xref: /illumos-gate/usr/src/uts/sun4v/cpu/niagara2.c (revision 102033aa92edf302ad31b3bdd7c6fcd2d6910903)
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  * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 #include <sys/types.h>
29 #include <sys/systm.h>
30 #include <sys/archsystm.h>
31 #include <sys/machparam.h>
32 #include <sys/machsystm.h>
33 #include <sys/cpu.h>
34 #include <sys/elf_SPARC.h>
35 #include <vm/hat_sfmmu.h>
36 #include <vm/page.h>
37 #include <vm/vm_dep.h>
38 #include <sys/cpuvar.h>
39 #include <sys/async.h>
40 #include <sys/cmn_err.h>
41 #include <sys/debug.h>
42 #include <sys/dditypes.h>
43 #include <sys/sunddi.h>
44 #include <sys/cpu_module.h>
45 #include <sys/prom_debug.h>
46 #include <sys/vmsystm.h>
47 #include <sys/prom_plat.h>
48 #include <sys/sysmacros.h>
49 #include <sys/intreg.h>
50 #include <sys/machtrap.h>
51 #include <sys/ontrap.h>
52 #include <sys/ivintr.h>
53 #include <sys/atomic.h>
54 #include <sys/panic.h>
55 #include <sys/dtrace.h>
56 #include <sys/simulate.h>
57 #include <sys/fault.h>
58 #include <sys/niagara2regs.h>
59 #include <sys/hsvc.h>
60 #include <sys/trapstat.h>
61 
62 uint_t root_phys_addr_lo_mask = 0xffffffffU;
63 char cpu_module_name[] = "SUNW,UltraSPARC-T2";
64 
65 /*
66  * Hypervisor services information for the NIAGARA2 CPU module
67  */
68 static boolean_t niagara2_hsvc_available = B_TRUE;
69 static uint64_t niagara2_sup_minor;		/* Supported minor number */
70 static hsvc_info_t niagara2_hsvc = {
71 	HSVC_REV_1, NULL, HSVC_GROUP_NIAGARA2_CPU, NIAGARA2_HSVC_MAJOR,
72 	NIAGARA2_HSVC_MINOR, cpu_module_name
73 };
74 
75 void
76 cpu_setup(void)
77 {
78 	extern int mmu_exported_pagesize_mask;
79 	extern int cpc_has_overflow_intr;
80 	int status;
81 
82 	/*
83 	 * Negotiate the API version for Niagara2 specific hypervisor
84 	 * services.
85 	 */
86 	status = hsvc_register(&niagara2_hsvc, &niagara2_sup_minor);
87 	if (status != 0) {
88 		cmn_err(CE_WARN, "%s: cannot negotiate hypervisor services "
89 		    "group: 0x%lx major: 0x%lx minor: 0x%lx errno: %d",
90 		    niagara2_hsvc.hsvc_modname, niagara2_hsvc.hsvc_group,
91 		    niagara2_hsvc.hsvc_major, niagara2_hsvc.hsvc_minor, status);
92 		niagara2_hsvc_available = B_FALSE;
93 	}
94 
95 	/*
96 	 * The setup common to all CPU modules is done in cpu_setup_common
97 	 * routine.
98 	 */
99 	cpu_setup_common(NULL);
100 
101 	cache |= (CACHE_PTAG | CACHE_IOCOHERENT);
102 
103 	if ((mmu_exported_pagesize_mask &
104 	    DEFAULT_SUN4V_MMU_PAGESIZE_MASK) !=
105 	    DEFAULT_SUN4V_MMU_PAGESIZE_MASK)
106 		cmn_err(CE_PANIC, "machine description"
107 		    " does not have required sun4v page sizes"
108 		    " 8K, 64K and 4M: MD mask is 0x%x",
109 		    mmu_exported_pagesize_mask);
110 
111 	cpu_hwcap_flags = AV_SPARC_VIS | AV_SPARC_VIS2 | AV_SPARC_ASI_BLK_INIT;
112 
113 	/*
114 	 * Niagara2 supports a 48-bit subset of the full 64-bit virtual
115 	 * address space. Virtual addresses between 0x0000800000000000
116 	 * and 0xffff.7fff.ffff.ffff inclusive lie within a "VA Hole"
117 	 * and must never be mapped. In addition, software must not use
118 	 * pages within 4GB of the VA hole as instruction pages to
119 	 * avoid problems with prefetching into the VA hole.
