Skip to contentSkip to conversation

mm, proc: Implement /proc/<pid>/totmaps

PATCH In review [email protected] 7 participants 22 messages Sep 5, 2016 – Sep 20, 2016 2 messages not in the archive

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

v5 cover mm, proc: Implement /proc/<pid>/totmaps

From: Robert Foss <[email protected]>

This series provides the /proc/PID/totmaps feature, which summarizes the information provided by /proc/PID/smaps for improved performance and usability reasons.

A 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.

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.

/proc/PID/totmaps provides roughly a 2x speedup compared to parsing /proc/PID/smaps with awk.

$ 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

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()

Changes since v2:
- Removed struct mem_size_stats *mss from struct proc_maps_private
- Removed priv->task assignment in totmaps_open() call
- Moved some assignements calls totmaps_open() around to increase code clarity
- Moved some function calls to unlock data structures before printing

Changes since v3:
- Fixed typo in totmaps documentation
- Fixed issue where proc_map_release wasn't called on error
- Fixed put_task_struct not being called during .release()

Changes since v4:
- Prevent access to invalid processes

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                 |   2 +
 fs/proc/task_mmu.c                 | 148 +++++++++++++++++++++++++++++++++++++
 4 files changed, 173 insertions(+), 1 deletion(-)
Signature · 1 line
2.7.4
#1 lore raw

PATCH v5 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 |   2 +
 fs/proc/task_mmu.c | 148 +++++++++++++++++++++++++++++++++++++++++++++++++++++
 3 files changed, 151 insertions(+)
fs/proc/base.c +1 -0
index ac0df4d..dc7e81b7 100644
@@ -2854,6 +2854,7 @@ 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 +2 -0
index 7931c55..3bdafe8 100644
@@ -298,6 +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 +148 -0

172 lines not on this page · raw message

Signature · 1 line
2.7.4
#2 lore raw

fs/proc/base.c static const struct pid_entry tgid_base_stuff[] = { · line 2857

Quoted 6 lines · On 09/05, robert.foss-ZGY8ohtN/[email protected] wrote:
fs/proc/base.c +1 -0
@@ -2854,6 +2854,7 @@ 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),

I must have missed something, but I fail to understand why this patch is so complicated.

Just use ONE("totmaps", S_IRUGO, proc_totmaps_operations) ?

fs/proc/task_mmu.c static int show_smap(struct seq_file *m, void *v, int is_pid) · line 837

Quoted 10 lines
fs/proc/task_mmu.c +10 -0
+static int totmaps_proc_show(struct seq_file *m, void *data)
+{
+	struct proc_maps_private *priv = m->private;
+	struct mm_struct *mm = priv->mm;
+	struct vm_area_struct *vma;
+	struct mem_size_stats mss_sum;
+
+	memset(&mss_sum, 0, sizeof(mss_sum));
+	down_read(&mm->mmap_sem);
+	hold_task_mempolicy(priv);
        ^^^^^^^^^^^^^^^^^^^^^^^^^

why?

fs/proc/task_mmu.c static int show_smap(struct seq_file *m, void *v, int is_pid) · line 839

Quoted 1 line
fs/proc/task_mmu.c +1 -0
+	for (vma = mm->mmap; vma != priv->tail_vma; vma = vma->vm_next) {

Hmm. the usage of ->tail_vma looks just wrong. I guess the code should work because it is NULL but still.

fs/proc/task_mmu.c static int show_smap(struct seq_file *m, void *v, int is_pid) · line 852

Quoted 13 lines
fs/proc/task_mmu.c +13 -0
+		struct mem_size_stats mss;
+		struct mm_walk smaps_walk = {
+			.pmd_entry = smaps_pte_range,
+			.mm = vma->vm_mm,
+			.private = &mss,
+		};
+
+		if (vma->vm_mm && !is_vm_hugetlb_page(vma)) {
+			memset(&mss, 0, sizeof(mss));
+			walk_page_vma(vma, &smaps_walk);
+			add_smaps_sum(&mss, &mss_sum);
+		}
+	}

Why? I mean, why not walk_page_range() ? You do not need this for-each-vma loop at all? At least if you change this patch to use the ONE() helper, and everything else looks unneeded in this case.

