spu_base.c 18 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515516517518519520521522523524525526527528529530531532533534535536537538539540541542543544545546547548549550551552553554555556557558559560561562563564565566567568569570571572573574575576577578579580581582583584585586587588589590591592593594595596597598599600601602603604605606607608609610611612613614615616617618619620621622623624625626627628629630631632633634635636637638639640641642643644645646647648649650651652653654655656657658659660661662663664665666667668669670671672673674675676677678679680681682683684685686687688689690691692693694695696697698699700701702703704705706707708709710711712713714715716717718719720721722723724725726727728729730731732733734735736737738739740741742743744745746747748749750751752753754755756757758759760761762763
  1. /*
  2. * Low-level SPU handling
  3. *
  4. * (C) Copyright IBM Deutschland Entwicklung GmbH 2005
  5. *
  6. * Author: Arnd Bergmann <arndb@de.ibm.com>
  7. *
  8. * This program is free software; you can redistribute it and/or modify
  9. * it under the terms of the GNU General Public License as published by
  10. * the Free Software Foundation; either version 2, or (at your option)
  11. * any later version.
  12. *
  13. * This program is distributed in the hope that it will be useful,
  14. * but WITHOUT ANY WARRANTY; without even the implied warranty of
  15. * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
  16. * GNU General Public License for more details.
  17. *
  18. * You should have received a copy of the GNU General Public License
  19. * along with this program; if not, write to the Free Software
  20. * Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
  21. */
  22. #undef DEBUG
  23. #include <linux/interrupt.h>
  24. #include <linux/list.h>
  25. #include <linux/module.h>
  26. #include <linux/ptrace.h>
  27. #include <linux/slab.h>
  28. #include <linux/wait.h>
  29. #include <linux/mm.h>
  30. #include <linux/io.h>
  31. #include <linux/mutex.h>
  32. #include <linux/linux_logo.h>
  33. #include <asm/spu.h>
  34. #include <asm/spu_priv1.h>
  35. #include <asm/spu_csa.h>
  36. #include <asm/xmon.h>
  37. #include <asm/prom.h>
  38. const struct spu_management_ops *spu_management_ops;
  39. EXPORT_SYMBOL_GPL(spu_management_ops);
  40. const struct spu_priv1_ops *spu_priv1_ops;
  41. EXPORT_SYMBOL_GPL(spu_priv1_ops);
  42. struct cbe_spu_info cbe_spu_info[MAX_NUMNODES];
  43. EXPORT_SYMBOL_GPL(cbe_spu_info);
  44. /*
  45. * The spufs fault-handling code needs to call force_sig_info to raise signals
  46. * on DMA errors. Export it here to avoid general kernel-wide access to this
  47. * function
  48. */
  49. EXPORT_SYMBOL_GPL(force_sig_info);
  50. /*
  51. * Protects cbe_spu_info and spu->number.
  52. */
  53. static DEFINE_SPINLOCK(spu_lock);
  54. /*
  55. * List of all spus in the system.
  56. *
  57. * This list is iterated by callers from irq context and callers that
  58. * want to sleep. Thus modifications need to be done with both
  59. * spu_full_list_lock and spu_full_list_mutex held, while iterating
  60. * through it requires either of these locks.
  61. *
  62. * In addition spu_full_list_lock protects all assignmens to
  63. * spu->mm.
  64. */
  65. static LIST_HEAD(spu_full_list);
  66. static DEFINE_SPINLOCK(spu_full_list_lock);
  67. static DEFINE_MUTEX(spu_full_list_mutex);
  68. struct spu_slb {
  69. u64 esid, vsid;
  70. };
  71. void spu_invalidate_slbs(struct spu *spu)
  72. {
  73. struct spu_priv2 __iomem *priv2 = spu->priv2;
  74. unsigned long flags;
  75. spin_lock_irqsave(&spu->register_lock, flags);
  76. if (spu_mfc_sr1_get(spu) & MFC_STATE1_RELOCATE_MASK)
  77. out_be64(&priv2->slb_invalidate_all_W, 0UL);
  78. spin_unlock_irqrestore(&spu->register_lock, flags);
  79. }
  80. EXPORT_SYMBOL_GPL(spu_invalidate_slbs);
  81. /* This is called by the MM core when a segment size is changed, to
  82. * request a flush of all the SPEs using a given mm
  83. */
  84. void spu_flush_all_slbs(struct mm_struct *mm)
  85. {
  86. struct spu *spu;
  87. unsigned long flags;
  88. spin_lock_irqsave(&spu_full_list_lock, flags);
  89. list_for_each_entry(spu, &spu_full_list, full_list) {
  90. if (spu->mm == mm)
  91. spu_invalidate_slbs(spu);
  92. }
  93. spin_unlock_irqrestore(&spu_full_list_lock, flags);
  94. }
  95. /* The hack below stinks... try to do something better one of
  96. * these days... Does it even work properly with NR_CPUS == 1 ?
