[PATCH] ia64: switch to perfmon2

This patch contains a major rewrite of the perfmon subsystem to
bring it to version 2.0. This version is NOT compatible with the
existing perfmon-1.x version which was in 2.5 and still is in 2.4
kernels. This new codebase brings a lot of new features including
the ability to attach to already running tasks, the ability the
follow clone2, the ability to write your own sampling buffer format
via kernel modules. It is also much more robust than its 1.x
counter-part. This version supports the Itanium, McKinley and
Madison PMUs. This is beta quality code and extensions to the
interface are planned.

arch/ia64/kernel/Makefile

@@ -5,8 +5,8 @@
 extra-y	:= head.o init_task.o
 
 obj-y := acpi.o entry.o efi.o efi_stub.o gate.o ia64_ksyms.o irq.o irq_ia64.o irq_lsapic.o	\
-	 ivt.o machvec.o pal.o perfmon.o process.o ptrace.o sal.o semaphore.o setup.o signal.o	\
-	 sys_ia64.o time.o traps.o unaligned.o unwind.o
+	 ivt.o machvec.o pal.o perfmon.o perfmon_default_smpl.o process.o ptrace.o sal.o \
+	 semaphore.o setup.o signal.o sys_ia64.o time.o traps.o unaligned.o unwind.o
 
 obj-$(CONFIG_EFI_VARS)		+= efivars.o
 obj-$(CONFIG_FSYS)		+= fsys.o

arch/ia64/kernel/fsys.S

@@ -376,7 +376,8 @@ GLOBAL_ENTRY(fsys_bubble_down)
 	 *	- r29: psr
 	 */
 #	define PSR_PRESERVED_BITS	(IA64_PSR_UP | IA64_PSR_MFL | IA64_PSR_MFH | IA64_PSR_PK \
-					 | IA64_PSR_DT | IA64_PSR_PP | IA64_PSR_RT | IA64_PSR_IC)
+					 | IA64_PSR_DT | IA64_PSR_PP | IA64_PSR_SP | IA64_PSR_RT \
+					 | IA64_PSR_IC)
 	/*
 	 * Reading psr.l gives us only bits 0-31, psr.it, and psr.mc.  The rest we have
 	 * to synthesize.

arch/ia64/kernel/ia64_ksyms.c

@@ -162,8 +162,11 @@ EXPORT_SYMBOL(machvec_noop);
 EXPORT_SYMBOL(zero_page_memmap_ptr);
 #ifdef CONFIG_PERFMON
 #include <asm/perfmon.h>
-EXPORT_SYMBOL(pfm_install_alternate_syswide_subsystem);
-EXPORT_SYMBOL(pfm_remove_alternate_syswide_subsystem);
+EXPORT_SYMBOL(pfm_register_buffer_fmt);
+EXPORT_SYMBOL(pfm_unregister_buffer_fmt);
+EXPORT_SYMBOL(pfm_mod_fast_read_pmds);
+EXPORT_SYMBOL(pfm_mod_read_pmds);
+EXPORT_SYMBOL(pfm_mod_write_pmcs);
 #endif
 
