| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | /* | ||
| 2 | * Copyright (c) 2026 Tiger Data, Inc. | ||
| 3 | * Licensed under the PostgreSQL License. See LICENSE for details. | ||
| 4 | * | ||
| 5 | * posting_convert.c - Convert AoS posting pages to fastscan format | ||
| 6 | * | ||
| 7 | * Walks an AoS posting chain, reads pre-encoded entries, and feeds | ||
| 8 | * them to PrismPostingBuilder via _add_encoded(). The builder | ||
| 9 | * handles group packing, page layout, and chain linking. | ||
| 10 | */ | ||
| 11 | |||
| 12 | #include <string.h> | ||
| 13 | |||
| 14 | #include "core/memory.h" | ||
| 15 | #include "index/posting_build.h" | ||
| 16 | #include "index/posting_convert.h" | ||
| 17 | #include "quant/fastscan.h" | ||
| 18 | #include "quant/rabitq.h" | ||
| 19 | |||
| 20 | /* | ||
| 21 | * Entries staged out of the AoS chain before any of them is written back. | ||
| 22 | * | ||
| 23 | * The PG storage layer holds one pinned buffer at a time, so reads and | ||
| 24 | * writes cannot interleave: appending to the builder mid-walk would take | ||
| 25 | * over the buffer the walk is reading from. Staging everything first is what | ||
| 26 | * keeps the two apart. | ||
| 27 | */ | ||
| 28 | typedef struct StagedEntry | ||
| 29 | { | ||
| 30 | ItemPointerData tid; | ||
| 31 | float f_add; | ||
| 32 | float f_rescale; | ||
| 33 | float f_error; | ||
| 34 | } StagedEntry; | ||
| 35 | |||
| 36 | typedef struct StageCtx | ||
| 37 | { | ||
| 38 | Dimension dim; | ||
| 39 | uint32_t packed_bytes; | ||
| 40 | uint32_t total_entries; | ||
| 41 | uint32_t entries_cap; | ||
| 42 | StagedEntry *staged; | ||
| 43 | uint8_t *all_bits; | ||
| 44 | } StageCtx; | ||
| 45 | |||
| 46 | static bool | ||
| 47 | 321 | stage_page(PrismPostingChainPos *pos, void *state) | |
| 48 | { | ||
| 49 | 321 | StageCtx *ctx = state; | |
| 50 | 321 | const PrismPostingPageOpaque *op = prism_posting_opaque(pos->page); | |
| 51 | 321 | char *content = prism_posting_page_content(pos->page, ctx->dim); | |
| 52 | |||
| 53 | 321 | prism_posting_check_count(pos->blkno, op, ctx->dim); | |
| 54 | |||
| 55 |
3/3✓ Branch 0 taken 44 times.
✓ Branch 1 taken 3848 times.
✓ Branch 2 taken 319 times.
|
4211 | for (uint32_t i = 0; i < op->entry_count; i++) |
| 56 | { | ||
| 57 | 3846 | PrismPostingEntryHeader *src = | |
| 58 |
2/2✓ Branch 0 taken 13 times.
✓ Branch 1 taken 3833 times.
|
3890 | prism_posting_entry_at(content, i, ctx->dim); |
| 59 | |||
| 60 | /* | ||
| 61 | * Leave behind the entries VACUUM has marked dead. A fastscan page | ||
| 62 | * packs codes with no per-entry flag -- deletion there is | ||
| 63 | * page-granular -- so a dead entry copied into one comes back as | ||
| 64 | * live and can never be marked again: a later VACUUM can only | ||
| 65 | * tombstone the page once *every* entry on it is dead, which a page | ||
| 66 | * holding live entries never is. The head's live_count, which the | ||
| 67 | * builder stamps from what it was given, would be wrong by the same | ||
| 68 | * number. | ||
| 69 | */ | ||
| 70 |
2/2✓ Branch 0 taken 13 times.
