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 "aspm.h"
62 #include "mmu_rb.h"
63 
64 #undef pr_fmt
65 #define pr_fmt(fmt) DRIVER_NAME ": " fmt
66 
67 #define SEND_CTXT_HALT_TIMEOUT 1000 /* msecs */
68 
69 /*
70  * File operation functions
71  */
72 static int hfi1_file_open(struct inode *, struct file *);
73 static int hfi1_file_close(struct inode *, struct file *);
74 static ssize_t hfi1_write_iter(struct kiocb *, struct iov_iter *);
75 static unsigned int hfi1_poll(struct file *, struct poll_table_struct *);
76 static int hfi1_file_mmap(struct file *, struct vm_area_struct *);
77 
78 static u64 kvirt_to_phys(void *);
79 static int assign_ctxt(struct file *, struct hfi1_user_info *);
80 static int init_subctxts(struct hfi1_ctxtdata *, const struct hfi1_user_info *);
81 static int user_init(struct file *);
82 static int get_ctxt_info(struct file *, void __user *, __u32);
83 static int get_base_info(struct file *, void __user *, __u32);
84 static int setup_ctxt(struct file *);
85 static int setup_subctxt(struct hfi1_ctxtdata *);
86 static int get_user_context(struct file *, struct hfi1_user_info *, int);
87 static int find_shared_ctxt(struct file *, const struct hfi1_user_info *);
88 static int allocate_ctxt(struct file *, struct hfi1_devdata *,
89 			 struct hfi1_user_info *);
90 static unsigned int poll_urgent(struct file *, struct poll_table_struct *);
91 static unsigned int poll_next(struct file *, struct poll_table_struct *);
92 static int user_event_ack(struct hfi1_ctxtdata *, int, unsigned long);
93 static int set_ctxt_pkey(struct hfi1_ctxtdata *, unsigned, u16);
94 static int manage_rcvq(struct hfi1_ctxtdata *, unsigned, int);
95 static int vma_fault(struct vm_area_struct *, struct vm_fault *);
96 static long hfi1_file_ioctl(struct file *fp, unsigned int cmd,
97 			    unsigned long arg);
98 
99 static const struct file_operations hfi1_file_ops = {
100 	.owner = THIS_MODULE,
101 	.write_iter = hfi1_write_iter,
102 	.open = hfi1_file_open,
103 	.release = hfi1_file_close,
104 	.unlocked_ioctl = hfi1_file_ioctl,
105 	.poll = hfi1_poll,
106 	.mmap = hfi1_file_mmap,
107 	.llseek = noop_llseek,
108 };
109 
110 static struct vm_operations_struct vm_ops = {
111 	.fault = vma_fault,
112 };
113 
114 /*
115  * Types of memories mapped into user processes' space
116  */
117 enum mmap_types {
118 	PIO_BUFS = 1,
119 	PIO_BUFS_SOP,
120 	PIO_CRED,
121 	RCV_HDRQ,
122 	RCV_EGRBUF,
123 	UREGS,
124 	EVENTS,
125 	STATUS,
126 	RTAIL,
127 	SUBCTXT_UREGS,
128 	SUBCTXT_RCV_HDRQ,
129 	SUBCTXT_EGRBUF,
130 	SDMA_COMP
131 };
132 
133 /*
134  * Masks and offsets defining the mmap tokens
135  */
136 #define HFI1_MMAP_OFFSET_MASK   0xfffULL
137 #define HFI1_MMAP_OFFSET_SHIFT  0
138 #define HFI1_MMAP_SUBCTXT_MASK  0xfULL
139 #define HFI1_MMAP_SUBCTXT_SHIFT 12
140 #define HFI1_MMAP_CTXT_MASK     0xffULL
141 #define HFI1_MMAP_CTXT_SHIFT    16
142 #define HFI1_MMAP_TYPE_MASK     0xfULL
143 #define HFI1_MMAP_TYPE_SHIFT    24
144 #define HFI1_MMAP_MAGIC_MASK    0xffffffffULL
145 #define HFI1_MMAP_MAGIC_SHIFT   32
146 
147 #define HFI1_MMAP_MAGIC         0xdabbad00
148 
149 #define HFI1_MMAP_TOKEN_SET(field, val)	\
150 	(((val) & HFI1_MMAP_##field##_MASK) << HFI1_MMAP_##field##_SHIFT)
151 #define HFI1_MMAP_TOKEN_GET(field, token) \
152 	(((token) >> HFI1_MMAP_##field##_SHIFT) & HFI1_MMAP_##field##_MASK)
153 #define HFI1_MMAP_TOKEN(type, ctxt, subctxt, addr)   \
154 	(HFI1_MMAP_TOKEN_SET(MAGIC, HFI1_MMAP_MAGIC) | \
155 	HFI1_MMAP_TOKEN_SET(TYPE, type) | \
156 	HFI1_MMAP_TOKEN_SET(CTXT, ctxt) | \
157 	HFI1_MMAP_TOKEN_SET(SUBCTXT, subctxt) | \
158 	HFI1_MMAP_TOKEN_SET(OFFSET, (offset_in_page(addr))))
159 
160 #define dbg(fmt, ...)				\
161 	pr_info(fmt, ##__VA_ARGS__)
162 
163 static inline int is_valid_mmap(u64 token)
164 {
165 	return (HFI1_MMAP_TOKEN_GET(MAGIC, token) == HFI1_MMAP_MAGIC);
166 }
167 
168 static int hfi1_file_open(struct inode *inode, struct file *fp)
169 {
170 	struct hfi1_filedata *fd;
171 	struct hfi1_devdata *dd = container_of(inode->i_cdev,
172 					       struct hfi1_devdata,
173 					       user_cdev);
174 
175 	/* Just take a ref now. Not all opens result in a context assign */
176 	kobject_get(&dd->kobj);
177 
178 	/* The real work is performed later in assign_ctxt() */
179 
180 	fd = kzalloc(sizeof(*fd), GFP_KERNEL);
181 
182 	if (fd) {
183 		fd->rec_cpu_num = -1; /* no cpu affinity by default */
184 		fd->mm = current->mm;
185 		atomic_inc(&fd->mm->mm_count);
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 		ret = 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;
443 	u64 token = vma->vm_pgoff << PAGE_SHIFT,
444 		memaddr = 0;
445 	void *memvirt = NULL;
446 	u8 subctxt, mapio = 0, vmf = 0, type;
447 	ssize_t memlen = 0;
448 	int ret = 0;
449 	u16 ctxt;
450 
451 	if (!is_valid_mmap(token) || !uctxt ||
452 	    !(vma->vm_flags & VM_SHARED)) {
453 		ret = -EINVAL;
454 		goto done;
455 	}
456 	dd = uctxt->dd;
457 	ctxt = HFI1_MMAP_TOKEN_GET(CTXT, token);
458 	subctxt = HFI1_MMAP_TOKEN_GET(SUBCTXT, token);
459 	type = HFI1_MMAP_TOKEN_GET(TYPE, token);
460 	if (ctxt != uctxt->ctxt || subctxt != fd->subctxt) {
461 		ret = -EINVAL;
462 		goto done;
463 	}
464 
465 	flags = vma->vm_flags;
466 
467 	switch (type) {
468 	case PIO_BUFS:
469 	case PIO_BUFS_SOP:
470 		memaddr = ((dd->physaddr + TXE_PIO_SEND) +
471 				/* chip pio base */
472 			   (uctxt->sc->hw_context * BIT(16))) +
473 				/* 64K PIO space / ctxt */
474 			(type == PIO_BUFS_SOP ?
