xref: /titanic_50/usr/src/uts/sun4u/cpu/us3_cheetahplus_asm.s (revision 45462bf898cce4257293567d9170f6cf79d0ea1d)
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 2008 Sun Microsystems, Inc.  All rights reserved.
23 * Use is subject to license terms.
24 *
25 * Assembly code support for the Cheetah+ module
26 */
27
28#pragma ident	"%Z%%M%	%I%	%E% SMI"
29
30#if !defined(lint)
31#include "assym.h"
32#endif	/* lint */
33
34#include <sys/asm_linkage.h>
35#include <sys/mmu.h>
36#include <vm/hat_sfmmu.h>
37#include <sys/machparam.h>
38#include <sys/machcpuvar.h>
39#include <sys/machthread.h>
40#include <sys/machtrap.h>
41#include <sys/privregs.h>
42#include <sys/asm_linkage.h>
43#include <sys/trap.h>
44#include <sys/cheetahregs.h>
45#include <sys/us3_module.h>
46#include <sys/xc_impl.h>
47#include <sys/intreg.h>
48#include <sys/async.h>
49#include <sys/clock.h>
50#include <sys/cheetahasm.h>
51#include <sys/cmpregs.h>
52
53#ifdef TRAPTRACE
54#include <sys/traptrace.h>
55#endif /* TRAPTRACE */
56
57
58#if !defined(lint)
59
60/* BEGIN CSTYLED */
61
62/*
63 * Cheetah+ version to reflush an Ecache line by index.
64 *
65 * By default we assume the Ecache is 2-way so we flush both
66 * ways. Even if the cache is direct-mapped no harm will come
67 * from performing the flush twice, apart from perhaps a performance
68 * penalty.
69 *
70 * XXX - scr2 not used.
71 */
72#define	ECACHE_REFLUSH_LINE(ec_set_size, index, scr2)			\
73	ldxa	[index]ASI_EC_DIAG, %g0;				\
74	ldxa	[index + ec_set_size]ASI_EC_DIAG, %g0;
75
76/*
77 * Cheetah+ version of ecache_flush_line.  Uses Cheetah+ Ecache Displacement
78 * Flush feature.
79 */
80#define	ECACHE_FLUSH_LINE(physaddr, ec_set_size, scr1, scr2)		\
81	sub	ec_set_size, 1, scr1;					\
82	and	physaddr, scr1, scr1;					\
83	set	CHP_ECACHE_IDX_DISP_FLUSH, scr2;			\
84	or	scr2, scr1, scr1;					\
85	ECACHE_REFLUSH_LINE(ec_set_size, scr1, scr2)
86
87/* END CSTYLED */
88
89/*
90 * Panther version to reflush a line from both the L2 cache and L3
91 * cache by the respective indexes. Flushes all ways of the line from
92 * each cache.
93 *
94 * l2_index	Index into the L2$ of the line to be flushed. This
95 *		register will not be modified by this routine.
96 * l3_index	Index into the L3$ of the line to be flushed. This
97 *		register will not be modified by this routine.
98 * scr2		scratch register.
99 * scr3		scratch register.
100 *
101 */
102#define	PN_ECACHE_REFLUSH_LINE(l2_index, l3_index, scr2, scr3)		\
103	set	PN_L2_MAX_SET, scr2;					\
104	set	PN_L2_SET_SIZE, scr3;					\
1051:									\
106	ldxa	[l2_index + scr2]ASI_L2_TAG, %g0;			\
107	cmp	scr2, %g0;						\
108	bg,a	1b;							\
109	  sub	scr2, scr3, scr2;					\
110	mov	6, scr2;						\
1117:									\
112	cmp	scr2, %g0;						\
113	bg,a	7b;							\
114	  sub	scr2, 1, scr2;						\
115	set	PN_L3_MAX_SET, scr2;					\
116	set	PN_L3_SET_SIZE, scr3;					\
1172:									\
118	ldxa	[l3_index + scr2]ASI_EC_DIAG, %g0;			\
119	cmp	scr2, %g0;						\
120	bg,a	2b;							\
121	  sub	scr2, scr3, scr2;
122
123/*
124 * Panther version of ecache_flush_line. Flushes the line corresponding
125 * to physaddr from both the L2 cache and the L3 cache.
126 *
127 * physaddr	Input: Physical address to flush.
128 *              Output: Physical address to flush (preserved).
129 * l2_idx_out	Input: scratch register.
130 *              Output: Index into the L2$ of the line to be flushed.
131 * l3_idx_out	Input: scratch register.
132 *              Output: Index into the L3$ of the line to be flushed.
133 * scr3		scratch register.
134 * scr4		scratch register.
135 *
136 */
137#define	PN_ECACHE_FLUSH_LINE(physaddr, l2_idx_out, l3_idx_out, scr3, scr4)	\
138	set	PN_L3_SET_SIZE, l2_idx_out;					\
139	sub	l2_idx_out, 1, l2_idx_out;					\
140	and	physaddr, l2_idx_out, l3_idx_out;				\
141	set	PN_L3_IDX_DISP_FLUSH, l2_idx_out;				\
142	or	l2_idx_out, l3_idx_out, l3_idx_out;				\
143	set	PN_L2_SET_SIZE, l2_idx_out;					\
144	sub	l2_idx_out, 1, l2_idx_out;					\
145	and	physaddr, l2_idx_out, l2_idx_out;				\
146	set	PN_L2_IDX_DISP_FLUSH, scr3;					\
147	or	l2_idx_out, scr3, l2_idx_out;					\
148	PN_ECACHE_REFLUSH_LINE(l2_idx_out, l3_idx_out, scr3, scr4)
149
150#endif	/* !lint */
151
152/*
153 * Fast ECC error at TL>0 handler
154 * We get here via trap 70 at TL>0->Software trap 0 at TL>0.  We enter
155 * this routine with %g1 and %g2 already saved in %tpc, %tnpc and %tstate.
