[linux.git] / arch / mips / kernel / octeon_switch.S

/*
 * This file is subject to the terms and conditions of the GNU General Public
 * License.  See the file "COPYING" in the main directory of this archive
 * for more details.
 *
 * Copyright (C) 1994, 1995, 1996, 1998, 1999, 2002, 2003 Ralf Baechle
 * Copyright (C) 1996 David S. Miller ([email protected])
 * Copyright (C) 1994, 1995, 1996, by Andreas Busse
 * Copyright (C) 1999 Silicon Graphics, Inc.
 * Copyright (C) 2000 MIPS Technologies, Inc.
 *    written by Carsten Langgaard, [email protected]
 */
#include <asm/asm.h>
#include <asm/cachectl.h>
#include <asm/fpregdef.h>
#include <asm/mipsregs.h>
#include <asm/asm-offsets.h>
#include <asm/pgtable-bits.h>
#include <asm/regdef.h>
#include <asm/stackframe.h>
#include <asm/thread_info.h>

#include <asm/asmmacro.h>

/*
 * Offset to the current process status flags, the first 32 bytes of the
 * stack are not used.
 */
#define ST_OFF (_THREAD_SIZE - 32 - PT_SIZE + PT_STATUS)

/*
 * task_struct *resume(task_struct *prev, task_struct *next,
 *		       struct thread_info *next_ti, int usedfpu)
 */
	.align	7
	LEAF(resume)
	.set arch=octeon
	mfc0	t1, CP0_STATUS
	LONG_S	t1, THREAD_STATUS(a0)
	cpu_save_nonscratch a0
	LONG_S	ra, THREAD_REG31(a0)

	/* check if we need to save COP2 registers */
	PTR_L	t2, TASK_THREAD_INFO(a0)
	LONG_L	t0, ST_OFF(t2)
	bbit0	t0, 30, 1f

	/* Disable COP2 in the stored process state */
	li	t1, ST0_CU2
	xor	t0, t1
	LONG_S	t0, ST_OFF(t2)

	/* Enable COP2 so we can save it */
	mfc0	t0, CP0_STATUS
	or	t0, t1
	mtc0	t0, CP0_STATUS

	/* Save COP2 */
	daddu	a0, THREAD_CP2
	jal octeon_cop2_save
	dsubu	a0, THREAD_CP2

	/* Disable COP2 now that we are done */
	mfc0	t0, CP0_STATUS
	li	t1, ST0_CU2
	xor	t0, t1
	mtc0	t0, CP0_STATUS

1:
#if CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE > 0
	/* Check if we need to store CVMSEG state */
	mfc0	t0, $11,7	/* CvmMemCtl */
	bbit0	t0, 6, 3f	/* Is user access enabled? */

	/* Store the CVMSEG state */
	/* Extract the size of CVMSEG */
	andi	t0, 0x3f
	/* Multiply * (cache line size/sizeof(long)/2) */
	sll	t0, 7-LONGLOG-1
	li	t1, -32768	/* Base address of CVMSEG */
	LONG_ADDI t2, a0, THREAD_CVMSEG /* Where to store CVMSEG to */
	synciobdma
2:
	.set noreorder
	LONG_L	t8, 0(t1)	/* Load from CVMSEG */
	subu	t0, 1		/* Decrement loop var */
	LONG_L	t9, LONGSIZE(t1)/* Load from CVMSEG */
	LONG_ADDU t1, LONGSIZE*2 /* Increment loc in CVMSEG */
	LONG_S	t8, 0(t2)	/* Store CVMSEG to thread storage */
	LONG_ADDU t2, LONGSIZE*2 /* Increment loc in thread storage */
	bnez	t0, 2b		/* Loop until we've copied it all */
	 LONG_S t9, -LONGSIZE(t2)/* Store CVMSEG to thread storage */
	.set reorder

	/* Disable access to CVMSEG */
	mfc0	t0, $11,7	/* CvmMemCtl */
	xori	t0, t0, 0x40	/* Bit 6 is CVMSEG user enable */
	mtc0	t0, $11,7	/* CvmMemCtl */
#endif
3:
	/*
	 * The order of restoring the registers takes care of the race
	 * updating $28, $29 and kernelsp without disabling ints.
	 */
	move	$28, a2
	cpu_restore_nonscratch a1

#if (_THREAD_SIZE - 32) < 0x8000
	PTR_ADDIU	t0, $28, _THREAD_SIZE - 32
#else
	PTR_LI		t0, _THREAD_SIZE - 32
	PTR_ADDU	t0, $28
#endif
	set_saved_sp	t0, t1, t2

	mfc0	t1, CP0_STATUS		/* Do we really need this? */
	li	a3, 0xff01
	and	t1, a3
	LONG_L	a2, THREAD_STATUS(a1)
	nor	a3, $0, a3
	and	a2, a3
	or	a2, t1
	mtc0	a2, CP0_STATUS
	move	v0, a0
	jr	ra
	END(resume)

/*
 * void octeon_cop2_save(struct octeon_cop2_state *a0)
 */
	.align	7
	LEAF(octeon_cop2_save)

	dmfc0	t9, $9,7	/* CvmCtl register. */

	/* Save the COP2 CRC state */
	dmfc2	t0, 0x0201
	dmfc2	t1, 0x0202
	dmfc2	t2, 0x0200
	sd	t0, OCTEON_CP2_CRC_IV(a0)
	sd	t1, OCTEON_CP2_CRC_LENGTH(a0)
	sd	t2, OCTEON_CP2_CRC_POLY(a0)
	/* Skip next instructions if CvmCtl[NODFA_CP2] set */
	bbit1	t9, 28, 1f

	/* Save the LLM state */
	dmfc2	t0, 0x0402
	dmfc2	t1, 0x040A
	sd	t0, OCTEON_CP2_LLM_DAT(a0)
	sd	t1, OCTEON_CP2_LLM_DAT+8(a0)

