[PATCH] fast vdso implementation for CLOCK_THREAD_CPUTIME_ID

The extract cpu time instruction (ectg) instruction allows the user
process to get the current thread cputime without calling into the
kernel. The code that uses the instruction needs to switch to the
access registers mode to get access to the per-cpu info page that
contains the two base values that are needed to calculate the current
cputime from the CPU timer with the ectg instruction.

Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>

arch/s390/include/asm/lowcore.h

@@ -106,8 +106,10 @@
 #define __LC_IPLDEV                     0xDB8
 #define __LC_CURRENT			0xDD8
 #define __LC_INT_CLOCK			0xDE8
+#define __LC_VDSO_PER_CPU		0xE38
 #endif /* __s390x__ */
 
+#define __LC_PASTE			0xE40
 
 #define __LC_PANIC_MAGIC		0xE00
 #ifndef __s390x__
@@ -381,7 +383,12 @@ struct _lowcore
         /* whether the kernel died with panic() or not */
         __u32        panic_magic;              /* 0xe00 */
 
-	__u8         pad13[0x11b8-0xe04];      /* 0xe04 */
+	/* Per cpu primary space access list */
+	__u8	     pad_0xe04[0xe3c-0xe04];   /* 0xe04 */
+	__u32	     vdso_per_cpu_data;	       /* 0xe3c */
+	__u32	     paste[16];		       /* 0xe40 */
+
+	__u8	     pad13[0x11b8-0xe80];      /* 0xe80 */
 
 	/* 64 bit extparam used for pfault, diag 250 etc  */
 	__u64        ext_params2;               /* 0x11B8 */

arch/s390/include/asm/vdso.h

@@ -12,9 +12,9 @@
 #ifndef __ASSEMBLY__
 
 /*
- * Note about this structure:
+ * Note about the vdso_data and vdso_per_cpu_data structures:
  *
- * NEVER USE THIS IN USERSPACE CODE DIRECTLY. The layout of this
+ * NEVER USE THEM IN USERSPACE CODE DIRECTLY. The layout of the
  * structure is supposed to be known only to the function in the vdso
  * itself and may change without notice.
  */
@@ -28,10 +28,21 @@ struct vdso_data {
 	__u64 wtom_clock_nsec;		/*				0x28 */
 	__u32 tz_minuteswest;		/* Minutes west of Greenwich	0x30 */
 	__u32 tz_dsttime;		/* Type of dst correction	0x34 */
+	__u32 ectg_available;
+};
+
+struct vdso_per_cpu_data {
+	__u64 ectg_timer_base;
+	__u64 ectg_user_time;
 };
 
 extern struct vdso_data *vdso_data;
 
+#ifdef CONFIG_64BIT
+int vdso_alloc_per_cpu(int cpu, struct _lowcore *lowcore);
+void vdso_free_per_cpu(int cpu, struct _lowcore *lowcore);
+#endif
+
 #endif /* __ASSEMBLY__ */
 
 #endif /* __KERNEL__ */

arch/s390/kernel/asm-offsets.c

@@ -48,6 +48,11 @@ int main(void)
 	DEFINE(__VDSO_WTOM_SEC, offsetof(struct vdso_data, wtom_clock_sec));
 	DEFINE(__VDSO_WTOM_NSEC, offsetof(struct vdso_data, wtom_clock_nsec));
 	DEFINE(__VDSO_TIMEZONE, offsetof(struct vdso_data, tz_minuteswest));
+	DEFINE(__VDSO_ECTG_OK, offsetof(struct vdso_data, ectg_available));
+	DEFINE(__VDSO_ECTG_BASE,
+	       offsetof(struct vdso_per_cpu_data, ectg_timer_base));
+	DEFINE(__VDSO_ECTG_USER,
+	       offsetof(struct vdso_per_cpu_data, ectg_user_time));
 	/* constants used by the vdso */
 	DEFINE(CLOCK_REALTIME, CLOCK_REALTIME);
 	DEFINE(CLOCK_MONOTONIC, CLOCK_MONOTONIC);