156 * For a complete description of the Fast ECC at TL>0 handling see the
157 * comment block "Cheetah/Cheetah+ Fast ECC at TL>0 trap strategy" in
158 * us3_common_asm.s
159 */
160#if defined(lint)
161
162void
163fast_ecc_tl1_err(void)
164{}
165
166#else	/* lint */
167
168	.section ".text"
169	.align	64
170	ENTRY_NP(fast_ecc_tl1_err)
171
172	/*
173	 * This macro turns off the D$/I$ if they are on and saves their
174	 * original state in ch_err_tl1_tmp, saves all the %g registers in the
175	 * ch_err_tl1_data structure, updates the ch_err_tl1_flags and saves
176	 * the %tpc in ch_err_tl1_tpc.  At the end of this macro, %g1 will
177	 * point to the ch_err_tl1_data structure and the original D$/I$ state
178	 * will be saved in ch_err_tl1_tmp.  All %g registers except for %g1
179	 * will be available.
180	 */
181	CH_ERR_TL1_FECC_ENTER;
182
183	/*
184	 * Get the diagnostic logout data.  %g4 must be initialized to
185	 * current CEEN state, %g5 must point to logout structure in
186	 * ch_err_tl1_data_t.  %g3 will contain the nesting count upon
187	 * return.
188	 */
189	ldxa	[%g0]ASI_ESTATE_ERR, %g4
190	and	%g4, EN_REG_CEEN, %g4
191	add	%g1, CH_ERR_TL1_LOGOUT, %g5
192	DO_TL1_CPU_LOGOUT(%g3, %g2, %g4, %g5, %g6, %g3, %g4)
193
194	/*
195	 * If the logout nesting count is exceeded, we're probably
196	 * not making any progress, try to panic instead.
197	 */
198	cmp	%g3, CLO_NESTING_MAX
199	bge	fecc_tl1_err
200	  nop
201
202	/*
203	 * Save the current CEEN and NCEEN state in %g7 and turn them off
204	 * before flushing the Ecache.
205	 */
206	ldxa	[%g0]ASI_ESTATE_ERR, %g7
207	andn	%g7, EN_REG_CEEN | EN_REG_NCEEN, %g5
208	stxa	%g5, [%g0]ASI_ESTATE_ERR
209	membar	#Sync
210
211	/*
212	 * Flush the Ecache, using the largest possible cache size with the
213	 * smallest possible line size since we can't get the actual sizes
214	 * from the cpu_node due to DTLB misses.
215	 */
216	PN_L2_FLUSHALL(%g3, %g4, %g5)
217
218	set	CH_ECACHE_MAX_SIZE, %g4
219	set	CH_ECACHE_MIN_LSIZE, %g5
220
221	GET_CPU_IMPL(%g6)
222	cmp	%g6, PANTHER_IMPL
223	bne	%xcc, 2f
224	  nop
225	set	PN_L3_SIZE, %g4
2262:
227	mov	%g6, %g3
228	CHP_ECACHE_FLUSHALL(%g4, %g5, %g3)
229
230	/*
231	 * Restore CEEN and NCEEN to the previous state.
232	 */
233	stxa	%g7, [%g0]ASI_ESTATE_ERR
234	membar	#Sync
235
236	/*
237	 * If we turned off the D$, then flush it and turn it back on.
238	 */
239	ldxa	[%g1 + CH_ERR_TL1_TMP]%asi, %g3
240	andcc	%g3, CH_ERR_TSTATE_DC_ON, %g0
241	bz	%xcc, 3f
242	  nop
243
244	/*
245	 * Flush the D$.
246	 */
247	ASM_LD(%g4, dcache_size)
248	ASM_LD(%g5, dcache_linesize)
249	CH_DCACHE_FLUSHALL(%g4, %g5, %g6)
250
251	/*
252	 * Turn the D$ back on.
253	 */
254	ldxa	[%g0]ASI_DCU, %g3
255	or	%g3, DCU_DC, %g3
256	stxa	%g3, [%g0]ASI_DCU
257	membar	#Sync
2583:
259	/*
260	 * If we turned off the I$, then flush it and turn it back on.
261	 */
262	ldxa	[%g1 + CH_ERR_TL1_TMP]%asi, %g3
263	andcc	%g3, CH_ERR_TSTATE_IC_ON, %g0
264	bz	%xcc, 4f
265	  nop
266
267	/*
268	 * Flush the I$.  Panther has different I$ parameters, and we
269	 * can't access the logout I$ params without possibly generating
270	 * a MMU miss.
271	 */
272	GET_CPU_IMPL(%g6)
273	set	PN_ICACHE_SIZE, %g3
274	set	CH_ICACHE_SIZE, %g4
275	mov	CH_ICACHE_LSIZE, %g5
276	cmp	%g6, PANTHER_IMPL
277	movz	%xcc, %g3, %g4
278	movz	%xcc, PN_ICACHE_LSIZE, %g5
279	CH_ICACHE_FLUSHALL(%g4, %g5, %g6, %g3)
280
281	/*
282	 * Turn the I$ back on.  Changing DCU_IC requires flush.
