On 2020-01-05 17:36, Aurelien Jarno wrote:
> > diff --git a/target/riscv/csr.c b/target/riscv/csr.c
> > index e0d4586760..2789215b5e 100644
> > --- a/target/riscv/csr.c
> > +++ b/target/riscv/csr.c
>
> [ snip ]
>
> > @@ -307,6 +307,7 @@ static int write_mstatus(CPURISCVState *env, int csrno,
> > target_ulong val)
> > {
> > target_ulong mstatus = env->mstatus;
> > target_ulong mask = 0;
> > + int dirty;
> >
> > /* flush tlb on mstatus fields that affect VM */
> > if (env->priv_ver <= PRIV_VERSION_1_09_1) {
> > @@ -340,8 +341,9 @@ static int write_mstatus(CPURISCVState *env, int csrno,
> > target_ulong val)
> >
> > mstatus = (mstatus & ~mask) | (val & mask);
> >
> > - int dirty = ((mstatus & MSTATUS_FS) == MSTATUS_FS) |
> > - ((mstatus & MSTATUS_XS) == MSTATUS_XS);
> > + dirty = (riscv_cpu_fp_enabled(env) &&
> > + ((mstatus & MSTATUS_FS) == MSTATUS_FS)) |
> > + ((mstatus & MSTATUS_XS) == MSTATUS_XS);
> > mstatus = set_field(mstatus, MSTATUS_SD, dirty);
> > env->mstatus = mstatus;
>
> This patch, and more precisely the above two hunks broke
> qemu-system-riscv64. More precisely, when running a Debian sid system
> inside QEMU, sshd hangs during key exchange.
The problem is that at this stage, mstatus != env->status. Prior to that
patch, dirty was computed exclusively on the new mstatus status, after
the update by val. With this patch, riscv_cpu_fp_enabled() refers to the
old value of mstatus. Therefore when FS is changed from "Off" (FS = 00)
to "Dirty" (FS == 11), the SD bit is not set.
--
Aurelien Jarno GPG: 4096R/1DDD8C9B
aurel...@aurel32.net http://www.aurel32.net
