Re: [PATCH] iommu/dma: limit iova free size to unmmaped iova

Re: [PATCH] iommu/dma: limit iova free size to unmmaped iova

[guptap]

On 2020-05-22 01:46, Robin Murphy wrote:
> On 2020-05-21 12:30, Prakash Gupta wrote:
>> Limit the iova size while freeing based on unmapped size. In absence 
>> of
>> this even with unmap failure, invalid iova is pushed to iova rcache 
>> and
>> subsequently can cause panic while rcache magazine is freed.
> 
> Can you elaborate on that panic?
> 
We have seen couple of stability issues around this.
Below is one such example:

kernel BUG at kernel/msm-4.19/drivers/iommu/iova.c:904!
iova_magazine_free_pfns
iova_rcache_insert
free_iova_fast
__iommu_unmap_page
iommu_dma_unmap_page


It turned out an iova pfn 0 got into iova_rcache. One possibility I see 
is
where client unmap with invalid dma_addr. The unmap call will fail and 
warn on
and still try to free iova. This will cause invalid pfn to be inserted 
into
rcache. As and when the magazine with invalid pfn will be freed
private_find_iova() will return NULL for invalid iova and meet bug 
condition.

>> Signed-off-by: Prakash Gupta <guptap@codeaurora.org>
>> 
>> :100644 100644 4959f5df21bd 098f7d377e04 M	drivers/iommu/dma-iommu.c
>> 
>> diff --git a/drivers/iommu/dma-iommu.c b/drivers/iommu/dma-iommu.c
>> index 4959f5df21bd..098f7d377e04 100644
>> --- a/drivers/iommu/dma-iommu.c
>> +++ b/drivers/iommu/dma-iommu.c
>> @@ -472,7 +472,8 @@ static void __iommu_dma_unmap(struct device *dev, 
>> dma_addr_t dma_addr,
>>     	if (!cookie->fq_domain)
>>   		iommu_tlb_sync(domain, &iotlb_gather);
>> -	iommu_dma_free_iova(cookie, dma_addr, size);
>> +	if (unmapped)
>> +		iommu_dma_free_iova(cookie, dma_addr, unmapped);
> 
> Frankly, if any part of the unmap fails then things have gone
> catastrophically wrong already, but either way this isn't right. The
> IOVA API doesn't support partial freeing - an IOVA *must* be freed
> with its original size, or not freed at all, otherwise it will corrupt
> the state of the rcaches and risk a cascade of further misbehaviour
> for future callers.
> 
I agree, we shouldn't be freeing the partial iova. Instead just making
sure if unmap was successful should be sufficient before freeing iova. 
So change
can instead be something like this:

-	iommu_dma_free_iova(cookie, dma_addr, size);
+	if (unmapped)
+		iommu_dma_free_iova(cookie, dma_addr, size);

> TBH my gut feeling here is that you're really just trying to treat a
> symptom of another bug elsewhere, namely some driver calling
> dma_unmap_* or dma_free_* with the wrong address or size in the first
> place.
> 
This condition would arise only if driver calling dma_unmap/free_* with 
0
iova_pfn. This will be flagged with a warning during unmap but will 
trigger
panic later on while doing unrelated dma_map/unmap_*. If unmapped has 
already
failed for invalid iova, there is no reason we should consider this as 
valid
iova and free. This part should be fixed.

On 2020-05-22 00:19, Andrew Morton wrote:
> I think we need a cc:stable here?
> 
Added now.

Thanks,
Prakash

Re: [PATCH] iommu/dma: limit iova free size to unmmaped iova

[Robin Murphy]

On 2020-05-22 07:25, guptap@codeaurora.org wrote:
> On 2020-05-22 01:46, Robin Murphy wrote:
>> On 2020-05-21 12:30, Prakash Gupta wrote:
>>> Limit the iova size while freeing based on unmapped size. In absence of
>>> this even with unmap failure, invalid iova is pushed to iova rcache and
>>> subsequently can cause panic while rcache magazine is freed.
>>
>> Can you elaborate on that panic?
>>
> We have seen couple of stability issues around this.
> Below is one such example:
> 
> kernel BUG at kernel/msm-4.19/drivers/iommu/iova.c:904!
> iova_magazine_free_pfns
> iova_rcache_insert
> free_iova_fast
> __iommu_unmap_page
> iommu_dma_unmap_page

OK, so it's not some NULL dereference or anything completely unexpected, 
that's good.

