Skip to contentSkip to conversation

perf: Introduce extended syscall error reporting

RFC PATCH Superseded Alexander Shishkin 3 participants 16 messages Jul 24, 2015 – Aug 27, 2015

Subsystems (2)

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore

Alexander Shishkin

author reviewer
#0 lore raw

cover perf: Introduce extended syscall error reporting

Hi Peter and Ingo and everybody,

Here's my second stab at improving perf's error reporting by attaching arbitrary strings to the integer error codes. This is largely based off of the previous email thread [1].

This time around, I employed a linker trick to convert the structures containing extended error information into integers, which are then made to look just like normal error codes so that IS_ERR_VALUE() and friends would still work correctly on them. So no extra pointers in the struct perf_event or anywhere else; the extended error codes are passed around like normal error codes. They only need to be converted in syscalls' topmost return statements. This is done in 1/4.

Then, 2/4 illustrates how error sites can be extended to include more information such as file names and line numbers [1].

The other two patches add perf_err() annotation to a few semi-randomly picked places in perf core (3/4) and x86 bits (4/4).

[1] http://marc.info/?l=linux-kernel&m=141470811013082
[2] http://marc.info/?l=linux-kernel&m=141471816115946

Alexander Shishkin (4):
perf: Introduce extended syscall error reporting
perf: Add file name and line number to perf extended error reports
perf: Annotate some of the error codes with perf_err() perf/x86: Annotate some of the error codes with perf_err()

 arch/x86/kernel/cpu/perf_event.c           |   8 ++-
 arch/x86/kernel/cpu/perf_event_intel_lbr.c |   2 +-
 include/linux/perf_event.h                 |  87 +++++++++++++++++++++++
 include/uapi/linux/perf_event.h            |   8 ++-
 kernel/events/core.c                       | 108 +++++++++++++++++++++++++----
 5 files changed, 195 insertions(+), 18 deletions(-)
Signature · 1 line
2.1.4
author reviewer
#1 lore raw

RFC PATCH 1/4 perf: Introduce extended syscall error reporting

It has been pointed several times out that perf syscall error reporting leaves a lot to be desired [1].

This patch introduces a fairly simple extension that allows call sites to annotate their error codes with arbitrary strings, which will then be copied to userspace (if they asked for it) along with the module name that produced the error message in JSON format. This way, we can provide both human-readable and machine-parsable information to user and leave room for extensions in the future, such as file name and line number if kernel debugging is enabled.

Each error "site" is referred to by its index, which is folded into an integer error value within the range of [-PERF_ERRNO, -MAX_ERRNO], where PERF_ERRNO is chosen to be below any known error codes, but still leaving enough room to enumerate error sites. This way, all the traditional macros will still handle these as error codes and we'd only have to convert them to their original values right before returning to userspace. This way we also don't have to worry about keeping a separate pointer to the error message inside a perf_event.

[1] http://marc.info/?l=linux-kernel&m=141470811013082

---
 include/linux/perf_event.h      | 76 ++++++++++++++++++++++++++++++++++++++
 include/uapi/linux/perf_event.h |  8 +++-
 kernel/events/core.c            | 81 ++++++++++++++++++++++++++++++++++++++---
 3 files changed, 159 insertions(+), 6 deletions(-)
include/linux/perf_event.h +76 -0
index 2027809433..9e9af962f6 100644
@@ -56,6 +56,82 @@ struct perf_guest_info_callbacks {
 #include <linux/cgroup.h>
 #include <asm/local.h>
 
