xref: /illumos-gate/usr/src/uts/sun4v/cpu/niagara2.c (revision aaa10e6791d1614700651df2821f84d490c094bf)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2007 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 #include <sys/types.h>
30 #include <sys/systm.h>
31 #include <sys/archsystm.h>
32 #include <sys/machparam.h>
33 #include <sys/machsystm.h>
34 #include <sys/cpu.h>
35 #include <sys/elf_SPARC.h>
36 #include <vm/hat_sfmmu.h>
37 #include <vm/page.h>
38 #include <vm/vm_dep.h>
39 #include <sys/cpuvar.h>
40 #include <sys/async.h>
41 #include <sys/cmn_err.h>
42 #include <sys/debug.h>
43 #include <sys/dditypes.h>
44 #include <sys/sunddi.h>
45 #include <sys/cpu_module.h>
46 #include <sys/prom_debug.h>
47 #include <sys/vmsystm.h>
48 #include <sys/prom_plat.h>
49 #include <sys/sysmacros.h>
50 #include <sys/intreg.h>
51 #include <sys/machtrap.h>
52 #include <sys/ontrap.h>
53 #include <sys/ivintr.h>
54 #include <sys/atomic.h>
55 #include <sys/panic.h>
56 #include <sys/dtrace.h>
57 #include <sys/simulate.h>
58 #include <sys/fault.h>
59 #include <sys/niagara2regs.h>
60 #include <sys/hsvc.h>
61 #include <sys/trapstat.h>
62 
63 uint_t root_phys_addr_lo_mask = 0xffffffffU;
64 char cpu_module_name[] = "SUNW,UltraSPARC-T2";
65 
66 /*
67  * Hypervisor services information for the NIAGARA2 CPU module
68  */
69 static boolean_t niagara2_hsvc_available = B_TRUE;
70 static uint64_t niagara2_sup_minor;		/* Supported minor number */
71 static hsvc_info_t niagara2_hsvc = {
72 	HSVC_REV_1, NULL, HSVC_GROUP_NIAGARA2_CPU, NIAGARA2_HSVC_MAJOR,
73 	NIAGARA2_HSVC_MINOR, cpu_module_name
74 };
75 
76 void
77 cpu_setup(void)
78 {
79 	extern int mmu_exported_pagesize_mask;
80 	extern int cpc_has_overflow_intr;
81 	extern size_t contig_mem_prealloc_base;
82 	int status;
83 
84 	/*
85 	 * Negotiate the API version for Niagara2 specific hypervisor
86 	 * services.
87 	 */
88 	status = hsvc_register(&niagara2_hsvc, &niagara2_sup_minor);
89 	if (status != 0) {
90 		cmn_err(CE_WARN, "%s: cannot negotiate hypervisor services "
91 		    "group: 0x%lx major: 0x%lx minor: 0x%lx errno: %d",
92 		    niagara2_hsvc.hsvc_modname, niagara2_hsvc.hsvc_group,
93 		    niagara2_hsvc.hsvc_major, niagara2_hsvc.hsvc_minor, status);
94 		niagara2_hsvc_available = B_FALSE;
95 	}
96 
97 	/*
98 	 * The setup common to all CPU modules is done in cpu_setup_common
99 	 * routine.
100 	 */
101 	cpu_setup_common(NULL);
102 
103 	cache |= (CACHE_PTAG | CACHE_IOCOHERENT);
104 
105 	if ((mmu_exported_pagesize_mask &
106 	    DEFAULT_SUN4V_MMU_PAGESIZE_MASK) !=
107 	    DEFAULT_SUN4V_MMU_PAGESIZE_MASK)
108 		cmn_err(CE_PANIC, "machine description"
109 		    " does not have required sun4v page sizes"
110 		    " 8K, 64K and 4M: MD mask is 0x%x",
111 		    mmu_exported_pagesize_mask);
112 
113 	cpu_hwcap_flags = AV_SPARC_VIS | AV_SPARC_VIS2 | AV_SPARC_ASI_BLK_INIT;
114 
115 	/*
116 	 * Niagara2 supports a 48-bit subset of the full 64-bit virtual
117 	 * address space. Virtual addresses between 0x0000800000000000
118 	 * and 0xffff.7fff.ffff.ffff inclusive lie within a "VA Hole"
119 	 * and must never be mapped. In addition, software must not use
120 	 * pages within 4GB of the VA hole as instruction pages to
121 	 * avoid problems with prefetching into the VA hole.
