projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
IB/hfi1: Re-factor MMU notification code
[deliverable/linux.git] / drivers / staging / rdma / hfi1 / file_ops.c
1 /*
2 * Copyright(c) 2015, 2016 Intel Corporation.
3 *
4 * This file is provided under a dual BSD/GPLv2 license. When using or
5 * redistributing this file, you may do so under either license.
6 *
7 * GPL LICENSE SUMMARY
8 *
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of version 2 of the GNU General Public License as
11 * published by the Free Software Foundation.
12 *
13 * This program is distributed in the hope that it will be useful, but
14 * WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * General Public License for more details.
17 *
18 * BSD LICENSE
19 *
20 * Redistribution and use in source and binary forms, with or without
21 * modification, are permitted provided that the following conditions
22 * are met:
23 *
24 * - Redistributions of source code must retain the above copyright
25 * notice, this list of conditions and the following disclaimer.
26 * - Redistributions in binary form must reproduce the above copyright
27 * notice, this list of conditions and the following disclaimer in
28 * the documentation and/or other materials provided with the
29 * distribution.
30 * - Neither the name of Intel Corporation nor the names of its
31 * contributors may be used to endorse or promote products derived
32 * from this software without specific prior written permission.
33 *
34 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
45 *
46 */
47 #include <linux/poll.h>
48 #include <linux/cdev.h>
49 #include <linux/vmalloc.h>
50 #include <linux/io.h>
51
52 #include "hfi.h"
53 #include "pio.h"
54 #include "device.h"
55 #include "common.h"
56 #include "trace.h"
57 #include "user_sdma.h"
58 #include "user_exp_rcv.h"
59 #include "eprom.h"
60 #include "aspm.h"
61 #include "mmu_rb.h"
62
63 #undef pr_fmt
64 #define pr_fmt(fmt) DRIVER_NAME ": " fmt
65
66 #define SEND_CTXT_HALT_TIMEOUT 1000 /* msecs */
67
68 /*
69 * File operation functions
70 */
71 static int hfi1_file_open(struct inode *, struct file *);
72 static int hfi1_file_close(struct inode *, struct file *);
73 static ssize_t hfi1_file_write(struct file *, const char __user *,
74 size_t, loff_t *);
75 static ssize_t hfi1_write_iter(struct kiocb *, struct iov_iter *);
76 static unsigned int hfi1_poll(struct file *, struct poll_table_struct *);
77 static int hfi1_file_mmap(struct file *, struct vm_area_struct *);
78
79 static u64 kvirt_to_phys(void *);
80 static int assign_ctxt(struct file *, struct hfi1_user_info *);
81 static int init_subctxts(struct hfi1_ctxtdata *, const struct hfi1_user_info *);
82 static int user_init(struct file *);
83 static int get_ctxt_info(struct file *, void __user *, __u32);
84 static int get_base_info(struct file *, void __user *, __u32);
85 static int setup_ctxt(struct file *);
86 static int setup_subctxt(struct hfi1_ctxtdata *);
87 static int get_user_context(struct file *, struct hfi1_user_info *,
88 int, unsigned);
89 static int find_shared_ctxt(struct file *, const struct hfi1_user_info *);
90 static int allocate_ctxt(struct file *, struct hfi1_devdata *,
91 struct hfi1_user_info *);
92 static unsigned int poll_urgent(struct file *, struct poll_table_struct *);
93 static unsigned int poll_next(struct file *, struct poll_table_struct *);
94 static int user_event_ack(struct hfi1_ctxtdata *, int, unsigned long);
95 static int set_ctxt_pkey(struct hfi1_ctxtdata *, unsigned, u16);
96 static int manage_rcvq(struct hfi1_ctxtdata *, unsigned, int);
97 static int vma_fault(struct vm_area_struct *, struct vm_fault *);
98
99 static const struct file_operations hfi1_file_ops = {
100 .owner = THIS_MODULE,
101 .write = hfi1_file_write,
102 .write_iter = hfi1_write_iter,
103 .open = hfi1_file_open,
104 .release = hfi1_file_close,
105 .poll = hfi1_poll,
106 .mmap = hfi1_file_mmap,
107 .llseek = noop_llseek,
108 };
109
110 static struct vm_operations_struct vm_ops = {
111 .fault = vma_fault,
112 };
113
114 /*
115 * Types of memories mapped into user processes' space
116 */
117 enum mmap_types {
118 PIO_BUFS = 1,
119 PIO_BUFS_SOP,
120 PIO_CRED,
121 RCV_HDRQ,
122 RCV_EGRBUF,
123 UREGS,
124 EVENTS,
125 STATUS,
126 RTAIL,
127 SUBCTXT_UREGS,
128 SUBCTXT_RCV_HDRQ,
129 SUBCTXT_EGRBUF,
130 SDMA_COMP
131 };
132
133 /*
134 * Masks and offsets defining the mmap tokens
135 */
136 #define HFI1_MMAP_OFFSET_MASK 0xfffULL
137 #define HFI1_MMAP_OFFSET_SHIFT 0
138 #define HFI1_MMAP_SUBCTXT_MASK 0xfULL
139 #define HFI1_MMAP_SUBCTXT_SHIFT 12
140 #define HFI1_MMAP_CTXT_MASK 0xffULL
141 #define HFI1_MMAP_CTXT_SHIFT 16
142 #define HFI1_MMAP_TYPE_MASK 0xfULL
143 #define HFI1_MMAP_TYPE_SHIFT 24
144 #define HFI1_MMAP_MAGIC_MASK 0xffffffffULL
145 #define HFI1_MMAP_MAGIC_SHIFT 32
146
147 #define HFI1_MMAP_MAGIC 0xdabbad00
148
149 #define HFI1_MMAP_TOKEN_SET(field, val) \
150 (((val) & HFI1_MMAP_##field##_MASK) << HFI1_MMAP_##field##_SHIFT)
151 #define HFI1_MMAP_TOKEN_GET(field, token) \
152 (((token) >> HFI1_MMAP_##field##_SHIFT) & HFI1_MMAP_##field##_MASK)
153 #define HFI1_MMAP_TOKEN(type, ctxt, subctxt, addr) \
154 (HFI1_MMAP_TOKEN_SET(MAGIC, HFI1_MMAP_MAGIC) | \
155 HFI1_MMAP_TOKEN_SET(TYPE, type) | \
156 HFI1_MMAP_TOKEN_SET(CTXT, ctxt) | \
157 HFI1_MMAP_TOKEN_SET(SUBCTXT, subctxt) | \
158 HFI1_MMAP_TOKEN_SET(OFFSET, (offset_in_page(addr))))
159
160 #define dbg(fmt, ...) \
161 pr_info(fmt, ##__VA_ARGS__)
162
163 static inline int is_valid_mmap(u64 token)
164 {
165 return (HFI1_MMAP_TOKEN_GET(MAGIC, token) == HFI1_MMAP_MAGIC);
166 }
167
168 static int hfi1_file_open(struct inode *inode, struct file *fp)
169 {
170 /* The real work is performed later in assign_ctxt() */
171 fp->private_data = kzalloc(sizeof(struct hfi1_filedata), GFP_KERNEL);
172 if (fp->private_data) /* no cpu affinity by default */
173 ((struct hfi1_filedata *)fp->private_data)->rec_cpu_num = -1;
174 return fp->private_data ? 0 : -ENOMEM;
175 }
176
177 static ssize_t hfi1_file_write(struct file *fp, const char __user *data,
178 size_t count, loff_t *offset)
179 {
180 const struct hfi1_cmd __user *ucmd;
181 struct hfi1_filedata *fd = fp->private_data;
182 struct hfi1_ctxtdata *uctxt = fd->uctxt;
183 struct hfi1_cmd cmd;
184 struct hfi1_user_info uinfo;
185 struct hfi1_tid_info tinfo;
186 unsigned long addr;
187 ssize_t consumed = 0, copy = 0, ret = 0;
188 void *dest = NULL;
189 __u64 user_val = 0;
190 int uctxt_required = 1;
191 int must_be_root = 0;
192
193 if (count < sizeof(cmd)) {
194 ret = -EINVAL;
195 goto bail;
196 }
197
198 ucmd = (const struct hfi1_cmd __user *)data;
199 if (copy_from_user(&cmd, ucmd, sizeof(cmd))) {
200 ret = -EFAULT;
201 goto bail;
202 }
203
204 consumed = sizeof(cmd);
205
206 switch (cmd.type) {
207 case HFI1_CMD_ASSIGN_CTXT:
208 uctxt_required = 0; /* assigned user context not required */
209 copy = sizeof(uinfo);
210 dest = &uinfo;
211 break;
212 case HFI1_CMD_SDMA_STATUS_UPD:
213 case HFI1_CMD_CREDIT_UPD:
214 copy = 0;
215 break;
216 case HFI1_CMD_TID_UPDATE:
217 case HFI1_CMD_TID_FREE:
218 case HFI1_CMD_TID_INVAL_READ:
219 copy = sizeof(tinfo);
220 dest = &tinfo;
221 break;
222 case HFI1_CMD_USER_INFO:
223 case HFI1_CMD_RECV_CTRL:
224 case HFI1_CMD_POLL_TYPE:
225 case HFI1_CMD_ACK_EVENT:
226 case HFI1_CMD_CTXT_INFO:
227 case HFI1_CMD_SET_PKEY:
228 case HFI1_CMD_CTXT_RESET:
229 copy = 0;
230 user_val = cmd.addr;
231 break;
232 case HFI1_CMD_EP_INFO:
233 case HFI1_CMD_EP_ERASE_CHIP:
234 case HFI1_CMD_EP_ERASE_RANGE:
235 case HFI1_CMD_EP_READ_RANGE:
236 case HFI1_CMD_EP_WRITE_RANGE:
237 uctxt_required = 0; /* assigned user context not required */
238 must_be_root = 1; /* validate user */
239 copy = 0;
240 break;
241 default:
242 ret = -EINVAL;
243 goto bail;
244 }
245
246 /* If the command comes with user data, copy it. */
247 if (copy) {
248 if (copy_from_user(dest, (void __user *)cmd.addr, copy)) {
249 ret = -EFAULT;
250 goto bail;
251 }
252 consumed += copy;
253 }
254
255 /*
256 * Make sure there is a uctxt when needed.
