1 /*
2  * Copyright(c) 2015-2017 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 #include <linux/sched/mm.h>
52 #include <linux/bitmap.h>
53 
54 #include <rdma/ib.h>
55 
56 #include "hfi.h"
57 #include "pio.h"
58 #include "device.h"
59 #include "common.h"
60 #include "trace.h"
61 #include "mmu_rb.h"
62 #include "user_sdma.h"
63 #include "user_exp_rcv.h"
64 #include "aspm.h"
65 
66 #undef pr_fmt
67 #define pr_fmt(fmt) DRIVER_NAME ": " fmt
68 
69 #define SEND_CTXT_HALT_TIMEOUT 1000 /* msecs */
70 
71 /*
72  * File operation functions
73  */
74 static int hfi1_file_open(struct inode *inode, struct file *fp);
75 static int hfi1_file_close(struct inode *inode, struct file *fp);
76 static ssize_t hfi1_write_iter(struct kiocb *kiocb, struct iov_iter *from);
77 static unsigned int hfi1_poll(struct file *fp, struct poll_table_struct *pt);
78 static int hfi1_file_mmap(struct file *fp, struct vm_area_struct *vma);
79 
80 static u64 kvirt_to_phys(void *addr);
81 static int assign_ctxt(struct hfi1_filedata *fd, struct hfi1_user_info *uinfo);
82 static void init_subctxts(struct hfi1_ctxtdata *uctxt,
83 			  const struct hfi1_user_info *uinfo);
84 static int init_user_ctxt(struct hfi1_filedata *fd,
85 			  struct hfi1_ctxtdata *uctxt);
86 static void user_init(struct hfi1_ctxtdata *uctxt);
87 static int get_ctxt_info(struct hfi1_filedata *fd, void __user *ubase,
88 			 __u32 len);
89 static int get_base_info(struct hfi1_filedata *fd, void __user *ubase,
90 			 __u32 len);
91 static int setup_base_ctxt(struct hfi1_filedata *fd,
92 			   struct hfi1_ctxtdata *uctxt);
93 static int setup_subctxt(struct hfi1_ctxtdata *uctxt);
94 
95 static int find_sub_ctxt(struct hfi1_filedata *fd,
96 			 const struct hfi1_user_info *uinfo);
97 static int allocate_ctxt(struct hfi1_filedata *fd, struct hfi1_devdata *dd,
98 			 struct hfi1_user_info *uinfo,
99 			 struct hfi1_ctxtdata **cd);
100 static void deallocate_ctxt(struct hfi1_ctxtdata *uctxt);
101 static unsigned int poll_urgent(struct file *fp, struct poll_table_struct *pt);
102 static unsigned int poll_next(struct file *fp, struct poll_table_struct *pt);
103 static int user_event_ack(struct hfi1_ctxtdata *uctxt, u16 subctxt,
104 			  unsigned long events);
105 static int set_ctxt_pkey(struct hfi1_ctxtdata *uctxt, u16 subctxt, u16 pkey);
106 static int manage_rcvq(struct hfi1_ctxtdata *uctxt, u16 subctxt,
107 		       int start_stop);
108 static int vma_fault(struct vm_fault *vmf);
109 static long hfi1_file_ioctl(struct file *fp, unsigned int cmd,
110 			    unsigned long arg);
111 
112 static const struct file_operations hfi1_file_ops = {
113 	.owner = THIS_MODULE,
114 	.write_iter = hfi1_write_iter,
115 	.open = hfi1_file_open,
116 	.release = hfi1_file_close,
117 	.unlocked_ioctl = hfi1_file_ioctl,
118 	.poll = hfi1_poll,
119 	.mmap = hfi1_file_mmap,
120 	.llseek = noop_llseek,
121 };
122 
123 static const struct vm_operations_struct vm_ops = {
124 	.fault = vma_fault,
125 };
126 
127 /*
128  * Types of memories mapped into user processes' space
129  */
130 enum mmap_types {
131 	PIO_BUFS = 1,
132 	PIO_BUFS_SOP,
133 	PIO_CRED,
134 	RCV_HDRQ,
135 	RCV_EGRBUF,
136 	UREGS,
137 	EVENTS,
138 	STATUS,
139 	RTAIL,
140 	SUBCTXT_UREGS,
141 	SUBCTXT_RCV_HDRQ,
142 	SUBCTXT_EGRBUF,
143 	SDMA_COMP
144 };
145 
146 /*
147  * Masks and offsets defining the mmap tokens
148  */
149 #define HFI1_MMAP_OFFSET_MASK   0xfffULL
150 #define HFI1_MMAP_OFFSET_SHIFT  0
151 #define HFI1_MMAP_SUBCTXT_MASK  0xfULL
152 #define HFI1_MMAP_SUBCTXT_SHIFT 12
153 #define HFI1_MMAP_CTXT_MASK     0xffULL
154 #define HFI1_MMAP_CTXT_SHIFT    16
155 #define HFI1_MMAP_TYPE_MASK     0xfULL
156 #define HFI1_MMAP_TYPE_SHIFT    24
157 #define HFI1_MMAP_MAGIC_MASK    0xffffffffULL
158 #define HFI1_MMAP_MAGIC_SHIFT   32
159 
160 #define HFI1_MMAP_MAGIC         0xdabbad00
161 
162 #define HFI1_MMAP_TOKEN_SET(field, val)	\
163 	(((val) & HFI1_MMAP_##field##_MASK) << HFI1_MMAP_##field##_SHIFT)
164 #define HFI1_MMAP_TOKEN_GET(field, token) \
165 	(((token) >> HFI1_MMAP_##field##_SHIFT) & HFI1_MMAP_##field##_MASK)
166 #define HFI1_MMAP_TOKEN(type, ctxt, subctxt, addr)   \
167 	(HFI1_MMAP_TOKEN_SET(MAGIC, HFI1_MMAP_MAGIC) | \
168 	HFI1_MMAP_TOKEN_SET(TYPE, type) | \
169 	HFI1_MMAP_TOKEN_SET(CTXT, ctxt) | \
170 	HFI1_MMAP_TOKEN_SET(SUBCTXT, subctxt) | \
171 	HFI1_MMAP_TOKEN_SET(OFFSET, (offset_in_page(addr))))
172 
173 #define dbg(fmt, ...)				\
174 	pr_info(fmt, ##__VA_ARGS__)
175 
176 static inline int is_valid_mmap(u64 token)
177 {
178 	return (HFI1_MMAP_TOKEN_GET(MAGIC, token) == HFI1_MMAP_MAGIC);
179 }
180 
181 static int hfi1_file_open(struct inode *inode, struct file *fp)
182 {
183 	struct hfi1_filedata *fd;
184 	struct hfi1_devdata *dd = container_of(inode->i_cdev,
185 					       struct hfi1_devdata,
186 					       user_cdev);
187 
188 	if (!((dd->flags & HFI1_PRESENT) && dd->kregbase1))
189 		return -EINVAL;
190 
191 	if (!atomic_inc_not_zero(&dd->user_refcount))
192 		return -ENXIO;
193 
194 	/* Just take a ref now. Not all opens result in a context assign */
195 	kobject_get(&dd->kobj);
196 
197 	/* The real work is performed later in assign_ctxt() */
198 
199 	fd = kzalloc(sizeof(*fd), GFP_KERNEL);
200 
201 	if (fd) {
202 		fd->rec_cpu_num = -1; /* no cpu affinity by default */
203 		fd->mm = current->mm;
204 		mmgrab(fd->mm);
205 		fd->dd = dd;
206 		fp->private_data = fd;
207 	} else {
208 		fp->private_data = NULL;
209 
210 		if (atomic_dec_and_test(&dd->user_refcount))
211 			complete(&dd->user_comp);
212 
213 		return -ENOMEM;
214 	}
215 
216 	return 0;
217 }
218 
219 static long hfi1_file_ioctl(struct file *fp, unsigned int cmd,
220 			    unsigned long arg)
221 {
222 	struct hfi1_filedata *fd = fp->private_data;
223 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
224 	struct hfi1_user_info uinfo;
225 	struct hfi1_tid_info tinfo;
226 	int ret = 0;
227 	unsigned long addr;
228 	int uval = 0;
229 	unsigned long ul_uval = 0;
230 	u16 uval16 = 0;
231 
232 	hfi1_cdbg(IOCTL, "IOCTL recv: 0x%x", cmd);
233 	if (cmd != HFI1_IOCTL_ASSIGN_CTXT &&
234 	    cmd != HFI1_IOCTL_GET_VERS &&
235 	    !uctxt)
236 		return -EINVAL;
237 
238 	switch (cmd) {
239 	case HFI1_IOCTL_ASSIGN_CTXT:
240 		if (uctxt)
241 			return -EINVAL;
242 
243 		if (copy_from_user(&uinfo,
244 				   (struct hfi1_user_info __user *)arg,
245 				   sizeof(uinfo)))
246 			return -EFAULT;
247 
248 		ret = assign_ctxt(fd, &uinfo);
249 		break;
250 	case HFI1_IOCTL_CTXT_INFO:
251 		ret = get_ctxt_info(fd, (void __user *)(unsigned long)arg,
252 				    sizeof(struct hfi1_ctxt_info));
253 		break;
254 	case HFI1_IOCTL_USER_INFO:
255 		ret = get_base_info(fd, (void __user *)(unsigned long)arg,
256 				    sizeof(struct hfi1_base_info));
257 		break;
258 	case HFI1_IOCTL_CREDIT_UPD:
259 		if (uctxt)
260 			sc_return_credits(uctxt->sc);
261 		break;
262 
263 	case HFI1_IOCTL_TID_UPDATE:
264 		if (copy_from_user(&tinfo,
265 				   (struct hfi11_tid_info __user *)arg,
266 				   sizeof(tinfo)))
267 			return -EFAULT;
268 
269 		ret = hfi1_user_exp_rcv_setup(fd, &tinfo);
270 		if (!ret) {
271 			/*
272 			 * Copy the number of tidlist entries we used
273 			 * and the length of the buffer we registered.