122 	 */
123 	hole_start = (caddr_t)((1ull << (va_bits - 1)) - (1ull << 32));
124 	hole_end = (caddr_t)((0ull - (1ull << (va_bits - 1))) + (1ull << 32));
125 
126 	/*
127 	 * Niagara2 has a performance counter overflow interrupt
128 	 */
129 	cpc_has_overflow_intr = 1;
130 
131 	/*
132 	 * Enable 4M pages for OOB.
133 	 */
134 	max_uheap_lpsize = MMU_PAGESIZE4M;
135 	max_ustack_lpsize = MMU_PAGESIZE4M;
136 	max_privmap_lpsize = MMU_PAGESIZE4M;
137 
138 	contig_mem_prealloc_base = NIAGARA2_PREALLOC_BASE;
139 }
140 
141 /*
142  * Set the magic constants of the implementation.
143  */
144 void
145 cpu_fiximp(struct cpu_node *cpunode)
146 {
147 	/*
148 	 * The Cache node is optional in MD. Therefore in case "Cache"
149 	 * node does not exists in MD, set the default L2 cache associativity,
150 	 * size, linesize.
151 	 */
152 	if (cpunode->ecache_size == 0)
153 		cpunode->ecache_size = L2CACHE_SIZE;
154 	if (cpunode->ecache_linesize == 0)
155 		cpunode->ecache_linesize = L2CACHE_LINESIZE;
156 	if (cpunode->ecache_associativity == 0)
157 		cpunode->ecache_associativity = L2CACHE_ASSOCIATIVITY;
158 }
159 
160 void
161 cpu_map_exec_units(struct cpu *cp)
162 {
163 	ASSERT(MUTEX_HELD(&cpu_lock));
164 
165 	/*
166 	 * The cpu_ipipe and cpu_fpu fields are initialized based on
167 	 * the execution unit sharing information from the MD. They
168 	 * default to the CPU id in the absence of such information.
169 	 */
170 	cp->cpu_m.cpu_ipipe = cpunodes[cp->cpu_id].exec_unit_mapping;
171 	if (cp->cpu_m.cpu_ipipe == NO_EU_MAPPING_FOUND)
172 		cp->cpu_m.cpu_ipipe = (id_t)(cp->cpu_id);
173 
174 	cp->cpu_m.cpu_fpu = cpunodes[cp->cpu_id].fpu_mapping;
175 	if (cp->cpu_m.cpu_fpu == NO_EU_MAPPING_FOUND)
176 		cp->cpu_m.cpu_fpu = (id_t)(cp->cpu_id);
177 
178 	/*
179 	 * Niagara 2 defines the core to be at the FPU level
180 	 */
181 	cp->cpu_m.cpu_core = cp->cpu_m.cpu_fpu;
182 }
183 
184 static int niagara2_cpucnt;
185 
186 void
187 cpu_init_private(struct cpu *cp)
188 {
189 	extern void niagara_kstat_init(void);
190 
191 	ASSERT(MUTEX_HELD(&cpu_lock));
192 
193 	cpu_map_exec_units(cp);
194 
195 	if ((niagara2_cpucnt++ == 0) && (niagara2_hsvc_available == B_TRUE))
196 		niagara_kstat_init();
197 }
198 
199 /*ARGSUSED*/
200 void
201 cpu_uninit_private(struct cpu *cp)
202 {
203 	extern void niagara_kstat_fini(void);
204 
205 	ASSERT(MUTEX_HELD(&cpu_lock));
206 
207 	if ((--niagara2_cpucnt == 0) && (niagara2_hsvc_available == B_TRUE))
208 		niagara_kstat_fini();
209 }
210 
211 /*
212  * On Niagara2, any flush will cause all preceding stores to be
213  * synchronized wrt the i$, regardless of address or ASI.  In fact,
214  * the address is ignored, so we always flush address 0.
