Add pg_buffercache_numa view with NUMA node info

Introduces a new view pg_buffercache_numa, showing NUMA memory nodes
for individual buffers. For each buffer the view returns an entry for
each memory page, with the associated NUMA node.

The database blocks and OS memory pages may have different size - the
default block size is 8KB, while the memory page is 4K (on x86). But
other combinations are possible, depending on configure parameters,
platform, etc. This means buffers may overlap with multiple memory
pages, each associated with a different NUMA node.

To determine the NUMA node for a buffer, we first need to touch the
memory pages using pg_numa_touch_mem_if_required, otherwise we might get
status -2 (ENOENT = The page is not present), indicating the page is
either unmapped or unallocated.

The view may be relatively expensive, especially when accessed for the
first time in a backend, as it touches all memory pages to get reliable
information about the NUMA node. This may also force allocation of the
shared memory.

Author: Jakub Wartak <[email protected]>
Reviewed-by: Andres Freund <[email protected]>
Reviewed-by: Bertrand Drouvot <[email protected]>
Reviewed-by: Tomas Vondra <[email protected]>
Discussion: https://postgr.es/m/CAKZiRmxh6KWo0aqRqvmcoaX2jUxZYb4kGp3N%3Dq1w%2BDiH-696Xw%40mail.gmail.com

8 files changed, 366 insertions, 3 deletions

diff --git a/contrib/pg_buffercache/Makefile b/contrib/pg_buffercache/Makefile
index eae65ead9e5..5f748543e2e 100644
--- a/contrib/pg_buffercache/Makefile
+++ b/contrib/pg_buffercache/Makefile
@@ -8,10 +8,11 @@ OBJS = \
 EXTENSION = pg_buffercache
 DATA = pg_buffercache--1.2.sql pg_buffercache--1.2--1.3.sql \
 	pg_buffercache--1.1--1.2.sql pg_buffercache--1.0--1.1.sql \
-	pg_buffercache--1.3--1.4.sql pg_buffercache--1.4--1.5.sql
+	pg_buffercache--1.3--1.4.sql pg_buffercache--1.4--1.5.sql \
+	pg_buffercache--1.5--1.6.sql
 PGFILEDESC = "pg_buffercache - monitoring of shared buffer cache in real-time"
 
