staging/rdma/hfi1: Remove else after break
[deliverable/linux.git] / drivers / staging / rdma / hfi1 / file_ops.c
CommitLineData
77241056
MM
1/*
2 *
3 * This file is provided under a dual BSD/GPLv2 license. When using or
4 * redistributing this file, you may do so under either license.
5 *
6 * GPL LICENSE SUMMARY
7 *
8 * Copyright(c) 2015 Intel Corporation.
9 *
10 * This program is free software; you can redistribute it and/or modify
11 * it under the terms of version 2 of the GNU General Public License as
12 * published by the Free Software Foundation.
13 *
14 * This program is distributed in the hope that it will be useful, but
15 * WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * General Public License for more details.
18 *
19 * BSD LICENSE
20 *
21 * Copyright(c) 2015 Intel Corporation.
22 *
23 * Redistribution and use in source and binary forms, with or without
24 * modification, are permitted provided that the following conditions
25 * are met:
26 *
27 * - Redistributions of source code must retain the above copyright
28 * notice, this list of conditions and the following disclaimer.
29 * - Redistributions in binary form must reproduce the above copyright
30 * notice, this list of conditions and the following disclaimer in
31 * the documentation and/or other materials provided with the
32 * distribution.
33 * - Neither the name of Intel Corporation nor the names of its
34 * contributors may be used to endorse or promote products derived
35 * from this software without specific prior written permission.
36 *
37 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
38 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
39 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
40 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
41 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
42 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
43 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
44 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
45 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
46 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
47 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
48 *
49 */
77241056
MM
50#include <linux/poll.h>
51#include <linux/cdev.h>
77241056 52#include <linux/vmalloc.h>
77241056 53#include <linux/io.h>
77241056
MM
54
55#include "hfi.h"
56#include "pio.h"
57#include "device.h"
58#include "common.h"
59#include "trace.h"
60#include "user_sdma.h"
701e441d 61#include "user_exp_rcv.h"
77241056 62#include "eprom.h"
affa48de 63#include "aspm.h"
77241056
MM
64
65#undef pr_fmt
66#define pr_fmt(fmt) DRIVER_NAME ": " fmt
67
68#define SEND_CTXT_HALT_TIMEOUT 1000 /* msecs */
69
70/*
71 * File operation functions
72 */
73static int hfi1_file_open(struct inode *, struct file *);
74static int hfi1_file_close(struct inode *, struct file *);
75static ssize_t hfi1_file_write(struct file *, const char __user *,
76 size_t, loff_t *);
77static ssize_t hfi1_write_iter(struct kiocb *, struct iov_iter *);
78static unsigned int hfi1_poll(struct file *, struct poll_table_struct *);
79static int hfi1_file_mmap(struct file *, struct vm_area_struct *);
80
81static u64 kvirt_to_phys(void *);
82static int assign_ctxt(struct file *, struct hfi1_user_info *);
83static int init_subctxts(struct hfi1_ctxtdata *, const struct hfi1_user_info *);
84static int user_init(struct file *);
85static int get_ctxt_info(struct file *, void __user *, __u32);
86static int get_base_info(struct file *, void __user *, __u32);
87static int setup_ctxt(struct file *);
88static int setup_subctxt(struct hfi1_ctxtdata *);
89static int get_user_context(struct file *, struct hfi1_user_info *,
90 int, unsigned);
91static int find_shared_ctxt(struct file *, const struct hfi1_user_info *);
92static int allocate_ctxt(struct file *, struct hfi1_devdata *,
93 struct hfi1_user_info *);
94static unsigned int poll_urgent(struct file *, struct poll_table_struct *);
95static unsigned int poll_next(struct file *, struct poll_table_struct *);
96static int user_event_ack(struct hfi1_ctxtdata *, int, unsigned long);
97static int set_ctxt_pkey(struct hfi1_ctxtdata *, unsigned, u16);
98static int manage_rcvq(struct hfi1_ctxtdata *, unsigned, int);
99static int vma_fault(struct vm_area_struct *, struct vm_fault *);
77241056
MM
100
101static const struct file_operations hfi1_file_ops = {
102 .owner = THIS_MODULE,
103 .write = hfi1_file_write,
104 .write_iter = hfi1_write_iter,
105 .open = hfi1_file_open,
106 .release = hfi1_file_close,
107 .poll = hfi1_poll,
108 .mmap = hfi1_file_mmap,
109 .llseek = noop_llseek,
110};
111
112static struct vm_operations_struct vm_ops = {
113 .fault = vma_fault,
114};
115
116/*
117 * Types of memories mapped into user processes' space
118 */
119enum mmap_types {
120 PIO_BUFS = 1,
121 PIO_BUFS_SOP,
122 PIO_CRED,
123 RCV_HDRQ,
124 RCV_EGRBUF,
125 UREGS,
126 EVENTS,
127 STATUS,
128 RTAIL,
129 SUBCTXT_UREGS,
130 SUBCTXT_RCV_HDRQ,
131 SUBCTXT_EGRBUF,
132 SDMA_COMP
133};
134
135/*
136 * Masks and offsets defining the mmap tokens
137 */
138#define HFI1_MMAP_OFFSET_MASK 0xfffULL
139#define HFI1_MMAP_OFFSET_SHIFT 0
140#define HFI1_MMAP_SUBCTXT_MASK 0xfULL
141#define HFI1_MMAP_SUBCTXT_SHIFT 12
142#define HFI1_MMAP_CTXT_MASK 0xffULL
143#define HFI1_MMAP_CTXT_SHIFT 16
144#define HFI1_MMAP_TYPE_MASK 0xfULL
145#define HFI1_MMAP_TYPE_SHIFT 24
146#define HFI1_MMAP_MAGIC_MASK 0xffffffffULL
147#define HFI1_MMAP_MAGIC_SHIFT 32
148
149#define HFI1_MMAP_MAGIC 0xdabbad00
150
151#define HFI1_MMAP_TOKEN_SET(field, val) \
152 (((val) & HFI1_MMAP_##field##_MASK) << HFI1_MMAP_##field##_SHIFT)
153#define HFI1_MMAP_TOKEN_GET(field, token) \
154 (((token) >> HFI1_MMAP_##field##_SHIFT) & HFI1_MMAP_##field##_MASK)
155#define HFI1_MMAP_TOKEN(type, ctxt, subctxt, addr) \
156 (HFI1_MMAP_TOKEN_SET(MAGIC, HFI1_MMAP_MAGIC) | \
157 HFI1_MMAP_TOKEN_SET(TYPE, type) | \
158 HFI1_MMAP_TOKEN_SET(CTXT, ctxt) | \
159 HFI1_MMAP_TOKEN_SET(SUBCTXT, subctxt) | \
e260e404 160 HFI1_MMAP_TOKEN_SET(OFFSET, (offset_in_page(addr))))
77241056 161
77241056
MM
162#define dbg(fmt, ...) \
163 pr_info(fmt, ##__VA_ARGS__)
164
77241056
MM
165static inline int is_valid_mmap(u64 token)
166{
167 return (HFI1_MMAP_TOKEN_GET(MAGIC, token) == HFI1_MMAP_MAGIC);
168}
169
170static int hfi1_file_open(struct inode *inode, struct file *fp)
171{
172 /* The real work is performed later in assign_ctxt() */
173 fp->private_data = kzalloc(sizeof(struct hfi1_filedata), GFP_KERNEL);
174 if (fp->private_data) /* no cpu affinity by default */
175 ((struct hfi1_filedata *)fp->private_data)->rec_cpu_num = -1;
176 return fp->private_data ? 0 : -ENOMEM;
177}
178
179static ssize_t hfi1_file_write(struct file *fp, const char __user *data,
180 size_t count, loff_t *offset)
181{
182 const struct hfi1_cmd __user *ucmd;
9e10af47
IW
183 struct hfi1_filedata *fd = fp->private_data;
184 struct hfi1_ctxtdata *uctxt = fd->uctxt;
77241056
MM
185 struct hfi1_cmd cmd;
186 struct hfi1_user_info uinfo;
187 struct hfi1_tid_info tinfo;
0b091fb3 188 unsigned long addr;
77241056
MM
189 ssize_t consumed = 0, copy = 0, ret = 0;
190 void *dest = NULL;
191 __u64 user_val = 0;
192 int uctxt_required = 1;
193 int must_be_root = 0;
194
195 if (count < sizeof(cmd)) {
196 ret = -EINVAL;
197 goto bail;
198 }
199
200 ucmd = (const struct hfi1_cmd __user *)data;
201 if (copy_from_user(&cmd, ucmd, sizeof(cmd))) {
202 ret = -EFAULT;
203 goto bail;
204 }
205
206 consumed = sizeof(cmd);
207
208 switch (cmd.type) {
209 case HFI1_CMD_ASSIGN_CTXT:
210 uctxt_required = 0; /* assigned user context not required */
211 copy = sizeof(uinfo);
212 dest = &uinfo;
213 break;
214 case HFI1_CMD_SDMA_STATUS_UPD:
215 case HFI1_CMD_CREDIT_UPD:
216 copy = 0;
217 break;
218 case HFI1_CMD_TID_UPDATE:
219 case HFI1_CMD_TID_FREE:
0b091fb3 220 case HFI1_CMD_TID_INVAL_READ:
77241056
MM
221 copy = sizeof(tinfo);
222 dest = &tinfo;
223 break;
224 case HFI1_CMD_USER_INFO:
225 case HFI1_CMD_RECV_CTRL:
226 case HFI1_CMD_POLL_TYPE:
227 case HFI1_CMD_ACK_EVENT:
228 case HFI1_CMD_CTXT_INFO:
229 case HFI1_CMD_SET_PKEY:
230 case HFI1_CMD_CTXT_RESET:
231 copy = 0;
232 user_val = cmd.addr;
233 break;
234 case HFI1_CMD_EP_INFO:
235 case HFI1_CMD_EP_ERASE_CHIP:
cd371e09
DL
236 case HFI1_CMD_EP_ERASE_RANGE:
237 case HFI1_CMD_EP_READ_RANGE:
238 case HFI1_CMD_EP_WRITE_RANGE:
77241056
MM
239 uctxt_required = 0; /* assigned user context not required */
240 must_be_root = 1; /* validate user */
241 copy = 0;
242 break;
243 default:
244 ret = -EINVAL;
245 goto bail;
246 }
247
248 /* If the command comes with user data, copy it. */
249 if (copy) {
250 if (copy_from_user(dest, (void __user *)cmd.addr, copy)) {
251 ret = -EFAULT;
252 goto bail;
253 }
254 consumed += copy;
255 }
256
257 /*
258 * Make sure there is a uctxt when needed.