215  */
216 /*ARGSUSED*/
217 void
218 dtrace_flush_sec(uintptr_t addr)
219 {
220 	doflush(0);
221 }
222 
223 /*
224  * Trapstat support for Niagara2 processor
225  * The Niagara2 provides HWTW support for TSB lookup and with HWTW
226  * enabled no TSB hit information will be available. Therefore setting
227  * the time spent in TLB miss handler for TSB hits to 0.
228  */
229 int
230 cpu_trapstat_conf(int cmd)
231 {
232 	int status = 0;
233 
234 	switch (cmd) {
235 	case CPU_TSTATCONF_INIT:
236 	case CPU_TSTATCONF_FINI:
237 	case CPU_TSTATCONF_ENABLE:
238 	case CPU_TSTATCONF_DISABLE:
239 		break;
240 	default:
241 		status = EINVAL;
242 		break;
243 	}
244 	return (status);
245 }
246 
247 void
248 cpu_trapstat_data(void *buf, uint_t tstat_pgszs)
249 {
250 	tstat_pgszdata_t	*tstatp = (tstat_pgszdata_t *)buf;
251 	int	i;
252 
253 	for (i = 0; i < tstat_pgszs; i++, tstatp++) {
254 		tstatp->tpgsz_kernel.tmode_itlb.ttlb_tlb.tmiss_count = 0;
255 		tstatp->tpgsz_kernel.tmode_itlb.ttlb_tlb.tmiss_time = 0;
256 		tstatp->tpgsz_user.tmode_itlb.ttlb_tlb.tmiss_count = 0;
257 		tstatp->tpgsz_user.tmode_itlb.ttlb_tlb.tmiss_time = 0;
258 		tstatp->tpgsz_kernel.tmode_dtlb.ttlb_tlb.tmiss_count = 0;
259 		tstatp->tpgsz_kernel.tmode_dtlb.ttlb_tlb.tmiss_time = 0;
260 		tstatp->tpgsz_user.tmode_dtlb.ttlb_tlb.tmiss_count = 0;
261 		tstatp->tpgsz_user.tmode_dtlb.ttlb_tlb.tmiss_time = 0;
262 	}
263 }
264 
265 /* NI2 L2$ index is pa[32:28]^pa[17:13].pa[19:18]^pa[12:11].pa[10:6] */
266 uint_t
267 page_pfn_2_color_cpu(pfn_t pfn, uchar_t szc)
268 {
269 	uint_t color;
270 
271 	ASSERT(szc <= TTE256M);
272 
273 	pfn = PFN_BASE(pfn, szc);
274 	color = ((pfn >> 15) ^ pfn) & 0x1f;
275 	if (szc >= TTE4M)
276 		return (color);
277 
278 	color = (color << 2) | ((pfn >> 5) & 0x3);
279 
280 	return (szc <= TTE64K ? color : (color >> 1));
281 }
282 
283 #if TTE256M != 5
284 #error TTE256M is not 5
285 #endif
286 
287 uint_t
288 page_get_nsz_color_mask_cpu(uchar_t szc, uint_t mask)
289 {
290 	static uint_t ni2_color_masks[5] = {0x63, 0x1e, 0x3e, 0x1f, 0x1f};
291 	ASSERT(szc < TTE256M);
292 
293 	mask &= ni2_color_masks[szc];
294 	return ((szc == TTE64K || szc == TTE512K) ? (mask >> 1) : mask);
295 }
296 
297 uint_t
298 page_get_nsz_color_cpu(uchar_t szc, uint_t color)
299 {
300 	ASSERT(szc < TTE256M);
301 	return ((szc == TTE64K || szc == TTE512K) ? (color >> 1) : color);
302 }
303 
304 uint_t
305 page_get_color_shift_cpu(uchar_t szc, uchar_t nszc)
306 {
307 	ASSERT(nszc > szc);
308 	ASSERT(nszc <= TTE256M);
309 
310 	if (szc <= TTE64K)
311 		return ((nszc >= TTE4M) ? 2 : ((nszc >= TTE512K) ? 1 : 0));
312 	if (szc == TTE512K)
313 		return (1);
314 
315 	return (0);
316 }
317 