283	 */
284	ldxa	[%g0]ASI_DCU, %g3
285	or	%g3, DCU_IC, %g3
286	stxa	%g3, [%g0]ASI_DCU
287	flush	%g0
2884:
289
290#ifdef TRAPTRACE
291	/*
292	 * Get current trap trace entry physical pointer.
293	 */
294	CPU_INDEX(%g6, %g5)
295	sll	%g6, TRAPTR_SIZE_SHIFT, %g6
296	set	trap_trace_ctl, %g5
297	add	%g6, %g5, %g6
298	ld	[%g6 + TRAPTR_LIMIT], %g5
299	tst	%g5
300	be	%icc, skip_traptrace
301	  nop
302	ldx	[%g6 + TRAPTR_PBASE], %g5
303	ld	[%g6 + TRAPTR_OFFSET], %g4
304	add	%g5, %g4, %g5
305
306	/*
307	 * Create trap trace entry.
308	 */
309	rd	%asi, %g7
310	wr	%g0, TRAPTR_ASI, %asi
311	rd	STICK, %g4
312	stxa	%g4, [%g5 + TRAP_ENT_TICK]%asi
313	rdpr	%tl, %g4
314	stha	%g4, [%g5 + TRAP_ENT_TL]%asi
315	rdpr	%tt, %g4
316	stha	%g4, [%g5 + TRAP_ENT_TT]%asi
317	rdpr	%tpc, %g4
318	stna	%g4, [%g5 + TRAP_ENT_TPC]%asi
319	rdpr	%tstate, %g4
320	stxa	%g4, [%g5 + TRAP_ENT_TSTATE]%asi
321	stna	%sp, [%g5 + TRAP_ENT_SP]%asi
322	stna	%g0, [%g5 + TRAP_ENT_TR]%asi
323	wr	%g0, %g7, %asi
324	ldxa	[%g1 + CH_ERR_TL1_SDW_AFAR]%asi, %g3
325	ldxa	[%g1 + CH_ERR_TL1_SDW_AFSR]%asi, %g4
326	wr	%g0, TRAPTR_ASI, %asi
327	stna	%g3, [%g5 + TRAP_ENT_F1]%asi
328	stna	%g4, [%g5 + TRAP_ENT_F2]%asi
329	wr	%g0, %g7, %asi
330	ldxa	[%g1 + CH_ERR_TL1_AFAR]%asi, %g3
331	ldxa	[%g1 + CH_ERR_TL1_AFSR]%asi, %g4
332	wr	%g0, TRAPTR_ASI, %asi
333	stna	%g3, [%g5 + TRAP_ENT_F3]%asi
334	stna	%g4, [%g5 + TRAP_ENT_F4]%asi
335	wr	%g0, %g7, %asi
336
337	/*
338	 * Advance trap trace pointer.
339	 */
340	ld	[%g6 + TRAPTR_OFFSET], %g5
341	ld	[%g6 + TRAPTR_LIMIT], %g4
342	st	%g5, [%g6 + TRAPTR_LAST_OFFSET]
343	add	%g5, TRAP_ENT_SIZE, %g5
344	sub	%g4, TRAP_ENT_SIZE, %g4
345	cmp	%g5, %g4
346	movge	%icc, 0, %g5
347	st	%g5, [%g6 + TRAPTR_OFFSET]
348skip_traptrace:
349#endif	/* TRAPTRACE */
350
351	/*
352	 * If nesting count is not zero, skip all the AFSR/AFAR
353	 * handling and just do the necessary cache-flushing.
354	 */
355	ldxa	[%g1 + CH_ERR_TL1_NEST_CNT]%asi, %g2
356	brnz	%g2, 6f
357	  nop
358
359	/*
360	 * If a UCU or L3_UCU followed by a WDU has occurred go ahead
361	 * and panic since a UE will occur (on the retry) before the
362	 * UCU and WDU messages are enqueued.
363	 */
364	ldxa	[%g1 + CH_ERR_TL1_SDW_AFSR]%asi, %g3
365	set	1, %g4
366	sllx	%g4, C_AFSR_UCU_SHIFT, %g4
367	btst	%g4, %g3		! UCU in original shadow AFSR?
368	bnz	%xcc, 5f
369	  mov	1, %g4
370	ldxa	[%g1 + CH_ERR_TL1_SDW_AFSR_EXT]%asi, %g3
371	sllx	%g4, C_AFSR_L3_UCU_SHIFT, %g4
372	btst	%g4, %g3		! L3_UCU in original shadow AFSR_EXT?
373	bz	%xcc, 6f
374	  nop
3755:
376	ldxa	[%g1 + CH_ERR_TL1_AFSR]%asi, %g4	! original AFSR
377	ldxa	[%g0]ASI_AFSR, %g3	! current AFSR
378	or	%g3, %g4, %g3		! %g3 = original + current AFSR
379	set	1, %g4
380	sllx	%g4, C_AFSR_WDU_SHIFT, %g4
381	btst	%g4, %g3		! WDU in original or current AFSR?
382	bnz	%xcc, fecc_tl1_err
383	  nop
384
3856:
386	/*
387	 * We fall into this macro if we've successfully logged the error in
388	 * the ch_err_tl1_data structure and want the PIL15 softint to pick
389	 * it up and log it.  %g1 must point to the ch_err_tl1_data structure.
390	 * Restores the %g registers and issues retry.
391	 */
392	CH_ERR_TL1_EXIT;
393
394	/*
395	 * Establish panic exit label.