-REGRESS = pg_buffercache
+REGRESS = pg_buffercache pg_buffercache_numa
 
 ifdef USE_PGXS
 PG_CONFIG = pg_config
diff --git a/contrib/pg_buffercache/expected/pg_buffercache_numa.out b/contrib/pg_buffercache/expected/pg_buffercache_numa.out
new file mode 100644
index 00000000000..a10b331a552
--- /dev/null
+++ b/contrib/pg_buffercache/expected/pg_buffercache_numa.out
@@ -0,0 +1,29 @@
+SELECT NOT(pg_numa_available()) AS skip_test \gset
+\if :skip_test
+\quit
+\endif
+-- We expect at least one entry for each buffer
+select count(*) >= (select setting::bigint
+                    from pg_settings
+                    where name = 'shared_buffers')
+from pg_buffercache_numa;
+ ?column? 
+----------
+ t
+(1 row)
+
+-- Check that the functions / views can't be accessed by default. To avoid
+-- having to create a dedicated user, use the pg_database_owner pseudo-role.
+SET ROLE pg_database_owner;
+SELECT count(*) > 0 FROM pg_buffercache_numa;
+ERROR:  permission denied for view pg_buffercache_numa
+RESET role;
+-- Check that pg_monitor is allowed to query view / function
+SET ROLE pg_monitor;
+SELECT count(*) > 0 FROM pg_buffercache_numa;
+ ?column? 
+----------
+ t
+(1 row)
+
+RESET role;
diff --git a/contrib/pg_buffercache/expected/pg_buffercache_numa_1.out b/contrib/pg_buffercache/expected/pg_buffercache_numa_1.out
new file mode 100644
index 00000000000..6dd6824b4e4
--- /dev/null
+++ b/contrib/pg_buffercache/expected/pg_buffercache_numa_1.out
@@ -0,0 +1,3 @@
+SELECT NOT(pg_numa_available()) AS skip_test \gset
+\if :skip_test
+\quit
diff --git a/contrib/pg_buffercache/meson.build b/contrib/pg_buffercache/meson.build
index 12d1fe48717..7cd039a1df9 100644
--- a/contrib/pg_buffercache/meson.build
+++ b/contrib/pg_buffercache/meson.build
@@ -23,6 +23,7 @@ install_data(
   'pg_buffercache--1.2.sql',
   'pg_buffercache--1.3--1.4.sql',
   'pg_buffercache--1.4--1.5.sql',
+  'pg_buffercache--1.5--1.6.sql',
   'pg_buffercache.control',
   kwargs: contrib_data_args,
 )
@@ -34,6 +35,7 @@ tests += {
   'regress': {
     'sql': [
       'pg_buffercache',
+      'pg_buffercache_numa',
     ],
   },
 }
diff --git a/contrib/pg_buffercache/pg_buffercache--1.5--1.6.sql b/contrib/pg_buffercache/pg_buffercache--1.5--1.6.sql
new file mode 100644
index 00000000000..f6668e41b37
--- /dev/null
+++ b/contrib/pg_buffercache/pg_buffercache--1.5--1.6.sql
@@ -0,0 +1,22 @@
+/* contrib/pg_buffercache/pg_buffercache--1.5--1.6.sql */
+
+-- complain if script is sourced in psql, rather than via CREATE EXTENSION
+\echo Use "ALTER EXTENSION pg_buffercache UPDATE TO '1.6'" to load this file. \quit
+
+-- Register the new functions.
+CREATE OR REPLACE FUNCTION pg_buffercache_numa_pages()
+RETURNS SETOF RECORD
+AS 'MODULE_PATHNAME', 'pg_buffercache_numa_pages'
+LANGUAGE C PARALLEL SAFE;
+
+-- Create a view for convenient access.
+CREATE VIEW pg_buffercache_numa AS
+	SELECT P.* FROM pg_buffercache_numa_pages() AS P
+	(bufferid integer, os_page_num int4, numa_node int4);
+
+-- Don't want these to be available to public.
+REVOKE ALL ON FUNCTION pg_buffercache_numa_pages() FROM PUBLIC;
+REVOKE ALL ON pg_buffercache_numa FROM PUBLIC;
+
+GRANT EXECUTE ON FUNCTION pg_buffercache_numa_pages() TO pg_monitor;
+GRANT SELECT ON pg_buffercache_numa TO pg_monitor;
diff --git a/contrib/pg_buffercache/pg_buffercache.control b/contrib/pg_buffercache/pg_buffercache.control
index 5ee875f77dd..b030ba3a6fa 100644
--- a/contrib/pg_buffercache/pg_buffercache.control
+++ b/contrib/pg_buffercache/pg_buffercache.control
@@ -1,5 +1,5 @@
 # pg_buffercache extension
 comment = 'examine the shared buffer cache'
-default_version = '1.5'
+default_version = '1.6'
 module_pathname = '$libdir/pg_buffercache'
 relocatable = true
diff --git a/contrib/pg_buffercache/pg_buffercache_pages.c b/contrib/pg_buffercache/pg_buffercache_pages.c
index 62602af1775..a702a47efe9 100644
--- a/contrib/pg_buffercache/pg_buffercache_pages.c
+++ b/contrib/pg_buffercache/pg_buffercache_pages.c
@@ -11,6 +11,7 @@
 #include "access/htup_details.h"
 #include "catalog/pg_type.h"
 #include "funcapi.h"
+#include "port/pg_numa.h"
 #include "storage/buf_internals.h"
 #include "storage/bufmgr.h"
 
@@ -20,6 +21,8 @@
 #define NUM_BUFFERCACHE_SUMMARY_ELEM 5
 #define NUM_BUFFERCACHE_USAGE_COUNTS_ELEM 4
 
+#define NUM_BUFFERCACHE_NUMA_ELEM	3
+
 PG_MODULE_MAGIC_EXT(
 					.name = "pg_buffercache",
 					.version = PG_VERSION
@@ -58,16 +61,44 @@ typedef struct
 	BufferCachePagesRec *record;
 } BufferCachePagesContext;
 
+/*
+ * Record structure holding the to be exposed cache data.
+ */
+typedef struct
+{
+	uint32		bufferid;
+	int32		page_num;
+	int32		numa_node;
+} BufferCacheNumaRec;
+
+/*
+ * Function context for data persisting over repeated calls.
+ */
+typedef struct
+{
+	TupleDesc	tupdesc;
+	int			buffers_per_page;
+	int			pages_per_buffer;
+	int			os_page_size;
+	BufferCacheNumaRec *record;
+} BufferCacheNumaContext;
+
 
