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, unsigned long arg, u32 len);
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, unsigned long arg, u32 len);
88 static int get_base_info(struct hfi1_filedata *fd, unsigned long arg, u32 len);
89 static int user_exp_rcv_setup(struct hfi1_filedata *fd, unsigned long arg,
90 			      u32 len);
91 static int user_exp_rcv_clear(struct hfi1_filedata *fd, unsigned long arg,
92 			      u32 len);
93 static int user_exp_rcv_invalid(struct hfi1_filedata *fd, unsigned long arg,
94 				u32 len);
95 static int setup_base_ctxt(struct hfi1_filedata *fd,
96 			   struct hfi1_ctxtdata *uctxt);
97 static int setup_subctxt(struct hfi1_ctxtdata *uctxt);
98 
99 static int find_sub_ctxt(struct hfi1_filedata *fd,
100 			 const struct hfi1_user_info *uinfo);
101 static int allocate_ctxt(struct hfi1_filedata *fd, struct hfi1_devdata *dd,
102 			 struct hfi1_user_info *uinfo,
103 			 struct hfi1_ctxtdata **cd);
104 static void deallocate_ctxt(struct hfi1_ctxtdata *uctxt);
105 static unsigned int poll_urgent(struct file *fp, struct poll_table_struct *pt);
106 static unsigned int poll_next(struct file *fp, struct poll_table_struct *pt);
107 static int user_event_ack(struct hfi1_ctxtdata *uctxt, u16 subctxt,
108 			  unsigned long arg);
109 static int set_ctxt_pkey(struct hfi1_ctxtdata *uctxt, unsigned long arg);
110 static int ctxt_reset(struct hfi1_ctxtdata *uctxt);
111 static int manage_rcvq(struct hfi1_ctxtdata *uctxt, u16 subctxt,
112 		       unsigned long arg);
113 static int vma_fault(struct vm_fault *vmf);
114 static long hfi1_file_ioctl(struct file *fp, unsigned int cmd,
115 			    unsigned long arg);
116 
117 static const struct file_operations hfi1_file_ops = {
118 	.owner = THIS_MODULE,
119 	.write_iter = hfi1_write_iter,
120 	.open = hfi1_file_open,
121 	.release = hfi1_file_close,
122 	.unlocked_ioctl = hfi1_file_ioctl,
123 	.poll = hfi1_poll,
124 	.mmap = hfi1_file_mmap,
125 	.llseek = noop_llseek,
126 };
127 
128 static const struct vm_operations_struct vm_ops = {
129 	.fault = vma_fault,
130 };
131 
132 /*
133  * Types of memories mapped into user processes' space
134  */
135 enum mmap_types {
136 	PIO_BUFS = 1,
137 	PIO_BUFS_SOP,
138 	PIO_CRED,
139 	RCV_HDRQ,
140 	RCV_EGRBUF,
141 	UREGS,
142 	EVENTS,
143 	STATUS,
144 	RTAIL,
145 	SUBCTXT_UREGS,
146 	SUBCTXT_RCV_HDRQ,
147 	SUBCTXT_EGRBUF,
148 	SDMA_COMP
149 };
150 
151 /*
152  * Masks and offsets defining the mmap tokens
153  */
154 #define HFI1_MMAP_OFFSET_MASK   0xfffULL
155 #define HFI1_MMAP_OFFSET_SHIFT  0
156 #define HFI1_MMAP_SUBCTXT_MASK  0xfULL
157 #define HFI1_MMAP_SUBCTXT_SHIFT 12
158 #define HFI1_MMAP_CTXT_MASK     0xffULL
159 #define HFI1_MMAP_CTXT_SHIFT    16
160 #define HFI1_MMAP_TYPE_MASK     0xfULL
161 #define HFI1_MMAP_TYPE_SHIFT    24
162 #define HFI1_MMAP_MAGIC_MASK    0xffffffffULL
163 #define HFI1_MMAP_MAGIC_SHIFT   32
164 
165 #define HFI1_MMAP_MAGIC         0xdabbad00
166 
167 #define HFI1_MMAP_TOKEN_SET(field, val)	\
168 	(((val) & HFI1_MMAP_##field##_MASK) << HFI1_MMAP_##field##_SHIFT)
169 #define HFI1_MMAP_TOKEN_GET(field, token) \
170 	(((token) >> HFI1_MMAP_##field##_SHIFT) & HFI1_MMAP_##field##_MASK)
171 #define HFI1_MMAP_TOKEN(type, ctxt, subctxt, addr)   \
172 	(HFI1_MMAP_TOKEN_SET(MAGIC, HFI1_MMAP_MAGIC) | \
173 	HFI1_MMAP_TOKEN_SET(TYPE, type) | \
174 	HFI1_MMAP_TOKEN_SET(CTXT, ctxt) | \
175 	HFI1_MMAP_TOKEN_SET(SUBCTXT, subctxt) | \
176 	HFI1_MMAP_TOKEN_SET(OFFSET, (offset_in_page(addr))))
177 
178 #define dbg(fmt, ...)				\
179 	pr_info(fmt, ##__VA_ARGS__)
180 
181 static inline int is_valid_mmap(u64 token)
182 {
183 	return (HFI1_MMAP_TOKEN_GET(MAGIC, token) == HFI1_MMAP_MAGIC);
184 }
185 
186 static int hfi1_file_open(struct inode *inode, struct file *fp)
187 {
188 	struct hfi1_filedata *fd;
189 	struct hfi1_devdata *dd = container_of(inode->i_cdev,
190 					       struct hfi1_devdata,
191 					       user_cdev);
192 
193 	if (!((dd->flags & HFI1_PRESENT) && dd->kregbase1))
194 		return -EINVAL;
195 
196 	if (!atomic_inc_not_zero(&dd->user_refcount))
197 		return -ENXIO;
198 
199 	/* Just take a ref now. Not all opens result in a context assign */
200 	kobject_get(&dd->kobj);
201 
202 	/* The real work is performed later in assign_ctxt() */
203 
204 	fd = kzalloc(sizeof(*fd), GFP_KERNEL);
205 
206 	if (fd) {
207 		fd->rec_cpu_num = -1; /* no cpu affinity by default */
208 		fd->mm = current->mm;
209 		mmgrab(fd->mm);
210 		fd->dd = dd;
211 		fp->private_data = fd;
212 	} else {
213 		fp->private_data = NULL;
214 
215 		if (atomic_dec_and_test(&dd->user_refcount))
216 			complete(&dd->user_comp);
217 
218 		return -ENOMEM;
219 	}
220 
221 	return 0;
222 }
223 
224 static long hfi1_file_ioctl(struct file *fp, unsigned int cmd,
225 			    unsigned long arg)
226 {
227 	struct hfi1_filedata *fd = fp->private_data;
228 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
229 	int ret = 0;
230 	int uval = 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 		ret = assign_ctxt(fd, arg, _IOC_SIZE(cmd));
241 		break;
242 
243 	case HFI1_IOCTL_CTXT_INFO:
244 		ret = get_ctxt_info(fd, arg, _IOC_SIZE(cmd));
245 		break;
246 
247 	case HFI1_IOCTL_USER_INFO:
248 		ret = get_base_info(fd, arg, _IOC_SIZE(cmd));
249 		break;
250 
251 	case HFI1_IOCTL_CREDIT_UPD:
252 		if (uctxt)
253 			sc_return_credits(uctxt->sc);
254 		break;
255 
256 	case HFI1_IOCTL_TID_UPDATE:
257 		ret = user_exp_rcv_setup(fd, arg, _IOC_SIZE(cmd));
258 		break;
259 
260 	case HFI1_IOCTL_TID_FREE:
261 		ret = user_exp_rcv_clear(fd, arg, _IOC_SIZE(cmd));
262 		break;
263 
264 	case HFI1_IOCTL_TID_INVAL_READ:
265 		ret = user_exp_rcv_invalid(fd, arg, _IOC_SIZE(cmd));
266 		break;
267 
268 	case HFI1_IOCTL_RECV_CTRL:
269 		ret = manage_rcvq(uctxt, fd->subctxt, arg);
270 		break;
271 
272 	case HFI1_IOCTL_POLL_TYPE:
273 		if (get_user(uval, (int __user *)arg))
274 			return -EFAULT;
275 		uctxt->poll_type = (typeof(uctxt->poll_type))uval;
276 		break;
277 
