Commit | Line | Data |
---|---|---|
7f510b46 | 1 | /* |
759d5768 | 2 | * Copyright (c) 2006 QLogic, Inc. All rights reserved. |
7f510b46 BS |
3 | * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved. |
4 | * | |
5 | * This software is available to you under a choice of one of two | |
6 | * licenses. You may choose to be licensed under the terms of the GNU | |
7 | * General Public License (GPL) Version 2, available from the file | |
8 | * COPYING in the main directory of this source tree, or the | |
9 | * OpenIB.org BSD license below: | |
10 | * | |
11 | * Redistribution and use in source and binary forms, with or | |
12 | * without modification, are permitted provided that the following | |
13 | * conditions are met: | |
14 | * | |
15 | * - Redistributions of source code must retain the above | |
16 | * copyright notice, this list of conditions and the following | |
17 | * disclaimer. | |
18 | * | |
19 | * - Redistributions in binary form must reproduce the above | |
20 | * copyright notice, this list of conditions and the following | |
21 | * disclaimer in the documentation and/or other materials | |
22 | * provided with the distribution. | |
23 | * | |
24 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, | |
25 | * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF | |
26 | * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND | |
27 | * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS | |
28 | * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN | |
29 | * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN | |
30 | * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE | |
31 | * SOFTWARE. | |
32 | */ | |
33 | ||
34 | #include <linux/pci.h> | |
35 | #include <linux/poll.h> | |
36 | #include <linux/cdev.h> | |
37 | #include <linux/swap.h> | |
38 | #include <linux/vmalloc.h> | |
39 | #include <asm/pgtable.h> | |
40 | ||
41 | #include "ipath_kernel.h" | |
7f510b46 | 42 | #include "ipath_layer.h" |
27b678dd | 43 | #include "ipath_common.h" |
7f510b46 BS |
44 | |
45 | static int ipath_open(struct inode *, struct file *); | |
46 | static int ipath_close(struct inode *, struct file *); | |
47 | static ssize_t ipath_write(struct file *, const char __user *, size_t, | |
48 | loff_t *); | |
49 | static unsigned int ipath_poll(struct file *, struct poll_table_struct *); | |
50 | static int ipath_mmap(struct file *, struct vm_area_struct *); | |
51 | ||
52 | static struct file_operations ipath_file_ops = { | |
53 | .owner = THIS_MODULE, | |
54 | .write = ipath_write, | |
55 | .open = ipath_open, | |
56 | .release = ipath_close, | |
57 | .poll = ipath_poll, | |
58 | .mmap = ipath_mmap | |
59 | }; | |
60 | ||
61 | static int ipath_get_base_info(struct ipath_portdata *pd, | |
62 | void __user *ubase, size_t ubase_size) | |
63 | { | |
64 | int ret = 0; | |
65 | struct ipath_base_info *kinfo = NULL; | |
66 | struct ipath_devdata *dd = pd->port_dd; | |
67 | ||
68 | if (ubase_size < sizeof(*kinfo)) { | |
69 | ipath_cdbg(PROC, | |
70 | "Base size %lu, need %lu (version mismatch?)\n", | |
71 | (unsigned long) ubase_size, | |
72 | (unsigned long) sizeof(*kinfo)); | |
73 | ret = -EINVAL; | |
74 | goto bail; | |
75 | } | |
76 | ||
77 | kinfo = kzalloc(sizeof(*kinfo), GFP_KERNEL); | |
78 | if (kinfo == NULL) { | |
79 | ret = -ENOMEM; | |
80 | goto bail; | |
81 | } | |
82 | ||
83 | ret = dd->ipath_f_get_base_info(pd, kinfo); | |
84 | if (ret < 0) | |
85 | goto bail; | |
86 | ||
87 | kinfo->spi_rcvhdr_cnt = dd->ipath_rcvhdrcnt; | |
88 | kinfo->spi_rcvhdrent_size = dd->ipath_rcvhdrentsize; | |
89 | kinfo->spi_tidegrcnt = dd->ipath_rcvegrcnt; | |
90 | kinfo->spi_rcv_egrbufsize = dd->ipath_rcvegrbufsize; | |
91 | /* | |
92 | * have to mmap whole thing | |
93 | */ | |
94 | kinfo->spi_rcv_egrbuftotlen = | |
95 | pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size; | |
96 | kinfo->spi_rcv_egrperchunk = pd->port_rcvegrbufs_perchunk; | |
97 | kinfo->spi_rcv_egrchunksize = kinfo->spi_rcv_egrbuftotlen / | |
98 | pd->port_rcvegrbuf_chunks; | |
99 | kinfo->spi_tidcnt = dd->ipath_rcvtidcnt; | |
100 | /* | |
101 | * for this use, may be ipath_cfgports summed over all chips that | |
102 | * are are configured and present | |
103 | */ | |
104 | kinfo->spi_nports = dd->ipath_cfgports; | |
105 | /* unit (chip/board) our port is on */ | |
106 | kinfo->spi_unit = dd->ipath_unit; | |
107 | /* for now, only a single page */ | |
108 | kinfo->spi_tid_maxsize = PAGE_SIZE; | |
109 | ||
110 | /* | |
111 | * Doing this per port, and based on the skip value, etc. This has | |
112 | * to be the actual buffer size, since the protocol code treats it | |
113 | * as an array. | |
114 | * | |
115 | * These have to be set to user addresses in the user code via mmap. | |
116 | * These values are used on return to user code for the mmap target | |
117 | * addresses only. For 32 bit, same 44 bit address problem, so use | |
118 | * the physical address, not virtual. Before 2.6.11, using the | |
119 | * page_address() macro worked, but in 2.6.11, even that returns the | |
120 | * full 64 bit address (upper bits all 1's). So far, using the | |
121 | * physical addresses (or chip offsets, for chip mapping) works, but | |
122 | * no doubt some future kernel release will chang that, and we'll be | |
123 | * on to yet another method of dealing with this | |
124 | */ | |
125 | kinfo->spi_rcvhdr_base = (u64) pd->port_rcvhdrq_phys; | |
f37bda92 | 126 | kinfo->spi_rcvhdr_tailaddr = (u64)pd->port_rcvhdrqtailaddr_phys; |
7f510b46 BS |
127 | kinfo->spi_rcv_egrbufs = (u64) pd->port_rcvegr_phys; |
128 | kinfo->spi_pioavailaddr = (u64) dd->ipath_pioavailregs_phys; | |
129 | kinfo->spi_status = (u64) kinfo->spi_pioavailaddr + | |
130 | (void *) dd->ipath_statusp - | |
131 | (void *) dd->ipath_pioavailregs_dma; | |
132 | kinfo->spi_piobufbase = (u64) pd->port_piobufs; | |
133 | kinfo->__spi_uregbase = | |
134 | dd->ipath_uregbase + dd->ipath_palign * pd->port_port; | |
135 | ||
136 | kinfo->spi_pioindex = dd->ipath_pbufsport * (pd->port_port - 1); | |
137 | kinfo->spi_piocnt = dd->ipath_pbufsport; | |
138 | kinfo->spi_pioalign = dd->ipath_palign; | |
139 | ||
140 | kinfo->spi_qpair = IPATH_KD_QP; | |
141 | kinfo->spi_piosize = dd->ipath_ibmaxlen; | |
142 | kinfo->spi_mtu = dd->ipath_ibmaxlen; /* maxlen, not ibmtu */ | |
143 | kinfo->spi_port = pd->port_port; | |
eaf6733b | 144 | kinfo->spi_sw_version = IPATH_KERN_SWVERSION; |
7f510b46 BS |
145 | kinfo->spi_hw_version = dd->ipath_revision; |
146 | ||
147 | if (copy_to_user(ubase, kinfo, sizeof(*kinfo))) | |
148 | ret = -EFAULT; | |
149 | ||
150 | bail: | |
151 | kfree(kinfo); | |
152 | return ret; | |
153 | } | |
154 | ||
155 | /** | |
156 | * ipath_tid_update - update a port TID | |
157 | * @pd: the port | |
158 | * @ti: the TID information | |
159 | * | |
160 | * The new implementation as of Oct 2004 is that the driver assigns | |
161 | * the tid and returns it to the caller. To make it easier to | |
162 | * catch bugs, and to reduce search time, we keep a cursor for | |
163 | * each port, walking the shadow tid array to find one that's not | |
164 | * in use. | |
165 | * | |
166 | * For now, if we can't allocate the full list, we fail, although | |
167 | * in the long run, we'll allocate as many as we can, and the | |
168 | * caller will deal with that by trying the remaining pages later. | |
169 | * That means that when we fail, we have to mark the tids as not in | |
170 | * use again, in our shadow copy. | |
171 | * | |
172 | * It's up to the caller to free the tids when they are done. | |
173 | * We'll unlock the pages as they free them. | |
174 | * | |
175 | * Also, right now we are locking one page at a time, but since | |
176 | * the intended use of this routine is for a single group of | |
177 | * virtually contiguous pages, that should change to improve | |
178 | * performance. | |
179 | */ | |
180 | static int ipath_tid_update(struct ipath_portdata *pd, | |
181 | const struct ipath_tid_info *ti) | |
182 | { | |
183 | int ret = 0, ntids; | |
184 | u32 tid, porttid, cnt, i, tidcnt; | |
185 | u16 *tidlist; | |
186 | struct ipath_devdata *dd = pd->port_dd; | |
187 | u64 physaddr; | |
188 | unsigned long vaddr; | |
189 | u64 __iomem *tidbase; | |
190 | unsigned long tidmap[8]; | |
191 | struct page **pagep = NULL; | |
192 | ||
193 | if (!dd->ipath_pageshadow) { | |
194 | ret = -ENOMEM; | |
195 | goto done; | |
196 | } | |
197 | ||
198 | cnt = ti->tidcnt; | |
199 | if (!cnt) { | |
200 | ipath_dbg("After copyin, tidcnt 0, tidlist %llx\n", | |
201 | (unsigned long long) ti->tidlist); | |
202 | /* | |
203 | * Should we treat as success? likely a bug | |
204 | */ | |
205 | ret = -EFAULT; | |