arch/s390/kernel/entry64.S

@@ -177,8 +177,11 @@ _TIF_WORK_INT = (_TIF_SIGPENDING | _TIF_NOTIFY_RESUME | _TIF_NEED_RESCHED | \
 	.if !\sync
 	ni	\psworg+1,0xfd		# clear wait state bit
 	.endif
-	lmg	%r0,%r15,SP_R0(%r15)	# load gprs 0-15 of user
+	lg	%r14,__LC_VDSO_PER_CPU
+	lmg	%r0,%r13,SP_R0(%r15)	# load gprs 0-13 of user
 	stpt	__LC_EXIT_TIMER
+	mvc	__VDSO_ECTG_BASE(16,%r14),__LC_EXIT_TIMER
+	lmg	%r14,%r15,SP_R14(%r15)	# load grps 14-15 of user
 	lpswe	\psworg			# back to caller
 	.endm
 
@@ -980,23 +983,23 @@ cleanup_sysc_return:
 
 cleanup_sysc_leave:
 	clc	8(8,%r12),BASED(cleanup_sysc_leave_insn)
-	je	2f
-	mvc	__LC_EXIT_TIMER(8),__LC_ASYNC_ENTER_TIMER
+	je	3f
 	clc	8(8,%r12),BASED(cleanup_sysc_leave_insn+8)
-	je	2f
-	mvc	__LC_RETURN_PSW(16),SP_PSW(%r15)
+	jhe	0f
+	mvc	__LC_EXIT_TIMER(8),__LC_ASYNC_ENTER_TIMER
+0:	mvc	__LC_RETURN_PSW(16),SP_PSW(%r15)
 	cghi	%r12,__LC_MCK_OLD_PSW
-	jne	0f
+	jne	1f
 	mvc	__LC_SAVE_AREA+64(32),SP_R12(%r15)
-	j	1f
-0:	mvc	__LC_SAVE_AREA+32(32),SP_R12(%r15)
-1:	lmg	%r0,%r11,SP_R0(%r15)
+	j	2f
+1:	mvc	__LC_SAVE_AREA+32(32),SP_R12(%r15)
+2:	lmg	%r0,%r11,SP_R0(%r15)
 	lg	%r15,SP_R15(%r15)
-2:	la	%r12,__LC_RETURN_PSW
+3:	la	%r12,__LC_RETURN_PSW
 	br	%r14
 cleanup_sysc_leave_insn:
 	.quad	sysc_done - 4
-	.quad	sysc_done - 8
+	.quad	sysc_done - 16
 
 cleanup_io_return:
 	mvc	__LC_RETURN_PSW(8),0(%r12)
@@ -1006,23 +1009,23 @@ cleanup_io_return:
 
 cleanup_io_leave:
 	clc	8(8,%r12),BASED(cleanup_io_leave_insn)
-	je	2f
-	mvc	__LC_EXIT_TIMER(8),__LC_ASYNC_ENTER_TIMER
+	je	3f
 	clc	8(8,%r12),BASED(cleanup_io_leave_insn+8)
-	je	2f
-	mvc	__LC_RETURN_PSW(16),SP_PSW(%r15)
+	jhe	0f
+	mvc	__LC_EXIT_TIMER(8),__LC_ASYNC_ENTER_TIMER
+0:	mvc	__LC_RETURN_PSW(16),SP_PSW(%r15)
 	cghi	%r12,__LC_MCK_OLD_PSW
-	jne	0f
+	jne	1f
 	mvc	__LC_SAVE_AREA+64(32),SP_R12(%r15)
-	j	1f
-0:	mvc	__LC_SAVE_AREA+32(32),SP_R12(%r15)
-1:	lmg	%r0,%r11,SP_R0(%r15)
+	j	2f
+1:	mvc	__LC_SAVE_AREA+32(32),SP_R12(%r15)
+2:	lmg	%r0,%r11,SP_R0(%r15)
 	lg	%r15,SP_R15(%r15)
-2:	la	%r12,__LC_RETURN_PSW
+3:	la	%r12,__LC_RETURN_PSW
 	br	%r14
 cleanup_io_leave_insn:
 	.quad	io_done - 4
-	.quad	io_done - 8
+	.quad	io_done - 16
 
