1 /*
2  * Copyright(c) 2015, 2016 Intel Corporation.
3  *
4  * This file is provided under a dual BSD/GPLv2 license.  When using or
5  * redistributing this file, you may do so under either license.
6  *
7  * GPL LICENSE SUMMARY
8  *
9  * This program is free software; you can redistribute it and/or modify
10  * it under the terms of version 2 of the GNU General Public License as
11  * published by the Free Software Foundation.
12  *
13  * This program is distributed in the hope that it will be useful, but
14  * WITHOUT ANY WARRANTY; without even the implied warranty of
15  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
16  * General Public License for more details.
17  *
18  * BSD LICENSE
19  *
20  * Redistribution and use in source and binary forms, with or without
21  * modification, are permitted provided that the following conditions
22  * are met:
23  *
24  *  - Redistributions of source code must retain the above copyright
25  *    notice, this list of conditions and the following disclaimer.
26  *  - Redistributions in binary form must reproduce the above copyright
27  *    notice, this list of conditions and the following disclaimer in
28  *    the documentation and/or other materials provided with the
29  *    distribution.
30  *  - Neither the name of Intel Corporation nor the names of its
31  *    contributors may be used to endorse or promote products derived
32  *    from this software without specific prior written permission.
33  *
34  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37  * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38  * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39  * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40  * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41  * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42  * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43  * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44  * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
45  *
46  */
47 #include <linux/poll.h>
48 #include <linux/cdev.h>
49 #include <linux/vmalloc.h>
50 #include <linux/io.h>
51 
52 #include <rdma/ib.h>
53 
54 #include "hfi.h"
55 #include "pio.h"
56 #include "device.h"
57 #include "common.h"
58 #include "trace.h"
59 #include "user_sdma.h"
60 #include "user_exp_rcv.h"
61 #include "eprom.h"
62 #include "aspm.h"
63 #include "mmu_rb.h"
64 
65 #undef pr_fmt
66 #define pr_fmt(fmt) DRIVER_NAME ": " fmt
67 
68 #define SEND_CTXT_HALT_TIMEOUT 1000 /* msecs */
69 
70 /*
71  * File operation functions
72  */
73 static int hfi1_file_open(struct inode *, struct file *);
74 static int hfi1_file_close(struct inode *, struct file *);
75 static ssize_t hfi1_write_iter(struct kiocb *, struct iov_iter *);
76 static unsigned int hfi1_poll(struct file *, struct poll_table_struct *);
77 static int hfi1_file_mmap(struct file *, struct vm_area_struct *);
78 
79 static u64 kvirt_to_phys(void *);
80 static int assign_ctxt(struct file *, struct hfi1_user_info *);
81 static int init_subctxts(struct hfi1_ctxtdata *, const struct hfi1_user_info *);
82 static int user_init(struct file *);
83 static int get_ctxt_info(struct file *, void __user *, __u32);
84 static int get_base_info(struct file *, void __user *, __u32);
85 static int setup_ctxt(struct file *);
86 static int setup_subctxt(struct hfi1_ctxtdata *);
87 static int get_user_context(struct file *, struct hfi1_user_info *, int);
88 static int find_shared_ctxt(struct file *, const struct hfi1_user_info *);
89 static int allocate_ctxt(struct file *, struct hfi1_devdata *,
90 			 struct hfi1_user_info *);
91 static unsigned int poll_urgent(struct file *, struct poll_table_struct *);
92 static unsigned int poll_next(struct file *, struct poll_table_struct *);
93 static int user_event_ack(struct hfi1_ctxtdata *, int, unsigned long);
94 static int set_ctxt_pkey(struct hfi1_ctxtdata *, unsigned, u16);
95 static int manage_rcvq(struct hfi1_ctxtdata *, unsigned, int);
96 static int vma_fault(struct vm_area_struct *, struct vm_fault *);
97 static long hfi1_file_ioctl(struct file *fp, unsigned int cmd,
98 			    unsigned long arg);
99 
100 static const struct file_operations hfi1_file_ops = {
101 	.owner = THIS_MODULE,
102 	.write_iter = hfi1_write_iter,
103 	.open = hfi1_file_open,
104 	.release = hfi1_file_close,
105 	.unlocked_ioctl = hfi1_file_ioctl,
106 	.poll = hfi1_poll,
107 	.mmap = hfi1_file_mmap,
108 	.llseek = noop_llseek,
109 };
110 
111 static struct vm_operations_struct vm_ops = {
112 	.fault = vma_fault,
113 };
114 
115 /*
116  * Types of memories mapped into user processes' space
117  */
118 enum mmap_types {
119 	PIO_BUFS = 1,
120 	PIO_BUFS_SOP,
121 	PIO_CRED,
122 	RCV_HDRQ,
123 	RCV_EGRBUF,
124 	UREGS,
125 	EVENTS,
126 	STATUS,
127 	RTAIL,
128 	SUBCTXT_UREGS,
129 	SUBCTXT_RCV_HDRQ,
130 	SUBCTXT_EGRBUF,
131 	SDMA_COMP
132 };
133 
134 /*
135  * Masks and offsets defining the mmap tokens
136  */
137 #define HFI1_MMAP_OFFSET_MASK   0xfffULL
138 #define HFI1_MMAP_OFFSET_SHIFT  0
139 #define HFI1_MMAP_SUBCTXT_MASK  0xfULL
140 #define HFI1_MMAP_SUBCTXT_SHIFT 12
141 #define HFI1_MMAP_CTXT_MASK     0xffULL
142 #define HFI1_MMAP_CTXT_SHIFT    16
143 #define HFI1_MMAP_TYPE_MASK     0xfULL
144 #define HFI1_MMAP_TYPE_SHIFT    24
145 #define HFI1_MMAP_MAGIC_MASK    0xffffffffULL
146 #define HFI1_MMAP_MAGIC_SHIFT   32
147 
148 #define HFI1_MMAP_MAGIC         0xdabbad00
149 
150 #define HFI1_MMAP_TOKEN_SET(field, val)	\
151 	(((val) & HFI1_MMAP_##field##_MASK) << HFI1_MMAP_##field##_SHIFT)
152 #define HFI1_MMAP_TOKEN_GET(field, token) \
153 	(((token) >> HFI1_MMAP_##field##_SHIFT) & HFI1_MMAP_##field##_MASK)
154 #define HFI1_MMAP_TOKEN(type, ctxt, subctxt, addr)   \
155 	(HFI1_MMAP_TOKEN_SET(MAGIC, HFI1_MMAP_MAGIC) | \
156 	HFI1_MMAP_TOKEN_SET(TYPE, type) | \
157 	HFI1_MMAP_TOKEN_SET(CTXT, ctxt) | \