475 				(TXE_PIO_SIZE / 2) : 0); /* sop? */
476 		/*
477 		 * Map only the amount allocated to the context, not the
478 		 * entire available context's PIO space.
479 		 */
480 		memlen = PAGE_ALIGN(uctxt->sc->credits * PIO_BLOCK_SIZE);
481 		flags &= ~VM_MAYREAD;
482 		flags |= VM_DONTCOPY | VM_DONTEXPAND;
483 		vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
484 		mapio = 1;
485 		break;
486 	case PIO_CRED:
487 		if (flags & VM_WRITE) {
488 			ret = -EPERM;
489 			goto done;
490 		}
491 		/*
492 		 * The credit return location for this context could be on the
493 		 * second or third page allocated for credit returns (if number
494 		 * of enabled contexts > 64 and 128 respectively).
495 		 */
496 		memvirt = dd->cr_base[uctxt->numa_id].va;
497 		memaddr = virt_to_phys(memvirt) +
498 			(((u64)uctxt->sc->hw_free -
499 			  (u64)dd->cr_base[uctxt->numa_id].va) & PAGE_MASK);
500 		memlen = PAGE_SIZE;
501 		flags &= ~VM_MAYWRITE;
502 		flags |= VM_DONTCOPY | VM_DONTEXPAND;
503 		/*
504 		 * The driver has already allocated memory for credit
505 		 * returns and programmed it into the chip. Has that
506 		 * memory been flagged as non-cached?
507 		 */
508 		/* vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); */
509 		mapio = 1;
510 		break;
511 	case RCV_HDRQ:
512 		memlen = uctxt->rcvhdrq_size;
513 		memvirt = uctxt->rcvhdrq;
514 		break;
515 	case RCV_EGRBUF: {
516 		unsigned long addr;
517 		int i;
518 		/*
519 		 * The RcvEgr buffer need to be handled differently
520 		 * as multiple non-contiguous pages need to be mapped
521 		 * into the user process.
522 		 */
523 		memlen = uctxt->egrbufs.size;
524 		if ((vma->vm_end - vma->vm_start) != memlen) {
525 			dd_dev_err(dd, "Eager buffer map size invalid (%lu != %lu)\n",
526 				   (vma->vm_end - vma->vm_start), memlen);
527 			ret = -EINVAL;
528 			goto done;
529 		}
530 		if (vma->vm_flags & VM_WRITE) {
531 			ret = -EPERM;
532 			goto done;
533 		}
534 		vma->vm_flags &= ~VM_MAYWRITE;
535 		addr = vma->vm_start;
536 		for (i = 0 ; i < uctxt->egrbufs.numbufs; i++) {
537 			memlen = uctxt->egrbufs.buffers[i].len;
538 			memvirt = uctxt->egrbufs.buffers[i].addr;
539 			ret = remap_pfn_range(
540 				vma, addr,
541 				/*
542 				 * virt_to_pfn() does the same, but
543 				 * it's not available on x86_64
544 				 * when CONFIG_MMU is enabled.
545 				 */
546 				PFN_DOWN(__pa(memvirt)),
547 				memlen,
548 				vma->vm_page_prot);
549 			if (ret < 0)
550 				goto done;
551 			addr += memlen;
552 		}
553 		ret = 0;
554 		goto done;
555 	}
556 	case UREGS:
557 		/*
558 		 * Map only the page that contains this context's user
559 		 * registers.
560 		 */
561 		memaddr = (unsigned long)
562 			(dd->physaddr + RXE_PER_CONTEXT_USER)
563 			+ (uctxt->ctxt * RXE_PER_CONTEXT_SIZE);
564 		/*
565 		 * TidFlow table is on the same page as the rest of the
566 		 * user registers.
567 		 */
568 		memlen = PAGE_SIZE;
569 		flags |= VM_DONTCOPY | VM_DONTEXPAND;
570 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
571 		mapio = 1;
572 		break;
573 	case EVENTS:
574 		/*
575 		 * Use the page where this context's flags are. User level
576 		 * knows where it's own bitmap is within the page.
577 		 */
578 		memaddr = (unsigned long)(dd->events +
579 					  ((uctxt->ctxt - dd->first_user_ctxt) *
580 					   HFI1_MAX_SHARED_CTXTS)) & PAGE_MASK;
581 		memlen = PAGE_SIZE;
582 		/*
583 		 * v3.7 removes VM_RESERVED but the effect is kept by
584 		 * using VM_IO.
