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_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 	if (!atomic_inc_not_zero(&dd->user_refcount))
176 		return -ENXIO;
177 
178 	/* Just take a ref now. Not all opens result in a context assign */
179 	kobject_get(&dd->kobj);
180 
181 	/* The real work is performed later in assign_ctxt() */
182 
183 	fd = kzalloc(sizeof(*fd), GFP_KERNEL);
184 
185 	if (fd) {
186 		fd->rec_cpu_num = -1; /* no cpu affinity by default */
187 		fd->mm = current->mm;
188 		atomic_inc(&fd->mm->mm_count);
189 		fp->private_data = fd;
190 	} else {
191 		fp->private_data = NULL;
192 
193 		if (atomic_dec_and_test(&dd->user_refcount))
194 			complete(&dd->user_comp);
195 
196 		return -ENOMEM;
197 	}
198 
199 	return 0;
200 }
201 
202 static long hfi1_file_ioctl(struct file *fp, unsigned int cmd,
203 			    unsigned long arg)
204 {
205 	struct hfi1_filedata *fd = fp->private_data;
206 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
207 	struct hfi1_user_info uinfo;
208 	struct hfi1_tid_info tinfo;
209 	int ret = 0;
210 	unsigned long addr;
211 	int uval = 0;
212 	unsigned long ul_uval = 0;
213 	u16 uval16 = 0;
214 
215 	hfi1_cdbg(IOCTL, "IOCTL recv: 0x%x", cmd);
216 	if (cmd != HFI1_IOCTL_ASSIGN_CTXT &&
217 	    cmd != HFI1_IOCTL_GET_VERS &&
218 	    !uctxt)
219 		return -EINVAL;
220 
221 	switch (cmd) {
222 	case HFI1_IOCTL_ASSIGN_CTXT:
223 		if (uctxt)
224 			return -EINVAL;
225 
226 		if (copy_from_user(&uinfo,
227 				   (struct hfi1_user_info __user *)arg,
228 				   sizeof(uinfo)))
229 			return -EFAULT;
230 
231 		ret = assign_ctxt(fp, &uinfo);
232 		if (ret < 0)
233 			return ret;
234 		ret = setup_ctxt(fp);
235 		if (ret)
236 			return ret;
237 		ret = user_init(fp);
238 		break;
239 	case HFI1_IOCTL_CTXT_INFO:
240 		ret = get_ctxt_info(fp, (void __user *)(unsigned long)arg,
241 				    sizeof(struct hfi1_ctxt_info));
242 		break;
243 	case HFI1_IOCTL_USER_INFO:
244 		ret = get_base_info(fp, (void __user *)(unsigned long)arg,
245 				    sizeof(struct hfi1_base_info));
246 		break;
247 	case HFI1_IOCTL_CREDIT_UPD:
248 		if (uctxt)
249 			sc_return_credits(uctxt->sc);
250 		break;
251 
252 	case HFI1_IOCTL_TID_UPDATE:
253 		if (copy_from_user(&tinfo,
254 				   (struct hfi11_tid_info __user *)arg,
255 				   sizeof(tinfo)))
256 			return -EFAULT;
257 
258 		ret = hfi1_user_exp_rcv_setup(fp, &tinfo);
259 		if (!ret) {
260 			/*
261 			 * Copy the number of tidlist entries we used
262 			 * and the length of the buffer we registered.
263 			 * These fields are adjacent in the structure so
264 			 * we can copy them at the same time.
265 			 */
266 			addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
267 			if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
268 					 sizeof(tinfo.tidcnt) +
269 					 sizeof(tinfo.length)))
270 				ret = -EFAULT;
271 		}
272 		break;
273 
274 	case HFI1_IOCTL_TID_FREE:
275 		if (copy_from_user(&tinfo,
276 				   (struct hfi11_tid_info __user *)arg,
277 				   sizeof(tinfo)))
278 			return -EFAULT;
279 
280 		ret = hfi1_user_exp_rcv_clear(fp, &tinfo);
281 		if (ret)
282 			break;
283 		addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
284 		if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
285 				 sizeof(tinfo.tidcnt)))
286 			ret = -EFAULT;
287 		break;
288 
289 	case HFI1_IOCTL_TID_INVAL_READ:
290 		if (copy_from_user(&tinfo,
291 				   (struct hfi11_tid_info __user *)arg,
292 				   sizeof(tinfo)))
293 			return -EFAULT;
294 
295 		ret = hfi1_user_exp_rcv_invalid(fp, &tinfo);
296 		if (ret)
297 			break;
298 		addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
299 		if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
300 				 sizeof(tinfo.tidcnt)))
301 			ret = -EFAULT;
302 		break;
303 
304 	case HFI1_IOCTL_RECV_CTRL:
305 		ret = get_user(uval, (int __user *)arg);
306 		if (ret != 0)
307 			return -EFAULT;
308 		ret = manage_rcvq(uctxt, fd->subctxt, uval);
309 		break;
310 
311 	case HFI1_IOCTL_POLL_TYPE:
312 		ret = get_user(uval, (int __user *)arg);
313 		if (ret != 0)
314 			return -EFAULT;
315 		uctxt->poll_type = (typeof(uctxt->poll_type))uval;
316 		break;
317 
318 	case HFI1_IOCTL_ACK_EVENT:
319 		ret = get_user(ul_uval, (unsigned long __user *)arg);
320 		if (ret != 0)
321 			return -EFAULT;
322 		ret = user_event_ack(uctxt, fd->subctxt, ul_uval);
323 		break;
324 
325 	case HFI1_IOCTL_SET_PKEY:
326 		ret = get_user(uval16, (u16 __user *)arg);
327 		if (ret != 0)
328 			return -EFAULT;
329 		if (HFI1_CAP_IS_USET(PKEY_CHECK))
330 			ret = set_ctxt_pkey(uctxt, fd->subctxt, uval16);
331 		else
332 			return -EPERM;
333 		break;
334 
335 	case HFI1_IOCTL_CTXT_RESET: {
336 		struct send_context *sc;
337 		struct hfi1_devdata *dd;
338 
339 		if (!uctxt || !uctxt->dd || !uctxt->sc)
340 			return -EINVAL;
341 
342 		/*
343 		 * There is no protection here. User level has to
344 		 * guarantee that no one will be writing to the send
345 		 * context while it is being re-initialized.
346 		 * If user level breaks that guarantee, it will break
347 		 * it's own context and no one else's.
348 		 */
349 		dd = uctxt->dd;
350 		sc = uctxt->sc;
351 		/*
352 		 * Wait until the interrupt handler has marked the
353 		 * context as halted or frozen. Report error if we time
354 		 * out.
355 		 */
356 		wait_event_interruptible_timeout(
357 			sc->halt_wait, (sc->flags & SCF_HALTED),
358 			msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
359 		if (!(sc->flags & SCF_HALTED))
360 			return -ENOLCK;
361 
362 		/*
363 		 * If the send context was halted due to a Freeze,
364 		 * wait until the device has been "unfrozen" before
365 		 * resetting the context.