+#ifndef PERF_MODNAME
+#define PERF_MODNAME "perf"
+#endif
+
+/*
+ * Extended error reporting: annotate an error code with a string
+ * and a module name to help users diagnase problems with their
+ * attributes and whatnot.
+ */
+struct perf_err_site {
+	const char		*message;
+	const char		*owner;
+	const int		code;
+};
+
+#ifdef CONFIG_PERF_EVENTS
+
+/*
+ * Place all occurrences of struct perf_err_site into a special section,
+ * so that we can find out their offsets, which we'll use to refer back
+ * to the error sites.
+ */
+extern const struct perf_err_site __start___perf_err[], __stop___perf_err[];
+
+#define __perf_err(__e, __c, __m) ({				\
+	static struct perf_err_site				\
+	__attribute__ ((unused,__section__("__perf_err")))	\
+	__err_site = {						\
+		.message	= (__m),			\
+		.owner		= PERF_MODNAME,			\
+		.code		= __builtin_constant_p((__c)) ?	\
+		(__c) : 0,					\
+	};							\
+	(__e) = &__err_site;					\
+})
+
+/*
+ * Use part of the [-1, -MAX_ERRNO] errno range for perf's extended error
+ * reporting. Anything within [-PERF_ERRNO, -MAX_ERRNO] is an index of a
+ * perf_err_site structure within __perf_err section. 3.5k should be enough
+ * for everybody, but let's add a boot-time warning just in case it overflows
+ * one day.
+ */
+#define PERF_ERRNO 512
+
+static inline int perf_errno(const struct perf_err_site *site)
+{
+	unsigned long err = site - __start___perf_err;
+
+	trace_printk("[%ld] %s:%d, %d\n", err, site->file, site->line, site->code);
+	return -(int)err - PERF_ERRNO;
+}
+
+static inline const struct perf_err_site *perf_errno_to_site(int err)
+{
+	return __start___perf_err - err - PERF_ERRNO;
+}
+
+#ifdef MODULE
+/*
+ * Module support is a tad trickier, but far from rocket surgery. Let's
+ * bypass it for now.
+ */
+#define perf_err(__c, __m) (__c)
+#else
+#define perf_err(__c, __m) ({					\
+	struct perf_err_site *s;				\
+	__perf_err(s, (__c), (__m));				\
+	perf_errno(s);						\
+})
+#endif
+
+#define PERF_ERR_PTR(__e, __m)	(ERR_PTR(perf_err(__e, __m)))
+
+#endif
+
 struct perf_callchain_entry {
 	__u64				nr;
 	__u64				ip[PERF_MAX_STACK_DEPTH];
include/uapi/linux/perf_event.h +7 -1
index d97f84c080..d1ae1a079c 100644
@@ -264,6 +264,7 @@ enum perf_event_read_format {
 					/* add: sample_stack_user */
 #define PERF_ATTR_SIZE_VER4	104	/* add: sample_regs_intr */
 #define PERF_ATTR_SIZE_VER5	112	/* add: aux_watermark */
+#define PERF_ATTR_SIZE_VER6	120	/* add: perf_err */
 
 /*
  * Hardware event_id to monitor via a performance monitoring event:
@@ -374,7 +375,12 @@ struct perf_event_attr {
 	 * Wakeup watermark for AUX area
 	 */
 	__u32	aux_watermark;
-	__u32	__reserved_2;	/* align to __u64 */
+
+	/*
+	 * Extended error reporting buffer
+	 */
+	__u32	perf_err_size;
+	__u64	perf_err;
 };
 
 #define perf_flags(attr)	(*(&(attr)->read_format + 1))
kernel/events/core.c +76 -5

123 lines not on this page · raw message

Signature · 1 line
2.1.4
author reviewer
#2 lore raw
Quoted 6 lines · Alexander Shishkin
include/linux/perf_event.h +5 -0
+static inline int perf_errno(const struct perf_err_site *site)
+{
+	unsigned long err = site - __start___perf_err;
+
+	trace_printk("[%ld] %s:%d, %d\n", err, site->file, site->line,

site->code);

I should mention that this trace_printk() was unintentional.

Regards,
--
Alex

maintainer
#3 lore raw
In reply to Alexander Shishkin

include/linux/perf_event.h struct perf_guest_info_callbacks { · line 102

Quoted 8 lines · Alexander Shishkin
include/linux/perf_event.h +8 -0
+/*
+ * Use part of the [-1, -MAX_ERRNO] errno range for perf's extended error
+ * reporting. Anything within [-PERF_ERRNO, -MAX_ERRNO] is an index of a
+ * perf_err_site structure within __perf_err section. 3.5k should be enough
+ * for everybody, but let's add a boot-time warning just in case it overflows
+ * one day.
+ */
+#define PERF_ERRNO 512

linux/errno.h has 512-529 in use, so this range overlaps with 'normal' errno values.