585 		 */
586 		flags |= VM_IO | VM_DONTEXPAND;
587 		vmf = 1;
588 		break;
589 	case STATUS:
590 		memaddr = kvirt_to_phys((void *)dd->status);
591 		memlen = PAGE_SIZE;
592 		flags |= VM_IO | VM_DONTEXPAND;
593 		break;
594 	case RTAIL:
595 		if (!HFI1_CAP_IS_USET(DMA_RTAIL)) {
596 			/*
597 			 * If the memory allocation failed, the context alloc
598 			 * also would have failed, so we would never get here
599 			 */
600 			ret = -EINVAL;
601 			goto done;
602 		}
603 		if (flags & VM_WRITE) {
604 			ret = -EPERM;
605 			goto done;
606 		}
607 		memlen = PAGE_SIZE;
608 		memvirt = (void *)uctxt->rcvhdrtail_kvaddr;
609 		flags &= ~VM_MAYWRITE;
610 		break;
611 	case SUBCTXT_UREGS:
612 		memaddr = (u64)uctxt->subctxt_uregbase;
613 		memlen = PAGE_SIZE;
614 		flags |= VM_IO | VM_DONTEXPAND;
615 		vmf = 1;
616 		break;
617 	case SUBCTXT_RCV_HDRQ:
618 		memaddr = (u64)uctxt->subctxt_rcvhdr_base;
619 		memlen = uctxt->rcvhdrq_size * uctxt->subctxt_cnt;
620 		flags |= VM_IO | VM_DONTEXPAND;
621 		vmf = 1;
622 		break;
623 	case SUBCTXT_EGRBUF:
624 		memaddr = (u64)uctxt->subctxt_rcvegrbuf;
625 		memlen = uctxt->egrbufs.size * uctxt->subctxt_cnt;
626 		flags |= VM_IO | VM_DONTEXPAND;
627 		flags &= ~VM_MAYWRITE;
628 		vmf = 1;
629 		break;
630 	case SDMA_COMP: {
631 		struct hfi1_user_sdma_comp_q *cq = fd->cq;
632 
633 		if (!cq) {
634 			ret = -EFAULT;
635 			goto done;
636 		}
637 		memaddr = (u64)cq->comps;
638 		memlen = PAGE_ALIGN(sizeof(*cq->comps) * cq->nentries);
639 		flags |= VM_IO | VM_DONTEXPAND;
640 		vmf = 1;
641 		break;
642 	}
643 	default:
644 		ret = -EINVAL;
645 		break;
646 	}
647 
648 	if ((vma->vm_end - vma->vm_start) != memlen) {
649 		hfi1_cdbg(PROC, "%u:%u Memory size mismatch %lu:%lu",
650 			  uctxt->ctxt, fd->subctxt,
651 			  (vma->vm_end - vma->vm_start), memlen);
652 		ret = -EINVAL;
653 		goto done;
654 	}
655 
656 	vma->vm_flags = flags;
657 	hfi1_cdbg(PROC,
658 		  "%u:%u type:%u io/vf:%d/%d, addr:0x%llx, len:%lu(%lu), flags:0x%lx\n",
659 		    ctxt, subctxt, type, mapio, vmf, memaddr, memlen,
660 		    vma->vm_end - vma->vm_start, vma->vm_flags);
661 	if (vmf) {
662 		vma->vm_pgoff = PFN_DOWN(memaddr);
663 		vma->vm_ops = &vm_ops;
664 		ret = 0;
665 	} else if (mapio) {
666 		ret = io_remap_pfn_range(vma, vma->vm_start,
667 					 PFN_DOWN(memaddr),
668 					 memlen,
669 					 vma->vm_page_prot);
670 	} else if (memvirt) {
671 		ret = remap_pfn_range(vma, vma->vm_start,
672 				      PFN_DOWN(__pa(memvirt)),
673 				      memlen,
674 				      vma->vm_page_prot);
675 	} else {
676 		ret = remap_pfn_range(vma, vma->vm_start,
677 				      PFN_DOWN(memaddr),
678 				      memlen,
679 				      vma->vm_page_prot);
680 	}
681 done:
682 	return ret;
683 }
684 
685 /*
686  * Local (non-chip) user memory is not mapped right away but as it is
687  * accessed by the user-level code.
688  */
689 static int vma_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
690 {
691 	struct page *page;
692 
693 	page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
694 	if (!page)
695 		return VM_FAULT_SIGBUS;
696 
697 	get_page(page);
698 	vmf->page = page;
699 
700 	return 0;
701 }
702 
703 static unsigned int hfi1_poll(struct file *fp, struct poll_table_struct *pt)
704 {
705 	struct hfi1_ctxtdata *uctxt;
706 	unsigned pollflag;
707 
708 	uctxt = ((struct hfi1_filedata *)fp->private_data)->uctxt;
709 	if (!uctxt)
710 		pollflag = POLLERR;
711 	else if (uctxt->poll_type == HFI1_POLL_TYPE_URGENT)
712 		pollflag = poll_urgent(fp, pt);
713 	else  if (uctxt->poll_type == HFI1_POLL_TYPE_ANYRCV)
714 		pollflag = poll_next(fp, pt);
715 	else /* invalid */
716 		pollflag = POLLERR;
717 
718 	return pollflag;
719 }
720 
721 static int hfi1_file_close(struct inode *inode, struct file *fp)
722 {
723 	struct hfi1_filedata *fdata = fp->private_data;
724 	struct hfi1_ctxtdata *uctxt = fdata->uctxt;
725 	struct hfi1_devdata *dd = container_of(inode->i_cdev,
726 					       struct hfi1_devdata,
727 					       user_cdev);
728 	unsigned long flags, *ev;
729 
730 	fp->private_data = NULL;
731 
732 	if (!uctxt)
733 		goto done;
734 
735 	hfi1_cdbg(PROC, "freeing ctxt %u:%u", uctxt->ctxt, fdata->subctxt);
736 	mutex_lock(&hfi1_mutex);
737 
738 	flush_wc();
739 	/* drain user sdma queue */
740 	hfi1_user_sdma_free_queues(fdata);
741 
742 	/* release the cpu */
743 	hfi1_put_proc_affinity(fdata->rec_cpu_num);
744 
745 	/*
746 	 * Clear any left over, unhandled events so the next process that
747 	 * gets this context doesn't get confused.
748 	 */
749 	ev = dd->events + ((uctxt->ctxt - dd->first_user_ctxt) *
750 			   HFI1_MAX_SHARED_CTXTS) + fdata->subctxt;
751 	*ev = 0;
752 
753 	if (--uctxt->cnt) {
754 		uctxt->active_slaves &= ~(1 << fdata->subctxt);
755 		mutex_unlock(&hfi1_mutex);
756 		goto done;
757 	}
758 
759 	spin_lock_irqsave(&dd->uctxt_lock, flags);
760 	/*
761 	 * Disable receive context and interrupt available, reset all
762 	 * RcvCtxtCtrl bits to default values.
763 	 */
764 	hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_DIS |
765 		     HFI1_RCVCTRL_TIDFLOW_DIS |
766 		     HFI1_RCVCTRL_INTRAVAIL_DIS |
767 		     HFI1_RCVCTRL_TAILUPD_DIS |
768 		     HFI1_RCVCTRL_ONE_PKT_EGR_DIS |
769 		     HFI1_RCVCTRL_NO_RHQ_DROP_DIS |
770 		     HFI1_RCVCTRL_NO_EGR_DROP_DIS, uctxt->ctxt);
771 	/* Clear the context's J_KEY */
772 	hfi1_clear_ctxt_jkey(dd, uctxt->ctxt);
773 	/*
774 	 * Reset context integrity checks to default.
