[U-Boot] [PATCH v6 03/13] efi: sandbox: Adjust memory usage for sandbox

Thu Jun 14 16:55:43 UTC 2018

Hi Alex,

On 14 June 2018 at 10:42, Alexander Graf <agraf at suse.de> wrote:
> On 06/14/2018 06:33 PM, Simon Glass wrote:
>>
>> Hi Alex,
>>
>> On 14 June 2018 at 10:26, Alexander Graf <agraf at suse.de> wrote:
>>>
>>> On 06/14/2018 06:13 PM, Simon Glass wrote:
>>>>
>>>> Hi Alex,
>>>>
>>>> On 14 June 2018 at 10:07, Alexander Graf <agraf at suse.de> wrote:
>>>>>
>>>>> On 06/14/2018 05:53 PM, Simon Glass wrote:
>>>>>>
>>>>>> Hi Alex,
>>>>>>
>>>>>> On 14 June 2018 at 09:47, Alexander Graf <agraf at suse.de> wrote:
>>>>>>>
>>>>>>>
>>>>>>>> Am 14.06.2018 um 17:43 schrieb Simon Glass <sjg at chromium.org>:
>>>>>>>>
>>>>>>>> Hi Alex,
>>>>>>>>
>>>>>>>>> On 14 June 2018 at 08:22, Alexander Graf <agraf at suse.de> wrote:
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>> Am 14.06.2018 um 16:12 schrieb Simon Glass <sjg at chromium.org>:
>>>>>>>>>>
>>>>>>>>>> Hi Alex,
>>>>>>>>>>
>>>>>>>>>>>> On 14 June 2018 at 07:41, Alexander Graf <agraf at suse.de> wrote:
>>>>>>>>>>>> On 06/14/2018 02:58 PM, Simon Glass wrote:
>>>>>>>>>>>>
>>>>>>>>>>>> Hi Alex,
>>>>>>>>>>>>
>>>>>>>>>>>>>> On 14 June 2018 at 04:12, Alexander Graf <agraf at suse.de>
>>>>>>>>>>>>>> wrote:
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> On 06/13/2018 04:37 AM, Simon Glass wrote:
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> With sandbox the U-Boot code is not mapped into the sandbox
>>>>>>>>>>>>>> memory
>>>>>>>>>>>>>> range
>>>>>>>>>>>>>> so does not need to be excluded when allocating EFI memory.
>>>>>>>>>>>>>> Update
>>>>>>>>>>>>>> the
>>>>>>>>>>>>>> EFI
>>>>>>>>>>>>>> memory init code to take account of that.
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> Also use mapmem instead of a cast to convert a memory address
>>>>>>>>>>>>>> to
>>>>>>>>>>>>>> a
>>>>>>>>>>>>>> pointer.
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> Signed-off-by: Simon Glass <sjg at chromium.org>
>>>>>>>>>>>>>> ---
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> Changes in v6: None
>>>>>>>>>>>>>> Changes in v5: None
>>>>>>>>>>>>>> Changes in v4: None
>>>>>>>>>>>>>> Changes in v3: None
>>>>>>>>>>>>>> Changes in v2:
>>>>>>>>>>>>>> - Update to use mapmem instead of a cast
>>>>>>>>>>>>>>
>>>>>>>>>>>>>>     lib/efi_loader/efi_memory.c | 31
>>>>>>>>>>>>>> ++++++++++++++++++-------------
>>>>>>>>>>>>>>     1 file changed, 18 insertions(+), 13 deletions(-)
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> diff --git a/lib/efi_loader/efi_memory.c
>>>>>>>>>>>>>> b/lib/efi_loader/efi_memory.c
>>>>>>>>>>>>>> index ec66af98ea..210e49ee8b 100644
>>>>>>>>>>>>>> --- a/lib/efi_loader/efi_memory.c
>>>>>>>>>>>>>> +++ b/lib/efi_loader/efi_memory.c
>>>>>>>>>>>>>> @@ -9,6 +9,7 @@
>>>>>>>>>>>>>>     #include <efi_loader.h>
>>>>>>>>>>>>>>     #include <inttypes.h>
>>>>>>>>>>>>>>     #include <malloc.h>
>>>>>>>>>>>>>> +#include <mapmem.h>
>>>>>>>>>>>>>>     #include <watchdog.h>
>>>>>>>>>>>>>>     #include <linux/list_sort.h>
>>>>>>>>>>>>>>     @@ -393,7 +394,7 @@ efi_status_t efi_allocate_pool(int
>>>>>>>>>>>>>> pool_type,
>>>>>>>>>>>>>> efi_uintn_t size, void **buffer)
>>>>>>>>>>>>>>                                  &t);
>>>>>>>>>>>>>>           if (r == EFI_SUCCESS) {
>>>>>>>>>>>>>> -               struct efi_pool_allocation *alloc = (void
>>>>>>>>>>>>>> *)(uintptr_t)t;
>>>>>>>>>>>>>> +               struct efi_pool_allocation *alloc =
>>>>>>>>>>>>>> map_sysmem(t,
>>>>>>>>>>>>>> size);
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>> This is still the wrong spot. We don't want the conversion to