366 		 */
367 		if (sc->flags & SCF_FROZEN) {
368 			wait_event_interruptible_timeout(
369 				dd->event_queue,
370 				!(ACCESS_ONCE(dd->flags) & HFI1_FROZEN),
371 				msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
372 			if (dd->flags & HFI1_FROZEN)
373 				return -ENOLCK;
374 
375 			if (dd->flags & HFI1_FORCED_FREEZE)
376 				/*
377 				 * Don't allow context reset if we are into
378 				 * forced freeze
379 				 */
380 				return -ENODEV;
381 
382 			sc_disable(sc);
383 			ret = sc_enable(sc);
384 			hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_ENB,
385 				     uctxt->ctxt);
386 		} else {
387 			ret = sc_restart(sc);
388 		}
389 		if (!ret)
390 			sc_return_credits(sc);
391 		break;
392 	}
393 
394 	case HFI1_IOCTL_GET_VERS:
395 		uval = HFI1_USER_SWVERSION;
396 		if (put_user(uval, (int __user *)arg))
397 			return -EFAULT;
398 		break;
399 
400 	default:
401 		return -EINVAL;
402 	}
403 
404 	return ret;
405 }
406 
407 static ssize_t hfi1_write_iter(struct kiocb *kiocb, struct iov_iter *from)
408 {
409 	struct hfi1_filedata *fd = kiocb->ki_filp->private_data;
410 	struct hfi1_user_sdma_pkt_q *pq = fd->pq;
411 	struct hfi1_user_sdma_comp_q *cq = fd->cq;
412 	int done = 0, reqs = 0;
413 	unsigned long dim = from->nr_segs;
414 
415 	if (!cq || !pq)
416 		return -EIO;
417 
418 	if (!iter_is_iovec(from) || !dim)
419 		return -EINVAL;
420 
421 	hfi1_cdbg(SDMA, "SDMA request from %u:%u (%lu)",
422 		  fd->uctxt->ctxt, fd->subctxt, dim);
423 
424 	if (atomic_read(&pq->n_reqs) == pq->n_max_reqs)
425 		return -ENOSPC;
426 
427 	while (dim) {
428 		int ret;
429 		unsigned long count = 0;
430 
431 		ret = hfi1_user_sdma_process_request(
432 			kiocb->ki_filp,	(struct iovec *)(from->iov + done),
433 			dim, &count);
434 		if (ret) {
435 			reqs = ret;
436 			break;
437 		}
438 		dim -= count;
439 		done += count;
440 		reqs++;
441 	}
442 
443 	return reqs;
444 }
445 
446 static int hfi1_file_mmap(struct file *fp, struct vm_area_struct *vma)
447 {
448 	struct hfi1_filedata *fd = fp->private_data;
449 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
450 	struct hfi1_devdata *dd;
451 	unsigned long flags;
452 	u64 token = vma->vm_pgoff << PAGE_SHIFT,
453 		memaddr = 0;
454 	void *memvirt = NULL;
455 	u8 subctxt, mapio = 0, vmf = 0, type;
456 	ssize_t memlen = 0;
457 	int ret = 0;
458 	u16 ctxt;
459 
460 	if (!is_valid_mmap(token) || !uctxt ||
461 	    !(vma->vm_flags & VM_SHARED)) {
462 		ret = -EINVAL;
463 		goto done;
464 	}
465 	dd = uctxt->dd;
466 	ctxt = HFI1_MMAP_TOKEN_GET(CTXT, token);
467 	subctxt = HFI1_MMAP_TOKEN_GET(SUBCTXT, token);
468 	type = HFI1_MMAP_TOKEN_GET(TYPE, token);
469 	if (ctxt != uctxt->ctxt || subctxt != fd->subctxt) {
470 		ret = -EINVAL;
471 		goto done;
472 	}
473 
474 	flags = vma->vm_flags;
475 
476 	switch (type) {
477 	case PIO_BUFS:
478 	case PIO_BUFS_SOP:
479 		memaddr = ((dd->physaddr + TXE_PIO_SEND) +
480 				/* chip pio base */
481 			   (uctxt->sc->hw_context * BIT(16))) +
482 				/* 64K PIO space / ctxt */
483 			(type == PIO_BUFS_SOP ?
484 				(TXE_PIO_SIZE / 2) : 0); /* sop? */
485 		/*
486 		 * Map only the amount allocated to the context, not the
487 		 * entire available context's PIO space.
488 		 */
489 		memlen = PAGE_ALIGN(uctxt->sc->credits * PIO_BLOCK_SIZE);
490 		flags &= ~VM_MAYREAD;
491 		flags |= VM_DONTCOPY | VM_DONTEXPAND;
492 		vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
493 		mapio = 1;
494 		break;
495 	case PIO_CRED:
496 		if (flags & VM_WRITE) {
497 			ret = -EPERM;
498 			goto done;
499 		}
500 		/*
501 		 * The credit return location for this context could be on the
502 		 * second or third page allocated for credit returns (if number
503 		 * of enabled contexts > 64 and 128 respectively).
504 		 */
505 		memvirt = dd->cr_base[uctxt->numa_id].va;
506 		memaddr = virt_to_phys(memvirt) +
507 			(((u64)uctxt->sc->hw_free -
508 			  (u64)dd->cr_base[uctxt->numa_id].va) & PAGE_MASK);
509 		memlen = PAGE_SIZE;
510 		flags &= ~VM_MAYWRITE;
511 		flags |= VM_DONTCOPY | VM_DONTEXPAND;
512 		/*
513 		 * The driver has already allocated memory for credit
514 		 * returns and programmed it into the chip. Has that
515 		 * memory been flagged as non-cached?
516 		 */
517 		/* vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); */
518 		mapio = 1;
519 		break;
520 	case RCV_HDRQ:
521 		memlen = uctxt->rcvhdrq_size;
522 		memvirt = uctxt->rcvhdrq;
523 		break;
524 	case RCV_EGRBUF: {
525 		unsigned long addr;
526 		int i;
527 		/*
528 		 * The RcvEgr buffer need to be handled differently
529 		 * as multiple non-contiguous pages need to be mapped
530 		 * into the user process.
531 		 */
532 		memlen = uctxt->egrbufs.size;
533 		if ((vma->vm_end - vma->vm_start) != memlen) {
534 			dd_dev_err(dd, "Eager buffer map size invalid (%lu != %lu)\n",
535 				   (vma->vm_end - vma->vm_start), memlen);
536 			ret = -EINVAL;
537 			goto done;
538 		}
539 		if (vma->vm_flags & VM_WRITE) {
540 			ret = -EPERM;
541 			goto done;
542 		}
543 		vma->vm_flags &= ~VM_MAYWRITE;
544 		addr = vma->vm_start;
545 		for (i = 0 ; i < uctxt->egrbufs.numbufs; i++) {
546 			memlen = uctxt->egrbufs.buffers[i].len;
547 			memvirt = uctxt->egrbufs.buffers[i].addr;
548 			ret = remap_pfn_range(
549 				vma, addr,
550 				/*
551 				 * virt_to_pfn() does the same, but
552 				 * it's not available on x86_64
553 				 * when CONFIG_MMU is enabled.