> It turned out an iova pfn 0 got into iova_rcache. One possibility I see is
> where client unmap with invalid dma_addr. The unmap call will fail and 
> warn on
> and still try to free iova. This will cause invalid pfn to be inserted into
> rcache. As and when the magazine with invalid pfn will be freed
> private_find_iova() will return NULL for invalid iova and meet bug 
> condition.

That would indeed be a bug in whatever driver made the offending 
dma_unmap call.

>>> Signed-off-by: Prakash Gupta <guptap@codeaurora.org>
>>>
>>> :100644 100644 4959f5df21bd 098f7d377e04 M    drivers/iommu/dma-iommu.c
>>>
>>> diff --git a/drivers/iommu/dma-iommu.c b/drivers/iommu/dma-iommu.c
>>> index 4959f5df21bd..098f7d377e04 100644
>>> --- a/drivers/iommu/dma-iommu.c
>>> +++ b/drivers/iommu/dma-iommu.c
>>> @@ -472,7 +472,8 @@ static void __iommu_dma_unmap(struct device *dev, 
>>> dma_addr_t dma_addr,
>>>         if (!cookie->fq_domain)
>>>           iommu_tlb_sync(domain, &iotlb_gather);
>>> -    iommu_dma_free_iova(cookie, dma_addr, size);
>>> +    if (unmapped)
>>> +        iommu_dma_free_iova(cookie, dma_addr, unmapped);
>>
>> Frankly, if any part of the unmap fails then things have gone
>> catastrophically wrong already, but either way this isn't right. The
>> IOVA API doesn't support partial freeing - an IOVA *must* be freed
>> with its original size, or not freed at all, otherwise it will corrupt
>> the state of the rcaches and risk a cascade of further misbehaviour
>> for future callers.
>>
> I agree, we shouldn't be freeing the partial iova. Instead just making
> sure if unmap was successful should be sufficient before freeing iova. 
> So change
> can instead be something like this:
> 
> -    iommu_dma_free_iova(cookie, dma_addr, size);
> +    if (unmapped)
> +        iommu_dma_free_iova(cookie, dma_addr, size);
> 
>> TBH my gut feeling here is that you're really just trying to treat a
>> symptom of another bug elsewhere, namely some driver calling
>> dma_unmap_* or dma_free_* with the wrong address or size in the first
>> place.
>>
> This condition would arise only if driver calling dma_unmap/free_* with 0
> iova_pfn. This will be flagged with a warning during unmap but will trigger
> panic later on while doing unrelated dma_map/unmap_*. If unmapped has 
> already
> failed for invalid iova, there is no reason we should consider this as 
> valid
> iova and free. This part should be fixed.

I disagree. In general, if drivers call the DMA API incorrectly it is 
liable to lead to data loss, memory corruption, and various other 
unpleasant misbehaviour - it is not the DMA layer's job to attempt to 
paper over driver bugs.

There *is* an argument for downgrading the BUG_ON() in 
iova_magazine_free_pfns() to a WARN_ON(), since frankly it isn't a 
sufficiently serious condition to justify killing the whole machine 
immediately, but NAK to bodging the iommu-dma mid-layer to "fix" that. A 
serious bug already happened elsewhere, so trying to hide the fallout 
really doesn't help anyone.

Robin.

> On 2020-05-22 00:19, Andrew Morton wrote:
>> I think we need a cc:stable here?
>>
> Added now.
> 
> Thanks,
> Prakash

Re: [PATCH] iommu/dma: limit iova free size to unmmaped iova

[guptap]