  97. */
  98. static inline void mm_needs_global_tlbie(struct mm_struct *mm)
  99. {
  100. int nr = (NR_CPUS > 1) ? NR_CPUS : NR_CPUS + 1;
  101. /* Global TLBIE broadcast required with SPEs. */
  102. __cpus_setall(&mm->cpu_vm_mask, nr);
  103. }
  104. void spu_associate_mm(struct spu *spu, struct mm_struct *mm)
  105. {
  106. unsigned long flags;
  107. spin_lock_irqsave(&spu_full_list_lock, flags);
  108. spu->mm = mm;
  109. spin_unlock_irqrestore(&spu_full_list_lock, flags);
  110. if (mm)
  111. mm_needs_global_tlbie(mm);
  112. }
  113. EXPORT_SYMBOL_GPL(spu_associate_mm);
  114. int spu_64k_pages_available(void)
  115. {
  116. return mmu_psize_defs[MMU_PAGE_64K].shift != 0;
  117. }
  118. EXPORT_SYMBOL_GPL(spu_64k_pages_available);
  119. static void spu_restart_dma(struct spu *spu)
  120. {
  121. struct spu_priv2 __iomem *priv2 = spu->priv2;
  122. if (!test_bit(SPU_CONTEXT_SWITCH_PENDING, &spu->flags))
  123. out_be64(&priv2->mfc_control_RW, MFC_CNTL_RESTART_DMA_COMMAND);
  124. }
  125. static inline void spu_load_slb(struct spu *spu, int slbe, struct spu_slb *slb)
  126. {
  127. struct spu_priv2 __iomem *priv2 = spu->priv2;
  128. pr_debug("%s: adding SLB[%d] 0x%016lx 0x%016lx\n",
  129. __func__, slbe, slb->vsid, slb->esid);
  130. out_be64(&priv2->slb_index_W, slbe);
  131. /* set invalid before writing vsid */
  132. out_be64(&priv2->slb_esid_RW, 0);
  133. /* now it's safe to write the vsid */
  134. out_be64(&priv2->slb_vsid_RW, slb->vsid);
  135. /* setting the new esid makes the entry valid again */
  136. out_be64(&priv2->slb_esid_RW, slb->esid);
  137. }
  138. static int __spu_trap_data_seg(struct spu *spu, unsigned long ea)
  139. {
  140. struct mm_struct *mm = spu->mm;
  141. struct spu_slb slb;
  142. int psize;
  143. pr_debug("%s\n", __func__);
  144. slb.esid = (ea & ESID_MASK) | SLB_ESID_V;
  145. switch(REGION_ID(ea)) {
  146. case USER_REGION_ID:
  147. #ifdef CONFIG_PPC_MM_SLICES
  148. psize = get_slice_psize(mm, ea);
  149. #else
  150. psize = mm->context.user_psize;
  151. #endif
  152. slb.vsid = (get_vsid(mm->context.id, ea, MMU_SEGSIZE_256M)
  153. << SLB_VSID_SHIFT) | SLB_VSID_USER;
  154. break;
  155. case VMALLOC_REGION_ID:
  156. if (ea < VMALLOC_END)
  157. psize = mmu_vmalloc_psize;
  158. else
  159. psize = mmu_io_psize;
  160. slb.vsid = (get_kernel_vsid(ea, MMU_SEGSIZE_256M)
  161. << SLB_VSID_SHIFT) | SLB_VSID_KERNEL;
  162. break;
  163. case KERNEL_REGION_ID:
  164. psize = mmu_linear_psize;
  165. slb.vsid = (get_kernel_vsid(ea, MMU_SEGSIZE_256M)
  166. << SLB_VSID_SHIFT) | SLB_VSID_KERNEL;
  167. break;
  168. default:
  169. /* Future: support kernel segments so that drivers
  170. * can use SPUs.