1:	bbit1	t9, 26, 3f	/* done if CvmCtl[NOCRYPTO] set */

	/* Save the COP2 crypto state */
	/* this part is mostly common to both pass 1 and later revisions */
	dmfc2	t0, 0x0084
	dmfc2	t1, 0x0080
	dmfc2	t2, 0x0081
	dmfc2	t3, 0x0082
	sd	t0, OCTEON_CP2_3DES_IV(a0)
	dmfc2	t0, 0x0088
	sd	t1, OCTEON_CP2_3DES_KEY(a0)
	dmfc2	t1, 0x0111			/* only necessary for pass 1 */
	sd	t2, OCTEON_CP2_3DES_KEY+8(a0)
	dmfc2	t2, 0x0102
	sd	t3, OCTEON_CP2_3DES_KEY+16(a0)
	dmfc2	t3, 0x0103
	sd	t0, OCTEON_CP2_3DES_RESULT(a0)
	dmfc2	t0, 0x0104
	sd	t1, OCTEON_CP2_AES_INP0(a0)	/* only necessary for pass 1 */
	dmfc2	t1, 0x0105
	sd	t2, OCTEON_CP2_AES_IV(a0)
	dmfc2	t2, 0x0106
	sd	t3, OCTEON_CP2_AES_IV+8(a0)
	dmfc2	t3, 0x0107
	sd	t0, OCTEON_CP2_AES_KEY(a0)
	dmfc2	t0, 0x0110
	sd	t1, OCTEON_CP2_AES_KEY+8(a0)
	dmfc2	t1, 0x0100
	sd	t2, OCTEON_CP2_AES_KEY+16(a0)
	dmfc2	t2, 0x0101
	sd	t3, OCTEON_CP2_AES_KEY+24(a0)
	mfc0	t3, $15,0	/* Get the processor ID register */
	sd	t0, OCTEON_CP2_AES_KEYLEN(a0)
	li	t0, 0x000d0000	/* This is the processor ID of Octeon Pass1 */
	sd	t1, OCTEON_CP2_AES_RESULT(a0)
	sd	t2, OCTEON_CP2_AES_RESULT+8(a0)
	/* Skip to the Pass1 version of the remainder of the COP2 state */
	beq	t3, t0, 2f

	/* the non-pass1 state when !CvmCtl[NOCRYPTO] */
	dmfc2	t1, 0x0240
	dmfc2	t2, 0x0241
	dmfc2	t3, 0x0242
	dmfc2	t0, 0x0243
	sd	t1, OCTEON_CP2_HSH_DATW(a0)
	dmfc2	t1, 0x0244
	sd	t2, OCTEON_CP2_HSH_DATW+8(a0)
	dmfc2	t2, 0x0245
	sd	t3, OCTEON_CP2_HSH_DATW+16(a0)
	dmfc2	t3, 0x0246
	sd	t0, OCTEON_CP2_HSH_DATW+24(a0)
	dmfc2	t0, 0x0247
	sd	t1, OCTEON_CP2_HSH_DATW+32(a0)
	dmfc2	t1, 0x0248
	sd	t2, OCTEON_CP2_HSH_DATW+40(a0)
	dmfc2	t2, 0x0249
	sd	t3, OCTEON_CP2_HSH_DATW+48(a0)
	dmfc2	t3, 0x024A
	sd	t0, OCTEON_CP2_HSH_DATW+56(a0)
	dmfc2	t0, 0x024B
	sd	t1, OCTEON_CP2_HSH_DATW+64(a0)
	dmfc2	t1, 0x024C
	sd	t2, OCTEON_CP2_HSH_DATW+72(a0)
	dmfc2	t2, 0x024D
	sd	t3, OCTEON_CP2_HSH_DATW+80(a0)
	dmfc2	t3, 0x024E
	sd	t0, OCTEON_CP2_HSH_DATW+88(a0)
	dmfc2	t0, 0x0250
	sd	t1, OCTEON_CP2_HSH_DATW+96(a0)
	dmfc2	t1, 0x0251
	sd	t2, OCTEON_CP2_HSH_DATW+104(a0)
	dmfc2	t2, 0x0252
	sd	t3, OCTEON_CP2_HSH_DATW+112(a0)
	dmfc2	t3, 0x0253
	sd	t0, OCTEON_CP2_HSH_IVW(a0)
	dmfc2	t0, 0x0254
	sd	t1, OCTEON_CP2_HSH_IVW+8(a0)
	dmfc2	t1, 0x0255
	sd	t2, OCTEON_CP2_HSH_IVW+16(a0)
	dmfc2	t2, 0x0256
	sd	t3, OCTEON_CP2_HSH_IVW+24(a0)
	dmfc2	t3, 0x0257
	sd	t0, OCTEON_CP2_HSH_IVW+32(a0)
	dmfc2	t0, 0x0258
	sd	t1, OCTEON_CP2_HSH_IVW+40(a0)
	dmfc2	t1, 0x0259
	sd	t2, OCTEON_CP2_HSH_IVW+48(a0)
	dmfc2	t2, 0x025E
	sd	t3, OCTEON_CP2_HSH_IVW+56(a0)
	dmfc2	t3, 0x025A
	sd	t0, OCTEON_CP2_GFM_MULT(a0)
	dmfc2	t0, 0x025B
	sd	t1, OCTEON_CP2_GFM_MULT+8(a0)
	sd	t2, OCTEON_CP2_GFM_POLY(a0)
	sd	t3, OCTEON_CP2_GFM_RESULT(a0)
	sd	t0, OCTEON_CP2_GFM_RESULT+8(a0)
	jr	ra

2:	/* pass 1 special stuff when !CvmCtl[NOCRYPTO] */
	dmfc2	t3, 0x0040
	dmfc2	t0, 0x0041
	dmfc2	t1, 0x0042
	dmfc2	t2, 0x0043
	sd	t3, OCTEON_CP2_HSH_DATW(a0)
	dmfc2	t3, 0x0044
	sd	t0, OCTEON_CP2_HSH_DATW+8(a0)
	dmfc2	t0, 0x0045
	sd	t1, OCTEON_CP2_HSH_DATW+16(a0)
	dmfc2	t1, 0x0046
	sd	t2, OCTEON_CP2_HSH_DATW+24(a0)
	dmfc2	t2, 0x0048
	sd	t3, OCTEON_CP2_HSH_DATW+32(a0)
	dmfc2	t3, 0x0049
	sd	t0, OCTEON_CP2_HSH_DATW+40(a0)
	dmfc2	t0, 0x004A
	sd	t1, OCTEON_CP2_HSH_DATW+48(a0)
	sd	t2, OCTEON_CP2_HSH_IVW(a0)
	sd	t3, OCTEON_CP2_HSH_IVW+8(a0)
	sd	t0, OCTEON_CP2_HSH_IVW+16(a0)

