[linux-block.git] / arch / sparc / include / asm / trap_block.h

/* SPDX-License-Identifier: GPL-2.0 */
#ifndef _SPARC_TRAP_BLOCK_H
#define _SPARC_TRAP_BLOCK_H

#include <asm/hypervisor.h>
#include <asm/asi.h>

#ifndef __ASSEMBLY__

/* Trap handling code needs to get at a few critical values upon
 * trap entry and to process TSB misses.  These cannot be in the
 * per_cpu() area as we really need to lock them into the TLB and
 * thus make them part of the main kernel image.  As a result we
 * try to make this as small as possible.
 *
 * This is padded out and aligned to 64-bytes to avoid false sharing
 * on SMP.
 */

/* If you modify the size of this structure, please update
 * TRAP_BLOCK_SZ_SHIFT below.
 */
struct thread_info;
struct trap_per_cpu {
/* D-cache line 1: Basic thread information, cpu and device mondo queues */
	struct thread_info	*thread;
	unsigned long		pgd_paddr;
	unsigned long		cpu_mondo_pa;
	unsigned long		dev_mondo_pa;

/* D-cache line 2: Error Mondo Queue and kernel buffer pointers */
	unsigned long		resum_mondo_pa;
	unsigned long		resum_kernel_buf_pa;
	unsigned long		nonresum_mondo_pa;
	unsigned long		nonresum_kernel_buf_pa;

/* Dcache lines 3, 4, 5, and 6: Hypervisor Fault Status */
	struct hv_fault_status	fault_info;

/* Dcache line 7: Physical addresses of CPU send mondo block and CPU list.  */
	unsigned long		cpu_mondo_block_pa;
	unsigned long		cpu_list_pa;
	unsigned long		tsb_huge;
	unsigned long		tsb_huge_temp;

/* Dcache line 8: IRQ work list, and keep trap_block a power-of-2 in size.  */
	unsigned long		irq_worklist_pa;
	unsigned int		cpu_mondo_qmask;
	unsigned int		dev_mondo_qmask;
	unsigned int		resum_qmask;
	unsigned int		nonresum_qmask;
	unsigned long		__per_cpu_base;
} __attribute__((aligned(64)));
extern struct trap_per_cpu trap_block[NR_CPUS];
void init_cur_cpu_trap(struct thread_info *);
void setup_tba(void);
extern int ncpus_probed;
extern u64 cpu_mondo_counter[NR_CPUS];

unsigned long real_hard_smp_processor_id(void);

struct cpuid_patch_entry {
	unsigned int	addr;
	unsigned int	cheetah_safari[4];
	unsigned int	cheetah_jbus[4];
	unsigned int	starfire[4];
	unsigned int	sun4v[4];
};
extern struct cpuid_patch_entry __cpuid_patch, __cpuid_patch_end;

struct sun4v_1insn_patch_entry {
	unsigned int	addr;
	unsigned int	insn;
};
extern struct sun4v_1insn_patch_entry __sun4v_1insn_patch,
	__sun4v_1insn_patch_end;
extern struct sun4v_1insn_patch_entry __fast_win_ctrl_1insn_patch,
	__fast_win_ctrl_1insn_patch_end;

struct sun4v_2insn_patch_entry {
	unsigned int	addr;
	unsigned int	insns[2];
};
extern struct sun4v_2insn_patch_entry __sun4v_2insn_patch,
	__sun4v_2insn_patch_end;
extern struct sun4v_2insn_patch_entry __sun_m7_2insn_patch,
	__sun_m7_2insn_patch_end;


#endif /* !(__ASSEMBLY__) */

#define TRAP_PER_CPU_THREAD		0x00
#define TRAP_PER_CPU_PGD_PADDR		0x08
#define TRAP_PER_CPU_CPU_MONDO_PA	0x10
#define TRAP_PER_CPU_DEV_MONDO_PA	0x18
#define TRAP_PER_CPU_RESUM_MONDO_PA	0x20
#define TRAP_PER_CPU_RESUM_KBUF_PA	0x28
#define TRAP_PER_CPU_NONRESUM_MONDO_PA	0x30
#define TRAP_PER_CPU_NONRESUM_KBUF_PA	0x38
#define TRAP_PER_CPU_FAULT_INFO		0x40
#define TRAP_PER_CPU_CPU_MONDO_BLOCK_PA	0xc0
#define TRAP_PER_CPU_CPU_LIST_PA	0xc8
#define TRAP_PER_CPU_TSB_HUGE		0xd0
#define TRAP_PER_CPU_TSB_HUGE_TEMP	0xd8
#define TRAP_PER_CPU_IRQ_WORKLIST_PA	0xe0
#define TRAP_PER_CPU_CPU_MONDO_QMASK	0xe8
#define TRAP_PER_CPU_DEV_MONDO_QMASK	0xec
#define TRAP_PER_CPU_RESUM_QMASK	0xf0
#define TRAP_PER_CPU_NONRESUM_QMASK	0xf4
#define TRAP_PER_CPU_PER_CPU_BASE	0xf8