278 	case HFI1_IOCTL_ACK_EVENT:
279 		ret = user_event_ack(uctxt, fd->subctxt, arg);
280 		break;
281 
282 	case HFI1_IOCTL_SET_PKEY:
283 		ret = set_ctxt_pkey(uctxt, arg);
284 		break;
285 
286 	case HFI1_IOCTL_CTXT_RESET:
287 		ret = ctxt_reset(uctxt);
288 		break;
289 
290 	case HFI1_IOCTL_GET_VERS:
291 		uval = HFI1_USER_SWVERSION;
292 		if (put_user(uval, (int __user *)arg))
293 			return -EFAULT;
294 		break;
295 
296 	default:
297 		return -EINVAL;
298 	}
299 
300 	return ret;
301 }
302 
303 static ssize_t hfi1_write_iter(struct kiocb *kiocb, struct iov_iter *from)
304 {
305 	struct hfi1_filedata *fd = kiocb->ki_filp->private_data;
306 	struct hfi1_user_sdma_pkt_q *pq = fd->pq;
307 	struct hfi1_user_sdma_comp_q *cq = fd->cq;
308 	int done = 0, reqs = 0;
309 	unsigned long dim = from->nr_segs;
310 
311 	if (!cq || !pq)
312 		return -EIO;
313 
314 	if (!iter_is_iovec(from) || !dim)
315 		return -EINVAL;
316 
317 	trace_hfi1_sdma_request(fd->dd, fd->uctxt->ctxt, fd->subctxt, dim);
318 
319 	if (atomic_read(&pq->n_reqs) == pq->n_max_reqs)
320 		return -ENOSPC;
321 
322 	while (dim) {
323 		int ret;
324 		unsigned long count = 0;
325 
326 		ret = hfi1_user_sdma_process_request(
327 			fd, (struct iovec *)(from->iov + done),
328 			dim, &count);
329 		if (ret) {
330 			reqs = ret;
331 			break;
332 		}
333 		dim -= count;
334 		done += count;
335 		reqs++;
336 	}
337 
338 	return reqs;
339 }
340 
341 static int hfi1_file_mmap(struct file *fp, struct vm_area_struct *vma)
342 {
343 	struct hfi1_filedata *fd = fp->private_data;
344 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
345 	struct hfi1_devdata *dd;
346 	unsigned long flags;
347 	u64 token = vma->vm_pgoff << PAGE_SHIFT,
348 		memaddr = 0;
349 	void *memvirt = NULL;
350 	u8 subctxt, mapio = 0, vmf = 0, type;
351 	ssize_t memlen = 0;
352 	int ret = 0;
353 	u16 ctxt;
354 
355 	if (!is_valid_mmap(token) || !uctxt ||
356 	    !(vma->vm_flags & VM_SHARED)) {
357 		ret = -EINVAL;
358 		goto done;
359 	}
360 	dd = uctxt->dd;
361 	ctxt = HFI1_MMAP_TOKEN_GET(CTXT, token);
362 	subctxt = HFI1_MMAP_TOKEN_GET(SUBCTXT, token);
363 	type = HFI1_MMAP_TOKEN_GET(TYPE, token);
364 	if (ctxt != uctxt->ctxt || subctxt != fd->subctxt) {
365 		ret = -EINVAL;
366 		goto done;
367 	}
368 
369 	flags = vma->vm_flags;
370 
371 	switch (type) {
372 	case PIO_BUFS:
373 	case PIO_BUFS_SOP:
374 		memaddr = ((dd->physaddr + TXE_PIO_SEND) +
375 				/* chip pio base */
376 			   (uctxt->sc->hw_context * BIT(16))) +
377 				/* 64K PIO space / ctxt */
378 			(type == PIO_BUFS_SOP ?
379 				(TXE_PIO_SIZE / 2) : 0); /* sop? */
380 		/*
381 		 * Map only the amount allocated to the context, not the
382 		 * entire available context's PIO space.
383 		 */
384 		memlen = PAGE_ALIGN(uctxt->sc->credits * PIO_BLOCK_SIZE);
385 		flags &= ~VM_MAYREAD;
386 		flags |= VM_DONTCOPY | VM_DONTEXPAND;
387 		vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
388 		mapio = 1;
389 		break;
390 	case PIO_CRED:
391 		if (flags & VM_WRITE) {
392 			ret = -EPERM;
393 			goto done;
394 		}
395 		/*
396 		 * The credit return location for this context could be on the
397 		 * second or third page allocated for credit returns (if number
398 		 * of enabled contexts > 64 and 128 respectively).
399 		 */
400 		memvirt = dd->cr_base[uctxt->numa_id].va;
401 		memaddr = virt_to_phys(memvirt) +
402 			(((u64)uctxt->sc->hw_free -
403 			  (u64)dd->cr_base[uctxt->numa_id].va) & PAGE_MASK);
404 		memlen = PAGE_SIZE;
405 		flags &= ~VM_MAYWRITE;
406 		flags |= VM_DONTCOPY | VM_DONTEXPAND;
407 		/*
408 		 * The driver has already allocated memory for credit
409 		 * returns and programmed it into the chip. Has that
410 		 * memory been flagged as non-cached?
411 		 */
412 		/* vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); */
413 		mapio = 1;
414 		break;
415 	case RCV_HDRQ:
416 		memlen = uctxt->rcvhdrq_size;
417 		memvirt = uctxt->rcvhdrq;
418 		break;
419 	case RCV_EGRBUF: {
420 		unsigned long addr;
421 		int i;
422 		/*
423 		 * The RcvEgr buffer need to be handled differently
424 		 * as multiple non-contiguous pages need to be mapped
425 		 * into the user process.
426 		 */
427 		memlen = uctxt->egrbufs.size;
428 		if ((vma->vm_end - vma->vm_start) != memlen) {
429 			dd_dev_err(dd, "Eager buffer map size invalid (%lu != %lu)\n",
430 				   (vma->vm_end - vma->vm_start), memlen);
431 			ret = -EINVAL;
432 			goto done;
433 		}
434 		if (vma->vm_flags & VM_WRITE) {
435 			ret = -EPERM;
436 			goto done;
437 		}
438 		vma->vm_flags &= ~VM_MAYWRITE;
439 		addr = vma->vm_start;
440 		for (i = 0 ; i < uctxt->egrbufs.numbufs; i++) {
441 			memlen = uctxt->egrbufs.buffers[i].len;
442 			memvirt = uctxt->egrbufs.buffers[i].addr;
443 			ret = remap_pfn_range(
444 				vma, addr,
445 				/*
446 				 * virt_to_pfn() does the same, but
447 				 * it's not available on x86_64
448 				 * when CONFIG_MMU is enabled.
449 				 */
450 				PFN_DOWN(__pa(memvirt)),
451 				memlen,
452 				vma->vm_page_prot);
453 			if (ret < 0)
454 				goto done;
455 			addr += memlen;
456 		}
457 		ret = 0;
458 		goto done;
459 	}
460 	case UREGS:
461 		/*
462 		 * Map only the page that contains this context's user
463 		 * registers.
464 		 */
465 		memaddr = (unsigned long)
466 			(dd->physaddr + RXE_PER_CONTEXT_USER)
467 			+ (uctxt->ctxt * RXE_PER_CONTEXT_SIZE);
468 		/*
469 		 * TidFlow table is on the same page as the rest of the
470 		 * user registers.
471 		 */
472 		memlen = PAGE_SIZE;
473 		flags |= VM_DONTCOPY | VM_DONTEXPAND;
474 		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
475 		mapio = 1;
476 		break;
477 	case EVENTS:
478 		/*
479 		 * Use the page where this context's flags are. User level
480 		 * knows where it's own bitmap is within the page.
481 		 */
482 		memaddr = (unsigned long)
483 			(dd->events + uctxt_offset(uctxt)) & PAGE_MASK;
484 		memlen = PAGE_SIZE;
485 		/*
486 		 * v3.7 removes VM_RESERVED but the effect is kept by
487 		 * using VM_IO.