259 */
260 if (uctxt_required && !uctxt) {
261 ret = -EINVAL;
262 goto bail;
263 }
264
265 /* only root can do these operations */
266 if (must_be_root && !capable(CAP_SYS_ADMIN)) {
267 ret = -EPERM;
268 goto bail;
269 }
270
271 switch (cmd.type) {
272 case HFI1_CMD_ASSIGN_CTXT:
273 ret = assign_ctxt(fp, &uinfo);
274 if (ret < 0)
275 goto bail;
276 ret = setup_ctxt(fp);
277 if (ret)
278 goto bail;
279 ret = user_init(fp);
280 break;
281 case HFI1_CMD_CTXT_INFO:
282 ret = get_ctxt_info(fp, (void __user *)(unsigned long)
283 user_val, cmd.len);
284 break;
285 case HFI1_CMD_USER_INFO:
286 ret = get_base_info(fp, (void __user *)(unsigned long)
287 user_val, cmd.len);
288 break;
289 case HFI1_CMD_SDMA_STATUS_UPD:
290 break;
291 case HFI1_CMD_CREDIT_UPD:
292 if (uctxt && uctxt->sc)
293 sc_return_credits(uctxt->sc);
294 break;
295 case HFI1_CMD_TID_UPDATE:
0b091fb3 296 ret = hfi1_user_exp_rcv_setup(fp, &tinfo);
77241056 297 if (!ret) {
77241056
MM
298 /*
299 * Copy the number of tidlist entries we used
300 * and the length of the buffer we registered.
301 * These fields are adjacent in the structure so
302 * we can copy them at the same time.
303 */
304 addr = (unsigned long)cmd.addr +
305 offsetof(struct hfi1_tid_info, tidcnt);
306 if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
307 sizeof(tinfo.tidcnt) +
308 sizeof(tinfo.length)))
309 ret = -EFAULT;
310 }
311 break;
0b091fb3
MH
312 case HFI1_CMD_TID_INVAL_READ:
313 ret = hfi1_user_exp_rcv_invalid(fp, &tinfo);
314 if (ret)
315 break;
316 addr = (unsigned long)cmd.addr +
317 offsetof(struct hfi1_tid_info, tidcnt);
318 if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
319 sizeof(tinfo.tidcnt)))
320 ret = -EFAULT;
321 break;
77241056 322 case HFI1_CMD_TID_FREE:
0b091fb3
MH
323 ret = hfi1_user_exp_rcv_clear(fp, &tinfo);
324 if (ret)
325 break;
326 addr = (unsigned long)cmd.addr +
327 offsetof(struct hfi1_tid_info, tidcnt);
328 if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
329 sizeof(tinfo.tidcnt)))
330 ret = -EFAULT;
77241056
MM
331 break;
332 case HFI1_CMD_RECV_CTRL:
9e10af47 333 ret = manage_rcvq(uctxt, fd->subctxt, (int)user_val);
77241056
MM
334 break;
335 case HFI1_CMD_POLL_TYPE:
336 uctxt->poll_type = (typeof(uctxt->poll_type))user_val;
337 break;
338 case HFI1_CMD_ACK_EVENT:
9e10af47 339 ret = user_event_ack(uctxt, fd->subctxt, user_val);
77241056
MM
340 break;
341 case HFI1_CMD_SET_PKEY:
342 if (HFI1_CAP_IS_USET(PKEY_CHECK))
9e10af47 343 ret = set_ctxt_pkey(uctxt, fd->subctxt, user_val);
77241056
MM
344 else
345 ret = -EPERM;
346 break;
347 case HFI1_CMD_CTXT_RESET: {
348 struct send_context *sc;
349 struct hfi1_devdata *dd;
350
351 if (!uctxt || !uctxt->dd || !uctxt->sc) {
352 ret = -EINVAL;
353 break;
354 }
355 /*
356 * There is no protection here. User level has to
357 * guarantee that no one will be writing to the send
358 * context while it is being re-initialized.
359 * If user level breaks that guarantee, it will break
360 * it's own context and no one else's.
361 */
362 dd = uctxt->dd;
363 sc = uctxt->sc;
364 /*
365 * Wait until the interrupt handler has marked the
366 * context as halted or frozen. Report error if we time
367 * out.
368 */
369 wait_event_interruptible_timeout(
370 sc->halt_wait, (sc->flags & SCF_HALTED),
371 msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
372 if (!(sc->flags & SCF_HALTED)) {
373 ret = -ENOLCK;
374 break;
375 }
376 /*
377 * If the send context was halted due to a Freeze,
378 * wait until the device has been "unfrozen" before
379 * resetting the context.
380 */
381 if (sc->flags & SCF_FROZEN) {
382 wait_event_interruptible_timeout(
383 dd->event_queue,
384 !(ACCESS_ONCE(dd->flags) & HFI1_FROZEN),
385 msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
386 if (dd->flags & HFI1_FROZEN) {
387 ret = -ENOLCK;
388 break;
389 }
390 if (dd->flags & HFI1_FORCED_FREEZE) {
4d114fdd
JJ
391 /*
392 * Don't allow context reset if we are into
393 * forced freeze
394 */
77241056
MM
395 ret = -ENODEV;
396 break;
397 }
398 sc_disable(sc);
399 ret = sc_enable(sc);
400 hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_ENB,
401 uctxt->ctxt);
e490974e 402 } else {
77241056 403 ret = sc_restart(sc);
e490974e 404 }
77241056
MM
405 if (!ret)
406 sc_return_credits(sc);
407 break;
408 }
409 case HFI1_CMD_EP_INFO:
410 case HFI1_CMD_EP_ERASE_CHIP:
cd371e09
DL
411 case HFI1_CMD_EP_ERASE_RANGE:
412 case HFI1_CMD_EP_READ_RANGE:
413 case HFI1_CMD_EP_WRITE_RANGE:
d24bc648 414 ret = handle_eprom_command(fp, &cmd);
77241056
MM
415 break;
416 }
417
418 if (ret >= 0)
419 ret = consumed;
420bail:
421 return ret;
422}
423
424static ssize_t hfi1_write_iter(struct kiocb *kiocb, struct iov_iter *from)
425{
9e10af47
IW
426 struct hfi1_filedata *fd = kiocb->ki_filp->private_data;
427 struct hfi1_user_sdma_pkt_q *pq = fd->pq;
428 struct hfi1_user_sdma_comp_q *cq = fd->cq;
77241056
MM
429 int ret = 0, done = 0, reqs = 0;
430 unsigned long dim = from->nr_segs;
431
9e10af47 432 if (!cq || !pq) {
77241056
MM
433 ret = -EIO;
434 goto done;
435 }
436
437 if (!iter_is_iovec(from) || !dim) {
438 ret = -EINVAL;
439 goto done;
440 }
441
442 hfi1_cdbg(SDMA, "SDMA request from %u:%u (%lu)",
9e10af47 443 fd->uctxt->ctxt, fd->subctxt, dim);
77241056
MM
444
445 if (atomic_read(&pq->n_reqs) == pq->n_max_reqs) {
446 ret = -ENOSPC;
447 goto done;
448 }
449
450 while (dim) {
451 unsigned long count = 0;
452
453 ret = hfi1_user_sdma_process_request(
454 kiocb->ki_filp, (struct iovec *)(from->iov + done),
455 dim, &count);
456 if (ret)
457 goto done;
458 dim -= count;
459 done += count;
460 reqs++;
461 }
462done:
463 return ret ? ret : reqs;
464}
465
466static int hfi1_file_mmap(struct file *fp, struct vm_area_struct *vma)
467{
9e10af47
IW
468 struct hfi1_filedata *fd = fp->private_data;
469 struct hfi1_ctxtdata *uctxt = fd->uctxt;
77241056
MM
470 struct hfi1_devdata *dd;
471 unsigned long flags, pfn;
472 u64 token = vma->vm_pgoff << PAGE_SHIFT,
473 memaddr = 0;
474 u8 subctxt, mapio = 0, vmf = 0, type;
475 ssize_t memlen = 0;
476 int ret = 0;
477 u16 ctxt;
478
77241056
MM
479 if (!is_valid_mmap(token) || !uctxt ||
480 !(vma->vm_flags & VM_SHARED)) {
481 ret = -EINVAL;
482 goto done;
483 }
484 dd = uctxt->dd;
485 ctxt = HFI1_MMAP_TOKEN_GET(CTXT, token);
486 subctxt = HFI1_MMAP_TOKEN_GET(SUBCTXT, token);
487 type = HFI1_MMAP_TOKEN_GET(TYPE, token);
9e10af47 488 if (ctxt != uctxt->ctxt || subctxt != fd->subctxt) {
77241056
MM
489 ret = -EINVAL;
490 goto done;
491 }
492
493 flags = vma->vm_flags;
494
495 switch (type) {
496 case PIO_BUFS:
497 case PIO_BUFS_SOP:
498 memaddr = ((dd->physaddr + TXE_PIO_SEND) +
499 /* chip pio base */
d32cf44a 500 (uctxt->sc->hw_context * BIT(16))) +
77241056
MM
501 /* 64K PIO space / ctxt */
502 (type == PIO_BUFS_SOP ?