Oleg.

#3 lore raw

Hey Oleg!

Thanks for the feedback, I'll keep it in mind, but currently it looks like the patch is on ice for non-implementation related reasons.

Rob.

Quoted 50 lines

fs/proc/base.c static const struct pid_entry tgid_base_stuff[] = { · line 2857

Quoted 6 lines
fs/proc/base.c +1 -0
@@ -2854,6 +2854,7 @@ 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),

I must have missed something, but I fail to understand why this patch is so complicated.

Just use ONE("totmaps", S_IRUGO, proc_totmaps_operations) ?

fs/proc/task_mmu.c static int show_smap(struct seq_file *m, void *v, int is_pid) · line 837

Quoted 10 lines
fs/proc/task_mmu.c +10 -0
+static int totmaps_proc_show(struct seq_file *m, void *data)
+{
+	struct proc_maps_private *priv = m->private;
+	struct mm_struct *mm = priv->mm;
+	struct vm_area_struct *vma;
+	struct mem_size_stats mss_sum;
+
+	memset(&mss_sum, 0, sizeof(mss_sum));
+	down_read(&mm->mmap_sem);
+	hold_task_mempolicy(priv);
        ^^^^^^^^^^^^^^^^^^^^^^^^^

why?

fs/proc/task_mmu.c static int show_smap(struct seq_file *m, void *v, int is_pid) · line 839

Quoted 1 line
fs/proc/task_mmu.c +1 -0
+	for (vma = mm->mmap; vma != priv->tail_vma; vma = vma->vm_next) {

Hmm. the usage of ->tail_vma looks just wrong. I guess the code should work because it is NULL but still.

fs/proc/task_mmu.c static int show_smap(struct seq_file *m, void *v, int is_pid) · line 852

Quoted 13 lines
fs/proc/task_mmu.c +13 -0
+		struct mem_size_stats mss;
+		struct mm_walk smaps_walk = {
+			.pmd_entry = smaps_pte_range,
+			.mm = vma->vm_mm,
+			.private = &mss,
+		};
+
+		if (vma->vm_mm && !is_vm_hugetlb_page(vma)) {
+			memset(&mss, 0, sizeof(mss));
+			walk_page_vma(vma, &smaps_walk);
+			add_smaps_sum(&mss, &mss_sum);
+		}
+	}

Why? I mean, why not walk_page_range() ? You do not need this for-each-vma loop at all? At least if you change this patch to use the ONE() helper, and everything else looks unneeded in this case.

Oleg.

#4 lore raw

PATCH v5 2/3 Documentation/filesystems: Fixed typo

In reply to [email protected]

From: Robert Foss <robert.foss-ZGY8ohtN/[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 68080ad..fcc1ac0 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
Acked-by on v5 2/3
#5 lore raw

Re: [PATCH v5 2/3] Documentation/filesystems: Fixed typo

Quoted 5 lines · On Mon, Sep 5, 2016 at 1:14 PM, <[email protected]> wrote:

From: Robert Foss <[email protected]>

Fixed a -> an typo.

This could be taken directly into the docs tree, I think -- no reason to make it depend on the rest of the series.

-Kees

Quoted 20 lines
---
 Documentation/filesystems/proc.txt | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
Documentation/filesystems/proc.txt +1 -2
index 68080ad..fcc1ac0 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.
--

2.7.4

Signature · 2 lines
Kees Cook
Nexus Security
#6 lore raw

Re: [PATCH v5 2/3] Documentation/filesystems: Fixed typo

Quoted 11 lines · Kees Cook
Quoted 5 lines · On Mon, Sep 5, 2016 at 1:14 PM, <robert.foss-ZGY8ohtN/[email protected]…

From: Robert Foss <robert.foss-ZGY8ohtN/[email protected]>

Fixed a -> an typo.