206 | goto done; | |
207 | } | |
208 | tidcnt = dd->ipath_rcvtidcnt; | |
209 | if (cnt >= tidcnt) { | |
210 | /* make sure it all fits in port_tid_pg_list */ | |
211 | dev_info(&dd->pcidev->dev, "Process tried to allocate %u " | |
212 | "TIDs, only trying max (%u)\n", cnt, tidcnt); | |
213 | cnt = tidcnt; | |
214 | } | |
215 | pagep = (struct page **)pd->port_tid_pg_list; | |
216 | tidlist = (u16 *) (&pagep[cnt]); | |
217 | ||
218 | memset(tidmap, 0, sizeof(tidmap)); | |
219 | tid = pd->port_tidcursor; | |
220 | /* before decrement; chip actual # */ | |
221 | porttid = pd->port_port * tidcnt; | |
222 | ntids = tidcnt; | |
223 | tidbase = (u64 __iomem *) (((char __iomem *) dd->ipath_kregbase) + | |
224 | dd->ipath_rcvtidbase + | |
225 | porttid * sizeof(*tidbase)); | |
226 | ||
227 | ipath_cdbg(VERBOSE, "Port%u %u tids, cursor %u, tidbase %p\n", | |
228 | pd->port_port, cnt, tid, tidbase); | |
229 | ||
230 | /* virtual address of first page in transfer */ | |
231 | vaddr = ti->tidvaddr; | |
232 | if (!access_ok(VERIFY_WRITE, (void __user *) vaddr, | |
233 | cnt * PAGE_SIZE)) { | |
234 | ipath_dbg("Fail vaddr %p, %u pages, !access_ok\n", | |
235 | (void *)vaddr, cnt); | |
236 | ret = -EFAULT; | |
237 | goto done; | |
238 | } | |
239 | ret = ipath_get_user_pages(vaddr, cnt, pagep); | |
240 | if (ret) { | |
241 | if (ret == -EBUSY) { | |
242 | ipath_dbg("Failed to lock addr %p, %u pages " | |
243 | "(already locked)\n", | |
244 | (void *) vaddr, cnt); | |
245 | /* | |
246 | * for now, continue, and see what happens but with | |
247 | * the new implementation, this should never happen, | |
248 | * unless perhaps the user has mpin'ed the pages | |
249 | * themselves (something we need to test) | |
250 | */ | |
251 | ret = 0; | |
252 | } else { | |
253 | dev_info(&dd->pcidev->dev, | |
254 | "Failed to lock addr %p, %u pages: " | |
255 | "errno %d\n", (void *) vaddr, cnt, -ret); | |
256 | goto done; | |
257 | } | |
258 | } | |
259 | for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) { | |
260 | for (; ntids--; tid++) { | |
261 | if (tid == tidcnt) | |
262 | tid = 0; | |
263 | if (!dd->ipath_pageshadow[porttid + tid]) | |
264 | break; | |
265 | } | |
266 | if (ntids < 0) { | |
267 | /* | |
268 | * oops, wrapped all the way through their TIDs, | |
269 | * and didn't have enough free; see comments at | |
270 | * start of routine | |
271 | */ | |
272 | ipath_dbg("Not enough free TIDs for %u pages " | |
273 | "(index %d), failing\n", cnt, i); | |
274 | i--; /* last tidlist[i] not filled in */ | |
275 | ret = -ENOMEM; | |
276 | break; | |
277 | } | |
278 | tidlist[i] = tid; | |
279 | ipath_cdbg(VERBOSE, "Updating idx %u to TID %u, " | |
280 | "vaddr %lx\n", i, tid, vaddr); | |
281 | /* we "know" system pages and TID pages are same size */ | |
282 | dd->ipath_pageshadow[porttid + tid] = pagep[i]; | |
283 | /* | |
284 | * don't need atomic or it's overhead | |
285 | */ | |
286 | __set_bit(tid, tidmap); | |
287 | physaddr = page_to_phys(pagep[i]); | |
288 | ipath_stats.sps_pagelocks++; | |
289 | ipath_cdbg(VERBOSE, | |
290 | "TID %u, vaddr %lx, physaddr %llx pgp %p\n", | |
291 | tid, vaddr, (unsigned long long) physaddr, | |
292 | pagep[i]); | |
293 | dd->ipath_f_put_tid(dd, &tidbase[tid], 1, physaddr); | |
294 | /* | |
295 | * don't check this tid in ipath_portshadow, since we | |
296 | * just filled it in; start with the next one. | |
297 | */ | |
298 | tid++; | |
299 | } | |
300 | ||
301 | if (ret) { | |
302 | u32 limit; | |
303 | cleanup: | |
304 | /* jump here if copy out of updated info failed... */ | |
305 | ipath_dbg("After failure (ret=%d), undo %d of %d entries\n", | |
306 | -ret, i, cnt); | |
307 | /* same code that's in ipath_free_tid() */ | |
308 | limit = sizeof(tidmap) * BITS_PER_BYTE; | |
309 | if (limit > tidcnt) | |
310 | /* just in case size changes in future */ | |
311 | limit = tidcnt; | |
312 | tid = find_first_bit((const unsigned long *)tidmap, limit); | |
313 | for (; tid < limit; tid++) { | |
314 | if (!test_bit(tid, tidmap)) | |
315 | continue; | |
316 | if (dd->ipath_pageshadow[porttid + tid]) { | |
317 | ipath_cdbg(VERBOSE, "Freeing TID %u\n", | |
318 | tid); | |
319 | dd->ipath_f_put_tid(dd, &tidbase[tid], 1, | |
320 | dd->ipath_tidinvalid); | |
321 | dd->ipath_pageshadow[porttid + tid] = NULL; | |
322 | ipath_stats.sps_pageunlocks++; | |
323 | } | |
324 | } | |
325 | ipath_release_user_pages(pagep, cnt); | |
326 | } else { | |
327 | /* | |
328 | * Copy the updated array, with ipath_tid's filled in, back | |
329 | * to user. Since we did the copy in already, this "should | |
330 | * never fail" If it does, we have to clean up... | |
331 | */ | |
332 | if (copy_to_user((void __user *) | |
333 | (unsigned long) ti->tidlist, | |
334 | tidlist, cnt * sizeof(*tidlist))) { | |
335 | ret = -EFAULT; | |
336 | goto cleanup; | |
337 | } | |
338 | if (copy_to_user((void __user *) (unsigned long) ti->tidmap, | |
339 | tidmap, sizeof tidmap)) { | |
340 | ret = -EFAULT; | |
341 | goto cleanup; | |
342 | } | |
343 | if (tid == tidcnt) | |
344 | tid = 0; | |
345 | pd->port_tidcursor = tid; | |
346 | } | |
347 | ||
348 | done: | |
349 | if (ret) | |
350 | ipath_dbg("Failed to map %u TID pages, failing with %d\n", | |
351 | ti->tidcnt, -ret); | |
352 | return ret; | |
353 | } | |
354 | ||
355 | /** | |
356 | * ipath_tid_free - free a port TID | |
357 | * @pd: the port | |
358 | * @ti: the TID info | |
359 | * | |
360 | * right now we are unlocking one page at a time, but since | |
361 | * the intended use of this routine is for a single group of | |
362 | * virtually contiguous pages, that should change to improve | |
363 | * performance. We check that the TID is in range for this port | |
364 | * but otherwise don't check validity; if user has an error and | |
365 | * frees the wrong tid, it's only their own data that can thereby | |
366 | * be corrupted. We do check that the TID was in use, for sanity | |
367 | * We always use our idea of the saved address, not the address that | |
368 | * they pass in to us. | |
369 | */ | |
370 | ||
371 | static int ipath_tid_free(struct ipath_portdata *pd, | |
372 | const struct ipath_tid_info *ti) | |
373 | { | |
374 | int ret = 0; | |
375 | u32 tid, porttid, cnt, limit, tidcnt; | |
376 | struct ipath_devdata *dd = pd->port_dd; | |
377 | u64 __iomem *tidbase; | |
378 | unsigned long tidmap[8]; | |
379 | ||
380 | if (!dd->ipath_pageshadow) { | |
381 | ret = -ENOMEM; | |
382 | goto done; | |
383 | } | |
384 | ||
385 | if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap, | |
386 | sizeof tidmap)) { | |
387 | ret = -EFAULT; | |
388 | goto done; | |
389 | } | |
390 | ||
391 | porttid = pd->port_port * dd->ipath_rcvtidcnt; | |
392 | tidbase = (u64 __iomem *) ((char __iomem *)(dd->ipath_kregbase) + | |
393 | dd->ipath_rcvtidbase + | |
394 | porttid * sizeof(*tidbase)); | |
395 | ||
396 | tidcnt = dd->ipath_rcvtidcnt; | |
397 | limit = sizeof(tidmap) * BITS_PER_BYTE; | |
398 | if (limit > tidcnt) | |
399 | /* just in case size changes in future */ | |
400 | limit = tidcnt; | |
401 | tid = find_first_bit(tidmap, limit); | |
402 | ipath_cdbg(VERBOSE, "Port%u free %u tids; first bit (max=%d) " | |
403 | "set is %d, porttid %u\n", pd->port_port, ti->tidcnt, | |
404 | limit, tid, porttid); | |
405 | for (cnt = 0; tid < limit; tid++) { | |
406 | /* | |
407 | * small optimization; if we detect a run of 3 or so without | |
408 | * any set, use find_first_bit again. That's mainly to | |
409 | * accelerate the case where we wrapped, so we have some at | |
410 | * the beginning, and some at the end, and a big gap | |
411 | * in the middle. | |
412 | */ | |
413 | if (!test_bit(tid, tidmap)) | |
414 | continue; | |
415 | cnt++; | |
416 | if (dd->ipath_pageshadow[porttid + tid]) { | |
417 | ipath_cdbg(VERBOSE, "PID %u freeing TID %u\n", | |
418 | pd->port_pid, tid); | |
419 | dd->ipath_f_put_tid(dd, &tidbase[tid], 1, | |
420 | dd->ipath_tidinvalid); | |
421 | ipath_release_user_pages( | |
422 | &dd->ipath_pageshadow[porttid + tid], 1); | |
423 | dd->ipath_pageshadow[porttid + tid] = NULL; | |
424 | ipath_stats.sps_pageunlocks++; | |
425 | } else | |
426 | ipath_dbg("Unused tid %u, ignoring\n", tid); | |
427 | } | |
428 | if (cnt != ti->tidcnt) | |
429 | ipath_dbg("passed in tidcnt %d, only %d bits set in map\n", | |
430 | ti->tidcnt, cnt); | |
431 | done: | |
432 | if (ret) | |
433 | ipath_dbg("Failed to unmap %u TID pages, failing with %d\n", | |
434 | ti->tidcnt, -ret); | |
435 | return ret; | |
436 | } | |
437 | ||
438 | /** | |
439 | * ipath_set_part_key - set a partition key | |
440 | * @pd: the port | |
441 | * @key: the key | |
442 | * | |
