Allow on-access pruning to set pages all-visible

Many queries do not modify the underlying relation. For such queries, if
on-access pruning occurs during the scan, we can check whether the page
has become all-visible and update the visibility map accordingly.
Previously, only vacuum and COPY FREEZE marked pages as all-visible or
all-frozen.

This commit implements on-access VM setting for sequential scans, tid
range scans, sample scans, bitmap heap scans, and the underlying heap
relation in index scans.

Setting the visibility map on-access can avoid write amplification
caused by vacuum later needing to set the page all-visible, which could
trigger a write and potentially an FPI. It also allows more frequent
index-only scans, since they require pages to be marked all-visible in
the VM.

Author: Melanie Plageman <melanieplageman@gmail.com>
Reviewed-by: Andres Freund <andres@anarazel.de>
Reviewed-by: Kirill Reshke <reshkekirill@gmail.com>
Reviewed-by: Chao Li <li.evan.chao@gmail.com>
Discussion: https://postgr.es/m/flat/CAAKRu_ZMw6Npd_qm2KM%2BFwQ3cMOMx1Dh3VMhp8-V7SOLxdK9-g%40mail.gmail.com
This commit is contained in:
Melanie Plageman 2026-03-30 15:47:07 -04:00
parent e3637a05dc
commit b46e1e54d0
5 changed files with 59 additions and 21 deletions

View file

@ -633,7 +633,8 @@ heap_prepare_pagescan(TableScanDesc sscan)
/*
* Prune and repair fragmentation for the whole page, if possible.
*/
heap_page_prune_opt(scan->rs_base.rs_rd, buffer, &scan->rs_vmbuffer);
heap_page_prune_opt(scan->rs_base.rs_rd, buffer, &scan->rs_vmbuffer,
sscan->rs_flags & SO_HINT_REL_READ_ONLY);
/*
* We must hold share lock on the buffer content while examining tuple

View file

@ -149,7 +149,8 @@ heapam_index_fetch_tuple(struct IndexFetchTableData *scan,
*/
if (prev_buf != hscan->xs_cbuf)
heap_page_prune_opt(hscan->xs_base.rel, hscan->xs_cbuf,
&hscan->xs_vmbuffer);
&hscan->xs_vmbuffer,
hscan->xs_base.flags & SO_HINT_REL_READ_ONLY);
}
/* Obtain share-lock on the buffer so we can examine visibility */
@ -2546,7 +2547,8 @@ BitmapHeapScanNextBlock(TableScanDesc scan,
/*
* Prune and repair fragmentation for the whole page, if possible.
*/
heap_page_prune_opt(scan->rs_rd, buffer, &hscan->rs_vmbuffer);
heap_page_prune_opt(scan->rs_rd, buffer, &hscan->rs_vmbuffer,
scan->rs_flags & SO_HINT_REL_READ_ONLY);
/*
* We must hold share lock on the buffer content while examining tuple

