-shared or master ctxts, this is a physical address.
997 	 * We only do one mmap for each space mapped.
998 	 */
999 	pgaddr = vma->vm_pgoff << PAGE_SHIFT;
1000 
1001 	/*
1002 	 * Check for 0 in case one of the allocations failed, but user
1003 	 * called mmap anyway.
1004 	 */
1005 	if (!pgaddr)  {
1006 		ret = -EINVAL;
1007 		goto bail;
1008 	}
1009 
1010 	/*
1011 	 * Physical addresses must fit in 40 bits for our hardware.
1012 	 * Check for kernel virtual addresses first, anything else must
1013 	 * match a HW or memory address.
1014 	 */
1015 	ret = mmap_kvaddr(vma, pgaddr, rcd, subctxt_fp(fp));
1016 	if (ret) {
1017 		if (ret > 0)
1018 			ret = 0;
1019 		goto bail;
1020 	}
1021 
1022 	ureg = dd->uregbase + dd->ureg_align * rcd->ctxt;
1023 	if (!rcd->subctxt_cnt) {
1024 		/* ctxt is not shared */
1025 		piocnt = rcd->piocnt;
1026 		piobufs = rcd->piobufs;
1027 	} else if (!subctxt_fp(fp)) {
1028 		/* caller is the master */
1029 		piocnt = (rcd->piocnt / rcd->subctxt_cnt) +
1030 			 (rcd->piocnt % rcd->subctxt_cnt);
1031 		piobufs = rcd->piobufs +
1032 			dd->palign * (rcd->piocnt - piocnt);
1033 	} else {
1034 		unsigned slave = subctxt_fp(fp) - 1;
1035 
1036 		/* caller is a slave */
1037 		piocnt = rcd->piocnt / rcd->subctxt_cnt;
1038 		piobufs = rcd->piobufs + dd->palign * piocnt * slave;
1039 	}
1040 
1041 	if (pgaddr == ureg)
1042 		ret = mmap_ureg(vma, dd, ureg);
1043 	else if (pgaddr == piobufs)
1044 		ret = mmap_piobufs(vma, dd, rcd, piobufs, piocnt);
1045 	else if (pgaddr == dd->pioavailregs_phys)
1046 		/* in-memory copy of pioavail registers */
1047 		ret = qib_mmap_mem(vma, rcd, PAGE_SIZE,
1048 				   (void *) dd->pioavailregs_dma, 0,
1049 				   "pioavail registers");
1050 	else if (pgaddr == rcd->rcvegr_phys)
1051 		ret = mmap_rcvegrbufs(vma, rcd);
1052 	else if (pgaddr == (u64) rcd->rcvhdrq_phys)
1053 		/*
1054 		 * The rcvhdrq itself; multiple pages, contiguous
1055 		 * from an i/o perspective.  Shared contexts need
1056 		 * to map r/w, so we allow writing.
1057 		 */
1058 		ret = qib_mmap_mem(vma, rcd, rcd->rcvhdrq_size,
1059 				   rcd->rcvhdrq, 1, "rcvhdrq");
1060 	else if (pgaddr == (u64) rcd->rcvhdrqtailaddr_phys)
1061 		/* in-memory copy of rcvhdrq tail register */
1062 		ret = qib_mmap_mem(vma, rcd, PAGE_SIZE,
1063 				   rcd->rcvhdrtail_kvaddr, 0,
1064 				   "rcvhdrq tail");
1065 	else
1066 		match = 0;
1067 	if (!match)
1068 		ret = -EINVAL;
1069 
1070 	vma->vm_private_data = NULL;
1071 
1072 	if (ret < 0)
1073 		qib_devinfo(dd->pcidev,
1074 			 "mmap Failure %d: off %llx len %lx\n",
1075 			 -ret, (unsigned long long)pgaddr,
1076 			 vma->vm_end - vma->vm_start);
1077 bail:
1078 	return ret;
1079 }
1080 
1081 static __poll_t qib_poll_urgent(struct qib_ctxtdata *rcd,
1082 				    struct file *fp,
1083 				    struct poll_table_struct *pt)
1084 {
1085 	struct qib_devdata *dd = rcd->dd;
1086 	__poll_t pollflag;
1087 
1088 	poll_wait(fp, &rcd->wait, pt);
1089 
1090 	spin_lock_irq(&dd->uctxt_lock);
1091 	if (rcd->urgent != rcd->urgent_poll) {
1092 		pollflag = EPOLLIN | EPOLLRDNORM;
1093 		rcd->urgent_poll = rcd->urgent;
1094 	} else {
1095 		pollflag = 0;
1096 		set_bit(QIB_CTXT_WAITING_URG, &rcd->flag);
1097 	}
1098 	spin_unlock_irq(&dd->uctxt_lock);
1099 
1100 	return pollflag;
1101 }
1102 
1103 static __poll_t qib_poll_next(struct qib_ctxtdata *rcd,
1104 				  struct file *fp,
1105 				  struct poll_table_struct *pt)
1106 {
1107 	struct qib_devdata *dd = rcd->dd;
1108 	__poll_t pollflag;
1109 
1110 	poll_wait(fp, &rcd->wait, pt);
1111 
1112 	spin_lock_irq(&dd->uctxt_lock);
1113 	if (dd->f_hdrqempty(rcd)) {
1114 		set_bit(QIB_CTXT_WAITING_RCV, &rcd->flag);
1115 		dd->f_rcvctrl(rcd->ppd, QIB_RCVCTRL_INTRAVAIL_ENB, rcd->ctxt);
1116 		pollflag = 0;
1117 	} else
1118 		pollflag = EPOLLIN | EPOLLRDNORM;
1119 	spin_unlock_irq(&dd->uctxt_lock);
1120 
1121 	return pollflag;
1122 }
1123 
1124 static __poll_t qib_poll(struct file *fp, struct poll_table_struct *pt)
1125 {
1126 	struct qib_ctxtdata *rcd;
1127 	__poll_t pollflag;
1128 
1129 	rcd = ctxt_fp(fp);
1130 	if (!rcd)
1131 		pollflag = EPOLLERR;
1132 	else if (rcd->poll_type == QIB_POLL_TYPE_URGENT)
1133 		pollflag = qib_poll_urgent(rcd, fp, pt);
1134 	else  if (rcd->poll_type == QIB_POLL_TYPE_ANYRCV)
1135 		pollflag = qib_poll_next(rcd, fp, pt);
1136 	else /* invalid */
1137 		pollflag = EPOLLERR;
1138 
1139 	return pollflag;
1140 }
1141 
1142 static void assign_ctxt_affinity(struct file *fp, struct qib_devdata *dd)
1143 {
1144 	struct qib_filedata *fd = fp->private_data;
1145 	const unsigned int weight = current->nr_cpus_allowed;
1146 	const struct cpumask *local_mask = cpumask_of_pcibus(dd->pcidev->bus);
1147 	int local_cpu;
1148 
1149 	/*
1150 	 * If process has NOT already set it's affinity, select and
1151 	 * reserve a processor for it on the local NUMA node.