158 	HFI1_MMAP_TOKEN_SET(SUBCTXT, subctxt) | \
159 	HFI1_MMAP_TOKEN_SET(OFFSET, (offset_in_page(addr))))
160 
161 #define dbg(fmt, ...)				\
162 	pr_info(fmt, ##__VA_ARGS__)
163 
164 static inline int is_valid_mmap(u64 token)
165 {
166 	return (HFI1_MMAP_TOKEN_GET(MAGIC, token) == HFI1_MMAP_MAGIC);
167 }
168 
169 static int hfi1_file_open(struct inode *inode, struct file *fp)
170 {
171 	struct hfi1_filedata *fd;
172 	struct hfi1_devdata *dd = container_of(inode->i_cdev,
173 					       struct hfi1_devdata,
174 					       user_cdev);
175 
176 	/* Just take a ref now. Not all opens result in a context assign */
177 	kobject_get(&dd->kobj);
178 
179 	/* The real work is performed later in assign_ctxt() */
180 
181 	fd = kzalloc(sizeof(*fd), GFP_KERNEL);
182 
183 	if (fd) {
184 		fd->rec_cpu_num = -1; /* no cpu affinity by default */
185 		fd->mm = current->mm;
186 	}
187 
188 	fp->private_data = fd;
189 
190 	return fd ? 0 : -ENOMEM;
191 }
192 
193 static long hfi1_file_ioctl(struct file *fp, unsigned int cmd,
194 			    unsigned long arg)
195 {
196 	struct hfi1_filedata *fd = fp->private_data;
197 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
198 	struct hfi1_user_info uinfo;
199 	struct hfi1_tid_info tinfo;
200 	int ret = 0;
201 	unsigned long addr;
202 	int uval = 0;
203 	unsigned long ul_uval = 0;
204 	u16 uval16 = 0;
205 
206 	hfi1_cdbg(IOCTL, "IOCTL recv: 0x%x", cmd);
207 	if (cmd != HFI1_IOCTL_ASSIGN_CTXT &&
208 	    cmd != HFI1_IOCTL_GET_VERS &&
209 	    !uctxt)
210 		return -EINVAL;
211 
212 	switch (cmd) {
213 	case HFI1_IOCTL_ASSIGN_CTXT:
214 		if (uctxt)
215 			return -EINVAL;
216 
217 		if (copy_from_user(&uinfo,
218 				   (struct hfi1_user_info __user *)arg,
219 				   sizeof(uinfo)))
220 			return -EFAULT;
221 
222 		ret = assign_ctxt(fp, &uinfo);
223 		if (ret < 0)
224 			return ret;
225 		setup_ctxt(fp);
226 		if (ret)
227 			return ret;
228 		ret = user_init(fp);
229 		break;
230 	case HFI1_IOCTL_CTXT_INFO:
231 		ret = get_ctxt_info(fp, (void __user *)(unsigned long)arg,
232 				    sizeof(struct hfi1_ctxt_info));
233 		break;
234 	case HFI1_IOCTL_USER_INFO:
235 		ret = get_base_info(fp, (void __user *)(unsigned long)arg,
236 				    sizeof(struct hfi1_base_info));
237 		break;
238 	case HFI1_IOCTL_CREDIT_UPD:
239 		if (uctxt)
240 			sc_return_credits(uctxt->sc);
241 		break;
242 
243 	case HFI1_IOCTL_TID_UPDATE:
244 		if (copy_from_user(&tinfo,
245 				   (struct hfi11_tid_info __user *)arg,
246 				   sizeof(tinfo)))
247 			return -EFAULT;
248 
249 		ret = hfi1_user_exp_rcv_setup(fp, &tinfo);
250 		if (!ret) {
251 			/*
252 			 * Copy the number of tidlist entries we used
253 			 * and the length of the buffer we registered.
254 			 * These fields are adjacent in the structure so
255 			 * we can copy them at the same time.
256 			 */
257 			addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
258 			if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
259 					 sizeof(tinfo.tidcnt) +
260 					 sizeof(tinfo.length)))
261 				ret = -EFAULT;
262 		}
263 		break;
264 
265 	case HFI1_IOCTL_TID_FREE:
266 		if (copy_from_user(&tinfo,
267 				   (struct hfi11_tid_info __user *)arg,
268 				   sizeof(tinfo)))
269 			return -EFAULT;
270 
271 		ret = hfi1_user_exp_rcv_clear(fp, &tinfo);
272 		if (ret)
273 			break;
274 		addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
275 		if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
276 				 sizeof(tinfo.tidcnt)))
277 			ret = -EFAULT;
278 		break;
279 
280 	case HFI1_IOCTL_TID_INVAL_READ:
281 		if (copy_from_user(&tinfo,
282 				   (struct hfi11_tid_info __user *)arg,
283 				   sizeof(tinfo)))
284 			return -EFAULT;
285 
286 		ret = hfi1_user_exp_rcv_invalid(fp, &tinfo);
287 		if (ret)
288 			break;
289 		addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
290 		if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
291 				 sizeof(tinfo.tidcnt)))
292 			ret = -EFAULT;
293 		break;
294 
295 	case HFI1_IOCTL_RECV_CTRL:
296 		ret = get_user(uval, (int __user *)arg);
297 		if (ret != 0)
298 			return -EFAULT;
299 		ret = manage_rcvq(uctxt, fd->subctxt, uval);
300 		break;
301 
302 	case HFI1_IOCTL_POLL_TYPE:
303 		ret = get_user(uval, (int __user *)arg);
304 		if (ret != 0)
305 			return -EFAULT;
306 		uctxt->poll_type = (typeof(uctxt->poll_type))uval;
307 		break;
308 
309 	case HFI1_IOCTL_ACK_EVENT:
310 		ret = get_user(ul_uval, (unsigned long __user *)arg);
311 		if (ret != 0)
312 			return -EFAULT;
313 		ret = user_event_ack(uctxt, fd->subctxt, ul_uval);
314 		break;
315 
316 	case HFI1_IOCTL_SET_PKEY:
317 		ret = get_user(uval16, (u16 __user *)arg);
318 		if (ret != 0)
319 			return -EFAULT;
320 		if (HFI1_CAP_IS_USET(PKEY_CHECK))
321 			ret = set_ctxt_pkey(uctxt, fd->subctxt, uval16);
322 		else
323 			return -EPERM;
324 		break;
325 
326 	case HFI1_IOCTL_CTXT_RESET: {
327 		struct send_context *sc;
328 		struct hfi1_devdata *dd;
329 
330 		if (!uctxt || !uctxt->dd || !uctxt->sc)
331 			return -EINVAL;
332 
333 		/*
334 		 * There is no protection here. User level has to
335 		 * guarantee that no one will be writing to the send
336 		 * context while it is being re-initialized.
337 		 * If user level breaks that guarantee, it will break
338 		 * it's own context and no one else's.
339 		 */
340 		dd = uctxt->dd;
341 		sc = uctxt->sc;
342 		/*
343 		 * Wait until the interrupt handler has marked the
344 		 * context as halted or frozen. Report error if we time
345 		 * out.
346 		 */
347 		wait_event_interruptible_timeout(
348 			sc->halt_wait, (sc->flags & SCF_HALTED),
349 			msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
350 		if (!(sc->flags & SCF_HALTED))
351 			return -ENOLCK;
352 
353 		/*
354 		 * If the send context was halted due to a Freeze,
355 		 * wait until the device has been "unfrozen" before
356 		 * resetting the context.