257 */
258 if (uctxt_required && !uctxt) {
259 ret = -EINVAL;
260 goto bail;
261 }
262
263 /* only root can do these operations */
264 if (must_be_root && !capable(CAP_SYS_ADMIN)) {
265 ret = -EPERM;
266 goto bail;
267 }
268
269 switch (cmd.type) {
270 case HFI1_CMD_ASSIGN_CTXT:
271 ret = assign_ctxt(fp, &uinfo);
272 if (ret < 0)
273 goto bail;
274 ret = setup_ctxt(fp);
275 if (ret)
276 goto bail;
277 ret = user_init(fp);
278 break;
279 case HFI1_CMD_CTXT_INFO:
280 ret = get_ctxt_info(fp, (void __user *)(unsigned long)
281 user_val, cmd.len);
282 break;
283 case HFI1_CMD_USER_INFO:
284 ret = get_base_info(fp, (void __user *)(unsigned long)
285 user_val, cmd.len);
286 break;
287 case HFI1_CMD_SDMA_STATUS_UPD:
288 break;
289 case HFI1_CMD_CREDIT_UPD:
290 if (uctxt && uctxt->sc)
291 sc_return_credits(uctxt->sc);
292 break;
293 case HFI1_CMD_TID_UPDATE:
294 ret = hfi1_user_exp_rcv_setup(fp, &tinfo);
295 if (!ret) {
296 /*
297 * Copy the number of tidlist entries we used
298 * and the length of the buffer we registered.
299 * These fields are adjacent in the structure so
300 * we can copy them at the same time.
301 */
302 addr = (unsigned long)cmd.addr +
303 offsetof(struct hfi1_tid_info, tidcnt);
304 if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
305 sizeof(tinfo.tidcnt) +
306 sizeof(tinfo.length)))
307 ret = -EFAULT;
308 }
309 break;
310 case HFI1_CMD_TID_INVAL_READ:
311 ret = hfi1_user_exp_rcv_invalid(fp, &tinfo);
312 if (ret)
313 break;
314 addr = (unsigned long)cmd.addr +
315 offsetof(struct hfi1_tid_info, tidcnt);
316 if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
317 sizeof(tinfo.tidcnt)))
318 ret = -EFAULT;
319 break;
320 case HFI1_CMD_TID_FREE:
321 ret = hfi1_user_exp_rcv_clear(fp, &tinfo);
322 if (ret)
323 break;
324 addr = (unsigned long)cmd.addr +
325 offsetof(struct hfi1_tid_info, tidcnt);
326 if (copy_to_user((void __user *)addr, &tinfo.tidcnt,
327 sizeof(tinfo.tidcnt)))
328 ret = -EFAULT;
329 break;
330 case HFI1_CMD_RECV_CTRL:
331 ret = manage_rcvq(uctxt, fd->subctxt, (int)user_val);
332 break;
333 case HFI1_CMD_POLL_TYPE:
334 uctxt->poll_type = (typeof(uctxt->poll_type))user_val;
335 break;
336 case HFI1_CMD_ACK_EVENT:
337 ret = user_event_ack(uctxt, fd->subctxt, user_val);
338 break;
339 case HFI1_CMD_SET_PKEY:
340 if (HFI1_CAP_IS_USET(PKEY_CHECK))
341 ret = set_ctxt_pkey(uctxt, fd->subctxt, user_val);
342 else
343 ret = -EPERM;
344 break;
345 case HFI1_CMD_CTXT_RESET: {
346 struct send_context *sc;
347 struct hfi1_devdata *dd;
348
349 if (!uctxt || !uctxt->dd || !uctxt->sc) {
350 ret = -EINVAL;
351 break;
352 }
353 /*
354 * There is no protection here. User level has to
355 * guarantee that no one will be writing to the send
356 * context while it is being re-initialized.
357 * If user level breaks that guarantee, it will break
358 * it's own context and no one else's.
359 */
360 dd = uctxt->dd;
361 sc = uctxt->sc;
362 /*
363 * Wait until the interrupt handler has marked the
364 * context as halted or frozen. Report error if we time
365 * out.
366 */
367 wait_event_interruptible_timeout(
368 sc->halt_wait, (sc->flags & SCF_HALTED),
369 msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
370 if (!(sc->flags & SCF_HALTED)) {
371 ret = -ENOLCK;
372 break;
373 }
374 /*
375 * If the send context was halted due to a Freeze,
376 * wait until the device has been "unfrozen" before
377 * resetting the context.
378 */
379 if (sc->flags & SCF_FROZEN) {
380 wait_event_interruptible_timeout(
381 dd->event_queue,
382 !(ACCESS_ONCE(dd->flags) & HFI1_FROZEN),
383 msecs_to_jiffies(SEND_CTXT_HALT_TIMEOUT));
384 if (dd->flags & HFI1_FROZEN) {
385 ret = -ENOLCK;
386 break;
387 }
388 if (dd->flags & HFI1_FORCED_FREEZE) {
389 /*
390 * Don't allow context reset if we are into
391 * forced freeze
392 */
393 ret = -ENODEV;
394 break;
395 }
396 sc_disable(sc);
397 ret = sc_enable(sc);
398 hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_ENB,
399 uctxt->ctxt);
400 } else {
401 ret = sc_restart(sc);
402 }
403 if (!ret)
404 sc_return_credits(sc);
405 break;
406 }
407 case HFI1_CMD_EP_INFO:
408 case HFI1_CMD_EP_ERASE_CHIP:
409 case HFI1_CMD_EP_ERASE_RANGE:
410 case HFI1_CMD_EP_READ_RANGE:
411 case HFI1_CMD_EP_WRITE_RANGE:
412 ret = handle_eprom_command(fp, &cmd);
413 break;
414 }
415
416 if (ret >= 0)
417 ret = consumed;
418 bail:
419 return ret;
420 }
421
422 static ssize_t hfi1_write_iter(struct kiocb *kiocb, struct iov_iter *from)
423 {
424 struct hfi1_filedata *fd = kiocb->ki_filp->private_data;
425 struct hfi1_user_sdma_pkt_q *pq = fd->pq;
426 struct hfi1_user_sdma_comp_q *cq = fd->cq;
427 int ret = 0, done = 0, reqs = 0;
428 unsigned long dim = from->nr_segs;
429
430 if (!cq || !pq) {
431 ret = -EIO;
432 goto done;
433 }
434
435 if (!iter_is_iovec(from) || !dim) {
436 ret = -EINVAL;
437 goto done;
438 }
439
440 hfi1_cdbg(SDMA, "SDMA request from %u:%u (%lu)",
441 fd->uctxt->ctxt, fd->subctxt, dim);
442
443 if (atomic_read(&pq->n_reqs) == pq->n_max_reqs) {
444 ret = -ENOSPC;
445 goto done;
446 }
447
448 while (dim) {
449 unsigned long count = 0;
450
451 ret = hfi1_user_sdma_process_request(
452 kiocb->ki_filp, (struct iovec *)(from->iov + done),
453 dim, &count);
454 if (ret)
455 goto done;
456 dim -= count;
457 done += count;
458 reqs++;
459 }
460 done:
461 return ret ? ret : reqs;
462 }
463
464 static int hfi1_file_mmap(struct file *fp, struct vm_area_struct *vma)
465 {
466 struct hfi1_filedata *fd = fp->private_data;
467 struct hfi1_ctxtdata *uctxt = fd->uctxt;
468 struct hfi1_devdata *dd;
469 unsigned long flags, pfn;
470 u64 token = vma->vm_pgoff << PAGE_SHIFT,
471 memaddr = 0;
472 u8 subctxt, mapio = 0, vmf = 0, type;
473 ssize_t memlen = 0;
474 int ret = 0;
475 u16 ctxt;
476
477 if (!is_valid_mmap(token) || !uctxt ||
478 !(vma->vm_flags & VM_SHARED)) {
479 ret = -EINVAL;
480 goto done;
481 }
482 dd = uctxt->dd;
483 ctxt = HFI1_MMAP_TOKEN_GET(CTXT, token);
484 subctxt = HFI1_MMAP_TOKEN_GET(SUBCTXT, token);
485 type = HFI1_MMAP_TOKEN_GET(TYPE, token);
486 if (ctxt != uctxt->ctxt || subctxt != fd->subctxt) {
487 ret = -EINVAL;
488 goto done;
489 }
490
491 flags = vma->vm_flags;
492
493 switch (type) {
494 case PIO_BUFS:
495 case PIO_BUFS_SOP:
496 memaddr = ((dd->physaddr + TXE_PIO_SEND) +
497 /* chip pio base */
498 (uctxt->sc->hw_context * BIT(16))) +
499 /* 64K PIO space / ctxt */
500 (type == PIO_BUFS_SOP ?