View file

@ -44,6 +44,8 @@ typedef struct
bool mark_unused_now;
/* whether to attempt freezing tuples */
bool attempt_freeze;
/* whether to attempt setting the VM */
bool attempt_set_vm;
struct VacuumCutoffs *cutoffs;
Relation relation;
@ -75,7 +77,8 @@ typedef struct
/*
* set_all_visible and set_all_frozen indicate if the all-visible and
* all-frozen bits in the visibility map can be set for this page after
* pruning.
* pruning. They are only tracked when the caller requests VM updates
* (attempt_set_vm); otherwise they remain false throughout.
*
* NOTE: set_all_visible and set_all_frozen initially don't include
* LP_DEAD items. That's convenient for heap_page_prune_and_freeze() to
@ -232,7 +235,8 @@ static void page_verify_redirects(Page page);
static bool heap_page_will_freeze(bool did_tuple_hint_fpi, bool do_prune, bool do_hint_prune,
PruneState *prstate);
static bool heap_page_will_set_vm(PruneState *prstate, PruneReason reason);
static bool heap_page_will_set_vm(PruneState *prstate, PruneReason reason,
bool do_prune, bool do_freeze);
/*
@ -251,9 +255,21 @@ static bool heap_page_will_set_vm(PruneState *prstate, PruneReason reason);
* reuse the pin across calls, avoiding repeated pin/unpin cycles. If we find
* VM corruption during pruning, we will fix it. Caller is responsible for
* unpinning *vmbuffer.
*
* rel_read_only is true if we determined at plan time that the query does not
* modify the relation. It is counterproductive to set the VM if the query
* will immediately clear it.
*
* As noted in ScanRelIsReadOnly(), INSERT ... SELECT from the same table will
* report the scan relation as read-only. This is usually harmless in
* practice. It is useful to set scanned pages all-visible that won't be
* inserted into. Pages it does insert to will rarely meet the criteria for
* pruning, and those that do are likely to contain in-progress inserts which
* make the page not fully all-visible.
*/
void
heap_page_prune_opt(Relation relation, Buffer buffer, Buffer *vmbuffer)
heap_page_prune_opt(Relation relation, Buffer buffer, Buffer *vmbuffer,
bool rel_read_only)
{
Page page = BufferGetPage(buffer);
TransactionId prune_xid;
@ -336,6 +352,8 @@ heap_page_prune_opt(Relation relation, Buffer buffer, Buffer *vmbuffer)
* current implementation.
*/
params.options = HEAP_PAGE_PRUNE_ALLOW_FAST_PATH;
if (rel_read_only)
params.options |= HEAP_PAGE_PRUNE_SET_VM;
heap_page_prune_and_freeze(&params, &presult, &dummy_off_loc,
NULL, NULL);
@ -392,6 +410,7 @@ prune_freeze_setup(PruneFreezeParams *params,
/* cutoffs must be provided if we will attempt freezing */
Assert(!(params->options & HEAP_PAGE_PRUNE_FREEZE) || params->cutoffs);
prstate->attempt_freeze = (params->options & HEAP_PAGE_PRUNE_FREEZE) != 0;
prstate->attempt_set_vm = (params->options & HEAP_PAGE_PRUNE_SET_VM) != 0;
prstate->cutoffs = params->cutoffs;
prstate->relation = params->relation;
prstate->block = BufferGetBlockNumber(params->buffer);
@ -461,14 +480,13 @@ prune_freeze_setup(PruneFreezeParams *params,
* We track whether the page will be all-visible/all-frozen at the end of
* pruning and freezing. While examining tuple visibility, we'll set
* set_all_visible to false if there are tuples on the page not visible to
* all running and future transactions. set_all_visible is always
* maintained but only VACUUM will set the VM if the page ends up being
* all-visible.
* all running and future transactions. If setting the VM is enabled for
* this scan, we will do so if the page ends up being all-visible.
*
* We also keep track of the newest live XID, which is used to calculate
* the snapshot conflict horizon for a WAL record setting the VM.
*/
prstate->set_all_visible = true;
prstate->set_all_visible = prstate->attempt_set_vm;
prstate->newest_live_xid = InvalidTransactionId;
/*
@ -477,7 +495,9 @@ prune_freeze_setup(PruneFreezeParams *params,
* caller passed HEAP_PAGE_PRUNE_FREEZE, because if they did not, we won't
* call heap_prepare_freeze_tuple() for each tuple, and set_all_frozen
* will never be cleared for tuples that need freezing. This would lead to
* incorrectly setting the visibility map all-frozen for this page.
* incorrectly setting the visibility map all-frozen for this page. We
* can't set the page all-frozen in the VM if the caller didn't pass
* HEAP_PAGE_PRUNE_SET_VM.
*
* When freezing is not required (no XIDs/MXIDs older than the freeze
* cutoff), we may still choose to "opportunistically" freeze if doing so
@ -494,7 +514,7 @@ prune_freeze_setup(PruneFreezeParams *params,
* whether to freeze, but before updating the VM, to avoid setting the VM
* bits incorrectly.
*/
prstate->set_all_frozen = prstate->attempt_freeze;
prstate->set_all_frozen = prstate->attempt_freeze && prstate->attempt_set_vm;
}
/*
@ -920,21 +940,34 @@ heap_page_fix_vm_corruption(PruneState *prstate, OffsetNumber offnum,
* This function does not actually set the VM bits or page-level visibility
* hint, PD_ALL_VISIBLE.
*
* This should be called only after do_freeze has been decided (and do_prune
* has been set), as these factor into our heuristic-based decision.
*
* Returns true if one or both VM bits should be set and false otherwise.
*/
static bool
heap_page_will_set_vm(PruneState *prstate, PruneReason reason)
heap_page_will_set_vm(PruneState *prstate, PruneReason reason,
bool do_prune, bool do_freeze)
{
/*
* Though on-access pruning maintains prstate->set_all_visible, we don't
* set the VM on-access for now.
*/
if (reason == PRUNE_ON_ACCESS)
if (!prstate->attempt_set_vm)
return false;
if (!prstate->set_all_visible)
return false;
/*
* If this is an on-access call and we're not actually pruning, avoid
* setting the visibility map if it would newly dirty the heap page or, if
* the page is already dirty, if doing so would require including a
* full-page image (FPI) of the heap page in the WAL.
*/
if (reason == PRUNE_ON_ACCESS && !do_prune && !do_freeze &&
(!BufferIsDirty(prstate->buffer) || XLogCheckBufferNeedsBackup(prstate->buffer)))
{
prstate->set_all_visible = prstate->set_all_frozen = false;
return false;
}
prstate->new_vmbits = VISIBILITYMAP_ALL_VISIBLE;
if (prstate->set_all_frozen)
@ -1165,9 +1198,10 @@ heap_page_prune_and_freeze(PruneFreezeParams *params,
prstate.set_all_visible = prstate.set_all_frozen = false;
Assert(!prstate.set_all_frozen || prstate.set_all_visible);
Assert(!prstate.set_all_visible || prstate.attempt_set_vm);
Assert(!prstate.set_all_visible || (prstate.lpdead_items == 0));
do_set_vm = heap_page_will_set_vm(&prstate, params->reason);
do_set_vm = heap_page_will_set_vm(&prstate, params->reason, do_prune, do_freeze);
/*
* new_vmbits should be 0 regardless of whether or not the page is