554 				 */
555 				PFN_DOWN(__pa(memvirt)),
556 				memlen,
557 				vma->vm_page_prot);
558 			if (ret < 0)
559 				goto done;
560 			addr += memlen;
561 		}
562 		ret = 0;
563 		goto done;
564 	}
565 	case UREGS:
566 		/*
567 		 * Map only the page that contains this context's user
568 		 * registers.
569 		 */
570 		memaddr = (unsigned long)
571 			(dd->physaddr + RXE_PER_CONTEXT_USER)
572 			+ (uctxt->ctxt * RXE_PER_CONTEXT_SIZE);
573 		/*
574 		 * TidFlow table is on the same page as the rest of the
575 		 * user registers.
576 		 */
577 		memlen = PAGE_SIZE;
578 		flags |= VM_DONTCOPY | VM_DONTEXPAND;
579 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
580 		mapio = 1;
581 		break;
582 	case EVENTS:
583 		/*
584 		 * Use the page where this context's flags are. User level
585 		 * knows where it's own bitmap is within the page.
586 		 */
587 		memaddr = (unsigned long)(dd->events +
588 					  ((uctxt->ctxt - dd->first_user_ctxt) *
589 					   HFI1_MAX_SHARED_CTXTS)) & PAGE_MASK;
590 		memlen = PAGE_SIZE;
591 		/*
592 		 * v3.7 removes VM_RESERVED but the effect is kept by
593 		 * using VM_IO.
594 		 */
595 		flags |= VM_IO | VM_DONTEXPAND;
596 		vmf = 1;
597 		break;
598 	case STATUS:
599 		memaddr = kvirt_to_phys((void *)dd->status);
600 		memlen = PAGE_SIZE;
601 		flags |= VM_IO | VM_DONTEXPAND;
602 		break;
603 	case RTAIL:
604 		if (!HFI1_CAP_IS_USET(DMA_RTAIL)) {
605 			/*
606 			 * If the memory allocation failed, the context alloc
607 			 * also would have failed, so we would never get here
608 			 */
609 			ret = -EINVAL;
610 			goto done;
611 		}
612 		if (flags & VM_WRITE) {
613 			ret = -EPERM;
614 			goto done;
615 		}
616 		memlen = PAGE_SIZE;
617 		memvirt = (void *)uctxt->rcvhdrtail_kvaddr;
618 		flags &= ~VM_MAYWRITE;
619 		break;
620 	case SUBCTXT_UREGS:
621 		memaddr = (u64)uctxt->subctxt_uregbase;
622 		memlen = PAGE_SIZE;
623 		flags |= VM_IO | VM_DONTEXPAND;
624 		vmf = 1;
625 		break;
626 	case SUBCTXT_RCV_HDRQ:
627 		memaddr = (u64)uctxt->subctxt_rcvhdr_base;
628 		memlen = uctxt->rcvhdrq_size * uctxt->subctxt_cnt;
629 		flags |= VM_IO | VM_DONTEXPAND;
630 		vmf = 1;
631 		break;
632 	case SUBCTXT_EGRBUF:
633 		memaddr = (u64)uctxt->subctxt_rcvegrbuf;
634 		memlen = uctxt->egrbufs.size * uctxt->subctxt_cnt;
635 		flags |= VM_IO | VM_DONTEXPAND;
636 		flags &= ~VM_MAYWRITE;
637 		vmf = 1;
638 		break;
639 	case SDMA_COMP: {
640 		struct hfi1_user_sdma_comp_q *cq = fd->cq;
641 
642 		if (!cq) {
643 			ret = -EFAULT;
644 			goto done;
645 		}
646 		memaddr = (u64)cq->comps;
647 		memlen = PAGE_ALIGN(sizeof(*cq->comps) * cq->nentries);
648 		flags |= VM_IO | VM_DONTEXPAND;
649 		vmf = 1;
650 		break;
651 	}
652 	default:
653 		ret = -EINVAL;
654 		break;
655 	}
656 
657 	if ((vma->vm_end - vma->vm_start) != memlen) {
658 		hfi1_cdbg(PROC, "%u:%u Memory size mismatch %lu:%lu",
659 			  uctxt->ctxt, fd->subctxt,
660 			  (vma->vm_end - vma->vm_start), memlen);
661 		ret = -EINVAL;
662 		goto done;
663 	}
664 
665 	vma->vm_flags = flags;
666 	hfi1_cdbg(PROC,
667 		  "%u:%u type:%u io/vf:%d/%d, addr:0x%llx, len:%lu(%lu), flags:0x%lx\n",
668 		    ctxt, subctxt, type, mapio, vmf, memaddr, memlen,
669 		    vma->vm_end - vma->vm_start, vma->vm_flags);
670 	if (vmf) {
671 		vma->vm_pgoff = PFN_DOWN(memaddr);
672 		vma->vm_ops = &vm_ops;
673 		ret = 0;
674 	} else if (mapio) {
675 		ret = io_remap_pfn_range(vma, vma->vm_start,
676 					 PFN_DOWN(memaddr),
677 					 memlen,
678 					 vma->vm_page_prot);
679 	} else if (memvirt) {
680 		ret = remap_pfn_range(vma, vma->vm_start,
681 				      PFN_DOWN(__pa(memvirt)),
682 				      memlen,
683 				      vma->vm_page_prot);
684 	} else {
685 		ret = remap_pfn_range(vma, vma->vm_start,
686 				      PFN_DOWN(memaddr),
687 				      memlen,
688 				      vma->vm_page_prot);
689 	}
690 done:
691 	return ret;
692 }
693 
694 /*
695  * Local (non-chip) user memory is not mapped right away but as it is
696  * accessed by the user-level code.
697  */
698 static int vma_fault(struct vm_fault *vmf)
699 {
700 	struct page *page;
701 
702 	page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
703 	if (!page)
704 		return VM_FAULT_SIGBUS;
705 
706 	get_page(page);
707 	vmf->page = page;
708 
709 	return 0;
710 }
711 
712 static unsigned int hfi1_poll(struct file *fp, struct poll_table_struct *pt)
713 {
714 	struct hfi1_ctxtdata *uctxt;
715 	unsigned pollflag;
716 
717 	uctxt = ((struct hfi1_filedata *)fp->private_data)->uctxt;
718 	if (!uctxt)
719 		pollflag = POLLERR;
720 	else if (uctxt->poll_type == HFI1_POLL_TYPE_URGENT)
721 		pollflag = poll_urgent(fp, pt);
722 	else  if (uctxt->poll_type == HFI1_POLL_TYPE_ANYRCV)
723 		pollflag = poll_next(fp, pt);
724 	else /* invalid */
725 		pollflag = POLLERR;
726 
727 	return pollflag;
728 }
729 
730 static int hfi1_file_close(struct inode *inode, struct file *fp)
731 {
732 	struct hfi1_filedata *fdata = fp->private_data;
733 	struct hfi1_ctxtdata *uctxt = fdata->uctxt;
734 	struct hfi1_devdata *dd = container_of(inode->i_cdev,
735 					       struct hfi1_devdata,
736 					       user_cdev);
737 	unsigned long flags, *ev;
738 
739 	fp->private_data = NULL;
740 
741 	if (!uctxt)
742 		goto done;
743 
744 	hfi1_cdbg(PROC, "freeing ctxt %u:%u", uctxt->ctxt, fdata->subctxt);
745 	mutex_lock(&hfi1_mutex);
746 
747 	flush_wc();
748 	/* drain user sdma queue */
749 	hfi1_user_sdma_free_queues(fdata);
750 
751 	/* release the cpu */
752 	hfi1_put_proc_affinity(fdata->rec_cpu_num);
753 
754 	/*
755 	 * Clear any left over, unhandled events so the next process that
756 	 * gets this context doesn't get confused.
