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" | |
27b678dd | 42 | #include "ipath_common.h" |
7f510b46 BS |
43 | |
44 | static int ipath_open(struct inode *, struct file *); | |
45 | static int ipath_close(struct inode *, struct file *); | |
46 | static ssize_t ipath_write(struct file *, const char __user *, size_t, | |
47 | loff_t *); | |
48 | static unsigned int ipath_poll(struct file *, struct poll_table_struct *); | |
49 | static int ipath_mmap(struct file *, struct vm_area_struct *); | |
50 | ||
2b8693c0 | 51 | static const struct file_operations ipath_file_ops = { |
7f510b46 BS |
52 | .owner = THIS_MODULE, |
53 | .write = ipath_write, | |
54 | .open = ipath_open, | |
55 | .release = ipath_close, | |
56 | .poll = ipath_poll, | |
57 | .mmap = ipath_mmap | |
58 | }; | |
59 | ||
0a5a83cf RC |
60 | /* |
61 | * Convert kernel virtual addresses to physical addresses so they don't | |
62 | * potentially conflict with the chip addresses used as mmap offsets. | |
63 | * It doesn't really matter what mmap offset we use as long as we can | |
64 | * interpret it correctly. | |
65 | */ | |
66 | static u64 cvt_kvaddr(void *p) | |
67 | { | |
68 | struct page *page; | |
69 | u64 paddr = 0; | |
70 | ||
71 | page = vmalloc_to_page(p); | |
72 | if (page) | |
73 | paddr = page_to_pfn(page) << PAGE_SHIFT; | |
74 | ||
75 | return paddr; | |
76 | } | |
77 | ||
9929b0fb | 78 | static int ipath_get_base_info(struct file *fp, |
7f510b46 BS |
79 | void __user *ubase, size_t ubase_size) |
80 | { | |
9929b0fb | 81 | struct ipath_portdata *pd = port_fp(fp); |
7f510b46 BS |
82 | int ret = 0; |
83 | struct ipath_base_info *kinfo = NULL; | |
84 | struct ipath_devdata *dd = pd->port_dd; | |
9929b0fb BS |
85 | unsigned subport_cnt; |
86 | int shared, master; | |
87 | size_t sz; | |
88 | ||
89 | subport_cnt = pd->port_subport_cnt; | |
90 | if (!subport_cnt) { | |
91 | shared = 0; | |
92 | master = 0; | |
93 | subport_cnt = 1; | |
94 | } else { | |
95 | shared = 1; | |
96 | master = !subport_fp(fp); | |
97 | } | |
7f510b46 | 98 | |
9929b0fb BS |
99 | sz = sizeof(*kinfo); |
100 | /* If port sharing is not requested, allow the old size structure */ | |
101 | if (!shared) | |
102 | sz -= 3 * sizeof(u64); | |
103 | if (ubase_size < sz) { | |
7f510b46 | 104 | ipath_cdbg(PROC, |
9929b0fb BS |
105 | "Base size %zu, need %zu (version mismatch?)\n", |
106 | ubase_size, sz); | |
7f510b46 BS |
107 | ret = -EINVAL; |
108 | goto bail; | |
109 | } | |
110 | ||
111 | kinfo = kzalloc(sizeof(*kinfo), GFP_KERNEL); | |
112 | if (kinfo == NULL) { | |
113 | ret = -ENOMEM; | |
114 | goto bail; | |
115 | } | |
116 | ||
117 | ret = dd->ipath_f_get_base_info(pd, kinfo); | |
118 | if (ret < 0) | |
119 | goto bail; | |
120 | ||
121 | kinfo->spi_rcvhdr_cnt = dd->ipath_rcvhdrcnt; | |
122 | kinfo->spi_rcvhdrent_size = dd->ipath_rcvhdrentsize; | |
123 | kinfo->spi_tidegrcnt = dd->ipath_rcvegrcnt; | |
124 | kinfo->spi_rcv_egrbufsize = dd->ipath_rcvegrbufsize; | |
125 | /* | |
126 | * have to mmap whole thing | |
127 | */ | |
128 | kinfo->spi_rcv_egrbuftotlen = | |
129 | pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size; | |
130 | kinfo->spi_rcv_egrperchunk = pd->port_rcvegrbufs_perchunk; | |
131 | kinfo->spi_rcv_egrchunksize = kinfo->spi_rcv_egrbuftotlen / | |
132 | pd->port_rcvegrbuf_chunks; | |
9929b0fb BS |
133 | kinfo->spi_tidcnt = dd->ipath_rcvtidcnt / subport_cnt; |
134 | if (master) | |
135 | kinfo->spi_tidcnt += dd->ipath_rcvtidcnt % subport_cnt; | |
7f510b46 BS |
136 | /* |
137 | * for this use, may be ipath_cfgports summed over all chips that | |
138 | * are are configured and present | |
139 | */ | |
140 | kinfo->spi_nports = dd->ipath_cfgports; | |
141 | /* unit (chip/board) our port is on */ | |
142 | kinfo->spi_unit = dd->ipath_unit; | |
143 | /* for now, only a single page */ | |
144 | kinfo->spi_tid_maxsize = PAGE_SIZE; | |
145 | ||
146 | /* | |
147 | * Doing this per port, and based on the skip value, etc. This has | |
148 | * to be the actual buffer size, since the protocol code treats it | |
149 | * as an array. | |
150 | * | |
151 | * These have to be set to user addresses in the user code via mmap. | |
152 | * These values are used on return to user code for the mmap target | |
153 | * addresses only. For 32 bit, same 44 bit address problem, so use | |
154 | * the physical address, not virtual. Before 2.6.11, using the | |
155 | * page_address() macro worked, but in 2.6.11, even that returns the | |
156 | * full 64 bit address (upper bits all 1's). So far, using the | |
157 | * physical addresses (or chip offsets, for chip mapping) works, but | |
9929b0fb BS |
158 | * no doubt some future kernel release will change that, and we'll be |
159 | * on to yet another method of dealing with this. | |
7f510b46 BS |
160 | */ |
161 | kinfo->spi_rcvhdr_base = (u64) pd->port_rcvhdrq_phys; | |
9929b0fb | 162 | kinfo->spi_rcvhdr_tailaddr = (u64) pd->port_rcvhdrqtailaddr_phys; |
7f510b46 BS |
163 | kinfo->spi_rcv_egrbufs = (u64) pd->port_rcvegr_phys; |
164 | kinfo->spi_pioavailaddr = (u64) dd->ipath_pioavailregs_phys; | |
165 | kinfo->spi_status = (u64) kinfo->spi_pioavailaddr + | |
166 | (void *) dd->ipath_statusp - | |
167 | (void *) dd->ipath_pioavailregs_dma; | |
9929b0fb BS |
168 | if (!shared) { |
169 | kinfo->spi_piocnt = dd->ipath_pbufsport; | |
170 | kinfo->spi_piobufbase = (u64) pd->port_piobufs; | |
171 | kinfo->__spi_uregbase = (u64) dd->ipath_uregbase + | |
172 | dd->ipath_palign * pd->port_port; | |
173 | } else if (master) { | |
174 | kinfo->spi_piocnt = (dd->ipath_pbufsport / subport_cnt) + | |
175 | (dd->ipath_pbufsport % subport_cnt); | |
176 | /* Master's PIO buffers are after all the slave's */ | |
177 | kinfo->spi_piobufbase = (u64) pd->port_piobufs + | |
178 | dd->ipath_palign * | |
179 | (dd->ipath_pbufsport - kinfo->spi_piocnt); | |
180 | kinfo->__spi_uregbase = (u64) dd->ipath_uregbase + | |
181 | dd->ipath_palign * pd->port_port; | |
182 | } else { | |
183 | unsigned slave = subport_fp(fp) - 1; | |
7f510b46 | 184 | |
9929b0fb BS |
185 | kinfo->spi_piocnt = dd->ipath_pbufsport / subport_cnt; |
186 | kinfo->spi_piobufbase = (u64) pd->port_piobufs + | |
187 | dd->ipath_palign * kinfo->spi_piocnt * slave; | |
0a5a83cf RC |
188 | kinfo->__spi_uregbase = cvt_kvaddr(pd->subport_uregbase + |
189 | PAGE_SIZE * slave); | |
9929b0fb | 190 | |
0a5a83cf RC |
191 | kinfo->spi_rcvhdr_base = cvt_kvaddr(pd->subport_rcvhdr_base + |
192 | pd->port_rcvhdrq_size * slave); | |
947d7617 | 193 | kinfo->spi_rcvhdr_tailaddr = 0; |
0a5a83cf RC |
194 | kinfo->spi_rcv_egrbufs = cvt_kvaddr(pd->subport_rcvegrbuf + |
195 | dd->ipath_rcvegrcnt * dd->ipath_rcvegrbufsize * slave); | |
9929b0fb BS |
196 | } |
197 | ||
198 | kinfo->spi_pioindex = (kinfo->spi_piobufbase - dd->ipath_piobufbase) / | |
199 | dd->ipath_palign; | |
7f510b46 BS |
200 | kinfo->spi_pioalign = dd->ipath_palign; |
201 | ||
202 | kinfo->spi_qpair = IPATH_KD_QP; | |
203 | kinfo->spi_piosize = dd->ipath_ibmaxlen; | |
204 | kinfo->spi_mtu = dd->ipath_ibmaxlen; /* maxlen, not ibmtu */ | |
205 | kinfo->spi_port = pd->port_port; | |
9929b0fb | 206 | kinfo->spi_subport = subport_fp(fp); |
eaf6733b | 207 | kinfo->spi_sw_version = IPATH_KERN_SWVERSION; |
7f510b46 BS |
208 | kinfo->spi_hw_version = dd->ipath_revision; |
209 | ||
9929b0fb BS |
210 | if (master) { |
211 | kinfo->spi_runtime_flags |= IPATH_RUNTIME_MASTER; | |
212 | kinfo->spi_subport_uregbase = | |
0a5a83cf | 213 | cvt_kvaddr(pd->subport_uregbase); |
9929b0fb | 214 | kinfo->spi_subport_rcvegrbuf = |
0a5a83cf | 215 | cvt_kvaddr(pd->subport_rcvegrbuf); |
9929b0fb | 216 | kinfo->spi_subport_rcvhdr_base = |
0a5a83cf | 217 | cvt_kvaddr(pd->subport_rcvhdr_base); |
9929b0fb | 218 | ipath_cdbg(PROC, "port %u flags %x %llx %llx %llx\n", |
0624b072 BS |
219 | kinfo->spi_port, kinfo->spi_runtime_flags, |
220 | (unsigned long long) kinfo->spi_subport_uregbase, | |
221 | (unsigned long long) kinfo->spi_subport_rcvegrbuf, | |
222 | (unsigned long long) kinfo->spi_subport_rcvhdr_base); | |
9929b0fb BS |
223 | } |
224 | ||
7f510b46 BS |
225 | if (copy_to_user(ubase, kinfo, sizeof(*kinfo))) |
226 | ret = -EFAULT; | |
227 | ||
228 | bail: | |
229 | kfree(kinfo); | |
230 | return ret; | |
231 | } | |
232 | ||
233 | /** | |
234 | * ipath_tid_update - update a port TID | |
235 | * @pd: the port | |
9929b0fb | 236 | * @fp: the ipath device file |
7f510b46 BS |
237 | * @ti: the TID information |
238 | * | |
239 | * The new implementation as of Oct 2004 is that the driver assigns | |
240 | * the tid and returns it to the caller. To make it easier to | |
241 | * catch bugs, and to reduce search time, we keep a cursor for | |
242 | * each port, walking the shadow tid array to find one that's not | |
243 | * in use. | |
244 | * | |
245 | * For now, if we can't allocate the full list, we fail, although | |
246 | * in the long run, we'll allocate as many as we can, and the | |
247 | * caller will deal with that by trying the remaining pages later. | |
248 | * That means that when we fail, we have to mark the tids as not in | |
249 | * use again, in our shadow copy. | |
250 | * | |
251 | * It's up to the caller to free the tids when they are done. | |
252 | * We'll unlock the pages as they free them. | |
253 | * | |
254 | * Also, right now we are locking one page at a time, but since | |
255 | * the intended use of this routine is for a single group of | |
256 | * virtually contiguous pages, that should change to improve | |
257 | * performance. | |
258 | */ | |
9929b0fb | 259 | static int ipath_tid_update(struct ipath_portdata *pd, struct file *fp, |
7f510b46 BS |
260 | const struct ipath_tid_info *ti) |
261 | { | |
262 | int ret = 0, ntids; | |
9929b0fb | 263 | u32 tid, porttid, cnt, i, tidcnt, tidoff; |
7f510b46 BS |
264 | u16 *tidlist; |
265 | struct ipath_devdata *dd = pd->port_dd; | |
266 | u64 physaddr; | |
267 | unsigned long vaddr; | |
268 | u64 __iomem *tidbase; | |
269 | unsigned long tidmap[8]; | |
270 | struct page **pagep = NULL; | |
9929b0fb | 271 | unsigned subport = subport_fp(fp); |
7f510b46 BS |
272 | |
273 | if (!dd->ipath_pageshadow) { | |
274 | ret = -ENOMEM; | |
275 | goto done; | |
276 | } | |
277 | ||
278 | cnt = ti->tidcnt; | |
279 | if (!cnt) { | |
280 | ipath_dbg("After copyin, tidcnt 0, tidlist %llx\n", | |
281 | (unsigned long long) ti->tidlist); | |
282 | /* | |
283 | * Should we treat as success? likely a bug | |
284 | */ | |
285 | ret = -EFAULT; | |
286 | goto done; | |
287 | } | |
9929b0fb BS |
288 | porttid = pd->port_port * dd->ipath_rcvtidcnt; |
289 | if (!pd->port_subport_cnt) { | |
290 | tidcnt = dd->ipath_rcvtidcnt; | |
291 | tid = pd->port_tidcursor; | |
292 | tidoff = 0; | |
293 | } else if (!subport) { | |
294 | tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) + | |
295 | (dd->ipath_rcvtidcnt % pd->port_subport_cnt); | |
296 | tidoff = dd->ipath_rcvtidcnt - tidcnt; | |
297 | porttid += tidoff; | |
298 | tid = tidcursor_fp(fp); | |
299 | } else { | |
300 | tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt; | |
301 | tidoff = tidcnt * (subport - 1); | |