775 	 * (writes to CSRs probably belong in chip.c)
776 	 */
777 	write_kctxt_csr(dd, uctxt->sc->hw_context, SEND_CTXT_CHECK_ENABLE,
778 			hfi1_pkt_default_send_ctxt_mask(dd, uctxt->sc->type));
779 	sc_disable(uctxt->sc);
780 	spin_unlock_irqrestore(&dd->uctxt_lock, flags);
781 
782 	dd->rcd[uctxt->ctxt] = NULL;
783 
784 	hfi1_user_exp_rcv_free(fdata);
785 	hfi1_clear_ctxt_pkey(dd, uctxt->ctxt);
786 
787 	uctxt->rcvwait_to = 0;
788 	uctxt->piowait_to = 0;
789 	uctxt->rcvnowait = 0;
790 	uctxt->pionowait = 0;
791 	uctxt->event_flags = 0;
792 
793 	hfi1_stats.sps_ctxts--;
794 	if (++dd->freectxts == dd->num_user_contexts)
795 		aspm_enable_all(dd);
796 	mutex_unlock(&hfi1_mutex);
797 	hfi1_free_ctxtdata(dd, uctxt);
798 done:
799 	mmdrop(fdata->mm);
800 	kobject_put(&dd->kobj);
801 	kfree(fdata);
802 	return 0;
803 }
804 
805 /*
806  * Convert kernel *virtual* addresses to physical addresses.
807  * This is used to vmalloc'ed addresses.
808  */
809 static u64 kvirt_to_phys(void *addr)
810 {
811 	struct page *page;
812 	u64 paddr = 0;
813 
814 	page = vmalloc_to_page(addr);
815 	if (page)
816 		paddr = page_to_pfn(page) << PAGE_SHIFT;
817 
818 	return paddr;
819 }
820 
821 static int assign_ctxt(struct file *fp, struct hfi1_user_info *uinfo)
822 {
823 	int i_minor, ret = 0;
824 	unsigned int swmajor, swminor;
825 
826 	swmajor = uinfo->userversion >> 16;
827 	if (swmajor != HFI1_USER_SWMAJOR) {
828 		ret = -ENODEV;
829 		goto done;
830 	}
831 
832 	swminor = uinfo->userversion & 0xffff;
833 
834 	mutex_lock(&hfi1_mutex);
835 	/* First, lets check if we need to setup a shared context? */
836 	if (uinfo->subctxt_cnt) {
837 		struct hfi1_filedata *fd = fp->private_data;
838 
839 		ret = find_shared_ctxt(fp, uinfo);
840 		if (ret < 0)
841 			goto done_unlock;
842 		if (ret) {
843 			fd->rec_cpu_num =
844 				hfi1_get_proc_affinity(fd->uctxt->numa_id);
845 		}
846 	}
847 
848 	/*
849 	 * We execute the following block if we couldn't find a
850 	 * shared context or if context sharing is not required.
851 	 */
852 	if (!ret) {
853 		i_minor = iminor(file_inode(fp)) - HFI1_USER_MINOR_BASE;
854 		ret = get_user_context(fp, uinfo, i_minor);
855 	}
856 done_unlock:
857 	mutex_unlock(&hfi1_mutex);
858 done:
859 	return ret;
860 }
861 
862 static int get_user_context(struct file *fp, struct hfi1_user_info *uinfo,
863 			    int devno)
864 {
865 	struct hfi1_devdata *dd = NULL;
866 	int devmax, npresent, nup;
867 
868 	devmax = hfi1_count_units(&npresent, &nup);
869 	if (!npresent)
870 		return -ENXIO;
871 
872 	if (!nup)
873 		return -ENETDOWN;
874 
875 	dd = hfi1_lookup(devno);
876 	if (!dd)
877 		return -ENODEV;
878 	else if (!dd->freectxts)
879 		return -EBUSY;
880 
881 	return allocate_ctxt(fp, dd, uinfo);
882 }
883 
884 static int find_shared_ctxt(struct file *fp,
885 			    const struct hfi1_user_info *uinfo)
886 {
887 	int devmax, ndev, i;
888 	int ret = 0;
889 	struct hfi1_filedata *fd = fp->private_data;
890 
891 	devmax = hfi1_count_units(NULL, NULL);
892 
893 	for (ndev = 0; ndev < devmax; ndev++) {
894 		struct hfi1_devdata *dd = hfi1_lookup(ndev);
895 
896 		if (!(dd && (dd->flags & HFI1_PRESENT) && dd->kregbase))
897 			continue;
898 		for (i = dd->first_user_ctxt; i < dd->num_rcv_contexts; i++) {
899 			struct hfi1_ctxtdata *uctxt = dd->rcd[i];
900 
901 			/* Skip ctxts which are not yet open */
902 			if (!uctxt || !uctxt->cnt)
903 				continue;
904 			/* Skip ctxt if it doesn't match the requested one */
905 			if (memcmp(uctxt->uuid, uinfo->uuid,
906 				   sizeof(uctxt->uuid)) ||
907 			    uctxt->jkey != generate_jkey(current_uid()) ||
908 			    uctxt->subctxt_id != uinfo->subctxt_id ||
909 			    uctxt->subctxt_cnt != uinfo->subctxt_cnt)
910 				continue;
911 
912 			/* Verify the sharing process matches the master */
913 			if (uctxt->userversion != uinfo->userversion ||
914 			    uctxt->cnt >= uctxt->subctxt_cnt) {
915 				ret = -EINVAL;
916 				goto done;
917 			}
918 			fd->uctxt = uctxt;
919 			fd->subctxt  = uctxt->cnt++;
920 			uctxt->active_slaves |= 1 << fd->subctxt;
921 			ret = 1;
922 			goto done;
923 		}
924 	}
925 
926 done:
927 	return ret;
928 }
929 
930 static int allocate_ctxt(struct file *fp, struct hfi1_devdata *dd,
931 			 struct hfi1_user_info *uinfo)
932 {
933 	struct hfi1_filedata *fd = fp->private_data;
934 	struct hfi1_ctxtdata *uctxt;
935 	unsigned ctxt;
936 	int ret, numa;
937 
938 	if (dd->flags & HFI1_FROZEN) {
939 		/*
940 		 * Pick an error that is unique from all other errors
941 		 * that are returned so the user process knows that
942 		 * it tried to allocate while the SPC was frozen.  It
943 		 * it should be able to retry with success in a short
944 		 * while.
945 		 */
946 		return -EIO;
947 	}
948 
949 	for (ctxt = dd->first_user_ctxt; ctxt < dd->num_rcv_contexts; ctxt++)
950 		if (!dd->rcd[ctxt])
951 			break;
952 
953 	if (ctxt == dd->num_rcv_contexts)
954 		return -EBUSY;
955 
956 	/*
957 	 * If we don't have a NUMA node requested, preference is towards
958 	 * device NUMA node.