1152 	 */
1153 	if ((weight >= qib_cpulist_count) &&
1154 		(cpumask_weight(local_mask) <= qib_cpulist_count)) {
1155 		for_each_cpu(local_cpu, local_mask)
1156 			if (!test_and_set_bit(local_cpu, qib_cpulist)) {
1157 				fd->rec_cpu_num = local_cpu;
1158 				return;
1159 			}
1160 	}
1161 
1162 	/*
1163 	 * If process has NOT already set it's affinity, select and
1164 	 * reserve a processor for it, as a rendevous for all
1165 	 * users of the driver.  If they don't actually later
1166 	 * set affinity to this cpu, or set it to some other cpu,
1167 	 * it just means that sooner or later we don't recommend
1168 	 * a cpu, and let the scheduler do it's best.
1169 	 */
1170 	if (weight >= qib_cpulist_count) {
1171 		int cpu;
1172 
1173 		cpu = find_first_zero_bit(qib_cpulist,
1174 					  qib_cpulist_count);
1175 		if (cpu == qib_cpulist_count)
1176 			qib_dev_err(dd,
1177 			"no cpus avail for affinity PID %u\n",
1178 			current->pid);
1179 		else {
1180 			__set_bit(cpu, qib_cpulist);
1181 			fd->rec_cpu_num = cpu;
1182 		}
1183 	}
1184 }
1185 
1186 /*
1187  * Check that userland and driver are compatible for subcontexts.
1188  */
1189 static int qib_compatible_subctxts(int user_swmajor, int user_swminor)
1190 {
1191 	/* this code is written long-hand for clarity */
1192 	if (QIB_USER_SWMAJOR != user_swmajor) {
1193 		/* no promise of compatibility if major mismatch */
1194 		return 0;
1195 	}
1196 	if (QIB_USER_SWMAJOR == 1) {
1197 		switch (QIB_USER_SWMINOR) {
1198 		case 0:
1199 		case 1:
1200 		case 2:
1201 			/* no subctxt implementation so cannot be compatible */
1202 			return 0;
1203 		case 3:
1204 			/* 3 is only compatible with itself */
1205 			return user_swminor == 3;
1206 		default:
1207 			/* >= 4 are compatible (or are expected to be) */
1208 			return user_swminor <= QIB_USER_SWMINOR;
1209 		}
1210 	}
1211 	/* make no promises yet for future major versions */
1212 	return 0;
1213 }
1214 
1215 static int init_subctxts(struct qib_devdata *dd,
1216 			 struct qib_ctxtdata *rcd,
1217 			 const struct qib_user_info *uinfo)
1218 {
1219 	int ret = 0;
1220 	unsigned num_subctxts;
1221 	size_t size;
1222 
1223 	/*
1224 	 * If the user is requesting zero subctxts,
1225 	 * skip the subctxt allocation.
1226 	 */
1227 	if (uinfo->spu_subctxt_cnt <= 0)
1228 		goto bail;
1229 	num_subctxts = uinfo->spu_subctxt_cnt;
1230 
1231 	/* Check for subctxt compatibility */
1232 	if (!qib_compatible_subctxts(uinfo->spu_userversion >> 16,
1233 		uinfo->spu_userversion & 0xffff)) {
1234 		qib_devinfo(dd->pcidev,
1235 			 "Mismatched user version (%d.%d) and driver version (%d.%d) while context sharing. Ensure that driver and library are from the same release.\n",
1236 			 (int) (uinfo->spu_userversion >> 16),
1237 			 (int) (uinfo->spu_userversion & 0xffff),
1238 			 QIB_USER_SWMAJOR, QIB_USER_SWMINOR);
1239 		goto bail;
1240 	}
1241 	if (num_subctxts > QLOGIC_IB_MAX_SUBCTXT) {
1242 		ret = -EINVAL;
1243 		goto bail;
1244 	}
1245 
1246 	rcd->subctxt_uregbase = vmalloc_user(PAGE_SIZE * num_subctxts);
1247 	if (!rcd->subctxt_uregbase) {
1248 		ret = -ENOMEM;
1249 		goto bail;
1250 	}
1251 	/* Note: rcd->rcvhdrq_size isn't initialized yet. */
1252 	size = ALIGN(dd->rcvhdrcnt * dd->rcvhdrentsize *
1253 		     sizeof(u32), PAGE_SIZE) * num_subctxts;
1254 	rcd->subctxt_rcvhdr_base = vmalloc_user(size);
1255 	if (!rcd->subctxt_rcvhdr_base) {
1256 		ret = -ENOMEM;
1257 		goto bail_ureg;
1258 	}
1259 
1260 	rcd->subctxt_rcvegrbuf = vmalloc_user(rcd->rcvegrbuf_chunks *
1261 					      rcd->rcvegrbuf_size *
1262 					      num_subctxts);
1263 	if (!rcd->subctxt_rcvegrbuf) {
1264 		ret = -ENOMEM;
1265 		goto bail_rhdr;
1266 	}
1267 
1268 	rcd->subctxt_cnt = uinfo->spu_subctxt_cnt;
1269 	rcd->subctxt_id = uinfo->spu_subctxt_id;
1270 	rcd->active_slaves = 1;
1271 	rcd->redirect_seq_cnt = 1;
1272 	set_bit(QIB_CTXT_MASTER_UNINIT, &rcd->flag);
1273 	goto bail;
1274 
1275 bail_rhdr:
1276 	vfree(rcd->subctxt_rcvhdr_base);
1277 bail_ureg:
1278 	vfree(rcd->subctxt_uregbase);
1279 	rcd->subctxt_uregbase = NULL;
1280 bail:
1281 	return ret;
1282 }
1283 
1284 static int setup_ctxt(struct qib_pportdata *ppd, int ctxt,
1285 		      struct file *fp, const struct qib_user_info *uinfo)
1286 {
1287 	struct qib_filedata *fd = fp->private_data;
1288 	struct qib_devdata *dd = ppd->dd;
1289 	struct qib_ctxtdata *rcd;
1290 	void *ptmp = NULL;
1291 	int ret;
1292 	int numa_id;
1293 
1294 	assign_ctxt_affinity(fp, dd);
1295 
1296 	numa_id = qib_numa_aware ? ((fd->rec_cpu_num != -1) ?