488 		 */
489 		flags |= VM_IO | VM_DONTEXPAND;
490 		vmf = 1;
491 		break;
492 	case STATUS:
493 		if (flags & (unsigned long)(VM_WRITE | VM_EXEC)) {
494 			ret = -EPERM;
495 			goto done;
496 		}
497 		memaddr = kvirt_to_phys((void *)dd->status);
498 		memlen = PAGE_SIZE;
499 		flags |= VM_IO | VM_DONTEXPAND;
500 		break;
501 	case RTAIL:
502 		if (!HFI1_CAP_IS_USET(DMA_RTAIL)) {
503 			/*
504 			 * If the memory allocation failed, the context alloc
505 			 * also would have failed, so we would never get here
506 			 */
507 			ret = -EINVAL;
508 			goto done;
509 		}
510 		if (flags & VM_WRITE) {
511 			ret = -EPERM;
512 			goto done;
513 		}
514 		memlen = PAGE_SIZE;
515 		memvirt = (void *)uctxt->rcvhdrtail_kvaddr;
516 		flags &= ~VM_MAYWRITE;
517 		break;
518 	case SUBCTXT_UREGS:
519 		memaddr = (u64)uctxt->subctxt_uregbase;
520 		memlen = PAGE_SIZE;
521 		flags |= VM_IO | VM_DONTEXPAND;
522 		vmf = 1;
523 		break;
524 	case SUBCTXT_RCV_HDRQ:
525 		memaddr = (u64)uctxt->subctxt_rcvhdr_base;
526 		memlen = uctxt->rcvhdrq_size * uctxt->subctxt_cnt;
527 		flags |= VM_IO | VM_DONTEXPAND;
528 		vmf = 1;
529 		break;
530 	case SUBCTXT_EGRBUF:
531 		memaddr = (u64)uctxt->subctxt_rcvegrbuf;
532 		memlen = uctxt->egrbufs.size * uctxt->subctxt_cnt;
533 		flags |= VM_IO | VM_DONTEXPAND;
534 		flags &= ~VM_MAYWRITE;
535 		vmf = 1;
536 		break;
537 	case SDMA_COMP: {
538 		struct hfi1_user_sdma_comp_q *cq = fd->cq;
539 
540 		if (!cq) {
541 			ret = -EFAULT;
542 			goto done;
543 		}
544 		memaddr = (u64)cq->comps;
545 		memlen = PAGE_ALIGN(sizeof(*cq->comps) * cq->nentries);
546 		flags |= VM_IO | VM_DONTEXPAND;
547 		vmf = 1;
548 		break;
549 	}
550 	default:
551 		ret = -EINVAL;
552 		break;
553 	}
554 
555 	if ((vma->vm_end - vma->vm_start) != memlen) {
556 		hfi1_cdbg(PROC, "%u:%u Memory size mismatch %lu:%lu",
557 			  uctxt->ctxt, fd->subctxt,
558 			  (vma->vm_end - vma->vm_start), memlen);
559 		ret = -EINVAL;
560 		goto done;
561 	}
562 
563 	vma->vm_flags = flags;
564 	hfi1_cdbg(PROC,
565 		  "%u:%u type:%u io/vf:%d/%d, addr:0x%llx, len:%lu(%lu), flags:0x%lx\n",
566 		    ctxt, subctxt, type, mapio, vmf, memaddr, memlen,
567 		    vma->vm_end - vma->vm_start, vma->vm_flags);
568 	if (vmf) {
569 		vma->vm_pgoff = PFN_DOWN(memaddr);
570 		vma->vm_ops = &vm_ops;
571 		ret = 0;
572 	} else if (mapio) {
573 		ret = io_remap_pfn_range(vma, vma->vm_start,
574 					 PFN_DOWN(memaddr),
575 					 memlen,
576 					 vma->vm_page_prot);
577 	} else if (memvirt) {
578 		ret = remap_pfn_range(vma, vma->vm_start,
579 				      PFN_DOWN(__pa(memvirt)),
580 				      memlen,
581 				      vma->vm_page_prot);
582 	} else {
583 		ret = remap_pfn_range(vma, vma->vm_start,
584 				      PFN_DOWN(memaddr),
585 				      memlen,
586 				      vma->vm_page_prot);
587 	}
588 done:
589 	return ret;
590 }
591 
592 /*
593  * Local (non-chip) user memory is not mapped right away but as it is
594  * accessed by the user-level code.
595  */
596 static int vma_fault(struct vm_fault *vmf)
597 {
598 	struct page *page;
599 
600 	page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
601 	if (!page)
602 		return VM_FAULT_SIGBUS;
603 
604 	get_page(page);
605 	vmf->page = page;
606 
607 	return 0;
608 }
609 
610 static unsigned int hfi1_poll(struct file *fp, struct poll_table_struct *pt)
611 {
612 	struct hfi1_ctxtdata *uctxt;
613 	unsigned pollflag;
614 
615 	uctxt = ((struct hfi1_filedata *)fp->private_data)->uctxt;
616 	if (!uctxt)
617 		pollflag = POLLERR;
618 	else if (uctxt->poll_type == HFI1_POLL_TYPE_URGENT)
619 		pollflag = poll_urgent(fp, pt);
620 	else  if (uctxt->poll_type == HFI1_POLL_TYPE_ANYRCV)
621 		pollflag = poll_next(fp, pt);
622 	else /* invalid */
623 		pollflag = POLLERR;
624 
625 	return pollflag;
626 }
627 
628 static int hfi1_file_close(struct inode *inode, struct file *fp)
629 {
630 	struct hfi1_filedata *fdata = fp->private_data;
631 	struct hfi1_ctxtdata *uctxt = fdata->uctxt;
632 	struct hfi1_devdata *dd = container_of(inode->i_cdev,
633 					       struct hfi1_devdata,
634 					       user_cdev);
635 	unsigned long flags, *ev;
636 
637 	fp->private_data = NULL;
638 
639 	if (!uctxt)
640 		goto done;
641 
642 	hfi1_cdbg(PROC, "closing ctxt %u:%u", uctxt->ctxt, fdata->subctxt);
643 
644 	flush_wc();
645 	/* drain user sdma queue */
646 	hfi1_user_sdma_free_queues(fdata, uctxt);
647 
648 	/* release the cpu */
649 	hfi1_put_proc_affinity(fdata->rec_cpu_num);
650 
651 	/* clean up rcv side */
652 	hfi1_user_exp_rcv_free(fdata);
653 
654 	/*
655 	 * fdata->uctxt is used in the above cleanup.  It is not ready to be
656 	 * removed until here.
657 	 */
658 	fdata->uctxt = NULL;
659 	hfi1_rcd_put(uctxt);
660 
661 	/*
662 	 * Clear any left over, unhandled events so the next process that
663 	 * gets this context doesn't get confused.
664 	 */
665 	ev = dd->events + uctxt_offset(uctxt) + fdata->subctxt;
666 	*ev = 0;
667 
668 	spin_lock_irqsave(&dd->uctxt_lock, flags);
669 	__clear_bit(fdata->subctxt, uctxt->in_use_ctxts);
670 	if (!bitmap_empty(uctxt->in_use_ctxts, HFI1_MAX_SHARED_CTXTS)) {
671 		spin_unlock_irqrestore(&dd->uctxt_lock, flags);
672 		goto done;
673 	}
674 	spin_unlock_irqrestore(&dd->uctxt_lock, flags);
675 
676 	/*
677 	 * Disable receive context and interrupt available, reset all
678 	 * RcvCtxtCtrl bits to default values.
679 	 */
680 	hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_DIS |
681 		     HFI1_RCVCTRL_TIDFLOW_DIS |
682 		     HFI1_RCVCTRL_INTRAVAIL_DIS |
683 		     HFI1_RCVCTRL_TAILUPD_DIS |
684 		     HFI1_RCVCTRL_ONE_PKT_EGR_DIS |
685 		     HFI1_RCVCTRL_NO_RHQ_DROP_DIS |
686 		     HFI1_RCVCTRL_NO_EGR_DROP_DIS, uctxt);
687 	/* Clear the context's J_KEY */
688 	hfi1_clear_ctxt_jkey(dd, uctxt);
689 	/*
690 	 * If a send context is allocated, reset context integrity
691 	 * checks to default and disable the send context.