 /*
  * Integer constants

arch/s390/kernel/head64.S

@@ -87,6 +87,8 @@ startup_continue:
 	lg	%r12,.Lparmaddr-.LPG1(%r13)	# pointer to parameter area
 					# move IPL device to lowcore
 	mvc	__LC_IPLDEV(4),IPL_DEVICE+4-PARMAREA(%r12)
+	lghi	%r0,__LC_PASTE
+	stg	%r0,__LC_VDSO_PER_CPU
 #
 # Setup stack
 #

arch/s390/kernel/setup.c

@@ -427,6 +427,8 @@ setup_lowcore(void)
 		/* enable extended save area */
 		__ctl_set_bit(14, 29);
 	}
+#else
+	lc->vdso_per_cpu_data = (unsigned long) &lc->paste[0];
 #endif
 	set_prefix((u32)(unsigned long) lc);
 }

arch/s390/kernel/smp.c

@@ -47,6 +47,7 @@
 #include <asm/lowcore.h>
 #include <asm/sclp.h>
 #include <asm/cpu.h>
+#include <asm/vdso.h>
 #include "entry.h"
 
 /*
@@ -506,6 +507,9 @@ static int __cpuinit smp_alloc_lowcore(int cpu)
 			goto out;
 		lowcore->extended_save_area_addr = (u32) save_area;
 	}
+#else
+	if (vdso_alloc_per_cpu(cpu, lowcore))
+		goto out;
 #endif
 	lowcore_ptr[cpu] = lowcore;
 	return 0;
@@ -528,6 +532,8 @@ static void smp_free_lowcore(int cpu)
 #ifndef CONFIG_64BIT
 	if (MACHINE_HAS_IEEE)
 		free_page((unsigned long) lowcore->extended_save_area_addr);
+#else
+	vdso_free_per_cpu(cpu, lowcore);
 #endif
 	free_page(lowcore->panic_stack - PAGE_SIZE);
 	free_pages(lowcore->async_stack - ASYNC_SIZE, ASYNC_ORDER);
@@ -670,6 +676,7 @@ void __init smp_prepare_cpus(unsigned int max_cpus)
 	lowcore = (void *) __get_free_pages(GFP_KERNEL | GFP_DMA, lc_order);
 	panic_stack = __get_free_page(GFP_KERNEL);
 	async_stack = __get_free_pages(GFP_KERNEL, ASYNC_ORDER);
+	BUG_ON(!lowcore || !panic_stack || !async_stack);
 #ifndef CONFIG_64BIT
 	if (MACHINE_HAS_IEEE)
 		save_area = get_zeroed_page(GFP_KERNEL);
@@ -683,6 +690,8 @@ void __init smp_prepare_cpus(unsigned int max_cpus)
 #ifndef CONFIG_64BIT
 	if (MACHINE_HAS_IEEE)
 		lowcore->extended_save_area_addr = (u32) save_area;
+#else
+	BUG_ON(vdso_alloc_per_cpu(smp_processor_id(), lowcore));
 #endif
 	set_prefix((u32)(unsigned long) lowcore);
 	local_mcck_enable();

arch/s390/kernel/vdso.c

@@ -31,9 +31,6 @@
 #include <asm/sections.h>
 #include <asm/vdso.h>
 
-/* Max supported size for symbol names */
-#define MAX_SYMNAME	64
-
 #if defined(CONFIG_32BIT) || defined(CONFIG_COMPAT)
 extern char vdso32_start, vdso32_end;
 static void *vdso32_kbase = &vdso32_start;
@@ -70,6 +67,119 @@ static union {
 } vdso_data_store __attribute__((__section__(".data.page_aligned")));
 struct vdso_data *vdso_data = &vdso_data_store.data;
 