>>>>>>>>>>>>> happen when
>>>>>>>>>>>>> going from an EFI internal address to an allocation, but when
>>>>>>>>>>>>> determining
>>>>>>>>>>>>> which addresses are usable in the first place.
>>>>>>>>>>>>
>>>>>>>>>>>> There seem to be two ways to do this:
>>>>>>>>>>>>
>>>>>>>>>>>> 1. Record addresses (ulong) in the EFI tables and use
>>>>>>>>>>>> map_sysmem()
>>>>>>>>>>>> before returning an address in the allocator
>>>>>>>>>>>> 2. Store pointers in the EFI tables using map_sysmem(), then do
>>>>>>>>>>>> no
>>>>>>>>>>>> mapping in the allocator
>>>>>>>>>>>>
>>>>>>>>>>>> I've gone with option 1 since:
>>>>>>>>>>>>
>>>>>>>>>>>> - the EFI addresses are not pointers
>>>>>>>>>>>> - it makes sandbox consistent with other architectures which use
>>>>>>>>>>>> an
>>>>>>>>>>>> address rather than a pointer in EFI tables
>>>>>>>>>>>> - we don't normally do pointer arithmetic on the results of
>>>>>>>>>>>> map_sysmem()
>>>>>>>>>>>> - we normally map the memory when it is used rather than when it
>>>>>>>>>>>> is
>>>>>>>>>>>> set up
>>>>>>>>>>>>
>>>>>>>>>>>> I think you are asking for option 2. I think that would get very
>>>>>>>>>>>> confusing. The addresses where things actually end up in sandbox
>>>>>>>>>>>> are
>>>>>>>>>>>> best kept to sandbox.
>>>>>>>>>>>>
>>>>>>>>>>>> Overall I feel that you are either missing the point of sandbox
>>>>>>>>>>>> addressing, or don't agree with how it is done. But it does work
>>>>>>>>>>>> pretty well and we don't get a lot of confusion with sandbox
>>>>>>>>>>>> pointers
>>>>>>>>>>>> since we typically use the address until the last moment.
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> I've assembled a quick tree for version 2. With that I'm able to
>>>>>>>>>>> run
>>>>>>>>>>> a
>>>>>>>>>>> simple hello world efi application. Grub refuses to start because
>>>>>>>>>>> it
>>>>>>>>>>> wants
>>>>>>>>>>> memory in the low 32bit and also emits random PIO accessing
>>>>>>>>>>> functions, which
>>>>>>>>>>> obviously don't work work from user space.
>>>>>>>>>>>
>>>>>>>>>>> But overall, I think this is the right path to tackle this:
>>>>>>>>>>>
>>>>>>>>>>> https://github.com/agraf/u-boot/tree/efi-sandbox
>>>>>>>>>>
>>>>>>>>>> What do you mean by version 2?
>>>>>>>>>
>>>>>>>>> Option 2 is what you called it. It's the only option we have to
>>>>>>>>> make
>>>>>>>>> efi binaries work.
>>>>>>>>>
>>>>>>>>>> It looks like you've added one patch,
>>>>>>>>>> so will you send that to the list?
>>>>>>>>>
>>>>>>>>> It's more than 1 patch. And yes, I'll send them.
>>>>>>>>>
>>>>>>>>>> Anyway, I hope I can convince you of the above, the way sandbox
>>>>>>>>>> memory
>>>>>>>>>> works.
>>>>>>>>>
>>>>>>>>> I still dislike option 1 :)
>>>>>>>>>
>>>>>>>>> The reason is simple: The efi memory map is available to efi
>>>>>>>>> payloads.
>>>>>>>>> It's perfectly legal for them to do a static allocation at a
>>>>>>>>> particular
>>>>>>>>> address. We also share a lot of (host) pointers for callbacks and
>>>>>>>>> structs
>>>>>>>>> already with efi applications, so there is no real point to have a
>>>>>>>>> split
>>>>>>>>> brain situation between u-boot and host pointers.
>>>>>>>>
>>>>>>>> OK so you mean they don't have to allocate memory before using it?
>>>>>>>> How
>>>>>>>> then do you make sure that there is no conflict? I thought EFI was
>>>>>>>> an
>>>>>>>> API?
>>>>>>>
>>>>>>> You allocate it, but payloads expect that the address you pass in as
>>>>>>> address you allocate at is the pointer/address that gets returned.
>>>>>>>
>>>>>> Can you please point me to that? Are you referring to this call?
>>>>>>
>>>>>> static efi_status_t EFIAPI efi_get_memory_map_ext(
