fix: AIX memory statistics: add Available field and improve swap calculations

[Dylan-M]

Summary

Fixes and improves AIX memory statistics for the nocgo path:

Memory (callSVMon)

• Available field: Populated from the real kernel-computed available column in svmon -G -O unit=KB, which includes free pages plus reclaimable cache. This value cannot be derived from raw page counts alone.
• Precision preserved: Primary stats (Total, Used, Free, swap) are read from raw svmon -G page output and multiplied by os.Getpagesize(), avoiding the sub-KB truncation that -O unit=KB introduces. A second svmon call with -O unit=KB is made only to read the available column.
• Swap calculations fixed: swap.Used is now explicitly set, and swap.UsedPercent is calculated.
• Parsing extracted for testability: parseSVMonPages() and parseSVMonAvailable() are separate functions with table-driven unit tests using real AIX sample output.

Swap Devices (SwapDevicesWithContext)

• GB/TB unit handling: parseLspsSize() now handles MB, GB, and TB suffixes in lsps -a output.
• Unit tests: parseLspsOutput() and parseLspsSize() have comprehensive tests covering MB, GB, NaNQ (NFS paging spaces), multiple devices, and edge cases.

CGO path (mem_aix_cgo.go)

• SwapDevicesWithContext: Implemented using perfstat.PagingSpaceStat().

Only affects AIX — no cross-platform changes.

[shirou]

Hi @Dylan-M, I've reviewed this PR more carefully. Unlike your other PRs, this one is relatively small, has a single concern, and is focused only on AIX — which made it much easier to review thoroughly. I appreciate that.

I'm happy to merge this with the following changes:

1. Handle GB (and potentially other units) in lsps -a parsing: The current code assumes the size column always ends with "MB", but lsps can report sizes in GB on larger systems. Please handle at least MB and GB.
2. Add a unit test for SwapDevicesWithContext (nocgo path): The lsps -a parsing logic should have a test with sample output, similar to how other packages test their command parsing.

These are small fixes — once addressed, this is good to merge. Thanks!

[Dylan-M]

Hi @Dylan-M, I've reviewed this PR more carefully. Unlike your other PRs, this one is relatively small, has a single concern, and is focused only on AIX — which made it much easier to review thoroughly. I appreciate that.

All of the rest should be resubmitted soon, in atomic PR "stacks".

I'm happy to merge this with the following changes:

1. Handle GB (and potentially other units) in `lsps -a` parsing: The current code assumes the size column always ends with "MB", but `lsps` can report sizes in GB on larger systems. Please handle at least MB and GB.

Thanks for catching that, none of my test systems have had larger. I've now updated it to match as high as AIX kernel docs show (TB). If we ever run into a case of needed higher than TB, we can revisit.

2. Add a unit test for `SwapDevicesWithContext` (nocgo path): The `lsps -a` parsing logic should have a test with sample output, similar to how other packages test their command parsing.

Yup, That was just me forgetting to add it. Thank you.

These are small fixes — once addressed, this is good to merge. Thanks!

No, thank you sir. Sorry for being such a pain - it is a lot of work and you were entirely correct it should have been split up better.

[Dylan-M]

@shirou This PR has been updated to address your review feedback:

1. GB/TB unit handling — parseLspsSize() now handles MB, GB, and TB suffixes in lsps -a output.
2. Unit test for nocgo path — Added mem_aix_nocgo_test.go with tests for parseLspsOutput() and parseLspsSize(), covering MB, GB, NaNQ (NFS paging spaces), and edge cases.
3. Bonus: real Available memory — Switched from svmon -G (raw pages, no available column) to svmon -G -O unit=KB which provides a real available value distinct from free.

Ready for re-review when you have a moment. Thanks!

[shirou]

unit=KB in callSVMon

Adding -O unit=KB to svmon causes it to convert from pages to KB internally, which truncates any sub-KB fractions. This loses precision compared to the current approach of reading raw page counts and multiplying by the page size.

I'd suggest keeping the default page-based output and replacing the hardcoded pagesize := uint64(4096) with os.Getpagesize() instead. The Available field addition and swap calculation improvements can be applied independently of this unit change.

PR description doesn't match implementation

The PR description says "The Available field is now explicitly set equal to Free memory", but the implementation actually parses the available column (p[6]) directly from svmon output — which is the correct approach. Please update the PR description to reflect what the code actually does.

Missing parse tests for callSVMon

There are good tests for parseLspsOutput, but the svmon output parsing logic in callSVMon has no corresponding test. Given the column index changes (now reading p[6]) and the unit discussion, it would be valuable to extract the parsing logic and add table-driven tests with sample svmon output, similar to what was done for lsps.

[Dylan-M]

@shirou All three points addressed:

1. Precision: Switched back to raw svmon -G page output with os.Getpagesize() for all primary stats (Total, Used, Free, swap). A separate svmon -G -O unit=KB call is made only to read the available column, which doesn't exist in the raw page output. This preserves full precision while still providing real available memory.

2. PR description: Updated to accurately reflect the implementation — available is read from the kernel-computed column, not set equal to free.

3. Parse tests: Extracted parseSVMonPages() and parseSVMonAvailable() as separate testable functions with table-driven tests using sample AIX output, matching the pattern used for parseLspsOutput().

[shirou]

Overall looks good. I'd like to see tightened before merge.

[Dylan-M]

Tested empirically on AIX 7.3.0.0 — across 30 sequential samples with a 500ms cadence, Available > Free held strictly with a stable gap of ~253 MB (range 252.5–254.6 MB). The kernel maintains a non-trivial reclaimable cache footprint in normal operation, so the equality case isn't hit in practice.

The relaxation was appropriate for an earlier iteration of this PR where Available was set equal to Free as a fallback, but became stale once the parsing switched to reading the kernel's real available value. Tightened back to > and dropped the AIX-specific note. Pushed.

[Dylan-M]

@shirou — going to follow your guidance and close out the AIX stack to restart with the one-PR-at-a-time discipline you asked for. The other 20+ PRs and the tracking issue close today.

One carve-out question on this PR specifically: it's been through three rounds of your review, the changes you asked for in the last round are in, and the diff is strictly AIX-only (mem/mem_aix*.go plus a single test-assertion edit). Would you prefer I close this one too and re-submit as the first PR under the new pattern, or keep it open since the review investment is already largely there?