Skip to content

Commit

Permalink
ia64/pv_ops: implement binary patching optimization for native.
Browse files Browse the repository at this point in the history 
implement binary patching optimization for pv_cpu_ops.
With this optimization, indirect call for pv_cpu_ops methods can be
converted into inline execution or direct call.

Signed-off-by: Isaku Yamahata <yamahata@valinux.co.jp>
Signed-off-by: Tony Luck <tony.luck@intel.com>
  • Loading branch information
Isaku Yamahata authored and Tony Luck committed Mar 26, 2009
1 parent bf7ab02 commit 03f511d
Show file tree
Hide file tree
Showing 7 changed files with 898 additions and 25 deletions.
6 changes: 5 additions & 1 deletion arch/ia64/include/asm/intrinsics.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -201,7 +201,11 @@ extern long ia64_cmpxchg_called_with_bad_pointer (void);

#ifndef __ASSEMBLY__
#if defined(CONFIG_PARAVIRT) && defined(__KERNEL__)
#define IA64_INTRINSIC_API(name) pv_cpu_ops.name
#ifdef ASM_SUPPORTED
# define IA64_INTRINSIC_API(name) paravirt_ ## name
#else
# define IA64_INTRINSIC_API(name) pv_cpu_ops.name
#endif
#define IA64_INTRINSIC_MACRO(name) paravirt_ ## name
#else
#define IA64_INTRINSIC_API(name) ia64_native_ ## name
Expand Down
8 changes: 8 additions & 0 deletions arch/ia64/include/asm/paravirt.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -118,6 +118,14 @@ struct pv_init_ops {
int (*arch_setup_nomca)(void);

void (*post_smp_prepare_boot_cpu)(void);

#ifdef ASM_SUPPORTED
unsigned long (*patch_bundle)(void *sbundle, void *ebundle,
unsigned long type);
unsigned long (*patch_inst)(unsigned long stag, unsigned long etag,
unsigned long type);
#endif
void (*patch_branch)(unsigned long tag, unsigned long type);
};

extern struct pv_init_ops pv_init_ops;
Expand Down
341 changes: 338 additions & 3 deletions arch/ia64/include/asm/paravirt_privop.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -60,20 +60,26 @@ extern unsigned long ia64_native_getreg_func(int regnum);
/* Instructions paravirtualized for performance */
/************************************************/

#ifndef ASM_SUPPORTED
#define paravirt_ssm_i() pv_cpu_ops.ssm_i()
#define paravirt_rsm_i() pv_cpu_ops.rsm_i()
#define __paravirt_getreg() pv_cpu_ops.getreg()
#endif

/* mask for ia64_native_ssm/rsm() must be constant.("i" constraing).
* static inline function doesn't satisfy it. */
#define paravirt_ssm(mask) \
do { \
if ((mask) == IA64_PSR_I) \
pv_cpu_ops.ssm_i(); \
paravirt_ssm_i(); \
else \
ia64_native_ssm(mask); \
} while (0)

#define paravirt_rsm(mask) \
do { \
if ((mask) == IA64_PSR_I) \
pv_cpu_ops.rsm_i(); \
paravirt_rsm_i(); \
else \
ia64_native_rsm(mask); \
} while (0)
Expand All @@ -86,7 +92,7 @@ extern unsigned long ia64_native_getreg_func(int regnum);
if ((reg) == _IA64_REG_IP) \
res = ia64_native_getreg(_IA64_REG_IP); \
else \
res = pv_cpu_ops.getreg(reg); \
res = __paravirt_getreg(reg); \
res; \
})

Expand Down Expand Up @@ -121,4 +127,333 @@ void paravirt_cpu_asm_init(const struct pv_cpu_asm_switch *cpu_asm_switch);
IA64_PARAVIRT_ASM_FUNC(work_processed_syscall)
#define ia64_leave_kernel IA64_PARAVIRT_ASM_FUNC(leave_kernel)


#if defined(CONFIG_PARAVIRT)
/******************************************************************************
* binary patching infrastructure
*/
#define PARAVIRT_PATCH_TYPE_FC 1
#define PARAVIRT_PATCH_TYPE_THASH 2
#define PARAVIRT_PATCH_TYPE_GET_CPUID 3
#define PARAVIRT_PATCH_TYPE_GET_PMD 4
#define PARAVIRT_PATCH_TYPE_PTCGA 5
#define PARAVIRT_PATCH_TYPE_GET_RR 6
#define PARAVIRT_PATCH_TYPE_SET_RR 7
#define PARAVIRT_PATCH_TYPE_SET_RR0_TO_RR4 8
#define PARAVIRT_PATCH_TYPE_SSM_I 9
#define PARAVIRT_PATCH_TYPE_RSM_I 10
#define PARAVIRT_PATCH_TYPE_GET_PSR_I 11
#define PARAVIRT_PATCH_TYPE_INTRIN_LOCAL_IRQ_RESTORE 12

