]>
Commit | Line | Data |
---|---|---|
f931551b RC |
1 | /* |
2 | * Copyright (c) 2006, 2007, 2008, 2009, 2010 QLogic Corporation. | |
3 | * All rights reserved. | |
4 | * Copyright (c) 2003, 2004, 2005, 2006 PathScale, Inc. All rights reserved. | |
5 | * | |
6 | * This software is available to you under a choice of one of two | |
7 | * licenses. You may choose to be licensed under the terms of the GNU | |
8 | * General Public License (GPL) Version 2, available from the file | |
9 | * COPYING in the main directory of this source tree, or the | |
10 | * OpenIB.org BSD license below: | |
11 | * | |
12 | * Redistribution and use in source and binary forms, with or | |
13 | * without modification, are permitted provided that the following | |
14 | * conditions are met: | |
15 | * | |
16 | * - Redistributions of source code must retain the above | |
17 | * copyright notice, this list of conditions and the following | |
18 | * disclaimer. | |
19 | * | |
20 | * - Redistributions in binary form must reproduce the above | |
21 | * copyright notice, this list of conditions and the following | |
22 | * disclaimer in the documentation and/or other materials | |
23 | * provided with the distribution. | |
24 | * | |
25 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, | |
26 | * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF | |
27 | * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND | |
28 | * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS | |
29 | * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN | |
30 | * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN | |
31 | * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE | |
32 | * SOFTWARE. | |
33 | */ | |
34 | ||
35 | #include <linux/pci.h> | |
36 | #include <linux/netdevice.h> | |
37 | #include <linux/vmalloc.h> | |
38 | #include <linux/delay.h> | |
39 | #include <linux/idr.h> | |
40 | ||
41 | #include "qib.h" | |
42 | #include "qib_common.h" | |
43 | ||
44 | /* | |
45 | * min buffers we want to have per context, after driver | |
46 | */ | |
47 | #define QIB_MIN_USER_CTXT_BUFCNT 7 | |
48 | ||
49 | #define QLOGIC_IB_R_SOFTWARE_MASK 0xFF | |
50 | #define QLOGIC_IB_R_SOFTWARE_SHIFT 24 | |
51 | #define QLOGIC_IB_R_EMULATOR_MASK (1ULL<<62) | |
52 | ||
53 | /* | |
54 | * Number of ctxts we are configured to use (to allow for more pio | |
55 | * buffers per ctxt, etc.) Zero means use chip value. | |
56 | */ | |
57 | ushort qib_cfgctxts; | |
58 | module_param_named(cfgctxts, qib_cfgctxts, ushort, S_IRUGO); | |
59 | MODULE_PARM_DESC(cfgctxts, "Set max number of contexts to use"); | |
60 | ||
61 | /* | |
62 | * If set, do not write to any regs if avoidable, hack to allow | |
63 | * check for deranged default register values. | |
64 | */ | |
65 | ushort qib_mini_init; | |
66 | module_param_named(mini_init, qib_mini_init, ushort, S_IRUGO); | |
67 | MODULE_PARM_DESC(mini_init, "If set, do minimal diag init"); | |
68 | ||
69 | unsigned qib_n_krcv_queues; | |
70 | module_param_named(krcvqs, qib_n_krcv_queues, uint, S_IRUGO); | |
71 | MODULE_PARM_DESC(krcvqs, "number of kernel receive queues per IB port"); | |
72 | ||
73 | /* | |
74 | * qib_wc_pat parameter: | |
75 | * 0 is WC via MTRR | |
76 | * 1 is WC via PAT | |
77 | * If PAT initialization fails, code reverts back to MTRR | |
78 | */ | |
79 | unsigned qib_wc_pat = 1; /* default (1) is to use PAT, not MTRR */ | |
80 | module_param_named(wc_pat, qib_wc_pat, uint, S_IRUGO); | |
81 | MODULE_PARM_DESC(wc_pat, "enable write-combining via PAT mechanism"); | |
82 | ||
83 | struct workqueue_struct *qib_wq; | |
84 | struct workqueue_struct *qib_cq_wq; | |
85 | ||
86 | static void verify_interrupt(unsigned long); | |
87 | ||
88 | static struct idr qib_unit_table; | |
89 | u32 qib_cpulist_count; | |
90 | unsigned long *qib_cpulist; | |
91 | ||
92 | /* set number of contexts we'll actually use */ | |
93 | void qib_set_ctxtcnt(struct qib_devdata *dd) | |
94 | { | |
95 | if (!qib_cfgctxts) | |
96 | dd->cfgctxts = dd->ctxtcnt; | |
97 | else if (qib_cfgctxts < dd->num_pports) | |
98 | dd->cfgctxts = dd->ctxtcnt; | |
99 | else if (qib_cfgctxts <= dd->ctxtcnt) | |
100 | dd->cfgctxts = qib_cfgctxts; | |
101 | else | |
102 | dd->cfgctxts = dd->ctxtcnt; | |
103 | } | |
104 | ||
105 | /* | |
106 | * Common code for creating the receive context array. | |
107 | */ | |
108 | int qib_create_ctxts(struct qib_devdata *dd) | |
109 | { | |
110 | unsigned i; | |
111 | int ret; | |
112 | ||
113 | /* | |
114 | * Allocate full ctxtcnt array, rather than just cfgctxts, because | |
115 | * cleanup iterates across all possible ctxts. | |
116 | */ | |
117 | dd->rcd = kzalloc(sizeof(*dd->rcd) * dd->ctxtcnt, GFP_KERNEL); | |
118 | if (!dd->rcd) { | |
119 | qib_dev_err(dd, "Unable to allocate ctxtdata array, " | |
120 | "failing\n"); | |
121 | ret = -ENOMEM; | |
122 | goto done; | |
123 | } | |
124 | ||
125 | /* create (one or more) kctxt */ | |
126 | for (i = 0; i < dd->first_user_ctxt; ++i) { | |
127 | struct qib_pportdata *ppd; | |
128 | struct qib_ctxtdata *rcd; | |
129 | ||
130 | if (dd->skip_kctxt_mask & (1 << i)) | |
131 | continue; | |
132 | ||
133 | ppd = dd->pport + (i % dd->num_pports); | |
134 | rcd = qib_create_ctxtdata(ppd, i); | |
135 | if (!rcd) { | |
136 | qib_dev_err(dd, "Unable to allocate ctxtdata" | |
137 | " for Kernel ctxt, failing\n"); | |
138 | ret = -ENOMEM; | |
139 | goto done; | |
140 | } | |
141 | rcd->pkeys[0] = QIB_DEFAULT_P_KEY; | |
142 | rcd->seq_cnt = 1; | |
143 | } | |
144 | ret = 0; | |
145 | done: | |
146 | return ret; | |
147 | } | |
148 | ||
149 | /* | |
150 | * Common code for user and kernel context setup. | |
151 | */ | |
152 | struct qib_ctxtdata *qib_create_ctxtdata(struct qib_pportdata *ppd, u32 ctxt) | |
153 | { | |
154 | struct qib_devdata *dd = ppd->dd; | |
155 | struct qib_ctxtdata *rcd; | |
156 | ||
157 | rcd = kzalloc(sizeof(*rcd), GFP_KERNEL); | |
158 | if (rcd) { | |
159 | INIT_LIST_HEAD(&rcd->qp_wait_list); | |
160 | rcd->ppd = ppd; | |
161 | rcd->dd = dd; | |
162 | rcd->cnt = 1; | |
163 | rcd->ctxt = ctxt; | |
164 | dd->rcd[ctxt] = rcd; | |
165 | ||
166 | dd->f_init_ctxt(rcd); | |
167 | ||
168 | /* | |
169 | * To avoid wasting a lot of memory, we allocate 32KB chunks | |
170 | * of physically contiguous memory, advance through it until | |
171 | * used up and then allocate more. Of course, we need | |
172 | * memory to store those extra pointers, now. 32KB seems to | |
173 | * be the most that is "safe" under memory pressure | |
174 | * (creating large files and then copying them over | |
175 | * NFS while doing lots of MPI jobs). The OOM killer can | |
176 | * get invoked, even though we say we can sleep and this can | |
177 | * cause significant system problems.... | |
178 | */ | |
179 | rcd->rcvegrbuf_size = 0x8000; | |
180 | rcd->rcvegrbufs_perchunk = | |
181 | rcd->rcvegrbuf_size / dd->rcvegrbufsize; | |
182 | rcd->rcvegrbuf_chunks = (rcd->rcvegrcnt + | |
183 | rcd->rcvegrbufs_perchunk - 1) / | |
184 | rcd->rcvegrbufs_perchunk; | |
185 | } | |
186 | return rcd; | |
187 | } | |
188 | ||