443 | * We can have up to 4 active at a time (other than the default, which is | |
444 | * always allowed). This is somewhat tricky, since multiple ports may set | |
445 | * the same key, so we reference count them, and clean up at exit. All 4 | |
446 | * partition keys are packed into a single infinipath register. It's an | |
447 | * error for a process to set the same pkey multiple times. We provide no | |
448 | * mechanism to de-allocate a pkey at this time, we may eventually need to | |
449 | * do that. I've used the atomic operations, and no locking, and only make | |
450 | * a single pass through what's available. This should be more than | |
451 | * adequate for some time. I'll think about spinlocks or the like if and as | |
452 | * it's necessary. | |
453 | */ | |
454 | static int ipath_set_part_key(struct ipath_portdata *pd, u16 key) | |
455 | { | |
456 | struct ipath_devdata *dd = pd->port_dd; | |
457 | int i, any = 0, pidx = -1; | |
458 | u16 lkey = key & 0x7FFF; | |
459 | int ret; | |
460 | ||
27b678dd | 461 | if (lkey == (IPATH_DEFAULT_P_KEY & 0x7FFF)) { |
7f510b46 BS |
462 | /* nothing to do; this key always valid */ |
463 | ret = 0; | |
464 | goto bail; | |
465 | } | |
466 | ||
467 | ipath_cdbg(VERBOSE, "p%u try to set pkey %hx, current keys " | |
468 | "%hx:%x %hx:%x %hx:%x %hx:%x\n", | |
469 | pd->port_port, key, dd->ipath_pkeys[0], | |
470 | atomic_read(&dd->ipath_pkeyrefs[0]), dd->ipath_pkeys[1], | |
471 | atomic_read(&dd->ipath_pkeyrefs[1]), dd->ipath_pkeys[2], | |
472 | atomic_read(&dd->ipath_pkeyrefs[2]), dd->ipath_pkeys[3], | |
473 | atomic_read(&dd->ipath_pkeyrefs[3])); | |
474 | ||
475 | if (!lkey) { | |
476 | ipath_cdbg(PROC, "p%u tries to set key 0, not allowed\n", | |
477 | pd->port_port); | |
478 | ret = -EINVAL; | |
479 | goto bail; | |
480 | } | |
481 | ||
482 | /* | |
483 | * Set the full membership bit, because it has to be | |
484 | * set in the register or the packet, and it seems | |
485 | * cleaner to set in the register than to force all | |
486 | * callers to set it. (see bug 4331) | |
487 | */ | |
488 | key |= 0x8000; | |
489 | ||
490 | for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) { | |
491 | if (!pd->port_pkeys[i] && pidx == -1) | |
492 | pidx = i; | |
493 | if (pd->port_pkeys[i] == key) { | |
494 | ipath_cdbg(VERBOSE, "p%u tries to set same pkey " | |
495 | "(%x) more than once\n", | |
496 | pd->port_port, key); | |
497 | ret = -EEXIST; | |
498 | goto bail; | |
499 | } | |
500 | } | |
501 | if (pidx == -1) { | |
502 | ipath_dbg("All pkeys for port %u already in use, " | |
503 | "can't set %x\n", pd->port_port, key); | |
504 | ret = -EBUSY; | |
505 | goto bail; | |
506 | } | |
507 | for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) { | |
508 | if (!dd->ipath_pkeys[i]) { | |
509 | any++; | |
510 | continue; | |
511 | } | |
512 | if (dd->ipath_pkeys[i] == key) { | |
513 | atomic_t *pkrefs = &dd->ipath_pkeyrefs[i]; | |
514 | ||
515 | if (atomic_inc_return(pkrefs) > 1) { | |
516 | pd->port_pkeys[pidx] = key; | |
517 | ipath_cdbg(VERBOSE, "p%u set key %x " | |
518 | "matches #%d, count now %d\n", | |
519 | pd->port_port, key, i, | |
520 | atomic_read(pkrefs)); | |
521 | ret = 0; | |
522 | goto bail; | |
523 | } else { | |
524 | /* | |
525 | * lost race, decrement count, catch below | |
526 | */ | |
527 | atomic_dec(pkrefs); | |
528 | ipath_cdbg(VERBOSE, "Lost race, count was " | |
529 | "0, after dec, it's %d\n", | |
530 | atomic_read(pkrefs)); | |
531 | any++; | |
532 | } | |
533 | } | |
534 | if ((dd->ipath_pkeys[i] & 0x7FFF) == lkey) { | |
535 | /* | |
536 | * It makes no sense to have both the limited and | |
537 | * full membership PKEY set at the same time since | |
538 | * the unlimited one will disable the limited one. | |
539 | */ | |
540 | ret = -EEXIST; | |
541 | goto bail; | |
542 | } | |
543 | } | |
544 | if (!any) { | |
545 | ipath_dbg("port %u, all pkeys already in use, " | |
546 | "can't set %x\n", pd->port_port, key); | |
547 | ret = -EBUSY; | |
548 | goto bail; | |
549 | } | |
550 | for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) { | |
551 | if (!dd->ipath_pkeys[i] && | |
552 | atomic_inc_return(&dd->ipath_pkeyrefs[i]) == 1) { | |
553 | u64 pkey; | |
554 | ||
555 | /* for ipathstats, etc. */ | |
556 | ipath_stats.sps_pkeys[i] = lkey; | |
557 | pd->port_pkeys[pidx] = dd->ipath_pkeys[i] = key; | |
558 | pkey = | |
559 | (u64) dd->ipath_pkeys[0] | | |
560 | ((u64) dd->ipath_pkeys[1] << 16) | | |
561 | ((u64) dd->ipath_pkeys[2] << 32) | | |
562 | ((u64) dd->ipath_pkeys[3] << 48); | |
563 | ipath_cdbg(PROC, "p%u set key %x in #%d, " | |
564 | "portidx %d, new pkey reg %llx\n", | |
565 | pd->port_port, key, i, pidx, | |
566 | (unsigned long long) pkey); | |
567 | ipath_write_kreg( | |
568 | dd, dd->ipath_kregs->kr_partitionkey, pkey); | |
569 | ||
570 | ret = 0; | |
571 | goto bail; | |
572 | } | |
573 | } | |
574 | ipath_dbg("port %u, all pkeys already in use 2nd pass, " | |
575 | "can't set %x\n", pd->port_port, key); | |
576 | ret = -EBUSY; | |
577 | ||
578 | bail: | |
579 | return ret; | |
580 | } | |
581 | ||
582 | /** | |
583 | * ipath_manage_rcvq - manage a port's receive queue | |
584 | * @pd: the port | |
585 | * @start_stop: action to carry out | |
586 | * | |
587 | * start_stop == 0 disables receive on the port, for use in queue | |
588 | * overflow conditions. start_stop==1 re-enables, to be used to | |
589 | * re-init the software copy of the head register | |
590 | */ | |
591 | static int ipath_manage_rcvq(struct ipath_portdata *pd, int start_stop) | |
592 | { | |
593 | struct ipath_devdata *dd = pd->port_dd; | |
594 | u64 tval; | |
595 | ||
596 | ipath_cdbg(PROC, "%sabling rcv for unit %u port %u\n", | |
597 | start_stop ? "en" : "dis", dd->ipath_unit, | |
598 | pd->port_port); | |
599 | /* atomically clear receive enable port. */ | |
600 | if (start_stop) { | |
601 | /* | |
602 | * On enable, force in-memory copy of the tail register to | |
603 | * 0, so that protocol code doesn't have to worry about | |
604 | * whether or not the chip has yet updated the in-memory | |
605 | * copy or not on return from the system call. The chip | |
606 | * always resets it's tail register back to 0 on a | |
607 | * transition from disabled to enabled. This could cause a | |
608 | * problem if software was broken, and did the enable w/o | |
609 | * the disable, but eventually the in-memory copy will be | |
610 | * updated and correct itself, even in the face of software | |
611 | * bugs. | |
612 | */ | |
613 | *pd->port_rcvhdrtail_kvaddr = 0; | |
614 | set_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port, | |
615 | &dd->ipath_rcvctrl); | |
616 | } else | |
617 | clear_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port, | |
618 | &dd->ipath_rcvctrl); | |
619 | ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl, | |
620 | dd->ipath_rcvctrl); | |
621 | /* now be sure chip saw it before we return */ | |
622 | tval = ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch); | |
623 | if (start_stop) { | |
624 | /* | |
625 | * And try to be sure that tail reg update has happened too. | |
626 | * This should in theory interlock with the RXE changes to | |
627 | * the tail register. Don't assign it to the tail register | |
628 | * in memory copy, since we could overwrite an update by the | |
629 | * chip if we did. | |
630 | */ | |
631 | tval = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port); | |
632 | } | |
633 | /* always; new head should be equal to new tail; see above */ | |
634 | return 0; | |
635 | } | |
636 | ||
637 | static void ipath_clean_part_key(struct ipath_portdata *pd, | |
638 | struct ipath_devdata *dd) | |
639 | { | |
640 | int i, j, pchanged = 0; | |
641 | u64 oldpkey; | |
642 | ||
643 | /* for debugging only */ | |
644 | oldpkey = (u64) dd->ipath_pkeys[0] | | |
645 | ((u64) dd->ipath_pkeys[1] << 16) | | |
646 | ((u64) dd->ipath_pkeys[2] << 32) | | |
647 | ((u64) dd->ipath_pkeys[3] << 48); | |
648 | ||
649 | for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) { | |
650 | if (!pd->port_pkeys[i]) | |
651 | continue; | |
652 | ipath_cdbg(VERBOSE, "look for key[%d] %hx in pkeys\n", i, | |
653 | pd->port_pkeys[i]); | |
654 | for (j = 0; j < ARRAY_SIZE(dd->ipath_pkeys); j++) { | |
655 | /* check for match independent of the global bit */ | |
656 | if ((dd->ipath_pkeys[j] & 0x7fff) != | |
657 | (pd->port_pkeys[i] & 0x7fff)) | |
658 | continue; | |
659 | if (atomic_dec_and_test(&dd->ipath_pkeyrefs[j])) { | |
660 | ipath_cdbg(VERBOSE, "p%u clear key " | |
661 | "%x matches #%d\n", | |
662 | pd->port_port, | |
663 | pd->port_pkeys[i], j); | |
664 | ipath_stats.sps_pkeys[j] = | |
