Hi Nick. > >This patch unfortunately appears like it may introduce an >uninitialized memory leak due to a data race between one >thread initializing a buffer then marking it uptodate, and >the other testing buffer uptodate then reading from the >buffer (buffer, read as: page memory covered by buffer head). > >For reference, this is basically the same class of data race >that I fixed 0ed361dec36945f3116ee1338638ada9a8920905 I think the same issue that your patch 0ed361dec36945f3116ee 1338638ada9a8920905 "fix PageUptodate data race" pointed out is there around buffer_head without my patch "vfs: pagecache usage optimization for pagesize!=blocksize". Because set_buffer_uptodate or buffer_uptodate are used without distinct locking facility (lock_buffer, or lock_page) in many places, so it would be needed to deal with memory ordering issue. Do you add wmb/rmb to BUFFER_FNS macros? > >I don't think this patch got quite enough justification to >warrant just blindly putting barriers in the buffer bitops. >The best-case numbers for it were reasonable enough when the >downside was only an extra branch or two in a relatively slow >path. I don't really know how best to go from here (maybe >someone can argue it is not a problem or come up with a better >fix?). > >Thanks, >Nick -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