1297 		cpu_to_node(fd->rec_cpu_num) :
1298 		numa_node_id()) : dd->assigned_node_id;
1299 
1300 	rcd = qib_create_ctxtdata(ppd, ctxt, numa_id);
1301 
1302 	/*
1303 	 * Allocate memory for use in qib_tid_update() at open to
1304 	 * reduce cost of expected send setup per message segment
1305 	 */
1306 	if (rcd)
1307 		ptmp = kmalloc(dd->rcvtidcnt * sizeof(u16) +
1308 			       dd->rcvtidcnt * sizeof(struct page **),
1309 			       GFP_KERNEL);
1310 
1311 	if (!rcd || !ptmp) {
1312 		qib_dev_err(dd,
1313 			"Unable to allocate ctxtdata memory, failing open\n");
1314 		ret = -ENOMEM;
1315 		goto bailerr;
1316 	}
1317 	rcd->userversion = uinfo->spu_userversion;
1318 	ret = init_subctxts(dd, rcd, uinfo);
1319 	if (ret)
1320 		goto bailerr;
1321 	rcd->tid_pg_list = ptmp;
1322 	rcd->pid = current->pid;
1323 	init_waitqueue_head(&dd->rcd[ctxt]->wait);
1324 	get_task_comm(rcd->comm, current);
1325 	ctxt_fp(fp) = rcd;
1326 	qib_stats.sps_ctxts++;
1327 	dd->freectxts--;
1328 	ret = 0;
1329 	goto bail;
1330 
1331 bailerr:
1332 	if (fd->rec_cpu_num != -1)
1333 		__clear_bit(fd->rec_cpu_num, qib_cpulist);
1334 
1335 	dd->rcd[ctxt] = NULL;
1336 	kfree(rcd);
1337 	kfree(ptmp);
1338 bail:
1339 	return ret;
1340 }
1341 
1342 static inline int usable(struct qib_pportdata *ppd)
1343 {
1344 	struct qib_devdata *dd = ppd->dd;
1345 
1346 	return dd && (dd->flags & QIB_PRESENT) && dd->kregbase && ppd->lid &&
1347 		(ppd->lflags & QIBL_LINKACTIVE);
1348 }
1349 
1350 /*
1351  * Select a context on the given device, either using a requested port
1352  * or the port based on the context number.
1353  */
1354 static int choose_port_ctxt(struct file *fp, struct qib_devdata *dd, u32 port,
1355 			    const struct qib_user_info *uinfo)
1356 {
1357 	struct qib_pportdata *ppd = NULL;
1358 	int ret, ctxt;
1359 
1360 	if (port) {
1361 		if (!usable(dd->pport + port - 1)) {
1362 			ret = -ENETDOWN;
1363 			goto done;
1364 		} else
1365 			ppd = dd->pport + port - 1;
1366 	}
1367 	for (ctxt = dd->first_user_ctxt; ctxt < dd->cfgctxts && dd->rcd[ctxt];
1368 	     ctxt++)
1369 		;
1370 	if (ctxt == dd->cfgctxts) {
1371 		ret = -EBUSY;
1372 		goto done;
1373 	}
1374 	if (!ppd) {
1375 		u32 pidx = ctxt % dd->num_pports;
1376 
1377 		if (usable(dd->pport + pidx))
1378 			ppd = dd->pport + pidx;
1379 		else {
1380 			for (pidx = 0; pidx < dd->num_pports && !ppd;
1381 			     pidx++)
1382 				if (usable(dd->pport + pidx))
1383 					ppd = dd->pport + pidx;
1384 		}
1385 	}
1386 	ret = ppd ? setup_ctxt(ppd, ctxt, fp, uinfo) : -ENETDOWN;
1387 done:
1388 	return ret;
1389 }
1390 
1391 static int find_free_ctxt(int unit, struct file *fp,
1392 			  const struct qib_user_info *uinfo)
1393 {
1394 	struct qib_devdata *dd = qib_lookup(unit);
1395 	int ret;
1396 
1397 	if (!dd || (uinfo->spu_port && uinfo->spu_port > dd->num_pports))
1398 		ret = -ENODEV;
1399 	else
1400 		ret = choose_port_ctxt(fp, dd, uinfo->spu_port, uinfo);
1401 
1402 	return ret;
1403 }
1404 
1405 static int get_a_ctxt(struct file *fp, const struct qib_user_info *uinfo,
1406 		      unsigned alg)
1407 {
1408 	struct qib_devdata *udd = NULL;
1409 	int ret = 0, devmax, npresent, nup, ndev, dusable = 0, i;
1410 	u32 port = uinfo->spu_port, ctxt;
1411 
1412 	devmax = qib_count_units(&npresent, &nup);
1413 	if (!npresent) {
1414 		ret = -ENXIO;
1415 		goto done;
1416 	}
1417 	if (nup == 0) {
1418 		ret = -ENETDOWN;
1419 		goto done;
1420 	}
1421 
1422 	if (alg == QIB_PORT_ALG_ACROSS) {
1423 		unsigned inuse = ~0U;
1424 
1425 		/* find device (with ACTIVE ports) with fewest ctxts in use */
1426 		for (ndev = 0; ndev < devmax; ndev++) {
1427 			struct qib_devdata *dd = qib_lookup(ndev);
1428 			unsigned cused = 0, cfree = 0, pusable = 0;
1429 
1430 			if (!dd)
1431 				continue;
1432 			if (port && port <= dd->num_pports &&
1433 			    usable(dd->pport + port - 1))
1434 				pusable = 1;
1435 			else
1436 				for (i = 0; i < dd->num_pports; i++)
1437 					if (usable(dd->pport + i))
1438 						pusable++;
1439 			if (!pusable)
1440 				continue;
1441 			for (ctxt = dd->first_user_ctxt; ctxt < dd->cfgctxts;
1442 			     ctxt++)
1443 				if (dd->rcd[ctxt])
1444 					cused++;
1445 				else
1446 					cfree++;
1447 			if (cfree && cused < inuse) {
1448 				udd = dd;
1449 				inuse = cused;
1450 			}
1451 		}
1452 		if (udd) {
1453 			ret = choose_port_ctxt(fp, udd, port, uinfo);
1454 			goto done;
1455 		}
1456 	} else {
1457 		for (ndev = 0; ndev < devmax; ndev++) {
1458 			struct qib_devdata *dd = qib_lookup(ndev);
1459 
1460 			if (dd) {
1461 				ret = choose_port_ctxt(fp, dd, port, uinfo);
1462 				if (!ret)
1463 					goto done;
1464 				if (ret == -EBUSY)
1465 					dusable++;
1466 			}
1467 		}
1468 	}
1469 	ret = dusable ? -EBUSY : -ENETDOWN;
1470 
1471 done:
1472 	return ret;
1473 }
1474 
1475 static int find_shared_ctxt(struct file *fp,
1476 			    const struct qib_user_info *uinfo)
1477 {
1478 	int devmax, ndev, i;
1479 	int ret = 0;
1480 
1481 	devmax = qib_count_units(NULL, NULL);
1482 
1483 	for (ndev = 0; ndev < devmax; ndev++) {
1484 		struct qib_devdata *dd = qib_lookup(ndev);
1485 
1486 		/* device portion of usable() */
1487 		if (!(dd && (dd->flags & QIB_PRESENT) && dd->kregbase))
1488 			continue;
1489 		for (i = dd->first_user_ctxt; i < dd->cfgctxts; i++) {
1490 			struct qib_ctxtdata *rcd = dd->rcd[i];
1491 
1492 			/* Skip ctxts which are not yet open */