959 	 */
960 	fd->rec_cpu_num = hfi1_get_proc_affinity(dd->node);
961 	if (fd->rec_cpu_num != -1)
962 		numa = cpu_to_node(fd->rec_cpu_num);
963 	else
964 		numa = numa_node_id();
965 	uctxt = hfi1_create_ctxtdata(dd->pport, ctxt, numa);
966 	if (!uctxt) {
967 		dd_dev_err(dd,
968 			   "Unable to allocate ctxtdata memory, failing open\n");
969 		return -ENOMEM;
970 	}
971 	hfi1_cdbg(PROC, "[%u:%u] pid %u assigned to CPU %d (NUMA %u)",
972 		  uctxt->ctxt, fd->subctxt, current->pid, fd->rec_cpu_num,
973 		  uctxt->numa_id);
974 
975 	/*
976 	 * Allocate and enable a PIO send context.
977 	 */
978 	uctxt->sc = sc_alloc(dd, SC_USER, uctxt->rcvhdrqentsize,
979 			     uctxt->dd->node);
980 	if (!uctxt->sc) {
981 		ret = -ENOMEM;
982 		goto ctxdata_free;
983 	}
984 	hfi1_cdbg(PROC, "allocated send context %u(%u)\n", uctxt->sc->sw_index,
985 		  uctxt->sc->hw_context);
986 	ret = sc_enable(uctxt->sc);
987 	if (ret)
988 		goto ctxdata_free;
989 
990 	/*
991 	 * Setup shared context resources if the user-level has requested
992 	 * shared contexts and this is the 'master' process.
993 	 * This has to be done here so the rest of the sub-contexts find the
994 	 * proper master.
995 	 */
996 	if (uinfo->subctxt_cnt && !fd->subctxt) {
997 		ret = init_subctxts(uctxt, uinfo);
998 		/*
999 		 * On error, we don't need to disable and de-allocate the
1000 		 * send context because it will be done during file close
1001 		 */
1002 		if (ret)
1003 			goto ctxdata_free;
1004 	}
1005 	uctxt->userversion = uinfo->userversion;
1006 	uctxt->flags = hfi1_cap_mask; /* save current flag state */
1007 	init_waitqueue_head(&uctxt->wait);
1008 	strlcpy(uctxt->comm, current->comm, sizeof(uctxt->comm));
1009 	memcpy(uctxt->uuid, uinfo->uuid, sizeof(uctxt->uuid));
1010 	uctxt->jkey = generate_jkey(current_uid());
1011 	INIT_LIST_HEAD(&uctxt->sdma_queues);
1012 	spin_lock_init(&uctxt->sdma_qlock);
1013 	hfi1_stats.sps_ctxts++;
1014 	/*
1015 	 * Disable ASPM when there are open user/PSM contexts to avoid
1016 	 * issues with ASPM L1 exit latency
1017 	 */
1018 	if (dd->freectxts-- == dd->num_user_contexts)
1019 		aspm_disable_all(dd);
1020 	fd->uctxt = uctxt;
1021 
1022 	return 0;
1023 
1024 ctxdata_free:
1025 	dd->rcd[ctxt] = NULL;
1026 	hfi1_free_ctxtdata(dd, uctxt);
1027 	return ret;
1028 }
1029 
1030 static int init_subctxts(struct hfi1_ctxtdata *uctxt,
1031 			 const struct hfi1_user_info *uinfo)
1032 {
1033 	unsigned num_subctxts;
1034 
1035 	num_subctxts = uinfo->subctxt_cnt;
1036 	if (num_subctxts > HFI1_MAX_SHARED_CTXTS)
1037 		return -EINVAL;
1038 
1039 	uctxt->subctxt_cnt = uinfo->subctxt_cnt;
1040 	uctxt->subctxt_id = uinfo->subctxt_id;
1041 	uctxt->active_slaves = 1;
1042 	uctxt->redirect_seq_cnt = 1;
1043 	set_bit(HFI1_CTXT_MASTER_UNINIT, &uctxt->event_flags);
1044 
1045 	return 0;
1046 }
1047 
1048 static int setup_subctxt(struct hfi1_ctxtdata *uctxt)
1049 {
1050 	int ret = 0;
1051 	unsigned num_subctxts = uctxt->subctxt_cnt;
1052 
1053 	uctxt->subctxt_uregbase = vmalloc_user(PAGE_SIZE);
1054 	if (!uctxt->subctxt_uregbase) {
1055 		ret = -ENOMEM;
1056 		goto bail;
1057 	}
1058 	/* We can take the size of the RcvHdr Queue from the master */
1059 	uctxt->subctxt_rcvhdr_base = vmalloc_user(uctxt->rcvhdrq_size *
1060 						  num_subctxts);
1061 	if (!uctxt->subctxt_rcvhdr_base) {
1062 		ret = -ENOMEM;
1063 		goto bail_ureg;
1064 	}
1065 
1066 	uctxt->subctxt_rcvegrbuf = vmalloc_user(uctxt->egrbufs.size *
1067 						num_subctxts);
1068 	if (!uctxt->subctxt_rcvegrbuf) {
1069 		ret = -ENOMEM;
1070 		goto bail_rhdr;
1071 	}
1072 	goto bail;
1073 bail_rhdr:
1074 	vfree(uctxt->subctxt_rcvhdr_base);
1075 bail_ureg:
1076 	vfree(uctxt->subctxt_uregbase);
1077 	uctxt->subctxt_uregbase = NULL;
1078 bail:
1079 	return ret;
1080 }
1081 
1082 static int user_init(struct file *fp)
1083 {
1084 	unsigned int rcvctrl_ops = 0;
1085 	struct hfi1_filedata *fd = fp->private_data;
1086 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1087 
1088 	/* make sure that the context has already been setup */
1089 	if (!test_bit(HFI1_CTXT_SETUP_DONE, &uctxt->event_flags))
1090 		return -EFAULT;
1091 
1092 	/* initialize poll variables... */
1093 	uctxt->urgent = 0;
1094 	uctxt->urgent_poll = 0;
1095 
1096 	/*
1097 	 * Now enable the ctxt for receive.
1098 	 * For chips that are set to DMA the tail register to memory
1099 	 * when they change (and when the update bit transitions from
1100 	 * 0 to 1.  So for those chips, we turn it off and then back on.
1101 	 * This will (very briefly) affect any other open ctxts, but the
1102 	 * duration is very short, and therefore isn't an issue.  We
1103 	 * explicitly set the in-memory tail copy to 0 beforehand, so we
1104 	 * don't have to wait to be sure the DMA update has happened
1105 	 * (chip resets head/tail to 0 on transition to enable).