692 	 */
693 	if (uctxt->sc) {
694 		set_pio_integrity(uctxt->sc);
695 		sc_disable(uctxt->sc);
696 	}
697 
698 	hfi1_free_ctxt_rcv_groups(uctxt);
699 	hfi1_clear_ctxt_pkey(dd, uctxt);
700 
701 	uctxt->event_flags = 0;
702 
703 	deallocate_ctxt(uctxt);
704 done:
705 	mmdrop(fdata->mm);
706 	kobject_put(&dd->kobj);
707 
708 	if (atomic_dec_and_test(&dd->user_refcount))
709 		complete(&dd->user_comp);
710 
711 	kfree(fdata);
712 	return 0;
713 }
714 
715 /*
716  * Convert kernel *virtual* addresses to physical addresses.
717  * This is used to vmalloc'ed addresses.
718  */
719 static u64 kvirt_to_phys(void *addr)
720 {
721 	struct page *page;
722 	u64 paddr = 0;
723 
724 	page = vmalloc_to_page(addr);
725 	if (page)
726 		paddr = page_to_pfn(page) << PAGE_SHIFT;
727 
728 	return paddr;
729 }
730 
731 /**
732  * complete_subctxt
733  * @fd: valid filedata pointer
734  *
735  * Sub-context info can only be set up after the base context
736  * has been completed.  This is indicated by the clearing of the
737  * HFI1_CTXT_BASE_UINIT bit.
738  *
739  * Wait for the bit to be cleared, and then complete the subcontext
740  * initialization.
741  *
742  */
743 static int complete_subctxt(struct hfi1_filedata *fd)
744 {
745 	int ret;
746 	unsigned long flags;
747 
748 	/*
749 	 * sub-context info can only be set up after the base context
750 	 * has been completed.
751 	 */
752 	ret = wait_event_interruptible(
753 		fd->uctxt->wait,
754 		!test_bit(HFI1_CTXT_BASE_UNINIT, &fd->uctxt->event_flags));
755 
756 	if (test_bit(HFI1_CTXT_BASE_FAILED, &fd->uctxt->event_flags))
757 		ret = -ENOMEM;
758 
759 	/* Finish the sub-context init */
760 	if (!ret) {
761 		fd->rec_cpu_num = hfi1_get_proc_affinity(fd->uctxt->numa_id);
762 		ret = init_user_ctxt(fd, fd->uctxt);
763 	}
764 
765 	if (ret) {
766 		hfi1_rcd_put(fd->uctxt);
767 		fd->uctxt = NULL;
768 		spin_lock_irqsave(&fd->dd->uctxt_lock, flags);
769 		__clear_bit(fd->subctxt, fd->uctxt->in_use_ctxts);
770 		spin_unlock_irqrestore(&fd->dd->uctxt_lock, flags);
771 	}
772 
773 	return ret;
774 }
775 
776 static int assign_ctxt(struct hfi1_filedata *fd, unsigned long arg, u32 len)
777 {
778 	int ret;
779 	unsigned int swmajor;
780 	struct hfi1_ctxtdata *uctxt = NULL;
781 	struct hfi1_user_info uinfo;
782 
783 	if (fd->uctxt)
784 		return -EINVAL;
785 
786 	if (sizeof(uinfo) != len)
787 		return -EINVAL;
788 
789 	if (copy_from_user(&uinfo, (void __user *)arg, sizeof(uinfo)))
790 		return -EFAULT;
791 
792 	swmajor = uinfo.userversion >> 16;
793 	if (swmajor != HFI1_USER_SWMAJOR)
794 		return -ENODEV;
795 
796 	if (uinfo.subctxt_cnt > HFI1_MAX_SHARED_CTXTS)
797 		return -EINVAL;
798 
799 	/*
800 	 * Acquire the mutex to protect against multiple creations of what
801 	 * could be a shared base context.
802 	 */
803 	mutex_lock(&hfi1_mutex);
804 	/*
805 	 * Get a sub context if available  (fd->uctxt will be set).
806 	 * ret < 0 error, 0 no context, 1 sub-context found
807 	 */
808 	ret = find_sub_ctxt(fd, &uinfo);
809 
810 	/*
811 	 * Allocate a base context if context sharing is not required or a
812 	 * sub context wasn't found.
813 	 */
814 	if (!ret)
815 		ret = allocate_ctxt(fd, fd->dd, &uinfo, &uctxt);
816 
817 	mutex_unlock(&hfi1_mutex);
818 
819 	/* Depending on the context type, finish the appropriate init */
820 	switch (ret) {
821 	case 0:
822 		ret = setup_base_ctxt(fd, uctxt);
823 		if (ret)
824 			deallocate_ctxt(uctxt);
825 		break;
826 	case 1:
827 		ret = complete_subctxt(fd);
828 		break;
829 	default:
830 		break;
831 	}
832 
833 	return ret;
834 }
835 
836 /**
837  * match_ctxt
838  * @fd: valid filedata pointer
839  * @uinfo: user info to compare base context with
840  * @uctxt: context to compare uinfo to.
841  *
842  * Compare the given context with the given information to see if it
843  * can be used for a sub context.
844  */
845 static int match_ctxt(struct hfi1_filedata *fd,
846 		      const struct hfi1_user_info *uinfo,
847 		      struct hfi1_ctxtdata *uctxt)
848 {
849 	struct hfi1_devdata *dd = fd->dd;
850 	unsigned long flags;
851 	u16 subctxt;
852 
853 	/* Skip dynamically allocated kernel contexts */
854 	if (uctxt->sc && (uctxt->sc->type == SC_KERNEL))
855 		return 0;
856 
857 	/* Skip ctxt if it doesn't match the requested one */
858 	if (memcmp(uctxt->uuid, uinfo->uuid, sizeof(uctxt->uuid)) ||
859 	    uctxt->jkey != generate_jkey(current_uid()) ||
860 	    uctxt->subctxt_id != uinfo->subctxt_id ||
861 	    uctxt->subctxt_cnt != uinfo->subctxt_cnt)
862 		return 0;
863 
864 	/* Verify the sharing process matches the base */
865 	if (uctxt->userversion != uinfo->userversion)
866 		return -EINVAL;
867 
868 	/* Find an unused sub context */
869 	spin_lock_irqsave(&dd->uctxt_lock, flags);
870 	if (bitmap_empty(uctxt->in_use_ctxts, HFI1_MAX_SHARED_CTXTS)) {
871 		/* context is being closed, do not use */
872 		spin_unlock_irqrestore(&dd->uctxt_lock, flags);
873 		return 0;
874 	}
875 
876 	subctxt = find_first_zero_bit(uctxt->in_use_ctxts,
877 				      HFI1_MAX_SHARED_CTXTS);
878 	if (subctxt >= uctxt->subctxt_cnt) {
879 		spin_unlock_irqrestore(&dd->uctxt_lock, flags);
880 		return -EBUSY;
881 	}
882 
883 	fd->subctxt = subctxt;
884 	__set_bit(fd->subctxt, uctxt->in_use_ctxts);
885 	spin_unlock_irqrestore(&dd->uctxt_lock, flags);
886 
887 	fd->uctxt = uctxt;
888 	hfi1_rcd_get(uctxt);
889 
890 	return 1;
891 }
892 
893 /**
894  * find_sub_ctxt
895  * @fd: valid filedata pointer
896  * @uinfo: matching info to use to find a possible context to share.
897  *
898  * The hfi1_mutex must be held when this function is called.  It is
899  * necessary to ensure serialized creation of shared contexts.
900  *
901  * Return:
902  *    0      No sub-context found
903  *    1      Subcontext found and allocated
904  *    errno  EINVAL (incorrect parameters)
905  *           EBUSY (all sub contexts in use)
906  */
907 static int find_sub_ctxt(struct hfi1_filedata *fd,
908 			 const struct hfi1_user_info *uinfo)
909 {
910 	struct hfi1_ctxtdata *uctxt;
911 	struct hfi1_devdata *dd = fd->dd;
912 	u16 i;
913 	int ret;
914 
915 	if (!uinfo->subctxt_cnt)
916 		return 0;
917 
918 	for (i = dd->first_dyn_alloc_ctxt; i < dd->num_rcv_contexts; i++) {
919 		uctxt = hfi1_rcd_get_by_index(dd, i);
920 		if (uctxt) {
921 			ret = match_ctxt(fd, uinfo, uctxt);
922 			hfi1_rcd_put(uctxt);
923 			/* value of != 0 will return */
924 			if (ret)
925 				return ret;
926 		}
927 	}
928 
929 	return 0;
930 }
931 
932 static int allocate_ctxt(struct hfi1_filedata *fd, struct hfi1_devdata *dd,
933 			 struct hfi1_user_info *uinfo,
934 			 struct hfi1_ctxtdata **rcd)
935 {
936 	struct hfi1_ctxtdata *uctxt;
937 	int ret, numa;
938 
939 	if (dd->flags & HFI1_FROZEN) {
940 		/*
941 		 * Pick an error that is unique from all other errors
942 		 * that are returned so the user process knows that
943 		 * it tried to allocate while the SPC was frozen.  It
944 		 * it should be able to retry with success in a short
945 		 * while.