/* PARAVIRT_PATY_TYPE_[GS]ETREG + _IA64_REG_xxx */
#define PARAVIRT_PATCH_TYPE_GETREG 0x10000000
#define PARAVIRT_PATCH_TYPE_SETREG 0x20000000

/*
* struct task_struct* (*ia64_switch_to)(void* next_task);
* void *ia64_leave_syscall;
* void *ia64_work_processed_syscall
* void *ia64_leave_kernel;
*/

#define PARAVIRT_PATCH_TYPE_BR_START 0x30000000
#define PARAVIRT_PATCH_TYPE_BR_SWITCH_TO \
(PARAVIRT_PATCH_TYPE_BR_START + 0)
#define PARAVIRT_PATCH_TYPE_BR_LEAVE_SYSCALL \
(PARAVIRT_PATCH_TYPE_BR_START + 1)
#define PARAVIRT_PATCH_TYPE_BR_WORK_PROCESSED_SYSCALL \
(PARAVIRT_PATCH_TYPE_BR_START + 2)
#define PARAVIRT_PATCH_TYPE_BR_LEAVE_KERNEL \
(PARAVIRT_PATCH_TYPE_BR_START + 3)

#ifdef ASM_SUPPORTED
#include <asm/paravirt_patch.h>

/*
* pv_cpu_ops calling stub.
* normal function call convension can't be written by gcc
* inline assembly.
*
* from the caller's point of view,
* the following registers will be clobbered.
* r2, r3
* r8-r15
* r16, r17
* b6, b7
* p6-p15
* ar.ccv
*
* from the callee's point of view ,
* the following registers can be used.
* r2, r3: scratch
* r8: scratch, input argument0 and return value
* r0-r15: scratch, input argument1-5
* b6: return pointer
* b7: scratch
* p6-p15: scratch
* ar.ccv: scratch
*
* other registers must not be changed. especially
* b0: rp: preserved. gcc ignores b0 in clobbered register.
* r16: saved gp
*/
/* 5 bundles */
#define __PARAVIRT_BR \
";;\n" \
"{ .mlx\n" \
"nop 0\n" \
"movl r2 = %[op_addr]\n"/* get function pointer address */ \
";;\n" \
"}\n" \
"1:\n" \
"{ .mii\n" \
"ld8 r2 = [r2]\n" /* load function descriptor address */ \
"mov r17 = ip\n" /* get ip to calc return address */ \
"mov r16 = gp\n" /* save gp */ \
";;\n" \
"}\n" \
"{ .mii\n" \
"ld8 r3 = [r2], 8\n" /* load entry address */ \
"adds r17 = 1f - 1b, r17\n" /* calculate return address */ \
";;\n" \
"mov b7 = r3\n" /* set entry address */ \
"}\n" \
"{ .mib\n" \
"ld8 gp = [r2]\n" /* load gp value */ \
"mov b6 = r17\n" /* set return address */ \
"br.cond.sptk.few b7\n" /* intrinsics are very short isns */ \
"}\n" \
"1:\n" \
"{ .mii\n" \
"mov gp = r16\n" /* restore gp value */ \
"nop 0\n" \
"nop 0\n" \
";;\n" \
"}\n"

#define PARAVIRT_OP(op) \
[op_addr] "i"(&pv_cpu_ops.op)

#define PARAVIRT_TYPE(type) \
PARAVIRT_PATCH_TYPE_ ## type

#define PARAVIRT_REG_CLOBBERS0 \
"r2", "r3", /*"r8",*/ "r9", "r10", "r11", "r14", \
"r15", "r16", "r17"

#define PARAVIRT_REG_CLOBBERS1 \
"r2","r3", /*"r8",*/ "r9", "r10", "r11", "r14", \
"r15", "r16", "r17"

#define PARAVIRT_REG_CLOBBERS2 \
"r2", "r3", /*"r8", "r9",*/ "r10", "r11", "r14", \
"r15", "r16", "r17"

#define PARAVIRT_REG_CLOBBERS5 \
"r2", "r3", /*"r8", "r9", "r10", "r11", "r14",*/ \
"r15", "r16", "r17"