120 	 */
121 	hole_start = (caddr_t)((1ull << (va_bits - 1)) - (1ull << 32));
122 	hole_end = (caddr_t)((0ull - (1ull << (va_bits - 1))) + (1ull << 32));
123 
124 	/*
125 	 * Niagara2 has a performance counter overflow interrupt
126 	 */
127 	cpc_has_overflow_intr = 1;
128 
129 	/*
130 	 * Enable 4M pages for OOB.
131 	 */
132 	max_uheap_lpsize = MMU_PAGESIZE4M;
133 	max_ustack_lpsize = MMU_PAGESIZE4M;
134 	max_privmap_lpsize = MMU_PAGESIZE4M;
135 }
136 
137 /*
138  * Set the magic constants of the implementation.
139  */
140 void
141 cpu_fiximp(struct cpu_node *cpunode)
142 {
143 	/*
144 	 * The Cache node is optional in MD. Therefore in case "Cache"
145 	 * node does not exists in MD, set the default L2 cache associativity,
146 	 * size, linesize.
147 	 */
148 	if (cpunode->ecache_size == 0)
149 		cpunode->ecache_size = L2CACHE_SIZE;
150 	if (cpunode->ecache_linesize == 0)
151 		cpunode->ecache_linesize = L2CACHE_LINESIZE;
152 	if (cpunode->ecache_associativity == 0)
153 		cpunode->ecache_associativity = L2CACHE_ASSOCIATIVITY;
154 }
155 
156 static int niagara2_cpucnt;
157 
158 void
159 cpu_init_private(struct cpu *cp)
160 {
161 	extern int niagara_kstat_init(void);
162 
163 	/*
164 	 * The cpu_ipipe field is initialized based on the execution
165 	 * unit sharing information from the MD. It defaults to the
166 	 * virtual CPU id in the absence of such information.
167 	 */
168 	cp->cpu_m.cpu_ipipe = cpunodes[cp->cpu_id].exec_unit_mapping;
169 	if (cp->cpu_m.cpu_ipipe == NO_EU_MAPPING_FOUND)
170 		cp->cpu_m.cpu_ipipe = (id_t)(cp->cpu_id);
171 
172 	ASSERT(MUTEX_HELD(&cpu_lock));
173 	if ((niagara2_cpucnt++ == 0) && (niagara2_hsvc_available == B_TRUE))
174 		(void) niagara_kstat_init();
175 }
176 
177 /*ARGSUSED*/
178 void
179 cpu_uninit_private(struct cpu *cp)
180 {
181 	extern int niagara_kstat_fini(void);
182 
183 	ASSERT(MUTEX_HELD(&cpu_lock));
184 	if ((--niagara2_cpucnt == 0) && (niagara2_hsvc_available == B_TRUE))
185 		(void) niagara_kstat_fini();
186 }
187 
188 /*
189  * On Niagara2, any flush will cause all preceding stores to be
190  * synchronized wrt the i$, regardless of address or ASI.  In fact,
191  * the address is ignored, so we always flush address 0.
192  */
193 /*ARGSUSED*/
194 void
195 dtrace_flush_sec(uintptr_t addr)
196 {
197 	doflush(0);
198 }
199 
200 /*
201  * Trapstat support for Niagara2 processor
202  * The Niagara2 provides HWTW support for TSB lookup and with HWTW
203  * enabled no TSB hit information will be available. Therefore setting
204  * the time spent in TLB miss handler for TSB hits to 0.