302 | porttid += tidoff; | |
303 | tid = tidcursor_fp(fp); | |
304 | } | |
305 | if (cnt > tidcnt) { | |
7f510b46 BS |
306 | /* make sure it all fits in port_tid_pg_list */ |
307 | dev_info(&dd->pcidev->dev, "Process tried to allocate %u " | |
308 | "TIDs, only trying max (%u)\n", cnt, tidcnt); | |
309 | cnt = tidcnt; | |
310 | } | |
9929b0fb BS |
311 | pagep = &((struct page **) pd->port_tid_pg_list)[tidoff]; |
312 | tidlist = &((u16 *) &pagep[dd->ipath_rcvtidcnt])[tidoff]; | |
7f510b46 BS |
313 | |
314 | memset(tidmap, 0, sizeof(tidmap)); | |
7f510b46 | 315 | /* before decrement; chip actual # */ |
7f510b46 BS |
316 | ntids = tidcnt; |
317 | tidbase = (u64 __iomem *) (((char __iomem *) dd->ipath_kregbase) + | |
318 | dd->ipath_rcvtidbase + | |
319 | porttid * sizeof(*tidbase)); | |
320 | ||
321 | ipath_cdbg(VERBOSE, "Port%u %u tids, cursor %u, tidbase %p\n", | |
322 | pd->port_port, cnt, tid, tidbase); | |
323 | ||
324 | /* virtual address of first page in transfer */ | |
325 | vaddr = ti->tidvaddr; | |
326 | if (!access_ok(VERIFY_WRITE, (void __user *) vaddr, | |
327 | cnt * PAGE_SIZE)) { | |
328 | ipath_dbg("Fail vaddr %p, %u pages, !access_ok\n", | |
329 | (void *)vaddr, cnt); | |
330 | ret = -EFAULT; | |
331 | goto done; | |
332 | } | |
333 | ret = ipath_get_user_pages(vaddr, cnt, pagep); | |
334 | if (ret) { | |
335 | if (ret == -EBUSY) { | |
336 | ipath_dbg("Failed to lock addr %p, %u pages " | |
337 | "(already locked)\n", | |
338 | (void *) vaddr, cnt); | |
339 | /* | |
340 | * for now, continue, and see what happens but with | |
341 | * the new implementation, this should never happen, | |
342 | * unless perhaps the user has mpin'ed the pages | |
343 | * themselves (something we need to test) | |
344 | */ | |
345 | ret = 0; | |
346 | } else { | |
347 | dev_info(&dd->pcidev->dev, | |
348 | "Failed to lock addr %p, %u pages: " | |
349 | "errno %d\n", (void *) vaddr, cnt, -ret); | |
350 | goto done; | |
351 | } | |
352 | } | |
353 | for (i = 0; i < cnt; i++, vaddr += PAGE_SIZE) { | |
354 | for (; ntids--; tid++) { | |
355 | if (tid == tidcnt) | |
356 | tid = 0; | |
357 | if (!dd->ipath_pageshadow[porttid + tid]) | |
358 | break; | |
359 | } | |
360 | if (ntids < 0) { | |
361 | /* | |
362 | * oops, wrapped all the way through their TIDs, | |
363 | * and didn't have enough free; see comments at | |
364 | * start of routine | |
365 | */ | |
366 | ipath_dbg("Not enough free TIDs for %u pages " | |
367 | "(index %d), failing\n", cnt, i); | |
368 | i--; /* last tidlist[i] not filled in */ | |
369 | ret = -ENOMEM; | |
370 | break; | |
371 | } | |
9929b0fb | 372 | tidlist[i] = tid + tidoff; |
7f510b46 | 373 | ipath_cdbg(VERBOSE, "Updating idx %u to TID %u, " |
9929b0fb | 374 | "vaddr %lx\n", i, tid + tidoff, vaddr); |
7f510b46 BS |
375 | /* we "know" system pages and TID pages are same size */ |
376 | dd->ipath_pageshadow[porttid + tid] = pagep[i]; | |
1fd3b40f BS |
377 | dd->ipath_physshadow[porttid + tid] = ipath_map_page( |
378 | dd->pcidev, pagep[i], 0, PAGE_SIZE, | |
379 | PCI_DMA_FROMDEVICE); | |
7f510b46 BS |
380 | /* |
381 | * don't need atomic or it's overhead | |
382 | */ | |
383 | __set_bit(tid, tidmap); | |
1fd3b40f | 384 | physaddr = dd->ipath_physshadow[porttid + tid]; |
7f510b46 BS |
385 | ipath_stats.sps_pagelocks++; |
386 | ipath_cdbg(VERBOSE, | |
387 | "TID %u, vaddr %lx, physaddr %llx pgp %p\n", | |
388 | tid, vaddr, (unsigned long long) physaddr, | |
389 | pagep[i]); | |
390 | dd->ipath_f_put_tid(dd, &tidbase[tid], 1, physaddr); | |
391 | /* | |
392 | * don't check this tid in ipath_portshadow, since we | |
393 | * just filled it in; start with the next one. | |
394 | */ | |
395 | tid++; | |
396 | } | |
397 | ||
398 | if (ret) { | |
399 | u32 limit; | |
400 | cleanup: | |
401 | /* jump here if copy out of updated info failed... */ | |
402 | ipath_dbg("After failure (ret=%d), undo %d of %d entries\n", | |
403 | -ret, i, cnt); | |
404 | /* same code that's in ipath_free_tid() */ | |
405 | limit = sizeof(tidmap) * BITS_PER_BYTE; | |
406 | if (limit > tidcnt) | |
407 | /* just in case size changes in future */ | |
408 | limit = tidcnt; | |
409 | tid = find_first_bit((const unsigned long *)tidmap, limit); | |
410 | for (; tid < limit; tid++) { | |
411 | if (!test_bit(tid, tidmap)) | |
412 | continue; | |
413 | if (dd->ipath_pageshadow[porttid + tid]) { | |
414 | ipath_cdbg(VERBOSE, "Freeing TID %u\n", | |
415 | tid); | |
416 | dd->ipath_f_put_tid(dd, &tidbase[tid], 1, | |
417 | dd->ipath_tidinvalid); | |
1fd3b40f BS |
418 | pci_unmap_page(dd->pcidev, |
419 | dd->ipath_physshadow[porttid + tid], | |
420 | PAGE_SIZE, PCI_DMA_FROMDEVICE); | |
7f510b46 BS |
421 | dd->ipath_pageshadow[porttid + tid] = NULL; |
422 | ipath_stats.sps_pageunlocks++; | |
423 | } | |
424 | } | |
425 | ipath_release_user_pages(pagep, cnt); | |
426 | } else { | |
427 | /* | |
428 | * Copy the updated array, with ipath_tid's filled in, back | |
429 | * to user. Since we did the copy in already, this "should | |
430 | * never fail" If it does, we have to clean up... | |
431 | */ | |
432 | if (copy_to_user((void __user *) | |
433 | (unsigned long) ti->tidlist, | |
434 | tidlist, cnt * sizeof(*tidlist))) { | |
435 | ret = -EFAULT; | |
436 | goto cleanup; | |
437 | } | |
438 | if (copy_to_user((void __user *) (unsigned long) ti->tidmap, | |
439 | tidmap, sizeof tidmap)) { | |
440 | ret = -EFAULT; | |
441 | goto cleanup; | |
442 | } | |
443 | if (tid == tidcnt) | |
444 | tid = 0; | |
9929b0fb BS |
445 | if (!pd->port_subport_cnt) |
446 | pd->port_tidcursor = tid; | |
447 | else | |
448 | tidcursor_fp(fp) = tid; | |
7f510b46 BS |
449 | } |
450 | ||
451 | done: | |
452 | if (ret) | |
453 | ipath_dbg("Failed to map %u TID pages, failing with %d\n", | |
454 | ti->tidcnt, -ret); | |
455 | return ret; | |
456 | } | |
457 | ||
458 | /** | |
459 | * ipath_tid_free - free a port TID | |
460 | * @pd: the port | |
9929b0fb | 461 | * @subport: the subport |
7f510b46 BS |
462 | * @ti: the TID info |
463 | * | |
464 | * right now we are unlocking one page at a time, but since | |
465 | * the intended use of this routine is for a single group of | |
466 | * virtually contiguous pages, that should change to improve | |
467 | * performance. We check that the TID is in range for this port | |
468 | * but otherwise don't check validity; if user has an error and | |
469 | * frees the wrong tid, it's only their own data that can thereby | |
470 | * be corrupted. We do check that the TID was in use, for sanity | |
471 | * We always use our idea of the saved address, not the address that | |
472 | * they pass in to us. | |
473 | */ | |
474 | ||
9929b0fb | 475 | static int ipath_tid_free(struct ipath_portdata *pd, unsigned subport, |
7f510b46 BS |
476 | const struct ipath_tid_info *ti) |
477 | { | |
478 | int ret = 0; | |
479 | u32 tid, porttid, cnt, limit, tidcnt; | |
480 | struct ipath_devdata *dd = pd->port_dd; | |
481 | u64 __iomem *tidbase; | |
482 | unsigned long tidmap[8]; | |
483 | ||
484 | if (!dd->ipath_pageshadow) { | |
485 | ret = -ENOMEM; | |
486 | goto done; | |
487 | } | |
488 | ||
489 | if (copy_from_user(tidmap, (void __user *)(unsigned long)ti->tidmap, | |
490 | sizeof tidmap)) { | |
491 | ret = -EFAULT; | |
492 | goto done; | |
493 | } | |
494 | ||
495 | porttid = pd->port_port * dd->ipath_rcvtidcnt; | |
9929b0fb BS |
496 | if (!pd->port_subport_cnt) |
497 | tidcnt = dd->ipath_rcvtidcnt; | |
498 | else if (!subport) { | |
499 | tidcnt = (dd->ipath_rcvtidcnt / pd->port_subport_cnt) + | |
500 | (dd->ipath_rcvtidcnt % pd->port_subport_cnt); | |
501 | porttid += dd->ipath_rcvtidcnt - tidcnt; | |
502 | } else { | |
503 | tidcnt = dd->ipath_rcvtidcnt / pd->port_subport_cnt; | |
504 | porttid += tidcnt * (subport - 1); | |
505 | } | |
7f510b46 BS |
506 | tidbase = (u64 __iomem *) ((char __iomem *)(dd->ipath_kregbase) + |
507 | dd->ipath_rcvtidbase + | |
508 | porttid * sizeof(*tidbase)); | |
509 | ||
7f510b46 BS |
510 | limit = sizeof(tidmap) * BITS_PER_BYTE; |
511 | if (limit > tidcnt) | |
512 | /* just in case size changes in future */ | |
513 | limit = tidcnt; | |
514 | tid = find_first_bit(tidmap, limit); | |
515 | ipath_cdbg(VERBOSE, "Port%u free %u tids; first bit (max=%d) " | |
516 | "set is %d, porttid %u\n", pd->port_port, ti->tidcnt, | |
517 | limit, tid, porttid); | |
518 | for (cnt = 0; tid < limit; tid++) { | |
519 | /* | |
520 | * small optimization; if we detect a run of 3 or so without | |
521 | * any set, use find_first_bit again. That's mainly to | |
522 | * accelerate the case where we wrapped, so we have some at | |
523 | * the beginning, and some at the end, and a big gap | |
524 | * in the middle. | |
525 | */ | |
526 | if (!test_bit(tid, tidmap)) | |
527 | continue; | |
528 | cnt++; | |
529 | if (dd->ipath_pageshadow[porttid + tid]) { | |
530 | ipath_cdbg(VERBOSE, "PID %u freeing TID %u\n", | |
531 | pd->port_pid, tid); | |
532 | dd->ipath_f_put_tid(dd, &tidbase[tid], 1, | |
533 | dd->ipath_tidinvalid); | |
1fd3b40f BS |
534 | pci_unmap_page(dd->pcidev, |
535 | dd->ipath_physshadow[porttid + tid], | |
536 | PAGE_SIZE, PCI_DMA_FROMDEVICE); | |
7f510b46 BS |
537 | ipath_release_user_pages( |
538 | &dd->ipath_pageshadow[porttid + tid], 1); | |
539 | dd->ipath_pageshadow[porttid + tid] = NULL; | |
540 | ipath_stats.sps_pageunlocks++; | |
541 | } else | |
542 | ipath_dbg("Unused tid %u, ignoring\n", tid); | |
543 | } | |
544 | if (cnt != ti->tidcnt) | |
545 | ipath_dbg("passed in tidcnt %d, only %d bits set in map\n", | |
546 | ti->tidcnt, cnt); | |
547 | done: | |
548 | if (ret) | |
549 | ipath_dbg("Failed to unmap %u TID pages, failing with %d\n", | |
550 | ti->tidcnt, -ret); | |
551 | return ret; | |
552 | } | |
553 | ||
554 | /** | |
555 | * ipath_set_part_key - set a partition key | |
556 | * @pd: the port | |
557 | * @key: the key | |
558 | * | |
559 | * We can have up to 4 active at a time (other than the default, which is | |
560 | * always allowed). This is somewhat tricky, since multiple ports may set | |
561 | * the same key, so we reference count them, and clean up at exit. All 4 | |
562 | * partition keys are packed into a single infinipath register. It's an | |
563 | * error for a process to set the same pkey multiple times. We provide no | |
564 | * mechanism to de-allocate a pkey at this time, we may eventually need to | |
565 | * do that. I've used the atomic operations, and no locking, and only make | |
566 | * a single pass through what's available. This should be more than | |
567 | * adequate for some time. I'll think about spinlocks or the like if and as | |
568 | * it's necessary. | |
569 | */ | |
570 | static int ipath_set_part_key(struct ipath_portdata *pd, u16 key) | |
571 | { | |
572 | struct ipath_devdata *dd = pd->port_dd; | |
573 | int i, any = 0, pidx = -1; | |
574 | u16 lkey = key & 0x7FFF; | |
575 | int ret; | |
576 | ||
27b678dd | 577 | if (lkey == (IPATH_DEFAULT_P_KEY & 0x7FFF)) { |
7f510b46 BS |
578 | /* nothing to do; this key always valid */ |
579 | ret = 0; | |
580 | goto bail; | |
581 | } | |
582 | ||
583 | ipath_cdbg(VERBOSE, "p%u try to set pkey %hx, current keys " | |
