Skip to contentSkip to conversation

Implement /proc/<pid>/totmaps

PATCH Superseded [email protected] 7 participants 49 messages Aug 12, 2016 – Sep 30, 2016

Subsystems (3)

v2 of 5 Superseded Latest is v5 →

lore
Share & tools
lore
#0 lore raw

v2 cover Implement /proc/<pid>/totmaps

From: Robert Foss <[email protected]>

This series implements /proc/PID/totmaps, a tool for retrieving summarized information about the mappings of a process.

Changes since v1:
- Removed IS_ERR check from get_task_mm() function
- Changed comment format
- Moved proc_totmaps_operations declaration inside internal.h
- Switched to using do_maps_open() in totmaps_open() function, which provides privilege checking
- Error handling reworked for totmaps_open() function
- Switched to stack allocated struct mem_size_stats mss_sum in totmaps_proc_show() function
- Removed get_task_mm() in totmaps_proc_show() since priv->mm already is available
- Added support to proc_map_release() fork priv==NULL, to allow function to be used for all failure cases
- Added proc_totmaps_op and for it helper functions
- Added documention in separate patch
- Removed totmaps_release() since it was just a wrapper for proc_map_release()

Robert Foss (3):
mm, proc: Implement /proc/<pid>/totmaps Documentation/filesystems: Fixed typo Documentation/filesystems: Added /proc/PID/totmaps documentation

 Documentation/filesystems/proc.txt |  23 ++++++-
 fs/proc/base.c                     |   1 +
 fs/proc/internal.h                 |   3 +
 fs/proc/task_mmu.c                 | 134 +++++++++++++++++++++++++++++++++++++
 4 files changed, 160 insertions(+), 1 deletion(-)
Signature · 1 line
2.7.4
#1 lore raw

PATCH v2 1/3 mm, proc: Implement /proc/<pid>/totmaps

From: Robert Foss <[email protected]>

This is based on earlier work by Thiago Goncales. It implements a new per process proc file which summarizes the contents of the smaps file but doesn't display any addresses. It gives more detailed information than statm like the PSS (proprotional set size). It differs from the original implementation in that it doesn't use the full blown set of seq operations, uses a different termination condition, and doesn't displayed "Locked" as that was broken on the original implemenation.

This new proc file provides information faster than parsing the potentially huge smaps file.

---
 fs/proc/base.c     |   1 +
 fs/proc/internal.h |   3 ++
 fs/proc/task_mmu.c | 134 +++++++++++++++++++++++++++++++++++++++++++++++++++++
 3 files changed, 138 insertions(+)
fs/proc/base.c +1 -0
index a11eb71..de3acdf 100644
@@ -2855,6 +2855,7 @@ static const struct pid_entry tgid_base_stuff[] = {
 	REG("clear_refs", S_IWUSR, proc_clear_refs_operations),
 	REG("smaps",      S_IRUGO, proc_pid_smaps_operations),
 	REG("pagemap",    S_IRUSR, proc_pagemap_operations),
+	REG("totmaps",    S_IRUGO, proc_totmaps_operations),
 #endif
 #ifdef CONFIG_SECURITY
 	DIR("attr",       S_IRUGO|S_IXUGO, proc_attr_dir_inode_operations, proc_attr_dir_operations),
fs/proc/internal.h +3 -0
index aa27810..c55e1fe 100644
@@ -281,6 +281,7 @@ struct proc_maps_private {
 	struct mm_struct *mm;
 #ifdef CONFIG_MMU
 	struct vm_area_struct *tail_vma;
+	struct mem_size_stats *mss;
 #endif
 #ifdef CONFIG_NUMA
 	struct mempolicy *task_mempolicy;
@@ -297,6 +298,8 @@ extern const struct file_operations proc_pid_smaps_operations;
 extern const struct file_operations proc_tid_smaps_operations;
 extern const struct file_operations proc_clear_refs_operations;
 extern const struct file_operations proc_pagemap_operations;
+extern const struct file_operations proc_totmaps_operations;
+
 
 extern unsigned long task_vsize(struct mm_struct *);
 extern unsigned long task_statm(struct mm_struct *,
fs/proc/task_mmu.c +134 -0

164 lines not on this page · raw message

Signature · 1 line
2.7.4
#2 lore raw

Re: [PACTH v2 1/3] mm, proc: Implement /proc/<pid>/totmaps

fs/proc/internal.h struct proc_maps_private { · line 284

Quoted 9 lines · On Fri, Aug 12, 2016 at 06:04:20PM -0400, [email protected] wrote:
fs/proc/internal.h +1 -0
index aa27810..c55e1fe 100644
@@ -281,6 +281,7 @@ struct proc_maps_private {
 	struct mm_struct *mm;
 #ifdef CONFIG_MMU
 	struct vm_area_struct *tail_vma;
+	struct mem_size_stats *mss;

This is unused now, right?

fs/proc/task_mmu.c static int proc_map_release(struct inode *inode, struct file *file) · line 251

Quoted 11 lines
fs/proc/task_mmu.c +3 -0
index 4648c7f..b7612e9 100644
@@ -246,6 +246,9 @@ static int proc_map_release(struct inode *inode, struct file *file)
 	struct seq_file *seq = file->private_data;
 	struct proc_maps_private *priv = seq->private;
 
+	if (!priv)
+		return 0;
+

You might want to get rid of this, see below.

Quoted 9 lines
fs/proc/task_mmu.c +9 -0
+static int totmaps_open(struct inode *inode, struct file *file)
+{
+	struct proc_maps_private *priv = NULL;
+	struct seq_file *seq;
+	int ret;
+
+	ret = do_maps_open(inode, file, &proc_totmaps_op);
+	if (ret)
+		goto error;

[...]

fs/proc/task_mmu.c static int tid_smaps_open(struct inode *inode, struct file *file) · line 963

Quoted 3 lines
fs/proc/task_mmu.c +3 -0
+error:
+	proc_map_release(inode, file);
+	return ret;

I don't think this is correct. Have a look at the other callers of do_maps_open() - none of them do any cleanup steps on error, they just return. I think the "goto error" here should be a return instead.

Have a look at the error cases that can cause do_maps_open() to
fail: do_maps_open() just calls proc_maps_open(). If the __seq_open_private() call fails because of memory pressure, file->private_data is still NULL, and your newly added NULL check in proc_map_release() causes proc_map_release() to be a no-op there. But if proc_maps_open() fails later on, things get nasty:
If, for example, proc_mem_open() fails because of a ptrace permission denial, __seq_open_file -> seq_open has already set file->private_data to a struct seq_file *, and then proc_maps_open(), prior to passing on the error code, calls seq_release_private -> seq_release, which frees that struct seq_file * without NULLing the private_data pointer. As far as I can tell, proc_map_release() would then run into a use-after-free scenario.

fs/proc/task_mmu.c static int tid_smaps_open(struct inode *inode, struct file *file) · line 957

Quoted 5 lines
fs/proc/task_mmu.c +5 -0
+	priv->task = get_proc_task(inode);
+	if (!priv->task) {
+		ret = -ESRCH;
+		goto error;
+	}

You're not actually using ->task anywhere in the current version, right? Can this be deleted?

Quoted 1 line

+const struct file_operations proc_totmaps_operations = {

[...]

fs/proc/task_mmu.c const struct file_operations proc_tid_smaps_operations = { · line 984

Quoted 1 line
fs/proc/task_mmu.c +1 -0
+	.release	= proc_map_release,

This won't release priv->task, causing a memory leak (exploitable through a reference counter overflow of the task_struct usage counter).

#3 lore raw

Re: [PACTH v2 1/3] mm, proc: Implement /proc/<pid>/totmaps

Quoted 12 lines · Jann Horn

fs/proc/internal.h struct proc_maps_private { · line 284

Quoted 9 lines · On Fri, Aug 12, 2016 at 06:04:20PM -0400, [email protected] wrote:
fs/proc/internal.h +1 -0
index aa27810..c55e1fe 100644
@@ -281,6 +281,7 @@ struct proc_maps_private {
 	struct mm_struct *mm;
 #ifdef CONFIG_MMU
 	struct vm_area_struct *tail_vma;
+	struct mem_size_stats *mss;

This is unused now, right?

Fixing it in v3.

Quoted 10 lines
Quoted 10 lines
fs/proc/task_mmu.c +0 -0
index 4648c7f..b7612e9 100644
@@ -246,6 +246,9 @@ static int proc_map_release(struct inode *inode, struct file *file)
 	struct seq_file *seq = file->private_data;
 	struct proc_maps_private *priv = seq->private;
 
Quoted 5 lines

fs/proc/task_mmu.c static int proc_map_release(struct inode *inode, struct file *file) · line 251

Quoted 3 lines
fs/proc/task_mmu.c +3 -0
+	if (!priv)
+		return 0;
+

You might want to get rid of this, see below.

Fixing it in v3.

Quoted 47 lines
Quoted 11 lines
fs/proc/task_mmu.c +9 -0
+static int totmaps_open(struct inode *inode, struct file *file)
+{
+	struct proc_maps_private *priv = NULL;
+	struct seq_file *seq;
+	int ret;
+
+	ret = do_maps_open(inode, file, &proc_totmaps_op);
+	if (ret)
+		goto error;

[...]

fs/proc/task_mmu.c static int tid_smaps_open(struct inode *inode, struct file *file) · line 963

Quoted 3 lines
fs/proc/task_mmu.c +3 -0
+error:
+	proc_map_release(inode, file);
+	return ret;

I don't think this is correct. Have a look at the other callers of do_maps_open() - none of them do any cleanup steps on error, they just return. I think the "goto error" here should be a return instead.

Have a look at the error cases that can cause do_maps_open() to
fail: do_maps_open() just calls proc_maps_open(). If the __seq_open_private() call fails because of memory pressure, file->private_data is still NULL, and your newly added NULL check in proc_map_release() causes proc_map_release() to be a no-op there. But if proc_maps_open() fails later on, things get nasty:
If, for example, proc_mem_open() fails because of a ptrace permission denial, __seq_open_file -> seq_open has already set file->private_data to a struct seq_file *, and then proc_maps_open(), prior to passing on the error code, calls seq_release_private -> seq_release, which frees that struct seq_file * without NULLing the private_data pointer. As far as I can tell, proc_map_release() would then run into a use-after-free scenario.

fs/proc/task_mmu.c static int tid_smaps_open(struct inode *inode, struct file *file) · line 957

Quoted 5 lines
fs/proc/task_mmu.c +5 -0
+	priv->task = get_proc_task(inode);
+	if (!priv->task) {
+		ret = -ESRCH;
+		goto error;
+	}

You're not actually using ->task anywhere in the current version, right? Can this be deleted?

Quoted 8 lines
Quoted 1 line

+const struct file_operations proc_totmaps_operations = {

[...]

fs/proc/task_mmu.c const struct file_operations proc_tid_smaps_operations = { · line 984

Quoted 1 line
fs/proc/task_mmu.c +1 -0
+	.release	= proc_map_release,

This won't release priv->task, causing a memory leak (exploitable through a reference counter overflow of the task_struct usage counter).

Thanks for the thorough walkthrough, it is much appreciated.

priv->task does not appear to be used any more, and can be removed. When "priv->task = get_proc_task(inode)" is removed, totmaps_open() starts to look just like the other XXX_open functions.

I'll send out v3 as soon as testing has been done.

#4 lore raw

Re: [PACTH v2 1/3] mm, proc: Implement /proc/<pid>/totmaps

Quoted 30 lines · Robert Foss
Quoted 13 lines · Jann Horn

fs/proc/internal.h struct proc_maps_private { · line 284

Quoted 9 lines · On Fri, Aug 12, 2016 at 06:04:20PM -0400, [email protected] wrote:
fs/proc/internal.h +1 -0
index aa27810..c55e1fe 100644
@@ -281,6 +281,7 @@ struct proc_maps_private {
     struct mm_struct *mm;
 #ifdef CONFIG_MMU
     struct vm_area_struct *tail_vma;
+    struct mem_size_stats *mss;

This is unused now, right?

Fixing it in v3.

Quoted 11 lines
Quoted 11 lines
fs/proc/task_mmu.c +0 -0
index 4648c7f..b7612e9 100644
@@ -246,6 +246,9 @@ static int proc_map_release(struct inode *inode,
 struct file *file)
     struct seq_file *seq = file->private_data;
     struct proc_maps_private *priv = seq->private;
 
Quoted 53 lines
Quoted 5 lines

fs/proc/task_mmu.c static int proc_map_release(struct inode *inode, struct file *file) · line 251

Quoted 3 lines
fs/proc/task_mmu.c +3 -0
+    if (!priv)
+        return 0;
+

You might want to get rid of this, see below.

Fixing it in v3.

Quoted 45 lines
Quoted 11 lines
fs/proc/task_mmu.c +9 -0
+static int totmaps_open(struct inode *inode, struct file *file)
+{
+    struct proc_maps_private *priv = NULL;
+    struct seq_file *seq;
+    int ret;
+
+    ret = do_maps_open(inode, file, &proc_totmaps_op);
+    if (ret)
+        goto error;

[...]

fs/proc/task_mmu.c static int tid_smaps_open(struct inode *inode, struct file *file) · line 963

Quoted 3 lines
fs/proc/task_mmu.c +3 -0
+error:
+    proc_map_release(inode, file);
+    return ret;

I don't think this is correct. Have a look at the other callers of do_maps_open() - none of them do any cleanup steps on error, they just return. I think the "goto error" here should be a return instead.

Have a look at the error cases that can cause do_maps_open() to
fail: do_maps_open() just calls proc_maps_open(). If the __seq_open_private() call fails because of memory pressure, file->private_data is still NULL, and your newly added NULL check in proc_map_release() causes proc_map_release() to be a no-op there. But if proc_maps_open() fails later on, things get nasty:
If, for example, proc_mem_open() fails because of a ptrace permission denial, __seq_open_file -> seq_open has already set file->private_data to a struct seq_file *, and then proc_maps_open(), prior to passing on the error code, calls seq_release_private -> seq_release, which frees that struct seq_file * without NULLing the private_data pointer. As far as I can tell, proc_map_release() would then run into a use-after-free scenario.

fs/proc/task_mmu.c static int tid_smaps_open(struct inode *inode, struct file *file) · line 957

Quoted 5 lines
fs/proc/task_mmu.c +5 -0
+    priv->task = get_proc_task(inode);
+    if (!priv->task) {
+        ret = -ESRCH;
+        goto error;
+    }

You're not actually using ->task anywhere in the current version, right? Can this be deleted?

Actually, priv->task is used by hold_task_mempolicy() in totmaps_proc_show(), which as far as I understand it is needed due to the "vma = mm->mmap" looping we do.

Quoted 18 lines
Quoted 10 lines
Quoted 3 lines

+const struct file_operations proc_totmaps_operations = {

[...]

Quoted 1 line

+ .release = proc_map_release,

This won't release priv->task, causing a memory leak (exploitable through a reference counter overflow of the task_struct usage counter).

Thanks for the thorough walkthrough, it is much appreciated.

priv->task does not appear to be used any more, and can be removed. When "priv->task = get_proc_task(inode)" is removed, totmaps_open() starts to look just like the other XXX_open functions.

I'll send out v3 as soon as testing has been done.

#5 lore raw

PATCH v2 2/3 Documentation/filesystems: Fixed typo

In reply to [email protected]

From: Robert Foss <[email protected]>

Fixed a -> an typo.

---
 Documentation/filesystems/proc.txt | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
Documentation/filesystems/proc.txt +1 -1
index e8d0075..7d001be 100644
@@ -145,7 +145,7 @@ Table 1-1: Process specific entries in /proc
 		symbol the task is blocked in - or "0" if not blocked.
  pagemap	Page table
  stack		Report full stack trace, enable via CONFIG_STACKTRACE
- smaps		a extension based on maps, showing the memory consumption of
+ smaps		an extension based on maps, showing the memory consumption of
 		each mapping and flags associated with it
  numa_maps	an extension based on maps, showing the memory locality and
 		binding policy as well as mem usage (in pages) of each mapping.
Signature · 1 line
2.7.4
#6 lore raw

PATCH v2 3/3 Documentation/filesystems: Added /proc/PID/totmaps documentation

In reply to [email protected]

From: Robert Foss <[email protected]>

Added documentation covering /proc/PID/totmaps.

---
 Documentation/filesystems/proc.txt | 21 +++++++++++++++++++++
 1 file changed, 21 insertions(+)
Documentation/filesystems/proc.txt +21 -0
index 7d001be..c06ff33 100644
@@ -11,6 +11,7 @@ Version 1.3                                              Kernel version 2.2.12
 					      Kernel version 2.4.0-test11-pre4
 ------------------------------------------------------------------------------
 fixes/update part 1.1  Stefani Seibold <[email protected]>       June 9 2009
+add totmaps            Robert Foss <[email protected]>  August 12 2016
 
 Table of Contents
 -----------------
@@ -147,6 +148,8 @@ Table 1-1: Process specific entries in /proc
  stack		Report full stack trace, enable via CONFIG_STACKTRACE
  smaps		an extension based on maps, showing the memory consumption of
 		each mapping and flags associated with it
+ totmaps        an extenssion based on maps, showing the total memory
+                consumption of all mappings
  numa_maps	an extension based on maps, showing the memory locality and
 		binding policy as well as mem usage (in pages) of each mapping.
 ..............................................................................
@@ -512,6 +515,24 @@ be vanished or the reverse -- new added.
 This file is only present if the CONFIG_MMU kernel configuration option is
 enabled.
 
+The /proc/PID/totmaps is an extension based on maps, showing the memory
+consumption totals for all of the process's mappings. It lists the sums of the
+same statistics as /proc/PID/smaps.
+
+The process' mappings will be summarized as a series of lines like the
+following:
+
+Rss:                4256 kB
+Pss:                1170 kB
+Shared_Clean:       2720 kB
+Shared_Dirty:       1136 kB
+Private_Clean:         0 kB
+Private_Dirty:       400 kB
+Referenced:         4256 kB
+Anonymous:          1536 kB
+AnonHugePages:         0 kB
+Swap:                  0 kB
+
 The /proc/PID/clear_refs is used to reset the PG_Referenced and ACCESSED/YOUNG
 bits on both physical and virtual pages associated with a process, and the
 soft-dirty bit on pte (see Documentation/vm/soft-dirty.txt for details).
Signature · 1 line
2.7.4
#7 lore raw
In reply to [email protected]
Quoted 4 lines · On Fri 12-08-16 18:04:19, [email protected] wrote:

From: Robert Foss <[email protected]>

This series implements /proc/PID/totmaps, a tool for retrieving summarized information about the mappings of a process.

The changelog is absolutely missing the usecase. Why do we need this? Why existing interfaces are not sufficient?

Quoted 33 lines

Changes since v1:
- Removed IS_ERR check from get_task_mm() function
- Changed comment format
- Moved proc_totmaps_operations declaration inside internal.h
- Switched to using do_maps_open() in totmaps_open() function, which provides privilege checking
- Error handling reworked for totmaps_open() function
- Switched to stack allocated struct mem_size_stats mss_sum in totmaps_proc_show() function
- Removed get_task_mm() in totmaps_proc_show() since priv->mm already is available
- Added support to proc_map_release() fork priv==NULL, to allow function to be used for all failure cases
- Added proc_totmaps_op and for it helper functions
- Added documention in separate patch
- Removed totmaps_release() since it was just a wrapper for proc_map_release()

Robert Foss (3):
mm, proc: Implement /proc/<pid>/totmaps Documentation/filesystems: Fixed typo Documentation/filesystems: Added /proc/PID/totmaps documentation

 Documentation/filesystems/proc.txt |  23 ++++++-
 fs/proc/base.c                     |   1 +
 fs/proc/internal.h                 |   3 +
 fs/proc/task_mmu.c                 | 134 +++++++++++++++++++++++++++++++++++++
 4 files changed, 160 insertions(+), 1 deletion(-)
Signature · 1 line
2.7.4
Signature · 2 lines
Michal Hocko
SUSE Labs
#8 lore raw
Quoted 8 lines · Michal Hocko
Quoted 4 lines · On Fri 12-08-16 18:04:19, [email protected] wrote:

From: Robert Foss <[email protected]>

This series implements /proc/PID/totmaps, a tool for retrieving summarized information about the mappings of a process.