#define PARAVIRT_BR_CLOBBERS \
"b6", "b7"

#define PARAVIRT_PR_CLOBBERS \
"p6", "p7", "p8", "p9", "p10", "p11", "p12", "p13", "p14", "p15"

#define PARAVIRT_AR_CLOBBERS \
"ar.ccv"

#define PARAVIRT_CLOBBERS0 \
PARAVIRT_REG_CLOBBERS0, \
PARAVIRT_BR_CLOBBERS, \
PARAVIRT_PR_CLOBBERS, \
PARAVIRT_AR_CLOBBERS, \
"memory"

#define PARAVIRT_CLOBBERS1 \
PARAVIRT_REG_CLOBBERS1, \
PARAVIRT_BR_CLOBBERS, \
PARAVIRT_PR_CLOBBERS, \
PARAVIRT_AR_CLOBBERS, \
"memory"

#define PARAVIRT_CLOBBERS2 \
PARAVIRT_REG_CLOBBERS2, \
PARAVIRT_BR_CLOBBERS, \
PARAVIRT_PR_CLOBBERS, \
PARAVIRT_AR_CLOBBERS, \
"memory"

#define PARAVIRT_CLOBBERS5 \
PARAVIRT_REG_CLOBBERS5, \
PARAVIRT_BR_CLOBBERS, \
PARAVIRT_PR_CLOBBERS, \
PARAVIRT_AR_CLOBBERS, \
"memory"

#define PARAVIRT_BR0(op, type) \
register unsigned long ia64_clobber asm ("r8"); \
asm volatile (paravirt_alt_bundle(__PARAVIRT_BR, \
PARAVIRT_TYPE(type)) \
: "=r"(ia64_clobber) \
: PARAVIRT_OP(op) \
: PARAVIRT_CLOBBERS0)

#define PARAVIRT_BR0_RET(op, type) \
register unsigned long ia64_intri_res asm ("r8"); \
asm volatile (paravirt_alt_bundle(__PARAVIRT_BR, \
PARAVIRT_TYPE(type)) \
: "=r"(ia64_intri_res) \
: PARAVIRT_OP(op) \
: PARAVIRT_CLOBBERS0)

#define PARAVIRT_BR1(op, type, arg1) \
register unsigned long __##arg1 asm ("r8") = arg1; \
register unsigned long ia64_clobber asm ("r8"); \
asm volatile (paravirt_alt_bundle(__PARAVIRT_BR, \
PARAVIRT_TYPE(type)) \
: "=r"(ia64_clobber) \
: PARAVIRT_OP(op), "0"(__##arg1) \
: PARAVIRT_CLOBBERS1)

#define PARAVIRT_BR1_RET(op, type, arg1) \
register unsigned long ia64_intri_res asm ("r8"); \
register unsigned long __##arg1 asm ("r8") = arg1; \
asm volatile (paravirt_alt_bundle(__PARAVIRT_BR, \
PARAVIRT_TYPE(type)) \
: "=r"(ia64_intri_res) \
: PARAVIRT_OP(op), "0"(__##arg1) \
: PARAVIRT_CLOBBERS1)

#define PARAVIRT_BR2(op, type, arg1, arg2) \
register unsigned long __##arg1 asm ("r8") = arg1; \
register unsigned long __##arg2 asm ("r9") = arg2; \
register unsigned long ia64_clobber1 asm ("r8"); \
register unsigned long ia64_clobber2 asm ("r9"); \
asm volatile (paravirt_alt_bundle(__PARAVIRT_BR, \
PARAVIRT_TYPE(type)) \
: "=r"(ia64_clobber1), "=r"(ia64_clobber2) \
: PARAVIRT_OP(op), "0"(__##arg1), "1"(__##arg2) \
: PARAVIRT_CLOBBERS2)


#define PARAVIRT_DEFINE_CPU_OP0(op, type) \
static inline void \
paravirt_ ## op (void) \
{ \
PARAVIRT_BR0(op, type); \
}

#define PARAVIRT_DEFINE_CPU_OP0_RET(op, type) \
static inline unsigned long \
paravirt_ ## op (void) \
{ \
PARAVIRT_BR0_RET(op, type); \
return ia64_intri_res; \
}

#define PARAVIRT_DEFINE_CPU_OP1(op, type) \
static inline void \
paravirt_ ## op (unsigned long arg1) \
{ \
PARAVIRT_BR1(op, type, arg1); \
}