1106 	 */
1107 	if (uctxt->rcvhdrtail_kvaddr)
1108 		clear_rcvhdrtail(uctxt);
1109 
1110 	/* Setup J_KEY before enabling the context */
1111 	hfi1_set_ctxt_jkey(uctxt->dd, uctxt->ctxt, uctxt->jkey);
1112 
1113 	rcvctrl_ops = HFI1_RCVCTRL_CTXT_ENB;
1114 	if (HFI1_CAP_UGET_MASK(uctxt->flags, HDRSUPP))
1115 		rcvctrl_ops |= HFI1_RCVCTRL_TIDFLOW_ENB;
1116 	/*
1117 	 * Ignore the bit in the flags for now until proper
1118 	 * support for multiple packet per rcv array entry is
1119 	 * added.
1120 	 */
1121 	if (!HFI1_CAP_UGET_MASK(uctxt->flags, MULTI_PKT_EGR))
1122 		rcvctrl_ops |= HFI1_RCVCTRL_ONE_PKT_EGR_ENB;
1123 	if (HFI1_CAP_UGET_MASK(uctxt->flags, NODROP_EGR_FULL))
1124 		rcvctrl_ops |= HFI1_RCVCTRL_NO_EGR_DROP_ENB;
1125 	if (HFI1_CAP_UGET_MASK(uctxt->flags, NODROP_RHQ_FULL))
1126 		rcvctrl_ops |= HFI1_RCVCTRL_NO_RHQ_DROP_ENB;
1127 	/*
1128 	 * The RcvCtxtCtrl.TailUpd bit has to be explicitly written.
1129 	 * We can't rely on the correct value to be set from prior
1130 	 * uses of the chip or ctxt. Therefore, add the rcvctrl op
1131 	 * for both cases.
1132 	 */
1133 	if (HFI1_CAP_UGET_MASK(uctxt->flags, DMA_RTAIL))
1134 		rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_ENB;
1135 	else
1136 		rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_DIS;
1137 	hfi1_rcvctrl(uctxt->dd, rcvctrl_ops, uctxt->ctxt);
1138 
1139 	/* Notify any waiting slaves */
1140 	if (uctxt->subctxt_cnt) {
1141 		clear_bit(HFI1_CTXT_MASTER_UNINIT, &uctxt->event_flags);
1142 		wake_up(&uctxt->wait);
1143 	}
1144 
1145 	return 0;
1146 }
1147 
1148 static int get_ctxt_info(struct file *fp, void __user *ubase, __u32 len)
1149 {
1150 	struct hfi1_ctxt_info cinfo;
1151 	struct hfi1_filedata *fd = fp->private_data;
1152 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1153 	int ret = 0;
1154 
1155 	memset(&cinfo, 0, sizeof(cinfo));
1156 	cinfo.runtime_flags = (((uctxt->flags >> HFI1_CAP_MISC_SHIFT) &
1157 				HFI1_CAP_MISC_MASK) << HFI1_CAP_USER_SHIFT) |
1158 			HFI1_CAP_UGET_MASK(uctxt->flags, MASK) |
1159 			HFI1_CAP_KGET_MASK(uctxt->flags, K2U);
1160 	/* adjust flag if this fd is not able to cache */
1161 	if (!fd->handler)
1162 		cinfo.runtime_flags |= HFI1_CAP_TID_UNMAP; /* no caching */
1163 
1164 	cinfo.num_active = hfi1_count_active_units();
1165 	cinfo.unit = uctxt->dd->unit;
1166 	cinfo.ctxt = uctxt->ctxt;
1167 	cinfo.subctxt = fd->subctxt;
1168 	cinfo.rcvtids = roundup(uctxt->egrbufs.alloced,
1169 				uctxt->dd->rcv_entries.group_size) +
1170 		uctxt->expected_count;
1171 	cinfo.credits = uctxt->sc->credits;
1172 	cinfo.numa_node = uctxt->numa_id;
1173 	cinfo.rec_cpu = fd->rec_cpu_num;
1174 	cinfo.send_ctxt = uctxt->sc->hw_context;
1175 
1176 	cinfo.egrtids = uctxt->egrbufs.alloced;
1177 	cinfo.rcvhdrq_cnt = uctxt->rcvhdrq_cnt;
1178 	cinfo.rcvhdrq_entsize = uctxt->rcvhdrqentsize << 2;
1179 	cinfo.sdma_ring_size = fd->cq->nentries;
1180 	cinfo.rcvegr_size = uctxt->egrbufs.rcvtid_size;
1181 
1182 	trace_hfi1_ctxt_info(uctxt->dd, uctxt->ctxt, fd->subctxt, cinfo);
1183 	if (copy_to_user(ubase, &cinfo, sizeof(cinfo)))
1184 		ret = -EFAULT;
1185 
1186 	return ret;
1187 }
1188 
1189 static int setup_ctxt(struct file *fp)
1190 {
1191 	struct hfi1_filedata *fd = fp->private_data;
1192 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1193 	struct hfi1_devdata *dd = uctxt->dd;
1194 	int ret = 0;
1195 
1196 	/*
1197 	 * Context should be set up only once, including allocation and
1198 	 * programming of eager buffers. This is done if context sharing
1199 	 * is not requested or by the master process.
1200 	 */
1201 	if (!uctxt->subctxt_cnt || !fd->subctxt) {
1202 		ret = hfi1_init_ctxt(uctxt->sc);
1203 		if (ret)
1204 			goto done;
1205 
1206 		/* Now allocate the RcvHdr queue and eager buffers. */
1207 		ret = hfi1_create_rcvhdrq(dd, uctxt);
1208 		if (ret)
1209 			goto done;
1210 		ret = hfi1_setup_eagerbufs(uctxt);
1211 		if (ret)
1212 			goto done;
1213 		if (uctxt->subctxt_cnt && !fd->subctxt) {
1214 			ret = setup_subctxt(uctxt);
1215 			if (ret)
1216 				goto done;
1217 		}
1218 	} else {
1219 		ret = wait_event_interruptible(uctxt->wait, !test_bit(
1220 					       HFI1_CTXT_MASTER_UNINIT,
1221 					       &uctxt->event_flags));
1222 		if (ret)
1223 			goto done;
1224 	}
1225 
1226 	ret = hfi1_user_sdma_alloc_queues(uctxt, fp);
1227 	if (ret)
1228 		goto done;
1229 	/*
1230 	 * Expected receive has to be setup for all processes (including
1231 	 * shared contexts). However, it has to be done after the master
1232 	 * context has been fully configured as it depends on the
1233 	 * eager/expected split of the RcvArray entries.
1234 	 * Setting it up here ensures that the subcontexts will be waiting
1235 	 * (due to the above wait_event_interruptible() until the master
1236 	 * is setup.