503 (TXE_PIO_SIZE / 2) : 0); /* sop? */
504 /*
505 * Map only the amount allocated to the context, not the
506 * entire available context's PIO space.
507 */
508 memlen = ALIGN(uctxt->sc->credits * PIO_BLOCK_SIZE,
509 PAGE_SIZE);
510 flags &= ~VM_MAYREAD;
511 flags |= VM_DONTCOPY | VM_DONTEXPAND;
512 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
513 mapio = 1;
514 break;
515 case PIO_CRED:
516 if (flags & VM_WRITE) {
517 ret = -EPERM;
518 goto done;
519 }
520 /*
521 * The credit return location for this context could be on the
522 * second or third page allocated for credit returns (if number
523 * of enabled contexts > 64 and 128 respectively).
524 */
525 memaddr = dd->cr_base[uctxt->numa_id].pa +
526 (((u64)uctxt->sc->hw_free -
527 (u64)dd->cr_base[uctxt->numa_id].va) & PAGE_MASK);
528 memlen = PAGE_SIZE;
529 flags &= ~VM_MAYWRITE;
530 flags |= VM_DONTCOPY | VM_DONTEXPAND;
531 /*
532 * The driver has already allocated memory for credit
533 * returns and programmed it into the chip. Has that
534 * memory been flagged as non-cached?
535 */
536 /* vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); */
537 mapio = 1;
538 break;
539 case RCV_HDRQ:
540 memaddr = uctxt->rcvhdrq_phys;
541 memlen = uctxt->rcvhdrq_size;
542 break;
543 case RCV_EGRBUF: {
544 unsigned long addr;
545 int i;
546 /*
547 * The RcvEgr buffer need to be handled differently
548 * as multiple non-contiguous pages need to be mapped
549 * into the user process.
550 */
551 memlen = uctxt->egrbufs.size;
552 if ((vma->vm_end - vma->vm_start) != memlen) {
553 dd_dev_err(dd, "Eager buffer map size invalid (%lu != %lu)\n",
554 (vma->vm_end - vma->vm_start), memlen);
555 ret = -EINVAL;
556 goto done;
557 }
558 if (vma->vm_flags & VM_WRITE) {
559 ret = -EPERM;
560 goto done;
561 }
562 vma->vm_flags &= ~VM_MAYWRITE;
563 addr = vma->vm_start;
564 for (i = 0 ; i < uctxt->egrbufs.numbufs; i++) {
565 ret = remap_pfn_range(
566 vma, addr,
567 uctxt->egrbufs.buffers[i].phys >> PAGE_SHIFT,
568 uctxt->egrbufs.buffers[i].len,
569 vma->vm_page_prot);
570 if (ret < 0)
571 goto done;
572 addr += uctxt->egrbufs.buffers[i].len;
573 }
574 ret = 0;
575 goto done;
576 }
577 case UREGS:
578 /*
579 * Map only the page that contains this context's user
580 * registers.
581 */
582 memaddr = (unsigned long)
583 (dd->physaddr + RXE_PER_CONTEXT_USER)
584 + (uctxt->ctxt * RXE_PER_CONTEXT_SIZE);
585 /*
586 * TidFlow table is on the same page as the rest of the
587 * user registers.
588 */
589 memlen = PAGE_SIZE;
590 flags |= VM_DONTCOPY | VM_DONTEXPAND;
591 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
592 mapio = 1;
593 break;
594 case EVENTS:
595 /*
596 * Use the page where this context's flags are. User level
597 * knows where it's own bitmap is within the page.
598 */
3c6c065a
MH
599 memaddr = (unsigned long)(dd->events +
600 ((uctxt->ctxt - dd->first_user_ctxt) *
601 HFI1_MAX_SHARED_CTXTS)) & PAGE_MASK;
77241056
MM
602 memlen = PAGE_SIZE;
603 /*
604 * v3.7 removes VM_RESERVED but the effect is kept by
605 * using VM_IO.
606 */
607 flags |= VM_IO | VM_DONTEXPAND;
608 vmf = 1;
609 break;
610 case STATUS:
611 memaddr = kvirt_to_phys((void *)dd->status);
612 memlen = PAGE_SIZE;
613 flags |= VM_IO | VM_DONTEXPAND;
614 break;
615 case RTAIL:
616 if (!HFI1_CAP_IS_USET(DMA_RTAIL)) {
617 /*
618 * If the memory allocation failed, the context alloc
619 * also would have failed, so we would never get here
620 */
621 ret = -EINVAL;
622 goto done;
623 }
624 if (flags & VM_WRITE) {
625 ret = -EPERM;
626 goto done;
627 }
628 memaddr = uctxt->rcvhdrqtailaddr_phys;
629 memlen = PAGE_SIZE;
630 flags &= ~VM_MAYWRITE;
631 break;
632 case SUBCTXT_UREGS:
633 memaddr = (u64)uctxt->subctxt_uregbase;
634 memlen = PAGE_SIZE;
635 flags |= VM_IO | VM_DONTEXPAND;
636 vmf = 1;
637 break;
638 case SUBCTXT_RCV_HDRQ:
639 memaddr = (u64)uctxt->subctxt_rcvhdr_base;
640 memlen = uctxt->rcvhdrq_size * uctxt->subctxt_cnt;
641 flags |= VM_IO | VM_DONTEXPAND;
642 vmf = 1;
643 break;
644 case SUBCTXT_EGRBUF:
645 memaddr = (u64)uctxt->subctxt_rcvegrbuf;
646 memlen = uctxt->egrbufs.size * uctxt->subctxt_cnt;
647 flags |= VM_IO | VM_DONTEXPAND;
648 flags &= ~VM_MAYWRITE;
649 vmf = 1;
650 break;
651 case SDMA_COMP: {
9e10af47 652 struct hfi1_user_sdma_comp_q *cq = fd->cq;
77241056 653
9e10af47 654 if (!cq) {
77241056
MM
655 ret = -EFAULT;
656 goto done;
657 }
77241056
MM
658 memaddr = (u64)cq->comps;
659 memlen = ALIGN(sizeof(*cq->comps) * cq->nentries, PAGE_SIZE);
660 flags |= VM_IO | VM_DONTEXPAND;
661 vmf = 1;
662 break;
663 }
664 default:
665 ret = -EINVAL;
666 break;
667 }
668
669 if ((vma->vm_end - vma->vm_start) != memlen) {
670 hfi1_cdbg(PROC, "%u:%u Memory size mismatch %lu:%lu",
9e10af47 671 uctxt->ctxt, fd->subctxt,
77241056
MM
672 (vma->vm_end - vma->vm_start), memlen);
673 ret = -EINVAL;
674 goto done;
675 }
676
677 vma->vm_flags = flags;
6c63e423
SS
678 hfi1_cdbg(PROC,
679 "%u:%u type:%u io/vf:%d/%d, addr:0x%llx, len:%lu(%lu), flags:0x%lx\n",
680 ctxt, subctxt, type, mapio, vmf, memaddr, memlen,
77241056
MM
681 vma->vm_end - vma->vm_start, vma->vm_flags);
682 pfn = (unsigned long)(memaddr >> PAGE_SHIFT);
683 if (vmf) {
684 vma->vm_pgoff = pfn;
685 vma->vm_ops = &vm_ops;
686 ret = 0;
687 } else if (mapio) {
688 ret = io_remap_pfn_range(vma, vma->vm_start, pfn, memlen,
689 vma->vm_page_prot);
690 } else {
691 ret = remap_pfn_range(vma, vma->vm_start, pfn, memlen,
692 vma->vm_page_prot);
693 }
694done:
695 return ret;
696}
697
698/*
699 * Local (non-chip) user memory is not mapped right away but as it is
700 * accessed by the user-level code.