665 | dd->ipath_pkeys[j] = 0; | |
666 | pchanged++; | |
667 | } | |
668 | else ipath_cdbg( | |
669 | VERBOSE, "p%u key %x matches #%d, " | |
670 | "but ref still %d\n", pd->port_port, | |
671 | pd->port_pkeys[i], j, | |
672 | atomic_read(&dd->ipath_pkeyrefs[j])); | |
673 | break; | |
674 | } | |
675 | pd->port_pkeys[i] = 0; | |
676 | } | |
677 | if (pchanged) { | |
678 | u64 pkey = (u64) dd->ipath_pkeys[0] | | |
679 | ((u64) dd->ipath_pkeys[1] << 16) | | |
680 | ((u64) dd->ipath_pkeys[2] << 32) | | |
681 | ((u64) dd->ipath_pkeys[3] << 48); | |
682 | ipath_cdbg(VERBOSE, "p%u old pkey reg %llx, " | |
683 | "new pkey reg %llx\n", pd->port_port, | |
684 | (unsigned long long) oldpkey, | |
685 | (unsigned long long) pkey); | |
686 | ipath_write_kreg(dd, dd->ipath_kregs->kr_partitionkey, | |
687 | pkey); | |
688 | } | |
689 | } | |
690 | ||
691 | /** | |
692 | * ipath_create_user_egr - allocate eager TID buffers | |
693 | * @pd: the port to allocate TID buffers for | |
694 | * | |
695 | * This routine is now quite different for user and kernel, because | |
696 | * the kernel uses skb's, for the accelerated network performance | |
697 | * This is the user port version | |
698 | * | |
699 | * Allocate the eager TID buffers and program them into infinipath | |
700 | * They are no longer completely contiguous, we do multiple allocation | |
701 | * calls. | |
702 | */ | |
703 | static int ipath_create_user_egr(struct ipath_portdata *pd) | |
704 | { | |
705 | struct ipath_devdata *dd = pd->port_dd; | |
706 | unsigned e, egrcnt, alloced, egrperchunk, chunk, egrsize, egroff; | |
707 | size_t size; | |
708 | int ret; | |
0ed9a4a0 BS |
709 | gfp_t gfp_flags; |
710 | ||
711 | /* | |
712 | * GFP_USER, but without GFP_FS, so buffer cache can be | |
713 | * coalesced (we hope); otherwise, even at order 4, | |
714 | * heavy filesystem activity makes these fail, and we can | |
715 | * use compound pages. | |
716 | */ | |
717 | gfp_flags = __GFP_WAIT | __GFP_IO | __GFP_COMP; | |
7f510b46 BS |
718 | |
719 | egrcnt = dd->ipath_rcvegrcnt; | |
720 | /* TID number offset for this port */ | |
721 | egroff = pd->port_port * egrcnt; | |
722 | egrsize = dd->ipath_rcvegrbufsize; | |
723 | ipath_cdbg(VERBOSE, "Allocating %d egr buffers, at egrtid " | |
724 | "offset %x, egrsize %u\n", egrcnt, egroff, egrsize); | |
725 | ||
726 | /* | |
727 | * to avoid wasting a lot of memory, we allocate 32KB chunks of | |
728 | * physically contiguous memory, advance through it until used up | |
729 | * and then allocate more. Of course, we need memory to store those | |
730 | * extra pointers, now. Started out with 256KB, but under heavy | |
731 | * memory pressure (creating large files and then copying them over | |
732 | * NFS while doing lots of MPI jobs), we hit some allocation | |
733 | * failures, even though we can sleep... (2.6.10) Still get | |
0ed9a4a0 BS |
734 | * failures at 64K. 32K is the lowest we can go without wasting |
735 | * additional memory. | |
7f510b46 BS |
736 | */ |
737 | size = 0x8000; | |
738 | alloced = ALIGN(egrsize * egrcnt, size); | |
739 | egrperchunk = size / egrsize; | |
740 | chunk = (egrcnt + egrperchunk - 1) / egrperchunk; | |
741 | pd->port_rcvegrbuf_chunks = chunk; | |
742 | pd->port_rcvegrbufs_perchunk = egrperchunk; | |
743 | pd->port_rcvegrbuf_size = size; | |
744 | pd->port_rcvegrbuf = vmalloc(chunk * sizeof(pd->port_rcvegrbuf[0])); | |
745 | if (!pd->port_rcvegrbuf) { | |
746 | ret = -ENOMEM; | |
747 | goto bail; | |
748 | } | |
749 | pd->port_rcvegrbuf_phys = | |
750 | vmalloc(chunk * sizeof(pd->port_rcvegrbuf_phys[0])); | |
751 | if (!pd->port_rcvegrbuf_phys) { | |
752 | ret = -ENOMEM; | |
753 | goto bail_rcvegrbuf; | |
754 | } | |
755 | for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) { | |
7f510b46 BS |
756 | |
757 | pd->port_rcvegrbuf[e] = dma_alloc_coherent( | |
758 | &dd->pcidev->dev, size, &pd->port_rcvegrbuf_phys[e], | |
759 | gfp_flags); | |
760 | ||
761 | if (!pd->port_rcvegrbuf[e]) { | |
762 | ret = -ENOMEM; | |
763 | goto bail_rcvegrbuf_phys; | |
764 | } | |
765 | } | |
766 | ||
767 | pd->port_rcvegr_phys = pd->port_rcvegrbuf_phys[0]; | |
768 | ||
769 | for (e = chunk = 0; chunk < pd->port_rcvegrbuf_chunks; chunk++) { | |
770 | dma_addr_t pa = pd->port_rcvegrbuf_phys[chunk]; | |
771 | unsigned i; | |
772 | ||
773 | for (i = 0; e < egrcnt && i < egrperchunk; e++, i++) { | |
774 | dd->ipath_f_put_tid(dd, e + egroff + | |
775 | (u64 __iomem *) | |
776 | ((char __iomem *) | |
777 | dd->ipath_kregbase + | |
778 | dd->ipath_rcvegrbase), 0, pa); | |
779 | pa += egrsize; | |
780 | } | |
781 | cond_resched(); /* don't hog the cpu */ | |
782 | } | |
783 | ||
784 | ret = 0; | |
785 | goto bail; | |
786 | ||
787 | bail_rcvegrbuf_phys: | |
788 | for (e = 0; e < pd->port_rcvegrbuf_chunks && | |
f37bda92 | 789 | pd->port_rcvegrbuf[e]; e++) { |
7f510b46 BS |
790 | dma_free_coherent(&dd->pcidev->dev, size, |
791 | pd->port_rcvegrbuf[e], | |
792 | pd->port_rcvegrbuf_phys[e]); | |
793 | ||
f37bda92 | 794 | } |
7f510b46 BS |
795 | vfree(pd->port_rcvegrbuf_phys); |
796 | pd->port_rcvegrbuf_phys = NULL; | |
797 | bail_rcvegrbuf: | |
798 | vfree(pd->port_rcvegrbuf); | |
799 | pd->port_rcvegrbuf = NULL; | |
800 | bail: | |
801 | return ret; | |
802 | } | |
803 | ||
804 | static int ipath_do_user_init(struct ipath_portdata *pd, | |
805 | const struct ipath_user_info *uinfo) | |
806 | { | |
807 | int ret = 0; | |
808 | struct ipath_devdata *dd = pd->port_dd; | |
7f510b46 | 809 | u32 head32; |
7f510b46 BS |
810 | |
811 | /* for now, if major version is different, bail */ | |
812 | if ((uinfo->spu_userversion >> 16) != IPATH_USER_SWMAJOR) { | |
813 | dev_info(&dd->pcidev->dev, | |
814 | "User major version %d not same as driver " | |
815 | "major %d\n", uinfo->spu_userversion >> 16, | |
816 | IPATH_USER_SWMAJOR); | |
817 | ret = -ENODEV; | |
818 | goto done; | |
819 | } | |
820 | ||
821 | if ((uinfo->spu_userversion & 0xffff) != IPATH_USER_SWMINOR) | |
822 | ipath_dbg("User minor version %d not same as driver " | |
823 | "minor %d\n", uinfo->spu_userversion & 0xffff, | |
824 | IPATH_USER_SWMINOR); | |
825 | ||
826 | if (uinfo->spu_rcvhdrsize) { | |
827 | ret = ipath_setrcvhdrsize(dd, uinfo->spu_rcvhdrsize); | |
828 | if (ret) | |
829 | goto done; | |
830 | } | |
831 | ||
832 | /* for now we do nothing with rcvhdrcnt: uinfo->spu_rcvhdrcnt */ | |
833 | ||
7f510b46 BS |
834 | /* for right now, kernel piobufs are at end, so port 1 is at 0 */ |
835 | pd->port_piobufs = dd->ipath_piobufbase + | |
836 | dd->ipath_pbufsport * (pd->port_port - | |
837 | 1) * dd->ipath_palign; | |
838 | ipath_cdbg(VERBOSE, "Set base of piobufs for port %u to 0x%x\n", | |
839 | pd->port_port, pd->port_piobufs); | |
840 | ||
841 | /* | |
842 | * Now allocate the rcvhdr Q and eager TIDs; skip the TID | |
843 | * array for time being. If pd->port_port > chip-supported, | |
844 | * we need to do extra stuff here to handle by handling overflow | |
845 | * through port 0, someday | |
846 | */ | |
847 | ret = ipath_create_rcvhdrq(dd, pd); | |
848 | if (!ret) | |
849 | ret = ipath_create_user_egr(pd); | |
850 | if (ret) | |
851 | goto done; | |
7f510b46 BS |
852 | |
853 | /* | |
f37bda92 | 854 | * set the eager head register for this port to the current values |
7f510b46 BS |
855 | * of the tail pointers, since we don't know if they were |
856 | * updated on last use of the port. | |
857 | */ | |
7f510b46 BS |
858 | head32 = ipath_read_ureg32(dd, ur_rcvegrindextail, pd->port_port); |
859 | ipath_write_ureg(dd, ur_rcvegrindexhead, head32, pd->port_port); | |
860 | dd->ipath_lastegrheads[pd->port_port] = -1; | |
861 | dd->ipath_lastrcvhdrqtails[pd->port_port] = -1; | |
f37bda92 BS |
862 | ipath_cdbg(VERBOSE, "Wrote port%d egrhead %x from tail regs\n", |
863 | pd->port_port, head32); | |
7f510b46 BS |
864 | pd->port_tidcursor = 0; /* start at beginning after open */ |
865 | /* | |
866 | * now enable the port; the tail registers will be written to memory | |
867 | * by the chip as soon as it sees the write to | |
868 | * dd->ipath_kregs->kr_rcvctrl. The update only happens on | |
869 | * transition from 0 to 1, so clear it first, then set it as part of | |
870 | * enabling the port. This will (very briefly) affect any other | |
871 | * open ports, but it shouldn't be long enough to be an issue. | |
f37bda92 BS |
872 | * We explictly set the in-memory copy to 0 beforehand, so we don't |
873 | * have to wait to be sure the DMA update has happened. | |
7f510b46 | 874 | */ |
f37bda92 BS |
875 | *pd->port_rcvhdrtail_kvaddr = 0ULL; |
876 | set_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port, | |
877 | &dd->ipath_rcvctrl); | |
