On Thu, Nov 30, 2023 at 08:11:52PM +0000, Pádraig Brady wrote: > Much clearer thanks. > > On my system: > > $ stat /sys/kernel/profiling > File: /sys/kernel/profiling > Size: 4096 Blocks: 0 IO Block: 4096 regular file > > I can easily repro by setting the buffer size < PAGE_SIZE.
Oh, clever. > So this patch handles the case where sysfs reports a file is a certain size, > but it isn't really. In that case seeking to anywhere other than the start > doesn't give an error, but reading returns nothing. So we use a buffer size > large enough (>= PAGE_SIZE as inferred from st_blksize) so that we'll be > reading from the start of the file in this case. Exactly. > Note st_blksize can have unusual values, so it might be better > to use the io_blksize() wrapper to sanitize the values. > Though that will generally give 128K, which is good when processing all of a > file, > but perhaps overkill when processing just the last part of a file. > > So perhaps it's better to use buffer_size = MAX (BUFSIZ, getpagesize ()) > and avoid all the st_blksize edge cases. Is there any performance concern about switching to 64K reads for "normal" files (on these systems)? > I'll think about it a little, and make the adjustments. Appreciated! -dann > thanks! > Pádraig.