501 (TXE_PIO_SIZE / 2) : 0); /* sop? */
502 /*
503 * Map only the amount allocated to the context, not the
504 * entire available context's PIO space.
505 */
506 memlen = ALIGN(uctxt->sc->credits * PIO_BLOCK_SIZE,
507 PAGE_SIZE);
508 flags &= ~VM_MAYREAD;
509 flags |= VM_DONTCOPY | VM_DONTEXPAND;
510 vma->vm_page_prot = pgprot_writecombine(vma->vm_page_prot);
511 mapio = 1;
512 break;
513 case PIO_CRED:
514 if (flags & VM_WRITE) {
515 ret = -EPERM;
516 goto done;
517 }
518 /*
519 * The credit return location for this context could be on the
520 * second or third page allocated for credit returns (if number
521 * of enabled contexts > 64 and 128 respectively).
522 */
523 memaddr = dd->cr_base[uctxt->numa_id].pa +
524 (((u64)uctxt->sc->hw_free -
525 (u64)dd->cr_base[uctxt->numa_id].va) & PAGE_MASK);
526 memlen = PAGE_SIZE;
527 flags &= ~VM_MAYWRITE;
528 flags |= VM_DONTCOPY | VM_DONTEXPAND;
529 /*
530 * The driver has already allocated memory for credit
531 * returns and programmed it into the chip. Has that
532 * memory been flagged as non-cached?
533 */
534 /* vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); */
535 mapio = 1;
536 break;
537 case RCV_HDRQ:
538 memaddr = uctxt->rcvhdrq_phys;
539 memlen = uctxt->rcvhdrq_size;
540 break;
541 case RCV_EGRBUF: {
542 unsigned long addr;
543 int i;
544 /*
545 * The RcvEgr buffer need to be handled differently
546 * as multiple non-contiguous pages need to be mapped
547 * into the user process.
548 */
549 memlen = uctxt->egrbufs.size;
550 if ((vma->vm_end - vma->vm_start) != memlen) {
551 dd_dev_err(dd, "Eager buffer map size invalid (%lu != %lu)\n",
552 (vma->vm_end - vma->vm_start), memlen);
553 ret = -EINVAL;
554 goto done;
555 }
556 if (vma->vm_flags & VM_WRITE) {
557 ret = -EPERM;
558 goto done;
559 }
560 vma->vm_flags &= ~VM_MAYWRITE;
561 addr = vma->vm_start;
562 for (i = 0 ; i < uctxt->egrbufs.numbufs; i++) {
563 ret = remap_pfn_range(
564 vma, addr,
565 uctxt->egrbufs.buffers[i].phys >> PAGE_SHIFT,
566 uctxt->egrbufs.buffers[i].len,
567 vma->vm_page_prot);
568 if (ret < 0)
569 goto done;
570 addr += uctxt->egrbufs.buffers[i].len;
571 }
572 ret = 0;
573 goto done;
574 }
575 case UREGS:
576 /*
577 * Map only the page that contains this context's user
578 * registers.
579 */
580 memaddr = (unsigned long)
581 (dd->physaddr + RXE_PER_CONTEXT_USER)
582 + (uctxt->ctxt * RXE_PER_CONTEXT_SIZE);
583 /*
584 * TidFlow table is on the same page as the rest of the
585 * user registers.
586 */
587 memlen = PAGE_SIZE;
588 flags |= VM_DONTCOPY | VM_DONTEXPAND;
589 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
590 mapio = 1;
591 break;
592 case EVENTS:
593 /*
594 * Use the page where this context's flags are. User level
595 * knows where it's own bitmap is within the page.
596 */
597 memaddr = (unsigned long)(dd->events +
598 ((uctxt->ctxt - dd->first_user_ctxt) *
599 HFI1_MAX_SHARED_CTXTS)) & PAGE_MASK;
600 memlen = PAGE_SIZE;
601 /*
602 * v3.7 removes VM_RESERVED but the effect is kept by
603 * using VM_IO.
604 */
605 flags |= VM_IO | VM_DONTEXPAND;
606 vmf = 1;
607 break;
608 case STATUS:
609 memaddr = kvirt_to_phys((void *)dd->status);
610 memlen = PAGE_SIZE;
611 flags |= VM_IO | VM_DONTEXPAND;
612 break;
613 case RTAIL:
614 if (!HFI1_CAP_IS_USET(DMA_RTAIL)) {
615 /*
616 * If the memory allocation failed, the context alloc
617 * also would have failed, so we would never get here
618 */
619 ret = -EINVAL;
620 goto done;
621 }
622 if (flags & VM_WRITE) {
623 ret = -EPERM;
624 goto done;
625 }
626 memaddr = uctxt->rcvhdrqtailaddr_phys;
627 memlen = PAGE_SIZE;
628 flags &= ~VM_MAYWRITE;
629 break;
630 case SUBCTXT_UREGS:
631 memaddr = (u64)uctxt->subctxt_uregbase;
632 memlen = PAGE_SIZE;
633 flags |= VM_IO | VM_DONTEXPAND;
634 vmf = 1;
635 break;
636 case SUBCTXT_RCV_HDRQ:
637 memaddr = (u64)uctxt->subctxt_rcvhdr_base;
638 memlen = uctxt->rcvhdrq_size * uctxt->subctxt_cnt;
639 flags |= VM_IO | VM_DONTEXPAND;
640 vmf = 1;
641 break;
642 case SUBCTXT_EGRBUF:
643 memaddr = (u64)uctxt->subctxt_rcvegrbuf;
644 memlen = uctxt->egrbufs.size * uctxt->subctxt_cnt;
645 flags |= VM_IO | VM_DONTEXPAND;
646 flags &= ~VM_MAYWRITE;
647 vmf = 1;
648 break;
649 case SDMA_COMP: {
650 struct hfi1_user_sdma_comp_q *cq = fd->cq;
651
652 if (!cq) {
653 ret = -EFAULT;
654 goto done;
655 }
656 memaddr = (u64)cq->comps;
657 memlen = ALIGN(sizeof(*cq->comps) * cq->nentries, PAGE_SIZE);
658 flags |= VM_IO | VM_DONTEXPAND;
659 vmf = 1;
660 break;
661 }
662 default:
663 ret = -EINVAL;
664 break;
665 }
666
667 if ((vma->vm_end - vma->vm_start) != memlen) {
668 hfi1_cdbg(PROC, "%u:%u Memory size mismatch %lu:%lu",
669 uctxt->ctxt, fd->subctxt,
670 (vma->vm_end - vma->vm_start), memlen);
671 ret = -EINVAL;
672 goto done;
673 }
674
675 vma->vm_flags = flags;
676 hfi1_cdbg(PROC,
677 "%u:%u type:%u io/vf:%d/%d, addr:0x%llx, len:%lu(%lu), flags:0x%lx\n",
678 ctxt, subctxt, type, mapio, vmf, memaddr, memlen,
679 vma->vm_end - vma->vm_start, vma->vm_flags);
680 pfn = (unsigned long)(memaddr >> PAGE_SHIFT);
681 if (vmf) {
682 vma->vm_pgoff = pfn;
683 vma->vm_ops = &vm_ops;
684 ret = 0;
685 } else if (mapio) {
686 ret = io_remap_pfn_range(vma, vma->vm_start, pfn, memlen,
687 vma->vm_page_prot);
688 } else {
689 ret = remap_pfn_range(vma, vma->vm_start, pfn, memlen,
690 vma->vm_page_prot);
691 }
692 done:
693 return ret;
694 }
695
696 /*
697 * Local (non-chip) user memory is not mapped right away but as it is
698 * accessed by the user-level code.