701 */
702static int vma_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
703{
704 struct page *page;
705
706 page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
707 if (!page)
708 return VM_FAULT_SIGBUS;
709
710 get_page(page);
711 vmf->page = page;
712
713 return 0;
714}
715
716static unsigned int hfi1_poll(struct file *fp, struct poll_table_struct *pt)
717{
718 struct hfi1_ctxtdata *uctxt;
719 unsigned pollflag;
720
9e10af47 721 uctxt = ((struct hfi1_filedata *)fp->private_data)->uctxt;
77241056
MM
722 if (!uctxt)
723 pollflag = POLLERR;
724 else if (uctxt->poll_type == HFI1_POLL_TYPE_URGENT)
725 pollflag = poll_urgent(fp, pt);
726 else if (uctxt->poll_type == HFI1_POLL_TYPE_ANYRCV)
727 pollflag = poll_next(fp, pt);
728 else /* invalid */
729 pollflag = POLLERR;
730
731 return pollflag;
732}
733
734static int hfi1_file_close(struct inode *inode, struct file *fp)
735{
736 struct hfi1_filedata *fdata = fp->private_data;
737 struct hfi1_ctxtdata *uctxt = fdata->uctxt;
738 struct hfi1_devdata *dd;
739 unsigned long flags, *ev;
740
741 fp->private_data = NULL;
742
743 if (!uctxt)
744 goto done;
745
746 hfi1_cdbg(PROC, "freeing ctxt %u:%u", uctxt->ctxt, fdata->subctxt);
747 dd = uctxt->dd;
748 mutex_lock(&hfi1_mutex);
749
750 flush_wc();
751 /* drain user sdma queue */
483119a7 752 hfi1_user_sdma_free_queues(fdata);
77241056 753
957558c9
MH
754 /* release the cpu */
755 hfi1_put_proc_affinity(dd, fdata->rec_cpu_num);
756
77241056
MM
757 /*
758 * Clear any left over, unhandled events so the next process that
759 * gets this context doesn't get confused.
760 */
761 ev = dd->events + ((uctxt->ctxt - dd->first_user_ctxt) *
762 HFI1_MAX_SHARED_CTXTS) + fdata->subctxt;
763 *ev = 0;
764
765 if (--uctxt->cnt) {
766 uctxt->active_slaves &= ~(1 << fdata->subctxt);
767 uctxt->subpid[fdata->subctxt] = 0;
768 mutex_unlock(&hfi1_mutex);
769 goto done;
770 }
771
772 spin_lock_irqsave(&dd->uctxt_lock, flags);
773 /*
774 * Disable receive context and interrupt available, reset all
775 * RcvCtxtCtrl bits to default values.
776 */
777 hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_DIS |
778 HFI1_RCVCTRL_TIDFLOW_DIS |
779 HFI1_RCVCTRL_INTRAVAIL_DIS |
566c157c 780 HFI1_RCVCTRL_TAILUPD_DIS |
77241056
MM
781 HFI1_RCVCTRL_ONE_PKT_EGR_DIS |
782 HFI1_RCVCTRL_NO_RHQ_DROP_DIS |
783 HFI1_RCVCTRL_NO_EGR_DROP_DIS, uctxt->ctxt);
784 /* Clear the context's J_KEY */
785 hfi1_clear_ctxt_jkey(dd, uctxt->ctxt);
786 /*
787 * Reset context integrity checks to default.
788 * (writes to CSRs probably belong in chip.c)
789 */
790 write_kctxt_csr(dd, uctxt->sc->hw_context, SEND_CTXT_CHECK_ENABLE,
791 hfi1_pkt_default_send_ctxt_mask(dd, uctxt->sc->type));
792 sc_disable(uctxt->sc);
793 uctxt->pid = 0;
794 spin_unlock_irqrestore(&dd->uctxt_lock, flags);
795
796 dd->rcd[uctxt->ctxt] = NULL;
797 uctxt->rcvwait_to = 0;
798 uctxt->piowait_to = 0;
799 uctxt->rcvnowait = 0;
800 uctxt->pionowait = 0;
801 uctxt->event_flags = 0;
802
0b091fb3 803 hfi1_user_exp_rcv_free(fdata);
77241056
MM
804 hfi1_clear_ctxt_pkey(dd, uctxt->ctxt);
805
77241056 806 hfi1_stats.sps_ctxts--;
affa48de
AD
807 if (++dd->freectxts == dd->num_user_contexts)
808 aspm_enable_all(dd);
77241056
MM
809 mutex_unlock(&hfi1_mutex);
810 hfi1_free_ctxtdata(dd, uctxt);
811done:
812 kfree(fdata);
813 return 0;
814}
815
816/*
817 * Convert kernel *virtual* addresses to physical addresses.
818 * This is used to vmalloc'ed addresses.
819 */
820static u64 kvirt_to_phys(void *addr)
821{
822 struct page *page;
823 u64 paddr = 0;
824
825 page = vmalloc_to_page(addr);
826 if (page)
827 paddr = page_to_pfn(page) << PAGE_SHIFT;
828
829 return paddr;
830}
831
832static int assign_ctxt(struct file *fp, struct hfi1_user_info *uinfo)
833{
834 int i_minor, ret = 0;
835 unsigned swmajor, swminor, alg = HFI1_ALG_ACROSS;
836
837 swmajor = uinfo->userversion >> 16;
838 if (swmajor != HFI1_USER_SWMAJOR) {
839 ret = -ENODEV;
840 goto done;
841 }
842
843 swminor = uinfo->userversion & 0xffff;
844
845 if (uinfo->hfi1_alg < HFI1_ALG_COUNT)
846 alg = uinfo->hfi1_alg;
847
848 mutex_lock(&hfi1_mutex);
849 /* First, lets check if we need to setup a shared context? */
957558c9
MH
850 if (uinfo->subctxt_cnt) {
851 struct hfi1_filedata *fd = fp->private_data;
852
77241056 853 ret = find_shared_ctxt(fp, uinfo);
957558c9
MH
854 if (ret < 0)
855 goto done_unlock;
856 if (ret)
857 fd->rec_cpu_num = hfi1_get_proc_affinity(
858 fd->uctxt->dd, fd->uctxt->numa_id);
859 }
77241056
MM
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 - 1, alg);
868 }
957558c9 869done_unlock:
77241056
MM
870 mutex_unlock(&hfi1_mutex);
871done:
872 return ret;
873}
874
70224973
DL
875/* return true if the device available for general use */
876static int usable_device(struct hfi1_devdata *dd)
877{
878 struct hfi1_pportdata *ppd = dd->pport;
879
880 return driver_lstate(ppd) == IB_PORT_ACTIVE;
881}
882
77241056
MM
883static int get_user_context(struct file *fp, struct hfi1_user_info *uinfo,
884 int devno, unsigned alg)
885{
886 struct hfi1_devdata *dd = NULL;
887 int ret = 0, devmax, npresent, nup, dev;
888
889 devmax = hfi1_count_units(&npresent, &nup);
890 if (!npresent) {
891 ret = -ENXIO;
892 goto done;
893 }
894 if (!nup) {
895 ret = -ENETDOWN;
896 goto done;
897 }
898 if (devno >= 0) {
899 dd = hfi1_lookup(devno);
900 if (!dd)
901 ret = -ENODEV;
902 else if (!dd->freectxts)
903 ret = -EBUSY;
904 } else {
905 struct hfi1_devdata *pdd;
906
907 if (alg == HFI1_ALG_ACROSS) {
908 unsigned free = 0U;
909
910 for (dev = 0; dev < devmax; dev++) {
911 pdd = hfi1_lookup(dev);
70224973
DL
912 if (!pdd)
913 continue;
914 if (!usable_device(pdd))
915 continue;
916 if (pdd->freectxts &&
77241056
MM
917 pdd->freectxts > free) {
918 dd = pdd;
919 free = pdd->freectxts;
920 }
921 }
922 } else {
923 for (dev = 0; dev < devmax; dev++) {
924 pdd = hfi1_lookup(dev);
70224973
DL
925 if (!pdd)
926 continue;
927 if (!usable_device(pdd))
928 continue;
929 if (pdd->freectxts) {
77241056
MM
930 dd = pdd;
931 break;
932 }
933 }
934 }
935 if (!dd)
936 ret = -EBUSY;
937 }
938done:
939 return ret ? ret : allocate_ctxt(fp, dd, uinfo);
940}
941
942static int find_shared_ctxt(struct file *fp,
943 const struct hfi1_user_info *uinfo)
944{
945 int devmax, ndev, i;
946 int ret = 0;
9e10af47 947 struct hfi1_filedata *fd = fp->private_data;
77241056
MM
948
949 devmax = hfi1_count_units(NULL, NULL);
950
951 for (ndev = 0; ndev < devmax; ndev++) {
952 struct hfi1_devdata *dd = hfi1_lookup(ndev);
953
77241056
MM
954 if (!(dd && (dd->flags & HFI1_PRESENT) && dd->kregbase))
955 continue;
956 for (i = dd->first_user_ctxt; i < dd->num_rcv_contexts; i++) {
957 struct hfi1_ctxtdata *uctxt = dd->rcd[i];
958
959 /* Skip ctxts which are not yet open */
960 if (!uctxt || !uctxt->cnt)
961 continue;
962 /* Skip ctxt if it doesn't match the requested one */
963 if (memcmp(uctxt->uuid, uinfo->uuid,
964 sizeof(uctxt->uuid)) ||
07839049 965 uctxt->jkey != generate_jkey(current_uid()) ||
77241056
MM
966 uctxt->subctxt_id != uinfo->subctxt_id ||
967 uctxt->subctxt_cnt != uinfo->subctxt_cnt)
968 continue;
969
970 /* Verify the sharing process matches the master */
971 if (uctxt->userversion != uinfo->userversion ||
972 uctxt->cnt >= uctxt->subctxt_cnt) {
973 ret = -EINVAL;
974 goto done;
975 }
9e10af47
IW
976 fd->uctxt = uctxt;
977 fd->subctxt = uctxt->cnt++;
978 uctxt->subpid[fd->subctxt] = current->pid;
979 uctxt->active_slaves |= 1 << fd->subctxt;
77241056
MM
980 ret = 1;
981 goto done;
982 }
983 }
984
985done:
986 return ret;
987}
988
989static int allocate_ctxt(struct file *fp, struct hfi1_devdata *dd,
990 struct hfi1_user_info *uinfo)
991{
9e10af47 992 struct hfi1_filedata *fd = fp->private_data;
77241056
MM
993 struct hfi1_ctxtdata *uctxt;
994 unsigned ctxt;
957558c9 995 int ret, numa;
77241056
MM
996
997 if (dd->flags & HFI1_FROZEN) {
998 /*
999 * Pick an error that is unique from all other errors
1000 * that are returned so the user process knows that
1001 * it tried to allocate while the SPC was frozen. It
1002 * it should be able to retry with success in a short
1003 * while.