  171. */
  172. pr_debug("invalid region access at %016lx\n", ea);
  173. return 1;
  174. }
  175. slb.vsid |= mmu_psize_defs[psize].sllp;
  176. spu_load_slb(spu, spu->slb_replace, &slb);
  177. spu->slb_replace++;
  178. if (spu->slb_replace >= 8)
  179. spu->slb_replace = 0;
  180. spu_restart_dma(spu);
  181. spu->stats.slb_flt++;
  182. return 0;
  183. }
  184. extern int hash_page(unsigned long ea, unsigned long access, unsigned long trap); //XXX
  185. static int __spu_trap_data_map(struct spu *spu, unsigned long ea, u64 dsisr)
  186. {
  187. pr_debug("%s, %lx, %lx\n", __func__, dsisr, ea);
  188. /* Handle kernel space hash faults immediately.
  189. User hash faults need to be deferred to process context. */
  190. if ((dsisr & MFC_DSISR_PTE_NOT_FOUND)
  191. && REGION_ID(ea) != USER_REGION_ID
  192. && hash_page(ea, _PAGE_PRESENT, 0x300) == 0) {
  193. spu_restart_dma(spu);
  194. return 0;
  195. }
  196. spu->class_0_pending = 0;
  197. spu->dar = ea;
  198. spu->dsisr = dsisr;
  199. spu->stop_callback(spu);
  200. return 0;
  201. }
  202. static void __spu_kernel_slb(void *addr, struct spu_slb *slb)
  203. {
  204. unsigned long ea = (unsigned long)addr;
  205. u64 llp;
  206. if (REGION_ID(ea) == KERNEL_REGION_ID)
  207. llp = mmu_psize_defs[mmu_linear_psize].sllp;
  208. else
  209. llp = mmu_psize_defs[mmu_virtual_psize].sllp;
  210. slb->vsid = (get_kernel_vsid(ea, MMU_SEGSIZE_256M) << SLB_VSID_SHIFT) |
  211. SLB_VSID_KERNEL | llp;
  212. slb->esid = (ea & ESID_MASK) | SLB_ESID_V;
  213. }
  214. /**
  215. * Given an array of @nr_slbs SLB entries, @slbs, return non-zero if the
  216. * address @new_addr is present.
  217. */
  218. static inline int __slb_present(struct spu_slb *slbs, int nr_slbs,
  219. void *new_addr)
  220. {
  221. unsigned long ea = (unsigned long)new_addr;
  222. int i;
  223. for (i = 0; i < nr_slbs; i++)
  224. if (!((slbs[i].esid ^ ea) & ESID_MASK))
  225. return 1;
  226. return 0;
  227. }
  228. /**
  229. * Setup the SPU kernel SLBs, in preparation for a context save/restore. We
  230. * need to map both the context save area, and the save/restore code.
  231. *
  232. * Because the lscsa and code may cross segment boundaires, we check to see
  233. * if mappings are required for the start and end of each range. We currently
  234. * assume that the mappings are smaller that one segment - if not, something
  235. * is seriously wrong.
