Re: [PATCH 0/2] PUD/PGDIR entries for linear mapping

[Alex Ghiti <alex@ghiti.fr>]

Le 6/12/20 à 8:59 AM, Alex Ghiti a écrit :
> Hi Atish,
>
> Le 6/11/20 à 1:29 PM, Atish Patra a écrit :
>> On Wed, Jun 10, 2020 at 11:51 PM Alex Ghiti<alex@ghiti.fr>  wrote:
>>> Hi Atish,
>>>
>>> Le 6/10/20 à 2:32 PM, Atish Patra a écrit :
>>>> On Wed, Jun 3, 2020 at 8:36 AM Alexandre Ghiti<alex@ghiti.fr>  wrote:
>>>>> This small patchset intends to use PUD/PGDIR entries for linear 
>>>>> mapping
>>>>> in order to better utilize TLB.
>>>>>
>>>>> At the moment, only PMD entries can be used since on common platforms
>>>>> (qemu/unleashed), the kernel is loaded at DRAM + 2MB which 
>>>>> dealigns virtual
>>>>> and physical addresses and then prevents the use of PUD/PGDIR 
>>>>> entries.
>>>>> So the kernel must be able to get those 2MB for PAGE_OFFSET to map 
>>>>> the
>>>>> beginning of the DRAM: this is achieved in patch 1.
>>>>>
>>>> I don't have in depth knowledge of how mm code works so this question
>>>> may be a completely
>>>> stupid one :). Just for my understanding,
>>>> As per my understanding, kernel will map those 2MB of memory but 
>>>> never use it.
>>>> How does the kernel ensure that it doesn't allocate any memory from 
>>>> those 2MB
>>>> memory if it is not marked as reserved?
>>> Yes, a 1GB hugepage will cover those 2MB: I rely on the previous boot
>>> stage to mark this region
>>> as reserved if there is something there (like opensbi). Otherwise, the
>>> kernel will indeed try to
>>> allocate memory from there :)
>>>
>> In that case, this patch mandates that the firmware region has to be
>> mark "reserved"
>> the device tree so that the Linux kernel doesn't try to allocate
>> memory from there.
>> OpenSBI is already doing it from v0.7. Thus, any user using latest
>> OpenSBI can leverage
>> this patch for a better TLB utilization.
>
>
> Note that *currently* OpenSBI v0.7 still adds the "no-map" property 
> which prevents such optimization.
>
>> However, legacy previous boot stages(BBL) do not reserve this area via
>> DT which may
>> result in an unexpected crash. I am not sure how many developers still
>> use BBL though.
>>
>> Few general suggestions to tackle this problem:
>> 1. This mandatory requirement should be added to the booting document
>> so that any other
>> SBI implementation is also aware of it.
>> 2. You may have to move the patch1 to a separate config so that any
>> users of legacy boot stages
>> can disable this feature.
>
>
> IMHO, the region occupied by runtime services should be marked as 
> reserved in the device-tree. So it seems redundant to add this as a 
> requirement, I would rather consider its absence as a bug.
>
> Even if I understand that this might break some system, I don't like 
> the idea of a new config to support old "buggy" bootloaders: when will 
> we be able to remove it ? We'll never know when people will stop using 
> those bootloaders, so it will stay here forever...Where can I find the 
> boot document you are talking about ? Can we simply state here that 
> this kernel version will not be compatible with those bootloaders 
> (we'll draw an exhaustive list here) ?


Ok, I have just found Documentation/riscv/boot-image-header.rst: could 
we imagine doing something like incrementing the version and use that as 
a hint in the kernel not to map the 2MB offset ? That's still legacy, 
but at least it does not require to recompile a kernel as the check 
would be done at runtime.


>
> Alex
>
>
>>> Alex
>>>
>>>
>>>>> But furthermore, at the moment, the firmware (opensbi) explicitly 
>>>>> asks the
>>>>> kernel not to map the region it occupies, which is on those common
>>>>> platforms at the very beginning of the DRAM and then it also dealigns
>>>>> virtual and physical addresses. I proposed a patch here:
>>>>>
>>>>> https://github.com/riscv/opensbi/pull/167
>>>>>
>>>>> that removes this 'constraint' but *not* all the time as it offers 
>>>>> some
>>>>> kind of protection in case PMP is not available. So sometimes, we may
>>>>> have a part of the memory below the kernel that is removed creating a
>>>>> misalignment between virtual and physical addresses. So for 
>>>>> performance
>>>>> reasons, we must at least make sure that PMD entries can be used: 
>>>>> that
>>>>> is guaranteed by patch 1 too.
>>>>>
>>>>> Finally the second patch simply improves best_map_size so that 
>>>>> whenever
>>>>> possible, PUD/PGDIR entries are used.
>>>>>
>>>>> Below is the kernel page table without this patch on a 6G platform:
>>>>>
>>>>> ---[ Linear mapping ]---
>>>>> 0xffffc00000000000-0xffffc00176e00000 0x0000000080200000 5998M 
>>>>> PMD     D A . . . W R V
>>>>>
>>>>> And with this patchset + opensbi patch:
>>>>>
>>>>> ---[ Linear mapping ]---
>>>>> 0xffffc00000000000-0xffffc00140000000 0x0000000080000000         
>>>>> 5G PUD     D A . . . W R V
>>>>> 0xffffc00140000000-0xffffc00177000000 0x00000001c0000000 880M 
>>>>> PMD     D A . . . W R V
>>>>>
>>>>> Alexandre Ghiti (2):
>>>>>     riscv: Get memory below load_pa while ensuring linear mapping 
>>>>> is PMD
>>>>>       aligned
>>>>>     riscv: Use PUD/PGDIR entries for linear mapping when possible
>>>>>
>>>>>    arch/riscv/include/asm/page.h |  8 ++++
>>>>>    arch/riscv/mm/init.c          | 69 
>>>>> +++++++++++++++++++++++++++++------
>>>>>    2 files changed, 65 insertions(+), 12 deletions(-)
>>>>>
>>>>> -- 
>>>>> 2.20.1
>>>>>
>>>>>
>>
>