On Mon, Sep 21, 2009 at 03:12:45PM +0400, Igor Sysoev wrote: > On Fri, Sep 18, 2009 at 10:40:27AM +0300, Kostik Belousov wrote: > > > On Thu, Sep 17, 2009 at 03:26:41PM -0700, Xin LI wrote: > > > -----BEGIN PGP SIGNED MESSAGE----- > > > Hash: SHA1 > > > > > > Hi, Igor, > > > > > > Igor Sysoev wrote: > > > > Hi, > > > > > > > > nginx-0.8.15 can use completely non-blocking sendfile() using > > > > SF_NODISKIO > > > > flag. When sendfile() returns EBUSY, nginx calls aio_read() to read > > > > single > > > > byte. The first aio_read() preloads the first 128K part of a file in VM > > > > cache, > > > > however, all successive aio_read()s preload just 16K parts of the file. > > > > This makes non-blocking sendfile() usage ineffective for files larger > > > > than 128K. > > > > > > > > I've created a small patch for Darwin compatible F_RDAHEAD fcntl: > > > > > > > > fcntl(fd, F_RDAHEAD, preload_size) > > > > > > > > There is small incompatibilty: Darwin's fcntl allows just to > > > > enable/disable > > > > read ahead, while the proposed patch allows to set exact preload size. > > > > > > > > Currently the preload size affects vn_read() code path only and does not > > > > affect on sendfile() code path. However, it can be easy extended on > > > > sendfile() part too. The preload size is still limited by sysctl > > > > vfs.read_max. > > > > > > > > The patch is against FreeBSD 7.2 and was tested on FreeBSD 7.2-STABLE > > > > only. > > > > > > I have ported this as a patch against -HEAD (should apply on 8.0-R but > > > it's too late for us to add a new feature) plus a manual page entry > > > documenting the feature. > > > > > > I've used F_READAHEAD as the name, but reading the manual page, it looks > > > like we can just use F_RDAHEAD since Darwin seems to just distinguish 0 > > > and !=0 case so that programmers won't have to use #ifdef or something > > > else to get code working on different platform? > > > > What I dislike about the patch is the new kernel-private flag that is > > eaten from the open(2) flags namespace. We do already have FHASLOCK, > > so far the only such flag. > > We can change > int f_seqcount; > to > u_int f_seqcount; > > and can use highest bit instead of O_READAHEAD: anyway f_seqcount is shifted > to 16 bits left.
Or do the same trick as was done for FHASLOCK and override some flag that is not saved after open, see FMASK. Or split f_seqcount into two u_short fields, one for f_seqcount, second for f_kflag, and use the later for FHASLOCK and FREADAHEAD. [We are trying to not grow struct file unless absolutely neccessary].
pgpKZwRUrHgnP.pgp
Description: PGP signature