 #ifdef CONFIG_NUMA

arch/ia64/kernel/perfmon_default_smpl.c

+/*
+ * Copyright (C) 2002-2003 Hewlett-Packard Co
+ *               Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This file implements the default sampling buffer format
+ * for the Linux/ia64 perfmon-2 subsystem.
+ */
+#include <linux/kernel.h>
+#include <linux/types.h>
+#include <linux/module.h>
+#include <linux/config.h>
+#include <linux/init.h>
+#include <asm/delay.h>
+#include <linux/smp.h>
+
+#include <asm/perfmon.h>
+#include <asm/perfmon_default_smpl.h>
+
+MODULE_AUTHOR("Stephane Eranian <eranian@hpl.hp.com>");
+MODULE_DESCRIPTION("perfmon default sampling format");
+MODULE_LICENSE("GPL");
+
+MODULE_PARM(debug, "i");
+MODULE_PARM_DESC(debug, "debug");
+
+MODULE_PARM(debug_ovfl, "i");
+MODULE_PARM_DESC(debug_ovfl, "debug ovfl");
+
+
+#define DEFAULT_DEBUG 1
+
+#ifdef DEFAULT_DEBUG
+#define DPRINT(a) \
+	do { \
+		if (unlikely(debug >0)) { printk("%s.%d: CPU%d ", __FUNCTION__, __LINE__, smp_processor_id()); printk a; } \
+	} while (0)
+
+#define DPRINT_ovfl(a) \
+	do { \
+		if (unlikely(debug_ovfl >0)) { printk("%s.%d: CPU%d ", __FUNCTION__, __LINE__, smp_processor_id()); printk a; } \
+	} while (0)
+
+#else
+#define DPRINT(a)
+#define DPRINT_ovfl(a)
+#endif
+
+static int debug, debug_ovfl;
+
+static int
+default_validate(struct task_struct *task, unsigned int flags, int cpu, void *data)
+{
+	pfm_default_smpl_arg_t *arg = (pfm_default_smpl_arg_t*)data;
+	int ret = 0;
+
+	if (data == NULL) {
+		DPRINT(("[%d] no argument passed\n", task->pid));
+		return -EINVAL;
+	}
+
+	DPRINT(("[%d] validate flags=0x%x CPU%d\n", task->pid, flags, cpu));
+
+	/*
+	 * must hold at least the buffer header + one minimally sized entry
+	 */
+	if (arg->buf_size < PFM_DEFAULT_SMPL_MIN_BUF_SIZE) return -EINVAL;
+
+	DPRINT(("buf_size=%lu\n", arg->buf_size));
+
+	return ret;
+}
+
+static int
+default_get_size(struct task_struct *task, unsigned int flags, int cpu, void *data, unsigned long *size)
+{
+	pfm_default_smpl_arg_t *arg = (pfm_default_smpl_arg_t *)data;
+
+	/*
+	 * size has been validated in default_validate
+	 */
+	*size = arg->buf_size;
+
+	return 0;
+}
+
+static int
+default_init(struct task_struct *task, void *buf, unsigned int flags, int cpu, void *data)
+{
+	pfm_default_smpl_hdr_t *hdr;
+	pfm_default_smpl_arg_t *arg = (pfm_default_smpl_arg_t *)data;
+
+	hdr = (pfm_default_smpl_hdr_t *)buf;
+
+	hdr->hdr_version      = PFM_DEFAULT_SMPL_VERSION;
+	hdr->hdr_buf_size     = arg->buf_size;
+	hdr->hdr_cur_pos      = (void *)((unsigned long)buf)+sizeof(*hdr);
+	hdr->hdr_last_pos     = (void *)((unsigned long)buf)+arg->buf_size;
+	hdr->hdr_overflows    = 0UL;
+	hdr->hdr_count        = 0UL;
+
+	DPRINT(("[%d] buffer=%p buf_size=%lu hdr_size=%lu hdr_version=%u\n",
+		task->pid,
+		buf,
+		hdr->hdr_buf_size,
+		sizeof(*hdr),
+		hdr->hdr_version));
+
+	return 0;
+}
+
+static int
+default_handler(struct task_struct *task, void *buf, pfm_ovfl_arg_t *arg, struct pt_regs *regs)
+{
+	pfm_default_smpl_hdr_t *hdr;
+	pfm_default_smpl_entry_t *ent;
+	void *cur, *last;
+	unsigned long *e;
+	unsigned long ovfl_mask;
+	unsigned long ovfl_notify;
+	unsigned long stamp;
+	unsigned int npmds, i;
+
+	/*
+	 * some time stamp
+	 */
+	stamp = ia64_get_itc();
+
+	if (unlikely(buf == NULL || arg == NULL|| regs == NULL || task == NULL)) {
+		DPRINT(("[%d] invalid arguments buf=%p arg=%p\n", task->pid, buf, arg));
+		return -EINVAL;
+	}
+
+	hdr         = (pfm_default_smpl_hdr_t *)buf;
+	cur         = hdr->hdr_cur_pos;
+	last        = hdr->hdr_last_pos;
+	ovfl_mask   = arg->ovfl_pmds[0];
+	ovfl_notify = arg->ovfl_notify[0];
+
+	/*
+	 * check for space against largest possibly entry.
+	 * We may waste space at the end of the buffer.
+	 */
+	if ((last - cur) < PFM_DEFAULT_MAX_ENTRY_SIZE) goto full;
+
+	npmds = hweight64(arg->smpl_pmds[0]);
+
+	ent = (pfm_default_smpl_entry_t *)cur;
+
+	prefetch(arg->smpl_pmds_values);
+
+	/* position for first pmd */
+	e = (unsigned long *)(ent+1);
+
+	hdr->hdr_count++;
+
+	DPRINT_ovfl(("[%d] count=%lu cur=%p last=%p free_bytes=%lu ovfl_pmds=0x%lx ovfl_notify=0x%lx npmds=%u\n",
+			task->pid,
+			hdr->hdr_count,
+			cur, last,
+			last-cur,
+			ovfl_mask,
+			ovfl_notify, npmds));
+
+	/*
+	 * current = task running at the time of the overflow.
+	 *
+	 * per-task mode:
+	 * 	- this is ususally the task being monitored.
+	 * 	  Under certain conditions, it might be a different task
+	 *
+	 * system-wide:
+	 * 	- this is not necessarily the task controlling the session
+	 */
+	ent->pid            = current->pid;
+	ent->cpu            = smp_processor_id();
+	ent->last_reset_val = arg->pmd_last_reset; //pmd[0].reg_last_reset_val;
+
+	/*
+	 * where did the fault happen (includes slot number)
+	 */
+	ent->ip = regs->cr_iip | ((regs->cr_ipsr >> 41) & 0x3);
+
+	/*
+	 * which registers overflowed
+	 */
+	ent->ovfl_pmds = ovfl_mask;
+	ent->tstamp    = stamp;
+	ent->set       = arg->active_set;
+	ent->reserved1 = 0;
+
+	/*
+	 * selectively store PMDs in increasing index number
+	 */
+	if (npmds) {
+		unsigned long *val = arg->smpl_pmds_values;
+		for(i=0; i < npmds; i++) {
+			*e++ = *val++;
+		}
+	}
+
+	/*
+	 * update position for next entry
+	 */
+	hdr->hdr_cur_pos = cur + sizeof(*ent) + (npmds << 3);
+
+	/*
+	 * keep same ovfl_pmds, ovfl_notify
+	 */
+	arg->ovfl_ctrl.notify_user     = 0;
+	arg->ovfl_ctrl.block           = 0;
+	arg->ovfl_ctrl.stop_monitoring = 0;
+	arg->ovfl_ctrl.reset_pmds      = 1;
+
+	return 0;
+full:
+	DPRINT_ovfl(("sampling buffer full free=%lu, count=%lu, ovfl_notify=0x%lx\n", last-cur, hdr->hdr_count, ovfl_notify));
+
+	/*
+	 * increment number of buffer overflow.
+	 * important to detect duplicate set of samples.
+	 */
+	hdr->hdr_overflows++;
+
+	/*
+	 * if no notification is needed, then we just reset the buffer index.
+	 */
+	if (ovfl_notify == 0UL) {
+		hdr->hdr_count = 0UL;
+		arg->ovfl_ctrl.notify_user     = 0;
+		arg->ovfl_ctrl.block           = 0;
+		arg->ovfl_ctrl.stop_monitoring = 0;
+		arg->ovfl_ctrl.reset_pmds      = 1;
+	} else {
+		/* keep same ovfl_pmds, ovfl_notify */
+		arg->ovfl_ctrl.notify_user     = 1;
+		arg->ovfl_ctrl.block           = 1;
+		arg->ovfl_ctrl.stop_monitoring = 1;
+		arg->ovfl_ctrl.reset_pmds      = 0;
+	}
+	return 0;
+}
+
+static int
+default_restart(struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs)
+{
+	pfm_default_smpl_hdr_t *hdr;
+
+	hdr = (pfm_default_smpl_hdr_t *)buf;
+
+	hdr->hdr_count   = 0UL;
+	hdr->hdr_cur_pos = (void *)((unsigned long)buf)+sizeof(*hdr);
+
+	ctrl->stop_monitoring = 0;
+	ctrl->reset_pmds      = PFM_PMD_LONG_RESET;
+
+	return 0;
+}
+
+static int
+default_exit(struct task_struct *task, void *buf, struct pt_regs *regs)
+{
+	DPRINT(("[%d] exit(%p)\n", task->pid, buf));
+	return 0;
+}
+
+static pfm_buffer_fmt_t default_fmt={
+	.fmt_name 	= "default_format",
+	.fmt_uuid	= PFM_DEFAULT_SMPL_UUID,
+	.fmt_arg_size	= sizeof(pfm_default_smpl_arg_t),
+	.fmt_validate	= default_validate,
+	.fmt_getsize	= default_get_size,
+	.fmt_init	= default_init,
+	.fmt_handler	= default_handler,
+	.fmt_restart	= default_restart,
+	.fmt_exit	= default_exit,
+};
+
+static int __init
+pfm_default_smpl_init_module(void)
+{
+	int ret;
+
+	ret = pfm_register_buffer_fmt(&default_fmt);
+	if (ret == 0) {
+		printk("perfmon_default_smpl: %s v%u.%u registered\n",
+			default_fmt.fmt_name,
+			PFM_DEFAULT_SMPL_VERSION_MAJ,
+			PFM_DEFAULT_SMPL_VERSION_MIN);
+	} else {
+		printk("perfmon_default_smpl: %s cannot register ret=%d\n",
+			default_fmt.fmt_name,
+			ret);
+	}
+
+	return ret;
+}
+
+static void __exit
+pfm_default_smpl_cleanup_module(void)
+{
+	int ret;
+	ret = pfm_unregister_buffer_fmt(default_fmt.fmt_uuid);
+
+	printk("perfmon_default_smpl: unregister %s=%d\n", default_fmt.fmt_name, ret);
+}
+
+module_init(pfm_default_smpl_init_module);
+module_exit(pfm_default_smpl_cleanup_module);
+

arch/ia64/kernel/perfmon_generic.h

 /*
- * This file contains the architected PMU register description tables
+ * This file contains the generic PMU register description tables
  * and pmc checker used by perfmon.c.
  *
- * Copyright (C) 2002  Hewlett Packard Co
+ * Copyright (C) 2002-2003  Hewlett Packard Co
  *               Stephane Eranian <eranian@hpl.hp.com>
  */
+
+
 #define RDEP(x)	(1UL<<(x))
 
 #if defined(CONFIG_ITANIUM) || defined (CONFIG_MCKINLEY)
@@ -39,10 +41,13 @@ static pfm_reg_desc_t pmd_gen_desc[PMU_MAX_PMDS]={
  * impl_pmcs, impl_pmds are computed at runtime to minimize errors!
  */
 static pmu_config_t pmu_conf={
-	.disabled = 1,
-	.ovfl_val = (1UL << 32) - 1,
-	.num_ibrs = 8,
-	.num_dbrs = 8,
-	.pmd_desc = pfm_gen_pmd_desc,
-	.pmc_desc = pfm_gen_pmc_desc
+	.pmu_name   = "Generic",
+	.pmu_family = 0xff, /* any */
+	.enabled    = 0,
+	.ovfl_val   = (1UL << 32) - 1,
+	.num_ibrs   = 0, /* does not use */
+	.num_dbrs   = 0, /* does not use */
+	.pmd_desc   = pfm_gen_pmd_desc,
+	.pmc_desc   = pfm_gen_pmc_desc
 };
+

arch/ia64/kernel/perfmon_itanium.h

@@ -2,7 +2,7 @@
  * This file contains the Itanium PMU register description tables
  * and pmc checker used by perfmon.c.
  *
- * Copyright (C) 2002  Hewlett Packard Co
+ * Copyright (C) 2002-2003  Hewlett Packard Co
  *               Stephane Eranian <eranian@hpl.hp.com>
  */
 
@@ -12,8 +12,8 @@
 #error "This file is only valid when CONFIG_ITANIUM is defined"
 #endif
 