396	 */
397	CH_ERR_TL1_PANIC_EXIT(fecc_tl1_err);
398
399	SET_SIZE(fast_ecc_tl1_err)
400
401#endif	/* lint */
402
403
404#if defined(lint)
405/*
406 * scrubphys - Pass in the aligned physical memory address
407 * that you want to scrub, along with the ecache set size.
408 *
409 *	1) Displacement flush the E$ line corresponding to %addr.
410 *	   The first ldxa guarantees that the %addr is no longer in
411 *	   M, O, or E (goes to I or S (if instruction fetch also happens).
412 *	2) "Write" the data using a CAS %addr,%g0,%g0.
413 *	   The casxa guarantees a transition from I to M or S to M.
414 *	3) Displacement flush the E$ line corresponding to %addr.
415 *	   The second ldxa pushes the M line out of the ecache, into the
416 *	   writeback buffers, on the way to memory.
417 *	4) The "membar #Sync" pushes the cache line out of the writeback
418 *	   buffers onto the bus, on the way to dram finally.
419 *
420 * This is a modified version of the algorithm suggested by Gary Lauterbach.
421 * In theory the CAS %addr,%g0,%g0 is supposed to mark the addr's cache line
422 * as modified, but then we found out that for spitfire, if it misses in the
423 * E$ it will probably install as an M, but if it hits in the E$, then it
424 * will stay E, if the store doesn't happen. So the first displacement flush
425 * should ensure that the CAS will miss in the E$.  Arrgh.
426 */
427/* ARGSUSED */
428void
429scrubphys(uint64_t paddr, int ecache_set_size)
430{}
431
432#else	/* lint */
433	ENTRY(scrubphys)
434	rdpr	%pstate, %o4
435	andn	%o4, PSTATE_IE | PSTATE_AM, %o5
436	wrpr	%o5, %g0, %pstate	! clear IE, AM bits
437
438	GET_CPU_IMPL(%o5)		! Panther Ecache is flushed differently
439	cmp	%o5, PANTHER_IMPL
440	bne	scrubphys_1
441	  nop
442	PN_ECACHE_FLUSH_LINE(%o0, %o1, %o2, %o3, %o5)
443	casxa	[%o0]ASI_MEM, %g0, %g0
444	PN_ECACHE_REFLUSH_LINE(%o1, %o2, %o3, %o0)
445	b	scrubphys_2
446	  nop
447scrubphys_1:
448	ECACHE_FLUSH_LINE(%o0, %o1, %o2, %o3)
449	casxa	[%o0]ASI_MEM, %g0, %g0
450	ECACHE_REFLUSH_LINE(%o1, %o2, %o3)
451scrubphys_2:
452	wrpr	%g0, %o4, %pstate	! restore earlier pstate register value
453
454	retl
455	membar	#Sync			! move the data out of the load buffer
456	SET_SIZE(scrubphys)
457
458#endif	/* lint */
459
460
461#if defined(lint)
462/*
463 * clearphys - Pass in the physical memory address of the checkblock
464 * that you want to push out, cleared with a recognizable pattern,
465 * from the ecache.
466 *
467 * To ensure that the ecc gets recalculated after the bad data is cleared,
468 * we must write out enough data to fill the w$ line (64 bytes). So we read
469 * in an entire ecache subblock's worth of data, and write it back out.
470 * Then we overwrite the 16 bytes of bad data with the pattern.
471 */
472/* ARGSUSED */
473void
474clearphys(uint64_t paddr, int ecache_set_size, int ecache_linesize)
475{
476}
477
478#else	/* lint */
479	ENTRY(clearphys)
480	/* turn off IE, AM bits */
481	rdpr	%pstate, %o4
482	andn	%o4, PSTATE_IE | PSTATE_AM, %o5
483	wrpr	%o5, %g0, %pstate
484
485	/* turn off NCEEN */
486	ldxa	[%g0]ASI_ESTATE_ERR, %o5
487	andn	%o5, EN_REG_NCEEN, %o3
488	stxa	%o3, [%g0]ASI_ESTATE_ERR
489	membar	#Sync
490
491	/* align address passed with 64 bytes subblock size */
492	mov	CH_ECACHE_SUBBLK_SIZE, %o2
493	andn	%o0, (CH_ECACHE_SUBBLK_SIZE - 1), %g1
494
495	/* move the good data into the W$ */
496clearphys_1:
497	subcc	%o2, 8, %o2
498	ldxa	[%g1 + %o2]ASI_MEM, %g2
499	bge	clearphys_1
500	  stxa	%g2, [%g1 + %o2]ASI_MEM
501
502	/* now overwrite the bad data */
503	setx	0xbadecc00badecc01, %g1, %g2
504	stxa	%g2, [%o0]ASI_MEM
505	mov	8, %g1
506	stxa	%g2, [%o0 + %g1]ASI_MEM
507
508	GET_CPU_IMPL(%o3)		! Panther Ecache is flushed differently
509	cmp	%o3, PANTHER_IMPL
510	bne	clearphys_2
511	  nop
512	PN_ECACHE_FLUSH_LINE(%o0, %o1, %o2, %o3, %g1)
513	casxa	[%o0]ASI_MEM, %g0, %g0
514	PN_ECACHE_REFLUSH_LINE(%o1, %o2, %o3, %o0)
515	b	clearphys_3
516	  nop
517clearphys_2:
518	ECACHE_FLUSH_LINE(%o0, %o1, %o2, %o3)
519	casxa	[%o0]ASI_MEM, %g0, %g0
520	ECACHE_REFLUSH_LINE(%o1, %o2, %o3)
521clearphys_3:
522	/* clear the AFSR */
523	ldxa	[%g0]ASI_AFSR, %o1
524	stxa	%o1, [%g0]ASI_AFSR
525	membar	#Sync
526
527	/* turn NCEEN back on */
528	stxa	%o5, [%g0]ASI_ESTATE_ERR
529	membar	#Sync
530
531	/* return and re-enable IE and AM */
532	retl
533	  wrpr	%g0, %o4, %pstate
534	SET_SIZE(clearphys)
535
536#endif	/* lint */
537
538
539#if defined(lint)
540/*
541 * Cheetah+ Ecache displacement flush the specified line from the E$
542 *
543 * For Panther, this means flushing the specified line from both the
544 * L2 cache and L3 cache.