189 | /* | |
190 | * Common code for initializing the physical port structure. | |
191 | */ | |
192 | void qib_init_pportdata(struct qib_pportdata *ppd, struct qib_devdata *dd, | |
193 | u8 hw_pidx, u8 port) | |
194 | { | |
195 | ppd->dd = dd; | |
196 | ppd->hw_pidx = hw_pidx; | |
197 | ppd->port = port; /* IB port number, not index */ | |
198 | ||
199 | spin_lock_init(&ppd->sdma_lock); | |
200 | spin_lock_init(&ppd->lflags_lock); | |
201 | init_waitqueue_head(&ppd->state_wait); | |
202 | ||
203 | init_timer(&ppd->symerr_clear_timer); | |
204 | ppd->symerr_clear_timer.function = qib_clear_symerror_on_linkup; | |
205 | ppd->symerr_clear_timer.data = (unsigned long)ppd; | |
206 | } | |
207 | ||
208 | static int init_pioavailregs(struct qib_devdata *dd) | |
209 | { | |
210 | int ret, pidx; | |
211 | u64 *status_page; | |
212 | ||
213 | dd->pioavailregs_dma = dma_alloc_coherent( | |
214 | &dd->pcidev->dev, PAGE_SIZE, &dd->pioavailregs_phys, | |
215 | GFP_KERNEL); | |
216 | if (!dd->pioavailregs_dma) { | |
217 | qib_dev_err(dd, "failed to allocate PIOavail reg area " | |
218 | "in memory\n"); | |
219 | ret = -ENOMEM; | |
220 | goto done; | |
221 | } | |
222 | ||
223 | /* | |
224 | * We really want L2 cache aligned, but for current CPUs of | |
225 | * interest, they are the same. | |
226 | */ | |
227 | status_page = (u64 *) | |
228 | ((char *) dd->pioavailregs_dma + | |
229 | ((2 * L1_CACHE_BYTES + | |
230 | dd->pioavregs * sizeof(u64)) & ~L1_CACHE_BYTES)); | |
231 | /* device status comes first, for backwards compatibility */ | |
232 | dd->devstatusp = status_page; | |
233 | *status_page++ = 0; | |
234 | for (pidx = 0; pidx < dd->num_pports; ++pidx) { | |
235 | dd->pport[pidx].statusp = status_page; | |
236 | *status_page++ = 0; | |
237 | } | |
238 | ||
239 | /* | |
240 | * Setup buffer to hold freeze and other messages, accessible to | |
241 | * apps, following statusp. This is per-unit, not per port. | |
242 | */ | |
243 | dd->freezemsg = (char *) status_page; | |
244 | *dd->freezemsg = 0; | |
245 | /* length of msg buffer is "whatever is left" */ | |
246 | ret = (char *) status_page - (char *) dd->pioavailregs_dma; | |
247 | dd->freezelen = PAGE_SIZE - ret; | |
248 | ||
249 | ret = 0; | |
250 | ||
251 | done: | |
252 | return ret; | |
253 | } | |
254 | ||
255 | /** | |
256 | * init_shadow_tids - allocate the shadow TID array | |
257 | * @dd: the qlogic_ib device | |
258 | * | |
259 | * allocate the shadow TID array, so we can qib_munlock previous | |
260 | * entries. It may make more sense to move the pageshadow to the | |
261 | * ctxt data structure, so we only allocate memory for ctxts actually | |
262 | * in use, since we at 8k per ctxt, now. | |
263 | * We don't want failures here to prevent use of the driver/chip, | |
264 | * so no return value. | |
265 | */ | |
266 | static void init_shadow_tids(struct qib_devdata *dd) | |
267 | { | |
268 | struct page **pages; | |
269 | dma_addr_t *addrs; | |
270 | ||
271 | pages = vmalloc(dd->cfgctxts * dd->rcvtidcnt * sizeof(struct page *)); | |
272 | if (!pages) { | |
273 | qib_dev_err(dd, "failed to allocate shadow page * " | |
274 | "array, no expected sends!\n"); | |
275 | goto bail; | |
276 | } | |
277 | ||
278 | addrs = vmalloc(dd->cfgctxts * dd->rcvtidcnt * sizeof(dma_addr_t)); | |
279 | if (!addrs) { | |
280 | qib_dev_err(dd, "failed to allocate shadow dma handle " | |
281 | "array, no expected sends!\n"); | |
282 | goto bail_free; | |
283 | } | |
284 | ||
285 | memset(pages, 0, dd->cfgctxts * dd->rcvtidcnt * sizeof(struct page *)); | |
286 | memset(addrs, 0, dd->cfgctxts * dd->rcvtidcnt * sizeof(dma_addr_t)); | |
287 | ||
288 | dd->pageshadow = pages; | |
289 | dd->physshadow = addrs; | |
290 | return; | |
291 | ||
292 | bail_free: | |
293 | vfree(pages); | |
294 | bail: | |
295 | dd->pageshadow = NULL; | |
296 | } | |
297 | ||
298 | /* | |
299 | * Do initialization for device that is only needed on | |
300 | * first detect, not on resets. | |
301 | */ | |
302 | static int loadtime_init(struct qib_devdata *dd) | |
303 | { | |
304 | int ret = 0; | |
305 | ||
306 | if (((dd->revision >> QLOGIC_IB_R_SOFTWARE_SHIFT) & | |
307 | QLOGIC_IB_R_SOFTWARE_MASK) != QIB_CHIP_SWVERSION) { | |
308 | qib_dev_err(dd, "Driver only handles version %d, " | |
309 | "chip swversion is %d (%llx), failng\n", | |
310 | QIB_CHIP_SWVERSION, | |
311 | (int)(dd->revision >> | |
312 | QLOGIC_IB_R_SOFTWARE_SHIFT) & | |
313 | QLOGIC_IB_R_SOFTWARE_MASK, | |
314 | (unsigned long long) dd->revision); | |
315 | ret = -ENOSYS; | |
316 | goto done; | |
317 | } | |
318 | ||
319 | if (dd->revision & QLOGIC_IB_R_EMULATOR_MASK) | |
320 | qib_devinfo(dd->pcidev, "%s", dd->boardversion); | |
321 | ||
322 | spin_lock_init(&dd->pioavail_lock); | |
323 | spin_lock_init(&dd->sendctrl_lock); | |
324 | spin_lock_init(&dd->uctxt_lock); | |
325 | spin_lock_init(&dd->qib_diag_trans_lock); | |
326 | spin_lock_init(&dd->eep_st_lock); | |
327 | mutex_init(&dd->eep_lock); | |
328 | ||
329 | if (qib_mini_init) | |
330 | goto done; | |
331 | ||
332 | ret = init_pioavailregs(dd); | |
333 | init_shadow_tids(dd); | |
334 | ||
335 | qib_get_eeprom_info(dd); | |
336 | ||
337 | /* setup time (don't start yet) to verify we got interrupt */ | |
338 | init_timer(&dd->intrchk_timer); | |
339 | dd->intrchk_timer.function = verify_interrupt; | |
340 | dd->intrchk_timer.data = (unsigned long) dd; | |
341 | ||
342 | done: | |
343 | return ret; | |
344 | } | |
345 | ||
346 | /** | |
347 | * init_after_reset - re-initialize after a reset | |
348 | * @dd: the qlogic_ib device | |
349 | * | |
350 | * sanity check at least some of the values after reset, and | |
351 | * ensure no receive or transmit (explictly, in case reset | |
352 | * failed | |
353 | */ | |
354 | static int init_after_reset(struct qib_devdata *dd) | |
355 | { | |
356 | int i; | |
357 | ||
358 | /* | |
359 | * Ensure chip does no sends or receives, tail updates, or | |
360 | * pioavail updates while we re-initialize. This is mostly | |
361 | * for the driver data structures, not chip registers. | |
362 | */ | |
363 | for (i = 0; i < dd->num_pports; ++i) { | |
364 | /* | |
365 | * ctxt == -1 means "all contexts". Only really safe for | |
366 | * _dis_abling things, as here. | |
367 | */ | |
368 | dd->f_rcvctrl(dd->pport + i, QIB_RCVCTRL_CTXT_DIS | | |
369 | QIB_RCVCTRL_INTRAVAIL_DIS | | |
370 | QIB_RCVCTRL_TAILUPD_DIS, -1); | |
371 | /* Redundant across ports for some, but no big deal. */ | |
372 | dd->f_sendctrl(dd->pport + i, QIB_SENDCTRL_SEND_DIS | | |
373 | QIB_SENDCTRL_AVAIL_DIS); | |
374 | } | |
375 | ||
376 | return 0; | |
377 | } | |
378 | ||
379 | static void enable_chip(struct qib_devdata *dd) | |
380 | { | |
381 | u64 rcvmask; | |
382 | int i; | |
383 | ||
384 | /* | |
385 | * Enable PIO send, and update of PIOavail regs to memory. | |
386 | */ | |
387 | for (i = 0; i < dd->num_pports; ++i) | |
388 | dd->f_sendctrl(dd->pport + i, QIB_SENDCTRL_SEND_ENB | | |
389 | QIB_SENDCTRL_AVAIL_ENB); | |
390 | /* | |
391 | * Enable kernel ctxts' receive and receive interrupt. | |
