Re: devel/llvm13 failed to reclaim memory on 8 GB Pi4 running -current [UFS context: used the whole swap space too]

From: Mark Millard <marklmi_at_yahoo.com>
Date: Fri, 28 Jan 2022 08:31:17 UTC
[Back to omitting Mark Johnston.]

On 2022-Jan-27, at 23:40, Mark Millard <marklmi@yahoo.com> wrote:

> [Mark Johnston included: handling was different than under ZFS.]
> 
> On 2022-Jan-27, at 22:35, Mark Millard <marklmi@yahoo.com> wrote:
> 
>> [Back to omitting Mark Johnston.]
>> 
>> On 2022-Jan-27, at 22:00, Mark Millard <marklmi@yahoo.com> wrote:
>> 
>>> On 2022-Jan-27, at 21:55, Mark Millard <marklmi@yahoo.com> wrote:
>>> 
>>>> On 2022-Jan-27, at 17:43, Mark Millard <marklmi@yahoo.com> wrote:
>>>> 
>>>>> On 2022-Jan-27, at 15:30, bob prohaska <fbsd@www.zefox.net> wrote:
>>>>> 
>>>>>> On Thu, Jan 27, 2022 at 02:21:44PM -0800, Mark Millard wrote:
>>>>>>> 
>>>>>>> Okay. I just started a poudriere bulk devel/llvm13 build
>>>>>>> in a ZFS context:
>>>>>>> 
>>>>>>> . . .
>>>>>>> [00:00:37] Pkg: +BE_AMDGPU -BE_FREEBSD +BE_NATIVE -BE_STANDARD +BE_WASM +CLANG +DOCS +EXTRAS -FLANG +LIT +LLD +LLDB +MLIR -OPENMP -PYCLANG
>>>>>>> [00:00:37] New: +BE_AMDGPU -BE_FREEBSD -BE_NATIVE +BE_STANDARD +BE_WASM +CLANG +DOCS +EXTRAS +FLANG +LIT +LLD +LLDB +MLIR +OPENMP +PYCLANG
>>>>>>> . . .
>>>>>>> [00:01:27] [01] [00:00:00] Building devel/llvm13 | llvm13-13.0.0_3
>>>>>>> 
>>>>>> 
>>>>>> Is this ARM hardware, or an emulator?
>>>>> 
>>>>> 8 GiByte RPi4B, USB3 NVMe media with a ZFS partition. The content
>>>>> is a slightly modified copy of the HoneyComb's PCIe slot Optane
>>>>> media.
>>>>> 
>>>>> The UFS-based 8 GiByte RPi4B is also based on copying from the
>>>>> same Optane media, both for the system materials and various
>>>>> ports/packages/pouriere related materials. (Not, necessarily,
>>>>> other things.)
>>>>> 
>>>>>> I've been using plain old make in /usr/ports/devel, 
>>>>>> might it be informative to try a poudriere build as well?
>>>>> 
>>>>> The Pkg:, New:, and llvm13 lines I listed are poudriere(-devel)
>>>>> output. I am doing my builds via poudriere. ALLOW_PARALLEL_JOBS=
>>>>> and USE_TMPFS="data" in use.
>>>>> 
>>>>> I have a context in which almost all prerequisites had already
>>>>> been built. (The change in options lead to 2 very small ports
>>>>> to build before devel/llvm13's started in a builder.)
>>>>> 
>>>>> (You might not have a jail that already has the prerequisites.)
>>>>> 
>>>>>> One would expect the added overhead to increase memory use.
>>>>>> 
>>>>> 
>>>>> Well, from the context I started in, only devel/llvm13 is being
>>>>> built once it starts. Once it gets to the build phase (after
>>>>> dependencies and such are set up), there is not much overhead
>>>>> because the only activity is the one builder and it is only
>>>>> building llvm13 --via make in the builder. At the end there
>>>>> would be extra activity as poudriere finishes up. During the
>>>>> build phase, I only expect minor overhead from poudriere
>>>>> monitoring the build logs and such.
>>>>> 
>>>>> I expect that the mere fact that a poudriere jail is in use
>>>>> for the builder to execute in does not contribute to
>>>>> significantly increasing the system's memory use or changing
>>>>> the system's memory use pattern.
>>>>> 
>>>>> 
>>>>> There are some other differences my context. The instances of
>>>>> main [so: 14] are non-debug builds (but with symbols). The
>>>>> builds are optimized for the RPi4B (and others) via use of
>>>>> -mcpu=cortex-a72 usage. My /usr/main-src/ does have some
