> On 8/2/23 06:08, Mikhail Tyutin wrote: > > The fix is to clear TLB_INVALID_MASK bit in tlb_addr, as it happens in > > other places e.g. > > load_helper(). > > > > Signed-off-by: Dmitriy Solovev <[email protected]> > > Signed-off-by: Mikhail Tyutin <[email protected]> > > --- > > accel/tcg/cputlb.c | 2 +- > > 1 file changed, 1 insertion(+), 1 deletion(-) > > The other places in load_helper happen only directly after tlb_fill has > succeeded. Here > you have no such guarantee. > > I think perhaps the save_iotlb_data() call should be applied to loads as > well, and then > tlb_plugin_lookup simplified. >
Hello Richard, We performed testing on more scenarios and noticed that patch when save_iotlb_data() call is added to io_readx (https://patchew.org/QEMU/[email protected]/). It doesn't work for addresses in OCRAM region. Those accessed bypass io_writex/io_readx function and therefore don’t invoke save_iotlb_data(). So we observe the wrong value of cpu->saved_iotlb for it. Would not be better to get back to initial v1 approach when we clean TLB_INVALID_MASK flag in tlb_plugin_lookup()? It works well for those regions. (https://patchew.org/QEMU/[email protected])