757 	 */
758 	ev = dd->events + ((uctxt->ctxt - dd->first_user_ctxt) *
759 			   HFI1_MAX_SHARED_CTXTS) + fdata->subctxt;
760 	*ev = 0;
761 
762 	if (--uctxt->cnt) {
763 		uctxt->active_slaves &= ~(1 << fdata->subctxt);
764 		mutex_unlock(&hfi1_mutex);
765 		goto done;
766 	}
767 
768 	spin_lock_irqsave(&dd->uctxt_lock, flags);
769 	/*
770 	 * Disable receive context and interrupt available, reset all
771 	 * RcvCtxtCtrl bits to default values.
772 	 */
773 	hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_DIS |
774 		     HFI1_RCVCTRL_TIDFLOW_DIS |
775 		     HFI1_RCVCTRL_INTRAVAIL_DIS |
776 		     HFI1_RCVCTRL_TAILUPD_DIS |
777 		     HFI1_RCVCTRL_ONE_PKT_EGR_DIS |
778 		     HFI1_RCVCTRL_NO_RHQ_DROP_DIS |
779 		     HFI1_RCVCTRL_NO_EGR_DROP_DIS, uctxt->ctxt);
780 	/* Clear the context's J_KEY */
781 	hfi1_clear_ctxt_jkey(dd, uctxt->ctxt);
782 	/*
783 	 * Reset context integrity checks to default.
784 	 * (writes to CSRs probably belong in chip.c)
785 	 */
786 	write_kctxt_csr(dd, uctxt->sc->hw_context, SEND_CTXT_CHECK_ENABLE,
787 			hfi1_pkt_default_send_ctxt_mask(dd, uctxt->sc->type));
788 	sc_disable(uctxt->sc);
789 	spin_unlock_irqrestore(&dd->uctxt_lock, flags);
790 
791 	dd->rcd[uctxt->ctxt] = NULL;
792 
793 	hfi1_user_exp_rcv_free(fdata);
794 	hfi1_clear_ctxt_pkey(dd, uctxt->ctxt);
795 
796 	uctxt->rcvwait_to = 0;
797 	uctxt->piowait_to = 0;
798 	uctxt->rcvnowait = 0;
799 	uctxt->pionowait = 0;
800 	uctxt->event_flags = 0;
801 
802 	hfi1_stats.sps_ctxts--;
803 	if (++dd->freectxts == dd->num_user_contexts)
804 		aspm_enable_all(dd);
805 	mutex_unlock(&hfi1_mutex);
806 	hfi1_free_ctxtdata(dd, uctxt);
807 done:
808 	mmdrop(fdata->mm);
809 	kobject_put(&dd->kobj);
810 
811 	if (atomic_dec_and_test(&dd->user_refcount))
812 		complete(&dd->user_comp);
813 
814 	kfree(fdata);
815 	return 0;
816 }
817 
818 /*
819  * Convert kernel *virtual* addresses to physical addresses.
820  * This is used to vmalloc'ed addresses.
821  */
822 static u64 kvirt_to_phys(void *addr)
823 {
824 	struct page *page;
825 	u64 paddr = 0;
826 
827 	page = vmalloc_to_page(addr);
828 	if (page)
829 		paddr = page_to_pfn(page) << PAGE_SHIFT;
830 
831 	return paddr;
832 }
833 
834 static int assign_ctxt(struct file *fp, struct hfi1_user_info *uinfo)
835 {
836 	int i_minor, ret = 0;
837 	unsigned int swmajor, swminor;
838 
839 	swmajor = uinfo->userversion >> 16;
840 	if (swmajor != HFI1_USER_SWMAJOR) {
841 		ret = -ENODEV;
842 		goto done;
843 	}
844 
845 	swminor = uinfo->userversion & 0xffff;
846 
847 	mutex_lock(&hfi1_mutex);
848 	/* First, lets check if we need to setup a shared context? */
849 	if (uinfo->subctxt_cnt) {
850 		struct hfi1_filedata *fd = fp->private_data;
851 
852 		ret = find_shared_ctxt(fp, uinfo);
853 		if (ret < 0)
854 			goto done_unlock;
855 		if (ret) {
856 			fd->rec_cpu_num =
857 				hfi1_get_proc_affinity(fd->uctxt->numa_id);
858 		}
859 	}
860 
861 	/*
862 	 * We execute the following block if we couldn't find a
863 	 * shared context or if context sharing is not required.
864 	 */
865 	if (!ret) {
866 		i_minor = iminor(file_inode(fp)) - HFI1_USER_MINOR_BASE;
867 		ret = get_user_context(fp, uinfo, i_minor);
868 	}
869 done_unlock:
870 	mutex_unlock(&hfi1_mutex);
871 done:
872 	return ret;
873 }
874 
875 static int get_user_context(struct file *fp, struct hfi1_user_info *uinfo,
876 			    int devno)
877 {
878 	struct hfi1_devdata *dd = NULL;
879 	int devmax, npresent, nup;
880 
881 	devmax = hfi1_count_units(&npresent, &nup);
882 	if (!npresent)
883 		return -ENXIO;
884 
885 	if (!nup)
886 		return -ENETDOWN;
887 
888 	dd = hfi1_lookup(devno);
889 	if (!dd)
890 		return -ENODEV;
891 	else if (!dd->freectxts)
892 		return -EBUSY;
893 
894 	return allocate_ctxt(fp, dd, uinfo);
895 }
896 
897 static int find_shared_ctxt(struct file *fp,
898 			    const struct hfi1_user_info *uinfo)
899 {
900 	int devmax, ndev, i;
901 	int ret = 0;
902 	struct hfi1_filedata *fd = fp->private_data;
903 
904 	devmax = hfi1_count_units(NULL, NULL);
905 
906 	for (ndev = 0; ndev < devmax; ndev++) {
907 		struct hfi1_devdata *dd = hfi1_lookup(ndev);
908 
909 		if (!(dd && (dd->flags & HFI1_PRESENT) && dd->kregbase))
910 			continue;
911 		for (i = dd->first_user_ctxt; i < dd->num_rcv_contexts; i++) {
912 			struct hfi1_ctxtdata *uctxt = dd->rcd[i];
913 
914 			/* Skip ctxts which are not yet open */
915 			if (!uctxt || !uctxt->cnt)
916 				continue;
917 			/* Skip ctxt if it doesn't match the requested one */
918 			if (memcmp(uctxt->uuid, uinfo->uuid,
919 				   sizeof(uctxt->uuid)) ||
920 			    uctxt->jkey != generate_jkey(current_uid()) ||
921 			    uctxt->subctxt_id != uinfo->subctxt_id ||
922 			    uctxt->subctxt_cnt != uinfo->subctxt_cnt)
923 				continue;
924 
925 			/* Verify the sharing process matches the master */
926 			if (uctxt->userversion != uinfo->userversion ||
927 			    uctxt->cnt >= uctxt->subctxt_cnt) {
928 				ret = -EINVAL;
929 				goto done;
930 			}
931 			fd->uctxt = uctxt;
932 			fd->subctxt  = uctxt->cnt++;
933 			uctxt->active_slaves |= 1 << fd->subctxt;
934 			ret = 1;
935 			goto done;
936 		}
937 	}
938 
939 done:
940 	return ret;
941 }
942 
943 static int allocate_ctxt(struct file *fp, struct hfi1_devdata *dd,
944 			 struct hfi1_user_info *uinfo)
945 {
946 	struct hfi1_filedata *fd = fp->private_data;
947 	struct hfi1_ctxtdata *uctxt;
948 	unsigned ctxt;
949 	int ret, numa;
950 
951 	if (dd->flags & HFI1_FROZEN) {
952 		/*
953 		 * Pick an error that is unique from all other errors
954 		 * that are returned so the user process knows that
955 		 * it tried to allocate while the SPC was frozen.  It
956 		 * it should be able to retry with success in a short
957 		 * while.