1237 	 */
1238 	ret = hfi1_user_exp_rcv_init(fp);
1239 	if (ret)
1240 		goto done;
1241 
1242 	set_bit(HFI1_CTXT_SETUP_DONE, &uctxt->event_flags);
1243 done:
1244 	return ret;
1245 }
1246 
1247 static int get_base_info(struct file *fp, void __user *ubase, __u32 len)
1248 {
1249 	struct hfi1_base_info binfo;
1250 	struct hfi1_filedata *fd = fp->private_data;
1251 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1252 	struct hfi1_devdata *dd = uctxt->dd;
1253 	ssize_t sz;
1254 	unsigned offset;
1255 	int ret = 0;
1256 
1257 	trace_hfi1_uctxtdata(uctxt->dd, uctxt);
1258 
1259 	memset(&binfo, 0, sizeof(binfo));
1260 	binfo.hw_version = dd->revision;
1261 	binfo.sw_version = HFI1_KERN_SWVERSION;
1262 	binfo.bthqp = kdeth_qp;
1263 	binfo.jkey = uctxt->jkey;
1264 	/*
1265 	 * If more than 64 contexts are enabled the allocated credit
1266 	 * return will span two or three contiguous pages. Since we only
1267 	 * map the page containing the context's credit return address,
1268 	 * we need to calculate the offset in the proper page.
1269 	 */
1270 	offset = ((u64)uctxt->sc->hw_free -
1271 		  (u64)dd->cr_base[uctxt->numa_id].va) % PAGE_SIZE;
1272 	binfo.sc_credits_addr = HFI1_MMAP_TOKEN(PIO_CRED, uctxt->ctxt,
1273 						fd->subctxt, offset);
1274 	binfo.pio_bufbase = HFI1_MMAP_TOKEN(PIO_BUFS, uctxt->ctxt,
1275 					    fd->subctxt,
1276 					    uctxt->sc->base_addr);
1277 	binfo.pio_bufbase_sop = HFI1_MMAP_TOKEN(PIO_BUFS_SOP,
1278 						uctxt->ctxt,
1279 						fd->subctxt,
1280 						uctxt->sc->base_addr);
1281 	binfo.rcvhdr_bufbase = HFI1_MMAP_TOKEN(RCV_HDRQ, uctxt->ctxt,
1282 					       fd->subctxt,
1283 					       uctxt->rcvhdrq);
1284 	binfo.rcvegr_bufbase = HFI1_MMAP_TOKEN(RCV_EGRBUF, uctxt->ctxt,
1285 					       fd->subctxt,
1286 					       uctxt->egrbufs.rcvtids[0].dma);
1287 	binfo.sdma_comp_bufbase = HFI1_MMAP_TOKEN(SDMA_COMP, uctxt->ctxt,
1288 						 fd->subctxt, 0);
1289 	/*
1290 	 * user regs are at
1291 	 * (RXE_PER_CONTEXT_USER + (ctxt * RXE_PER_CONTEXT_SIZE))
1292 	 */
1293 	binfo.user_regbase = HFI1_MMAP_TOKEN(UREGS, uctxt->ctxt,
1294 					    fd->subctxt, 0);
1295 	offset = offset_in_page((((uctxt->ctxt - dd->first_user_ctxt) *
1296 		    HFI1_MAX_SHARED_CTXTS) + fd->subctxt) *
1297 		  sizeof(*dd->events));
1298 	binfo.events_bufbase = HFI1_MMAP_TOKEN(EVENTS, uctxt->ctxt,
1299 					      fd->subctxt,
1300 					      offset);
1301 	binfo.status_bufbase = HFI1_MMAP_TOKEN(STATUS, uctxt->ctxt,
1302 					      fd->subctxt,
1303 					      dd->status);
1304 	if (HFI1_CAP_IS_USET(DMA_RTAIL))
1305 		binfo.rcvhdrtail_base = HFI1_MMAP_TOKEN(RTAIL, uctxt->ctxt,
1306 						       fd->subctxt, 0);
1307 	if (uctxt->subctxt_cnt) {
1308 		binfo.subctxt_uregbase = HFI1_MMAP_TOKEN(SUBCTXT_UREGS,
1309 							uctxt->ctxt,
1310 							fd->subctxt, 0);
1311 		binfo.subctxt_rcvhdrbuf = HFI1_MMAP_TOKEN(SUBCTXT_RCV_HDRQ,
1312 							 uctxt->ctxt,
1313 							 fd->subctxt, 0);
1314 		binfo.subctxt_rcvegrbuf = HFI1_MMAP_TOKEN(SUBCTXT_EGRBUF,
1315 							 uctxt->ctxt,
1316 							 fd->subctxt, 0);
1317 	}
1318 	sz = (len < sizeof(binfo)) ? len : sizeof(binfo);
1319 	if (copy_to_user(ubase, &binfo, sz))
1320 		ret = -EFAULT;
1321 	return ret;
1322 }
1323 
1324 static unsigned int poll_urgent(struct file *fp,
1325 				struct poll_table_struct *pt)
1326 {
1327 	struct hfi1_filedata *fd = fp->private_data;
1328 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1329 	struct hfi1_devdata *dd = uctxt->dd;
1330 	unsigned pollflag;
1331 
1332 	poll_wait(fp, &uctxt->wait, pt);
1333 
1334 	spin_lock_irq(&dd->uctxt_lock);
1335 	if (uctxt->urgent != uctxt->urgent_poll) {
1336 		pollflag = POLLIN | POLLRDNORM;
1337 		uctxt->urgent_poll = uctxt->urgent;
1338 	} else {
1339 		pollflag = 0;
1340 		set_bit(HFI1_CTXT_WAITING_URG, &uctxt->event_flags);
1341 	}
1342 	spin_unlock_irq(&dd->uctxt_lock);
1343 
1344 	return pollflag;
1345 }
1346 
1347 static unsigned int poll_next(struct file *fp,
1348 			      struct poll_table_struct *pt)
1349 {
1350 	struct hfi1_filedata *fd = fp->private_data;
1351 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1352 	struct hfi1_devdata *dd = uctxt->dd;
1353 	unsigned pollflag;
1354 
1355 	poll_wait(fp, &uctxt->wait, pt);
1356 
1357 	spin_lock_irq(&dd->uctxt_lock);
1358 	if (hdrqempty(uctxt)) {
1359 		set_bit(HFI1_CTXT_WAITING_RCV, &uctxt->event_flags);
1360 		hfi1_rcvctrl(dd, HFI1_RCVCTRL_INTRAVAIL_ENB, uctxt->ctxt);
1361 		pollflag = 0;
1362 	} else {
1363 		pollflag = POLLIN | POLLRDNORM;
1364 	}
1365 	spin_unlock_irq(&dd->uctxt_lock);
1366 
1367 	return pollflag;
1368 }
1369 
1370 /*
1371  * Find all user contexts in use, and set the specified bit in their
1372  * event mask.