7f510b46 BS |
878 | ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl, |
879 | dd->ipath_rcvctrl & ~INFINIPATH_R_TAILUPD); | |
880 | ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl, | |
881 | dd->ipath_rcvctrl); | |
7f510b46 BS |
882 | done: |
883 | return ret; | |
884 | } | |
885 | ||
f37bda92 BS |
886 | |
887 | /* common code for the mappings on dma_alloc_coherent mem */ | |
888 | static int ipath_mmap_mem(struct vm_area_struct *vma, | |
889 | struct ipath_portdata *pd, unsigned len, | |
890 | int write_ok, dma_addr_t addr, char *what) | |
891 | { | |
892 | struct ipath_devdata *dd = pd->port_dd; | |
893 | unsigned pfn = (unsigned long)addr >> PAGE_SHIFT; | |
894 | int ret; | |
895 | ||
896 | if ((vma->vm_end - vma->vm_start) > len) { | |
897 | dev_info(&dd->pcidev->dev, | |
898 | "FAIL on %s: len %lx > %x\n", what, | |
899 | vma->vm_end - vma->vm_start, len); | |
900 | ret = -EFAULT; | |
901 | goto bail; | |
902 | } | |
903 | ||
904 | if (!write_ok) { | |
905 | if (vma->vm_flags & VM_WRITE) { | |
906 | dev_info(&dd->pcidev->dev, | |
907 | "%s must be mapped readonly\n", what); | |
908 | ret = -EPERM; | |
909 | goto bail; | |
910 | } | |
911 | ||
912 | /* don't allow them to later change with mprotect */ | |
913 | vma->vm_flags &= ~VM_MAYWRITE; | |
914 | } | |
915 | ||
916 | ret = remap_pfn_range(vma, vma->vm_start, pfn, | |
917 | len, vma->vm_page_prot); | |
918 | if (ret) | |
919 | dev_info(&dd->pcidev->dev, | |
920 | "%s port%u mmap of %lx, %x bytes r%c failed: %d\n", | |
921 | what, pd->port_port, (unsigned long)addr, len, | |
922 | write_ok?'w':'o', ret); | |
923 | else | |
924 | ipath_cdbg(VERBOSE, "%s port%u mmaped %lx, %x bytes r%c\n", | |
925 | what, pd->port_port, (unsigned long)addr, len, | |
926 | write_ok?'w':'o'); | |
927 | bail: | |
928 | return ret; | |
929 | } | |
930 | ||
7f510b46 BS |
931 | static int mmap_ureg(struct vm_area_struct *vma, struct ipath_devdata *dd, |
932 | u64 ureg) | |
933 | { | |
934 | unsigned long phys; | |
935 | int ret; | |
936 | ||
f37bda92 BS |
937 | /* |
938 | * This is real hardware, so use io_remap. This is the mechanism | |
939 | * for the user process to update the head registers for their port | |
940 | * in the chip. | |
941 | */ | |
7f510b46 BS |
942 | if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) { |
943 | dev_info(&dd->pcidev->dev, "FAIL mmap userreg: reqlen " | |
944 | "%lx > PAGE\n", vma->vm_end - vma->vm_start); | |
945 | ret = -EFAULT; | |
946 | } else { | |
947 | phys = dd->ipath_physaddr + ureg; | |
948 | vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); | |
949 | ||
950 | vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND; | |
951 | ret = io_remap_pfn_range(vma, vma->vm_start, | |
952 | phys >> PAGE_SHIFT, | |
953 | vma->vm_end - vma->vm_start, | |
954 | vma->vm_page_prot); | |
955 | } | |
956 | return ret; | |
957 | } | |
958 | ||
959 | static int mmap_piobufs(struct vm_area_struct *vma, | |
960 | struct ipath_devdata *dd, | |
961 | struct ipath_portdata *pd) | |
962 | { | |
963 | unsigned long phys; | |
964 | int ret; | |
965 | ||
966 | /* | |
f37bda92 BS |
967 | * When we map the PIO buffers in the chip, we want to map them as |
968 | * writeonly, no read possible. This prevents access to previous | |
969 | * process data, and catches users who might try to read the i/o | |
970 | * space due to a bug. | |
7f510b46 | 971 | */ |
7f510b46 BS |
972 | if ((vma->vm_end - vma->vm_start) > |
973 | (dd->ipath_pbufsport * dd->ipath_palign)) { | |
974 | dev_info(&dd->pcidev->dev, "FAIL mmap piobufs: " | |
975 | "reqlen %lx > PAGE\n", | |
976 | vma->vm_end - vma->vm_start); | |
977 | ret = -EFAULT; | |
978 | goto bail; | |
979 | } | |
980 | ||
981 | phys = dd->ipath_physaddr + pd->port_piobufs; | |
f37bda92 | 982 | |
7f510b46 | 983 | /* |
f37bda92 | 984 | * Don't mark this as non-cached, or we don't get the |
7f510b46 | 985 | * write combining behavior we want on the PIO buffers! |
7f510b46 BS |
986 | */ |
987 | ||
eb9dc6f4 BS |
988 | #if defined(__powerpc__) |
989 | /* There isn't a generic way to specify writethrough mappings */ | |
990 | pgprot_val(vma->vm_page_prot) |= _PAGE_NO_CACHE; | |
991 | pgprot_val(vma->vm_page_prot) |= _PAGE_WRITETHRU; | |
992 | pgprot_val(vma->vm_page_prot) &= ~_PAGE_GUARDED; | |
993 | #endif | |
994 | ||
7f510b46 BS |
995 | if (vma->vm_flags & VM_READ) { |
996 | dev_info(&dd->pcidev->dev, | |
997 | "Can't map piobufs as readable (flags=%lx)\n", | |
998 | vma->vm_flags); | |
999 | ret = -EPERM; | |
1000 | goto bail; | |
1001 | } | |
1002 | ||
1003 | /* don't allow them to later change to readable with mprotect */ | |
f37bda92 | 1004 | vma->vm_flags &= ~VM_MAYREAD; |
7f510b46 BS |
1005 | vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND; |
1006 | ||
1007 | ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT, | |
1008 | vma->vm_end - vma->vm_start, | |
1009 | vma->vm_page_prot); | |
1010 | bail: | |
1011 | return ret; | |
1012 | } | |
1013 | ||
1014 | static int mmap_rcvegrbufs(struct vm_area_struct *vma, | |
1015 | struct ipath_portdata *pd) | |
1016 | { | |
1017 | struct ipath_devdata *dd = pd->port_dd; | |
1018 | unsigned long start, size; | |
1019 | size_t total_size, i; | |
1020 | dma_addr_t *phys; | |
1021 | int ret; | |
1022 | ||
7f510b46 BS |
1023 | size = pd->port_rcvegrbuf_size; |
1024 | total_size = pd->port_rcvegrbuf_chunks * size; | |
1025 | if ((vma->vm_end - vma->vm_start) > total_size) { | |
1026 | dev_info(&dd->pcidev->dev, "FAIL on egr bufs: " | |
1027 | "reqlen %lx > actual %lx\n", | |
1028 | vma->vm_end - vma->vm_start, | |
1029 | (unsigned long) total_size); | |
1030 | ret = -EFAULT; | |
1031 | goto bail; | |
1032 | } | |
1033 | ||
1034 | if (vma->vm_flags & VM_WRITE) { | |
1035 | dev_info(&dd->pcidev->dev, "Can't map eager buffers as " | |
1036 | "writable (flags=%lx)\n", vma->vm_flags); | |
1037 | ret = -EPERM; | |
1038 | goto bail; | |
1039 | } | |
f37bda92 BS |
1040 | /* don't allow them to later change to writeable with mprotect */ |
1041 | vma->vm_flags &= ~VM_MAYWRITE; | |
7f510b46 BS |
1042 | |
1043 | start = vma->vm_start; | |
1044 | phys = pd->port_rcvegrbuf_phys; | |
1045 | ||
7f510b46 BS |
1046 | for (i = 0; i < pd->port_rcvegrbuf_chunks; i++, start += size) { |
1047 | ret = remap_pfn_range(vma, start, phys[i] >> PAGE_SHIFT, | |
1048 | size, vma->vm_page_prot); | |
1049 | if (ret < 0) | |
1050 | goto bail; | |
1051 | } | |
1052 | ret = 0; | |
1053 | ||
1054 | bail: | |
1055 | return ret; | |
1056 | } | |
1057 | ||
7f510b46 BS |
1058 | /** |
1059 | * ipath_mmap - mmap various structures into user space | |
1060 | * @fp: the file pointer | |
1061 | * @vma: the VM area | |
1062 | * | |
1063 | * We use this to have a shared buffer between the kernel and the user code | |
1064 | * for the rcvhdr queue, egr buffers, and the per-port user regs and pio | |
1065 | * buffers in the chip. We have the open and close entries so we can bump | |
1066 | * the ref count and keep the driver from being unloaded while still mapped. | |
1067 | */ | |
1068 | static int ipath_mmap(struct file *fp, struct vm_area_struct *vma) | |
1069 | { | |
1070 | struct ipath_portdata *pd; | |
1071 | struct ipath_devdata *dd; | |
1072 | u64 pgaddr, ureg; | |
1073 | int ret; | |
1074 | ||
1075 | pd = port_fp(fp); | |
1076 | dd = pd->port_dd; | |
f37bda92 | 1077 | |
7f510b46 BS |
1078 | /* |
1079 | * This is the ipath_do_user_init() code, mapping the shared buffers | |
1080 | * into the user process. The address referred to by vm_pgoff is the | |
1081 | * virtual, not physical, address; we only do one mmap for each | |
1082 | * space mapped. | |
1083 | */ | |
1084 | pgaddr = vma->vm_pgoff << PAGE_SHIFT; | |
1085 | ||
1086 | /* | |
f37bda92 BS |
1087 | * Must fit in 40 bits for our hardware; some checked elsewhere, |
1088 | * but we'll be paranoid. Check for 0 is mostly in case one of the | |
1089 | * allocations failed, but user called mmap anyway. We want to catch | |
1090 | * that before it can match. | |
7f510b46 | 1091 | */ |
f37bda92 BS |
1092 | if (!pgaddr || pgaddr >= (1ULL<<40)) { |
1093 | ipath_dev_err(dd, "Bad phys addr %llx, start %lx, end %lx\n", | |
1094 | (unsigned long long)pgaddr, vma->vm_start, vma->vm_end); | |
1095 | return -EINVAL; | |
1096 | } | |
7f510b46 | 1097 | |
f37bda92 | 1098 | /* just the offset of the port user registers, not physical addr */ |
7f510b46 BS |
1099 | ureg = dd->ipath_uregbase + dd->ipath_palign * pd->port_port; |
1100 | ||
f37bda92 | 1101 | ipath_cdbg(MM, "ushare: pgaddr %llx vm_start=%lx, vmlen %lx\n", |
7f510b46 | 1102 | (unsigned long long) pgaddr, vma->vm_start, |
f37bda92 | 1103 | vma->vm_end - vma->vm_start); |