958 		 */
959 		return -EIO;
960 	}
961 
962 	for (ctxt = dd->first_user_ctxt; ctxt < dd->num_rcv_contexts; ctxt++)
963 		if (!dd->rcd[ctxt])
964 			break;
965 
966 	if (ctxt == dd->num_rcv_contexts)
967 		return -EBUSY;
968 
969 	/*
970 	 * If we don't have a NUMA node requested, preference is towards
971 	 * device NUMA node.
972 	 */
973 	fd->rec_cpu_num = hfi1_get_proc_affinity(dd->node);
974 	if (fd->rec_cpu_num != -1)
975 		numa = cpu_to_node(fd->rec_cpu_num);
976 	else
977 		numa = numa_node_id();
978 	uctxt = hfi1_create_ctxtdata(dd->pport, ctxt, numa);
979 	if (!uctxt) {
980 		dd_dev_err(dd,
981 			   "Unable to allocate ctxtdata memory, failing open\n");
982 		return -ENOMEM;
983 	}
984 	hfi1_cdbg(PROC, "[%u:%u] pid %u assigned to CPU %d (NUMA %u)",
985 		  uctxt->ctxt, fd->subctxt, current->pid, fd->rec_cpu_num,
986 		  uctxt->numa_id);
987 
988 	/*
989 	 * Allocate and enable a PIO send context.
990 	 */
991 	uctxt->sc = sc_alloc(dd, SC_USER, uctxt->rcvhdrqentsize,
992 			     uctxt->dd->node);
993 	if (!uctxt->sc) {
994 		ret = -ENOMEM;
995 		goto ctxdata_free;
996 	}
997 	hfi1_cdbg(PROC, "allocated send context %u(%u)\n", uctxt->sc->sw_index,
998 		  uctxt->sc->hw_context);
999 	ret = sc_enable(uctxt->sc);
1000 	if (ret)
1001 		goto ctxdata_free;
1002 
1003 	/*
1004 	 * Setup shared context resources if the user-level has requested
1005 	 * shared contexts and this is the 'master' process.
1006 	 * This has to be done here so the rest of the sub-contexts find the
1007 	 * proper master.
1008 	 */
1009 	if (uinfo->subctxt_cnt && !fd->subctxt) {
1010 		ret = init_subctxts(uctxt, uinfo);
1011 		/*
1012 		 * On error, we don't need to disable and de-allocate the
1013 		 * send context because it will be done during file close
1014 		 */
1015 		if (ret)
1016 			goto ctxdata_free;
1017 	}
1018 	uctxt->userversion = uinfo->userversion;
1019 	uctxt->flags = hfi1_cap_mask; /* save current flag state */
1020 	init_waitqueue_head(&uctxt->wait);
1021 	strlcpy(uctxt->comm, current->comm, sizeof(uctxt->comm));
1022 	memcpy(uctxt->uuid, uinfo->uuid, sizeof(uctxt->uuid));
1023 	uctxt->jkey = generate_jkey(current_uid());
1024 	INIT_LIST_HEAD(&uctxt->sdma_queues);
1025 	spin_lock_init(&uctxt->sdma_qlock);
1026 	hfi1_stats.sps_ctxts++;
1027 	/*
1028 	 * Disable ASPM when there are open user/PSM contexts to avoid
1029 	 * issues with ASPM L1 exit latency
1030 	 */
1031 	if (dd->freectxts-- == dd->num_user_contexts)
1032 		aspm_disable_all(dd);
1033 	fd->uctxt = uctxt;
1034 
1035 	return 0;
1036 
1037 ctxdata_free:
1038 	dd->rcd[ctxt] = NULL;
1039 	hfi1_free_ctxtdata(dd, uctxt);
1040 	return ret;
1041 }
1042 
1043 static int init_subctxts(struct hfi1_ctxtdata *uctxt,
1044 			 const struct hfi1_user_info *uinfo)
1045 {
1046 	unsigned num_subctxts;
1047 
1048 	num_subctxts = uinfo->subctxt_cnt;
1049 	if (num_subctxts > HFI1_MAX_SHARED_CTXTS)
1050 		return -EINVAL;
1051 
1052 	uctxt->subctxt_cnt = uinfo->subctxt_cnt;
1053 	uctxt->subctxt_id = uinfo->subctxt_id;
1054 	uctxt->active_slaves = 1;
1055 	uctxt->redirect_seq_cnt = 1;
1056 	set_bit(HFI1_CTXT_MASTER_UNINIT, &uctxt->event_flags);
1057 
1058 	return 0;
1059 }
1060 
1061 static int setup_subctxt(struct hfi1_ctxtdata *uctxt)
1062 {
1063 	int ret = 0;
1064 	unsigned num_subctxts = uctxt->subctxt_cnt;
1065 
1066 	uctxt->subctxt_uregbase = vmalloc_user(PAGE_SIZE);
1067 	if (!uctxt->subctxt_uregbase) {
1068 		ret = -ENOMEM;
1069 		goto bail;
1070 	}
1071 	/* We can take the size of the RcvHdr Queue from the master */
1072 	uctxt->subctxt_rcvhdr_base = vmalloc_user(uctxt->rcvhdrq_size *
1073 						  num_subctxts);
1074 	if (!uctxt->subctxt_rcvhdr_base) {
1075 		ret = -ENOMEM;
1076 		goto bail_ureg;
1077 	}
1078 
1079 	uctxt->subctxt_rcvegrbuf = vmalloc_user(uctxt->egrbufs.size *
1080 						num_subctxts);
1081 	if (!uctxt->subctxt_rcvegrbuf) {
1082 		ret = -ENOMEM;
1083 		goto bail_rhdr;
1084 	}
1085 	goto bail;
1086 bail_rhdr:
1087 	vfree(uctxt->subctxt_rcvhdr_base);
1088 bail_ureg:
1089 	vfree(uctxt->subctxt_uregbase);
1090 	uctxt->subctxt_uregbase = NULL;
1091 bail:
1092 	return ret;
1093 }
1094 
1095 static int user_init(struct file *fp)
1096 {
1097 	unsigned int rcvctrl_ops = 0;
1098 	struct hfi1_filedata *fd = fp->private_data;
1099 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1100 
1101 	/* make sure that the context has already been setup */
1102 	if (!test_bit(HFI1_CTXT_SETUP_DONE, &uctxt->event_flags))
1103 		return -EFAULT;
1104 
1105 	/* initialize poll variables... */
1106 	uctxt->urgent = 0;
1107 	uctxt->urgent_poll = 0;
1108 
1109 	/*
1110 	 * Now enable the ctxt for receive.