545 *
546 * Register usage:
547 *	%o0 - 64 bit physical address for flushing
548 *	%o1 - Ecache set size
549 */
550/*ARGSUSED*/
551void
552ecache_flush_line(uint64_t flushaddr, int ec_set_size)
553{
554}
555#else	/* lint */
556	ENTRY(ecache_flush_line)
557
558	GET_CPU_IMPL(%o3)		! Panther Ecache is flushed differently
559	cmp	%o3, PANTHER_IMPL
560	bne	ecache_flush_line_1
561	  nop
562
563	PN_ECACHE_FLUSH_LINE(%o0, %o1, %o2, %o3, %o4)
564	b	ecache_flush_line_2
565	  nop
566ecache_flush_line_1:
567	ECACHE_FLUSH_LINE(%o0, %o1, %o2, %o3)
568ecache_flush_line_2:
569	retl
570	  nop
571	SET_SIZE(ecache_flush_line)
572#endif	/* lint */
573
574#if defined(lint)
575void
576set_afsr_ext(uint64_t afsr_ext)
577{
578	afsr_ext = afsr_ext;
579}
580#else /* lint */
581
582	ENTRY(set_afsr_ext)
583	set	ASI_AFSR_EXT_VA, %o1
584	stxa	%o0, [%o1]ASI_AFSR		! afsr_ext reg
585	membar	#Sync
586	retl
587	nop
588	SET_SIZE(set_afsr_ext)
589
590#endif /* lint */
591
592
593#if defined(lint)
594/*
595 * The CPU jumps here from the MMU exception handler if an ITLB parity
596 * error is detected and we are running on Panther.
597 *
598 * In this routine we collect diagnostic information and write it to our
599 * logout structure (if possible) and clear all ITLB entries that may have
600 * caused our parity trap.
601 * Then we call cpu_tlb_parity_error via systrap in order to drop down to TL0
602 * and log any error messages. As for parameters to cpu_tlb_parity_error, we
603 * send two:
604 *
605 * %g2	- Contains the VA whose lookup in the ITLB caused the parity error
606 * %g3	- Contains the tlo_info field of the pn_tlb_logout logout struct,
607 *	  regardless of whether or not we actually used the logout struct.
608 *
609 * In the TL0 handler (cpu_tlb_parity_error) we will compare those two
610 * parameters to the data contained in the logout structure in order to
611 * determine whether the logout information is valid for this particular
612 * error or not.
613 */
614void
615itlb_parity_trap(void)
616{}
617
618#else	/* lint */
619
620	ENTRY_NP(itlb_parity_trap)
621	/*
622	 * Collect important information about the trap which will be
623	 * used as a parameter to the TL0 handler.
624	 */
625	wr	%g0, ASI_IMMU, %asi
626	rdpr	%tpc, %g2			! VA that caused the IMMU trap
627	ldxa	[MMU_TAG_ACCESS_EXT]%asi, %g3	! read the trap VA page size
628	set	PN_ITLB_PGSZ_MASK, %g4
629	and	%g3, %g4, %g3
630	ldxa	[MMU_TAG_ACCESS]%asi, %g4
631	set	TAGREAD_CTX_MASK, %g5
632	and	%g4, %g5, %g4
633	or	%g4, %g3, %g3			! 'or' in the trap context and
634	mov	1, %g4				! add the IMMU flag to complete
635	sllx	%g4, PN_TLO_INFO_IMMU_SHIFT, %g4
636	or	%g4, %g3, %g3			! the tlo_info field for logout
637	stxa	%g0,[MMU_SFSR]%asi		! clear the SFSR
638	membar	#Sync
639
640	/*
641	 * at this point:
642	 *    %g2 - contains the VA whose lookup caused the trap
643	 *    %g3 - contains the tlo_info field
644	 *
645	 * Next, we calculate the TLB index value for the failing VA.
646	 */
647	mov	%g2, %g4			! We need the ITLB index
648	set	PN_ITLB_PGSZ_MASK, %g5
649	and	%g3, %g5, %g5
650	srlx	%g5, PN_ITLB_PGSZ_SHIFT, %g5
651	PN_GET_TLB_INDEX(%g4, %g5)		! %g4 has the index
652	sllx	%g4, PN_TLB_ACC_IDX_SHIFT, %g4	! shift the index into place
653	set	PN_ITLB_T512, %g5
654	or	%g4, %g5, %g4			! and add in the TLB ID
655
656	/*
657	 * at this point:
658	 *    %g2 - contains the VA whose lookup caused the trap
659	 *    %g3 - contains the tlo_info field
660	 *    %g4 - contains the TLB access index value for the
661	 *          VA/PgSz in question
662	 *
663	 * Check to see if the logout structure is available.