584 | "%hx:%x %hx:%x %hx:%x %hx:%x\n", | |
585 | pd->port_port, key, dd->ipath_pkeys[0], | |
586 | atomic_read(&dd->ipath_pkeyrefs[0]), dd->ipath_pkeys[1], | |
587 | atomic_read(&dd->ipath_pkeyrefs[1]), dd->ipath_pkeys[2], | |
588 | atomic_read(&dd->ipath_pkeyrefs[2]), dd->ipath_pkeys[3], | |
589 | atomic_read(&dd->ipath_pkeyrefs[3])); | |
590 | ||
591 | if (!lkey) { | |
592 | ipath_cdbg(PROC, "p%u tries to set key 0, not allowed\n", | |
593 | pd->port_port); | |
594 | ret = -EINVAL; | |
595 | goto bail; | |
596 | } | |
597 | ||
598 | /* | |
599 | * Set the full membership bit, because it has to be | |
600 | * set in the register or the packet, and it seems | |
601 | * cleaner to set in the register than to force all | |
602 | * callers to set it. (see bug 4331) | |
603 | */ | |
604 | key |= 0x8000; | |
605 | ||
606 | for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) { | |
607 | if (!pd->port_pkeys[i] && pidx == -1) | |
608 | pidx = i; | |
609 | if (pd->port_pkeys[i] == key) { | |
610 | ipath_cdbg(VERBOSE, "p%u tries to set same pkey " | |
611 | "(%x) more than once\n", | |
612 | pd->port_port, key); | |
613 | ret = -EEXIST; | |
614 | goto bail; | |
615 | } | |
616 | } | |
617 | if (pidx == -1) { | |
618 | ipath_dbg("All pkeys for port %u already in use, " | |
619 | "can't set %x\n", pd->port_port, key); | |
620 | ret = -EBUSY; | |
621 | goto bail; | |
622 | } | |
623 | for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) { | |
624 | if (!dd->ipath_pkeys[i]) { | |
625 | any++; | |
626 | continue; | |
627 | } | |
628 | if (dd->ipath_pkeys[i] == key) { | |
629 | atomic_t *pkrefs = &dd->ipath_pkeyrefs[i]; | |
630 | ||
631 | if (atomic_inc_return(pkrefs) > 1) { | |
632 | pd->port_pkeys[pidx] = key; | |
633 | ipath_cdbg(VERBOSE, "p%u set key %x " | |
634 | "matches #%d, count now %d\n", | |
635 | pd->port_port, key, i, | |
636 | atomic_read(pkrefs)); | |
637 | ret = 0; | |
638 | goto bail; | |
639 | } else { | |
640 | /* | |
641 | * lost race, decrement count, catch below | |
642 | */ | |
643 | atomic_dec(pkrefs); | |
644 | ipath_cdbg(VERBOSE, "Lost race, count was " | |
645 | "0, after dec, it's %d\n", | |
646 | atomic_read(pkrefs)); | |
647 | any++; | |
648 | } | |
649 | } | |
650 | if ((dd->ipath_pkeys[i] & 0x7FFF) == lkey) { | |
651 | /* | |
652 | * It makes no sense to have both the limited and | |
653 | * full membership PKEY set at the same time since | |
654 | * the unlimited one will disable the limited one. | |
655 | */ | |
656 | ret = -EEXIST; | |
657 | goto bail; | |
658 | } | |
659 | } | |
660 | if (!any) { | |
661 | ipath_dbg("port %u, all pkeys already in use, " | |
662 | "can't set %x\n", pd->port_port, key); | |
663 | ret = -EBUSY; | |
664 | goto bail; | |
665 | } | |
666 | for (any = i = 0; i < ARRAY_SIZE(dd->ipath_pkeys); i++) { | |
667 | if (!dd->ipath_pkeys[i] && | |
668 | atomic_inc_return(&dd->ipath_pkeyrefs[i]) == 1) { | |
669 | u64 pkey; | |
670 | ||
671 | /* for ipathstats, etc. */ | |
672 | ipath_stats.sps_pkeys[i] = lkey; | |
673 | pd->port_pkeys[pidx] = dd->ipath_pkeys[i] = key; | |
674 | pkey = | |
675 | (u64) dd->ipath_pkeys[0] | | |
676 | ((u64) dd->ipath_pkeys[1] << 16) | | |
677 | ((u64) dd->ipath_pkeys[2] << 32) | | |
678 | ((u64) dd->ipath_pkeys[3] << 48); | |
679 | ipath_cdbg(PROC, "p%u set key %x in #%d, " | |
680 | "portidx %d, new pkey reg %llx\n", | |
681 | pd->port_port, key, i, pidx, | |
682 | (unsigned long long) pkey); | |
683 | ipath_write_kreg( | |
684 | dd, dd->ipath_kregs->kr_partitionkey, pkey); | |
685 | ||
686 | ret = 0; | |
687 | goto bail; | |
688 | } | |
689 | } | |
690 | ipath_dbg("port %u, all pkeys already in use 2nd pass, " | |
691 | "can't set %x\n", pd->port_port, key); | |
692 | ret = -EBUSY; | |
693 | ||
694 | bail: | |
695 | return ret; | |
696 | } | |
697 | ||
698 | /** | |
699 | * ipath_manage_rcvq - manage a port's receive queue | |
700 | * @pd: the port | |
9929b0fb | 701 | * @subport: the subport |
7f510b46 BS |
702 | * @start_stop: action to carry out |
703 | * | |
704 | * start_stop == 0 disables receive on the port, for use in queue | |
705 | * overflow conditions. start_stop==1 re-enables, to be used to | |
706 | * re-init the software copy of the head register | |
707 | */ | |
9929b0fb BS |
708 | static int ipath_manage_rcvq(struct ipath_portdata *pd, unsigned subport, |
709 | int start_stop) | |
7f510b46 BS |
710 | { |
711 | struct ipath_devdata *dd = pd->port_dd; | |
7f510b46 | 712 | |
9929b0fb | 713 | ipath_cdbg(PROC, "%sabling rcv for unit %u port %u:%u\n", |
7f510b46 | 714 | start_stop ? "en" : "dis", dd->ipath_unit, |
9929b0fb BS |
715 | pd->port_port, subport); |
716 | if (subport) | |
717 | goto bail; | |
7f510b46 BS |
718 | /* atomically clear receive enable port. */ |
719 | if (start_stop) { | |
720 | /* | |
721 | * On enable, force in-memory copy of the tail register to | |
722 | * 0, so that protocol code doesn't have to worry about | |
723 | * whether or not the chip has yet updated the in-memory | |
724 | * copy or not on return from the system call. The chip | |
725 | * always resets it's tail register back to 0 on a | |
726 | * transition from disabled to enabled. This could cause a | |
727 | * problem if software was broken, and did the enable w/o | |
728 | * the disable, but eventually the in-memory copy will be | |
729 | * updated and correct itself, even in the face of software | |
730 | * bugs. | |
731 | */ | |
1fd3b40f | 732 | *(volatile u64 *)pd->port_rcvhdrtail_kvaddr = 0; |
7f510b46 BS |
733 | set_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port, |
734 | &dd->ipath_rcvctrl); | |
735 | } else | |
736 | clear_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port, | |
737 | &dd->ipath_rcvctrl); | |
738 | ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl, | |
739 | dd->ipath_rcvctrl); | |
740 | /* now be sure chip saw it before we return */ | |
44f8e3f3 | 741 | ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch); |
7f510b46 BS |
742 | if (start_stop) { |
743 | /* | |
744 | * And try to be sure that tail reg update has happened too. | |
745 | * This should in theory interlock with the RXE changes to | |
746 | * the tail register. Don't assign it to the tail register | |
747 | * in memory copy, since we could overwrite an update by the | |
748 | * chip if we did. | |
749 | */ | |
44f8e3f3 | 750 | ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port); |
7f510b46 BS |
751 | } |
752 | /* always; new head should be equal to new tail; see above */ | |
9929b0fb | 753 | bail: |
7f510b46 BS |
754 | return 0; |
755 | } | |
756 | ||
757 | static void ipath_clean_part_key(struct ipath_portdata *pd, | |
758 | struct ipath_devdata *dd) | |
759 | { | |
760 | int i, j, pchanged = 0; | |
761 | u64 oldpkey; | |
762 | ||
763 | /* for debugging only */ | |
764 | oldpkey = (u64) dd->ipath_pkeys[0] | | |
765 | ((u64) dd->ipath_pkeys[1] << 16) | | |
766 | ((u64) dd->ipath_pkeys[2] << 32) | | |
767 | ((u64) dd->ipath_pkeys[3] << 48); | |
768 | ||
769 | for (i = 0; i < ARRAY_SIZE(pd->port_pkeys); i++) { | |
770 | if (!pd->port_pkeys[i]) | |
771 | continue; | |
772 | ipath_cdbg(VERBOSE, "look for key[%d] %hx in pkeys\n", i, | |
773 | pd->port_pkeys[i]); | |
774 | for (j = 0; j < ARRAY_SIZE(dd->ipath_pkeys); j++) { | |
775 | /* check for match independent of the global bit */ | |
776 | if ((dd->ipath_pkeys[j] & 0x7fff) != | |
777 | (pd->port_pkeys[i] & 0x7fff)) | |
778 | continue; | |
779 | if (atomic_dec_and_test(&dd->ipath_pkeyrefs[j])) { | |
780 | ipath_cdbg(VERBOSE, "p%u clear key " | |
781 | "%x matches #%d\n", | |
782 | pd->port_port, | |
783 | pd->port_pkeys[i], j); | |
784 | ipath_stats.sps_pkeys[j] = | |
785 | dd->ipath_pkeys[j] = 0; | |
786 | pchanged++; | |
787 | } | |
788 | else ipath_cdbg( | |
789 | VERBOSE, "p%u key %x matches #%d, " | |
790 | "but ref still %d\n", pd->port_port, | |
791 | pd->port_pkeys[i], j, | |
792 | atomic_read(&dd->ipath_pkeyrefs[j])); | |
793 | break; | |
794 | } | |
795 | pd->port_pkeys[i] = 0; | |
796 | } | |
797 | if (pchanged) { | |
798 | u64 pkey = (u64) dd->ipath_pkeys[0] | | |
799 | ((u64) dd->ipath_pkeys[1] << 16) | | |
800 | ((u64) dd->ipath_pkeys[2] << 32) | | |
801 | ((u64) dd->ipath_pkeys[3] << 48); | |
802 | ipath_cdbg(VERBOSE, "p%u old pkey reg %llx, " | |
803 | "new pkey reg %llx\n", pd->port_port, | |
804 | (unsigned long long) oldpkey, | |
805 | (unsigned long long) pkey); | |
806 | ipath_write_kreg(dd, dd->ipath_kregs->kr_partitionkey, | |
807 | pkey); | |
808 | } | |
809 | } | |
810 | ||
9929b0fb BS |
811 | /* |
812 | * Initialize the port data with the receive buffer sizes | |
813 | * so this can be done while the master port is locked. | |
814 | * Otherwise, there is a race with a slave opening the port | |
815 | * and seeing these fields uninitialized. | |
816 | */ | |
817 | static void init_user_egr_sizes(struct ipath_portdata *pd) | |
818 | { | |
819 | struct ipath_devdata *dd = pd->port_dd; | |
820 | unsigned egrperchunk, egrcnt, size; | |
821 | ||
822 | /* | |
823 | * to avoid wasting a lot of memory, we allocate 32KB chunks of | |
824 | * physically contiguous memory, advance through it until used up | |
825 | * and then allocate more. Of course, we need memory to store those | |
826 | * extra pointers, now. Started out with 256KB, but under heavy | |
827 | * memory pressure (creating large files and then copying them over | |
828 | * NFS while doing lots of MPI jobs), we hit some allocation | |
829 | * failures, even though we can sleep... (2.6.10) Still get | |
830 | * failures at 64K. 32K is the lowest we can go without wasting | |
831 | * additional memory. | |
832 | */ | |
833 | size = 0x8000; | |
834 | egrperchunk = size / dd->ipath_rcvegrbufsize; | |
835 | egrcnt = dd->ipath_rcvegrcnt; | |
836 | pd->port_rcvegrbuf_chunks = (egrcnt + egrperchunk - 1) / egrperchunk; | |
837 | pd->port_rcvegrbufs_perchunk = egrperchunk; | |
838 | pd->port_rcvegrbuf_size = size; | |
839 | } | |
840 | ||
7f510b46 BS |
841 | /** |
842 | * ipath_create_user_egr - allocate eager TID buffers | |
843 | * @pd: the port to allocate TID buffers for | |
844 | * | |
845 | * This routine is now quite different for user and kernel, because | |
846 | * the kernel uses skb's, for the accelerated network performance | |
847 | * This is the user port version | |
848 | * | |
849 | * Allocate the eager TID buffers and program them into infinipath | |
850 | * They are no longer completely contiguous, we do multiple allocation | |
851 | * calls. | |
852 | */ | |
853 | static int ipath_create_user_egr(struct ipath_portdata *pd) | |
854 | { | |
855 | struct ipath_devdata *dd = pd->port_dd; | |
9929b0fb | 856 | unsigned e, egrcnt, egrperchunk, chunk, egrsize, egroff; |
7f510b46 BS |
857 | size_t size; |
858 | int ret; | |
0ed9a4a0 BS |
859 | gfp_t gfp_flags; |
860 | ||
861 | /* | |
862 | * GFP_USER, but without GFP_FS, so buffer cache can be | |
863 | * coalesced (we hope); otherwise, even at order 4, | |
864 | * heavy filesystem activity makes these fail, and we can | |
865 | * use compound pages. | |
