Date: Tue, 20 Aug 2002 04:21:11 +1000 (EST) From: Bruce Evans <bde@zeta.org.au> To: Matthew Dillon <dillon@apollo.backplane.com> Cc: "Semen A. Ustimenko" <semenu@FreeBSD.ORG>, <freebsd-current@FreeBSD.ORG>, Maxim Konovalov <maxim@FreeBSD.ORG>, Robert Watson <rwatson@FreeBSD.ORG>, <dg@FreeBSD.ORG> Subject: Re: sendfile() change (Was: Re: cvs commit: src/sys/kern uipc_syscalls.c) Message-ID: <20020820040605.N20882-100000@gamplex.bde.org> In-Reply-To: <200208191649.g7JGnXIR077528@apollo.backplane.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Mon, 19 Aug 2002, Matthew Dillon wrote: > :Why can't we get rid of VOP_READ(.. UIO_NOCOPY...) call in sendfile()? > :Me, I can't quite understand what UIO_NOCOPY means... As long as > :sendfile() function already plays around pages, it can use VOP_GETPAGES(). > :The following patch looks works for me. Could anybody said if it has any > :benefits or not? > UIO_NOCOPY tells the filesystem to not bother copying the data > into the passed buffer but to instead simply load it into the buffer > cache / VM backing store. > > While this adds buffer cache management overhead, it ought to yield > far greater performance over doing a VOP_GETPAGES() at this point > because the filesystem will be able to do clustered reads and > read-ahead (potentially a 64K I/O) instead of a 4K I/O. > > An alternative would be to cluster the VM pages in sendfile() and > call VOP_GETPAGES() on a block of pages instead of just one. I'm > guessing that is not being done because it's about 100 lines of code > to do it right. It's easier just to call vn_rdwr() and let the > system do the clustering. This argument seems to apply to exec_map_first_pages() too. It does its own blocking to get up to VM_INITIAL_PAGEIN (normally 16) pages using only about 50 lines of code, but does this slightly wrong. It doesn't honor the device limit of si_iosize_max, unlike filesystem code. This used to break exec on zip drives (si_iosize_max used to be 32K, but VM_INITIAL_PAGEIN pages is 64K bytes on i386's). Bruce To Unsubscribe: send mail to majordomo@FreeBSD.org with "unsubscribe freebsd-current" in the body of the message
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20020820040605.N20882-100000>