3:	/* pass 1 or CvmCtl[NOCRYPTO] set */
	jr	ra
	END(octeon_cop2_save)

/*
 * void octeon_cop2_restore(struct octeon_cop2_state *a0)
 */
	.align	7
	.set push
	.set noreorder
	LEAF(octeon_cop2_restore)
	/* First cache line was prefetched before the call */
	pref	4,  128(a0)
	dmfc0	t9, $9,7	/* CvmCtl register. */

	pref	4,  256(a0)
	ld	t0, OCTEON_CP2_CRC_IV(a0)
	pref	4,  384(a0)
	ld	t1, OCTEON_CP2_CRC_LENGTH(a0)
	ld	t2, OCTEON_CP2_CRC_POLY(a0)

	/* Restore the COP2 CRC state */
	dmtc2	t0, 0x0201
	dmtc2	t1, 0x1202
	bbit1	t9, 28, 2f	/* Skip LLM if CvmCtl[NODFA_CP2] is set */
	 dmtc2	t2, 0x4200

	/* Restore the LLM state */
	ld	t0, OCTEON_CP2_LLM_DAT(a0)
	ld	t1, OCTEON_CP2_LLM_DAT+8(a0)
	dmtc2	t0, 0x0402
	dmtc2	t1, 0x040A

2:
	bbit1	t9, 26, done_restore	/* done if CvmCtl[NOCRYPTO] set */
	 nop

	/* Restore the COP2 crypto state common to pass 1 and pass 2 */
	ld	t0, OCTEON_CP2_3DES_IV(a0)
	ld	t1, OCTEON_CP2_3DES_KEY(a0)
	ld	t2, OCTEON_CP2_3DES_KEY+8(a0)
	dmtc2	t0, 0x0084
	ld	t0, OCTEON_CP2_3DES_KEY+16(a0)
	dmtc2	t1, 0x0080
	ld	t1, OCTEON_CP2_3DES_RESULT(a0)
	dmtc2	t2, 0x0081
	ld	t2, OCTEON_CP2_AES_INP0(a0) /* only really needed for pass 1 */
	dmtc2	t0, 0x0082
	ld	t0, OCTEON_CP2_AES_IV(a0)
	dmtc2	t1, 0x0098
	ld	t1, OCTEON_CP2_AES_IV+8(a0)
	dmtc2	t2, 0x010A		    /* only really needed for pass 1 */
	ld	t2, OCTEON_CP2_AES_KEY(a0)
	dmtc2	t0, 0x0102
	ld	t0, OCTEON_CP2_AES_KEY+8(a0)
	dmtc2	t1, 0x0103
	ld	t1, OCTEON_CP2_AES_KEY+16(a0)
	dmtc2	t2, 0x0104
	ld	t2, OCTEON_CP2_AES_KEY+24(a0)
	dmtc2	t0, 0x0105
	ld	t0, OCTEON_CP2_AES_KEYLEN(a0)
	dmtc2	t1, 0x0106
	ld	t1, OCTEON_CP2_AES_RESULT(a0)
	dmtc2	t2, 0x0107
	ld	t2, OCTEON_CP2_AES_RESULT+8(a0)
	mfc0	t3, $15,0	/* Get the processor ID register */
	dmtc2	t0, 0x0110
	li	t0, 0x000d0000	/* This is the processor ID of Octeon Pass1 */
	dmtc2	t1, 0x0100
	bne	t0, t3, 3f	/* Skip the next stuff for non-pass1 */
	 dmtc2	t2, 0x0101

	/* this code is specific for pass 1 */
	ld	t0, OCTEON_CP2_HSH_DATW(a0)
	ld	t1, OCTEON_CP2_HSH_DATW+8(a0)
	ld	t2, OCTEON_CP2_HSH_DATW+16(a0)
	dmtc2	t0, 0x0040
	ld	t0, OCTEON_CP2_HSH_DATW+24(a0)
	dmtc2	t1, 0x0041
	ld	t1, OCTEON_CP2_HSH_DATW+32(a0)
	dmtc2	t2, 0x0042
	ld	t2, OCTEON_CP2_HSH_DATW+40(a0)
	dmtc2	t0, 0x0043
	ld	t0, OCTEON_CP2_HSH_DATW+48(a0)
	dmtc2	t1, 0x0044
	ld	t1, OCTEON_CP2_HSH_IVW(a0)
	dmtc2	t2, 0x0045
	ld	t2, OCTEON_CP2_HSH_IVW+8(a0)
	dmtc2	t0, 0x0046
	ld	t0, OCTEON_CP2_HSH_IVW+16(a0)
	dmtc2	t1, 0x0048
	dmtc2	t2, 0x0049
	b done_restore	 /* unconditional branch */
	 dmtc2	t0, 0x004A