1004 */
1005 return -EIO;
1006 }
1007
1008 for (ctxt = dd->first_user_ctxt; ctxt < dd->num_rcv_contexts; ctxt++)
1009 if (!dd->rcd[ctxt])
1010 break;
1011
1012 if (ctxt == dd->num_rcv_contexts)
1013 return -EBUSY;
1014
957558c9
MH
1015 fd->rec_cpu_num = hfi1_get_proc_affinity(dd, -1);
1016 if (fd->rec_cpu_num != -1)
1017 numa = cpu_to_node(fd->rec_cpu_num);
1018 else
1019 numa = numa_node_id();
1020 uctxt = hfi1_create_ctxtdata(dd->pport, ctxt, numa);
77241056
MM
1021 if (!uctxt) {
1022 dd_dev_err(dd,
1023 "Unable to allocate ctxtdata memory, failing open\n");
1024 return -ENOMEM;
1025 }
957558c9
MH
1026 hfi1_cdbg(PROC, "[%u:%u] pid %u assigned to CPU %d (NUMA %u)",
1027 uctxt->ctxt, fd->subctxt, current->pid, fd->rec_cpu_num,
1028 uctxt->numa_id);
1029
77241056
MM
1030 /*
1031 * Allocate and enable a PIO send context.
1032 */
1033 uctxt->sc = sc_alloc(dd, SC_USER, uctxt->rcvhdrqentsize,
cc57236f 1034 uctxt->dd->node);
77241056
MM
1035 if (!uctxt->sc)
1036 return -ENOMEM;
1037
6c63e423
SS
1038 hfi1_cdbg(PROC, "allocated send context %u(%u)\n", uctxt->sc->sw_index,
1039 uctxt->sc->hw_context);
77241056
MM
1040 ret = sc_enable(uctxt->sc);
1041 if (ret)
1042 return ret;
1043 /*
1044 * Setup shared context resources if the user-level has requested
1045 * shared contexts and this is the 'master' process.
1046 * This has to be done here so the rest of the sub-contexts find the
1047 * proper master.
1048 */
9e10af47 1049 if (uinfo->subctxt_cnt && !fd->subctxt) {
77241056
MM
1050 ret = init_subctxts(uctxt, uinfo);
1051 /*
1052 * On error, we don't need to disable and de-allocate the
1053 * send context because it will be done during file close
1054 */
1055 if (ret)
1056 return ret;
1057 }
1058 uctxt->userversion = uinfo->userversion;
1059 uctxt->pid = current->pid;
1060 uctxt->flags = HFI1_CAP_UGET(MASK);
1061 init_waitqueue_head(&uctxt->wait);
1062 strlcpy(uctxt->comm, current->comm, sizeof(uctxt->comm));
1063 memcpy(uctxt->uuid, uinfo->uuid, sizeof(uctxt->uuid));
1064 uctxt->jkey = generate_jkey(current_uid());
1065 INIT_LIST_HEAD(&uctxt->sdma_queues);
1066 spin_lock_init(&uctxt->sdma_qlock);
1067 hfi1_stats.sps_ctxts++;
affa48de
AD
1068 /*
1069 * Disable ASPM when there are open user/PSM contexts to avoid
1070 * issues with ASPM L1 exit latency
1071 */
1072 if (dd->freectxts-- == dd->num_user_contexts)
1073 aspm_disable_all(dd);
9e10af47 1074 fd->uctxt = uctxt;
77241056
MM
1075
1076 return 0;
1077}
1078
1079static int init_subctxts(struct hfi1_ctxtdata *uctxt,
1080 const struct hfi1_user_info *uinfo)
1081{
77241056
MM
1082 unsigned num_subctxts;
1083
1084 num_subctxts = uinfo->subctxt_cnt;
acac10fd
MH
1085 if (num_subctxts > HFI1_MAX_SHARED_CTXTS)
1086 return -EINVAL;
77241056
MM
1087
1088 uctxt->subctxt_cnt = uinfo->subctxt_cnt;
1089 uctxt->subctxt_id = uinfo->subctxt_id;
1090 uctxt->active_slaves = 1;
1091 uctxt->redirect_seq_cnt = 1;
1092 set_bit(HFI1_CTXT_MASTER_UNINIT, &uctxt->event_flags);
acac10fd
MH
1093
1094 return 0;
77241056
MM
1095}
1096
1097static int setup_subctxt(struct hfi1_ctxtdata *uctxt)
1098{
1099 int ret = 0;
1100 unsigned num_subctxts = uctxt->subctxt_cnt;
1101
1102 uctxt->subctxt_uregbase = vmalloc_user(PAGE_SIZE);
1103 if (!uctxt->subctxt_uregbase) {
1104 ret = -ENOMEM;
1105 goto bail;
1106 }
1107 /* We can take the size of the RcvHdr Queue from the master */
1108 uctxt->subctxt_rcvhdr_base = vmalloc_user(uctxt->rcvhdrq_size *
1109 num_subctxts);
1110 if (!uctxt->subctxt_rcvhdr_base) {
1111 ret = -ENOMEM;
1112 goto bail_ureg;
1113 }
1114
1115 uctxt->subctxt_rcvegrbuf = vmalloc_user(uctxt->egrbufs.size *
1116 num_subctxts);
1117 if (!uctxt->subctxt_rcvegrbuf) {
1118 ret = -ENOMEM;
1119 goto bail_rhdr;
1120 }
1121 goto bail;
1122bail_rhdr:
1123 vfree(uctxt->subctxt_rcvhdr_base);
1124bail_ureg:
1125 vfree(uctxt->subctxt_uregbase);
1126 uctxt->subctxt_uregbase = NULL;
1127bail:
1128 return ret;
1129}
1130
1131static int user_init(struct file *fp)
1132{
1133 int ret;
1134 unsigned int rcvctrl_ops = 0;
9e10af47
IW
1135 struct hfi1_filedata *fd = fp->private_data;
1136 struct hfi1_ctxtdata *uctxt = fd->uctxt;
77241056
MM
1137
1138 /* make sure that the context has already been setup */
1139 if (!test_bit(HFI1_CTXT_SETUP_DONE, &uctxt->event_flags)) {
1140 ret = -EFAULT;
1141 goto done;
1142 }
1143
1144 /*
1145 * Subctxts don't need to initialize anything since master
1146 * has done it.
1147 */
9e10af47 1148 if (fd->subctxt) {
17fb4f29
JJ
1149 ret = wait_event_interruptible(uctxt->wait, !test_bit(
1150 HFI1_CTXT_MASTER_UNINIT,
1151 &uctxt->event_flags));
0b091fb3 1152 goto expected;
77241056
MM
1153 }
1154
1155 /* initialize poll variables... */
1156 uctxt->urgent = 0;
1157 uctxt->urgent_poll = 0;
1158
1159 /*
1160 * Now enable the ctxt for receive.
1161 * For chips that are set to DMA the tail register to memory
1162 * when they change (and when the update bit transitions from
1163 * 0 to 1. So for those chips, we turn it off and then back on.
1164 * This will (very briefly) affect any other open ctxts, but the
1165 * duration is very short, and therefore isn't an issue. We
1166 * explicitly set the in-memory tail copy to 0 beforehand, so we
1167 * don't have to wait to be sure the DMA update has happened
1168 * (chip resets head/tail to 0 on transition to enable).