  236. */
  237. void spu_setup_kernel_slbs(struct spu *spu, struct spu_lscsa *lscsa,
  238. void *code, int code_size)
  239. {
  240. struct spu_slb slbs[4];
  241. int i, nr_slbs = 0;
  242. /* start and end addresses of both mappings */
  243. void *addrs[] = {
  244. lscsa, (void *)lscsa + sizeof(*lscsa) - 1,
  245. code, code + code_size - 1
  246. };
  247. /* check the set of addresses, and create a new entry in the slbs array
  248. * if there isn't already a SLB for that address */
  249. for (i = 0; i < ARRAY_SIZE(addrs); i++) {
  250. if (__slb_present(slbs, nr_slbs, addrs[i]))
  251. continue;
  252. __spu_kernel_slb(addrs[i], &slbs[nr_slbs]);
  253. nr_slbs++;
  254. }
  255. spin_lock_irq(&spu->register_lock);
  256. /* Add the set of SLBs */
  257. for (i = 0; i < nr_slbs; i++)
  258. spu_load_slb(spu, i, &slbs[i]);
  259. spin_unlock_irq(&spu->register_lock);
  260. }
  261. EXPORT_SYMBOL_GPL(spu_setup_kernel_slbs);
  262. static irqreturn_t
  263. spu_irq_class_0(int irq, void *data)
  264. {
  265. struct spu *spu;
  266. unsigned long stat, mask;
  267. spu = data;
  268. spin_lock(&spu->register_lock);
  269. mask = spu_int_mask_get(spu, 0);
  270. stat = spu_int_stat_get(spu, 0) & mask;
  271. spu->class_0_pending |= stat;
  272. spu->dsisr = spu_mfc_dsisr_get(spu);
  273. spu->dar = spu_mfc_dar_get(spu);
  274. spin_unlock(&spu->register_lock);
  275. spu->stop_callback(spu);
  276. spu_int_stat_clear(spu, 0, stat);
  277. return IRQ_HANDLED;
  278. }
  279. static irqreturn_t
  280. spu_irq_class_1(int irq, void *data)
  281. {
  282. struct spu *spu;
  283. unsigned long stat, mask, dar, dsisr;
  284. spu = data;
  285. /* atomically read & clear class1 status. */
  286. spin_lock(&spu->register_lock);
  287. mask = spu_int_mask_get(spu, 1);
  288. stat = spu_int_stat_get(spu, 1) & mask;
  289. dar = spu_mfc_dar_get(spu);
  290. dsisr = spu_mfc_dsisr_get(spu);
  291. if (stat & CLASS1_STORAGE_FAULT_INTR)
  292. spu_mfc_dsisr_set(spu, 0ul);
  293. spu_int_stat_clear(spu, 1, stat);
  294. if (stat & CLASS1_SEGMENT_FAULT_INTR)
  295. __spu_trap_data_seg(spu, dar);
  296. spin_unlock(&spu->register_lock);
  297. pr_debug("%s: %lx %lx %lx %lx\n", __func__, mask, stat,
  298. dar, dsisr);
  299. if (stat & CLASS1_STORAGE_FAULT_INTR)
  300. __spu_trap_data_map(spu, dar, dsisr);
  301. if (stat & CLASS1_LS_COMPARE_SUSPEND_ON_GET_INTR)
  302. ;
  303. if (stat & CLASS1_LS_COMPARE_SUSPEND_ON_PUT_INTR)
  304. ;
  305. return stat ? IRQ_HANDLED : IRQ_NONE;
  306. }
  307. static irqreturn_t
  308. spu_irq_class_2(int irq, void *data)
  309. {
  310. struct spu *spu;
  311. unsigned long stat;
  312. unsigned long mask;
  313. const int mailbox_intrs =
  314. CLASS2_MAILBOX_THRESHOLD_INTR | CLASS2_MAILBOX_INTR;
  315. spu = data;
  316. spin_lock(&spu->register_lock);
  317. stat = spu_int_stat_get(spu, 2);
  318. mask = spu_int_mask_get(spu, 2);
  319. /* ignore interrupts we're not waiting for */
  320. stat &= mask;
  321. /* mailbox interrupts are level triggered. mask them now before
  322. * acknowledging */
  323. if (stat & mailbox_intrs)
  324. spu_int_mask_and(spu, 2, ~(stat & mailbox_intrs));
  325. /* acknowledge all interrupts before the callbacks */
  326. spu_int_stat_clear(spu, 2, stat);
  327. spin_unlock(&spu->register_lock);
  328. pr_debug("class 2 interrupt %d, %lx, %lx\n", irq, stat, mask);
  329. if (stat & CLASS2_MAILBOX_INTR)
  330. spu->ibox_callback(spu);
  331. if (stat & CLASS2_SPU_STOP_INTR)
  332. spu->stop_callback(spu);
  333. if (stat & CLASS2_SPU_HALT_INTR)
  334. spu->stop_callback(spu);
  335. if (stat & CLASS2_SPU_DMA_TAG_GROUP_COMPLETE_INTR)
  336. spu->mfc_callback(spu);
