Message-ID: <11a01f4e-9ae5-4001-9f9c-74a746f898cd@arm.com>
Date: Tue, 20 Jan 2026 17:35:50 +0000
From: Ryan Roberts <ryan.roberts@....com>
To: Yeoreum Yun <yeoreum.yun@....com>
Cc: Will Deacon <will@...nel.org>, linux-arm-kernel@...ts.infradead.org,
 linux-kernel@...r.kernel.org, linux-rt-devel@...ts.linux.dev,
 catalin.marinas@....com, akpm@...ux-oundation.org, david@...nel.org,
 kevin.brodsky@....com, quic_zhenhuah@...cinc.com, dev.jain@....com,
 yang@...amperecomputing.com, chaitanyas.prakash@....com,
 bigeasy@...utronix.de, clrkwllms@...nel.org, rostedt@...dmis.org,
 lorenzo.stoakes@...cle.com, ardb@...nel.org, jackmanb@...gle.com,
 vbabka@...e.cz, mhocko@...e.com
Subject: Re: [PATCH v5 2/3] arm64: mmu: avoid allocating pages while splitting
 the linear mapping

On 20/01/2026 16:31, Yeoreum Yun wrote:
> Hi Ryan,
> 
>> On 20/01/2026 15:53, Will Deacon wrote:
>>> On Tue, Jan 20, 2026 at 10:40:30AM +0000, Ryan Roberts wrote:
>>>> On 20/01/2026 09:29, Yeoreum Yun wrote:
>>>>> Hi Ryan
>>>>>> On 19/01/2026 21:24, Yeoreum Yun wrote:
>>>>>>> Hi Will,
>>>>>>>
>>>>>>>> On Mon, Jan 05, 2026 at 08:23:27PM +0000, Yeoreum Yun wrote:
>>>>>>>>> +static int __init linear_map_prealloc_split_pgtables(void)
>>>>>>>>> +{
>>>>>>>>> +	int ret, i;
>>>>>>>>> +	unsigned long lstart = _PAGE_OFFSET(vabits_actual);
>>>>>>>>> +	unsigned long lend = PAGE_END;
>>>>>>>>> +	unsigned long kstart = (unsigned long)lm_alias(_stext);
>>>>>>>>> +	unsigned long kend = (unsigned long)lm_alias(__init_begin);
>>>>>>>>> +
>>>>>>>>> +	const struct mm_walk_ops collect_to_split_ops = {
>>>>>>>>> +		.pud_entry	= collect_to_split_pud_entry,
>>>>>>>>> +		.pmd_entry	= collect_to_split_pmd_entry
>>>>>>>>> +	};
>>>>>>>>
>>>>>>>> Why do we need to rewalk the page-table here instead of collating the
>>>>>>>> number of block mappings we put down when creating the linear map in
>>>>>>>> the first place?linear_map_maybe_split_to_ptes(
>>>>>>
>>>>>> That's a good point; perhaps we can reuse the counters that this series introduces?
>>>>>>
>>>>>> https://lore.kernel.org/all/20260107002944.2940963-1-yang@os.amperecomputing.com/
>>>>>>
>>>>>>>
>>>>>>> First, linear alias of the [_text, __init_begin) is not a target for
>>>>>>> the split and it also seems strange to me to add code inside alloc_init_XXX()
>>>>>>> that both checks an address range and counts to get the number of block mappings.
>>>>>>>
>>>>>>> Second, for a future feature,
>>>>>>> I hope to add some code to split "specfic" area to be spilt e.x)
>>>>>>> to set a specific pkey for specific area.
>>>>>>
>>>>>> Could you give more detail on this? My working assumption is that either the
>>>>>> system supports BBML2 or it doesn't. If it doesn't, we need to split the whole
>>>>>> linear map. If it does, we already have logic to split parts of the linear map
>>>>>> when needed.
>>>>>
>>>>> This is not for a linear mapping case. but for a "kernel text area".
>>>>> As a draft, I want to mark some of kernel code can executable
>>>>> both kernel and eBPF program.
>>>>> (I'm trying to make eBPF program non-executable kernel code directly
>>>>> with POE feature).
>>>>> For this "executable area" both of kernel and eBPF program
>>>>> -- typical example is exception entry, It need to split that specific
>>>>> range and mark them with special POE index.
>>>>
>>>> Ahh yes, I recall you mentioning this a while back (although I confess all the
>>>> deatils have fallen out of my head). You'd need to make sure you're definitely
>>>> not splitting an area of text that the secondary CPUs are executing while they
>>>> are being held in the pen, since at least one of those CPUs doesn't support BBML2.
>>>>
>>>>>
>>>>>>
>>>>>>>
>>>>>>> In this case, it's useful to rewalk the page-table with the specific
>>>>>>> range to get the number of block mapping.
>>>>>>>
>>>>>>>>
>>>>>>>>> +	split_pgtables_idx = 0;
>>>>>>>>> +	split_pgtables_count = 0;
>>>>>>>>> +
>>>>>>>>> +	ret = walk_kernel_page_table_range_lockless(lstart, kstart,
>>>>>>>>> +						    &collect_to_split_ops,
>>>>>>>>> +						    NULL, NULL);
>>>>>>>>> +	if (!ret)
>>>>>>>>> +		ret = walk_kernel_page_table_range_lockless(kend, lend,
>>>>>>>>> +							    &collect_to_split_ops,
>>>>>>>>> +							    NULL, NULL);
>>>>>>>>> +	if (ret || !split_pgtables_count)
>>>>>>>>> +		goto error;
>>>
>>> Just noticed this, but why do we check '!split_pgtables_count' here?