This could be taken directly into the docs tree, I think -- no reason to make it depend on the rest of the series.

Agreed. Would you like a separate submission for that?

Rob.

maintainer
#7 lore raw

Re: [PATCH v5 2/3] Documentation/filesystems: Fixed typo

Quoted 4 lines · On Wed, 7 Sep 2016 20:22:00 -0400 Robert Foss <[email protected]> wrote:
Quoted 2 lines

This could be taken directly into the docs tree, I think -- no reason to make it depend on the rest of the series.

Agreed. Would you like a separate submission for that?

Please, I've lost track of the original...

Thanks,

jon

#8 lore raw

PATCH v5 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 fcc1ac0..49a8483 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 extension 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.
 ..............................................................................
@@ -515,6 +518,24 @@ 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
#9 lore raw
In reply to [email protected]
Quoted 16 lines · On Mon 05-09-16 16:14:06, [email protected] wrote:

From: Robert Foss <[email protected]>

This series provides the /proc/PID/totmaps feature, which summarizes the information provided by /proc/PID/smaps for improved performance and usability reasons.

A 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.

I still maintain my concerns about a single pss value. It might work in a very specific situations where the consumer knows what is shared but other than that the value can be more misleading than helpful. So a NACK from me until I am shown that this is usable in general and still helpful.

Signature · 2 lines
Michal Hocko
SUSE Labs
#10 lore raw
Quoted 22 lines · Michal Hocko
Quoted 16 lines · On Mon 05-09-16 16:14:06, robert.foss-ZGY8ohtN/[email protected]…

From: Robert Foss <robert.foss-ZGY8ohtN/[email protected]>

This series provides the /proc/PID/totmaps feature, which summarizes the information provided by /proc/PID/smaps for improved performance and usability reasons.

A 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.

Quoted 5 lines

I still maintain my concerns about a single pss value. It might work in a very specific situations where the consumer knows what is shared but other than that the value can be more misleading than helpful. So a NACK from me until I am shown that this is usable in general and still helpful.

I know you think Pss isn't useful in general (though I'll point out two other independent people said they found it useful) but how about the other fields like Swap, Private_Dirty and Private_Shared?

If we removed Pss would you still NACK it?

Quoted 4 lines

--
Michal Hocko
SUSE Labs

#11 lore raw
Quoted 23 lines · Sonny Rao
Quoted 22 lines · Michal Hocko
Quoted 20 lines · On Mon 05-09-16 16:14:06, [email protected] wrote:

From: Robert Foss <[email protected]>

This series provides the /proc/PID/totmaps feature, which summarizes the information provided by /proc/PID/smaps for improved performance and usability reasons.

A 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.

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.

Quoted 8 lines
Quoted 5 lines

I still maintain my concerns about a single pss value. It might work in a very specific situations where the consumer knows what is shared but other than that the value can be more misleading than helpful. So a NACK from me until I am shown that this is usable in general and still helpful.

I know you think Pss isn't useful in general (though I'll point out two other independent people said they found it useful)

sure, and one of them admitted that the value is useful because they _know_ the resource. The other was quite vague for me to understand all the details. Please, try to understand that once you provide a user API then it will carved in stone. If the interface is poor and ambigous it will bite us later. One very specific usecase doesn't justify something that might be really misleading for 90% of cases.

Quoted 2 lines

but how about the other fields like Swap, Private_Dirty and Private_Shared?

Private_Shared can be pretty confusing as well without the whole context as well see my other emails in the original thread (just to remind shmem/tmpfs makes all this really confusing).

Signature · 2 lines
Michal Hocko
SUSE Labs
#12 lore raw
Quoted 40 lines · Michal Hocko
Quoted 23 lines · Sonny Rao
Quoted 22 lines · Michal Hocko
Quoted 20 lines · On Mon 05-09-16 16:14:06, robert.foss-ZGY8ohtN/[email protected]…

From: Robert Foss <robert.foss-ZGY8ohtN/[email protected]>

This series provides the /proc/PID/totmaps feature, which summarizes the information provided by /proc/PID/smaps for improved performance and usability reasons.

