| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-ID: <n64tn2dai56kovscubbajub7ke26njnxnnqqb5n7nsh3rtzfmt@53xi25pcxlec>
Date: Mon, 19 Aug 2024 15:07:50 +0300
From: "Kirill A. Shutemov" <kirill.shutemov@...ux.intel.com>
To: Alexey Gladkov <legion@...nel.org>
Cc: linux-kernel@...r.kernel.org, linux-coco@...ts.linux.dev,
Thomas Gleixner <tglx@...utronix.de>, Ingo Molnar <mingo@...hat.com>, Borislav Petkov <bp@...en8.de>,
Dave Hansen <dave.hansen@...ux.intel.com>, "H. Peter Anvin" <hpa@...or.com>,
Andrew Morton <akpm@...ux-foundation.org>, Yuan Yao <yuan.yao@...el.com>,
Geert Uytterhoeven <geert@...ux-m68k.org>, Yuntao Wang <ytcoode@...il.com>, Kai Huang <kai.huang@...el.com>,
Baoquan He <bhe@...hat.com>, Oleg Nesterov <oleg@...hat.com>, cho@...rosoft.com,
decui@...rosoft.com, John.Starks@...rosoft.com
Subject: Re: [PATCH v3 02/10] x86/tdx: Add validation of userspace MMIO
instructions
On Mon, Aug 19, 2024 at 01:48:16PM +0200, Alexey Gladkov wrote:
> On Mon, Aug 19, 2024 at 01:39:17PM +0300, Kirill A. Shutemov wrote:
> > On Fri, Aug 16, 2024 at 03:43:52PM +0200, Alexey Gladkov wrote:
> > > From: "Alexey Gladkov (Intel)" <legion@...nel.org>
> > >
> > > Instructions from kernel space are considered trusted. If the MMIO
> > > instruction is from userspace it must be checked.
> > >
> > > For userspace instructions, it is need to check that the INSN has not
> > > changed at the time of #VE and before the execution of the instruction.
> >
> > Well, we cannot really check if the instruction changed under us. We can
> > only check if the parsed instruction does an MMIO operation that is
> > allowed for the process.
>
> We also check that the memory access (read/write) type matches. Yes, we
> can't check the instruction itself, but we check the arguments.
>
> > >
> > > Once the userspace instruction parsed is enforced that the address
> > > points to mapped memory of current process and that address does not
> > > point to private memory.
> > >
> > > After parsing the userspace instruction, it is necessary to ensure that:
> > >
> > > 1. the operation direction (read/write) corresponds to #VE info;
> > > 2. the address still points to mapped memory of current process;
> > > 3. the address does not point to private memory.
> >
> > I don't see where you check 3.
> >
> > I guess you can add pte_decrypted(pte) check to get_phys_addr().
> >
> > But I'm not sure it is strictly needed.
>
> (ve->gpa != cc_mkdec(phys_addr)
>
> The ve->gpa was checked in the virt_exception_user/kernel().
phys_addr doesn't have shared bit. It is masked out on pte_pfn(). That's
the reason you use cc_mkdec() to compare with ve->gpa. Otherwise it would
fail.
>
> >
> > > Signed-off-by: Alexey Gladkov (Intel) <legion@...nel.org>
> > > ---
> > > arch/x86/coco/tdx/tdx.c | 128 ++++++++++++++++++++++++++++++++++++----
> > > 1 file changed, 115 insertions(+), 13 deletions(-)
> > >
> > > diff --git a/arch/x86/coco/tdx/tdx.c b/arch/x86/coco/tdx/tdx.c
> > > index af0b6c1cacf7..86c22fec97fb 100644
> > > --- a/arch/x86/coco/tdx/tdx.c
> > > +++ b/arch/x86/coco/tdx/tdx.c
> > > @@ -8,6 +8,7 @@
> > > #include <linux/export.h>
> > > #include <linux/io.h>
> > > #include <linux/kexec.h>
> > > +#include <linux/mm.h>
> > > #include <asm/coco.h>
> > > #include <asm/tdx.h>
> > > #include <asm/vmx.h>
> > > @@ -405,6 +406,84 @@ static bool mmio_write(int size, unsigned long addr, unsigned long val)
> > > EPT_WRITE, addr, val);
> > > }
> > >
> > > +static inline bool is_private_gpa(u64 gpa)
> > > +{
> > > + return gpa == cc_mkenc(gpa);
> > > +}
> > > +
> > > +static int get_phys_addr(unsigned long addr, phys_addr_t *phys_addr, bool *writable)
> > > +{
> > > + unsigned int level;
> > > + pgd_t *pgdp;
> > > + pte_t *ptep;
> > > +
> > > + /*
> > > + * Address validation only makes sense for a user process. The lock must
> > > + * be obtained before validation can begin.
