RFC-0011: zx_object_get_info ZX_INFO_KMEM_STATS_EXTENDED

Summary

This RFC proposes adding a new topic - ZX_INFO_KMEM_STATS_EXTENDED, to the zx_object_get_info() syscall. The new topic will return a single record of type zx_info_kmem_stats_extended_t, a struct consisting of all the fields in zx_info_kmem_stats_t, plus some additional fields that are more expensive to collect.

Motivation

The ZX_INFO_KMEM_STATS topic does not expose any metrics for the amount of memory that can be reclaimed by the kernel under memory pressure. There is currently only a counter for free_bytes, which is the amount of physical free memory on the system. This number alone is not very useful, and can be misleading. In practice the amount of memory "available" is more than free memory, because the kernel can reclaim memory by evicting pages under memory pressure.

Exposing metrics for the various kinds of available memory on the system will allow for more useful memory diagnostics.

Design

The zx_info_kmem_stats_extended_t struct contains all of the fields present in zx_info_kmem_stats_t and the following additional fields:

    // The amount of memory committed to pager-backed VMOs.
    uint64_t vmo_pager_total_bytes;

    // The amount of memory committed to pager-backed VMOs, that has been most
    // recently accessed, and would not be eligible for eviction by the kernel
    // under memory pressure.
    uint64_t vmo_pager_newest_bytes;

    // The amount of memory committed to pager-backed VMOs, that has been least
    // recently accessed, and would be the first to be evicted by the kernel
    // under memory pressure.
    uint64_t vmo_pager_oldest_bytes;

    // The amount of memory committed to discardable VMOs that is currently
    // locked, or unreclaimable by the kernel under memory pressure.
    uint64_t vmo_discardable_locked_bytes;

    // The amount of memory committed to discardable VMOs that is currently
    // unlocked, or reclaimable by the kernel under memory pressure.
    uint64_t vmo_discardable_unlocked_bytes;

The ZX_INFO_KMEM_STATS_EXTENDED topic has the same constraints as ZX_INFO_KMEM_STATS, i.e. it requires the root resource.

The vmo_pager_* fields will be populated by computing the lengths of the pager-backed queues. The vmo_discardable_* fields are currently unimplemented and will be set to 0.

Implementation

The zx_object_get_info() syscall can be extended in a single CL.

Performance

This is a new topic, so the performance of existing callers of zx_object_get_info() will remain unaffected. New callers should choose between using ZX_INFO_KMEM_STATS and ZX_INFO_KMEM_STATS_EXTENDED depending on the level of detail they require. ZX_INFO_KMEM_STATS_EXTENDED provides more detailed metrics and is therefore more expensive, so it should be used only when the additional metrics are required.

Security considerations

None.

Privacy considerations

None.

Testing

Zircon core-tests will be written that query the ZX_INFO_KMEM_STATS_EXTENDED topic.

Documentation

The syscall documentation for zx_object_get_info() will need to be updated.

Drawbacks, alternatives, and unknowns

The alternative to adding a new topic is simply extending the existing ZX_INFO_KMEM_STATS topic to expose the additional metrics. This would involve extending the zx_info_kmem_stats_t struct to include the new fields. The downside with this approach is that existing users of ZX_INFO_KMEM_STATS would take a performance hit, since the new fields that are exposed by ZX_INFO_KMEM_STATS_EXTENDED can be expensive to compute. For example, computing the pager-backed counts is linear in the total number of pages across all VMOs that are backed by the pager. Existing clients of ZX_INFO_KMEM_STATS would end up unnecessarily incurring this additional performance cost for metrics they might not even require. Instead, the ZX_INFO_KMEM_STATS_EXTENDED topic is meant to be used only when clients require this additional level of detail.

Prior art and references

On Linux /proc/meminfo includes counters for MemAvailable, Active and Inactive.