>>>>>>            efi_uintn_t *memory_map_size,
>>>>>>            struct efi_mem_desc *memory_map,
>>>>>>            efi_uintn_t *map_key,
>>>>>>            efi_uintn_t *descriptor_size,
>>>>>>            uint32_t *descriptor_version)
>>>>>>
>>>>>> If so, we can still support that.
>>>>>>
>>>>>>> In a nutshell, what you're proposing is that the efi quivalent of
>>>>>>> mmap(addr, MAP_FIXED) returns something != addr. That will break
>>>>>>> things.
>>>>>>
>>>>>> I see this function:
>>>>>>
>>>>>> efi_status_t efi_allocate_pool(int pool_type, efi_uintn_t size, void
>>>>>> **buffer)
>>>>>>
>>>>>> So there appears to be no way for people to specify the address they
>>>>>> want. Is there another call somewhere?
>>>>>
>>>>>
>>>>> You can call efi_allocate_pages() with EFI_ALLOCATE_ADDRESS or even
>>>>> EFI_ALLOCATE_MAX_ADDRESS which you usually base on information you
>>>>> receive
>>>>> from efi_get_memory_map().
>>>>
>>>> That's fine though, I'm not worried about that. It will work correctly,
>>>> right?
>>>>
>>>>>> NOTE: I am not proposing that the address is made available to the EFI
>>>>>> application - that would always see the sandbox pointer.
>>>>>
>>>>>
>>>>> In that case it's *much* easier to just always store host pointers in
>>>>> the
>>>>> efi memory map. Everything else will be very intrusive.
>>>>
>>>> Yes there will be map_sysmem() and map_to_sysmem() calls at each
>>>> boundary. That 'intrusion' is what we have everywhere in sandbox and
>>>> it is how things work. But the nice thing is that the tables don't end
>>>> up with 'private' sandbox addresses in them. Also see me comments
>>>> above.
>>>
>>>
>>> I seem to be missing the point. efi_loader is just a tiny shim layer
>>> whose
>>> only purpose it is to interface between the binary efi interface and
>>> U-Boot
>>> internal code. So why shouldn't efi_loader treat addresses as pointers
>>> internally?
>>>
>>> So how about this: I'll send out a patch set that works the way I think
>>> is
>>> the right way forward. With this I can at least run binaries generated by
>>> gnu-efi. Then you can send a patch on top with a proposal how you think
>>> the
>>> mapping should happen. My guess is that your patch will just clutter the
>>> code, but I'll be happy to be persuaded otherwise.
>>
>> I'm just not keen on putting sandbox pointers in internal data
>> structures, sorry. It's not the way it is supposed to work and it
>> makes sandbox diffferent from all the other archs. Trying to minimise
>> the number of calls to map_sysmem() is not really the goal here.
>
>
> Well, you're not the one who needs to maintain efi_loader code after every
> 2nd line in it ends up being some random map/unmap call ;).
>
> But my proposal was serious. Putting virtual addresses in is trivial and
> gives us proper API interfaces. We can based on that still decide whether
> it's better to not have pointers in the map. I am not convinced yet, but
> open to the idea.
>
> And yes, code maintainability is my uttermost goal of this. Things have to
> be simple for sandbox, otherwise they will break.

A virtual address has a particular meaning. This is not what is going
on with sandbox...

Making sandbox store its internal addresses in an EFI table is not a
good idea. I've tried to explain why. It is not more maintainable - it
just adds confusion. At present all ulong addresses are normal U-Boot
addresses and pointers point into the emulated RAM. It's pretty clear
for people. Please don't change that for EFI.

Can we get my patches applied and then add yours on top? I will send
v7 which I think addresses all comments.

Regards,
Simon