On 4/4/22 17:31, Muchun Song wrote: > On Mon, Apr 4, 2022 at 5:25 PM Anshuman Khandual > <anshuman.khandual@xxxxxxx> wrote: >> >> Hello Muchun, >> >> On 3/31/22 12:26, Muchun Song wrote: >>> The feature of minimizing overhead of struct page associated with each >>> HugeTLB page aims to free its vmemmap pages (used as struct page) to >>> save memory, where is ~14GB/16GB per 1TB HugeTLB pages (2MB/1GB type). >> >> Enabling this feature saves us around 1.4/1.6 % memory but looking from >> other way around, unavailability of vmemmap backing pages (~1.4GB) when >> freeing up a corresponding HugeTLB page, could prevent ~1TB memory from >> being used as normal page form (requiring their own struct pages), thus >> forcing the HugeTLB page to remain as such ? Is not this problematic ? >> >> These additional 1TB memory in normal pages, from a HugeTLB dissolution >> could have eased the system's memory pressure without this feature being >> enabled. > > You are right. If the system is already under heavy memory pressure, it could > prevent the user from freeing HugeTLB pages to the buddy allocator. If the > HugeTLB page are allocated from non-movable zone, this scenario may be > not problematic since once a HugeTLB page is freed, then the system will But how can even the first HugeTLB page be freed without vmemmmap which is throttled due to lack of sufficient memory ? > have memory to be allocated to be used as vmemmap pages, subsequent > freeing of HugeTLB pages may be getting easier. However, if the HUgeTLB > pages are allocated from the movable zone, then the thing becomes terrible, > which is documented in Documentation/admin-guide/mm/memory-hotplug.rst. > > So there is a cmdline "hugetlb_free_vmemmap" to control if enabling this > feature. The user should enable/disable this depending on their workload. Should there also be a sysfs interface for this knob as well ? Perhaps the system usage might change on the way, without requiring a reboot.