7f510b46 | 1104 | |
f37bda92 BS |
1105 | if (vma->vm_start & (PAGE_SIZE-1)) { |
1106 | ipath_dev_err(dd, | |
1107 | "vm_start not aligned: %lx, end=%lx phys %lx\n", | |
1108 | vma->vm_start, vma->vm_end, (unsigned long)pgaddr); | |
1109 | ret = -EINVAL; | |
1110 | } | |
1111 | else if (pgaddr == ureg) | |
7f510b46 BS |
1112 | ret = mmap_ureg(vma, dd, ureg); |
1113 | else if (pgaddr == pd->port_piobufs) | |
1114 | ret = mmap_piobufs(vma, dd, pd); | |
1115 | else if (pgaddr == (u64) pd->port_rcvegr_phys) | |
1116 | ret = mmap_rcvegrbufs(vma, pd); | |
f37bda92 BS |
1117 | else if (pgaddr == (u64) pd->port_rcvhdrq_phys) { |
1118 | /* | |
1119 | * The rcvhdrq itself; readonly except on HT-400 (so have | |
1120 | * to allow writable mapping), multiple pages, contiguous | |
1121 | * from an i/o perspective. | |
1122 | */ | |
1123 | unsigned total_size = | |
1124 | ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize | |
1125 | * sizeof(u32), PAGE_SIZE); | |
1126 | ret = ipath_mmap_mem(vma, pd, total_size, 1, | |
1127 | pd->port_rcvhdrq_phys, | |
1128 | "rcvhdrq"); | |
1129 | } | |
1130 | else if (pgaddr == (u64)pd->port_rcvhdrqtailaddr_phys) | |
1131 | /* in-memory copy of rcvhdrq tail register */ | |
1132 | ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0, | |
1133 | pd->port_rcvhdrqtailaddr_phys, | |
1134 | "rcvhdrq tail"); | |
7f510b46 | 1135 | else if (pgaddr == dd->ipath_pioavailregs_phys) |
f37bda92 BS |
1136 | /* in-memory copy of pioavail registers */ |
1137 | ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0, | |
1138 | dd->ipath_pioavailregs_phys, | |
1139 | "pioavail registers"); | |
7f510b46 BS |
1140 | else |
1141 | ret = -EINVAL; | |
1142 | ||
1143 | vma->vm_private_data = NULL; | |
1144 | ||
1145 | if (ret < 0) | |
1146 | dev_info(&dd->pcidev->dev, | |
1147 | "Failure %d on addr %lx, off %lx\n", | |
1148 | -ret, vma->vm_start, vma->vm_pgoff); | |
1149 | ||
1150 | return ret; | |
1151 | } | |
1152 | ||
1153 | static unsigned int ipath_poll(struct file *fp, | |
1154 | struct poll_table_struct *pt) | |
1155 | { | |
1156 | struct ipath_portdata *pd; | |
1157 | u32 head, tail; | |
1158 | int bit; | |
1159 | struct ipath_devdata *dd; | |
1160 | ||
1161 | pd = port_fp(fp); | |
1162 | dd = pd->port_dd; | |
1163 | ||
1164 | bit = pd->port_port + INFINIPATH_R_INTRAVAIL_SHIFT; | |
1165 | set_bit(bit, &dd->ipath_rcvctrl); | |
1166 | ||
1167 | /* | |
1168 | * Before blocking, make sure that head is still == tail, | |
1169 | * reading from the chip, so we can be sure the interrupt | |
1170 | * enable has made it to the chip. If not equal, disable | |
1171 | * interrupt again and return immediately. This avoids races, | |
1172 | * and the overhead of the chip read doesn't matter much at | |
1173 | * this point, since we are waiting for something anyway. | |
1174 | */ | |
1175 | ||
1176 | ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl, | |
1177 | dd->ipath_rcvctrl); | |
1178 | ||
1179 | head = ipath_read_ureg32(dd, ur_rcvhdrhead, pd->port_port); | |
1180 | tail = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port); | |
1181 | ||
1182 | if (tail == head) { | |
1183 | set_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag); | |
9dcc0e58 BS |
1184 | if(dd->ipath_rhdrhead_intr_off) /* arm rcv interrupt */ |
1185 | (void)ipath_write_ureg(dd, ur_rcvhdrhead, | |
1186 | dd->ipath_rhdrhead_intr_off | |
1187 | | head, pd->port_port); | |
7f510b46 BS |
1188 | poll_wait(fp, &pd->port_wait, pt); |
1189 | ||
1190 | if (test_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag)) { | |
1191 | /* timed out, no packets received */ | |
1192 | clear_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag); | |
1193 | pd->port_rcvwait_to++; | |
1194 | } | |
1195 | } | |
1196 | else { | |
1197 | /* it's already happened; don't do wait_event overhead */ | |
1198 | pd->port_rcvnowait++; | |
1199 | } | |
1200 | ||
1201 | clear_bit(bit, &dd->ipath_rcvctrl); | |
1202 | ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl, | |
1203 | dd->ipath_rcvctrl); | |
1204 | ||
1205 | return 0; | |
1206 | } | |
1207 | ||
1208 | static int try_alloc_port(struct ipath_devdata *dd, int port, | |
1209 | struct file *fp) | |
1210 | { | |
1211 | int ret; | |
1212 | ||
1213 | if (!dd->ipath_pd[port]) { | |
1214 | void *p, *ptmp; | |
1215 | ||
1216 | p = kzalloc(sizeof(struct ipath_portdata), GFP_KERNEL); | |
1217 | ||
1218 | /* | |
1219 | * Allocate memory for use in ipath_tid_update() just once | |
1220 | * at open, not per call. Reduces cost of expected send | |
1221 | * setup. | |
1222 | */ | |
1223 | ptmp = kmalloc(dd->ipath_rcvtidcnt * sizeof(u16) + | |
1224 | dd->ipath_rcvtidcnt * sizeof(struct page **), | |
1225 | GFP_KERNEL); | |
1226 | if (!p || !ptmp) { | |
1227 | ipath_dev_err(dd, "Unable to allocate portdata " | |
1228 | "memory, failing open\n"); | |
1229 | ret = -ENOMEM; | |
1230 | kfree(p); | |
1231 | kfree(ptmp); | |
1232 | goto bail; | |
1233 | } | |
1234 | dd->ipath_pd[port] = p; | |
1235 | dd->ipath_pd[port]->port_port = port; | |
1236 | dd->ipath_pd[port]->port_dd = dd; | |
1237 | dd->ipath_pd[port]->port_tid_pg_list = ptmp; | |
1238 | init_waitqueue_head(&dd->ipath_pd[port]->port_wait); | |
1239 | } | |
1240 | if (!dd->ipath_pd[port]->port_cnt) { | |
1241 | dd->ipath_pd[port]->port_cnt = 1; | |
1242 | fp->private_data = (void *) dd->ipath_pd[port]; | |
1243 | ipath_cdbg(PROC, "%s[%u] opened unit:port %u:%u\n", | |
1244 | current->comm, current->pid, dd->ipath_unit, | |
1245 | port); | |
1246 | dd->ipath_pd[port]->port_pid = current->pid; | |
1247 | strncpy(dd->ipath_pd[port]->port_comm, current->comm, | |
1248 | sizeof(dd->ipath_pd[port]->port_comm)); | |
1249 | ipath_stats.sps_ports++; | |
1250 | ret = 0; | |
1251 | goto bail; | |
1252 | } | |
1253 | ret = -EBUSY; | |
1254 | ||
1255 | bail: | |
1256 | return ret; | |
1257 | } | |
1258 | ||
1259 | static inline int usable(struct ipath_devdata *dd) | |
1260 | { | |
1261 | return dd && | |
1262 | (dd->ipath_flags & IPATH_PRESENT) && | |
1263 | dd->ipath_kregbase && | |
1264 | dd->ipath_lid && | |
1265 | !(dd->ipath_flags & (IPATH_LINKDOWN | IPATH_DISABLED | |
1266 | | IPATH_LINKUNK)); | |
1267 | } | |
1268 | ||
1269 | static int find_free_port(int unit, struct file *fp) | |
1270 | { | |
1271 | struct ipath_devdata *dd = ipath_lookup(unit); | |
1272 | int ret, i; | |
1273 | ||
1274 | if (!dd) { | |
1275 | ret = -ENODEV; | |
1276 | goto bail; | |
1277 | } | |
1278 | ||
1279 | if (!usable(dd)) { | |
1280 | ret = -ENETDOWN; | |
1281 | goto bail; | |
1282 | } | |
1283 | ||
1284 | for (i = 0; i < dd->ipath_cfgports; i++) { | |
1285 | ret = try_alloc_port(dd, i, fp); | |
1286 | if (ret != -EBUSY) | |
1287 | goto bail; | |
1288 | } | |
1289 | ret = -EBUSY; | |
1290 | ||
1291 | bail: | |
1292 | return ret; | |
1293 | } | |
1294 | ||
1295 | static int find_best_unit(struct file *fp) | |
1296 | { | |
1297 | int ret = 0, i, prefunit = -1, devmax; | |
1298 | int maxofallports, npresent, nup; | |
1299 | int ndev; | |
1300 | ||
1301 | (void) ipath_count_units(&npresent, &nup, &maxofallports); | |
1302 | ||
1303 | /* | |
1304 | * This code is present to allow a knowledgeable person to | |
1305 | * specify the layout of processes to processors before opening | |
1306 | * this driver, and then we'll assign the process to the "closest" | |
1307 | * HT-400 to that processor (we assume reasonable connectivity, | |
1308 | * for now). This code assumes that if affinity has been set | |
1309 | * before this point, that at most one cpu is set; for now this | |
1310 | * is reasonable. I check for both cpus_empty() and cpus_full(), | |
1311 | * in case some kernel variant sets none of the bits when no | |
1312 | * affinity is set. 2.6.11 and 12 kernels have all present | |
1313 | * cpus set. Some day we'll have to fix it up further to handle | |
1314 | * a cpu subset. This algorithm fails for two HT-400's connected | |
1315 | * in tunnel fashion. Eventually this needs real topology | |
1316 | * information. There may be some issues with dual core numbering | |
1317 | * as well. This needs more work prior to release. | |
1318 | */ | |
1319 | if (!cpus_empty(current->cpus_allowed) && | |
1320 | !cpus_full(current->cpus_allowed)) { | |
1321 | int ncpus = num_online_cpus(), curcpu = -1; | |
1322 | for (i = 0; i < ncpus; i++) | |
1323 | if (cpu_isset(i, current->cpus_allowed)) { | |
1324 | ipath_cdbg(PROC, "%s[%u] affinity set for " | |
1325 | "cpu %d\n", current->comm, | |
1326 | current->pid, i); | |
1327 | curcpu = i; | |
1328 | } | |
1329 | if (curcpu != -1) { | |
1330 | if (npresent) { | |
1331 | prefunit = curcpu / (ncpus / npresent); | |