1169 */
1170 if (uctxt->rcvhdrtail_kvaddr)
1171 clear_rcvhdrtail(uctxt);
1172
1173 /* Setup J_KEY before enabling the context */
1174 hfi1_set_ctxt_jkey(uctxt->dd, uctxt->ctxt, uctxt->jkey);
1175
1176 rcvctrl_ops = HFI1_RCVCTRL_CTXT_ENB;
1177 if (HFI1_CAP_KGET_MASK(uctxt->flags, HDRSUPP))
1178 rcvctrl_ops |= HFI1_RCVCTRL_TIDFLOW_ENB;
1179 /*
1180 * Ignore the bit in the flags for now until proper
1181 * support for multiple packet per rcv array entry is
1182 * added.
1183 */
1184 if (!HFI1_CAP_KGET_MASK(uctxt->flags, MULTI_PKT_EGR))
1185 rcvctrl_ops |= HFI1_RCVCTRL_ONE_PKT_EGR_ENB;
1186 if (HFI1_CAP_KGET_MASK(uctxt->flags, NODROP_EGR_FULL))
1187 rcvctrl_ops |= HFI1_RCVCTRL_NO_EGR_DROP_ENB;
1188 if (HFI1_CAP_KGET_MASK(uctxt->flags, NODROP_RHQ_FULL))
1189 rcvctrl_ops |= HFI1_RCVCTRL_NO_RHQ_DROP_ENB;
566c157c
MH
1190 /*
1191 * The RcvCtxtCtrl.TailUpd bit has to be explicitly written.
1192 * We can't rely on the correct value to be set from prior
1193 * uses of the chip or ctxt. Therefore, add the rcvctrl op
1194 * for both cases.
1195 */
77241056
MM
1196 if (HFI1_CAP_KGET_MASK(uctxt->flags, DMA_RTAIL))
1197 rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_ENB;
566c157c
MH
1198 else
1199 rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_DIS;
77241056
MM
1200 hfi1_rcvctrl(uctxt->dd, rcvctrl_ops, uctxt->ctxt);
1201
1202 /* Notify any waiting slaves */
1203 if (uctxt->subctxt_cnt) {
1204 clear_bit(HFI1_CTXT_MASTER_UNINIT, &uctxt->event_flags);
1205 wake_up(&uctxt->wait);
1206 }
77241056 1207
0b091fb3
MH
1208expected:
1209 /*
1210 * Expected receive has to be setup for all processes (including
1211 * shared contexts). However, it has to be done after the master
1212 * context has been fully configured as it depends on the
1213 * eager/expected split of the RcvArray entries.
1214 * Setting it up here ensures that the subcontexts will be waiting
1215 * (due to the above wait_event_interruptible() until the master
1216 * is setup.
1217 */
1218 ret = hfi1_user_exp_rcv_init(fp);
77241056
MM
1219done:
1220 return ret;
1221}
1222
1223static int get_ctxt_info(struct file *fp, void __user *ubase, __u32 len)
1224{
1225 struct hfi1_ctxt_info cinfo;
77241056 1226 struct hfi1_filedata *fd = fp->private_data;
9e10af47 1227 struct hfi1_ctxtdata *uctxt = fd->uctxt;
77241056
MM
1228 int ret = 0;
1229
ebe6b2e8 1230 memset(&cinfo, 0, sizeof(cinfo));
77241056
MM
1231 ret = hfi1_get_base_kinfo(uctxt, &cinfo);
1232 if (ret < 0)
1233 goto done;
1234 cinfo.num_active = hfi1_count_active_units();
1235 cinfo.unit = uctxt->dd->unit;
1236 cinfo.ctxt = uctxt->ctxt;
9e10af47 1237 cinfo.subctxt = fd->subctxt;
77241056
MM
1238 cinfo.rcvtids = roundup(uctxt->egrbufs.alloced,
1239 uctxt->dd->rcv_entries.group_size) +
1240 uctxt->expected_count;
1241 cinfo.credits = uctxt->sc->credits;
1242 cinfo.numa_node = uctxt->numa_id;
1243 cinfo.rec_cpu = fd->rec_cpu_num;
1244 cinfo.send_ctxt = uctxt->sc->hw_context;
1245
1246 cinfo.egrtids = uctxt->egrbufs.alloced;
1247 cinfo.rcvhdrq_cnt = uctxt->rcvhdrq_cnt;
1248 cinfo.rcvhdrq_entsize = uctxt->rcvhdrqentsize << 2;
9e10af47 1249 cinfo.sdma_ring_size = fd->cq->nentries;
77241056
MM
1250 cinfo.rcvegr_size = uctxt->egrbufs.rcvtid_size;
1251
9e10af47 1252 trace_hfi1_ctxt_info(uctxt->dd, uctxt->ctxt, fd->subctxt, cinfo);
77241056
MM
1253 if (copy_to_user(ubase, &cinfo, sizeof(cinfo)))
1254 ret = -EFAULT;
1255done:
1256 return ret;
1257}
1258
1259static int setup_ctxt(struct file *fp)
1260{
9e10af47
IW
1261 struct hfi1_filedata *fd = fp->private_data;
1262 struct hfi1_ctxtdata *uctxt = fd->uctxt;
77241056
MM
1263 struct hfi1_devdata *dd = uctxt->dd;
1264 int ret = 0;
1265
1266 /*
1267 * Context should be set up only once (including allocation and
1268 * programming of eager buffers. This is done if context sharing
1269 * is not requested or by the master process.
1270 */
9e10af47 1271 if (!uctxt->subctxt_cnt || !fd->subctxt) {
77241056
MM
1272 ret = hfi1_init_ctxt(uctxt->sc);
1273 if (ret)
1274 goto done;
1275
1276 /* Now allocate the RcvHdr queue and eager buffers. */
1277 ret = hfi1_create_rcvhdrq(dd, uctxt);
1278 if (ret)
1279 goto done;
1280 ret = hfi1_setup_eagerbufs(uctxt);
1281 if (ret)
1282 goto done;
9e10af47 1283 if (uctxt->subctxt_cnt && !fd->subctxt) {
77241056
MM
1284 ret = setup_subctxt(uctxt);
1285 if (ret)
1286 goto done;
1287 }
77241056
MM
1288 }
1289 ret = hfi1_user_sdma_alloc_queues(uctxt, fp);
1290 if (ret)
1291 goto done;
1292
1293 set_bit(HFI1_CTXT_SETUP_DONE, &uctxt->event_flags);
1294done:
1295 return ret;
1296}
1297
1298static int get_base_info(struct file *fp, void __user *ubase, __u32 len)
1299{
1300 struct hfi1_base_info binfo;
9e10af47
IW
1301 struct hfi1_filedata *fd = fp->private_data;
1302 struct hfi1_ctxtdata *uctxt = fd->uctxt;
77241056
MM
1303 struct hfi1_devdata *dd = uctxt->dd;
1304 ssize_t sz;
1305 unsigned offset;
1306 int ret = 0;
1307
1308 trace_hfi1_uctxtdata(uctxt->dd, uctxt);
1309
1310 memset(&binfo, 0, sizeof(binfo));
1311 binfo.hw_version = dd->revision;
1312 binfo.sw_version = HFI1_KERN_SWVERSION;
1313 binfo.bthqp = kdeth_qp;
1314 binfo.jkey = uctxt->jkey;
1315 /*
1316 * If more than 64 contexts are enabled the allocated credit
1317 * return will span two or three contiguous pages. Since we only
1318 * map the page containing the context's credit return address,
1319 * we need to calculate the offset in the proper page.
