[patch 2/3] S390-HWBKPT v5: Enabling s390 kernel for kernel PER

[patch 2/3] S390-HWBKPT v5: Enabling s390 kernel for kernel PER

[Mahesh Salgaonkar]

Enable kernel PER tracing in the kernel entry point.
A new flag kernel_per_tracing has been added to lowcore for each CPU,
which is checked at every kernel entry point and set the PER bit in
current PSW.

Signed-off-by: Mahesh Salgaonkar <mahesh@linux.vnet.ibm.com>
---

 arch/s390/include/asm/lowcore.h  |   10 ++++++++--
 arch/s390/kernel/asm-offsets.c   |    2 ++
 arch/s390/kernel/entry.S         |   12 ++++++++++++
 arch/s390/kernel/entry64.S       |   12 ++++++++++++
 arch/s390/kernel/hw_breakpoint.c |   28 +++++++++++++++++++++-------
 5 files changed, 55 insertions(+), 9 deletions(-)


diff --git a/arch/s390/include/asm/lowcore.h b/arch/s390/include/asm/lowcore.h
index 05527c0..a9d3c5f 100644
--- a/arch/s390/include/asm/lowcore.h
+++ b/arch/s390/include/asm/lowcore.h
@@ -135,7 +135,10 @@ struct _lowcore {
 	__u64	clock_comparator;		/* 0x02d0 */
 	__u32	machine_flags;			/* 0x02d8 */
 	__u32	ftrace_func;			/* 0x02dc */
-	__u8	pad_0x02e0[0x0300-0x02e0];	/* 0x02e0 */
+
+	/* Kernel PER tracing mask */
+	__u32	kernel_per_tracing_mask;	/* 0x02e0 */
+	__u8	pad_0x02e4[0x0300-0x02e4];	/* 0x02e4 */
 
 	/* Interrupt response block */
 	__u8	irb[64];			/* 0x0300 */
@@ -265,7 +268,10 @@ struct _lowcore {
 	__u64	vdso_per_cpu_data;		/* 0x0350 */
 	__u64	machine_flags;			/* 0x0358 */
 	__u64	ftrace_func;			/* 0x0360 */
-	__u8	pad_0x0368[0x0380-0x0368];	/* 0x0368 */
+
+	/* Kernel PER tracing mask */
+	__u64	kernel_per_tracing_mask;	/* 0x0368 */
+	__u8	pad_0x0370[0x0380-0x0370];	/* 0x0370 */
 