The changelog is absolutely missing the usecase. Why do we need this? Why existing interfaces are not sufficient?

You are absolutely right, more info information is in 1/3.
But the gist of it is that it provides a faster and more convenient way of accessing the information in /proc/PID/smaps.

Quoted 35 lines
Quoted 34 lines

Changes since v1:
- Removed IS_ERR check from get_task_mm() function
- Changed comment format
- Moved proc_totmaps_operations declaration inside internal.h
- Switched to using do_maps_open() in totmaps_open() function, which provides privilege checking
- Error handling reworked for totmaps_open() function
- Switched to stack allocated struct mem_size_stats mss_sum in totmaps_proc_show() function
- Removed get_task_mm() in totmaps_proc_show() since priv->mm already is available
- Added support to proc_map_release() fork priv==NULL, to allow function to be used for all failure cases
- Added proc_totmaps_op and for it helper functions
- Added documention in separate patch
- Removed totmaps_release() since it was just a wrapper for proc_map_release()

Robert Foss (3):
mm, proc: Implement /proc/<pid>/totmaps Documentation/filesystems: Fixed typo Documentation/filesystems: Added /proc/PID/totmaps documentation

 Documentation/filesystems/proc.txt |  23 ++++++-
 fs/proc/base.c                     |   1 +
 fs/proc/internal.h                 |   3 +
 fs/proc/task_mmu.c                 | 134 +++++++++++++++++++++++++++++++++++++
 4 files changed, 160 insertions(+), 1 deletion(-)

--
2.7.4

#9 lore raw
Quoted 6 lines · Robert Foss
Quoted 3 lines · Michal Hocko
Quoted 2 lines · On Fri 12-08-16 18:04:19, [email protected] wrote:

From: Robert Foss <[email protected]>

Quoted 7 lines
Quoted 5 lines
Quoted 2 lines

This series implements /proc/PID/totmaps, a tool for retrieving summarized information about the mappings of a process.

The changelog is absolutely missing the usecase. Why do we need this? Why existing interfaces are not sufficient?

You are absolutely right, more info information is in 1/3.

Patch 1 is silent about the use case as well. It is usually recommended to describe the motivation for the change in the cover letter.

Quoted 2 lines

But the gist of it is that it provides a faster and more convenient way of accessing the information in /proc/PID/smaps.

I am sorry to insist but this is far from a description I was hoping for. Why do we need a more convenient API? Please note that this is a userspace API which we will have to maintain for ever. We have made many mistakes in the past where exporting some information made sense at the time while it turned out being a mistake only later on. So let's make sure we will not fall into the same trap again.

So please make sure you describe the use case, why the current API is insufficient and why it cannot be tweaked to provide the information you are looking for.

Signature · 2 lines
Michal Hocko
SUSE Labs
#10 lore raw
Quoted 10 lines · Michal Hocko
Quoted 9 lines · Robert Foss
Quoted 6 lines · Michal Hocko
Quoted 4 lines · On Fri 12-08-16 18:04:19, [email protected] wrote:

From: Robert Foss <[email protected]>

This series implements /proc/PID/totmaps, a tool for retrieving summarized information about the mappings of a process.

Quoted 7 lines
Quoted 4 lines
Quoted 2 lines

The changelog is absolutely missing the usecase. Why do we need this? Why existing interfaces are not sufficient?

You are absolutely right, more info information is in 1/3.

Patch 1 is silent about the use case as well. It is usually recommended to describe the motivation for the change in the cover letter.

I'll change it for v3.

Quoted 11 lines
Quoted 3 lines

But the gist of it is that it provides a faster and more convenient way of accessing the information in /proc/PID/smaps.

I am sorry to insist but this is far from a description I was hoping for. Why do we need a more convenient API? Please note that this is a userspace API which we will have to maintain for ever. We have made many mistakes in the past where exporting some information made sense at the time while it turned out being a mistake only later on. So let's make sure we will not fall into the same trap again.

Quoted 4 lines

So please make sure you describe the use case, why the current API is insufficient and why it cannot be tweaked to provide the information you are looking for.

I'll add a more elaborate description to the v3 cover letter.
In v1, there was a discussion which I think presented the practical applications rather well:

https://lkml.org/lkml/2016/8/9/628

or the qoute from Sonny Rao pasted below:

Quoted 10 lines

The use case is to speed up monitoring of
memory consumption in environments where RSS isn't precise.

For example Chrome tends to many processes which have hundreds of VMAs with a substantial amount of shared memory, and the error of using RSS rather than PSS tends to be very large when looking at overall memory consumption. PSS isn't kept as a single number that's exported like RSS, so to calculate PSS means having to parse a very large smaps file.

Quoted 4 lines

This process is slow and has to be repeated for many processes, and we found that the just act of doing the parsing was taking up a significant amount of CPU time, so this patch is an attempt to make that process cheaper.

If a reformatted version of this still isn't adequate or desirable for the cover-letter, please give me another heads up.

Thanks!

#11 lore raw
Quoted 3 lines · Robert Foss

On 2016-08-15 09:42 AM, Michal Hocko wrote:

[...]

Quoted 10 lines
Quoted 10 lines

The use case is to speed up monitoring of
memory consumption in environments where RSS isn't precise.

For example Chrome tends to many processes which have hundreds of VMAs with a substantial amount of shared memory, and the error of using RSS rather than PSS tends to be very large when looking at overall memory consumption. PSS isn't kept as a single number that's exported like RSS, so to calculate PSS means having to parse a very large smaps file.

Quoted 4 lines
Quoted 4 lines

This process is slow and has to be repeated for many processes, and we found that the just act of doing the parsing was taking up a significant amount of CPU time, so this patch is an attempt to make that process cheaper.

Well, this is slow because it requires the pte walk otherwise you cannot know how many ptes map the particular shared page. Your patch (totmaps_proc_show) does the very same page table walk because in fact it is unavoidable. So what exactly is the difference except for the userspace parsing which is quite trivial e.g. my currently running Firefox has

$ awk '/^[0-9a-f]/{print}' /proc/4950/smaps | wc -l
984

quite some VMAs, yet parsing it spends basically all the time in the kernel...

$ /usr/bin/time -v awk '/^Rss/{rss+=$2} /^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss}' /proc/4950/smaps 
rss:1112288 pss:1096435
        Command being timed: "awk /^Rss/{rss+=$2} /^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss} /proc/4950/smaps"
        User time (seconds): 0.00
        System time (seconds): 0.02
        Percent of CPU this job got: 91%
        Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.02

So I am not really sure I see the performance benefit.

Signature · 2 lines
Michal Hocko
SUSE Labs
#12 lore raw
Quoted 39 lines · Michal Hocko
Quoted 3 lines · Robert Foss

On 2016-08-15 09:42 AM, Michal Hocko wrote:

[...]

Quoted 10 lines
Quoted 10 lines

The use case is to speed up monitoring of
memory consumption in environments where RSS isn't precise.

For example Chrome tends to many processes which have hundreds of VMAs with a substantial amount of shared memory, and the error of using RSS rather than PSS tends to be very large when looking at overall memory consumption. PSS isn't kept as a single number that's exported like RSS, so to calculate PSS means having to parse a very large smaps file.

Quoted 4 lines
Quoted 4 lines

This process is slow and has to be repeated for many processes, and we found that the just act of doing the parsing was taking up a significant amount of CPU time, so this patch is an attempt to make that process cheaper.

Well, this is slow because it requires the pte walk otherwise you cannot know how many ptes map the particular shared page. Your patch (totmaps_proc_show) does the very same page table walk because in fact it is unavoidable. So what exactly is the difference except for the userspace parsing which is quite trivial e.g. my currently running Firefox has

$ awk '/^[0-9a-f]/{print}' /proc/4950/smaps | wc -l
984

quite some VMAs, yet parsing it spends basically all the time in the kernel...

$ /usr/bin/time -v awk '/^Rss/{rss+=$2} /^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss}' /proc/4950/smaps
rss:1112288 pss:1096435
        Command being timed: "awk /^Rss/{rss+=$2} /^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss} /proc/4950/smaps"
        User time (seconds): 0.00
        System time (seconds): 0.02
        Percent of CPU this job got: 91%
        Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.02
Quoted 2 lines

So I am not really sure I see the performance benefit.

I did some performance measurements of my own, and it would seem like there is about a 2x performance gain to be had. To me that is substantial, and a larger gain than commonly seen.

There naturally also the benefit that this is a lot easier to interact with programmatically.

$ ps aux | grep firefox
robertfoss   5025 24.3 13.7 3562820 2219616 ?     Rl   Aug15 277:44 /usr/lib/firefox/firefox https://allg.one/xpb
$ awk '/^[0-9a-f]/{print}' /proc/5025/smaps | wc -l
1503
$ /usr/bin/time -v -p zsh -c "(repeat 25 {cat /proc/5025/totmaps})"
[...]
	Command being timed: "zsh -c (repeat 25 {cat /proc/5025/totmaps})"
	User time (seconds): 0.00
	System time (seconds): 0.40
	Percent of CPU this job got: 90%
	Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.45
$ /usr/bin/time -v -p zsh -c "repeat 25 { awk '/^Rss/{rss+=\$2} /^Pss/{pss+=\$2} END {printf \"rss:%d pss:%d\n\", rss, pss}\' /proc/5025/smaps }"
[...]
	Command being timed: "zsh -c repeat 25 { awk '/^Rss/{rss+=$2} 
/^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss}\' /proc/5025/smaps }"
	User time (seconds): 0.37
	System time (seconds): 0.45
	Percent of CPU this job got: 92%
	Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.89
#13 lore raw

On Tue 16-08-16 12:46:51, Robert Foss wrote:
[...]

Quoted 11 lines
$ /usr/bin/time -v -p zsh -c "repeat 25 { awk '/^Rss/{rss+=\$2}
/^Pss/{pss+=\$2} END {printf \"rss:%d pss:%d\n\", rss, pss}\'
/proc/5025/smaps }"
[...]
	Command being timed: "zsh -c repeat 25 { awk '/^Rss/{rss+=$2}
/^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss}\' /proc/5025/smaps
}"
	User time (seconds): 0.37
	System time (seconds): 0.45
	Percent of CPU this job got: 92%
	Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.89

This is really unexpected. Where is the user time spent? Anyway, rather than measuring some random processes I've tried to measure something resembling the worst case. So I've created a simple program to mmap as much as possible:

#include <sys/mman.h>
#include <sys/types.h>
#include <unistd.h>
#include <stdio.h>
int main()

{
	while (mmap(NULL, 4096, PROT_READ|PROT_WRITE, MAP_ANON|MAP_SHARED|MAP_POPULATE, -1, 0) != MAP_FAILED)
		;

	printf("pid:%d\n", getpid());
	pause();
	return 0;
}

so depending on /proc/sys/vm/max_map_count you will get the maximum possible mmaps. I am using a default so 65k mappings. Then I have retried your 25x file parsing:

$ cat s.sh
#!/bin/sh

pid=$1
for i in $(seq 25)
do

	awk '/^Rss/{rss+=$2} /^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss}' /proc/$pid/smaps

done

But I am getting different results from you:

$ awk '/^[0-9a-f]/{print}' /proc/14808/smaps | wc -l
65532
[...]
        Command being timed: "sh s.sh 14808"
        User time (seconds): 0.00
        System time (seconds): 20.10
        Percent of CPU this job got: 99%
        Elapsed (wall clock) time (h:mm:ss or m:ss): 0:20.20

The results are stable when I try multiple times, in fact there shouldn't be any reason for them not to be. Then I went on to increase max_map_count to 250k and that behaves consistently:

$ awk '/^[0-9a-f]/{print}' /proc/16093/smaps | wc -l     
250002
[...]
        Command being timed: "sh s.sh 16093"
        User time (seconds): 0.00
        System time (seconds): 77.93
        Percent of CPU this job got: 98%
        Elapsed (wall clock) time (h:mm:ss or m:ss): 1:19.09

So with a reasonable user space the parsing is really not all that time consuming wrt. smaps handling. That being said I am still very skeptical about a dedicated proc file which accomplishes what userspace can done in a trivial way.

Signature · 2 lines
Michal Hocko
SUSE Labs
#14 lore raw
Quoted 24 lines · Michal Hocko

On Tue 16-08-16 12:46:51, Robert Foss wrote:
[...]

Quoted 11 lines
$ /usr/bin/time -v -p zsh -c "repeat 25 { awk '/^Rss/{rss+=\$2}
/^Pss/{pss+=\$2} END {printf \"rss:%d pss:%d\n\", rss, pss}\'
/proc/5025/smaps }"
[...]
	Command being timed: "zsh -c repeat 25 { awk '/^Rss/{rss+=$2}
/^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss}\' /proc/5025/smaps
}"
	User time (seconds): 0.37
	System time (seconds): 0.45
	Percent of CPU this job got: 92%
	Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.89

This is really unexpected. Where is the user time spent? Anyway, rather than measuring some random processes I've tried to measure something resembling the worst case. So I've created a simple program to mmap as much as possible:

#include <sys/mman.h>
#include <sys/types.h>
#include <unistd.h>
#include <stdio.h>
int main()

Quoted 8 lines
{
	while (mmap(NULL, 4096, PROT_READ|PROT_WRITE, MAP_ANON|MAP_SHARED|MAP_POPULATE, -1, 0) != MAP_FAILED)
		;

	printf("pid:%d\n", getpid());
	pause();
	return 0;
}

Ah, nice, that's a reasonable test program. :)

Quoted 4 lines

So with a reasonable user space the parsing is really not all that time consuming wrt. smaps handling. That being said I am still very skeptical about a dedicated proc file which accomplishes what userspace can done in a trivial way.

Now, since your numbers showed that all the time is spent in the kernel, also create this test program to just read that file over and over again:

$ cat justreadloop.c
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <sched.h>
#include <unistd.h>
#include <err.h>
#include <stdio.h>

char buf[1000000];

int main(int argc, char **argv) {
printf("pid:%d\n", getpid());

  while (1) {
    int fd = open(argv[1], O_RDONLY);
    if (fd < 0) continue;
    if (read(fd, buf, sizeof(buf)) < 0)
      err(1, "read");
    close(fd);
  }
}
$ gcc -Wall -o justreadloop justreadloop.c
$ 

Now launch your test:

$ ./mapstuff 
pid:29397

point justreadloop at it:

$ ./justreadloop /proc/29397/smaps
pid:32567

... and then check the performance stats of justreadloop:

# perf top -p 32567

This is what I see:

Samples: 232K of event 'cycles:ppp', Event count (approx.): 60448424325

Overhead  Shared Object     Symbol
  30,43%  [kernel]          [k] format_decode
   9,12%  [kernel]          [k] number
   7,66%  [kernel]          [k] vsnprintf
   7,06%  [kernel]          [k] __lock_acquire
   3,23%  [kernel]          [k] lock_release
   2,85%  [kernel]          [k] debug_lockdep_rcu_enabled
   2,25%  [kernel]          [k] skip_atoi
   2,13%  [kernel]          [k] lock_acquire
   2,05%  [kernel]          [k] show_smap

That's at least 30.43% + 9.12% + 7.66% = 47.21% of the task's kernel time spent on evaluating format strings. The new interface wouldn't have to spend that much time on format strings because there isn't so much text to format. (My kernel is built with a bunch of debug options - the results might look very different on distro kernels or so, so please try this yourself.)

I guess it could be argued that this is not just a problem with smaps, but also a problem with format strings (or text-based interfaces in general) just being slow in general.

(Here is a totally random and crazy thought: Can we put something into the kernel build process that replaces printf calls that use simple format strings with equivalent non-printf calls? Move the cost of evaluating the format string to compile time?)

#15 lore raw
Quoted 86 lines · Jann Horn
Quoted 24 lines · Michal Hocko

On Tue 16-08-16 12:46:51, Robert Foss wrote:
[...]

Quoted 11 lines
$ /usr/bin/time -v -p zsh -c "repeat 25 { awk '/^Rss/{rss+=\$2}
/^Pss/{pss+=\$2} END {printf \"rss:%d pss:%d\n\", rss, pss}\'
/proc/5025/smaps }"
[...]
	Command being timed: "zsh -c repeat 25 { awk '/^Rss/{rss+=$2}
/^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss}\' /proc/5025/smaps
}"
	User time (seconds): 0.37
	System time (seconds): 0.45
	Percent of CPU this job got: 92%
	Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.89

This is really unexpected. Where is the user time spent? Anyway, rather than measuring some random processes I've tried to measure something resembling the worst case. So I've created a simple program to mmap as much as possible:

#include <sys/mman.h>
#include <sys/types.h>
#include <unistd.h>
#include <stdio.h>
int main()

Quoted 8 lines
{
	while (mmap(NULL, 4096, PROT_READ|PROT_WRITE, MAP_ANON|MAP_SHARED|MAP_POPULATE, -1, 0) != MAP_FAILED)
		;

	printf("pid:%d\n", getpid());
	pause();
	return 0;
}

Ah, nice, that's a reasonable test program. :)

Quoted 4 lines

So with a reasonable user space the parsing is really not all that time consuming wrt. smaps handling. That being said I am still very skeptical about a dedicated proc file which accomplishes what userspace can done in a trivial way.

Now, since your numbers showed that all the time is spent in the kernel, also create this test program to just read that file over and over again:

$ cat justreadloop.c
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <sched.h>
#include <unistd.h>
#include <err.h>
#include <stdio.h>

char buf[1000000];

int main(int argc, char **argv) {
printf("pid:%d\n", getpid());

  while (1) {
    int fd = open(argv[1], O_RDONLY);
    if (fd < 0) continue;
    if (read(fd, buf, sizeof(buf)) < 0)
      err(1, "read");
    close(fd);
  }
}
$ gcc -Wall -o justreadloop justreadloop.c
$ 

Now launch your test:

$ ./mapstuff 
pid:29397

point justreadloop at it:

$ ./justreadloop /proc/29397/smaps
pid:32567

... and then check the performance stats of justreadloop:

# perf top -p 32567

This is what I see:

Samples: 232K of event 'cycles:ppp', Event count (approx.): 60448424325

Quoted 10 lines
Overhead  Shared Object     Symbol
  30,43%  [kernel]          [k] format_decode
   9,12%  [kernel]          [k] number
   7,66%  [kernel]          [k] vsnprintf
   7,06%  [kernel]          [k] __lock_acquire
   3,23%  [kernel]          [k] lock_release
   2,85%  [kernel]          [k] debug_lockdep_rcu_enabled
   2,25%  [kernel]          [k] skip_atoi
   2,13%  [kernel]          [k] lock_acquire
   2,05%  [kernel]          [k] show_smap

This is a lot! I would expect the rmap walk to consume more but it even doesn't show up in the top consumers.

Quoted 4 lines

That's at least 30.43% + 9.12% + 7.66% = 47.21% of the task's kernel time spent on evaluating format strings. The new interface wouldn't have to spend that much time on format strings because there isn't so much text to format.

well, this is true of course but I would much rather try to reduce the overhead of smaps file than add a new file. The following should help already. I've measured ~7% systime cut down. I guess there is still some room for improvements but I have to say I'm far from being convinced about a new proc file just because we suck at dumping information to the userspace. If this was something like /proc/<pid>/stat which is essentially read all the time then it would be a different question but is the rss, pss going to be all that often? If yes why? These are the questions which should be answered before we even start considering the implementation.
---

Quoted 1 line

From 2a6883a7278ff8979808cb8e2dbcefe5ea3bf672 Mon Sep 17 00:00:00 2001

From: Michal Hocko <[email protected]>
Date: Wed, 17 Aug 2016 14:00:13 +0200
Subject: [PATCH] proc, smaps: reduce printing overhead

seq_printf (used by show_smap) can be pretty expensive when dumping a lot of numbers. Say we would like to get Rss and Pss from a particular process. In order to measure a pathological case let's generate as many mappings as possible:

$ cat max_mmap.c
int main()
{
	while (mmap(NULL, 4096, PROT_READ|PROT_WRITE, MAP_ANON|MAP_SHARED|MAP_POPULATE, -1, 0) != MAP_FAILED)
		;
	printf("pid:%d\n", getpid());
	pause();
	return 0;
}
$ awk '/^Rss/{rss+=$2} /^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss}' /proc/$pid/smaps

would do a trick. The whole runtime is in the kernel space which is not that that unexpected because smaps is not the cheapest one (we have to do rmap walk etc.).

