[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [PATCH v4 6/7] xen/riscv: page table handling


  • To: oleksii.kurochko@xxxxxxxxx
  • From: Jan Beulich <jbeulich@xxxxxxxx>
  • Date: Thu, 15 Aug 2024 17:26:21 +0200
  • Autocrypt: addr=jbeulich@xxxxxxxx; keydata= xsDiBFk3nEQRBADAEaSw6zC/EJkiwGPXbWtPxl2xCdSoeepS07jW8UgcHNurfHvUzogEq5xk hu507c3BarVjyWCJOylMNR98Yd8VqD9UfmX0Hb8/BrA+Hl6/DB/eqGptrf4BSRwcZQM32aZK 7Pj2XbGWIUrZrd70x1eAP9QE3P79Y2oLrsCgbZJfEwCgvz9JjGmQqQkRiTVzlZVCJYcyGGsD /0tbFCzD2h20ahe8rC1gbb3K3qk+LpBtvjBu1RY9drYk0NymiGbJWZgab6t1jM7sk2vuf0Py O9Hf9XBmK0uE9IgMaiCpc32XV9oASz6UJebwkX+zF2jG5I1BfnO9g7KlotcA/v5ClMjgo6Gl MDY4HxoSRu3i1cqqSDtVlt+AOVBJBACrZcnHAUSuCXBPy0jOlBhxPqRWv6ND4c9PH1xjQ3NP nxJuMBS8rnNg22uyfAgmBKNLpLgAGVRMZGaGoJObGf72s6TeIqKJo/LtggAS9qAUiuKVnygo 3wjfkS9A3DRO+SpU7JqWdsveeIQyeyEJ/8PTowmSQLakF+3fote9ybzd880fSmFuIEJldWxp Y2ggPGpiZXVsaWNoQHN1c2UuY29tPsJgBBMRAgAgBQJZN5xEAhsDBgsJCAcDAgQVAggDBBYC AwECHgECF4AACgkQoDSui/t3IH4J+wCfQ5jHdEjCRHj23O/5ttg9r9OIruwAn3103WUITZee e7Sbg12UgcQ5lv7SzsFNBFk3nEQQCACCuTjCjFOUdi5Nm244F+78kLghRcin/awv+IrTcIWF hUpSs1Y91iQQ7KItirz5uwCPlwejSJDQJLIS+QtJHaXDXeV6NI0Uef1hP20+y8qydDiVkv6l IreXjTb7DvksRgJNvCkWtYnlS3mYvQ9NzS9PhyALWbXnH6sIJd2O9lKS1Mrfq+y0IXCP10eS FFGg+Av3IQeFatkJAyju0PPthyTqxSI4lZYuJVPknzgaeuJv/2NccrPvmeDg6Coe7ZIeQ8Yj t0ARxu2xytAkkLCel1Lz1WLmwLstV30g80nkgZf/wr+/BXJW/oIvRlonUkxv+IbBM3dX2OV8 AmRv1ySWPTP7AAMFB/9PQK/VtlNUJvg8GXj9ootzrteGfVZVVT4XBJkfwBcpC/XcPzldjv+3 HYudvpdNK3lLujXeA5fLOH+Z/G9WBc5pFVSMocI71I8bT8lIAzreg0WvkWg5V2WZsUMlnDL9 mpwIGFhlbM3gfDMs7MPMu8YQRFVdUvtSpaAs8OFfGQ0ia3LGZcjA6Ik2+xcqscEJzNH+qh8V m5jjp28yZgaqTaRbg3M/+MTbMpicpZuqF4rnB0AQD12/3BNWDR6bmh+EkYSMcEIpQmBM51qM EKYTQGybRCjpnKHGOxG0rfFY1085mBDZCH5Kx0cl0HVJuQKC+dV2ZY5AqjcKwAxpE75MLFkr wkkEGBECAAkFAlk3nEQCGwwACgkQoDSui/t3IH7nnwCfcJWUDUFKdCsBH/E5d+0ZnMQi+G0A nAuWpQkjM1ASeQwSHEeAWPgskBQL
  • Cc: Alistair Francis <alistair.francis@xxxxxxx>, Bob Eshleman <bobbyeshleman@xxxxxxxxx>, Connor Davis <connojdavis@xxxxxxxxx>, Andrew Cooper <andrew.cooper3@xxxxxxxxxx>, Julien Grall <julien@xxxxxxx>, Stefano Stabellini <sstabellini@xxxxxxxxxx>, xen-devel@xxxxxxxxxxxxxxxxxxxx
  • Delivery-date: Thu, 15 Aug 2024 15:26:26 +0000
  • List-id: Xen developer discussion <xen-devel.lists.xenproject.org>