3:	/* this is post-pass1 code */
	ld	t2, OCTEON_CP2_HSH_DATW(a0)
	ld	t0, OCTEON_CP2_HSH_DATW+8(a0)
	ld	t1, OCTEON_CP2_HSH_DATW+16(a0)
	dmtc2	t2, 0x0240
	ld	t2, OCTEON_CP2_HSH_DATW+24(a0)
	dmtc2	t0, 0x0241
	ld	t0, OCTEON_CP2_HSH_DATW+32(a0)
	dmtc2	t1, 0x0242
	ld	t1, OCTEON_CP2_HSH_DATW+40(a0)
	dmtc2	t2, 0x0243
	ld	t2, OCTEON_CP2_HSH_DATW+48(a0)
	dmtc2	t0, 0x0244
	ld	t0, OCTEON_CP2_HSH_DATW+56(a0)
	dmtc2	t1, 0x0245
	ld	t1, OCTEON_CP2_HSH_DATW+64(a0)
	dmtc2	t2, 0x0246
	ld	t2, OCTEON_CP2_HSH_DATW+72(a0)
	dmtc2	t0, 0x0247
	ld	t0, OCTEON_CP2_HSH_DATW+80(a0)
	dmtc2	t1, 0x0248
	ld	t1, OCTEON_CP2_HSH_DATW+88(a0)
	dmtc2	t2, 0x0249
	ld	t2, OCTEON_CP2_HSH_DATW+96(a0)
	dmtc2	t0, 0x024A
	ld	t0, OCTEON_CP2_HSH_DATW+104(a0)
	dmtc2	t1, 0x024B
	ld	t1, OCTEON_CP2_HSH_DATW+112(a0)
	dmtc2	t2, 0x024C
	ld	t2, OCTEON_CP2_HSH_IVW(a0)
	dmtc2	t0, 0x024D
	ld	t0, OCTEON_CP2_HSH_IVW+8(a0)
	dmtc2	t1, 0x024E
	ld	t1, OCTEON_CP2_HSH_IVW+16(a0)
	dmtc2	t2, 0x0250
	ld	t2, OCTEON_CP2_HSH_IVW+24(a0)
	dmtc2	t0, 0x0251
	ld	t0, OCTEON_CP2_HSH_IVW+32(a0)
	dmtc2	t1, 0x0252
	ld	t1, OCTEON_CP2_HSH_IVW+40(a0)
	dmtc2	t2, 0x0253
	ld	t2, OCTEON_CP2_HSH_IVW+48(a0)
	dmtc2	t0, 0x0254
	ld	t0, OCTEON_CP2_HSH_IVW+56(a0)
	dmtc2	t1, 0x0255
	ld	t1, OCTEON_CP2_GFM_MULT(a0)
	dmtc2	t2, 0x0256
	ld	t2, OCTEON_CP2_GFM_MULT+8(a0)
	dmtc2	t0, 0x0257
	ld	t0, OCTEON_CP2_GFM_POLY(a0)
	dmtc2	t1, 0x0258
	ld	t1, OCTEON_CP2_GFM_RESULT(a0)
	dmtc2	t2, 0x0259
	ld	t2, OCTEON_CP2_GFM_RESULT+8(a0)
	dmtc2	t0, 0x025E
	dmtc2	t1, 0x025A
	dmtc2	t2, 0x025B

done_restore:
	jr	ra
	 nop
	END(octeon_cop2_restore)
	.set pop

/*
 * void octeon_mult_save()
 * sp is assumed to point to a struct pt_regs
 *
 * NOTE: This is called in SAVE_SOME in stackframe.h. It can only
 *	 safely modify k0 and k1.
 */
	.align	7
	.set push
	.set noreorder
	LEAF(octeon_mult_save)
	dmfc0	k0, $9,7	/* CvmCtl register. */
	bbit1	k0, 27, 1f	/* Skip CvmCtl[NOMUL] */
	 nop

	/* Save the multiplier state */
	v3mulu	k0, $0, $0
	v3mulu	k1, $0, $0
	sd	k0, PT_MTP(sp)	      /* PT_MTP	   has P0 */
	v3mulu	k0, $0, $0
	sd	k1, PT_MTP+8(sp)      /* PT_MTP+8  has P1 */
	ori	k1, $0, 1
	v3mulu	k1, k1, $0
	sd	k0, PT_MTP+16(sp)     /* PT_MTP+16 has P2 */
	v3mulu	k0, $0, $0
	sd	k1, PT_MPL(sp)	      /* PT_MPL	   has MPL0 */
	v3mulu	k1, $0, $0
	sd	k0, PT_MPL+8(sp)      /* PT_MPL+8  has MPL1 */
	jr	ra
	 sd	k1, PT_MPL+16(sp)     /* PT_MPL+16 has MPL2 */

1:	/* Resume here if CvmCtl[NOMUL] */
	jr	ra
	END(octeon_mult_save)
	.set pop

/*
 * void octeon_mult_restore()
 * sp is assumed to point to a struct pt_regs
 *
 * NOTE: This is called in RESTORE_SOME in stackframe.h.
 */
	.align	7
	.set push
	.set noreorder
	LEAF(octeon_mult_restore)
	dmfc0	k1, $9,7		/* CvmCtl register. */
	ld	v0, PT_MPL(sp)		/* MPL0 */
	ld	v1, PT_MPL+8(sp)	/* MPL1 */
	ld	k0, PT_MPL+16(sp)	/* MPL2 */
	bbit1	k1, 27, 1f		/* Skip CvmCtl[NOMUL] */
	/* Normally falls through, so no time wasted here */
	nop

	/* Restore the multiplier state */
	ld	k1, PT_MTP+16(sp)	/* P2 */
	MTM0	v0			/* MPL0 */
	ld	v0, PT_MTP+8(sp)	/* P1 */
	MTM1	v1			/* MPL1 */
	ld	v1, PT_MTP(sp)		/* P0 */
	MTM2	k0			/* MPL2 */
	MTP2	k1			/* P2 */
	MTP1	v0			/* P1 */
	jr	ra
	 MTP0	v1			/* P0 */