 	/* Interrupt response block. */
 	__u8	irb[64];			/* 0x0380 */
diff --git a/arch/s390/kernel/asm-offsets.c b/arch/s390/kernel/asm-offsets.c
index a094089..abb33ec 100644
--- a/arch/s390/kernel/asm-offsets.c
+++ b/arch/s390/kernel/asm-offsets.c
@@ -130,6 +130,8 @@ int main(void)
 	DEFINE(__LC_INT_CLOCK, offsetof(struct _lowcore, int_clock));
 	DEFINE(__LC_MACHINE_FLAGS, offsetof(struct _lowcore, machine_flags));
 	DEFINE(__LC_FTRACE_FUNC, offsetof(struct _lowcore, ftrace_func));
+	DEFINE(__LC_KERNEL_PER_TRACING_MASK,
+			offsetof(struct _lowcore, kernel_per_tracing_mask));
 	DEFINE(__LC_IRB, offsetof(struct _lowcore, irb));
 	DEFINE(__LC_CPU_TIMER_SAVE_AREA, offsetof(struct _lowcore, cpu_timer_save_area));
 	DEFINE(__LC_CLOCK_COMP_SAVE_AREA, offsetof(struct _lowcore, clock_comp_save_area));
diff --git a/arch/s390/kernel/entry.S b/arch/s390/kernel/entry.S
index 6af7045..2f93424 100644
--- a/arch/s390/kernel/entry.S
+++ b/arch/s390/kernel/entry.S
@@ -266,6 +266,10 @@ sysc_do_restart:
 	sth	%r7,SP_SVCNR(%r15)
 	sll	%r7,2		  # svc number *4
 	l	%r8,BASED(.Lsysc_table)
+	tm	__LC_KERNEL_PER_TRACING_MASK,0x40  # kernel per tracing is on?
+	jz	sysc_noper
+	stosm	__SF_EMPTY(%r15),0x40	# enable per event recording
+sysc_noper:
 	tm	__TI_flags+2(%r9),_TIF_SYSCALL
 	l	%r8,0(%r7,%r8)	  # get system call addr.
 	bnz	BASED(sysc_tracesys)
@@ -604,6 +608,10 @@ io_int_handler:
 io_no_vtime:
 	l	%r9,__LC_THREAD_INFO	# load pointer to thread_info struct
 	TRACE_IRQS_OFF
+	tm	__LC_KERNEL_PER_TRACING_MASK,0x40  # kernel per tracing is on?
+	jz	io_noper
+	stosm	__SF_EMPTY(%r15),0x40	# enable per event recording
+io_noper:
 	l	%r1,BASED(.Ldo_IRQ)	# load address of do_IRQ
 	la	%r2,SP_PTREGS(%r15)	# address of register-save area
 	basr	%r14,%r1		# branch to standard irq handler
@@ -750,6 +758,10 @@ ext_int_handler:
 ext_no_vtime:
 	l	%r9,__LC_THREAD_INFO	# load pointer to thread_info struct
 	TRACE_IRQS_OFF
+	tm	__LC_KERNEL_PER_TRACING_MASK,0x40  # kernel per tracing is on?
+	jz	ext_noper
+	stosm	__SF_EMPTY(%r15),0x40	# enable per event recording
+ext_noper:
 	la	%r2,SP_PTREGS(%r15)	# address of register-save area
 	lh	%r3,__LC_EXT_INT_CODE	# get interruption code
 	l	%r1,BASED(.Ldo_extint)
diff --git a/arch/s390/kernel/entry64.S b/arch/s390/kernel/entry64.S
index 52106d5..adfb985 100644
--- a/arch/s390/kernel/entry64.S
+++ b/arch/s390/kernel/entry64.S
@@ -260,6 +260,10 @@ sysc_do_restart:
 	larl	%r10,sys_call_table_emu  # use 31 bit emulation system calls
 sysc_noemu:
 #endif
+	tm	__LC_KERNEL_PER_TRACING_MASK,0x40  # kernel per tracing is on?
+	jz	sysc_noper
+	stosm	__SF_EMPTY(%r15),0x40	# enable per event recording
+sysc_noper:
 	tm	__TI_flags+6(%r9),_TIF_SYSCALL
 	lgf	%r8,0(%r7,%r10) # load address of system call routine
 	jnz	sysc_tracesys
@@ -579,6 +583,10 @@ io_no_vtime:
 	lg	%r9,__LC_THREAD_INFO	# load pointer to thread_info struct
 	TRACE_IRQS_OFF
 	la	%r2,SP_PTREGS(%r15)	# address of register-save area
+	tm	__LC_KERNEL_PER_TRACING_MASK,0x40  # kernel per tracing is on?
+	jz	io_noper
+	stosm	__SF_EMPTY(%r15),0x40	# enable per event recording
+io_noper:
 	brasl	%r14,do_IRQ		# call standard irq handler
 io_return:
 	tm	__TI_flags+7(%r9),_TIF_WORK_INT
@@ -744,6 +752,10 @@ ext_no_vtime:
 	TRACE_IRQS_OFF
 	la	%r2,SP_PTREGS(%r15)	# address of register-save area
 	llgh	%r3,__LC_EXT_INT_CODE	# get interruption code
+	tm	__LC_KERNEL_PER_TRACING_MASK,0x40  # kernel per tracing is on?
+	jz	ext_noper
+	stosm	__SF_EMPTY(%r15),0x40	# enable per event recording
+ext_noper:
 	brasl	%r14,do_extint
 	j	io_return
 
diff --git a/arch/s390/kernel/hw_breakpoint.c b/arch/s390/kernel/hw_breakpoint.c
index 1cce08e..19c5bb5 100644
--- a/arch/s390/kernel/hw_breakpoint.c
+++ b/arch/s390/kernel/hw_breakpoint.c
@@ -82,6 +82,19 @@ int arch_install_hw_breakpoint(struct perf_event *bp)
 	per_regs[0].starting_addr = info->address;
 	per_regs[0].ending_addr = info->address + info->len - 1;
 