392 | * Other ctxts done as user opens and inits them. | |
393 | */ | |
394 | rcvmask = QIB_RCVCTRL_CTXT_ENB | QIB_RCVCTRL_INTRAVAIL_ENB; | |
395 | rcvmask |= (dd->flags & QIB_NODMA_RTAIL) ? | |
396 | QIB_RCVCTRL_TAILUPD_DIS : QIB_RCVCTRL_TAILUPD_ENB; | |
397 | for (i = 0; dd->rcd && i < dd->first_user_ctxt; ++i) { | |
398 | struct qib_ctxtdata *rcd = dd->rcd[i]; | |
399 | ||
400 | if (rcd) | |
401 | dd->f_rcvctrl(rcd->ppd, rcvmask, i); | |
402 | } | |
403 | } | |
404 | ||
405 | static void verify_interrupt(unsigned long opaque) | |
406 | { | |
407 | struct qib_devdata *dd = (struct qib_devdata *) opaque; | |
408 | ||
409 | if (!dd) | |
410 | return; /* being torn down */ | |
411 | ||
412 | /* | |
413 | * If we don't have a lid or any interrupts, let the user know and | |
414 | * don't bother checking again. | |
415 | */ | |
416 | if (dd->int_counter == 0) { | |
417 | if (!dd->f_intr_fallback(dd)) | |
418 | dev_err(&dd->pcidev->dev, "No interrupts detected, " | |
419 | "not usable.\n"); | |
420 | else /* re-arm the timer to see if fallback works */ | |
421 | mod_timer(&dd->intrchk_timer, jiffies + HZ/2); | |
422 | } | |
423 | } | |
424 | ||
425 | static void init_piobuf_state(struct qib_devdata *dd) | |
426 | { | |
427 | int i, pidx; | |
428 | u32 uctxts; | |
429 | ||
430 | /* | |
431 | * Ensure all buffers are free, and fifos empty. Buffers | |
432 | * are common, so only do once for port 0. | |
433 | * | |
434 | * After enable and qib_chg_pioavailkernel so we can safely | |
435 | * enable pioavail updates and PIOENABLE. After this, packets | |
436 | * are ready and able to go out. | |
437 | */ | |
438 | dd->f_sendctrl(dd->pport, QIB_SENDCTRL_DISARM_ALL); | |
439 | for (pidx = 0; pidx < dd->num_pports; ++pidx) | |
440 | dd->f_sendctrl(dd->pport + pidx, QIB_SENDCTRL_FLUSH); | |
441 | ||
442 | /* | |
443 | * If not all sendbufs are used, add the one to each of the lower | |
444 | * numbered contexts. pbufsctxt and lastctxt_piobuf are | |
445 | * calculated in chip-specific code because it may cause some | |
446 | * chip-specific adjustments to be made. | |
447 | */ | |
448 | uctxts = dd->cfgctxts - dd->first_user_ctxt; | |
449 | dd->ctxts_extrabuf = dd->pbufsctxt ? | |
450 | dd->lastctxt_piobuf - (dd->pbufsctxt * uctxts) : 0; | |
451 | ||
452 | /* | |
453 | * Set up the shadow copies of the piobufavail registers, | |
454 | * which we compare against the chip registers for now, and | |
455 | * the in memory DMA'ed copies of the registers. | |
456 | * By now pioavail updates to memory should have occurred, so | |
457 | * copy them into our working/shadow registers; this is in | |
458 | * case something went wrong with abort, but mostly to get the | |
459 | * initial values of the generation bit correct. | |
460 | */ | |
461 | for (i = 0; i < dd->pioavregs; i++) { | |
462 | __le64 tmp; | |
463 | ||
464 | tmp = dd->pioavailregs_dma[i]; | |
465 | /* | |
466 | * Don't need to worry about pioavailkernel here | |
467 | * because we will call qib_chg_pioavailkernel() later | |
468 | * in initialization, to busy out buffers as needed. | |
469 | */ | |
470 | dd->pioavailshadow[i] = le64_to_cpu(tmp); | |
471 | } | |
472 | while (i < ARRAY_SIZE(dd->pioavailshadow)) | |
473 | dd->pioavailshadow[i++] = 0; /* for debugging sanity */ | |
474 | ||
475 | /* after pioavailshadow is setup */ | |
476 | qib_chg_pioavailkernel(dd, 0, dd->piobcnt2k + dd->piobcnt4k, | |
477 | TXCHK_CHG_TYPE_KERN, NULL); | |
478 | dd->f_initvl15_bufs(dd); | |
479 | } | |
480 | ||
481 | /** | |
482 | * qib_init - do the actual initialization sequence on the chip | |
483 | * @dd: the qlogic_ib device | |
484 | * @reinit: reinitializing, so don't allocate new memory | |
485 | * | |
486 | * Do the actual initialization sequence on the chip. This is done | |
487 | * both from the init routine called from the PCI infrastructure, and | |
488 | * when we reset the chip, or detect that it was reset internally, | |
489 | * or it's administratively re-enabled. | |
490 | * | |
491 | * Memory allocation here and in called routines is only done in | |
492 | * the first case (reinit == 0). We have to be careful, because even | |
493 | * without memory allocation, we need to re-write all the chip registers | |
494 | * TIDs, etc. after the reset or enable has completed. | |
495 | */ | |
496 | int qib_init(struct qib_devdata *dd, int reinit) | |
497 | { | |
498 | int ret = 0, pidx, lastfail = 0; | |
499 | u32 portok = 0; | |
500 | unsigned i; | |
501 | struct qib_ctxtdata *rcd; | |
502 | struct qib_pportdata *ppd; | |
503 | unsigned long flags; | |
504 | ||
505 | /* Set linkstate to unknown, so we can watch for a transition. */ | |
506 | for (pidx = 0; pidx < dd->num_pports; ++pidx) { | |
507 | ppd = dd->pport + pidx; | |
508 | spin_lock_irqsave(&ppd->lflags_lock, flags); | |
509 | ppd->lflags &= ~(QIBL_LINKACTIVE | QIBL_LINKARMED | | |
510 | QIBL_LINKDOWN | QIBL_LINKINIT | | |
511 | QIBL_LINKV); | |
512 | spin_unlock_irqrestore(&ppd->lflags_lock, flags); | |
513 | } | |
514 | ||
515 | if (reinit) | |
516 | ret = init_after_reset(dd); | |
517 | else | |
518 | ret = loadtime_init(dd); | |
519 | if (ret) | |
520 | goto done; | |
521 | ||
522 | /* Bypass most chip-init, to get to device creation */ | |
523 | if (qib_mini_init) | |
524 | return 0; | |
525 | ||
526 | ret = dd->f_late_initreg(dd); | |
527 | if (ret) | |
528 | goto done; | |
529 | ||
530 | /* dd->rcd can be NULL if early init failed */ | |
531 | for (i = 0; dd->rcd && i < dd->first_user_ctxt; ++i) { | |
532 | /* | |
533 | * Set up the (kernel) rcvhdr queue and egr TIDs. If doing | |
534 | * re-init, the simplest way to handle this is to free | |
535 | * existing, and re-allocate. | |
536 | * Need to re-create rest of ctxt 0 ctxtdata as well. | |
537 | */ | |
538 | rcd = dd->rcd[i]; | |
539 | if (!rcd) | |
540 | continue; | |
541 | ||
542 | lastfail = qib_create_rcvhdrq(dd, rcd); | |
543 | if (!lastfail) | |
544 | lastfail = qib_setup_eagerbufs(rcd); | |
545 | if (lastfail) { | |
546 | qib_dev_err(dd, "failed to allocate kernel ctxt's " | |
547 | "rcvhdrq and/or egr bufs\n"); | |
548 | continue; | |
549 | } | |
550 | } | |
551 | ||
552 | for (pidx = 0; pidx < dd->num_pports; ++pidx) { | |
553 | int mtu; | |
554 | if (lastfail) | |
555 | ret = lastfail; | |
556 | ppd = dd->pport + pidx; | |
557 | mtu = ib_mtu_enum_to_int(qib_ibmtu); | |
558 | if (mtu == -1) { | |
559 | mtu = QIB_DEFAULT_MTU; | |
560 | qib_ibmtu = 0; /* don't leave invalid value */ | |
561 | } | |
562 | /* set max we can ever have for this driver load */ | |
563 | ppd->init_ibmaxlen = min(mtu > 2048 ? | |
564 | dd->piosize4k : dd->piosize2k, | |
565 | dd->rcvegrbufsize + | |
566 | (dd->rcvhdrentsize << 2)); | |
567 | /* | |
568 | * Have to initialize ibmaxlen, but this will normally | |
569 | * change immediately in qib_set_mtu(). | |
570 | */ | |
571 | ppd->ibmaxlen = ppd->init_ibmaxlen; | |
572 | qib_set_mtu(ppd, mtu); | |
573 | ||
574 | spin_lock_irqsave(&ppd->lflags_lock, flags); | |
575 | ppd->lflags |= QIBL_IB_LINK_DISABLED; | |
576 | spin_unlock_irqrestore(&ppd->lflags_lock, flags); | |
577 | ||
578 | lastfail = dd->f_bringup_serdes(ppd); | |
579 | if (lastfail) { | |
580 | qib_devinfo(dd->pcidev, | |
581 | "Failed to bringup IB port %u\n", ppd->port); | |
582 | lastfail = -ENETDOWN; | |
583 | continue; | |
584 | } | |
585 | ||
586 | /* let link come up, and enable IBC */ | |
587 | spin_lock_irqsave(&ppd->lflags_lock, flags); | |
588 | ppd->lflags &= ~QIBL_IB_LINK_DISABLED; | |
589 | spin_unlock_irqrestore(&ppd->lflags_lock, flags); | |
590 | portok++; | |
591 | } | |
592 | ||
593 | if (!portok) { | |
594 | /* none of the ports initialized */ | |
595 | if (!ret && lastfail) | |
596 | ret = lastfail; | |
597 | else if (!ret) | |
598 | ret = -ENETDOWN; | |
599 | /* but continue on, so we can debug cause */ | |
600 | } | |
601 | ||
602 | enable_chip(dd); | |
603 | ||
604 | init_piobuf_state(dd); | |
605 | ||
606 | done: | |
607 | if (!ret) { | |
608 | /* chip is OK for user apps; mark it as initialized */ | |
609 | for (pidx = 0; pidx < dd->num_pports; ++pidx) { | |
610 | ppd = dd->pport + pidx; | |
611 | /* | |
612 | * Set status even if port serdes is not initialized | |
613 | * so that diags will work. | |
614 | */ | |
615 | *ppd->statusp |= QIB_STATUS_CHIP_PRESENT | | |
616 | QIB_STATUS_INITTED; | |
617 | if (!ppd->link_speed_enabled) | |
618 | continue; | |
619 | if (dd->flags & QIB_HAS_SEND_DMA) | |
620 | ret = qib_setup_sdma(ppd); | |
621 | init_timer(&ppd->hol_timer); | |
622 | ppd->hol_timer.function = qib_hol_event; | |
623 | ppd->hol_timer.data = (unsigned long)ppd; | |
624 | ppd->hol_state = QIB_HOL_UP; | |
625 | } | |
626 | ||
627 | /* now we can enable all interrupts from the chip */ | |
628 | dd->f_set_intr_state(dd, 1); | |
629 | ||
630 | /* | |
631 | * Setup to verify we get an interrupt, and fallback | |
632 | * to an alternate if necessary and possible. | |
633 | */ | |
634 | mod_timer(&dd->intrchk_timer, jiffies + HZ/2); | |
635 | /* start stats retrieval timer */ | |
636 | mod_timer(&dd->stats_timer, jiffies + HZ * ACTIVITY_TIMER); | |
637 | } | |
638 | ||
639 | /* if ret is non-zero, we probably should do some cleanup here... */ | |
640 | return ret; | |
641 | } | |
642 | ||
643 | /* | |
644 | * These next two routines are placeholders in case we don't have per-arch | |
645 | * code for controlling write combining. If explicit control of write | |
646 | * combining is not available, performance will probably be awful. | |
647 | */ | |
648 | ||
649 | int __attribute__((weak)) qib_enable_wc(struct qib_devdata *dd) | |
650 | { | |
651 | return -EOPNOTSUPP; | |
652 | } | |
653 | ||
654 | void __attribute__((weak)) qib_disable_wc(struct qib_devdata *dd) | |
655 | { | |
656 | } | |
657 | ||
658 | static inline struct qib_devdata *__qib_lookup(int unit) | |
659 | { | |
660 | return idr_find(&qib_unit_table, unit); | |
661 | } | |
662 | ||
663 | struct qib_devdata *qib_lookup(int unit) | |
664 | { | |
665 | struct qib_devdata *dd; | |
666 | unsigned long flags; | |
667 | ||
668 | spin_lock_irqsave(&qib_devs_lock, flags); | |
669 | dd = __qib_lookup(unit); | |
670 | spin_unlock_irqrestore(&qib_devs_lock, flags); | |
671 | ||
672 | return dd; | |
673 | } | |
674 | ||
675 | /* | |
676 | * Stop the timers during unit shutdown, or after an error late | |
677 | * in initialization. | |
678 | */ | |
679 | static void qib_stop_timers(struct qib_devdata *dd) | |
680 | { | |
681 | struct qib_pportdata *ppd; | |
682 | int pidx; | |
683 | ||
684 | if (dd->stats_timer.data) { | |
685 | del_timer_sync(&dd->stats_timer); | |
686 | dd->stats_timer.data = 0; | |
687 | } | |
688 | if (dd->intrchk_timer.data) { | |
689 | del_timer_sync(&dd->intrchk_timer); | |
690 | dd->intrchk_timer.data = 0; | |
691 | } | |
692 | for (pidx = 0; pidx < dd->num_pports; ++pidx) { | |
693 | ppd = dd->pport + pidx; | |
694 | if (ppd->hol_timer.data) | |
695 | del_timer_sync(&ppd->hol_timer); | |
696 | if (ppd->led_override_timer.data) { | |
697 | del_timer_sync(&ppd->led_override_timer); | |
698 | atomic_set(&ppd->led_override_timer_active, 0); | |
699 | } | |
700 | if (ppd->symerr_clear_timer.data) | |
701 | del_timer_sync(&ppd->symerr_clear_timer); | |
702 | } | |
703 | } | |
704 | ||
705 | /** | |
706 | * qib_shutdown_device - shut down a device | |
707 | * @dd: the qlogic_ib device | |
708 | * | |
709 | * This is called to make the device quiet when we are about to | |
710 | * unload the driver, and also when the device is administratively | |
711 | * disabled. It does not free any data structures. | |
712 | * Everything it does has to be setup again by qib_init(dd, 1) | |
713 | */ | |
714 | static void qib_shutdown_device(struct qib_devdata *dd) | |
715 | { | |
716 | struct qib_pportdata *ppd; | |
717 | unsigned pidx; | |
718 | ||
719 | for (pidx = 0; pidx < dd->num_pports; ++pidx) { | |
720 | ppd = dd->pport + pidx; | |
721 | ||
722 | spin_lock_irq(&ppd->lflags_lock); | |
723 | ppd->lflags &= ~(QIBL_LINKDOWN | QIBL_LINKINIT | | |
724 | QIBL_LINKARMED | QIBL_LINKACTIVE | | |
725 | QIBL_LINKV); | |
726 | spin_unlock_irq(&ppd->lflags_lock); | |
727 | *ppd->statusp &= ~(QIB_STATUS_IB_CONF | QIB_STATUS_IB_READY); | |
728 | } | |
729 | dd->flags &= ~QIB_INITTED; | |
730 | ||
731 | /* mask interrupts, but not errors */ | |
732 | dd->f_set_intr_state(dd, 0); | |
733 | ||
734 | for (pidx = 0; pidx < dd->num_pports; ++pidx) { | |
735 | ppd = dd->pport + pidx; | |
736 | dd->f_rcvctrl(ppd, QIB_RCVCTRL_TAILUPD_DIS | | |
737 | QIB_RCVCTRL_CTXT_DIS | | |
738 | QIB_RCVCTRL_INTRAVAIL_DIS | | |
739 | QIB_RCVCTRL_PKEY_ENB, -1); | |
740 | /* | |
741 | * Gracefully stop all sends allowing any in progress to | |
742 | * trickle out first. | |
743 | */ | |
744 | dd->f_sendctrl(ppd, QIB_SENDCTRL_CLEAR); | |
745 | } | |
746 | ||
747 | /* | |
748 | * Enough for anything that's going to trickle out to have actually | |
749 | * done so. | |
750 | */ | |
751 | udelay(20); | |
752 | ||
753 | for (pidx = 0; pidx < dd->num_pports; ++pidx) { | |
754 | ppd = dd->pport + pidx; | |
755 | dd->f_setextled(ppd, 0); /* make sure LEDs are off */ | |
756 | ||
757 | if (dd->flags & QIB_HAS_SEND_DMA) | |
758 | qib_teardown_sdma(ppd); | |
759 | ||
760 | dd->f_sendctrl(ppd, QIB_SENDCTRL_AVAIL_DIS | | |
761 | QIB_SENDCTRL_SEND_DIS); | |
762 | /* | |
763 | * Clear SerdesEnable. | |
764 | * We can't count on interrupts since we are stopping. | |
765 | */ | |
766 | dd->f_quiet_serdes(ppd); | |
767 | } | |
768 | ||
769 | qib_update_eeprom_log(dd); | |
770 | } | |
771 | ||
772 | /** | |
773 | * qib_free_ctxtdata - free a context's allocated data | |
774 | * @dd: the qlogic_ib device | |
775 | * @rcd: the ctxtdata structure | |
776 | * | |
777 | * free up any allocated data for a context | |
778 | * This should not touch anything that would affect a simultaneous | |
779 | * re-allocation of context data, because it is called after qib_mutex | |
780 | * is released (and can be called from reinit as well). | |
781 | * It should never change any chip state, or global driver state. | |
782 | */ | |
783 | void qib_free_ctxtdata(struct qib_devdata *dd, struct qib_ctxtdata *rcd) | |