664	 */
665	set	CHPR_TLB_LOGOUT, %g6
666	GET_CPU_PRIVATE_PTR(%g6, %g1, %g5, itlb_parity_trap_1)
667	set	LOGOUT_INVALID_U32, %g6
668	sllx	%g6, 32, %g6			! if our logout structure is
669	set	LOGOUT_INVALID_L32, %g5		! unavailable or if it is
670	or	%g5, %g6, %g5			! already being used, then we
671	ldx	[%g1 + PN_TLO_ADDR], %g6	! don't collect any diagnostic
672	cmp	%g6, %g5			! information before clearing
673	bne	itlb_parity_trap_1		! and logging the error.
674	  nop
675
676	/*
677	 * Record the logout information. %g4 contains our index + TLB ID
678	 * for use in ASI_ITLB_ACCESS and ASI_ITLB_TAGREAD. %g1 contains
679	 * the pointer to our logout struct.
680	 */
681	stx	%g3, [%g1 + PN_TLO_INFO]
682	stx	%g2, [%g1 + PN_TLO_ADDR]
683	stx	%g2, [%g1 + PN_TLO_PC]		! %tpc == fault addr for IMMU
684
685	add	%g1, PN_TLO_ITLB_TTE, %g1	! move up the pointer
686
687	ldxa	[%g4]ASI_ITLB_ACCESS, %g5	! read the data
688	stx	%g5, [%g1 + CH_TLO_TTE_DATA]	! store it away
689	ldxa	[%g4]ASI_ITLB_TAGREAD, %g5	! read the tag
690	stx	%g5, [%g1 + CH_TLO_TTE_TAG]	! store it away
691
692	set	PN_TLB_ACC_WAY_BIT, %g6		! same thing again for way 1
693	or	%g4, %g6, %g4
694	add	%g1, CH_TLO_TTE_SIZE, %g1	! move up the pointer
695
696	ldxa	[%g4]ASI_ITLB_ACCESS, %g5	! read the data
697	stx	%g5, [%g1 + CH_TLO_TTE_DATA]	! store it away
698	ldxa	[%g4]ASI_ITLB_TAGREAD, %g5	! read the tag
699	stx	%g5, [%g1 + CH_TLO_TTE_TAG]	! store it away
700
701	andn	%g4, %g6, %g4			! back to way 0
702
703itlb_parity_trap_1:
704	/*
705	 * at this point:
706	 *    %g2 - contains the VA whose lookup caused the trap
707	 *    %g3 - contains the tlo_info field
708	 *    %g4 - contains the TLB access index value for the
709	 *          VA/PgSz in question
710	 *
711	 * Here we will clear the errors from the TLB.
712	 */
713	set	MMU_TAG_ACCESS, %g5		! We write a TTE tag value of
714	stxa	%g0, [%g5]ASI_IMMU		! 0 as it will be invalid.
715	stxa	%g0, [%g4]ASI_ITLB_ACCESS	! Write the data and tag
716	membar	#Sync
717
718	set	PN_TLB_ACC_WAY_BIT, %g6		! same thing again for way 1
719	or	%g4, %g6, %g4
720
721	stxa	%g0, [%g4]ASI_ITLB_ACCESS	! Write same data and tag
722	membar	#Sync
723
724	sethi	%hi(FLUSH_ADDR), %g6		! PRM says we need to issue a
725	flush   %g6				! flush after writing MMU regs
726
727	/*
728	 * at this point:
729	 *    %g2 - contains the VA whose lookup caused the trap
730	 *    %g3 - contains the tlo_info field
731	 *
732	 * Call cpu_tlb_parity_error via systrap at PIL 14 unless we're
733	 * already at PIL 15.	 */
734	set	cpu_tlb_parity_error, %g1
735	rdpr	%pil, %g4
736	cmp	%g4, PIL_14
737	movl	%icc, PIL_14, %g4
738	ba	sys_trap
739	  nop
740	SET_SIZE(itlb_parity_trap)
741
742#endif	/* lint */
743
744#if defined(lint)
745/*
746 * The CPU jumps here from the MMU exception handler if a DTLB parity
747 * error is detected and we are running on Panther.
748 *
749 * In this routine we collect diagnostic information and write it to our
750 * logout structure (if possible) and clear all DTLB entries that may have
751 * caused our parity trap.
752 * Then we call cpu_tlb_parity_error via systrap in order to drop down to TL0
753 * and log any error messages. As for parameters to cpu_tlb_parity_error, we
754 * send two:
755 *
756 * %g2	- Contains the VA whose lookup in the DTLB caused the parity error
757 * %g3	- Contains the tlo_info field of the pn_tlb_logout logout struct,
758 *	  regardless of whether or not we actually used the logout struct.
759 *
760 * In the TL0 handler (cpu_tlb_parity_error) we will compare those two
761 * parameters to the data contained in the logout structure in order to
762 * determine whether the logout information is valid for this particular
763 * error or not.
764 */
765void
766dtlb_parity_trap(void)
767{}
768
769#else	/* lint */
770
771	ENTRY_NP(dtlb_parity_trap)
772	/*
773	 * Collect important information about the trap which will be
774	 * used as a parameter to the TL0 handler.