357 		 */
358 		if (sc->flags & SCF_FROZEN) {
359 			wait_event_interruptible_timeout(
360 				dd->event_queue,
361 				!(ACCESS_ONCE(dd->flags) & HFI1_FROZEN),
362 				msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
363 			if (dd->flags & HFI1_FROZEN)
364 				return -ENOLCK;
365 
366 			if (dd->flags & HFI1_FORCED_FREEZE)
367 				/*
368 				 * Don't allow context reset if we are into
369 				 * forced freeze
370 				 */
371 				return -ENODEV;
372 
373 			sc_disable(sc);
374 			ret = sc_enable(sc);
375 			hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_ENB,
376 				     uctxt->ctxt);
377 		} else {
378 			ret = sc_restart(sc);
379 		}
380 		if (!ret)
381 			sc_return_credits(sc);
382 		break;
383 	}
384 
385 	case HFI1_IOCTL_GET_VERS:
386 		uval = HFI1_USER_SWVERSION;
387 		if (put_user(uval, (int __user *)arg))
388 			return -EFAULT;
389 		break;
390 
391 	default:
392 		return -EINVAL;
393 	}
394 
395 	return ret;
396 }
397 
398 static ssize_t hfi1_write_iter(struct kiocb *kiocb, struct iov_iter *from)
399 {
400 	struct hfi1_filedata *fd = kiocb->ki_filp->private_data;
401 	struct hfi1_user_sdma_pkt_q *pq = fd->pq;
402 	struct hfi1_user_sdma_comp_q *cq = fd->cq;
403 	int done = 0, reqs = 0;
404 	unsigned long dim = from->nr_segs;
405 
406 	if (!cq || !pq)
407 		return -EIO;
408 
409 	if (!iter_is_iovec(from) || !dim)
410 		return -EINVAL;
411 
412 	hfi1_cdbg(SDMA, "SDMA request from %u:%u (%lu)",
413 		  fd->uctxt->ctxt, fd->subctxt, dim);
414 
415 	if (atomic_read(&pq->n_reqs) == pq->n_max_reqs)
416 		return -ENOSPC;
417 
418 	while (dim) {
419 		int ret;
420 		unsigned long count = 0;
421 
422 		ret = hfi1_user_sdma_process_request(
423 			kiocb->ki_filp,	(struct iovec *)(from->iov + done),
424 			dim, &count);
425 		if (ret) {
426 			reqs = ret;
427 			break;
428 		}
429 		dim -= count;
430 		done += count;
431 		reqs++;
432 	}
433 
434 	return reqs;
435 }
436 
437 static int hfi1_file_mmap(struct file *fp, struct vm_area_struct *vma)
438 {
439 	struct hfi1_filedata *fd = fp->private_data;
440 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
441 	struct hfi1_devdata *dd;
442 	unsigned long flags, pfn;
443 	u64 token = vma->vm_pgoff << PAGE_SHIFT,
444 		memaddr = 0;
445 	u8 subctxt, mapio = 0, vmf = 0, type;
446 	ssize_t memlen = 0;
447 	int ret = 0;
448 	u16 ctxt;
449 
450 	if (!is_valid_mmap(token) || !uctxt ||
451 	    !(vma->vm_flags & VM_SHARED)) {
452 		ret = -EINVAL;
453 		goto done;
454 	}
455 	dd = uctxt->dd;
456 	ctxt = HFI1_MMAP_TOKEN_GET(CTXT, token);
457 	subctxt = HFI1_MMAP_TOKEN_GET(SUBCTXT, token);
458 	type = HFI1_MMAP_TOKEN_GET(TYPE, token);
459 	if (ctxt != uctxt->ctxt || subctxt != fd->subctxt) {
460 		ret = -EINVAL;
461 		goto done;
462 	}
463 
464 	flags = vma->vm_flags;
465 
466 	switch (type) {
467 	case PIO_BUFS:
468 	case PIO_BUFS_SOP:
469 		memaddr = ((dd->physaddr + TXE_PIO_SEND) +
470 				/* chip pio base */
471 			   (uctxt->sc->hw_context * BIT(16))) +
472 				/* 64K PIO space / ctxt */
473 			(type == PIO_BUFS_SOP ?
474 				(TXE_PIO_SIZE / 2) : 0); /* sop? */
475 		/*
476 		 * Map only the amount allocated to the context, not the
477 		 * entire available context's PIO space.
478 		 */
479 		memlen = PAGE_ALIGN(uctxt->sc->credits * PIO_BLOCK_SIZE);
480 		flags &= ~VM_MAYREAD;
481 		flags |= VM_DONTCOPY | VM_DONTEXPAND;
482 		vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
483 		mapio = 1;
484 		break;
485 	case PIO_CRED:
486 		if (flags & VM_WRITE) {
487 			ret = -EPERM;
488 			goto done;
489 		}
490 		/*
491 		 * The credit return location for this context could be on the
492 		 * second or third page allocated for credit returns (if number
493 		 * of enabled contexts > 64 and 128 respectively).
494 		 */
495 		memaddr = dd->cr_base[uctxt->numa_id].pa +
496 			(((u64)uctxt->sc->hw_free -
497 			  (u64)dd->cr_base[uctxt->numa_id].va) & PAGE_MASK);
498 		memlen = PAGE_SIZE;
499 		flags &= ~VM_MAYWRITE;
500 		flags |= VM_DONTCOPY | VM_DONTEXPAND;
501 		/*
502 		 * The driver has already allocated memory for credit
503 		 * returns and programmed it into the chip. Has that
504 		 * memory been flagged as non-cached?
505 		 */
506 		/* vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); */
507 		mapio = 1;
508 		break;
509 	case RCV_HDRQ:
510 		memaddr = uctxt->rcvhdrq_phys;
511 		memlen = uctxt->rcvhdrq_size;
512 		break;
513 	case RCV_EGRBUF: {
514 		unsigned long addr;
515 		int i;
516 		/*
517 		 * The RcvEgr buffer need to be handled differently
518 		 * as multiple non-contiguous pages need to be mapped
519 		 * into the user process.
520 		 */
521 		memlen = uctxt->egrbufs.size;
522 		if ((vma->vm_end - vma->vm_start) != memlen) {
523 			dd_dev_err(dd, "Eager buffer map size invalid (%lu != %lu)\n",
524 				   (vma->vm_end - vma->vm_start), memlen);
525 			ret = -EINVAL;
526 			goto done;
527 		}
528 		if (vma->vm_flags & VM_WRITE) {
529 			ret = -EPERM;
530 			goto done;
531 		}
532 		vma->vm_flags &= ~VM_MAYWRITE;
533 		addr = vma->vm_start;
534 		for (i = 0 ; i < uctxt->egrbufs.numbufs; i++) {
535 			ret = remap_pfn_range(
536 				vma, addr,
537 				uctxt->egrbufs.buffers[i].phys >> PAGE_SHIFT,
538 				uctxt->egrbufs.buffers[i].len,
539 				vma->vm_page_prot);
540 			if (ret < 0)
541 				goto done;
542 			addr += uctxt->egrbufs.buffers[i].len;
543 		}
544 		ret = 0;
545 		goto done;
546 	}
547 	case UREGS:
548 		/*
549 		 * Map only the page that contains this context's user
550 		 * registers.
551 		 */
552 		memaddr = (unsigned long)
553 			(dd->physaddr + RXE_PER_CONTEXT_USER)
554 			+ (uctxt->ctxt * RXE_PER_CONTEXT_SIZE);
555 		/*
556 		 * TidFlow table is on the same page as the rest of the
557 		 * user registers.
558 		 */
559 		memlen = PAGE_SIZE;
560 		flags |= VM_DONTCOPY | VM_DONTEXPAND;
561 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
562 		mapio = 1;
563 		break;
564 	case EVENTS:
565 		/*
566 		 * Use the page where this context's flags are. User level
567 		 * knows where it's own bitmap is within the page.
568 		 */
569 		memaddr = (unsigned long)(dd->events +
570 					  ((uctxt->ctxt - dd->first_user_ctxt) *
571 					   HFI1_MAX_SHARED_CTXTS)) & PAGE_MASK;
572 		memlen = PAGE_SIZE;
573 		/*
574 		 * v3.7 removes VM_RESERVED but the effect is kept by
575 		 * using VM_IO.