-static int pfm_ita_pmc_check(struct task_struct *task, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
-static int pfm_write_ibr_dbr(int mode, struct task_struct *task, void *arg, int count, struct pt_regs *regs);
+static int pfm_ita_pmc_check(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
+static int pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
 
 static pfm_reg_desc_t pfm_ita_pmc_desc[PMU_MAX_PMCS]={
 /* pmc0  */ { PFM_REG_CONTROL , 0, 0x1UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
@@ -59,52 +59,53 @@ static pfm_reg_desc_t pfm_ita_pmd_desc[PMU_MAX_PMDS]={
  * impl_pmcs, impl_pmds are computed at runtime to minimize errors!
  */
 static pmu_config_t pmu_conf={
-	.disabled = 1,
-	.ovfl_val = (1UL << 32) - 1,
-	.num_ibrs = 8,
-	.num_dbrs = 8,
-	.pmd_desc = pfm_ita_pmd_desc,
-	.pmc_desc = pfm_ita_pmc_desc
+	.pmu_name      = "Itanium",
+	.pmu_family    = 0x7,
+	.enabled       = 0,
+	.ovfl_val      = (1UL << 32) - 1,
+	.pmd_desc      = pfm_ita_pmd_desc,
+	.pmc_desc      = pfm_ita_pmc_desc,
+	.num_ibrs      = 8,
+	.num_dbrs      = 8,
+	.use_rr_dbregs = 1 /* debug register are use for range retrictions */
 };
 
-
 static int
-pfm_ita_pmc_check(struct task_struct *task, unsigned int cnum, unsigned long *val, struct pt_regs *regs)
+pfm_ita_pmc_check(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs)
 {
-	pfm_context_t *ctx = task->thread.pfm_context;
 	int ret;
 
 	/*
 	 * we must clear the (instruction) debug registers if pmc13.ta bit is cleared
-	 * before they are written (fl_using_dbreg==0) to avoid picking up stale information. 
+	 * before they are written (fl_using_dbreg==0) to avoid picking up stale information.
 	 */
 	if (cnum == 13 && ((*val & 0x1) == 0UL) && ctx->ctx_fl_using_dbreg == 0) {
 
 		/* don't mix debug with perfmon */
-		if ((task->thread.flags & IA64_THREAD_DBG_VALID) != 0) return -EINVAL;
+		if (task && (task->thread.flags & IA64_THREAD_DBG_VALID) != 0) return -EINVAL;
 
-		/* 
+		/*
 		 * a count of 0 will mark the debug registers as in use and also
 		 * ensure that they are properly cleared.
 		 */
-		ret = pfm_write_ibr_dbr(1, task, NULL, 0, regs);
+		ret = pfm_write_ibr_dbr(1, ctx, NULL, 0, regs);
 		if (ret) return ret;
 	}
 
 	/*
 	 * we must clear the (data) debug registers if pmc11.pt bit is cleared
-	 * before they are written (fl_using_dbreg==0) to avoid picking up stale information. 
+	 * before they are written (fl_using_dbreg==0) to avoid picking up stale information.
 	 */
 	if (cnum == 11 && ((*val >> 28)& 0x1) == 0 && ctx->ctx_fl_using_dbreg == 0) {
 
 		/* don't mix debug with perfmon */
-		if ((task->thread.flags & IA64_THREAD_DBG_VALID) != 0) return -EINVAL;
+		if (task && (task->thread.flags & IA64_THREAD_DBG_VALID) != 0) return -EINVAL;
 
-		/* 
+		/*
 		 * a count of 0 will mark the debug registers as in use and also
 		 * ensure that they are properly cleared.
 		 */
-		ret = pfm_write_ibr_dbr(0, task, NULL, 0, regs);
+		ret = pfm_write_ibr_dbr(0, ctx, NULL, 0, regs);
 		if (ret) return ret;
 	}
 	return 0;

arch/ia64/kernel/perfmon_mckinley.h

@@ -2,7 +2,7 @@
  * This file contains the McKinley PMU register description tables
  * and pmc checker used by perfmon.c.
  *
- * Copyright (C) 2002  Hewlett Packard Co
+ * Copyright (C) 2002-2003  Hewlett Packard Co
  *               Stephane Eranian <eranian@hpl.hp.com>
  */
 
@@ -12,9 +12,8 @@
 #error "This file is only valid when CONFIG_MCKINLEY is defined"
 #endif
 
-static int pfm_mck_reserved(struct task_struct *task, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
-static int pfm_mck_pmc_check(struct task_struct *task, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
-static int pfm_write_ibr_dbr(int mode, struct task_struct *task, void *arg, int count, struct pt_regs *regs);
+static int pfm_mck_pmc_check(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs);
+static int pfm_write_ibr_dbr(int mode, pfm_context_t *ctx, void *arg, int count, struct pt_regs *regs);
 
 static pfm_reg_desc_t pfm_mck_pmc_desc[PMU_MAX_PMCS]={
 /* pmc0  */ { PFM_REG_CONTROL , 0, 0x1UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
@@ -22,17 +21,17 @@ static pfm_reg_desc_t pfm_mck_pmc_desc[PMU_MAX_PMCS]={
 /* pmc2  */ { PFM_REG_CONTROL , 0, 0x0UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
 /* pmc3  */ { PFM_REG_CONTROL , 0, 0x0UL, -1UL, NULL, NULL, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
 /* pmc4  */ { PFM_REG_COUNTING, 6, 0x0000000000800000UL, 0xfffff7fUL, NULL, pfm_mck_pmc_check, {RDEP(4),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc5  */ { PFM_REG_COUNTING, 6, 0x0UL, 0xfffff7fUL, NULL,  pfm_mck_reserved, {RDEP(5),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc6  */ { PFM_REG_COUNTING, 6, 0x0UL, 0xfffff7fUL, NULL,  pfm_mck_reserved, {RDEP(6),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc7  */ { PFM_REG_COUNTING, 6, 0x0UL, 0xfffff7fUL, NULL,  pfm_mck_reserved, {RDEP(7),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc8  */ { PFM_REG_CONFIG  , 0, 0xffffffff3fffffffUL, 0xffffffff3fffffffUL, NULL, pfm_mck_pmc_check, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc9  */ { PFM_REG_CONFIG  , 0, 0xffffffff3ffffffcUL, 0xffffffff3fffffffUL, NULL, pfm_mck_pmc_check, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc10 */ { PFM_REG_MONITOR , 4, 0x0UL, 0xffffUL, NULL, pfm_mck_reserved, {RDEP(0)|RDEP(1),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc11 */ { PFM_REG_MONITOR , 6, 0x0UL, 0x30f01cf, NULL,  pfm_mck_reserved, {RDEP(2)|RDEP(3)|RDEP(17),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc12 */ { PFM_REG_MONITOR , 6, 0x0UL, 0xffffUL, NULL,  pfm_mck_reserved, {RDEP(8)|RDEP(9)|RDEP(10)|RDEP(11)|RDEP(12)|RDEP(13)|RDEP(14)|RDEP(15)|RDEP(16),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
+/* pmc5  */ { PFM_REG_COUNTING, 6, 0x0UL, 0xfffff7fUL, NULL,  pfm_mck_pmc_check, {RDEP(5),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
+/* pmc6  */ { PFM_REG_COUNTING, 6, 0x0UL, 0xfffff7fUL, NULL,  pfm_mck_pmc_check, {RDEP(6),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
+/* pmc7  */ { PFM_REG_COUNTING, 6, 0x0UL, 0xfffff7fUL, NULL,  pfm_mck_pmc_check, {RDEP(7),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
+/* pmc8  */ { PFM_REG_CONFIG  , 0, 0xffffffff3fffffffUL, 0xffffffff3ffffffbUL, NULL, pfm_mck_pmc_check, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
+/* pmc9  */ { PFM_REG_CONFIG  , 0, 0xffffffff3ffffffcUL, 0xffffffff3ffffffbUL, NULL, pfm_mck_pmc_check, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
+/* pmc10 */ { PFM_REG_MONITOR , 4, 0x0UL, 0xffffUL, NULL, pfm_mck_pmc_check, {RDEP(0)|RDEP(1),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
+/* pmc11 */ { PFM_REG_MONITOR , 6, 0x0UL, 0x30f01cf, NULL,  pfm_mck_pmc_check, {RDEP(2)|RDEP(3)|RDEP(17),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
+/* pmc12 */ { PFM_REG_MONITOR , 6, 0x0UL, 0xffffUL, NULL,  pfm_mck_pmc_check, {RDEP(8)|RDEP(9)|RDEP(10)|RDEP(11)|RDEP(12)|RDEP(13)|RDEP(14)|RDEP(15)|RDEP(16),0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
 /* pmc13 */ { PFM_REG_CONFIG  , 0, 0x00002078fefefefeUL, 0x1e00018181818UL, NULL, pfm_mck_pmc_check, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
 /* pmc14 */ { PFM_REG_CONFIG  , 0, 0x0db60db60db60db6UL, 0x2492UL, NULL, pfm_mck_pmc_check, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
-/* pmc15 */ { PFM_REG_CONFIG  , 0, 0x00000000fffffff0UL, 0xfUL, NULL, pfm_mck_reserved, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
+/* pmc15 */ { PFM_REG_CONFIG  , 0, 0x00000000fffffff0UL, 0xfUL, NULL, pfm_mck_pmc_check, {0UL,0UL, 0UL, 0UL}, {0UL,0UL, 0UL, 0UL}},
 	    { PFM_REG_END     , 0, 0x0UL, -1UL, NULL, NULL, {0,}, {0,}}, /* end marker */
 };
 
