On 05.03.21 16:43, Catalin Marinas wrote:
On Fri, Mar 05, 2021 at 11:19:08AM +0000, Catalin Marinas wrote:
On Tue, Mar 02, 2021 at 08:21:03PM -0800, Patrick Daly wrote:
In a system which supports MTE, the linear kernel region must allow
reading/writing allocation tags. For memory present at boot this
is already being done in map_mem(). Add the same in arch_add_memory().
Signed-off-by: Patrick Daly <pdaly@xxxxxxxxxxxxxx>
---
arch/arm64/mm/mmu.c | 12 +++++++++++-
1 file changed, 11 insertions(+), 1 deletion(-)
diff --git a/arch/arm64/mm/mmu.c b/arch/arm64/mm/mmu.c
index 9b25d60b..0fcfe90 100644
--- a/arch/arm64/mm/mmu.c
+++ b/arch/arm64/mm/mmu.c
@@ -1463,6 +1463,7 @@ int arch_add_memory(int nid, u64 start, u64 size,
struct mhp_params *params)
{
int ret, flags = 0;
+ pgprot_t pgprot;
if (!inside_linear_region(start, size)) {
pr_err("[%llx %llx] is outside linear mapping region\n", start, start + size);
@@ -1477,8 +1478,17 @@ int arch_add_memory(int nid, u64 start, u64 size,
IS_ENABLED(CONFIG_KFENCE))
flags = NO_BLOCK_MAPPINGS | NO_CONT_MAPPINGS;
+ /*
+ * The linear map must allow allocation tags reading/writing
+ * if MTE is present. Otherwise, it has the same attributes as
+ * PAGE_KERNEL.
+ */
+ pgprot = params->pgprot;
+ if (pgprot_val(pgprot) == pgprot_val(PAGE_KERNEL))
+ pgprot = PAGE_KERNEL_TAGGED;
+
__create_pgd_mapping(swapper_pg_dir, start, __phys_to_virt(start),
- size, params->pgprot, __pgd_pgtable_alloc,
+ size, pgprot, __pgd_pgtable_alloc,
flags);
We'll need a similar pattern for vmalloc() once we have the khwasan
support in place. So we could add a pgprot_tagged() function (similar to
pgprot_writecombine() etc.) which does the above check and returns
PAGE_KERNEL_TAGGED, maybe only checking the PTE_ATTRINDX_MASK bits
rather than the whole prot bits.
However, the bigger problem is that arch_add_memory() is also called for
ZONE_DEVICE mappings and we can't always guarantee that such range
supports tagging (most likely it doesn't, e.g. persistent memory),
leading to potential external aborts.
One option is to expand mhp_params to pass additional information so
that the arch code can make the right decision. Another option is to
make PAGE_KERNEL_TAGGED global which is just PAGE_KERNEL for all the
other architectures and use it in the core code.
Yet another option which we haven't fully explored with MTE is to have
PAGE_KERNEL always tagged but add a new PAGE_KERNEL_DEVICE (or
_UNTAGGED) for specific cases like ZONE_DEVICE. We need to make sure
that PAGE_KERNEL doesn't end up in places where the backing memory does
not support tags.
I'll give the last option a quick try and see if it falls apart (just
changing PAGE_KERNEL to tagged). In terms of tag cache usage, it
probably won't have much of an impact since the whole of the linear map
is tagged already.
I played with this a bit and the last option is not really feasible.
There are several places in the kernel where PAGE_KERNEL is used just
because the expectation is for write-back memory. For MTE, we need
tagged memory either because the kernel allocators need it (with
KASAN_HW_TAGS) or because it is shared with the user, potentially mapped
as tagged in user space and such tags need to be preserved by the
kernel. So for the latter, only the linear map needs to be tagged (and
memory hotplug falls into this category).
This leaves us with one of the first two options. I think the easiest
is:
---------------8<-------------------------------
diff --git a/arch/arm64/include/asm/pgtable.h b/arch/arm64/include/asm/pgtable.h
index e17b96d0e4b5..5c78b92d9ec5 100644
--- a/arch/arm64/include/asm/pgtable.h
+++ b/arch/arm64/include/asm/pgtable.h
@@ -486,6 +486,8 @@ static inline pmd_t pmd_mkdevmap(pmd_t pmd)
__pgprot_modify(prot, PTE_ATTRINDX_MASK, PTE_ATTRINDX(MT_NORMAL_NC) | PTE_PXN | PTE_UXN)
#define pgprot_device(prot) \
__pgprot_modify(prot, PTE_ATTRINDX_MASK, PTE_ATTRINDX(MT_DEVICE_nGnRE) | PTE_PXN | PTE_UXN)
+#define pgprot_tagged(prot) \
+ __pgprot_modify(prot, PTE_ATTRINDX_MASK, PTE_ATTRINDX(MT_NORMAL_TAGGED))
/*
* DMA allocations for non-coherent devices use what the Arm architecture calls
* "Normal non-cacheable" memory, which permits speculation, unaligned accesses
diff --git a/include/linux/pgtable.h b/include/linux/pgtable.h
index cdfc4e9f253e..f5f5044db2ce 100644
--- a/include/linux/pgtable.h
+++ b/include/linux/pgtable.h
@@ -904,6 +904,10 @@ static inline void ptep_modify_prot_commit(struct vm_area_struct *vma,
#define pgprot_device pgprot_noncached
#endif
+#ifndef pgprot_tagged
+#define pgprot_tagged(prot) (prot)
+#endif
+
#ifdef CONFIG_MMU
#ifndef pgprot_modify
#define pgprot_modify pgprot_modify
diff --git a/mm/memory_hotplug.c b/mm/memory_hotplug.c
index 5ba51a8bdaeb..4253d80a59ba 100644
--- a/mm/memory_hotplug.c
+++ b/mm/memory_hotplug.c
@@ -1072,7 +1072,7 @@ static int online_memory_block(struct memory_block *mem, void *arg)
*/
int __ref add_memory_resource(int nid, struct resource *res, mhp_t mhp_flags)
{
- struct mhp_params params = { .pgprot = PAGE_KERNEL };
+ struct mhp_params params = { .pgprot = pgprot_tagged(PAGE_KERNEL) };
This looks like we're pushing arch specific stuff ("tagged") in here.
Can't we generalize this to something like
pgprot_mhp_default
(or a better name)
that defaults to PAGE_KERNEL on all architectures except arm64 which
overwrites this somehow?
--
Thanks,
David / dhildenb