274 			 */
275 			addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
276 			if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
277 					 sizeof(tinfo.tidcnt)))
278 				return -EFAULT;
279 
280 			addr = arg + offsetof(struct hfi1_tid_info, length);
281 			if (copy_to_user((void __user *)addr, &tinfo.length,
282 					 sizeof(tinfo.length)))
283 				ret = -EFAULT;
284 		}
285 		break;
286 
287 	case HFI1_IOCTL_TID_FREE:
288 		if (copy_from_user(&tinfo,
289 				   (struct hfi11_tid_info __user *)arg,
290 				   sizeof(tinfo)))
291 			return -EFAULT;
292 
293 		ret = hfi1_user_exp_rcv_clear(fd, &tinfo);
294 		if (ret)
295 			break;
296 		addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
297 		if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
298 				 sizeof(tinfo.tidcnt)))
299 			ret = -EFAULT;
300 		break;
301 
302 	case HFI1_IOCTL_TID_INVAL_READ:
303 		if (copy_from_user(&tinfo,
304 				   (struct hfi11_tid_info __user *)arg,
305 				   sizeof(tinfo)))
306 			return -EFAULT;
307 
308 		ret = hfi1_user_exp_rcv_invalid(fd, &tinfo);
309 		if (ret)
310 			break;
311 		addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
312 		if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
313 				 sizeof(tinfo.tidcnt)))
314 			ret = -EFAULT;
315 		break;
316 
317 	case HFI1_IOCTL_RECV_CTRL:
318 		ret = get_user(uval, (int __user *)arg);
319 		if (ret != 0)
320 			return -EFAULT;
321 		ret = manage_rcvq(uctxt, fd->subctxt, uval);
322 		break;
323 
324 	case HFI1_IOCTL_POLL_TYPE:
325 		ret = get_user(uval, (int __user *)arg);
326 		if (ret != 0)
327 			return -EFAULT;
328 		uctxt->poll_type = (typeof(uctxt->poll_type))uval;
329 		break;
330 
331 	case HFI1_IOCTL_ACK_EVENT:
332 		ret = get_user(ul_uval, (unsigned long __user *)arg);
333 		if (ret != 0)
334 			return -EFAULT;
335 		ret = user_event_ack(uctxt, fd->subctxt, ul_uval);
336 		break;
337 
338 	case HFI1_IOCTL_SET_PKEY:
339 		ret = get_user(uval16, (u16 __user *)arg);
340 		if (ret != 0)
341 			return -EFAULT;
342 		if (HFI1_CAP_IS_USET(PKEY_CHECK))
343 			ret = set_ctxt_pkey(uctxt, fd->subctxt, uval16);
344 		else
345 			return -EPERM;
346 		break;
347 
348 	case HFI1_IOCTL_CTXT_RESET: {
349 		struct send_context *sc;
350 		struct hfi1_devdata *dd;
351 
352 		if (!uctxt || !uctxt->dd || !uctxt->sc)
353 			return -EINVAL;
354 
355 		/*
356 		 * There is no protection here. User level has to
357 		 * guarantee that no one will be writing to the send
358 		 * context while it is being re-initialized.
359 		 * If user level breaks that guarantee, it will break
360 		 * it's own context and no one else's.
361 		 */
362 		dd = uctxt->dd;
363 		sc = uctxt->sc;
364 		/*
365 		 * Wait until the interrupt handler has marked the
366 		 * context as halted or frozen. Report error if we time
367 		 * out.
368 		 */
369 		wait_event_interruptible_timeout(
370 			sc->halt_wait, (sc->flags & SCF_HALTED),
371 			msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
372 		if (!(sc->flags & SCF_HALTED))
373 			return -ENOLCK;
374 
375 		/*
376 		 * If the send context was halted due to a Freeze,
377 		 * wait until the device has been "unfrozen" before
378 		 * resetting the context.
379 		 */
380 		if (sc->flags & SCF_FROZEN) {
381 			wait_event_interruptible_timeout(
382 				dd->event_queue,
383 				!(ACCESS_ONCE(dd->flags) & HFI1_FROZEN),
384 				msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
385 			if (dd->flags & HFI1_FROZEN)
386 				return -ENOLCK;
387 
388 			if (dd->flags & HFI1_FORCED_FREEZE)
389 				/*
390 				 * Don't allow context reset if we are into
391 				 * forced freeze
392 				 */
393 				return -ENODEV;
394 
395 			sc_disable(sc);
396 			ret = sc_enable(sc);
397 			hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_ENB, uctxt);
398 		} else {
399 			ret = sc_restart(sc);
400 		}
401 		if (!ret)
402 			sc_return_credits(sc);
403 		break;
404 	}
405 
406 	case HFI1_IOCTL_GET_VERS:
407 		uval = HFI1_USER_SWVERSION;
408 		if (put_user(uval, (int __user *)arg))
409 			return -EFAULT;
410 		break;
411 
412 	default:
413 		return -EINVAL;
414 	}
415 
416 	return ret;
417 }
418 
419 static ssize_t hfi1_write_iter(struct kiocb *kiocb, struct iov_iter *from)
420 {
421 	struct hfi1_filedata *fd = kiocb->ki_filp->private_data;
422 	struct hfi1_user_sdma_pkt_q *pq = fd->pq;
423 	struct hfi1_user_sdma_comp_q *cq = fd->cq;
424 	int done = 0, reqs = 0;
425 	unsigned long dim = from->nr_segs;
426 
427 	if (!cq || !pq)
428 		return -EIO;
429 
430 	if (!iter_is_iovec(from) || !dim)
431 		return -EINVAL;
432 
433 	trace_hfi1_sdma_request(fd->dd, fd->uctxt->ctxt, fd->subctxt, dim);
434 
435 	if (atomic_read(&pq->n_reqs) == pq->n_max_reqs)
436 		return -ENOSPC;
437 
438 	while (dim) {
439 		int ret;
440 		unsigned long count = 0;
441 
442 		ret = hfi1_user_sdma_process_request(
443 			fd, (struct iovec *)(from->iov + done),
444 			dim, &count);
445 		if (ret) {
446 			reqs = ret;
447 			break;
448 		}
449 		dim -= count;
450 		done += count;
451 		reqs++;
452 	}
453 
454 	return reqs;
455 }
456 
457 static int hfi1_file_mmap(struct file *fp, struct vm_area_struct *vma)
458 {
459 	struct hfi1_filedata *fd = fp->private_data;
460 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
461 	struct hfi1_devdata *dd;
462 	unsigned long flags;
463 	u64 token = vma->vm_pgoff << PAGE_SHIFT,
464 		memaddr = 0;
465 	void *memvirt = NULL;
466 	u8 subctxt, mapio = 0, vmf = 0, type;
467 	ssize_t memlen = 0;
468 	int ret = 0;
469 	u16 ctxt;
470 
471 	if (!is_valid_mmap(token) || !uctxt ||
472 	    !(vma->vm_flags & VM_SHARED)) {
473 		ret = -EINVAL;
474 		goto done;
475 	}
476 	dd = uctxt->dd;
477 	ctxt = HFI1_MMAP_TOKEN_GET(CTXT, token);
478 	subctxt = HFI1_MMAP_TOKEN_GET(SUBCTXT, token);
479 	type = HFI1_MMAP_TOKEN_GET(TYPE, token);
480 	if (ctxt != uctxt->ctxt || subctxt != fd->subctxt) {
481 		ret = -EINVAL;
482 		goto done;
483 	}
484 
485 	flags = vma->vm_flags;
486 
487 	switch (type) {
488 	case PIO_BUFS:
489 	case PIO_BUFS_SOP:
490 		memaddr = ((dd->physaddr + TXE_PIO_SEND) +
491 				/* chip pio base */
492 			   (uctxt->sc->hw_context * BIT(16))) +
493 				/* 64K PIO space / ctxt */
494 			(type == PIO_BUFS_SOP ?