866 | */ | |
867 | gfp_flags = __GFP_WAIT | __GFP_IO | __GFP_COMP; | |
7f510b46 BS |
868 | |
869 | egrcnt = dd->ipath_rcvegrcnt; | |
870 | /* TID number offset for this port */ | |
871 | egroff = pd->port_port * egrcnt; | |
872 | egrsize = dd->ipath_rcvegrbufsize; | |
873 | ipath_cdbg(VERBOSE, "Allocating %d egr buffers, at egrtid " | |
874 | "offset %x, egrsize %u\n", egrcnt, egroff, egrsize); | |
875 | ||
9929b0fb BS |
876 | chunk = pd->port_rcvegrbuf_chunks; |
877 | egrperchunk = pd->port_rcvegrbufs_perchunk; | |
878 | size = pd->port_rcvegrbuf_size; | |
879 | pd->port_rcvegrbuf = kmalloc(chunk * sizeof(pd->port_rcvegrbuf[0]), | |
880 | GFP_KERNEL); | |
7f510b46 BS |
881 | if (!pd->port_rcvegrbuf) { |
882 | ret = -ENOMEM; | |
883 | goto bail; | |
884 | } | |
885 | pd->port_rcvegrbuf_phys = | |
9929b0fb BS |
886 | kmalloc(chunk * sizeof(pd->port_rcvegrbuf_phys[0]), |
887 | GFP_KERNEL); | |
7f510b46 BS |
888 | if (!pd->port_rcvegrbuf_phys) { |
889 | ret = -ENOMEM; | |
890 | goto bail_rcvegrbuf; | |
891 | } | |
892 | for (e = 0; e < pd->port_rcvegrbuf_chunks; e++) { | |
7f510b46 BS |
893 | |
894 | pd->port_rcvegrbuf[e] = dma_alloc_coherent( | |
895 | &dd->pcidev->dev, size, &pd->port_rcvegrbuf_phys[e], | |
896 | gfp_flags); | |
897 | ||
898 | if (!pd->port_rcvegrbuf[e]) { | |
899 | ret = -ENOMEM; | |
900 | goto bail_rcvegrbuf_phys; | |
901 | } | |
902 | } | |
903 | ||
904 | pd->port_rcvegr_phys = pd->port_rcvegrbuf_phys[0]; | |
905 | ||
906 | for (e = chunk = 0; chunk < pd->port_rcvegrbuf_chunks; chunk++) { | |
907 | dma_addr_t pa = pd->port_rcvegrbuf_phys[chunk]; | |
908 | unsigned i; | |
909 | ||
910 | for (i = 0; e < egrcnt && i < egrperchunk; e++, i++) { | |
911 | dd->ipath_f_put_tid(dd, e + egroff + | |
912 | (u64 __iomem *) | |
913 | ((char __iomem *) | |
914 | dd->ipath_kregbase + | |
915 | dd->ipath_rcvegrbase), 0, pa); | |
916 | pa += egrsize; | |
917 | } | |
918 | cond_resched(); /* don't hog the cpu */ | |
919 | } | |
920 | ||
921 | ret = 0; | |
922 | goto bail; | |
923 | ||
924 | bail_rcvegrbuf_phys: | |
925 | for (e = 0; e < pd->port_rcvegrbuf_chunks && | |
f37bda92 | 926 | pd->port_rcvegrbuf[e]; e++) { |
7f510b46 BS |
927 | dma_free_coherent(&dd->pcidev->dev, size, |
928 | pd->port_rcvegrbuf[e], | |
929 | pd->port_rcvegrbuf_phys[e]); | |
930 | ||
f37bda92 | 931 | } |
9929b0fb | 932 | kfree(pd->port_rcvegrbuf_phys); |
7f510b46 BS |
933 | pd->port_rcvegrbuf_phys = NULL; |
934 | bail_rcvegrbuf: | |
9929b0fb | 935 | kfree(pd->port_rcvegrbuf); |
7f510b46 BS |
936 | pd->port_rcvegrbuf = NULL; |
937 | bail: | |
938 | return ret; | |
939 | } | |
940 | ||
f37bda92 BS |
941 | |
942 | /* common code for the mappings on dma_alloc_coherent mem */ | |
943 | static int ipath_mmap_mem(struct vm_area_struct *vma, | |
1fd3b40f BS |
944 | struct ipath_portdata *pd, unsigned len, int write_ok, |
945 | void *kvaddr, char *what) | |
f37bda92 BS |
946 | { |
947 | struct ipath_devdata *dd = pd->port_dd; | |
1fd3b40f | 948 | unsigned long pfn; |
f37bda92 BS |
949 | int ret; |
950 | ||
951 | if ((vma->vm_end - vma->vm_start) > len) { | |
952 | dev_info(&dd->pcidev->dev, | |
953 | "FAIL on %s: len %lx > %x\n", what, | |
954 | vma->vm_end - vma->vm_start, len); | |
955 | ret = -EFAULT; | |
956 | goto bail; | |
957 | } | |
958 | ||
959 | if (!write_ok) { | |
960 | if (vma->vm_flags & VM_WRITE) { | |
961 | dev_info(&dd->pcidev->dev, | |
962 | "%s must be mapped readonly\n", what); | |
963 | ret = -EPERM; | |
964 | goto bail; | |
965 | } | |
966 | ||
967 | /* don't allow them to later change with mprotect */ | |
968 | vma->vm_flags &= ~VM_MAYWRITE; | |
969 | } | |
970 | ||
1fd3b40f | 971 | pfn = virt_to_phys(kvaddr) >> PAGE_SHIFT; |
f37bda92 BS |
972 | ret = remap_pfn_range(vma, vma->vm_start, pfn, |
973 | len, vma->vm_page_prot); | |
974 | if (ret) | |
1fd3b40f BS |
975 | dev_info(&dd->pcidev->dev, "%s port%u mmap of %lx, %x " |
976 | "bytes r%c failed: %d\n", what, pd->port_port, | |
977 | pfn, len, write_ok?'w':'o', ret); | |
f37bda92 | 978 | else |
1fd3b40f BS |
979 | ipath_cdbg(VERBOSE, "%s port%u mmaped %lx, %x bytes " |
980 | "r%c\n", what, pd->port_port, pfn, len, | |
981 | write_ok?'w':'o'); | |
f37bda92 BS |
982 | bail: |
983 | return ret; | |
984 | } | |
985 | ||
7f510b46 BS |
986 | static int mmap_ureg(struct vm_area_struct *vma, struct ipath_devdata *dd, |
987 | u64 ureg) | |
988 | { | |
989 | unsigned long phys; | |
990 | int ret; | |
991 | ||
f37bda92 BS |
992 | /* |
993 | * This is real hardware, so use io_remap. This is the mechanism | |
994 | * for the user process to update the head registers for their port | |
995 | * in the chip. | |
996 | */ | |
7f510b46 BS |
997 | if ((vma->vm_end - vma->vm_start) > PAGE_SIZE) { |
998 | dev_info(&dd->pcidev->dev, "FAIL mmap userreg: reqlen " | |
999 | "%lx > PAGE\n", vma->vm_end - vma->vm_start); | |
1000 | ret = -EFAULT; | |
1001 | } else { | |
1002 | phys = dd->ipath_physaddr + ureg; | |
1003 | vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot); | |
1004 | ||
1005 | vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND; | |
1006 | ret = io_remap_pfn_range(vma, vma->vm_start, | |
1007 | phys >> PAGE_SHIFT, | |
1008 | vma->vm_end - vma->vm_start, | |
1009 | vma->vm_page_prot); | |
1010 | } | |
1011 | return ret; | |
1012 | } | |
1013 | ||
1014 | static int mmap_piobufs(struct vm_area_struct *vma, | |
1015 | struct ipath_devdata *dd, | |
9929b0fb BS |
1016 | struct ipath_portdata *pd, |
1017 | unsigned piobufs, unsigned piocnt) | |
7f510b46 BS |
1018 | { |
1019 | unsigned long phys; | |
1020 | int ret; | |
1021 | ||
1022 | /* | |
f37bda92 BS |
1023 | * When we map the PIO buffers in the chip, we want to map them as |
1024 | * writeonly, no read possible. This prevents access to previous | |
1025 | * process data, and catches users who might try to read the i/o | |
1026 | * space due to a bug. | |
7f510b46 | 1027 | */ |
9929b0fb | 1028 | if ((vma->vm_end - vma->vm_start) > (piocnt * dd->ipath_palign)) { |
7f510b46 BS |
1029 | dev_info(&dd->pcidev->dev, "FAIL mmap piobufs: " |
1030 | "reqlen %lx > PAGE\n", | |
1031 | vma->vm_end - vma->vm_start); | |
9929b0fb | 1032 | ret = -EINVAL; |
7f510b46 BS |
1033 | goto bail; |
1034 | } | |
1035 | ||
9929b0fb | 1036 | phys = dd->ipath_physaddr + piobufs; |
f37bda92 | 1037 | |
7f510b46 | 1038 | /* |
f37bda92 | 1039 | * Don't mark this as non-cached, or we don't get the |
7f510b46 | 1040 | * write combining behavior we want on the PIO buffers! |
7f510b46 BS |
1041 | */ |
1042 | ||
eb9dc6f4 BS |
1043 | #if defined(__powerpc__) |
1044 | /* There isn't a generic way to specify writethrough mappings */ | |
1045 | pgprot_val(vma->vm_page_prot) |= _PAGE_NO_CACHE; | |
1046 | pgprot_val(vma->vm_page_prot) |= _PAGE_WRITETHRU; | |
1047 | pgprot_val(vma->vm_page_prot) &= ~_PAGE_GUARDED; | |
1048 | #endif | |
1049 | ||
367fe711 BS |
1050 | /* |
1051 | * don't allow them to later change to readable with mprotect (for when | |
1052 | * not initially mapped readable, as is normally the case) | |
1053 | */ | |
f37bda92 | 1054 | vma->vm_flags &= ~VM_MAYREAD; |
7f510b46 BS |
1055 | vma->vm_flags |= VM_DONTCOPY | VM_DONTEXPAND; |
1056 | ||
1057 | ret = io_remap_pfn_range(vma, vma->vm_start, phys >> PAGE_SHIFT, | |
1058 | vma->vm_end - vma->vm_start, | |
1059 | vma->vm_page_prot); | |
1060 | bail: | |
1061 | return ret; | |
1062 | } | |
1063 | ||
1064 | static int mmap_rcvegrbufs(struct vm_area_struct *vma, | |
1065 | struct ipath_portdata *pd) | |
1066 | { | |
1067 | struct ipath_devdata *dd = pd->port_dd; | |
1068 | unsigned long start, size; | |
1069 | size_t total_size, i; | |
1fd3b40f | 1070 | unsigned long pfn; |
7f510b46 BS |
1071 | int ret; |
1072 | ||
7f510b46 BS |
1073 | size = pd->port_rcvegrbuf_size; |
1074 | total_size = pd->port_rcvegrbuf_chunks * size; | |
1075 | if ((vma->vm_end - vma->vm_start) > total_size) { | |
1076 | dev_info(&dd->pcidev->dev, "FAIL on egr bufs: " | |
1077 | "reqlen %lx > actual %lx\n", | |
1078 | vma->vm_end - vma->vm_start, | |
1079 | (unsigned long) total_size); | |
9929b0fb | 1080 | ret = -EINVAL; |
7f510b46 BS |
1081 | goto bail; |
1082 | } | |
1083 | ||
1084 | if (vma->vm_flags & VM_WRITE) { | |
1085 | dev_info(&dd->pcidev->dev, "Can't map eager buffers as " | |
1086 | "writable (flags=%lx)\n", vma->vm_flags); | |
1087 | ret = -EPERM; | |
1088 | goto bail; | |
1089 | } | |
f37bda92 BS |
1090 | /* don't allow them to later change to writeable with mprotect */ |
1091 | vma->vm_flags &= ~VM_MAYWRITE; | |
7f510b46 BS |
1092 | |
1093 | start = vma->vm_start; | |
7f510b46 | 1094 | |
7f510b46 | 1095 | for (i = 0; i < pd->port_rcvegrbuf_chunks; i++, start += size) { |
1fd3b40f BS |
1096 | pfn = virt_to_phys(pd->port_rcvegrbuf[i]) >> PAGE_SHIFT; |
1097 | ret = remap_pfn_range(vma, start, pfn, size, | |
1098 | vma->vm_page_prot); | |
7f510b46 BS |
1099 | if (ret < 0) |
1100 | goto bail; | |
1101 | } | |
1102 | ret = 0; | |
1103 | ||
1104 | bail: | |
1105 | return ret; | |
1106 | } | |
1107 | ||
9929b0fb BS |
1108 | /* |
1109 | * ipath_file_vma_nopage - handle a VMA page fault. | |
1110 | */ | |
1111 | static struct page *ipath_file_vma_nopage(struct vm_area_struct *vma, | |
1112 | unsigned long address, int *type) | |
1113 | { | |
1114 | unsigned long offset = address - vma->vm_start; | |
1115 | struct page *page = NOPAGE_SIGBUS; | |
1116 | void *pageptr; | |
1117 | ||
1118 | /* | |
1119 | * Convert the vmalloc address into a struct page. | |
1120 | */ | |
1121 | pageptr = (void *)(offset + (vma->vm_pgoff << PAGE_SHIFT)); | |
1122 | page = vmalloc_to_page(pageptr); | |
1123 | if (!page) | |
1124 | goto out; | |
1125 | ||
1126 | /* Increment the reference count. */ | |
1127 | get_page(page); | |
1128 | if (type) | |
1129 | *type = VM_FAULT_MINOR; | |
1130 | out: | |
1131 | return page; | |
1132 | } | |
1133 | ||
1134 | static struct vm_operations_struct ipath_file_vm_ops = { | |
1135 | .nopage = ipath_file_vma_nopage, | |
1136 | }; | |
1137 | ||
1138 | static int mmap_kvaddr(struct vm_area_struct *vma, u64 pgaddr, | |
1139 | struct ipath_portdata *pd, unsigned subport) | |
1140 | { | |
1141 | unsigned long len; | |
1142 | struct ipath_devdata *dd; | |
1143 | void *addr; | |
1144 | size_t size; | |
0a5a83cf | 1145 | int ret = 0; |
9929b0fb BS |
1146 | |
1147 | /* If the port is not shared, all addresses should be physical */ | |
0a5a83cf | 1148 | if (!pd->port_subport_cnt) |
9929b0fb | 1149 | goto bail; |
9929b0fb BS |
1150 | |
1151 | dd = pd->port_dd; | |
1152 | size = pd->port_rcvegrbuf_chunks * pd->port_rcvegrbuf_size; | |
1153 | ||
1154 | /* | |
1155 | * Master has all the slave uregbase, rcvhdrq, and | |
1156 | * rcvegrbufs mmapped. | |
1157 | */ | |
1158 | if (subport == 0) { | |
1159 | unsigned num_slaves = pd->port_subport_cnt - 1; | |
1160 | ||
0a5a83cf | 1161 | if (pgaddr == cvt_kvaddr(pd->subport_uregbase)) { |
9929b0fb BS |
1162 | addr = pd->subport_uregbase; |
1163 | size = PAGE_SIZE * num_slaves; | |
0a5a83cf | 1164 | } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base)) { |
9929b0fb BS |
1165 | addr = pd->subport_rcvhdr_base; |
1166 | size = pd->port_rcvhdrq_size * num_slaves; | |
0a5a83cf | 1167 | } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf)) { |
9929b0fb BS |
1168 | addr = pd->subport_rcvegrbuf; |
1169 | size *= num_slaves; | |
0a5a83cf | 1170 | } else |
9929b0fb | 1171 | goto bail; |
0a5a83cf RC |
1172 | } else if (pgaddr == cvt_kvaddr(pd->subport_uregbase + |
1173 | PAGE_SIZE * (subport - 1))) { | |
9929b0fb BS |
1174 | addr = pd->subport_uregbase + PAGE_SIZE * (subport - 1); |
1175 | size = PAGE_SIZE; | |
0a5a83cf RC |
1176 | } else if (pgaddr == cvt_kvaddr(pd->subport_rcvhdr_base + |
1177 | pd->port_rcvhdrq_size * (subport - 1))) { | |
9929b0fb BS |
1178 | addr = pd->subport_rcvhdr_base + |
1179 | pd->port_rcvhdrq_size * (subport - 1); | |
1180 | size = pd->port_rcvhdrq_size; | |
0a5a83cf RC |
1181 | } else if (pgaddr == cvt_kvaddr(pd->subport_rcvegrbuf + |
1182 | size * (subport - 1))) { | |
9929b0fb BS |
1183 | addr = pd->subport_rcvegrbuf + size * (subport - 1); |
1184 | /* rcvegrbufs are read-only on the slave */ | |
1185 | if (vma->vm_flags & VM_WRITE) { | |
1186 | dev_info(&dd->pcidev->dev, | |
1187 | "Can't map eager buffers as " | |
1188 | "writable (flags=%lx)\n", vma->vm_flags); | |
1189 | ret = -EPERM; | |
1190 | goto bail; | |
1191 | } | |
1192 | /* | |
1193 | * Don't allow permission to later change to writeable | |
1194 | * with mprotect. | |
1195 | */ | |
1196 | vma->vm_flags &= ~VM_MAYWRITE; | |
0a5a83cf | 1197 | } else |
9929b0fb | 1198 | goto bail; |
9929b0fb BS |
1199 | len = vma->vm_end - vma->vm_start; |
1200 | if (len > size) { | |
1201 | ipath_cdbg(MM, "FAIL: reqlen %lx > %zx\n", len, size); | |
1202 | ret = -EINVAL; | |
1203 | goto bail; | |
1204 | } | |
1205 | ||
1206 | vma->vm_pgoff = (unsigned long) addr >> PAGE_SHIFT; | |
1207 | vma->vm_ops = &ipath_file_vm_ops; | |
1208 | vma->vm_flags |= VM_RESERVED | VM_DONTEXPAND; | |
0a5a83cf | 1209 | ret = 1; |
9929b0fb BS |
1210 | |
1211 | bail: | |
1212 | return ret; | |
1213 | } | |
1214 | ||
7f510b46 BS |
1215 | /** |
1216 | * ipath_mmap - mmap various structures into user space | |
1217 | * @fp: the file pointer | |
1218 | * @vma: the VM area | |
1219 | * | |
1220 | * We use this to have a shared buffer between the kernel and the user code | |
1221 | * for the rcvhdr queue, egr buffers, and the per-port user regs and pio | |
1222 | * buffers in the chip. We have the open and close entries so we can bump | |
1223 | * the ref count and keep the driver from being unloaded while still mapped. | |
1224 | */ | |
1225 | static int ipath_mmap(struct file *fp, struct vm_area_struct *vma) | |
1226 | { | |
1227 | struct ipath_portdata *pd; | |
1228 | struct ipath_devdata *dd; | |
1229 | u64 pgaddr, ureg; | |
9929b0fb | 1230 | unsigned piobufs, piocnt; |
7f510b46 BS |
1231 | int ret; |
1232 | ||
1233 | pd = port_fp(fp); | |
9929b0fb BS |
1234 | if (!pd) { |
1235 | ret = -EINVAL; | |
1236 | goto bail; | |
1237 | } | |
7f510b46 | 1238 | dd = pd->port_dd; |
f37bda92 | 1239 | |
7f510b46 BS |
1240 | /* |
1241 | * This is the ipath_do_user_init() code, mapping the shared buffers | |
1242 | * into the user process. The address referred to by vm_pgoff is the | |
9929b0fb BS |
1243 | * file offset passed via mmap(). For shared ports, this is the |
1244 | * kernel vmalloc() address of the pages to share with the master. | |
1245 | * For non-shared or master ports, this is a physical address. | |
1246 | * We only do one mmap for each space mapped. | |
7f510b46 BS |
1247 | */ |
1248 | pgaddr = vma->vm_pgoff << PAGE_SHIFT; | |
1249 | ||
1250 | /* | |
9929b0fb BS |
1251 | * Check for 0 in case one of the allocations failed, but user |
1252 | * called mmap anyway. | |
7f510b46 | 1253 | */ |
9929b0fb BS |
1254 | if (!pgaddr) { |
1255 | ret = -EINVAL; | |
1256 | goto bail; | |
f37bda92 | 1257 | } |
7f510b46 | 1258 | |
9929b0fb | 1259 | ipath_cdbg(MM, "pgaddr %llx vm_start=%lx len %lx port %u:%u:%u\n", |
7f510b46 | 1260 | (unsigned long long) pgaddr, vma->vm_start, |
9929b0fb BS |
1261 | vma->vm_end - vma->vm_start, dd->ipath_unit, |
1262 | pd->port_port, subport_fp(fp)); | |
7f510b46 | 1263 | |
9929b0fb BS |
1264 | /* |
1265 | * Physical addresses must fit in 40 bits for our hardware. | |
1266 | * Check for kernel virtual addresses first, anything else must | |
1267 | * match a HW or memory address. | |
1268 | */ | |
0a5a83cf RC |
1269 | ret = mmap_kvaddr(vma, pgaddr, pd, subport_fp(fp)); |
1270 | if (ret) { | |
1271 | if (ret > 0) | |
1272 | ret = 0; | |
9929b0fb BS |
1273 | goto bail; |
1274 | } | |
1275 | ||
1276 | if (!pd->port_subport_cnt) { | |
1277 | /* port is not shared */ | |
1278 | ureg = dd->ipath_uregbase + dd->ipath_palign * pd->port_port; | |
1279 | piocnt = dd->ipath_pbufsport; | |
1280 | piobufs = pd->port_piobufs; | |
1281 | } else if (!subport_fp(fp)) { | |
1282 | /* caller is the master */ | |
1283 | ureg = dd->ipath_uregbase + dd->ipath_palign * pd->port_port; | |
1284 | piocnt = (dd->ipath_pbufsport / pd->port_subport_cnt) + | |
1285 | (dd->ipath_pbufsport % pd->port_subport_cnt); | |
1286 | piobufs = pd->port_piobufs + | |
1287 | dd->ipath_palign * (dd->ipath_pbufsport - piocnt); | |
1288 | } else { | |
1289 | unsigned slave = subport_fp(fp) - 1; | |
1290 | ||
1291 | /* caller is a slave */ | |
1292 | ureg = 0; | |
1293 | piocnt = dd->ipath_pbufsport / pd->port_subport_cnt; | |
1294 | piobufs = pd->port_piobufs + dd->ipath_palign * piocnt * slave; | |
f37bda92 | 1295 | } |
9929b0fb BS |
1296 | |
1297 | if (pgaddr == ureg) | |
7f510b46 | 1298 | ret = mmap_ureg(vma, dd, ureg); |
9929b0fb BS |
1299 | else if (pgaddr == piobufs) |
1300 | ret = mmap_piobufs(vma, dd, pd, piobufs, piocnt); | |
1301 | else if (pgaddr == dd->ipath_pioavailregs_phys) | |
1302 | /* in-memory copy of pioavail registers */ | |
1303 | ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0, | |
1fd3b40f | 1304 | (void *) dd->ipath_pioavailregs_dma, |
9929b0fb BS |
1305 | "pioavail registers"); |
1306 | else if (subport_fp(fp)) | |
1307 | /* Subports don't mmap the physical receive buffers */ | |
1308 | ret = -EINVAL; | |
1309 | else if (pgaddr == pd->port_rcvegr_phys) | |
7f510b46 | 1310 | ret = mmap_rcvegrbufs(vma, pd); |
9929b0fb | 1311 | else if (pgaddr == (u64) pd->port_rcvhdrq_phys) |
f37bda92 | 1312 | /* |
525d0ca1 | 1313 | * The rcvhdrq itself; readonly except on HT (so have |
f37bda92 BS |
1314 | * to allow writable mapping), multiple pages, contiguous |
1315 | * from an i/o perspective. | |
1316 | */ | |
9929b0fb | 1317 | ret = ipath_mmap_mem(vma, pd, pd->port_rcvhdrq_size, 1, |
1fd3b40f | 1318 | pd->port_rcvhdrq, |
f37bda92 | 1319 | "rcvhdrq"); |
9929b0fb | 1320 | else if (pgaddr == (u64) pd->port_rcvhdrqtailaddr_phys) |
f37bda92 BS |
1321 | /* in-memory copy of rcvhdrq tail register */ |
1322 | ret = ipath_mmap_mem(vma, pd, PAGE_SIZE, 0, | |
1fd3b40f | 1323 | pd->port_rcvhdrtail_kvaddr, |
f37bda92 | 1324 | "rcvhdrq tail"); |
7f510b46 BS |
1325 | else |
1326 | ret = -EINVAL; | |
1327 | ||
1328 | vma->vm_private_data = NULL; | |
1329 | ||
1330 | if (ret < 0) | |
1331 | dev_info(&dd->pcidev->dev, | |
9929b0fb BS |
1332 | "Failure %d on off %llx len %lx\n", |
1333 | -ret, (unsigned long long)pgaddr, | |
1334 | vma->vm_end - vma->vm_start); | |
1335 | bail: | |
7f510b46 BS |
1336 | return ret; |
1337 | } | |
1338 | ||
1339 | static unsigned int ipath_poll(struct file *fp, | |
1340 | struct poll_table_struct *pt) | |
1341 | { | |
1342 | struct ipath_portdata *pd; | |
1343 | u32 head, tail; | |
1344 | int bit; | |
e35d710d | 1345 | unsigned pollflag = 0; |
7f510b46 BS |
1346 | struct ipath_devdata *dd; |
1347 | ||
1348 | pd = port_fp(fp); | |
9929b0fb BS |
1349 | if (!pd) |
1350 | goto bail; | |
7f510b46 BS |
1351 | dd = pd->port_dd; |
1352 | ||
1353 | bit = pd->port_port + INFINIPATH_R_INTRAVAIL_SHIFT; | |
1354 | set_bit(bit, &dd->ipath_rcvctrl); | |
1355 | ||
1356 | /* | |
1357 | * Before blocking, make sure that head is still == tail, | |
1358 | * reading from the chip, so we can be sure the interrupt | |
1359 | * enable has made it to the chip. If not equal, disable | |
1360 | * interrupt again and return immediately. This avoids races, | |
1361 | * and the overhead of the chip read doesn't matter much at | |
1362 | * this point, since we are waiting for something anyway. | |
1363 | */ | |
1364 | ||
1365 | ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl, | |
1366 | dd->ipath_rcvctrl); | |
1367 | ||
1368 | head = ipath_read_ureg32(dd, ur_rcvhdrhead, pd->port_port); | |
1369 | tail = ipath_read_ureg32(dd, ur_rcvhdrtail, pd->port_port); | |
1370 | ||
1371 | if (tail == head) { | |
1372 | set_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag); | |
9929b0fb | 1373 | if (dd->ipath_rhdrhead_intr_off) /* arm rcv interrupt */ |
9dcc0e58 BS |
1374 | (void)ipath_write_ureg(dd, ur_rcvhdrhead, |
1375 | dd->ipath_rhdrhead_intr_off | |
1376 | | head, pd->port_port); | |
7f510b46 BS |
1377 | poll_wait(fp, &pd->port_wait, pt); |
1378 | ||
1379 | if (test_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag)) { | |
1380 | /* timed out, no packets received */ | |
1381 | clear_bit(IPATH_PORT_WAITING_RCV, &pd->port_flag); | |
1382 | pd->port_rcvwait_to++; | |
1383 | } | |
e35d710d BS |
1384 | else |
1385 | pollflag = POLLIN | POLLRDNORM; | |
7f510b46 BS |
1386 | } |
1387 | else { | |
1388 | /* it's already happened; don't do wait_event overhead */ | |
e35d710d | 1389 | pollflag = POLLIN | POLLRDNORM; |
7f510b46 BS |
1390 | pd->port_rcvnowait++; |
1391 | } | |
1392 | ||
1393 | clear_bit(bit, &dd->ipath_rcvctrl); | |
1394 | ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl, | |
1395 | dd->ipath_rcvctrl); | |
1396 | ||
9929b0fb | 1397 | bail: |
e35d710d | 1398 | return pollflag; |
7f510b46 BS |
1399 | } |
1400 | ||
9929b0fb BS |
1401 | static int init_subports(struct ipath_devdata *dd, |
1402 | struct ipath_portdata *pd, | |
1403 | const struct ipath_user_info *uinfo) | |
1404 | { | |
1405 | int ret = 0; | |
1406 | unsigned num_slaves; | |
1407 | size_t size; | |
1408 | ||
1409 | /* Old user binaries don't know about subports */ | |
1410 | if ((uinfo->spu_userversion & 0xffff) != IPATH_USER_SWMINOR) | |
1411 | goto bail; | |
1412 | /* | |
1413 | * If the user is requesting zero or one port, | |
1414 | * skip the subport allocation. | |
1415 | */ | |
1416 | if (uinfo->spu_subport_cnt <= 1) | |
1417 | goto bail; | |
0a5a83cf | 1418 | if (uinfo->spu_subport_cnt > INFINIPATH_MAX_SUBPORT) { |
9929b0fb BS |
1419 | ret = -EINVAL; |
1420 | goto bail; | |
1421 | } | |
1422 | ||
1423 | num_slaves = uinfo->spu_subport_cnt - 1; | |
1424 | pd->subport_uregbase = vmalloc(PAGE_SIZE * num_slaves); | |
1425 | if (!pd->subport_uregbase) { | |
1426 | ret = -ENOMEM; | |
1427 | goto bail; | |
1428 | } | |
1429 | /* Note: pd->port_rcvhdrq_size isn't initialized yet. */ | |
1430 | size = ALIGN(dd->ipath_rcvhdrcnt * dd->ipath_rcvhdrentsize * | |
1431 | sizeof(u32), PAGE_SIZE) * num_slaves; | |
1432 | pd->subport_rcvhdr_base = vmalloc(size); | |
1433 | if (!pd->subport_rcvhdr_base) { | |
1434 | ret = -ENOMEM; | |
1435 | goto bail_ureg; | |
1436 | } | |
1437 | ||