#define TRAP_BLOCK_SZ_SHIFT		8

#include <asm/scratchpad.h>

#define __GET_CPUID(REG)				\
	/* Spitfire implementation (default). */	\
661:	ldxa		[%g0] ASI_UPA_CONFIG, REG;	\
	srlx		REG, 17, REG;			\
	 and		REG, 0x1f, REG;			\
	nop;						\
	.section	.cpuid_patch, "ax";		\
	/* Instruction location. */			\
	.word		661b;				\
	/* Cheetah Safari implementation. */		\
	ldxa		[%g0] ASI_SAFARI_CONFIG, REG;	\
	srlx		REG, 17, REG;			\
	and		REG, 0x3ff, REG;		\
	nop;						\
	/* Cheetah JBUS implementation. */		\
	ldxa		[%g0] ASI_JBUS_CONFIG, REG;	\
	srlx		REG, 17, REG;			\
	and		REG, 0x1f, REG;			\
	nop;						\
	/* Starfire implementation. */			\
	sethi		%hi(0x1fff40000d0 >> 9), REG;	\
	sllx		REG, 9, REG;			\
	or		REG, 0xd0, REG;			\
	lduwa		[REG] ASI_PHYS_BYPASS_EC_E, REG;\
	/* sun4v implementation. */			\
	mov		SCRATCHPAD_CPUID, REG;		\
	ldxa		[REG] ASI_SCRATCHPAD, REG;	\
	nop;						\
	nop;						\
	.previous;

#ifdef CONFIG_SMP

#define TRAP_LOAD_TRAP_BLOCK(DEST, TMP)		\
	__GET_CPUID(TMP)			\
	sethi	%hi(trap_block), DEST;		\
	sllx	TMP, TRAP_BLOCK_SZ_SHIFT, TMP;	\
	or	DEST, %lo(trap_block), DEST;	\
	add	DEST, TMP, DEST;		\

/* Clobbers TMP, current address space PGD phys address into DEST.  */
#define TRAP_LOAD_PGD_PHYS(DEST, TMP)		\
	TRAP_LOAD_TRAP_BLOCK(DEST, TMP)		\
	ldx	[DEST + TRAP_PER_CPU_PGD_PADDR], DEST;

/* Clobbers TMP, loads local processor's IRQ work area into DEST.  */
#define TRAP_LOAD_IRQ_WORK_PA(DEST, TMP)	\
	TRAP_LOAD_TRAP_BLOCK(DEST, TMP)		\
	add	DEST, TRAP_PER_CPU_IRQ_WORKLIST_PA, DEST;

/* Clobbers TMP, loads DEST with current thread info pointer.  */
#define TRAP_LOAD_THREAD_REG(DEST, TMP)		\
	TRAP_LOAD_TRAP_BLOCK(DEST, TMP)		\
	ldx	[DEST + TRAP_PER_CPU_THREAD], DEST;

/* Given the current thread info pointer in THR, load the per-cpu
 * area base of the current processor into DEST.  REG1, REG2, and REG3 are
 * clobbered.
 *
 * You absolutely cannot use DEST as a temporary in this code.  The
 * reason is that traps can happen during execution, and return from
 * trap will load the fully resolved DEST per-cpu base.  This can corrupt
 * the calculations done by the macro mid-stream.
 */
#define LOAD_PER_CPU_BASE(DEST, THR, REG1, REG2, REG3)	\
	lduh	[THR + TI_CPU], REG1;			\
	sethi	%hi(trap_block), REG2;			\
	sllx	REG1, TRAP_BLOCK_SZ_SHIFT, REG1;	\
	or	REG2, %lo(trap_block), REG2;		\
	add	REG2, REG1, REG2;			\
	ldx	[REG2 + TRAP_PER_CPU_PER_CPU_BASE], DEST;