 /*
  * Function returning data from the shared buffer cache - buffer number,
  * relation node/tablespace/database/blocknum and dirty indicator.
  */
 PG_FUNCTION_INFO_V1(pg_buffercache_pages);
+PG_FUNCTION_INFO_V1(pg_buffercache_numa_pages);
 PG_FUNCTION_INFO_V1(pg_buffercache_summary);
 PG_FUNCTION_INFO_V1(pg_buffercache_usage_counts);
 PG_FUNCTION_INFO_V1(pg_buffercache_evict);
 
+
+/* Only need to touch memory once per backend process lifetime */
+static bool firstNumaTouch = true;
+
+
 Datum
 pg_buffercache_pages(PG_FUNCTION_ARGS)
 {
@@ -246,6 +277,260 @@ pg_buffercache_pages(PG_FUNCTION_ARGS)
 		SRF_RETURN_DONE(funcctx);
 }
 
+/*
+ * Inquire about NUMA memory mappings for shared buffers.
+ *
+ * Returns NUMA node ID for each memory page used by the buffer. Buffers may
+ * be smaller or larger than OS memory pages. For each buffer we return one
+ * entry for each memory page used by the buffer (it fhe buffer is smaller,
+ * it only uses a part of one memory page).
+ *
+ * We expect both sizes (for buffers and memory pages) to be a power-of-2, so
+ * one is always a multiple of the other.
+ *
+ * In order to get reliable results we also need to touch memory pages, so
+ * that the inquiry about NUMA memory node doesn't return -2 (which indicates
+ * unmapped/unallocated pages).
+ */
+Datum
+pg_buffercache_numa_pages(PG_FUNCTION_ARGS)
+{
+	FuncCallContext *funcctx;
+	MemoryContext oldcontext;
+	BufferCacheNumaContext *fctx;	/* User function context. */
+	TupleDesc	tupledesc;
+	TupleDesc	expected_tupledesc;
+	HeapTuple	tuple;
+	Datum		result;
+
+	if (SRF_IS_FIRSTCALL())
+	{
+		int			i,
+					idx;
+		Size		os_page_size;
+		void	  **os_page_ptrs;
+		int		   *os_page_status;
+		uint64		os_page_count;
+		int			pages_per_buffer;
+		int			max_entries;
+		volatile uint64 touch pg_attribute_unused();
+		char	   *startptr,
+				   *endptr;
+
+		if (pg_numa_init() == -1)
+			elog(ERROR, "libnuma initialization failed or NUMA is not supported on this platform");
+
+		/*
+		 * The database block size and OS memory page size are unlikely to be
+		 * the same. The block size is 1-32KB, the memory page size depends on
+		 * platform. On x86 it's usually 4KB, on ARM it's 4KB or 64KB, but
+		 * there are also features like THP etc. Moreover, we don't quite know
+		 * how the pages and buffers "align" in memory - the buffers may be
+		 * shifted in some way, using more memory pages than necessary.
+		 *
+		 * So we need to be careful about mappping buffers to memory pages. We
+		 * calculate the maximum number of pages a buffer might use, so that
+		 * we allocate enough space for the entries. And then we count the
+		 * actual number of entries as we scan the buffers.
+		 *
+		 * This information is needed before calling move_pages() for NUMA
+		 * node id inquiry.
+		 */
+		os_page_size = pg_numa_get_pagesize();
+
+		/*
+		 * The pages and block size is expected to be 2^k, so one divides the
+		 * other (we don't know in which direction). This does not say
+		 * anything about relative alignment of pages/buffers.
+		 */
+		Assert((os_page_size % BLCKSZ == 0) || (BLCKSZ % os_page_size == 0));
+
+		/*
+		 * How many addresses we are going to query? Simply get the page for
+		 * the first buffer, and first page after the last buffer, and count
+		 * the pages from that.
+		 */
+		startptr = (char *) TYPEALIGN_DOWN(os_page_size,
+										   BufferGetBlock(1));
+		endptr = (char *) TYPEALIGN(os_page_size,