1493 			if (!rcd || !rcd->cnt)
1494 				continue;
1495 			/* Skip ctxt if it doesn't match the requested one */
1496 			if (rcd->subctxt_id != uinfo->spu_subctxt_id)
1497 				continue;
1498 			/* Verify the sharing process matches the master */
1499 			if (rcd->subctxt_cnt != uinfo->spu_subctxt_cnt ||
1500 			    rcd->userversion != uinfo->spu_userversion ||
1501 			    rcd->cnt >= rcd->subctxt_cnt) {
1502 				ret = -EINVAL;
1503 				goto done;
1504 			}
1505 			ctxt_fp(fp) = rcd;
1506 			subctxt_fp(fp) = rcd->cnt++;
1507 			rcd->subpid[subctxt_fp(fp)] = current->pid;
1508 			tidcursor_fp(fp) = 0;
1509 			rcd->active_slaves |= 1 << subctxt_fp(fp);
1510 			ret = 1;
1511 			goto done;
1512 		}
1513 	}
1514 
1515 done:
1516 	return ret;
1517 }
1518 
1519 static int qib_open(struct inode *in, struct file *fp)
1520 {
1521 	/* The real work is performed later in qib_assign_ctxt() */
1522 	fp->private_data = kzalloc(sizeof(struct qib_filedata), GFP_KERNEL);
1523 	if (fp->private_data) /* no cpu affinity by default */
1524 		((struct qib_filedata *)fp->private_data)->rec_cpu_num = -1;
1525 	return fp->private_data ? 0 : -ENOMEM;
1526 }
1527 
1528 static int find_hca(unsigned int cpu, int *unit)
1529 {
1530 	int ret = 0, devmax, npresent, nup, ndev;
1531 
1532 	*unit = -1;
1533 
1534 	devmax = qib_count_units(&npresent, &nup);
1535 	if (!npresent) {
1536 		ret = -ENXIO;
1537 		goto done;
1538 	}
1539 	if (!nup) {
1540 		ret = -ENETDOWN;
1541 		goto done;
1542 	}
1543 	for (ndev = 0; ndev < devmax; ndev++) {
1544 		struct qib_devdata *dd = qib_lookup(ndev);
1545 
1546 		if (dd) {
1547 			if (pcibus_to_node(dd->pcidev->bus) < 0) {
1548 				ret = -EINVAL;
1549 				goto done;
1550 			}
1551 			if (cpu_to_node(cpu) ==
1552 				pcibus_to_node(dd->pcidev->bus)) {
1553 				*unit = ndev;
1554 				goto done;
1555 			}
1556 		}
1557 	}
1558 done:
1559 	return ret;
1560 }
1561 
1562 static int do_qib_user_sdma_queue_create(struct file *fp)
1563 {
1564 	struct qib_filedata *fd = fp->private_data;
1565 	struct qib_ctxtdata *rcd = fd->rcd;
1566 	struct qib_devdata *dd = rcd->dd;
1567 
1568 	if (dd->flags & QIB_HAS_SEND_DMA) {
1569 
1570 		fd->pq = qib_user_sdma_queue_create(&dd->pcidev->dev,
1571 						    dd->unit,
1572 						    rcd->ctxt,
1573 						    fd->subctxt);
1574 		if (!fd->pq)
1575 			return -ENOMEM;
1576 	}
1577 
1578 	return 0;
1579 }
1580 
1581 /*
1582  * Get ctxt early, so can set affinity prior to memory allocation.
1583  */
1584 static int qib_assign_ctxt(struct file *fp, const struct qib_user_info *uinfo)
1585 {
1586 	int ret;
1587 	int i_minor;
1588 	unsigned swmajor, swminor, alg = QIB_PORT_ALG_ACROSS;
1589 
1590 	/* Check to be sure we haven't already initialized this file */
1591 	if (ctxt_fp(fp)) {
1592 		ret = -EINVAL;
1593 		goto done;
1594 	}
1595 
1596 	/* for now, if major version is different, bail */
1597 	swmajor = uinfo->spu_userversion >> 16;
1598 	if (swmajor != QIB_USER_SWMAJOR) {
1599 		ret = -ENODEV;
1600 		goto done;
1601 	}
1602 
1603 	swminor = uinfo->spu_userversion & 0xffff;
1604 
1605 	if (swminor >= 11 && uinfo->spu_port_alg < QIB_PORT_ALG_COUNT)
1606 		alg = uinfo->spu_port_alg;
1607 
1608 	mutex_lock(&qib_mutex);
1609 
1610 	if (qib_compatible_subctxts(swmajor, swminor) &&
1611 	    uinfo->spu_subctxt_cnt) {
1612 		ret = find_shared_ctxt(fp, uinfo);
1613 		if (ret > 0) {
1614 			ret = do_qib_user_sdma_queue_create(fp);
1615 			if (!ret)
1616 				assign_ctxt_affinity(fp, (ctxt_fp(fp))->dd);
1617 			goto done_ok;
1618 		}
1619 	}
1620 
1621 	i_minor = iminor(file_inode(fp)) - QIB_USER_MINOR_BASE;
1622 	if (i_minor)
1623 		ret = find_free_ctxt(i_minor - 1, fp, uinfo);
1624 	else {
1625 		int unit;
1626 		const unsigned int cpu = cpumask_first(current->cpus_ptr);
1627 		const unsigned int weight = current->nr_cpus_allowed;
1628 
1629 		if (weight == 1 && !test_bit(cpu, qib_cpulist))
1630 			if (!find_hca(cpu, &unit) && unit >= 0)
1631 				if (!find_free_ctxt(unit, fp, uinfo)) {
1632 					ret = 0;
1633 					goto done_chk_sdma;
1634 				}
1635 		ret = get_a_ctxt(fp, uinfo, alg);
1636 	}
1637 
1638 done_chk_sdma:
1639 	if (!ret)
1640 		ret = do_qib_user_sdma_queue_create(fp);
1641 done_ok:
1642 	mutex_unlock(&qib_mutex);
1643 
1644 done:
1645 	return ret;
1646 }
1647 
1648 
1649 static int qib_do_user_init(struct file *fp,
1650 			    const struct qib_user_info *uinfo)
1651 {
1652 	int ret;
1653 	struct qib_ctxtdata *rcd = ctxt_fp(fp);
1654 	struct qib_devdata *dd;
1655 	unsigned uctxt;
1656 
1657 	/* Subctxts don't need to initialize anything since master did it. */
1658 	if (subctxt_fp(fp)) {
1659 		ret = wait_event_interruptible(rcd->wait,
1660 			!test_bit(QIB_CTXT_MASTER_UNINIT, &rcd->flag));
1661 		goto bail;
1662 	}
1663 
1664 	dd = rcd->dd;
1665 
1666 	/* some ctxts may get extra buffers, calculate that here */
1667 	uctxt = rcd->ctxt - dd->first_user_ctxt;
1668 	if (uctxt < dd->ctxts_extrabuf) {
1669 		rcd->piocnt = dd->pbufsctxt + 1;
1670 		rcd->pio_base = rcd->piocnt * uctxt;
1671 	} else {
1672 		rcd->piocnt = dd->pbufsctxt;
1673 		rcd->pio_base = rcd->piocnt * uctxt +
1674 			dd->ctxts_extrabuf;
1675 	}
1676 
1677 	/*
1678 	 * All user buffers are 2KB buffers.  If we ever support
1679 	 * giving 4KB buffers to user processes, this will need some
1680 	 * work.  Can't use piobufbase directly, because it has
1681 	 * both 2K and 4K buffer base values.  So check and handle.