maintainer
#4 lore raw
In reply to Alexander Shishkin

include/linux/perf_event.h struct perf_guest_info_callbacks { · line 93

Quoted 11 lines · Alexander Shishkin
include/linux/perf_event.h +11 -0
+#define __perf_err(__e, __c, __m) ({				\
+	static struct perf_err_site				\
+	__attribute__ ((unused,__section__("__perf_err")))	\
+	__err_site = {						\
+		.message	= (__m),			\
+		.owner		= PERF_MODNAME,			\
+		.code		= __builtin_constant_p((__c)) ?	\
+		(__c) : 0,					\
+	};							\
+	(__e) = &__err_site;					\
+})

Why is __e an argument at all? Why not simply return &__err_site ?

maintainer
#5 lore raw
In reply to Alexander Shishkin
Quoted 15 lines · Alexander Shishkin
kernel/events/core.c +15 -0
+static void perf_error_report_site(struct perf_event_attr *attr,
+				   const struct perf_err_site *site)
+{
+	void *buffer;
+
+	if (!site || !extended_reporting_enabled(attr))
+		return;
+
+	/* in case of nested perf_err()s, which you shouldn't really do */
+	while (site->code <= -PERF_ERRNO)
+		site = perf_errno_to_site(site->code);
+
+	buffer = kasprintf(GFP_KERNEL,
+			   "{\n"
+			   "\t\"code\": %d,\n"

kernel/events/core.c line 88

Quoted 10 lines
kernel/events/core.c +10 -0
+			   "\t\"module\": \"%s\",\n"
+			   "\t\"message\": \"%s\"\n"
+			   "}\n",
+			   site->code, site->owner, site->message
+			   );
+	if (!buffer)
+		return;
+
+	if (copy_to_user((void __user *)attr->perf_err, buffer,
+			 attr->perf_err_size)) {

Should that not be min(attr->perf_err_size, strlen(buffer)) ?

Also, should we not '\0' the last char in attr->perf_err in case buffer is longer.

kernel/events/core.c line 91

Quoted 3 lines
kernel/events/core.c +3 -0
+		/* if we failed to copy once, don't bother later */
+		attr->perf_err_size = 0;
+	}

So we want update the user's perf_err_size with the actual size we copied in?

Quoted 3 lines
kernel/events/core.c +3 -0
+
+	kfree(buffer);
+}
author reviewer
#6 lore raw
Quoted 20 lines · Peter Zijlstra
Quoted 19 lines · Alexander Shishkin
kernel/events/core.c +19 -0
+static void perf_error_report_site(struct perf_event_attr *attr,
+				   const struct perf_err_site *site)
+{
+	void *buffer;
+
+	if (!site || !extended_reporting_enabled(attr))
+		return;
+
+	/* in case of nested perf_err()s, which you shouldn't really do */
+	while (site->code <= -PERF_ERRNO)
+		site = perf_errno_to_site(site->code);
+
+	buffer = kasprintf(GFP_KERNEL,
+			   "{\n"
+			   "\t\"code\": %d,\n"
+			   "\t\"module\": \"%s\",\n"
+			   "\t\"message\": \"%s\"\n"
+			   "}\n",
+			   site->code, site->owner, site->message
Quoted 8 lines

kernel/events/core.c line 88

Quoted 6 lines
kernel/events/core.c +6 -0
+			   );
+	if (!buffer)
+		return;
+
+	if (copy_to_user((void __user *)attr->perf_err, buffer,
+			 attr->perf_err_size)) {

Should that not be min(attr->perf_err_size, strlen(buffer)) ?

Indeed.

Quoted 2 lines

Also, should we not '\0' the last char in attr->perf_err in case buffer is longer.

I'm guessing we should, because otherwise the user is likely to get all confused. They can tell, however, if the buffer they allocated was too small by looking for (and not finding) the closing bracket.

Quoted 7 lines

kernel/events/core.c line 91

Quoted 4 lines
kernel/events/core.c +3 -0
+		/* if we failed to copy once, don't bother later */
+		attr->perf_err_size = 0;
+	}

So we want update the user's perf_err_size with the actual size we copied in?

We *could* do that for the perf_event_open() case where we have the user's copy of the attribute and can mess around with it, but as you noticed we're also using this error reporting in the other syscalls that involve perf events and with those we won't have the user's attribute any more.