#else

#define TRAP_LOAD_TRAP_BLOCK(DEST, TMP)		\
	sethi	%hi(trap_block), DEST;		\
	or	DEST, %lo(trap_block), DEST;	\

/* Uniprocessor versions, we know the cpuid is zero.  */
#define TRAP_LOAD_PGD_PHYS(DEST, TMP)		\
	TRAP_LOAD_TRAP_BLOCK(DEST, TMP)		\
	ldx	[DEST + TRAP_PER_CPU_PGD_PADDR], DEST;

/* Clobbers TMP, loads local processor's IRQ work area into DEST.  */
#define TRAP_LOAD_IRQ_WORK_PA(DEST, TMP)	\
	TRAP_LOAD_TRAP_BLOCK(DEST, TMP)		\
	add	DEST, TRAP_PER_CPU_IRQ_WORKLIST_PA, DEST;

#define TRAP_LOAD_THREAD_REG(DEST, TMP)		\
	TRAP_LOAD_TRAP_BLOCK(DEST, TMP)		\
	ldx	[DEST + TRAP_PER_CPU_THREAD], DEST;

/* No per-cpu areas on uniprocessor, so no need to load DEST.  */
#define LOAD_PER_CPU_BASE(DEST, THR, REG1, REG2, REG3)

#endif /* !(CONFIG_SMP) */

#endif /* _SPARC_TRAP_BLOCK_H */
Commit	Line	Data
b2441318	1	/* SPDX-License-Identifier: GPL-2.0 */
19f0fa3f DM	2	#ifndef _SPARC_TRAP_BLOCK_H
	3	#define _SPARC_TRAP_BLOCK_H
	4
	5	#include <asm/hypervisor.h>
	6	#include <asm/asi.h>
	7
	8	#ifndef __ASSEMBLY__
	9
	10	/* Trap handling code needs to get at a few critical values upon
	11	* trap entry and to process TSB misses. These cannot be in the
	12	* per_cpu() area as we really need to lock them into the TLB and
	13	* thus make them part of the main kernel image. As a result we
	14	* try to make this as small as possible.
	15	*
	16	* This is padded out and aligned to 64-bytes to avoid false sharing
	17	* on SMP.
	18	*/
	19
	20	/* If you modify the size of this structure, please update
	21	* TRAP_BLOCK_SZ_SHIFT below.
	22	*/
	23	struct thread_info;
	24	struct trap_per_cpu {
	25	/* D-cache line 1: Basic thread information, cpu and device mondo queues */
	26	struct thread_info *thread;
	27	unsigned long pgd_paddr;
	28	unsigned long cpu_mondo_pa;
	29	unsigned long dev_mondo_pa;
	30
	31	/* D-cache line 2: Error Mondo Queue and kernel buffer pointers */
	32	unsigned long resum_mondo_pa;
	33	unsigned long resum_kernel_buf_pa;
	34	unsigned long nonresum_mondo_pa;
	35	unsigned long nonresum_kernel_buf_pa;
	36
	37	/* Dcache lines 3, 4, 5, and 6: Hypervisor Fault Status */
	38	struct hv_fault_status fault_info;
	39
	40	/* Dcache line 7: Physical addresses of CPU send mondo block and CPU list. */
	41	unsigned long cpu_mondo_block_pa;
	42	unsigned long cpu_list_pa;
	43	unsigned long tsb_huge;
	44	unsigned long tsb_huge_temp;
	45
	46	/* Dcache line 8: IRQ work list, and keep trap_block a power-of-2 in size. */
	47	unsigned long irq_worklist_pa;
	48	unsigned int cpu_mondo_qmask;
	49	unsigned int dev_mondo_qmask;
	50	unsigned int resum_qmask;
	51	unsigned int nonresum_qmask;
5a5488d3	52	unsigned long __per_cpu_base;
19f0fa3f DM	53	} __attribute__((aligned(64)));
19f0fa3f DM	54	extern struct trap_per_cpu trap_block[NR_CPUS];
f05a6865 SR	55	void init_cur_cpu_trap(struct thread_info *);