>>>>> personal changes in it. (Some messaging about the kills is
>>>>> part of that.)
>>>>> 
>>>>> The RPi4B's are using:
>>>>> 
>>>>> over_voltage=6 
>>>>> arm_freq=2000 
>>>>> sdram_freq_min=3200 
>>>>> force_turbo=1 
>>>>> 
>>>>> (There are heat-sinks, fans, and good power supplies.)
>>>>> 
>>>>> The media in use are USB3 1 TB Samsung Portable SSD T7
>>>>> Touch's. I'm unlikely to see "swap_pager: indefinite
>>>>> wait buffer:" notices if the cause was based on the
>>>>> media performance. (You have spinning rust, if I
>>>>> remember right.)
>>>>> 
>>>>> I do not have a monitoring script making a huge log file
>>>>> during the build. So less is competing for media access
>>>>> or leading to other overheads. (But, as I remember,
>>>>> you have gotten the problem without having such a script
>>>>> running.)
>>>> 
>>>> 
>>>> ZFS context:
>>>> 
>>>> Well, the ZFS example used up all the swap space, according
>>>> to my patched top. This means that my setting of
>>>> vm.pfault_oom_attempts is not appropriate for this context:
>>>> 
>>>> # Delay when persistent low free RAM leads to
>>>> # Out Of Memory killing of processes:
>>>> vm.pageout_oom_seq=120
>>>> #
>>>> # For plunty of swap/paging space (will not
>>>> # run out), avoid pageout delays leading to
>>>> # Out Of Memory killing of processes:
>>>> vm.pfault_oom_attempts=-1
>>>> #
>>>> # For possibly insufficient swap/paging space
>>>> # (might run out), increase the pageout delay
>>>> # that leads to Out Of Memory killing of
>>>> # processes (showing defaults at the time):
>>>> #vm.pfault_oom_attempts= 3
>>>> #vm.pfault_oom_wait= 10
>>>> # (The multiplication is the total but there
>>>> # are other potential tradoffs in the factors
>>>> # multiplied, even for nearly the same total.)
>>>> 
>>>> I'll need to retest with something more like the
>>>> commented out vm.pfault_oom_attempts and
>>>> vm.pfault_oom_wait figures in order to see the
>>>> intended handling of the test case.
>>>> 
>>>> What are you using for each of:
>>>> vm.pageout_oom_seq ?
>>>> vm.pfault_oom_attempts ?
>>>> vm.pfault_oom_wait ?
>>>> 
>>>> 
>>>> For reference, for ZFS:
>>>> 
>>>> last pid:   380;  load averages:   1.50,   3.07,   3.93 MaxObs:   5.71,   4.92,   4.76                                                                                          up 0+07:23:14  21:23:43
>>>> 68 threads:    1 running, 65 sleeping, 2 waiting, 19 MaxObsRunning
>>>> CPU: 13.3% user,  0.0% nice,  4.9% system,  0.9% interrupt, 80.8% idle
>>>> Mem: 4912Mi Active, 167936B Inact, 1193Mi Laundry, 1536Mi Wired, 40960B Buf, 33860Ki Free, 6179Mi MaxObsActive, 6476Mi MaxObsWired, 7820Mi MaxObs(Act+Wir+Lndry)
>>>> ARC: 777086Ki Total, 132156Ki MFU, 181164Ki MRU, 147456B Anon, 5994Ki Header, 457626Ki Other
>>>>  59308Ki Compressed, 254381Ki Uncompressed, 4.29:1 Ratio
>>>> Swap: 8192Mi Total, 8192Mi Used, K Free, 100% Inuse, 19572Ki In, 3436Ki Out, 8192Mi MaxObsUsed, 14458Mi MaxObs(Act+Lndry+SwapUsed), 15993Mi MaxObs(Act+Wir+Lndry+SwapUsed)
>>>> 
>>>> Console:
>>>> (Looks like I misremembered adjusting the "out of swap space"
>>>> wording for the misnomer message.)
>>>> 
>>>> swap_pager: out of swap space
>>>> swp_pager_getswapspace(18): failed
>>>> swap_pager: out of swap space
>>>> swp_pager_getswapspace(1): failed
>>>> swp_pager_getswapspace(1): failed
>>>> swap_pager: out of swap space
>>>> swp_pager_getswapspace(1): failed
>>>> swp_pager_getswapspace(7): failed
>>>> swp_pager_getswapspace(24): failed
>>>> swp_pager_getswapspace(3): failed