  337. if (stat & CLASS2_MAILBOX_THRESHOLD_INTR)
  338. spu->wbox_callback(spu);
  339. spu->stats.class2_intr++;
  340. return stat ? IRQ_HANDLED : IRQ_NONE;
  341. }
  342. static int spu_request_irqs(struct spu *spu)
  343. {
  344. int ret = 0;
  345. if (spu->irqs[0] != NO_IRQ) {
  346. snprintf(spu->irq_c0, sizeof (spu->irq_c0), "spe%02d.0",
  347. spu->number);
  348. ret = request_irq(spu->irqs[0], spu_irq_class_0,
  349. IRQF_DISABLED,
  350. spu->irq_c0, spu);
  351. if (ret)
  352. goto bail0;
  353. }
  354. if (spu->irqs[1] != NO_IRQ) {
  355. snprintf(spu->irq_c1, sizeof (spu->irq_c1), "spe%02d.1",
  356. spu->number);
  357. ret = request_irq(spu->irqs[1], spu_irq_class_1,
  358. IRQF_DISABLED,
  359. spu->irq_c1, spu);
  360. if (ret)
  361. goto bail1;
  362. }
  363. if (spu->irqs[2] != NO_IRQ) {
  364. snprintf(spu->irq_c2, sizeof (spu->irq_c2), "spe%02d.2",
  365. spu->number);
  366. ret = request_irq(spu->irqs[2], spu_irq_class_2,
  367. IRQF_DISABLED,
  368. spu->irq_c2, spu);
  369. if (ret)
  370. goto bail2;
  371. }
  372. return 0;
  373. bail2:
  374. if (spu->irqs[1] != NO_IRQ)
  375. free_irq(spu->irqs[1], spu);
  376. bail1:
  377. if (spu->irqs[0] != NO_IRQ)
  378. free_irq(spu->irqs[0], spu);
  379. bail0:
  380. return ret;
  381. }
  382. static void spu_free_irqs(struct spu *spu)
  383. {
  384. if (spu->irqs[0] != NO_IRQ)
  385. free_irq(spu->irqs[0], spu);
  386. if (spu->irqs[1] != NO_IRQ)
  387. free_irq(spu->irqs[1], spu);
  388. if (spu->irqs[2] != NO_IRQ)
  389. free_irq(spu->irqs[2], spu);
  390. }
  391. void spu_init_channels(struct spu *spu)
  392. {
  393. static const struct {
  394. unsigned channel;
  395. unsigned count;
  396. } zero_list[] = {
  397. { 0x00, 1, }, { 0x01, 1, }, { 0x03, 1, }, { 0x04, 1, },
  398. { 0x18, 1, }, { 0x19, 1, }, { 0x1b, 1, }, { 0x1d, 1, },
  399. }, count_list[] = {
  400. { 0x00, 0, }, { 0x03, 0, }, { 0x04, 0, }, { 0x15, 16, },
  401. { 0x17, 1, }, { 0x18, 0, }, { 0x19, 0, }, { 0x1b, 0, },
  402. { 0x1c, 1, }, { 0x1d, 0, }, { 0x1e, 1, },
  403. };
  404. struct spu_priv2 __iomem *priv2;
  405. int i;
  406. priv2 = spu->priv2;
  407. /* initialize all channel data to zero */
  408. for (i = 0; i < ARRAY_SIZE(zero_list); i++) {
  409. int count;
  410. out_be64(&priv2->spu_chnlcntptr_RW, zero_list[i].channel);
  411. for (count = 0; count < zero_list[i].count; count++)
  412. out_be64(&priv2->spu_chnldata_RW, 0);
  413. }
  414. /* initialize channel counts to meaningful values */
  415. for (i = 0; i < ARRAY_SIZE(count_list); i++) {
  416. out_be64(&priv2->spu_chnlcntptr_RW, count_list[i].channel);
  417. out_be64(&priv2->spu_chnlcnt_RW, count_list[i].count);
  418. }
  419. }
  420. EXPORT_SYMBOL_GPL(spu_init_channels);
  421. static int spu_shutdown(struct sys_device *sysdev)
  422. {
  423. struct spu *spu = container_of(sysdev, struct spu, sysdev);
  424. spu_free_irqs(spu);
  425. spu_destroy_spu(spu);
  426. return 0;
  427. }
  428. static struct sysdev_class spu_sysdev_class = {
  429. .name = "spu",
  430. .shutdown = spu_shutdown,
  431. };
  432. int spu_add_sysdev_attr(struct sysdev_attribute *attr)
  433. {
  434. struct spu *spu;
  435. mutex_lock(&spu_full_list_mutex);
  436. list_for_each_entry(spu, &spu_full_list, full_list)
  437. sysdev_create_file(&spu->sysdev, attr);
  438. mutex_unlock(&spu_full_list_mutex);
  439. return 0;
  440. }
  441. EXPORT_SYMBOL_GPL(spu_add_sysdev_attr);
  442. int spu_add_sysdev_attr_group(struct attribute_group *attrs)
  443. {
  444. struct spu *spu;
  445. int rc = 0;
  446. mutex_lock(&spu_full_list_mutex);
  447. list_for_each_entry(spu, &spu_full_list, full_list) {