1682 	 */
1683 	if ((rcd->pio_base + rcd->piocnt) > dd->piobcnt2k) {
1684 		if (rcd->pio_base >= dd->piobcnt2k) {
1685 			qib_dev_err(dd,
1686 				    "%u:ctxt%u: no 2KB buffers available\n",
1687 				    dd->unit, rcd->ctxt);
1688 			ret = -ENOBUFS;
1689 			goto bail;
1690 		}
1691 		rcd->piocnt = dd->piobcnt2k - rcd->pio_base;
1692 		qib_dev_err(dd, "Ctxt%u: would use 4KB bufs, using %u\n",
1693 			    rcd->ctxt, rcd->piocnt);
1694 	}
1695 
1696 	rcd->piobufs = dd->pio2k_bufbase + rcd->pio_base * dd->palign;
1697 	qib_chg_pioavailkernel(dd, rcd->pio_base, rcd->piocnt,
1698 			       TXCHK_CHG_TYPE_USER, rcd);
1699 	/*
1700 	 * try to ensure that processes start up with consistent avail update
1701 	 * for their own range, at least.   If system very quiet, it might
1702 	 * have the in-memory copy out of date at startup for this range of
1703 	 * buffers, when a context gets re-used.  Do after the chg_pioavail
1704 	 * and before the rest of setup, so it's "almost certain" the dma
1705 	 * will have occurred (can't 100% guarantee, but should be many
1706 	 * decimals of 9s, with this ordering), given how much else happens
1707 	 * after this.
1708 	 */
1709 	dd->f_sendctrl(dd->pport, QIB_SENDCTRL_AVAIL_BLIP);
1710 
1711 	/*
1712 	 * Now allocate the rcvhdr Q and eager TIDs; skip the TID
1713 	 * array for time being.  If rcd->ctxt > chip-supported,
1714 	 * we need to do extra stuff here to handle by handling overflow
1715 	 * through ctxt 0, someday
1716 	 */
1717 	ret = qib_create_rcvhdrq(dd, rcd);
1718 	if (!ret)
1719 		ret = qib_setup_eagerbufs(rcd);
1720 	if (ret)
1721 		goto bail_pio;
1722 
1723 	rcd->tidcursor = 0; /* start at beginning after open */
1724 
1725 	/* initialize poll variables... */
1726 	rcd->urgent = 0;
1727 	rcd->urgent_poll = 0;
1728 
1729 	/*
1730 	 * Now enable the ctxt for receive.
1731 	 * For chips that are set to DMA the tail register to memory
1732 	 * when they change (and when the update bit transitions from
1733 	 * 0 to 1.  So for those chips, we turn it off and then back on.
1734 	 * This will (very briefly) affect any other open ctxts, but the
1735 	 * duration is very short, and therefore isn't an issue.  We
1736 	 * explicitly set the in-memory tail copy to 0 beforehand, so we
1737 	 * don't have to wait to be sure the DMA update has happened
1738 	 * (chip resets head/tail to 0 on transition to enable).
1739 	 */
1740 	if (rcd->rcvhdrtail_kvaddr)
1741 		qib_clear_rcvhdrtail(rcd);
1742 
1743 	dd->f_rcvctrl(rcd->ppd, QIB_RCVCTRL_CTXT_ENB | QIB_RCVCTRL_TIDFLOW_ENB,
1744 		      rcd->ctxt);
1745 
1746 	/* Notify any waiting slaves */
1747 	if (rcd->subctxt_cnt) {
1748 		clear_bit(QIB_CTXT_MASTER_UNINIT, &rcd->flag);
1749 		wake_up(&rcd->wait);
1750 	}
1751 	return 0;
1752 
1753 bail_pio:
1754 	qib_chg_pioavailkernel(dd, rcd->pio_base, rcd->piocnt,
1755 			       TXCHK_CHG_TYPE_KERN, rcd);
1756 bail:
1757 	return ret;
1758 }
1759 
1760 /**
1761  * unlock_expected_tids - unlock any expected TID entries context still had
1762  * in use
1763  * @rcd: ctxt
1764  *
1765  * We don't actually update the chip here, because we do a bulk update
1766  * below, using f_clear_tids.
1767  */
1768 static void unlock_expected_tids(struct qib_ctxtdata *rcd)
1769 {
1770 	struct qib_devdata *dd = rcd->dd;
1771 	int ctxt_tidbase = rcd->ctxt * dd->rcvtidcnt;
1772 	int i, cnt = 0, maxtid = ctxt_tidbase + dd->rcvtidcnt;
1773 
1774 	for (i = ctxt_tidbase; i < maxtid; i++) {
1775 		struct page *p = dd->pageshadow[i];
1776 		dma_addr_t phys;
1777 
1778 		if (!p)
1779 			continue;
1780 
1781 		phys = dd->physshadow[i];
1782 		dd->physshadow[i] = dd->tidinvalid;
1783 		dd->pageshadow[i] = NULL;
1784 		dma_unmap_page(&dd->pcidev->dev, phys, PAGE_SIZE,
1785 			       DMA_FROM_DEVICE);
1786 		qib_release_user_pages(&p, 1);
1787 		cnt++;
1788 	}
1789 }
1790 
1791 static int qib_close(struct inode *in, struct file *fp)
1792 {
1793 	struct qib_filedata *fd;
1794 	struct qib_ctxtdata *rcd;
1795 	struct qib_devdata *dd;
1796 	unsigned long flags;
1797 	unsigned ctxt;
1798 
1799 	mutex_lock(&qib_mutex);
1800 
1801 	fd = fp->private_data;
1802 	fp->private_data = NULL;
1803 	rcd = fd->rcd;
1804 	if (!rcd) {
1805 		mutex_unlock(&qib_mutex);
1806 		goto bail;
1807 	}
1808 
1809 	dd = rcd->dd;
1810 
1811 	/* ensure all pio buffer writes in progress are flushed */
1812 	qib_flush_wc();
1813 
1814 	/* drain user sdma queue */
1815 	if (fd->pq) {
1816 		qib_user_sdma_queue_drain(rcd->ppd, fd->pq);
1817 		qib_user_sdma_queue_destroy(fd->pq);
1818 	}
1819 
1820 	if (fd->rec_cpu_num != -1)
1821 		__clear_bit(fd->rec_cpu_num, qib_cpulist);
1822 
1823 	if (--rcd->cnt) {
1824 		/*
1825 		 * XXX If the master closes the context before the slave(s),
1826 		 * revoke the mmap for the eager receive queue so
1827 		 * the slave(s) don't wait for receive data forever.