495 				(TXE_PIO_SIZE / 2) : 0); /* sop? */
496 		/*
497 		 * Map only the amount allocated to the context, not the
498 		 * entire available context's PIO space.
499 		 */
500 		memlen = PAGE_ALIGN(uctxt->sc->credits * PIO_BLOCK_SIZE);
501 		flags &= ~VM_MAYREAD;
502 		flags |= VM_DONTCOPY | VM_DONTEXPAND;
503 		vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
504 		mapio = 1;
505 		break;
506 	case PIO_CRED:
507 		if (flags & VM_WRITE) {
508 			ret = -EPERM;
509 			goto done;
510 		}
511 		/*
512 		 * The credit return location for this context could be on the
513 		 * second or third page allocated for credit returns (if number
514 		 * of enabled contexts > 64 and 128 respectively).
515 		 */
516 		memvirt = dd->cr_base[uctxt->numa_id].va;
517 		memaddr = virt_to_phys(memvirt) +
518 			(((u64)uctxt->sc->hw_free -
519 			  (u64)dd->cr_base[uctxt->numa_id].va) & PAGE_MASK);
520 		memlen = PAGE_SIZE;
521 		flags &= ~VM_MAYWRITE;
522 		flags |= VM_DONTCOPY | VM_DONTEXPAND;
523 		/*
524 		 * The driver has already allocated memory for credit
525 		 * returns and programmed it into the chip. Has that
526 		 * memory been flagged as non-cached?
527 		 */
528 		/* vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); */
529 		mapio = 1;
530 		break;
531 	case RCV_HDRQ:
532 		memlen = uctxt->rcvhdrq_size;
533 		memvirt = uctxt->rcvhdrq;
534 		break;
535 	case RCV_EGRBUF: {
536 		unsigned long addr;
537 		int i;
538 		/*
539 		 * The RcvEgr buffer need to be handled differently
540 		 * as multiple non-contiguous pages need to be mapped
541 		 * into the user process.
542 		 */
543 		memlen = uctxt->egrbufs.size;
544 		if ((vma->vm_end - vma->vm_start) != memlen) {
545 			dd_dev_err(dd, "Eager buffer map size invalid (%lu != %lu)\n",
546 				   (vma->vm_end - vma->vm_start), memlen);
547 			ret = -EINVAL;
548 			goto done;
549 		}
550 		if (vma->vm_flags & VM_WRITE) {
551 			ret = -EPERM;
552 			goto done;
553 		}
554 		vma->vm_flags &= ~VM_MAYWRITE;
555 		addr = vma->vm_start;
556 		for (i = 0 ; i < uctxt->egrbufs.numbufs; i++) {
557 			memlen = uctxt->egrbufs.buffers[i].len;
558 			memvirt = uctxt->egrbufs.buffers[i].addr;
559 			ret = remap_pfn_range(
560 				vma, addr,
561 				/*
562 				 * virt_to_pfn() does the same, but
563 				 * it's not available on x86_64
564 				 * when CONFIG_MMU is enabled.
565 				 */
566 				PFN_DOWN(__pa(memvirt)),
567 				memlen,
568 				vma->vm_page_prot);
569 			if (ret < 0)
570 				goto done;
571 			addr += memlen;
572 		}
573 		ret = 0;
574 		goto done;
575 	}
576 	case UREGS:
577 		/*
578 		 * Map only the page that contains this context's user
579 		 * registers.
580 		 */
581 		memaddr = (unsigned long)
582 			(dd->physaddr + RXE_PER_CONTEXT_USER)
583 			+ (uctxt->ctxt * RXE_PER_CONTEXT_SIZE);
584 		/*
585 		 * TidFlow table is on the same page as the rest of the
586 		 * user registers.
587 		 */
588 		memlen = PAGE_SIZE;
589 		flags |= VM_DONTCOPY | VM_DONTEXPAND;
590 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
591 		mapio = 1;
592 		break;
593 	case EVENTS:
594 		/*
595 		 * Use the page where this context's flags are. User level
596 		 * knows where it's own bitmap is within the page.
597 		 */
598 		memaddr = (unsigned long)(dd->events +
599 				  ((uctxt->ctxt - dd->first_dyn_alloc_ctxt) *
600 				   HFI1_MAX_SHARED_CTXTS)) & PAGE_MASK;
601 		memlen = PAGE_SIZE;
602 		/*
603 		 * v3.7 removes VM_RESERVED but the effect is kept by
604 		 * using VM_IO.
605 		 */
606 		flags |= VM_IO | VM_DONTEXPAND;
607 		vmf = 1;
608 		break;
609 	case STATUS:
610 		if (flags & (unsigned long)(VM_WRITE | VM_EXEC)) {
611 			ret = -EPERM;
612 			goto done;
613 		}
614 		memaddr = kvirt_to_phys((void *)dd->status);
615 		memlen = PAGE_SIZE;
616 		flags |= VM_IO | VM_DONTEXPAND;
617 		break;
618 	case RTAIL:
619 		if (!HFI1_CAP_IS_USET(DMA_RTAIL)) {
620 			/*
621 			 * If the memory allocation failed, the context alloc
622 			 * also would have failed, so we would never get here
623 			 */
624 			ret = -EINVAL;
625 			goto done;
626 		}
627 		if (flags & VM_WRITE) {
628 			ret = -EPERM;
629 			goto done;
630 		}
631 		memlen = PAGE_SIZE;
632 		memvirt = (void *)uctxt->rcvhdrtail_kvaddr;
633 		flags &= ~VM_MAYWRITE;
634 		break;
635 	case SUBCTXT_UREGS:
636 		memaddr = (u64)uctxt->subctxt_uregbase;
637 		memlen = PAGE_SIZE;
638 		flags |= VM_IO | VM_DONTEXPAND;
639 		vmf = 1;
640 		break;
641 	case SUBCTXT_RCV_HDRQ:
642 		memaddr = (u64)uctxt->subctxt_rcvhdr_base;
643 		memlen = uctxt->rcvhdrq_size * uctxt->subctxt_cnt;
644 		flags |= VM_IO | VM_DONTEXPAND;
645 		vmf = 1;
646 		break;
647 	case SUBCTXT_EGRBUF:
648 		memaddr = (u64)uctxt->subctxt_rcvegrbuf;
649 		memlen = uctxt->egrbufs.size * uctxt->subctxt_cnt;
650 		flags |= VM_IO | VM_DONTEXPAND;
651 		flags &= ~VM_MAYWRITE;
652 		vmf = 1;
653 		break;
654 	case SDMA_COMP: {
655 		struct hfi1_user_sdma_comp_q *cq = fd->cq;
656 
657 		if (!cq) {
658 			ret = -EFAULT;
659 			goto done;
660 		}
661 		memaddr = (u64)cq->comps;
662 		memlen = PAGE_ALIGN(sizeof(*cq->comps) * cq->nentries);
663 		flags |= VM_IO | VM_DONTEXPAND;
664 		vmf = 1;
665 		break;
666 	}
667 	default:
668 		ret = -EINVAL;
669 		break;
670 	}
671 
672 	if ((vma->vm_end - vma->vm_start) != memlen) {
673 		hfi1_cdbg(PROC, "%u:%u Memory size mismatch %lu:%lu",
674 			  uctxt->ctxt, fd->subctxt,
675 			  (vma->vm_end - vma->vm_start), memlen);
676 		ret = -EINVAL;
677 		goto done;
678 	}
679 
680 	vma->vm_flags = flags;
681 	hfi1_cdbg(PROC,
682 		  "%u:%u type:%u io/vf:%d/%d, addr:0x%llx, len:%lu(%lu), flags:0x%lx\n",
683 		    ctxt, subctxt, type, mapio, vmf, memaddr, memlen,
684 		    vma->vm_end - vma->vm_start, vma->vm_flags);
685 	if (vmf) {
686 		vma->vm_pgoff = PFN_DOWN(memaddr);
687 		vma->vm_ops = &vm_ops;
688 		ret = 0;
689 	} else if (mapio) {
690 		ret = io_remap_pfn_range(vma, vma->vm_start,
691 					 PFN_DOWN(memaddr),
692 					 memlen,
693 					 vma->vm_page_prot);
694 	} else if (memvirt) {
695 		ret = remap_pfn_range(vma, vma->vm_start,
696 				      PFN_DOWN(__pa(memvirt)),
697 				      memlen,
698 				      vma->vm_page_prot);
699 	} else {
700 		ret = remap_pfn_range(vma, vma->vm_start,
701 				      PFN_DOWN(memaddr),
702 				      memlen,
703 				      vma->vm_page_prot);
704 	}
705 done:
706 	return ret;
707 }
708 
709 /*
710  * Local (non-chip) user memory is not mapped right away but as it is
711  * accessed by the user-level code.