@@ -62,20 +61,22 @@ static pfm_reg_desc_t pfm_mck_pmd_desc[PMU_MAX_PMDS]={
  * impl_pmcs, impl_pmds are computed at runtime to minimize errors!
  */
 static pmu_config_t pmu_conf={
-	.disabled = 1,
-	.ovfl_val = (1UL << 47) - 1,
-	.num_ibrs = 8,
-	.num_dbrs = 8,
-	.pmd_desc = pfm_mck_pmd_desc,
-	.pmc_desc = pfm_mck_pmc_desc
+	.pmu_name      = "Itanium 2",
+	.pmu_family    = 0x1f,
+	.enabled       = 0,
+	.ovfl_val      = (1UL << 47) - 1,
+	.pmd_desc      = pfm_mck_pmd_desc,
+	.pmc_desc      = pfm_mck_pmc_desc,
+	.num_ibrs       = 8,
+	.num_dbrs       = 8,
+	.use_rr_dbregs = 1 /* debug register are use for range retrictions */
 };
 
-
 /*
  * PMC reserved fields must have their power-up values preserved
  */
 static int
-pfm_mck_reserved(struct task_struct *task, unsigned int cnum, unsigned long *val, struct pt_regs *regs)
+pfm_mck_reserved(unsigned int cnum, unsigned long *val, struct pt_regs *regs)
 {
 	unsigned long tmp1, tmp2, ival = *val;
 
@@ -87,52 +88,51 @@ pfm_mck_reserved(struct task_struct *task, unsigned int cnum, unsigned long *val
 
 	*val = tmp1 | tmp2;
 
-	DBprintk(("pmc[%d]=0x%lx, mask=0x%lx, reset=0x%lx, val=0x%lx\n", 
-		  cnum, ival, PMC_RSVD_MASK(cnum), PMC_DFL_VAL(cnum), *val)); 
+	DPRINT(("pmc[%d]=0x%lx, mask=0x%lx, reset=0x%lx, val=0x%lx\n",
+		  cnum, ival, PMC_RSVD_MASK(cnum), PMC_DFL_VAL(cnum), *val));
 	return 0;
 }
 
 static int
-pfm_mck_pmc_check(struct task_struct *task, unsigned int cnum, unsigned long *val, struct pt_regs *regs)
+pfm_mck_pmc_check(struct task_struct *task, pfm_context_t *ctx, unsigned int cnum, unsigned long *val, struct pt_regs *regs)
 {
-	struct thread_struct *th = &task->thread;
-	pfm_context_t *ctx = task->thread.pfm_context;
 	int ret = 0, check_case1 = 0;
+	struct thread_struct *th = &task->thread;
 	unsigned long val8 = 0, val14 = 0, val13 = 0;
 
 	/* first preserve the reserved fields */
-	pfm_mck_reserved(task, cnum, val, regs);
+	pfm_mck_reserved(cnum, val, regs);
 
 	/*
-	 * we must clear the debug registers if any pmc13.ena_dbrpX bit is enabled 
-	 * before they are written (fl_using_dbreg==0) to avoid picking up stale information. 
+	 * we must clear the debug registers if any pmc13.ena_dbrpX bit is enabled
+	 * before they are written (fl_using_dbreg==0) to avoid picking up stale information.
 	 */
 	if (cnum == 13 && (*val & (0xfUL << 45)) && ctx->ctx_fl_using_dbreg == 0) {
 
 		/* don't mix debug with perfmon */
-		if ((task->thread.flags & IA64_THREAD_DBG_VALID) != 0) return -EINVAL;
+		if (task && (task->thread.flags & IA64_THREAD_DBG_VALID) != 0) return -EINVAL;
 
-		/* 
+		/*
 		 * a count of 0 will mark the debug registers as in use and also
 		 * ensure that they are properly cleared.
 		 */
-		ret = pfm_write_ibr_dbr(1, task, NULL, 0, regs);
+		ret = pfm_write_ibr_dbr(1, ctx, NULL, 0, regs);
 		if (ret) return ret;
 	}
-	/* 
-	 * we must clear the (instruction) debug registers if any pmc14.ibrpX bit is enabled 
-	 * before they are (fl_using_dbreg==0) to avoid picking up stale information. 
+	/*
+	 * we must clear the (instruction) debug registers if any pmc14.ibrpX bit is enabled
+	 * before they are (fl_using_dbreg==0) to avoid picking up stale information.
 	 */
 	if (cnum == 14 && ((*val & 0x2222) != 0x2222) && ctx->ctx_fl_using_dbreg == 0) {
 
 		/* don't mix debug with perfmon */
-		if ((task->thread.flags & IA64_THREAD_DBG_VALID) != 0) return -EINVAL;
+		if (task && (task->thread.flags & IA64_THREAD_DBG_VALID) != 0) return -EINVAL;
 
-		/* 
+		/*
 		 * a count of 0 will mark the debug registers as in use and also
 		 * ensure that they are properly cleared.
 		 */
-		ret = pfm_write_ibr_dbr(0, task, NULL, 0, regs);
+		ret = pfm_write_ibr_dbr(0, ctx, NULL, 0, regs);
 		if (ret) return ret;
 
 	}
@@ -141,17 +141,17 @@ pfm_mck_pmc_check(struct task_struct *task, unsigned int cnum, unsigned long *va
 		case  4: *val |= 1UL << 23; /* force power enable bit */
 			 break;
 		case  8: val8 = *val;
-			 val13 = th->pmc[13];
-			 val14 = th->pmc[14];
+			 val13 = th->pmcs[13];
+			 val14 = th->pmcs[14];
 			 check_case1 = 1;
 			 break;
-		case 13: val8  = th->pmc[8];
+		case 13: val8  = th->pmcs[8];
 			 val13 = *val;
-			 val14 = th->pmc[14];
+			 val14 = th->pmcs[14];
 			 check_case1 = 1;
 			 break;
-		case 14: val8  = th->pmc[13];
-			 val13 = th->pmc[13];
+		case 14: val8  = th->pmcs[13];
+			 val13 = th->pmcs[13];
 			 val14 = *val;
 			 check_case1 = 1;
 			 break;
@@ -165,7 +165,7 @@ pfm_mck_pmc_check(struct task_struct *task, unsigned int cnum, unsigned long *va
 		   && ((((val14>>1) & 0x3) == 0x2 || ((val14>>1) & 0x3) == 0x0)
 		       ||(((val14>>4) & 0x3) == 0x2 || ((val14>>4) & 0x3) == 0x0));
 
-		if (ret) printk(KERN_DEBUG "perfmon: failure check_case1\n");
+		if (ret) printk("perfmon: failure check_case1\n");
 	}
 
 	return ret ? -EINVAL : 0;

arch/ia64/kernel/process.c

@@ -148,8 +148,8 @@ do_notify_resume_user (sigset_t *oldset, struct sigscratch *scr, long in_syscall
 #endif
 
 #ifdef CONFIG_PERFMON
-	if (current->thread.pfm_ovfl_block_reset)
-		pfm_ovfl_block_reset();
+	if (current->thread.pfm_needs_checking)
+		pfm_handle_work();
 #endif
 
 	/* deal with pending signal delivery */
@@ -387,16 +387,8 @@ copy_thread (int nr, unsigned long clone_flags,
 #endif
 