1332 | ipath_dbg("%s[%u] %d chips, %d cpus, " | |
1333 | "%d cpus/chip, select unit %d\n", | |
1334 | current->comm, current->pid, | |
1335 | npresent, ncpus, ncpus / npresent, | |
1336 | prefunit); | |
1337 | } | |
1338 | } | |
1339 | } | |
1340 | ||
1341 | /* | |
1342 | * user ports start at 1, kernel port is 0 | |
1343 | * For now, we do round-robin access across all chips | |
1344 | */ | |
1345 | ||
1346 | if (prefunit != -1) | |
1347 | devmax = prefunit + 1; | |
1348 | else | |
1349 | devmax = ipath_count_units(NULL, NULL, NULL); | |
1350 | recheck: | |
1351 | for (i = 1; i < maxofallports; i++) { | |
1352 | for (ndev = prefunit != -1 ? prefunit : 0; ndev < devmax; | |
1353 | ndev++) { | |
1354 | struct ipath_devdata *dd = ipath_lookup(ndev); | |
1355 | ||
1356 | if (!usable(dd)) | |
1357 | continue; /* can't use this unit */ | |
1358 | if (i >= dd->ipath_cfgports) | |
1359 | /* | |
1360 | * Maxed out on users of this unit. Try | |
1361 | * next. | |
1362 | */ | |
1363 | continue; | |
1364 | ret = try_alloc_port(dd, i, fp); | |
1365 | if (!ret) | |
1366 | goto done; | |
1367 | } | |
1368 | } | |
1369 | ||
1370 | if (npresent) { | |
1371 | if (nup == 0) { | |
1372 | ret = -ENETDOWN; | |
1373 | ipath_dbg("No ports available (none initialized " | |
1374 | "and ready)\n"); | |
1375 | } else { | |
1376 | if (prefunit > 0) { | |
1377 | /* if started above 0, retry from 0 */ | |
1378 | ipath_cdbg(PROC, | |
1379 | "%s[%u] no ports on prefunit " | |
1380 | "%d, clear and re-check\n", | |
1381 | current->comm, current->pid, | |
1382 | prefunit); | |
1383 | devmax = ipath_count_units(NULL, NULL, | |
1384 | NULL); | |
1385 | prefunit = -1; | |
1386 | goto recheck; | |
1387 | } | |
1388 | ret = -EBUSY; | |
1389 | ipath_dbg("No ports available\n"); | |
1390 | } | |
1391 | } else { | |
1392 | ret = -ENXIO; | |
1393 | ipath_dbg("No boards found\n"); | |
1394 | } | |
1395 | ||
1396 | done: | |
1397 | return ret; | |
1398 | } | |
1399 | ||
1400 | static int ipath_open(struct inode *in, struct file *fp) | |
1401 | { | |
a2acb2ff | 1402 | int ret, user_minor; |
7f510b46 BS |
1403 | |
1404 | mutex_lock(&ipath_mutex); | |
1405 | ||
a2acb2ff | 1406 | user_minor = iminor(in) - IPATH_USER_MINOR_BASE; |
7f510b46 | 1407 | ipath_cdbg(VERBOSE, "open on dev %lx (minor %d)\n", |
a2acb2ff | 1408 | (long)in->i_rdev, user_minor); |
7f510b46 | 1409 | |
a2acb2ff BS |
1410 | if (user_minor) |
1411 | ret = find_free_port(user_minor - 1, fp); | |
7f510b46 BS |
1412 | else |
1413 | ret = find_best_unit(fp); | |
1414 | ||
1415 | mutex_unlock(&ipath_mutex); | |
1416 | return ret; | |
1417 | } | |
1418 | ||
1419 | /** | |
1420 | * unlock_exptid - unlock any expected TID entries port still had in use | |
1421 | * @pd: port | |
1422 | * | |
1423 | * We don't actually update the chip here, because we do a bulk update | |
1424 | * below, using ipath_f_clear_tids. | |
1425 | */ | |
1426 | static void unlock_expected_tids(struct ipath_portdata *pd) | |
1427 | { | |
1428 | struct ipath_devdata *dd = pd->port_dd; | |
1429 | int port_tidbase = pd->port_port * dd->ipath_rcvtidcnt; | |
1430 | int i, cnt = 0, maxtid = port_tidbase + dd->ipath_rcvtidcnt; | |
1431 | ||
1432 | ipath_cdbg(VERBOSE, "Port %u unlocking any locked expTID pages\n", | |
1433 | pd->port_port); | |
1434 | for (i = port_tidbase; i < maxtid; i++) { | |
1435 | if (!dd->ipath_pageshadow[i]) | |
1436 | continue; | |
1437 | ||
1438 | ipath_release_user_pages_on_close(&dd->ipath_pageshadow[i], | |
1439 | 1); | |
1440 | dd->ipath_pageshadow[i] = NULL; | |
1441 | cnt++; | |
1442 | ipath_stats.sps_pageunlocks++; | |
1443 | } | |
1444 | if (cnt) | |
1445 | ipath_cdbg(VERBOSE, "Port %u locked %u expTID entries\n", | |
1446 | pd->port_port, cnt); | |
1447 | ||
1448 | if (ipath_stats.sps_pagelocks || ipath_stats.sps_pageunlocks) | |
1449 | ipath_cdbg(VERBOSE, "%llu pages locked, %llu unlocked\n", | |
1450 | (unsigned long long) ipath_stats.sps_pagelocks, | |
1451 | (unsigned long long) | |
1452 | ipath_stats.sps_pageunlocks); | |
1453 | } | |
1454 | ||
1455 | static int ipath_close(struct inode *in, struct file *fp) | |
1456 | { | |
1457 | int ret = 0; | |
1458 | struct ipath_portdata *pd; | |
1459 | struct ipath_devdata *dd; | |
1460 | unsigned port; | |
1461 | ||
1462 | ipath_cdbg(VERBOSE, "close on dev %lx, private data %p\n", | |
1463 | (long)in->i_rdev, fp->private_data); | |
1464 | ||
1465 | mutex_lock(&ipath_mutex); | |
1466 | ||
1467 | pd = port_fp(fp); | |
1468 | port = pd->port_port; | |
1469 | fp->private_data = NULL; | |
1470 | dd = pd->port_dd; | |
1471 | ||
1472 | if (pd->port_hdrqfull) { | |
1473 | ipath_cdbg(PROC, "%s[%u] had %u rcvhdrqfull errors " | |
1474 | "during run\n", pd->port_comm, pd->port_pid, | |
1475 | pd->port_hdrqfull); | |
1476 | pd->port_hdrqfull = 0; | |
1477 | } | |
1478 | ||
1479 | if (pd->port_rcvwait_to || pd->port_piowait_to | |
1480 | || pd->port_rcvnowait || pd->port_pionowait) { | |
1481 | ipath_cdbg(VERBOSE, "port%u, %u rcv, %u pio wait timeo; " | |
1482 | "%u rcv %u, pio already\n", | |
1483 | pd->port_port, pd->port_rcvwait_to, | |
1484 | pd->port_piowait_to, pd->port_rcvnowait, | |
1485 | pd->port_pionowait); | |
1486 | pd->port_rcvwait_to = pd->port_piowait_to = | |
1487 | pd->port_rcvnowait = pd->port_pionowait = 0; | |
1488 | } | |
1489 | if (pd->port_flag) { | |
1490 | ipath_dbg("port %u port_flag still set to 0x%lx\n", | |
1491 | pd->port_port, pd->port_flag); | |
1492 | pd->port_flag = 0; | |
1493 | } | |
1494 | ||
1495 | if (dd->ipath_kregbase) { | |
35783ec0 BS |
1496 | int i; |
1497 | /* atomically clear receive enable port. */ | |
1498 | clear_bit(INFINIPATH_R_PORTENABLE_SHIFT + port, | |
1499 | &dd->ipath_rcvctrl); | |
1500 | ipath_write_kreg( dd, dd->ipath_kregs->kr_rcvctrl, | |
1501 | dd->ipath_rcvctrl); | |
1502 | /* and read back from chip to be sure that nothing | |
1503 | * else is in flight when we do the rest */ | |
1504 | (void)ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch); | |
7f510b46 BS |
1505 | |
1506 | /* clean up the pkeys for this port user */ | |
1507 | ipath_clean_part_key(pd, dd); | |
1508 | ||
7f510b46 | 1509 | |
35783ec0 BS |
1510 | /* |
1511 | * be paranoid, and never write 0's to these, just use an | |
1512 | * unused part of the port 0 tail page. Of course, | |
1513 | * rcvhdraddr points to a large chunk of memory, so this | |
1514 | * could still trash things, but at least it won't trash | |
1515 | * page 0, and by disabling the port, it should stop "soon", | |
1516 | * even if a packet or two is in already in flight after we | |
1517 | * disabled the port. | |
1518 | */ | |
1519 | ipath_write_kreg_port(dd, | |
1520 | dd->ipath_kregs->kr_rcvhdrtailaddr, port, | |
1521 | dd->ipath_dummy_hdrq_phys); | |
1522 | ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdraddr, | |
1523 | pd->port_port, dd->ipath_dummy_hdrq_phys); | |
1524 | ||
1525 | i = dd->ipath_pbufsport * (port - 1); | |
1526 | ipath_disarm_piobufs(dd, i, dd->ipath_pbufsport); | |
1527 | ||
1528 | if (dd->ipath_pageshadow) | |
1529 | unlock_expected_tids(pd); | |
1530 | ipath_stats.sps_ports--; | |
1531 | ipath_cdbg(PROC, "%s[%u] closed port %u:%u\n", | |
1532 | pd->port_comm, pd->port_pid, | |
1533 | dd->ipath_unit, port); | |
1534 | ||
1535 | dd->ipath_f_clear_tids(dd, pd->port_port); | |
7f510b46 BS |
1536 | } |
1537 | ||
1538 | pd->port_cnt = 0; | |
1539 | pd->port_pid = 0; | |
1540 | ||
f37bda92 | 1541 | dd->ipath_pd[pd->port_port] = NULL; /* before releasing mutex */ |
7f510b46 | 1542 | mutex_unlock(&ipath_mutex); |
f37bda92 | 1543 | ipath_free_pddata(dd, pd); /* after releasing the mutex */ |
7f510b46 BS |
1544 | |
1545 | return ret; | |
1546 | } | |
1547 | ||
1548 | static int ipath_port_info(struct ipath_portdata *pd, | |
1549 | struct ipath_port_info __user *uinfo) | |
1550 | { | |
1551 | struct ipath_port_info info; | |
1552 | int nup; | |
1553 | int ret; | |
1554 | ||
1555 | (void) ipath_count_units(NULL, &nup, NULL); | |
1556 | info.num_active = nup; | |
1557 | info.unit = pd->port_dd->ipath_unit; | |
1558 | info.port = pd->port_port; | |
1559 | ||
1560 | if (copy_to_user(uinfo, &info, sizeof(info))) { | |
1561 | ret = -EFAULT; | |
1562 | goto bail; | |
1563 | } | |
1564 | ret = 0; | |
1565 | ||
1566 | bail: | |
1567 | return ret; | |
1568 | } | |
1569 | ||
1570 | static ssize_t ipath_write(struct file *fp, const char __user *data, | |
1571 | size_t count, loff_t *off) | |
1572 | { | |
1573 | const struct ipath_cmd __user *ucmd; | |
1574 | struct ipath_portdata *pd; | |
1575 | const void __user *src; | |
1576 | size_t consumed, copy; | |
1577 | struct ipath_cmd cmd; | |
1578 | ssize_t ret = 0; | |
1579 | void *dest; | |
1580 | ||