f05a6865 SR	56	void setup_tba(void);
19f0fa3f	57	extern int ncpus_probed;
9d53caec	58	extern u64 cpu_mondo_counter[NR_CPUS];
19f0fa3f	59
f05a6865	60	unsigned long real_hard_smp_processor_id(void);
19f0fa3f DM	61
	62	struct cpuid_patch_entry {
	63	unsigned int addr;
	64	unsigned int cheetah_safari[4];
	65	unsigned int cheetah_jbus[4];
	66	unsigned int starfire[4];
	67	unsigned int sun4v[4];
	68	};
	69	extern struct cpuid_patch_entry __cpuid_patch, __cpuid_patch_end;
	70
	71	struct sun4v_1insn_patch_entry {
	72	unsigned int addr;
	73	unsigned int insn;
	74	};
	75	extern struct sun4v_1insn_patch_entry __sun4v_1insn_patch,
	76	__sun4v_1insn_patch_end;
a7159a87 AY	77	extern struct sun4v_1insn_patch_entry __fast_win_ctrl_1insn_patch,
a7159a87 AY	78	__fast_win_ctrl_1insn_patch_end;
19f0fa3f DM	79
	80	struct sun4v_2insn_patch_entry {
	81	unsigned int addr;
	82	unsigned int insns[2];
	83	};
	84	extern struct sun4v_2insn_patch_entry __sun4v_2insn_patch,
	85	__sun4v_2insn_patch_end;
494e5b6f KA	86	extern struct sun4v_2insn_patch_entry __sun_m7_2insn_patch,
494e5b6f KA	87	__sun_m7_2insn_patch_end;
19f0fa3f DM	88
	89
	90	#endif /* !(__ASSEMBLY__) */
	91
	92	#define TRAP_PER_CPU_THREAD 0x00
	93	#define TRAP_PER_CPU_PGD_PADDR 0x08
	94	#define TRAP_PER_CPU_CPU_MONDO_PA 0x10
	95	#define TRAP_PER_CPU_DEV_MONDO_PA 0x18
	96	#define TRAP_PER_CPU_RESUM_MONDO_PA 0x20
	97	#define TRAP_PER_CPU_RESUM_KBUF_PA 0x28
	98	#define TRAP_PER_CPU_NONRESUM_MONDO_PA 0x30
	99	#define TRAP_PER_CPU_NONRESUM_KBUF_PA 0x38
	100	#define TRAP_PER_CPU_FAULT_INFO 0x40
	101	#define TRAP_PER_CPU_CPU_MONDO_BLOCK_PA 0xc0
	102	#define TRAP_PER_CPU_CPU_LIST_PA 0xc8
	103	#define TRAP_PER_CPU_TSB_HUGE 0xd0
	104	#define TRAP_PER_CPU_TSB_HUGE_TEMP 0xd8
	105	#define TRAP_PER_CPU_IRQ_WORKLIST_PA 0xe0
	106	#define TRAP_PER_CPU_CPU_MONDO_QMASK 0xe8
	107	#define TRAP_PER_CPU_DEV_MONDO_QMASK 0xec
	108	#define TRAP_PER_CPU_RESUM_QMASK 0xf0
	109	#define TRAP_PER_CPU_NONRESUM_QMASK 0xf4
5a5488d3	110	#define TRAP_PER_CPU_PER_CPU_BASE 0xf8
19f0fa3f DM	111
	112	#define TRAP_BLOCK_SZ_SHIFT 8
	113
	114	#include <asm/scratchpad.h>
	115
	116	#define __GET_CPUID(REG) \
	117	/* Spitfire implementation (default). */ \
	118	661: ldxa [%g0] ASI_UPA_CONFIG, REG; \
	119	srlx REG, 17, REG; \
	120	and REG, 0x1f, REG; \
	121	nop; \
	122	.section .cpuid_patch, "ax"; \
	123	/* Instruction location. */ \
	124	.word 661b; \
	125	/* Cheetah Safari implementation. */ \
	126	ldxa [%g0] ASI_SAFARI_CONFIG, REG; \
	127	srlx REG, 17, REG; \
	128	and REG, 0x3ff, REG; \
	129	nop; \
	130	/* Cheetah JBUS implementation. */ \
	131	ldxa [%g0] ASI_JBUS_CONFIG, REG; \
	132	srlx REG, 17, REG; \
	133	and REG, 0x1f, REG; \
	134	nop; \
	135	/* Starfire implementation. */ \
	136	sethi %hi(0x1fff40000d0 >> 9), REG; \
	137	sllx REG, 9, REG; \