 #ifdef CONFIG_PERFMON
-	/*
-	 * reset notifiers and owner check (may not have a perfmon context)
-	 */
-	atomic_set(&p->thread.pfm_notifiers_check, 0);
-	atomic_set(&p->thread.pfm_owners_check, 0);
-	/* clear list of sampling buffer to free for new task */
-	p->thread.pfm_smpl_buf_list = NULL;
-
 	if (current->thread.pfm_context)
-		retval = pfm_inherit(p, child_ptregs);
+		pfm_inherit(p, child_ptregs);
 #endif
 	return retval;
 }
@@ -609,34 +601,6 @@ flush_thread (void)
 	ia64_drop_fpu(current);
 }
 
-#ifdef CONFIG_PERFMON
-/*
- * by the time we get here, the task is detached from the tasklist. This is important
- * because it means that no other tasks can ever find it as a notified task, therfore there
- * is no race condition between this code and let's say a pfm_context_create().
- * Conversely, the pfm_cleanup_notifiers() cannot try to access a task's pfm context if this
- * other task is in the middle of its own pfm_context_exit() because it would already be out of
- * the task list. Note that this case is very unlikely between a direct child and its parents
- * (if it is the notified process) because of the way the exit is notified via SIGCHLD.
- */
-
-void
-release_thread (struct task_struct *task)
-{
-	if (task->thread.pfm_context)
-		pfm_context_exit(task);
-
-	if (atomic_read(&task->thread.pfm_notifiers_check) > 0)
-		pfm_cleanup_notifiers(task);
-
-	if (atomic_read(&task->thread.pfm_owners_check) > 0)
-		pfm_cleanup_owners(task);
-
-	if (task->thread.pfm_smpl_buf_list)
-		pfm_cleanup_smpl_buf(task);
-}
-#endif
-
 /*
  * Clean up state associated with current thread.  This is called when
  * the thread calls exit().
@@ -648,7 +612,7 @@ exit_thread (void)
 #ifdef CONFIG_PERFMON
        /* if needed, stop monitoring and flush state to perfmon context */
 	if (current->thread.pfm_context)
-		pfm_flush_regs(current);
+		pfm_exit_thread(current);
 
 	/* free debug register resources */
 	if (current->thread.flags & IA64_THREAD_DBG_VALID)

arch/ia64/kernel/signal.c

@@ -189,24 +189,15 @@ copy_siginfo_to_user (siginfo_t *to, siginfo_t *from)
 			err |= __put_user(from->si_addr, &to->si_addr);
 			err |= __put_user(from->si_imm, &to->si_imm);
 			break;
-		      case __SI_CHLD >> 16:
-			err |= __put_user(from->si_utime, &to->si_utime);
-			err |= __put_user(from->si_stime, &to->si_stime);
-			err |= __put_user(from->si_status, &to->si_status);
-		      case __SI_PROF >> 16:
-			err |= __put_user(from->si_uid, &to->si_uid);
-			err |= __put_user(from->si_pid, &to->si_pid);
-			if (from->si_code == PROF_OVFL) {
-				err |= __put_user(from->si_pfm_ovfl[0], &to->si_pfm_ovfl[0]);
-				err |= __put_user(from->si_pfm_ovfl[1], &to->si_pfm_ovfl[1]);
-				err |= __put_user(from->si_pfm_ovfl[2], &to->si_pfm_ovfl[2]);
-				err |= __put_user(from->si_pfm_ovfl[3], &to->si_pfm_ovfl[3]);
-			}
 		      case __SI_TIMER >> 16:
 			err |= __put_user(from->si_tid, &to->si_tid);
 			err |= __put_user(from->si_overrun, &to->si_overrun);
 			err |= __put_user(from->si_value, &to->si_value);
 			break;
+		      case __SI_CHLD >> 16:
+			err |= __put_user(from->si_utime, &to->si_utime);
+			err |= __put_user(from->si_stime, &to->si_stime);
+			err |= __put_user(from->si_status, &to->si_status);
 		      default:
 			err |= __put_user(from->si_uid, &to->si_uid);
 			err |= __put_user(from->si_pid, &to->si_pid);
@@ -243,10 +234,6 @@ copy_siginfo_from_user (siginfo_t *to, siginfo_t *from)
 			to->si_code |= __SI_POLL;
 			break;
 
-		      case SIGPROF:
-			to->si_code |= __SI_PROF;
-			break;
-
 		      default:
 			break;
 		}

include/asm-ia64/kregs.h

@@ -72,7 +72,7 @@
 #define IA64_PSR_BITS_TO_CLEAR	(IA64_PSR_MFL | IA64_PSR_MFH | IA64_PSR_DB | IA64_PSR_LP | \
 				 IA64_PSR_TB  | IA64_PSR_ID  | IA64_PSR_DA | IA64_PSR_DD | \
 				 IA64_PSR_SS  | IA64_PSR_ED  | IA64_PSR_IA)
-#define IA64_PSR_BITS_TO_SET	(IA64_PSR_DFH)
+#define IA64_PSR_BITS_TO_SET	(IA64_PSR_DFH | IA64_PSR_SP)
 
 #define IA64_PSR_BE	(__IA64_UL(1) << IA64_PSR_BE_BIT)
 #define IA64_PSR_UP	(__IA64_UL(1) << IA64_PSR_UP_BIT)

include/asm-ia64/perfmon.h

@@ -14,20 +14,21 @@
 #define PFM_READ_PMDS		0x03
 #define PFM_STOP		0x04
 #define PFM_START		0x05
-#define PFM_ENABLE		0x06
-#define PFM_DISABLE		0x07
+#define PFM_ENABLE		0x06 /* obsolete */
+#define PFM_DISABLE		0x07 /* obsolete */
 #define PFM_CREATE_CONTEXT	0x08
-#define PFM_DESTROY_CONTEXT	0x09
+#define PFM_DESTROY_CONTEXT	0x09 /* obsolete use close() */
 #define PFM_RESTART		0x0a
-#define PFM_PROTECT_CONTEXT	0x0b
+#define PFM_PROTECT_CONTEXT	0x0b /* obsolete */
 #define PFM_GET_FEATURES	0x0c
 #define PFM_DEBUG		0x0d
-#define PFM_UNPROTECT_CONTEXT	0x0e
+#define PFM_UNPROTECT_CONTEXT	0x0e /* obsolete */
 #define PFM_GET_PMC_RESET_VAL	0x0f
-
+#define PFM_LOAD_CONTEXT	0x10
+#define PFM_UNLOAD_CONTEXT	0x11
 
 /*
- * CPU model specific commands (may not be supported on all models)
+ * PMU model specific commands (may not be supported on all PMU models)
  */
 #define PFM_WRITE_IBRS		0x20
 #define PFM_WRITE_DBRS		0x21
@@ -35,19 +36,21 @@
 /*
  * context flags
  */
-#define PFM_FL_INHERIT_NONE	 0x00	/* never inherit a context across fork (default) */
-#define PFM_FL_INHERIT_ONCE	 0x01	/* clone pfm_context only once across fork() */
-#define PFM_FL_INHERIT_ALL	 0x02	/* always clone pfm_context across fork() */
-#define PFM_FL_NOTIFY_BLOCK    	 0x04	/* block task on user level notifications */
-#define PFM_FL_SYSTEM_WIDE	 0x08	/* create a system wide context */
-#define PFM_FL_EXCL_IDLE         0x20   /* exclude idle task from system wide session */
-#define PFM_FL_UNSECURE		 0x40   /* allow unsecure monitoring for non self-monitoring task */
+#define PFM_FL_NOTIFY_BLOCK    	 0x01	/* block task on user level notifications */
+#define PFM_FL_SYSTEM_WIDE	 0x02	/* create a system wide context */
+#define PFM_FL_UNSECURE		 0x04   /* allow unsecure monitoring for non self-monitoring task */
+#define PFM_FL_OVFL_NO_MSG	 0x80   /* do not post overflow/end messages for notification */
+
+/*
+ * event set flags
+ */
+#define PFM_SETFL_EXCL_IDLE      0x01   /* exclude idle task (syswide only) XXX: DO NOT USE YET */
 
 /*
  * PMC flags
  */
 #define PFM_REGFL_OVFL_NOTIFY	0x1	/* send notification on overflow */
-#define PFM_REGFL_RANDOM	0x2	/* randomize sampling periods    */
+#define PFM_REGFL_RANDOM	0x2	/* randomize sampling interval   */
 