+									(char *) BufferGetBlock(NBuffers) + BLCKSZ);
+		os_page_count = (endptr - startptr) / os_page_size;
+
+		/* Used to determine the NUMA node for all OS pages at once */
+		os_page_ptrs = palloc0(sizeof(void *) * os_page_count);
+		os_page_status = palloc(sizeof(uint64) * os_page_count);
+
+		/* Fill pointers for all the memory pages. */
+		idx = 0;
+		for (char *ptr = startptr; ptr < endptr; ptr += os_page_size)
+		{
+			os_page_ptrs[idx++] = ptr;
+
+			/* Only need to touch memory once per backend process lifetime */
+			if (firstNumaTouch)
+				pg_numa_touch_mem_if_required(touch, ptr);
+		}
+
+		Assert(idx == os_page_count);
+
+		elog(DEBUG1, "NUMA: NBuffers=%d os_page_count=" UINT64_FORMAT " "
+			 "os_page_size=%zu", NBuffers, os_page_count, os_page_size);
+
+		/*
+		 * If we ever get 0xff back from kernel inquiry, then we probably have
+		 * bug in our buffers to OS page mapping code here.
+		 */
+		memset(os_page_status, 0xff, sizeof(int) * os_page_count);
+
+		/* Query NUMA status for all the pointers */
+		if (pg_numa_query_pages(0, os_page_count, os_page_ptrs, os_page_status) == -1)
+			elog(ERROR, "failed NUMA pages inquiry: %m");
+
+		/* Initialize the multi-call context, load entries about buffers */
+
+		funcctx = SRF_FIRSTCALL_INIT();
+
+		/* Switch context when allocating stuff to be used in later calls */
+		oldcontext = MemoryContextSwitchTo(funcctx->multi_call_memory_ctx);
+
+		/* Create a user function context for cross-call persistence */
+		fctx = (BufferCacheNumaContext *) palloc(sizeof(BufferCacheNumaContext));
+
+		if (get_call_result_type(fcinfo, NULL, &expected_tupledesc) != TYPEFUNC_COMPOSITE)
+			elog(ERROR, "return type must be a row type");
+
+		if (expected_tupledesc->natts != NUM_BUFFERCACHE_NUMA_ELEM)
+			elog(ERROR, "incorrect number of output arguments");
+
+		/* Construct a tuple descriptor for the result rows. */
+		tupledesc = CreateTemplateTupleDesc(expected_tupledesc->natts);
+		TupleDescInitEntry(tupledesc, (AttrNumber) 1, "bufferid",
+						   INT4OID, -1, 0);
+		TupleDescInitEntry(tupledesc, (AttrNumber) 2, "os_page_num",
+						   INT4OID, -1, 0);
+		TupleDescInitEntry(tupledesc, (AttrNumber) 3, "numa_node",
+						   INT4OID, -1, 0);
+
+		fctx->tupdesc = BlessTupleDesc(tupledesc);
+
+		/*
+		 * Each buffer needs at least one entry, but it might be offset in
+		 * some way, and use one extra entry. So we allocate space for the
+		 * maximum number of entries we might need, and then count the exact
+		 * number as we're walking buffers. That way we can do it in one pass,
+		 * without reallocating memory.
+		 */
+		pages_per_buffer = Max(1, BLCKSZ / os_page_size) + 1;
+		max_entries = NBuffers * pages_per_buffer;
+
+		/* Allocate entries for BufferCachePagesRec records. */
+		fctx->record = (BufferCacheNumaRec *)
+			MemoryContextAllocHuge(CurrentMemoryContext,
+								   sizeof(BufferCacheNumaRec) * max_entries);
+
+		/* Return to original context when allocating transient memory */
+		MemoryContextSwitchTo(oldcontext);
+
+		if (firstNumaTouch)
+			elog(DEBUG1, "NUMA: page-faulting the buffercache for proper NUMA readouts");
+
+		/*
+		 * Scan through all the buffers, saving the relevant fields in the
+		 * fctx->record structure.
+		 *
+		 * We don't hold the partition locks, so we don't get a consistent
+		 * snapshot across all buffers, but we do grab the buffer header
+		 * locks, so the information of each buffer is self-consistent.
+		 *
+		 * This loop touches and stores addresses into os_page_ptrs[] as input