576 		 */
577 		flags |= VM_IO | VM_DONTEXPAND;
578 		vmf = 1;
579 		break;
580 	case STATUS:
581 		memaddr = kvirt_to_phys((void *)dd->status);
582 		memlen = PAGE_SIZE;
583 		flags |= VM_IO | VM_DONTEXPAND;
584 		break;
585 	case RTAIL:
586 		if (!HFI1_CAP_IS_USET(DMA_RTAIL)) {
587 			/*
588 			 * If the memory allocation failed, the context alloc
589 			 * also would have failed, so we would never get here
590 			 */
591 			ret = -EINVAL;
592 			goto done;
593 		}
594 		if (flags & VM_WRITE) {
595 			ret = -EPERM;
596 			goto done;
597 		}
598 		memaddr = uctxt->rcvhdrqtailaddr_phys;
599 		memlen = PAGE_SIZE;
600 		flags &= ~VM_MAYWRITE;
601 		break;
602 	case SUBCTXT_UREGS:
603 		memaddr = (u64)uctxt->subctxt_uregbase;
604 		memlen = PAGE_SIZE;
605 		flags |= VM_IO | VM_DONTEXPAND;
606 		vmf = 1;
607 		break;
608 	case SUBCTXT_RCV_HDRQ:
609 		memaddr = (u64)uctxt->subctxt_rcvhdr_base;
610 		memlen = uctxt->rcvhdrq_size * uctxt->subctxt_cnt;
611 		flags |= VM_IO | VM_DONTEXPAND;
612 		vmf = 1;
613 		break;
614 	case SUBCTXT_EGRBUF:
615 		memaddr = (u64)uctxt->subctxt_rcvegrbuf;
616 		memlen = uctxt->egrbufs.size * uctxt->subctxt_cnt;
617 		flags |= VM_IO | VM_DONTEXPAND;
618 		flags &= ~VM_MAYWRITE;
619 		vmf = 1;
620 		break;
621 	case SDMA_COMP: {
622 		struct hfi1_user_sdma_comp_q *cq = fd->cq;
623 
624 		if (!cq) {
625 			ret = -EFAULT;
626 			goto done;
627 		}
628 		memaddr = (u64)cq->comps;
629 		memlen = PAGE_ALIGN(sizeof(*cq->comps) * cq->nentries);
630 		flags |= VM_IO | VM_DONTEXPAND;
631 		vmf = 1;
632 		break;
633 	}
634 	default:
635 		ret = -EINVAL;
636 		break;
637 	}
638 
639 	if ((vma->vm_end - vma->vm_start) != memlen) {
640 		hfi1_cdbg(PROC, "%u:%u Memory size mismatch %lu:%lu",
641 			  uctxt->ctxt, fd->subctxt,
642 			  (vma->vm_end - vma->vm_start), memlen);
643 		ret = -EINVAL;
644 		goto done;
645 	}
646 
647 	vma->vm_flags = flags;
648 	hfi1_cdbg(PROC,
649 		  "%u:%u type:%u io/vf:%d/%d, addr:0x%llx, len:%lu(%lu), flags:0x%lx\n",
650 		    ctxt, subctxt, type, mapio, vmf, memaddr, memlen,
651 		    vma->vm_end - vma->vm_start, vma->vm_flags);
652 	pfn = (unsigned long)(memaddr >> PAGE_SHIFT);
653 	if (vmf) {
654 		vma->vm_pgoff = pfn;
655 		vma->vm_ops = &vm_ops;
656 		ret = 0;
657 	} else if (mapio) {
658 		ret = io_remap_pfn_range(vma, vma->vm_start, pfn, memlen,
659 					 vma->vm_page_prot);
660 	} else {
661 		ret = remap_pfn_range(vma, vma->vm_start, pfn, memlen,
662 				      vma->vm_page_prot);
663 	}
664 done:
665 	return ret;
666 }
667 
668 /*
669  * Local (non-chip) user memory is not mapped right away but as it is
670  * accessed by the user-level code.
671  */
672 static int vma_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
673 {
674 	struct page *page;
675 
676 	page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
677 	if (!page)
678 		return VM_FAULT_SIGBUS;
679 
680 	get_page(page);
681 	vmf->page = page;
682 
683 	return 0;
684 }
685 
686 static unsigned int hfi1_poll(struct file *fp, struct poll_table_struct *pt)
687 {
688 	struct hfi1_ctxtdata *uctxt;
689 	unsigned pollflag;
690 
691 	uctxt = ((struct hfi1_filedata *)fp->private_data)->uctxt;
692 	if (!uctxt)
693 		pollflag = POLLERR;
694 	else if (uctxt->poll_type == HFI1_POLL_TYPE_URGENT)
695 		pollflag = poll_urgent(fp, pt);
696 	else  if (uctxt->poll_type == HFI1_POLL_TYPE_ANYRCV)
697 		pollflag = poll_next(fp, pt);
698 	else /* invalid */
699 		pollflag = POLLERR;
700 
701 	return pollflag;
702 }
703 
704 static int hfi1_file_close(struct inode *inode, struct file *fp)
705 {
706 	struct hfi1_filedata *fdata = fp->private_data;
707 	struct hfi1_ctxtdata *uctxt = fdata->uctxt;
708 	struct hfi1_devdata *dd = container_of(inode->i_cdev,
709 					       struct hfi1_devdata,
710 					       user_cdev);
711 	unsigned long flags, *ev;
712 
713 	fp->private_data = NULL;
714 
715 	if (!uctxt)
716 		goto done;
717 
718 	hfi1_cdbg(PROC, "freeing ctxt %u:%u", uctxt->ctxt, fdata->subctxt);
719 	mutex_lock(&hfi1_mutex);
720 
721 	flush_wc();
722 	/* drain user sdma queue */
723 	hfi1_user_sdma_free_queues(fdata);
724 
725 	/* release the cpu */
726 	hfi1_put_proc_affinity(fdata->rec_cpu_num);
727 
728 	/*
729 	 * Clear any left over, unhandled events so the next process that
730 	 * gets this context doesn't get confused.
731 	 */
732 	ev = dd->events + ((uctxt->ctxt - dd->first_user_ctxt) *
733 			   HFI1_MAX_SHARED_CTXTS) + fdata->subctxt;
734 	*ev = 0;
735 
736 	if (--uctxt->cnt) {
737 		uctxt->active_slaves &= ~(1 << fdata->subctxt);
738 		mutex_unlock(&hfi1_mutex);
739 		goto done;
740 	}
741 
742 	spin_lock_irqsave(&dd->uctxt_lock, flags);
743 	/*
744 	 * Disable receive context and interrupt available, reset all
745 	 * RcvCtxtCtrl bits to default values.
746 	 */
747 	hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_DIS |
748 		     HFI1_RCVCTRL_TIDFLOW_DIS |
749 		     HFI1_RCVCTRL_INTRAVAIL_DIS |
750 		     HFI1_RCVCTRL_TAILUPD_DIS |
751 		     HFI1_RCVCTRL_ONE_PKT_EGR_DIS |
752 		     HFI1_RCVCTRL_NO_RHQ_DROP_DIS |
753 		     HFI1_RCVCTRL_NO_EGR_DROP_DIS, uctxt->ctxt);
754 	/* Clear the context's J_KEY */
755 	hfi1_clear_ctxt_jkey(dd, uctxt->ctxt);
756 	/*
757 	 * Reset context integrity checks to default.
758 	 * (writes to CSRs probably belong in chip.c)
759 	 */
760 	write_kctxt_csr(dd, uctxt->sc->hw_context, SEND_CTXT_CHECK_ENABLE,
761 			hfi1_pkt_default_send_ctxt_mask(dd, uctxt->sc->type));
762 	sc_disable(uctxt->sc);
763 	spin_unlock_irqrestore(&dd->uctxt_lock, flags);
764 
765 	dd->rcd[uctxt->ctxt] = NULL;
766 
767 	hfi1_user_exp_rcv_free(fdata);
768 	hfi1_clear_ctxt_pkey(dd, uctxt->ctxt);
769 
770 	uctxt->rcvwait_to = 0;
771 	uctxt->piowait_to = 0;
772 	uctxt->rcvnowait = 0;
773 	uctxt->pionowait = 0;
774 	uctxt->event_flags = 0;
775 
776 	hfi1_stats.sps_ctxts--;
777 	if (++dd->freectxts == dd->num_user_contexts)
778 		aspm_enable_all(dd);
779 	mutex_unlock(&hfi1_mutex);
780 	hfi1_free_ctxtdata(dd, uctxt);
781 done:
782 	kobject_put(&dd->kobj);
783 	kfree(fdata);
784 	return 0;
785 }
786 
787 /*
788  * Convert kernel *virtual* addresses to physical addresses.
789  * This is used to vmalloc'ed addresses.