 /*
  * PMD/PMC/IBR/DBR return flags (ignored on input)
@@ -55,153 +58,197 @@
  * Those flags are used on output and must be checked in case EAGAIN is returned
  * by any of the calls using a pfarg_reg_t or pfarg_dbreg_t structure.
  */
-#define PFM_REG_RETFL_NOTAVAIL	(1U<<31) /* set if register is implemented but not available */
-#define PFM_REG_RETFL_EINVAL	(1U<<30) /* set if register entry is invalid */
+#define PFM_REG_RETFL_NOTAVAIL	(1UL<<31) /* set if register is implemented but not available */
+#define PFM_REG_RETFL_EINVAL	(1UL<<30) /* set if register entry is invalid */
 #define PFM_REG_RETFL_MASK	(PFM_REG_RETFL_NOTAVAIL|PFM_REG_RETFL_EINVAL)
 
 #define PFM_REG_HAS_ERROR(flag)	(((flag) & PFM_REG_RETFL_MASK) != 0)
 
+typedef unsigned char pfm_uuid_t[16];	/* custom sampling buffer identifier type */
+
 /*
  * Request structure used to define a context
  */
 typedef struct {
-	unsigned long ctx_smpl_entries;	/* how many entries in sampling buffer */
-	unsigned long ctx_smpl_regs[4];	/* which pmds to record on overflow */
-
-	pid_t	      ctx_notify_pid;	/* which process to notify on overflow */
-	int	      ctx_flags;	/* noblock/block, inherit flags */
-	void	      *ctx_smpl_vaddr;	/* returns address of BTB buffer */
-
-	unsigned long ctx_cpu_mask;	/* on which CPU to enable perfmon (systemwide) */
-
-	unsigned long reserved[8];	/* for future use */
+	pfm_uuid_t    ctx_smpl_buf_id;	/* which buffer format to use (if needed) */
+	unsigned long ctx_flags;	/* noblock/block */
+	unsigned int  ctx_nextra_sets;	/* number of extra event sets (you always get 1) */
+	int	      ctx_fd;		/* return arg: unique identification for context */
+	void	      *ctx_smpl_vaddr;	/* return arg: virtual address of sampling buffer, is used */
+	unsigned long ctx_reserved[11];	/* for future use */
 } pfarg_context_t;
 
 /*
  * Request structure used to write/read a PMC or PMD
  */
 typedef struct {
-	unsigned int	reg_num;	/* which register */
-	unsigned int	reg_flags;	/* PMC: notify/don't notify. PMD/PMC: return flags */
-	unsigned long	reg_value;	/* configuration (PMC) or initial value (PMD) */
+	unsigned int	reg_num;	   /* which register                             */
+	unsigned int	reg_set;	   /* event set for this register                */
+
+	unsigned long	reg_value;	   /* initial pmc/pmd value                      */
+	unsigned long	reg_flags;	   /* input: pmc/pmd flags, return: reg error    */
+
+	unsigned long	reg_long_reset;	   /* reset after buffer overflow notification   */
+	unsigned long	reg_short_reset;   /* reset after counter overflow               */
 
-	unsigned long	reg_long_reset;	/* reset after sampling buffer overflow (large) */
-	unsigned long	reg_short_reset;/* reset after counter overflow (small) */
+	unsigned long	reg_reset_pmds[4]; /* which other counters to reset on overflow  */
+	unsigned long	reg_random_seed;   /* seed value when randomization is used      */
+	unsigned long	reg_random_mask;   /* bitmask used to limit random value         */
+	unsigned long   reg_last_reset_val;/* return: PMD last reset value               */
 
-	unsigned long	reg_reset_pmds[4];   /* which other counters to reset on overflow */
-	unsigned long	reg_random_seed;     /* seed value when randomization is used */
-	unsigned long	reg_random_mask;     /* bitmask used to limit random value */
-	unsigned long	reg_last_reset_value;/* last value used to reset the PMD (PFM_READ_PMDS) */
+	unsigned long	reg_smpl_pmds[4];  /* which pmds are accessed when PMC overflows */
+	unsigned long	reg_smpl_eventid;  /* opaque sampling event identifier           */
 
-	unsigned long   reserved[13];	/* for future use */
+	unsigned long   reserved[3];	   /* for future use                             */
 } pfarg_reg_t;
 
 typedef struct {
-	unsigned int	dbreg_num;	/* which register */
-	unsigned int	dbreg_flags;	/* dbregs return flags */
-	unsigned long	dbreg_value;	/* configuration (PMC) or initial value (PMD) */
-	unsigned long	reserved[6];
+	unsigned int	dbreg_num;		/* which debug register        */
+	unsigned int 	dbreg_set;		/* event set for this register */
+	unsigned long	dbreg_value;		/* value for debug register    */
+	unsigned long	dbreg_flags;		/* return: dbreg error         */
+	unsigned long	dbreg_reserved[1];	/* for future use              */
 } pfarg_dbreg_t;
 
-typedef struct {			
+typedef struct {
 	unsigned int	ft_version;	/* perfmon: major [16-31], minor [0-15] */
-	unsigned int	ft_smpl_version;/* sampling format: major [16-31], minor [0-15] */
-	unsigned long	reserved[4];	/* for future use */
+	unsigned int	ft_reserved;	/* reserved for future use              */
+	unsigned long	reserved[4];	/* for future use                       */
 } pfarg_features_t;
 
-/*
- * This header is at the beginning of the sampling buffer returned to the user.
- * It is exported as Read-Only at this point. It is directly followed by the
- * first record.
- */
 typedef struct {
-	unsigned int	hdr_version;		/* contains perfmon version (smpl format diffs) */
-	unsigned int	reserved;
-	unsigned long	hdr_entry_size;		/* size of one entry in bytes */
-	unsigned long	hdr_count;		/* how many valid entries */
-	unsigned long	hdr_pmds[4];		/* which pmds are recorded */
-} perfmon_smpl_hdr_t;
+	pid_t		load_pid;	  /* process to load the context into */
+	unsigned int	load_set;	  /* first event set to load          */
+	unsigned long	load_reserved[2]; /* for future use                   */
+} pfarg_load_t;
 
-/*
- * Define the version numbers for both perfmon as a whole and the sampling buffer format.
- */
-#define PFM_VERSION_MAJ		1U
-#define PFM_VERSION_MIN		4U
-#define PFM_VERSION		(((PFM_VERSION_MAJ&0xffff)<<16)|(PFM_VERSION_MIN & 0xffff))
+typedef struct {
+	int		msg_type;		/* generic message header */
+	int		msg_ctx_fd;		/* generic message header */
+	unsigned long	msg_tstamp;		/* for perf tuning */
+	unsigned int	msg_active_set;		/* active set at the time of overflow */
+	unsigned long	msg_ovfl_pmds[4];	/* which PMDs overflowed */
+} pfm_ovfl_msg_t;
 
-#define PFM_SMPL_VERSION_MAJ	1U
-#define PFM_SMPL_VERSION_MIN	0U
-#define PFM_SMPL_VERSION	(((PFM_SMPL_VERSION_MAJ&0xffff)<<16)|(PFM_SMPL_VERSION_MIN & 0xffff))
+typedef struct {
+	int		msg_type;		/* generic message header */
+	int		msg_ctx_fd;		/* generic message header */
+	unsigned long	msg_tstamp;		/* for perf tuning */
+} pfm_end_msg_t;
 
+typedef struct {
+	int		msg_type;		/* type of the message */
+	int		msg_ctx_fd;		/* unique identifier for the context */
+	unsigned long	msg_tstamp;		/* for perf tuning */
+} pfm_gen_msg_t;
 
-#define PFM_VERSION_MAJOR(x)	(((x)>>16) & 0xffff)
-#define PFM_VERSION_MINOR(x)	((x) & 0xffff)
+#define PFM_MSG_OVFL	1	/* an overflow happened */
+#define PFM_MSG_END	2	/* task to which context was attached ended */
+
+typedef union {
+	pfm_ovfl_msg_t	pfm_ovfl_msg;
+	pfm_end_msg_t	pfm_end_msg;
+	pfm_gen_msg_t	pfm_gen_msg;
+} pfm_msg_t;
 