1828 		 */
1829 		rcd->active_slaves &= ~(1 << fd->subctxt);
1830 		rcd->subpid[fd->subctxt] = 0;
1831 		mutex_unlock(&qib_mutex);
1832 		goto bail;
1833 	}
1834 
1835 	/* early; no interrupt users after this */
1836 	spin_lock_irqsave(&dd->uctxt_lock, flags);
1837 	ctxt = rcd->ctxt;
1838 	dd->rcd[ctxt] = NULL;
1839 	rcd->pid = 0;
1840 	spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1841 
1842 	if (rcd->rcvwait_to || rcd->piowait_to ||
1843 	    rcd->rcvnowait || rcd->pionowait) {
1844 		rcd->rcvwait_to = 0;
1845 		rcd->piowait_to = 0;
1846 		rcd->rcvnowait = 0;
1847 		rcd->pionowait = 0;
1848 	}
1849 	if (rcd->flag)
1850 		rcd->flag = 0;
1851 
1852 	if (dd->kregbase) {
1853 		/* atomically clear receive enable ctxt and intr avail. */
1854 		dd->f_rcvctrl(rcd->ppd, QIB_RCVCTRL_CTXT_DIS |
1855 				  QIB_RCVCTRL_INTRAVAIL_DIS, ctxt);
1856 
1857 		/* clean up the pkeys for this ctxt user */
1858 		qib_clean_part_key(rcd, dd);
1859 		qib_disarm_piobufs(dd, rcd->pio_base, rcd->piocnt);
1860 		qib_chg_pioavailkernel(dd, rcd->pio_base,
1861 				       rcd->piocnt, TXCHK_CHG_TYPE_KERN, NULL);
1862 
1863 		dd->f_clear_tids(dd, rcd);
1864 
1865 		if (dd->pageshadow)
1866 			unlock_expected_tids(rcd);
1867 		qib_stats.sps_ctxts--;
1868 		dd->freectxts++;
1869 	}
1870 
1871 	mutex_unlock(&qib_mutex);
1872 	qib_free_ctxtdata(dd, rcd); /* after releasing the mutex */
1873 
1874 bail:
1875 	kfree(fd);
1876 	return 0;
1877 }
1878 
1879 static int qib_ctxt_info(struct file *fp, struct qib_ctxt_info __user *uinfo)
1880 {
1881 	struct qib_ctxt_info info;
1882 	int ret;
1883 	size_t sz;
1884 	struct qib_ctxtdata *rcd = ctxt_fp(fp);
1885 	struct qib_filedata *fd;
1886 
1887 	fd = fp->private_data;
1888 
1889 	info.num_active = qib_count_active_units();
1890 	info.unit = rcd->dd->unit;
1891 	info.port = rcd->ppd->port;
1892 	info.ctxt = rcd->ctxt;
1893 	info.subctxt =  subctxt_fp(fp);
1894 	/* Number of user ctxts available for this device. */
1895 	info.num_ctxts = rcd->dd->cfgctxts - rcd->dd->first_user_ctxt;
1896 	info.num_subctxts = rcd->subctxt_cnt;
1897 	info.rec_cpu = fd->rec_cpu_num;
1898 	sz = sizeof(info);
1899 
1900 	if (copy_to_user(uinfo, &info, sz)) {
1901 		ret = -EFAULT;
1902 		goto bail;
1903 	}
1904 	ret = 0;
1905 
1906 bail:
1907 	return ret;
1908 }
1909 
1910 static int qib_sdma_get_inflight(struct qib_user_sdma_queue *pq,
1911 				 u32 __user *inflightp)
1912 {
1913 	const u32 val = qib_user_sdma_inflight_counter(pq);
1914 
1915 	if (put_user(val, inflightp))
1916 		return -EFAULT;
1917 
1918 	return 0;
1919 }
1920 
1921 static int qib_sdma_get_complete(struct qib_pportdata *ppd,
1922 				 struct qib_user_sdma_queue *pq,
1923 				 u32 __user *completep)
1924 {
1925 	u32 val;
1926 	int err;
1927 
1928 	if (!pq)
1929 		return -EINVAL;
1930 
1931 	err = qib_user_sdma_make_progress(ppd, pq);
1932 	if (err < 0)
1933 		return err;
1934 
1935 	val = qib_user_sdma_complete_counter(pq);
1936 	if (put_user(val, completep))
1937 		return -EFAULT;
1938 
1939 	return 0;
1940 }
1941 
1942 static int disarm_req_delay(struct qib_ctxtdata *rcd)
1943 {
1944 	int ret = 0;
1945 
1946 	if (!usable(rcd->ppd)) {
1947 		int i;
1948 		/*
1949 		 * if link is down, or otherwise not usable, delay
1950 		 * the caller up to 30 seconds, so we don't thrash
1951 		 * in trying to get the chip back to ACTIVE, and
1952 		 * set flag so they make the call again.
1953 		 */
1954 		if (rcd->user_event_mask) {
1955 			/*
1956 			 * subctxt_cnt is 0 if not shared, so do base
1957 			 * separately, first, then remaining subctxt, if any
1958 			 */
1959 			set_bit(_QIB_EVENT_DISARM_BUFS_BIT,
1960 				&rcd->user_event_mask[0]);
1961 			for (i = 1; i < rcd->subctxt_cnt; i++)
1962 				set_bit(_QIB_EVENT_DISARM_BUFS_BIT,
1963 					&rcd->user_event_mask[i]);
1964 		}
1965 		for (i = 0; !usable(rcd->ppd) && i < 300; i++)
1966 			msleep(100);
1967 		ret = -ENETDOWN;
1968 	}
1969 	return ret;
1970 }
1971 
1972 /*
1973  * Find all user contexts in use, and set the specified bit in their
1974  * event mask.
1975  * See also find_ctxt() for a similar use, that is specific to send buffers.
1976  */
1977 int qib_set_uevent_bits(struct qib_pportdata *ppd, const int evtbit)
1978 {
1979 	struct qib_ctxtdata *rcd;
1980 	unsigned ctxt;
1981 	int ret = 0;
1982 	unsigned long flags;
1983 
1984 	spin_lock_irqsave(&ppd->dd->uctxt_lock, flags);
1985 	for (ctxt = ppd->dd->first_user_ctxt; ctxt < ppd->dd->cfgctxts;
1986 	     ctxt++) {
1987 		rcd = ppd->dd->rcd[ctxt];
1988 		if (!rcd)
1989 			continue;
1990 		if (rcd->user_event_mask) {
1991 			int i;
1992 			/*
1993 			 * subctxt_cnt is 0 if not shared, so do base
1994 			 * separately, first, then remaining subctxt, if any
1995 			 */
1996 			set_bit(evtbit, &rcd->user_event_mask[0]);
1997 			for (i = 1; i < rcd->subctxt_cnt; i++)
1998 				set_bit(evtbit, &rcd->user_event_mask[i]);
1999 		}
2000 		ret = 1;
2001 		break;
2002 	}
2003 	spin_unlock_irqrestore(&ppd->dd->uctxt_lock, flags);
2004 
2005 	return ret;
2006 }
2007 
2008 /*
2009  * clear the event notifier events for this context.
2010  * For the DISARM_BUFS case, we also take action (this obsoletes
2011  * the older QIB_CMD_DISARM_BUFS, but we keep it for backwards
2012  * compatibility.
2013  * Other bits don't currently require actions, just atomically clear.
2014  * User process then performs actions appropriate to bit having been
2015  * set, if desired, and checks again in future.