        Command being timed: "awk /^Rss/{rss+=$2} /^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss} /proc/3050/smaps"
        User time (seconds): 0.01
        System time (seconds): 0.44
        Percent of CPU this job got: 99%
        Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.47

But the perf says:

    22.55%  awk      [kernel.kallsyms]  [k] format_decode
    14.65%  awk      [kernel.kallsyms]  [k] vsnprintf
     6.40%  awk      [kernel.kallsyms]  [k] number
     2.53%  awk      [kernel.kallsyms]  [k] shmem_mapping
     2.53%  awk      [kernel.kallsyms]  [k] show_smap
     1.81%  awk      [kernel.kallsyms]  [k] lock_acquire

we are spending most of the time actually generating the output which is quite lame. Let's replace seq_printf by seq_puts and seq_put_decimal_ull. This will give us:

        Command being timed: "awk /^Rss/{rss+=$2} /^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss} /proc/3067/smaps"
        User time (seconds): 0.00
        System time (seconds): 0.41
        Percent of CPU this job got: 99%
        Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.42

which will give us ~7% improvement. Perf says:

    28.87%  awk      [kernel.kallsyms]  [k] seq_puts
     5.30%  awk      [kernel.kallsyms]  [k] vsnprintf
     4.54%  awk      [kernel.kallsyms]  [k] format_decode
     3.73%  awk      [kernel.kallsyms]  [k] show_smap
     2.56%  awk      [kernel.kallsyms]  [k] shmem_mapping
     1.92%  awk      [kernel.kallsyms]  [k] number
     1.80%  awk      [kernel.kallsyms]  [k] lock_acquire
     1.75%  awk      [kernel.kallsyms]  [k] print_name_value_kb
---
 fs/proc/task_mmu.c | 63 ++++++++++++++++++++++--------------------------------
 1 file changed, 25 insertions(+), 38 deletions(-)
fs/proc/task_mmu.c +25 -38
index 187d84ef9de9..41c24c0811da 100644
@@ -721,6 +721,13 @@ void __weak arch_show_smap(struct seq_file *m, struct vm_area_struct *vma)
 {
 }
 
+static void print_name_value_kb(struct seq_file *m, const char *name, unsigned long val)
+{
+	seq_puts(m, name);
+	seq_put_decimal_ull(m, 0, val);
+	seq_puts(m, " kB\n");
+}
+
 static int show_smap(struct seq_file *m, void *v, int is_pid)
 {
 	struct vm_area_struct *vma = v;
@@ -765,45 +772,25 @@ static int show_smap(struct seq_file *m, void *v, int is_pid)
 
 	show_map_vma(m, vma, is_pid);
 
-	seq_printf(m,
-		   "Size:           %8lu kB\n"
-		   "Rss:            %8lu kB\n"
-		   "Pss:            %8lu kB\n"
-		   "Shared_Clean:   %8lu kB\n"
-		   "Shared_Dirty:   %8lu kB\n"
-		   "Private_Clean:  %8lu kB\n"
-		   "Private_Dirty:  %8lu kB\n"
-		   "Referenced:     %8lu kB\n"
-		   "Anonymous:      %8lu kB\n"
-		   "AnonHugePages:  %8lu kB\n"
-		   "ShmemPmdMapped: %8lu kB\n"
-		   "Shared_Hugetlb: %8lu kB\n"
-		   "Private_Hugetlb: %7lu kB\n"
-		   "Swap:           %8lu kB\n"
-		   "SwapPss:        %8lu kB\n"
-		   "KernelPageSize: %8lu kB\n"
-		   "MMUPageSize:    %8lu kB\n"
-		   "Locked:         %8lu kB\n",
-		   (vma->vm_end - vma->vm_start) >> 10,
-		   mss.resident >> 10,
-		   (unsigned long)(mss.pss >> (10 + PSS_SHIFT)),
-		   mss.shared_clean  >> 10,
-		   mss.shared_dirty  >> 10,
-		   mss.private_clean >> 10,
-		   mss.private_dirty >> 10,
-		   mss.referenced >> 10,
-		   mss.anonymous >> 10,
-		   mss.anonymous_thp >> 10,
-		   mss.shmem_thp >> 10,
-		   mss.shared_hugetlb >> 10,
-		   mss.private_hugetlb >> 10,
-		   mss.swap >> 10,
-		   (unsigned long)(mss.swap_pss >> (10 + PSS_SHIFT)),
-		   vma_kernel_pagesize(vma) >> 10,
-		   vma_mmu_pagesize(vma) >> 10,
-		   (vma->vm_flags & VM_LOCKED) ?
+	print_name_value_kb(m, "Size:           ", (vma->vm_end - vma->vm_start) >> 10);
+	print_name_value_kb(m, "Rss:            ", mss.resident >> 10);
+	print_name_value_kb(m, "Pss:            ", (unsigned long)(mss.pss >> (10 + PSS_SHIFT)));
+	print_name_value_kb(m, "Shared_Clean:   ", mss.shared_clean  >> 10);
+	print_name_value_kb(m, "Shared_Dirty:   ", mss.shared_dirty  >> 10);
+	print_name_value_kb(m, "Private_Clean:  ", mss.private_clean >> 10);
+	print_name_value_kb(m, "Private_Dirty:  ", mss.private_dirty >> 10);
+	print_name_value_kb(m, "Referenced:     ", mss.referenced >> 10);
+	print_name_value_kb(m, "Anonymous:      ", mss.anonymous >> 10);
+	print_name_value_kb(m, "AnonHugePages:  ", mss.anonymous_thp >> 10);
+	print_name_value_kb(m, "ShmemPmdMapped: ", mss.shmem_thp >> 10);
+	print_name_value_kb(m, "Shared_Hugetlb: ", mss.shared_hugetlb >> 10);
+	print_name_value_kb(m, "Private_Hugetlb: ", mss.private_hugetlb >> 10);
+	print_name_value_kb(m, "Swap:           ", mss.swap >> 10);
+	print_name_value_kb(m, "SwapPss:        ", (unsigned long)(mss.swap_pss >> (10 + PSS_SHIFT)));
+	print_name_value_kb(m, "KernelPageSize: ", vma_kernel_pagesize(vma) >> 10);
+	print_name_value_kb(m, "MMUPageSize:    ", vma_mmu_pagesize(vma) >> 10);
+	print_name_value_kb(m, "Locked:         ", (vma->vm_flags & VM_LOCKED) ?
 			(unsigned long)(mss.pss >> (10 + PSS_SHIFT)) : 0);
-
 	arch_show_smap(m, vma);
 	show_smap_vma_flags(m, vma);
 	m_cache_vma(m, vma);
Signature · 5 lines
2.8.1

-- 
Michal Hocko
SUSE Labs
#16 lore raw
Quoted 106 lines · Michal Hocko
Quoted 86 lines · Jann Horn
Quoted 24 lines · Michal Hocko

On Tue 16-08-16 12:46:51, Robert Foss wrote:
[...]

Quoted 11 lines
$ /usr/bin/time -v -p zsh -c "repeat 25 { awk '/^Rss/{rss+=\$2}
/^Pss/{pss+=\$2} END {printf \"rss:%d pss:%d\n\", rss, pss}\'
/proc/5025/smaps }"
[...]
	Command being timed: "zsh -c repeat 25 { awk '/^Rss/{rss+=$2}
/^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss}\' /proc/5025/smaps
}"
	User time (seconds): 0.37
	System time (seconds): 0.45
	Percent of CPU this job got: 92%
	Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.89

This is really unexpected. Where is the user time spent? Anyway, rather than measuring some random processes I've tried to measure something resembling the worst case. So I've created a simple program to mmap as much as possible:

#include <sys/mman.h>
#include <sys/types.h>
#include <unistd.h>
#include <stdio.h>
int main()

Quoted 8 lines
{
	while (mmap(NULL, 4096, PROT_READ|PROT_WRITE, MAP_ANON|MAP_SHARED|MAP_POPULATE, -1, 0) != MAP_FAILED)
		;

	printf("pid:%d\n", getpid());
	pause();
	return 0;
}

Ah, nice, that's a reasonable test program. :)

Quoted 4 lines

So with a reasonable user space the parsing is really not all that time consuming wrt. smaps handling. That being said I am still very skeptical about a dedicated proc file which accomplishes what userspace can done in a trivial way.

Now, since your numbers showed that all the time is spent in the kernel, also create this test program to just read that file over and over again:

$ cat justreadloop.c
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <sched.h>
#include <unistd.h>
#include <err.h>
#include <stdio.h>

char buf[1000000];

int main(int argc, char **argv) {
printf("pid:%d\n", getpid());

  while (1) {
    int fd = open(argv[1], O_RDONLY);
    if (fd < 0) continue;
    if (read(fd, buf, sizeof(buf)) < 0)
      err(1, "read");
    close(fd);
  }
}
$ gcc -Wall -o justreadloop justreadloop.c
$

Now launch your test:

$ ./mapstuff
pid:29397

point justreadloop at it:

$ ./justreadloop /proc/29397/smaps
pid:32567

... and then check the performance stats of justreadloop:

# perf top -p 32567

This is what I see:

Samples: 232K of event 'cycles:ppp', Event count (approx.): 60448424325

Quoted 10 lines
Overhead  Shared Object     Symbol
  30,43%  [kernel]          [k] format_decode
   9,12%  [kernel]          [k] number
   7,66%  [kernel]          [k] vsnprintf
   7,06%  [kernel]          [k] __lock_acquire
   3,23%  [kernel]          [k] lock_release
   2,85%  [kernel]          [k] debug_lockdep_rcu_enabled
   2,25%  [kernel]          [k] skip_atoi
   2,13%  [kernel]          [k] lock_acquire
   2,05%  [kernel]          [k] show_smap

This is a lot! I would expect the rmap walk to consume more but it even doesn't show up in the top consumers.

Quoted 4 lines

That's at least 30.43% + 9.12% + 7.66% = 47.21% of the task's kernel time spent on evaluating format strings. The new interface wouldn't have to spend that much time on format strings because there isn't so much text to format.

Quoted 10 lines

well, this is true of course but I would much rather try to reduce the overhead of smaps file than add a new file. The following should help already. I've measured ~7% systime cut down. I guess there is still some room for improvements but I have to say I'm far from being convinced about a new proc file just because we suck at dumping information to the userspace. If this was something like /proc/<pid>/stat which is essentially read all the time then it would be a different question but is the rss, pss going to be all that often? If yes why? These are the questions which should be answered before we even start considering the implementation.

@Sonny Rao: Maybe you can comment on how often, for how many processes this information is needed and for which reasons this information is useful.

Quoted 150 lines

---
From 2a6883a7278ff8979808cb8e2dbcefe5ea3bf672 Mon Sep 17 00:00:00 2001
From: Michal Hocko <[email protected]>
Date: Wed, 17 Aug 2016 14:00:13 +0200
Subject: [PATCH] proc, smaps: reduce printing overhead

seq_printf (used by show_smap) can be pretty expensive when dumping a lot of numbers. Say we would like to get Rss and Pss from a particular process. In order to measure a pathological case let's generate as many mappings as possible:

$ cat max_mmap.c
int main()
{
	while (mmap(NULL, 4096, PROT_READ|PROT_WRITE, MAP_ANON|MAP_SHARED|MAP_POPULATE, -1, 0) != MAP_FAILED)
		;
	printf("pid:%d\n", getpid());
	pause();
	return 0;
}
$ awk '/^Rss/{rss+=$2} /^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss}' /proc/$pid/smaps

would do a trick. The whole runtime is in the kernel space which is not that that unexpected because smaps is not the cheapest one (we have to do rmap walk etc.).

        Command being timed: "awk /^Rss/{rss+=$2} /^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss} /proc/3050/smaps"
        User time (seconds): 0.01
        System time (seconds): 0.44
        Percent of CPU this job got: 99%
        Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.47

But the perf says:

    22.55%  awk      [kernel.kallsyms]  [k] format_decode
    14.65%  awk      [kernel.kallsyms]  [k] vsnprintf
     6.40%  awk      [kernel.kallsyms]  [k] number
     2.53%  awk      [kernel.kallsyms]  [k] shmem_mapping
     2.53%  awk      [kernel.kallsyms]  [k] show_smap
     1.81%  awk      [kernel.kallsyms]  [k] lock_acquire

we are spending most of the time actually generating the output which is quite lame. Let's replace seq_printf by seq_puts and seq_put_decimal_ull. This will give us:

        Command being timed: "awk /^Rss/{rss+=$2} /^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss} /proc/3067/smaps"
        User time (seconds): 0.00
        System time (seconds): 0.41
        Percent of CPU this job got: 99%
        Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.42

which will give us ~7% improvement. Perf says:

    28.87%  awk      [kernel.kallsyms]  [k] seq_puts
     5.30%  awk      [kernel.kallsyms]  [k] vsnprintf
     4.54%  awk      [kernel.kallsyms]  [k] format_decode
     3.73%  awk      [kernel.kallsyms]  [k] show_smap
     2.56%  awk      [kernel.kallsyms]  [k] shmem_mapping
     1.92%  awk      [kernel.kallsyms]  [k] number
     1.80%  awk      [kernel.kallsyms]  [k] lock_acquire
     1.75%  awk      [kernel.kallsyms]  [k] print_name_value_kb
---
 fs/proc/task_mmu.c | 63 ++++++++++++++++++++++--------------------------------
 1 file changed, 25 insertions(+), 38 deletions(-)
fs/proc/task_mmu.c +25 -38
index 187d84ef9de9..41c24c0811da 100644
@@ -721,6 +721,13 @@ void __weak arch_show_smap(struct seq_file *m, struct vm_area_struct *vma)
 {
 }
 
+static void print_name_value_kb(struct seq_file *m, const char *name, unsigned long val)
+{
+	seq_puts(m, name);
+	seq_put_decimal_ull(m, 0, val);
+	seq_puts(m, " kB\n");
+}
+
 static int show_smap(struct seq_file *m, void *v, int is_pid)
 {
 	struct vm_area_struct *vma = v;
@@ -765,45 +772,25 @@ static int show_smap(struct seq_file *m, void *v, int is_pid)
 
 	show_map_vma(m, vma, is_pid);
 
-	seq_printf(m,
-		   "Size:           %8lu kB\n"
-		   "Rss:            %8lu kB\n"
-		   "Pss:            %8lu kB\n"
-		   "Shared_Clean:   %8lu kB\n"
-		   "Shared_Dirty:   %8lu kB\n"
-		   "Private_Clean:  %8lu kB\n"
-		   "Private_Dirty:  %8lu kB\n"
-		   "Referenced:     %8lu kB\n"
-		   "Anonymous:      %8lu kB\n"
-		   "AnonHugePages:  %8lu kB\n"
-		   "ShmemPmdMapped: %8lu kB\n"
-		   "Shared_Hugetlb: %8lu kB\n"
-		   "Private_Hugetlb: %7lu kB\n"
-		   "Swap:           %8lu kB\n"
-		   "SwapPss:        %8lu kB\n"
-		   "KernelPageSize: %8lu kB\n"
-		   "MMUPageSize:    %8lu kB\n"
-		   "Locked:         %8lu kB\n",
-		   (vma->vm_end - vma->vm_start) >> 10,
-		   mss.resident >> 10,
-		   (unsigned long)(mss.pss >> (10 + PSS_SHIFT)),
-		   mss.shared_clean  >> 10,
-		   mss.shared_dirty  >> 10,
-		   mss.private_clean >> 10,
-		   mss.private_dirty >> 10,
-		   mss.referenced >> 10,
-		   mss.anonymous >> 10,
-		   mss.anonymous_thp >> 10,
-		   mss.shmem_thp >> 10,
-		   mss.shared_hugetlb >> 10,
-		   mss.private_hugetlb >> 10,
-		   mss.swap >> 10,
-		   (unsigned long)(mss.swap_pss >> (10 + PSS_SHIFT)),
-		   vma_kernel_pagesize(vma) >> 10,
-		   vma_mmu_pagesize(vma) >> 10,
-		   (vma->vm_flags & VM_LOCKED) ?
+	print_name_value_kb(m, "Size:           ", (vma->vm_end - vma->vm_start) >> 10);
+	print_name_value_kb(m, "Rss:            ", mss.resident >> 10);
+	print_name_value_kb(m, "Pss:            ", (unsigned long)(mss.pss >> (10 + PSS_SHIFT)));
+	print_name_value_kb(m, "Shared_Clean:   ", mss.shared_clean  >> 10);
+	print_name_value_kb(m, "Shared_Dirty:   ", mss.shared_dirty  >> 10);
+	print_name_value_kb(m, "Private_Clean:  ", mss.private_clean >> 10);
+	print_name_value_kb(m, "Private_Dirty:  ", mss.private_dirty >> 10);
+	print_name_value_kb(m, "Referenced:     ", mss.referenced >> 10);
+	print_name_value_kb(m, "Anonymous:      ", mss.anonymous >> 10);
+	print_name_value_kb(m, "AnonHugePages:  ", mss.anonymous_thp >> 10);
+	print_name_value_kb(m, "ShmemPmdMapped: ", mss.shmem_thp >> 10);
+	print_name_value_kb(m, "Shared_Hugetlb: ", mss.shared_hugetlb >> 10);
+	print_name_value_kb(m, "Private_Hugetlb: ", mss.private_hugetlb >> 10);
+	print_name_value_kb(m, "Swap:           ", mss.swap >> 10);
+	print_name_value_kb(m, "SwapPss:        ", (unsigned long)(mss.swap_pss >> (10 + PSS_SHIFT)));
+	print_name_value_kb(m, "KernelPageSize: ", vma_kernel_pagesize(vma) >> 10);
+	print_name_value_kb(m, "MMUPageSize:    ", vma_mmu_pagesize(vma) >> 10);
+	print_name_value_kb(m, "Locked:         ", (vma->vm_flags & VM_LOCKED) ?
 			(unsigned long)(mss.pss >> (10 + PSS_SHIFT)) : 0);
-
 	arch_show_smap(m, vma);
 	show_smap_vma_flags(m, vma);
 	m_cache_vma(m, vma);
#17 lore raw
In reply to Michal Hocko
Quoted 106 lines · Michal Hocko
Quoted 86 lines · Jann Horn
Quoted 24 lines · Michal Hocko

On Tue 16-08-16 12:46:51, Robert Foss wrote:
[...]

Quoted 11 lines
$ /usr/bin/time -v -p zsh -c "repeat 25 { awk '/^Rss/{rss+=\$2}
/^Pss/{pss+=\$2} END {printf \"rss:%d pss:%d\n\", rss, pss}\'
/proc/5025/smaps }"
[...]
  Command being timed: "zsh -c repeat 25 { awk '/^Rss/{rss+=$2}
/^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss}\' /proc/5025/smaps
}"
  User time (seconds): 0.37
  System time (seconds): 0.45
  Percent of CPU this job got: 92%
  Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.89

This is really unexpected. Where is the user time spent? Anyway, rather than measuring some random processes I've tried to measure something resembling the worst case. So I've created a simple program to mmap as much as possible:

#include <sys/mman.h>
#include <sys/types.h>
#include <unistd.h>
#include <stdio.h>
int main()

Quoted 8 lines
{
    while (mmap(NULL, 4096, PROT_READ|PROT_WRITE, MAP_ANON|MAP_SHARED|MAP_POPULATE, -1, 0) != MAP_FAILED)
            ;

    printf("pid:%d\n", getpid());
    pause();
    return 0;
}

Ah, nice, that's a reasonable test program. :)

Quoted 4 lines

So with a reasonable user space the parsing is really not all that time consuming wrt. smaps handling. That being said I am still very skeptical about a dedicated proc file which accomplishes what userspace can done in a trivial way.

Now, since your numbers showed that all the time is spent in the kernel, also create this test program to just read that file over and over again:

$ cat justreadloop.c
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <sched.h>
#include <unistd.h>
#include <err.h>
#include <stdio.h>

char buf[1000000];

int main(int argc, char **argv) {
printf("pid:%d\n", getpid());

  while (1) {
    int fd = open(argv[1], O_RDONLY);
    if (fd < 0) continue;
    if (read(fd, buf, sizeof(buf)) < 0)
      err(1, "read");
    close(fd);
  }
}
$ gcc -Wall -o justreadloop justreadloop.c
$

Now launch your test:

$ ./mapstuff
pid:29397

point justreadloop at it:

$ ./justreadloop /proc/29397/smaps
pid:32567

... and then check the performance stats of justreadloop:

# perf top -p 32567

This is what I see:

Samples: 232K of event 'cycles:ppp', Event count (approx.): 60448424325

Quoted 10 lines
Overhead  Shared Object     Symbol
  30,43%  [kernel]          [k] format_decode
   9,12%  [kernel]          [k] number
   7,66%  [kernel]          [k] vsnprintf
   7,06%  [kernel]          [k] __lock_acquire
   3,23%  [kernel]          [k] lock_release
   2,85%  [kernel]          [k] debug_lockdep_rcu_enabled
   2,25%  [kernel]          [k] skip_atoi
   2,13%  [kernel]          [k] lock_acquire
   2,05%  [kernel]          [k] show_smap

This is a lot! I would expect the rmap walk to consume more but it even doesn't show up in the top consumers.

Quoted 4 lines

That's at least 30.43% + 9.12% + 7.66% = 47.21% of the task's kernel time spent on evaluating format strings. The new interface wouldn't have to spend that much time on format strings because there isn't so much text to format.

Quoted 9 lines

well, this is true of course but I would much rather try to reduce the overhead of smaps file than add a new file. The following should help already. I've measured ~7% systime cut down. I guess there is still some room for improvements but I have to say I'm far from being convinced about a new proc file just because we suck at dumping information to the userspace.
If this was something like /proc/<pid>/stat which is essentially read all the time then it would be a different question but is the rss, pss going to be all that often? If yes why?

If the question is why do we need to read RSS, PSS, Private_*, Swap and the other fields so often?

I have two use cases so far involving monitoring per-process memory usage, and we usually need to read stats for about 25 processes.

Here's a timing example on an fairly recent ARM system 4 core RK3288 running at 1.8Ghz

localhost ~ # time cat /proc/25946/smaps > /dev/null

real    0m0.036s
user    0m0.020s
sys     0m0.020s

localhost ~ # time cat /proc/25946/totmaps > /dev/null

real    0m0.027s
user    0m0.010s
sys     0m0.010s

localhost ~ #

I'll ignore the user time for now, and we see about 20 ms of system time with smaps and 10 ms with totmaps, with 20 similar processes it would be 400 milliseconds of cpu time for the kernel to get this information from smaps vs 200 milliseconds with totmaps. Even totmaps is still pretty slow, but much better than smaps.

Use cases:
1) Basic task monitoring -- like "top" that shows memory consumption including PSS, Private, Swap
1 second update means about 40% of one CPU is spent in the kernel
gathering the data with smaps

2) User space OOM handling -- we'd rather do a more graceful shutdown than let the kernel's OOM killer activate and need to gather this information
and we'd like to be able to get this information to make the
decision much faster than 400ms

Quoted 158 lines

These are the
questions which should be answered before we even start considering the implementation.
---
From 2a6883a7278ff8979808cb8e2dbcefe5ea3bf672 Mon Sep 17 00:00:00 2001
From: Michal Hocko <[email protected]>
Date: Wed, 17 Aug 2016 14:00:13 +0200
Subject: [PATCH] proc, smaps: reduce printing overhead

seq_printf (used by show_smap) can be pretty expensive when dumping a lot of numbers. Say we would like to get Rss and Pss from a particular process. In order to measure a pathological case let's generate as many mappings as possible:

$ cat max_mmap.c
int main()
{
        while (mmap(NULL, 4096, PROT_READ|PROT_WRITE, MAP_ANON|MAP_SHARED|MAP_POPULATE, -1, 0) != MAP_FAILED)
                ;
        printf("pid:%d\n", getpid());
        pause();
        return 0;
}
$ awk '/^Rss/{rss+=$2} /^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss}' /proc/$pid/smaps

would do a trick. The whole runtime is in the kernel space which is not that that unexpected because smaps is not the cheapest one (we have to do rmap walk etc.).

        Command being timed: "awk /^Rss/{rss+=$2} /^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss} /proc/3050/smaps"
        User time (seconds): 0.01
        System time (seconds): 0.44
        Percent of CPU this job got: 99%
        Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.47

But the perf says:

    22.55%  awk      [kernel.kallsyms]  [k] format_decode
    14.65%  awk      [kernel.kallsyms]  [k] vsnprintf
     6.40%  awk      [kernel.kallsyms]  [k] number
     2.53%  awk      [kernel.kallsyms]  [k] shmem_mapping
     2.53%  awk      [kernel.kallsyms]  [k] show_smap
     1.81%  awk      [kernel.kallsyms]  [k] lock_acquire

we are spending most of the time actually generating the output which is quite lame. Let's replace seq_printf by seq_puts and seq_put_decimal_ull. This will give us:

        Command being timed: "awk /^Rss/{rss+=$2} /^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss} /proc/3067/smaps"
        User time (seconds): 0.00
        System time (seconds): 0.41
        Percent of CPU this job got: 99%
        Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.42

which will give us ~7% improvement. Perf says:

    28.87%  awk      [kernel.kallsyms]  [k] seq_puts
     5.30%  awk      [kernel.kallsyms]  [k] vsnprintf
     4.54%  awk      [kernel.kallsyms]  [k] format_decode
     3.73%  awk      [kernel.kallsyms]  [k] show_smap
     2.56%  awk      [kernel.kallsyms]  [k] shmem_mapping
     1.92%  awk      [kernel.kallsyms]  [k] number
     1.80%  awk      [kernel.kallsyms]  [k] lock_acquire
     1.75%  awk      [kernel.kallsyms]  [k] print_name_value_kb
---
 fs/proc/task_mmu.c | 63 ++++++++++++++++++++++--------------------------------
 1 file changed, 25 insertions(+), 38 deletions(-)
fs/proc/task_mmu.c +25 -39
index 187d84ef9de9..41c24c0811da 100644
@@ -721,6 +721,13 @@ void __weak arch_show_smap(struct seq_file *m, struct vm_area_struct *vma)
 {
 }
 
+static void print_name_value_kb(struct seq_file *m, const char *name, unsigned long val)
+{
+       seq_puts(m, name);
+       seq_put_decimal_ull(m, 0, val);
+       seq_puts(m, " kB\n");
+}
+
 static int show_smap(struct seq_file *m, void *v, int is_pid)
 {
        struct vm_area_struct *vma = v;
@@ -765,45 +772,25 @@ static int show_smap(struct seq_file *m, void *v, int is_pid)
 
        show_map_vma(m, vma, is_pid);
 
-       seq_printf(m,
-                  "Size:           %8lu kB\n"
-                  "Rss:            %8lu kB\n"
-                  "Pss:            %8lu kB\n"
-                  "Shared_Clean:   %8lu kB\n"
-                  "Shared_Dirty:   %8lu kB\n"
-                  "Private_Clean:  %8lu kB\n"
-                  "Private_Dirty:  %8lu kB\n"
-                  "Referenced:     %8lu kB\n"
-                  "Anonymous:      %8lu kB\n"
-                  "AnonHugePages:  %8lu kB\n"
-                  "ShmemPmdMapped: %8lu kB\n"
-                  "Shared_Hugetlb: %8lu kB\n"
-                  "Private_Hugetlb: %7lu kB\n"
-                  "Swap:           %8lu kB\n"
-                  "SwapPss:        %8lu kB\n"
-                  "KernelPageSize: %8lu kB\n"
-                  "MMUPageSize:    %8lu kB\n"
-                  "Locked:         %8lu kB\n",
-                  (vma->vm_end - vma->vm_start) >> 10,
-                  mss.resident >> 10,
-                  (unsigned long)(mss.pss >> (10 + PSS_SHIFT)),
-                  mss.shared_clean  >> 10,
-                  mss.shared_dirty  >> 10,
-                  mss.private_clean >> 10,
-                  mss.private_dirty >> 10,
-                  mss.referenced >> 10,
-                  mss.anonymous >> 10,
-                  mss.anonymous_thp >> 10,
-                  mss.shmem_thp >> 10,
-                  mss.shared_hugetlb >> 10,
-                  mss.private_hugetlb >> 10,
-                  mss.swap >> 10,
-                  (unsigned long)(mss.swap_pss >> (10 + PSS_SHIFT)),
-                  vma_kernel_pagesize(vma) >> 10,
-                  vma_mmu_pagesize(vma) >> 10,
-                  (vma->vm_flags & VM_LOCKED) ?
+       print_name_value_kb(m, "Size:           ", (vma->vm_end - vma->vm_start) >> 10);
+       print_name_value_kb(m, "Rss:            ", mss.resident >> 10);
+       print_name_value_kb(m, "Pss:            ", (unsigned long)(mss.pss >> (10 + PSS_SHIFT)));
+       print_name_value_kb(m, "Shared_Clean:   ", mss.shared_clean  >> 10);
+       print_name_value_kb(m, "Shared_Dirty:   ", mss.shared_dirty  >> 10);
+       print_name_value_kb(m, "Private_Clean:  ", mss.private_clean >> 10);
+       print_name_value_kb(m, "Private_Dirty:  ", mss.private_dirty >> 10);
+       print_name_value_kb(m, "Referenced:     ", mss.referenced >> 10);
+       print_name_value_kb(m, "Anonymous:      ", mss.anonymous >> 10);
+       print_name_value_kb(m, "AnonHugePages:  ", mss.anonymous_thp >> 10);
+       print_name_value_kb(m, "ShmemPmdMapped: ", mss.shmem_thp >> 10);
+       print_name_value_kb(m, "Shared_Hugetlb: ", mss.shared_hugetlb >> 10);
+       print_name_value_kb(m, "Private_Hugetlb: ", mss.private_hugetlb >> 10);
+       print_name_value_kb(m, "Swap:           ", mss.swap >> 10);
+       print_name_value_kb(m, "SwapPss:        ", (unsigned long)(mss.swap_pss >> (10 + PSS_SHIFT)));
+       print_name_value_kb(m, "KernelPageSize: ", vma_kernel_pagesize(vma) >> 10);
+       print_name_value_kb(m, "MMUPageSize:    ", vma_mmu_pagesize(vma) >> 10);
+       print_name_value_kb(m, "Locked:         ", (vma->vm_flags & VM_LOCKED) ?
                        (unsigned long)(mss.pss >> (10 + PSS_SHIFT)) : 0);
-
        arch_show_smap(m, vma);
        show_smap_vma_flags(m, vma);
        m_cache_vma(m, vma);
--

2.8.1

--
Michal Hocko
SUSE Labs

#18 lore raw
Quoted 2 lines · Sonny Rao
Quoted 1 line · Michal Hocko

On Wed 17-08-16 11:31:25, Jann Horn wrote:

[...]

Quoted 43 lines
Quoted 5 lines
Quoted 4 lines

That's at least 30.43% + 9.12% + 7.66% = 47.21% of the task's kernel time spent on evaluating format strings. The new interface wouldn't have to spend that much time on format strings because there isn't so much text to format.

Quoted 9 lines

well, this is true of course but I would much rather try to reduce the overhead of smaps file than add a new file. The following should help already. I've measured ~7% systime cut down. I guess there is still some room for improvements but I have to say I'm far from being convinced about a new proc file just because we suck at dumping information to the userspace.
If this was something like /proc/<pid>/stat which is essentially read all the time then it would be a different question but is the rss, pss going to be all that often? If yes why?

If the question is why do we need to read RSS, PSS, Private_*, Swap and the other fields so often?

I have two use cases so far involving monitoring per-process memory usage, and we usually need to read stats for about 25 processes.

Here's a timing example on an fairly recent ARM system 4 core RK3288 running at 1.8Ghz

localhost ~ # time cat /proc/25946/smaps > /dev/null

real    0m0.036s
user    0m0.020s
sys     0m0.020s

localhost ~ # time cat /proc/25946/totmaps > /dev/null

real    0m0.027s
user    0m0.010s
sys     0m0.010s

localhost ~ #

I'll ignore the user time for now, and we see about 20 ms of system time with smaps and 10 ms with totmaps, with 20 similar processes it would be 400 milliseconds of cpu time for the kernel to get this information from smaps vs 200 milliseconds with totmaps. Even totmaps is still pretty slow, but much better than smaps.

Quoted 5 lines

Use cases:
1) Basic task monitoring -- like "top" that shows memory consumption including PSS, Private, Swap
1 second update means about 40% of one CPU is spent in the kernel
gathering the data with smaps

I would argue that even 20% is way too much for such a monitoring. What is the value to do it so often tha 20 vs 40ms really matters?

Quoted 4 lines

2) User space OOM handling -- we'd rather do a more graceful shutdown than let the kernel's OOM killer activate and need to gather this information and we'd like to be able to get this information to make the decision much faster than 400ms

Global OOM handling in userspace is really dubious if you ask me. I understand you want something better than SIGKILL and in fact this is already possible with memory cgroup controller (btw. memcg will give you a cheap access to rss, amount of shared, swapped out memory as well). Anyway if you are getting close to the OOM your system will most probably be really busy and chances are that also reading your new file will take much more time. I am also not quite sure how is pss useful for oom decisions.

Don't take me wrong, /proc/<pid>/totmaps might be suitable for your specific usecase but so far I haven't heard any sound argument for it to be generally usable. It is true that smaps is unnecessarily costly but at least I can see some room for improvements. A simple patch I've posted cut the formatting overhead by 7%. Maybe we can do more.

Signature · 2 lines
Michal Hocko
SUSE Labs
#19 lore raw
Quoted 47 lines · Michal Hocko
Quoted 2 lines · Sonny Rao
Quoted 1 line · Michal Hocko

On Wed 17-08-16 11:31:25, Jann Horn wrote:

[...]

Quoted 43 lines
Quoted 5 lines
Quoted 4 lines

That's at least 30.43% + 9.12% + 7.66% = 47.21% of the task's kernel time spent on evaluating format strings. The new interface wouldn't have to spend that much time on format strings because there isn't so much text to format.

Quoted 9 lines

well, this is true of course but I would much rather try to reduce the overhead of smaps file than add a new file. The following should help already. I've measured ~7% systime cut down. I guess there is still some room for improvements but I have to say I'm far from being convinced about a new proc file just because we suck at dumping information to the userspace.
If this was something like /proc/<pid>/stat which is essentially read all the time then it would be a different question but is the rss, pss going to be all that often? If yes why?

If the question is why do we need to read RSS, PSS, Private_*, Swap and the other fields so often?

I have two use cases so far involving monitoring per-process memory usage, and we usually need to read stats for about 25 processes.

Here's a timing example on an fairly recent ARM system 4 core RK3288 running at 1.8Ghz

localhost ~ # time cat /proc/25946/smaps > /dev/null

real    0m0.036s
user    0m0.020s
sys     0m0.020s

localhost ~ # time cat /proc/25946/totmaps > /dev/null

real    0m0.027s
user    0m0.010s
sys     0m0.010s

localhost ~ #

I'll ignore the user time for now, and we see about 20 ms of system time with smaps and 10 ms with totmaps, with 20 similar processes it would be 400 milliseconds of cpu time for the kernel to get this information from smaps vs 200 milliseconds with totmaps. Even totmaps is still pretty slow, but much better than smaps.

Quoted 8 lines
Quoted 5 lines

Use cases:
1) Basic task monitoring -- like "top" that shows memory consumption including PSS, Private, Swap
1 second update means about 40% of one CPU is spent in the kernel
gathering the data with smaps

I would argue that even 20% is way too much for such a monitoring. What is the value to do it so often tha 20 vs 40ms really matters?

Yeah it is too much (I believe I said that) but it's significantly better.

Quoted 5 lines
Quoted 4 lines

2) User space OOM handling -- we'd rather do a more graceful shutdown than let the kernel's OOM killer activate and need to gather this information and we'd like to be able to get this information to make the decision much faster than 400ms

Quoted 8 lines

Global OOM handling in userspace is really dubious if you ask me. I understand you want something better than SIGKILL and in fact this is already possible with memory cgroup controller (btw. memcg will give you a cheap access to rss, amount of shared, swapped out memory as well). Anyway if you are getting close to the OOM your system will most probably be really busy and chances are that also reading your new file will take much more time. I am also not quite sure how is pss useful for oom decisions.

I mentioned it before, but based on experience RSS just isn't good enough -- there's too much sharing going on in our use case to make the correct decision based on RSS. If RSS were good enough, simply put, this patch wouldn't exist. So even with memcg I think we'd have the same problem?

Quoted 6 lines

Don't take me wrong, /proc/<pid>/totmaps might be suitable for your specific usecase but so far I haven't heard any sound argument for it to be generally usable. It is true that smaps is unnecessarily costly but at least I can see some room for improvements. A simple patch I've posted cut the formatting overhead by 7%. Maybe we can do more.

It seems like a general problem that if you want these values the existing kernel interface can be very expensive, so it would be generally usable by any application which wants a per process PSS,
private data, dirty data or swap value. I mentioned two use cases,
but I guess I don't understand the comment about why it's not usable by other use cases.

Quoted 3 lines

--
Michal Hocko
SUSE Labs

#20 lore raw
Quoted 2 lines · Sonny Rao
Quoted 1 line · Michal Hocko

On Wed 17-08-16 11:57:56, Sonny Rao wrote:

[...]

Quoted 14 lines
Quoted 5 lines
Quoted 4 lines

2) User space OOM handling -- we'd rather do a more graceful shutdown than let the kernel's OOM killer activate and need to gather this information and we'd like to be able to get this information to make the decision much faster than 400ms

Quoted 8 lines

Global OOM handling in userspace is really dubious if you ask me. I understand you want something better than SIGKILL and in fact this is already possible with memory cgroup controller (btw. memcg will give you a cheap access to rss, amount of shared, swapped out memory as well). Anyway if you are getting close to the OOM your system will most probably be really busy and chances are that also reading your new file will take much more time. I am also not quite sure how is pss useful for oom decisions.

Quoted 4 lines

I mentioned it before, but based on experience RSS just isn't good enough -- there's too much sharing going on in our use case to make the correct decision based on RSS. If RSS were good enough, simply put, this patch wouldn't exist.

But that doesn't answer my question, I am afraid. So how exactly do you use pss for oom decisions?

Quoted 1 line

So even with memcg I think we'd have the same problem?

memcg will give you instant anon, shared counters for all processes in the memcg.

Quoted 10 lines
Quoted 5 lines

Don't take me wrong, /proc/<pid>/totmaps might be suitable for your specific usecase but so far I haven't heard any sound argument for it to be generally usable. It is true that smaps is unnecessarily costly but at least I can see some room for improvements. A simple patch I've posted cut the formatting overhead by 7%. Maybe we can do more.

It seems like a general problem that if you want these values the existing kernel interface can be very expensive, so it would be generally usable by any application which wants a per process PSS, private data, dirty data or swap value.

yes this is really unfortunate. And if at all possible we should address that. Precise values require the expensive rmap walk. We can introduce some caching to help that. But so far it seems the biggest overhead is to simply format the output and that should be addressed before any new proc file is added.

Quoted 2 lines

I mentioned two use cases, but I guess I don't understand the comment about why it's not usable by other use cases.

I might be wrong here but a use of pss is quite limited and I do not remember anybody asking for large optimizations in that area. I still do not understand your use cases properly so I am quite skeptical about a general usefulness of a new file.

Signature · 2 lines
Michal Hocko
SUSE Labs
#21 lore raw
Quoted 23 lines · Michal Hocko
Quoted 2 lines · Sonny Rao
Quoted 1 line · Michal Hocko

On Wed 17-08-16 11:57:56, Sonny Rao wrote:

[...]

Quoted 14 lines
Quoted 5 lines
Quoted 4 lines

2) User space OOM handling -- we'd rather do a more graceful shutdown than let the kernel's OOM killer activate and need to gather this information and we'd like to be able to get this information to make the decision much faster than 400ms

Quoted 8 lines

Global OOM handling in userspace is really dubious if you ask me. I understand you want something better than SIGKILL and in fact this is already possible with memory cgroup controller (btw. memcg will give you a cheap access to rss, amount of shared, swapped out memory as well). Anyway if you are getting close to the OOM your system will most probably be really busy and chances are that also reading your new file will take much more time. I am also not quite sure how is pss useful for oom decisions.