+	if (!bp->attr.exclude_kernel
+		&& per_regs[0].em_storage_alteration) {
+		/* For kernel wide breakpoints the Storage-Alteration-Space
+		 * Control bit as well as the Storage-Alternation-Event bit
+		 * in the ASCE needs to be set. Otherwise we might end up
+		 * tracing copy_to_user events as well.
+		 */
+		per_regs[0].storage_alt_space_ctl = 1;
+		__ctl_set_bit(1, 7);
+		__ctl_set_bit(7, 7);
+		__ctl_set_bit(13, 7);
+	}
+
 	/* Load the control register 9, 10 and 11 with per info */
 	__ctl_load(per_regs, 9, 11);
 	__get_cpu_var(cpu_per_regs[0]) = per_regs[0];
@@ -93,13 +106,7 @@ int arch_install_hw_breakpoint(struct perf_event *bp)
 		 */
 		if (!bp->attr.exclude_kernel) {
 			/* Enable wide breakpoint in the kernel */
-			/* FIXME:
-			 * It's not good idea to use existing flag in lowcore
-			 * for turning on/off PER tracing in kernel. instead
-			 * define a new flag and handle PER tracing checks in
-			 * entry*.S
-			 */
-
+			S390_lowcore.kernel_per_tracing_mask |= PSW_MASK_PER;
 
 			/* set PER bit int psw_kernel_bits to avoid loosing it
 			 * accidently if someone modifies PSW bit directly.
@@ -148,7 +155,14 @@ void arch_uninstall_hw_breakpoint(struct perf_event *bp)
 
 	if (!bp->attr.exclude_kernel) {
 		/* clear wide breakpoint in the kernel */
+		S390_lowcore.kernel_per_tracing_mask &= ~PSW_MASK_PER;
 		psw_kernel_bits &= ~PSW_MASK_PER;
+		per_regs[0] = __get_cpu_var(cpu_per_regs[0]);
+		if (per_regs[0].em_storage_alteration) {
+			__ctl_clear_bit(1, 7);
+			__ctl_clear_bit(7, 7);
+			__ctl_clear_bit(13, 7);
+		}
 	}
 	if (tsk) {
 		/* clear breakpoint bound to a task context */

Re: [patch 2/3] S390-HWBKPT v5: Enabling s390 kernel for kernel

[Heiko Carstens]

On Thu, May 27, 2010 at 09:13:01AM +0530, Mahesh Salgaonkar wrote:
> --- a/arch/s390/kernel/hw_breakpoint.c
> +++ b/arch/s390/kernel/hw_breakpoint.c
> @@ -82,6 +82,19 @@ int arch_install_hw_breakpoint(struct perf_event *bp)
>  	per_regs[0].starting_addr = info->address;
>  	per_regs[0].ending_addr = info->address + info->len - 1;
> 
> +	if (!bp->attr.exclude_kernel
> +		&& per_regs[0].em_storage_alteration) {
> +		/* For kernel wide breakpoints the Storage-Alteration-Space
> +		 * Control bit as well as the Storage-Alternation-Event bit
> +		 * in the ASCE needs to be set. Otherwise we might end up
> +		 * tracing copy_to_user events as well.
> +		 */
> +		per_regs[0].storage_alt_space_ctl = 1;
> +		__ctl_set_bit(1, 7);
> +		__ctl_set_bit(7, 7);
> +		__ctl_set_bit(13, 7);
> +	}

If !exclude_kernel has the same meaning as "only_kernel" then this could work
if the Storage-Alteration-Event would only be set in the kernel space ASCE.

The corresponding bit in the user space ASCE would be lost on next context
switch. But since that is hopefully irrelevant that shouldn't be a problem.