2016  */
2017 static int qib_user_event_ack(struct qib_ctxtdata *rcd, int subctxt,
2018 			      unsigned long events)
2019 {
2020 	int ret = 0, i;
2021 
2022 	for (i = 0; i <= _QIB_MAX_EVENT_BIT; i++) {
2023 		if (!test_bit(i, &events))
2024 			continue;
2025 		if (i == _QIB_EVENT_DISARM_BUFS_BIT) {
2026 			(void)qib_disarm_piobufs_ifneeded(rcd);
2027 			ret = disarm_req_delay(rcd);
2028 		} else
2029 			clear_bit(i, &rcd->user_event_mask[subctxt]);
2030 	}
2031 	return ret;
2032 }
2033 
2034 static ssize_t qib_write(struct file *fp, const char __user *data,
2035 			 size_t count, loff_t *off)
2036 {
2037 	const struct qib_cmd __user *ucmd;
2038 	struct qib_ctxtdata *rcd;
2039 	const void __user *src;
2040 	size_t consumed, copy = 0;
2041 	struct qib_cmd cmd;
2042 	ssize_t ret = 0;
2043 	void *dest;
2044 
2045 	if (!ib_safe_file_access(fp)) {
2046 		pr_err_once("qib_write: process %d (%s) changed security contexts after opening file descriptor, this is not allowed.\n",
2047 			    task_tgid_vnr(current), current->comm);
2048 		return -EACCES;
2049 	}
2050 
2051 	if (count < sizeof(cmd.type)) {
2052 		ret = -EINVAL;
2053 		goto bail;
2054 	}
2055 
2056 	ucmd = (const struct qib_cmd __user *) data;
2057 
2058 	if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) {
2059 		ret = -EFAULT;
2060 		goto bail;
2061 	}
2062 
2063 	consumed = sizeof(cmd.type);
2064 
2065 	switch (cmd.type) {
2066 	case QIB_CMD_ASSIGN_CTXT:
2067 	case QIB_CMD_USER_INIT:
2068 		copy = sizeof(cmd.cmd.user_info);
2069 		dest = &cmd.cmd.user_info;
2070 		src = &ucmd->cmd.user_info;
2071 		break;
2072 
2073 	case QIB_CMD_RECV_CTRL:
2074 		copy = sizeof(cmd.cmd.recv_ctrl);
2075 		dest = &cmd.cmd.recv_ctrl;
2076 		src = &ucmd->cmd.recv_ctrl;
2077 		break;
2078 
2079 	case QIB_CMD_CTXT_INFO:
2080 		copy = sizeof(cmd.cmd.ctxt_info);
2081 		dest = &cmd.cmd.ctxt_info;
2082 		src = &ucmd->cmd.ctxt_info;
2083 		break;
2084 
2085 	case QIB_CMD_TID_UPDATE:
2086 	case QIB_CMD_TID_FREE:
2087 		copy = sizeof(cmd.cmd.tid_info);
2088 		dest = &cmd.cmd.tid_info;
2089 		src = &ucmd->cmd.tid_info;
2090 		break;
2091 
2092 	case QIB_CMD_SET_PART_KEY:
2093 		copy = sizeof(cmd.cmd.part_key);
2094 		dest = &cmd.cmd.part_key;
2095 		src = &ucmd->cmd.part_key;
2096 		break;
2097 
2098 	case QIB_CMD_DISARM_BUFS:
2099 	case QIB_CMD_PIOAVAILUPD: /* force an update of PIOAvail reg */
2100 		copy = 0;
2101 		src = NULL;
2102 		dest = NULL;
2103 		break;
2104 
2105 	case QIB_CMD_POLL_TYPE:
2106 		copy = sizeof(cmd.cmd.poll_type);
2107 		dest = &cmd.cmd.poll_type;
2108 		src = &ucmd->cmd.poll_type;
2109 		break;
2110 
2111 	case QIB_CMD_ARMLAUNCH_CTRL:
2112 		copy = sizeof(cmd.cmd.armlaunch_ctrl);
2113 		dest = &cmd.cmd.armlaunch_ctrl;
2114 		src = &ucmd->cmd.armlaunch_ctrl;
2115 		break;
2116 
2117 	case QIB_CMD_SDMA_INFLIGHT:
2118 		copy = sizeof(cmd.cmd.sdma_inflight);
2119 		dest = &cmd.cmd.sdma_inflight;
2120 		src = &ucmd->cmd.sdma_inflight;
2121 		break;
2122 
2123 	case QIB_CMD_SDMA_COMPLETE:
2124 		copy = sizeof(cmd.cmd.sdma_complete);
2125 		dest = &cmd.cmd.sdma_complete;
2126 		src = &ucmd->cmd.sdma_complete;
2127 		break;
2128 
2129 	case QIB_CMD_ACK_EVENT:
2130 		copy = sizeof(cmd.cmd.event_mask);
2131 		dest = &cmd.cmd.event_mask;
2132 		src = &ucmd->cmd.event_mask;
2133 		break;
2134 
2135 	default:
2136 		ret = -EINVAL;
2137 		goto bail;
2138 	}
2139 
2140 	if (copy) {
2141 		if ((count - consumed) < copy) {
2142 			ret = -EINVAL;
2143 			goto bail;
2144 		}
2145 		if (copy_from_user(dest, src, copy)) {
2146 			ret = -EFAULT;
2147 			goto bail;
2148 		}
2149 		consumed += copy;
2150 	}
2151 
2152 	rcd = ctxt_fp(fp);
2153 	if (!rcd && cmd.type != QIB_CMD_ASSIGN_CTXT) {
2154 		ret = -EINVAL;
2155 		goto bail;
2156 	}
2157 
2158 	switch (cmd.type) {
2159 	case QIB_CMD_ASSIGN_CTXT:
2160 		if (rcd) {
2161 			ret = -EINVAL;
2162 			goto bail;
2163 		}
2164 
2165 		ret = qib_assign_ctxt(fp, &cmd.cmd.user_info);
2166 		if (ret)
2167 			goto bail;
2168 		break;
2169 
2170 	case QIB_CMD_USER_INIT:
2171 		ret = qib_do_user_init(fp, &cmd.cmd.user_info);
2172 		if (ret)
2173 			goto bail;
2174 		ret = qib_get_base_info(fp, u64_to_user_ptr(
2175 					  cmd.cmd.user_info.spu_base_info),
2176 					cmd.cmd.user_info.spu_base_info_size);
2177 		break;
2178 
2179 	case QIB_CMD_RECV_CTRL:
2180 		ret = qib_manage_rcvq(rcd, subctxt_fp(fp), cmd.cmd.recv_ctrl);
2181 		break;
2182 
2183 	case QIB_CMD_CTXT_INFO:
2184 		ret = qib_ctxt_info(fp, (struct qib_ctxt_info __user *)
2185 				    (unsigned long) cmd.cmd.ctxt_info);
2186 		break;
2187 
2188 	case QIB_CMD_TID_UPDATE:
2189 		ret = qib_tid_update(rcd, fp, &cmd.cmd.tid_info);
2190 		break;
2191 
2192 	case QIB_CMD_TID_FREE:
2193 		ret = qib_tid_free(rcd, subctxt_fp(fp), &cmd.cmd.tid_info);
2194 		break;
2195 
2196 	case QIB_CMD_SET_PART_KEY:
2197 		ret = qib_set_part_key(rcd, cmd.cmd.part_key);
2198 		break;
2199 
2200 	case QIB_CMD_DISARM_BUFS:
2201 		(void)qib_disarm_piobufs_ifneeded(rcd);
2202 		ret = disarm_req_delay(rcd);
2203 		break;
2204 
2205 	case QIB_CMD_PIOAVAILUPD:
2206 		qib_force_pio_avail_update(rcd->dd);