Regards,
--
Alex

maintainer
#7 lore raw
In reply to Alexander Shishkin
Quoted 10 lines · Alexander Shishkin
kernel/events/core.c +1 -1
@@ -3890,7 +3958,7 @@ perf_read(struct file *file, char __user *buf, size_t count, loff_t *ppos)
 	ret = perf_read_hw(event, buf, count);
 	perf_event_ctx_unlock(event, ctx);
 
-	return ret;
+	return perf_error_report(&event->attr, ret);
 }
 
 static unsigned int perf_poll(struct file *file, poll_table *wait)

kernel/events/core.c static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg) · line 4177

Quoted 8 lines
kernel/events/core.c +1 -1
@@ -4103,7 +4171,7 @@ static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
 	ret = _perf_ioctl(event, cmd, arg);
 	perf_event_ctx_unlock(event, ctx);
 
-	return ret;
+	return perf_error_report(&event->attr, ret);
 }
 
 #ifdef CONFIG_COMPAT

Oooh, shiny.. So not only does it work for SYSCALL time fails, it keeps the pointer about and reports all subsequent fails in there too.

author reviewer
#8 lore raw

RFC PATCH 2/4 perf: Add file name and line number to perf extended error reports

In reply to Alexander Shishkin

If kernel debugging is enabled, include additional information to extended syscall error reports: file name and line number, to make error hunting even easier.

And in really desperate times, this allows for such things (a variation on Hugh Dickins' trick [1]):

	#undef EINVAL
	#define EINVAL perf_err(-22, "")

However, CONFIG_DEBUG_KERNEL might not be the right option for this as I notice that, for example, debian enables it in its kernels.

[1] http://marc.info/?l=linux-kernel&m=141215166009542

---
 include/linux/perf_event.h | 11 +++++++++++
 kernel/events/core.c       |  7 +++++++
 2 files changed, 18 insertions(+)
include/linux/perf_event.h +11 -0
index 9e9af962f6..d84a2914f5 100644
@@ -68,11 +68,21 @@ struct perf_guest_info_callbacks {
 struct perf_err_site {
 	const char		*message;
 	const char		*owner;
+#ifdef CONFIG_DEBUG_KERNEL
+	const char		*file;
+	const int		line;
+#endif
 	const int		code;
 };
 
 #ifdef CONFIG_PERF_EVENTS
 
+#ifdef CONFIG_DEBUG_KERNEL
+#define DEBUG_PERF_ERR	.file = __FILE__, .line = __LINE__,
+#else
+#define DEBUG_PERF_ERR
+#endif
+
 /*
  * Place all occurrences of struct perf_err_site into a special section,
  * so that we can find out their offsets, which we'll use to refer back
@@ -88,6 +98,7 @@ extern const struct perf_err_site __start___perf_err[], __stop___perf_err[];
 		.owner		= PERF_MODNAME,			\
 		.code		= __builtin_constant_p((__c)) ?	\
 		(__c) : 0,					\
+		DEBUG_PERF_ERR					\
 	};							\
 	(__e) = &__err_site;					\
 })
kernel/events/core.c +7 -0
index 85bcf3a5f9..4ff09fdc67 100644
@@ -75,10 +75,17 @@ static void perf_error_report_site(struct perf_event_attr *attr,
 
 	buffer = kasprintf(GFP_KERNEL,
 			   "{\n"
+#ifdef CONFIG_DEBUG_KERNEL
+			   "\t\"file\": \"%s\",\n"
+			   "\t\"line\": %d,\n"
+#endif
 			   "\t\"code\": %d,\n"
 			   "\t\"module\": \"%s\",\n"
 			   "\t\"message\": \"%s\"\n"
 			   "}\n",
+#ifdef CONFIG_DEBUG_KERNEL
+			   site->file, site->line,
+#endif
 			   site->code, site->owner, site->message
 			   );
 	if (!buffer)
Signature · 1 line
2.1.4
author reviewer
#9 lore raw

RFC PATCH 3/4 perf: Annotate some of the error codes with perf_err()

In reply to Alexander Shishkin

This patch annotates a few semi-random error paths in perf core to illustrate the extended error reporting facility. Most of them can be triggered from perf tools.