790  */
791 static u64 kvirt_to_phys(void *addr)
792 {
793 	struct page *page;
794 	u64 paddr = 0;
795 
796 	page = vmalloc_to_page(addr);
797 	if (page)
798 		paddr = page_to_pfn(page) << PAGE_SHIFT;
799 
800 	return paddr;
801 }
802 
803 static int assign_ctxt(struct file *fp, struct hfi1_user_info *uinfo)
804 {
805 	int i_minor, ret = 0;
806 	unsigned int swmajor, swminor;
807 
808 	swmajor = uinfo->userversion >> 16;
809 	if (swmajor != HFI1_USER_SWMAJOR) {
810 		ret = -ENODEV;
811 		goto done;
812 	}
813 
814 	swminor = uinfo->userversion & 0xffff;
815 
816 	mutex_lock(&hfi1_mutex);
817 	/* First, lets check if we need to setup a shared context? */
818 	if (uinfo->subctxt_cnt) {
819 		struct hfi1_filedata *fd = fp->private_data;
820 
821 		ret = find_shared_ctxt(fp, uinfo);
822 		if (ret < 0)
823 			goto done_unlock;
824 		if (ret) {
825 			fd->rec_cpu_num =
826 				hfi1_get_proc_affinity(fd->uctxt->numa_id);
827 		}
828 	}
829 
830 	/*
831 	 * We execute the following block if we couldn't find a
832 	 * shared context or if context sharing is not required.
833 	 */
834 	if (!ret) {
835 		i_minor = iminor(file_inode(fp)) - HFI1_USER_MINOR_BASE;
836 		ret = get_user_context(fp, uinfo, i_minor);
837 	}
838 done_unlock:
839 	mutex_unlock(&hfi1_mutex);
840 done:
841 	return ret;
842 }
843 
844 static int get_user_context(struct file *fp, struct hfi1_user_info *uinfo,
845 			    int devno)
846 {
847 	struct hfi1_devdata *dd = NULL;
848 	int devmax, npresent, nup;
849 
850 	devmax = hfi1_count_units(&npresent, &nup);
851 	if (!npresent)
852 		return -ENXIO;
853 
854 	if (!nup)
855 		return -ENETDOWN;
856 
857 	dd = hfi1_lookup(devno);
858 	if (!dd)
859 		return -ENODEV;
860 	else if (!dd->freectxts)
861 		return -EBUSY;
862 
863 	return allocate_ctxt(fp, dd, uinfo);
864 }
865 
866 static int find_shared_ctxt(struct file *fp,
867 			    const struct hfi1_user_info *uinfo)
868 {
869 	int devmax, ndev, i;
870 	int ret = 0;
871 	struct hfi1_filedata *fd = fp->private_data;
872 
873 	devmax = hfi1_count_units(NULL, NULL);
874 
875 	for (ndev = 0; ndev < devmax; ndev++) {
876 		struct hfi1_devdata *dd = hfi1_lookup(ndev);
877 
878 		if (!(dd && (dd->flags & HFI1_PRESENT) && dd->kregbase))
879 			continue;
880 		for (i = dd->first_user_ctxt; i < dd->num_rcv_contexts; i++) {
881 			struct hfi1_ctxtdata *uctxt = dd->rcd[i];
882 
883 			/* Skip ctxts which are not yet open */
884 			if (!uctxt || !uctxt->cnt)
885 				continue;
886 			/* Skip ctxt if it doesn't match the requested one */
887 			if (memcmp(uctxt->uuid, uinfo->uuid,
888 				   sizeof(uctxt->uuid)) ||
889 			    uctxt->jkey != generate_jkey(current_uid()) ||
890 			    uctxt->subctxt_id != uinfo->subctxt_id ||
891 			    uctxt->subctxt_cnt != uinfo->subctxt_cnt)
892 				continue;
893 
894 			/* Verify the sharing process matches the master */
895 			if (uctxt->userversion != uinfo->userversion ||
896 			    uctxt->cnt >= uctxt->subctxt_cnt) {
897 				ret = -EINVAL;
898 				goto done;
899 			}
900 			fd->uctxt = uctxt;
901 			fd->subctxt  = uctxt->cnt++;
902 			uctxt->active_slaves |= 1 << fd->subctxt;
903 			ret = 1;
904 			goto done;
905 		}
906 	}
907 
908 done:
909 	return ret;
910 }
911 
912 static int allocate_ctxt(struct file *fp, struct hfi1_devdata *dd,
913 			 struct hfi1_user_info *uinfo)
914 {
915 	struct hfi1_filedata *fd = fp->private_data;
916 	struct hfi1_ctxtdata *uctxt;
917 	unsigned ctxt;
918 	int ret, numa;
919 
920 	if (dd->flags & HFI1_FROZEN) {
921 		/*
922 		 * Pick an error that is unique from all other errors
923 		 * that are returned so the user process knows that
924 		 * it tried to allocate while the SPC was frozen.  It
925 		 * it should be able to retry with success in a short
926 		 * while.
927 		 */
928 		return -EIO;
929 	}
930 
931 	for (ctxt = dd->first_user_ctxt; ctxt < dd->num_rcv_contexts; ctxt++)
932 		if (!dd->rcd[ctxt])
933 			break;
934 
935 	if (ctxt == dd->num_rcv_contexts)
936 		return -EBUSY;
937 
938 	/*
939 	 * If we don't have a NUMA node requested, preference is towards
940 	 * device NUMA node.
941 	 */
942 	fd->rec_cpu_num = hfi1_get_proc_affinity(dd->node);
943 	if (fd->rec_cpu_num != -1)
944 		numa = cpu_to_node(fd->rec_cpu_num);
945 	else
946 		numa = numa_node_id();
947 	uctxt = hfi1_create_ctxtdata(dd->pport, ctxt, numa);
948 	if (!uctxt) {
949 		dd_dev_err(dd,
950 			   "Unable to allocate ctxtdata memory, failing open\n");
951 		return -ENOMEM;
952 	}
953 	hfi1_cdbg(PROC, "[%u:%u] pid %u assigned to CPU %d (NUMA %u)",
954 		  uctxt->ctxt, fd->subctxt, current->pid, fd->rec_cpu_num,
955 		  uctxt->numa_id);
956 
957 	/*
958 	 * Allocate and enable a PIO send context.
959 	 */
960 	uctxt->sc = sc_alloc(dd, SC_USER, uctxt->rcvhdrqentsize,
961 			     uctxt->dd->node);
962 	if (!uctxt->sc)
963 		return -ENOMEM;
964 
965 	hfi1_cdbg(PROC, "allocated send context %u(%u)\n", uctxt->sc->sw_index,
966 		  uctxt->sc->hw_context);
967 	ret = sc_enable(uctxt->sc);
968 	if (ret)
969 		return ret;
970 	/*
971 	 * Setup shared context resources if the user-level has requested
972 	 * shared contexts and this is the 'master' process.
973 	 * This has to be done here so the rest of the sub-contexts find the
974 	 * proper master.