699 */
700 static int vma_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
701 {
702 struct page *page;
703
704 page = vmalloc_to_page((void *)(vmf->pgoff << PAGE_SHIFT));
705 if (!page)
706 return VM_FAULT_SIGBUS;
707
708 get_page(page);
709 vmf->page = page;
710
711 return 0;
712 }
713
714 static unsigned int hfi1_poll(struct file *fp, struct poll_table_struct *pt)
715 {
716 struct hfi1_ctxtdata *uctxt;
717 unsigned pollflag;
718
719 uctxt = ((struct hfi1_filedata *)fp->private_data)->uctxt;
720 if (!uctxt)
721 pollflag = POLLERR;
722 else if (uctxt->poll_type == HFI1_POLL_TYPE_URGENT)
723 pollflag = poll_urgent(fp, pt);
724 else if (uctxt->poll_type == HFI1_POLL_TYPE_ANYRCV)
725 pollflag = poll_next(fp, pt);
726 else /* invalid */
727 pollflag = POLLERR;
728
729 return pollflag;
730 }
731
732 static int hfi1_file_close(struct inode *inode, struct file *fp)
733 {
734 struct hfi1_filedata *fdata = fp->private_data;
735 struct hfi1_ctxtdata *uctxt = fdata->uctxt;
736 struct hfi1_devdata *dd;
737 unsigned long flags, *ev;
738
739 fp->private_data = NULL;
740
741 if (!uctxt)
742 goto done;
743
744 hfi1_cdbg(PROC, "freeing ctxt %u:%u", uctxt->ctxt, fdata->subctxt);
745 dd = uctxt->dd;
746 mutex_lock(&hfi1_mutex);
747
748 flush_wc();
749 /* drain user sdma queue */
750 hfi1_user_sdma_free_queues(fdata);
751
752 /* release the cpu */
753 hfi1_put_proc_affinity(dd, fdata->rec_cpu_num);
754
755 /*
756 * Clear any left over, unhandled events so the next process that
757 * gets this context doesn't get confused.
758 */
759 ev = dd->events + ((uctxt->ctxt - dd->first_user_ctxt) *
760 HFI1_MAX_SHARED_CTXTS) + fdata->subctxt;
761 *ev = 0;
762
763 if (--uctxt->cnt) {
764 uctxt->active_slaves &= ~(1 << fdata->subctxt);
765 uctxt->subpid[fdata->subctxt] = 0;
766 mutex_unlock(&hfi1_mutex);
767 goto done;
768 }
769
770 spin_lock_irqsave(&dd->uctxt_lock, flags);
771 /*
772 * Disable receive context and interrupt available, reset all
773 * RcvCtxtCtrl bits to default values.
774 */
775 hfi1_rcvctrl(dd, HFI1_RCVCTRL_CTXT_DIS |
776 HFI1_RCVCTRL_TIDFLOW_DIS |
777 HFI1_RCVCTRL_INTRAVAIL_DIS |
778 HFI1_RCVCTRL_TAILUPD_DIS |
779 HFI1_RCVCTRL_ONE_PKT_EGR_DIS |
780 HFI1_RCVCTRL_NO_RHQ_DROP_DIS |
781 HFI1_RCVCTRL_NO_EGR_DROP_DIS, uctxt->ctxt);
782 /* Clear the context's J_KEY */
783 hfi1_clear_ctxt_jkey(dd, uctxt->ctxt);
784 /*
785 * Reset context integrity checks to default.
786 * (writes to CSRs probably belong in chip.c)
787 */
788 write_kctxt_csr(dd, uctxt->sc->hw_context, SEND_CTXT_CHECK_ENABLE,
789 hfi1_pkt_default_send_ctxt_mask(dd, uctxt->sc->type));
790 sc_disable(uctxt->sc);
791 uctxt->pid = 0;
792 spin_unlock_irqrestore(&dd->uctxt_lock, flags);
793
794 dd->rcd[uctxt->ctxt] = NULL;
795 uctxt->rcvwait_to = 0;
796 uctxt->piowait_to = 0;
797 uctxt->rcvnowait = 0;
798 uctxt->pionowait = 0;
799 uctxt->event_flags = 0;
800
801 hfi1_user_exp_rcv_free(fdata);
802 hfi1_clear_ctxt_pkey(dd, uctxt->ctxt);
803
804 hfi1_stats.sps_ctxts--;
805 if (++dd->freectxts == dd->num_user_contexts)
806 aspm_enable_all(dd);
807 mutex_unlock(&hfi1_mutex);
808 hfi1_free_ctxtdata(dd, uctxt);
809 done:
810 kfree(fdata);
811 return 0;
812 }
813
814 /*
815 * Convert kernel *virtual* addresses to physical addresses.
816 * This is used to vmalloc'ed addresses.
817 */
818 static u64 kvirt_to_phys(void *addr)
819 {
820 struct page *page;
821 u64 paddr = 0;
822
823 page = vmalloc_to_page(addr);
824 if (page)
825 paddr = page_to_pfn(page) << PAGE_SHIFT;
826
827 return paddr;
828 }
829
830 static int assign_ctxt(struct file *fp, struct hfi1_user_info *uinfo)
831 {
832 int i_minor, ret = 0;
833 unsigned swmajor, swminor, alg = HFI1_ALG_ACROSS;
834
835 swmajor = uinfo->userversion >> 16;
836 if (swmajor != HFI1_USER_SWMAJOR) {
837 ret = -ENODEV;
838 goto done;
839 }
840
841 swminor = uinfo->userversion & 0xffff;
842
843 if (uinfo->hfi1_alg < HFI1_ALG_COUNT)
844 alg = uinfo->hfi1_alg;
845
846 mutex_lock(&hfi1_mutex);
847 /* First, lets check if we need to setup a shared context? */
848 if (uinfo->subctxt_cnt) {
849 struct hfi1_filedata *fd = fp->private_data;
850
851 ret = find_shared_ctxt(fp, uinfo);
852 if (ret < 0)
853 goto done_unlock;
854 if (ret)
855 fd->rec_cpu_num = hfi1_get_proc_affinity(
856 fd->uctxt->dd, fd->uctxt->numa_id);
857 }
858
859 /*
860 * We execute the following block if we couldn't find a
861 * shared context or if context sharing is not required.
862 */
863 if (!ret) {
864 i_minor = iminor(file_inode(fp)) - HFI1_USER_MINOR_BASE;
865 ret = get_user_context(fp, uinfo, i_minor - 1, alg);
866 }
867 done_unlock:
868 mutex_unlock(&hfi1_mutex);
869 done:
870 return ret;
871 }
872
873 /* return true if the device available for general use */
874 static int usable_device(struct hfi1_devdata *dd)
875 {
876 struct hfi1_pportdata *ppd = dd->pport;
877
878 return driver_lstate(ppd) == IB_PORT_ACTIVE;
879 }
880
881 static int get_user_context(struct file *fp, struct hfi1_user_info *uinfo,
882 int devno, unsigned alg)
883 {
884 struct hfi1_devdata *dd = NULL;
885 int ret = 0, devmax, npresent, nup, dev;
886
887 devmax = hfi1_count_units(&npresent, &nup);
888 if (!npresent) {
889 ret = -ENXIO;
890 goto done;
891 }
892 if (!nup) {
893 ret = -ENETDOWN;
894 goto done;
895 }
896 if (devno >= 0) {
897 dd = hfi1_lookup(devno);
898 if (!dd)
899 ret = -ENODEV;
900 else if (!dd->freectxts)
901 ret = -EBUSY;
902 } else {
903 struct hfi1_devdata *pdd;
904
905 if (alg == HFI1_ALG_ACROSS) {
906 unsigned free = 0U;
907
908 for (dev = 0; dev < devmax; dev++) {
909 pdd = hfi1_lookup(dev);
910 if (!pdd)
911 continue;
912 if (!usable_device(pdd))
913 continue;
914 if (pdd->freectxts &&
915 pdd->freectxts > free) {
916 dd = pdd;
917 free = pdd->freectxts;
918 }
919 }
920 } else {
921 for (dev = 0; dev < devmax; dev++) {
922 pdd = hfi1_lookup(dev);
923 if (!pdd)
924 continue;
925 if (!usable_device(pdd))
926 continue;
927 if (pdd->freectxts) {
928 dd = pdd;
929 break;
930 }
931 }
932 }
933 if (!dd)
934 ret = -EBUSY;
935 }
936 done:
937 return ret ? ret : allocate_ctxt(fp, dd, uinfo);
938 }
939
940 static int find_shared_ctxt(struct file *fp,
941 const struct hfi1_user_info *uinfo)
942 {
943 int devmax, ndev, i;
944 int ret = 0;
945 struct hfi1_filedata *fd = fp->private_data;
946
947 devmax = hfi1_count_units(NULL, NULL);
948
949 for (ndev = 0; ndev < devmax; ndev++) {
950 struct hfi1_devdata *dd = hfi1_lookup(ndev);
951
952 if (!(dd && (dd->flags & HFI1_PRESENT) && dd->kregbase))
953 continue;
954 for (i = dd->first_user_ctxt; i < dd->num_rcv_contexts; i++) {
955 struct hfi1_ctxtdata *uctxt = dd->rcd[i];
956
957 /* Skip ctxts which are not yet open */
958 if (!uctxt || !uctxt->cnt)
959 continue;
960 /* Skip ctxt if it doesn't match the requested one */
961 if (memcmp(uctxt->uuid, uinfo->uuid,
962 sizeof(uctxt->uuid)) ||
963 uctxt->jkey != generate_jkey(current_uid()) ||
964 uctxt->subctxt_id != uinfo->subctxt_id ||
965 uctxt->subctxt_cnt != uinfo->subctxt_cnt)
966 continue;
967
968 /* Verify the sharing process matches the master */
969 if (uctxt->userversion != uinfo->userversion ||
970 uctxt->cnt >= uctxt->subctxt_cnt) {
971 ret = -EINVAL;
972 goto done;
973 }
974 fd->uctxt = uctxt;
975 fd->subctxt = uctxt->cnt++;
976 uctxt->subpid[fd->subctxt] = current->pid;
977 uctxt->active_slaves |= 1 << fd->subctxt;
978 ret = 1;
979 goto done;
980 }
981 }
982
983 done:
984 return ret;
985 }
986
987 static int allocate_ctxt(struct file *fp, struct hfi1_devdata *dd,
988 struct hfi1_user_info *uinfo)
989 {
990 struct hfi1_filedata *fd = fp->private_data;
991 struct hfi1_ctxtdata *uctxt;
992 unsigned ctxt;
993 int ret, numa;
994
995 if (dd->flags & HFI1_FROZEN) {
996 /*
997 * Pick an error that is unique from all other errors
998 * that are returned so the user process knows that
999 * it tried to allocate while the SPC was frozen. It
1000 * it should be able to retry with success in a short
1001 * while.