A 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.

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.

Quoted 8 lines
Quoted 5 lines

I still maintain my concerns about a single pss value. It might work in a very specific situations where the consumer knows what is shared but other than that the value can be more misleading than helpful. So a NACK from me until I am shown that this is usable in general and still helpful.

I know you think Pss isn't useful in general (though I'll point out two other independent people said they found it useful)

sure, and one of them admitted that the value is useful because they _know_ the resource. The other was quite vague for me to understand all the details. Please, try to understand that once you provide a user API then it will carved in stone. If the interface is poor and ambigous it will bite us later. One very specific usecase doesn't justify something that might be really misleading for 90% of cases.

Quoted 6 lines
Quoted 2 lines

but how about the other fields like Swap, Private_Dirty and Private_Shared?

Private_Shared can be pretty confusing as well without the whole context as well see my other emails in the original thread (just to remind shmem/tmpfs makes all this really confusing).

But this is exactly the issue -- RSS is can be just as confusing if you don't know something about the application. I think the issue is how common that situation is, and you seem to believe that it's so uncommon that it's actually better to keep the information more difficult to get for those of us who know something about our systems.

That's fine, I guess we just have to disagree here, thanks for look at this.

Quoted 4 lines

--
Michal Hocko
SUSE Labs

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

On Mon 12-09-16 08:31:36, Sonny Rao wrote:

[...]

Quoted 9 lines
Quoted 6 lines
Quoted 2 lines

but how about the other fields like Swap, Private_Dirty and Private_Shared?

Private_Shared can be pretty confusing as well without the whole context as well see my other emails in the original thread (just to remind shmem/tmpfs makes all this really confusing).

But this is exactly the issue -- RSS is can be just as confusing if you don't know something about the application.

I agree that rss can be confusing but we will not make the situation any better if we add yet another confusing metric.

Quoted 6 lines

I think the issue is
how common that situation is, and you seem to believe that it's so uncommon that it's actually better to keep the information more difficult to get for those of us who know something about our systems.

That's fine, I guess we just have to disagree here, thanks for look at this.

I think you should just step back and think more about what exactly you expect from the counter(s). I believe what you want is an estimate of a freeable memory when the particular process dies or is killed. That would mean resident single mapped private anonymous memory
+ unlinked single mapped shareable mappings + single mapped swapped out memory. Maybe I've missed something but it should be something along those lines. Definitely something that the current smaps infrastructure doesn't give you, though.

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

On Mon 12-09-16 08:31:36, Sonny Rao wrote:

[...]

Quoted 9 lines
Quoted 6 lines
Quoted 2 lines

but how about the other fields like Swap, Private_Dirty and Private_Shared?

Private_Shared can be pretty confusing as well without the whole context as well see my other emails in the original thread (just to remind shmem/tmpfs makes all this really confusing).

But this is exactly the issue -- RSS is can be just as confusing if you don't know something about the application.

I agree that rss can be confusing but we will not make the situation any better if we add yet another confusing metric.

Quoted 6 lines

I think the issue is
how common that situation is, and you seem to believe that it's so uncommon that it's actually better to keep the information more difficult to get for those of us who know something about our systems.

That's fine, I guess we just have to disagree here, thanks for look at this.

Quoted 8 lines

I think you should just step back and think more about what exactly you expect from the counter(s). I believe what you want is an estimate of a freeable memory when the particular process dies or is killed. That would mean resident single mapped private anonymous memory
+ unlinked single mapped shareable mappings + single mapped swapped out memory. Maybe I've missed something but it should be something along those lines. Definitely something that the current smaps infrastructure doesn't give you, though.