 /*
- * Entry header in the sampling buffer.  The header is directly followed
- * with the PMDs saved in increasing index order: PMD4, PMD5, .... How
- * many PMDs are present is determined by the user program during
- * context creation.
- *
- * XXX: in this version of the entry, only up to 64 registers can be
- * recorded. This should be enough for quite some time. Always check
- * sampling format before parsing entries!
- *
- * In the case where multiple counters overflow at the same time, the
- * last_reset_value member indicates the initial value of the PMD with
- * the smallest index.  For instance, if PMD2 and PMD5 have overflowed,
- * the last_reset_value member contains the initial value of PMD2.
+ * Define the version numbers for both perfmon as a whole and the sampling buffer format.
  */
-typedef struct {
-	int		pid;			/* identification of process */
-	int		cpu;			/* which cpu was used */
-	unsigned long	last_reset_value;	/* initial value of counter that overflowed */
-	unsigned long	stamp;			/* timestamp */
-	unsigned long	ip;			/* where did the overflow interrupt happened */
-	unsigned long	regs;			/* bitmask of which registers overflowed */
-	unsigned long   reserved;		/* unused */
-} perfmon_smpl_entry_t;
+#define PFM_VERSION_MAJ		 2U
+#define PFM_VERSION_MIN		 0U
+#define PFM_SMPL_HDR_VERSION_MAJ 2U
+#define PFM_SMPL_HDR_VERSION_MIN 0U
+#define PFM_VERSION		 (((PFM_VERSION_MAJ&0xffff)<<16)|(PFM_VERSION_MIN & 0xffff))
+#define PFM_VERSION_MAJOR(x)	 (((x)>>16) & 0xffff)
+#define PFM_VERSION_MINOR(x)	 ((x) & 0xffff)
 
-extern long perfmonctl(pid_t pid, int cmd, void *arg, int narg);
+
+/*
+ * miscellaneous architected definitions
+ */
+#define PMU_FIRST_COUNTER	4	/* first counting monitor (PMC/PMD) */
+#define PMU_MAX_PMCS		256	/* maximum architected number of PMC registers */
+#define PMU_MAX_PMDS		256	/* maximum architected number of PMD registers */
 
 #ifdef __KERNEL__
 
-typedef struct {
-	void (*handler)(int irq, void *arg, struct pt_regs *regs);
-} pfm_intr_handler_desc_t;
+extern long perfmonctl(int fd, int cmd, void *arg, int narg);
 
 extern void pfm_save_regs (struct task_struct *);
 extern void pfm_load_regs (struct task_struct *);
 
-extern int  pfm_inherit (struct task_struct *, struct pt_regs *);
-extern void pfm_context_exit (struct task_struct *);
-extern void pfm_flush_regs (struct task_struct *);
-extern void pfm_cleanup_notifiers (struct task_struct *);
-extern void pfm_cleanup_owners (struct task_struct *);
+extern void pfm_exit_thread(struct task_struct *);
 extern int  pfm_use_debug_registers(struct task_struct *);
 extern int  pfm_release_debug_registers(struct task_struct *);
-extern int  pfm_cleanup_smpl_buf(struct task_struct *);
 extern void pfm_syst_wide_update_task(struct task_struct *, unsigned long info, int is_ctxswin);
-extern void pfm_ovfl_block_reset(void);
+extern void pfm_inherit(struct task_struct *task, struct pt_regs *regs);
 extern void pfm_init_percpu(void);
+extern void pfm_handle_work(void);
 
-/* 
- * hooks to allow VTune/Prospect to cooperate with perfmon.
- * (reserved for system wide monitoring modules only)
+/*
+ * Reset PMD register flags
  */
-extern int pfm_install_alternate_syswide_subsystem(pfm_intr_handler_desc_t *h);
-extern int pfm_remove_alternate_syswide_subsystem(pfm_intr_handler_desc_t *h);
+#define PFM_PMD_NO_RESET	0
+#define PFM_PMD_LONG_RESET	1
+#define PFM_PMD_SHORT_RESET	2
+
+typedef struct {
+	unsigned int notify_user:1;	/* notify user program of overflow                           */
+	unsigned int reset_pmds :2;	/* PFM_PMD_NO_RESET, PFM_PMD_LONG_RESET, PFM_PMD_SHORT_RESET */
+	unsigned int block:1;		/* block monitored task on kernel exit                       */
+	unsigned int stop_monitoring:1; /* will mask monitoring via PMCx.plm                         */
+	unsigned int reserved:26;	/* for future use                                            */
+} pfm_ovfl_ctrl_t;
+
+typedef struct {
+	unsigned long   ovfl_pmds[4];	/* bitmask of overflowed pmds                            */
+	unsigned long   ovfl_notify[4];	/* bitmask of overflow pmds which asked for notification */
+	unsigned long   pmd_value;	/* current 64-bit value of 1st pmd which overflowed      */
+	unsigned long   pmd_last_reset;	/* last reset value of 1st pmd which overflowed          */
+	unsigned long	pmd_eventid;	/* eventid associated with 1st pmd which overflowed      */
+	unsigned int    active_set;	/* event set active at the time of the overflow          */
+	unsigned int    reserved1;
+	unsigned long	smpl_pmds[4];
+	unsigned long   smpl_pmds_values[PMU_MAX_PMDS];
+	pfm_ovfl_ctrl_t ovfl_ctrl;	/* return: perfmon controls to set by handler            */
+} pfm_ovfl_arg_t;
+
+
+typedef struct _pfm_buffer_fmt_t {
+	char		*fmt_name;
+	pfm_uuid_t	fmt_uuid;
+	size_t		fmt_arg_size;
+	unsigned long	fmt_flags;
+
+	int		(*fmt_validate)(struct task_struct *task, unsigned int flags, int cpu, void *arg);
+	int		(*fmt_getsize)(struct task_struct *task, unsigned int flags, int cpu, void *arg, unsigned long *size);
+	int 		(*fmt_init)(struct task_struct *task, void *buf, unsigned int flags, int cpu, void *arg);
+	int		(*fmt_handler)(struct task_struct *task, void *buf, pfm_ovfl_arg_t *arg, struct pt_regs *regs);
+	int		(*fmt_restart)(struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs);
+	int		(*fmt_restart_active)(struct task_struct *task, pfm_ovfl_ctrl_t *ctrl, void *buf, struct pt_regs *regs);
+	int		(*fmt_exit)(struct task_struct *task, void *buf, struct pt_regs *regs);
+
+	struct _pfm_buffer_fmt_t *fmt_next;
+	struct _pfm_buffer_fmt_t *fmt_prev;
+} pfm_buffer_fmt_t;
+
+extern int pfm_register_buffer_fmt(pfm_buffer_fmt_t *fmt);
+extern int pfm_unregister_buffer_fmt(pfm_uuid_t uuid);
+
+/*
+ * perfmon interface exported to modules
+ */
+extern long pfm_mod_fast_read_pmds(struct task_struct *, unsigned long mask[4], unsigned long *addr, struct pt_regs *regs);
+extern long pfm_mod_read_pmds(struct task_struct *, pfarg_reg_t *req, unsigned int nreq, struct pt_regs *regs);
+extern long pfm_mod_write_pmcs(struct task_struct *, pfarg_reg_t *req, unsigned int nreq, struct pt_regs *regs);
 
 /*
  * describe the content of the local_cpu_date->pfm_syst_info field
  */
-#define PFM_CPUINFO_SYST_WIDE	0x1	/* if set a system wide session exist */
+#define PFM_CPUINFO_SYST_WIDE	0x1	/* if set a system wide session exists */
 #define PFM_CPUINFO_DCR_PP	0x2	/* if set the system wide session has started */
 #define PFM_CPUINFO_EXCL_IDLE	0x4	/* the system wide session excludes the idle task */
 
-#define PFM_CPUINFO_CLEAR(v)	__get_cpu_var(pfm_syst_info) &= ~(v)
-#define PFM_CPUINFO_SET(v)	__get_cpu_var(pfm_syst_info) |= (v)
-
 #endif /* __KERNEL__ */
 