1002 */
1003 return -EIO;
1004 }
1005
1006 for (ctxt = dd->first_user_ctxt; ctxt < dd->num_rcv_contexts; ctxt++)
1007 if (!dd->rcd[ctxt])
1008 break;
1009
1010 if (ctxt == dd->num_rcv_contexts)
1011 return -EBUSY;
1012
1013 fd->rec_cpu_num = hfi1_get_proc_affinity(dd, -1);
1014 if (fd->rec_cpu_num != -1)
1015 numa = cpu_to_node(fd->rec_cpu_num);
1016 else
1017 numa = numa_node_id();
1018 uctxt = hfi1_create_ctxtdata(dd->pport, ctxt, numa);
1019 if (!uctxt) {
1020 dd_dev_err(dd,
1021 "Unable to allocate ctxtdata memory, failing open\n");
1022 return -ENOMEM;
1023 }
1024 hfi1_cdbg(PROC, "[%u:%u] pid %u assigned to CPU %d (NUMA %u)",
1025 uctxt->ctxt, fd->subctxt, current->pid, fd->rec_cpu_num,
1026 uctxt->numa_id);
1027
1028 /*
1029 * Allocate and enable a PIO send context.
1030 */
1031 uctxt->sc = sc_alloc(dd, SC_USER, uctxt->rcvhdrqentsize,
1032 uctxt->dd->node);
1033 if (!uctxt->sc)
1034 return -ENOMEM;
1035
1036 hfi1_cdbg(PROC, "allocated send context %u(%u)\n", uctxt->sc->sw_index,
1037 uctxt->sc->hw_context);
1038 ret = sc_enable(uctxt->sc);
1039 if (ret)
1040 return ret;
1041 /*
1042 * Setup shared context resources if the user-level has requested
1043 * shared contexts and this is the 'master' process.
1044 * This has to be done here so the rest of the sub-contexts find the
1045 * proper master.
1046 */
1047 if (uinfo->subctxt_cnt && !fd->subctxt) {
1048 ret = init_subctxts(uctxt, uinfo);
1049 /*
1050 * On error, we don't need to disable and de-allocate the
1051 * send context because it will be done during file close
1052 */
1053 if (ret)
1054 return ret;
1055 }
1056 uctxt->userversion = uinfo->userversion;
1057 uctxt->pid = current->pid;
1058 uctxt->flags = HFI1_CAP_UGET(MASK);
1059 init_waitqueue_head(&uctxt->wait);
1060 strlcpy(uctxt->comm, current->comm, sizeof(uctxt->comm));
1061 memcpy(uctxt->uuid, uinfo->uuid, sizeof(uctxt->uuid));
1062 uctxt->jkey = generate_jkey(current_uid());
1063 INIT_LIST_HEAD(&uctxt->sdma_queues);
1064 spin_lock_init(&uctxt->sdma_qlock);
1065 hfi1_stats.sps_ctxts++;
1066 /*
1067 * Disable ASPM when there are open user/PSM contexts to avoid
1068 * issues with ASPM L1 exit latency
1069 */
1070 if (dd->freectxts-- == dd->num_user_contexts)
1071 aspm_disable_all(dd);
1072 fd->uctxt = uctxt;
1073
1074 return 0;
1075 }
1076
1077 static int init_subctxts(struct hfi1_ctxtdata *uctxt,
1078 const struct hfi1_user_info *uinfo)
1079 {
1080 unsigned num_subctxts;
1081
1082 num_subctxts = uinfo->subctxt_cnt;
1083 if (num_subctxts > HFI1_MAX_SHARED_CTXTS)
1084 return -EINVAL;
1085
1086 uctxt->subctxt_cnt = uinfo->subctxt_cnt;
1087 uctxt->subctxt_id = uinfo->subctxt_id;
1088 uctxt->active_slaves = 1;
1089 uctxt->redirect_seq_cnt = 1;
1090 set_bit(HFI1_CTXT_MASTER_UNINIT, &uctxt->event_flags);
1091
1092 return 0;
1093 }
1094
1095 static int setup_subctxt(struct hfi1_ctxtdata *uctxt)
1096 {
1097 int ret = 0;
1098 unsigned num_subctxts = uctxt->subctxt_cnt;
1099
1100 uctxt->subctxt_uregbase = vmalloc_user(PAGE_SIZE);
1101 if (!uctxt->subctxt_uregbase) {
1102 ret = -ENOMEM;
1103 goto bail;
1104 }
1105 /* We can take the size of the RcvHdr Queue from the master */
1106 uctxt->subctxt_rcvhdr_base = vmalloc_user(uctxt->rcvhdrq_size *
1107 num_subctxts);
1108 if (!uctxt->subctxt_rcvhdr_base) {
1109 ret = -ENOMEM;
1110 goto bail_ureg;
1111 }
1112
1113 uctxt->subctxt_rcvegrbuf = vmalloc_user(uctxt->egrbufs.size *
1114 num_subctxts);
1115 if (!uctxt->subctxt_rcvegrbuf) {
1116 ret = -ENOMEM;
1117 goto bail_rhdr;
1118 }
1119 goto bail;
1120 bail_rhdr:
1121 vfree(uctxt->subctxt_rcvhdr_base);
1122 bail_ureg:
1123 vfree(uctxt->subctxt_uregbase);
1124 uctxt->subctxt_uregbase = NULL;
1125 bail:
1126 return ret;
1127 }
1128
1129 static int user_init(struct file *fp)
1130 {
1131 int ret;
1132 unsigned int rcvctrl_ops = 0;
1133 struct hfi1_filedata *fd = fp->private_data;
1134 struct hfi1_ctxtdata *uctxt = fd->uctxt;
1135
1136 /* make sure that the context has already been setup */
1137 if (!test_bit(HFI1_CTXT_SETUP_DONE, &uctxt->event_flags)) {
1138 ret = -EFAULT;
1139 goto done;
1140 }
1141
1142 /*
1143 * Subctxts don't need to initialize anything since master
1144 * has done it.
1145 */
1146 if (fd->subctxt) {
1147 ret = wait_event_interruptible(uctxt->wait, !test_bit(
1148 HFI1_CTXT_MASTER_UNINIT,
1149 &uctxt->event_flags));
1150 goto expected;
1151 }
1152
1153 /* initialize poll variables... */
1154 uctxt->urgent = 0;
1155 uctxt->urgent_poll = 0;
1156
1157 /*
1158 * Now enable the ctxt for receive.
1159 * For chips that are set to DMA the tail register to memory
1160 * when they change (and when the update bit transitions from
1161 * 0 to 1. So for those chips, we turn it off and then back on.