784 | { | |
785 | if (!rcd) | |
786 | return; | |
787 | ||
788 | if (rcd->rcvhdrq) { | |
789 | dma_free_coherent(&dd->pcidev->dev, rcd->rcvhdrq_size, | |
790 | rcd->rcvhdrq, rcd->rcvhdrq_phys); | |
791 | rcd->rcvhdrq = NULL; | |
792 | if (rcd->rcvhdrtail_kvaddr) { | |
793 | dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE, | |
794 | rcd->rcvhdrtail_kvaddr, | |
795 | rcd->rcvhdrqtailaddr_phys); | |
796 | rcd->rcvhdrtail_kvaddr = NULL; | |
797 | } | |
798 | } | |
799 | if (rcd->rcvegrbuf) { | |
800 | unsigned e; | |
801 | ||
802 | for (e = 0; e < rcd->rcvegrbuf_chunks; e++) { | |
803 | void *base = rcd->rcvegrbuf[e]; | |
804 | size_t size = rcd->rcvegrbuf_size; | |
805 | ||
806 | dma_free_coherent(&dd->pcidev->dev, size, | |
807 | base, rcd->rcvegrbuf_phys[e]); | |
808 | } | |
809 | kfree(rcd->rcvegrbuf); | |
810 | rcd->rcvegrbuf = NULL; | |
811 | kfree(rcd->rcvegrbuf_phys); | |
812 | rcd->rcvegrbuf_phys = NULL; | |
813 | rcd->rcvegrbuf_chunks = 0; | |
814 | } | |
815 | ||
816 | kfree(rcd->tid_pg_list); | |
817 | vfree(rcd->user_event_mask); | |
818 | vfree(rcd->subctxt_uregbase); | |
819 | vfree(rcd->subctxt_rcvegrbuf); | |
820 | vfree(rcd->subctxt_rcvhdr_base); | |
821 | kfree(rcd); | |
822 | } | |
823 | ||
824 | /* | |
825 | * Perform a PIO buffer bandwidth write test, to verify proper system | |
826 | * configuration. Even when all the setup calls work, occasionally | |
827 | * BIOS or other issues can prevent write combining from working, or | |
828 | * can cause other bandwidth problems to the chip. | |
829 | * | |
830 | * This test simply writes the same buffer over and over again, and | |
831 | * measures close to the peak bandwidth to the chip (not testing | |
832 | * data bandwidth to the wire). On chips that use an address-based | |
833 | * trigger to send packets to the wire, this is easy. On chips that | |
834 | * use a count to trigger, we want to make sure that the packet doesn't | |
835 | * go out on the wire, or trigger flow control checks. | |
836 | */ | |
837 | static void qib_verify_pioperf(struct qib_devdata *dd) | |
838 | { | |
839 | u32 pbnum, cnt, lcnt; | |
840 | u32 __iomem *piobuf; | |
841 | u32 *addr; | |
842 | u64 msecs, emsecs; | |
843 | ||
844 | piobuf = dd->f_getsendbuf(dd->pport, 0ULL, &pbnum); | |
845 | if (!piobuf) { | |
846 | qib_devinfo(dd->pcidev, | |
847 | "No PIObufs for checking perf, skipping\n"); | |
848 | return; | |
849 | } | |
850 | ||
851 | /* | |
852 | * Enough to give us a reasonable test, less than piobuf size, and | |
853 | * likely multiple of store buffer length. | |
854 | */ | |
855 | cnt = 1024; | |
856 | ||
857 | addr = vmalloc(cnt); | |
858 | if (!addr) { | |
859 | qib_devinfo(dd->pcidev, | |
860 | "Couldn't get memory for checking PIO perf," | |
861 | " skipping\n"); | |
862 | goto done; | |
863 | } | |
864 | ||
865 | preempt_disable(); /* we want reasonably accurate elapsed time */ | |
866 | msecs = 1 + jiffies_to_msecs(jiffies); | |
867 | for (lcnt = 0; lcnt < 10000U; lcnt++) { | |
868 | /* wait until we cross msec boundary */ | |
869 | if (jiffies_to_msecs(jiffies) >= msecs) | |
870 | break; | |
871 | udelay(1); | |
872 | } | |
873 | ||
874 | dd->f_set_armlaunch(dd, 0); | |
875 | ||
876 | /* | |
877 | * length 0, no dwords actually sent | |
878 | */ | |
879 | writeq(0, piobuf); | |
880 | qib_flush_wc(); | |
881 | ||
882 | /* | |
883 | * This is only roughly accurate, since even with preempt we | |
884 | * still take interrupts that could take a while. Running for | |
885 | * >= 5 msec seems to get us "close enough" to accurate values. | |
886 | */ | |
887 | msecs = jiffies_to_msecs(jiffies); | |
888 | for (emsecs = lcnt = 0; emsecs <= 5UL; lcnt++) { | |
889 | qib_pio_copy(piobuf + 64, addr, cnt >> 2); | |
890 | emsecs = jiffies_to_msecs(jiffies) - msecs; | |
891 | } | |
892 | ||
893 | /* 1 GiB/sec, slightly over IB SDR line rate */ | |
894 | if (lcnt < (emsecs * 1024U)) | |
895 | qib_dev_err(dd, | |
896 | "Performance problem: bandwidth to PIO buffers is " | |
897 | "only %u MiB/sec\n", | |
898 | lcnt / (u32) emsecs); | |
899 | ||
900 | preempt_enable(); | |
901 | ||
902 | vfree(addr); | |
903 | ||
904 | done: | |
905 | /* disarm piobuf, so it's available again */ | |
906 | dd->f_sendctrl(dd->pport, QIB_SENDCTRL_DISARM_BUF(pbnum)); | |
907 | qib_sendbuf_done(dd, pbnum); | |
908 | dd->f_set_armlaunch(dd, 1); | |
909 | } | |
910 | ||
911 | ||
912 | void qib_free_devdata(struct qib_devdata *dd) | |
913 | { | |
914 | unsigned long flags; | |
915 | ||
916 | spin_lock_irqsave(&qib_devs_lock, flags); | |
917 | idr_remove(&qib_unit_table, dd->unit); | |
918 | list_del(&dd->list); | |
919 | spin_unlock_irqrestore(&qib_devs_lock, flags); | |
920 | ||
921 | ib_dealloc_device(&dd->verbs_dev.ibdev); | |
922 | } | |
923 | ||
924 | /* | |
925 | * Allocate our primary per-unit data structure. Must be done via verbs | |
926 | * allocator, because the verbs cleanup process both does cleanup and | |
927 | * free of the data structure. | |
928 | * "extra" is for chip-specific data. | |
929 | * | |
930 | * Use the idr mechanism to get a unit number for this unit. | |
931 | */ | |
932 | struct qib_devdata *qib_alloc_devdata(struct pci_dev *pdev, size_t extra) | |
933 | { | |
934 | unsigned long flags; | |
935 | struct qib_devdata *dd; | |
936 | int ret; | |
937 | ||
938 | if (!idr_pre_get(&qib_unit_table, GFP_KERNEL)) { | |
939 | dd = ERR_PTR(-ENOMEM); | |
940 | goto bail; | |
941 | } | |
942 | ||
943 | dd = (struct qib_devdata *) ib_alloc_device(sizeof(*dd) + extra); | |
944 | if (!dd) { | |
945 | dd = ERR_PTR(-ENOMEM); | |
946 | goto bail; | |
947 | } | |
948 | ||
949 | spin_lock_irqsave(&qib_devs_lock, flags); | |
950 | ret = idr_get_new(&qib_unit_table, dd, &dd->unit); | |
951 | if (ret >= 0) | |
952 | list_add(&dd->list, &qib_dev_list); | |
953 | spin_unlock_irqrestore(&qib_devs_lock, flags); | |
954 | ||
955 | if (ret < 0) { | |
956 | qib_early_err(&pdev->dev, | |
957 | "Could not allocate unit ID: error %d\n", -ret); | |
958 | ib_dealloc_device(&dd->verbs_dev.ibdev); | |
959 | dd = ERR_PTR(ret); | |
960 | goto bail; | |
961 | } | |
962 | ||
963 | if (!qib_cpulist_count) { | |
964 | u32 count = num_online_cpus(); | |
965 | qib_cpulist = kzalloc(BITS_TO_LONGS(count) * | |
966 | sizeof(long), GFP_KERNEL); | |
967 | if (qib_cpulist) | |
968 | qib_cpulist_count = count; | |
969 | else | |
970 | qib_early_err(&pdev->dev, "Could not alloc cpulist " | |
971 | "info, cpu affinity might be wrong\n"); | |
972 | } | |
973 | ||
974 | bail: | |
975 | return dd; | |
976 | } | |
977 | ||
978 | /* | |
979 | * Called from freeze mode handlers, and from PCI error | |
980 | * reporting code. Should be paranoid about state of | |
981 | * system and data structures. | |
982 | */ | |
983 | void qib_disable_after_error(struct qib_devdata *dd) | |
984 | { | |
985 | if (dd->flags & QIB_INITTED) { | |
986 | u32 pidx; | |
987 | ||
988 | dd->flags &= ~QIB_INITTED; | |
989 | if (dd->pport) | |
990 | for (pidx = 0; pidx < dd->num_pports; ++pidx) { | |