946 		 */
947 		return -EIO;
948 	}
949 
950 	if (!dd->freectxts)
951 		return -EBUSY;
952 
953 	/*
954 	 * If we don't have a NUMA node requested, preference is towards
955 	 * device NUMA node.
956 	 */
957 	fd->rec_cpu_num = hfi1_get_proc_affinity(dd->node);
958 	if (fd->rec_cpu_num != -1)
959 		numa = cpu_to_node(fd->rec_cpu_num);
960 	else
961 		numa = numa_node_id();
962 	ret = hfi1_create_ctxtdata(dd->pport, numa, &uctxt);
963 	if (ret < 0) {
964 		dd_dev_err(dd, "user ctxtdata allocation failed\n");
965 		return ret;
966 	}
967 	hfi1_cdbg(PROC, "[%u:%u] pid %u assigned to CPU %d (NUMA %u)",
968 		  uctxt->ctxt, fd->subctxt, current->pid, fd->rec_cpu_num,
969 		  uctxt->numa_id);
970 
971 	/*
972 	 * Allocate and enable a PIO send context.
973 	 */
974 	uctxt->sc = sc_alloc(dd, SC_USER, uctxt->rcvhdrqentsize, dd->node);
975 	if (!uctxt->sc) {
976 		ret = -ENOMEM;
977 		goto ctxdata_free;
978 	}
979 	hfi1_cdbg(PROC, "allocated send context %u(%u)\n", uctxt->sc->sw_index,
980 		  uctxt->sc->hw_context);
981 	ret = sc_enable(uctxt->sc);
982 	if (ret)
983 		goto ctxdata_free;
984 
985 	/*
986 	 * Setup sub context information if the user-level has requested
987 	 * sub contexts.
988 	 * This has to be done here so the rest of the sub-contexts find the
989 	 * proper base context.
990 	 */
991 	if (uinfo->subctxt_cnt)
992 		init_subctxts(uctxt, uinfo);
993 	uctxt->userversion = uinfo->userversion;
994 	uctxt->flags = hfi1_cap_mask; /* save current flag state */
995 	init_waitqueue_head(&uctxt->wait);
996 	strlcpy(uctxt->comm, current->comm, sizeof(uctxt->comm));
997 	memcpy(uctxt->uuid, uinfo->uuid, sizeof(uctxt->uuid));
998 	uctxt->jkey = generate_jkey(current_uid());
999 	hfi1_stats.sps_ctxts++;
1000 	/*
1001 	 * Disable ASPM when there are open user/PSM contexts to avoid
1002 	 * issues with ASPM L1 exit latency
1003 	 */
1004 	if (dd->freectxts-- == dd->num_user_contexts)
1005 		aspm_disable_all(dd);
1006 
1007 	*rcd = uctxt;
1008 
1009 	return 0;
1010 
1011 ctxdata_free:
1012 	hfi1_free_ctxt(uctxt);
1013 	return ret;
1014 }
1015 
1016 static void deallocate_ctxt(struct hfi1_ctxtdata *uctxt)
1017 {
1018 	mutex_lock(&hfi1_mutex);
1019 	hfi1_stats.sps_ctxts--;
1020 	if (++uctxt->dd->freectxts == uctxt->dd->num_user_contexts)
1021 		aspm_enable_all(uctxt->dd);
1022 	mutex_unlock(&hfi1_mutex);
1023 
1024 	hfi1_free_ctxt(uctxt);
1025 }
1026 
1027 static void init_subctxts(struct hfi1_ctxtdata *uctxt,
1028 			  const struct hfi1_user_info *uinfo)
1029 {
1030 	uctxt->subctxt_cnt = uinfo->subctxt_cnt;
1031 	uctxt->subctxt_id = uinfo->subctxt_id;
1032 	set_bit(HFI1_CTXT_BASE_UNINIT, &uctxt->event_flags);
1033 }
1034 
1035 static int setup_subctxt(struct hfi1_ctxtdata *uctxt)
1036 {
1037 	int ret = 0;
1038 	u16 num_subctxts = uctxt->subctxt_cnt;
1039 
1040 	uctxt->subctxt_uregbase = vmalloc_user(PAGE_SIZE);
1041 	if (!uctxt->subctxt_uregbase)
1042 		return -ENOMEM;
1043 
1044 	/* We can take the size of the RcvHdr Queue from the master */
1045 	uctxt->subctxt_rcvhdr_base = vmalloc_user(uctxt->rcvhdrq_size *
1046 						  num_subctxts);
1047 	if (!uctxt->subctxt_rcvhdr_base) {
1048 		ret = -ENOMEM;
1049 		goto bail_ureg;
1050 	}
1051 
1052 	uctxt->subctxt_rcvegrbuf = vmalloc_user(uctxt->egrbufs.size *
1053 						num_subctxts);
1054 	if (!uctxt->subctxt_rcvegrbuf) {
1055 		ret = -ENOMEM;
1056 		goto bail_rhdr;
1057 	}
1058 
1059 	return 0;
1060 
1061 bail_rhdr:
1062 	vfree(uctxt->subctxt_rcvhdr_base);
1063 	uctxt->subctxt_rcvhdr_base = NULL;
1064 bail_ureg:
1065 	vfree(uctxt->subctxt_uregbase);
1066 	uctxt->subctxt_uregbase = NULL;
1067 
1068 	return ret;
1069 }
1070 
1071 static void user_init(struct hfi1_ctxtdata *uctxt)
1072 {
1073 	unsigned int rcvctrl_ops = 0;
1074 
1075 	/* initialize poll variables... */
1076 	uctxt->urgent = 0;
1077 	uctxt->urgent_poll = 0;
1078 
1079 	/*
1080 	 * Now enable the ctxt for receive.
1081 	 * For chips that are set to DMA the tail register to memory
1082 	 * when they change (and when the update bit transitions from
1083 	 * 0 to 1.  So for those chips, we turn it off and then back on.
1084 	 * This will (very briefly) affect any other open ctxts, but the
1085 	 * duration is very short, and therefore isn't an issue.  We
1086 	 * explicitly set the in-memory tail copy to 0 beforehand, so we
1087 	 * don't have to wait to be sure the DMA update has happened
1088 	 * (chip resets head/tail to 0 on transition to enable).
1089 	 */
1090 	if (uctxt->rcvhdrtail_kvaddr)
1091 		clear_rcvhdrtail(uctxt);
1092 
1093 	/* Setup J_KEY before enabling the context */
1094 	hfi1_set_ctxt_jkey(uctxt->dd, uctxt, uctxt->jkey);
1095 
1096 	rcvctrl_ops = HFI1_RCVCTRL_CTXT_ENB;
1097 	if (HFI1_CAP_UGET_MASK(uctxt->flags, HDRSUPP))
1098 		rcvctrl_ops |= HFI1_RCVCTRL_TIDFLOW_ENB;
1099 	/*
1100 	 * Ignore the bit in the flags for now until proper
1101 	 * support for multiple packet per rcv array entry is
1102 	 * added.
1103 	 */
1104 	if (!HFI1_CAP_UGET_MASK(uctxt->flags, MULTI_PKT_EGR))
1105 		rcvctrl_ops |= HFI1_RCVCTRL_ONE_PKT_EGR_ENB;
1106 	if (HFI1_CAP_UGET_MASK(uctxt->flags, NODROP_EGR_FULL))
1107 		rcvctrl_ops |= HFI1_RCVCTRL_NO_EGR_DROP_ENB;
1108 	if (HFI1_CAP_UGET_MASK(uctxt->flags, NODROP_RHQ_FULL))
1109 		rcvctrl_ops |= HFI1_RCVCTRL_NO_RHQ_DROP_ENB;
1110 	/*
1111 	 * The RcvCtxtCtrl.TailUpd bit has to be explicitly written.