1162 * This will (very briefly) affect any other open ctxts, but the
1163 * duration is very short, and therefore isn't an issue. We
1164 * explicitly set the in-memory tail copy to 0 beforehand, so we
1165 * don't have to wait to be sure the DMA update has happened
1166 * (chip resets head/tail to 0 on transition to enable).
1167 */
1168 if (uctxt->rcvhdrtail_kvaddr)
1169 clear_rcvhdrtail(uctxt);
1170
1171 /* Setup J_KEY before enabling the context */
1172 hfi1_set_ctxt_jkey(uctxt->dd, uctxt->ctxt, uctxt->jkey);
1173
1174 rcvctrl_ops = HFI1_RCVCTRL_CTXT_ENB;
1175 if (HFI1_CAP_KGET_MASK(uctxt->flags, HDRSUPP))
1176 rcvctrl_ops |= HFI1_RCVCTRL_TIDFLOW_ENB;
1177 /*
1178 * Ignore the bit in the flags for now until proper
1179 * support for multiple packet per rcv array entry is
1180 * added.
1181 */
1182 if (!HFI1_CAP_KGET_MASK(uctxt->flags, MULTI_PKT_EGR))
1183 rcvctrl_ops |= HFI1_RCVCTRL_ONE_PKT_EGR_ENB;
1184 if (HFI1_CAP_KGET_MASK(uctxt->flags, NODROP_EGR_FULL))
1185 rcvctrl_ops |= HFI1_RCVCTRL_NO_EGR_DROP_ENB;
1186 if (HFI1_CAP_KGET_MASK(uctxt->flags, NODROP_RHQ_FULL))
1187 rcvctrl_ops |= HFI1_RCVCTRL_NO_RHQ_DROP_ENB;
1188 /*
1189 * The RcvCtxtCtrl.TailUpd bit has to be explicitly written.
1190 * We can't rely on the correct value to be set from prior
1191 * uses of the chip or ctxt. Therefore, add the rcvctrl op
1192 * for both cases.
1193 */
1194 if (HFI1_CAP_KGET_MASK(uctxt->flags, DMA_RTAIL))
1195 rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_ENB;
1196 else
1197 rcvctrl_ops |= HFI1_RCVCTRL_TAILUPD_DIS;
1198 hfi1_rcvctrl(uctxt->dd, rcvctrl_ops, uctxt->ctxt);
1199
1200 /* Notify any waiting slaves */
1201 if (uctxt->subctxt_cnt) {
1202 clear_bit(HFI1_CTXT_MASTER_UNINIT, &uctxt->event_flags);
1203 wake_up(&uctxt->wait);
1204 }
1205
1206 expected:
1207 /*
1208 * Expected receive has to be setup for all processes (including
1209 * shared contexts). However, it has to be done after the master
1210 * context has been fully configured as it depends on the
1211 * eager/expected split of the RcvArray entries.
1212 * Setting it up here ensures that the subcontexts will be waiting
1213 * (due to the above wait_event_interruptible() until the master
1214 * is setup.
1215 */
1216 ret = hfi1_user_exp_rcv_init(fp);
1217 done:
1218 return ret;
1219 }
1220
1221 static int get_ctxt_info(struct file *fp, void __user *ubase, __u32 len)
1222 {
1223 struct hfi1_ctxt_info cinfo;
1224 struct hfi1_filedata *fd = fp->private_data;
1225 struct hfi1_ctxtdata *uctxt = fd->uctxt;
1226 int ret = 0;
1227
1228 memset(&cinfo, 0, sizeof(cinfo));
1229 ret = hfi1_get_base_kinfo(uctxt, &cinfo);
1230 if (ret < 0)
1231 goto done;
1232 cinfo.num_active = hfi1_count_active_units();
1233 cinfo.unit = uctxt->dd->unit;
1234 cinfo.ctxt = uctxt->ctxt;
1235 cinfo.subctxt = fd->subctxt;
1236 cinfo.rcvtids = roundup(uctxt->egrbufs.alloced,
1237 uctxt->dd->rcv_entries.group_size) +
1238 uctxt->expected_count;
1239 cinfo.credits = uctxt->sc->credits;
1240 cinfo.numa_node = uctxt->numa_id;
1241 cinfo.rec_cpu = fd->rec_cpu_num;
1242 cinfo.send_ctxt = uctxt->sc->hw_context;
1243
1244 cinfo.egrtids = uctxt->egrbufs.alloced;
1245 cinfo.rcvhdrq_cnt = uctxt->rcvhdrq_cnt;
1246 cinfo.rcvhdrq_entsize = uctxt->rcvhdrqentsize << 2;
1247 cinfo.sdma_ring_size = fd->cq->nentries;
1248 cinfo.rcvegr_size = uctxt->egrbufs.rcvtid_size;
1249
1250 trace_hfi1_ctxt_info(uctxt->dd, uctxt->ctxt, fd->subctxt, cinfo);
1251 if (copy_to_user(ubase, &cinfo, sizeof(cinfo)))
1252 ret = -EFAULT;
1253 done:
1254 return ret;
1255 }
1256
1257 static int setup_ctxt(struct file *fp)
1258 {
1259 struct hfi1_filedata *fd = fp->private_data;
1260 struct hfi1_ctxtdata *uctxt = fd->uctxt;
1261 struct hfi1_devdata *dd = uctxt->dd;
1262 int ret = 0;
1263
1264 /*
1265 * Context should be set up only once (including allocation and
1266 * programming of eager buffers. This is done if context sharing
1267 * is not requested or by the master process.
1268 */
1269 if (!uctxt->subctxt_cnt || !fd->subctxt) {
1270 ret = hfi1_init_ctxt(uctxt->sc);
1271 if (ret)
1272 goto done;
1273
1274 /* Now allocate the RcvHdr queue and eager buffers. */
1275 ret = hfi1_create_rcvhdrq(dd, uctxt);
1276 if (ret)
1277 goto done;
1278 ret = hfi1_setup_eagerbufs(uctxt);
1279 if (ret)
1280 goto done;
1281 if (uctxt->subctxt_cnt && !fd->subctxt) {
1282 ret = setup_subctxt(uctxt);
1283 if (ret)
1284 goto done;
1285 }
1286 }
1287 ret = hfi1_user_sdma_alloc_queues(uctxt, fp);
1288 if (ret)
1289 goto done;
1290
1291 set_bit(HFI1_CTXT_SETUP_DONE, &uctxt->event_flags);
1292 done:
1293 return ret;
1294 }
1295
1296 static int get_base_info(struct file *fp, void __user *ubase, __u32 len)
1297 {
1298 struct hfi1_base_info binfo;
1299 struct hfi1_filedata *fd = fp->private_data;
1300 struct hfi1_ctxtdata *uctxt = fd->uctxt;
1301 struct hfi1_devdata *dd = uctxt->dd;
1302 ssize_t sz;
1303 unsigned offset;
1304 int ret = 0;
1305
1306 trace_hfi1_uctxtdata(uctxt->dd, uctxt);
1307
1308 memset(&binfo, 0, sizeof(binfo));
1309 binfo.hw_version = dd->revision;
1310 binfo.sw_version = HFI1_KERN_SWVERSION;
1311 binfo.bthqp = kdeth_qp;
1312 binfo.jkey = uctxt->jkey;
1313 /*
1314 * If more than 64 contexts are enabled the allocated credit
1315 * return will span two or three contiguous pages. Since we only
1316 * map the page containing the context's credit return address,
1317 * we need to calculate the offset in the proper page.