991 | struct qib_pportdata *ppd; | |
992 | ||
993 | ppd = dd->pport + pidx; | |
994 | if (dd->flags & QIB_PRESENT) { | |
995 | qib_set_linkstate(ppd, | |
996 | QIB_IB_LINKDOWN_DISABLE); | |
997 | dd->f_setextled(ppd, 0); | |
998 | } | |
999 | *ppd->statusp &= ~QIB_STATUS_IB_READY; | |
1000 | } | |
1001 | } | |
1002 | ||
1003 | /* | |
1004 | * Mark as having had an error for driver, and also | |
1005 | * for /sys and status word mapped to user programs. | |
1006 | * This marks unit as not usable, until reset. | |
1007 | */ | |
1008 | if (dd->devstatusp) | |
1009 | *dd->devstatusp |= QIB_STATUS_HWERROR; | |
1010 | } | |
1011 | ||
1012 | static void __devexit qib_remove_one(struct pci_dev *); | |
1013 | static int __devinit qib_init_one(struct pci_dev *, | |
1014 | const struct pci_device_id *); | |
1015 | ||
1016 | #define DRIVER_LOAD_MSG "QLogic " QIB_DRV_NAME " loaded: " | |
1017 | #define PFX QIB_DRV_NAME ": " | |
1018 | ||
1019 | static const struct pci_device_id qib_pci_tbl[] = { | |
1020 | { PCI_DEVICE(PCI_VENDOR_ID_PATHSCALE, PCI_DEVICE_ID_QLOGIC_IB_6120) }, | |
1021 | { PCI_DEVICE(PCI_VENDOR_ID_QLOGIC, PCI_DEVICE_ID_QLOGIC_IB_7220) }, | |
1022 | { PCI_DEVICE(PCI_VENDOR_ID_QLOGIC, PCI_DEVICE_ID_QLOGIC_IB_7322) }, | |
1023 | { 0, } | |
1024 | }; | |
1025 | ||
1026 | MODULE_DEVICE_TABLE(pci, qib_pci_tbl); | |
1027 | ||
1028 | struct pci_driver qib_driver = { | |
1029 | .name = QIB_DRV_NAME, | |
1030 | .probe = qib_init_one, | |
1031 | .remove = __devexit_p(qib_remove_one), | |
1032 | .id_table = qib_pci_tbl, | |
1033 | .err_handler = &qib_pci_err_handler, | |
1034 | }; | |
1035 | ||
1036 | /* | |
1037 | * Do all the generic driver unit- and chip-independent memory | |
1038 | * allocation and initialization. | |
1039 | */ | |
1040 | static int __init qlogic_ib_init(void) | |
1041 | { | |
1042 | int ret; | |
1043 | ||
1044 | ret = qib_dev_init(); | |
1045 | if (ret) | |
1046 | goto bail; | |
1047 | ||
1048 | /* | |
1049 | * We create our own workqueue mainly because we want to be | |
1050 | * able to flush it when devices are being removed. We can't | |
1051 | * use schedule_work()/flush_scheduled_work() because both | |
1052 | * unregister_netdev() and linkwatch_event take the rtnl lock, | |
1053 | * so flush_scheduled_work() can deadlock during device | |
1054 | * removal. | |
1055 | */ | |
1056 | qib_wq = create_workqueue("qib"); | |
1057 | if (!qib_wq) { | |
1058 | ret = -ENOMEM; | |
1059 | goto bail_dev; | |
1060 | } | |
1061 | ||
950aff53 | 1062 | qib_cq_wq = create_singlethread_workqueue("qib_cq"); |
f931551b RC |
1063 | if (!qib_cq_wq) { |
1064 | ret = -ENOMEM; | |
1065 | goto bail_wq; | |
1066 | } | |
1067 | ||
1068 | /* | |
1069 | * These must be called before the driver is registered with | |
1070 | * the PCI subsystem. | |
1071 | */ | |
1072 | idr_init(&qib_unit_table); | |
1073 | if (!idr_pre_get(&qib_unit_table, GFP_KERNEL)) { | |
1074 | printk(KERN_ERR QIB_DRV_NAME ": idr_pre_get() failed\n"); | |
1075 | ret = -ENOMEM; | |
1076 | goto bail_cq_wq; | |
1077 | } | |
1078 | ||
1079 | ret = pci_register_driver(&qib_driver); | |
1080 | if (ret < 0) { | |
1081 | printk(KERN_ERR QIB_DRV_NAME | |
1082 | ": Unable to register driver: error %d\n", -ret); | |
1083 | goto bail_unit; | |
1084 | } | |
1085 | ||
1086 | /* not fatal if it doesn't work */ | |
1087 | if (qib_init_qibfs()) | |
1088 | printk(KERN_ERR QIB_DRV_NAME ": Unable to register ipathfs\n"); | |
1089 | goto bail; /* all OK */ | |
1090 | ||
1091 | bail_unit: | |
1092 | idr_destroy(&qib_unit_table); | |
1093 | bail_cq_wq: | |
1094 | destroy_workqueue(qib_cq_wq); | |
1095 | bail_wq: | |
1096 | destroy_workqueue(qib_wq); | |
1097 | bail_dev: | |
1098 | qib_dev_cleanup(); | |
1099 | bail: | |
1100 | return ret; | |
1101 | } | |
1102 | ||
1103 | module_init(qlogic_ib_init); | |
1104 | ||
1105 | /* | |
1106 | * Do the non-unit driver cleanup, memory free, etc. at unload. | |
1107 | */ | |
1108 | static void __exit qlogic_ib_cleanup(void) | |
1109 | { | |
1110 | int ret; | |
1111 | ||
1112 | ret = qib_exit_qibfs(); | |
1113 | if (ret) | |
1114 | printk(KERN_ERR QIB_DRV_NAME ": " | |
1115 | "Unable to cleanup counter filesystem: " | |
1116 | "error %d\n", -ret); | |
1117 | ||
1118 | pci_unregister_driver(&qib_driver); | |
1119 | ||
1120 | destroy_workqueue(qib_wq); | |
1121 | destroy_workqueue(qib_cq_wq); | |
1122 | ||
1123 | qib_cpulist_count = 0; | |
1124 | kfree(qib_cpulist); | |
1125 | ||
1126 | idr_destroy(&qib_unit_table); | |
1127 | qib_dev_cleanup(); | |
1128 | } | |
1129 | ||
1130 | module_exit(qlogic_ib_cleanup); | |
1131 | ||
1132 | /* this can only be called after a successful initialization */ | |
1133 | static void cleanup_device_data(struct qib_devdata *dd) | |
1134 | { | |
1135 | int ctxt; | |
1136 | int pidx; | |
1137 | struct qib_ctxtdata **tmp; | |
1138 | unsigned long flags; | |
1139 | ||
1140 | /* users can't do anything more with chip */ | |
1141 | for (pidx = 0; pidx < dd->num_pports; ++pidx) | |
1142 | if (dd->pport[pidx].statusp) | |
1143 | *dd->pport[pidx].statusp &= ~QIB_STATUS_CHIP_PRESENT; | |
1144 | ||
1145 | if (!qib_wc_pat) | |
1146 | qib_disable_wc(dd); | |
1147 | ||
1148 | if (dd->pioavailregs_dma) { | |
1149 | dma_free_coherent(&dd->pcidev->dev, PAGE_SIZE, | |
1150 | (void *) dd->pioavailregs_dma, | |
1151 | dd->pioavailregs_phys); | |
1152 | dd->pioavailregs_dma = NULL; | |
1153 | } | |
1154 | ||
1155 | if (dd->pageshadow) { | |
1156 | struct page **tmpp = dd->pageshadow; | |
1157 | dma_addr_t *tmpd = dd->physshadow; | |
1158 | int i, cnt = 0; | |
1159 | ||
1160 | for (ctxt = 0; ctxt < dd->cfgctxts; ctxt++) { | |
1161 | int ctxt_tidbase = ctxt * dd->rcvtidcnt; | |
1162 | int maxtid = ctxt_tidbase + dd->rcvtidcnt; | |
1163 | ||
1164 | for (i = ctxt_tidbase; i < maxtid; i++) { | |
1165 | if (!tmpp[i]) | |
1166 | continue; | |
1167 | pci_unmap_page(dd->pcidev, tmpd[i], | |
1168 | PAGE_SIZE, PCI_DMA_FROMDEVICE); | |
1169 | qib_release_user_pages(&tmpp[i], 1); | |
1170 | tmpp[i] = NULL; | |
1171 | cnt++; | |
1172 | } | |
1173 | } | |
1174 | ||
1175 | tmpp = dd->pageshadow; | |
1176 | dd->pageshadow = NULL; | |
1177 | vfree(tmpp); | |
1178 | } | |
1179 | ||
1180 | /* | |
1181 | * Free any resources still in use (usually just kernel contexts) | |
1182 | * at unload; we do for ctxtcnt, because that's what we allocate. | |
1183 | * We acquire lock to be really paranoid that rcd isn't being | |
1184 | * accessed from some interrupt-related code (that should not happen, | |
1185 | * but best to be sure). | |
1186 | */ | |
1187 | spin_lock_irqsave(&dd->uctxt_lock, flags); | |
1188 | tmp = dd->rcd; | |
1189 | dd->rcd = NULL; | |
1190 | spin_unlock_irqrestore(&dd->uctxt_lock, flags); | |
1191 | for (ctxt = 0; tmp && ctxt < dd->ctxtcnt; ctxt++) { | |
1192 | struct qib_ctxtdata *rcd = tmp[ctxt]; | |
1193 | ||
1194 | tmp[ctxt] = NULL; /* debugging paranoia */ | |
1195 | qib_free_ctxtdata(dd, rcd); | |
1196 | } | |
1197 | kfree(tmp); | |
1198 | kfree(dd->boardname); | |
1199 | } | |
1200 | ||
1201 | /* | |
1202 | * Clean up on unit shutdown, or error during unit load after | |
