Garbage Collection in Flash Storage

NAND updates do not simply overwrite the old physical page. A controller can place the new version elsewhere, leaving an obsolete page behind. Before reusing an erase block, it preserves any remaining current pages in other space and erases the victim. An invalid page is therefore not yet an independently reusable erased page. [2]

Reclamation needs headroom: the live data has to fit somewhere before its source block is erased. Selecting a victim with fewer live pages reduces copying for that cleaning event. Spare capacity and placement affect the available choices, but no single free-space percentage guarantees a particular latency. [1]

GC is not necessarily an idle-time job. Kingston describes specific SandForce-based SSDs using foreground collection without background collection. That is a documented implementation choice, not proof that foreground cleaning always wins. Scheduling decides when work competes with requests; it does not remove the work. [2]

TRIM supplies information that selected logical ranges are no longer needed. When the stack conveys and honors that information, a collector can avoid preserving their old contents. The notification does not require immediate physical erasure. Free space reported by a filesystem and erased space available inside a device are different observations. [3]

Nor is an apparently empty logical range evidence of secure removal. Wei and colleagues distinguish host-visible addresses from physical flash copies and warn that TRIM is not a reliable security guarantee. Their 2011 device measurements are historical evidence, not recovery rates for current SSDs. [4]

Garbage collection and wear leveling have different objectives: recover reusable space versus distribute wear. A policy that repeatedly chooses easy-to-clean hot blocks can leave cold blocks comparatively untouched. Efficient reclamation does not by itself demonstrate balanced wear. [1]

The accounting below is Hesela's hypothetical example, not a device benchmark. Assume one fully programmed block of N equal-sized pages, V still valid. Each valid page is copied once into already-erased destination space, then the victim is successfully erased. Ignore metadata, parity, failures and all other traffic.

Erasure makes N victim pages free, but relocation consumed V free destination pages. The net gain is N - V pages. Copying work per net reclaimed page is V / (N - V), defined only for V < N. This denominator is reclaimed space, not host writes: the ratio is not whole-device write amplification.

With no valid pages, no live-page copy is needed, although erasure still is. With all pages valid, moving them and erasing the victim gives zero net free-space gain in this model. That boundary helps explain why counting erased blocks alone cannot measure useful reclamation.

When comparing implementations, record workload, fill history, discard delivery, spare capacity and whether reported counters represent host writes, physical programming or only copies. Copyback can change the transfer path without making relocation equivalent to zero physical writes; the linked analysis examines that separate boundary.