>>>> swp_pager_getswapspace(18): failed
>>>> swp_pager_getswapspace(17): failed
>>>> swp_pager_getswapspace(1): failed
>>>> swp_pager_getswapspace(12): failed
>>>> swp_pager_getswapspace(23): failed
>>>> swp_pager_getswapspace(30): failed
>>>> swp_pager_getswapspace(3): failed
>>>> swp_pager_getswapspace(2): failed
>>>> 
>>>> . . . Then a bunch of time with no messages . . .
>>>> 
>>>> swp_pager_getswapspace(5): failed
>>>> swp_pager_getswapspace(28): failed
>>>> 
>>>> . . . Then a bunch of time with no messages . . .
>>>> 
>>>> 
>>>> Top again:
>>>> 
>>>> last pid:   382;  load averages:   0.73,   1.00,   2.40 MaxObs:   5.71,   4.92,   4.76                                                                                          up 0+07:31:26  21:31:55
>>>> 70 threads:    1 running, 65 sleeping, 4 waiting, 19 MaxObsRunning
>>>> CPU:  0.1% user,  0.0% nice,  5.6% system,  0.0% interrupt, 94.3% idle
>>>> Mem: 3499Mi Active, 4096B Inact, 2612Mi Laundry, 1457Mi Wired, 40960B Buf, 34676Ki Free, 6179Mi MaxObsActive, 6476Mi MaxObsWired, 7820Mi MaxObs(Act+Wir+Lndry)
>>>> ARC: 777154Ki Total, 135196Ki MFU, 178330Ki MRU, 5995Ki Header, 457631Ki Other
>>>>  59520Ki Compressed, 254231Ki Uncompressed, 4.27:1 Ratio
>>>> Swap: 8192Mi Total, 8192Mi Used, K Free, 100% Inuse, 409600B In, 4096B Out, 8192Mi MaxObsUsed, 14458Mi MaxObs(Act+Lndry+SwapUsed), 15993Mi MaxObs(Act+Wir+Lndry+SwapUsed)
>>>> 
>>>> 
>>>> I then used top to kill ninja and the 4 large compiles
>>>> that were going on. I'll change:
>>>> 
>>>> vm.pfault_oom_attempts
>>>> vm.pfault_oom_wait
>>>> 
>>>> and reboot and start over.
>>>> 
>>>> 
>>>> I expect that the ongoing UFS test will likely end up
>>>> similarly and that similar adjustments and restarts
>>>> will be needed because of actually running out of
>>>> swap space.
>>>> 
>>> 
>>> I forgot to report:
>>> 
>>> [00:01:27] [01] [00:00:00] Building devel/llvm13 | llvm13-13.0.0_3
>>> [07:49:17] [01] [07:47:50] Finished devel/llvm13 | llvm13-13.0.0_3: Failed: build
>>> 
>>> So the swap space filling happened somewhat before
>>> that much time had passed.
>> 
>> ZFS context:
>> 
>> I will not start the next bulk until just before bed. I do not
>> want it to fail while I'm not monitoring it.
>> 
>> The last 4 reported compile starts reported in the log are:
>> 
>> [ 64% 4725/7265] . . . flang/lib/Evaluate/fold.cpp
>> [ 65% 4726/7265] . . . flang/lib/Evaluate/fold-character.cpp
>> [ 65% 4727/7265] . . . flang/lib/Evaluate/check-expression.cpp
>> [ 65% 4728/7265] . . . flang/lib/Evaluate/fold-designator.cpp
>> 
>> But it is possible one or more of these completed and some
>> earlier one(s) was(were) still running.
>> 
>> So, if you do not need the Fortran compiler, you can probably
>> avoid the problem by setting the options for devel/llvm13 to
>> not build flang.
> 
> UFS context:
> 
> . . .;  load averages:   . . . MaxObs:   5.47,   4.99,   4.82
> . . . threads:    . . ., 14 MaxObsRunning
> . . .
> Mem: . . ., 6457Mi MaxObsActive, 1263Mi MaxObsWired, 7830Mi MaxObs(Act+Wir+Lndry)
> Swap: 8192Mi Total, 8192Mi Used, K Free, 100% Inuse, 8192Mi MaxObsUsed, 14758Mi MaxObs(Act+Lndry+SwapUsed), 16017Mi MaxObs(Act+Wir+Lndry+SwapUsed)
> 
> 
> Console:
> 
> swap_pager: out of swap space
> swp_pager_getswapspace(4): failed
> swp_pager_getswapspace(1): failed
> swp_pager_getswapspace(1): failed
> swp_pager_getswapspace(2): failed
> swp_pager_getswapspace(2): failed
> swp_pager_getswapspace(4): failed
> swp_pager_getswapspace(1): failed
> swp_pager_getswapspace(9): failed