205  */
206 int
207 cpu_trapstat_conf(int cmd)
208 {
209 	int status = 0;
210 
211 	switch (cmd) {
212 	case CPU_TSTATCONF_INIT:
213 	case CPU_TSTATCONF_FINI:
214 	case CPU_TSTATCONF_ENABLE:
215 	case CPU_TSTATCONF_DISABLE:
216 		break;
217 	default:
218 		status = EINVAL;
219 		break;
220 	}
221 	return (status);
222 }
223 
224 void
225 cpu_trapstat_data(void *buf, uint_t tstat_pgszs)
226 {
227 	tstat_pgszdata_t	*tstatp = (tstat_pgszdata_t *)buf;
228 	int	i;
229 
230 	for (i = 0; i < tstat_pgszs; i++, tstatp++) {
231 		tstatp->tpgsz_kernel.tmode_itlb.ttlb_tlb.tmiss_count = 0;
232 		tstatp->tpgsz_kernel.tmode_itlb.ttlb_tlb.tmiss_time = 0;
233 		tstatp->tpgsz_user.tmode_itlb.ttlb_tlb.tmiss_count = 0;
234 		tstatp->tpgsz_user.tmode_itlb.ttlb_tlb.tmiss_time = 0;
235 		tstatp->tpgsz_kernel.tmode_dtlb.ttlb_tlb.tmiss_count = 0;
236 		tstatp->tpgsz_kernel.tmode_dtlb.ttlb_tlb.tmiss_time = 0;
237 		tstatp->tpgsz_user.tmode_dtlb.ttlb_tlb.tmiss_count = 0;
238 		tstatp->tpgsz_user.tmode_dtlb.ttlb_tlb.tmiss_time = 0;
239 	}
240 }
241 
242 /* NI2 L2$ index is pa[32:28]^pa[17:13].pa[19:18]^pa[12:11].pa[10:6] */
243 uint_t
244 page_pfn_2_color_cpu(pfn_t pfn, uchar_t szc)
245 {
246 	uint_t color;
247 
248 	ASSERT(szc <= TTE256M);
249 
250 	pfn = PFN_BASE(pfn, szc);
251 	color = ((pfn >> 15) ^ pfn) & 0x1f;
252 	if (szc >= TTE4M)
253 		return (color);
254 
255 	color = (color << 2) | ((pfn >> 5) & 0x3);
256 
257 	return (szc <= TTE64K ? color : (color >> 1));
258 }
259 
260 #if TTE256M != 5
261 #error TTE256M is not 5
262 #endif
263 
264 uint_t
265 page_get_nsz_color_mask_cpu(uchar_t szc, uint_t mask)
266 {
267 	static uint_t ni2_color_masks[5] = {0x63, 0x1e, 0x3e, 0x1f, 0x1f};
268 	ASSERT(szc < TTE256M);
269 
270 	mask &= ni2_color_masks[szc];
271 	return ((szc == TTE64K || szc == TTE512K) ? (mask >> 1) : mask);
272 }
273 
274 uint_t
275 page_get_nsz_color_cpu(uchar_t szc, uint_t color)
276 {
277 	ASSERT(szc < TTE256M);
278 	return ((szc == TTE64K || szc == TTE512K) ? (color >> 1) : color);
279 }
280 
281 uint_t
282 page_get_color_shift_cpu(uchar_t szc, uchar_t nszc)
283 {
284 	ASSERT(nszc > szc);
285 	ASSERT(nszc <= TTE256M);
286 
287 	if (szc <= TTE64K)
288 		return ((nszc >= TTE4M) ? 2 : ((nszc >= TTE512K) ? 1 : 0));
289 	if (szc == TTE512K)
290 		return (1);
291 
292 	return (0);
293 }
294 
295 /*ARGSUSED*/
296 pfn_t
297 page_next_pfn_for_color_cpu(pfn_t pfn, uchar_t szc, uint_t color,
298     uint_t ceq_mask, uint_t color_mask)
299 {
300 	pfn_t pstep = PNUM_SIZE(szc);
301 	pfn_t npfn, pfn_ceq_mask, pfn_color;
302 	pfn_t tmpmask, mask = (pfn_t)-1;
303 
304 	ASSERT((color & ~ceq_mask) == 0);
305 
306 	if (((page_pfn_2_color_cpu(pfn, szc) ^ color) & ceq_mask) == 0) {
307 
308 		/* we start from the page with correct color */
309 		if (szc >= TTE512K) {
310 			if (szc >= TTE4M) {
311 				/* page color is PA[32:28] */
312 				pfn_ceq_mask = ceq_mask << 15;
313 			} else {
314 				/* page color is PA[32:28].PA[19:19] */
315 				pfn_ceq_mask = ((ceq_mask & 1) << 6) |
316 				    ((ceq_mask >> 1) << 15);
317 			}
318 			pfn = ADD_MASKED(pfn, pstep, pfn_ceq_mask, mask);
319 			return (pfn);
320 		} else {
321 			/*
322 			 * We deal 64K or 8K page. Check if we could the
323 			 * satisfy the request without changing PA[32:28]
324 			 */
325 			pfn_ceq_mask = ((ceq_mask & 3) << 5) | (ceq_mask >> 2);
326 			npfn = ADD_MASKED(pfn, pstep, pfn_ceq_mask, mask);
327 
328 			if ((((npfn ^ pfn) >> 15) & 0x1f) == 0)
329 				return (npfn);
330 