On 2020-05-22 14:54, Robin Murphy wrote:
> On 2020-05-22 07:25, guptap@codeaurora.org wrote:
>> On 2020-05-22 01:46, Robin Murphy wrote:
>>> On 2020-05-21 12:30, Prakash Gupta wrote:
>> I agree, we shouldn't be freeing the partial iova. Instead just making
>> sure if unmap was successful should be sufficient before freeing iova. 
>> So change
>> can instead be something like this:
>> 
>> -    iommu_dma_free_iova(cookie, dma_addr, size);
>> +    if (unmapped)
>> +        iommu_dma_free_iova(cookie, dma_addr, size);
>> 
>>> TBH my gut feeling here is that you're really just trying to treat a
>>> symptom of another bug elsewhere, namely some driver calling
>>> dma_unmap_* or dma_free_* with the wrong address or size in the first
>>> place.
>>> 
>> This condition would arise only if driver calling dma_unmap/free_* 
>> with 0
>> iova_pfn. This will be flagged with a warning during unmap but will 
>> trigger
>> panic later on while doing unrelated dma_map/unmap_*. If unmapped has 
>> already
>> failed for invalid iova, there is no reason we should consider this as 
>> valid
>> iova and free. This part should be fixed.
> 
> I disagree. In general, if drivers call the DMA API incorrectly it is
> liable to lead to data loss, memory corruption, and various other
> unpleasant misbehaviour - it is not the DMA layer's job to attempt to
> paper over driver bugs.
> 
> There *is* an argument for downgrading the BUG_ON() in
> iova_magazine_free_pfns() to a WARN_ON(), since frankly it isn't a
> sufficiently serious condition to justify killing the whole machine
> immediately, but NAK to bodging the iommu-dma mid-layer to "fix" that.
> A serious bug already happened elsewhere, so trying to hide the
> fallout really doesn't help anyone.
> 
Sorry for delayed response, it was a long weekend.
I agree that invalid DMA API call can result in unexpected issues and 
client
should fix it, but then the present behavior makes it difficult to catch 
cases
when driver is making wrong DMA API calls. When invalid iova pfn is 
passed it
doesn't fail then and there, though DMA layer is aware of iova being 
invalid. It
fails much after that in the context of an valid map/unmap, with 
BUG_ON().

Downgrading BUG_ON() to WARN_ON() in iova_magazine_free_pfns() will not 
help
much as invalid iova will cause NULL pointer dereference.

I see no reason why DMA layer wants to free an iova for which unmapped 
failed.
IMHO queuing an invalid iova (which already failed unmap) to rcache 
which
eventually going to crash the system looks like iommu-dma layer issue.

Thanks,
Prakash

Re: [PATCH] iommu/dma: limit iova free size to unmmaped iova

[guptap]

On 2020-05-26 12:49, guptap@codeaurora.org wrote:
> On 2020-05-22 14:54, Robin Murphy wrote:
>> On 2020-05-22 07:25, guptap@codeaurora.org wrote:
>>> On 2020-05-22 01:46, Robin Murphy wrote:
>>>> On 2020-05-21 12:30, Prakash Gupta wrote:
>>> I agree, we shouldn't be freeing the partial iova. Instead just 
>>> making
>>> sure if unmap was successful should be sufficient before freeing 
>>> iova. So change
>>> can instead be something like this:
>>> 
>>> -    iommu_dma_free_iova(cookie, dma_addr, size);
>>> +    if (unmapped)
>>> +        iommu_dma_free_iova(cookie, dma_addr, size);
>>> 
>>>> TBH my gut feeling here is that you're really just trying to treat a
>>>> symptom of another bug elsewhere, namely some driver calling
>>>> dma_unmap_* or dma_free_* with the wrong address or size in the 
>>>> first
>>>> place.
>>>> 
>>> This condition would arise only if driver calling dma_unmap/free_* 
>>> with 0
>>> iova_pfn. This will be flagged with a warning during unmap but will 
>>> trigger
>>> panic later on while doing unrelated dma_map/unmap_*. If unmapped has 
>>> already
>>> failed for invalid iova, there is no reason we should consider this 
>>> as valid
>>> iova and free. This part should be fixed.
>> 
>> I disagree. In general, if drivers call the DMA API incorrectly it is
>> liable to lead to data loss, memory corruption, and various other
>> unpleasant misbehaviour - it is not the DMA layer's job to attempt to
>> paper over driver bugs.
>> 
>> There *is* an argument for downgrading the BUG_ON() in
>> iova_magazine_free_pfns() to a WARN_ON(), since frankly it isn't a
>> sufficiently serious condition to justify killing the whole machine
>> immediately, but NAK to bodging the iommu-dma mid-layer to "fix" that.
>> A serious bug already happened elsewhere, so trying to hide the
>> fallout really doesn't help anyone.
>> 
> Sorry for delayed response, it was a long weekend.
> I agree that invalid DMA API call can result in unexpected issues and 
> client
> should fix it, but then the present behavior makes it difficult to 
> catch cases
> when driver is making wrong DMA API calls. When invalid iova pfn is 
> passed it
> doesn't fail then and there, though DMA layer is aware of iova being 
> invalid. It
> fails much after that in the context of an valid map/unmap, with 
> BUG_ON().
> 
> Downgrading BUG_ON() to WARN_ON() in iova_magazine_free_pfns() will not 
> help
> much as invalid iova will cause NULL pointer dereference.
> 
> I see no reason why DMA layer wants to free an iova for which unmapped 
> failed.
> IMHO queuing an invalid iova (which already failed unmap) to rcache 
> which
> eventually going to crash the system looks like iommu-dma layer issue.
> 
> Thanks,
> Prakash

ping?