1438 | pd->subport_rcvegrbuf = vmalloc(pd->port_rcvegrbuf_chunks * | |
1439 | pd->port_rcvegrbuf_size * | |
1440 | num_slaves); | |
1441 | if (!pd->subport_rcvegrbuf) { | |
1442 | ret = -ENOMEM; | |
1443 | goto bail_rhdr; | |
1444 | } | |
1445 | ||
1446 | pd->port_subport_cnt = uinfo->spu_subport_cnt; | |
1447 | pd->port_subport_id = uinfo->spu_subport_id; | |
1448 | pd->active_slaves = 1; | |
947d7617 | 1449 | set_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag); |
9929b0fb BS |
1450 | goto bail; |
1451 | ||
1452 | bail_rhdr: | |
1453 | vfree(pd->subport_rcvhdr_base); | |
1454 | bail_ureg: | |
1455 | vfree(pd->subport_uregbase); | |
1456 | pd->subport_uregbase = NULL; | |
1457 | bail: | |
1458 | return ret; | |
1459 | } | |
1460 | ||
7f510b46 | 1461 | static int try_alloc_port(struct ipath_devdata *dd, int port, |
9929b0fb BS |
1462 | struct file *fp, |
1463 | const struct ipath_user_info *uinfo) | |
7f510b46 | 1464 | { |
9929b0fb | 1465 | struct ipath_portdata *pd; |
7f510b46 BS |
1466 | int ret; |
1467 | ||
9929b0fb BS |
1468 | if (!(pd = dd->ipath_pd[port])) { |
1469 | void *ptmp; | |
7f510b46 | 1470 | |
9929b0fb | 1471 | pd = kzalloc(sizeof(struct ipath_portdata), GFP_KERNEL); |
7f510b46 BS |
1472 | |
1473 | /* | |
1474 | * Allocate memory for use in ipath_tid_update() just once | |
1475 | * at open, not per call. Reduces cost of expected send | |
1476 | * setup. | |
1477 | */ | |
1478 | ptmp = kmalloc(dd->ipath_rcvtidcnt * sizeof(u16) + | |
1479 | dd->ipath_rcvtidcnt * sizeof(struct page **), | |
1480 | GFP_KERNEL); | |
9929b0fb | 1481 | if (!pd || !ptmp) { |
7f510b46 BS |
1482 | ipath_dev_err(dd, "Unable to allocate portdata " |
1483 | "memory, failing open\n"); | |
1484 | ret = -ENOMEM; | |
9929b0fb | 1485 | kfree(pd); |
7f510b46 BS |
1486 | kfree(ptmp); |
1487 | goto bail; | |
1488 | } | |
9929b0fb | 1489 | dd->ipath_pd[port] = pd; |
7f510b46 BS |
1490 | dd->ipath_pd[port]->port_port = port; |
1491 | dd->ipath_pd[port]->port_dd = dd; | |
1492 | dd->ipath_pd[port]->port_tid_pg_list = ptmp; | |
1493 | init_waitqueue_head(&dd->ipath_pd[port]->port_wait); | |
1494 | } | |
9929b0fb BS |
1495 | if (!pd->port_cnt) { |
1496 | pd->userversion = uinfo->spu_userversion; | |
1497 | init_user_egr_sizes(pd); | |
1498 | if ((ret = init_subports(dd, pd, uinfo)) != 0) | |
1499 | goto bail; | |
7f510b46 BS |
1500 | ipath_cdbg(PROC, "%s[%u] opened unit:port %u:%u\n", |
1501 | current->comm, current->pid, dd->ipath_unit, | |
1502 | port); | |
9929b0fb BS |
1503 | pd->port_cnt = 1; |
1504 | port_fp(fp) = pd; | |
1505 | pd->port_pid = current->pid; | |
1506 | strncpy(pd->port_comm, current->comm, sizeof(pd->port_comm)); | |
7f510b46 BS |
1507 | ipath_stats.sps_ports++; |
1508 | ret = 0; | |
9929b0fb BS |
1509 | } else |
1510 | ret = -EBUSY; | |
7f510b46 BS |
1511 | |
1512 | bail: | |
1513 | return ret; | |
1514 | } | |
1515 | ||
1516 | static inline int usable(struct ipath_devdata *dd) | |
1517 | { | |
1518 | return dd && | |
1519 | (dd->ipath_flags & IPATH_PRESENT) && | |
1520 | dd->ipath_kregbase && | |
1521 | dd->ipath_lid && | |
1522 | !(dd->ipath_flags & (IPATH_LINKDOWN | IPATH_DISABLED | |
1523 | | IPATH_LINKUNK)); | |
1524 | } | |
1525 | ||
9929b0fb BS |
1526 | static int find_free_port(int unit, struct file *fp, |
1527 | const struct ipath_user_info *uinfo) | |
7f510b46 BS |
1528 | { |
1529 | struct ipath_devdata *dd = ipath_lookup(unit); | |
1530 | int ret, i; | |
1531 | ||
1532 | if (!dd) { | |
1533 | ret = -ENODEV; | |
1534 | goto bail; | |
1535 | } | |
1536 | ||
1537 | if (!usable(dd)) { | |
1538 | ret = -ENETDOWN; | |
1539 | goto bail; | |
1540 | } | |
1541 | ||
9929b0fb BS |
1542 | for (i = 1; i < dd->ipath_cfgports; i++) { |
1543 | ret = try_alloc_port(dd, i, fp, uinfo); | |
7f510b46 BS |
1544 | if (ret != -EBUSY) |
1545 | goto bail; | |
1546 | } | |
1547 | ret = -EBUSY; | |
1548 | ||
1549 | bail: | |
1550 | return ret; | |
1551 | } | |
1552 | ||
9929b0fb BS |
1553 | static int find_best_unit(struct file *fp, |
1554 | const struct ipath_user_info *uinfo) | |
7f510b46 BS |
1555 | { |
1556 | int ret = 0, i, prefunit = -1, devmax; | |
1557 | int maxofallports, npresent, nup; | |
1558 | int ndev; | |
1559 | ||
9929b0fb | 1560 | devmax = ipath_count_units(&npresent, &nup, &maxofallports); |
7f510b46 BS |
1561 | |
1562 | /* | |
1563 | * This code is present to allow a knowledgeable person to | |
1564 | * specify the layout of processes to processors before opening | |
1565 | * this driver, and then we'll assign the process to the "closest" | |
525d0ca1 | 1566 | * InfiniPath chip to that processor (we assume reasonable connectivity, |
7f510b46 BS |
1567 | * for now). This code assumes that if affinity has been set |
1568 | * before this point, that at most one cpu is set; for now this | |
1569 | * is reasonable. I check for both cpus_empty() and cpus_full(), | |
1570 | * in case some kernel variant sets none of the bits when no | |
1571 | * affinity is set. 2.6.11 and 12 kernels have all present | |
1572 | * cpus set. Some day we'll have to fix it up further to handle | |
525d0ca1 | 1573 | * a cpu subset. This algorithm fails for two HT chips connected |
7f510b46 BS |
1574 | * in tunnel fashion. Eventually this needs real topology |
1575 | * information. There may be some issues with dual core numbering | |
1576 | * as well. This needs more work prior to release. | |
1577 | */ | |
1578 | if (!cpus_empty(current->cpus_allowed) && | |
1579 | !cpus_full(current->cpus_allowed)) { | |
1580 | int ncpus = num_online_cpus(), curcpu = -1; | |
1581 | for (i = 0; i < ncpus; i++) | |
1582 | if (cpu_isset(i, current->cpus_allowed)) { | |
1583 | ipath_cdbg(PROC, "%s[%u] affinity set for " | |
1584 | "cpu %d\n", current->comm, | |
1585 | current->pid, i); | |
1586 | curcpu = i; | |
1587 | } | |
1588 | if (curcpu != -1) { | |
1589 | if (npresent) { | |
1590 | prefunit = curcpu / (ncpus / npresent); | |
1591 | ipath_dbg("%s[%u] %d chips, %d cpus, " | |
1592 | "%d cpus/chip, select unit %d\n", | |
1593 | current->comm, current->pid, | |
1594 | npresent, ncpus, ncpus / npresent, | |
1595 | prefunit); | |
1596 | } | |
1597 | } | |
1598 | } | |
1599 | ||
1600 | /* | |
1601 | * user ports start at 1, kernel port is 0 | |
1602 | * For now, we do round-robin access across all chips | |
1603 | */ | |
1604 | ||
1605 | if (prefunit != -1) | |
1606 | devmax = prefunit + 1; | |
7f510b46 BS |
1607 | recheck: |
1608 | for (i = 1; i < maxofallports; i++) { | |
1609 | for (ndev = prefunit != -1 ? prefunit : 0; ndev < devmax; | |
1610 | ndev++) { | |
1611 | struct ipath_devdata *dd = ipath_lookup(ndev); | |
1612 | ||
1613 | if (!usable(dd)) | |
1614 | continue; /* can't use this unit */ | |
1615 | if (i >= dd->ipath_cfgports) | |
1616 | /* | |
1617 | * Maxed out on users of this unit. Try | |
1618 | * next. | |
1619 | */ | |
1620 | continue; | |
9929b0fb | 1621 | ret = try_alloc_port(dd, i, fp, uinfo); |
7f510b46 BS |
1622 | if (!ret) |
1623 | goto done; | |
1624 | } | |
1625 | } | |
1626 | ||
1627 | if (npresent) { | |
1628 | if (nup == 0) { | |
1629 | ret = -ENETDOWN; | |
1630 | ipath_dbg("No ports available (none initialized " | |
1631 | "and ready)\n"); | |
1632 | } else { | |
1633 | if (prefunit > 0) { | |
1634 | /* if started above 0, retry from 0 */ | |
1635 | ipath_cdbg(PROC, | |
1636 | "%s[%u] no ports on prefunit " | |
1637 | "%d, clear and re-check\n", | |
1638 | current->comm, current->pid, | |
1639 | prefunit); | |
1640 | devmax = ipath_count_units(NULL, NULL, | |
1641 | NULL); | |
1642 | prefunit = -1; | |
1643 | goto recheck; | |
1644 | } | |
1645 | ret = -EBUSY; | |
1646 | ipath_dbg("No ports available\n"); | |
1647 | } | |
1648 | } else { | |
1649 | ret = -ENXIO; | |
1650 | ipath_dbg("No boards found\n"); | |
1651 | } | |
1652 | ||
1653 | done: | |
1654 | return ret; | |
1655 | } | |
1656 | ||
9929b0fb BS |
1657 | static int find_shared_port(struct file *fp, |
1658 | const struct ipath_user_info *uinfo) | |
1659 | { | |
1660 | int devmax, ndev, i; | |
1661 | int ret = 0; | |
1662 | ||
1663 | devmax = ipath_count_units(NULL, NULL, NULL); | |
1664 | ||
1665 | for (ndev = 0; ndev < devmax; ndev++) { | |
1666 | struct ipath_devdata *dd = ipath_lookup(ndev); | |
1667 | ||
1668 | if (!dd) | |
1669 | continue; | |
1670 | for (i = 1; i < dd->ipath_cfgports; i++) { | |
1671 | struct ipath_portdata *pd = dd->ipath_pd[i]; | |
1672 | ||
1673 | /* Skip ports which are not yet open */ | |
1674 | if (!pd || !pd->port_cnt) | |
1675 | continue; | |
1676 | /* Skip port if it doesn't match the requested one */ | |
1677 | if (pd->port_subport_id != uinfo->spu_subport_id) | |
1678 | continue; | |
1679 | /* Verify the sharing process matches the master */ | |
1680 | if (pd->port_subport_cnt != uinfo->spu_subport_cnt || | |
1681 | pd->userversion != uinfo->spu_userversion || | |
1682 | pd->port_cnt >= pd->port_subport_cnt) { | |
1683 | ret = -EINVAL; | |
1684 | goto done; | |
1685 | } | |
1686 | port_fp(fp) = pd; | |
1687 | subport_fp(fp) = pd->port_cnt++; | |
1688 | tidcursor_fp(fp) = 0; | |
1689 | pd->active_slaves |= 1 << subport_fp(fp); | |
1690 | ipath_cdbg(PROC, | |
1691 | "%s[%u] %u sharing %s[%u] unit:port %u:%u\n", | |
1692 | current->comm, current->pid, | |
1693 | subport_fp(fp), | |
1694 | pd->port_comm, pd->port_pid, | |
1695 | dd->ipath_unit, pd->port_port); | |
1696 | ret = 1; | |
1697 | goto done; | |
1698 | } | |
1699 | } | |
1700 | ||
1701 | done: | |
1702 | return ret; | |
1703 | } | |
1704 | ||
7f510b46 BS |
1705 | static int ipath_open(struct inode *in, struct file *fp) |
1706 | { | |
c97d27d8 | 1707 | /* The real work is performed later in ipath_assign_port() */ |
9929b0fb BS |
1708 | fp->private_data = kzalloc(sizeof(struct ipath_filedata), GFP_KERNEL); |
1709 | return fp->private_data ? 0 : -ENOMEM; | |
1710 | } | |
1711 | ||
c97d27d8 BS |
1712 | |
1713 | /* Get port early, so can set affinity prior to memory allocation */ | |
1714 | static int ipath_assign_port(struct file *fp, | |
9929b0fb BS |
1715 | const struct ipath_user_info *uinfo) |
1716 | { | |
1717 | int ret; | |
9929b0fb BS |
1718 | int i_minor; |
1719 | unsigned swminor; | |
1720 | ||
1721 | /* Check to be sure we haven't already initialized this file */ | |
1722 | if (port_fp(fp)) { | |
1723 | ret = -EINVAL; | |
1724 | goto done; | |
1725 | } | |
1726 | ||
1727 | /* for now, if major version is different, bail */ | |
1728 | if ((uinfo->spu_userversion >> 16) != IPATH_USER_SWMAJOR) { | |
1729 | ipath_dbg("User major version %d not same as driver " | |
1730 | "major %d\n", uinfo->spu_userversion >> 16, | |
1731 | IPATH_USER_SWMAJOR); | |
1732 | ret = -ENODEV; | |
1733 | goto done; | |
1734 | } | |
1735 | ||
1736 | swminor = uinfo->spu_userversion & 0xffff; | |
1737 | if (swminor != IPATH_USER_SWMINOR) | |
1738 | ipath_dbg("User minor version %d not same as driver " | |
1739 | "minor %d\n", swminor, IPATH_USER_SWMINOR); | |
7f510b46 BS |
1740 | |
1741 | mutex_lock(&ipath_mutex); | |
1742 | ||
9929b0fb BS |
1743 | if (swminor == IPATH_USER_SWMINOR && uinfo->spu_subport_cnt && |
1744 | (ret = find_shared_port(fp, uinfo))) { | |