 #endif /* _ASM_IA64_PERFMON_H */

include/asm-ia64/perfmon_default_smpl.h

+/*
+ * Copyright (C) 2002-2003 Hewlett-Packard Co
+ *               Stephane Eranian <eranian@hpl.hp.com>
+ *
+ * This file implements the default sampling buffer format
+ * for Linux/ia64 perfmon subsystem.
+ */
+#ifndef __PERFMON_DEFAULT_SMPL_H__
+#define __PERFMON_DEFAULT_SMPL_H__ 1
+
+#define PFM_DEFAULT_SMPL_UUID { \
+		0x4d, 0x72, 0xbe, 0xc0, 0x06, 0x64, 0x41, 0x43, 0x82, 0xb4, 0xd3, 0xfd, 0x27, 0x24, 0x3c, 0x97}
+
+/*
+ * format specific parameters (passed at context creation)
+ */
+typedef struct {
+	unsigned long buf_size;		/* size of the buffer in bytes */
+	unsigned long reserved[3];	/* for future use */
+} pfm_default_smpl_arg_t;
+
+/*
+ * combined context+format specific structure. Can be passed
+ * to PFM_CONTEXT_CREATE
+ */
+typedef struct {
+	pfarg_context_t		ctx_arg;
+	pfm_default_smpl_arg_t	buf_arg;
+} pfm_default_smpl_ctx_arg_t;
+
+/*
+ * This header is at the beginning of the sampling buffer returned to the user.
+ * It is directly followed by the first record.
+ */
+typedef struct {
+	unsigned long	hdr_count;		/* how many valid entries */
+	void		*hdr_cur_pos;		/* current position in the buffer */
+	void		*hdr_last_pos;		/* first byte beyond buffer */
+
+	unsigned long	hdr_overflows;		/* how many times the buffer overflowed */
+	unsigned long   hdr_buf_size;		/* how many bytes in the buffer */
+	unsigned int	hdr_version;		/* contains perfmon version (smpl format diffs) */
+	unsigned int	hdr_reserved1;		/* for future use */
+	unsigned long	hdr_reserved[10];	/* for future use */
+} pfm_default_smpl_hdr_t;
+
+/*
+ * Entry header in the sampling buffer.  The header is directly followed
+ * with the PMDs saved in increasing index order: PMD4, PMD5, .... How
+ * many PMDs are present depends on how the session was programmed.
+ *
+ * XXX: in this version of the entry, only up to 64 registers can be
+ * recorded. This should be enough for quite some time. Always check
+ * sampling format before parsing entries!
+ *
+ * In the case where multiple counters overflow at the same time, the
+ * last_reset_value member indicates the initial value of the
+ * overflowed PMD with the smallest index.  For instance, if PMD2 and
+ * PMD5 have overflowed, the last_reset_value member contains the
+ * initial value of PMD2.
+ */
+typedef struct {
+	int		pid;			/* current process at PMU interrupt point */
+	int		cpu;			/* cpu on which the overfow occured */
+	unsigned long	last_reset_val;		/* initial value of 1st overflowed PMD */
+	unsigned long	ip;			/* where did the overflow interrupt happened */
+	unsigned long	ovfl_pmds;		/* which PMDS registers overflowed (64 max) */
+	unsigned long   tstamp;			/* ar.itc on the CPU that took the overflow */
+	unsigned int	set;			/* event set active when overflow ocurred   */
+	unsigned int	reserved1;		/* for future use */
+} pfm_default_smpl_entry_t;
+
+#define PFM_DEFAULT_MAX_PMDS		64 /* how many pmds supported by data structures (sizeof(unsigned long) */
+#define PFM_DEFAULT_MAX_ENTRY_SIZE	(sizeof(pfm_default_smpl_entry_t)+(sizeof(unsigned long)*PFM_DEFAULT_MAX_PMDS))
+#define PFM_DEFAULT_SMPL_MIN_BUF_SIZE	(sizeof(pfm_default_smpl_hdr_t)+PFM_DEFAULT_MAX_ENTRY_SIZE)
+
+#define PFM_DEFAULT_SMPL_VERSION_MAJ	2U
+#define PFM_DEFAULT_SMPL_VERSION_MIN	0U
+#define PFM_DEFAULT_SMPL_VERSION	(((PFM_DEFAULT_SMPL_VERSION_MAJ&0xffff)<<16)|(PFM_DEFAULT_SMPL_VERSION_MIN & 0xffff))
+
+#endif /* __PERFMON_DEFAULT_SMPL_H__ */

include/asm-ia64/processor.h

@@ -258,20 +258,14 @@ struct thread_struct {
 # define INIT_THREAD_IA32
 #endif /* CONFIG_IA32_SUPPORT */
 #ifdef CONFIG_PERFMON
-	__u64 pmc[IA64_NUM_PMC_REGS];
-	__u64 pmd[IA64_NUM_PMD_REGS];
-	unsigned long pfm_ovfl_block_reset;/* non-zero if we need to block or reset regs on ovfl */
-	void *pfm_context;		/* pointer to detailed PMU context */
-	atomic_t pfm_notifiers_check;	/* when >0, will cleanup ctx_notify_task in tasklist */
-	atomic_t pfm_owners_check;	/* when >0, will cleanup ctx_owner in tasklist */
-	void *pfm_smpl_buf_list;	/* list of sampling buffers to vfree */
-# define INIT_THREAD_PM		.pmc =			{0, },	\
-				.pmd =			{0, },	\
-				.pfm_ovfl_block_reset =	0,	\
-				.pfm_context =		NULL,	\
-				.pfm_notifiers_check =	{ 0 },	\
-				.pfm_owners_check =	{ 0 },	\
-				.pfm_smpl_buf_list =	NULL,
+	__u64 pmcs[IA64_NUM_PMC_REGS];
+	__u64 pmds[IA64_NUM_PMD_REGS];
+	void *pfm_context;		     /* pointer to detailed PMU context */
+	unsigned long pfm_needs_checking;    /* when >0, pending perfmon work on kernel exit */
+# define INIT_THREAD_PM		.pmcs =			{0UL, },  \
+				.pmds =			{0UL, },  \
+				.pfm_context =		NULL,     \
+				.pfm_needs_checking =	0UL,
 #else
 # define INIT_THREAD_PM
 #endif
@@ -326,11 +320,7 @@ struct task_struct;
  * parent of DEAD_TASK has collected the exist status of the task via
  * wait().
  */
-#ifdef CONFIG_PERFMON
-  extern void release_thread (struct task_struct *task);
-#else
-# define release_thread(dead_task)
-#endif
+#define release_thread(dead_task)
 
 /* Prepare to copy thread state - unlazy all lazy status */
 #define prepare_to_copy(tsk)	do { } while (0)

include/asm-ia64/siginfo.h

@@ -69,13 +69,6 @@ typedef struct siginfo {
 			long _band;	/* POLL_IN, POLL_OUT, POLL_MSG (XPG requires a "long") */
 			int _fd;
 		} _sigpoll;
-
-		/* SIGPROF */
-		struct {
-			pid_t _pid;		/* which child */
-			uid_t _uid;		/* sender's uid */
-			unsigned long _pfm_ovfl_counters[4]; /* which PMU counter overflowed */
-		} _sigprof;
 	} _sifields;
 } siginfo_t;
 
@@ -94,14 +87,6 @@ typedef struct siginfo {
 #define __ISR_VALID_BIT	0
 #define __ISR_VALID	(1 << __ISR_VALID_BIT)
 
-/*
- * si_code values
- * Positive values for kernel-generated signals.
- */
-#ifdef __KERNEL__
-#define __SI_PROF	(6 << 16)
-#endif
-
 /*
  * SIGILL si_codes
  */
@@ -137,11 +122,6 @@ typedef struct siginfo {
 #undef NSIGTRAP
 #define NSIGTRAP	4
 
-/*
- * SIGPROF si_codes
- */
-#define PROF_OVFL	(__SI_PROF|1)  /* some counters overflowed */
-
 #ifdef __KERNEL__
 #include <linux/string.h>
 
@@ -151,8 +131,8 @@ copy_siginfo (siginfo_t *to, siginfo_t *from)
 	if (from->si_code < 0)
 		memcpy(to, from, sizeof(siginfo_t));
 	else
-		/* _sigprof is currently the largest know union member */
-		memcpy(to, from, 4*sizeof(int) + sizeof(from->_sifields._sigprof));
+		/* _sigchld is currently the largest know union member */
+		memcpy(to, from, 4*sizeof(int) + sizeof(from->_sifields._sigchld));
 }
 
 extern int copy_siginfo_from_user(siginfo_t *to, siginfo_t *from);