1320 */
1321 offset = ((u64)uctxt->sc->hw_free -
1322 (u64)dd->cr_base[uctxt->numa_id].va) % PAGE_SIZE;
1323 binfo.sc_credits_addr = HFI1_MMAP_TOKEN(PIO_CRED, uctxt->ctxt,
9e10af47 1324 fd->subctxt, offset);
77241056 1325 binfo.pio_bufbase = HFI1_MMAP_TOKEN(PIO_BUFS, uctxt->ctxt,
9e10af47 1326 fd->subctxt,
77241056
MM
1327 uctxt->sc->base_addr);
1328 binfo.pio_bufbase_sop = HFI1_MMAP_TOKEN(PIO_BUFS_SOP,
1329 uctxt->ctxt,
9e10af47 1330 fd->subctxt,
77241056
MM
1331 uctxt->sc->base_addr);
1332 binfo.rcvhdr_bufbase = HFI1_MMAP_TOKEN(RCV_HDRQ, uctxt->ctxt,
9e10af47 1333 fd->subctxt,
77241056
MM
1334 uctxt->rcvhdrq);
1335 binfo.rcvegr_bufbase = HFI1_MMAP_TOKEN(RCV_EGRBUF, uctxt->ctxt,
9e10af47 1336 fd->subctxt,
77241056
MM
1337 uctxt->egrbufs.rcvtids[0].phys);
1338 binfo.sdma_comp_bufbase = HFI1_MMAP_TOKEN(SDMA_COMP, uctxt->ctxt,
9e10af47 1339 fd->subctxt, 0);
77241056
MM
1340 /*
1341 * user regs are at
1342 * (RXE_PER_CONTEXT_USER + (ctxt * RXE_PER_CONTEXT_SIZE))
1343 */
1344 binfo.user_regbase = HFI1_MMAP_TOKEN(UREGS, uctxt->ctxt,
9e10af47 1345 fd->subctxt, 0);
e260e404 1346 offset = offset_in_page((((uctxt->ctxt - dd->first_user_ctxt) *
9e10af47 1347 HFI1_MAX_SHARED_CTXTS) + fd->subctxt) *
e260e404 1348 sizeof(*dd->events));
77241056 1349 binfo.events_bufbase = HFI1_MMAP_TOKEN(EVENTS, uctxt->ctxt,
9e10af47 1350 fd->subctxt,
77241056
MM
1351 offset);
1352 binfo.status_bufbase = HFI1_MMAP_TOKEN(STATUS, uctxt->ctxt,
9e10af47 1353 fd->subctxt,
77241056
MM
1354 dd->status);
1355 if (HFI1_CAP_IS_USET(DMA_RTAIL))
1356 binfo.rcvhdrtail_base = HFI1_MMAP_TOKEN(RTAIL, uctxt->ctxt,
9e10af47 1357 fd->subctxt, 0);
77241056
MM
1358 if (uctxt->subctxt_cnt) {
1359 binfo.subctxt_uregbase = HFI1_MMAP_TOKEN(SUBCTXT_UREGS,
1360 uctxt->ctxt,
9e10af47 1361 fd->subctxt, 0);
77241056
MM
1362 binfo.subctxt_rcvhdrbuf = HFI1_MMAP_TOKEN(SUBCTXT_RCV_HDRQ,
1363 uctxt->ctxt,
9e10af47 1364 fd->subctxt, 0);
77241056
MM
1365 binfo.subctxt_rcvegrbuf = HFI1_MMAP_TOKEN(SUBCTXT_EGRBUF,
1366 uctxt->ctxt,
9e10af47 1367 fd->subctxt, 0);
77241056
MM
1368 }
1369 sz = (len < sizeof(binfo)) ? len : sizeof(binfo);
1370 if (copy_to_user(ubase, &binfo, sz))
1371 ret = -EFAULT;
1372 return ret;
1373}
1374
1375static unsigned int poll_urgent(struct file *fp,
1376 struct poll_table_struct *pt)
1377{
9e10af47
IW
1378 struct hfi1_filedata *fd = fp->private_data;
1379 struct hfi1_ctxtdata *uctxt = fd->uctxt;
77241056
MM
1380 struct hfi1_devdata *dd = uctxt->dd;
1381 unsigned pollflag;
1382
1383 poll_wait(fp, &uctxt->wait, pt);
1384
1385 spin_lock_irq(&dd->uctxt_lock);
1386 if (uctxt->urgent != uctxt->urgent_poll) {
1387 pollflag = POLLIN | POLLRDNORM;
1388 uctxt->urgent_poll = uctxt->urgent;
1389 } else {
1390 pollflag = 0;
1391 set_bit(HFI1_CTXT_WAITING_URG, &uctxt->event_flags);
1392 }
1393 spin_unlock_irq(&dd->uctxt_lock);
1394
1395 return pollflag;
1396}
1397
1398static unsigned int poll_next(struct file *fp,
1399 struct poll_table_struct *pt)
1400{
9e10af47
IW
1401 struct hfi1_filedata *fd = fp->private_data;
1402 struct hfi1_ctxtdata *uctxt = fd->uctxt;
77241056
MM
1403 struct hfi1_devdata *dd = uctxt->dd;
1404 unsigned pollflag;
1405
1406 poll_wait(fp, &uctxt->wait, pt);
1407
1408 spin_lock_irq(&dd->uctxt_lock);
1409 if (hdrqempty(uctxt)) {
1410 set_bit(HFI1_CTXT_WAITING_RCV, &uctxt->event_flags);
1411 hfi1_rcvctrl(dd, HFI1_RCVCTRL_INTRAVAIL_ENB, uctxt->ctxt);
1412 pollflag = 0;
e490974e 1413 } else {
77241056 1414 pollflag = POLLIN | POLLRDNORM;
e490974e 1415 }
77241056
MM
1416 spin_unlock_irq(&dd->uctxt_lock);
1417
1418 return pollflag;
1419}
1420
1421/*
1422 * Find all user contexts in use, and set the specified bit in their
1423 * event mask.
1424 * See also find_ctxt() for a similar use, that is specific to send buffers.
1425 */
1426int hfi1_set_uevent_bits(struct hfi1_pportdata *ppd, const int evtbit)
1427{
1428 struct hfi1_ctxtdata *uctxt;
1429 struct hfi1_devdata *dd = ppd->dd;
1430 unsigned ctxt;
1431 int ret = 0;
1432 unsigned long flags;
1433
1434 if (!dd->events) {
1435 ret = -EINVAL;
1436 goto done;
1437 }
1438
1439 spin_lock_irqsave(&dd->uctxt_lock, flags);
1440 for (ctxt = dd->first_user_ctxt; ctxt < dd->num_rcv_contexts;
1441 ctxt++) {
1442 uctxt = dd->rcd[ctxt];
1443 if (uctxt) {
1444 unsigned long *evs = dd->events +
1445 (uctxt->ctxt - dd->first_user_ctxt) *
1446 HFI1_MAX_SHARED_CTXTS;
1447 int i;
1448 /*
1449 * subctxt_cnt is 0 if not shared, so do base
1450 * separately, first, then remaining subctxt, if any
1451 */
1452 set_bit(evtbit, evs);
1453 for (i = 1; i < uctxt->subctxt_cnt; i++)
1454 set_bit(evtbit, evs + i);
1455 }
1456 }
1457 spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1458done:
1459 return ret;
1460}
1461
1462/**
1463 * manage_rcvq - manage a context's receive queue
1464 * @uctxt: the context
1465 * @subctxt: the sub-context
1466 * @start_stop: action to carry out
1467 *
1468 * start_stop == 0 disables receive on the context, for use in queue
1469 * overflow conditions. start_stop==1 re-enables, to be used to
1470 * re-init the software copy of the head register
1471 */
1472static int manage_rcvq(struct hfi1_ctxtdata *uctxt, unsigned subctxt,
1473 int start_stop)
1474{
1475 struct hfi1_devdata *dd = uctxt->dd;
1476 unsigned int rcvctrl_op;
1477
1478 if (subctxt)
1479 goto bail;
1480 /* atomically clear receive enable ctxt. */
1481 if (start_stop) {
1482 /*
1483 * On enable, force in-memory copy of the tail register to
1484 * 0, so that protocol code doesn't have to worry about
1485 * whether or not the chip has yet updated the in-memory
1486 * copy or not on return from the system call. The chip
1487 * always resets it's tail register back to 0 on a
1488 * transition from disabled to enabled.
1489 */
1490 if (uctxt->rcvhdrtail_kvaddr)
1491 clear_rcvhdrtail(uctxt);
1492 rcvctrl_op = HFI1_RCVCTRL_CTXT_ENB;
e490974e 1493 } else {
77241056 1494 rcvctrl_op = HFI1_RCVCTRL_CTXT_DIS;
e490974e 1495 }
77241056
MM
1496 hfi1_rcvctrl(dd, rcvctrl_op, uctxt->ctxt);
1497 /* always; new head should be equal to new tail; see above */
1498bail:
1499 return 0;
1500}
1501
1502/*
1503 * clear the event notifier events for this context.
1504 * User process then performs actions appropriate to bit having been
1505 * set, if desired, and checks again in future.