+		 * to one big big move_pages(2) inquiry system call. Basically we ask
+		 * for all memory pages for NBuffers.
+		 */
+		startptr = (char *) TYPEALIGN_DOWN(os_page_size, (char *) BufferGetBlock(1));
+		idx = 0;
+		for (i = 0; i < NBuffers; i++)
+		{
+			char	   *buffptr = (char *) BufferGetBlock(i + 1);
+			BufferDesc *bufHdr;
+			uint32		buf_state;
+			uint32		bufferid;
+			int32		page_num;
+			char	   *startptr_buff,
+					   *endptr_buff;
+
+			CHECK_FOR_INTERRUPTS();
+
+			bufHdr = GetBufferDescriptor(i);
+
+			/* Lock each buffer header before inspecting. */
+			buf_state = LockBufHdr(bufHdr);
+			bufferid = BufferDescriptorGetBuffer(bufHdr);
+			UnlockBufHdr(bufHdr, buf_state);
+
+			/* start of the first page of this buffer */
+			startptr_buff = (char *) TYPEALIGN_DOWN(os_page_size, buffptr);
+
+			/* end of the buffer (no need to align to memory page) */
+			endptr_buff = buffptr + BLCKSZ;
+
+			Assert(startptr_buff < endptr_buff);
+
+			/* calculate ID of the first page for this buffer */
+			page_num = (startptr_buff - startptr) / os_page_size;
+
+			/* Add an entry for each OS page overlapping with this buffer. */
+			for (char *ptr = startptr_buff; ptr < endptr_buff; ptr += os_page_size)
+			{
+				fctx->record[idx].bufferid = bufferid;
+				fctx->record[idx].page_num = page_num;
+				fctx->record[idx].numa_node = os_page_status[page_num];
+
+				/* advance to the next entry/page */
+				++idx;
+				++page_num;
+			}
+		}
+
+		Assert((idx >= os_page_count) && (idx <= max_entries));
+
+		/* Set max calls and remember the user function context. */
+		funcctx->max_calls = idx;
+		funcctx->user_fctx = fctx;
+
+		/* Remember this backend touched the pages */
+		firstNumaTouch = false;
+	}
+
+	funcctx = SRF_PERCALL_SETUP();
+
+	/* Get the saved state */
+	fctx = funcctx->user_fctx;
+
+	if (funcctx->call_cntr < funcctx->max_calls)
+	{
+		uint32		i = funcctx->call_cntr;
+		Datum		values[NUM_BUFFERCACHE_NUMA_ELEM];
+		bool		nulls[NUM_BUFFERCACHE_NUMA_ELEM];
+
+		values[0] = Int32GetDatum(fctx->record[i].bufferid);
+		nulls[0] = false;
+
+		values[1] = Int32GetDatum(fctx->record[i].page_num);
+		nulls[1] = false;
+
+		values[2] = Int32GetDatum(fctx->record[i].numa_node);
+		nulls[2] = false;
+
+		/* Build and return the tuple. */
+		tuple = heap_form_tuple(fctx->tupdesc, values, nulls);
+		result = HeapTupleGetDatum(tuple);
+
+		SRF_RETURN_NEXT(funcctx, result);
+	}
+	else
+		SRF_RETURN_DONE(funcctx);
+}
+
 Datum
 pg_buffercache_summary(PG_FUNCTION_ARGS)
 {
diff --git a/contrib/pg_buffercache/sql/pg_buffercache_numa.sql b/contrib/pg_buffercache/sql/pg_buffercache_numa.sql
new file mode 100644
index 00000000000..837f3d64e21
--- /dev/null
+++ b/contrib/pg_buffercache/sql/pg_buffercache_numa.sql
@@ -0,0 +1,21 @@
+SELECT NOT(pg_numa_available()) AS skip_test \gset
+\if :skip_test
+\quit
+\endif
+
+-- We expect at least one entry for each buffer
+select count(*) >= (select setting::bigint
+                    from pg_settings
+                    where name = 'shared_buffers')
+from pg_buffercache_numa;
+
+-- Check that the functions / views can't be accessed by default. To avoid
+-- having to create a dedicated user, use the pg_database_owner pseudo-role.
+SET ROLE pg_database_owner;
+SELECT count(*) > 0 FROM pg_buffercache_numa;
+RESET role;
+
+-- Check that pg_monitor is allowed to query view / function
+SET ROLE pg_monitor;
+SELECT count(*) > 0 FROM pg_buffercache_numa;
+RESET role;