Re: [PATCH] tools/libxl: Fixes to stream v2 task joining logic

[Andrew Cooper <andrew.cooper3@citrix.com>]

On 23/07/15 13:03, Wei Liu wrote:
> On Thu, Jul 23, 2015 at 12:42:25PM +0100, Andrew Cooper wrote:
>> On 23/07/15 12:38, Wei Liu wrote:
>>> On Thu, Jul 23, 2015 at 12:09:38PM +0100, Andrew Cooper wrote:
>>>> During review of the libxl migration v2 series, I changes the task
>>>> joining logic, but clearly didn't think the result through
>>>> properly. This patch fixes several errors.
>>>>
>>>> 1) Do not call check_all_finished() in the success cases of
>>>> libxl__xc_domain_{save,restore}_done().  It serves no specific purpose
>>>> as the save helper state will report itself as inactive by this point,
>>>> and avoids triggering a second stream->completion_callback() in the case
>>>> that write_toolstack_record()/stream_continue() record errors
>>>> synchronously themselves.
>>>>
>>>> 2) Only ever set stream->rc in stream_complete().  The first version of
>>>> the migration v2 series had separate rc and joined_rc parameters, where
>>>> this logic worked.  However when combining the two, the teardown path
>>>> fails to trigger if stream_done() records stream->rc itself.  A side
>>>> effect of this is that stream_done() needs to take an rc parameter.
>>>>
>>>> 3) Avoid stacking of check_all_finished() via synchronous teardown of
>>>> tasks.  If the _abort() functions call back synchronously,
>>>> stream->completion_callback() ends up getting called twice, as first and
>>>> last check_all_finished() frames observe each task being finished.
>>>>
>>>> Signed-off-by: Andrew Cooper <andrew.cooper3@citrix.com>
>>>> CC: Ian Campbell <Ian.Campbell@citrix.com>
>>>> CC: Ian Jackson <Ian.Jackson@eu.citrix.com>
>>>> CC: Wei Liu <wei.liu2@citrix.com>
>>>>
>>>> ---
>>>> I found this while working to fix the toolstack record issue, but am
>>>> posting this bugfix ahead of the other work as OSSTest has tripped over
>>>> the issue.
>>> This change itself doesn't seem to have anything to do with libxc. In
>>> OSSTest the error that triggers this knock on effect is the failure of
>>> xc_map_foreign_bulk. Does that mean this patch only fix half of the
>>> problem seen in OSSTest?
>> Saving to image new xl format (info 0x3/0x0/1797)
>> xc: info: In experimental xc_domain_save2
>> xc: info: Saving domain 2, type x86 HVM
>> xc: progress: Memory: 67584/1344513    5%
>> xc: progress: Memory: 135168/1344513   10%
>> xc: progress: Memory: 201728/1344513   15%
>> xc: progress: Memory: 269312/1344513   20%
>> xc: progress: Memory: 336896/1344513   25%
>> xc: progress: Memory: 403456/1344513   30%
>> xc: progress: Memory: 471040/1344513   35%
>> xc: progress: Memory: 538624/1344513   40%
>> xc: progress: Memory: 605184/1344513   45%
>> xc: progress: Memory: 672768/1344513   50%
>> xc: progress: Memory: 740352/1344513   55%
>> xc: error: xc_map_foreign_bulk: mmap failed (12 = Cannot allocate
>> memory): Internal error
>> xc: error: Failed to map guest pages (12 = Cannot allocate memory):
>> Internal error
>> xc: error: Save failed (12 = Cannot allocate memory): Internal error
>> libxl: error: libxl_stream_write.c:267:libxl__xc_domain_save_done:
>> saving domain: domain responded to suspend request: Cannot allocate memory
>> xl: libxl_event.c:1866: libxl__ao_inprogress_gc: Assertion
>> `!ao->complete' failed.
>>
>>
>> I don't know why xc_map_foreign_bulk() failed, but at a guess I would
>> day dom0 is out of RAM.
>>
> I don't think so. There is no OOM error in serial log and dmesg.
>
> A similar failure is discovered in QEMU mainline test. That one is using
> OVMF + QEMU upstream. So this failure is not related to firmware
> or device model.
>
> http://logs.test-lab.xenproject.org/osstest/logs/59819/test-amd64-i386-xl-qemuu-debianhvm-amd64-xsm/info.html
>
> It should be a real bug manifests only on 32bit toolstack + guest with
> large amount of ram (5000 in this case).
>
> Looks like we have one more bug to hunt down.

>From linux_privcmd_map_foreign_bulk()

addr = mmap(NULL, (unsigned long)num << XC_PAGE_SHIFT, prot, MAP_SHARED,
            fd, 0);
if ( addr == MAP_FAILED )
{
    PERROR("xc_map_foreign_bulk: mmap failed");
    return NULL;
}

So the mmap() call did genuinely fail with ENOMEM.

Migration v2 only ever has 1024 guest pages ever mapped at once (see
xc_sr_save.c:write_batch() ), and doesn't leak the mapping, which means
that the kernel decided that it was out of memory.

~Andrew