1:	/* Resume here if CvmCtl[NOMUL] */
	jr	ra
	 nop
	END(octeon_mult_restore)
	.set pop
Commit	Line	Data
5b3b1688 DD	1	/*
	2	* This file is subject to the terms and conditions of the GNU General Public
	3	* License. See the file "COPYING" in the main directory of this archive
	4	* for more details.
	5	*
	6	* Copyright (C) 1994, 1995, 1996, 1998, 1999, 2002, 2003 Ralf Baechle
79add627	7	* Copyright (C) 1996 David S. Miller ([email protected])
5b3b1688 DD	8	* Copyright (C) 1994, 1995, 1996, by Andreas Busse
	9	* Copyright (C) 1999 Silicon Graphics, Inc.
	10	* Copyright (C) 2000 MIPS Technologies, Inc.
	11	* written by Carsten Langgaard, [email protected]
	12	*/
	13	#include <asm/asm.h>
	14	#include <asm/cachectl.h>
	15	#include <asm/fpregdef.h>
	16	#include <asm/mipsregs.h>
	17	#include <asm/asm-offsets.h>
5b3b1688 DD	18	#include <asm/pgtable-bits.h>
	19	#include <asm/regdef.h>
	20	#include <asm/stackframe.h>
	21	#include <asm/thread_info.h>
	22
	23	#include <asm/asmmacro.h>
	24
	25	/*
	26	* Offset to the current process status flags, the first 32 bytes of the
	27	* stack are not used.
	28	*/
	29	#define ST_OFF (_THREAD_SIZE - 32 - PT_SIZE + PT_STATUS)
	30
	31	/*
	32	* task_struct resume(task_struct prev, task_struct *next,
70342287	33	* struct thread_info *next_ti, int usedfpu)
5b3b1688 DD	34	*/
	35	.align 7
	36	LEAF(resume)
	37	.set arch=octeon
5b3b1688 DD	38	mfc0 t1, CP0_STATUS
	39	LONG_S t1, THREAD_STATUS(a0)
	40	cpu_save_nonscratch a0
	41	LONG_S ra, THREAD_REG31(a0)
	42
	43	/* check if we need to save COP2 registers */
	44	PTR_L t2, TASK_THREAD_INFO(a0)
	45	LONG_L t0, ST_OFF(t2)
	46	bbit0 t0, 30, 1f
	47
	48	/* Disable COP2 in the stored process state */
	49	li t1, ST0_CU2
	50	xor t0, t1
	51	LONG_S t0, ST_OFF(t2)
	52
	53	/* Enable COP2 so we can save it */
	54	mfc0 t0, CP0_STATUS
	55	or t0, t1
	56	mtc0 t0, CP0_STATUS
	57
	58	/* Save COP2 */
	59	daddu a0, THREAD_CP2
	60	jal octeon_cop2_save
	61	dsubu a0, THREAD_CP2
	62
	63	/* Disable COP2 now that we are done */
	64	mfc0 t0, CP0_STATUS
	65	li t1, ST0_CU2
	66	xor t0, t1
	67	mtc0 t0, CP0_STATUS
	68
	69	1:
	70	#if CONFIG_CAVIUM_OCTEON_CVMSEG_SIZE > 0
	71	/* Check if we need to store CVMSEG state */
70342287	72	mfc0 t0, $11,7 /* CvmMemCtl */
5b3b1688 DD	73	bbit0 t0, 6, 3f /* Is user access enabled? */
	74
	75	/* Store the CVMSEG state */
	76	/* Extract the size of CVMSEG */
	77	andi t0, 0x3f
	78	/* Multiply * (cache line size/sizeof(long)/2) */
	79	sll t0, 7-LONGLOG-1
70342287 RB	80	li t1, -32768 /* Base address of CVMSEG */
70342287 RB	81	LONG_ADDI t2, a0, THREAD_CVMSEG /* Where to store CVMSEG to */
5b3b1688 DD	82	synciobdma
	83	2:
	84	.set noreorder
	85	LONG_L t8, 0(t1) /* Load from CVMSEG */
	86	subu t0, 1 /* Decrement loop var */
	87	LONG_L t9, LONGSIZE(t1)/* Load from CVMSEG */
	88	LONG_ADDU t1, LONGSIZE2 / Increment loc in CVMSEG */
	89	LONG_S t8, 0(t2) /* Store CVMSEG to thread storage */
	90	LONG_ADDU t2, LONGSIZE2 / Increment loc in thread storage */
	91	bnez t0, 2b /* Loop until we've copied it all */
70342287	92	LONG_S t9, -LONGSIZE(t2)/* Store CVMSEG to thread storage */
5b3b1688 DD	93	.set reorder
	94
	95	/* Disable access to CVMSEG */
70342287	96	mfc0 t0, $11,7 /* CvmMemCtl */
5b3b1688	97	xori t0, t0, 0x40 /* Bit 6 is CVMSEG user enable */
70342287	98	mtc0 t0, $11,7 /* CvmMemCtl */
5b3b1688 DD	99	#endif
	100	3:
	101	/*
	102	* The order of restoring the registers takes care of the race
	103	* updating $28, $29 and kernelsp without disabling ints.
	104	*/
	105	move $28, a2
	106	cpu_restore_nonscratch a1
	107
	108	#if (_THREAD_SIZE - 32) < 0x8000
	109	PTR_ADDIU t0, $28, _THREAD_SIZE - 32
	110	#else
	111	PTR_LI t0, _THREAD_SIZE - 32
	112	PTR_ADDU t0, $28
	113	#endif
	114	set_saved_sp t0, t1, t2
	115
	116	mfc0 t1, CP0_STATUS /* Do we really need this? */
	117	li a3, 0xff01
	118	and t1, a3
	119	LONG_L a2, THREAD_STATUS(a1)
	120	nor a3, $0, a3
	121	and a2, a3
	122	or a2, t1