> swp_pager_getswapspace(4): failed
> swp_pager_getswapspace(7): failed
> swp_pager_getswapspace(29): failed
> swp_pager_getswapspace(9): failed
> swp_pager_getswapspace(1): failed
> swp_pager_getswapspace(2): failed
> swp_pager_getswapspace(1): failed
> swp_pager_getswapspace(4): failed
> swp_pager_getswapspace(1): failed
> swp_pager_getswapspace(10): failed
> 
> . . . Then some time with no messages . . .
> 
> vm_pageout_mightbe_oom: kill context: v_free_count: 7740, v_inactive_count: 1
> Jan 27 23:01:07 CA72_UFS kernel: pid 57238 (c++), jid 3, uid 0, was killed: failed to reclaim memory
> swp_pager_getswapspace(2): failed
> 
> 
> Note: The "vm_pageout_mightbe_oom: kill context:"
> notice is one of the few parts of an old reporting
> patch Mark J. had supplied (long ago) that still
> fits in the modern code (or that I was able to keep
> updated enough to fit, anyway). It is another of the
> personal updates that I keep in my source trees,
> such as in /usr/main-src/ .
> 
> diff --git a/sys/vm/vm_pageout.c b/sys/vm/vm_pageout.c
> index 36d5f3275800..f345e2d4a2d4 100644
> --- a/sys/vm/vm_pageout.c
> +++ b/sys/vm/vm_pageout.c
> @@ -1828,6 +1828,8 @@ vm_pageout_mightbe_oom(struct vm_domain *vmd, int page_shortage,
>         * start OOM.  Initiate the selection and signaling of the
>         * victim.
>         */
> +       printf("vm_pageout_mightbe_oom: kill context: v_free_count: %u, v_inactive_count: %u\n",
> +          vmd->vmd_free_count, vmd->vmd_pagequeues[PQ_INACTIVE].pq_cnt);
>        vm_pageout_oom(VM_OOM_MEM);
> 
>        /*
> 
> 
> Again, I'd used vm.pfault_oom_attempts inappropriately
> for running out of swap (although with UFS it did do
> a kill fairly soon):
> 
> # Delay when persistent low free RAM leads to
> # Out Of Memory killing of processes:
> vm.pageout_oom_seq=120
> #
> # For plunty of swap/paging space (will not
> # run out), avoid pageout delays leading to
> # Out Of Memory killing of processes:
> vm.pfault_oom_attempts=-1
> #
> # For possibly insufficient swap/paging space
> # (might run out), increase the pageout delay
> # that leads to Out Of Memory killing of
> # processes (showing defaults at the time):
> #vm.pfault_oom_attempts= 3
> #vm.pfault_oom_wait= 10
> # (The multiplication is the total but there
> # are other potential tradoffs in the factors
> # multiplied, even for nearly the same total.)
> 
> I'll change:
> 
> vm.pfault_oom_attempts
> vm.pfault_oom_wait
> 
> and reboot --and start the bulk somewhat before
> going to bed.
> 
> 
> For reference:
> 
> [00:02:13] [01] [00:00:00] Building devel/llvm13 | llvm13-13.0.0_3
> [07:37:05] [01] [07:34:52] Finished devel/llvm13 | llvm13-13.0.0_3: Failed: build
> 
> 
> [ 65% 4728/7265] . . . flang/lib/Evaluate/fold-designator.cpp
> [ 65% 4729/7265] . . . flang/lib/Evaluate/fold-integer.cpp
> FAILED: tools/flang/lib/Evaluate/CMakeFiles/obj.FortranEvaluate.dir/fold-integer.cpp.o 
> [ 65% 4729/7265] . . . flang/lib/Evaluate/fold-logical.cpp
> [ 65% 4729/7265] . . . flang/lib/Evaluate/fold-complex.cpp
> [ 65% 4729/7265] . . . flang/lib/Evaluate/fold-real.cpp
> 
> So the flang/lib/Evaluate/fold-integer.cpp one was the one killed.
> 
> Notably, the specific sources being compiled are different
> than in the ZFS context report. But this might be because
> of my killing ninja explicitly in the ZFS context, before
> killing the running compilers.
> 
> Again, using the options to avoid building the Fortran
> compiler probably avoids such memory use --if you do not
> need the Fortran compiler.
> 