	138	or REG, 0xd0, REG; \
	139	lduwa [REG] ASI_PHYS_BYPASS_EC_E, REG;\
	140	/* sun4v implementation. */ \
	141	mov SCRATCHPAD_CPUID, REG; \
	142	ldxa [REG] ASI_SCRATCHPAD, REG; \
	143	nop; \
	144	nop; \
	145	.previous;
	146
	147	#ifdef CONFIG_SMP
	148
	149	#define TRAP_LOAD_TRAP_BLOCK(DEST, TMP) \
	150	__GET_CPUID(TMP) \
	151	sethi %hi(trap_block), DEST; \
	152	sllx TMP, TRAP_BLOCK_SZ_SHIFT, TMP; \
	153	or DEST, %lo(trap_block), DEST; \
	154	add DEST, TMP, DEST; \
	155
	156	/* Clobbers TMP, current address space PGD phys address into DEST. */
	157	#define TRAP_LOAD_PGD_PHYS(DEST, TMP) \
	158	TRAP_LOAD_TRAP_BLOCK(DEST, TMP) \
	159	ldx [DEST + TRAP_PER_CPU_PGD_PADDR], DEST;
	160
	161	/* Clobbers TMP, loads local processor's IRQ work area into DEST. */
	162	#define TRAP_LOAD_IRQ_WORK_PA(DEST, TMP) \
	163	TRAP_LOAD_TRAP_BLOCK(DEST, TMP) \
	164	add DEST, TRAP_PER_CPU_IRQ_WORKLIST_PA, DEST;
	165
	166	/* Clobbers TMP, loads DEST with current thread info pointer. */
	167	#define TRAP_LOAD_THREAD_REG(DEST, TMP) \
	168	TRAP_LOAD_TRAP_BLOCK(DEST, TMP) \
	169	ldx [DEST + TRAP_PER_CPU_THREAD], DEST;
	170
	171	/* Given the current thread info pointer in THR, load the per-cpu
	172	* area base of the current processor into DEST. REG1, REG2, and REG3 are
	173	* clobbered.
	174	*
175	* You absolutely cannot use DEST as a temporary in this code. The
176	* reason is that traps can happen during execution, and return from
177	* trap will load the fully resolved DEST per-cpu base. This can corrupt
178	* the calculations done by the macro mid-stream.
179	*/
180	#define LOAD_PER_CPU_BASE(DEST, THR, REG1, REG2, REG3) \
181	lduh [THR + TI_CPU], REG1; \
5a5488d3 DM	182	sethi %hi(trap_block), REG2; \
	183	sllx REG1, TRAP_BLOCK_SZ_SHIFT, REG1; \
	184	or REG2, %lo(trap_block), REG2; \
	185	add REG2, REG1, REG2; \
	186	ldx [REG2 + TRAP_PER_CPU_PER_CPU_BASE], DEST;
19f0fa3f DM	187
	188	#else
	189
	190	#define TRAP_LOAD_TRAP_BLOCK(DEST, TMP) \
	191	sethi %hi(trap_block), DEST; \
	192	or DEST, %lo(trap_block), DEST; \
	193
	194	/* Uniprocessor versions, we know the cpuid is zero. */
	195	#define TRAP_LOAD_PGD_PHYS(DEST, TMP) \
	196	TRAP_LOAD_TRAP_BLOCK(DEST, TMP) \
	197	ldx [DEST + TRAP_PER_CPU_PGD_PADDR], DEST;
	198
	199	/* Clobbers TMP, loads local processor's IRQ work area into DEST. */
	200	#define TRAP_LOAD_IRQ_WORK_PA(DEST, TMP) \
	201	TRAP_LOAD_TRAP_BLOCK(DEST, TMP) \
	202	add DEST, TRAP_PER_CPU_IRQ_WORKLIST_PA, DEST;
	203
	204	#define TRAP_LOAD_THREAD_REG(DEST, TMP) \
	205	TRAP_LOAD_TRAP_BLOCK(DEST, TMP) \
	206	ldx [DEST + TRAP_PER_CPU_THREAD], DEST;
	207
	208	/* No per-cpu areas on uniprocessor, so no need to load DEST. */
	209	#define LOAD_PER_CPU_BASE(DEST, THR, REG1, REG2, REG3)
	210
	211	#endif /* !(CONFIG_SMP) */
	212
	213	#endif /* _SPARC_TRAP_BLOCK_H */