1112 	 * We can't rely on the correct value to be set from prior
1113 	 * uses of the chip or ctxt. Therefore, add the rcvctrl op
1114 	 * for both cases.
1115 	 */
1116 	if (HFI1_CAP_UGET_MASK(uctxt->flags, DMA_RTAIL))
1117 		rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_ENB;
1118 	else
1119 		rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_DIS;
1120 	hfi1_rcvctrl(uctxt->dd, rcvctrl_ops, uctxt);
1121 }
1122 
1123 static int get_ctxt_info(struct hfi1_filedata *fd, unsigned long arg, u32 len)
1124 {
1125 	struct hfi1_ctxt_info cinfo;
1126 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1127 
1128 	if (sizeof(cinfo) != len)
1129 		return -EINVAL;
1130 
1131 	memset(&cinfo, 0, sizeof(cinfo));
1132 	cinfo.runtime_flags = (((uctxt->flags >> HFI1_CAP_MISC_SHIFT) &
1133 				HFI1_CAP_MISC_MASK) << HFI1_CAP_USER_SHIFT) |
1134 			HFI1_CAP_UGET_MASK(uctxt->flags, MASK) |
1135 			HFI1_CAP_KGET_MASK(uctxt->flags, K2U);
1136 	/* adjust flag if this fd is not able to cache */
1137 	if (!fd->handler)
1138 		cinfo.runtime_flags |= HFI1_CAP_TID_UNMAP; /* no caching */
1139 
1140 	cinfo.num_active = hfi1_count_active_units();
1141 	cinfo.unit = uctxt->dd->unit;
1142 	cinfo.ctxt = uctxt->ctxt;
1143 	cinfo.subctxt = fd->subctxt;
1144 	cinfo.rcvtids = roundup(uctxt->egrbufs.alloced,
1145 				uctxt->dd->rcv_entries.group_size) +
1146 		uctxt->expected_count;
1147 	cinfo.credits = uctxt->sc->credits;
1148 	cinfo.numa_node = uctxt->numa_id;
1149 	cinfo.rec_cpu = fd->rec_cpu_num;
1150 	cinfo.send_ctxt = uctxt->sc->hw_context;
1151 
1152 	cinfo.egrtids = uctxt->egrbufs.alloced;
1153 	cinfo.rcvhdrq_cnt = uctxt->rcvhdrq_cnt;
1154 	cinfo.rcvhdrq_entsize = uctxt->rcvhdrqentsize << 2;
1155 	cinfo.sdma_ring_size = fd->cq->nentries;
1156 	cinfo.rcvegr_size = uctxt->egrbufs.rcvtid_size;
1157 
1158 	trace_hfi1_ctxt_info(uctxt->dd, uctxt->ctxt, fd->subctxt, cinfo);
1159 	if (copy_to_user((void __user *)arg, &cinfo, len))
1160 		return -EFAULT;
1161 
1162 	return 0;
1163 }
1164 
1165 static int init_user_ctxt(struct hfi1_filedata *fd,
1166 			  struct hfi1_ctxtdata *uctxt)
1167 {
1168 	int ret;
1169 
1170 	ret = hfi1_user_sdma_alloc_queues(uctxt, fd);
1171 	if (ret)
1172 		return ret;
1173 
1174 	ret = hfi1_user_exp_rcv_init(fd, uctxt);
1175 	if (ret)
1176 		hfi1_user_sdma_free_queues(fd, uctxt);
1177 
1178 	return ret;
1179 }
1180 
1181 static int setup_base_ctxt(struct hfi1_filedata *fd,
1182 			   struct hfi1_ctxtdata *uctxt)
1183 {
1184 	struct hfi1_devdata *dd = uctxt->dd;
1185 	int ret = 0;
1186 
1187 	hfi1_init_ctxt(uctxt->sc);
1188 
1189 	/* Now allocate the RcvHdr queue and eager buffers. */
1190 	ret = hfi1_create_rcvhdrq(dd, uctxt);
1191 	if (ret)
1192 		goto done;
1193 
1194 	ret = hfi1_setup_eagerbufs(uctxt);
1195 	if (ret)
1196 		goto done;
1197 
1198 	/* If sub-contexts are enabled, do the appropriate setup */
1199 	if (uctxt->subctxt_cnt)
1200 		ret = setup_subctxt(uctxt);
1201 	if (ret)
1202 		goto done;
1203 
1204 	ret = hfi1_alloc_ctxt_rcv_groups(uctxt);
1205 	if (ret)
1206 		goto done;
1207 
1208 	ret = init_user_ctxt(fd, uctxt);
1209 	if (ret)
1210 		goto done;
1211 
1212 	user_init(uctxt);
1213 
1214 	/* Now that the context is set up, the fd can get a reference. */
1215 	fd->uctxt = uctxt;
1216 	hfi1_rcd_get(uctxt);
1217 
1218 done:
1219 	if (uctxt->subctxt_cnt) {
1220 		/*
1221 		 * On error, set the failed bit so sub-contexts will clean up
1222 		 * correctly.
1223 		 */
1224 		if (ret)
1225 			set_bit(HFI1_CTXT_BASE_FAILED, &uctxt->event_flags);
1226 
1227 		/*
1228 		 * Base context is done (successfully or not), notify anybody
1229 		 * using a sub-context that is waiting for this completion.
1230 		 */
1231 		clear_bit(HFI1_CTXT_BASE_UNINIT, &uctxt->event_flags);
1232 		wake_up(&uctxt->wait);
1233 	}
1234 
1235 	return ret;
1236 }
1237 
1238 static int get_base_info(struct hfi1_filedata *fd, unsigned long arg, u32 len)
1239 {
1240 	struct hfi1_base_info binfo;
1241 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1242 	struct hfi1_devdata *dd = uctxt->dd;
1243 	unsigned offset;
1244 
1245 	trace_hfi1_uctxtdata(uctxt->dd, uctxt, fd->subctxt);
1246 
1247 	if (sizeof(binfo) != len)
1248 		return -EINVAL;
1249 
1250 	memset(&binfo, 0, sizeof(binfo));
1251 	binfo.hw_version = dd->revision;
1252 	binfo.sw_version = HFI1_KERN_SWVERSION;
1253 	binfo.bthqp = kdeth_qp;
1254 	binfo.jkey = uctxt->jkey;
1255 	/*
1256 	 * If more than 64 contexts are enabled the allocated credit
1257 	 * return will span two or three contiguous pages. Since we only
1258 	 * map the page containing the context's credit return address,
1259 	 * we need to calculate the offset in the proper page.