Notes on the 2 types of "out of swap space" notices
in the main [so: 14] code:

One place that can lead to a variation of such notices is:

void
vm_pageout_oom(int shortage)
{
. . .
        if (bigproc != NULL) {
                switch (shortage) {
                case VM_OOM_MEM:
                        reason = "failed to reclaim memory";
                        break;
                case VM_OOM_MEM_PF:
                        reason = "a thread waited too long to allocate a page";
                        break;
                case VM_OOM_SWAPZ:
                        reason = "out of swap space";
                        break;
                default:
                        panic("unknown OOM reason %d", shortage);
                }
                if (vm_panic_on_oom != 0 && --vm_panic_on_oom == 0)
                        panic("%s", reason);
                PROC_LOCK(bigproc);
                killproc(bigproc, reason);
. . .
}

It is the VM_OOM_SWAPZ path that produces the misnomer
text during killproc(bigproc, reason) .

But there is another place that produces an accurate message
that does not, of itself, mention kills:

static void
swp_sizecheck(void)
{
  
        if (swap_pager_avail < nswap_lowat) {
                if (swap_pager_almost_full == 0) {
                        printf("swap_pager: out of swap space\n");
                        swap_pager_almost_full = 1;
                }
        } else {
                swap_pager_full = 0;
                if (swap_pager_avail > nswap_hiwat)
                        swap_pager_almost_full = 0;
        }
}

It looks like you were getting the 2nd (accurate) form
of message generation. At this point kills have not
started and the message is not about kills at all.

Thus my initial notes about misnomer messages for your
specific context were wrong. Sorry.

===
Mark Millard
marklmi at yahoo.com