1111 	 * For chips that are set to DMA the tail register to memory
1112 	 * when they change (and when the update bit transitions from
1113 	 * 0 to 1.  So for those chips, we turn it off and then back on.
1114 	 * This will (very briefly) affect any other open ctxts, but the
1115 	 * duration is very short, and therefore isn't an issue.  We
1116 	 * explicitly set the in-memory tail copy to 0 beforehand, so we
1117 	 * don't have to wait to be sure the DMA update has happened
1118 	 * (chip resets head/tail to 0 on transition to enable).
1119 	 */
1120 	if (uctxt->rcvhdrtail_kvaddr)
1121 		clear_rcvhdrtail(uctxt);
1122 
1123 	/* Setup J_KEY before enabling the context */
1124 	hfi1_set_ctxt_jkey(uctxt->dd, uctxt->ctxt, uctxt->jkey);
1125 
1126 	rcvctrl_ops = HFI1_RCVCTRL_CTXT_ENB;
1127 	if (HFI1_CAP_UGET_MASK(uctxt->flags, HDRSUPP))
1128 		rcvctrl_ops |= HFI1_RCVCTRL_TIDFLOW_ENB;
1129 	/*
1130 	 * Ignore the bit in the flags for now until proper
1131 	 * support for multiple packet per rcv array entry is
1132 	 * added.
1133 	 */
1134 	if (!HFI1_CAP_UGET_MASK(uctxt->flags, MULTI_PKT_EGR))
1135 		rcvctrl_ops |= HFI1_RCVCTRL_ONE_PKT_EGR_ENB;
1136 	if (HFI1_CAP_UGET_MASK(uctxt->flags, NODROP_EGR_FULL))
1137 		rcvctrl_ops |= HFI1_RCVCTRL_NO_EGR_DROP_ENB;
1138 	if (HFI1_CAP_UGET_MASK(uctxt->flags, NODROP_RHQ_FULL))
1139 		rcvctrl_ops |= HFI1_RCVCTRL_NO_RHQ_DROP_ENB;
1140 	/*
1141 	 * The RcvCtxtCtrl.TailUpd bit has to be explicitly written.
1142 	 * We can't rely on the correct value to be set from prior
1143 	 * uses of the chip or ctxt. Therefore, add the rcvctrl op
1144 	 * for both cases.
1145 	 */
1146 	if (HFI1_CAP_UGET_MASK(uctxt->flags, DMA_RTAIL))
1147 		rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_ENB;
1148 	else
1149 		rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_DIS;
1150 	hfi1_rcvctrl(uctxt->dd, rcvctrl_ops, uctxt->ctxt);
1151 
1152 	/* Notify any waiting slaves */
1153 	if (uctxt->subctxt_cnt) {
1154 		clear_bit(HFI1_CTXT_MASTER_UNINIT, &uctxt->event_flags);
1155 		wake_up(&uctxt->wait);
1156 	}
1157 
1158 	return 0;
1159 }
1160 
1161 static int get_ctxt_info(struct file *fp, void __user *ubase, __u32 len)
1162 {
1163 	struct hfi1_ctxt_info cinfo;
1164 	struct hfi1_filedata *fd = fp->private_data;
1165 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1166 	int ret = 0;
1167 
1168 	memset(&cinfo, 0, sizeof(cinfo));
1169 	cinfo.runtime_flags = (((uctxt->flags >> HFI1_CAP_MISC_SHIFT) &
1170 				HFI1_CAP_MISC_MASK) << HFI1_CAP_USER_SHIFT) |
1171 			HFI1_CAP_UGET_MASK(uctxt->flags, MASK) |
1172 			HFI1_CAP_KGET_MASK(uctxt->flags, K2U);
1173 	/* adjust flag if this fd is not able to cache */
1174 	if (!fd->handler)
1175 		cinfo.runtime_flags |= HFI1_CAP_TID_UNMAP; /* no caching */
1176 
1177 	cinfo.num_active = hfi1_count_active_units();
1178 	cinfo.unit = uctxt->dd->unit;
1179 	cinfo.ctxt = uctxt->ctxt;
1180 	cinfo.subctxt = fd->subctxt;
1181 	cinfo.rcvtids = roundup(uctxt->egrbufs.alloced,
1182 				uctxt->dd->rcv_entries.group_size) +
1183 		uctxt->expected_count;
1184 	cinfo.credits = uctxt->sc->credits;
1185 	cinfo.numa_node = uctxt->numa_id;
1186 	cinfo.rec_cpu = fd->rec_cpu_num;
1187 	cinfo.send_ctxt = uctxt->sc->hw_context;
1188 
1189 	cinfo.egrtids = uctxt->egrbufs.alloced;
1190 	cinfo.rcvhdrq_cnt = uctxt->rcvhdrq_cnt;
1191 	cinfo.rcvhdrq_entsize = uctxt->rcvhdrqentsize << 2;
1192 	cinfo.sdma_ring_size = fd->cq->nentries;
1193 	cinfo.rcvegr_size = uctxt->egrbufs.rcvtid_size;
1194 
1195 	trace_hfi1_ctxt_info(uctxt->dd, uctxt->ctxt, fd->subctxt, cinfo);
1196 	if (copy_to_user(ubase, &cinfo, sizeof(cinfo)))
1197 		ret = -EFAULT;
1198 
1199 	return ret;
1200 }
1201 
1202 static int setup_ctxt(struct file *fp)
1203 {
1204 	struct hfi1_filedata *fd = fp->private_data;
1205 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1206 	struct hfi1_devdata *dd = uctxt->dd;
1207 	int ret = 0;
1208 
1209 	/*
1210 	 * Context should be set up only once, including allocation and
1211 	 * programming of eager buffers. This is done if context sharing
1212 	 * is not requested or by the master process.