712  */
713 static int vma_fault(struct vm_fault *vmf)
714 {
715 	struct page *page;
716 
717 	page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
718 	if (!page)
719 		return VM_FAULT_SIGBUS;
720 
721 	get_page(page);
722 	vmf->page = page;
723 
724 	return 0;
725 }
726 
727 static unsigned int hfi1_poll(struct file *fp, struct poll_table_struct *pt)
728 {
729 	struct hfi1_ctxtdata *uctxt;
730 	unsigned pollflag;
731 
732 	uctxt = ((struct hfi1_filedata *)fp->private_data)->uctxt;
733 	if (!uctxt)
734 		pollflag = POLLERR;
735 	else if (uctxt->poll_type == HFI1_POLL_TYPE_URGENT)
736 		pollflag = poll_urgent(fp, pt);
737 	else  if (uctxt->poll_type == HFI1_POLL_TYPE_ANYRCV)
738 		pollflag = poll_next(fp, pt);
739 	else /* invalid */
740 		pollflag = POLLERR;
741 
742 	return pollflag;
743 }
744 
745 static int hfi1_file_close(struct inode *inode, struct file *fp)
746 {
747 	struct hfi1_filedata *fdata = fp->private_data;
748 	struct hfi1_ctxtdata *uctxt = fdata->uctxt;
749 	struct hfi1_devdata *dd = container_of(inode->i_cdev,
750 					       struct hfi1_devdata,
751 					       user_cdev);
752 	unsigned long flags, *ev;
753 
754 	fp->private_data = NULL;
755 
756 	if (!uctxt)
757 		goto done;
758 
759 	hfi1_cdbg(PROC, "closing ctxt %u:%u", uctxt->ctxt, fdata->subctxt);
760 
761 	flush_wc();
762 	/* drain user sdma queue */
763 	hfi1_user_sdma_free_queues(fdata, uctxt);
764 
765 	/* release the cpu */
766 	hfi1_put_proc_affinity(fdata->rec_cpu_num);
767 
768 	/* clean up rcv side */
769 	hfi1_user_exp_rcv_free(fdata);
770 
771 	/*
772 	 * fdata->uctxt is used in the above cleanup.  It is not ready to be
773 	 * removed until here.
774 	 */
775 	fdata->uctxt = NULL;
776 	hfi1_rcd_put(uctxt);
777 
778 	/*
779 	 * Clear any left over, unhandled events so the next process that
780 	 * gets this context doesn't get confused.
781 	 */
782 	ev = dd->events + ((uctxt->ctxt - dd->first_dyn_alloc_ctxt) *
783 			   HFI1_MAX_SHARED_CTXTS) + fdata->subctxt;
784 	*ev = 0;
785 
786 	spin_lock_irqsave(&dd->uctxt_lock, flags);
787 	__clear_bit(fdata->subctxt, uctxt->in_use_ctxts);
788 	if (!bitmap_empty(uctxt->in_use_ctxts, HFI1_MAX_SHARED_CTXTS)) {
789 		spin_unlock_irqrestore(&dd->uctxt_lock, flags);
790 		goto done;
791 	}
792 	spin_unlock_irqrestore(&dd->uctxt_lock, flags);
793 
794 	/*
795 	 * Disable receive context and interrupt available, reset all
796 	 * RcvCtxtCtrl bits to default values.
797 	 */
798 	hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_DIS |
799 		     HFI1_RCVCTRL_TIDFLOW_DIS |
800 		     HFI1_RCVCTRL_INTRAVAIL_DIS |
801 		     HFI1_RCVCTRL_TAILUPD_DIS |
802 		     HFI1_RCVCTRL_ONE_PKT_EGR_DIS |
803 		     HFI1_RCVCTRL_NO_RHQ_DROP_DIS |
804 		     HFI1_RCVCTRL_NO_EGR_DROP_DIS, uctxt);
805 	/* Clear the context's J_KEY */
806 	hfi1_clear_ctxt_jkey(dd, uctxt);
807 	/*
808 	 * If a send context is allocated, reset context integrity
809 	 * checks to default and disable the send context.
810 	 */
811 	if (uctxt->sc) {
812 		set_pio_integrity(uctxt->sc);
813 		sc_disable(uctxt->sc);
814 	}
815 
816 	hfi1_free_ctxt_rcv_groups(uctxt);
817 	hfi1_clear_ctxt_pkey(dd, uctxt);
818 
819 	uctxt->event_flags = 0;
820 
821 	deallocate_ctxt(uctxt);
822 done:
823 	mmdrop(fdata->mm);
824 	kobject_put(&dd->kobj);
825 
826 	if (atomic_dec_and_test(&dd->user_refcount))
827 		complete(&dd->user_comp);
828 
829 	kfree(fdata);
830 	return 0;
831 }
832 
833 /*
834  * Convert kernel *virtual* addresses to physical addresses.
835  * This is used to vmalloc'ed addresses.
836  */
837 static u64 kvirt_to_phys(void *addr)
838 {
839 	struct page *page;
840 	u64 paddr = 0;
841 
842 	page = vmalloc_to_page(addr);
843 	if (page)
844 		paddr = page_to_pfn(page) << PAGE_SHIFT;
845 
846 	return paddr;
847 }
848 
849 /**
850  * complete_subctxt
851  * @fd: valid filedata pointer
852  *
853  * Sub-context info can only be set up after the base context
854  * has been completed.  This is indicated by the clearing of the
855  * HFI1_CTXT_BASE_UINIT bit.
856  *
857  * Wait for the bit to be cleared, and then complete the subcontext
858  * initialization.
859  *
860  */
861 static int complete_subctxt(struct hfi1_filedata *fd)
862 {
863 	int ret;
864 	unsigned long flags;
865 
866 	/*
867 	 * sub-context info can only be set up after the base context
868 	 * has been completed.