  448. rc = sysfs_create_group(&spu->sysdev.kobj, attrs);
  449. /* we're in trouble here, but try unwinding anyway */
  450. if (rc) {
  451. printk(KERN_ERR "%s: can't create sysfs group '%s'\n",
  452. __func__, attrs->name);
  453. list_for_each_entry_continue_reverse(spu,
  454. &spu_full_list, full_list)
  455. sysfs_remove_group(&spu->sysdev.kobj, attrs);
  456. break;
  457. }
  458. }
  459. mutex_unlock(&spu_full_list_mutex);
  460. return rc;
  461. }
  462. EXPORT_SYMBOL_GPL(spu_add_sysdev_attr_group);
  463. void spu_remove_sysdev_attr(struct sysdev_attribute *attr)
  464. {
  465. struct spu *spu;
  466. mutex_lock(&spu_full_list_mutex);
  467. list_for_each_entry(spu, &spu_full_list, full_list)
  468. sysdev_remove_file(&spu->sysdev, attr);
  469. mutex_unlock(&spu_full_list_mutex);
  470. }
  471. EXPORT_SYMBOL_GPL(spu_remove_sysdev_attr);
  472. void spu_remove_sysdev_attr_group(struct attribute_group *attrs)
  473. {
  474. struct spu *spu;
  475. mutex_lock(&spu_full_list_mutex);
  476. list_for_each_entry(spu, &spu_full_list, full_list)
  477. sysfs_remove_group(&spu->sysdev.kobj, attrs);
  478. mutex_unlock(&spu_full_list_mutex);
  479. }
  480. EXPORT_SYMBOL_GPL(spu_remove_sysdev_attr_group);
  481. static int spu_create_sysdev(struct spu *spu)
  482. {
  483. int ret;
  484. spu->sysdev.id = spu->number;
  485. spu->sysdev.cls = &spu_sysdev_class;
  486. ret = sysdev_register(&spu->sysdev);
  487. if (ret) {
  488. printk(KERN_ERR "Can't register SPU %d with sysfs\n",
  489. spu->number);
  490. return ret;
  491. }
  492. sysfs_add_device_to_node(&spu->sysdev, spu->node);
  493. return 0;
  494. }
  495. static int __init create_spu(void *data)
  496. {
  497. struct spu *spu;
  498. int ret;
  499. static int number;
  500. unsigned long flags;
  501. struct timespec ts;
  502. ret = -ENOMEM;
  503. spu = kzalloc(sizeof (*spu), GFP_KERNEL);
  504. if (!spu)
  505. goto out;
  506. spu->alloc_state = SPU_FREE;
  507. spin_lock_init(&spu->register_lock);
  508. spin_lock(&spu_lock);
  509. spu->number = number++;
  510. spin_unlock(&spu_lock);
  511. ret = spu_create_spu(spu, data);
  512. if (ret)
  513. goto out_free;
  514. spu_mfc_sdr_setup(spu);
  515. spu_mfc_sr1_set(spu, 0x33);
  516. ret = spu_request_irqs(spu);
  517. if (ret)
  518. goto out_destroy;
  519. ret = spu_create_sysdev(spu);
  520. if (ret)
  521. goto out_free_irqs;
  522. mutex_lock(&cbe_spu_info[spu->node].list_mutex);
  523. list_add(&spu->cbe_list, &cbe_spu_info[spu->node].spus);
  524. cbe_spu_info[spu->node].n_spus++;
  525. mutex_unlock(&cbe_spu_info[spu->node].list_mutex);
  526. mutex_lock(&spu_full_list_mutex);
  527. spin_lock_irqsave(&spu_full_list_lock, flags);
  528. list_add(&spu->full_list, &spu_full_list);
  529. spin_unlock_irqrestore(&spu_full_list_lock, flags);
  530. mutex_unlock(&spu_full_list_mutex);
  531. spu->stats.util_state = SPU_UTIL_IDLE_LOADED;
  532. ktime_get_ts(&ts);
  533. spu->stats.tstamp = timespec_to_ns(&ts);
  534. INIT_LIST_HEAD(&spu->aff_list);
  535. goto out;
  536. out_free_irqs:
  537. spu_free_irqs(spu);
  538. out_destroy:
  539. spu_destroy_spu(spu);
  540. out_free:
  541. kfree(spu);
  542. out:
  543. return ret;
  544. }
  545. static const char *spu_state_names[] = {
  546. "user", "system", "iowait", "idle"
  547. };
  548. static unsigned long long spu_acct_time(struct spu *spu,
  549. enum spu_utilization_state state)
  550. {
  551. struct timespec ts;
  552. unsigned long long time = spu->stats.times[state];
  553. /*
  554. * If the spu is idle or the context is stopped, utilization
  555. * statistics are not updated. Apply the time delta from the
  556. * last recorded state of the spu.