1745 | mutex_unlock(&ipath_mutex); | |
1746 | if (ret > 0) | |
1747 | ret = 0; | |
1748 | goto done; | |
1749 | } | |
1750 | ||
1cfd6e64 | 1751 | i_minor = iminor(fp->f_path.dentry->d_inode) - IPATH_USER_MINOR_BASE; |
7f510b46 | 1752 | ipath_cdbg(VERBOSE, "open on dev %lx (minor %d)\n", |
1cfd6e64 | 1753 | (long)fp->f_path.dentry->d_inode->i_rdev, i_minor); |
7f510b46 | 1754 | |
9929b0fb BS |
1755 | if (i_minor) |
1756 | ret = find_free_port(i_minor - 1, fp, uinfo); | |
7f510b46 | 1757 | else |
9929b0fb | 1758 | ret = find_best_unit(fp, uinfo); |
7f510b46 BS |
1759 | |
1760 | mutex_unlock(&ipath_mutex); | |
9929b0fb | 1761 | |
c97d27d8 BS |
1762 | done: |
1763 | return ret; | |
1764 | } | |
1765 | ||
1766 | ||
1767 | static int ipath_do_user_init(struct file *fp, | |
1768 | const struct ipath_user_info *uinfo) | |
1769 | { | |
1770 | int ret; | |
947d7617 | 1771 | struct ipath_portdata *pd = port_fp(fp); |
c97d27d8 BS |
1772 | struct ipath_devdata *dd; |
1773 | u32 head32; | |
9929b0fb | 1774 | |
947d7617 RC |
1775 | /* Subports don't need to initialize anything since master did it. */ |
1776 | if (subport_fp(fp)) { | |
1777 | ret = wait_event_interruptible(pd->port_wait, | |
1778 | !test_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag)); | |
1779 | goto done; | |
1780 | } | |
1781 | ||
9929b0fb BS |
1782 | dd = pd->port_dd; |
1783 | ||
1784 | if (uinfo->spu_rcvhdrsize) { | |
1785 | ret = ipath_setrcvhdrsize(dd, uinfo->spu_rcvhdrsize); | |
1786 | if (ret) | |
1787 | goto done; | |
1788 | } | |
1789 | ||
1790 | /* for now we do nothing with rcvhdrcnt: uinfo->spu_rcvhdrcnt */ | |
1791 | ||
1792 | /* for right now, kernel piobufs are at end, so port 1 is at 0 */ | |
1793 | pd->port_piobufs = dd->ipath_piobufbase + | |
1794 | dd->ipath_pbufsport * (pd->port_port - 1) * dd->ipath_palign; | |
1795 | ipath_cdbg(VERBOSE, "Set base of piobufs for port %u to 0x%x\n", | |
1796 | pd->port_port, pd->port_piobufs); | |
1797 | ||
1798 | /* | |
1799 | * Now allocate the rcvhdr Q and eager TIDs; skip the TID | |
1800 | * array for time being. If pd->port_port > chip-supported, | |
1801 | * we need to do extra stuff here to handle by handling overflow | |
1802 | * through port 0, someday | |
1803 | */ | |
1804 | ret = ipath_create_rcvhdrq(dd, pd); | |
1805 | if (!ret) | |
1806 | ret = ipath_create_user_egr(pd); | |
1807 | if (ret) | |
1808 | goto done; | |
1809 | ||
1810 | /* | |
1811 | * set the eager head register for this port to the current values | |
1812 | * of the tail pointers, since we don't know if they were | |
1813 | * updated on last use of the port. | |
1814 | */ | |
1815 | head32 = ipath_read_ureg32(dd, ur_rcvegrindextail, pd->port_port); | |
1816 | ipath_write_ureg(dd, ur_rcvegrindexhead, head32, pd->port_port); | |
1817 | dd->ipath_lastegrheads[pd->port_port] = -1; | |
1818 | dd->ipath_lastrcvhdrqtails[pd->port_port] = -1; | |
1819 | ipath_cdbg(VERBOSE, "Wrote port%d egrhead %x from tail regs\n", | |
1820 | pd->port_port, head32); | |
1821 | pd->port_tidcursor = 0; /* start at beginning after open */ | |
1822 | /* | |
1823 | * now enable the port; the tail registers will be written to memory | |
1824 | * by the chip as soon as it sees the write to | |
1825 | * dd->ipath_kregs->kr_rcvctrl. The update only happens on | |
1826 | * transition from 0 to 1, so clear it first, then set it as part of | |
1827 | * enabling the port. This will (very briefly) affect any other | |
1828 | * open ports, but it shouldn't be long enough to be an issue. | |
1829 | * We explictly set the in-memory copy to 0 beforehand, so we don't | |
1830 | * have to wait to be sure the DMA update has happened. | |
1831 | */ | |
1fd3b40f | 1832 | *(volatile u64 *)pd->port_rcvhdrtail_kvaddr = 0ULL; |
9929b0fb BS |
1833 | set_bit(INFINIPATH_R_PORTENABLE_SHIFT + pd->port_port, |
1834 | &dd->ipath_rcvctrl); | |
1835 | ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl, | |
1836 | dd->ipath_rcvctrl & ~INFINIPATH_R_TAILUPD); | |
1837 | ipath_write_kreg(dd, dd->ipath_kregs->kr_rcvctrl, | |
1838 | dd->ipath_rcvctrl); | |
947d7617 RC |
1839 | /* Notify any waiting slaves */ |
1840 | if (pd->port_subport_cnt) { | |
1841 | clear_bit(IPATH_PORT_MASTER_UNINIT, &pd->port_flag); | |
1842 | wake_up(&pd->port_wait); | |
1843 | } | |
9929b0fb | 1844 | done: |
7f510b46 BS |
1845 | return ret; |
1846 | } | |
1847 | ||
1848 | /** | |
1849 | * unlock_exptid - unlock any expected TID entries port still had in use | |
1850 | * @pd: port | |
1851 | * | |
1852 | * We don't actually update the chip here, because we do a bulk update | |
1853 | * below, using ipath_f_clear_tids. | |
1854 | */ | |
1855 | static void unlock_expected_tids(struct ipath_portdata *pd) | |
1856 | { | |
1857 | struct ipath_devdata *dd = pd->port_dd; | |
1858 | int port_tidbase = pd->port_port * dd->ipath_rcvtidcnt; | |
1859 | int i, cnt = 0, maxtid = port_tidbase + dd->ipath_rcvtidcnt; | |
1860 | ||
1861 | ipath_cdbg(VERBOSE, "Port %u unlocking any locked expTID pages\n", | |
1862 | pd->port_port); | |
1863 | for (i = port_tidbase; i < maxtid; i++) { | |
1864 | if (!dd->ipath_pageshadow[i]) | |
1865 | continue; | |
1866 | ||
1fd3b40f BS |
1867 | pci_unmap_page(dd->pcidev, dd->ipath_physshadow[i], |
1868 | PAGE_SIZE, PCI_DMA_FROMDEVICE); | |
7f510b46 BS |
1869 | ipath_release_user_pages_on_close(&dd->ipath_pageshadow[i], |
1870 | 1); | |
1871 | dd->ipath_pageshadow[i] = NULL; | |
1872 | cnt++; | |
1873 | ipath_stats.sps_pageunlocks++; | |
1874 | } | |
1875 | if (cnt) | |
1876 | ipath_cdbg(VERBOSE, "Port %u locked %u expTID entries\n", | |
1877 | pd->port_port, cnt); | |
1878 | ||
1879 | if (ipath_stats.sps_pagelocks || ipath_stats.sps_pageunlocks) | |
1880 | ipath_cdbg(VERBOSE, "%llu pages locked, %llu unlocked\n", | |
1881 | (unsigned long long) ipath_stats.sps_pagelocks, | |
1882 | (unsigned long long) | |
1883 | ipath_stats.sps_pageunlocks); | |
1884 | } | |
1885 | ||
1886 | static int ipath_close(struct inode *in, struct file *fp) | |
1887 | { | |
1888 | int ret = 0; | |
9929b0fb | 1889 | struct ipath_filedata *fd; |
7f510b46 BS |
1890 | struct ipath_portdata *pd; |
1891 | struct ipath_devdata *dd; | |
1892 | unsigned port; | |
1893 | ||
1894 | ipath_cdbg(VERBOSE, "close on dev %lx, private data %p\n", | |
1895 | (long)in->i_rdev, fp->private_data); | |
1896 | ||
1897 | mutex_lock(&ipath_mutex); | |
1898 | ||
9929b0fb | 1899 | fd = (struct ipath_filedata *) fp->private_data; |
7f510b46 | 1900 | fp->private_data = NULL; |
9929b0fb BS |
1901 | pd = fd->pd; |
1902 | if (!pd) { | |
1903 | mutex_unlock(&ipath_mutex); | |
1904 | goto bail; | |
1905 | } | |
1906 | if (--pd->port_cnt) { | |
1907 | /* | |
1908 | * XXX If the master closes the port before the slave(s), | |
1909 | * revoke the mmap for the eager receive queue so | |
1910 | * the slave(s) don't wait for receive data forever. | |
1911 | */ | |
1912 | pd->active_slaves &= ~(1 << fd->subport); | |
1913 | mutex_unlock(&ipath_mutex); | |
1914 | goto bail; | |
1915 | } | |
1916 | port = pd->port_port; | |
7f510b46 BS |
1917 | dd = pd->port_dd; |
1918 | ||
1919 | if (pd->port_hdrqfull) { | |
1920 | ipath_cdbg(PROC, "%s[%u] had %u rcvhdrqfull errors " | |
1921 | "during run\n", pd->port_comm, pd->port_pid, | |
1922 | pd->port_hdrqfull); | |
1923 | pd->port_hdrqfull = 0; | |
1924 | } | |
1925 | ||
1926 | if (pd->port_rcvwait_to || pd->port_piowait_to | |
1927 | || pd->port_rcvnowait || pd->port_pionowait) { | |
1928 | ipath_cdbg(VERBOSE, "port%u, %u rcv, %u pio wait timeo; " | |
1929 | "%u rcv %u, pio already\n", | |
1930 | pd->port_port, pd->port_rcvwait_to, | |
1931 | pd->port_piowait_to, pd->port_rcvnowait, | |
1932 | pd->port_pionowait); | |
1933 | pd->port_rcvwait_to = pd->port_piowait_to = | |
1934 | pd->port_rcvnowait = pd->port_pionowait = 0; | |
1935 | } | |
1936 | if (pd->port_flag) { | |
1937 | ipath_dbg("port %u port_flag still set to 0x%lx\n", | |
1938 | pd->port_port, pd->port_flag); | |
1939 | pd->port_flag = 0; | |
1940 | } | |
1941 | ||
1942 | if (dd->ipath_kregbase) { | |
35783ec0 BS |
1943 | int i; |
1944 | /* atomically clear receive enable port. */ | |
1945 | clear_bit(INFINIPATH_R_PORTENABLE_SHIFT + port, | |
1946 | &dd->ipath_rcvctrl); | |
1947 | ipath_write_kreg( dd, dd->ipath_kregs->kr_rcvctrl, | |
1948 | dd->ipath_rcvctrl); | |
1949 | /* and read back from chip to be sure that nothing | |
1950 | * else is in flight when we do the rest */ | |
1951 | (void)ipath_read_kreg64(dd, dd->ipath_kregs->kr_scratch); | |
7f510b46 BS |
1952 | |
1953 | /* clean up the pkeys for this port user */ | |
1954 | ipath_clean_part_key(pd, dd); | |
35783ec0 BS |
1955 | /* |
1956 | * be paranoid, and never write 0's to these, just use an | |
1957 | * unused part of the port 0 tail page. Of course, | |
1958 | * rcvhdraddr points to a large chunk of memory, so this | |
1959 | * could still trash things, but at least it won't trash | |
1960 | * page 0, and by disabling the port, it should stop "soon", | |
1961 | * even if a packet or two is in already in flight after we | |
1962 | * disabled the port. | |
1963 | */ | |
1964 | ipath_write_kreg_port(dd, | |
1965 | dd->ipath_kregs->kr_rcvhdrtailaddr, port, | |
1966 | dd->ipath_dummy_hdrq_phys); | |
1967 | ipath_write_kreg_port(dd, dd->ipath_kregs->kr_rcvhdraddr, | |
1968 | pd->port_port, dd->ipath_dummy_hdrq_phys); | |
1969 | ||
1970 | i = dd->ipath_pbufsport * (port - 1); | |
1971 | ipath_disarm_piobufs(dd, i, dd->ipath_pbufsport); | |
1972 | ||
1fd3b40f BS |
1973 | dd->ipath_f_clear_tids(dd, pd->port_port); |
1974 | ||
35783ec0 BS |
1975 | if (dd->ipath_pageshadow) |
1976 | unlock_expected_tids(pd); | |
1977 | ipath_stats.sps_ports--; | |
1978 | ipath_cdbg(PROC, "%s[%u] closed port %u:%u\n", | |
1979 | pd->port_comm, pd->port_pid, | |
1980 | dd->ipath_unit, port); | |
7f510b46 BS |
1981 | } |
1982 | ||
7f510b46 | 1983 | pd->port_pid = 0; |
f37bda92 | 1984 | dd->ipath_pd[pd->port_port] = NULL; /* before releasing mutex */ |
7f510b46 | 1985 | mutex_unlock(&ipath_mutex); |
f37bda92 | 1986 | ipath_free_pddata(dd, pd); /* after releasing the mutex */ |
7f510b46 | 1987 | |
9929b0fb BS |
1988 | bail: |
1989 | kfree(fd); | |
7f510b46 BS |
1990 | return ret; |
1991 | } | |
1992 | ||
9929b0fb | 1993 | static int ipath_port_info(struct ipath_portdata *pd, u16 subport, |
7f510b46 BS |
1994 | struct ipath_port_info __user *uinfo) |
1995 | { | |
1996 | struct ipath_port_info info; | |
1997 | int nup; | |
1998 | int ret; | |
9929b0fb | 1999 | size_t sz; |
7f510b46 BS |
2000 | |
2001 | (void) ipath_count_units(NULL, &nup, NULL); | |
2002 | info.num_active = nup; | |
2003 | info.unit = pd->port_dd->ipath_unit; | |
2004 | info.port = pd->port_port; | |
9929b0fb BS |
2005 | info.subport = subport; |
2006 | /* Don't return new fields if old library opened the port. */ | |
2007 | if ((pd->userversion & 0xffff) == IPATH_USER_SWMINOR) { | |
2008 | /* Number of user ports available for this device. */ | |
2009 | info.num_ports = pd->port_dd->ipath_cfgports - 1; | |
2010 | info.num_subports = pd->port_subport_cnt; | |
2011 | sz = sizeof(info); | |