869 	 */
870 	ret = wait_event_interruptible(
871 		fd->uctxt->wait,
872 		!test_bit(HFI1_CTXT_BASE_UNINIT, &fd->uctxt->event_flags));
873 
874 	if (test_bit(HFI1_CTXT_BASE_FAILED, &fd->uctxt->event_flags))
875 		ret = -ENOMEM;
876 
877 	/* Finish the sub-context init */
878 	if (!ret) {
879 		fd->rec_cpu_num = hfi1_get_proc_affinity(fd->uctxt->numa_id);
880 		ret = init_user_ctxt(fd, fd->uctxt);
881 	}
882 
883 	if (ret) {
884 		hfi1_rcd_put(fd->uctxt);
885 		fd->uctxt = NULL;
886 		spin_lock_irqsave(&fd->dd->uctxt_lock, flags);
887 		__clear_bit(fd->subctxt, fd->uctxt->in_use_ctxts);
888 		spin_unlock_irqrestore(&fd->dd->uctxt_lock, flags);
889 	}
890 
891 	return ret;
892 }
893 
894 static int assign_ctxt(struct hfi1_filedata *fd, struct hfi1_user_info *uinfo)
895 {
896 	int ret;
897 	unsigned int swmajor, swminor;
898 	struct hfi1_ctxtdata *uctxt = NULL;
899 
900 	swmajor = uinfo->userversion >> 16;
901 	if (swmajor != HFI1_USER_SWMAJOR)
902 		return -ENODEV;
903 
904 	if (uinfo->subctxt_cnt > HFI1_MAX_SHARED_CTXTS)
905 		return -EINVAL;
906 
907 	swminor = uinfo->userversion & 0xffff;
908 
909 	/*
910 	 * Acquire the mutex to protect against multiple creations of what
911 	 * could be a shared base context.
912 	 */
913 	mutex_lock(&hfi1_mutex);
914 	/*
915 	 * Get a sub context if available  (fd->uctxt will be set).
916 	 * ret < 0 error, 0 no context, 1 sub-context found
917 	 */
918 	ret = find_sub_ctxt(fd, uinfo);
919 
920 	/*
921 	 * Allocate a base context if context sharing is not required or a
922 	 * sub context wasn't found.
923 	 */
924 	if (!ret)
925 		ret = allocate_ctxt(fd, fd->dd, uinfo, &uctxt);
926 
927 	mutex_unlock(&hfi1_mutex);
928 
929 	/* Depending on the context type, finish the appropriate init */
930 	switch (ret) {
931 	case 0:
932 		ret = setup_base_ctxt(fd, uctxt);
933 		if (uctxt->subctxt_cnt) {
934 			/*
935 			 * Base context is done (successfully or not), notify
936 			 * anybody using a sub-context that is waiting for
937 			 * this completion.
938 			 */
939 			clear_bit(HFI1_CTXT_BASE_UNINIT, &uctxt->event_flags);
940 			wake_up(&uctxt->wait);
941 		}
942 		break;
943 	case 1:
944 		ret = complete_subctxt(fd);
945 		break;
946 	default:
947 		break;
948 	}
949 
950 	return ret;
951 }
952 
953 /**
954  * match_ctxt
955  * @fd: valid filedata pointer
956  * @uinfo: user info to compare base context with
957  * @uctxt: context to compare uinfo to.
958  *
959  * Compare the given context with the given information to see if it
960  * can be used for a sub context.
961  */
962 static int match_ctxt(struct hfi1_filedata *fd,
963 		      const struct hfi1_user_info *uinfo,
964 		      struct hfi1_ctxtdata *uctxt)
965 {
966 	struct hfi1_devdata *dd = fd->dd;
967 	unsigned long flags;
968 	u16 subctxt;
969 
970 	/* Skip dynamically allocated kernel contexts */
971 	if (uctxt->sc && (uctxt->sc->type == SC_KERNEL))
972 		return 0;
973 
974 	/* Skip ctxt if it doesn't match the requested one */
975 	if (memcmp(uctxt->uuid, uinfo->uuid, sizeof(uctxt->uuid)) ||
976 	    uctxt->jkey != generate_jkey(current_uid()) ||
977 	    uctxt->subctxt_id != uinfo->subctxt_id ||
978 	    uctxt->subctxt_cnt != uinfo->subctxt_cnt)
979 		return 0;
980 
981 	/* Verify the sharing process matches the base */
982 	if (uctxt->userversion != uinfo->userversion)
983 		return -EINVAL;
984 
985 	/* Find an unused sub context */
986 	spin_lock_irqsave(&dd->uctxt_lock, flags);
987 	if (bitmap_empty(uctxt->in_use_ctxts, HFI1_MAX_SHARED_CTXTS)) {
988 		/* context is being closed, do not use */
989 		spin_unlock_irqrestore(&dd->uctxt_lock, flags);
990 		return 0;
991 	}
992 
993 	subctxt = find_first_zero_bit(uctxt->in_use_ctxts,
994 				      HFI1_MAX_SHARED_CTXTS);
995 	if (subctxt >= uctxt->subctxt_cnt) {
996 		spin_unlock_irqrestore(&dd->uctxt_lock, flags);
997 		return -EBUSY;
998 	}
999 
1000 	fd->subctxt = subctxt;
1001 	__set_bit(fd->subctxt, uctxt->in_use_ctxts);
1002 	spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1003 
1004 	fd->uctxt = uctxt;
1005 	hfi1_rcd_get(uctxt);
1006 
1007 	return 1;
1008 }
1009 
1010 /**
1011  * find_sub_ctxt
1012  * @fd: valid filedata pointer
1013  * @uinfo: matching info to use to find a possible context to share.
1014  *
1015  * The hfi1_mutex must be held when this function is called.  It is
1016  * necessary to ensure serialized creation of shared contexts.
1017  *
1018  * Return:
1019  *    0      No sub-context found
1020  *    1      Subcontext found and allocated
1021  *    errno  EINVAL (incorrect parameters)
1022  *           EBUSY (all sub contexts in use)
1023  */
1024 static int find_sub_ctxt(struct hfi1_filedata *fd,
1025 			 const struct hfi1_user_info *uinfo)
1026 {
1027 	struct hfi1_ctxtdata *uctxt;
1028 	struct hfi1_devdata *dd = fd->dd;
1029 	u16 i;
1030 	int ret;
1031 
1032 	if (!uinfo->subctxt_cnt)
1033 		return 0;
1034 
1035 	for (i = dd->first_dyn_alloc_ctxt; i < dd->num_rcv_contexts; i++) {
1036 		uctxt = hfi1_rcd_get_by_index(dd, i);
1037 		if (uctxt) {
1038 			ret = match_ctxt(fd, uinfo, uctxt);
1039 			hfi1_rcd_put(uctxt);
1040 			/* value of != 0 will return */
1041 			if (ret)
1042 				return ret;
1043 		}
1044 	}
1045 
1046 	return 0;
1047 }
1048 
1049 static int allocate_ctxt(struct hfi1_filedata *fd, struct hfi1_devdata *dd,
1050 			 struct hfi1_user_info *uinfo,
1051 			 struct hfi1_ctxtdata **rcd)
1052 {
1053 	struct hfi1_ctxtdata *uctxt;
1054 	int ret, numa;
1055 
1056 	if (dd->flags & HFI1_FROZEN) {
1057 		/*
1058 		 * Pick an error that is unique from all other errors
1059 		 * that are returned so the user process knows that
1060 		 * it tried to allocate while the SPC was frozen.  It
1061 		 * it should be able to retry with success in a short
1062 		 * while.
1063 		 */
1064 		return -EIO;
1065 	}
1066 
1067 	if (!dd->freectxts)
1068 		return -EBUSY;
1069 
1070 	/*
1071 	 * If we don't have a NUMA node requested, preference is towards
1072 	 * device NUMA node.
1073 	 */
1074 	fd->rec_cpu_num = hfi1_get_proc_affinity(dd->node);
1075 	if (fd->rec_cpu_num != -1)
1076 		numa = cpu_to_node(fd->rec_cpu_num);
1077 	else
1078 		numa = numa_node_id();
1079 	ret = hfi1_create_ctxtdata(dd->pport, numa, &uctxt);
1080 	if (ret < 0) {
1081 		dd_dev_err(dd, "user ctxtdata allocation failed\n");
1082 		return ret;
1083 	}
1084 	hfi1_cdbg(PROC, "[%u:%u] pid %u assigned to CPU %d (NUMA %u)",
1085 		  uctxt->ctxt, fd->subctxt, current->pid, fd->rec_cpu_num,
1086 		  uctxt->numa_id);
1087 
1088 	/*
1089 	 * Allocate and enable a PIO send context.
