Cache management under low memory

[Justin Mattson]

When the kernel has signaled a low memory state, how is this managed for filesystems? For example, are in-memory caches trimmed and is the total count of pages of pageable memory clamped at some level until the kernel signals memory conditions have improved?

Thanks,

Justin

[Chris Suter]

Except for the fact that loaded blobs for Blobfs are discardable memory (which is managed by the kernel), I don't think we do anything at the moment. When we land write-back support for mutable filesystems, that'll help (since that memory can be marked discardable too), but that's some way off. It might be a while before we start looking at doing something for metadata caches.

— Chris

--
All posts must follow the Fuchsia Code of Conduct https://fuchsia.dev/fuchsia-src/CODE_OF_CONDUCT or may be removed.
---
You received this message because you are subscribed to the Google Groups "storage-dev" group.
To unsubscribe from this group and stop receiving emails from it, send an email to storage-dev...@fuchsia.dev.
To view this discussion on the web visit https://groups.google.com/a/fuchsia.dev/d/msgid/storage-dev/CA%2BX8dX9r1VDGiVLbttKL5f%3Dev_SLSR%3DuXQaHbXSvdVJBmx7kxQ%40mail.gmail.com.

[Adam Barth]

[ +Étienne J. Membrives +David Moore / Note: Public thread]

Looking at some memory profiles, I noticed zxcrypt-buffer is fairly large:

https://cs.opensource.google/fuchsia/fuchsia/+/main:src/devices/block/drivers/zxcrypt/device-info.cc;l=85

Looks like it's size is initialized to a constant:

https://cs.opensource.google/fuchsia/fuchsia/+/main:src/security/zxcrypt/volume.cc;l=49

I wonder if there's something more sophisticated we should do for sizing that buffer.  I guess I should file an issue:

https://bugs.fuchsia.dev/p/fuchsia/issues/detail?id=92849

Adam

To view this discussion on the web visit https://groups.google.com/a/fuchsia.dev/d/msgid/storage-dev/CABpLAtwMP22dNVp_AES8qJgPF-yawwe5BzK9pLFKXYbLRwYhvQ%40mail.gmail.com.

[Seokhwan Kim]

As a set of CLs for writeback has been merged, f2fs is able to adjust the number of committed pages in a writeback cache. Basically, f2fs uses discardable vmo to present a file block in memory, and it keeps vmo_lock for dirty vmos as well as ones in use. In order to limit the total amount of locked discardable pages, it writes out dirty vmos when the amount exceed a threshold.
If it is notified of the current memory state from kernel, I think we can add a CL to adjust the threshold and the number of locked committed pages according to the state. 

https://fuchsia-review.googlesource.com/c/fuchsia/+/626194

Best regards,

Seokhwan

2022년 2월 3일 목요일 오후 12시 30분 34초 UTC+9에 Adam Barth님이 작성: