bug#49656: more data on #43389

bug#49656: more data on #43389

[Madhu]

[-- Attachment #1: Type: Text/Plain, Size: 809 bytes --]

This is the emacs memory bloat issue. I thought I had unarchived the
bug earlier but it seems to be closed up again. Should I open a new
bug? Its already got a 1000 posts on the bugzilla page

Over the past few months, I've experienced the memory problem a few
times.  I wanted to share the malloc info data I collected on 3
instance and am attaching a tarball with 3 directories named 4 5 6
which include malloc-info, memory-usage, and memory-report data from
emacs exhibiting the behaviour - after all buffers are killed and
after an M-x gc.

Directories 4 and 6 show the common pattern where memory is used
unknown to GC. data in directory 5 probably another issue where there
are a lot of dead strings.

I'd be glad if/when experts can find the time to take a look and see
if it offers any further clues.

[-- Attachment #2: madhu-emacs-malloc-info-data-20210719.tar.gz --]
[-- Type: Application/Octet-Stream, Size: 7042 bytes --]

bug#49656: more data on #43389

[Eli Zaretskii]

> Date: Tue, 20 Jul 2021 12:43:31 +0530 (IST)
> From: Madhu <enometh@meer.net>
> 
> This is the emacs memory bloat issue. I thought I had unarchived the
> bug earlier but it seems to be closed up again. Should I open a new
> bug? Its already got a 1000 posts on the bugzilla page

I'm not yet sure this is a bug, FWIW.

> Over the past few months, I've experienced the memory problem a few
> times.  I wanted to share the malloc info data I collected on 3
> instance and am attaching a tarball with 3 directories named 4 5 6
> which include malloc-info, memory-usage, and memory-report data from
> emacs exhibiting the behaviour - after all buffers are killed and
> after an M-x gc.
> 
> Directories 4 and 6 show the common pattern where memory is used
> unknown to GC. data in directory 5 probably another issue where there
> are a lot of dead strings.

I've looked at the data, thanks.  But the most important information
is missing from the data you presented.  The case where you have 570MB
in strings could be interesting, but without knowing what you did in
that session, it is impossible to say whether there's an issue there.
It could be explained by the live Lisp data you produced in that
session.

So what we need is the following, for each case separately:

  . the size of the memory footprint of the Emacs session
  . how long was the Emacs session running since it started
  . what were the main activities in that session
  . do you have some customizations related to GC, and if you do, what
    are they
  . what Emacs version is that, and what are its configuration
    parameters and features compiled into it

In general, that old bug is closed, because after making a single
change that did cause memory bloat in some usage patterns, we
concluded that the rest were user problems caused by messing with GC
thresholds.  The glibc developers looked at the data we collected and
found nothing that could be interpreted as a bug in glibc.

TIA

bug#49656: more data on #43389

[Lars Ingebrigtsen]

Madhu <enometh@meer.net> writes:

> This is the emacs memory bloat issue. I thought I had unarchived the
> bug earlier but it seems to be closed up again.

Unarchiving a bug doesn't reopen it -- you have to do both.

> Should I open a new bug? Its already got a 1000 posts on the bugzilla
> page

You did open a new bug report.  :-)  bug#49656.

-- 
(domestic pets only, the antidote for overdose, milk.)
   bloggy blog: http://lars.ingebrigtsen.no

bug#49656: more data on #43389

[Madhu]

*  Eli Zaretskii <eliz@gnu.org> <83o8ax5hxp.fsf@gnu.org>
Wrote on Tue, 20 Jul 2021 14:43:46 +0300

>> This is the emacs memory bloat issue. I thought I had unarchived the
>> bug earlier but it seems to be closed up again. Should I open a new
>> bug? Its already got a 1000 posts on the bugzilla page
> I'm not yet sure this is a bug, FWIW.

It's been bugging me for a few years now, but I won't mind if you
close it.

>> Over the past few months, I've experienced the memory problem a few
>> times.  I wanted to share the malloc info data I collected on 3
>> instance and am attaching a tarball with 3 directories named 4 5 6
>> which include malloc-info, memory-usage, and memory-report data from
>> emacs exhibiting the behaviour - after all buffers are killed and
>> after an M-x gc.
>> Directories 4 and 6 show the common pattern where memory is used
>> unknown to GC. data in directory 5 probably another issue where there
>> are a lot of dead strings.
> I've looked at the data, thanks.  But the most important information
> is missing from the data you presented.  The case where you have 570MB
> in strings could be interesting, but without knowing what you did in
> that session, it is impossible to say whether there's an issue there.
> It could be explained by the live Lisp data you produced in that
> session.

That 570MB case was the exceptional case - i noticed that only
once.

In all other cases I hit the memory leak gc thinks it only has 200M
while the process has 1.5 to 3.5 GB resident.

I don't see how this can  be a user error.

> So what we need is the following, for each case separately:
>   . the size of the memory footprint of the Emacs session

I have this data for a few runs.  The sizes from `ps' correspond
exactly to what malloc info shows so i decided to omit it.

>   . how long was the Emacs session running since it started

2-3 days usually. I always have to kill emacs because I cant suspend
the OS to disk because all of it is resident.

>   . what were the main activities in that session . do you have some
>   customizations related to GC, and if you do, what are they . what
>   Emacs version is that, and what are its configuration parameters
>   and features compiled into it

I am not doing anything special with GC. All the recent reports are on
Emacs 28, within a month of two of master.  The problem is independent
of the toolkit, I get it in motif, gtk, and xt.  I havent tried a no-x
build in 2 years.

if you remember I asked on emacs-help once and you informed me of a GC
fix and it fixed that problem.

> In general, that old bug is closed, because after making a single
> change that did cause memory bloat in some usage patterns, we
> concluded that the rest were user problems caused by messing with GC
> thresholds.  The glibc developers looked at the data we collected
> and found nothing that could be interpreted as a bug in glibc.

But doesn't the malloc info allocation and the gc reports indicate a
clear discrepancy?  Does it not indicate a leak which the GC is not
aware of?

bug#49656: more data on #43389

[Eli Zaretskii]

> Date: Tue, 20 Jul 2021 21:38:37 +0530 (IST)
> Cc: 49656@debbugs.gnu.org
> From: Madhu <enometh@meer.net>
> 
> In all other cases I hit the memory leak gc thinks it only has 200M
> while the process has 1.5 to 3.5 GB resident.
> 
> I don't see how this can  be a user error.

It could be if, for example, you are using packages or your own code
which plays with GC thresholds.  E.g., I'm told that lsp-mode does
that; are you per chance using it?

> > So what we need is the following, for each case separately:
> >   . the size of the memory footprint of the Emacs session
> 
> I have this data for a few runs.  The sizes from `ps' correspond
> exactly to what malloc info shows so i decided to omit it.

Ah, but gleaning this from the malloc info takes some time, so telling
the number explicitly will help making the report more easily
understandable.

> >   . how long was the Emacs session running since it started
> 
> 2-3 days usually. I always have to kill emacs because I cant suspend
> the OS to disk because all of it is resident.

How much memory and swap do you have there?  Can you enlarge the total
VM size (by enlarging swap) so that you could run Emacs longer when it
gets to such large sizes?

Btw, 1.5GB is not too large for several days worth of work.

> >   . what were the main activities in that session . do you have some
> >   customizations related to GC, and if you do, what are they . what
> >   Emacs version is that, and what are its configuration parameters
> >   and features compiled into it
> 
> I am not doing anything special with GC.

Are you sure no package you use does something like that?  What were
the values of GC thresholds when the memory was large?  Was the value
of gcs-done increasing with time, or did it stay put (which would be
an evidence that Emacs doesn't run GC at all)?  If GC was running, how
frequently did it run?  (You could answer the last question either by
looking at the rate of growth in gcs-done, or by setting
garbage-collection-messages non-nil, which will cause Emacs announce
ever GC in the echo area.)

> if you remember I asked on emacs-help once and you informed me of a GC
> fix and it fixed that problem.

But that problem was fixed, so it cannot possibly affect you now.  And
we found that problem because the user reported lack of GC cycles
after some point.

> But doesn't the malloc info allocation and the gc reports indicate a
> clear discrepancy?

Not to me, it doesn't.  It means the malloc arena holds to a lot of
memory that it cannot release to the OS, but it is known that this can
happen for certain patterns of memory allocation and deallocation.
You can find explanations of why this happens on the net.

> Does it not indicate a leak which the GC is not aware of?

No.  Emacs allocates a lot of memory for purposes other than Lisp
objects, and that memory is not visible to GC nor to memory-report.

bug#49656: more data on #43389

[Madhu]

*  Eli Zaretskii <eliz@gnu.org> <837dhk6hby.fsf@gnu.org>
Wrote on Tue, 20 Jul 2021 20:11:29 +0300
>> Date: Tue, 20 Jul 2021 21:38:37 +0530 (IST)
>> From: Madhu <enometh@meer.net>
>>
>> In all other cases I hit the memory leak gc thinks it only has 200M
>> while the process has 1.5 to 3.5 GB resident.
>>
>> I don't see how this can  be a user error.
>
> It could be if, for example, you are using packages or your own code
> which plays with GC thresholds.  E.g., I'm told that lsp-mode does
> that; are you per chance using it?

No I haven't gotten around to trying that i'm sure none of my packages
mess with gc-threshold.  I will double check that.

> How much memory and swap do you have there?  Can you enlarge the total
> VM size (by enlarging swap) so that you could run Emacs longer when it
> gets to such large sizes?

The problem is not that emacs won't run.  I am unable to use the
memory in the machine for other purposes and am obliged to kill emacs.

> Btw, 1.5GB is not too large for several days worth of work.

I also have emacs running for several weeks when the problem doesn't
happen with RES never above say 500M for the same workloads and usage
patterns.


> Are you sure no package you use does something like that?  What were
> the values of GC thresholds when the memory was large?

Yes I do not run a lot of packages and certainly nothing that I don't
audit first - I am aware of all the packages that are loaded at least
in the exceptional cases.  There is no reason for me to suspect those
packages because I use them all the time in emacs sessions where the
problem does not manifest.

>  Was the value of gcs-done increasing with time, or did it stay put
> (which would be an evidence that Emacs doesn't run GC at all)?  If
> GC was running, how frequently did it run?  (You could answer the
> last question either by looking at the rate of growth in gcs-done,
> or by setting garbage-collection-messages non-nil, which will cause
> Emacs announce ever GC in the echo area.)

There is nothing exceptional with GC. GC usually completes quickly
because it doesnt access much memory.


>> But doesn't the malloc info allocation and the gc reports indicate a
>> clear discrepancy?
>
> Not to me, it doesn't.  It means the malloc arena holds to a lot of
> memory that it cannot release to the OS, but it is known that this can
> happen for certain patterns of memory allocation and deallocation.
> You can find explanations of why this happens on the net.

But that is the point - I've been asking from the first time I posted
on this - WHAT is emacs allocating in these exceptional cases.  I
understand my usage patterns over the years of constant emacs use and
the "random" memory bloat in some sessions makes no sense and it only
suggests a memory leak in emacs code.

>> Does it not indicate a leak which the GC is not aware of?
>
> No.  Emacs allocates a lot of memory for purposes other than Lisp
> objects, and that memory is not visible to GC nor to memory-report.

This is too vague.  This 2GB active memory is not legitimate and the
point is to figure out where and why emacs is holding on to this
memory.


This is where I need assistance.  my usage patterns are pretty
deterministic The memory usage of the qlisp packages I use would show
up in GC.

SO what could that be, which it is NOT allocating under identical
usage patterns?  It isn't fonts or external images or anything else I
can see.

I'm sure others should be hitting the problem too

bug#49656: more data on #43389

[Eli Zaretskii]

> Date: Wed, 21 Jul 2021 07:16:23 +0530 (IST)
> Cc: 49656@debbugs.gnu.org
> From: Madhu <enometh@meer.net>
> 
> > It could be if, for example, you are using packages or your own code
> > which plays with GC thresholds.  E.g., I'm told that lsp-mode does
> > that; are you per chance using it?
> 
> No I haven't gotten around to trying that i'm sure none of my packages
> mess with gc-threshold.  I will double check that.

Thanks, please do.  It's important to know that.

> > How much memory and swap do you have there?  Can you enlarge the total
> > VM size (by enlarging swap) so that you could run Emacs longer when it
> > gets to such large sizes?
> 
> The problem is not that emacs won't run.  I am unable to use the
> memory in the machine for other purposes and am obliged to kill emacs.

If you enlarge the swap space, your system will still be workable even
when Emacs has such a large footprint.  That's the purpose of my
suggestion.  It is important to have a usable system when these
situations happen because if we come up with some experiments to
conduct in such cases, you need a usable system to be able to do that.

> > Btw, 1.5GB is not too large for several days worth of work.
> 
> I also have emacs running for several weeks when the problem doesn't
> happen with RES never above say 500M for the same workloads and usage
> patterns.

Then one thing to try to figure out is what was the difference between
these two classes of sessions -- what and how did you do differently
in one that you didn't in the other?  That could point us in the right
direction.

> >  Was the value of gcs-done increasing with time, or did it stay put
> > (which would be an evidence that Emacs doesn't run GC at all)?  If
> > GC was running, how frequently did it run?  (You could answer the
> > last question either by looking at the rate of growth in gcs-done,
> > or by setting garbage-collection-messages non-nil, which will cause
> > Emacs announce ever GC in the echo area.)
> 
> There is nothing exceptional with GC. GC usually completes quickly
> because it doesnt access much memory.

I'd like quantitative measures, please: what is the rate of GC cycles
when the memory footprint is measurd in GBs, and how much time does
each GC cycle take?  Also, does the memory footprint becomes smaller
after a GC cycle, and by how much?

> >> But doesn't the malloc info allocation and the gc reports indicate a
> >> clear discrepancy?
> >
> > Not to me, it doesn't.  It means the malloc arena holds to a lot of
> > memory that it cannot release to the OS, but it is known that this can
> > happen for certain patterns of memory allocation and deallocation.
> > You can find explanations of why this happens on the net.
> 
> But that is the point - I've been asking from the first time I posted
> on this - WHAT is emacs allocating in these exceptional cases.

It's impractical to try to answer these questions.  If you put a
breakpoint on 'malloc' and 'free' and then run Emacs, you will see
that it calls these functions extremely frequently, with widely
different block sizes.  We don't have infrastructure that would record
each allocation and deallocation with enough info to be able to
analyze that, and even if we did, finding the callers which are
responsible for the memory that's not returned to the OS is a very
large and hard job.  We tried that previously, with the help of the
glibc developers using their debugging features -- it didn't really
help us with the diagnostics.

> I understand my usage patterns over the years of constant emacs use
> and the "random" memory bloat in some sessions makes no sense and it
> only suggests a memory leak in emacs code.

It is impossible to look for memory leaks in Emacs without some clues
about where those leaks could be.  If you can provide the information
I asked above, we might be able to come up with such clues.

Alternatively, you could try running Emacs under Valgrind (see
instructions in etc/DEBUG), although this will probably catch memory
leaks only on the C level, not on the Lisp level.

> >> Does it not indicate a leak which the GC is not aware of?
> >
> > No.  Emacs allocates a lot of memory for purposes other than Lisp
> > objects, and that memory is not visible to GC nor to memory-report.
> 
> This is too vague.

But that's all I can say, given the information you provided till now.

> This 2GB active memory is not legitimate and the point is to figure
> out where and why emacs is holding on to this memory.
> 
> This is where I need assistance.

This is me assisting you.  I'm asking you to help us analyze this
problem by providing more information.  I don't think we will be able
to make progress here without that additional info.

> SO what could that be, which it is NOT allocating under identical
> usage patterns?

I don't know, sorry.

> I'm sure others should be hitting the problem too

We could, of course, wait for others to report similar problems, and
then hope the information they provide would point us to the right
direction.  IME, though, this is not a very efficient method, because
in many cases the reasons for the memory growth are not the same.

bug#49656: Abnormal Emacs memory usage (bug#43389)

[Lars Ingebrigtsen]

Eli Zaretskii <eliz@gnu.org> writes:

>> No I haven't gotten around to trying that i'm sure none of my packages
>> mess with gc-threshold.  I will double check that.
>
> Thanks, please do.  It's important to know that.

(I'm going through old bug reports that unfortunately weren't resolved
at the time.)

This was a year ago -- Madhu, did you make any progress in identifying
the problem (if you're still seeing it)?

bug#49656: Abnormal Emacs memory usage (bug#43389)

[Madhu]

*  Lars Ingebrigtsen <larsi@gnus.org> <871qt9bhe1.fsf_-_@gnus.org>
Wrote on Sun, 21 Aug 2022 20:06:46 +0200
> Eli Zaretskii <eliz@gnu.org> writes:
>>> No I haven't gotten around to trying that i'm sure none of my packages
>>> mess with gc-threshold.  I will double check that.
>>
>> Thanks, please do.  It's important to know that.
>
> (I'm going through old bug reports that unfortunately weren't resolved
> at the time.)
>
> This was a year ago -- Madhu, did you make any progress in identifying
> the problem (if you're still seeing it)?

No, I haven't triggered this problem recently with emacs-29.  Maybe
I've stopped using the packages that triggered it.  The bug should be
probably closed, Thanks

bug#49656: Abnormal Emacs memory usage (bug#43389)

[Lars Ingebrigtsen]

Madhu <enometh@meer.net> writes:

> No, I haven't triggered this problem recently with emacs-29.  Maybe
> I've stopped using the packages that triggered it.  The bug should be
> probably closed, Thanks

OK; done.