lkml.org 
[lkml]   [2021]   [Sep]   [1]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: [PATCH v2 0/9] Free user PTE page table pages
Date
On 19.08.21 05:18, Qi Zheng wrote:
> Hi,
>
> This patch series aims to free user PTE page table pages when all PTE entries
> are empty.
>
> The beginning of this story is that some malloc libraries(e.g. jemalloc or
> tcmalloc) usually allocate the amount of VAs by mmap() and do not unmap those VAs.
> They will use madvise(MADV_DONTNEED) to free physical memory if they want.
> But the page tables do not be freed by madvise(), so it can produce many
> page tables when the process touches an enormous virtual address space.
>
> The following figures are a memory usage snapshot of one process which actually
> happened on our server:
>
> VIRT: 55t
> RES: 590g
> VmPTE: 110g
>
> As we can see, the PTE page tables size is 110g, while the RES is 590g. In
> theory, the process only need 1.2g PTE page tables to map those physical
> memory. The reason why PTE page tables occupy a lot of memory is that
> madvise(MADV_DONTNEED) only empty the PTE and free physical memory but
> doesn't free the PTE page table pages. So we can free those empty PTE page
> tables to save memory. In the above cases, we can save memory about 108g(best
> case). And the larger the difference between the size of VIRT and RES, the
> more memory we save.
>
> In this patch series, we add a pte_refcount field to the struct page of page
> table to track how many users of PTE page table. Similar to the mechanism of
> page refcount, the user of PTE page table should hold a refcount to it before
> accessing. The PTE page table page will be freed when the last refcount is
> dropped.
>
> Testing:
>
> The following code snippet can show the effect of optimization:
>
> mmap 50G
> while (1) {
> for (; i < 1024 * 25; i++) {
> touch 2M memory
> madvise MADV_DONTNEED 2M
> }
> }
>
> As we can see, the memory usage of VmPTE is reduced:
>
> before after
> VIRT 50.0 GB 50.0 GB
> RES 3.1 MB 3.6 MB
> VmPTE 102640 kB 248 kB
>
> I also have tested the stability by LTP[1] for several weeks. I have not seen
> any crash so far.
>
> The performance of page fault can be affected because of the allocation/freeing
> of PTE page table pages. The following is the test result by using a micro
> benchmark[2]:
>
> root@~# perf stat -e page-faults --repeat 5 ./multi-fault $threads:
>
> threads before (pf/min) after (pf/min)
> 1 32,085,255 31,880,833 (-0.64%)
> 8 101,674,967 100,588,311 (-1.17%)
> 16 113,207,000 112,801,832 (-0.36%)
>
> (The "pfn/min" means how many page faults in one minute.)
>
> The performance of page fault is ~1% slower than before.
>
> This series is based on next-20210812.
>
> Comments and suggestions are welcome.
>
> Thanks,
> Qi.
>


Some high-level feedback after studying the code:

1. Try introducing the new dummy primitives ("API") first, and then
convert each subsystem individually; especially, maybe convert the whole
pagefault handling in a single patch, because it's far from trivial.
This will make this series much easier to digest.

Then, have a patch that adds actual logic to the dummy primitives via a
config option.

2. Minimize the API.

a) pte_alloc_get{,_map,_map_lock}() is really ugly. Maybe restrict it to
pte_alloc_get().

b) pmd_trans_unstable_or_pte_try_get() and friends are really ugly.

Handle it independently for now, even if it implies duplicate runtime
checks.

if (pmd_trans_unstable() || !pte_try_get()) ...

We can always optimize later, once we can come up with something cleaner.

3. Merge #6, and #7, after factoring out all changes to other subsystems
to use the API

4. Merge #8 into #6. There is a lot of unnecessary code churn back and
forth, and IMHO the whole approach might not make sense without RCU due
to the additional locking overhead.

Or at least, try to not modify the API you introduced in patch #6 or #7
in #8 again. Converting all call sites back and forth just makes review
quite hard.


I am preparing some some cleanups that will make get_locked_pte() and
similar a little nicer to handle. I'll send them out this or next week.

--
Thanks,

David / dhildenb

\
 
 \ /
  Last update: 2021-09-01 14:37    [W:0.319 / U:0.148 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site