Re: [PATCH] iommu/dma: limit iova free size to unmmaped iova

[Robin Murphy]

On 2020-05-26 08:19, guptap@codeaurora.org wrote:
> On 2020-05-22 14:54, Robin Murphy wrote:
>> On 2020-05-22 07:25, guptap@codeaurora.org wrote:
>>> On 2020-05-22 01:46, Robin Murphy wrote:
>>>> On 2020-05-21 12:30, Prakash Gupta wrote:
>>> I agree, we shouldn't be freeing the partial iova. Instead just making
>>> sure if unmap was successful should be sufficient before freeing 
>>> iova. So change
>>> can instead be something like this:
>>>
>>> -    iommu_dma_free_iova(cookie, dma_addr, size);
>>> +    if (unmapped)
>>> +        iommu_dma_free_iova(cookie, dma_addr, size);
>>>
>>>> TBH my gut feeling here is that you're really just trying to treat a
>>>> symptom of another bug elsewhere, namely some driver calling
>>>> dma_unmap_* or dma_free_* with the wrong address or size in the first
>>>> place.
>>>>
>>> This condition would arise only if driver calling dma_unmap/free_* 
>>> with 0
>>> iova_pfn. This will be flagged with a warning during unmap but will 
>>> trigger
>>> panic later on while doing unrelated dma_map/unmap_*. If unmapped has 
>>> already
>>> failed for invalid iova, there is no reason we should consider this 
>>> as valid
>>> iova and free. This part should be fixed.
>>
>> I disagree. In general, if drivers call the DMA API incorrectly it is
>> liable to lead to data loss, memory corruption, and various other
>> unpleasant misbehaviour - it is not the DMA layer's job to attempt to
>> paper over driver bugs.
>>
>> There *is* an argument for downgrading the BUG_ON() in
>> iova_magazine_free_pfns() to a WARN_ON(), since frankly it isn't a
>> sufficiently serious condition to justify killing the whole machine
>> immediately, but NAK to bodging the iommu-dma mid-layer to "fix" that.
>> A serious bug already happened elsewhere, so trying to hide the
>> fallout really doesn't help anyone.
>>
> Sorry for delayed response, it was a long weekend.
> I agree that invalid DMA API call can result in unexpected issues and 
> client
> should fix it, but then the present behavior makes it difficult to catch 
> cases
> when driver is making wrong DMA API calls. When invalid iova pfn is 
> passed it
> doesn't fail then and there, though DMA layer is aware of iova being 
> invalid. It
> fails much after that in the context of an valid map/unmap, with BUG_ON().
> 
> Downgrading BUG_ON() to WARN_ON() in iova_magazine_free_pfns() will not 
> help
> much as invalid iova will cause NULL pointer dereference.

Obviously I didn't mean a literal s/BUG/WARN/ substitution - some 
additional control flow to actually handle the error case was implied.

I'll write up the patch myself, since it's easier than further debating.

> I see no reason why DMA layer wants to free an iova for which unmapped 
> failed.
> IMHO queuing an invalid iova (which already failed unmap) to rcache which
> eventually going to crash the system looks like iommu-dma layer issue.

What if the unmap fails because the address range is already entirely 
unmapped? Freeing the IOVA (or at least attempting to) would be 
logically appropriate in that case. In fact some IOMMU drivers might not 
even consider that a failure, so the DMA layer may not even be aware 
that it's been handed a bogus unallocated address.

The point is that unmapping *doesn't* fail under normal and correct 
operation, so the DMA layer should not expect to have to handle it. Even 
if it does happen, that's a highly exceptional case that the DMA layer 
cannot recover from by itself; at best it can just push the problem 
elsewhere. It's pretty hard to justify doing extra work to simply move 
an exceptional problem around without really addressing it.