	123	mtc0 a2, CP0_STATUS
	124	move v0, a0
	125	jr ra
	126	END(resume)
	127
	128	/*
	129	* void octeon_cop2_save(struct octeon_cop2_state *a0)
	130	*/
	131	.align 7
	132	LEAF(octeon_cop2_save)
	133
	134	dmfc0 t9, $9,7 /* CvmCtl register. */
	135
70342287	136	/* Save the COP2 CRC state */
5b3b1688 DD	137	dmfc2 t0, 0x0201
	138	dmfc2 t1, 0x0202
	139	dmfc2 t2, 0x0200
	140	sd t0, OCTEON_CP2_CRC_IV(a0)
	141	sd t1, OCTEON_CP2_CRC_LENGTH(a0)
	142	sd t2, OCTEON_CP2_CRC_POLY(a0)
	143	/* Skip next instructions if CvmCtl[NODFA_CP2] set */
	144	bbit1 t9, 28, 1f
	145
	146	/* Save the LLM state */
	147	dmfc2 t0, 0x0402
	148	dmfc2 t1, 0x040A
	149	sd t0, OCTEON_CP2_LLM_DAT(a0)
	150	sd t1, OCTEON_CP2_LLM_DAT+8(a0)
	151
70342287	152	1: bbit1 t9, 26, 3f /* done if CvmCtl[NOCRYPTO] set */
5b3b1688 DD	153
5b3b1688 DD	154	/* Save the COP2 crypto state */
70342287 RB	155	/* this part is mostly common to both pass 1 and later revisions */
	156	dmfc2 t0, 0x0084
	157	dmfc2 t1, 0x0080
	158	dmfc2 t2, 0x0081
	159	dmfc2 t3, 0x0082
5b3b1688	160	sd t0, OCTEON_CP2_3DES_IV(a0)
70342287	161	dmfc2 t0, 0x0088
5b3b1688	162	sd t1, OCTEON_CP2_3DES_KEY(a0)
70342287	163	dmfc2 t1, 0x0111 /* only necessary for pass 1 */
5b3b1688	164	sd t2, OCTEON_CP2_3DES_KEY+8(a0)
70342287	165	dmfc2 t2, 0x0102
5b3b1688	166	sd t3, OCTEON_CP2_3DES_KEY+16(a0)
70342287	167	dmfc2 t3, 0x0103
5b3b1688	168	sd t0, OCTEON_CP2_3DES_RESULT(a0)
70342287 RB	169	dmfc2 t0, 0x0104
	170	sd t1, OCTEON_CP2_AES_INP0(a0) /* only necessary for pass 1 */
	171	dmfc2 t1, 0x0105
5b3b1688 DD	172	sd t2, OCTEON_CP2_AES_IV(a0)
	173	dmfc2 t2, 0x0106
	174	sd t3, OCTEON_CP2_AES_IV+8(a0)
70342287	175	dmfc2 t3, 0x0107
5b3b1688 DD	176	sd t0, OCTEON_CP2_AES_KEY(a0)
	177	dmfc2 t0, 0x0110
	178	sd t1, OCTEON_CP2_AES_KEY+8(a0)
	179	dmfc2 t1, 0x0100
	180	sd t2, OCTEON_CP2_AES_KEY+16(a0)
	181	dmfc2 t2, 0x0101
	182	sd t3, OCTEON_CP2_AES_KEY+24(a0)
70342287	183	mfc0 t3, $15,0 /* Get the processor ID register */
5b3b1688 DD	184	sd t0, OCTEON_CP2_AES_KEYLEN(a0)
	185	li t0, 0x000d0000 /* This is the processor ID of Octeon Pass1 */
	186	sd t1, OCTEON_CP2_AES_RESULT(a0)
	187	sd t2, OCTEON_CP2_AES_RESULT+8(a0)
	188	/* Skip to the Pass1 version of the remainder of the COP2 state */
	189	beq t3, t0, 2f
	190
70342287	191	/* the non-pass1 state when !CvmCtl[NOCRYPTO] */
5b3b1688 DD	192	dmfc2 t1, 0x0240
	193	dmfc2 t2, 0x0241
	194	dmfc2 t3, 0x0242
	195	dmfc2 t0, 0x0243
	196	sd t1, OCTEON_CP2_HSH_DATW(a0)
	197	dmfc2 t1, 0x0244
	198	sd t2, OCTEON_CP2_HSH_DATW+8(a0)
	199	dmfc2 t2, 0x0245
	200	sd t3, OCTEON_CP2_HSH_DATW+16(a0)
	201	dmfc2 t3, 0x0246
	202	sd t0, OCTEON_CP2_HSH_DATW+24(a0)
	203	dmfc2 t0, 0x0247
	204	sd t1, OCTEON_CP2_HSH_DATW+32(a0)
	205	dmfc2 t1, 0x0248
	206	sd t2, OCTEON_CP2_HSH_DATW+40(a0)
	207	dmfc2 t2, 0x0249
	208	sd t3, OCTEON_CP2_HSH_DATW+48(a0)
	209	dmfc2 t3, 0x024A
	210	sd t0, OCTEON_CP2_HSH_DATW+56(a0)
	211	dmfc2 t0, 0x024B
	212	sd t1, OCTEON_CP2_HSH_DATW+64(a0)
	213	dmfc2 t1, 0x024C
	214	sd t2, OCTEON_CP2_HSH_DATW+72(a0)
	215	dmfc2 t2, 0x024D
	216	sd t3, OCTEON_CP2_HSH_DATW+80(a0)
70342287	217	dmfc2 t3, 0x024E
5b3b1688 DD	218	sd t0, OCTEON_CP2_HSH_DATW+88(a0)
	219	dmfc2 t0, 0x0250
	220	sd t1, OCTEON_CP2_HSH_DATW+96(a0)
	221	dmfc2 t1, 0x0251
	222	sd t2, OCTEON_CP2_HSH_DATW+104(a0)
	223	dmfc2 t2, 0x0252
	224	sd t3, OCTEON_CP2_HSH_DATW+112(a0)
	225	dmfc2 t3, 0x0253
	226	sd t0, OCTEON_CP2_HSH_IVW(a0)
	227	dmfc2 t0, 0x0254
	228	sd t1, OCTEON_CP2_HSH_IVW+8(a0)
	229	dmfc2 t1, 0x0255
	230	sd t2, OCTEON_CP2_HSH_IVW+16(a0)
	231	dmfc2 t2, 0x0256
	232	sd t3, OCTEON_CP2_HSH_IVW+24(a0)
	233	dmfc2 t3, 0x0257
	234	sd t0, OCTEON_CP2_HSH_IVW+32(a0)
70342287	235	dmfc2 t0, 0x0258
5b3b1688	236	sd t1, OCTEON_CP2_HSH_IVW+40(a0)
70342287	237	dmfc2 t1, 0x0259
5b3b1688 DD	238	sd t2, OCTEON_CP2_HSH_IVW+48(a0)
	239	dmfc2 t2, 0x025E
	240	sd t3, OCTEON_CP2_HSH_IVW+56(a0)
	241	dmfc2 t3, 0x025A
	242	sd t0, OCTEON_CP2_GFM_MULT(a0)
	243	dmfc2 t0, 0x025B
	244	sd t1, OCTEON_CP2_GFM_MULT+8(a0)