1506 */
1507static int user_event_ack(struct hfi1_ctxtdata *uctxt, int subctxt,
1508 unsigned long events)
1509{
1510 int i;
1511 struct hfi1_devdata *dd = uctxt->dd;
1512 unsigned long *evs;
1513
1514 if (!dd->events)
1515 return 0;
1516
1517 evs = dd->events + ((uctxt->ctxt - dd->first_user_ctxt) *
1518 HFI1_MAX_SHARED_CTXTS) + subctxt;
1519
1520 for (i = 0; i <= _HFI1_MAX_EVENT_BIT; i++) {
1521 if (!test_bit(i, &events))
1522 continue;
1523 clear_bit(i, evs);
1524 }
1525 return 0;
1526}
1527
77241056
MM
1528static int set_ctxt_pkey(struct hfi1_ctxtdata *uctxt, unsigned subctxt,
1529 u16 pkey)
1530{
1531 int ret = -ENOENT, i, intable = 0;
1532 struct hfi1_pportdata *ppd = uctxt->ppd;
1533 struct hfi1_devdata *dd = uctxt->dd;
1534
1535 if (pkey == LIM_MGMT_P_KEY || pkey == FULL_MGMT_P_KEY) {
1536 ret = -EINVAL;
1537 goto done;
1538 }
1539
1540 for (i = 0; i < ARRAY_SIZE(ppd->pkeys); i++)
1541 if (pkey == ppd->pkeys[i]) {
1542 intable = 1;
1543 break;
1544 }
1545
1546 if (intable)
1547 ret = hfi1_set_ctxt_pkey(dd, uctxt->ctxt, pkey);
1548done:
1549 return ret;
1550}
1551
1552static int ui_open(struct inode *inode, struct file *filp)
1553{
1554 struct hfi1_devdata *dd;
1555
1556 dd = container_of(inode->i_cdev, struct hfi1_devdata, ui_cdev);
1557 filp->private_data = dd; /* for other methods */
1558 return 0;
1559}
1560
1561static int ui_release(struct inode *inode, struct file *filp)
1562{
1563 /* nothing to do */
1564 return 0;
1565}
1566
1567static loff_t ui_lseek(struct file *filp, loff_t offset, int whence)
1568{
1569 struct hfi1_devdata *dd = filp->private_data;
1570
1571 switch (whence) {
1572 case SEEK_SET:
1573 break;
1574 case SEEK_CUR:
1575 offset += filp->f_pos;
1576 break;
1577 case SEEK_END:
1578 offset = ((dd->kregend - dd->kregbase) + DC8051_DATA_MEM_SIZE) -
1579 offset;
1580 break;
1581 default:
1582 return -EINVAL;
1583 }
1584
1585 if (offset < 0)
1586 return -EINVAL;
1587
1588 if (offset >= (dd->kregend - dd->kregbase) + DC8051_DATA_MEM_SIZE)
1589 return -EINVAL;
1590
1591 filp->f_pos = offset;
1592
1593 return filp->f_pos;
1594}
1595
77241056
MM
1596/* NOTE: assumes unsigned long is 8 bytes */
1597static ssize_t ui_read(struct file *filp, char __user *buf, size_t count,
17fb4f29 1598 loff_t *f_pos)
77241056
MM
1599{
1600 struct hfi1_devdata *dd = filp->private_data;
1601 void __iomem *base = dd->kregbase;
1602 unsigned long total, csr_off,
1603 barlen = (dd->kregend - dd->kregbase);
1604 u64 data;
1605
1606 /* only read 8 byte quantities */
1607 if ((count % 8) != 0)
1608 return -EINVAL;
1609 /* offset must be 8-byte aligned */
1610 if ((*f_pos % 8) != 0)
1611 return -EINVAL;
1612 /* destination buffer must be 8-byte aligned */
1613 if ((unsigned long)buf % 8 != 0)
1614 return -EINVAL;
1615 /* must be in range */
1616 if (*f_pos + count > (barlen + DC8051_DATA_MEM_SIZE))
1617 return -EINVAL;
1618 /* only set the base if we are not starting past the BAR */
1619 if (*f_pos < barlen)
1620 base += *f_pos;
1621 csr_off = *f_pos;
1622 for (total = 0; total < count; total += 8, csr_off += 8) {
1623 /* accessing LCB CSRs requires more checks */
1624 if (is_lcb_offset(csr_off)) {
1625 if (read_lcb_csr(dd, csr_off, (u64 *)&data))
1626 break; /* failed */
1627 }
1628 /*
1629 * Cannot read ASIC GPIO/QSFP* clear and force CSRs without a
1630 * false parity error. Avoid the whole issue by not reading
1631 * them. These registers are defined as having a read value
1632 * of 0.
1633 */
d0d236ea
JJ
1634 else if (csr_off == ASIC_GPIO_CLEAR ||
1635 csr_off == ASIC_GPIO_FORCE ||
1636 csr_off == ASIC_QSFP1_CLEAR ||
1637 csr_off == ASIC_QSFP1_FORCE ||
1638 csr_off == ASIC_QSFP2_CLEAR ||
1639 csr_off == ASIC_QSFP2_FORCE)
77241056
MM
1640 data = 0;
1641 else if (csr_off >= barlen) {
1642 /*
1643 * read_8051_data can read more than just 8 bytes at
1644 * a time. However, folding this into the loop and
1645 * handling the reads in 8 byte increments allows us
1646 * to smoothly transition from chip memory to 8051
1647 * memory.
1648 */
1649 if (read_8051_data(dd,
1650 (u32)(csr_off - barlen),
1651 sizeof(data), &data))
1652 break; /* failed */
1653 } else
1654 data = readq(base + total);
1655 if (put_user(data, (unsigned long __user *)(buf + total)))
1656 break;
1657 }
1658 *f_pos += total;
1659 return total;
1660}
1661
1662/* NOTE: assumes unsigned long is 8 bytes */
1663static ssize_t ui_write(struct file *filp, const char __user *buf,
1664 size_t count, loff_t *f_pos)
1665{
1666 struct hfi1_devdata *dd = filp->private_data;
1667 void __iomem *base;
1668 unsigned long total, data, csr_off;
1669 int in_lcb;
1670
1671 /* only write 8 byte quantities */
1672 if ((count % 8) != 0)
1673 return -EINVAL;
1674 /* offset must be 8-byte aligned */
1675 if ((*f_pos % 8) != 0)
1676 return -EINVAL;
1677 /* source buffer must be 8-byte aligned */
1678 if ((unsigned long)buf % 8 != 0)
1679 return -EINVAL;
1680 /* must be in range */
1681 if (*f_pos + count > dd->kregend - dd->kregbase)
1682 return -EINVAL;
1683
1684 base = (void __iomem *)dd->kregbase + *f_pos;
1685 csr_off = *f_pos;
1686 in_lcb = 0;
1687 for (total = 0; total < count; total += 8, csr_off += 8) {
1688 if (get_user(data, (unsigned long __user *)(buf + total)))
1689 break;
1690 /* accessing LCB CSRs requires a special procedure */
1691 if (is_lcb_offset(csr_off)) {
1692 if (!in_lcb) {
1693 int ret = acquire_lcb_access(dd, 1);
1694
1695 if (ret)
1696 break;
1697 in_lcb = 1;
1698 }
1699 } else {
1700 if (in_lcb) {
1701 release_lcb_access(dd, 1);
1702 in_lcb = 0;
1703 }
1704 }
1705 writeq(data, base + total);
1706 }
1707 if (in_lcb)
1708 release_lcb_access(dd, 1);
1709 *f_pos += total;
1710 return total;
1711}
1712
1713static const struct file_operations ui_file_ops = {
1714 .owner = THIS_MODULE,
1715 .llseek = ui_lseek,
1716 .read = ui_read,
1717 .write = ui_write,
1718 .open = ui_open,
1719 .release = ui_release,
1720};
b91cc573 1721
77241056
MM
1722#define UI_OFFSET 192 /* device minor offset for UI devices */
1723static int create_ui = 1;
1724
1725static struct cdev wildcard_cdev;
1726static struct device *wildcard_device;
1727
1728static atomic_t user_count = ATOMIC_INIT(0);
1729
1730static void user_remove(struct hfi1_devdata *dd)
1731{
1732 if (atomic_dec_return(&user_count) == 0)
1733 hfi1_cdev_cleanup(&wildcard_cdev, &wildcard_device);
1734
1735 hfi1_cdev_cleanup(&dd->user_cdev, &dd->user_device);
1736 hfi1_cdev_cleanup(&dd->ui_cdev, &dd->ui_device);
1737}
1738
1739static int user_add(struct hfi1_devdata *dd)
1740{
1741 char name[10];
1742 int ret;
1743
1744 if (atomic_inc_return(&user_count) == 1) {
1745 ret = hfi1_cdev_init(0, class_name(), &hfi1_file_ops,
e116a64f
IW
1746 &wildcard_cdev, &wildcard_device,
1747 true);
77241056
MM
1748 if (ret)
1749 goto done;
1750 }
1751
1752 snprintf(name, sizeof(name), "%s_%d", class_name(), dd->unit);
1753 ret = hfi1_cdev_init(dd->unit + 1, name, &hfi1_file_ops,
e116a64f
IW
1754 &dd->user_cdev, &dd->user_device,
1755 true);
77241056
MM
1756 if (ret)
1757 goto done;
1758
1759 if (create_ui) {
1760 snprintf(name, sizeof(name),
1761 "%s_ui%d", class_name(), dd->unit);
1762 ret = hfi1_cdev_init(dd->unit + UI_OFFSET, name, &ui_file_ops,
e116a64f
IW
1763 &dd->ui_cdev, &dd->ui_device,
1764 false);
77241056
MM
1765 if (ret)
1766 goto done;
1767 }
1768
1769 return 0;
1770done:
1771 user_remove(dd);
1772 return ret;
1773}
1774
1775/*
1776 * Create per-unit files in /dev
1777 */
1778int hfi1_device_create(struct hfi1_devdata *dd)
1779{
1780 int r, ret;
1781
1782 r = user_add(dd);
1783 ret = hfi1_diag_add(dd);
1784 if (r && !ret)
1785 ret = r;
1786 return ret;
1787}
1788
1789/*
1790 * Remove per-unit files in /dev
1791 * void, core kernel returns no errors for this stuff
1792 */
1793void hfi1_device_remove(struct hfi1_devdata *dd)
1794{
1795 user_remove(dd);
1796 hfi1_diag_remove(dd);
1797}
This page took 0.177567 seconds and 5 git commands to generate.