1213 	 */
1214 	if (!uctxt->subctxt_cnt || !fd->subctxt) {
1215 		ret = hfi1_init_ctxt(uctxt->sc);
1216 		if (ret)
1217 			goto done;
1218 
1219 		/* Now allocate the RcvHdr queue and eager buffers. */
1220 		ret = hfi1_create_rcvhdrq(dd, uctxt);
1221 		if (ret)
1222 			goto done;
1223 		ret = hfi1_setup_eagerbufs(uctxt);
1224 		if (ret)
1225 			goto done;
1226 		if (uctxt->subctxt_cnt && !fd->subctxt) {
1227 			ret = setup_subctxt(uctxt);
1228 			if (ret)
1229 				goto done;
1230 		}
1231 	} else {
1232 		ret = wait_event_interruptible(uctxt->wait, !test_bit(
1233 					       HFI1_CTXT_MASTER_UNINIT,
1234 					       &uctxt->event_flags));
1235 		if (ret)
1236 			goto done;
1237 	}
1238 
1239 	ret = hfi1_user_sdma_alloc_queues(uctxt, fp);
1240 	if (ret)
1241 		goto done;
1242 	/*
1243 	 * Expected receive has to be setup for all processes (including
1244 	 * shared contexts). However, it has to be done after the master
1245 	 * context has been fully configured as it depends on the
1246 	 * eager/expected split of the RcvArray entries.
1247 	 * Setting it up here ensures that the subcontexts will be waiting
1248 	 * (due to the above wait_event_interruptible() until the master
1249 	 * is setup.
1250 	 */
1251 	ret = hfi1_user_exp_rcv_init(fp);
1252 	if (ret)
1253 		goto done;
1254 
1255 	set_bit(HFI1_CTXT_SETUP_DONE, &uctxt->event_flags);
1256 done:
1257 	return ret;
1258 }
1259 
1260 static int get_base_info(struct file *fp, void __user *ubase, __u32 len)
1261 {
1262 	struct hfi1_base_info binfo;
1263 	struct hfi1_filedata *fd = fp->private_data;
1264 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1265 	struct hfi1_devdata *dd = uctxt->dd;
1266 	ssize_t sz;
1267 	unsigned offset;
1268 	int ret = 0;
1269 
1270 	trace_hfi1_uctxtdata(uctxt->dd, uctxt);
1271 
1272 	memset(&binfo, 0, sizeof(binfo));
1273 	binfo.hw_version = dd->revision;
1274 	binfo.sw_version = HFI1_KERN_SWVERSION;
1275 	binfo.bthqp = kdeth_qp;
1276 	binfo.jkey = uctxt->jkey;
1277 	/*
1278 	 * If more than 64 contexts are enabled the allocated credit
1279 	 * return will span two or three contiguous pages. Since we only
1280 	 * map the page containing the context's credit return address,
1281 	 * we need to calculate the offset in the proper page.
1282 	 */
1283 	offset = ((u64)uctxt->sc->hw_free -
1284 		  (u64)dd->cr_base[uctxt->numa_id].va) % PAGE_SIZE;
1285 	binfo.sc_credits_addr = HFI1_MMAP_TOKEN(PIO_CRED, uctxt->ctxt,
1286 						fd->subctxt, offset);
1287 	binfo.pio_bufbase = HFI1_MMAP_TOKEN(PIO_BUFS, uctxt->ctxt,
1288 					    fd->subctxt,
1289 					    uctxt->sc->base_addr);
1290 	binfo.pio_bufbase_sop = HFI1_MMAP_TOKEN(PIO_BUFS_SOP,
1291 						uctxt->ctxt,
1292 						fd->subctxt,
1293 						uctxt->sc->base_addr);
1294 	binfo.rcvhdr_bufbase = HFI1_MMAP_TOKEN(RCV_HDRQ, uctxt->ctxt,
1295 					       fd->subctxt,
1296 					       uctxt->rcvhdrq);
1297 	binfo.rcvegr_bufbase = HFI1_MMAP_TOKEN(RCV_EGRBUF, uctxt->ctxt,
1298 					       fd->subctxt,
1299 					       uctxt->egrbufs.rcvtids[0].dma);
1300 	binfo.sdma_comp_bufbase = HFI1_MMAP_TOKEN(SDMA_COMP, uctxt->ctxt,
1301 						 fd->subctxt, 0);
1302 	/*
1303 	 * user regs are at
1304 	 * (RXE_PER_CONTEXT_USER + (ctxt * RXE_PER_CONTEXT_SIZE))
1305 	 */
1306 	binfo.user_regbase = HFI1_MMAP_TOKEN(UREGS, uctxt->ctxt,
1307 					    fd->subctxt, 0);
1308 	offset = offset_in_page((((uctxt->ctxt - dd->first_user_ctxt) *
1309 		    HFI1_MAX_SHARED_CTXTS) + fd->subctxt) *
1310 		  sizeof(*dd->events));
1311 	binfo.events_bufbase = HFI1_MMAP_TOKEN(EVENTS, uctxt->ctxt,
1312 					      fd->subctxt,
1313 					      offset);
1314 	binfo.status_bufbase = HFI1_MMAP_TOKEN(STATUS, uctxt->ctxt,
1315 					      fd->subctxt,
1316 					      dd->status);
1317 	if (HFI1_CAP_IS_USET(DMA_RTAIL))
1318 		binfo.rcvhdrtail_base = HFI1_MMAP_TOKEN(RTAIL, uctxt->ctxt,
1319 						       fd->subctxt, 0);
1320 	if (uctxt->subctxt_cnt) {
1321 		binfo.subctxt_uregbase = HFI1_MMAP_TOKEN(SUBCTXT_UREGS,
1322 							uctxt->ctxt,
1323 							fd->subctxt, 0);
1324 		binfo.subctxt_rcvhdrbuf = HFI1_MMAP_TOKEN(SUBCTXT_RCV_HDRQ,
1325 							 uctxt->ctxt,
1326 							 fd->subctxt, 0);
1327 		binfo.subctxt_rcvegrbuf = HFI1_MMAP_TOKEN(SUBCTXT_EGRBUF,
1328 							 uctxt->ctxt,
1329 							 fd->subctxt, 0);
1330 	}
1331 	sz = (len < sizeof(binfo)) ? len : sizeof(binfo);
1332 	if (copy_to_user(ubase, &binfo, sz))
1333 		ret = -EFAULT;
1334 	return ret;
1335 }
1336 
1337 static unsigned int poll_urgent(struct file *fp,
1338 				struct poll_table_struct *pt)
1339 {
1340 	struct hfi1_filedata *fd = fp->private_data;
1341 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1342 	struct hfi1_devdata *dd = uctxt->dd;
1343 	unsigned pollflag;
1344 
1345 	poll_wait(fp, &uctxt->wait, pt);
1346 
1347 	spin_lock_irq(&dd->uctxt_lock);
1348 	if (uctxt->urgent != uctxt->urgent_poll) {
1349 		pollflag = POLLIN | POLLRDNORM;
1350 		uctxt->urgent_poll = uctxt->urgent;
1351 	} else {
1352 		pollflag = 0;
1353 		set_bit(HFI1_CTXT_WAITING_URG, &uctxt->event_flags);
1354 	}
1355 	spin_unlock_irq(&dd->uctxt_lock);
1356 
1357 	return pollflag;
1358 }
1359 
1360 static unsigned int poll_next(struct file *fp,
1361 			      struct poll_table_struct *pt)
1362 {
1363 	struct hfi1_filedata *fd = fp->private_data;
1364 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1365 	struct hfi1_devdata *dd = uctxt->dd;
1366 	unsigned pollflag;
1367 
1368 	poll_wait(fp, &uctxt->wait, pt);
1369 
1370 	spin_lock_irq(&dd->uctxt_lock);
1371 	if (hdrqempty(uctxt)) {
1372 		set_bit(HFI1_CTXT_WAITING_RCV, &uctxt->event_flags);
1373 		hfi1_rcvctrl(dd, HFI1_RCVCTRL_INTRAVAIL_ENB, uctxt->ctxt);
1374 		pollflag = 0;
1375 	} else {
1376 		pollflag = POLLIN | POLLRDNORM;
1377 	}
1378 	spin_unlock_irq(&dd->uctxt_lock);
1379 
1380 	return pollflag;
1381 }
1382 
1383 /*
1384  * Find all user contexts in use, and set the specified bit in their
1385  * event mask.
