On Tue, Aug 22, 2017 at 02:56:00PM +1000, Alexey Kardashevskiy wrote: > On 19/08/17 12:46, Alexey Kardashevskiy wrote: > > On 19/08/17 01:18, Eric Blake wrote: > >> On 08/18/2017 08:31 AM, Stefan Hajnoczi wrote: > >>> Most qcow2 files are uncompressed so it is wasteful to allocate (32 + 1) > >>> * cluster_size + 512 bytes upfront. Allocate s->cluster_cache and > >>> s->cluster_data when the first read operation is performance on a > >>> compressed cluster. > >>> > >>> The buffers are freed in .bdrv_close(). .bdrv_open() no longer has any > >>> code paths that can allocate these buffers, so remove the free functions > >>> in the error code path. > >>> > >>> Reported-by: Alexey Kardashevskiy <a...@ozlabs.ru> > >>> Cc: Kevin Wolf <kw...@redhat.com> > >>> Signed-off-by: Stefan Hajnoczi <stefa...@redhat.com> > >>> --- > >>> Alexey: Does this improve your memory profiling results? > >> > >> Is this a regression from earlier versions? > > > > Hm, I have not thought about this. > > > > So. I did bisect and this started happening from > > 9a4c0e220d8a4f82b5665d0ee95ef94d8e1509d5 > > "hw/virtio-pci: fix virtio behaviour" > > > > Before that, the very same command line would take less than 1GB of > > resident memory. That thing basically enforces virtio-1.0 for QEMU <=2.6 > > which means that upstream with "-machine pseries-2.6" works fine (less than > > 1GB), "-machine pseries-2.7" does not (close to 7GB, sometime even 9GB). > > > > Then I tried bisecting again, with > > "scsi=off,disable-modern=off,disable-legacy=on" on my 150 virtio-block > > devices, started from > > e266d421490e0 "virtio-pci: add flags to enable/disable legacy/modern" (it > > added the disable-modern switch) which uses 2GB of memory. > > > > I ended up with ada434cd0b44 "virtio-pci: implement cfg capability". > > > > Then I removed proxy->modern_as on v2.10.0-rc3 (see below) and got 1.5GB of > > used memory (yay!) > > > > I do not really know how to reinterpret all of this, do you? > > > Anyone, ping? Should I move the conversation to the original thread? Any > hacks to try with libc?
I suggest a new top-level thread with Michael Tsirkin CCed. Stefan