Quoted 4 lines

I mentioned it before, but based on experience RSS just isn't good enough -- there's too much sharing going on in our use case to make the correct decision based on RSS. If RSS were good enough, simply put, this patch wouldn't exist.

Quoted 7 lines

But that doesn't answer my question, I am afraid. So how exactly do you use pss for oom decisions?

Quoted 1 line

So even with memcg I think we'd have the same problem?

memcg will give you instant anon, shared counters for all processes in the memcg.

Is it technically feasible to add instant pss support to memcg?

@Sonny Rao: Would using cgroups be acceptable for chromiumos?

Quoted 26 lines
Quoted 11 lines
Quoted 6 lines

Don't take me wrong, /proc/<pid>/totmaps might be suitable for your specific usecase but so far I haven't heard any sound argument for it to be generally usable. It is true that smaps is unnecessarily costly but at least I can see some room for improvements. A simple patch I've posted cut the formatting overhead by 7%. Maybe we can do more.

It seems like a general problem that if you want these values the existing kernel interface can be very expensive, so it would be generally usable by any application which wants a per process PSS, private data, dirty data or swap value.

yes this is really unfortunate. And if at all possible we should address that. Precise values require the expensive rmap walk. We can introduce some caching to help that. But so far it seems the biggest overhead is to simply format the output and that should be addressed before any new proc file is added.

Quoted 2 lines

I mentioned two use cases, but I guess I don't understand the comment about why it's not usable by other use cases.

I might be wrong here but a use of pss is quite limited and I do not remember anybody asking for large optimizations in that area. I still do not understand your use cases properly so I am quite skeptical about a general usefulness of a new file.

#22 lore raw
Quoted 40 lines · Robert Foss
Quoted 37 lines · Michal Hocko
Quoted 4 lines · Sonny Rao
Quoted 2 lines · Michal Hocko

On Wed 17-08-16 11:57:56, Sonny Rao wrote:

[...]

Quoted 17 lines
Quoted 7 lines
Quoted 5 lines

2) User space OOM handling -- we'd rather do a more graceful shutdown than let the kernel's OOM killer activate and need to gather this information and we'd like to be able to get this information to make the decision much faster than 400ms

Quoted 8 lines

Global OOM handling in userspace is really dubious if you ask me. I understand you want something better than SIGKILL and in fact this is already possible with memory cgroup controller (btw. memcg will give you a cheap access to rss, amount of shared, swapped out memory as well). Anyway if you are getting close to the OOM your system will most probably be really busy and chances are that also reading your new file will take much more time. I am also not quite sure how is pss useful for oom decisions.

Quoted 4 lines

I mentioned it before, but based on experience RSS just isn't good enough -- there's too much sharing going on in our use case to make the correct decision based on RSS. If RSS were good enough, simply put, this patch wouldn't exist.

But that doesn't answer my question, I am afraid. So how exactly do you use pss for oom decisions?

Quoted 1 line

So even with memcg I think we'd have the same problem?

Quoted 7 lines
Quoted 2 lines

memcg will give you instant anon, shared counters for all processes in the memcg.

Is it technically feasible to add instant pss support to memcg?

@Sonny Rao: Would using cgroups be acceptable for chromiumos?

It's possible, though I think we'd end up putting each renderer in it's own cgroup to get the PSS stat, so it seems a bit like overkill. I think memcg also has some overhead that we'd need to quantify but I could be mistaken about this.

Quoted 32 lines
Quoted 31 lines
Quoted 10 lines
Quoted 8 lines

Don't take me wrong, /proc/<pid>/totmaps might be suitable for your specific usecase but so far I haven't heard any sound argument for it to be generally usable. It is true that smaps is unnecessarily costly but at least I can see some room for improvements. A simple patch I've posted cut the formatting overhead by 7%. Maybe we can do more.

Quoted 4 lines

It seems like a general problem that if you want these values the existing kernel interface can be very expensive, so it would be generally usable by any application which wants a per process PSS, private data, dirty data or swap value.

yes this is really unfortunate. And if at all possible we should address that. Precise values require the expensive rmap walk. We can introduce some caching to help that. But so far it seems the biggest overhead is to simply format the output and that should be addressed before any new proc file is added.

Quoted 2 lines

I mentioned two use cases, but I guess I don't understand the comment about why it's not usable by other use cases.

I might be wrong here but a use of pss is quite limited and I do not remember anybody asking for large optimizations in that area. I still do not understand your use cases properly so I am quite skeptical about a general usefulness of a new file.

#23 lore raw
In reply to Michal Hocko

Hi Michal,

Quoted 18 lines · Michal Hocko
Quoted 2 lines · Sonny Rao
Quoted 1 line · Michal Hocko

On Wed 17-08-16 11:57:56, Sonny Rao wrote:

[...]

Quoted 14 lines
Quoted 13 lines
Quoted 4 lines

2) User space OOM handling -- we'd rather do a more graceful shutdown than let the kernel's OOM killer activate and need to gather this information and we'd like to be able to get this information to make the decision much faster than 400ms

Global OOM handling in userspace is really dubious if you ask me. I understand you want something better than SIGKILL and in fact this is already possible with memory cgroup controller (btw. memcg will give you a cheap access to rss, amount of shared, swapped out memory as well). Anyway if you are getting close to the OOM your system will most probably be really busy and chances are that also reading your new file will take much more time. I am also not quite sure how is pss useful for oom decisions.

Quoted 7 lines
Quoted 4 lines

I mentioned it before, but based on experience RSS just isn't good enough -- there's too much sharing going on in our use case to make the correct decision based on RSS. If RSS were good enough, simply put, this patch wouldn't exist.

But that doesn't answer my question, I am afraid. So how exactly do you use pss for oom decisions?

My case is not for OOM decision but I agree it would be great if we can get *fast* smap summary information.

PSS is really great tool to figure out how processes consume memory more exactly rather than RSS. We have been used it for monitoring of memory for per-process. Although it is not used for OOM decision, it would be great if it is speed up because we don't want to spend many CPU time for just monitoring.

For our usecase, we don't need AnonHugePages, ShmemPmdMapped, Shared_Hugetlb, Private_Hugetlb, KernelPageSize, MMUPageSize because we never enable THP and hugetlb. Additionally, Locked can be known via vma flags so we don't need it, either. Even, we don't need address range for just monitoring when we don't investigate in detail.

Although they are not severe overhead, why does it emit the useless information? Even bloat day by day. :( With that, userspace tools should spend more time to parse which is pointless.

Having said that, I'm not fan of creating new stat knob for that, either. How about appending summary information in the end of smap? So, monitoring users can just open the file and lseek to the (end - 1) and read the summary only.

Thanks.

#24 lore raw
Quoted 44 lines · Minchan Kim

Hi Michal,

Quoted 18 lines · Michal Hocko
Quoted 2 lines · Sonny Rao
Quoted 1 line · Michal Hocko

On Wed 17-08-16 11:57:56, Sonny Rao wrote:

[...]

Quoted 14 lines
Quoted 13 lines
Quoted 4 lines

2) User space OOM handling -- we'd rather do a more graceful shutdown than let the kernel's OOM killer activate and need to gather this information and we'd like to be able to get this information to make the decision much faster than 400ms

Global OOM handling in userspace is really dubious if you ask me. I understand you want something better than SIGKILL and in fact this is already possible with memory cgroup controller (btw. memcg will give you a cheap access to rss, amount of shared, swapped out memory as well). Anyway if you are getting close to the OOM your system will most probably be really busy and chances are that also reading your new file will take much more time. I am also not quite sure how is pss useful for oom decisions.

Quoted 7 lines
Quoted 4 lines

I mentioned it before, but based on experience RSS just isn't good enough -- there's too much sharing going on in our use case to make the correct decision based on RSS. If RSS were good enough, simply put, this patch wouldn't exist.

But that doesn't answer my question, I am afraid. So how exactly do you use pss for oom decisions?

My case is not for OOM decision but I agree it would be great if we can get *fast* smap summary information.

PSS is really great tool to figure out how processes consume memory more exactly rather than RSS. We have been used it for monitoring of memory for per-process. Although it is not used for OOM decision, it would be great if it is speed up because we don't want to spend many CPU time for just monitoring.

For our usecase, we don't need AnonHugePages, ShmemPmdMapped, Shared_Hugetlb, Private_Hugetlb, KernelPageSize, MMUPageSize because we never enable THP and hugetlb. Additionally, Locked can be known via vma flags so we don't need it, either. Even, we don't need address range for just monitoring when we don't investigate in detail.

Quoted 9 lines

Although they are not severe overhead, why does it emit the useless information? Even bloat day by day. :( With that, userspace tools should spend more time to parse which is pointless.

Having said that, I'm not fan of creating new stat knob for that, either. How about appending summary information in the end of smap? So, monitoring users can just open the file and lseek to the (end - 1) and read the summary only.

That would work fine for us as long as it's fast -- i.e. we don't still have to do all the expensive per-VMA format conversion in the kernel.

Quoted 1 line

Thanks.

#25 lore raw
In reply to Minchan Kim
Quoted 44 lines · Minchan Kim

Hi Michal,

Quoted 18 lines · Michal Hocko
Quoted 2 lines · Sonny Rao
Quoted 1 line · Michal Hocko

On Wed 17-08-16 11:57:56, Sonny Rao wrote:

[...]

Quoted 14 lines
Quoted 13 lines
Quoted 4 lines

2) User space OOM handling -- we'd rather do a more graceful shutdown than let the kernel's OOM killer activate and need to gather this information and we'd like to be able to get this information to make the decision much faster than 400ms

Global OOM handling in userspace is really dubious if you ask me. I understand you want something better than SIGKILL and in fact this is already possible with memory cgroup controller (btw. memcg will give you a cheap access to rss, amount of shared, swapped out memory as well). Anyway if you are getting close to the OOM your system will most probably be really busy and chances are that also reading your new file will take much more time. I am also not quite sure how is pss useful for oom decisions.

Quoted 7 lines
Quoted 4 lines

I mentioned it before, but based on experience RSS just isn't good enough -- there's too much sharing going on in our use case to make the correct decision based on RSS. If RSS were good enough, simply put, this patch wouldn't exist.

But that doesn't answer my question, I am afraid. So how exactly do you use pss for oom decisions?

My case is not for OOM decision but I agree it would be great if we can get *fast* smap summary information.

PSS is really great tool to figure out how processes consume memory more exactly rather than RSS. We have been used it for monitoring of memory for per-process. Although it is not used for OOM decision, it would be great if it is speed up because we don't want to spend many CPU time for just monitoring.

For our usecase, we don't need AnonHugePages, ShmemPmdMapped, Shared_Hugetlb, Private_Hugetlb, KernelPageSize, MMUPageSize because we never enable THP and hugetlb. Additionally, Locked can be known via vma flags so we don't need it, either. Even, we don't need address range for just monitoring when we don't investigate in detail.

Quoted 3 lines

Although they are not severe overhead, why does it emit the useless information? Even bloat day by day. :( With that, userspace tools should spend more time to parse which is pointless.

So far it doesn't really seem that the parsing is the biggest problem. The major cycles killer is the output formatting and that doesn't sound like a problem we are not able to address. And I would even argue that we want to address it in a generic way as much as possible.

Quoted 4 lines

Having said that, I'm not fan of creating new stat knob for that, either. How about appending summary information in the end of smap? So, monitoring users can just open the file and lseek to the (end - 1) and read the summary only.

That might confuse existing parsers. Besides that we already have /proc/<pid>/statm which gives cumulative numbers already. I am not sure how often it is used and whether the pte walk is too expensive for existing users but that should be explored and evaluated before a new file is created.

The /proc became a dump of everything people found interesting just because we were to easy to allow those additions. Do not repeat those mistakes, please!

Signature · 2 lines
Michal Hocko
SUSE Labs
#26 lore raw
Quoted 65 lines · Michal Hocko
Quoted 44 lines · Minchan Kim

Hi Michal,

Quoted 18 lines · Michal Hocko
Quoted 2 lines · Sonny Rao
Quoted 1 line · Michal Hocko

On Wed 17-08-16 11:57:56, Sonny Rao wrote:

[...]

Quoted 14 lines
Quoted 13 lines
Quoted 4 lines

2) User space OOM handling -- we'd rather do a more graceful shutdown than let the kernel's OOM killer activate and need to gather this information and we'd like to be able to get this information to make the decision much faster than 400ms

Global OOM handling in userspace is really dubious if you ask me. I understand you want something better than SIGKILL and in fact this is already possible with memory cgroup controller (btw. memcg will give you a cheap access to rss, amount of shared, swapped out memory as well). Anyway if you are getting close to the OOM your system will most probably be really busy and chances are that also reading your new file will take much more time. I am also not quite sure how is pss useful for oom decisions.

Quoted 7 lines
Quoted 4 lines

I mentioned it before, but based on experience RSS just isn't good enough -- there's too much sharing going on in our use case to make the correct decision based on RSS. If RSS were good enough, simply put, this patch wouldn't exist.

But that doesn't answer my question, I am afraid. So how exactly do you use pss for oom decisions?

My case is not for OOM decision but I agree it would be great if we can get *fast* smap summary information.

PSS is really great tool to figure out how processes consume memory more exactly rather than RSS. We have been used it for monitoring of memory for per-process. Although it is not used for OOM decision, it would be great if it is speed up because we don't want to spend many CPU time for just monitoring.

For our usecase, we don't need AnonHugePages, ShmemPmdMapped, Shared_Hugetlb, Private_Hugetlb, KernelPageSize, MMUPageSize because we never enable THP and hugetlb. Additionally, Locked can be known via vma flags so we don't need it, either. Even, we don't need address range for just monitoring when we don't investigate in detail.

Quoted 3 lines

Although they are not severe overhead, why does it emit the useless information? Even bloat day by day. :( With that, userspace tools should spend more time to parse which is pointless.

So far it doesn't really seem that the parsing is the biggest problem. The major cycles killer is the output formatting and that doesn't sound like a problem we are not able to address. And I would even argue that we want to address it in a generic way as much as possible.

Quoted 4 lines

Having said that, I'm not fan of creating new stat knob for that, either. How about appending summary information in the end of smap? So, monitoring users can just open the file and lseek to the (end - 1) and read the summary only.

That might confuse existing parsers. Besides that we already have /proc/<pid>/statm which gives cumulative numbers already. I am not sure how often it is used and whether the pte walk is too expensive for existing users but that should be explored and evaluated before a new file is created.

Quoted 3 lines

The /proc became a dump of everything people found interesting just because we were to easy to allow those additions. Do not repeat those mistakes, please!

Another thing I noticed was that we lock down smaps on Chromium OS. I think this is to avoid exposing more information than necessary via proc. The totmaps file gives us just the information we need and
nothing else. I certainly don't think we need a proc file for this
use case -- do you think a new system call is better or something else?

Quoted 3 lines

--
Michal Hocko
SUSE Labs

#27 lore raw
In reply to Michal Hocko
Quoted 45 lines · Michal Hocko
Quoted 44 lines · Minchan Kim

Hi Michal,

Quoted 18 lines · Michal Hocko
Quoted 2 lines · Sonny Rao
Quoted 1 line · Michal Hocko

On Wed 17-08-16 11:57:56, Sonny Rao wrote:

[...]

Quoted 14 lines
Quoted 13 lines
Quoted 4 lines

2) User space OOM handling -- we'd rather do a more graceful shutdown than let the kernel's OOM killer activate and need to gather this information and we'd like to be able to get this information to make the decision much faster than 400ms

Global OOM handling in userspace is really dubious if you ask me. I understand you want something better than SIGKILL and in fact this is already possible with memory cgroup controller (btw. memcg will give you a cheap access to rss, amount of shared, swapped out memory as well). Anyway if you are getting close to the OOM your system will most probably be really busy and chances are that also reading your new file will take much more time. I am also not quite sure how is pss useful for oom decisions.

Quoted 7 lines
Quoted 4 lines

I mentioned it before, but based on experience RSS just isn't good enough -- there's too much sharing going on in our use case to make the correct decision based on RSS. If RSS were good enough, simply put, this patch wouldn't exist.

But that doesn't answer my question, I am afraid. So how exactly do you use pss for oom decisions?

My case is not for OOM decision but I agree it would be great if we can get *fast* smap summary information.

PSS is really great tool to figure out how processes consume memory more exactly rather than RSS. We have been used it for monitoring of memory for per-process. Although it is not used for OOM decision, it would be great if it is speed up because we don't want to spend many CPU time for just monitoring.

For our usecase, we don't need AnonHugePages, ShmemPmdMapped, Shared_Hugetlb, Private_Hugetlb, KernelPageSize, MMUPageSize because we never enable THP and hugetlb. Additionally, Locked can be known via vma flags so we don't need it, either. Even, we don't need address range for just monitoring when we don't investigate in detail.

Quoted 6 lines
Quoted 3 lines

Although they are not severe overhead, why does it emit the useless information? Even bloat day by day. :( With that, userspace tools should spend more time to parse which is pointless.

So far it doesn't really seem that the parsing is the biggest problem. The major cycles killer is the output formatting and that doesn't sound

I cannot understand how kernel space is more expensive. Hmm. I tested your test program on my machine.

#!/bin/sh
./smap_test &
pid=$!

for i in $(seq 25)
do
cat /proc/$pid/smaps > /dev/null
done
kill $pid

root@bbox:/home/barrios/test/smap# time ./s_v.sh
pid:21925
real    0m3.365s
user    0m0.031s
sys     0m3.046s

vs.

#!/bin/sh
./smap_test &
pid=$!

for i in $(seq 25)
do

        awk '/^Rss/{rss+=$2} /^Pss/{pss+=$2} END {}' \
         /proc/$pid/smaps

done
kill $pid

root@bbox:/home/barrios/test/smap# time ./s.sh 
pid:21973
real    0m17.812s
user    0m12.612s
sys     0m5.187s

perf report says

    39.56%  awk        gawk               [.] dfaexec                             
     7.61%  awk        [kernel.kallsyms]  [k] format_decode                       
     6.37%  awk        gawk               [.] avoid_dfa                           
     5.85%  awk        gawk               [.] interpret                           
     5.69%  awk        [kernel.kallsyms]  [k] __memcpy                            
     4.37%  awk        [kernel.kallsyms]  [k] vsnprintf                           
     2.69%  awk        [kernel.kallsyms]  [k] number.isra.13                      
     2.10%  awk        gawk               [.] research                            
     1.91%  awk        gawk               [.] 0x00000000000351d0                  
     1.49%  awk        gawk               [.] free_wstr                           
     1.27%  awk        gawk               [.] unref                               
     1.19%  awk        gawk               [.] reset_record                        
     0.95%  awk        gawk               [.] set_record                          
     0.95%  awk        gawk               [.] get_field                           
     0.94%  awk        [kernel.kallsyms]  [k] show_smap                           

Parsing is much expensive than kernel.
Could you retest your test program?

Quoted 2 lines

like a problem we are not able to address. And I would even argue that we want to address it in a generic way as much as possible.

Sure. What solution do you think as generic way?

Quoted 18 lines
Quoted 5 lines

Having said that, I'm not fan of creating new stat knob for that, either. How about appending summary information in the end of smap? So, monitoring users can just open the file and lseek to the (end - 1) and read the summary only.

That might confuse existing parsers. Besides that we already have /proc/<pid>/statm which gives cumulative numbers already. I am not sure how often it is used and whether the pte walk is too expensive for existing users but that should be explored and evaluated before a new file is created.

The /proc became a dump of everything people found interesting just because we were to easy to allow those additions. Do not repeat those mistakes, please!

Signature · 2 lines
Michal Hocko
SUSE Labs
#28 lore raw

On Mon 22-08-16 09:07:45, Minchan Kim wrote:
[...]