Yes your description of what we want is pretty good. Having a reasonable lower bound on the estimate is fine, though we probably want to break out swapped out memory separately. Given that smaps doesn't provide this in a straightforward way, what do you think is the right way to provide this information?

Quoted 3 lines

--
Michal Hocko
SUSE Labs

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

On Mon 12-09-16 08:31:36, Sonny Rao wrote:

[...]

Quoted 9 lines
Quoted 6 lines
Quoted 2 lines

but how about the other fields like Swap, Private_Dirty and Private_Shared?

Private_Shared can be pretty confusing as well without the whole context as well see my other emails in the original thread (just to remind shmem/tmpfs makes all this really confusing).

But this is exactly the issue -- RSS is can be just as confusing if you don't know something about the application.

I agree that rss can be confusing but we will not make the situation any better if we add yet another confusing metric.

Quoted 6 lines

I think the issue is
how common that situation is, and you seem to believe that it's so uncommon that it's actually better to keep the information more difficult to get for those of us who know something about our systems.

That's fine, I guess we just have to disagree here, thanks for look at this.

Quoted 8 lines

I think you should just step back and think more about what exactly you expect from the counter(s). I believe what you want is an estimate of a freeable memory when the particular process dies or is killed. That would mean resident single mapped private anonymous memory
+ unlinked single mapped shareable mappings + single mapped swapped out memory. Maybe I've missed something but it should be something along those lines. Definitely something that the current smaps infrastructure doesn't give you, though.

Quoted 3 lines

Yes your description of what we want is pretty good. Having a reasonable lower bound on the estimate is fine, though we probably want to break out swapped out memory separately.

Why would you want to separate that?

Quoted 3 lines

Given that smaps
doesn't provide this in a straightforward way, what do you think is the right way to provide this information?

I would be tempted to sneak it into /proc/<pid>/statm because that looks like a proper place but getting this information is not for free performance wise so I am not really sure something that relies on this file would see unexpected stalls. Maybe this could be worked around by some caching... I would suggest to check who is actually using this file (top/ps etc...)

If this would be unacceptable then a new file could be considered.

Signature · 2 lines
Michal Hocko
SUSE Labs
#16 lore raw
Quoted 45 lines · Michal Hocko
Quoted 34 lines · Sonny Rao
Quoted 24 lines · Michal Hocko
Quoted 2 lines · Sonny Rao
Quoted 1 line · Michal Hocko

On Mon 12-09-16 08:31:36, Sonny Rao wrote:

[...]

Quoted 9 lines
Quoted 6 lines
Quoted 2 lines

but how about the other fields like Swap, Private_Dirty and Private_Shared?

Private_Shared can be pretty confusing as well without the whole context as well see my other emails in the original thread (just to remind shmem/tmpfs makes all this really confusing).

But this is exactly the issue -- RSS is can be just as confusing if you don't know something about the application.

I agree that rss can be confusing but we will not make the situation any better if we add yet another confusing metric.

Quoted 6 lines

I think the issue is
how common that situation is, and you seem to believe that it's so uncommon that it's actually better to keep the information more difficult to get for those of us who know something about our systems.

That's fine, I guess we just have to disagree here, thanks for look at this.

Quoted 8 lines

I think you should just step back and think more about what exactly you expect from the counter(s). I believe what you want is an estimate of a freeable memory when the particular process dies or is killed. That would mean resident single mapped private anonymous memory
+ unlinked single mapped shareable mappings + single mapped swapped out memory. Maybe I've missed something but it should be something along those lines. Definitely something that the current smaps infrastructure doesn't give you, though.

Quoted 3 lines

Yes your description of what we want is pretty good. Having a reasonable lower bound on the estimate is fine, though we probably want to break out swapped out memory separately.

Why would you want to separate that?

Quoted 3 lines

Given that smaps
doesn't provide this in a straightforward way, what do you think is the right way to provide this information?

