On Tue, Sep 10, 2013 at 06:09:54PM +0300, Konstantin Belousov wrote: > On Tue, Sep 10, 2013 at 07:29:51AM -0700, Matthew Fleming wrote: > > I don't know exactly what the stock malloc(9) will return, but memguard(9), > > under its default mode with vm.memguard.options having MG_GUARD_AROUND set > > will align the returned pointer to only 16 bytes. When I added that > > feature I almost made it 8 bytes, but I think I saw that uma(9) had a > > 16-byte alignment so I preserved that. I.e., this code does still work > > with malloc(9) and memguard(9). > > > > But why does this need 16 byte alignment? Especially when one of the > > comments says 4-byte alignment? > > The comment about 4-byte alignment is for VIA. > > For Intel, citing the IA-32 SDM rev. 47, Vol.3, 9.11.6 Microcode Update > Loader: > "The microcode update data requires a 16-byte boundary alignment". > I.e. CPU takes the linear address of the buffer, which must be aligned > on 16 bytes.
Sent too fast. There is at least one more instance of the same use, see bus_dmamem_alloc() in sys/x86/x86/busdma_machdep.c.
pgp_xbvmMXtdw.pgp
Description: PGP signature