318 /*ARGSUSED*/
319 pfn_t
320 page_next_pfn_for_color_cpu(pfn_t pfn, uchar_t szc, uint_t color,
321     uint_t ceq_mask, uint_t color_mask)
322 {
323 	pfn_t pstep = PNUM_SIZE(szc);
324 	pfn_t npfn, pfn_ceq_mask, pfn_color;
325 	pfn_t tmpmask, mask = (pfn_t)-1;
326 
327 	ASSERT((color & ~ceq_mask) == 0);
328 
329 	if (((page_pfn_2_color_cpu(pfn, szc) ^ color) & ceq_mask) == 0) {
330 
331 		/* we start from the page with correct color */
332 		if (szc >= TTE512K) {
333 			if (szc >= TTE4M) {
334 				/* page color is PA[32:28] */
335 				pfn_ceq_mask = ceq_mask << 15;
336 			} else {
337 				/* page color is PA[32:28].PA[19:19] */
338 				pfn_ceq_mask = ((ceq_mask & 1) << 6) |
339 				    ((ceq_mask >> 1) << 15);
340 			}
341 			pfn = ADD_MASKED(pfn, pstep, pfn_ceq_mask, mask);
342 			return (pfn);
343 		} else {
344 			/*
345 			 * We deal 64K or 8K page. Check if we could the
346 			 * satisfy the request without changing PA[32:28]
347 			 */
348 			pfn_ceq_mask = ((ceq_mask & 3) << 5) | (ceq_mask >> 2);
349 			npfn = ADD_MASKED(pfn, pstep, pfn_ceq_mask, mask);
350 
351 			if ((((npfn ^ pfn) >> 15) & 0x1f) == 0)
352 				return (npfn);
353 
354 			/*
355 			 * for next pfn we have to change bits PA[32:28]
356 			 * set PA[63:28] and PA[19:18] of the next pfn
357 			 */
358 			npfn = (pfn >> 15) << 15;
359 			npfn |= (ceq_mask & color & 3) << 5;
360 			pfn_ceq_mask = (szc == TTE8K) ? 0 :
361 			    (ceq_mask & 0x1c) << 13;
362 			npfn = ADD_MASKED(npfn, (1 << 15), pfn_ceq_mask, mask);
363 
364 			/*
365 			 * set bits PA[17:13] to match the color
366 			 */
367 			ceq_mask >>= 2;
368 			color = (color >> 2) & ceq_mask;
369 			npfn |= ((npfn >> 15) ^ color) & ceq_mask;
370 			return (npfn);
371 		}
372 	}
373 
374 	/*
375 	 * we start from the page with incorrect color - rare case
376 	 */
377 	if (szc >= TTE512K) {
378 		if (szc >= TTE4M) {
379 			/* page color is in bits PA[32:28] */
380 			npfn = ((pfn >> 20) << 20) | (color << 15);
381 			pfn_ceq_mask = (ceq_mask << 15) | 0x7fff;
382 		} else {
383 			/* try get the right color by changing bit PA[19:19] */
384 			npfn = pfn + pstep;
385 			if (((page_pfn_2_color_cpu(npfn, szc) ^ color) &
386 			    ceq_mask) == 0)
387 				return (npfn);
388 
389 			/* page color is PA[32:28].PA[19:19] */
390 			pfn_ceq_mask = ((ceq_mask & 1) << 6) |
391 			    ((ceq_mask >> 1) << 15) | (0xff << 7);
392 			pfn_color = ((color & 1) << 6) | ((color >> 1) << 15);
393 			npfn = ((pfn >> 20) << 20) | pfn_color;
394 		}
395 
396 		while (npfn <= pfn) {
397 			npfn = ADD_MASKED(npfn, pstep, pfn_ceq_mask, mask);
398 		}
399 		return (npfn);
400 	}
401 
402 	/*
403 	 * We deal 64K or 8K page of incorrect color.
