Re: [PATCH] target/riscv/kvm.c: fix mvendorid size in vcpu_set_machine_ids()

[Andrew Jones <ajones@ventanamicro.com>]

On Thu, Aug 03, 2023 at 08:36:57AM -0300, Daniel Henrique Barboza wrote:
> 
> 
> On 8/3/23 06:29, Andrew Jones wrote:
> > On Wed, Aug 02, 2023 at 03:00:58PM -0300, Daniel Henrique Barboza wrote:
> > > cpu->cfg.mvendorid is a 32 bit field and kvm_set_one_reg() always write
> > > a target_ulong val, i.e. a 64 bit field in a 64 bit host.
> > > 
> > > Given that we're passing a pointer to the mvendorid field, the reg is
> > > reading 64 bits starting from mvendorid and going 32 bits in the next
> > > field, marchid. Here's an example:
> > > 
> > > $ ./qemu-system-riscv64 -machine virt,accel=kvm -m 2G -smp 1 \
> > >     -cpu rv64,marchid=0xab,mvendorid=0xcd,mimpid=0xef(...)
> > > 
> > > (inside the guest)
> > >   # cat /proc/cpuinfo
> > > processor	: 0
> > > hart		: 0
> > > isa		: rv64imafdc_zicbom_zicboz_zihintpause_zbb_sstc
> > > mmu		: sv57
> > > mvendorid	: 0xab000000cd
> > > marchid		: 0xab
> > > mimpid		: 0xef
> > > 
> > > 'mvendorid' was written as a combination of 0xab (the value from the
> > > adjacent field, marchid) and its intended value 0xcd.
> > > 
> > > Fix it by assigning cpu->cfg.mvendorid to a target_ulong var 'reg' and
> > > use it as input for kvm_set_one_reg(). Here's the result with this patch
> > > applied and using the same QEMU command line:
> > > 
> > >   # cat /proc/cpuinfo
> > > processor	: 0
> > > hart		: 0
> > > isa		: rv64imafdc_zicbom_zicboz_zihintpause_zbb_sstc
> > > mmu		: sv57
> > > mvendorid	: 0xcd
> > > marchid		: 0xab
> > > mimpid		: 0xef
> > > 
> > > This bug affects only the generic (rv64) CPUs when running with KVM in a
> > > 64 bit env since the 'host' CPU does not allow the machine IDs to be
> > > changed via command line.
> > > 
> > > Fixes: 1fb5a622f7 ("target/riscv: handle mvendorid/marchid/mimpid for KVM CPUs")
> > > Signed-off-by: Daniel Henrique Barboza <dbarboza@ventanamicro.com>
> > > ---
> > >   target/riscv/kvm.c | 9 ++++++++-
> > >   1 file changed, 8 insertions(+), 1 deletion(-)
> > > 
> > > diff --git a/target/riscv/kvm.c b/target/riscv/kvm.c
> > > index 9d8a8982f9..b1fd2233c0 100644
> > > --- a/target/riscv/kvm.c
> > > +++ b/target/riscv/kvm.c
> > > @@ -852,12 +852,19 @@ void kvm_arch_init_irq_routing(KVMState *s)
> > >   static int kvm_vcpu_set_machine_ids(RISCVCPU *cpu, CPUState *cs)
> > >   {
> > >       CPURISCVState *env = &cpu->env;
> > > +    target_ulong reg;
> > 
> > We can use the type of cfg since KVM just gets an address and uses the
> > KVM register type to determine the size. So here,
> > 
> >   uint32_t reg = cpu->cfg.mvendorid;
> > 
> > and then...
> > 
> > >       uint64_t id;
> > >       int ret;
> > >       id = kvm_riscv_reg_id(env, KVM_REG_RISCV_CONFIG,
> > >                             KVM_REG_RISCV_CONFIG_REG(mvendorid));
> > > -    ret = kvm_set_one_reg(cs, id, &cpu->cfg.mvendorid);
> > > +    /*
> > > +     * cfg.mvendorid is an uint32 but a target_ulong will
> > > +     * be written. Assign it to a target_ulong var to avoid
> > > +     * writing pieces of other cpu->cfg fields in the reg.
> > > +     */
> > 
> > ...we don't need this comment since we're not doing anything special.
> 
> I tried it out and it doesn't seem to work. Here's the result:
> 
> / # cat /proc/cpuinfo
> processor	: 0
> hart		: 0
> isa		: rv64imafdc_zicbom_zicboz_zihintpause_zbb_sstc
> mmu		: sv57
> mvendorid	: 0xaaaaaa000000cd
> marchid		: 0xab
> mimpid		: 0xef
> 
> 
> The issue here is that the kernel considers 'mvendorid' as an unsigned long (or
> what QEMU calls target_ulong). kvm_set_one_reg() will write an unsigned long
> regardless of the uint32_t typing of 'reg', meaning that it'll end up writing
> 32 bits of uninitialized stuff from the stack.

Indeed, I managed to reverse the problem in my head. We need to to worry
about KVM's notion of the type, not QEMU's. I feel like we still need some
sort of helper, but one that takes the type of the KVM register into
account. KVM_REG_RISCV_CONFIG registers are KVM ulongs, which may be
different than QEMU's ulongs, if we ever supported 32-bit userspace on
64-bit kernels. Also, not all KVM register types are ulong, some are
explicitly u32s and others u64s. I see kvm_riscv_reg_id() is used to try
and get the right KVM reg size set, but it's broken for RISCV_FP_F_REG(),
since those are all u32s, even when KVM has 64-bit ulong (I guess nobody
is testing get/set-one-reg with F registers using that helper, otherwise
we'd be getting EINVAL from KVM). And KVM_REG_RISCV_FP_D_REG(fcsr) is also
broken and RISCV_TIMER_REG() looks broken with 32-bit KVM, since it should
always be u64. I guess all that stuff needs an audit.

So, I think we need a helper that has a switch on the KVM register type
and provides the right sized buffer for each case.

Thanks,
drew


> 
> target_ulong seems that the right choice here. We could perhaps work with
> uint64_t (other parts of the code does that) but target_ulong is nicer with
> 32-bit setups.
> 
> 
> Thanks,
> 
> Daniel
> 
> > 
> > > +    reg = cpu->cfg.mvendorid;
> > > +    ret = kvm_set_one_reg(cs, id, &reg);
> > >       if (ret != 0) {
> > >           return ret;
> > >       }
> > > -- 
> > > 2.41.0
> > > 
> > 
> > We should audit and fix all uses of &cpu->cfg.* with KVM ioctls. We can
> > also consider introducing wrappers like
> > 
> > #define kvm_set_one_reg_safe(cs, id, addr)	\
> > ({						\
> > 	typeof(*(addr)) _addr = *(addr);	\
> > 	kvm_set_one_reg(cs, id, &_addr)		\
> > })
> > 
> > Thanks,
> > drew