[PATCH] hv: hv_balloon: avoid memory leak on alloc_error of 2MB memory block

[Dexuan Cui]

> -----Original Message-----
> From: Jason Wang [mailto:jasowang at redhat.com]
> Sent: Monday, November 24, 2014 16:48 PM
> To: Dexuan Cui; gregkh at linuxfoundation.org; linux-kernel at vger.kernel.org;
> driverdev-devel at linuxdriverproject.org; olaf at aepfle.de;
> apw at canonical.com; KY Srinivasan
> Cc: Haiyang Zhang
> Subject: Re: [PATCH] hv: hv_balloon: avoid memory leak on alloc_error of
> 2MB memory block
> 
> On 11/24/2014 03:54 PM, Dexuan Cui wrote:
> >> -----Original Message-----
> >> From: Jason Wang [mailto:jasowang at redhat.com]
> >> Sent: Monday, November 24, 2014 15:28 PM
> >> To: Dexuan Cui; gregkh at linuxfoundation.org; linux-
> kernel at vger.kernel.org;
> >> driverdev-devel at linuxdriverproject.org; olaf at aepfle.de;
> >> apw at canonical.com; KY Srinivasan
> >> Cc: Haiyang Zhang
> >> Subject: Re: [PATCH] hv: hv_balloon: avoid memory leak on alloc_error
> of
> >> 2MB memory block
> >>
> >> On 11/24/2014 02:08 PM, Dexuan Cui wrote:
> >>>> -----Original Message-----
> >>>>> From: Jason Wang [mailto:jasowang at redhat.com]
> >>>>> Sent: Monday, November 24, 2014 13:18 PM
> >>>>> To: Dexuan Cui; gregkh at linuxfoundation.org; linux-
> >> kernel at vger.kernel.org;
> >>>>> driverdev-devel at linuxdriverproject.org; olaf at aepfle.de;
> >>>>> apw at canonical.com; KY Srinivasan
> >>>>> Cc: Haiyang Zhang
> >>>>> Subject: Re: [PATCH] hv: hv_balloon: avoid memory leak on
> >> alloc_error of
> >>>>> 2MB memory block
> >>>>>
> >>>>> On 11/24/2014 01:56 PM, Dexuan Cui wrote:
> >>>>>>> If num_ballooned is not 0, we shouldn't neglect the already-
> >> allocated
> >>>>> 2MB
> >>>>>>> memory block(s).
> >>>>>>>
> >>>>>>> Cc: K. Y. Srinivasan <kys at microsoft.com>
> >>>>>>> Cc: <stable at vger.kernel.org>
> >>>>>>> Signed-off-by: Dexuan Cui <decui at microsoft.com>
> >>>>>>> ---
> >>>>>>>  drivers/hv/hv_balloon.c | 4 +++-
> >>>>>>>  1 file changed, 3 insertions(+), 1 deletion(-)
> >>>>>>>
> >>>>>>> diff --git a/drivers/hv/hv_balloon.c b/drivers/hv/hv_balloon.c
> >>>>>>> index 5e90c5d..cba2d3b 100644
> >>>>>>> --- a/drivers/hv/hv_balloon.c
> >>>>>>> +++ b/drivers/hv/hv_balloon.c
> >>>>>>> @@ -1091,6 +1091,8 @@ static void balloon_up(struct
> >> work_struct
> >>>>> *dummy)
> >>>>>>>  	bool done = false;
> >>>>>>>  	int i;
> >>>>>>>
> >>>>>>> +	/* The host does balloon_up in 2MB. */
> >>>>>>> +	WARN_ON(num_pages % PAGES_IN_2M != 0);
> >>>>>>>
> >>>>>>>  	/*
> >>>>>>>  	 * We will attempt 2M allocations. However, if we fail to
> >>>>>>> @@ -1111,7 +1113,7 @@ static void balloon_up(struct
> >> work_struct
> >>>>> *dummy)
> >>>>>>>  						bl_resp, alloc_unit,
> >>>>>>>  						 &alloc_error);
> >>>>>>>
> >>>>>>> -		if ((alloc_error) && (alloc_unit != 1)) {
> >>>>>>> +		if (alloc_error && (alloc_unit != 1) &&
> >> num_ballooned == 0)
> >>>>> {
> >>>>>>>  			alloc_unit = 1;
> >>>>>>>  			continue;
> >>>>>>>  		}
> >>>>> Before the change, we may retry the 4K allocation when part or all
> 2M
> >>>>> allocations were failed. This makes sense when memory is
> fragmented.
> >> But
> >>> Yes, but all the partially-allocated 2MB memory blocks are lost(mem
> leak).
> >>>
> >>>>> after the change, if part of 2M allocation were failed, we won't retry
> >>>>> 4K allocation. Is this expected?
> >>> Hi Jason,
> >>> The patch doesn't break the "try 2MB first; then try 4K" logic:
> >>>
> >>> With the change, we'll retry the 2MB allocation in the next iteration of
> the
> >>> same while (!done) loop -- we expect this retry will cause
> >>> "alloc_error && (alloc_unit != 1) && num_ballooned == 0" to be true,
> >>> so we'll later try 4K allocation, as we did before.
> >> If I read the code correctly, if part of 2M allocation fails.
> >> alloc_balloon_pages() will have a non zero return value with alloc_error
> >> set. Then it will match the following check:
> >>
> >>                 if ((alloc_error) || (num_ballooned == num_pages))
> >> {
> >>
> >> which will set done to true. So there's no second iteration of while
> >> (!done) loop?
> > Oh... I see the issue in my patch.
> > Thanks for pointing this out, Jason!
> >
> >> Probably you need something like:
> >>
> >> if ((alloc_unit != 1) && (num_ballooned == 0)) {
> >>     alloc_unit = 1;
> >>     continue;
> >> }
> >>
> >> if ((alloc_unit == 1) || (num_ballooned == num_pages)) {
> >>     ...
> >> }
> > Your code is good, except for one thing:
> > alloc_balloon_pages() can return due to:
> >
> > if (bl_resp->hdr.size + sizeof(union dm_mem_page_range) >
> >                          PAGE_SIZE)
> >                         return i * alloc_unit;
> >
> > In this case, "alloc_unit == 1" is true, but we shouldn't run "done = true".
> >
> > How do you like this? I made a few changes based on your code.
> >
> > @@ -1086,16 +1088,18 @@ static void balloon_up(struct work_struct
> *dummy)
> >                 num_pages -= num_ballooned;
> > +               alloc_error = false;
> >                 num_ballooned = alloc_balloon_pages(&dm_device, num_pages,
> >                                                 bl_resp, alloc_unit,
> >                                                  &alloc_error);
> >
> > -               if ((alloc_error) && (alloc_unit != 1)) {
> > +               if (alloc_unit != 1 && num_ballooned == 0) {
> >                         alloc_unit = 1;
> >                         continue;
> >                 }
> >
> > -               if ((alloc_error) || (num_ballooned == num_pages)) {
> > +               if ((alloc_unit == 1 && alloc_error) ||
> > +                       (num_ballooned == num_pages)) {
> >                         bl_resp->more_pages = 0;
> >                         done = true;
> >                         dm_device.state = DM_INITIALIZED;
> >
> >
> > If you're Ok with this, I'll send out a v2 patch.
> > -- Dexuan
> 
> Looks good.
Thanks, Jason!

Let's wait for KY's clarification about if the host can balloon up 1MB.

I'll send the v2 thereafter.

-- Dexuan