404 	 * Try correcting color without changing PA[32:28]
405 	 */
406 
407 	pfn_ceq_mask = ((ceq_mask & 3) << 5) | (ceq_mask >> 2);
408 	pfn_color = ((color & 3) << 5) | (color >> 2);
409 	npfn = (pfn & ~(pfn_t)0x7f);
410 	npfn |= (((pfn >> 15) & 0x1f) ^ pfn_color) & pfn_ceq_mask;
411 	npfn = (szc == TTE64K) ? (npfn & ~(pfn_t)0x7) : npfn;
412 
413 	if (((page_pfn_2_color_cpu(npfn, szc) ^ color) & ceq_mask) == 0) {
414 
415 		/* the color is fixed - find the next page */
416 		while (npfn <= pfn) {
417 			npfn = ADD_MASKED(npfn, pstep, pfn_ceq_mask, mask);
418 		}
419 		if ((((npfn ^ pfn) >> 15) & 0x1f) == 0)
420 			return (npfn);
421 	}
422 
423 	/* to fix the color need to touch PA[32:28] */
424 	npfn = (szc == TTE8K) ? ((pfn >> 15) << 15) :
425 	    (((pfn >> 18) << 18) | ((color & 0x1c) << 13));
426 	tmpmask = (szc == TTE8K) ? 0 : (ceq_mask & 0x1c) << 13;
427 
428 	while (npfn <= pfn) {
429 		npfn = ADD_MASKED(npfn, (1 << 15), tmpmask, mask);
430 	}
431 
432 	/* set bits PA[19:13] to match the color */
433 	npfn |= (((npfn >> 15) & 0x1f) ^ pfn_color) & pfn_ceq_mask;
434 	npfn = (szc == TTE64K) ? (npfn & ~(pfn_t)0x7) : npfn;
435 
436 	ASSERT(((page_pfn_2_color_cpu(npfn, szc) ^ color) & ceq_mask) == 0);
437 
438 	return (npfn);
439 }
440 
441 /*
442  * init page coloring
443  */
444 void
445 page_coloring_init_cpu()
446 {
447 	int i;
448 
449 	hw_page_array[0].hp_colors = 1 << 7;
450 	hw_page_array[1].hp_colors = 1 << 7;
451 	hw_page_array[2].hp_colors = 1 << 6;
452 
453 	for (i = 3; i < mmu_page_sizes; i++) {
454 		hw_page_array[i].hp_colors = 1 << 5;
455 	}
456 }
457 
458 /*
459  * group colorequiv colors on N2 by low order bits of the color first
460  */
461 void
462 page_set_colorequiv_arr_cpu(void)
463 {
464 	static uint_t nequiv_shades_log2[MMU_PAGE_SIZES] = {2, 5, 0, 0, 0, 0};
465 
466 	if (colorequiv > 1) {
467 		int i;
468 		uint_t sv_a = lowbit(colorequiv) - 1;
469 
470 		if (sv_a > 15)
471 			sv_a = 15;
472 
473 		for (i = 0; i < MMU_PAGE_SIZES; i++) {
474 			uint_t colors;
475 			uint_t a = sv_a;
476 
477 			if ((colors = hw_page_array[i].hp_colors) <= 1)
478 				continue;
479 			while ((colors >> a) == 0)
480 				a--;
481 			if (a > (colorequivszc[i] & 0xf) +
482 			    (colorequivszc[i] >> 4)) {
483 				if (a <= nequiv_shades_log2[i]) {
484 					colorequivszc[i] = a;
485 				} else {
486 					colorequivszc[i] =
487 					    ((a - nequiv_shades_log2[i]) << 4) |
488 					    nequiv_shades_log2[i];
489 				}
490 			}
491 		}
492 	}
493 }
494