2012 | } else | |
2013 | sz = sizeof(info) - 2 * sizeof(u16); | |
7f510b46 | 2014 | |
9929b0fb | 2015 | if (copy_to_user(uinfo, &info, sz)) { |
7f510b46 BS |
2016 | ret = -EFAULT; |
2017 | goto bail; | |
2018 | } | |
2019 | ret = 0; | |
2020 | ||
2021 | bail: | |
2022 | return ret; | |
2023 | } | |
2024 | ||
9929b0fb BS |
2025 | static int ipath_get_slave_info(struct ipath_portdata *pd, |
2026 | void __user *slave_mask_addr) | |
2027 | { | |
2028 | int ret = 0; | |
2029 | ||
2030 | if (copy_to_user(slave_mask_addr, &pd->active_slaves, sizeof(u32))) | |
2031 | ret = -EFAULT; | |
2032 | return ret; | |
2033 | } | |
2034 | ||
7f510b46 BS |
2035 | static ssize_t ipath_write(struct file *fp, const char __user *data, |
2036 | size_t count, loff_t *off) | |
2037 | { | |
2038 | const struct ipath_cmd __user *ucmd; | |
2039 | struct ipath_portdata *pd; | |
2040 | const void __user *src; | |
2041 | size_t consumed, copy; | |
2042 | struct ipath_cmd cmd; | |
2043 | ssize_t ret = 0; | |
2044 | void *dest; | |
2045 | ||
2046 | if (count < sizeof(cmd.type)) { | |
2047 | ret = -EINVAL; | |
2048 | goto bail; | |
2049 | } | |
2050 | ||
2051 | ucmd = (const struct ipath_cmd __user *) data; | |
2052 | ||
2053 | if (copy_from_user(&cmd.type, &ucmd->type, sizeof(cmd.type))) { | |
2054 | ret = -EFAULT; | |
2055 | goto bail; | |
2056 | } | |
2057 | ||
2058 | consumed = sizeof(cmd.type); | |
2059 | ||
2060 | switch (cmd.type) { | |
c97d27d8 BS |
2061 | case IPATH_CMD_ASSIGN_PORT: |
2062 | case __IPATH_CMD_USER_INIT: | |
7f510b46 BS |
2063 | case IPATH_CMD_USER_INIT: |
2064 | copy = sizeof(cmd.cmd.user_info); | |
2065 | dest = &cmd.cmd.user_info; | |
2066 | src = &ucmd->cmd.user_info; | |
2067 | break; | |
2068 | case IPATH_CMD_RECV_CTRL: | |
2069 | copy = sizeof(cmd.cmd.recv_ctrl); | |
2070 | dest = &cmd.cmd.recv_ctrl; | |
2071 | src = &ucmd->cmd.recv_ctrl; | |
2072 | break; | |
2073 | case IPATH_CMD_PORT_INFO: | |
2074 | copy = sizeof(cmd.cmd.port_info); | |
2075 | dest = &cmd.cmd.port_info; | |
2076 | src = &ucmd->cmd.port_info; | |
2077 | break; | |
2078 | case IPATH_CMD_TID_UPDATE: | |
2079 | case IPATH_CMD_TID_FREE: | |
2080 | copy = sizeof(cmd.cmd.tid_info); | |
2081 | dest = &cmd.cmd.tid_info; | |
2082 | src = &ucmd->cmd.tid_info; | |
2083 | break; | |
2084 | case IPATH_CMD_SET_PART_KEY: | |
2085 | copy = sizeof(cmd.cmd.part_key); | |
2086 | dest = &cmd.cmd.part_key; | |
2087 | src = &ucmd->cmd.part_key; | |
2088 | break; | |
9929b0fb BS |
2089 | case IPATH_CMD_SLAVE_INFO: |
2090 | copy = sizeof(cmd.cmd.slave_mask_addr); | |
2091 | dest = &cmd.cmd.slave_mask_addr; | |
2092 | src = &ucmd->cmd.slave_mask_addr; | |
2093 | break; | |
7f510b46 BS |
2094 | default: |
2095 | ret = -EINVAL; | |
2096 | goto bail; | |
2097 | } | |
2098 | ||
2099 | if ((count - consumed) < copy) { | |
2100 | ret = -EINVAL; | |
2101 | goto bail; | |
2102 | } | |
2103 | ||
2104 | if (copy_from_user(dest, src, copy)) { | |
2105 | ret = -EFAULT; | |
2106 | goto bail; | |
2107 | } | |
2108 | ||
2109 | consumed += copy; | |
2110 | pd = port_fp(fp); | |
c97d27d8 BS |
2111 | if (!pd && cmd.type != __IPATH_CMD_USER_INIT && |
2112 | cmd.type != IPATH_CMD_ASSIGN_PORT) { | |
9929b0fb BS |
2113 | ret = -EINVAL; |
2114 | goto bail; | |
2115 | } | |
7f510b46 BS |
2116 | |
2117 | switch (cmd.type) { | |
c97d27d8 BS |
2118 | case IPATH_CMD_ASSIGN_PORT: |
2119 | ret = ipath_assign_port(fp, &cmd.cmd.user_info); | |
2120 | if (ret) | |
2121 | goto bail; | |
2122 | break; | |
2123 | case __IPATH_CMD_USER_INIT: | |
2124 | /* backwards compatibility, get port first */ | |
2125 | ret = ipath_assign_port(fp, &cmd.cmd.user_info); | |
2126 | if (ret) | |
2127 | goto bail; | |
2128 | /* and fall through to current version. */ | |
7f510b46 | 2129 | case IPATH_CMD_USER_INIT: |
9929b0fb BS |
2130 | ret = ipath_do_user_init(fp, &cmd.cmd.user_info); |
2131 | if (ret) | |
7f510b46 BS |
2132 | goto bail; |
2133 | ret = ipath_get_base_info( | |
9929b0fb | 2134 | fp, (void __user *) (unsigned long) |
7f510b46 BS |
2135 | cmd.cmd.user_info.spu_base_info, |
2136 | cmd.cmd.user_info.spu_base_info_size); | |
2137 | break; | |
2138 | case IPATH_CMD_RECV_CTRL: | |
9929b0fb | 2139 | ret = ipath_manage_rcvq(pd, subport_fp(fp), cmd.cmd.recv_ctrl); |
7f510b46 BS |
2140 | break; |
2141 | case IPATH_CMD_PORT_INFO: | |
9929b0fb | 2142 | ret = ipath_port_info(pd, subport_fp(fp), |
7f510b46 BS |
2143 | (struct ipath_port_info __user *) |
2144 | (unsigned long) cmd.cmd.port_info); | |
2145 | break; | |
2146 | case IPATH_CMD_TID_UPDATE: | |
9929b0fb | 2147 | ret = ipath_tid_update(pd, fp, &cmd.cmd.tid_info); |
7f510b46 BS |
2148 | break; |
2149 | case IPATH_CMD_TID_FREE: | |
9929b0fb | 2150 | ret = ipath_tid_free(pd, subport_fp(fp), &cmd.cmd.tid_info); |
7f510b46 BS |
2151 | break; |
2152 | case IPATH_CMD_SET_PART_KEY: | |
2153 | ret = ipath_set_part_key(pd, cmd.cmd.part_key); | |
2154 | break; | |
9929b0fb BS |
2155 | case IPATH_CMD_SLAVE_INFO: |
2156 | ret = ipath_get_slave_info(pd, | |
2157 | (void __user *) (unsigned long) | |
2158 | cmd.cmd.slave_mask_addr); | |
2159 | break; | |
7f510b46 BS |
2160 | } |
2161 | ||
2162 | if (ret >= 0) | |
2163 | ret = consumed; | |
2164 | ||
2165 | bail: | |
2166 | return ret; | |
2167 | } | |
2168 | ||
2169 | static struct class *ipath_class; | |
2170 | ||
2b8693c0 | 2171 | static int init_cdev(int minor, char *name, const struct file_operations *fops, |
7f510b46 BS |
2172 | struct cdev **cdevp, struct class_device **class_devp) |
2173 | { | |
2174 | const dev_t dev = MKDEV(IPATH_MAJOR, minor); | |
2175 | struct cdev *cdev = NULL; | |
2176 | struct class_device *class_dev = NULL; | |
2177 | int ret; | |
2178 | ||
2179 | cdev = cdev_alloc(); | |
2180 | if (!cdev) { | |
2181 | printk(KERN_ERR IPATH_DRV_NAME | |
2182 | ": Could not allocate cdev for minor %d, %s\n", | |
2183 | minor, name); | |
2184 | ret = -ENOMEM; | |
2185 | goto done; | |
2186 | } | |
2187 | ||
2188 | cdev->owner = THIS_MODULE; | |
2189 | cdev->ops = fops; | |
2190 | kobject_set_name(&cdev->kobj, name); | |
2191 | ||
2192 | ret = cdev_add(cdev, dev, 1); | |
2193 | if (ret < 0) { | |
2194 | printk(KERN_ERR IPATH_DRV_NAME | |
2195 | ": Could not add cdev for minor %d, %s (err %d)\n", | |
2196 | minor, name, -ret); | |
2197 | goto err_cdev; | |
2198 | } | |
2199 | ||
2200 | class_dev = class_device_create(ipath_class, NULL, dev, NULL, name); | |
2201 | ||
2202 | if (IS_ERR(class_dev)) { | |
2203 | ret = PTR_ERR(class_dev); | |
2204 | printk(KERN_ERR IPATH_DRV_NAME ": Could not create " | |
2205 | "class_dev for minor %d, %s (err %d)\n", | |
2206 | minor, name, -ret); | |
2207 | goto err_cdev; | |
2208 | } | |
2209 | ||
2210 | goto done; | |
2211 | ||
2212 | err_cdev: | |
2213 | cdev_del(cdev); | |
2214 | cdev = NULL; | |
2215 | ||
2216 | done: | |
2217 | if (ret >= 0) { | |
2218 | *cdevp = cdev; | |
2219 | *class_devp = class_dev; | |
2220 | } else { | |
2221 | *cdevp = NULL; | |
2222 | *class_devp = NULL; | |
2223 | } | |
2224 | ||
2225 | return ret; | |
2226 | } | |
2227 | ||
2b8693c0 | 2228 | int ipath_cdev_init(int minor, char *name, const struct file_operations *fops, |
7f510b46 BS |
2229 | struct cdev **cdevp, struct class_device **class_devp) |
2230 | { | |
2231 | return init_cdev(minor, name, fops, cdevp, class_devp); | |
2232 | } | |
2233 | ||
2234 | static void cleanup_cdev(struct cdev **cdevp, | |
2235 | struct class_device **class_devp) | |
2236 | { | |
2237 | struct class_device *class_dev = *class_devp; | |
2238 | ||
2239 | if (class_dev) { | |
2240 | class_device_unregister(class_dev); | |
2241 | *class_devp = NULL; | |
2242 | } | |
2243 | ||
2244 | if (*cdevp) { | |
2245 | cdev_del(*cdevp); | |
2246 | *cdevp = NULL; | |
2247 | } | |
2248 | } | |
2249 | ||
2250 | void ipath_cdev_cleanup(struct cdev **cdevp, | |
2251 | struct class_device **class_devp) | |
2252 | { | |
2253 | cleanup_cdev(cdevp, class_devp); | |
2254 | } | |
2255 | ||
2256 | static struct cdev *wildcard_cdev; | |
2257 | static struct class_device *wildcard_class_dev; | |
2258 | ||
2259 | static const dev_t dev = MKDEV(IPATH_MAJOR, 0); | |
2260 | ||
2261 | static int user_init(void) | |
2262 | { | |
2263 | int ret; | |
2264 | ||
2265 | ret = register_chrdev_region(dev, IPATH_NMINORS, IPATH_DRV_NAME); | |
2266 | if (ret < 0) { | |
2267 | printk(KERN_ERR IPATH_DRV_NAME ": Could not register " | |
2268 | "chrdev region (err %d)\n", -ret); | |
2269 | goto done; | |
2270 | } | |
2271 | ||
2272 | ipath_class = class_create(THIS_MODULE, IPATH_DRV_NAME); | |
2273 | ||
2274 | if (IS_ERR(ipath_class)) { | |
2275 | ret = PTR_ERR(ipath_class); | |
2276 | printk(KERN_ERR IPATH_DRV_NAME ": Could not create " | |
2277 | "device class (err %d)\n", -ret); | |
2278 | goto bail; | |
2279 | } | |
2280 | ||
2281 | goto done; | |
2282 | bail: | |
2283 | unregister_chrdev_region(dev, IPATH_NMINORS); | |
2284 | done: | |
2285 | return ret; | |
2286 | } | |
2287 | ||
2288 | static void user_cleanup(void) | |
2289 | { | |
2290 | if (ipath_class) { | |
2291 | class_destroy(ipath_class); | |
2292 | ipath_class = NULL; | |
2293 | } | |
2294 | ||
2295 | unregister_chrdev_region(dev, IPATH_NMINORS); | |
2296 | } | |
2297 | ||
2298 | static atomic_t user_count = ATOMIC_INIT(0); | |
2299 | static atomic_t user_setup = ATOMIC_INIT(0); | |
2300 | ||
2301 | int ipath_user_add(struct ipath_devdata *dd) | |
2302 | { | |
2303 | char name[10]; | |
2304 | int ret; | |
2305 | ||
2306 | if (atomic_inc_return(&user_count) == 1) { | |
2307 | ret = user_init(); | |
2308 | if (ret < 0) { | |
2309 | ipath_dev_err(dd, "Unable to set up user support: " | |
2310 | "error %d\n", -ret); | |
2311 | goto bail; | |
2312 | } | |
7f510b46 BS |
2313 | ret = init_cdev(0, "ipath", &ipath_file_ops, &wildcard_cdev, |
2314 | &wildcard_class_dev); | |
2315 | if (ret < 0) { | |
2316 | ipath_dev_err(dd, "Could not create wildcard " | |
2317 | "minor: error %d\n", -ret); | |
0fd41363 | 2318 | goto bail_user; |
7f510b46 BS |
2319 | } |
2320 | ||
2321 | atomic_set(&user_setup, 1); | |
2322 | } | |
2323 | ||
2324 | snprintf(name, sizeof(name), "ipath%d", dd->ipath_unit); | |
2325 | ||
2326 | ret = init_cdev(dd->ipath_unit + 1, name, &ipath_file_ops, | |
a2acb2ff | 2327 | &dd->user_cdev, &dd->user_class_dev); |
7f510b46 BS |
2328 | if (ret < 0) |
2329 | ipath_dev_err(dd, "Could not create user minor %d, %s\n", | |
2330 | dd->ipath_unit + 1, name); | |
2331 | ||
2332 | goto bail; | |
2333 | ||
0fd41363 | 2334 | bail_user: |
7f510b46 BS |
2335 | user_cleanup(); |
2336 | bail: | |
2337 | return ret; | |
2338 | } | |
2339 | ||
a2acb2ff | 2340 | void ipath_user_remove(struct ipath_devdata *dd) |
7f510b46 | 2341 | { |
a2acb2ff | 2342 | cleanup_cdev(&dd->user_cdev, &dd->user_class_dev); |
7f510b46 BS |
2343 | |
2344 | if (atomic_dec_return(&user_count) == 0) { | |
2345 | if (atomic_read(&user_setup) == 0) | |
2346 | goto bail; | |
2347 | ||
2348 | cleanup_cdev(&wildcard_cdev, &wildcard_class_dev); | |
7f510b46 BS |
2349 | user_cleanup(); |
2350 | ||
2351 | atomic_set(&user_setup, 0); | |
2352 | } | |
2353 | bail: | |
2354 | return; | |
2355 | } |