lkml.org 
[lkml]   [2019]   [Oct]   [9]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH] iommu/mediatek: Move the tlb_sync into tlb_flush
From
Date
On Wed, 2019-10-09 at 16:56 +0900, Tomasz Figa wrote:
> On Tue, Oct 8, 2019 at 5:09 PM Yong Wu <yong.wu@mediatek.com> wrote:
> >
> > Hi Tomasz,
> >
> > Sorry for reply late.
> >
> > On Wed, 2019-10-02 at 14:18 +0900, Tomasz Figa wrote:
> > > Hi Yong,
> > >
> > > On Mon, Sep 30, 2019 at 2:42 PM Yong Wu <yong.wu@mediatek.com> wrote:
> > > >
> > > > The commit 4d689b619445 ("iommu/io-pgtable-arm-v7s: Convert to IOMMU API
> > > > TLB sync") help move the tlb_sync of unmap from v7s into the iommu
> > > > framework. It helps add a new function "mtk_iommu_iotlb_sync", But it
> > > > lacked the dom->pgtlock, then it will cause the variable
> > > > "tlb_flush_active" may be changed unexpectedly, we could see this warning
> > > > log randomly:
> > > >
> > >
> > > Thanks for the patch! Please see my comments inline.
> > >
> > > > mtk-iommu 10205000.iommu: Partial TLB flush timed out, falling back to
> > > > full flush
> > > >
> > > > To fix this issue, we can add dom->pgtlock in the "mtk_iommu_iotlb_sync".
> > > > And when checking this issue, we find that __arm_v7s_unmap call
> > > > io_pgtable_tlb_add_flush consecutively when it is supersection/largepage,
> > > > this also is potential unsafe for us. There is no tlb flush queue in the
> > > > MediaTek M4U HW. The HW always expect the tlb_flush/tlb_sync one by one.
> > > > If v7s don't always gurarantee the sequence, Thus, In this patch I move
> > > > the tlb_sync into tlb_flush(also rename the function deleting "_nosync").
> > > > and we don't care if it is leaf, rearrange the callback functions. Also,
> > > > the tlb flush/sync was already finished in v7s, then iotlb_sync and
> > > > iotlb_sync_all is unnecessary.
> > >
> > > Performance-wise, we could do much better. Instead of synchronously
> > > syncing at the end of mtk_iommu_tlb_add_flush(), we could sync at the
> > > beginning, if there was any previous flush still pending. We would
> > > also have to keep the .iotlb_sync() callback, to take care of waiting
> > > for the last flush. That would allow better pipelining with CPU in
> > > cases like this:
> > >
> > > for (all pages in range) {
> > > change page table();
> > > flush();
> > > }
> > >
> > > "change page table()" could execute while the IOMMU is flushing the
> > > previous change.
> >
> > Do you mean adding a new tlb_sync before tlb_flush_no_sync, like below:
> >
> > mtk_iommu_tlb_add_flush_nosync {
> > + mtk_iommu_tlb_sync();
> > tlb_flush_no_sync();
> > data->tlb_flush_active = true;
> > }
> >
> > mtk_iommu_tlb_sync {
> > if (!data->tlb_flush_active)
> > return;
> > tlb_sync();
> > data->tlb_flush_active = false;
> > }
> >
> > This way look improve the flow, But adjusting the flow is not the root
> > cause of this issue. the problem is "data->tlb_flush_active" may be
> > changed from mtk_iommu_iotlb_sync which don't have a dom->pglock.
>
> That was not the only problem with existing code. Existing code also
> assumed that add_flush and sync always go in pairs, but that's not
> true.

Yes. Thus I put the tlb_flush always followed by tlb_sync to make sure
they always go in pairs.

>
> My suggestion is to fix the locking in the driver and keep the sync
> deferred as much as possible, so that performance is not degraded. I

I really didn't get this timeout warning log in previous kernel(Many
tlb_flush followed by one tlb_sync), But deferring the sync is not
suggested by our DE, thus I still would like to fix the sequence in this
patch with putting them together.

> changed my mind, though. I think we would need to make more changes to
> the driver to make it implement the flushing efficiently, so let's go
> with the current simple approach for now and improve incrementally.
>
> >
> > Currently the synchronisation of the tlb_flush/tlb_sync flow are
> > controlled by the variable "data->tlb_flush_active".
> >
> > In this patch putting the tlb_flush/tlb_sync together looks make
> > the flow simpler:
> > a) Don't need the sensitive variable "tlb_flush_active".
> > b) Remove mtk_iommu_iotlb_sync, Don't need add lock in it.
> > c) Simplify the tlb_flush_walk/tlb_flush_leaf.
> > is it ok?
> >
>
> Okay, let's do so as a first step to fix the issue. Then we can
> optimize in follow up patches.

Thanks the confirm, I have sent a quick v2.

>
> > >
> > > >
> > > > Besides, there are two minor changes:
> > > > a) Use writel for the register F_MMU_INV_RANGE which is for triggering the
> > > > HW work. We expect all the setting(iova_start/iova_end...) have already
> > > > been finished before F_MMU_INV_RANGE.
> > > > b) Reduce the tlb timeout value from 100000us to 1000us. the original value
> > > > is so long that affect the multimedia performance.
> > >
> > > By definition, timeout is something that should not normally happen.
> > > Too long timeout affecting multimedia performance would suggest that
> > > the timeout was actually happening, which is the core problem, not the
> > > length of the timeout. Could you provide more details on this?
> >
> > As description above, this issue is because there is no dom->pgtlock in
> > the mtk_iommu_iotlb_sync. I have tried that the issue will disappear
> > after adding lock in it.
> >
> > Although the issue is fixed after this patch, I still would like to
> > reduce the timeout value for somehow error happen in the future. 100ms
> > is unnecessary for us. It looks a minor improvement rather than fixing
> > the issue. I will use a new patch for it.
> >
>
> Okay, makes sense.
>
> Best regards,
> Tomasz


\
 
 \ /
  Last update: 2019-10-09 15:39    [W:0.038 / U:0.200 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site