---
 kernel/events/core.c | 20 +++++++++++---------
 1 file changed, 11 insertions(+), 9 deletions(-)
kernel/events/core.c +11 -9
index 4ff09fdc67..e1d9fe6f23 100644
@@ -3367,10 +3367,10 @@ find_lively_task_by_vpid(pid_t vpid)
 	rcu_read_unlock();
 
 	if (!task)
-		return ERR_PTR(-ESRCH);
+		return PERF_ERR_PTR(-ESRCH, "task not found");
 
 	/* Reuse ptrace permission checks for now. */
-	err = -EACCES;
+	err = perf_err(-EACCES, "insufficient permissions for tracing this task");
 	if (!ptrace_may_access(task, PTRACE_MODE_READ))
 		goto errout;
 
@@ -3398,7 +3398,8 @@ find_get_context(struct pmu *pmu, struct task_struct *task,
 	if (!task) {
 		/* Must be root to operate on a CPU event: */
 		if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
-			return ERR_PTR(-EACCES);
+			return PERF_ERR_PTR(-EACCES,
+					    "must be root to operate on a CPU event");
 
 		/*
 		 * We could be clever and allow to attach a event to an
@@ -3406,7 +3407,7 @@ find_get_context(struct pmu *pmu, struct task_struct *task,
 		 * that's for later.
 		 */
 		if (!cpu_online(cpu))
-			return ERR_PTR(-ENODEV);
+			return PERF_ERR_PTR(-ENODEV, "cpu is offline");
 
 		cpuctx = per_cpu_ptr(pmu->pmu_cpu_context, cpu);
 		ctx = &cpuctx->ctx;
@@ -7985,15 +7986,16 @@ SYSCALL_DEFINE5(perf_event_open,
 
 	if (!attr.exclude_kernel) {
 		if (perf_paranoid_kernel() && !capable(CAP_SYS_ADMIN))
-			return -EACCES;
+			return perf_err_sync(&attr, -EACCES,
+					     "kernel tracing forbidden for the unprivileged");
 	}
 
 	if (attr.freq) {
 		if (attr.sample_freq > sysctl_perf_event_sample_rate)
-			return -EINVAL;
+			return perf_err_sync(&attr, -EINVAL, "sample_freq too high");
 	} else {
 		if (attr.sample_period & (1ULL << 63))
-			return -EINVAL;
+			return perf_err_sync(&attr, -EINVAL, "sample_period too high");
 	}
 
 	/*
@@ -8003,14 +8005,14 @@ SYSCALL_DEFINE5(perf_event_open,
 	 * cgroup.
 	 */
 	if ((flags & PERF_FLAG_PID_CGROUP) && (pid == -1 || cpu == -1))
-		return -EINVAL;
+		return perf_err_sync(&attr, -EINVAL, "pid and cpu need to be set in cgroup mode");
 
 	if (flags & PERF_FLAG_FD_CLOEXEC)
 		f_flags |= O_CLOEXEC;
 
 	event_fd = get_unused_fd_flags(f_flags);
 	if (event_fd < 0)
-		return event_fd;
+		return perf_err_sync(&attr, event_fd, "can't obtain a file descriptor");
 
 	if (group_fd != -1) {
 		err = perf_fget_light(group_fd, &group);
Signature · 1 line
2.1.4
author reviewer
#10 lore raw

RFC PATCH 4/4 perf/x86: Annotate some of the error codes with perf_err()

In reply to Alexander Shishkin

This patch annotates a few x86-specific error paths with perf's extended error reporting facility.

---
 arch/x86/kernel/cpu/perf_event.c           | 8 ++++++--
 arch/x86/kernel/cpu/perf_event_intel_lbr.c | 2 +-
 2 files changed, 7 insertions(+), 3 deletions(-)
arch/x86/kernel/cpu/perf_event.c +6 -2
index 3658de4790..2b22e57100 100644
@@ -12,6 +12,8 @@
  *  For licencing details see kernel-base/COPYING
  */
 
+#define PERF_MODNAME	"perf/x86"
+
 #include <linux/perf_event.h>
 #include <linux/capability.h>
 #include <linux/notifier.h>
@@ -426,11 +428,13 @@ int x86_setup_perfctr(struct perf_event *event)
 
 		/* BTS is currently only allowed for user-mode. */
 		if (!attr->exclude_kernel)
-			return -EOPNOTSUPP;
+			return perf_err(-EOPNOTSUPP,
+					"BTS sampling not allowed for kernel space");
 