1260 	 */
1261 	offset = ((u64)uctxt->sc->hw_free -
1262 		  (u64)dd->cr_base[uctxt->numa_id].va) % PAGE_SIZE;
1263 	binfo.sc_credits_addr = HFI1_MMAP_TOKEN(PIO_CRED, uctxt->ctxt,
1264 						fd->subctxt, offset);
1265 	binfo.pio_bufbase = HFI1_MMAP_TOKEN(PIO_BUFS, uctxt->ctxt,
1266 					    fd->subctxt,
1267 					    uctxt->sc->base_addr);
1268 	binfo.pio_bufbase_sop = HFI1_MMAP_TOKEN(PIO_BUFS_SOP,
1269 						uctxt->ctxt,
1270 						fd->subctxt,
1271 						uctxt->sc->base_addr);
1272 	binfo.rcvhdr_bufbase = HFI1_MMAP_TOKEN(RCV_HDRQ, uctxt->ctxt,
1273 					       fd->subctxt,
1274 					       uctxt->rcvhdrq);
1275 	binfo.rcvegr_bufbase = HFI1_MMAP_TOKEN(RCV_EGRBUF, uctxt->ctxt,
1276 					       fd->subctxt,
1277 					       uctxt->egrbufs.rcvtids[0].dma);
1278 	binfo.sdma_comp_bufbase = HFI1_MMAP_TOKEN(SDMA_COMP, uctxt->ctxt,
1279 						  fd->subctxt, 0);
1280 	/*
1281 	 * user regs are at
1282 	 * (RXE_PER_CONTEXT_USER + (ctxt * RXE_PER_CONTEXT_SIZE))
1283 	 */
1284 	binfo.user_regbase = HFI1_MMAP_TOKEN(UREGS, uctxt->ctxt,
1285 					     fd->subctxt, 0);
1286 	offset = offset_in_page((uctxt_offset(uctxt) + fd->subctxt) *
1287 				sizeof(*dd->events));
1288 	binfo.events_bufbase = HFI1_MMAP_TOKEN(EVENTS, uctxt->ctxt,
1289 					       fd->subctxt,
1290 					       offset);
1291 	binfo.status_bufbase = HFI1_MMAP_TOKEN(STATUS, uctxt->ctxt,
1292 					       fd->subctxt,
1293 					       dd->status);
1294 	if (HFI1_CAP_IS_USET(DMA_RTAIL))
1295 		binfo.rcvhdrtail_base = HFI1_MMAP_TOKEN(RTAIL, uctxt->ctxt,
1296 							fd->subctxt, 0);
1297 	if (uctxt->subctxt_cnt) {
1298 		binfo.subctxt_uregbase = HFI1_MMAP_TOKEN(SUBCTXT_UREGS,
1299 							 uctxt->ctxt,
1300 							 fd->subctxt, 0);
1301 		binfo.subctxt_rcvhdrbuf = HFI1_MMAP_TOKEN(SUBCTXT_RCV_HDRQ,
1302 							  uctxt->ctxt,
1303 							  fd->subctxt, 0);
1304 		binfo.subctxt_rcvegrbuf = HFI1_MMAP_TOKEN(SUBCTXT_EGRBUF,
1305 							  uctxt->ctxt,
1306 							  fd->subctxt, 0);
1307 	}
1308 
1309 	if (copy_to_user((void __user *)arg, &binfo, len))
1310 		return -EFAULT;
1311 
1312 	return 0;
1313 }
1314 
1315 /**
1316  * user_exp_rcv_setup - Set up the given tid rcv list
1317  * @fd: file data of the current driver instance
1318  * @arg: ioctl argumnent for user space information
1319  * @len: length of data structure associated with ioctl command
1320  *
1321  * Wrapper to validate ioctl information before doing _rcv_setup.
1322  *
1323  */
1324 static int user_exp_rcv_setup(struct hfi1_filedata *fd, unsigned long arg,
1325 			      u32 len)
1326 {
1327 	int ret;
1328 	unsigned long addr;
1329 	struct hfi1_tid_info tinfo;
1330 
1331 	if (sizeof(tinfo) != len)
1332 		return -EINVAL;
1333 
1334 	if (copy_from_user(&tinfo, (void __user *)arg, (sizeof(tinfo))))
1335 		return -EFAULT;
1336 
1337 	ret = hfi1_user_exp_rcv_setup(fd, &tinfo);
1338 	if (!ret) {
1339 		/*
1340 		 * Copy the number of tidlist entries we used
1341 		 * and the length of the buffer we registered.
1342 		 */
1343 		addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
1344 		if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
1345 				 sizeof(tinfo.tidcnt)))
1346 			return -EFAULT;
1347 
1348 		addr = arg + offsetof(struct hfi1_tid_info, length);
1349 		if (copy_to_user((void __user *)addr, &tinfo.length,
1350 				 sizeof(tinfo.length)))
1351 			ret = -EFAULT;
1352 	}
1353 
1354 	return ret;
1355 }
1356 
1357 /**
1358  * user_exp_rcv_clear - Clear the given tid rcv list
1359  * @fd: file data of the current driver instance
1360  * @arg: ioctl argumnent for user space information
1361  * @len: length of data structure associated with ioctl command
1362  *
1363  * The hfi1_user_exp_rcv_clear() can be called from the error path.  Because
1364  * of this, we need to use this wrapper to copy the user space information
1365  * before doing the clear.
1366  */
1367 static int user_exp_rcv_clear(struct hfi1_filedata *fd, unsigned long arg,
1368 			      u32 len)
1369 {
1370 	int ret;
1371 	unsigned long addr;
1372 	struct hfi1_tid_info tinfo;
1373 
1374 	if (sizeof(tinfo) != len)
1375 		return -EINVAL;
1376 
1377 	if (copy_from_user(&tinfo, (void __user *)arg, (sizeof(tinfo))))
1378 		return -EFAULT;
1379 
1380 	ret = hfi1_user_exp_rcv_clear(fd, &tinfo);
1381 	if (!ret) {
1382 		addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
1383 		if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
1384 				 sizeof(tinfo.tidcnt)))
1385 			return -EFAULT;
1386 	}
1387 
1388 	return ret;
1389 }
1390 
1391 /**
1392  * user_exp_rcv_invalid - Invalidate the given tid rcv list
1393  * @fd: file data of the current driver instance
1394  * @arg: ioctl argumnent for user space information
1395  * @len: length of data structure associated with ioctl command
1396  *
1397  * Wrapper to validate ioctl information before doing _rcv_invalid.
1398  *
1399  */
1400 static int user_exp_rcv_invalid(struct hfi1_filedata *fd, unsigned long arg,
1401 				u32 len)
1402 {
1403 	int ret;
1404 	unsigned long addr;
1405 	struct hfi1_tid_info tinfo;
1406 
1407 	if (sizeof(tinfo) != len)
1408 		return -EINVAL;
1409 
1410 	if (!fd->invalid_tids)
1411 		return -EINVAL;
1412 
1413 	if (copy_from_user(&tinfo, (void __user *)arg, (sizeof(tinfo))))
1414 		return -EFAULT;
1415 
1416 	ret = hfi1_user_exp_rcv_invalid(fd, &tinfo);
1417 	if (ret)
1418 		return ret;
1419 
1420 	addr = arg + offsetof(struct hfi1_tid_info, tidcnt);
1421 	if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
1422 			 sizeof(tinfo.tidcnt)))
1423 		ret = -EFAULT;
1424 
1425 	return ret;
1426 }
1427 
1428 static unsigned int poll_urgent(struct file *fp,
1429 				struct poll_table_struct *pt)
1430 {
1431 	struct hfi1_filedata *fd = fp->private_data;
1432 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1433 	struct hfi1_devdata *dd = uctxt->dd;
1434 	unsigned pollflag;
1435 
1436 	poll_wait(fp, &uctxt->wait, pt);
1437 
1438 	spin_lock_irq(&dd->uctxt_lock);
1439 	if (uctxt->urgent != uctxt->urgent_poll) {
1440 		pollflag = POLLIN | POLLRDNORM;
1441 		uctxt->urgent_poll = uctxt->urgent;
1442 	} else {
1443 		pollflag = 0;
1444 		set_bit(HFI1_CTXT_WAITING_URG, &uctxt->event_flags);
1445 	}
1446 	spin_unlock_irq(&dd->uctxt_lock);
1447 
1448 	return pollflag;
1449 }
1450 
1451 static unsigned int poll_next(struct file *fp,
1452 			      struct poll_table_struct *pt)
1453 {
1454 	struct hfi1_filedata *fd = fp->private_data;
1455 	struct hfi1_ctxtdata *uctxt = fd->uctxt;
1456 	struct hfi1_devdata *dd = uctxt->dd;
1457 	unsigned pollflag;
1458 
1459 	poll_wait(fp, &uctxt->wait, pt);
1460 
1461 	spin_lock_irq(&dd->uctxt_lock);
1462 	if (hdrqempty(uctxt)) {
1463 		set_bit(HFI1_CTXT_WAITING_RCV, &uctxt->event_flags);
1464 		hfi1_rcvctrl(dd, HFI1_RCVCTRL_INTRAVAIL_ENB, uctxt);
1465 		pollflag = 0;
1466 	} else {
1467 		pollflag = POLLIN | POLLRDNORM;
1468 	}
1469 	spin_unlock_irq(&dd->uctxt_lock);
1470 
1471 	return pollflag;
1472 }
1473 
1474 /*
1475  * Find all user contexts in use, and set the specified bit in their
1476  * event mask.
1477  * See also find_ctxt() for a similar use, that is specific to send buffers.