1203 | * successful initialization. | |
1204 | */ | |
1205 | static void qib_postinit_cleanup(struct qib_devdata *dd) | |
1206 | { | |
1207 | /* | |
1208 | * Clean up chip-specific stuff. | |
1209 | * We check for NULL here, because it's outside | |
1210 | * the kregbase check, and we need to call it | |
1211 | * after the free_irq. Thus it's possible that | |
1212 | * the function pointers were never initialized. | |
1213 | */ | |
1214 | if (dd->f_cleanup) | |
1215 | dd->f_cleanup(dd); | |
1216 | ||
1217 | qib_pcie_ddcleanup(dd); | |
1218 | ||
1219 | cleanup_device_data(dd); | |
1220 | ||
1221 | qib_free_devdata(dd); | |
1222 | } | |
1223 | ||
1224 | static int __devinit qib_init_one(struct pci_dev *pdev, | |
1225 | const struct pci_device_id *ent) | |
1226 | { | |
1227 | int ret, j, pidx, initfail; | |
1228 | struct qib_devdata *dd = NULL; | |
1229 | ||
1230 | ret = qib_pcie_init(pdev, ent); | |
1231 | if (ret) | |
1232 | goto bail; | |
1233 | ||
1234 | /* | |
1235 | * Do device-specific initialiation, function table setup, dd | |
1236 | * allocation, etc. | |
1237 | */ | |
1238 | switch (ent->device) { | |
1239 | case PCI_DEVICE_ID_QLOGIC_IB_6120: | |
7e3a1f4a | 1240 | #ifdef CONFIG_PCI_MSI |
f931551b | 1241 | dd = qib_init_iba6120_funcs(pdev, ent); |
7e3a1f4a RC |
1242 | #else |
1243 | qib_early_err(&pdev->dev, "QLogic PCIE device 0x%x cannot " | |
1244 | "work if CONFIG_PCI_MSI is not enabled\n", | |
1245 | ent->device); | |
1246 | #endif | |
f931551b RC |
1247 | break; |
1248 | ||
1249 | case PCI_DEVICE_ID_QLOGIC_IB_7220: | |
1250 | dd = qib_init_iba7220_funcs(pdev, ent); | |
1251 | break; | |
1252 | ||
1253 | case PCI_DEVICE_ID_QLOGIC_IB_7322: | |
1254 | dd = qib_init_iba7322_funcs(pdev, ent); | |
1255 | break; | |
1256 | ||
1257 | default: | |
1258 | qib_early_err(&pdev->dev, "Failing on unknown QLogic " | |
1259 | "deviceid 0x%x\n", ent->device); | |
1260 | ret = -ENODEV; | |
1261 | } | |
1262 | ||
1263 | if (IS_ERR(dd)) | |
1264 | ret = PTR_ERR(dd); | |
1265 | if (ret) | |
1266 | goto bail; /* error already printed */ | |
1267 | ||
1268 | /* do the generic initialization */ | |
1269 | initfail = qib_init(dd, 0); | |
1270 | ||
1271 | ret = qib_register_ib_device(dd); | |
1272 | ||
1273 | /* | |
1274 | * Now ready for use. this should be cleared whenever we | |
1275 | * detect a reset, or initiate one. If earlier failure, | |
1276 | * we still create devices, so diags, etc. can be used | |
1277 | * to determine cause of problem. | |
1278 | */ | |
1279 | if (!qib_mini_init && !initfail && !ret) | |
1280 | dd->flags |= QIB_INITTED; | |
1281 | ||
1282 | j = qib_device_create(dd); | |
1283 | if (j) | |
1284 | qib_dev_err(dd, "Failed to create /dev devices: %d\n", -j); | |
1285 | j = qibfs_add(dd); | |
1286 | if (j) | |
1287 | qib_dev_err(dd, "Failed filesystem setup for counters: %d\n", | |
1288 | -j); | |
1289 | ||
1290 | if (qib_mini_init || initfail || ret) { | |
1291 | qib_stop_timers(dd); | |
756a33b8 | 1292 | flush_scheduled_work(); |
f931551b RC |
1293 | for (pidx = 0; pidx < dd->num_pports; ++pidx) |
1294 | dd->f_quiet_serdes(dd->pport + pidx); | |
756a33b8 RC |
1295 | if (qib_mini_init) |
1296 | goto bail; | |
1297 | if (!j) { | |
1298 | (void) qibfs_remove(dd); | |
1299 | qib_device_remove(dd); | |
1300 | } | |
1301 | if (!ret) | |
1302 | qib_unregister_ib_device(dd); | |
1303 | qib_postinit_cleanup(dd); | |
f931551b RC |
1304 | if (initfail) |
1305 | ret = initfail; | |
1306 | goto bail; | |
1307 | } | |
1308 | ||
1309 | if (!qib_wc_pat) { | |
1310 | ret = qib_enable_wc(dd); | |
1311 | if (ret) { | |
1312 | qib_dev_err(dd, "Write combining not enabled " | |
1313 | "(err %d): performance may be poor\n", | |
1314 | -ret); | |
1315 | ret = 0; | |
1316 | } | |
1317 | } | |
1318 | ||
1319 | qib_verify_pioperf(dd); | |
1320 | bail: | |
1321 | return ret; | |
1322 | } | |
1323 | ||
1324 | static void __devexit qib_remove_one(struct pci_dev *pdev) | |
1325 | { | |
1326 | struct qib_devdata *dd = pci_get_drvdata(pdev); | |
1327 | int ret; | |
1328 | ||
1329 | /* unregister from IB core */ | |
1330 | qib_unregister_ib_device(dd); | |
1331 | ||
1332 | /* | |
1333 | * Disable the IB link, disable interrupts on the device, | |
1334 | * clear dma engines, etc. | |
1335 | */ | |
1336 | if (!qib_mini_init) | |
1337 | qib_shutdown_device(dd); | |
1338 | ||
1339 | qib_stop_timers(dd); | |
1340 | ||
1341 | /* wait until all of our (qsfp) schedule_work() calls complete */ | |
1342 | flush_scheduled_work(); | |
1343 | ||
1344 | ret = qibfs_remove(dd); | |
1345 | if (ret) | |
1346 | qib_dev_err(dd, "Failed counters filesystem cleanup: %d\n", | |
1347 | -ret); | |
1348 | ||
1349 | qib_device_remove(dd); | |
1350 | ||
1351 | qib_postinit_cleanup(dd); | |
1352 | } | |
1353 | ||
1354 | /** | |
1355 | * qib_create_rcvhdrq - create a receive header queue | |
1356 | * @dd: the qlogic_ib device | |
1357 | * @rcd: the context data | |
1358 | * | |
1359 | * This must be contiguous memory (from an i/o perspective), and must be | |
1360 | * DMA'able (which means for some systems, it will go through an IOMMU, | |
1361 | * or be forced into a low address range). | |
1362 | */ | |
1363 | int qib_create_rcvhdrq(struct qib_devdata *dd, struct qib_ctxtdata *rcd) | |
1364 | { | |
1365 | unsigned amt; | |
1366 | ||
1367 | if (!rcd->rcvhdrq) { | |
1368 | dma_addr_t phys_hdrqtail; | |
1369 | gfp_t gfp_flags; | |
1370 | ||
1371 | amt = ALIGN(dd->rcvhdrcnt * dd->rcvhdrentsize * | |
1372 | sizeof(u32), PAGE_SIZE); | |
1373 | gfp_flags = (rcd->ctxt >= dd->first_user_ctxt) ? | |
1374 | GFP_USER : GFP_KERNEL; | |
1375 | rcd->rcvhdrq = dma_alloc_coherent( | |
1376 | &dd->pcidev->dev, amt, &rcd->rcvhdrq_phys, | |
1377 | gfp_flags | __GFP_COMP); | |
1378 | ||
1379 | if (!rcd->rcvhdrq) { | |
1380 | qib_dev_err(dd, "attempt to allocate %d bytes " | |
1381 | "for ctxt %u rcvhdrq failed\n", | |
1382 | amt, rcd->ctxt); | |
1383 | goto bail; | |
1384 | } | |
1385 | ||
1386 | if (rcd->ctxt >= dd->first_user_ctxt) { | |
1387 | rcd->user_event_mask = vmalloc_user(PAGE_SIZE); | |
1388 | if (!rcd->user_event_mask) | |
1389 | goto bail_free_hdrq; | |
1390 | } | |
1391 | ||
1392 | if (!(dd->flags & QIB_NODMA_RTAIL)) { | |
1393 | rcd->rcvhdrtail_kvaddr = dma_alloc_coherent( | |
1394 | &dd->pcidev->dev, PAGE_SIZE, &phys_hdrqtail, | |
1395 | gfp_flags); | |
1396 | if (!rcd->rcvhdrtail_kvaddr) | |
1397 | goto bail_free; | |
1398 | rcd->rcvhdrqtailaddr_phys = phys_hdrqtail; | |
1399 | } | |
1400 | ||
1401 | rcd->rcvhdrq_size = amt; | |
1402 | } | |
1403 | ||
1404 | /* clear for security and sanity on each use */ | |
1405 | memset(rcd->rcvhdrq, 0, rcd->rcvhdrq_size); | |
1406 | if (rcd->rcvhdrtail_kvaddr) | |
1407 | memset(rcd->rcvhdrtail_kvaddr, 0, PAGE_SIZE); | |
1408 | return 0; | |
1409 | ||
1410 | bail_free: | |
1411 | qib_dev_err(dd, "attempt to allocate 1 page for ctxt %u " | |
1412 | "rcvhdrqtailaddr failed\n", rcd->ctxt); | |
1413 | vfree(rcd->user_event_mask); | |
1414 | rcd->user_event_mask = NULL; | |
1415 | bail_free_hdrq: | |