Quoted 11 lines

#!/bin/sh
./smap_test &
pid=$!

for i in $(seq 25)
do

        awk '/^Rss/{rss+=$2} /^Pss/{pss+=$2} END {}' \
         /proc/$pid/smaps

done
kill $pid

Quoted 6 lines
root@bbox:/home/barrios/test/smap# time ./s.sh 
pid:21973
real    0m17.812s
user    0m12.612s
sys     0m5.187s

retested on the bare metal (x86_64 - 2CPUs)

        Command being timed: "sh s.sh"
        User time (seconds): 0.00
        System time (seconds): 18.08
        Percent of CPU this job got: 98%
        Elapsed (wall clock) time (h:mm:ss or m:ss): 0:18.29

multiple runs are quite consistent in those numbers. I am running with

$ awk --version
GNU Awk 4.1.3, API: 1.1 (GNU MPFR 3.1.4, GNU MP 6.1.0)
Quoted 4 lines
Quoted 2 lines

like a problem we are not able to address. And I would even argue that we want to address it in a generic way as much as possible.

Sure. What solution do you think as generic way?

either optimize seq_printf or replace it with something faster.

Signature · 2 lines
Michal Hocko
SUSE Labs
#29 lore raw
Quoted 31 lines · Michal Hocko

On Mon 22-08-16 09:07:45, Minchan Kim wrote:
[...]

Quoted 11 lines

#!/bin/sh
./smap_test &
pid=$!

for i in $(seq 25)
do

        awk '/^Rss/{rss+=$2} /^Pss/{pss+=$2} END {}' \
         /proc/$pid/smaps

done
kill $pid

Quoted 6 lines
root@bbox:/home/barrios/test/smap# time ./s.sh 
pid:21973
real    0m17.812s
user    0m12.612s
sys     0m5.187s

retested on the bare metal (x86_64 - 2CPUs)

        Command being timed: "sh s.sh"
        User time (seconds): 0.00
        System time (seconds): 18.08
        Percent of CPU this job got: 98%
        Elapsed (wall clock) time (h:mm:ss or m:ss): 0:18.29

multiple runs are quite consistent in those numbers. I am running with

$ awk --version
GNU Awk 4.1.3, API: 1.1 (GNU MPFR 3.1.4, GNU MP 6.1.0)
Quoted 6 lines
Quoted 4 lines
Quoted 2 lines

like a problem we are not able to address. And I would even argue that we want to address it in a generic way as much as possible.

Sure. What solution do you think as generic way?

either optimize seq_printf or replace it with something faster.

If it's real culprit, I agree. However, I tested your test program on my 2 x86 machines and my friend's machine.

Ubuntu, Fedora, Arch

They have awk 4.0.1 and 4.1.3.

Result are same. Userspace speand more times I mentioned.

[root@blaptop smap_test]# time awk '/^Rss/{rss+=$2} /^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss}' /proc/3552/smaps rss:263484 pss:262188

real    0m0.770s
user    0m0.574s
sys     0m0.197s

I will attach my test progrma source.
I hope you guys test and repost the result because it's the key for direction of patchset.

Thanks.

#include <sys/mman.h>

int main()

{
	unsigned long nr_vma = 0;

        while (1) {
		if (mmap(0, 4096, PROT_READ|PROT_WRITE, MAP_ANONYMOUS|MAP_SHARED|MAP_POPULATE, -1, 0) == MAP_FAILED)
			break;
		nr_vma++;
        };

        printf("pid:%d nr_vma:%lu\n", getpid(), nr_vma);
        pause();
        return 0;
}
#30 lore raw
Quoted 28 lines · Minchan Kim
Quoted 27 lines · Michal Hocko

On Mon 22-08-16 09:07:45, Minchan Kim wrote:
[...]

Quoted 11 lines

#!/bin/sh
./smap_test &
pid=$!

for i in $(seq 25)
do

        awk '/^Rss/{rss+=$2} /^Pss/{pss+=$2} END {}' \
         /proc/$pid/smaps

done
kill $pid

Quoted 6 lines
root@bbox:/home/barrios/test/smap# time ./s.sh
pid:21973
real    0m17.812s
user    0m12.612s
sys     0m5.187s

retested on the bare metal (x86_64 - 2CPUs)

        Command being timed: "sh s.sh"
        User time (seconds): 0.00
        System time (seconds): 18.08
        Percent of CPU this job got: 98%
        Elapsed (wall clock) time (h:mm:ss or m:ss): 0:18.29
Quoted 4 lines
Quoted 4 lines

multiple runs are quite consistent in those numbers. I am running with

$ awk --version
GNU Awk 4.1.3, API: 1.1 (GNU MPFR 3.1.4, GNU MP 6.1.0)
$ ./smap_test &
pid:19658 nr_vma:65514
$ time awk '/^Rss/{rss+=$2} /^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss}' /proc/19658/smaps
rss:263452 pss:262151

real 0m0.625s
user 0m0.404s
sys 0m0.216s

$ awk --version
GNU Awk 4.1.3, API: 1.1 (GNU MPFR 3.1.4, GNU MP 6.1.0)
Quoted 29 lines
Quoted 6 lines
Quoted 4 lines
Quoted 2 lines

like a problem we are not able to address. And I would even argue that we want to address it in a generic way as much as possible.

Sure. What solution do you think as generic way?

either optimize seq_printf or replace it with something faster.

If it's real culprit, I agree. However, I tested your test program on my 2 x86 machines and my friend's machine.

Ubuntu, Fedora, Arch

They have awk 4.0.1 and 4.1.3.

Result are same. Userspace speand more times I mentioned.

[root@blaptop smap_test]# time awk '/^Rss/{rss+=$2} /^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss}' /proc/3552/smaps rss:263484 pss:262188

real    0m0.770s
user    0m0.574s
sys     0m0.197s

I will attach my test progrma source.
I hope you guys test and repost the result because it's the key for direction of patchset.

Thanks.

#31 lore raw
In reply to Minchan Kim
Quoted 51 lines · Minchan Kim
Quoted 31 lines · Michal Hocko

On Mon 22-08-16 09:07:45, Minchan Kim wrote:
[...]

Quoted 11 lines

#!/bin/sh
./smap_test &
pid=$!

for i in $(seq 25)
do

        awk '/^Rss/{rss+=$2} /^Pss/{pss+=$2} END {}' \
         /proc/$pid/smaps

done
kill $pid

Quoted 6 lines
root@bbox:/home/barrios/test/smap# time ./s.sh 
pid:21973
real    0m17.812s
user    0m12.612s
sys     0m5.187s

retested on the bare metal (x86_64 - 2CPUs)

        Command being timed: "sh s.sh"
        User time (seconds): 0.00
        System time (seconds): 18.08
        Percent of CPU this job got: 98%
        Elapsed (wall clock) time (h:mm:ss or m:ss): 0:18.29

multiple runs are quite consistent in those numbers. I am running with

$ awk --version
GNU Awk 4.1.3, API: 1.1 (GNU MPFR 3.1.4, GNU MP 6.1.0)
Quoted 6 lines
Quoted 4 lines
Quoted 2 lines

like a problem we are not able to address. And I would even argue that we want to address it in a generic way as much as possible.

Sure. What solution do you think as generic way?

either optimize seq_printf or replace it with something faster.

If it's real culprit, I agree. However, I tested your test program on my 2 x86 machines and my friend's machine.

Ubuntu, Fedora, Arch

They have awk 4.0.1 and 4.1.3.

Result are same. Userspace speand more times I mentioned.

[root@blaptop smap_test]# time awk '/^Rss/{rss+=$2} /^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss}' /proc/3552/smaps rss:263484 pss:262188

Quoted 7 lines
real    0m0.770s
user    0m0.574s
sys     0m0.197s

I will attach my test progrma source.
I hope you guys test and repost the result because it's the key for direction of patchset.

Hmm, this is really interesting. I have checked a different machine and it shows different results. Same code, slightly different version of awk (4.1.0) and the results are different

        Command being timed: "awk /^Rss/{rss+=$2} /^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss} /proc/48925/smaps"
        User time (seconds): 0.43
        System time (seconds): 0.27

I have no idea why those numbers are so different on my laptop yet. It surely looks suspicious. I will try to debug this further tomorrow. Anyway, the performance is just one side of the problem. I have tried to express my concerns about a single exported pss value in other email. Please try to step back and think about how useful is this information without the knowing which resource we are talking about.

Signature · 2 lines
Michal Hocko
SUSE Labs
#32 lore raw

On Mon 22-08-16 18:45:54, Michal Hocko wrote:
[...]

Quoted 3 lines

I have no idea why those numbers are so different on my laptop yet. It surely looks suspicious. I will try to debug this further tomorrow.

Hmm, so I've tried to use my version of awk on other machine and vice
versa and it didn't make any difference. So this is independent on the
awk version it seems. So I've tried to strace /usr/bin/time and
wait4(-1, [{WIFEXITED(s) && WEXITSTATUS(s) == 0}], 0, {ru_utime={0, 0}, ru_stime={0, 688438}, ...}) = 9128

so the kernel indeed reports 0 user time for some reason. Note I was testing with 4.7 and right now with 4.8.0-rc3 kernel (no local modifications). The other machine which reports non-0 utime is 3.12 SLES kernel. Maybe I am hitting some accounting bug. At first I was suspecting CONFIG_NO_HZ_FULL because that is the main difference between my and the other machine but then I've noticed that the tests I was doing in kvm have this disabled too.. so it must be something else.

Weird...

Signature · 2 lines
Michal Hocko
SUSE Labs
#33 lore raw
Quoted 11 lines · Michal Hocko

On Mon 22-08-16 18:45:54, Michal Hocko wrote:
[...]

Quoted 3 lines

I have no idea why those numbers are so different on my laptop yet. It surely looks suspicious. I will try to debug this further tomorrow.

Hmm, so I've tried to use my version of awk on other machine and vice
versa and it didn't make any difference. So this is independent on the
awk version it seems. So I've tried to strace /usr/bin/time and
wait4(-1, [{WIFEXITED(s) && WEXITSTATUS(s) == 0}], 0, {ru_utime={0, 0}, ru_stime={0, 688438}, ...}) = 9128

Quoted 7 lines

so the kernel indeed reports 0 user time for some reason. Note I was testing with 4.7 and right now with 4.8.0-rc3 kernel (no local modifications). The other machine which reports non-0 utime is 3.12 SLES kernel. Maybe I am hitting some accounting bug. At first I was suspecting CONFIG_NO_HZ_FULL because that is the main difference between my and the other machine but then I've noticed that the tests I was doing in kvm have this disabled too.. so it must be something else.

4.5 reports non-0 while 4.6 zero utime. NO_HZ configuration is the same in both kernels.

Signature · 2 lines
Michal Hocko
SUSE Labs
#34 lore raw
Quoted 20 lines · Michal Hocko
Quoted 11 lines · Michal Hocko

On Mon 22-08-16 18:45:54, Michal Hocko wrote:
[...]

Quoted 3 lines

I have no idea why those numbers are so different on my laptop yet. It surely looks suspicious. I will try to debug this further tomorrow.

Hmm, so I've tried to use my version of awk on other machine and vice
versa and it didn't make any difference. So this is independent on the
awk version it seems. So I've tried to strace /usr/bin/time and
wait4(-1, [{WIFEXITED(s) && WEXITSTATUS(s) == 0}], 0, {ru_utime={0, 0}, ru_stime={0, 688438}, ...}) = 9128

Quoted 7 lines

so the kernel indeed reports 0 user time for some reason. Note I was testing with 4.7 and right now with 4.8.0-rc3 kernel (no local modifications). The other machine which reports non-0 utime is 3.12 SLES kernel. Maybe I am hitting some accounting bug. At first I was suspecting CONFIG_NO_HZ_FULL because that is the main difference between my and the other machine but then I've noticed that the tests I was doing in kvm have this disabled too.. so it must be something else.

Quoted 2 lines

4.5 reports non-0 while 4.6 zero utime. NO_HZ configuration is the same in both kernels.

and one more thing. It is not like utime accounting would be completely broken and always report 0. Other commands report non-0 values even on 4.6 kernels. I will try to bisect this down later today.

Signature · 2 lines
Michal Hocko
SUSE Labs
#35 lore raw

utime accounting regression since 4.6 (was: Re: [PACTH v2 0/3] Implement /proc/<pid>/totmaps)

Quoted 21 lines · Michal Hocko
Quoted 20 lines · Michal Hocko
Quoted 18 lines · Michal Hocko

On Mon 22-08-16 18:45:54, Michal Hocko wrote:
[...]

Quoted 3 lines

I have no idea why those numbers are so different on my laptop yet. It surely looks suspicious. I will try to debug this further tomorrow.

Hmm, so I've tried to use my version of awk on other machine and vice
versa and it didn't make any difference. So this is independent on the
awk version it seems. So I've tried to strace /usr/bin/time and
wait4(-1, [{WIFEXITED(s) && WEXITSTATUS(s) == 0}], 0, {ru_utime={0, 0}, ru_stime={0, 688438}, ...}) = 9128

so the kernel indeed reports 0 user time for some reason. Note I was testing with 4.7 and right now with 4.8.0-rc3 kernel (no local modifications). The other machine which reports non-0 utime is 3.12 SLES kernel. Maybe I am hitting some accounting bug. At first I was suspecting CONFIG_NO_HZ_FULL because that is the main difference between my and the other machine but then I've noticed that the tests I was doing in kvm have this disabled too.. so it must be something else.

Quoted 6 lines
Quoted 2 lines

4.5 reports non-0 while 4.6 zero utime. NO_HZ configuration is the same in both kernels.

and one more thing. It is not like utime accounting would be completely broken and always report 0. Other commands report non-0 values even on 4.6 kernels. I will try to bisect this down later today.

OK, so it seems I found it. I was quite lucky because account_user_time is not all that popular function and there were basically no changes besides Riks ff9a9b4c4334 ("sched, time: Switch VIRT_CPU_ACCOUNTING_GEN to jiffy granularity") and that seems to cause the regression. Reverting the commit on top of the current mmotm seems to fix the issue for me.

And just to give Rik more context. While debugging overhead of the /proc/<pid>/smaps I am getting a misleading output from /usr/bin/time -v (source for ./max_mmap is [1])

root@test1:~# uname -r
4.5.0-rc6-bisect1-00025-gff9a9b4c4334
root@test1:~# ./max_map 
pid:2990 maps:65515
root@test1:~# /usr/bin/time -v awk '/^Rss/{rss+=$2} /^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss}' /proc/2990/smaps
rss:263368 pss:262203
        Command being timed: "awk /^Rss/{rss+=$2} /^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss} /proc/2990/smaps"
        User time (seconds): 0.00
        System time (seconds): 0.45
        Percent of CPU this job got: 98%
        Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.46
        Average shared text size (kbytes): 0
        Average unshared data size (kbytes): 0
        Average stack size (kbytes): 0
        Average total size (kbytes): 0
        Maximum resident set size (kbytes): 1796
        Average resident set size (kbytes): 0
        Major (requiring I/O) page faults: 1
        Minor (reclaiming a frame) page faults: 83
        Voluntary context switches: 6
        Involuntary context switches: 6
        Swaps: 0
        File system inputs: 248
        File system outputs: 0
        Socket messages sent: 0
        Socket messages received: 0
        Signals delivered: 0
        Page size (bytes): 4096
        Exit status: 0

See the User time being 0 (as you can see above in the quoted text it is not a rounding error in userspace or something similar because wait4 really returns 0). Now with the revert

root@test1:~# uname -r
4.5.0-rc6-revert-00026-g7fc86f968bf5
root@test1:~# /usr/bin/time -v awk '/^Rss/{rss+=$2} /^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss}' /proc/3015/smaps
rss:263316 pss:262199
        Command being timed: "awk /^Rss/{rss+=$2} /^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss} /proc/3015/smaps"
        User time (seconds): 0.18
        System time (seconds): 0.29
        Percent of CPU this job got: 97%
        Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.50
        Average shared text size (kbytes): 0
        Average unshared data size (kbytes): 0
        Average stack size (kbytes): 0
        Average total size (kbytes): 0
        Maximum resident set size (kbytes): 1760
        Average resident set size (kbytes): 0
        Major (requiring I/O) page faults: 1
        Minor (reclaiming a frame) page faults: 79
        Voluntary context switches: 5
        Involuntary context switches: 7
        Swaps: 0
        File system inputs: 248
        File system outputs: 0
        Socket messages sent: 0
        Socket messages received: 0
        Signals delivered: 0
        Page size (bytes): 4096
        Exit status: 0

So it looks like the whole user time is accounted as the system time. My config is attached and yes I do have CONFIG_VIRT_CPU_ACCOUNTING_GEN enabled. Could you have a look please?

[1] http://lkml.kernel.org/r/[email protected]

Signature · 2 lines
Michal Hocko
SUSE Labs
#36 lore raw

Re: utime accounting regression since 4.6 (was: Re: [PACTH v2 0/3] Implement /proc/<pid>/totmaps)

Quoted 72 lines · Michal Hocko
Quoted 37 lines · Michal Hocko
Quoted 32 lines · Michal Hocko
Quoted 16 lines · Michal Hocko

On Mon 22-08-16 18:45:54, Michal Hocko wrote:
[...]

Quoted 5 lines

I have no idea why those numbers are so different on my laptop
yet. It surely looks suspicious. I will try to debug this further
tomorrow.

Hmm, so I've tried to use my version of awk on other machine and vice
versa and it didn't make any difference. So this is independent on the
awk version it seems. So I've tried to strace /usr/bin/time and wait4(-1, [{WIFEXITED(s) && WEXITSTATUS(s) == 0}], 0, {ru_utime={0, 0}, ru_stime={0, 688438}, ...}) = 9128

Quoted 14 lines

so the kernel indeed reports 0 user time for some reason. Note I
was testing with 4.7 and right now with 4.8.0-rc3 kernel (no local
modifications). The other machine which reports non-0 utime is 3.12
SLES kernel. Maybe I am hitting some accounting bug. At first I was
suspecting CONFIG_NO_HZ_FULL because that is the main difference between
my and the other machine but then I've noticed that the tests I was
doing in kvm have this disabled too.. so it must be something else.

Quoted 3 lines

4.5 reports non-0 while 4.6 zero utime. NO_HZ configuration is the same
in both kernels.

Quoted 5 lines

and one more thing. It is not like utime accounting would be completely
broken and always report 0. Other commands report non-0 values even on
4.6 kernels. I will try to bisect this down later today.

OK, so it seems I found it. I was quite lucky because account_user_time
is not all that popular function and there were basically no changes besides Riks ff9a9b4c4334 ("sched, time: Switch VIRT_CPU_ACCOUNTING_GEN
to jiffy granularity") and that seems to cause the regression. Reverting
the commit on top of the current mmotm seems to fix the issue for me.

And just to give Rik more context. While debugging overhead of the /proc/<pid>/smaps I am getting a misleading output from /usr/bin/time -v
(source for ./max_mmap is [1])

root@test1:~# uname -r
4.5.0-rc6-bisect1-00025-gff9a9b4c4334
root@test1:~# ./max_map 
pid:2990 maps:65515
root@test1:~# /usr/bin/time -v awk '/^Rss/{rss+=$2} /^Pss/{pss+=$2}
END {printf "rss:%d pss:%d\n", rss, pss}' /proc/2990/smaps
rss:263368 pss:262203
        Command being timed: "awk /^Rss/{rss+=$2} /^Pss/{pss+=$2} END
{printf "rss:%d pss:%d\n", rss, pss} /proc/2990/smaps"
        User time (seconds): 0.00
        System time (seconds): 0.45
        Percent of CPU this job got: 98%
Quoted 8 lines
root@test1:~# /usr/bin/time -v awk '/^Rss/{rss+=$2} /^Pss/{pss+=$2}
END {printf "rss:%d pss:%d\n", rss, pss}' /proc/3015/smaps
rss:263316 pss:262199
        Command being timed: "awk /^Rss/{rss+=$2} /^Pss/{pss+=$2} END
{printf "rss:%d pss:%d\n", rss, pss} /proc/3015/smaps"
        User time (seconds): 0.18
        System time (seconds): 0.29
        Percent of CPU this job got: 97%

The patch in question makes user and system time accounting essentially tick-based. If jiffies changes while the task is in user mode, time gets accounted as user time, if jiffies changes while the task is in system mode, time gets accounted as system time.

If you get "unlucky", with a job like the above, it is possible all time gets accounted to system time.

