Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
From:   <yf.wang@mediatek.com>
To:     <robin.murphy@arm.com>
CC:     <Libo.Kang@mediatek.com>, <Ning.Li@mediatek.com>,
        <iommu@lists.linux.dev>, <jianjiao.zeng@mediatek.com>,
        <joro@8bytes.org>, <linux-arm-kernel@lists.infradead.org>,
        <linux-kernel@vger.kernel.org>,
        <linux-mediatek@lists.infradead.org>, <matthias.bgg@gmail.com>,
        <will@kernel.org>, <wsd_upstream@mediatek.com>,
        <yf.wang@mediatek.com>, <yong.wu@mediatek.com>
Subject: RE: [PATCH] iommu/iova: Fix alloc iova overflows issue
Date:   Wed, 11 Jan 2023 13:57:49 +0800
Message-ID: <20230111055749.11519-1-yf.wang@mediatek.com>
In-Reply-To: <99695d77-e54f-3270-ee9a-9b3cbb3296e5@arm.com>
References: <99695d77-e54f-3270-ee9a-9b3cbb3296e5@arm.com>
MIME-Version: 1.0
Content-Type: text/plain
Precedence: bulk

On Mon, 2023-01-09 at 11:47 +0000, Robin Murphy wrote:
> On 2023-01-09 08:34, yf.wang@mediatek.com wrote:
> > From: Yunfei Wang <yf.wang@mediatek.com>
> > 
> > In __alloc_and_insert_iova_range, there is an issue that retry_pfn
> > overflows. The value of iovad->anchor.pfn_hi is ~0UL, then when
> > iovad->cached_node is iovad->anchor, curr_iova->pfn_hi + 1 will
> > overflow. As a result, if the retry logic is executed, low_pfn is
> > updated to 0, and then new_pfn < low_pfn returns false to make the
> > allocation successful.
> > 
> > This issue occurs in the following two situations:
> > 1. The first iova size exceeds the domain size. When initializing
> > iova domain, iovad->cached_node is assigned as iovad->anchor. For
> > example, the iova domain size is 10M, start_pfn is 0x1_F000_0000,
> > and the iova size allocated for the first time is 11M. The
> > following is the log information, new->pfn_lo is smaller than
> > iovad->cached_node.
> > 
> > Example log:
> > [  223.798112][T1705487] sh:
> > [name:iova&]__alloc_and_insert_iova_range
> > start_pfn:0x1f0000,retry_pfn:0x0,size:0xb00,limit_pfn:0x1f0a00
> > [  223.799590][T1705487] sh:
> > [name:iova&]__alloc_and_insert_iova_range
> > success start_pfn:0x1f0000,new->pfn_lo:0x1efe00,new-
> > >pfn_hi:0x1f08ff
> > 
> > 2. The node with the largest iova->pfn_lo value in the iova domain
> > is deleted, iovad->cached_node will be updated to iovad->anchor,
> > and then the alloc iova size exceeds the maximum iova size that can
> > be allocated in the domain.
> > 
> > Adding judgment that retry_pfn must be greater than iovad-
> > >start_pfn
> > can fix this issue.
> 
> Hmm, indeed that's a sneaky little bug - thanks for the thorough 
> analysis - but couldn't we avoid the overflow entirely? I don't have
> the 
> complete logic paged in, but superficially it seems like:
> 
> 	retry_pfn = curr_iova->pfn_hi;
> 	...
> 		retry_pfn <= limit_pfn
> 	...
> 		low_pfn = retry_pfn + 1;
> 
> should still work, shouldn't it?
> 
> Thanks,
> Robin.
> 

Hi Robin,
Thanks for your suggestion, your solution is more perfect,
PATCH v2 version will modify it.

Thanks,
Yunfei.

> > Signed-off-by: jianjiao zeng <jianjiao.zeng@mediatek.com>
> > Signed-off-by: Yunfei Wang <yf.wang@mediatek.com>
> > ---
> >   drivers/iommu/iova.c | 3 ++-
> >   1 file changed, 2 insertions(+), 1 deletion(-)
> > 
> > diff --git a/drivers/iommu/iova.c b/drivers/iommu/iova.c
> > index a44ad92fc5eb..0073206c2a95 100644
> > --- a/drivers/iommu/iova.c
> > +++ b/drivers/iommu/iova.c
> > @@ -209,7 +209,8 @@ static int __alloc_and_insert_iova_range(struct
> > iova_domain *iovad,
> >   	} while (curr && new_pfn <= curr_iova->pfn_hi && new_pfn >=
> > low_pfn);
> >   
> >   	if (high_pfn < size || new_pfn < low_pfn) {
> > -		if (low_pfn == iovad->start_pfn && retry_pfn <
> > limit_pfn) {
> > +		if (low_pfn == iovad->start_pfn &&
> > +		    retry_pfn >= iovad->start_pfn && retry_pfn <
> > limit_pfn) {
> >   			high_pfn = limit_pfn;
> >   			low_pfn = retry_pfn;
> >   			curr = iova_find_limit(iovad, limit_pfn);