1090 	 */
1091 	uctxt->sc = sc_alloc(dd, SC_USER, uctxt->rcvhdrqentsize, dd->node);
1092 	if (!uctxt->sc) {
1093 		ret = -ENOMEM;
1094 		goto ctxdata_free;
1095 	}
1096 	hfi1_cdbg(PROC, "allocated send context %u(%u)\n", uctxt->sc->sw_index,
1097 		  uctxt->sc->hw_context);
1098 	ret = sc_enable(uctxt->sc);
1099 	if (ret)
1100 		goto ctxdata_free;
1101 
1102 	/*
1103 	 * Setup sub context information if the user-level has requested
1104 	 * sub contexts.
1105 	 * This has to be done here so the rest of the sub-contexts find the
1106 	 * proper base context.
1107 	 */
1108 	if (uinfo->subctxt_cnt)
1109 		init_subctxts(uctxt, uinfo);
1110 	uctxt->userversion = uinfo->userversion;
1111 	uctxt->flags = hfi1_cap_mask; /* save current flag state */
1112 	init_waitqueue_head(&uctxt->wait);
1113 	strlcpy(uctxt->comm, current->comm, sizeof(uctxt->comm));
1114 	memcpy(uctxt->uuid, uinfo->uuid, sizeof(uctxt->uuid));
1115 	uctxt->jkey = generate_jkey(current_uid());
1116 	hfi1_stats.sps_ctxts++;
1117 	/*
1118 	 * Disable ASPM when there are open user/PSM contexts to avoid
1119 	 * issues with ASPM L1 exit latency
1120 	 */
1121 	if (dd->freectxts-- == dd->num_user_contexts)
1122 		aspm_disable_all(dd);
1123 
1124 	*rcd = uctxt;
1125 
1126 	return 0;
1127 
1128 ctxdata_free:
1129 	hfi1_free_ctxt(uctxt);
1130 	return ret;
1131 }
1132 
1133 static void deallocate_ctxt(struct hfi1_ctxtdata *uctxt)
1134 {
1135 	mutex_lock(&hfi1_mutex);
1136 	hfi1_stats.sps_ctxts--;
1137 	if (++uctxt->dd->freectxts == uctxt->dd->num_user_contexts)
1138 		aspm_enable_all(uctxt->dd);
1139 	mutex_unlock(&hfi1_mutex);
1140 
1141 	hfi1_free_ctxt(uctxt);
1142 }
1143 
1144 static void init_subctxts(struct hfi1_ctxtdata *uctxt,
1145 			  const struct hfi1_user_info *uinfo)
1146 {
1147 	uctxt->subctxt_cnt = uinfo->subctxt_cnt;
1148 	uctxt->subctxt_id = uinfo->subctxt_id;
1149 	set_bit(HFI1_CTXT_BASE_UNINIT, &uctxt->event_flags);
1150 }
1151 
1152 static int setup_subctxt(struct hfi1_ctxtdata *uctxt)
1153 {
1154 	int ret = 0;
1155 	u16 num_subctxts = uctxt->subctxt_cnt;
1156 
1157 	uctxt->subctxt_uregbase = vmalloc_user(PAGE_SIZE);
1158 	if (!uctxt->subctxt_uregbase)
1159 		return -ENOMEM;
1160 
1161 	/* We can take the size of the RcvHdr Queue from the master */
1162 	uctxt->subctxt_rcvhdr_base = vmalloc_user(uctxt->rcvhdrq_size *
1163 						  num_subctxts);
1164 	if (!uctxt->subctxt_rcvhdr_base) {
1165 		ret = -ENOMEM;
1166 		goto bail_ureg;
1167 	}
1168 
1169 	uctxt->subctxt_rcvegrbuf = vmalloc_user(uctxt->egrbufs.size *
1170 						num_subctxts);
1171 	if (!uctxt->subctxt_rcvegrbuf) {
1172 		ret = -ENOMEM;
1173 		goto bail_rhdr;
1174 	}
1175 
1176 	return 0;
1177 
1178 bail_rhdr:
1179 	vfree(uctxt->subctxt_rcvhdr_base);
1180 	uctxt->subctxt_rcvhdr_base = NULL;
1181 bail_ureg:
1182 	vfree(uctxt->subctxt_uregbase);
1183 	uctxt->subctxt_uregbase = NULL;
1184 
1185 	return ret;
1186 }
1187 
1188 static void user_init(struct hfi1_ctxtdata *uctxt)
1189 {
1190 	unsigned int rcvctrl_ops = 0;
1191 
1192 	/* initialize poll variables... */
1193 	uctxt->urgent = 0;
1194 	uctxt->urgent_poll = 0;
1195 
1196 	/*
1197 	 * Now enable the ctxt for receive.
1198 	 * For chips that are set to DMA the tail register to memory
1199 	 * when they change (and when the update bit transitions from
1200 	 * 0 to 1.  So for those chips, we turn it off and then back on.
1201 	 * This will (very briefly) affect any other open ctxts, but the
1202 	 * duration is very short, and therefore isn't an issue.  We
1203 	 * explicitly set the in-memory tail copy to 0 beforehand, so we
1204 	 * don't have to wait to be sure the DMA update has happened
1205 	 * (chip resets head/tail to 0 on transition to enable).
1206 	 */
1207 	if (uctxt->rcvhdrtail_kvaddr)
1208 		clear_rcvhdrtail(uctxt);
1209 
1210 	/* Setup J_KEY before enabling the context */
1211 	hfi1_set_ctxt_jkey(uctxt->dd, uctxt, uctxt->jkey);
1212 
1213 	rcvctrl_ops = HFI1_RCVCTRL_CTXT_ENB;
1214 	if (HFI1_CAP_UGET_MASK(uctxt->flags, HDRSUPP))
1215 		rcvctrl_ops |= HFI1_RCVCTRL_TIDFLOW_ENB;
1216 	/*
1217 	 * Ignore the bit in the flags for now until proper
1218 	 * support for multiple packet per rcv array entry is
1219 	 * added.
1220 	 */
1221 	if (!HFI1_CAP_UGET_MASK(uctxt->flags, MULTI_PKT_EGR))
1222 		rcvctrl_ops |= HFI1_RCVCTRL_ONE_PKT_EGR_ENB;
1223 	if (HFI1_CAP_UGET_MASK(uctxt->flags, NODROP_EGR_FULL))
1224 		rcvctrl_ops |= HFI1_RCVCTRL_NO_EGR_DROP_ENB;
1225 	if (HFI1_CAP_UGET_MASK(uctxt->flags, NODROP_RHQ_FULL))
1226 		rcvctrl_ops |= HFI1_RCVCTRL_NO_RHQ_DROP_ENB;
1227 	/*
1228 	 * The RcvCtxtCtrl.TailUpd bit has to be explicitly written.
1229 	 * We can't rely on the correct value to be set from prior
1230 	 * uses of the chip or ctxt. Therefore, add the rcvctrl op
1231 	 * for both cases.
