Some drivers who call ioremap_prot without setting base flags like
ioremap_prot(addr, len, 0) may work well before
commit 56f3c1413f5c ("powerpc/mm: properly set PAGE_KERNEL flags in
ioremap()"), but now they will get a virtual address "successfully"
from ioremap_prot and badly fault on memory access later because that
patch also dropped the hack adding of base flags for ioremap_prot.
So return NULL and throw a warning if the caller of ioremap_prot did
not set base flags properly. Why not just hack adding PAGE_KERNEL flags
in the ioremap_prot, because most scenarios can be covered by high level
functions like ioremap(), ioremap_coherent(), ioremap_cache()...
so it is better to keep max flexibility for this low level api.
Signed-off-by: Nanyong Sun <[email protected]>
---
arch/powerpc/mm/ioremap.c | 4 ++++
1 file changed, 4 insertions(+)
diff --git a/arch/powerpc/mm/ioremap.c b/arch/powerpc/mm/ioremap.c
index 57342154d2b0..b7eda0f0d04d 100644
--- a/arch/powerpc/mm/ioremap.c
+++ b/arch/powerpc/mm/ioremap.c
@@ -46,6 +46,10 @@ void __iomem *ioremap_prot(phys_addr_t addr, unsigned long size, unsigned long f
pte_t pte = __pte(flags);
void *caller = __builtin_return_address(0);
+ /* The caller should set base page flags properly */
+ if (WARN_ON((flags & _PAGE_PRESENT) == 0))
+ return NULL;
+
/* writeable implies dirty for kernel addresses */
if (pte_write(pte))
pte = pte_mkdirty(pte);
--
2.18.0.huawei.25
Le 03/09/2021 à 11:03, Nanyong Sun a écrit :
> Some drivers who call ioremap_prot without setting base flags like
> ioremap_prot(addr, len, 0) may work well before
> commit 56f3c1413f5c ("powerpc/mm: properly set PAGE_KERNEL flags in
> ioremap()"), but now they will get a virtual address "successfully"
> from ioremap_prot and badly fault on memory access later because that
> patch also dropped the hack adding of base flags for ioremap_prot.
>
> So return NULL and throw a warning if the caller of ioremap_prot did
> not set base flags properly. Why not just hack adding PAGE_KERNEL flags
> in the ioremap_prot, because most scenarios can be covered by high level
> functions like ioremap(), ioremap_coherent(), ioremap_cache()...
> so it is better to keep max flexibility for this low level api.
As far as I can see, there is no user of this fonction that sets flags to 0 in the kernel tree.
Did you find any ? If you did, I think it is better to fix the caller.
Christophe
>
> Signed-off-by: Nanyong Sun <[email protected]>
> ---
> arch/powerpc/mm/ioremap.c | 4 ++++
> 1 file changed, 4 insertions(+)
>
> diff --git a/arch/powerpc/mm/ioremap.c b/arch/powerpc/mm/ioremap.c
> index 57342154d2b0..b7eda0f0d04d 100644
> --- a/arch/powerpc/mm/ioremap.c
> +++ b/arch/powerpc/mm/ioremap.c
> @@ -46,6 +46,10 @@ void __iomem *ioremap_prot(phys_addr_t addr, unsigned long size, unsigned long f
> pte_t pte = __pte(flags);
> void *caller = __builtin_return_address(0);
>
> + /* The caller should set base page flags properly */
> + if (WARN_ON((flags & _PAGE_PRESENT) == 0))
This probably doesn't work for some plateforms like book3s/64. You should use helpers like
pte_present().
See the comment at
https://elixir.bootlin.com/linux/v5.14/source/arch/powerpc/include/asm/book3s/64/pgtable.h#L591
> + return NULL;
> +
> /* writeable implies dirty for kernel addresses */
> if (pte_write(pte))
> pte = pte_mkdirty(pte);
>
Nanyong Sun <[email protected]> writes:
> On 2021/9/3 17:16, Christophe Leroy wrote:
>> Le 03/09/2021 à 11:03, Nanyong Sun a écrit :
>>> Some drivers who call ioremap_prot without setting base flags like
>>> ioremap_prot(addr, len, 0) may work well before
>>> commit 56f3c1413f5c ("powerpc/mm: properly set PAGE_KERNEL flags in
>>> ioremap()"), but now they will get a virtual address "successfully"
>>> from ioremap_prot and badly fault on memory access later because that
>>> patch also dropped the hack adding of base flags for ioremap_prot.
>>>
>>> So return NULL and throw a warning if the caller of ioremap_prot did
>>> not set base flags properly. Why not just hack adding PAGE_KERNEL flags
>>> in the ioremap_prot, because most scenarios can be covered by high level
>>> functions like ioremap(), ioremap_coherent(), ioremap_cache()...
>>> so it is better to keep max flexibility for this low level api.
>>
>> As far as I can see, there is no user of this fonction that sets flags
>> to 0 in the kernel tree.
>>
>> Did you find any ? If you did, I think it is better to fix the caller.
>>
>> Christophe
>>
> I see some vendor's drivers which are not on upstream ...
Sorry, but we don't carry extraneous checks in upstream for the sake of
out-of-tree drivers.
cheers