✓ Branch 1 taken 3877 times.
|
3890 | if (src->meta.flags & PRISM_POSTING_FLAG_DELETED) |
| 71 | 13 | continue; | |
| 72 | |||
| 73 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 3877 times.
|
3877 | if (ctx->total_entries >= ctx->entries_cap) |
| 74 | { | ||
| 75 | ✗ | ctx->entries_cap *= 2; | |
| 76 | ✗ | ctx->staged = vs_realloc( | |
| 77 | ✗ | ctx->staged, ctx->entries_cap * sizeof(StagedEntry)); | |
| 78 | ✗ | ctx->all_bits = vs_realloc( | |
| 79 | ✗ | ctx->all_bits, | |
| 80 | ✗ | ctx->entries_cap * (size_t)ctx->packed_bytes); | |
| 81 | } | ||
| 82 | |||
| 83 | 3877 | StagedEntry *dst = &ctx->staged[ctx->total_entries]; | |
| 84 | |||
| 85 | 3877 | dst->tid = src->meta.tid; | |
| 86 | 3877 | dst->f_add = src->f_add; | |
| 87 | 3877 | dst->f_rescale = src->f_rescale; | |
| 88 | 3877 | dst->f_error = src->f_error; | |
| 89 | 3877 | memcpy(ctx->all_bits + (size_t)ctx->total_entries * ctx->packed_bytes, | |
| 90 | 3877 | src->bits, | |
| 91 | 3877 | ctx->packed_bytes); | |
| 92 | 3877 | ctx->total_entries++; | |
| 93 | } | ||
| 94 | |||
| 95 | 321 | return true; | |
| 96 | } | ||
| 97 | |||
| 98 | BlockNumber | ||
| 99 | 321 | prism_posting_convert_to_fastscan( | |
| 100 | VsStorage *storage, BlockNumber aos_head, Dimension dim) | ||
| 101 | { | ||
| 102 |
2/2✓ Branch 0 taken 319 times.
✓ Branch 1 taken 2 times.
|
321 | if (aos_head == InvalidBlockNumber) |
| 103 | ✗ | return InvalidBlockNumber; | |
| 104 | |||
| 105 | 321 | uint32_t packed_bytes = VS_RABITQ_BYTES(dim); | |
| 106 | |||
| 107 | /* Read pt_centroid from AoS first page */ | ||
| 108 | 321 | Page first_page = vs_storage_read_page(storage, aos_head); | |
| 109 | 321 | float *pt_centroid = vs_alloc(dim * sizeof(float)); | |
| 110 | 323 | memcpy(pt_centroid, | |
| 111 | 321 | prism_posting_pt_centroid(first_page), | |
| 112 | dim * sizeof(float)); | ||
| 113 | |||
| 114 | 321 | PrismPostingPageOpaque *first_op = prism_posting_opaque(first_page); | |
| 115 | 321 | uint32_t cluster_id = first_op->cluster_id; | |
| 116 | 321 | vs_storage_release_page(storage, aos_head); | |
| 117 | |||
| 118 | 319 | PrismPostingBuilder builder; | |
| 119 | 321 | prism_posting_builder_init_fastscan( | |
| 120 | &builder, storage, NULL, dim, cluster_id, NULL, pt_centroid); | ||
| 121 | |||
| 122 | 321 | StageCtx ctx = { | |
| 123 | .dim = dim, | ||
| 124 | .packed_bytes = packed_bytes, | ||
| 125 | .entries_cap = 256, | ||
| 126 | }; | ||
| 127 | 321 | ctx.staged = vs_alloc(ctx.entries_cap * sizeof(StagedEntry)); | |
| 128 | 321 | ctx.all_bits = vs_alloc(ctx.entries_cap * (size_t)packed_bytes); | |
| 129 | |||
| 130 | 321 | prism_posting_chain_walk(storage, aos_head, stage_page, &ctx); | |
| 131 | |||
| 132 | /* Feed staged entries to the fastscan builder */ | ||
| 133 |
3/3✓ Branch 0 taken 44 times.
✓ Branch 1 taken 3835 times.
✓ Branch 2 taken 319 times.
|
4198 | for (uint32_t i = 0; i < ctx.total_entries; i++) |
| 134 | { | ||
| 135 | 3877 | prism_posting_builder_add_encoded( | |
| 136 | &builder, | ||
| 137 | 44 | ctx.staged[i].tid, | |
| 138 | 44 | ctx.staged[i].f_add, | |
| 139 | 44 | ctx.staged[i].f_rescale, | |
| 140 | 3877 | ctx.staged[i].f_error, | |
| 141 | 3877 | ctx.all_bits + (size_t)i * packed_bytes); | |
| 142 | } | ||
| 143 | |||
| 144 | 321 | vs_free(ctx.staged); | |
| 145 | 321 | vs_free(ctx.all_bits); | |
| 146 | |||
| 147 | 321 | BlockNumber result = prism_posting_builder_finish(&builder); | |
| 148 | 321 | prism_posting_builder_cleanup(&builder); | |
| 149 | 321 | vs_free(pt_centroid); | |
| 150 | |||
| 151 | 321 | return result; | |
| 152 | } | ||
| 153 |