[RFC] Add GNU_PROPERTY_1_NEEDED_THP_ON_[READ_ONLY_SEGMENTS|MALLOC]

[Jan Beulich]

On 20.04.2026 09:55, H.J. Lu wrote:
> On Mon, Apr 20, 2026 at 1:56 PM Jan Beulich <jbeulich@suse.com> wrote:
>> On 18.04.2026 01:24, H.J. Lu wrote:
>>> On Fri, Apr 17, 2026 at 9:36 PM Jan Beulich <jbeulich@suse.com> wrote:
>>>> On 17.04.2026 10:42, H.J. Lu wrote:
>>>>> Transparent Huge Pages (THP) has been enabled in glibc to use THP on
>>>>> THP eligible read-only segments and malloc to improve application
>>>>> performance.  But there are some common issues & trade-offs of THP:
>>>>>
>>>>> 1. Latency Spikes: The background process (khugepaged) that consolidates
>>>>> pages can cause temporary system freezes or high CPU usage.
>>>>> 2. Memory Waste: Using a 2MB page for only 4KB of data can lead to
>>>>> internal fragmentation and higher memory consumption.
>>>>> 3. Database Incompatibility: Many databases (e.g., Oracle, MongoDB,
>>>>> SAP ASE) recommend disabling THP because their non-contiguous access
>>>>> patterns often trigger performance degradations.
>>>>>
>>>>> THP shouldn't be always enabled on THP eligible read-only segments nor
>>>>> in malloc automatically regardless if THP usage will improve or hurt
>>>>> application performance:
>>>>>
>>>>> https://sourceware.org/bugzilla/show_bug.cgi?id=34081
>>>>>
>>>>> To enable developers per-application control over THP usage under madvise
>>>>> THP mode, add the following bits to GNU_PROPERTY_1_NEEDED:
>>>>>
>>>>> 1. GNU_PROPERTY_1_NEEDED_THP_ON_READ_ONLY_SEGMENTS
>>>>>
>>>>>  #define GNU_PROPERTY_1_NEEDED_THP_ON_READ_ONLY_SEGMENTS (1U << 1)
>>>>>
>>>>> to opt-in THP usage on read-only segments in an application.  When this
>>>>> bit is set on executable, run-time loader should call madvise with
>>>>> MADV_HUGEPAGE on all Transparent Huge Pages (THP) eligible read-only
>>>>> segments in executable and its loaded shared libraries after they are
>>>>> mapped into memory under madvise THP mode.
>>>>
>>>> Why would one (the main) executable control behavior for all involved
>>>> shared libraries? Shouldn't each binary get its own say?
>>>
>>> Normally library developers don't know if THP usage in the library will
>>> improve or hurt application performance.
>>
>> "Normally" as in what exactly? An entirely private shared library may be
>> used by both an application wanting THP and one that doesn't. It
>> containing, say, huge tables of data, it may very well benefit from THP
>> itself, irrespective of the application's choice.
> 
> Application developers should evaluate the overall impact of THP on the
> application to decide if THP should be used for it.

And that impact may also depend on the data the application acts upon. Plus,
as said, the developers of a shared library may be in a better position to
do such an evaluation for that library. And the "overall impact" may also
be an entirely abstract thing: E.g. the set of shared libraries loaded may
also vary across runs. I continue to think that this is too coarse grained.
Even a per-binary bit may be too little, as different segments may want
different settings.

>>>>> 2. GNU_PROPERTY_1_NEEDED_THP_ON_MALLOC
>>>>>
>>>>>  #define GNU_PROPERTY_1_NEEDED_THP_ON_MALLOC (1U << 2)
>>>>>
>>>>> to opt-in THP usage in malloc in an application.  When this bit is set on
>>>>> executable, malloc should call madvise with MADV_HUGEPAGE if the
>>>>> allocated size is greater than or equal to THP page size after calling
>>>>> mmap to allocate memory under madvise THP mode.
>>>>
>>>> Things are yet more fuzzy here, as you can't really control this by
>>>> origin of a call to an allocation function.
>>>
>>> This will be checked by malloc in glibc to decide if madvise should be
>>> called.
>>
>> "This" being what here? The new property? If so, then yes, sure, but how
>> does that relate to the point I made?
> 
> The malloc implementation in glibc checks the
> GNU_PROPERTY_1_NEEDED_THP_ON_MALLOC bit
> on the application.  The malloc implementation will use
> madvise if the bit is set.

IOW again "one bit controls everything".

Jan