 		/* disallow bts if conflicting events are present */
 		if (x86_add_exclusive(x86_lbr_exclusive_lbr))
-			return -EBUSY;
+			return perf_err(-EBUSY,
+					"LBR conflicts with active events");
 
 		event->destroy = hw_perf_lbr_event_destroy;
 	}
arch/x86/kernel/cpu/perf_event_intel_lbr.c +1 -1
index 452a7bd2de..b004fb310f 100644
@@ -577,7 +577,7 @@ int intel_pmu_setup_lbr_filter(struct perf_event *event)
 	 * no LBR on this PMU
 	 */
 	if (!x86_pmu.lbr_nr)
-		return -EOPNOTSUPP;
+		return perf_err(-EOPNOTSUPP, "LBR is not supported by this cpu");
 
 	/*
 	 * setup SW LBR filter
Signature · 1 line
2.1.4
maintainer
#11 lore raw
In reply to Alexander Shishkin
Quoted 14 lines · Alexander Shishkin

Hi Peter and Ingo and everybody,

Here's my second stab at improving perf's error reporting by attaching arbitrary strings to the integer error codes. This is largely based off of the previous email thread [1].

This time around, I employed a linker trick to convert the structures containing extended error information into integers, which are then made to look just like normal error codes so that IS_ERR_VALUE() and friends would still work correctly on them. So no extra pointers in the struct perf_event or anywhere else; the extended error codes are passed around like normal error codes. They only need to be converted in syscalls' topmost return statements. This is done in 1/4.

Quoted 5 lines

Then, 2/4 illustrates how error sites can be extended to include more information such as file names and line numbers [1].

The other two patches add perf_err() annotation to a few semi-randomly picked places in perf core (3/4) and x86 bits (4/4).

Looks generally ok to me. Thanks for doing this.

#12 lore raw
Quoted 15 lines · Peter Zijlstra
Quoted 14 lines · Alexander Shishkin

Hi Peter and Ingo and everybody,

Here's my second stab at improving perf's error reporting by attaching arbitrary strings to the integer error codes. This is largely based off of the previous email thread [1].

This time around, I employed a linker trick to convert the structures containing extended error information into integers, which are then made to look just like normal error codes so that IS_ERR_VALUE() and friends would still work correctly on them. So no extra pointers in the struct perf_event or anywhere else; the extended error codes are passed around like normal error codes. They only need to be converted in syscalls' topmost return statements. This is done in 1/4.

Quoted 7 lines
Quoted 5 lines

Then, 2/4 illustrates how error sites can be extended to include more information such as file names and line numbers [1].

The other two patches add perf_err() annotation to a few semi-randomly picked places in perf core (3/4) and x86 bits (4/4).

Looks generally ok to me. Thanks for doing this.

I like this too.

Alexander, mind sending a finalized, signed off version?

Thanks,

Ingo

author reviewer
#13 lore raw
Quoted 20 lines · Ingo Molnar
Quoted 18 lines · Peter Zijlstra
Quoted 17 lines · Alexander Shishkin

Hi Peter and Ingo and everybody,

Here's my second stab at improving perf's error reporting by attaching arbitrary strings to the integer error codes. This is largely based off of the previous email thread [1].

This time around, I employed a linker trick to convert the structures containing extended error information into integers, which are then made to look just like normal error codes so that IS_ERR_VALUE() and friends would still work correctly on them. So no extra pointers in the struct perf_event or anywhere else; the extended error codes are passed around like normal error codes. They only need to be converted in syscalls' topmost return statements. This is done in 1/4.

Then, 2/4 illustrates how error sites can be extended to include more information such as file names and line numbers [1].

Quoted 8 lines
Quoted 4 lines
Quoted 2 lines

The other two patches add perf_err() annotation to a few semi-randomly picked places in perf core (3/4) and x86 bits (4/4).

Looks generally ok to me. Thanks for doing this.

I like this too.

Alexander, mind sending a finalized, signed off version?

Sure, I have everything ready, except for what about 2/4 (using CONFIG_DEBUG_KERNEL to extend output with file name and line number)? Should I leave it out or can we pick a more specific kconfig option or add a new one?