775	 */
776	wr	%g0, ASI_DMMU, %asi
777	ldxa	[MMU_SFAR]%asi, %g2		! VA that caused the IMMU trap
778	ldxa	[MMU_TAG_ACCESS_EXT]%asi, %g3	! read the trap VA page sizes
779	set	PN_DTLB_PGSZ_MASK, %g4
780	and	%g3, %g4, %g3
781	ldxa	[MMU_TAG_ACCESS]%asi, %g4
782	set	TAGREAD_CTX_MASK, %g5		! 'or' in the trap context
783	and	%g4, %g5, %g4			! to complete the tlo_info
784	or	%g4, %g3, %g3			! field for logout
785	stxa	%g0,[MMU_SFSR]%asi		! clear the SFSR
786	membar	#Sync
787
788	/*
789	 * at this point:
790	 *    %g2 - contains the VA whose lookup caused the trap
791	 *    %g3 - contains the tlo_info field
792	 *
793	 * Calculate the TLB index values for the failing VA. Since the T512
794	 * TLBs can be configured for different page sizes, we need to find
795	 * the index into each one separately.
796	 */
797	mov	%g2, %g4			! First we get the DTLB_0 index
798	set	PN_DTLB_PGSZ0_MASK, %g5
799	and	%g3, %g5, %g5
800	srlx	%g5, PN_DTLB_PGSZ0_SHIFT, %g5
801	PN_GET_TLB_INDEX(%g4, %g5)		! %g4 has the DTLB_0 index
802	sllx	%g4, PN_TLB_ACC_IDX_SHIFT, %g4	! shift the index into place
803	set	PN_DTLB_T512_0, %g5
804	or	%g4, %g5, %g4			! and add in the TLB ID
805
806	mov	%g2, %g7			! Next we get the DTLB_1 index
807	set	PN_DTLB_PGSZ1_MASK, %g5
808	and	%g3, %g5, %g5
809	srlx	%g5, PN_DTLB_PGSZ1_SHIFT, %g5
810	PN_GET_TLB_INDEX(%g7, %g5)		! %g7 has the DTLB_1 index
811	sllx	%g7, PN_TLB_ACC_IDX_SHIFT, %g7	! shift the index into place
812	set	PN_DTLB_T512_1, %g5
813	or	%g7, %g5, %g7			! and add in the TLB ID
814
815	/*
816	 * at this point:
817	 *    %g2 - contains the VA whose lookup caused the trap
818	 *    %g3 - contains the tlo_info field
819	 *    %g4 - contains the T512_0 access index value for the
820	 *          VA/PgSz in question
821	 *    %g7 - contains the T512_1 access index value for the
822	 *          VA/PgSz in question
823	 *
824	 * If this trap happened at TL>0, then we don't want to mess
825	 * with the normal logout struct since that could caused a TLB
826	 * miss.
827	 */
828	rdpr	%tl, %g6			! read current trap level
829	cmp	%g6, 1				! skip over the tl>1 code
830	ble	dtlb_parity_trap_1		! if TL <= 1.
831	  nop
832
833	/*
834	 * If we are here, then the trap happened at TL>1. Simply
835	 * update our tlo_info field and then skip to the TLB flush
836	 * code.
837	 */
838	mov	1, %g6
839	sllx	%g6, PN_TLO_INFO_TL1_SHIFT, %g6
840	or	%g6, %g3, %g3
841	ba	dtlb_parity_trap_2
842	  nop
843
844dtlb_parity_trap_1:
845	/*
846	 * at this point:
847	 *    %g2 - contains the VA whose lookup caused the trap
848	 *    %g3 - contains the tlo_info field
849	 *    %g4 - contains the T512_0 access index value for the
850	 *          VA/PgSz in question
851	 *    %g7 - contains the T512_1 access index value for the
852	 *          VA/PgSz in question
853	 *
854	 * Check to see if the logout structure is available.
855	 */
856	set	CHPR_TLB_LOGOUT, %g6
857	GET_CPU_PRIVATE_PTR(%g6, %g1, %g5, dtlb_parity_trap_2)
858	set	LOGOUT_INVALID_U32, %g6
859	sllx	%g6, 32, %g6			! if our logout structure is
860	set	LOGOUT_INVALID_L32, %g5		! unavailable or if it is
861	or	%g5, %g6, %g5			! already being used, then we
862	ldx	[%g1 + PN_TLO_ADDR], %g6	! don't collect any diagnostic
863	cmp	%g6, %g5			! information before clearing
864	bne	dtlb_parity_trap_2		! and logging the error.
865	  nop
866
867	/*
868	 * Record the logout information. %g4 contains our DTLB_0
869	 * index + TLB ID and %g7 contains our DTLB_1 index + TLB ID
870	 * both of which will be used for ASI_DTLB_ACCESS and
871	 * ASI_DTLB_TAGREAD. %g1 contains the pointer to our logout
872	 * struct.
873	 */
874	stx	%g3, [%g1 + PN_TLO_INFO]
875	stx	%g2, [%g1 + PN_TLO_ADDR]
876	rdpr	%tpc, %g5
877	stx	%g5, [%g1 + PN_TLO_PC]
878
879	add	%g1, PN_TLO_DTLB_TTE, %g1	! move up the pointer
880
881	ldxa	[%g4]ASI_DTLB_ACCESS, %g5	! read the data from DTLB_0
882	stx	%g5, [%g1 + CH_TLO_TTE_DATA]	! way 0 and store it away
883	ldxa	[%g4]ASI_DTLB_TAGREAD, %g5	! read the tag from DTLB_0
884	stx	%g5, [%g1 + CH_TLO_TTE_TAG]	! way 0 and store it away
885
886	ldxa	[%g7]ASI_DTLB_ACCESS, %g5	! now repeat for DTLB_1 way 0
887	stx	%g5, [%g1 + (CH_TLO_TTE_DATA + (CH_TLO_TTE_SIZE * 2))]
888	ldxa	[%g7]ASI_DTLB_TAGREAD, %g5
889	stx	%g5, [%g1 + (CH_TLO_TTE_TAG + (CH_TLO_TTE_SIZE * 2))]
890
891	set	PN_TLB_ACC_WAY_BIT, %g6		! same thing again for way 1
892	or	%g4, %g6, %g4			! of each TLB.