1581 | if (count < sizeof(cmd.type)) { | |
1582 | ret = -EINVAL; | |
1583 | goto bail; | |
1584 | } | |
1585 | ||
1586 | ucmd = (const struct ipath_cmd __user *) data; | |
1587 | ||
1588 | if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) { | |
1589 | ret = -EFAULT; | |
1590 | goto bail; | |
1591 | } | |
1592 | ||
1593 | consumed = sizeof(cmd.type); | |
1594 | ||
1595 | switch (cmd.type) { | |
1596 | case IPATH_CMD_USER_INIT: | |
1597 | copy = sizeof(cmd.cmd.user_info); | |
1598 | dest = &cmd.cmd.user_info; | |
1599 | src = &ucmd->cmd.user_info; | |
1600 | break; | |
1601 | case IPATH_CMD_RECV_CTRL: | |
1602 | copy = sizeof(cmd.cmd.recv_ctrl); | |
1603 | dest = &cmd.cmd.recv_ctrl; | |
1604 | src = &ucmd->cmd.recv_ctrl; | |
1605 | break; | |
1606 | case IPATH_CMD_PORT_INFO: | |
1607 | copy = sizeof(cmd.cmd.port_info); | |
1608 | dest = &cmd.cmd.port_info; | |
1609 | src = &ucmd->cmd.port_info; | |
1610 | break; | |
1611 | case IPATH_CMD_TID_UPDATE: | |
1612 | case IPATH_CMD_TID_FREE: | |
1613 | copy = sizeof(cmd.cmd.tid_info); | |
1614 | dest = &cmd.cmd.tid_info; | |
1615 | src = &ucmd->cmd.tid_info; | |
1616 | break; | |
1617 | case IPATH_CMD_SET_PART_KEY: | |
1618 | copy = sizeof(cmd.cmd.part_key); | |
1619 | dest = &cmd.cmd.part_key; | |
1620 | src = &ucmd->cmd.part_key; | |
1621 | break; | |
1622 | default: | |
1623 | ret = -EINVAL; | |
1624 | goto bail; | |
1625 | } | |
1626 | ||
1627 | if ((count - consumed) < copy) { | |
1628 | ret = -EINVAL; | |
1629 | goto bail; | |
1630 | } | |
1631 | ||
1632 | if (copy_from_user(dest, src, copy)) { | |
1633 | ret = -EFAULT; | |
1634 | goto bail; | |
1635 | } | |
1636 | ||
1637 | consumed += copy; | |
1638 | pd = port_fp(fp); | |
1639 | ||
1640 | switch (cmd.type) { | |
1641 | case IPATH_CMD_USER_INIT: | |
1642 | ret = ipath_do_user_init(pd, &cmd.cmd.user_info); | |
1643 | if (ret < 0) | |
1644 | goto bail; | |
1645 | ret = ipath_get_base_info( | |
1646 | pd, (void __user *) (unsigned long) | |
1647 | cmd.cmd.user_info.spu_base_info, | |
1648 | cmd.cmd.user_info.spu_base_info_size); | |
1649 | break; | |
1650 | case IPATH_CMD_RECV_CTRL: | |
1651 | ret = ipath_manage_rcvq(pd, cmd.cmd.recv_ctrl); | |
1652 | break; | |
1653 | case IPATH_CMD_PORT_INFO: | |
1654 | ret = ipath_port_info(pd, | |
1655 | (struct ipath_port_info __user *) | |
1656 | (unsigned long) cmd.cmd.port_info); | |
1657 | break; | |
1658 | case IPATH_CMD_TID_UPDATE: | |
1659 | ret = ipath_tid_update(pd, &cmd.cmd.tid_info); | |
1660 | break; | |
1661 | case IPATH_CMD_TID_FREE: | |
1662 | ret = ipath_tid_free(pd, &cmd.cmd.tid_info); | |
1663 | break; | |
1664 | case IPATH_CMD_SET_PART_KEY: | |
1665 | ret = ipath_set_part_key(pd, cmd.cmd.part_key); | |
1666 | break; | |
1667 | } | |
1668 | ||
1669 | if (ret >= 0) | |
1670 | ret = consumed; | |
1671 | ||
1672 | bail: | |
1673 | return ret; | |
1674 | } | |
1675 | ||
1676 | static struct class *ipath_class; | |
1677 | ||
1678 | static int init_cdev(int minor, char *name, struct file_operations *fops, | |
1679 | struct cdev **cdevp, struct class_device **class_devp) | |
1680 | { | |
1681 | const dev_t dev = MKDEV(IPATH_MAJOR, minor); | |
1682 | struct cdev *cdev = NULL; | |
1683 | struct class_device *class_dev = NULL; | |
1684 | int ret; | |
1685 | ||
1686 | cdev = cdev_alloc(); | |
1687 | if (!cdev) { | |
1688 | printk(KERN_ERR IPATH_DRV_NAME | |
1689 | ": Could not allocate cdev for minor %d, %s\n", | |
1690 | minor, name); | |
1691 | ret = -ENOMEM; | |
1692 | goto done; | |
1693 | } | |
1694 | ||
1695 | cdev->owner = THIS_MODULE; | |
1696 | cdev->ops = fops; | |
1697 | kobject_set_name(&cdev->kobj, name); | |
1698 | ||
1699 | ret = cdev_add(cdev, dev, 1); | |
1700 | if (ret < 0) { | |
1701 | printk(KERN_ERR IPATH_DRV_NAME | |
1702 | ": Could not add cdev for minor %d, %s (err %d)\n", | |
1703 | minor, name, -ret); | |
1704 | goto err_cdev; | |
1705 | } | |
1706 | ||
1707 | class_dev = class_device_create(ipath_class, NULL, dev, NULL, name); | |
1708 | ||
1709 | if (IS_ERR(class_dev)) { | |
1710 | ret = PTR_ERR(class_dev); | |
1711 | printk(KERN_ERR IPATH_DRV_NAME ": Could not create " | |
1712 | "class_dev for minor %d, %s (err %d)\n", | |
1713 | minor, name, -ret); | |
1714 | goto err_cdev; | |
1715 | } | |
1716 | ||
1717 | goto done; | |
1718 | ||
1719 | err_cdev: | |
1720 | cdev_del(cdev); | |
1721 | cdev = NULL; | |
1722 | ||
1723 | done: | |
1724 | if (ret >= 0) { | |
1725 | *cdevp = cdev; | |
1726 | *class_devp = class_dev; | |
1727 | } else { | |
1728 | *cdevp = NULL; | |
1729 | *class_devp = NULL; | |
1730 | } | |
1731 | ||
1732 | return ret; | |
1733 | } | |
1734 | ||
1735 | int ipath_cdev_init(int minor, char *name, struct file_operations *fops, | |
1736 | struct cdev **cdevp, struct class_device **class_devp) | |
1737 | { | |
1738 | return init_cdev(minor, name, fops, cdevp, class_devp); | |
1739 | } | |
1740 | ||
1741 | static void cleanup_cdev(struct cdev **cdevp, | |
1742 | struct class_device **class_devp) | |
1743 | { | |
1744 | struct class_device *class_dev = *class_devp; | |
1745 | ||
1746 | if (class_dev) { | |
1747 | class_device_unregister(class_dev); | |
1748 | *class_devp = NULL; | |
1749 | } | |
1750 | ||
1751 | if (*cdevp) { | |
1752 | cdev_del(*cdevp); | |
1753 | *cdevp = NULL; | |
1754 | } | |
1755 | } | |
1756 | ||
1757 | void ipath_cdev_cleanup(struct cdev **cdevp, | |
1758 | struct class_device **class_devp) | |
1759 | { | |
1760 | cleanup_cdev(cdevp, class_devp); | |
1761 | } | |
1762 | ||
1763 | static struct cdev *wildcard_cdev; | |
1764 | static struct class_device *wildcard_class_dev; | |
1765 | ||
1766 | static const dev_t dev = MKDEV(IPATH_MAJOR, 0); | |
1767 | ||
1768 | static int user_init(void) | |
1769 | { | |
1770 | int ret; | |
1771 | ||
1772 | ret = register_chrdev_region(dev, IPATH_NMINORS, IPATH_DRV_NAME); | |
1773 | if (ret < 0) { | |
1774 | printk(KERN_ERR IPATH_DRV_NAME ": Could not register " | |
1775 | "chrdev region (err %d)\n", -ret); | |
1776 | goto done; | |
1777 | } | |
1778 | ||
1779 | ipath_class = class_create(THIS_MODULE, IPATH_DRV_NAME); | |
1780 | ||
1781 | if (IS_ERR(ipath_class)) { | |
1782 | ret = PTR_ERR(ipath_class); | |
1783 | printk(KERN_ERR IPATH_DRV_NAME ": Could not create " | |
1784 | "device class (err %d)\n", -ret); | |
1785 | goto bail; | |
1786 | } | |
1787 | ||
1788 | goto done; | |
1789 | bail: | |
1790 | unregister_chrdev_region(dev, IPATH_NMINORS); | |
1791 | done: | |
1792 | return ret; | |
1793 | } | |
1794 | ||
1795 | static void user_cleanup(void) | |
1796 | { | |
1797 | if (ipath_class) { | |
1798 | class_destroy(ipath_class); | |
1799 | ipath_class = NULL; | |
1800 | } | |
1801 | ||
1802 | unregister_chrdev_region(dev, IPATH_NMINORS); | |
1803 | } | |
1804 | ||
1805 | static atomic_t user_count = ATOMIC_INIT(0); | |
1806 | static atomic_t user_setup = ATOMIC_INIT(0); | |
1807 | ||
1808 | int ipath_user_add(struct ipath_devdata *dd) | |
1809 | { | |
1810 | char name[10]; | |
1811 | int ret; | |
1812 | ||
1813 | if (atomic_inc_return(&user_count) == 1) { | |
1814 | ret = user_init(); | |
1815 | if (ret < 0) { | |
1816 | ipath_dev_err(dd, "Unable to set up user support: " | |
1817 | "error %d\n", -ret); | |
1818 | goto bail; | |
1819 | } | |
7f510b46 BS |
1820 | ret = init_cdev(0, "ipath", &ipath_file_ops, &wildcard_cdev, |
1821 | &wildcard_class_dev); | |
1822 | if (ret < 0) { | |
1823 | ipath_dev_err(dd, "Could not create wildcard " | |
1824 | "minor: error %d\n", -ret); | |
a2acb2ff | 1825 | goto bail_sma; |
7f510b46 BS |
1826 | } |
1827 | ||
1828 | atomic_set(&user_setup, 1); | |
1829 | } | |
1830 | ||
1831 | snprintf(name, sizeof(name), "ipath%d", dd->ipath_unit); | |
1832 | ||
1833 | ret = init_cdev(dd->ipath_unit + 1, name, &ipath_file_ops, | |
a2acb2ff | 1834 | &dd->user_cdev, &dd->user_class_dev); |
7f510b46 BS |
1835 | if (ret < 0) |
1836 | ipath_dev_err(dd, "Could not create user minor %d, %s\n", | |
1837 | dd->ipath_unit + 1, name); | |
1838 | ||
1839 | goto bail; | |
1840 | ||
7f510b46 BS |
1841 | bail_sma: |
1842 | user_cleanup(); | |
1843 | bail: | |
1844 | return ret; | |
1845 | } | |
1846 | ||
a2acb2ff | 1847 | void ipath_user_remove(struct ipath_devdata *dd) |
7f510b46 | 1848 | { |
a2acb2ff | 1849 | cleanup_cdev(&dd->user_cdev, &dd->user_class_dev); |
7f510b46 BS |
1850 | |
1851 | if (atomic_dec_return(&user_count) == 0) { | |
1852 | if (atomic_read(&user_setup) == 0) | |
1853 | goto bail; | |
1854 | ||
1855 | cleanup_cdev(&wildcard_cdev, &wildcard_class_dev); | |
7f510b46 BS |
1856 | user_cleanup(); |
1857 | ||
1858 | atomic_set(&user_setup, 0); | |
1859 | } | |
1860 | bail: | |
1861 | return; | |
1862 | } | |
f37bda92 | 1863 |