> > > + */
> > > + mmap_assert_locked(current->mm);
> > > +
> > > + pgdp = pgd_offset(current->mm, addr);
> > > +
> > > + if (!pgd_none(*pgdp)) {
> > > + ptep = lookup_address_in_pgd(pgdp, addr, &level);
> > > + if (ptep) {
> > > + unsigned long offset;
> > > +
> > > + offset = addr & ~page_level_mask(level);
> > > + *phys_addr = PFN_PHYS(pte_pfn(*ptep));
> > > + *phys_addr |= offset;
> > > +
> > > + *writable = pte_write(*ptep);
> > > +
> > > + return 0;
> > > + }
> > > + }
> > > +
> > > + return -EFAULT;
> > > +}
> > > +
> > > +static int valid_vaddr(struct ve_info *ve, enum insn_mmio_type mmio, int size,
> > > + unsigned long vaddr)
> > > +{
> > > + phys_addr_t phys_addr;
> > > + bool writable = false;
> > > +
> > > + /* It's not fatal. This can happen due to swap out or page migration. */
> > > + if (get_phys_addr(vaddr, &phys_addr, &writable) || (ve->gpa != cc_mkdec(phys_addr)))
> >
> > Too long line?
>
> All patches pass checkpatch without warnings.
Checkpatch is not the ultimate authority. But I am neither. :P
> >
> > > + return -EAGAIN;
> > > +
> > > + /*
> > > + * Re-check whether #VE info matches the instruction that was decoded.
> > > + *
> > > + * The ve->gpa was valid at the time ve_info was received. But this code
> > > + * executed with interrupts enabled, allowing tlb shootdown and therefore
> > > + * munmap() to be executed in the parallel thread.
> > > + *
> > > + * By the time MMIO emulation is performed, ve->gpa may be already
> > > + * unmapped from the process, the device it belongs to removed from
> > > + * system and something else could be plugged in its place.
> > > + */
> > > + switch (mmio) {
> > > + case INSN_MMIO_WRITE:
> > > + case INSN_MMIO_WRITE_IMM:
> > > + if (!writable || !(ve->exit_qual & EPT_VIOLATION_ACC_WRITE))
> > > + return -EFAULT;
> > > + break;
> > > + case INSN_MMIO_READ:
> > > + case INSN_MMIO_READ_ZERO_EXTEND:
> > > + case INSN_MMIO_READ_SIGN_EXTEND:
> > > + if (!(ve->exit_qual & EPT_VIOLATION_ACC_READ))
> > > + return -EFAULT;
> > > + break;
> > > + default:
> > > + WARN_ONCE(1, "Unsupported mmio instruction: %d", mmio);
> > > + return -EINVAL;
> > > + }
> > > +
> > > + return 0;
> > > +}
> > > +
> > > static int handle_mmio_write(struct insn *insn, enum insn_mmio_type mmio, int size,
> > > struct pt_regs *regs, struct ve_info *ve)
> > > {
> > > @@ -489,7 +568,7 @@ static int handle_mmio(struct pt_regs *regs, struct ve_info *ve)
> > > enum insn_mmio_type mmio;
> > > struct insn insn = {};
> > > unsigned long vaddr;
> > > - int size;
> > > + int size, ret;
> > >
> > > /* Only in-kernel MMIO is supported */
> > > if (WARN_ON_ONCE(user_mode(regs)))
> > > @@ -505,6 +584,17 @@ static int handle_mmio(struct pt_regs *regs, struct ve_info *ve)
> > > if (WARN_ON_ONCE(mmio == INSN_MMIO_DECODE_FAILED))
> > > return -EINVAL;
> > >
> > > + vaddr = (unsigned long)insn_get_addr_ref(&insn, regs);
> > > +
> > > + if (current->mm) {
> >
> > Hm. This path will be taken for any MMIO if it is done in context of a
> > process, even in-kernel only. I don't think we want it. It is useless
> > overhead.
>
> The kthread do not have a current->mm.
I am not talking about kthread. I am talking about initiating MMIO from
kernel, but within a process context. Like, you call an ioctl() on a
device fd and it triggers MMIO in kernel. This scenario would have
current->mm, but it is not userspace MMIO.
> > Use user_mode(regs) instead.
>
> I can't use this. When nested exception happens in the handle_mmio_movs()
> the regs will be not in the user mode.
>
> I can make a flag that will be set either for user_mode or if we have a
> nested exception.
Hm. Yeah. This is ugly. Let me think about it.
--
Kiryl Shutsemau / Kirill A. Shutemov
Powered by blists - more mailing lists