Quoted 6 lines

I would be tempted to sneak it into /proc/<pid>/statm because that looks like a proper place but getting this information is not for free performance wise so I am not really sure something that relies on this file would see unexpected stalls. Maybe this could be worked around by some caching... I would suggest to check who is actually using this file (top/ps etc...)

What would this caching look like? Can any information be re-used between vma walks?

Quoted 3 lines

If this would be unacceptable then a new file could be considered.

#17 lore raw
Quoted 2 lines · Robert Foss
Quoted 1 line · Michal Hocko

On Tue 13-09-16 13:27:39, Sonny Rao wrote:

[...]

Quoted 11 lines
Quoted 10 lines
Quoted 3 lines

Given that smaps
doesn't provide this in a straightforward way, what do you think is the right way to provide this information?

I would be tempted to sneak it into /proc/<pid>/statm because that looks like a proper place but getting this information is not for free performance wise so I am not really sure something that relies on this file would see unexpected stalls. Maybe this could be worked around by some caching... I would suggest to check who is actually using this file (top/ps etc...)

Quoted 2 lines

What would this caching look like? Can any information be re-used between vma walks?

yes basically return the same value if called within HZ or something similar. But that assumes that statm latency really matters and it is called often enough.

Signature · 2 lines
Michal Hocko
SUSE Labs

A message that is not in the archive (sent off-list or not yet received): [email protected]

A message that is not in the archive (sent off-list or not yet received): [email protected]

#20 lore raw
Quoted 3 lines · Michal Hocko

[not sure why the CC list was trimmed - do no do that please unless you
have a strong reason for that - if this was not intentional please restpre it]

Ah, sorry, pressed the wrong key.

Quoted 41 lines · Jann Horn
Quoted 28 lines · Michal Hocko
Quoted 15 lines · Robert Foss
Quoted 2 lines · Michal Hocko
Quoted 1 line · Sonny Rao

[...]

Quoted 11 lines
Quoted 10 lines
Quoted 3 lines

Given that smaps
doesn't provide this in a straightforward way, what do you think is the right way to provide this information?

I would be tempted to sneak it into /proc/<pid>/statm because that looks like a proper place but getting this information is not for free performance wise so I am not really sure something that relies on this file would see unexpected stalls. Maybe this could be worked around by some caching... I would suggest to check who is actually using this file (top/ps etc...)

Quoted 6 lines
Quoted 2 lines

What would this caching look like? Can any information be re-used between vma walks?

yes basically return the same value if called within HZ or something similar. But that assumes that statm latency really matters and it is called often enough.

That sounds horrible. If some application decides that they want to check statm directly after some action or so (like after program startup), this is going to give them a very bad time. That probably doesn't happen often - but still.

Quoted 6 lines

I can already imagine some developer going "yeah, that usleep()... that's because the kernel API returns stale information for a couple milliseconds after we do something *shrug*".

What are you trying to optimize for? Ten users on the same machine, each of which is running "top" because it looks so great?

Please try to read what I wrote again. I didn't say this would be needed. The idea was that _if_ /proc/<pid>/statm is used very _often_ than some caching might help to reduce the overhead. Especially when you consider that the information is not precise anyway. It can change anytime while you are doing the address space walk.

#21 lore raw
Quoted 49 lines · Jann Horn
Quoted 3 lines · Michal Hocko

[not sure why the CC list was trimmed - do no do that please unless you
have a strong reason for that - if this was not intentional please restpre it]

Ah, sorry, pressed the wrong key.

Quoted 41 lines · Jann Horn
Quoted 28 lines · Michal Hocko
Quoted 15 lines · Robert Foss
Quoted 2 lines · Michal Hocko
Quoted 1 line · Sonny Rao

[...]

Quoted 11 lines
Quoted 10 lines
Quoted 3 lines

Given that smaps
doesn't provide this in a straightforward way, what do you think is the right way to provide this information?

I would be tempted to sneak it into /proc/<pid>/statm because that looks like a proper place but getting this information is not for free performance wise so I am not really sure something that relies on this file would see unexpected stalls. Maybe this could be worked around by some caching... I would suggest to check who is actually using this file (top/ps etc...)