331 			/*
332 			 * for next pfn we have to change bits PA[32:28]
333 			 * set PA[63:28] and PA[19:18] of the next pfn
334 			 */
335 			npfn = (pfn >> 15) << 15;
336 			npfn |= (ceq_mask & color & 3) << 5;
337 			pfn_ceq_mask = (szc == TTE8K) ? 0 :
338 			    (ceq_mask & 0x1c) << 13;
339 			npfn = ADD_MASKED(npfn, (1 << 15), pfn_ceq_mask, mask);
340 
341 			/*
342 			 * set bits PA[17:13] to match the color
343 			 */
344 			ceq_mask >>= 2;
345 			color = (color >> 2) & ceq_mask;
346 			npfn |= ((npfn >> 15) ^ color) & ceq_mask;
347 			return (npfn);
348 		}
349 	}
350 
351 	/*
352 	 * we start from the page with incorrect color - rare case
353 	 */
354 	if (szc >= TTE512K) {
355 		if (szc >= TTE4M) {
356 			/* page color is in bits PA[32:28] */
357 			npfn = ((pfn >> 20) << 20) | (color << 15);
358 			pfn_ceq_mask = (ceq_mask << 15) | 0x7fff;
359 		} else {
360 			/* try get the right color by changing bit PA[19:19] */
361 			npfn = pfn + pstep;
362 			if (((page_pfn_2_color_cpu(npfn, szc) ^ color) &
363 			    ceq_mask) == 0)
364 				return (npfn);
365 
366 			/* page color is PA[32:28].PA[19:19] */
367 			pfn_ceq_mask = ((ceq_mask & 1) << 6) |
368 			    ((ceq_mask >> 1) << 15) | (0xff << 7);
369 			pfn_color = ((color & 1) << 6) | ((color >> 1) << 15);
370 			npfn = ((pfn >> 20) << 20) | pfn_color;
371 		}
372 
373 		while (npfn <= pfn) {
374 			npfn = ADD_MASKED(npfn, pstep, pfn_ceq_mask, mask);
375 		}
376 		return (npfn);
377 	}
378 
379 	/*
380 	 * We deal 64K or 8K page of incorrect color.
381 	 * Try correcting color without changing PA[32:28]
382 	 */
383 
384 	pfn_ceq_mask = ((ceq_mask & 3) << 5) | (ceq_mask >> 2);
385 	pfn_color = ((color & 3) << 5) | (color >> 2);
386 	npfn = (pfn & ~(pfn_t)0x7f);
387 	npfn |= (((pfn >> 15) & 0x1f) ^ pfn_color) & pfn_ceq_mask;
388 	npfn = (szc == TTE64K) ? (npfn & ~(pfn_t)0x7) : npfn;
389 
390 	if (((page_pfn_2_color_cpu(npfn, szc) ^ color) & ceq_mask) == 0) {
391 
392 		/* the color is fixed - find the next page */
393 		while (npfn <= pfn) {
394 			npfn = ADD_MASKED(npfn, pstep, pfn_ceq_mask, mask);
395 		}
396 		if ((((npfn ^ pfn) >> 15) & 0x1f) == 0)
397 			return (npfn);
398 	}
399 
400 	/* to fix the color need to touch PA[32:28] */
401 	npfn = (szc == TTE8K) ? ((pfn >> 15) << 15) :
402 	    (((pfn >> 18) << 18) | ((color & 0x1c) << 13));
403 	tmpmask = (szc == TTE8K) ? 0 : (ceq_mask & 0x1c) << 13;
404 
405 	while (npfn <= pfn) {
406 		npfn = ADD_MASKED(npfn, (1 << 15), tmpmask, mask);
407 	}
408 
409 	/* set bits PA[19:13] to match the color */
410 	npfn |= (((npfn >> 15) & 0x1f) ^ pfn_color) & pfn_ceq_mask;
411 	npfn = (szc == TTE64K) ? (npfn & ~(pfn_t)0x7) : npfn;
412 
413 	ASSERT(((page_pfn_2_color_cpu(npfn, szc) ^ color) & ceq_mask) == 0);
414 
415 	return (npfn);
416 }
417 
418 /*
419  * init page coloring
420  */
421 void
422 page_coloring_init_cpu()
423 {
424 	int i;
425 	uint_t colors;
426 
427 	hw_page_array[0].hp_colors = 1 << 7;
428 	hw_page_array[1].hp_colors = 1 << 7;
429 	hw_page_array[2].hp_colors = 1 << 6;
430 
431 	for (i = 3; i < mmu_page_sizes; i++) {
432 		hw_page_array[i].hp_colors = 1 << 5;
433 	}
434 
435 	if (colorequiv > 1) {
436 		int a = lowbit(colorequiv) - 1;
437 
438 		if (a > 15)
439 			a = 15;
440 
441 		for (i = 0; i < mmu_page_sizes; i++) {
442 			if ((colors = hw_page_array[i].hp_colors) <= 1) {
443 				continue;
444 			}
445 			while ((colors >> a) == 0)
446 				a--;
447 			colorequivszc[i] = (a << 4);
448 		}
449 	}
450 }
451