From mboxrd@z Thu Jan 1 00:00:00 1970 From: Denys Vlasenko Subject: [PATCH] vfs: reduce stack usage by shrinking struct kiocb Date: Sun, 27 Apr 2008 06:17:36 +0200 Message-ID: <200804270617.36629.vda.linux@googlemail.com> Mime-Version: 1.0 Content-Type: Multipart/Mixed; boundary="Boundary-00=_g5/EI0ZahsCaZ9p" Cc: xfs@oss.sgi.com, Eric Sandeen , Adrian Bunk , linux-kernel@vger.kernel.org, linux-fsdevel@vger.kernel.org, linux-aio@kvack.org To: David Chinner , Benjamin LaHaise Return-path: Received: from fg-out-1718.google.com ([72.14.220.155]:11554 "EHLO fg-out-1718.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751149AbYD0ESk (ORCPT ); Sun, 27 Apr 2008 00:18:40 -0400 Received: by fg-out-1718.google.com with SMTP id l27so4334398fgb.17 for ; Sat, 26 Apr 2008 21:18:37 -0700 (PDT) Sender: linux-fsdevel-owner@vger.kernel.org List-ID: --Boundary-00=_g5/EI0ZahsCaZ9p Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit Content-Disposition: inline Hi Al, Benjamin, David, struct kiocb is placed on stack by, for example, do_sync_write(). Eventually it contributes to xfs writeout path's stack usage, among others. This is *the* path which causes 4k stack overflows on i386 with xfs. This patch trivially reorders fields of this structure, and makes some of them smaller. Reordering helps 64-bit architectures: int, void*, int, void* - bad, int, int, void*, void* - better. These fields are made smaller: ki_flags: long -> short: possible values are 0,1,2, so short is enough. ki_nr_segs: ulong -> uint: nobody uses 4 billion element writev's (and it would not work anyway) ki_cur_seg: same For 32bit x86, it makes this struct only 4 bytes smaller. This isn't much, but it helps not only xfs, but all filesystems. For 64-bit case savings are a bit more significant, as ulong -> uint actually makes a difference, and reordering of 64-bit fields eliminates some padding. Only compile tested. Observed stack reductions on 32 bits: -sock_recvmsg [vmlinux]: 196 -sock_sendmsg [vmlinux]: 196 +sock_recvmsg [vmlinux]: 192 +sock_sendmsg [vmlinux]: 192 -do_sync_write [vmlinux]: 140 -do_sync_read [vmlinux]: 140 +do_sync_write [vmlinux]: 136 +do_sync_read [vmlinux]: 136 -do_sync_readv_writev [vmlinux]: 132 +do_sync_readv_writev [vmlinux]: 128 Signed-off-by: Denys Vlasenko -- vda --Boundary-00=_g5/EI0ZahsCaZ9p Content-Type: text/x-diff; charset="us-ascii"; name="stk4.diff" Content-Transfer-Encoding: 7bit Content-Disposition: attachment; filename="stk4.diff" diff -urpN linux-2.6-xfs1/include/linux/aio.h linux-2.6-xfs1.stk4/include/linux/aio.h --- linux-2.6-xfs1/include/linux/aio.h 2008-03-30 03:27:54.000000000 +0200 +++ linux-2.6-xfs1.stk4/include/linux/aio.h 2008-04-27 05:47:09.000000000 +0200 @@ -86,7 +86,8 @@ struct kioctx; */ struct kiocb { struct list_head ki_run_list; - unsigned long ki_flags; + unsigned short ki_flags; /* range: 0..2 */ + unsigned short ki_opcode; int ki_users; unsigned ki_key; /* id of this request */ @@ -102,23 +103,22 @@ struct kiocb { } ki_obj; __u64 ki_user_data; /* user's data for completion */ - wait_queue_t ki_wait; loff_t ki_pos; + wait_queue_t ki_wait; void *private; /* State that we remember to be able to restart/retry */ - unsigned short ki_opcode; + /*unsigned short ki_opcode; - moved up for denser packing */ size_t ki_nbytes; /* copy of iocb->aio_nbytes */ - char __user *ki_buf; /* remaining iocb->aio_buf */ size_t ki_left; /* remaining bytes */ + unsigned ki_nr_segs; + unsigned ki_cur_seg; struct iovec ki_inline_vec; /* inline vector */ + char __user *ki_buf; /* remaining iocb->aio_buf */ struct iovec *ki_iovec; - unsigned long ki_nr_segs; - unsigned long ki_cur_seg; struct list_head ki_list; /* the aio core uses this * for cancellation */ - /* * If the aio_resfd field of the userspace iocb is not zero, * this is the underlying file* to deliver event to. --Boundary-00=_g5/EI0ZahsCaZ9p--