View file

@ -2021,7 +2021,7 @@ lazy_scan_prune(LVRelState *vacrel,
.buffer = buf,
.vmbuffer = vmbuffer,
.reason = PRUNE_VACUUM_SCAN,
.options = HEAP_PAGE_PRUNE_FREEZE,
.options = HEAP_PAGE_PRUNE_FREEZE | HEAP_PAGE_PRUNE_SET_VM,
.vistest = vacrel->vistest,
.cutoffs = &vacrel->cutoffs,
};

View file

@ -43,6 +43,7 @@
#define HEAP_PAGE_PRUNE_MARK_UNUSED_NOW (1 << 0)
#define HEAP_PAGE_PRUNE_FREEZE (1 << 1)
#define HEAP_PAGE_PRUNE_ALLOW_FAST_PATH (1 << 2)
#define HEAP_PAGE_PRUNE_SET_VM (1 << 3)
typedef struct BulkInsertStateData *BulkInsertState;
typedef struct GlobalVisState GlobalVisState;
@ -431,7 +432,7 @@ extern TransactionId heap_index_delete_tuples(Relation rel,
/* in heap/pruneheap.c */
extern void heap_page_prune_opt(Relation relation, Buffer buffer,
Buffer *vmbuffer);
Buffer *vmbuffer, bool rel_read_only);
extern void heap_page_prune_and_freeze(PruneFreezeParams *params,
PruneFreezeResult *presult,
OffsetNumber *off_loc,