975 	 */
976 	if (uinfo->subctxt_cnt && !fd->subctxt) {
977 		ret = init_subctxts(uctxt, uinfo);
978 		/*
979 		 * On error, we don't need to disable and de-allocate the
980 		 * send context because it will be done during file close
981 		 */
982 		if (ret)
983 			return ret;
984 	}
985 	uctxt->userversion = uinfo->userversion;
986 	uctxt->flags = hfi1_cap_mask; /* save current flag state */
987 	init_waitqueue_head(&uctxt->wait);
988 	strlcpy(uctxt->comm, current->comm, sizeof(uctxt->comm));
989 	memcpy(uctxt->uuid, uinfo->uuid, sizeof(uctxt->uuid));
990 	uctxt->jkey = generate_jkey(current_uid());
991 	INIT_LIST_HEAD(&uctxt->sdma_queues);
992 	spin_lock_init(&uctxt->sdma_qlock);
993 	hfi1_stats.sps_ctxts++;
994 	/*
995 	 * Disable ASPM when there are open user/PSM contexts to avoid
996 	 * issues with ASPM L1 exit latency
997 	 */
998 	if (dd->freectxts-- == dd->num_user_contexts)
999 		aspm_disable_all(dd);
1000 	fd->uctxt = uctxt;
1001 
1002 	return 0;
1003 }
1004 
1005 static int init_subctxts(struct hfi1_ctxtdata *uctxt,
1006 			 const struct hfi1_user_info *uinfo)
1007 {
1008 	unsigned num_subctxts;
1009 
1010 	num_subctxts = uinfo->subctxt_cnt;
1011 	if (num_subctxts > HFI1_MAX_SHARED_CTXTS)
1012 		return -EINVAL;
1013 
1014 	uctxt->subctxt_cnt = uinfo->subctxt_cnt;
1015 	uctxt->subctxt_id = uinfo->subctxt_id;
1016 	uctxt->active_slaves = 1;
1017 	uctxt->redirect_seq_cnt = 1;
1018 	set_bit(HFI1_CTXT_MASTER_UNINIT, &uctxt->event_flags);
1019 
1020 	return 0;
1021 }
1022 
1023 static int setup_subctxt(struct hfi1_ctxtdata *uctxt)
1024 {
1025 	int ret = 0;
1026 	unsigned num_subctxts = uctxt->subctxt_cnt;
1027 
1028 	uctxt->subctxt_uregbase = vmalloc_user(PAGE_SIZE);
1029 	if (!uctxt->subctxt_uregbase) {
1030 		ret = -ENOMEM;
1031 		goto bail;
1032 	}
1033 	/* We can take the size of the RcvHdr Queue from the master */
1034 	uctxt->subctxt_rcvhdr_base = vmalloc_user(uctxt->rcvhdrq_size *
1035 						  num_subctxts);
1036 	if (!uctxt->subctxt_rcvhdr_base) {
1037 		ret = -ENOMEM;
1038 		goto bail_ureg;
1039 	}
1040 
1041 	uctxt->subctxt_rcvegrbuf = vmalloc_user(uctxt->egrbufs.size *
1042 						num_subctxts);
1043 	if (!uctxt->subctxt_rcvegrbuf) {
1044 		ret = -ENOMEM;
1045 		goto bail_rhdr;
1046 	}
1047 	goto bail;
1048 bail_rhdr:
1049 	vfree(uctxt->subctxt_rcvhdr_base);
1050 bail_ureg:
1051 	vfree(uctxt->subctxt_uregbase);
1052 	uctxt->subctxt_uregbase = NULL;
1053 bail:
1054 	return ret;
1055 }
1056 
1057 static int user_init(struct file *fp)
1058 {
1059 	unsigned int rcvctrl_ops = 0;
1060 	struct hfi1_filedata *fd = fp->private_data;
1061 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1062 
1063 	/* make sure that the context has already been setup */
1064 	if (!test_bit(HFI1_CTXT_SETUP_DONE, &uctxt->event_flags))
1065 		return -EFAULT;
1066 
1067 	/* initialize poll variables... */
1068 	uctxt->urgent = 0;
1069 	uctxt->urgent_poll = 0;
1070 
1071 	/*
1072 	 * Now enable the ctxt for receive.
1073 	 * For chips that are set to DMA the tail register to memory
1074 	 * when they change (and when the update bit transitions from
1075 	 * 0 to 1.  So for those chips, we turn it off and then back on.
1076 	 * This will (very briefly) affect any other open ctxts, but the
1077 	 * duration is very short, and therefore isn't an issue.  We
1078 	 * explicitly set the in-memory tail copy to 0 beforehand, so we
1079 	 * don't have to wait to be sure the DMA update has happened
1080 	 * (chip resets head/tail to 0 on transition to enable).
1081 	 */
1082 	if (uctxt->rcvhdrtail_kvaddr)
1083 		clear_rcvhdrtail(uctxt);
1084 
1085 	/* Setup J_KEY before enabling the context */
1086 	hfi1_set_ctxt_jkey(uctxt->dd, uctxt->ctxt, uctxt->jkey);
1087 
1088 	rcvctrl_ops = HFI1_RCVCTRL_CTXT_ENB;
1089 	if (HFI1_CAP_UGET_MASK(uctxt->flags, HDRSUPP))
1090 		rcvctrl_ops |= HFI1_RCVCTRL_TIDFLOW_ENB;
1091 	/*
1092 	 * Ignore the bit in the flags for now until proper
1093 	 * support for multiple packet per rcv array entry is
1094 	 * added.
1095 	 */
1096 	if (!HFI1_CAP_UGET_MASK(uctxt->flags, MULTI_PKT_EGR))
1097 		rcvctrl_ops |= HFI1_RCVCTRL_ONE_PKT_EGR_ENB;
1098 	if (HFI1_CAP_UGET_MASK(uctxt->flags, NODROP_EGR_FULL))
1099 		rcvctrl_ops |= HFI1_RCVCTRL_NO_EGR_DROP_ENB;
1100 	if (HFI1_CAP_UGET_MASK(uctxt->flags, NODROP_RHQ_FULL))
1101 		rcvctrl_ops |= HFI1_RCVCTRL_NO_RHQ_DROP_ENB;
1102 	/*
1103 	 * The RcvCtxtCtrl.TailUpd bit has to be explicitly written.
1104 	 * We can't rely on the correct value to be set from prior
1105 	 * uses of the chip or ctxt. Therefore, add the rcvctrl op
1106 	 * for both cases.
1107 	 */
1108 	if (HFI1_CAP_UGET_MASK(uctxt->flags, DMA_RTAIL))
1109 		rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_ENB;
1110 	else
1111 		rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_DIS;
1112 	hfi1_rcvctrl(uctxt->dd, rcvctrl_ops, uctxt->ctxt);
1113 
1114 	/* Notify any waiting slaves */
1115 	if (uctxt->subctxt_cnt) {
1116 		clear_bit(HFI1_CTXT_MASTER_UNINIT, &uctxt->event_flags);
1117 		wake_up(&uctxt->wait);
1118 	}
1119 
1120 	return 0;
1121 }
1122 
1123 static int get_ctxt_info(struct file *fp, void __user *ubase, __u32 len)
1124 {
1125 	struct hfi1_ctxt_info cinfo;
1126 	struct hfi1_filedata *fd = fp->private_data;
1127 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1128 	int ret = 0;
1129 
1130 	memset(&cinfo, 0, sizeof(cinfo));
1131 	cinfo.runtime_flags = (((uctxt->flags >> HFI1_CAP_MISC_SHIFT) &
1132 				HFI1_CAP_MISC_MASK) << HFI1_CAP_USER_SHIFT) |
1133 			HFI1_CAP_UGET_MASK(uctxt->flags, MASK) |
1134 			HFI1_CAP_KGET_MASK(uctxt->flags, K2U);
1135 	/* adjust flag if this fd is not able to cache */
1136 	if (!fd->handler)
1137 		cinfo.runtime_flags |= HFI1_CAP_TID_UNMAP; /* no caching */
1138 
1139 	cinfo.num_active = hfi1_count_active_units();
1140 	cinfo.unit = uctxt->dd->unit;
1141 	cinfo.ctxt = uctxt->ctxt;
1142 	cinfo.subctxt = fd->subctxt;
1143 	cinfo.rcvtids = roundup(uctxt->egrbufs.alloced,
1144 				uctxt->dd->rcv_entries.group_size) +
1145 		uctxt->expected_count;
1146 	cinfo.credits = uctxt->sc->credits;
1147 	cinfo.numa_node = uctxt->numa_id;
1148 	cinfo.rec_cpu = fd->rec_cpu_num;
1149 	cinfo.send_ctxt = uctxt->sc->hw_context;
1150 
1151 	cinfo.egrtids = uctxt->egrbufs.alloced;
1152 	cinfo.rcvhdrq_cnt = uctxt->rcvhdrq_cnt;
1153 	cinfo.rcvhdrq_entsize = uctxt->rcvhdrqentsize << 2;
1154 	cinfo.sdma_ring_size = fd->cq->nentries;
1155 	cinfo.rcvegr_size = uctxt->egrbufs.rcvtid_size;
1156 
1157 	trace_hfi1_ctxt_info(uctxt->dd, uctxt->ctxt, fd->subctxt, cinfo);
1158 	if (copy_to_user(ubase, &cinfo, sizeof(cinfo)))
1159 		ret = -EFAULT;
1160 
1161 	return ret;
1162 }
1163 
1164 static int setup_ctxt(struct file *fp)
1165 {
1166 	struct hfi1_filedata *fd = fp->private_data;
1167 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1168 	struct hfi1_devdata *dd = uctxt->dd;
1169 	int ret = 0;
1170 
1171 	/*
1172 	 * Context should be set up only once, including allocation and
1173 	 * programming of eager buffers. This is done if context sharing
1174 	 * is not requested or by the master process.