+/*
+ * Setup vdso data page.
+ */
+static void vdso_init_data(struct vdso_data *vd)
+{
+	unsigned int facility_list;
+
+	facility_list = stfl();
+	vd->ectg_available = switch_amode && (facility_list & 1);
+}
+
+#ifdef CONFIG_64BIT
+/*
+ * Setup per cpu vdso data page.
+ */
+static void vdso_init_per_cpu_data(int cpu, struct vdso_per_cpu_data *vpcd)
+{
+}
+
+/*
+ * Allocate/free per cpu vdso data.
+ */
+#ifdef CONFIG_64BIT
+#define SEGMENT_ORDER	2
+#else
+#define SEGMENT_ORDER	1
+#endif
+
+int vdso_alloc_per_cpu(int cpu, struct _lowcore *lowcore)
+{
+	unsigned long segment_table, page_table, page_frame;
+	u32 *psal, *aste;
+	int i;
+
+	lowcore->vdso_per_cpu_data = __LC_PASTE;
+
+	if (!switch_amode || !vdso_enabled)
+		return 0;
+
+	segment_table = __get_free_pages(GFP_KERNEL, SEGMENT_ORDER);
+	page_table = get_zeroed_page(GFP_KERNEL | GFP_DMA);
+	page_frame = get_zeroed_page(GFP_KERNEL);
+	if (!segment_table || !page_table || !page_frame)
+		goto out;
+
+	clear_table((unsigned long *) segment_table, _SEGMENT_ENTRY_EMPTY,
+		    PAGE_SIZE << SEGMENT_ORDER);
+	clear_table((unsigned long *) page_table, _PAGE_TYPE_EMPTY,
+		    256*sizeof(unsigned long));
+
+	*(unsigned long *) segment_table = _SEGMENT_ENTRY + page_table;
+	*(unsigned long *) page_table = _PAGE_RO + page_frame;
+
+	psal = (u32 *) (page_table + 256*sizeof(unsigned long));
+	aste = psal + 32;
+
+	for (i = 4; i < 32; i += 4)
+		psal[i] = 0x80000000;
+
+	lowcore->paste[4] = (u32)(addr_t) psal;
+	psal[0] = 0x20000000;
+	psal[2] = (u32)(addr_t) aste;
+	*(unsigned long *) (aste + 2) = segment_table +
+		_ASCE_TABLE_LENGTH + _ASCE_USER_BITS + _ASCE_TYPE_SEGMENT;
+	aste[4] = (u32)(addr_t) psal;
+	lowcore->vdso_per_cpu_data = page_frame;
+
+	vdso_init_per_cpu_data(cpu, (struct vdso_per_cpu_data *) page_frame);
+	return 0;
+
+out:
+	free_page(page_frame);
+	free_page(page_table);
+	free_pages(segment_table, SEGMENT_ORDER);
+	return -ENOMEM;
+}
+
+#ifdef CONFIG_HOTPLUG_CPU
+void vdso_free_per_cpu(int cpu, struct _lowcore *lowcore)
+{
+	unsigned long segment_table, page_table, page_frame;
+	u32 *psal, *aste;
+
+	if (!switch_amode || !vdso_enabled)
+		return;
+
+	psal = (u32 *)(addr_t) lowcore->paste[4];
+	aste = (u32 *)(addr_t) psal[2];
+	segment_table = *(unsigned long *)(aste + 2) & PAGE_MASK;
+	page_table = *(unsigned long *) segment_table;
+	page_frame = *(unsigned long *) page_table;
+
+	free_page(page_frame);
+	free_page(page_table);
+	free_pages(segment_table, SEGMENT_ORDER);
+}
+#endif /* CONFIG_HOTPLUG_CPU */
+
+static void __vdso_init_cr5(void *dummy)
+{
+	unsigned long cr5;
+
+	cr5 = offsetof(struct _lowcore, paste);
+	__ctl_load(cr5, 5, 5);
+}
+
+static void vdso_init_cr5(void)
+{
+	if (switch_amode && vdso_enabled)
+		on_each_cpu(__vdso_init_cr5, NULL, 1);
+}
+#endif /* CONFIG_64BIT */
+
 /*
  * This is called from binfmt_elf, we create the special vma for the
  * vDSO and insert it into the mm struct tree
@@ -172,6 +282,9 @@ static int __init vdso_init(void)
 {
 	int i;
 
