site_archiver@lists.apple.com Delivered-To: darwin-dev@lists.apple.com -- Terry _______________________________________________ Do not post admin requests to the list. They will be ignored. Darwin-dev mailing list (Darwin-dev@lists.apple.com) Help/Unsubscribe/Update your Subscription: http://lists.apple.com/mailman/options/darwin-dev/site_archiver%40lists.appl... On Jan 7, 2008, at 12:51 PM, David Hoerl wrote: Date: Mon, 7 Jan 2008 09:49:19 -0800 From: Norm Green <norm.green@gemstone.com> Subject: RE: fread/fwrite/fflush bug ... Someone pointed out the fpurge() function. Calling fpurge() on the reader before the 2nd read fixes the problem. This might be one fix for your code, but how about other code not yet ported to OSX that is looking for "proper" behavior (making an assumption here). Reading the man page for fpurge, it says: "For input streams this discards any input read from the underlying object but not yet obtained via getc(3)...". I interpret this to mean the following would fail (but not your example): - writer writes say 1000 bytes - reader is told to read one byte at location 0 (but fread gets a buffer full) - writer changes all the data - all 1000 bytes - reader reads location 1, and gets old data (which it already had in the buffer) - reader keeps reading a byte at a time. When the internal buffer is exhausted THEN it would start reading the actual file again, and get a buffer of correct data. Based on the man page, one could assume that the seek backwards should cause the data in the buffer to get flushed as if fpurge was called. Apparently fread looks to see if the requested bytes are located in its buffer in all cases, and if so, uses them. The presumption by the man page here is it's the same program doing the reading and writing, and therefore the buffers are shared because you are dealing with the same descriptor. For other descriptors/buffers, you are effectively assuming a distributed cache coherency for the data cached by the stdio subsystem, which is not a valid assumption, under almost any circumstances. There's actually a stdio implementation that was done in university research that uses mmap()'ed pages rather than local buffers, but it pretty much never went anywhere outside of research because of the performance penalties for anything but small files. This email sent to site_archiver@lists.apple.com