1318 */
1319 offset = ((u64)uctxt->sc->hw_free -
1320 (u64)dd->cr_base[uctxt->numa_id].va) % PAGE_SIZE;
1321 binfo.sc_credits_addr = HFI1_MMAP_TOKEN(PIO_CRED, uctxt->ctxt,
1322 fd->subctxt, offset);
1323 binfo.pio_bufbase = HFI1_MMAP_TOKEN(PIO_BUFS, uctxt->ctxt,
1324 fd->subctxt,
1325 uctxt->sc->base_addr);
1326 binfo.pio_bufbase_sop = HFI1_MMAP_TOKEN(PIO_BUFS_SOP,
1327 uctxt->ctxt,
1328 fd->subctxt,
1329 uctxt->sc->base_addr);
1330 binfo.rcvhdr_bufbase = HFI1_MMAP_TOKEN(RCV_HDRQ, uctxt->ctxt,
1331 fd->subctxt,
1332 uctxt->rcvhdrq);
1333 binfo.rcvegr_bufbase = HFI1_MMAP_TOKEN(RCV_EGRBUF, uctxt->ctxt,
1334 fd->subctxt,
1335 uctxt->egrbufs.rcvtids[0].phys);
1336 binfo.sdma_comp_bufbase = HFI1_MMAP_TOKEN(SDMA_COMP, uctxt->ctxt,
1337 fd->subctxt, 0);
1338 /*
1339 * user regs are at
1340 * (RXE_PER_CONTEXT_USER + (ctxt * RXE_PER_CONTEXT_SIZE))
1341 */
1342 binfo.user_regbase = HFI1_MMAP_TOKEN(UREGS, uctxt->ctxt,
1343 fd->subctxt, 0);
1344 offset = offset_in_page((((uctxt->ctxt - dd->first_user_ctxt) *
1345 HFI1_MAX_SHARED_CTXTS) + fd->subctxt) *
1346 sizeof(*dd->events));
1347 binfo.events_bufbase = HFI1_MMAP_TOKEN(EVENTS, uctxt->ctxt,
1348 fd->subctxt,
1349 offset);
1350 binfo.status_bufbase = HFI1_MMAP_TOKEN(STATUS, uctxt->ctxt,
1351 fd->subctxt,
1352 dd->status);
1353 if (HFI1_CAP_IS_USET(DMA_RTAIL))
1354 binfo.rcvhdrtail_base = HFI1_MMAP_TOKEN(RTAIL, uctxt->ctxt,
1355 fd->subctxt, 0);
1356 if (uctxt->subctxt_cnt) {
1357 binfo.subctxt_uregbase = HFI1_MMAP_TOKEN(SUBCTXT_UREGS,
1358 uctxt->ctxt,
1359 fd->subctxt, 0);
1360 binfo.subctxt_rcvhdrbuf = HFI1_MMAP_TOKEN(SUBCTXT_RCV_HDRQ,
1361 uctxt->ctxt,
1362 fd->subctxt, 0);
1363 binfo.subctxt_rcvegrbuf = HFI1_MMAP_TOKEN(SUBCTXT_EGRBUF,
1364 uctxt->ctxt,
1365 fd->subctxt, 0);
1366 }
1367 sz = (len < sizeof(binfo)) ? len : sizeof(binfo);
1368 if (copy_to_user(ubase, &binfo, sz))
1369 ret = -EFAULT;
1370 return ret;
1371 }
1372
1373 static unsigned int poll_urgent(struct file *fp,
1374 struct poll_table_struct *pt)
1375 {
1376 struct hfi1_filedata *fd = fp->private_data;
1377 struct hfi1_ctxtdata *uctxt = fd->uctxt;
1378 struct hfi1_devdata *dd = uctxt->dd;
1379 unsigned pollflag;
1380
1381 poll_wait(fp, &uctxt->wait, pt);
1382
1383 spin_lock_irq(&dd->uctxt_lock);
1384 if (uctxt->urgent != uctxt->urgent_poll) {
1385 pollflag = POLLIN | POLLRDNORM;
1386 uctxt->urgent_poll = uctxt->urgent;
1387 } else {
1388 pollflag = 0;
1389 set_bit(HFI1_CTXT_WAITING_URG, &uctxt->event_flags);
1390 }
1391 spin_unlock_irq(&dd->uctxt_lock);
1392
1393 return pollflag;
1394 }
1395
1396 static unsigned int poll_next(struct file *fp,
1397 struct poll_table_struct *pt)
1398 {
1399 struct hfi1_filedata *fd = fp->private_data;
1400 struct hfi1_ctxtdata *uctxt = fd->uctxt;
1401 struct hfi1_devdata *dd = uctxt->dd;
1402 unsigned pollflag;
1403
1404 poll_wait(fp, &uctxt->wait, pt);
1405
1406 spin_lock_irq(&dd->uctxt_lock);
1407 if (hdrqempty(uctxt)) {
1408 set_bit(HFI1_CTXT_WAITING_RCV, &uctxt->event_flags);
1409 hfi1_rcvctrl(dd, HFI1_RCVCTRL_INTRAVAIL_ENB, uctxt->ctxt);
1410 pollflag = 0;
1411 } else {
1412 pollflag = POLLIN | POLLRDNORM;
1413 }
1414 spin_unlock_irq(&dd->uctxt_lock);
1415
1416 return pollflag;
1417 }
1418
1419 /*
1420 * Find all user contexts in use, and set the specified bit in their
1421 * event mask.
1422 * See also find_ctxt() for a similar use, that is specific to send buffers.
1423 */
1424 int hfi1_set_uevent_bits(struct hfi1_pportdata *ppd, const int evtbit)
1425 {
1426 struct hfi1_ctxtdata *uctxt;
1427 struct hfi1_devdata *dd = ppd->dd;
1428 unsigned ctxt;
1429 int ret = 0;
1430 unsigned long flags;
1431
1432 if (!dd->events) {
1433 ret = -EINVAL;
1434 goto done;
1435 }
1436
1437 spin_lock_irqsave(&dd->uctxt_lock, flags);
1438 for (ctxt = dd->first_user_ctxt; ctxt < dd->num_rcv_contexts;
1439 ctxt++) {
1440 uctxt = dd->rcd[ctxt];
1441 if (uctxt) {
1442 unsigned long *evs = dd->events +
1443 (uctxt->ctxt - dd->first_user_ctxt) *
1444 HFI1_MAX_SHARED_CTXTS;
1445 int i;
1446 /*
1447 * subctxt_cnt is 0 if not shared, so do base
1448 * separately, first, then remaining subctxt, if any
1449 */
1450 set_bit(evtbit, evs);
1451 for (i = 1; i < uctxt->subctxt_cnt; i++)
1452 set_bit(evtbit, evs + i);
1453 }
1454 }
1455 spin_unlock_irqrestore(&dd->uctxt_lock, flags);
1456 done:
1457 return ret;
1458 }
1459
1460 /**
1461 * manage_rcvq - manage a context's receive queue
1462 * @uctxt: the context
1463 * @subctxt: the sub-context
1464 * @start_stop: action to carry out
1465 *
1466 * start_stop == 0 disables receive on the context, for use in queue
1467 * overflow conditions. start_stop==1 re-enables, to be used to
1468 * re-init the software copy of the head register
1469 */
1470 static int manage_rcvq(struct hfi1_ctxtdata *uctxt, unsigned subctxt,
1471 int start_stop)
1472 {
1473 struct hfi1_devdata *dd = uctxt->dd;
1474 unsigned int rcvctrl_op;
1475
1476 if (subctxt)
1477 goto bail;
1478 /* atomically clear receive enable ctxt. */
1479 if (start_stop) {
1480 /*
1481 * On enable, force in-memory copy of the tail register to
1482 * 0, so that protocol code doesn't have to worry about
1483 * whether or not the chip has yet updated the in-memory
1484 * copy or not on return from the system call. The chip
1485 * always resets it's tail register back to 0 on a
1486 * transition from disabled to enabled.
1487 */
1488 if (uctxt->rcvhdrtail_kvaddr)
1489 clear_rcvhdrtail(uctxt);
1490 rcvctrl_op = HFI1_RCVCTRL_CTXT_ENB;
1491 } else {
1492 rcvctrl_op = HFI1_RCVCTRL_CTXT_DIS;
1493 }
1494 hfi1_rcvctrl(dd, rcvctrl_op, uctxt->ctxt);
1495 /* always; new head should be equal to new tail; see above */
1496 bail:
1497 return 0;
1498 }
1499
1500 /*
1501 * clear the event notifier events for this context.
1502 * User process then performs actions appropriate to bit having been
1503 * set, if desired, and checks again in future.
