From 33986e97b7d8819b8dc613d1ba878bc6db4f2163 Mon Sep 17 00:00:00 2001 From: Kim Alvefur Date: Wed, 7 Jun 2023 00:39:30 +0200 Subject: util.datamanager: Pad list writes to avoid crossing block boundaries By padding items so that they do not cross block boundaries, it becomes eaiser to delete whole blocks with fallocate() without cutting items in half, improving efficiency of such operations. Since list stores are used for message archives, where the most common deletion operation would be of the oldest entires, at the top of the file. With this, all blocks that contain items to be removed could be deleted without needing to read, delete and write out the whole file. --- util/datamanager.lua | 7 +++++++ 1 file changed, 7 insertions(+) diff --git a/util/datamanager.lua b/util/datamanager.lua index 630353a9..6759592a 100644 --- a/util/datamanager.lua +++ b/util/datamanager.lua @@ -32,6 +32,7 @@ local path_separator = assert ( package.config:match ( "^([^\n]+)" ) , "package. local prosody = prosody; +local blocksize = 0x1000; local raw_mkdir = lfs.mkdir; local atomic_append; local remove_blocks; @@ -244,6 +245,12 @@ local function append(username, host, datastore, ext, data) end local pos = f:seek("end"); + if (blocksize-(pos%blocksize)) < (#data%blocksize) then + -- pad to blocksize with newlines so that the next item is both on a new + -- block and a new line + atomic_append(f, ("\n"):rep(blocksize-(pos%blocksize))); + pos = f:seek("end"); + end local ok, msg = atomic_append(f, data); -- cgit v1.2.3