1386  * See also find_ctxt() for a similar use, that is specific to send buffers.
1387  */
1388 int hfi1_set_uevent_bits(struct hfi1_pportdata *ppd, const int evtbit)
1389 {
1390 	struct hfi1_ctxtdata *uctxt;
1391 	struct hfi1_devdata *dd = ppd->dd;
1392 	unsigned ctxt;
1393 	int ret = 0;
1394 	unsigned long flags;
1395 
1396 	if (!dd->events) {
1397 		ret = -EINVAL;
1398 		goto done;
1399 	}
1400 
1401 	spin_lock_irqsave(&dd->uctxt_lock, flags);
1402 	for (ctxt = dd->first_user_ctxt; ctxt < dd->num_rcv_contexts;
1403 	     ctxt++) {
1404 		uctxt = dd->rcd[ctxt];
1405 		if (uctxt) {
1406 			unsigned long *evs = dd->events +
1407 				(uctxt->ctxt - dd->first_user_ctxt) *
1408 				HFI1_MAX_SHARED_CTXTS;
1409 			int i;
1410 			/*
1411 			 * subctxt_cnt is 0 if not shared, so do base
1412 			 * separately, first, then remaining subctxt, if any
1413 			 */
1414 			set_bit(evtbit, evs);
1415 			for (i = 1; i < uctxt->subctxt_cnt; i++)
1416 				set_bit(evtbit, evs + i);
1417 		}
1418 	}
1419 	spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1420 done:
1421 	return ret;
1422 }
1423 
1424 /**
1425  * manage_rcvq - manage a context's receive queue
1426  * @uctxt: the context
1427  * @subctxt: the sub-context
1428  * @start_stop: action to carry out
1429  *
1430  * start_stop == 0 disables receive on the context, for use in queue
1431  * overflow conditions.  start_stop==1 re-enables, to be used to
1432  * re-init the software copy of the head register
1433  */
1434 static int manage_rcvq(struct hfi1_ctxtdata *uctxt, unsigned subctxt,
1435 		       int start_stop)
1436 {
1437 	struct hfi1_devdata *dd = uctxt->dd;
1438 	unsigned int rcvctrl_op;
1439 
1440 	if (subctxt)
1441 		goto bail;
1442 	/* atomically clear receive enable ctxt. */
1443 	if (start_stop) {
1444 		/*
1445 		 * On enable, force in-memory copy of the tail register to
1446 		 * 0, so that protocol code doesn't have to worry about
1447 		 * whether or not the chip has yet updated the in-memory
1448 		 * copy or not on return from the system call. The chip
1449 		 * always resets it's tail register back to 0 on a
1450 		 * transition from disabled to enabled.
1451 		 */
1452 		if (uctxt->rcvhdrtail_kvaddr)
1453 			clear_rcvhdrtail(uctxt);
1454 		rcvctrl_op = HFI1_RCVCTRL_CTXT_ENB;
1455 	} else {
1456 		rcvctrl_op = HFI1_RCVCTRL_CTXT_DIS;
1457 	}
1458 	hfi1_rcvctrl(dd, rcvctrl_op, uctxt->ctxt);
1459 	/* always; new head should be equal to new tail; see above */
1460 bail:
1461 	return 0;
1462 }
1463 
1464 /*
1465  * clear the event notifier events for this context.
1466  * User process then performs actions appropriate to bit having been
1467  * set, if desired, and checks again in future.
1468  */
1469 static int user_event_ack(struct hfi1_ctxtdata *uctxt, int subctxt,
1470 			  unsigned long events)
1471 {
1472 	int i;
1473 	struct hfi1_devdata *dd = uctxt->dd;
1474 	unsigned long *evs;
1475 
1476 	if (!dd->events)
1477 		return 0;
1478 
1479 	evs = dd->events + ((uctxt->ctxt - dd->first_user_ctxt) *
1480 			    HFI1_MAX_SHARED_CTXTS) + subctxt;
1481 
1482 	for (i = 0; i <= _HFI1_MAX_EVENT_BIT; i++) {
1483 		if (!test_bit(i, &events))
1484 			continue;
1485 		clear_bit(i, evs);
1486 	}
1487 	return 0;
1488 }
1489 
1490 static int set_ctxt_pkey(struct hfi1_ctxtdata *uctxt, unsigned subctxt,
1491 			 u16 pkey)
1492 {
1493 	int ret = -ENOENT, i, intable = 0;
1494 	struct hfi1_pportdata *ppd = uctxt->ppd;
1495 	struct hfi1_devdata *dd = uctxt->dd;
1496 
1497 	if (pkey == LIM_MGMT_P_KEY || pkey == FULL_MGMT_P_KEY) {
1498 		ret = -EINVAL;
1499 		goto done;
1500 	}
1501 
1502 	for (i = 0; i < ARRAY_SIZE(ppd->pkeys); i++)
1503 		if (pkey == ppd->pkeys[i]) {
1504 			intable = 1;
1505 			break;
1506 		}
1507 
1508 	if (intable)
1509 		ret = hfi1_set_ctxt_pkey(dd, uctxt->ctxt, pkey);
1510 done:
1511 	return ret;
1512 }
1513 
1514 static void user_remove(struct hfi1_devdata *dd)
1515 {
1516 
1517 	hfi1_cdev_cleanup(&dd->user_cdev, &dd->user_device);
1518 }
1519 
1520 static int user_add(struct hfi1_devdata *dd)
1521 {
1522 	char name[10];
1523 	int ret;
1524 
1525 	snprintf(name, sizeof(name), "%s_%d", class_name(), dd->unit);
1526 	ret = hfi1_cdev_init(dd->unit, name, &hfi1_file_ops,
1527 			     &dd->user_cdev, &dd->user_device,
1528 			     true, &dd->kobj);
1529 	if (ret)
1530 		user_remove(dd);
1531 
1532 	return ret;
1533 }
1534 
1535 /*
1536  * Create per-unit files in /dev
1537  */
1538 int hfi1_device_create(struct hfi1_devdata *dd)
1539 {
1540 	return user_add(dd);
1541 }
1542 
1543 /*
1544  * Remove per-unit files in /dev
1545  * void, core kernel returns no errors for this stuff
1546  */
1547 void hfi1_device_remove(struct hfi1_devdata *dd)
1548 {
1549 	user_remove(dd);
1550 }
1551