And in this particular case, personally I would *much* rather see 
warnings spewing from both the pagetable and IOVA code as early as 
possible to clearly indicate that the DMA layer itself has been thrown 
out of sync, than just have warnings that might represent some other 
source of pagetable corruption (or at worst, depending on the pagetable 
code, no warnings at all and only have dma_map_*() calls quietly start 
failing much, much later due to all the IOVA space having been leaked by 
bad unmaps).

Robin.

Re: [PATCH] iommu/dma: limit iova free size to unmmaped iova

[guptap]

On 2020-06-02 18:37, Robin Murphy wrote:
> On 2020-05-26 08:19, guptap@codeaurora.org wrote:
>> On 2020-05-22 14:54, Robin Murphy wrote:
>>> On 2020-05-22 07:25, guptap@codeaurora.org wrote:
>>>> On 2020-05-22 01:46, Robin Murphy wrote:
>>>>> On 2020-05-21 12:30, Prakash Gupta wrote:
>> Sorry for delayed response, it was a long weekend.
>> I agree that invalid DMA API call can result in unexpected issues and 
>> client
>> should fix it, but then the present behavior makes it difficult to 
>> catch cases
>> when driver is making wrong DMA API calls. When invalid iova pfn is 
>> passed it
>> doesn't fail then and there, though DMA layer is aware of iova being 
>> invalid. It
>> fails much after that in the context of an valid map/unmap, with 
>> BUG_ON().
>> 
>> Downgrading BUG_ON() to WARN_ON() in iova_magazine_free_pfns() will 
>> not help
>> much as invalid iova will cause NULL pointer dereference.
> 
> Obviously I didn't mean a literal s/BUG/WARN/ substitution - some
> additional control flow to actually handle the error case was implied.
> 
> I'll write up the patch myself, since it's easier than further 
> debating.
> 

I think it will address the issue I am facing as well. I will wait for 
your patch.

>> I see no reason why DMA layer wants to free an iova for which unmapped 
>> failed.
>> IMHO queuing an invalid iova (which already failed unmap) to rcache 
>> which
>> eventually going to crash the system looks like iommu-dma layer issue.
> 
> What if the unmap fails because the address range is already entirely
> unmapped? Freeing the IOVA (or at least attempting to) would be
> logically appropriate in that case. In fact some IOMMU drivers might
> not even consider that a failure, so the DMA layer may not even be
> aware that it's been handed a bogus unallocated address.
> 
> The point is that unmapping *doesn't* fail under normal and correct
> operation, so the DMA layer should not expect to have to handle it.
> Even if it does happen, that's a highly exceptional case that the DMA
> layer cannot recover from by itself; at best it can just push the
> problem elsewhere. It's pretty hard to justify doing extra work to
> simply move an exceptional problem around without really addressing
> it.
> 

iommu_unmap() expects that all areas within unmap size are mapped. 
infact It
abandons further unmap if it find any chunk not mapped.  So if an IOMMU
implementation is not returning failure for already unmapped area,
then it's prone to issue where DMA API reuse the IOVA, which is already 
mapped.

In this case where iommu implementation returns error for already 
unmapped area,
currently there is no way to distinguish an unmap failure due to range 
already
unmapped or say while partially unmapping a section map, memory 
allocation fails
and thus unmap returns failure.
In second case, we will free the iova even with mapping present. And 
subsequent
dma mapping will keep on failing if it uses this freed iova. For managed 
iova
both unmap/map should be done with dma APIs, it should be safe to expect 
if a
range is unmapped with DMA APIs corresponding iova is also freed, so 
there
shouldn't be a real need to free iova where unmap fails due to range 
already
entirely unmapped.

> And in this particular case, personally I would *much* rather see
> warnings spewing from both the pagetable and IOVA code as early as
> possible to clearly indicate that the DMA layer itself has been thrown
> out of sync, than just have warnings that might represent some other
> source of pagetable corruption (or at worst, depending on the
> pagetable code, no warnings at all and only have dma_map_*() calls
> quietly start failing much, much later due to all the IOVA space
> having been leaked by bad unmaps).
> 

I am not sure how useful is this freed iova if corresponding mappings 
are not
unmapped. We won't be able to use those iova. The subsequent iommu_map 
will fail
if using this freed iova. So it's important to ensure we only free iova 
which is
unmapped successfully.

Thanks,
Prakash