>>> if the page-table is already somehow mapped at page granularity, that
>>> doesn't necessarily sound like a fatal error to me.
>>>
>>>>>>>>> +
>>>>>>>>> +	ret = -ENOMEM;
>>>>>>>>> +
>>>>>>>>> +	split_pgtables = kvmalloc(split_pgtables_count * sizeof(struct ptdesc *),
>>>>>>>>> +				  GFP_KERNEL | __GFP_ZERO);
>>>>>>>>> +	if (!split_pgtables)
>>>>>>>>> +		goto error;
>>>>>>>>> +
>>>>>>>>> +	for (i = 0; i < split_pgtables_count; i++) {
>>>>>>>>> +		/* The page table will be filled during splitting, so zeroing it is unnecessary. */
>>>>>>>>> +		split_pgtables[i] = pagetable_alloc(GFP_PGTABLE_KERNEL & ~__GFP_ZERO, 0);
>>>>>>>>> +		if (!split_pgtables[i])
>>>>>>>>> +			goto error;
>>>>>>>>
>>>>>>>> This looks potentially expensive on the boot path and only gets worse as
>>>>>>>> the amount of memory grows. Maybe we should predicate this preallocation
>>>>>>>> on preempt-rt?
>>>>>>>
>>>>>>> Agree. then I'll apply pre-allocation with PREEMPT_RT only.
>>>>>>
>>>>>> I guess I'm missing something obvious but I don't understand the problem here...
>>>>>> We are only deferring the allocation of all these pgtables, so the cost is
>>>>>> neutral surely? Had we correctly guessed that the system doesn't support BBML2
>>>>>> earlier, we would have had to allocate all these pgtables earlier.
>>>>>>
>>>>>> Another way to look at it is that we are still allocating the same number of
>>>>>> pgtables in the existing fallback path, it's just that we are doing it inside
>>>>>> the stop_machine().
>>>>>>
>>>>>> My vote would be _not_ to have a separate path for PREEMPT_RT, which will end up
>>>>>> with significantly less testing...
>>>>>
>>>>> IIUC, Will's mention is additional memory allocation for
>>>>> "split_pgtables" where saved "pre-allocate" page tables.
>>>>> As the memory increase, definitely this size would increase the cost.
>>>>
>>>> Err, so you're referring to the extra kvmalloc()? I don't think that's a big
>>>> deal is it? you get 512 pointers per page. So the amortized cost is 1/512= 0.2%?
>>>
>>> Right, it was the page-table pages I was worried about not the array of
>>> pointers.
>>>
>>>> I suspect we have both misunderstood Will's point...
>>>
>>> I probably just got confused by linear_map_free_split_pgtables() as it
>>> has logic to free unused page-table pages between 'split_pgtables_idx'
>>> and 'split_pgtables_count', implying that we can over-allocate.
>>>
>>> If that is only needed for the error path in
>>> linear_map_prealloc_split_pgtables(), then perhaps that part should be
>>> inlined to deal with the case where we fail to allocate part way through.
>>
>> I was originally concerned [1] that there could be a race where another CPU
>> caused the normal splitting machinery to kick in after this cpu determined the
>> number of required page tables, so there could be some left over in that case.
>>
>> On reflection, I guess (hope) that's not possible because we've determined that
>> some CPUs don't support BBML2. I'm guessing the secondaries haven't been
>> released to do general work yet?
> 
> I don't think so, since the linear_map_maybe_split_to_ptes() called
> in smp_cpus_done() but in here, secondary cpus already on and
> it seems schedulable.
> 
> That's why although, This is unlikely, after collecting the number of
> splitiing by other cpu have a possibility to *split* which was counted
> and at that time I agreed for your comments because of this *low
> possiblity*.
> 
>>
>> In which case, I agree, this could be simplified and we could just assert that
>> all pre-allocated pages get used up if there is no error?
>>
>> [1] https://lore.kernel.org/all/73ced1db-a2e2-49ea-927e-9fc4a30e771e@arm.com/
> 
> So with above reason, I still think it need to sustain the free
> unused pagetable.
> 
> Am I missing something?

My concern is that if a secondary CPU can race and cause a split, that is
unsound because we have determined that although the primary CPU supports BBML2,
at least one of the secondary CPUs does not. So splitting a live mapping is unsafe.

I just had a brief chat with Rutland, and he agrees that this _could_ be a
problem. Basically there is a window between onlining the secondary cpus and
entering the stop_machine() where one of those cpus _could_ end up doing
something that causes us to split the linear map.

I'm not immediately sure how to solve that.

> 
> --
> Sincerely,
> Yeoreum Yun

Powered by blists - more mailing lists