#define PARAVIRT_DEFINE_CPU_OP1_RET(op, type) \
static inline unsigned long \
paravirt_ ## op (unsigned long arg1) \
{ \
PARAVIRT_BR1_RET(op, type, arg1); \
return ia64_intri_res; \
}

#define PARAVIRT_DEFINE_CPU_OP2(op, type) \
static inline void \
paravirt_ ## op (unsigned long arg1, \
unsigned long arg2) \
{ \
PARAVIRT_BR2(op, type, arg1, arg2); \
}


PARAVIRT_DEFINE_CPU_OP1(fc, FC);
PARAVIRT_DEFINE_CPU_OP1_RET(thash, THASH)
PARAVIRT_DEFINE_CPU_OP1_RET(get_cpuid, GET_CPUID)
PARAVIRT_DEFINE_CPU_OP1_RET(get_pmd, GET_PMD)
PARAVIRT_DEFINE_CPU_OP2(ptcga, PTCGA)
PARAVIRT_DEFINE_CPU_OP1_RET(get_rr, GET_RR)
PARAVIRT_DEFINE_CPU_OP2(set_rr, SET_RR)
PARAVIRT_DEFINE_CPU_OP0(ssm_i, SSM_I)
PARAVIRT_DEFINE_CPU_OP0(rsm_i, RSM_I)
PARAVIRT_DEFINE_CPU_OP0_RET(get_psr_i, GET_PSR_I)
PARAVIRT_DEFINE_CPU_OP1(intrin_local_irq_restore, INTRIN_LOCAL_IRQ_RESTORE)

static inline void
paravirt_set_rr0_to_rr4(unsigned long val0, unsigned long val1,
unsigned long val2, unsigned long val3,
unsigned long val4)
{
register unsigned long __val0 asm ("r8") = val0;
register unsigned long __val1 asm ("r9") = val1;
register unsigned long __val2 asm ("r10") = val2;
register unsigned long __val3 asm ("r11") = val3;
register unsigned long __val4 asm ("r14") = val4;

register unsigned long ia64_clobber0 asm ("r8");
register unsigned long ia64_clobber1 asm ("r9");
register unsigned long ia64_clobber2 asm ("r10");
register unsigned long ia64_clobber3 asm ("r11");
register unsigned long ia64_clobber4 asm ("r14");

asm volatile (paravirt_alt_bundle(__PARAVIRT_BR,
PARAVIRT_TYPE(SET_RR0_TO_RR4))
: "=r"(ia64_clobber0),
"=r"(ia64_clobber1),
"=r"(ia64_clobber2),
"=r"(ia64_clobber3),
"=r"(ia64_clobber4)
: PARAVIRT_OP(set_rr0_to_rr4),
"0"(__val0), "1"(__val1), "2"(__val2),
"3"(__val3), "4"(__val4)
: PARAVIRT_CLOBBERS5);
}

/* unsigned long paravirt_getreg(int reg) */
#define __paravirt_getreg(reg) \
({ \
register unsigned long ia64_intri_res asm ("r8"); \
register unsigned long __reg asm ("r8") = (reg); \
\
BUILD_BUG_ON(!__builtin_constant_p(reg)); \
asm volatile (paravirt_alt_bundle(__PARAVIRT_BR, \
PARAVIRT_TYPE(GETREG) \
+ (reg)) \
: "=r"(ia64_intri_res) \
: PARAVIRT_OP(getreg), "0"(__reg) \
: PARAVIRT_CLOBBERS1); \
\
ia64_intri_res; \
})

/* void paravirt_setreg(int reg, unsigned long val) */
#define paravirt_setreg(reg, val) \
do { \
register unsigned long __val asm ("r8") = val; \
register unsigned long __reg asm ("r9") = reg; \
register unsigned long ia64_clobber1 asm ("r8"); \
register unsigned long ia64_clobber2 asm ("r9"); \
\
BUILD_BUG_ON(!__builtin_constant_p(reg)); \
asm volatile (paravirt_alt_bundle(__PARAVIRT_BR, \
PARAVIRT_TYPE(SETREG) \
+ (reg)) \
: "=r"(ia64_clobber1), \
"=r"(ia64_clobber2) \
: PARAVIRT_OP(setreg), \
"1"(__reg), "0"(__val) \
: PARAVIRT_CLOBBERS2); \
} while (0)

#endif /* ASM_SUPPORTED */
#endif /* CONFIG_PARAVIRT && ASM_SUPPOTED */

#endif /* _ASM_IA64_PARAVIRT_PRIVOP_H */
Loading

0 comments on commit 03f511d

Please sign in to comment.