On 15.08.2024 15:34, oleksii.kurochko@xxxxxxxxx wrote:
> On Thu, 2024-08-15 at 14:16 +0200, Jan Beulich wrote:
>> On 15.08.2024 13:21, oleksii.kurochko@xxxxxxxxx wrote:
>>> On Thu, 2024-08-15 at 10:09 +0200, Jan Beulich wrote:
>>>> On 14.08.2024 18:50, oleksii.kurochko@xxxxxxxxx wrote:
>>>>> On Tue, 2024-08-13 at 12:31 +0200, Jan Beulich wrote:
>>>>>> On 09.08.2024 18:19, Oleksii Kurochko wrote:
>>>>>>> RISC-V detects superpages using `pte.x` and `pte.r`, as
>>>>>>> there
>>>>>>> is no specific bit in the PTE for this purpose. From the
>>>>>>> RISC-V
>>>>>>> spec:
>>>>>>> ```
>>>>>>>   ...
>>>>>>>   4. Otherwise, the PTE is valid. If pte.r = 1 or pte.x =
>>>>>>> 1, go
>>>>>>> to
>>>>>>> step 5.
>>>>>>>      Otherwise, this PTE is a pointer to the next level of
>>>>>>> the
>>>>>>> page
>>>>>>> table.
>>>>>>>      ... .
>>>>>>>   5. A leaf PTE has been found.
>>>>>>>      ...
>>>>>>>   ...
>>>>>>> ```
>>>>>>>
>>>>>>> The code doesn’t support super page shattering so 4KB pages
>>>>>>> are
>>>>>>> used as
>>>>>>> default.
>>>>>>
>>>>>> ... you have this. Yet still callers expecting re-mapping in
>>>>>> the
>>>>>> (large)
>>>>>> range they map can request small-page mappings right away.
>>>>> I am not sure that I fully understand what do you mean by
>>>>> "expcting
>>>>> re-
>>>>> mapping".
>>>>
>>>> Right now you have callers pass PTE_BLOCK when they know that no
>>>> small
>>>> page re-mappings are going to occur for an area. What I'm
>>>> suggesting
>>>> is
>>>> that you invert this logic: Have callers pass PTE_SMALL when
>>>> there is
>>>> a possibility that re-mapping requests may be issued later. Then,
>>>> later, by simply grep-ing for PTE_SMALL you'll be able to easily
>>>> find
>>>> all candidates that possibly can be relaxed when super-page
>>>> shattering
>>>> starts being supported. That's going to be easier than finding
>>>> all
>>>> instances where PTE_BLOCK is _not_used.
>>> So if I understand correctly. Actually nothing will change in
>>> algorithm
>>> of pt_update() and only PTE_SMALL should be introduced instead of
>>> PTE_BLOCK. And if I will know that something will be better to map
>>> as
>>> PTE_SMALL to not face shattering in case of unmap (for example) I
>>> just
>>> can map this memory as PTE_SMALL and that is it?
>>
>> That is it.
>>
>>>>>>> +    spin_unlock(&xen_pt_lock);
>>>>>>
>>>>>> Does this really need to come after fence and flush?
>>>>> I think yes, as page table should be updated only by 1 CPU at
>>>>> the
>>>>> same
>>>>> time. And before give ability to other CPU to update page table
>>>>> we
>>>>> have
>>>>> to finish a work on current CPU.
>>>>
>>>> Can you then explain to me, perhaps by way of an example, what
>>>> will
>>>> go
>>>> wrong if the unlock is ahead of the flush? (I'm less certain
>>>> about
>>>> the
>>>> fence, and that's also less expensive.)
>>> pt_update() will be called for interleaved region, for example, by
>>> different CPUs:
>>>
>>>                      pt_update():
>>> CPU1:                                    CPU2:
>>>  ...                                spin_lock(&xen_pt_lock);
>>> RISCV_FENCE(rw, rw);                 ....
>>>
>>> /* After this function will be
>>>    executed the following thing
>>>    can happen ------------------>  start to update page table
>>> */                                 entries which was partially     
>>> spin_unlock(&xen_pt_lock);         created during CPU1 but
>>> CPU2       
>>> ....                               doesn't know about them
>>> yet        
>>> ....                               because flush_tlb() ( sfence.vma
>>> ) 
>>> ....                               wasn't done      
>>> ....                                                               
>>>    
>>> flush_tlb_range_va();
>>
>> Not exactly: CPU2 knows about them as far as the memory used /
>> modified
>> goes, and that's all that matters for further page table
>> modifications.
>> CPU2 only doesn't know about the new page table entries yet when it
>> comes
>> to using them for a translation (by the hardware page walker). Yet
>> this
>> aspect is irrelevant here, if I'm not mistaken.
> And it isn't an issue that CPU2 will add these new page table entries
> again during execution of CPU2's pt_update()?

Add these page table entries again? That's only going to happen due to
another bug somewhere, I suppose. And it would be as much (or as little)
of an issue if that happened right after dropping the lock.

Jan



 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.