893	or	%g7, %g6, %g7
894	add	%g1, CH_TLO_TTE_SIZE, %g1	! move up the pointer
895
896	ldxa	[%g4]ASI_DTLB_ACCESS, %g5	! read the data from DTLB_0
897	stx	%g5, [%g1 + CH_TLO_TTE_DATA]	! way 1 and store it away
898	ldxa	[%g4]ASI_DTLB_TAGREAD, %g5	! read the tag from DTLB_0
899	stx	%g5, [%g1 + CH_TLO_TTE_TAG]	! way 1 and store it away
900
901	ldxa	[%g7]ASI_DTLB_ACCESS, %g5	! now repeat for DTLB_1 way 1
902	stx	%g5, [%g1 + (CH_TLO_TTE_DATA + (CH_TLO_TTE_SIZE * 2))]
903	ldxa	[%g7]ASI_DTLB_TAGREAD, %g5
904	stx	%g5, [%g1 + (CH_TLO_TTE_TAG + (CH_TLO_TTE_SIZE * 2))]
905
906	andn	%g4, %g6, %g4			! back to way 0
907	andn	%g7, %g6, %g7			! back to way 0
908
909dtlb_parity_trap_2:
910	/*
911	 * at this point:
912	 *    %g2 - contains the VA whose lookup caused the trap
913	 *    %g3 - contains the tlo_info field
914	 *    %g4 - contains the T512_0 access index value for the
915	 *          VA/PgSz in question
916	 *    %g7 - contains the T512_1 access index value for the
917	 *          VA/PgSz in question
918	 *
919	 * Here we will clear the errors from the DTLB.
920	 */
921	set	MMU_TAG_ACCESS, %g5		! We write a TTE tag value of
922	stxa	%g0, [%g5]ASI_DMMU		! 0 as it will be invalid.
923	stxa	%g0, [%g4]ASI_DTLB_ACCESS	! Write the data and tag.
924	stxa	%g0, [%g7]ASI_DTLB_ACCESS	! Now repeat for DTLB_1 way 0
925	membar	#Sync
926
927	set	PN_TLB_ACC_WAY_BIT, %g6		! same thing again for way 1
928	or	%g4, %g6, %g4
929	or	%g7, %g6, %g7
930
931	stxa	%g0, [%g4]ASI_DTLB_ACCESS	! Write same data and tag.
932	stxa	%g0, [%g7]ASI_DTLB_ACCESS	! Now repeat for DTLB_1 way 0
933	membar	#Sync
934
935	sethi	%hi(FLUSH_ADDR), %g6		! PRM says we need to issue a
936	flush   %g6				! flush after writing MMU regs
937
938	/*
939	 * at this point:
940	 *    %g2 - contains the VA whose lookup caused the trap
941	 *    %g3 - contains the tlo_info field
942	 *
943	 * Call cpu_tlb_parity_error via systrap at PIL 14 unless we're
944	 * already at PIL 15. We do this even for TL>1 traps since
945	 * those will lead to a system panic.
946	 */
947	set	cpu_tlb_parity_error, %g1
948	rdpr	%pil, %g4
949	cmp	%g4, PIL_14
950	movl	%icc, PIL_14, %g4
951	ba	sys_trap
952	  nop
953	SET_SIZE(dtlb_parity_trap)
954
955#endif	/* lint */
956
957
958#if defined(lint)
959/*
960 * Calculates the Panther TLB index based on a virtual address and page size
961 *
962 * Register usage:
963 *	%o0 - virtual address whose index we want
964 *	%o1 - Page Size of the TLB in question as encoded in the
965 *	      ASI_[D|I]MMU_TAG_ACCESS_EXT register.
966 */
967uint64_t
968pn_get_tlb_index(uint64_t va, uint64_t pg_sz)
969{
970	return ((va + pg_sz)-(va + pg_sz));
971}
972#else	/* lint */
973	ENTRY(pn_get_tlb_index)
974
975	PN_GET_TLB_INDEX(%o0, %o1)
976
977	retl
978	  nop
979	SET_SIZE(pn_get_tlb_index)
980#endif	/* lint */
981
982
983#if defined(lint)
984/*
985 * For Panther CPUs we need to flush the IPB after any I$ or D$
986 * parity errors are detected.
987 */
988void
989flush_ipb(void)
990{ return; }
991
992#else	/* lint */
993
994	ENTRY(flush_ipb)
995	clr	%o0
996
997flush_ipb_1:
998	stxa	%g0, [%o0]ASI_IPB_TAG
999	membar	#Sync
1000	cmp	%o0, PN_IPB_TAG_ADDR_MAX
1001	blt	flush_ipb_1
1002	  add	%o0, PN_IPB_TAG_ADDR_LINESIZE, 	%o0
1003
1004	sethi	%hi(FLUSH_ADDR), %o0
1005	flush   %o0
1006	retl
1007	nop
1008	SET_SIZE(flush_ipb)
1009
1010#endif	/* lint */
1011
1012
1013