[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: [for-4.15][PATCH v3 3/3] xen/iommu: x86: Harden the IOMMU page-table allocator
On 18/02/2021 13:10, Jan Beulich wrote:
On 17.02.2021 17:29, Julien Grall wrote:
On 17/02/2021 15:13, Jan Beulich wrote:
On 17.02.2021 15:24, Julien Grall wrote:> --- a/xen/drivers/passthrough/x86/iommu.c> +++
b/xen/drivers/passthrough/x86/iommu.c> @@ -149,6 +149,13 @@ int arch_iommu_domain_init(struct domain
*d)> > void arch_iommu_domain_destroy(struct domain *d)> {> + /*> + * There
should be not page-tables left allocated by the time the
Nit: s/not/no/ ?
+ * domain is destroyed. Note that arch_iommu_domain_destroy() is
+ * called unconditionally, so pgtables may be unitialized.
+ */
+ ASSERT(dom_iommu(d)->platform_ops == NULL ||
+ page_list_empty(&dom_iommu(d)->arch.pgtables.list));
}
static bool __hwdom_init hwdom_iommu_map(const struct domain *d,
@@ -279,6 +286,9 @@ int iommu_free_pgtables(struct domain *d)
*/
hd->platform_ops->clear_root_pgtable(d);
+ /* After this barrier no new page allocations can occur. */
+ spin_barrier(&hd->arch.pgtables.lock);
Didn't patch 2 utilize the call to ->clear_root_pgtable() itself as
the barrier? Why introduce another one (with a similar comment)
explicitly now?
The barriers act differently, one will get against any IOMMU page-tables
modification. The other one will gate against allocation.
There is no guarantee that the former will prevent the latter.
Oh, right - different locks. I got confused here because in both
cases the goal is to prevent allocations.
@@ -315,9 +326,29 @@ struct page_info *iommu_alloc_pgtable(struct domain *d)
unmap_domain_page(p);
spin_lock(&hd->arch.pgtables.lock);
- page_list_add(pg, &hd->arch.pgtables.list);
+ /*
+ * The IOMMU page-tables are freed when relinquishing the domain, but
+ * nothing prevent allocation to happen afterwards. There is no valid
+ * reasons to continue to update the IOMMU page-tables while the
+ * domain is dying.
+ *
+ * So prevent page-table allocation when the domain is dying.
+ *
+ * We relying on &hd->arch.pgtables.lock to synchronize d->is_dying.
+ */
+ if ( likely(!d->is_dying) )
+ {
+ alive = true;
+ page_list_add(pg, &hd->arch.pgtables.list);
+ }
spin_unlock(&hd->arch.pgtables.lock);
+ if ( unlikely(!alive) )
+ {
+ free_domheap_page(pg);
+ pg = NULL;
+ }
+
return pg;
}
As before I'm concerned of this forcing error paths to be taken
elsewhere, in case an allocation still happens (e.g. from unmap
once super page mappings are supported). Considering some of the
error handling in the IOMMU code is to invoke domain_crash(), it
would be quite unfortunate if we ended up crashing a domain
while it is being cleaned up after.
It is unfortunate, but I think this is better than having to leak page
tables.
Additionally, the (at present still hypothetical) unmap case, if
failing because of the change here, would then again chance to
leave mappings in place while the underlying pages get freed. As
this would likely require an XSA, the change doesn't feel like
"hardening" to me.
I would agree with this if memory allocations could never fail. That's
not that case and will become worse as we use IOMMU pool.
Do you have callers in mind that doesn't check the returns of iommu_unmap()?
The function is marked __must_check, so there won't be any direct
callers ignoring errors (albeit I may be wrong here - we used to
have cases where we simply suppressed the resulting compiler
diagnostic, without really handling errors; not sure if all of
these are gone by now). Risks might be elsewhere.
But this is not a new risk. So I don't understand why you think my patch
is the one that may lead to an XSA in the future.
What my patch could do is expose such issues more easily rather than
waiting until an OOM condition.
Cheers,
--
Julien Grall
|