1478  */
1479 int hfi1_set_uevent_bits(struct hfi1_pportdata *ppd, const int evtbit)
1480 {
1481 	struct hfi1_ctxtdata *uctxt;
1482 	struct hfi1_devdata *dd = ppd->dd;
1483 	u16 ctxt;
1484 
1485 	if (!dd->events)
1486 		return -EINVAL;
1487 
1488 	for (ctxt = dd->first_dyn_alloc_ctxt; ctxt < dd->num_rcv_contexts;
1489 	     ctxt++) {
1490 		uctxt = hfi1_rcd_get_by_index(dd, ctxt);
1491 		if (uctxt) {
1492 			unsigned long *evs;
1493 			int i;
1494 			/*
1495 			 * subctxt_cnt is 0 if not shared, so do base
1496 			 * separately, first, then remaining subctxt, if any
1497 			 */
1498 			evs = dd->events + uctxt_offset(uctxt);
1499 			set_bit(evtbit, evs);
1500 			for (i = 1; i < uctxt->subctxt_cnt; i++)
1501 				set_bit(evtbit, evs + i);
1502 			hfi1_rcd_put(uctxt);
1503 		}
1504 	}
1505 
1506 	return 0;
1507 }
1508 
1509 /**
1510  * manage_rcvq - manage a context's receive queue
1511  * @uctxt: the context
1512  * @subctxt: the sub-context
1513  * @start_stop: action to carry out
1514  *
1515  * start_stop == 0 disables receive on the context, for use in queue
1516  * overflow conditions.  start_stop==1 re-enables, to be used to
1517  * re-init the software copy of the head register
1518  */
1519 static int manage_rcvq(struct hfi1_ctxtdata *uctxt, u16 subctxt,
1520 		       unsigned long arg)
1521 {
1522 	struct hfi1_devdata *dd = uctxt->dd;
1523 	unsigned int rcvctrl_op;
1524 	int start_stop;
1525 
1526 	if (subctxt)
1527 		return 0;
1528 
1529 	if (get_user(start_stop, (int __user *)arg))
1530 		return -EFAULT;
1531 
1532 	/* atomically clear receive enable ctxt. */
1533 	if (start_stop) {
1534 		/*
1535 		 * On enable, force in-memory copy of the tail register to
1536 		 * 0, so that protocol code doesn't have to worry about
1537 		 * whether or not the chip has yet updated the in-memory
1538 		 * copy or not on return from the system call. The chip
1539 		 * always resets it's tail register back to 0 on a
1540 		 * transition from disabled to enabled.
1541 		 */
1542 		if (uctxt->rcvhdrtail_kvaddr)
1543 			clear_rcvhdrtail(uctxt);
1544 		rcvctrl_op = HFI1_RCVCTRL_CTXT_ENB;
1545 	} else {
1546 		rcvctrl_op = HFI1_RCVCTRL_CTXT_DIS;
1547 	}
1548 	hfi1_rcvctrl(dd, rcvctrl_op, uctxt);
1549 	/* always; new head should be equal to new tail; see above */
1550 
1551 	return 0;
1552 }
1553 
1554 /*
1555  * clear the event notifier events for this context.
1556  * User process then performs actions appropriate to bit having been
1557  * set, if desired, and checks again in future.
1558  */
1559 static int user_event_ack(struct hfi1_ctxtdata *uctxt, u16 subctxt,
1560 			  unsigned long arg)
1561 {
1562 	int i;
1563 	struct hfi1_devdata *dd = uctxt->dd;
1564 	unsigned long *evs;
1565 	unsigned long events;
1566 
1567 	if (!dd->events)
1568 		return 0;
1569 
1570 	if (get_user(events, (unsigned long __user *)arg))
1571 		return -EFAULT;
1572 
1573 	evs = dd->events + uctxt_offset(uctxt) + subctxt;
1574 
1575 	for (i = 0; i <= _HFI1_MAX_EVENT_BIT; i++) {
1576 		if (!test_bit(i, &events))
1577 			continue;
1578 		clear_bit(i, evs);
1579 	}
1580 	return 0;
1581 }
1582 
1583 static int set_ctxt_pkey(struct hfi1_ctxtdata *uctxt, unsigned long arg)
1584 {
1585 	int i;
1586 	struct hfi1_pportdata *ppd = uctxt->ppd;
1587 	struct hfi1_devdata *dd = uctxt->dd;
1588 	u16 pkey;
1589 
1590 	if (!HFI1_CAP_IS_USET(PKEY_CHECK))
1591 		return -EPERM;
1592 
1593 	if (get_user(pkey, (u16 __user *)arg))
1594 		return -EFAULT;
1595 
1596 	if (pkey == LIM_MGMT_P_KEY || pkey == FULL_MGMT_P_KEY)
1597 		return -EINVAL;
1598 
1599 	for (i = 0; i < ARRAY_SIZE(ppd->pkeys); i++)
1600 		if (pkey == ppd->pkeys[i])
1601 			return hfi1_set_ctxt_pkey(dd, uctxt, pkey);
1602 
1603 	return -ENOENT;
1604 }
1605 
1606 /**
1607  * ctxt_reset - Reset the user context
1608  * @uctxt: valid user context
1609  */
1610 static int ctxt_reset(struct hfi1_ctxtdata *uctxt)
1611 {
1612 	struct send_context *sc;
1613 	struct hfi1_devdata *dd;
1614 	int ret = 0;
1615 
1616 	if (!uctxt || !uctxt->dd || !uctxt->sc)
1617 		return -EINVAL;
1618 
1619 	/*
1620 	 * There is no protection here. User level has to guarantee that
1621 	 * no one will be writing to the send context while it is being
1622 	 * re-initialized.  If user level breaks that guarantee, it will
1623 	 * break it's own context and no one else's.
1624 	 */
1625 	dd = uctxt->dd;
1626 	sc = uctxt->sc;
1627 
1628 	/*
1629 	 * Wait until the interrupt handler has marked the context as
1630 	 * halted or frozen. Report error if we time out.
1631 	 */
1632 	wait_event_interruptible_timeout(
1633 		sc->halt_wait, (sc->flags & SCF_HALTED),
1634 		msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
1635 	if (!(sc->flags & SCF_HALTED))
1636 		return -ENOLCK;
1637 
1638 	/*
1639 	 * If the send context was halted due to a Freeze, wait until the
1640 	 * device has been "unfrozen" before resetting the context.
1641 	 */
1642 	if (sc->flags & SCF_FROZEN) {
1643 		wait_event_interruptible_timeout(
1644 			dd->event_queue,
1645 			!(READ_ONCE(dd->flags) & HFI1_FROZEN),
1646 			msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
1647 		if (dd->flags & HFI1_FROZEN)
1648 			return -ENOLCK;
1649 
1650 		if (dd->flags & HFI1_FORCED_FREEZE)
1651 			/*
1652 			 * Don't allow context reset if we are into
1653 			 * forced freeze
1654 			 */
1655 			return -ENODEV;
1656 
1657 		sc_disable(sc);
1658 		ret = sc_enable(sc);
1659 		hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_ENB, uctxt);
1660 	} else {
1661 		ret = sc_restart(sc);
1662 	}
1663 	if (!ret)
1664 		sc_return_credits(sc);
1665 
1666 	return ret;
1667 }
1668 
1669 static void user_remove(struct hfi1_devdata *dd)
1670 {
1671 
1672 	hfi1_cdev_cleanup(&dd->user_cdev, &dd->user_device);
1673 }
1674 
1675 static int user_add(struct hfi1_devdata *dd)
1676 {
1677 	char name[10];
1678 	int ret;
1679 
1680 	snprintf(name, sizeof(name), "%s_%d", class_name(), dd->unit);
1681 	ret = hfi1_cdev_init(dd->unit, name, &hfi1_file_ops,
1682 			     &dd->user_cdev, &dd->user_device,
1683 			     true, &dd->kobj);
1684 	if (ret)
1685 		user_remove(dd);
1686 
1687 	return ret;
1688 }
1689 
1690 /*
1691  * Create per-unit files in /dev
1692  */
1693 int hfi1_device_create(struct hfi1_devdata *dd)
1694 {
1695 	return user_add(dd);
1696 }
1697 
1698 /*
1699  * Remove per-unit files in /dev
1700  * void, core kernel returns no errors for this stuff
1701  */
1702 void hfi1_device_remove(struct hfi1_devdata *dd)
1703 {
1704 	user_remove(dd);
1705 }
1706