This would be true both with the system running with a periodic timer tick (before and after my patch is applied), and in nohz_idle mode (after my patch).

However, it does seem quite unlikely that you get zero user time, since you have 125 timer ticks in half a second. Furthermore, you do not even have NO_HZ_FULL enabled...

Does the workload consistently get zero user time?

If so, we need to dig further to see under what precise circumstances that happens.

On my laptop, with kernel 4.6.3-300.fc24.x86_64 I get this:

$ /usr/bin/time -v awk '/^Rss/{rss+=$2} /^Pss/{pss+=$2} END {printf
"rss:%d pss:%d\n", rss, pss}' /proc/19825/smaps
rss:263368 pss:262145
	Command being timed: "awk /^Rss/{rss+=$2} /^Pss/{pss+=$2} END
{printf "rss:%d pss:%d\n", rss, pss} /proc/19825/smaps"
	User time (seconds): 0.64
	System time (seconds): 0.19
	Percent of CPU this job got: 99%
	Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.83

The main difference between your and my NO_HZ config seems to be that NO_HZ_FULL is set here. However, it is not enabled at run time, so both of our systems should only really get NO_HZ_IDLE effectively.

Running tasks should get sampled with the regular timer tick, while they are running.

In other words, vtime accounting should be disabled in both of our tests, for everything except the idle task.

Do I need to do anything special to reproduce your bug, besides running the max mmap program and the awk script?

#37 lore raw

Re: utime accounting regression since 4.6 (was: Re: [PACTH v2 0/3] Implement /proc/<pid>/totmaps)

On Tue 23-08-16 17:46:11, Rik van Riel wrote:
[...]

Quoted 2 lines

Does the workload consistently get zero user
time?

root@test1:~# for i in $(seq 20); do /usr/bin/time -v awk '/^Rss/{rss+=$2} /^Pss/{pss+=$2} END {printf "rss:%d pss:%d\n", rss, pss}' /proc/3028/smaps; done 2>&1 | grep "User time" | sort | uniq -c
     20         User time (seconds): 0.00

sounds like pretty much consistent bad luck to me. I've attached the single awk strace run, maybe this will help you understand the timing...

[...]

Quoted 3 lines

Do I need to do anything special to reproduce
your bug, besides running the max mmap program
and the awk script?

That's how I noticed this. Nothing special really need. I am running that in a kvm virt. machine but the bare metal gives me the same results.

Signature · 2 lines
Michal Hocko
SUSE Labs
#38 lore raw

Re: utime accounting regression since 4.6 (was: Re: [PACTH v2 0/3] Implement /proc/<pid>/totmaps)

In reply to Rik van Riel

[CC Mike and Mel as they have seen some accounting oddities
when doing performance testing. They can share details but essentially the system time just gets too high]

For your reference the email thread started
http://lkml.kernel.org/r/[email protected]

I suspect this is mainly for short lived processes - like kernel compile

$ /usr/bin/time -v make mm/mmap.o
[...]
        User time (seconds): 0.45
        System time (seconds): 0.82
        Percent of CPU this job got: 111%
        Elapsed (wall clock) time (h:mm:ss or m:ss): 0:01.14
$ rm mm/mmap.o
$ /usr/bin/time -v make mm/mmap.o
[...]
        User time (seconds): 0.47
        System time (seconds): 1.55
        Percent of CPU this job got: 107%
        Elapsed (wall clock) time (h:mm:ss or m:ss): 0:01.88

This is quite unexpected for a cache hot compile. I would expect most of the time being spent in userspace.

$ perf report | grep kernel.vmlinux
     2.01%  as        [kernel.vmlinux]         [k] page_fault
     0.59%  cc1       [kernel.vmlinux]         [k] page_fault
     0.15%  git       [kernel.vmlinux]         [k] page_fault
     0.12%  bash      [kernel.vmlinux]         [k] page_fault
     0.11%  sh        [kernel.vmlinux]         [k] page_fault
     0.08%  gcc       [kernel.vmlinux]         [k] page_fault
     0.06%  make      [kernel.vmlinux]         [k] page_fault
     0.04%  rm        [kernel.vmlinux]         [k] page_fault
     0.03%  ld        [kernel.vmlinux]         [k] page_fault
     0.02%  bash      [kernel.vmlinux]         [k] entry_SYSCALL_64
     0.01%  git       [kernel.vmlinux]         [k] entry_SYSCALL_64
     0.01%  cat       [kernel.vmlinux]         [k] page_fault
     0.01%  collect2  [kernel.vmlinux]         [k] page_fault
     0.00%  sh        [kernel.vmlinux]         [k] entry_SYSCALL_64
     0.00%  rm        [kernel.vmlinux]         [k] entry_SYSCALL_64
     0.00%  grep      [kernel.vmlinux]         [k] page_fault

doesn't show anything unexpected.

Original Rik's reply follows:

Quoted 1 line · Rik van Riel

On Tue, 2016-08-23 at 16:33 +0200, Michal Hocko wrote:

[...]

Quoted 96 lines
Quoted 28 lines

OK, so it seems I found it. I was quite lucky because account_user_time
is not all that popular function and there were basically no changes besides Riks ff9a9b4c4334 ("sched, time: Switch VIRT_CPU_ACCOUNTING_GEN
to jiffy granularity") and that seems to cause the regression. Reverting
the commit on top of the current mmotm seems to fix the issue for me.

And just to give Rik more context. While debugging overhead of the /proc/<pid>/smaps I am getting a misleading output from /usr/bin/time -v
(source for ./max_mmap is [1])

root@test1:~# uname -r
4.5.0-rc6-bisect1-00025-gff9a9b4c4334
root@test1:~# ./max_map 
pid:2990 maps:65515
root@test1:~# /usr/bin/time -v awk '/^Rss/{rss+=$2} /^Pss/{pss+=$2}
END {printf "rss:%d pss:%d\n", rss, pss}' /proc/2990/smaps
rss:263368 pss:262203
        Command being timed: "awk /^Rss/{rss+=$2} /^Pss/{pss+=$2} END
{printf "rss:%d pss:%d\n", rss, pss} /proc/2990/smaps"
        User time (seconds): 0.00
        System time (seconds): 0.45
        Percent of CPU this job got: 98%
Quoted 8 lines
root@test1:~# /usr/bin/time -v awk '/^Rss/{rss+=$2} /^Pss/{pss+=$2}
END {printf "rss:%d pss:%d\n", rss, pss}' /proc/3015/smaps
rss:263316 pss:262199
        Command being timed: "awk /^Rss/{rss+=$2} /^Pss/{pss+=$2} END
{printf "rss:%d pss:%d\n", rss, pss} /proc/3015/smaps"
        User time (seconds): 0.18
        System time (seconds): 0.29
        Percent of CPU this job got: 97%

The patch in question makes user and system time accounting essentially tick-based. If jiffies changes while the task is in user mode, time gets accounted as user time, if jiffies changes while the task is in system mode, time gets accounted as system time.

If you get "unlucky", with a job like the above, it is possible all time gets accounted to system time.

This would be true both with the system running with a periodic timer tick (before and after my patch is applied), and in nohz_idle mode (after my patch).

However, it does seem quite unlikely that you get zero user time, since you have 125 timer ticks in half a second. Furthermore, you do not even have NO_HZ_FULL enabled...

Does the workload consistently get zero user time?

If so, we need to dig further to see under what precise circumstances that happens.

On my laptop, with kernel 4.6.3-300.fc24.x86_64 I get this:

$ /usr/bin/time -v awk '/^Rss/{rss+=$2} /^Pss/{pss+=$2} END {printf
"rss:%d pss:%d\n", rss, pss}' /proc/19825/smaps
rss:263368 pss:262145
	Command being timed: "awk /^Rss/{rss+=$2} /^Pss/{pss+=$2} END
{printf "rss:%d pss:%d\n", rss, pss} /proc/19825/smaps"
	User time (seconds): 0.64
	System time (seconds): 0.19
	Percent of CPU this job got: 99%
	Elapsed (wall clock) time (h:mm:ss or m:ss): 0:00.83

The main difference between your and my NO_HZ config seems to be that NO_HZ_FULL is set here. However, it is not enabled at run time, so both of our systems should only really get NO_HZ_IDLE effectively.

Running tasks should get sampled with the regular timer tick, while they are running.

In other words, vtime accounting should be disabled in both of our tests, for everything except the idle task.

Do I need to do anything special to reproduce your bug, besides running the max mmap program and the awk script?

Signature · 2 lines
Michal Hocko
SUSE Labs
#39 lore raw

Re: utime accounting regression since 4.6 (was: Re: [PACTH v2 0/3] Implement /proc/<pid>/totmaps)

Quoted 9 lines · Michal Hocko

[CC Mike and Mel as they have seen some accounting oddities when doing performance testing. They can share details but essentially the system time just gets too high]

For your reference the email thread started
http://lkml.kernel.org/r/[email protected]

I suspect this is mainly for short lived processes - like kernel compile

Quoted 13 lines
$ /usr/bin/time -v make mm/mmap.o
[...]
        User time (seconds): 0.45
        System time (seconds): 0.82
        Percent of CPU this job got: 111%
        Elapsed (wall clock) time (h:mm:ss or m:ss): 0:01.14
$ rm mm/mmap.o
$ /usr/bin/time -v make mm/mmap.o
[...]
        User time (seconds): 0.47
        System time (seconds): 1.55
        Percent of CPU this job got: 107%
        Elapsed (wall clock) time (h:mm:ss or m:ss): 0:01.88

I was not able to get the "expected" results from
your last reproducer, but this one does happen on
my system, too.

The bad news is, I still have no clue what is causing it...

Signature · 1 line
All Rights Reversed.
#40 lore raw
In reply to Michal Hocko
Quoted 18 lines · Michal Hocko
Quoted 2 lines · Sonny Rao
Quoted 1 line · Michal Hocko

On Wed 17-08-16 11:57:56, Sonny Rao wrote:

[...]

Quoted 14 lines
Quoted 13 lines
Quoted 4 lines

2) User space OOM handling -- we'd rather do a more graceful shutdown than let the kernel's OOM killer activate and need to gather this information and we'd like to be able to get this information to make the decision much faster than 400ms

Global OOM handling in userspace is really dubious if you ask me. I understand you want something better than SIGKILL and in fact this is already possible with memory cgroup controller (btw. memcg will give you a cheap access to rss, amount of shared, swapped out memory as well). Anyway if you are getting close to the OOM your system will most probably be really busy and chances are that also reading your new file will take much more time. I am also not quite sure how is pss useful for oom decisions.

Quoted 7 lines
Quoted 4 lines

I mentioned it before, but based on experience RSS just isn't good enough -- there's too much sharing going on in our use case to make the correct decision based on RSS. If RSS were good enough, simply put, this patch wouldn't exist.

But that doesn't answer my question, I am afraid. So how exactly do you use pss for oom decisions?

We use PSS to calculate the memory used by a process among all the processes in the system, in the case of Chrome this tells us how much each renderer process (which is roughly tied to a particular "tab" in Chrome) is using and how much it has swapped out, so we know what the worst offenders are -- I'm not sure what's unclear about that?

Chrome tends to use a lot of shared memory so we found PSS to be better than RSS, and I can give you examples of the RSS and PSS on real systems to illustrate the magnitude of the difference between those two numbers if that would be useful.

Quoted 6 lines
Quoted 2 lines

So even with memcg I think we'd have the same problem?

memcg will give you instant anon, shared counters for all processes in the memcg.

We want to be able to get per-process granularity quickly. I'm not sure if memcg provides that exactly?

Quoted 17 lines
Quoted 10 lines
Quoted 5 lines

Don't take me wrong, /proc/<pid>/totmaps might be suitable for your specific usecase but so far I haven't heard any sound argument for it to be generally usable. It is true that smaps is unnecessarily costly but at least I can see some room for improvements. A simple patch I've posted cut the formatting overhead by 7%. Maybe we can do more.

It seems like a general problem that if you want these values the existing kernel interface can be very expensive, so it would be generally usable by any application which wants a per process PSS, private data, dirty data or swap value.

yes this is really unfortunate. And if at all possible we should address that. Precise values require the expensive rmap walk. We can introduce some caching to help that. But so far it seems the biggest overhead is to simply format the output and that should be addressed before any new proc file is added.

Quoted 7 lines
Quoted 2 lines

I mentioned two use cases, but I guess I don't understand the comment about why it's not usable by other use cases.

I might be wrong here but a use of pss is quite limited and I do not remember anybody asking for large optimizations in that area. I still do not understand your use cases properly so I am quite skeptical about a general usefulness of a new file.

How do you know that usage of PSS is quite limited? I can only say that we've been using it on Chromium OS for at least four years and have found it very valuable, and I think I've explained the use cases in this thread. If you have more specific questions then I can try to clarify.

Quoted 4 lines

--
Michal Hocko
SUSE Labs

#41 lore raw
Quoted 24 lines · Sonny Rao
Quoted 23 lines · Michal Hocko
Quoted 2 lines · Sonny Rao
Quoted 1 line · Michal Hocko

On Wed 17-08-16 11:57:56, Sonny Rao wrote:

[...]

Quoted 18 lines
Quoted 5 lines
Quoted 4 lines

2) User space OOM handling -- we'd rather do a more graceful shutdown than let the kernel's OOM killer activate and need to gather this information and we'd like to be able to get this information to make the decision much faster than 400ms

Quoted 8 lines

Global OOM handling in userspace is really dubious if you ask me. I understand you want something better than SIGKILL and in fact this is already possible with memory cgroup controller (btw. memcg will give you a cheap access to rss, amount of shared, swapped out memory as well). Anyway if you are getting close to the OOM your system will most probably be really busy and chances are that also reading your new file will take much more time. I am also not quite sure how is pss useful for oom decisions.

I mentioned it before, but based on experience RSS just isn't good enough -- there's too much sharing going on in our use case to make the correct decision based on RSS. If RSS were good enough, simply put, this patch wouldn't exist.

Quoted 8 lines
Quoted 2 lines

But that doesn't answer my question, I am afraid. So how exactly do you use pss for oom decisions?

We use PSS to calculate the memory used by a process among all the processes in the system, in the case of Chrome this tells us how much each renderer process (which is roughly tied to a particular "tab" in Chrome) is using and how much it has swapped out, so we know what the worst offenders are -- I'm not sure what's unclear about that?

So let me ask more specifically. How can you make any decision based on the pss when you do not know _what_ is the shared resource. In other words if you select a task to terminate based on the pss then you have to kill others who share the same resource otherwise you do not release that shared resource. Not to mention that such a shared resource might be on tmpfs/shmem and it won't get released even after all processes which map it are gone.

I am sorry for being dense but it is still not clear to me how the single pss number can be used for oom or, in general, any serious decisions. The counter might be useful of course for debugging purposes or to have a general overview but then arguing about 40 vs 20ms sounds a bit strange to me.

Quoted 6 lines

Chrome tends to use a lot of shared memory so we found PSS to be better than RSS, and I can give you examples of the RSS and PSS on real systems to illustrate the magnitude of the difference between those two numbers if that would be useful.

Quoted 1 line
Quoted 8 lines
Quoted 5 lines
Quoted 1 line

So even with memcg I think we'd have the same problem?

memcg will give you instant anon, shared counters for all processes in the memcg.

We want to be able to get per-process granularity quickly. I'm not sure if memcg provides that exactly?

I will give you that information if you do process-per-memcg but that doesn't sound ideal. I thought those 20-something processes you were talking about are treated together but it seems I misunderstood.

Signature · 2 lines
Michal Hocko
SUSE Labs
#42 lore raw
Quoted 34 lines · Michal Hocko
Quoted 24 lines · Sonny Rao
Quoted 23 lines · Michal Hocko
Quoted 2 lines · Sonny Rao
Quoted 1 line · Michal Hocko

On Wed 17-08-16 11:57:56, Sonny Rao wrote:

[...]

Quoted 18 lines
Quoted 5 lines
Quoted 4 lines

2) User space OOM handling -- we'd rather do a more graceful shutdown than let the kernel's OOM killer activate and need to gather this information and we'd like to be able to get this information to make the decision much faster than 400ms

Quoted 8 lines

Global OOM handling in userspace is really dubious if you ask me. I understand you want something better than SIGKILL and in fact this is already possible with memory cgroup controller (btw. memcg will give you a cheap access to rss, amount of shared, swapped out memory as well). Anyway if you are getting close to the OOM your system will most probably be really busy and chances are that also reading your new file will take much more time. I am also not quite sure how is pss useful for oom decisions.

I mentioned it before, but based on experience RSS just isn't good enough -- there's too much sharing going on in our use case to make the correct decision based on RSS. If RSS were good enough, simply put, this patch wouldn't exist.

Quoted 8 lines
Quoted 2 lines

But that doesn't answer my question, I am afraid. So how exactly do you use pss for oom decisions?

We use PSS to calculate the memory used by a process among all the processes in the system, in the case of Chrome this tells us how much each renderer process (which is roughly tied to a particular "tab" in Chrome) is using and how much it has swapped out, so we know what the worst offenders are -- I'm not sure what's unclear about that?

Quoted 7 lines

So let me ask more specifically. How can you make any decision based on the pss when you do not know _what_ is the shared resource. In other words if you select a task to terminate based on the pss then you have to kill others who share the same resource otherwise you do not release that shared resource. Not to mention that such a shared resource might be on tmpfs/shmem and it won't get released even after all processes which map it are gone.

Ok I see why you're confused now, sorry.

In our case that we do know what is being shared in general because the sharing is mostly between those processes that we're looking at and not other random processes or tmpfs, so PSS gives us useful data in the context of these processes which are sharing the data especially for monitoring between the set of these renderer processes.

We also use the private clean and private dirty and swap fields to make a few metrics for the processes and charge each process for it's private, shared, and swap data. Private clean and dirty are used for estimating a lower bound on how much memory would be freed. Swap and PSS also give us some indication of additional memory which might get freed up.

Quoted 6 lines

I am sorry for being dense but it is still not clear to me how the single pss number can be used for oom or, in general, any serious decisions. The counter might be useful of course for debugging purposes or to have a general overview but then arguing about 40 vs 20ms sounds a bit strange to me.

Yeah so it's more than just the single PSS number, it's PSS, Private_Clean, Private_dirty, Swap are all interesting numbers to make these decisions.

Quoted 22 lines
Quoted 7 lines

Chrome tends to use a lot of shared memory so we found PSS to be better than RSS, and I can give you examples of the RSS and PSS on real systems to illustrate the magnitude of the difference between those two numbers if that would be useful.

Quoted 1 line
Quoted 8 lines
Quoted 5 lines
Quoted 1 line

So even with memcg I think we'd have the same problem?

memcg will give you instant anon, shared counters for all processes in the memcg.

We want to be able to get per-process granularity quickly. I'm not sure if memcg provides that exactly?

I will give you that information if you do process-per-memcg but that doesn't sound ideal. I thought those 20-something processes you were talking about are treated together but it seems I misunderstood. --
Michal Hocko
SUSE Labs

#43 lore raw
Quoted 43 lines · Sonny Rao
Quoted 34 lines · Michal Hocko
Quoted 24 lines · Sonny Rao
Quoted 23 lines · Michal Hocko
Quoted 2 lines · Sonny Rao
Quoted 1 line · Michal Hocko

On Wed 17-08-16 11:57:56, Sonny Rao wrote:

[...]

Quoted 18 lines
Quoted 5 lines
Quoted 4 lines

2) User space OOM handling -- we'd rather do a more graceful shutdown than let the kernel's OOM killer activate and need to gather this information and we'd like to be able to get this information to make the decision much faster than 400ms

Quoted 8 lines

Global OOM handling in userspace is really dubious if you ask me. I understand you want something better than SIGKILL and in fact this is already possible with memory cgroup controller (btw. memcg will give you a cheap access to rss, amount of shared, swapped out memory as well). Anyway if you are getting close to the OOM your system will most probably be really busy and chances are that also reading your new file will take much more time. I am also not quite sure how is pss useful for oom decisions.

I mentioned it before, but based on experience RSS just isn't good enough -- there's too much sharing going on in our use case to make the correct decision based on RSS. If RSS were good enough, simply put, this patch wouldn't exist.