	245	sd t2, OCTEON_CP2_GFM_POLY(a0)
	246	sd t3, OCTEON_CP2_GFM_RESULT(a0)
	247	sd t0, OCTEON_CP2_GFM_RESULT+8(a0)
	248	jr ra
	249
70342287	250	2: /* pass 1 special stuff when !CvmCtl[NOCRYPTO] */
5b3b1688 DD	251	dmfc2 t3, 0x0040
	252	dmfc2 t0, 0x0041
	253	dmfc2 t1, 0x0042
	254	dmfc2 t2, 0x0043
	255	sd t3, OCTEON_CP2_HSH_DATW(a0)
	256	dmfc2 t3, 0x0044
	257	sd t0, OCTEON_CP2_HSH_DATW+8(a0)
	258	dmfc2 t0, 0x0045
	259	sd t1, OCTEON_CP2_HSH_DATW+16(a0)
	260	dmfc2 t1, 0x0046
	261	sd t2, OCTEON_CP2_HSH_DATW+24(a0)
	262	dmfc2 t2, 0x0048
	263	sd t3, OCTEON_CP2_HSH_DATW+32(a0)
	264	dmfc2 t3, 0x0049
	265	sd t0, OCTEON_CP2_HSH_DATW+40(a0)
	266	dmfc2 t0, 0x004A
	267	sd t1, OCTEON_CP2_HSH_DATW+48(a0)
	268	sd t2, OCTEON_CP2_HSH_IVW(a0)
	269	sd t3, OCTEON_CP2_HSH_IVW+8(a0)
	270	sd t0, OCTEON_CP2_HSH_IVW+16(a0)
	271
70342287	272	3: /* pass 1 or CvmCtl[NOCRYPTO] set */
5b3b1688 DD	273	jr ra
	274	END(octeon_cop2_save)
	275
	276	/*
	277	* void octeon_cop2_restore(struct octeon_cop2_state *a0)
	278	*/
	279	.align 7
	280	.set push
	281	.set noreorder
	282	LEAF(octeon_cop2_restore)
70342287 RB	283	/* First cache line was prefetched before the call */
70342287 RB	284	pref 4, 128(a0)
5b3b1688 DD	285	dmfc0 t9, $9,7 /* CvmCtl register. */
5b3b1688 DD	286
70342287	287	pref 4, 256(a0)
5b3b1688	288	ld t0, OCTEON_CP2_CRC_IV(a0)
70342287	289	pref 4, 384(a0)
5b3b1688 DD	290	ld t1, OCTEON_CP2_CRC_LENGTH(a0)
	291	ld t2, OCTEON_CP2_CRC_POLY(a0)
	292
	293	/* Restore the COP2 CRC state */
	294	dmtc2 t0, 0x0201
70342287	295	dmtc2 t1, 0x1202
5b3b1688 DD	296	bbit1 t9, 28, 2f /* Skip LLM if CvmCtl[NODFA_CP2] is set */
	297	dmtc2 t2, 0x4200
	298
	299	/* Restore the LLM state */
	300	ld t0, OCTEON_CP2_LLM_DAT(a0)
	301	ld t1, OCTEON_CP2_LLM_DAT+8(a0)
	302	dmtc2 t0, 0x0402
	303	dmtc2 t1, 0x040A
	304
	305	2:
	306	bbit1 t9, 26, done_restore /* done if CvmCtl[NOCRYPTO] set */
	307	nop
	308
	309	/* Restore the COP2 crypto state common to pass 1 and pass 2 */
	310	ld t0, OCTEON_CP2_3DES_IV(a0)
	311	ld t1, OCTEON_CP2_3DES_KEY(a0)
	312	ld t2, OCTEON_CP2_3DES_KEY+8(a0)
70342287	313	dmtc2 t0, 0x0084
5b3b1688	314	ld t0, OCTEON_CP2_3DES_KEY+16(a0)
70342287	315	dmtc2 t1, 0x0080
5b3b1688	316	ld t1, OCTEON_CP2_3DES_RESULT(a0)
70342287	317	dmtc2 t2, 0x0081
5b3b1688 DD	318	ld t2, OCTEON_CP2_AES_INP0(a0) /* only really needed for pass 1 */
	319	dmtc2 t0, 0x0082
	320	ld t0, OCTEON_CP2_AES_IV(a0)
70342287	321	dmtc2 t1, 0x0098
5b3b1688	322	ld t1, OCTEON_CP2_AES_IV+8(a0)
70342287	323	dmtc2 t2, 0x010A /* only really needed for pass 1 */
5b3b1688	324	ld t2, OCTEON_CP2_AES_KEY(a0)
70342287	325	dmtc2 t0, 0x0102
5b3b1688 DD	326	ld t0, OCTEON_CP2_AES_KEY+8(a0)
	327	dmtc2 t1, 0x0103
	328	ld t1, OCTEON_CP2_AES_KEY+16(a0)
	329	dmtc2 t2, 0x0104
	330	ld t2, OCTEON_CP2_AES_KEY+24(a0)
	331	dmtc2 t0, 0x0105
	332	ld t0, OCTEON_CP2_AES_KEYLEN(a0)
	333	dmtc2 t1, 0x0106
	334	ld t1, OCTEON_CP2_AES_RESULT(a0)
	335	dmtc2 t2, 0x0107
	336	ld t2, OCTEON_CP2_AES_RESULT+8(a0)
70342287	337	mfc0 t3, $15,0 /* Get the processor ID register */
5b3b1688 DD	338	dmtc2 t0, 0x0110
	339	li t0, 0x000d0000 /* This is the processor ID of Octeon Pass1 */
	340	dmtc2 t1, 0x0100
	341	bne t0, t3, 3f /* Skip the next stuff for non-pass1 */
	342	dmtc2 t2, 0x0101
	343
70342287	344	/* this code is specific for pass 1 */
5b3b1688 DD	345	ld t0, OCTEON_CP2_HSH_DATW(a0)
	346	ld t1, OCTEON_CP2_HSH_DATW+8(a0)
	347	ld t2, OCTEON_CP2_HSH_DATW+16(a0)
	348	dmtc2 t0, 0x0040
	349	ld t0, OCTEON_CP2_HSH_DATW+24(a0)
	350	dmtc2 t1, 0x0041
	351	ld t1, OCTEON_CP2_HSH_DATW+32(a0)
	352	dmtc2 t2, 0x0042
	353	ld t2, OCTEON_CP2_HSH_DATW+40(a0)
	354	dmtc2 t0, 0x0043
	355	ld t0, OCTEON_CP2_HSH_DATW+48(a0)
	356	dmtc2 t1, 0x0044
	357	ld t1, OCTEON_CP2_HSH_IVW(a0)
	358	dmtc2 t2, 0x0045
	359	ld t2, OCTEON_CP2_HSH_IVW+8(a0)
	360	dmtc2 t0, 0x0046
	361	ld t0, OCTEON_CP2_HSH_IVW+16(a0)
	362	dmtc2 t1, 0x0048
	363	dmtc2 t2, 0x0049
70342287	364	b done_restore /* unconditional branch */
