Anthony Liguori schrieb: > Kevin Wolf wrote: >> Anthony Liguori schrieb: >> >>> Kevin Wolf wrote: >>> >>>> Anthony Liguori schrieb: >>>> >>>> Yes, if it fails, the EINVAL is no surprise. I meant what code path it >>>> was using. Obviously we missed something in our patch and I'd like to >>>> fix that. Did the error occur on raw images or something like qcow2? >>>> >>> It's a raw image and the calls are being made via >>> bdrv_aio_read/bdrv_aio_write. It doesn't occur with a qcow2 but then >>> cache=off doesn't seem to do what it's supposed to with cache=off (I >>> believe the underlying backing file is not opened O_DIRECT?). >>> >> >> This is really strange. In raw_aio_read/write there is a check like this: >> >> if (unlikely(s->aligned_buf != NULL && ((uintptr_t) buf % 512))) { >> // emulate it using raw_pread/write which uses >> // s->aligned_buf for the request then >> } >> > > Something is goofy then. > >> For qcow2 I think O_DIRECT actually is in effect. Otherwise it would >> have worked even without our patch, and it didn't. And indeed, looking >> at the code, it passes flags to bdrv_file_open when it opens the image >> file. >> > > Something's broken then. Maybe -snapshot doesn't pick up the > O_DIRECT'ness? I'll have to check again. I was definitely seeing page > cache behavior with cache=off.
Right, qemu seems to drop the flags for the backing file when using BDRV_O_SNAPSHOT (bdrv2_open in block.c opens the file). So O_DIRECT applies only to new data. Have you been using -snapshot when you had trouble with the unaligned buffer, too? I don't think I have tested this one when I made the patch... Kevin -- To unsubscribe from this list: send the line "unsubscribe kvm" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html