Quoted 8 lines
Quoted 2 lines

But that doesn't answer my question, I am afraid. So how exactly do you use pss for oom decisions?

We use PSS to calculate the memory used by a process among all the processes in the system, in the case of Chrome this tells us how much each renderer process (which is roughly tied to a particular "tab" in Chrome) is using and how much it has swapped out, so we know what the worst offenders are -- I'm not sure what's unclear about that?

Quoted 7 lines

So let me ask more specifically. How can you make any decision based on the pss when you do not know _what_ is the shared resource. In other words if you select a task to terminate based on the pss then you have to kill others who share the same resource otherwise you do not release that shared resource. Not to mention that such a shared resource might be on tmpfs/shmem and it won't get released even after all processes which map it are gone.

Quoted 7 lines

Ok I see why you're confused now, sorry.

In our case that we do know what is being shared in general because the sharing is mostly between those processes that we're looking at and not other random processes or tmpfs, so PSS gives us useful data in the context of these processes which are sharing the data especially for monitoring between the set of these renderer processes.

OK, I see and agree that pss might be useful when you _know_ what is shared. But this sounds quite specific to a particular workload. How many users are in a similar situation? In other words, if we present a single number without the context, how much useful it will be in general? Is it possible that presenting such a number could be even misleading for somebody who doesn't have an idea which resources are shared? These are all questions which should be answered before we actually add this number (be it a new/existing proc file or a syscall). I still believe that the number without wider context is just not all that useful.

Quoted 4 lines

We also use the private clean and private dirty and swap fields to make a few metrics for the processes and charge each process for it's private, shared, and swap data. Private clean and dirty are used for estimating a lower bound on how much memory would be freed.

I can imagine that this kind of information might be useful and presented in /proc/<pid>/statm. The question is whether some of the existing consumers would see the performance impact due to he page table walk. Anyway even these counters might get quite tricky because even shareable resources are considered private if the process is the only one to map them (so again this might be a file on tmpfs...).

Quoted 3 lines

Swap and
PSS also give us some indication of additional memory which might get freed up.

Signature · 2 lines
Michal Hocko
SUSE Labs
#44 lore raw
Quoted 52 lines · Michal Hocko
Quoted 43 lines · Sonny Rao
Quoted 34 lines · Michal Hocko
Quoted 24 lines · Sonny Rao
Quoted 23 lines · Michal Hocko
Quoted 2 lines · Sonny Rao
Quoted 1 line · Michal Hocko

On Wed 17-08-16 11:57:56, Sonny Rao wrote:

[...]

Quoted 18 lines
Quoted 5 lines
Quoted 4 lines

2) User space OOM handling -- we'd rather do a more graceful shutdown than let the kernel's OOM killer activate and need to gather this information and we'd like to be able to get this information to make the decision much faster than 400ms

Quoted 8 lines

Global OOM handling in userspace is really dubious if you ask me. I understand you want something better than SIGKILL and in fact this is already possible with memory cgroup controller (btw. memcg will give you a cheap access to rss, amount of shared, swapped out memory as well). Anyway if you are getting close to the OOM your system will most probably be really busy and chances are that also reading your new file will take much more time. I am also not quite sure how is pss useful for oom decisions.

I mentioned it before, but based on experience RSS just isn't good enough -- there's too much sharing going on in our use case to make the correct decision based on RSS. If RSS were good enough, simply put, this patch wouldn't exist.

Quoted 8 lines
Quoted 2 lines

But that doesn't answer my question, I am afraid. So how exactly do you use pss for oom decisions?

We use PSS to calculate the memory used by a process among all the processes in the system, in the case of Chrome this tells us how much each renderer process (which is roughly tied to a particular "tab" in Chrome) is using and how much it has swapped out, so we know what the worst offenders are -- I'm not sure what's unclear about that?

Quoted 7 lines

So let me ask more specifically. How can you make any decision based on the pss when you do not know _what_ is the shared resource. In other words if you select a task to terminate based on the pss then you have to kill others who share the same resource otherwise you do not release that shared resource. Not to mention that such a shared resource might be on tmpfs/shmem and it won't get released even after all processes which map it are gone.

Quoted 7 lines

Ok I see why you're confused now, sorry.

In our case that we do know what is being shared in general because the sharing is mostly between those processes that we're looking at and not other random processes or tmpfs, so PSS gives us useful data in the context of these processes which are sharing the data especially for monitoring between the set of these renderer processes.

Quoted 10 lines

OK, I see and agree that pss might be useful when you _know_ what is shared. But this sounds quite specific to a particular workload. How many users are in a similar situation? In other words, if we present a single number without the context, how much useful it will be in general? Is it possible that presenting such a number could be even misleading for somebody who doesn't have an idea which resources are shared? These are all questions which should be answered before we actually add this number (be it a new/existing proc file or a syscall). I still believe that the number without wider context is just not all that useful.

I see the specific point about PSS -- because you need to know what is being shared or otherwise use it in a whole system context, but I still think the whole system context is a valid and generally useful thing. But what about the private_clean and private_dirty? Surely those are more generally useful for calculating a lower bound on process memory usage without additional knowledge?

At the end of the day all of these metrics are approximations, and it comes down to how far off the various approximations are and what trade offs we are willing to make.
RSS is the cheapest but the most coarse.

PSS (with the correct context) and Private data plus swap are much better but also more expensive due to the PT walk.
As far as I know, to get anything but RSS we have to go through smaps or use memcg. Swap seems to be available in /proc/<pid>/status.

I looked at the "shared" value in /proc/<pid>/statm but it doesn't seem to correlate well with the shared value in smaps -- not sure why?

It might be useful to show the magnitude of difference of using RSS vs PSS/Private in the case of the Chrome renderer processes. On the system I was looking at there were about 40 of these processes, but I picked a few to give an idea:

localhost ~ # cat /proc/21550/totmaps

Rss:               98972 kB
Pss:               54717 kB
Shared_Clean:      19020 kB
Shared_Dirty:      26352 kB
Private_Clean:         0 kB
Private_Dirty:     53600 kB
Referenced:        92184 kB
Anonymous:         46524 kB
AnonHugePages:     24576 kB
Swap:              13148 kB

RSS is 80% higher than PSS and 84% higher than private data

localhost ~ # cat /proc/21470/totmaps

Rss:              118420 kB
Pss:               70938 kB
Shared_Clean:      22212 kB
Shared_Dirty:      26520 kB
Private_Clean:         0 kB
Private_Dirty:     69688 kB
Referenced:       111500 kB
Anonymous:         79928 kB
AnonHugePages:     24576 kB
Swap:              12964 kB

RSS is 66% higher than RSS and 69% higher than private data

localhost ~ # cat /proc/21435/totmaps

Rss:               97156 kB
Pss:               50044 kB
Shared_Clean:      21920 kB
Shared_Dirty:      26400 kB
Private_Clean:         0 kB
Private_Dirty:     48836 kB
Referenced:        90012 kB
Anonymous:         75228 kB
AnonHugePages:     24576 kB
Swap:              13064 kB

RSS is 94% higher than PSS and 98% higher than private data.

It looks like there's a set of about 40MB of shared pages which cause the difference in this case.
Swap was roughly even on these but I don't think it's always going to be true.

Quoted 19 lines
Quoted 5 lines

We also use the private clean and private dirty and swap fields to make a few metrics for the processes and charge each process for it's private, shared, and swap data. Private clean and dirty are used for estimating a lower bound on how much memory would be freed.

I can imagine that this kind of information might be useful and presented in /proc/<pid>/statm. The question is whether some of the existing consumers would see the performance impact due to he page table walk. Anyway even these counters might get quite tricky because even shareable resources are considered private if the process is the only one to map them (so again this might be a file on tmpfs...).

Quoted 3 lines

Swap and
PSS also give us some indication of additional memory which might get freed up.

--
Michal Hocko
SUSE Labs

#45 lore raw
Quoted 131 lines · Sonny Rao
Quoted 52 lines · Michal Hocko
Quoted 43 lines · Sonny Rao
Quoted 34 lines · Michal Hocko
Quoted 24 lines · Sonny Rao
Quoted 23 lines · Michal Hocko
Quoted 2 lines · Sonny Rao
Quoted 1 line · Michal Hocko

On Wed 17-08-16 11:57:56, Sonny Rao wrote:

[...]

Quoted 18 lines
Quoted 5 lines
Quoted 4 lines

2) User space OOM handling -- we'd rather do a more graceful shutdown than let the kernel's OOM killer activate and need to gather this information and we'd like to be able to get this information to make the decision much faster than 400ms

Quoted 8 lines

Global OOM handling in userspace is really dubious if you ask me. I understand you want something better than SIGKILL and in fact this is already possible with memory cgroup controller (btw. memcg will give you a cheap access to rss, amount of shared, swapped out memory as well). Anyway if you are getting close to the OOM your system will most probably be really busy and chances are that also reading your new file will take much more time. I am also not quite sure how is pss useful for oom decisions.

I mentioned it before, but based on experience RSS just isn't good enough -- there's too much sharing going on in our use case to make the correct decision based on RSS. If RSS were good enough, simply put, this patch wouldn't exist.

Quoted 8 lines
Quoted 2 lines

But that doesn't answer my question, I am afraid. So how exactly do you use pss for oom decisions?

We use PSS to calculate the memory used by a process among all the processes in the system, in the case of Chrome this tells us how much each renderer process (which is roughly tied to a particular "tab" in Chrome) is using and how much it has swapped out, so we know what the worst offenders are -- I'm not sure what's unclear about that?

Quoted 7 lines

So let me ask more specifically. How can you make any decision based on the pss when you do not know _what_ is the shared resource. In other words if you select a task to terminate based on the pss then you have to kill others who share the same resource otherwise you do not release that shared resource. Not to mention that such a shared resource might be on tmpfs/shmem and it won't get released even after all processes which map it are gone.

Quoted 7 lines

Ok I see why you're confused now, sorry.

In our case that we do know what is being shared in general because the sharing is mostly between those processes that we're looking at and not other random processes or tmpfs, so PSS gives us useful data in the context of these processes which are sharing the data especially for monitoring between the set of these renderer processes.

Quoted 10 lines

OK, I see and agree that pss might be useful when you _know_ what is shared. But this sounds quite specific to a particular workload. How many users are in a similar situation? In other words, if we present a single number without the context, how much useful it will be in general? Is it possible that presenting such a number could be even misleading for somebody who doesn't have an idea which resources are shared? These are all questions which should be answered before we actually add this number (be it a new/existing proc file or a syscall). I still believe that the number without wider context is just not all that useful.

I see the specific point about PSS -- because you need to know what is being shared or otherwise use it in a whole system context, but I still think the whole system context is a valid and generally useful thing. But what about the private_clean and private_dirty? Surely those are more generally useful for calculating a lower bound on process memory usage without additional knowledge?

At the end of the day all of these metrics are approximations, and it comes down to how far off the various approximations are and what trade offs we are willing to make.
RSS is the cheapest but the most coarse.

PSS (with the correct context) and Private data plus swap are much better but also more expensive due to the PT walk.
As far as I know, to get anything but RSS we have to go through smaps or use memcg. Swap seems to be available in /proc/<pid>/status.

I looked at the "shared" value in /proc/<pid>/statm but it doesn't seem to correlate well with the shared value in smaps -- not sure why?

It might be useful to show the magnitude of difference of using RSS vs PSS/Private in the case of the Chrome renderer processes. On the system I was looking at there were about 40 of these processes, but I picked a few to give an idea:

localhost ~ # cat /proc/21550/totmaps

Rss:               98972 kB
Pss:               54717 kB
Shared_Clean:      19020 kB
Shared_Dirty:      26352 kB
Private_Clean:         0 kB
Private_Dirty:     53600 kB
Referenced:        92184 kB
Anonymous:         46524 kB
AnonHugePages:     24576 kB
Swap:              13148 kB

RSS is 80% higher than PSS and 84% higher than private data

localhost ~ # cat /proc/21470/totmaps

Rss:              118420 kB
Pss:               70938 kB
Shared_Clean:      22212 kB
Shared_Dirty:      26520 kB
Private_Clean:         0 kB
Private_Dirty:     69688 kB
Referenced:       111500 kB
Anonymous:         79928 kB
AnonHugePages:     24576 kB
Swap:              12964 kB

RSS is 66% higher than RSS and 69% higher than private data

localhost ~ # cat /proc/21435/totmaps

Rss:               97156 kB
Pss:               50044 kB
Shared_Clean:      21920 kB
Shared_Dirty:      26400 kB
Private_Clean:         0 kB
Private_Dirty:     48836 kB
Referenced:        90012 kB
Anonymous:         75228 kB
AnonHugePages:     24576 kB
Swap:              13064 kB
Quoted 7 lines

RSS is 94% higher than PSS and 98% higher than private data.

It looks like there's a set of about 40MB of shared pages which cause the difference in this case.
Swap was roughly even on these but I don't think it's always going to be true.

Sorry to hijack the thread, but I've found it recently
and I guess it's the best place to present our point.
We are working at our custom OS based on Linux and we also suffered much by /proc/<pid>/smaps file. As in Chrome we tried to improve our internal application memory management polices (Low Memory Killer) using data provided by smaps but we failed due to very long time needed for reading and parsing properly the file.

We've also observed that RSS measurement is often highly over PSS which seems to be more real memory usage for process. Using smaps we would be able to calculate USS usage and know exact minimum value of memory that would be freed after terminating some process. Those are very important sources of information as they give as the possibility to provide best possible app life-cycle.

We have also tried to use smaps in some application for OS developers as source of detailed information of memory usage of the system. For checking possible ways of improvement we tried totmaps from earlier version. On sample case for our app the CPU usage as presented by 'top' decreases from ~60% to ~4.5% only by changing source from smpas to tomaps.

So we are also very interested in using interface such as totmaps as it gives detailed and complete memory usage information for user-space and in our case much of information provided by smaps is for us not useful at all.

We are also using or tried using other interfaces like status, statm, cgroups.memory etc. but still totmaps/smaps are still the best interface to get all of the informations per process based in single place.

Quoted 21 lines
Quoted 19 lines
Quoted 5 lines

We also use the private clean and private dirty and swap fields to make a few metrics for the processes and charge each process for it's private, shared, and swap data. Private clean and dirty are used for estimating a lower bound on how much memory would be freed.

I can imagine that this kind of information might be useful and presented in /proc/<pid>/statm. The question is whether some of the existing consumers would see the performance impact due to he page table walk. Anyway even these counters might get quite tricky because even shareable resources are considered private if the process is the only one to map them (so again this might be a file on tmpfs...).

Quoted 3 lines

Swap and
PSS also give us some indication of additional memory which might get freed up.

--
Michal Hocko
SUSE Labs

Signature · 3 lines
Marcin Jabrzyk
Samsung R&D Institute Poland
Samsung Electronics
#46 lore raw

On Wed 24-08-16 12:14:06, Marcin Jabrzyk wrote:
[...]

Quoted 7 lines

Sorry to hijack the thread, but I've found it recently
and I guess it's the best place to present our point.
We are working at our custom OS based on Linux and we also suffered much by /proc/<pid>/smaps file. As in Chrome we tried to improve our internal application memory management polices (Low Memory Killer) using data provided by smaps but we failed due to very long time needed for reading and parsing properly the file.

I was already questioning Pss and also Private_* for any memory killer purpose earlier in the thread because cumulative numbers for all mappings can be really meaningless. Especially when you do not know about which resource is shared and by whom. Maybe you can describe how you are using those cumulative numbers for your decisions and prove me wrong but I simply haven't heard any sound arguments so far. Everything was just "we know what we are doing in our environment so we know those resouces and therefore those numbers make sense to us". But with all due respect this is not a reason to add a user visible API into the kernel.

Signature · 2 lines
Michal Hocko
SUSE Labs
#47 lore raw
In reply to Sonny Rao

[Sorry for a late reply, I was busy with other stuff]

Quoted 1 line · Sonny Rao

On Mon, Aug 22, 2016 at 12:54 AM, Michal Hocko <[email protected]> wrote:

[...]

Quoted 3 lines

But what about the private_clean and private_dirty? Surely those are more generally useful for calculating a lower bound on process memory usage without additional knowledge?

I guess private_clean can be used as a reasonable estimate. private_dirty less so because it may refer to e.g. tmpfs which is not mapped by other process and so no memory would be freed after unmap without removing the file.

Quoted 4 lines

At the end of the day all of these metrics are approximations, and it comes down to how far off the various approximations are and what trade offs we are willing to make.
RSS is the cheapest but the most coarse.

I agree on this part definitely. I also understand that what we provide currently is quite confusing and not really helpful. But I am afraid that the accounting is far from trivial to make right for all the possible cases.

Quoted 2 lines

PSS (with the correct context) and Private data plus swap are much better but also more expensive due to the PT walk.

Maybe we can be more clever and do some form of caching. I haven't thought that through to see how hard that could be. I mean we could cache some data per mm_struct and invalidate them only after the current value would get too much out of sync.

Quoted 5 lines

As far as I know, to get anything but RSS we have to go through smaps or use memcg. Swap seems to be available in /proc/<pid>/status.

I looked at the "shared" value in /proc/<pid>/statm but it doesn't seem to correlate well with the shared value in smaps -- not sure why?

task_statm() does only approximate to get_mm_counter(mm, MM_FILEPAGES) + get_mm_counter(mm, MM_SHMEMPAGES) so all the pages accounted to the mm. If they are not shared by anybody else they would be considered private by smaps.

Quoted 46 lines

It might be useful to show the magnitude of difference of using RSS vs PSS/Private in the case of the Chrome renderer processes. On the system I was looking at there were about 40 of these processes, but I picked a few to give an idea:

localhost ~ # cat /proc/21550/totmaps

Rss:               98972 kB
Pss:               54717 kB
Shared_Clean:      19020 kB
Shared_Dirty:      26352 kB
Private_Clean:         0 kB
Private_Dirty:     53600 kB
Referenced:        92184 kB
Anonymous:         46524 kB
AnonHugePages:     24576 kB
Swap:              13148 kB

RSS is 80% higher than PSS and 84% higher than private data

localhost ~ # cat /proc/21470/totmaps

Rss:              118420 kB
Pss:               70938 kB
Shared_Clean:      22212 kB
Shared_Dirty:      26520 kB
Private_Clean:         0 kB
Private_Dirty:     69688 kB
Referenced:       111500 kB
Anonymous:         79928 kB
AnonHugePages:     24576 kB
Swap:              12964 kB

RSS is 66% higher than RSS and 69% higher than private data

localhost ~ # cat /proc/21435/totmaps

Rss:               97156 kB
Pss:               50044 kB
Shared_Clean:      21920 kB
Shared_Dirty:      26400 kB
Private_Clean:         0 kB
Private_Dirty:     48836 kB
Referenced:        90012 kB
Anonymous:         75228 kB
AnonHugePages:     24576 kB
Swap:              13064 kB
Quoted 5 lines

RSS is 94% higher than PSS and 98% higher than private data.

It looks like there's a set of about 40MB of shared pages which cause the difference in this case.
Swap was roughly even on these but I don't think it's always going to be true.

OK, I see that those processes differ in the way how they are using memory but I am not really sure what kind of conclusion you can draw from that.

Signature · 2 lines
Michal Hocko
SUSE Labs
#48 lore raw
Quoted 10 lines · Michal Hocko

[Sorry for a late reply, I was busy with other stuff]

Quoted 1 line · Sonny Rao

On Mon, Aug 22, 2016 at 12:54 AM, Michal Hocko <[email protected]> wrote:

[...]

Quoted 3 lines

But what about the private_clean and private_dirty? Surely those are more generally useful for calculating a lower bound on process memory usage without additional knowledge?

I guess private_clean can be used as a reasonable estimate.

I was thinking about this more and I think I am wrong here. Even truly MAP_PRIVATE|MAP_ANON will be in private_dirty. So private_clean will become not all that interesting and similarly misleading as its _dirty variant (mmaped file after [m]sync should become _clean) and that doesn't mean the memory will get freed after the process which maps it terminates. Take shmem as an example again.

Quoted 3 lines

private_dirty less so because it may refer to e.g. tmpfs which is not mapped by other process and so no memory would be freed after unmap without removing the file.

Signature · 2 lines
Michal Hocko
SUSE Labs