2207 		break;
2208 
2209 	case QIB_CMD_POLL_TYPE:
2210 		rcd->poll_type = cmd.cmd.poll_type;
2211 		break;
2212 
2213 	case QIB_CMD_ARMLAUNCH_CTRL:
2214 		rcd->dd->f_set_armlaunch(rcd->dd, cmd.cmd.armlaunch_ctrl);
2215 		break;
2216 
2217 	case QIB_CMD_SDMA_INFLIGHT:
2218 		ret = qib_sdma_get_inflight(user_sdma_queue_fp(fp),
2219 					    (u32 __user *) (unsigned long)
2220 					    cmd.cmd.sdma_inflight);
2221 		break;
2222 
2223 	case QIB_CMD_SDMA_COMPLETE:
2224 		ret = qib_sdma_get_complete(rcd->ppd,
2225 					    user_sdma_queue_fp(fp),
2226 					    (u32 __user *) (unsigned long)
2227 					    cmd.cmd.sdma_complete);
2228 		break;
2229 
2230 	case QIB_CMD_ACK_EVENT:
2231 		ret = qib_user_event_ack(rcd, subctxt_fp(fp),
2232 					 cmd.cmd.event_mask);
2233 		break;
2234 	}
2235 
2236 	if (ret >= 0)
2237 		ret = consumed;
2238 
2239 bail:
2240 	return ret;
2241 }
2242 
2243 static ssize_t qib_write_iter(struct kiocb *iocb, struct iov_iter *from)
2244 {
2245 	struct qib_filedata *fp = iocb->ki_filp->private_data;
2246 	struct qib_ctxtdata *rcd = ctxt_fp(iocb->ki_filp);
2247 	struct qib_user_sdma_queue *pq = fp->pq;
2248 
2249 	if (!iter_is_iovec(from) || !from->nr_segs || !pq)
2250 		return -EINVAL;
2251 
2252 	return qib_user_sdma_writev(rcd, pq, from->iov, from->nr_segs);
2253 }
2254 
2255 static struct class *qib_class;
2256 static dev_t qib_dev;
2257 
2258 int qib_cdev_init(int minor, const char *name,
2259 		  const struct file_operations *fops,
2260 		  struct cdev **cdevp, struct device **devp)
2261 {
2262 	const dev_t dev = MKDEV(MAJOR(qib_dev), minor);
2263 	struct cdev *cdev;
2264 	struct device *device = NULL;
2265 	int ret;
2266 
2267 	cdev = cdev_alloc();
2268 	if (!cdev) {
2269 		pr_err("Could not allocate cdev for minor %d, %s\n",
2270 		       minor, name);
2271 		ret = -ENOMEM;
2272 		goto done;
2273 	}
2274 
2275 	cdev->owner = THIS_MODULE;
2276 	cdev->ops = fops;
2277 	kobject_set_name(&cdev->kobj, name);
2278 
2279 	ret = cdev_add(cdev, dev, 1);
2280 	if (ret < 0) {
2281 		pr_err("Could not add cdev for minor %d, %s (err %d)\n",
2282 		       minor, name, -ret);
2283 		goto err_cdev;
2284 	}
2285 
2286 	device = device_create(qib_class, NULL, dev, NULL, "%s", name);
2287 	if (!IS_ERR(device))
2288 		goto done;
2289 	ret = PTR_ERR(device);
2290 	device = NULL;
2291 	pr_err("Could not create device for minor %d, %s (err %d)\n",
2292 	       minor, name, -ret);
2293 err_cdev:
2294 	cdev_del(cdev);
2295 	cdev = NULL;
2296 done:
2297 	*cdevp = cdev;
2298 	*devp = device;
2299 	return ret;
2300 }
2301 
2302 void qib_cdev_cleanup(struct cdev **cdevp, struct device **devp)
2303 {
2304 	struct device *device = *devp;
2305 
2306 	if (device) {
2307 		device_unregister(device);
2308 		*devp = NULL;
2309 	}
2310 
2311 	if (*cdevp) {
2312 		cdev_del(*cdevp);
2313 		*cdevp = NULL;
2314 	}
2315 }
2316 
2317 static struct cdev *wildcard_cdev;
2318 static struct device *wildcard_device;
2319 
2320 int __init qib_dev_init(void)
2321 {
2322 	int ret;
2323 
2324 	ret = alloc_chrdev_region(&qib_dev, 0, QIB_NMINORS, QIB_DRV_NAME);
2325 	if (ret < 0) {
2326 		pr_err("Could not allocate chrdev region (err %d)\n", -ret);
2327 		goto done;
2328 	}
2329 
2330 	qib_class = class_create(THIS_MODULE, "ipath");
2331 	if (IS_ERR(qib_class)) {
2332 		ret = PTR_ERR(qib_class);
2333 		pr_err("Could not create device class (err %d)\n", -ret);
2334 		unregister_chrdev_region(qib_dev, QIB_NMINORS);
2335 	}
2336 
2337 done:
2338 	return ret;
2339 }
2340 
2341 void qib_dev_cleanup(void)
2342 {
2343 	if (qib_class) {
2344 		class_destroy(qib_class);
2345 		qib_class = NULL;
2346 	}
2347 
2348 	unregister_chrdev_region(qib_dev, QIB_NMINORS);
2349 }
2350 
2351 static atomic_t user_count = ATOMIC_INIT(0);
2352 
2353 static void qib_user_remove(struct qib_devdata *dd)
2354 {
2355 	if (atomic_dec_return(&user_count) == 0)
2356 		qib_cdev_cleanup(&wildcard_cdev, &wildcard_device);
2357 
2358 	qib_cdev_cleanup(&dd->user_cdev, &dd->user_device);
2359 }
2360 
2361 static int qib_user_add(struct qib_devdata *dd)
2362 {
2363 	char name[10];
2364 	int ret;
2365 
2366 	if (atomic_inc_return(&user_count) == 1) {
2367 		ret = qib_cdev_init(0, "ipath", &qib_file_ops,
2368 				    &wildcard_cdev, &wildcard_device);
2369 		if (ret)
2370 			goto done;
2371 	}
2372 
2373 	snprintf(name, sizeof(name), "ipath%d", dd->unit);
2374 	ret = qib_cdev_init(dd->unit + 1, name, &qib_file_ops,
2375 			    &dd->user_cdev, &dd->user_device);
2376 	if (ret)
2377 		qib_user_remove(dd);
2378 done:
2379 	return ret;
2380 }
2381 
2382 /*
2383  * Create per-unit files in /dev
2384  */
2385 int qib_device_create(struct qib_devdata *dd)
2386 {
2387 	int r, ret;
2388 
2389 	r = qib_user_add(dd);
2390 	ret = qib_diag_add(dd);
2391 	if (r && !ret)
2392 		ret = r;
2393 	return ret;
2394 }
2395 
2396 /*
2397  * Remove per-unit files in /dev
2398  * void, core kernel returns no errors for this stuff
2399  */
2400 void qib_device_remove(struct qib_devdata *dd)
2401 {
2402 	qib_user_remove(dd);
2403 	qib_diag_remove(dd);
2404 }
2405