1373  * See also find_ctxt() for a similar use, that is specific to send buffers.
1374  */
1375 int hfi1_set_uevent_bits(struct hfi1_pportdata *ppd, const int evtbit)
1376 {
1377 	struct hfi1_ctxtdata *uctxt;
1378 	struct hfi1_devdata *dd = ppd->dd;
1379 	unsigned ctxt;
1380 	int ret = 0;
1381 	unsigned long flags;
1382 
1383 	if (!dd->events) {
1384 		ret = -EINVAL;
1385 		goto done;
1386 	}
1387 
1388 	spin_lock_irqsave(&dd->uctxt_lock, flags);
1389 	for (ctxt = dd->first_user_ctxt; ctxt < dd->num_rcv_contexts;
1390 	     ctxt++) {
1391 		uctxt = dd->rcd[ctxt];
1392 		if (uctxt) {
1393 			unsigned long *evs = dd->events +
1394 				(uctxt->ctxt - dd->first_user_ctxt) *
1395 				HFI1_MAX_SHARED_CTXTS;
1396 			int i;
1397 			/*
1398 			 * subctxt_cnt is 0 if not shared, so do base
1399 			 * separately, first, then remaining subctxt, if any
1400 			 */
1401 			set_bit(evtbit, evs);
1402 			for (i = 1; i < uctxt->subctxt_cnt; i++)
1403 				set_bit(evtbit, evs + i);
1404 		}
1405 	}
1406 	spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1407 done:
1408 	return ret;
1409 }
1410 
1411 /**
1412  * manage_rcvq - manage a context's receive queue
1413  * @uctxt: the context
1414  * @subctxt: the sub-context
1415  * @start_stop: action to carry out
1416  *
1417  * start_stop == 0 disables receive on the context, for use in queue
1418  * overflow conditions.  start_stop==1 re-enables, to be used to
1419  * re-init the software copy of the head register
1420  */
1421 static int manage_rcvq(struct hfi1_ctxtdata *uctxt, unsigned subctxt,
1422 		       int start_stop)
1423 {
1424 	struct hfi1_devdata *dd = uctxt->dd;
1425 	unsigned int rcvctrl_op;
1426 
1427 	if (subctxt)
1428 		goto bail;
1429 	/* atomically clear receive enable ctxt. */
1430 	if (start_stop) {
1431 		/*
1432 		 * On enable, force in-memory copy of the tail register to
1433 		 * 0, so that protocol code doesn't have to worry about
1434 		 * whether or not the chip has yet updated the in-memory
1435 		 * copy or not on return from the system call. The chip
1436 		 * always resets it's tail register back to 0 on a
1437 		 * transition from disabled to enabled.
1438 		 */
1439 		if (uctxt->rcvhdrtail_kvaddr)
1440 			clear_rcvhdrtail(uctxt);
1441 		rcvctrl_op = HFI1_RCVCTRL_CTXT_ENB;
1442 	} else {
1443 		rcvctrl_op = HFI1_RCVCTRL_CTXT_DIS;
1444 	}
1445 	hfi1_rcvctrl(dd, rcvctrl_op, uctxt->ctxt);
1446 	/* always; new head should be equal to new tail; see above */
1447 bail:
1448 	return 0;
1449 }
1450 
1451 /*
1452  * clear the event notifier events for this context.
1453  * User process then performs actions appropriate to bit having been
1454  * set, if desired, and checks again in future.
1455  */
1456 static int user_event_ack(struct hfi1_ctxtdata *uctxt, int subctxt,
1457 			  unsigned long events)
1458 {
1459 	int i;
1460 	struct hfi1_devdata *dd = uctxt->dd;
1461 	unsigned long *evs;
1462 
1463 	if (!dd->events)
1464 		return 0;
1465 
1466 	evs = dd->events + ((uctxt->ctxt - dd->first_user_ctxt) *
1467 			    HFI1_MAX_SHARED_CTXTS) + subctxt;
1468 
1469 	for (i = 0; i <= _HFI1_MAX_EVENT_BIT; i++) {
1470 		if (!test_bit(i, &events))
1471 			continue;
1472 		clear_bit(i, evs);
1473 	}
1474 	return 0;
1475 }
1476 
1477 static int set_ctxt_pkey(struct hfi1_ctxtdata *uctxt, unsigned subctxt,
1478 			 u16 pkey)
1479 {
1480 	int ret = -ENOENT, i, intable = 0;
1481 	struct hfi1_pportdata *ppd = uctxt->ppd;
1482 	struct hfi1_devdata *dd = uctxt->dd;
1483 
1484 	if (pkey == LIM_MGMT_P_KEY || pkey == FULL_MGMT_P_KEY) {
1485 		ret = -EINVAL;
1486 		goto done;
1487 	}
1488 
1489 	for (i = 0; i < ARRAY_SIZE(ppd->pkeys); i++)
1490 		if (pkey == ppd->pkeys[i]) {
1491 			intable = 1;
1492 			break;
1493 		}
1494 
1495 	if (intable)
1496 		ret = hfi1_set_ctxt_pkey(dd, uctxt->ctxt, pkey);
1497 done:
1498 	return ret;
1499 }
1500 
1501 static void user_remove(struct hfi1_devdata *dd)
1502 {
1503 
1504 	hfi1_cdev_cleanup(&dd->user_cdev, &dd->user_device);
1505 }
1506 
1507 static int user_add(struct hfi1_devdata *dd)
1508 {
1509 	char name[10];
1510 	int ret;
1511 
1512 	snprintf(name, sizeof(name), "%s_%d", class_name(), dd->unit);
1513 	ret = hfi1_cdev_init(dd->unit, name, &hfi1_file_ops,
1514 			     &dd->user_cdev, &dd->user_device,
1515 			     true, &dd->kobj);
1516 	if (ret)
1517 		user_remove(dd);
1518 
1519 	return ret;
1520 }
1521 
1522 /*
1523  * Create per-unit files in /dev
1524  */
1525 int hfi1_device_create(struct hfi1_devdata *dd)
1526 {
1527 	return user_add(dd);
1528 }
1529 
1530 /*
1531  * Remove per-unit files in /dev
1532  * void, core kernel returns no errors for this stuff
1533  */
1534 void hfi1_device_remove(struct hfi1_devdata *dd)
1535 {
1536 	user_remove(dd);
1537 }
1538