  557. */
  558. if (spu->stats.util_state == state) {
  559. ktime_get_ts(&ts);
  560. time += timespec_to_ns(&ts) - spu->stats.tstamp;
  561. }
  562. return time / NSEC_PER_MSEC;
  563. }
  564. static ssize_t spu_stat_show(struct sys_device *sysdev, char *buf)
  565. {
  566. struct spu *spu = container_of(sysdev, struct spu, sysdev);
  567. return sprintf(buf, "%s %llu %llu %llu %llu "
  568. "%llu %llu %llu %llu %llu %llu %llu %llu\n",
  569. spu_state_names[spu->stats.util_state],
  570. spu_acct_time(spu, SPU_UTIL_USER),
  571. spu_acct_time(spu, SPU_UTIL_SYSTEM),
  572. spu_acct_time(spu, SPU_UTIL_IOWAIT),
  573. spu_acct_time(spu, SPU_UTIL_IDLE_LOADED),
  574. spu->stats.vol_ctx_switch,
  575. spu->stats.invol_ctx_switch,
  576. spu->stats.slb_flt,
  577. spu->stats.hash_flt,
  578. spu->stats.min_flt,
  579. spu->stats.maj_flt,
  580. spu->stats.class2_intr,
  581. spu->stats.libassist);
  582. }
  583. static SYSDEV_ATTR(stat, 0644, spu_stat_show, NULL);
  584. static int __init init_spu_base(void)
  585. {
  586. int i, ret = 0;
  587. for (i = 0; i < MAX_NUMNODES; i++) {
  588. mutex_init(&cbe_spu_info[i].list_mutex);
  589. INIT_LIST_HEAD(&cbe_spu_info[i].spus);
  590. }
  591. if (!spu_management_ops)
  592. goto out;
  593. /* create sysdev class for spus */
  594. ret = sysdev_class_register(&spu_sysdev_class);
  595. if (ret)
  596. goto out;
  597. ret = spu_enumerate_spus(create_spu);
  598. if (ret < 0) {
  599. printk(KERN_WARNING "%s: Error initializing spus\n",
  600. __func__);
  601. goto out_unregister_sysdev_class;
  602. }
  603. if (ret > 0) {
  604. /*
  605. * We cannot put the forward declaration in
  606. * <linux/linux_logo.h> because of conflicting session type
  607. * conflicts for const and __initdata with different compiler
  608. * versions
  609. */
  610. extern const struct linux_logo logo_spe_clut224;
  611. fb_append_extra_logo(&logo_spe_clut224, ret);
  612. }
  613. mutex_lock(&spu_full_list_mutex);
  614. xmon_register_spus(&spu_full_list);
  615. crash_register_spus(&spu_full_list);
  616. mutex_unlock(&spu_full_list_mutex);
  617. spu_add_sysdev_attr(&attr_stat);
  618. spu_init_affinity();
  619. return 0;
  620. out_unregister_sysdev_class:
  621. sysdev_class_unregister(&spu_sysdev_class);
  622. out:
  623. return ret;
  624. }
  625. module_init(init_spu_base);
  626. MODULE_LICENSE("GPL");
  627. MODULE_AUTHOR("Arnd Bergmann <arndb@de.ibm.com>");