+	if (!vdso_enabled)
+		return 0;
+	vdso_init_data(vdso_data);
 #if defined(CONFIG_32BIT) || defined(CONFIG_COMPAT)
 	/* Calculate the size of the 32 bit vDSO */
 	vdso32_pages = ((&vdso32_end - &vdso32_start
@@ -208,6 +321,10 @@ static int __init vdso_init(void)
 	}
 	vdso64_pagelist[vdso64_pages - 1] = virt_to_page(vdso_data);
 	vdso64_pagelist[vdso64_pages] = NULL;
+#ifndef CONFIG_SMP
+	BUG_ON(vdso_alloc_per_cpu(0, S390_lowcore));
+#endif
+	vdso_init_cr5();
 #endif /* CONFIG_64BIT */
 
 	get_page(virt_to_page(vdso_data));

arch/s390/kernel/vdso64/clock_getres.S

@@ -22,7 +22,12 @@ __kernel_clock_getres:
 	cghi	%r2,CLOCK_REALTIME
 	je	0f
 	cghi	%r2,CLOCK_MONOTONIC
+	je	0f
+	cghi	%r2,-2		/* CLOCK_THREAD_CPUTIME_ID for this thread */
 	jne	2f
+	larl	%r5,_vdso_data
+	icm	%r0,15,__LC_ECTG_OK(%r5)
+	jz	2f
 0:	ltgr	%r3,%r3
 	jz	1f				/* res == NULL */
 	larl	%r1,3f

arch/s390/kernel/vdso64/clock_gettime.S

@@ -22,8 +22,10 @@ __kernel_clock_gettime:
 	larl	%r5,_vdso_data
 	cghi	%r2,CLOCK_REALTIME
 	je	4f
+	cghi	%r2,-2		/* CLOCK_THREAD_CPUTIME_ID for this thread */
+	je	9f
 	cghi	%r2,CLOCK_MONOTONIC
-	jne	9f
+	jne	12f
 
 	/* CLOCK_MONOTONIC */
 	ltgr	%r3,%r3
@@ -42,7 +44,7 @@ __kernel_clock_gettime:
 	alg	%r0,__VDSO_WTOM_SEC(%r5)
 	clg	%r4,__VDSO_UPD_COUNT(%r5)	/* check update counter */
 	jne	0b
-	larl	%r5,10f
+	larl	%r5,13f
 1:	clg	%r1,0(%r5)
 	jl	2f
 	slg	%r1,0(%r5)
@@ -68,7 +70,7 @@ __kernel_clock_gettime:
 	lg	%r0,__VDSO_XTIME_SEC(%r5)
 	clg	%r4,__VDSO_UPD_COUNT(%r5)	/* check update counter */
 	jne	5b
-	larl	%r5,10f
+	larl	%r5,13f
 6:	clg	%r1,0(%r5)
 	jl	7f
 	slg	%r1,0(%r5)
@@ -79,11 +81,38 @@ __kernel_clock_gettime:
 8:	lghi	%r2,0
 	br	%r14
 
+	/* CLOCK_THREAD_CPUTIME_ID for this thread */
+9:	icm	%r0,15,__VDSO_ECTG_OK(%r5)
+	jz	12f
+	ear	%r2,%a4
+	llilh	%r4,0x0100
+	sar	%a4,%r4
+	lghi	%r4,0
+	sacf	512				/* Magic ectg instruction */
+	.insn	ssf,0xc80100000000,__VDSO_ECTG_BASE(4),__VDSO_ECTG_USER(4),4
+	sacf	0
+	sar	%a4,%r2
+	algr	%r1,%r0				/* r1 = cputime as TOD value */
+	mghi	%r1,1000			/* convert to nanoseconds */
+	srlg	%r1,%r1,12			/* r1 = cputime in nanosec */
+	lgr	%r4,%r1
+	larl	%r5,13f
+	srlg	%r1,%r1,9			/* divide by 1000000000 */
+	mlg	%r0,8(%r5)
+	srlg	%r0,%r0,11			/* r0 = tv_sec */
+	stg	%r0,0(%r3)
+	msg	%r0,0(%r5)			/* calculate tv_nsec */
+	slgr	%r4,%r0				/* r4 = tv_nsec */
+	stg	%r4,8(%r3)
+	lghi	%r2,0
+	br	%r14
+
 	/* Fallback to system call */
-9:	lghi	%r1,__NR_clock_gettime
+12:	lghi	%r1,__NR_clock_gettime
 	svc	0
 	br	%r14
 
-10:	.quad	1000000000
+13:	.quad	1000000000
+14:	.quad	19342813113834067
 	.cfi_endproc
 	.size	__kernel_clock_gettime,.-__kernel_clock_gettime