1175 	 */
1176 	if (!uctxt->subctxt_cnt || !fd->subctxt) {
1177 		ret = hfi1_init_ctxt(uctxt->sc);
1178 		if (ret)
1179 			goto done;
1180 
1181 		/* Now allocate the RcvHdr queue and eager buffers. */
1182 		ret = hfi1_create_rcvhdrq(dd, uctxt);
1183 		if (ret)
1184 			goto done;
1185 		ret = hfi1_setup_eagerbufs(uctxt);
1186 		if (ret)
1187 			goto done;
1188 		if (uctxt->subctxt_cnt && !fd->subctxt) {
1189 			ret = setup_subctxt(uctxt);
1190 			if (ret)
1191 				goto done;
1192 		}
1193 	} else {
1194 		ret = wait_event_interruptible(uctxt->wait, !test_bit(
1195 					       HFI1_CTXT_MASTER_UNINIT,
1196 					       &uctxt->event_flags));
1197 		if (ret)
1198 			goto done;
1199 	}
1200 
1201 	ret = hfi1_user_sdma_alloc_queues(uctxt, fp);
1202 	if (ret)
1203 		goto done;
1204 	/*
1205 	 * Expected receive has to be setup for all processes (including
1206 	 * shared contexts). However, it has to be done after the master
1207 	 * context has been fully configured as it depends on the
1208 	 * eager/expected split of the RcvArray entries.
1209 	 * Setting it up here ensures that the subcontexts will be waiting
1210 	 * (due to the above wait_event_interruptible() until the master
1211 	 * is setup.
1212 	 */
1213 	ret = hfi1_user_exp_rcv_init(fp);
1214 	if (ret)
1215 		goto done;
1216 
1217 	set_bit(HFI1_CTXT_SETUP_DONE, &uctxt->event_flags);
1218 done:
1219 	return ret;
1220 }
1221 
1222 static int get_base_info(struct file *fp, void __user *ubase, __u32 len)
1223 {
1224 	struct hfi1_base_info binfo;
1225 	struct hfi1_filedata *fd = fp->private_data;
1226 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1227 	struct hfi1_devdata *dd = uctxt->dd;
1228 	ssize_t sz;
1229 	unsigned offset;
1230 	int ret = 0;
1231 
1232 	trace_hfi1_uctxtdata(uctxt->dd, uctxt);
1233 
1234 	memset(&binfo, 0, sizeof(binfo));
1235 	binfo.hw_version = dd->revision;
1236 	binfo.sw_version = HFI1_KERN_SWVERSION;
1237 	binfo.bthqp = kdeth_qp;
1238 	binfo.jkey = uctxt->jkey;
1239 	/*
1240 	 * If more than 64 contexts are enabled the allocated credit
1241 	 * return will span two or three contiguous pages. Since we only
1242 	 * map the page containing the context's credit return address,
1243 	 * we need to calculate the offset in the proper page.
1244 	 */
1245 	offset = ((u64)uctxt->sc->hw_free -
1246 		  (u64)dd->cr_base[uctxt->numa_id].va) % PAGE_SIZE;
1247 	binfo.sc_credits_addr = HFI1_MMAP_TOKEN(PIO_CRED, uctxt->ctxt,
1248 						fd->subctxt, offset);
1249 	binfo.pio_bufbase = HFI1_MMAP_TOKEN(PIO_BUFS, uctxt->ctxt,
1250 					    fd->subctxt,
1251 					    uctxt->sc->base_addr);
1252 	binfo.pio_bufbase_sop = HFI1_MMAP_TOKEN(PIO_BUFS_SOP,
1253 						uctxt->ctxt,
1254 						fd->subctxt,
1255 						uctxt->sc->base_addr);
1256 	binfo.rcvhdr_bufbase = HFI1_MMAP_TOKEN(RCV_HDRQ, uctxt->ctxt,
1257 					       fd->subctxt,
1258 					       uctxt->rcvhdrq);
1259 	binfo.rcvegr_bufbase = HFI1_MMAP_TOKEN(RCV_EGRBUF, uctxt->ctxt,
1260 					       fd->subctxt,
1261 					       uctxt->egrbufs.rcvtids[0].phys);
1262 	binfo.sdma_comp_bufbase = HFI1_MMAP_TOKEN(SDMA_COMP, uctxt->ctxt,
1263 						 fd->subctxt, 0);
1264 	/*
1265 	 * user regs are at
1266 	 * (RXE_PER_CONTEXT_USER + (ctxt * RXE_PER_CONTEXT_SIZE))
1267 	 */
1268 	binfo.user_regbase = HFI1_MMAP_TOKEN(UREGS, uctxt->ctxt,
1269 					    fd->subctxt, 0);
1270 	offset = offset_in_page((((uctxt->ctxt - dd->first_user_ctxt) *
1271 		    HFI1_MAX_SHARED_CTXTS) + fd->subctxt) *
1272 		  sizeof(*dd->events));
1273 	binfo.events_bufbase = HFI1_MMAP_TOKEN(EVENTS, uctxt->ctxt,
1274 					      fd->subctxt,
1275 					      offset);
1276 	binfo.status_bufbase = HFI1_MMAP_TOKEN(STATUS, uctxt->ctxt,
1277 					      fd->subctxt,
1278 					      dd->status);
1279 	if (HFI1_CAP_IS_USET(DMA_RTAIL))
1280 		binfo.rcvhdrtail_base = HFI1_MMAP_TOKEN(RTAIL, uctxt->ctxt,
1281 						       fd->subctxt, 0);
1282 	if (uctxt->subctxt_cnt) {
1283 		binfo.subctxt_uregbase = HFI1_MMAP_TOKEN(SUBCTXT_UREGS,
1284 							uctxt->ctxt,
1285 							fd->subctxt, 0);
1286 		binfo.subctxt_rcvhdrbuf = HFI1_MMAP_TOKEN(SUBCTXT_RCV_HDRQ,
1287 							 uctxt->ctxt,
1288 							 fd->subctxt, 0);
1289 		binfo.subctxt_rcvegrbuf = HFI1_MMAP_TOKEN(SUBCTXT_EGRBUF,
1290 							 uctxt->ctxt,
1291 							 fd->subctxt, 0);
1292 	}
1293 	sz = (len < sizeof(binfo)) ? len : sizeof(binfo);
1294 	if (copy_to_user(ubase, &binfo, sz))
1295 		ret = -EFAULT;
1296 	return ret;
1297 }
1298 
1299 static unsigned int poll_urgent(struct file *fp,
1300 				struct poll_table_struct *pt)
1301 {
1302 	struct hfi1_filedata *fd = fp->private_data;
1303 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1304 	struct hfi1_devdata *dd = uctxt->dd;
1305 	unsigned pollflag;
1306 
1307 	poll_wait(fp, &uctxt->wait, pt);
1308 
1309 	spin_lock_irq(&dd->uctxt_lock);
1310 	if (uctxt->urgent != uctxt->urgent_poll) {
1311 		pollflag = POLLIN | POLLRDNORM;
1312 		uctxt->urgent_poll = uctxt->urgent;
1313 	} else {
1314 		pollflag = 0;
1315 		set_bit(HFI1_CTXT_WAITING_URG, &uctxt->event_flags);
1316 	}
1317 	spin_unlock_irq(&dd->uctxt_lock);
1318 
1319 	return pollflag;
1320 }
1321 
1322 static unsigned int poll_next(struct file *fp,
1323 			      struct poll_table_struct *pt)
1324 {
1325 	struct hfi1_filedata *fd = fp->private_data;
1326 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1327 	struct hfi1_devdata *dd = uctxt->dd;
1328 	unsigned pollflag;
1329 
1330 	poll_wait(fp, &uctxt->wait, pt);
1331 
1332 	spin_lock_irq(&dd->uctxt_lock);
1333 	if (hdrqempty(uctxt)) {
1334 		set_bit(HFI1_CTXT_WAITING_RCV, &uctxt->event_flags);
1335 		hfi1_rcvctrl(dd, HFI1_RCVCTRL_INTRAVAIL_ENB, uctxt->ctxt);
1336 		pollflag = 0;
1337 	} else {
1338 		pollflag = POLLIN | POLLRDNORM;
1339 	}
1340 	spin_unlock_irq(&dd->uctxt_lock);
1341 
1342 	return pollflag;
1343 }
1344 
1345 /*
1346  * Find all user contexts in use, and set the specified bit in their
1347  * event mask.