5b3b1688 DD	365	dmtc2 t0, 0x004A
5b3b1688 DD	366
70342287	367	3: /* this is post-pass1 code */
5b3b1688 DD	368	ld t2, OCTEON_CP2_HSH_DATW(a0)
	369	ld t0, OCTEON_CP2_HSH_DATW+8(a0)
	370	ld t1, OCTEON_CP2_HSH_DATW+16(a0)
	371	dmtc2 t2, 0x0240
	372	ld t2, OCTEON_CP2_HSH_DATW+24(a0)
	373	dmtc2 t0, 0x0241
	374	ld t0, OCTEON_CP2_HSH_DATW+32(a0)
	375	dmtc2 t1, 0x0242
	376	ld t1, OCTEON_CP2_HSH_DATW+40(a0)
	377	dmtc2 t2, 0x0243
	378	ld t2, OCTEON_CP2_HSH_DATW+48(a0)
	379	dmtc2 t0, 0x0244
	380	ld t0, OCTEON_CP2_HSH_DATW+56(a0)
	381	dmtc2 t1, 0x0245
	382	ld t1, OCTEON_CP2_HSH_DATW+64(a0)
	383	dmtc2 t2, 0x0246
	384	ld t2, OCTEON_CP2_HSH_DATW+72(a0)
	385	dmtc2 t0, 0x0247
	386	ld t0, OCTEON_CP2_HSH_DATW+80(a0)
	387	dmtc2 t1, 0x0248
	388	ld t1, OCTEON_CP2_HSH_DATW+88(a0)
	389	dmtc2 t2, 0x0249
	390	ld t2, OCTEON_CP2_HSH_DATW+96(a0)
	391	dmtc2 t0, 0x024A
	392	ld t0, OCTEON_CP2_HSH_DATW+104(a0)
	393	dmtc2 t1, 0x024B
	394	ld t1, OCTEON_CP2_HSH_DATW+112(a0)
	395	dmtc2 t2, 0x024C
	396	ld t2, OCTEON_CP2_HSH_IVW(a0)
	397	dmtc2 t0, 0x024D
	398	ld t0, OCTEON_CP2_HSH_IVW+8(a0)
	399	dmtc2 t1, 0x024E
	400	ld t1, OCTEON_CP2_HSH_IVW+16(a0)
	401	dmtc2 t2, 0x0250
	402	ld t2, OCTEON_CP2_HSH_IVW+24(a0)
	403	dmtc2 t0, 0x0251
	404	ld t0, OCTEON_CP2_HSH_IVW+32(a0)
	405	dmtc2 t1, 0x0252
	406	ld t1, OCTEON_CP2_HSH_IVW+40(a0)
	407	dmtc2 t2, 0x0253
	408	ld t2, OCTEON_CP2_HSH_IVW+48(a0)
	409	dmtc2 t0, 0x0254
	410	ld t0, OCTEON_CP2_HSH_IVW+56(a0)
	411	dmtc2 t1, 0x0255
	412	ld t1, OCTEON_CP2_GFM_MULT(a0)
	413	dmtc2 t2, 0x0256
	414	ld t2, OCTEON_CP2_GFM_MULT+8(a0)
	415	dmtc2 t0, 0x0257
	416	ld t0, OCTEON_CP2_GFM_POLY(a0)
	417	dmtc2 t1, 0x0258
	418	ld t1, OCTEON_CP2_GFM_RESULT(a0)
	419	dmtc2 t2, 0x0259
	420	ld t2, OCTEON_CP2_GFM_RESULT+8(a0)
	421	dmtc2 t0, 0x025E
	422	dmtc2 t1, 0x025A
	423	dmtc2 t2, 0x025B
	424
	425	done_restore:
	426	jr ra
	427	nop
	428	END(octeon_cop2_restore)
	429	.set pop
	430
	431	/*
432	* void octeon_mult_save()
433	* sp is assumed to point to a struct pt_regs
434	*
435	* NOTE: This is called in SAVE_SOME in stackframe.h. It can only
70342287	436	* safely modify k0 and k1.
5b3b1688 DD	437	*/
	438	.align 7
	439	.set push
	440	.set noreorder
	441	LEAF(octeon_mult_save)
	442	dmfc0 k0, $9,7 /* CvmCtl register. */
	443	bbit1 k0, 27, 1f /* Skip CvmCtl[NOMUL] */
	444	nop
	445
	446	/* Save the multiplier state */
	447	v3mulu k0, $0, $0
	448	v3mulu k1, $0, $0
70342287	449	sd k0, PT_MTP(sp) /* PT_MTP has P0 */
5b3b1688 DD	450	v3mulu k0, $0, $0
	451	sd k1, PT_MTP+8(sp) /* PT_MTP+8 has P1 */
	452	ori k1, $0, 1
	453	v3mulu k1, k1, $0
	454	sd k0, PT_MTP+16(sp) /* PT_MTP+16 has P2 */
	455	v3mulu k0, $0, $0
70342287	456	sd k1, PT_MPL(sp) /* PT_MPL has MPL0 */
5b3b1688 DD	457	v3mulu k1, $0, $0
	458	sd k0, PT_MPL+8(sp) /* PT_MPL+8 has MPL1 */
	459	jr ra
	460	sd k1, PT_MPL+16(sp) /* PT_MPL+16 has MPL2 */
	461
	462	1: /* Resume here if CvmCtl[NOMUL] */
	463	jr ra
	464	END(octeon_mult_save)
	465	.set pop
	466
	467	/*
	468	* void octeon_mult_restore()
	469	* sp is assumed to point to a struct pt_regs
	470	*
	471	* NOTE: This is called in RESTORE_SOME in stackframe.h.
	472	*/
	473	.align 7
	474	.set push
	475	.set noreorder
	476	LEAF(octeon_mult_restore)
	477	dmfc0 k1, $9,7 /* CvmCtl register. */
70342287 RB	478	ld v0, PT_MPL(sp) /* MPL0 */
	479	ld v1, PT_MPL+8(sp) /* MPL1 */
	480	ld k0, PT_MPL+16(sp) /* MPL2 */
5b3b1688 DD	481	bbit1 k1, 27, 1f /* Skip CvmCtl[NOMUL] */
	482	/* Normally falls through, so no time wasted here */
	483	nop
	484
	485	/* Restore the multiplier state */
70342287	486	ld k1, PT_MTP+16(sp) /* P2 */
5b3b1688 DD	487	MTM0 v0 /* MPL0 */
	488	ld v0, PT_MTP+8(sp) /* P1 */
	489	MTM1 v1 /* MPL1 */
70342287	490	ld v1, PT_MTP(sp) /* P0 */
5b3b1688 DD	491	MTM2 k0 /* MPL2 */
	492	MTP2 k1 /* P2 */
	493	MTP1 v0 /* P1 */
	494	jr ra
	495	MTP0 v1 /* P0 */
	496
	497	1: /* Resume here if CvmCtl[NOMUL] */
	498	jr ra
	499	nop
	500	END(octeon_mult_restore)
	501	.set pop