1416 | dma_free_coherent(&dd->pcidev->dev, amt, rcd->rcvhdrq, | |
1417 | rcd->rcvhdrq_phys); | |
1418 | rcd->rcvhdrq = NULL; | |
1419 | bail: | |
1420 | return -ENOMEM; | |
1421 | } | |
1422 | ||
1423 | /** | |
1424 | * allocate eager buffers, both kernel and user contexts. | |
1425 | * @rcd: the context we are setting up. | |
1426 | * | |
1427 | * Allocate the eager TID buffers and program them into hip. | |
1428 | * They are no longer completely contiguous, we do multiple allocation | |
1429 | * calls. Otherwise we get the OOM code involved, by asking for too | |
1430 | * much per call, with disastrous results on some kernels. | |
1431 | */ | |
1432 | int qib_setup_eagerbufs(struct qib_ctxtdata *rcd) | |
1433 | { | |
1434 | struct qib_devdata *dd = rcd->dd; | |
1435 | unsigned e, egrcnt, egrperchunk, chunk, egrsize, egroff; | |
1436 | size_t size; | |
1437 | gfp_t gfp_flags; | |
1438 | ||
1439 | /* | |
1440 | * GFP_USER, but without GFP_FS, so buffer cache can be | |
1441 | * coalesced (we hope); otherwise, even at order 4, | |
1442 | * heavy filesystem activity makes these fail, and we can | |
1443 | * use compound pages. | |
1444 | */ | |
1445 | gfp_flags = __GFP_WAIT | __GFP_IO | __GFP_COMP; | |
1446 | ||
1447 | egrcnt = rcd->rcvegrcnt; | |
1448 | egroff = rcd->rcvegr_tid_base; | |
1449 | egrsize = dd->rcvegrbufsize; | |
1450 | ||
1451 | chunk = rcd->rcvegrbuf_chunks; | |
1452 | egrperchunk = rcd->rcvegrbufs_perchunk; | |
1453 | size = rcd->rcvegrbuf_size; | |
1454 | if (!rcd->rcvegrbuf) { | |
1455 | rcd->rcvegrbuf = | |
1456 | kzalloc(chunk * sizeof(rcd->rcvegrbuf[0]), | |
1457 | GFP_KERNEL); | |
1458 | if (!rcd->rcvegrbuf) | |
1459 | goto bail; | |
1460 | } | |
1461 | if (!rcd->rcvegrbuf_phys) { | |
1462 | rcd->rcvegrbuf_phys = | |
1463 | kmalloc(chunk * sizeof(rcd->rcvegrbuf_phys[0]), | |
1464 | GFP_KERNEL); | |
1465 | if (!rcd->rcvegrbuf_phys) | |
1466 | goto bail_rcvegrbuf; | |
1467 | } | |
1468 | for (e = 0; e < rcd->rcvegrbuf_chunks; e++) { | |
1469 | if (rcd->rcvegrbuf[e]) | |
1470 | continue; | |
1471 | rcd->rcvegrbuf[e] = | |
1472 | dma_alloc_coherent(&dd->pcidev->dev, size, | |
1473 | &rcd->rcvegrbuf_phys[e], | |
1474 | gfp_flags); | |
1475 | if (!rcd->rcvegrbuf[e]) | |
1476 | goto bail_rcvegrbuf_phys; | |
1477 | } | |
1478 | ||
1479 | rcd->rcvegr_phys = rcd->rcvegrbuf_phys[0]; | |
1480 | ||
1481 | for (e = chunk = 0; chunk < rcd->rcvegrbuf_chunks; chunk++) { | |
1482 | dma_addr_t pa = rcd->rcvegrbuf_phys[chunk]; | |
1483 | unsigned i; | |
1484 | ||
5df4223a RC |
1485 | /* clear for security and sanity on each use */ |
1486 | memset(rcd->rcvegrbuf[chunk], 0, size); | |
1487 | ||
f931551b RC |
1488 | for (i = 0; e < egrcnt && i < egrperchunk; e++, i++) { |
1489 | dd->f_put_tid(dd, e + egroff + | |
1490 | (u64 __iomem *) | |
1491 | ((char __iomem *) | |
1492 | dd->kregbase + | |
1493 | dd->rcvegrbase), | |
1494 | RCVHQ_RCV_TYPE_EAGER, pa); | |
1495 | pa += egrsize; | |
1496 | } | |
1497 | cond_resched(); /* don't hog the cpu */ | |
1498 | } | |
1499 | ||
1500 | return 0; | |
1501 | ||
1502 | bail_rcvegrbuf_phys: | |
1503 | for (e = 0; e < rcd->rcvegrbuf_chunks && rcd->rcvegrbuf[e]; e++) | |
1504 | dma_free_coherent(&dd->pcidev->dev, size, | |
1505 | rcd->rcvegrbuf[e], rcd->rcvegrbuf_phys[e]); | |
1506 | kfree(rcd->rcvegrbuf_phys); | |
1507 | rcd->rcvegrbuf_phys = NULL; | |
1508 | bail_rcvegrbuf: | |
1509 | kfree(rcd->rcvegrbuf); | |
1510 | rcd->rcvegrbuf = NULL; | |
1511 | bail: | |
1512 | return -ENOMEM; | |
1513 | } | |
1514 | ||
fce24a9d DO |
1515 | /* |
1516 | * Note: Changes to this routine should be mirrored | |
1517 | * for the diagnostics routine qib_remap_ioaddr32(). | |
1518 | * There is also related code for VL15 buffers in qib_init_7322_variables(). | |
1519 | * The teardown code that unmaps is in qib_pcie_ddcleanup() | |
1520 | */ | |
f931551b RC |
1521 | int init_chip_wc_pat(struct qib_devdata *dd, u32 vl15buflen) |
1522 | { | |
1523 | u64 __iomem *qib_kregbase = NULL; | |
1524 | void __iomem *qib_piobase = NULL; | |
1525 | u64 __iomem *qib_userbase = NULL; | |
1526 | u64 qib_kreglen; | |
1527 | u64 qib_pio2koffset = dd->piobufbase & 0xffffffff; | |
1528 | u64 qib_pio4koffset = dd->piobufbase >> 32; | |
1529 | u64 qib_pio2klen = dd->piobcnt2k * dd->palign; | |
1530 | u64 qib_pio4klen = dd->piobcnt4k * dd->align4k; | |
1531 | u64 qib_physaddr = dd->physaddr; | |
1532 | u64 qib_piolen; | |
1533 | u64 qib_userlen = 0; | |
1534 | ||
1535 | /* | |
1536 | * Free the old mapping because the kernel will try to reuse the | |
1537 | * old mapping and not create a new mapping with the | |
1538 | * write combining attribute. | |
1539 | */ | |
1540 | iounmap(dd->kregbase); | |
1541 | dd->kregbase = NULL; | |
1542 | ||
1543 | /* | |
1544 | * Assumes chip address space looks like: | |
1545 | * - kregs + sregs + cregs + uregs (in any order) | |
1546 | * - piobufs (2K and 4K bufs in either order) | |
1547 | * or: | |
1548 | * - kregs + sregs + cregs (in any order) | |
1549 | * - piobufs (2K and 4K bufs in either order) | |
1550 | * - uregs | |
1551 | */ | |
1552 | if (dd->piobcnt4k == 0) { | |
1553 | qib_kreglen = qib_pio2koffset; | |
1554 | qib_piolen = qib_pio2klen; | |
1555 | } else if (qib_pio2koffset < qib_pio4koffset) { | |
1556 | qib_kreglen = qib_pio2koffset; | |
1557 | qib_piolen = qib_pio4koffset + qib_pio4klen - qib_kreglen; | |
1558 | } else { | |
1559 | qib_kreglen = qib_pio4koffset; | |
1560 | qib_piolen = qib_pio2koffset + qib_pio2klen - qib_kreglen; | |
1561 | } | |
1562 | qib_piolen += vl15buflen; | |
1563 | /* Map just the configured ports (not all hw ports) */ | |
1564 | if (dd->uregbase > qib_kreglen) | |
1565 | qib_userlen = dd->ureg_align * dd->cfgctxts; | |
1566 | ||
1567 | /* Sanity checks passed, now create the new mappings */ | |
1568 | qib_kregbase = ioremap_nocache(qib_physaddr, qib_kreglen); | |
1569 | if (!qib_kregbase) | |
1570 | goto bail; | |
1571 | ||
1572 | qib_piobase = ioremap_wc(qib_physaddr + qib_kreglen, qib_piolen); | |
1573 | if (!qib_piobase) | |
1574 | goto bail_kregbase; | |
1575 | ||
1576 | if (qib_userlen) { | |
1577 | qib_userbase = ioremap_nocache(qib_physaddr + dd->uregbase, | |
1578 | qib_userlen); | |
1579 | if (!qib_userbase) | |
1580 | goto bail_piobase; | |
1581 | } | |
1582 | ||
1583 | dd->kregbase = qib_kregbase; | |
1584 | dd->kregend = (u64 __iomem *) | |
1585 | ((char __iomem *) qib_kregbase + qib_kreglen); | |
1586 | dd->piobase = qib_piobase; | |
1587 | dd->pio2kbase = (void __iomem *) | |
1588 | (((char __iomem *) dd->piobase) + | |
1589 | qib_pio2koffset - qib_kreglen); | |
1590 | if (dd->piobcnt4k) | |
1591 | dd->pio4kbase = (void __iomem *) | |
1592 | (((char __iomem *) dd->piobase) + | |
1593 | qib_pio4koffset - qib_kreglen); | |
1594 | if (qib_userlen) | |
1595 | /* ureg will now be accessed relative to dd->userbase */ | |
1596 | dd->userbase = qib_userbase; | |
1597 | return 0; | |
1598 | ||
1599 | bail_piobase: | |
1600 | iounmap(qib_piobase); | |
1601 | bail_kregbase: | |
1602 | iounmap(qib_kregbase); | |
1603 | bail: | |
1604 | return -ENOMEM; | |
1605 | } |