Quoted 6 lines
Quoted 2 lines

What would this caching look like? Can any information be re-used between vma walks?

yes basically return the same value if called within HZ or something similar. But that assumes that statm latency really matters and it is called often enough.

That sounds horrible. If some application decides that they want to check statm directly after some action or so (like after program startup), this is going to give them a very bad time. That probably doesn't happen often - but still.

Quoted 6 lines

I can already imagine some developer going "yeah, that usleep()... that's because the kernel API returns stale information for a couple milliseconds after we do something *shrug*".

What are you trying to optimize for? Ten users on the same machine, each of which is running "top" because it looks so great?

Please try to read what I wrote again. I didn't say this would be needed. The idea was that _if_ /proc/<pid>/statm is used very _often_ than some caching might help to reduce the overhead. Especially when you consider that the information is not precise anyway. It can change anytime while you are doing the address space walk.

Just thinking out loud here -- I haven't looked closely at the code so please bear with me :-)

Instead of checking when the last read was and returning old data, what about a scheme where we still have a timestamp for last stat read on and any changes to that address space invalidate the timestamp.

The invalidation could be racy because we're not too concerned about
immediate accuracy -- so just a write. The main issue I could see
which this is that it could cause the cacheline holding this timestamp to bounce around a lot? Maybe there's an existing solution in the page table locking that could be leveraged here to at least maintain whatever scalability enhancements are present for this type of situation where there are many updates happening in parallel.

#22 lore raw
In reply to Michal Hocko
Quoted 15 lines · Michal Hocko
Quoted 2 lines · Robert Foss
Quoted 1 line · Michal Hocko

On Tue 13-09-16 13:27:39, Sonny Rao wrote:

[...]

Quoted 11 lines
Quoted 10 lines
Quoted 3 lines

Given that smaps
doesn't provide this in a straightforward way, what do you think is the right way to provide this information?

I would be tempted to sneak it into /proc/<pid>/statm because that looks like a proper place but getting this information is not for free performance wise so I am not really sure something that relies on this file would see unexpected stalls. Maybe this could be worked around by some caching... I would suggest to check who is actually using this file (top/ps etc...)

Quoted 6 lines
Quoted 2 lines

What would this caching look like? Can any information be re-used between vma walks?

yes basically return the same value if called within HZ or something similar. But that assumes that statm latency really matters and it is called often enough.

Any single application querying more often than HZ, would presumably do so for accuracy reasons.
However for multiple applications that combined query more often than HZ, this would most definitely be halpful in terms of performance.

@Sonny, does chromiumos fall into the first or second category?

#23 lore raw
Quoted 34 lines · Robert Foss
Quoted 22 lines · Michal Hocko
Quoted 4 lines · Robert Foss
Quoted 2 lines · Michal Hocko

On Tue 13-09-16 13:27:39, Sonny Rao wrote:

[...]

Quoted 14 lines
Quoted 12 lines
Quoted 4 lines

Given that smaps
doesn't provide this in a straightforward way, what do you think is the right way to provide this information?

I would be tempted to sneak it into /proc/<pid>/statm because that looks like a proper place but getting this information is not for free performance wise so I am not really sure something that relies on this file would see unexpected stalls. Maybe this could be worked around by some caching... I would suggest to check who is actually using this file (top/ps etc...)

Quoted 7 lines
Quoted 2 lines

What would this caching look like? Can any information be re-used between vma walks?

yes basically return the same value if called within HZ or something similar. But that assumes that statm latency really matters and it is called often enough.

Quoted 6 lines

Any single application querying more often than HZ, would presumably do so for accuracy reasons.
However for multiple applications that combined query more often than HZ, this would most definitely be halpful in terms of performance.

@Sonny, does chromiumos fall into the first or second category?

It's a single application -- and it definitely doesn't query at HZ -- especially given how long it takes to gather the data :-)