linux-kernel - Re: [PATCH 1/2] iommu/io-pgtable-arm: Implement .iotlb_sync

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives
Hash Suite for Android: free password hash cracker in your pocket
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <aQn_g85KI_uuYpJh@willie-the-truck>
Date: Tue, 4 Nov 2025 13:28:35 +0000
From: Will Deacon <will@...nel.org>
To: Mostafa Saleh <smostafa@...gle.com>
Cc: Daniel Mentz <danielmentz@...gle.com>, iommu@...ts.linux.dev,
	linux-kernel@...r.kernel.org,
	Pranjal Shrivastava <praan@...gle.com>,
	Liviu Dudau <liviu.dudau@....com>, Jason Gunthorpe <jgg@...dia.com>,
	Rob Clark <robin.clark@....qualcomm.com>
Subject: Re: [PATCH 1/2] iommu/io-pgtable-arm: Implement .iotlb_sync_map
 callback

On Tue, Sep 30, 2025 at 09:10:44AM +0000, Mostafa Saleh wrote:
> On Mon, Sep 29, 2025 at 02:00:09PM -0700, Daniel Mentz wrote:
> > On Mon, Sep 29, 2025 at 5:21 AM Mostafa Saleh <smostafa@...gle.com> wrote:
> > > On Sat, Sep 27, 2025 at 10:39:52PM +0000, Daniel Mentz wrote:
> > > > @@ -582,6 +582,69 @@ static int arm_lpae_map_pages(struct io_pgtable_ops *ops, unsigned long iova,
> > > >       return ret;
> > > >  }
> > > >
> > > > +static int __arm_lpae_iotlb_sync_map(struct arm_lpae_io_pgtable *data, unsigned long iova,
> > > > +                           size_t size, int lvl, arm_lpae_iopte *ptep)
> > > > +{
> > > > +     struct io_pgtable *iop = &data->iop;
> > > > +     size_t block_size = ARM_LPAE_BLOCK_SIZE(lvl, data);
> > > > +     int ret = 0, num_entries, max_entries;
> > > > +     unsigned long iova_offset, sync_idx_start, sync_idx_end;
> > > > +     int i, shift, synced_entries = 0;
> > > > +
> > > > +     shift = (ARM_LPAE_LVL_SHIFT(lvl - 1, data) + ARM_LPAE_PGD_IDX(lvl - 1, data));
> > > > +     iova_offset = iova & ((1ULL << shift) - 1);
> > > > +     sync_idx_start = ARM_LPAE_LVL_IDX(iova, lvl, data);
> > > > +     sync_idx_end = (iova_offset + size + block_size - ARM_LPAE_GRANULE(data)) >>
> > > > +             ARM_LPAE_LVL_SHIFT(lvl, data);
> > > > +     max_entries = arm_lpae_max_entries(sync_idx_start, data);
> > > > +     num_entries = min_t(unsigned long, sync_idx_end - sync_idx_start, max_entries);
> > > > +     ptep += sync_idx_start;
> > > > +
> > > > +     if (lvl < (ARM_LPAE_MAX_LEVELS - 1)) {
> > > > +             for (i = 0; i < num_entries; i++) {
> > > > +                     arm_lpae_iopte pte = READ_ONCE(ptep[i]);
> > > > +                     unsigned long synced;
> > > > +
> > > > +                     WARN_ON(!pte);
> > > > +
> > > > +                     if (iopte_type(pte) == ARM_LPAE_PTE_TYPE_TABLE) {
> > > > +                             int n = i - synced_entries;
> > > > +
> > > > +                             if (n) {
> > > > +                                     __arm_lpae_sync_pte(&ptep[synced_entries], n, &iop->cfg);
> > > > +                                     synced_entries += n;
> > > > +                             }
> > > > +                             ret = __arm_lpae_iotlb_sync_map(data, iova, size, lvl + 1,
> > > > +                                                             iopte_deref(pte, data));
> > > > +                             synced_entries++;
> > > > +                     }
> > > > +                     synced = block_size - (iova & (block_size - 1));
> > > > +                     size -= synced;
> > > > +                     iova += synced;
> > > > +             }
> > > > +     }
> > > > +
> > > > +     if (synced_entries != num_entries)
> > > > +             __arm_lpae_sync_pte(&ptep[synced_entries], num_entries - synced_entries, &iop->cfg);
> > > > +
> > > > +     return ret;
> > > > +}
> > >
> > > Can't we rely on the exisiting generic table walker "__arm_lpae_iopte_walk",
> > > instead writing a new one, that is already used for iova_to_phys and dirty bit.
> > 
> > The performance gains of .iotlb_sync_map are achieved by performing
> > CMOs on a range of descriptors as opposed to individually on each
> > descriptor in isolation. The function __arm_lpae_iopte_walk is
> > inherently incompatible with this, because it calls the .visit
> > callback once for each descriptor it finds in the specified range. I
> > guess I could work around this limitation by saving some state in
> > io_pgtable_walk_data and developing a .visit function that tries to
> > coalesce individual descriptors into contiguous ranges and delays CMOs
> > until it finds a break in continuity. I'm afraid, though, that that
> > might hurt performance significantly.
> 
> Exactly, I think that would be the way, I don’t have a strong opinion
> though, but I’d avoid open coding a new walker unless it’s necessary.
> Also, the current walker won’t do ranges, it needs some more changes,
> I did that as part of (half of the patch doesn’t apply for this case):
> https://lore.kernel.org/all/20241212180423.1578358-38-smostafa@google.com/

I'm inclined to agree that it would be better to avoid open-coding a
new walker here and if we're able to reuse/extend the generic walker
then that would be cleaner.

If that's not workable (due to Daniel's performance worries), another
option is to bring back the ->map_sg() hook (removed by d88e61faad52
("iommu: Remove the ->map_sg indirection")) and implement an optimised
version of that, preferably sharing as much code as possible with the
existing map path.

Will