1232 	 */
1233 	if (HFI1_CAP_UGET_MASK(uctxt->flags, DMA_RTAIL))
1234 		rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_ENB;
1235 	else
1236 		rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_DIS;
1237 	hfi1_rcvctrl(uctxt->dd, rcvctrl_ops, uctxt);
1238 }
1239 
1240 static int get_ctxt_info(struct hfi1_filedata *fd, void __user *ubase,
1241 			 __u32 len)
1242 {
1243 	struct hfi1_ctxt_info cinfo;
1244 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1245 	int ret = 0;
1246 
1247 	memset(&cinfo, 0, sizeof(cinfo));
1248 	cinfo.runtime_flags = (((uctxt->flags >> HFI1_CAP_MISC_SHIFT) &
1249 				HFI1_CAP_MISC_MASK) << HFI1_CAP_USER_SHIFT) |
1250 			HFI1_CAP_UGET_MASK(uctxt->flags, MASK) |
1251 			HFI1_CAP_KGET_MASK(uctxt->flags, K2U);
1252 	/* adjust flag if this fd is not able to cache */
1253 	if (!fd->handler)
1254 		cinfo.runtime_flags |= HFI1_CAP_TID_UNMAP; /* no caching */
1255 
1256 	cinfo.num_active = hfi1_count_active_units();
1257 	cinfo.unit = uctxt->dd->unit;
1258 	cinfo.ctxt = uctxt->ctxt;
1259 	cinfo.subctxt = fd->subctxt;
1260 	cinfo.rcvtids = roundup(uctxt->egrbufs.alloced,
1261 				uctxt->dd->rcv_entries.group_size) +
1262 		uctxt->expected_count;
1263 	cinfo.credits = uctxt->sc->credits;
1264 	cinfo.numa_node = uctxt->numa_id;
1265 	cinfo.rec_cpu = fd->rec_cpu_num;
1266 	cinfo.send_ctxt = uctxt->sc->hw_context;
1267 
1268 	cinfo.egrtids = uctxt->egrbufs.alloced;
1269 	cinfo.rcvhdrq_cnt = uctxt->rcvhdrq_cnt;
1270 	cinfo.rcvhdrq_entsize = uctxt->rcvhdrqentsize << 2;
1271 	cinfo.sdma_ring_size = fd->cq->nentries;
1272 	cinfo.rcvegr_size = uctxt->egrbufs.rcvtid_size;
1273 
1274 	trace_hfi1_ctxt_info(uctxt->dd, uctxt->ctxt, fd->subctxt, cinfo);
1275 	if (copy_to_user(ubase, &cinfo, sizeof(cinfo)))
1276 		ret = -EFAULT;
1277 
1278 	return ret;
1279 }
1280 
1281 static int init_user_ctxt(struct hfi1_filedata *fd,
1282 			  struct hfi1_ctxtdata *uctxt)
1283 {
1284 	int ret;
1285 
1286 	ret = hfi1_user_sdma_alloc_queues(uctxt, fd);
1287 	if (ret)
1288 		return ret;
1289 
1290 	ret = hfi1_user_exp_rcv_init(fd, uctxt);
1291 	if (ret)
1292 		hfi1_user_sdma_free_queues(fd, uctxt);
1293 
1294 	return ret;
1295 }
1296 
1297 static int setup_base_ctxt(struct hfi1_filedata *fd,
1298 			   struct hfi1_ctxtdata *uctxt)
1299 {
1300 	struct hfi1_devdata *dd = uctxt->dd;
1301 	int ret = 0;
1302 
1303 	hfi1_init_ctxt(uctxt->sc);
1304 
1305 	/* Now allocate the RcvHdr queue and eager buffers. */
1306 	ret = hfi1_create_rcvhdrq(dd, uctxt);
1307 	if (ret)
1308 		return ret;
1309 
1310 	ret = hfi1_setup_eagerbufs(uctxt);
1311 	if (ret)
1312 		goto setup_failed;
1313 
1314 	/* If sub-contexts are enabled, do the appropriate setup */
1315 	if (uctxt->subctxt_cnt)
1316 		ret = setup_subctxt(uctxt);
1317 	if (ret)
1318 		goto setup_failed;
1319 
1320 	ret = hfi1_alloc_ctxt_rcv_groups(uctxt);
1321 	if (ret)
1322 		goto setup_failed;
1323 
1324 	ret = init_user_ctxt(fd, uctxt);
1325 	if (ret)
1326 		goto setup_failed;
1327 
1328 	user_init(uctxt);
1329 
1330 	/* Now that the context is set up, the fd can get a reference. */
1331 	fd->uctxt = uctxt;
1332 	hfi1_rcd_get(uctxt);
1333 
1334 	return 0;
1335 
1336 setup_failed:
1337 	/* Set the failed bit so sub-context init can do the right thing */
1338 	set_bit(HFI1_CTXT_BASE_FAILED, &uctxt->event_flags);
1339 	deallocate_ctxt(uctxt);
1340 
1341 	return ret;
1342 }
1343 
1344 static int get_base_info(struct hfi1_filedata *fd, void __user *ubase,
1345 			 __u32 len)
1346 {
1347 	struct hfi1_base_info binfo;
1348 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1349 	struct hfi1_devdata *dd = uctxt->dd;
1350 	ssize_t sz;
1351 	unsigned offset;
1352 	int ret = 0;
1353 
1354 	trace_hfi1_uctxtdata(uctxt->dd, uctxt, fd->subctxt);
1355 
1356 	memset(&binfo, 0, sizeof(binfo));
1357 	binfo.hw_version = dd->revision;
1358 	binfo.sw_version = HFI1_KERN_SWVERSION;
1359 	binfo.bthqp = kdeth_qp;
1360 	binfo.jkey = uctxt->jkey;
1361 	/*
1362 	 * If more than 64 contexts are enabled the allocated credit
1363 	 * return will span two or three contiguous pages. Since we only
1364 	 * map the page containing the context's credit return address,
1365 	 * we need to calculate the offset in the proper page.
1366 	 */
1367 	offset = ((u64)uctxt->sc->hw_free -
1368 		  (u64)dd->cr_base[uctxt->numa_id].va) % PAGE_SIZE;
1369 	binfo.sc_credits_addr = HFI1_MMAP_TOKEN(PIO_CRED, uctxt->ctxt,
1370 						fd->subctxt, offset);
1371 	binfo.pio_bufbase = HFI1_MMAP_TOKEN(PIO_BUFS, uctxt->ctxt,
1372 					    fd->subctxt,
1373 					    uctxt->sc->base_addr);
1374 	binfo.pio_bufbase_sop = HFI1_MMAP_TOKEN(PIO_BUFS_SOP,
1375 						uctxt->ctxt,
1376 						fd->subctxt,
1377 						uctxt->sc->base_addr);
1378 	binfo.rcvhdr_bufbase = HFI1_MMAP_TOKEN(RCV_HDRQ, uctxt->ctxt,
1379 					       fd->subctxt,
1380 					       uctxt->rcvhdrq);
1381 	binfo.rcvegr_bufbase = HFI1_MMAP_TOKEN(RCV_EGRBUF, uctxt->ctxt,
1382 					       fd->subctxt,
1383 					       uctxt->egrbufs.rcvtids[0].dma);
1384 	binfo.sdma_comp_bufbase = HFI1_MMAP_TOKEN(SDMA_COMP, uctxt->ctxt,
1385 						 fd->subctxt, 0);
1386 	/*
1387 	 * user regs are at
1388 	 * (RXE_PER_CONTEXT_USER + (ctxt * RXE_PER_CONTEXT_SIZE))
1389 	 */
1390 	binfo.user_regbase = HFI1_MMAP_TOKEN(UREGS, uctxt->ctxt,
1391 					    fd->subctxt, 0);
1392 	offset = offset_in_page((((uctxt->ctxt - dd->first_dyn_alloc_ctxt) *
1393 		    HFI1_MAX_SHARED_CTXTS) + fd->subctxt) *
1394 		  sizeof(*dd->events));
1395 	binfo.events_bufbase = HFI1_MMAP_TOKEN(EVENTS, uctxt->ctxt,
1396 					      fd->subctxt,
1397 					      offset);
1398 	binfo.status_bufbase = HFI1_MMAP_TOKEN(STATUS, uctxt->ctxt,
1399 					      fd->subctxt,
1400 					      dd->status);
1401 	if (HFI1_CAP_IS_USET(DMA_RTAIL))
1402 		binfo.rcvhdrtail_base = HFI1_MMAP_TOKEN(RTAIL, uctxt->ctxt,
1403 						       fd->subctxt, 0);
1404 	if (uctxt->subctxt_cnt) {
1405 		binfo.subctxt_uregbase = HFI1_MMAP_TOKEN(SUBCTXT_UREGS,
1406 							uctxt->ctxt,
1407 							fd->subctxt, 0);
1408 		binfo.subctxt_rcvhdrbuf = HFI1_MMAP_TOKEN(SUBCTXT_RCV_HDRQ,
1409 							 uctxt->ctxt,
1410 							 fd->subctxt, 0);
1411 		binfo.subctxt_rcvegrbuf = HFI1_MMAP_TOKEN(SUBCTXT_EGRBUF,
1412 							 uctxt->ctxt,
1413 							 fd->subctxt, 0);
1414 	}
1415 	sz = (len < sizeof(binfo)) ? len : sizeof(binfo);
1416 	if (copy_to_user(ubase, &binfo, sz))
1417 		ret = -EFAULT;
1418 	return ret;
1419 }
1420 
1421 static unsigned int poll_urgent(struct file *fp,
1422 				struct poll_table_struct *pt)
1423 {
1424 	struct hfi1_filedata *fd = fp->private_data;
1425 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1426 	struct hfi1_devdata *dd = uctxt->dd;
1427 	unsigned pollflag;
1428 
1429 	poll_wait(fp, &uctxt->wait, pt);
1430 
1431 	spin_lock_irq(&dd->uctxt_lock);
1432 	if (uctxt->urgent != uctxt->urgent_poll) {
1433 		pollflag = POLLIN | POLLRDNORM;
1434 		uctxt->urgent_poll = uctxt->urgent;
1435 	} else {
1436 		pollflag = 0;
1437 		set_bit(HFI1_CTXT_WAITING_URG, &uctxt->event_flags);
1438 	}
1439 	spin_unlock_irq(&dd->uctxt_lock);
1440 
1441 	return pollflag;
1442 }
1443 
1444 static unsigned int poll_next(struct file *fp,
1445 			      struct poll_table_struct *pt)
1446 {
1447 	struct hfi1_filedata *fd = fp->private_data;
1448 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1449 	struct hfi1_devdata *dd = uctxt->dd;
1450 	unsigned pollflag;
1451 
1452 	poll_wait(fp, &uctxt->wait, pt);
1453 
1454 	spin_lock_irq(&dd->uctxt_lock);
1455 	if (hdrqempty(uctxt)) {
1456 		set_bit(HFI1_CTXT_WAITING_RCV, &uctxt->event_flags);
1457 		hfi1_rcvctrl(dd, HFI1_RCVCTRL_INTRAVAIL_ENB, uctxt);
1458 		pollflag = 0;
1459 	} else {
1460 		pollflag = POLLIN | POLLRDNORM;
1461 	}
1462 	spin_unlock_irq(&dd->uctxt_lock);
1463 
1464 	return pollflag;
1465 }
1466 
1467 /*
1468  * Find all user contexts in use, and set the specified bit in their
1469  * event mask.