Regards,
--
Alex

#14 lore raw
Quoted 27 lines · Alexander Shishkin
Quoted 25 lines · Ingo Molnar
Quoted 22 lines · Peter Zijlstra
Quoted 14 lines · Alexander Shishkin

Hi Peter and Ingo and everybody,

Here's my second stab at improving perf's error reporting by attaching arbitrary strings to the integer error codes. This is largely based off of the previous email thread [1].

This time around, I employed a linker trick to convert the structures containing extended error information into integers, which are then made to look just like normal error codes so that IS_ERR_VALUE() and friends would still work correctly on them. So no extra pointers in the struct perf_event or anywhere else; the extended error codes are passed around like normal error codes. They only need to be converted in syscalls' topmost return statements. This is done in 1/4.

Quoted 5 lines

Then, 2/4 illustrates how error sites can be extended to include more information such as file names and line numbers [1].

The other two patches add perf_err() annotation to a few semi-randomly picked places in perf core (3/4) and x86 bits (4/4).

Looks generally ok to me. Thanks for doing this.

Quoted 7 lines
Quoted 3 lines

I like this too.

Alexander, mind sending a finalized, signed off version?

Sure, I have everything ready, except for what about 2/4 (using CONFIG_DEBUG_KERNEL to extend output with file name and line number)? Should I leave it out or can we pick a more specific kconfig option or add a new one?

I'd make it unconditional. We'll see whether we compress the debug info some more if the number of callsites increases dramatically. Tooling can decide whether it wants to display it by default, or print it only if some verbosity option is activated.

Also, mind structuring it in a bit more generic way that makes it possible for other kernel subsystems to use this facility too? I.e. add a new header for it instead of embedding it in perf, etc. Nothing complex, just some common-sense generalization for a useful looking facility.

For example the scheduler could start using it in struct sched_attr and feed back the 15+ of failure causes in sys_sched_setattr() / __sched_setscheduler() in a bit more usable fashion.

Thanks,

Ingo

author reviewer
#15 lore raw
Quoted 33 lines · Ingo Molnar
Quoted 31 lines · Alexander Shishkin
Quoted 28 lines · Ingo Molnar
Quoted 15 lines · Peter Zijlstra
Quoted 14 lines · Alexander Shishkin

Hi Peter and Ingo and everybody,

Here's my second stab at improving perf's error reporting by attaching arbitrary strings to the integer error codes. This is largely based off of the previous email thread [1].

This time around, I employed a linker trick to convert the structures containing extended error information into integers, which are then made to look just like normal error codes so that IS_ERR_VALUE() and friends would still work correctly on them. So no extra pointers in the struct perf_event or anywhere else; the extended error codes are passed around like normal error codes. They only need to be converted in syscalls' topmost return statements. This is done in 1/4.

Quoted 7 lines
Quoted 5 lines

Then, 2/4 illustrates how error sites can be extended to include more information such as file names and line numbers [1].

The other two patches add perf_err() annotation to a few semi-randomly picked places in perf core (3/4) and x86 bits (4/4).

Looks generally ok to me. Thanks for doing this.

I like this too.

Alexander, mind sending a finalized, signed off version?

Quoted 8 lines
Quoted 3 lines

Sure, I have everything ready, except for what about 2/4 (using CONFIG_DEBUG_KERNEL to extend output with file name and line number)? Should I leave it out or can we pick a more specific kconfig option or add a new one?

I'd make it unconditional. We'll see whether we compress the debug info some more if the number of callsites increases dramatically. Tooling can decide whether it wants to display it by default, or print it only if some verbosity option is activated.

Done.

Quoted 4 lines

Also, mind structuring it in a bit more generic way that makes it possible for other kernel subsystems to use this facility too? I.e. add a new header for it instead of embedding it in perf, etc. Nothing complex, just some common-sense generalization for a useful looking facility.

Yes, now I have both a perf specific and a more generic implementation that perf then makes use of. Now let's arrive at a common denominator in the other thread, I guess, and then we can go ahead with either version.

Quoted 3 lines

For example the scheduler could start using it in struct sched_attr and feed back the 15+ of failure causes in sys_sched_setattr() / __sched_setscheduler() in a bit more usable fashion.

Yes, what I posted yesterday should be useful for this case as well.

Regards,
--
Alex