1348  * See also find_ctxt() for a similar use, that is specific to send buffers.
1349  */
1350 int hfi1_set_uevent_bits(struct hfi1_pportdata *ppd, const int evtbit)
1351 {
1352 	struct hfi1_ctxtdata *uctxt;
1353 	struct hfi1_devdata *dd = ppd->dd;
1354 	unsigned ctxt;
1355 	int ret = 0;
1356 	unsigned long flags;
1357 
1358 	if (!dd->events) {
1359 		ret = -EINVAL;
1360 		goto done;
1361 	}
1362 
1363 	spin_lock_irqsave(&dd->uctxt_lock, flags);
1364 	for (ctxt = dd->first_user_ctxt; ctxt < dd->num_rcv_contexts;
1365 	     ctxt++) {
1366 		uctxt = dd->rcd[ctxt];
1367 		if (uctxt) {
1368 			unsigned long *evs = dd->events +
1369 				(uctxt->ctxt - dd->first_user_ctxt) *
1370 				HFI1_MAX_SHARED_CTXTS;
1371 			int i;
1372 			/*
1373 			 * subctxt_cnt is 0 if not shared, so do base
1374 			 * separately, first, then remaining subctxt, if any
1375 			 */
1376 			set_bit(evtbit, evs);
1377 			for (i = 1; i < uctxt->subctxt_cnt; i++)
1378 				set_bit(evtbit, evs + i);
1379 		}
1380 	}
1381 	spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1382 done:
1383 	return ret;
1384 }
1385 
1386 /**
1387  * manage_rcvq - manage a context's receive queue
1388  * @uctxt: the context
1389  * @subctxt: the sub-context
1390  * @start_stop: action to carry out
1391  *
1392  * start_stop == 0 disables receive on the context, for use in queue
1393  * overflow conditions.  start_stop==1 re-enables, to be used to
1394  * re-init the software copy of the head register
1395  */
1396 static int manage_rcvq(struct hfi1_ctxtdata *uctxt, unsigned subctxt,
1397 		       int start_stop)
1398 {
1399 	struct hfi1_devdata *dd = uctxt->dd;
1400 	unsigned int rcvctrl_op;
1401 
1402 	if (subctxt)
1403 		goto bail;
1404 	/* atomically clear receive enable ctxt. */
1405 	if (start_stop) {
1406 		/*
1407 		 * On enable, force in-memory copy of the tail register to
1408 		 * 0, so that protocol code doesn't have to worry about
1409 		 * whether or not the chip has yet updated the in-memory
1410 		 * copy or not on return from the system call. The chip
1411 		 * always resets it's tail register back to 0 on a
1412 		 * transition from disabled to enabled.
1413 		 */
1414 		if (uctxt->rcvhdrtail_kvaddr)
1415 			clear_rcvhdrtail(uctxt);
1416 		rcvctrl_op = HFI1_RCVCTRL_CTXT_ENB;
1417 	} else {
1418 		rcvctrl_op = HFI1_RCVCTRL_CTXT_DIS;
1419 	}
1420 	hfi1_rcvctrl(dd, rcvctrl_op, uctxt->ctxt);
1421 	/* always; new head should be equal to new tail; see above */
1422 bail:
1423 	return 0;
1424 }
1425 
1426 /*
1427  * clear the event notifier events for this context.
1428  * User process then performs actions appropriate to bit having been
1429  * set, if desired, and checks again in future.
1430  */
1431 static int user_event_ack(struct hfi1_ctxtdata *uctxt, int subctxt,
1432 			  unsigned long events)
1433 {
1434 	int i;
1435 	struct hfi1_devdata *dd = uctxt->dd;
1436 	unsigned long *evs;
1437 
1438 	if (!dd->events)
1439 		return 0;
1440 
1441 	evs = dd->events + ((uctxt->ctxt - dd->first_user_ctxt) *
1442 			    HFI1_MAX_SHARED_CTXTS) + subctxt;
1443 
1444 	for (i = 0; i <= _HFI1_MAX_EVENT_BIT; i++) {
1445 		if (!test_bit(i, &events))
1446 			continue;
1447 		clear_bit(i, evs);
1448 	}
1449 	return 0;
1450 }
1451 
1452 static int set_ctxt_pkey(struct hfi1_ctxtdata *uctxt, unsigned subctxt,
1453 			 u16 pkey)
1454 {
1455 	int ret = -ENOENT, i, intable = 0;
1456 	struct hfi1_pportdata *ppd = uctxt->ppd;
1457 	struct hfi1_devdata *dd = uctxt->dd;
1458 
1459 	if (pkey == LIM_MGMT_P_KEY || pkey == FULL_MGMT_P_KEY) {
1460 		ret = -EINVAL;
1461 		goto done;
1462 	}
1463 
1464 	for (i = 0; i < ARRAY_SIZE(ppd->pkeys); i++)
1465 		if (pkey == ppd->pkeys[i]) {
1466 			intable = 1;
1467 			break;
1468 		}
1469 
1470 	if (intable)
1471 		ret = hfi1_set_ctxt_pkey(dd, uctxt->ctxt, pkey);
1472 done:
1473 	return ret;
1474 }
1475 
1476 static void user_remove(struct hfi1_devdata *dd)
1477 {
1478 
1479 	hfi1_cdev_cleanup(&dd->user_cdev, &dd->user_device);
1480 }
1481 
1482 static int user_add(struct hfi1_devdata *dd)
1483 {
1484 	char name[10];
1485 	int ret;
1486 
1487 	snprintf(name, sizeof(name), "%s_%d", class_name(), dd->unit);
1488 	ret = hfi1_cdev_init(dd->unit, name, &hfi1_file_ops,
1489 			     &dd->user_cdev, &dd->user_device,
1490 			     true, &dd->kobj);
1491 	if (ret)
1492 		user_remove(dd);
1493 
1494 	return ret;
1495 }
1496 
1497 /*
1498  * Create per-unit files in /dev
1499  */
1500 int hfi1_device_create(struct hfi1_devdata *dd)
1501 {
1502 	return user_add(dd);
1503 }
1504 
1505 /*
1506  * Remove per-unit files in /dev
1507  * void, core kernel returns no errors for this stuff
1508  */
1509 void hfi1_device_remove(struct hfi1_devdata *dd)
1510 {
1511 	user_remove(dd);
1512 }
1513