1504 */
1505 static int user_event_ack(struct hfi1_ctxtdata *uctxt, int subctxt,
1506 unsigned long events)
1507 {
1508 int i;
1509 struct hfi1_devdata *dd = uctxt->dd;
1510 unsigned long *evs;
1511
1512 if (!dd->events)
1513 return 0;
1514
1515 evs = dd->events + ((uctxt->ctxt - dd->first_user_ctxt) *
1516 HFI1_MAX_SHARED_CTXTS) + subctxt;
1517
1518 for (i = 0; i <= _HFI1_MAX_EVENT_BIT; i++) {
1519 if (!test_bit(i, &events))
1520 continue;
1521 clear_bit(i, evs);
1522 }
1523 return 0;
1524 }
1525
1526 static int set_ctxt_pkey(struct hfi1_ctxtdata *uctxt, unsigned subctxt,
1527 u16 pkey)
1528 {
1529 int ret = -ENOENT, i, intable = 0;
1530 struct hfi1_pportdata *ppd = uctxt->ppd;
1531 struct hfi1_devdata *dd = uctxt->dd;
1532
1533 if (pkey == LIM_MGMT_P_KEY || pkey == FULL_MGMT_P_KEY) {
1534 ret = -EINVAL;
1535 goto done;
1536 }
1537
1538 for (i = 0; i < ARRAY_SIZE(ppd->pkeys); i++)
1539 if (pkey == ppd->pkeys[i]) {
1540 intable = 1;
1541 break;
1542 }
1543
1544 if (intable)
1545 ret = hfi1_set_ctxt_pkey(dd, uctxt->ctxt, pkey);
1546 done:
1547 return ret;
1548 }
1549
1550 static int ui_open(struct inode *inode, struct file *filp)
1551 {
1552 struct hfi1_devdata *dd;
1553
1554 dd = container_of(inode->i_cdev, struct hfi1_devdata, ui_cdev);
1555 filp->private_data = dd; /* for other methods */
1556 return 0;
1557 }
1558
1559 static int ui_release(struct inode *inode, struct file *filp)
1560 {
1561 /* nothing to do */
1562 return 0;
1563 }
1564
1565 static loff_t ui_lseek(struct file *filp, loff_t offset, int whence)
1566 {
1567 struct hfi1_devdata *dd = filp->private_data;
1568
1569 switch (whence) {
1570 case SEEK_SET:
1571 break;
1572 case SEEK_CUR:
1573 offset += filp->f_pos;
1574 break;
1575 case SEEK_END:
1576 offset = ((dd->kregend - dd->kregbase) + DC8051_DATA_MEM_SIZE) -
1577 offset;
1578 break;
1579 default:
1580 return -EINVAL;
1581 }
1582
1583 if (offset < 0)
1584 return -EINVAL;
1585
1586 if (offset >= (dd->kregend - dd->kregbase) + DC8051_DATA_MEM_SIZE)
1587 return -EINVAL;
1588
1589 filp->f_pos = offset;
1590
1591 return filp->f_pos;
1592 }
1593
1594 /* NOTE: assumes unsigned long is 8 bytes */
1595 static ssize_t ui_read(struct file *filp, char __user *buf, size_t count,
1596 loff_t *f_pos)
1597 {
1598 struct hfi1_devdata *dd = filp->private_data;
1599 void __iomem *base = dd->kregbase;
1600 unsigned long total, csr_off,
1601 barlen = (dd->kregend - dd->kregbase);
1602 u64 data;
1603
1604 /* only read 8 byte quantities */
1605 if ((count % 8) != 0)
1606 return -EINVAL;
1607 /* offset must be 8-byte aligned */
1608 if ((*f_pos % 8) != 0)
1609 return -EINVAL;
1610 /* destination buffer must be 8-byte aligned */
1611 if ((unsigned long)buf % 8 != 0)
1612 return -EINVAL;
1613 /* must be in range */
1614 if (*f_pos + count > (barlen + DC8051_DATA_MEM_SIZE))
1615 return -EINVAL;
1616 /* only set the base if we are not starting past the BAR */
1617 if (*f_pos < barlen)
1618 base += *f_pos;
1619 csr_off = *f_pos;
1620 for (total = 0; total < count; total += 8, csr_off += 8) {
1621 /* accessing LCB CSRs requires more checks */
1622 if (is_lcb_offset(csr_off)) {
1623 if (read_lcb_csr(dd, csr_off, (u64 *)&data))
1624 break; /* failed */
1625 }
1626 /*
1627 * Cannot read ASIC GPIO/QSFP* clear and force CSRs without a
1628 * false parity error. Avoid the whole issue by not reading
1629 * them. These registers are defined as having a read value
1630 * of 0.
1631 */
1632 else if (csr_off == ASIC_GPIO_CLEAR ||
1633 csr_off == ASIC_GPIO_FORCE ||
1634 csr_off == ASIC_QSFP1_CLEAR ||
1635 csr_off == ASIC_QSFP1_FORCE ||
1636 csr_off == ASIC_QSFP2_CLEAR ||
1637 csr_off == ASIC_QSFP2_FORCE)
1638 data = 0;
1639 else if (csr_off >= barlen) {
1640 /*
1641 * read_8051_data can read more than just 8 bytes at
1642 * a time. However, folding this into the loop and
1643 * handling the reads in 8 byte increments allows us
1644 * to smoothly transition from chip memory to 8051
1645 * memory.
1646 */
1647 if (read_8051_data(dd,
1648 (u32)(csr_off - barlen),
1649 sizeof(data), &data))
1650 break; /* failed */
1651 } else
1652 data = readq(base + total);
1653 if (put_user(data, (unsigned long __user *)(buf + total)))
1654 break;
1655 }
1656 *f_pos += total;
1657 return total;
1658 }
1659
1660 /* NOTE: assumes unsigned long is 8 bytes */
1661 static ssize_t ui_write(struct file *filp, const char __user *buf,
1662 size_t count, loff_t *f_pos)
1663 {
1664 struct hfi1_devdata *dd = filp->private_data;
1665 void __iomem *base;
1666 unsigned long total, data, csr_off;
1667 int in_lcb;
1668
1669 /* only write 8 byte quantities */
1670 if ((count % 8) != 0)
1671 return -EINVAL;
1672 /* offset must be 8-byte aligned */
1673 if ((*f_pos % 8) != 0)
1674 return -EINVAL;
1675 /* source buffer must be 8-byte aligned */
1676 if ((unsigned long)buf % 8 != 0)
1677 return -EINVAL;
1678 /* must be in range */
1679 if (*f_pos + count > dd->kregend - dd->kregbase)
1680 return -EINVAL;
1681
1682 base = (void __iomem *)dd->kregbase + *f_pos;
1683 csr_off = *f_pos;
1684 in_lcb = 0;
1685 for (total = 0; total < count; total += 8, csr_off += 8) {
1686 if (get_user(data, (unsigned long __user *)(buf + total)))
1687 break;
1688 /* accessing LCB CSRs requires a special procedure */
1689 if (is_lcb_offset(csr_off)) {
1690 if (!in_lcb) {
1691 int ret = acquire_lcb_access(dd, 1);
1692
1693 if (ret)
1694 break;
1695 in_lcb = 1;
1696 }
1697 } else {
1698 if (in_lcb) {
1699 release_lcb_access(dd, 1);
1700 in_lcb = 0;
1701 }
1702 }
1703 writeq(data, base + total);
1704 }
1705 if (in_lcb)
1706 release_lcb_access(dd, 1);
1707 *f_pos += total;
1708 return total;
1709 }
1710
1711 static const struct file_operations ui_file_ops = {
1712 .owner = THIS_MODULE,
1713 .llseek = ui_lseek,
1714 .read = ui_read,
1715 .write = ui_write,
1716 .open = ui_open,
1717 .release = ui_release,
1718 };
1719
1720 #define UI_OFFSET 192 /* device minor offset for UI devices */
1721 static int create_ui = 1;
1722
1723 static struct cdev wildcard_cdev;
1724 static struct device *wildcard_device;
1725
1726 static atomic_t user_count = ATOMIC_INIT(0);
1727
1728 static void user_remove(struct hfi1_devdata *dd)
1729 {
1730 if (atomic_dec_return(&user_count) == 0)
1731 hfi1_cdev_cleanup(&wildcard_cdev, &wildcard_device);
1732
1733 hfi1_cdev_cleanup(&dd->user_cdev, &dd->user_device);
1734 hfi1_cdev_cleanup(&dd->ui_cdev, &dd->ui_device);
1735 }
1736
1737 static int user_add(struct hfi1_devdata *dd)
1738 {
1739 char name[10];
1740 int ret;
1741
1742 if (atomic_inc_return(&user_count) == 1) {
1743 ret = hfi1_cdev_init(0, class_name(), &hfi1_file_ops,
1744 &wildcard_cdev, &wildcard_device,
1745 true);
1746 if (ret)
1747 goto done;
1748 }
1749
1750 snprintf(name, sizeof(name), "%s_%d", class_name(), dd->unit);
1751 ret = hfi1_cdev_init(dd->unit + 1, name, &hfi1_file_ops,
1752 &dd->user_cdev, &dd->user_device,
1753 true);
1754 if (ret)
1755 goto done;
1756
1757 if (create_ui) {
1758 snprintf(name, sizeof(name),
1759 "%s_ui%d", class_name(), dd->unit);
1760 ret = hfi1_cdev_init(dd->unit + UI_OFFSET, name, &ui_file_ops,
1761 &dd->ui_cdev, &dd->ui_device,
1762 false);
1763 if (ret)
1764 goto done;
1765 }
1766
1767 return 0;
1768 done:
1769 user_remove(dd);
1770 return ret;
1771 }
1772
1773 /*
1774 * Create per-unit files in /dev
1775 */
1776 int hfi1_device_create(struct hfi1_devdata *dd)
1777 {
1778 int r, ret;
1779
1780 r = user_add(dd);
1781 ret = hfi1_diag_add(dd);
1782 if (r && !ret)
1783 ret = r;
1784 return ret;
1785 }
1786
1787 /*
1788 * Remove per-unit files in /dev
1789 * void, core kernel returns no errors for this stuff
1790 */
1791 void hfi1_device_remove(struct hfi1_devdata *dd)
1792 {
1793 user_remove(dd);
1794 hfi1_diag_remove(dd);
1795 }
Linux kernel - Deliverables
This page took 0.097417 seconds and 6 git commands to generate.