1470  * See also find_ctxt() for a similar use, that is specific to send buffers.
1471  */
1472 int hfi1_set_uevent_bits(struct hfi1_pportdata *ppd, const int evtbit)
1473 {
1474 	struct hfi1_ctxtdata *uctxt;
1475 	struct hfi1_devdata *dd = ppd->dd;
1476 	u16 ctxt;
1477 
1478 	if (!dd->events)
1479 		return -EINVAL;
1480 
1481 	for (ctxt = dd->first_dyn_alloc_ctxt; ctxt < dd->num_rcv_contexts;
1482 	     ctxt++) {
1483 		uctxt = hfi1_rcd_get_by_index(dd, ctxt);
1484 		if (uctxt) {
1485 			unsigned long *evs = dd->events +
1486 				(uctxt->ctxt - dd->first_dyn_alloc_ctxt) *
1487 				HFI1_MAX_SHARED_CTXTS;
1488 			int i;
1489 			/*
1490 			 * subctxt_cnt is 0 if not shared, so do base
1491 			 * separately, first, then remaining subctxt, if any
1492 			 */
1493 			set_bit(evtbit, evs);
1494 			for (i = 1; i < uctxt->subctxt_cnt; i++)
1495 				set_bit(evtbit, evs + i);
1496 			hfi1_rcd_put(uctxt);
1497 		}
1498 	}
1499 
1500 	return 0;
1501 }
1502 
1503 /**
1504  * manage_rcvq - manage a context's receive queue
1505  * @uctxt: the context
1506  * @subctxt: the sub-context
1507  * @start_stop: action to carry out
1508  *
1509  * start_stop == 0 disables receive on the context, for use in queue
1510  * overflow conditions.  start_stop==1 re-enables, to be used to
1511  * re-init the software copy of the head register
1512  */
1513 static int manage_rcvq(struct hfi1_ctxtdata *uctxt, u16 subctxt,
1514 		       int start_stop)
1515 {
1516 	struct hfi1_devdata *dd = uctxt->dd;
1517 	unsigned int rcvctrl_op;
1518 
1519 	if (subctxt)
1520 		goto bail;
1521 	/* atomically clear receive enable ctxt. */
1522 	if (start_stop) {
1523 		/*
1524 		 * On enable, force in-memory copy of the tail register to
1525 		 * 0, so that protocol code doesn't have to worry about
1526 		 * whether or not the chip has yet updated the in-memory
1527 		 * copy or not on return from the system call. The chip
1528 		 * always resets it's tail register back to 0 on a
1529 		 * transition from disabled to enabled.
1530 		 */
1531 		if (uctxt->rcvhdrtail_kvaddr)
1532 			clear_rcvhdrtail(uctxt);
1533 		rcvctrl_op = HFI1_RCVCTRL_CTXT_ENB;
1534 	} else {
1535 		rcvctrl_op = HFI1_RCVCTRL_CTXT_DIS;
1536 	}
1537 	hfi1_rcvctrl(dd, rcvctrl_op, uctxt);
1538 	/* always; new head should be equal to new tail; see above */
1539 bail:
1540 	return 0;
1541 }
1542 
1543 /*
1544  * clear the event notifier events for this context.
1545  * User process then performs actions appropriate to bit having been
1546  * set, if desired, and checks again in future.
1547  */
1548 static int user_event_ack(struct hfi1_ctxtdata *uctxt, u16 subctxt,
1549 			  unsigned long events)
1550 {
1551 	int i;
1552 	struct hfi1_devdata *dd = uctxt->dd;
1553 	unsigned long *evs;
1554 
1555 	if (!dd->events)
1556 		return 0;
1557 
1558 	evs = dd->events + ((uctxt->ctxt - dd->first_dyn_alloc_ctxt) *
1559 			    HFI1_MAX_SHARED_CTXTS) + subctxt;
1560 
1561 	for (i = 0; i <= _HFI1_MAX_EVENT_BIT; i++) {
1562 		if (!test_bit(i, &events))
1563 			continue;
1564 		clear_bit(i, evs);
1565 	}
1566 	return 0;
1567 }
1568 
1569 static int set_ctxt_pkey(struct hfi1_ctxtdata *uctxt, u16 subctxt, u16 pkey)
1570 {
1571 	int ret = -ENOENT, i, intable = 0;
1572 	struct hfi1_pportdata *ppd = uctxt->ppd;
1573 	struct hfi1_devdata *dd = uctxt->dd;
1574 
1575 	if (pkey == LIM_MGMT_P_KEY || pkey == FULL_MGMT_P_KEY) {
1576 		ret = -EINVAL;
1577 		goto done;
1578 	}
1579 
1580 	for (i = 0; i < ARRAY_SIZE(ppd->pkeys); i++)
1581 		if (pkey == ppd->pkeys[i]) {
1582 			intable = 1;
1583 			break;
1584 		}
1585 
1586 	if (intable)
1587 		ret = hfi1_set_ctxt_pkey(dd, uctxt, pkey);
1588 done:
1589 	return ret;
1590 }
1591 
1592 static void user_remove(struct hfi1_devdata *dd)
1593 {
1594 
1595 	hfi1_cdev_cleanup(&dd->user_cdev, &dd->user_device);
1596 }
1597 
1598 static int user_add(struct hfi1_devdata *dd)
1599 {
1600 	char name[10];
1601 	int ret;
1602 
1603 	snprintf(name, sizeof(name), "%s_%d", class_name(), dd->unit);
1604 	ret = hfi1_cdev_init(dd->unit, name, &hfi1_file_ops,
1605 			     &dd->user_cdev, &dd->user_device,
1606 			     true, &dd->kobj);
1607 	if (ret)
1608 		user_remove(dd);
1609 
1610 	return ret;
1611 }
1612 
1613 /*
1614  * Create per-unit files in /dev
1615  */
1616 int hfi1_device_create(struct hfi1_devdata *dd)
1617 {
1618 	return user_add(dd);
1619 }
1620 
1621 /*
1622  * Remove per-unit files in /dev
1623  * void, core kernel returns no errors for this stuff
1624  */
1625 void hfi1_device_remove(struct hfi1_devdata *dd)
1626 {
1627 	user_remove(dd);
1628 }
1629