Re: [Xen-devel] [PATCH 6/6] x86/hvm: Implement hvmemul_write() using real mappings

[Paul Durrant <Paul.Durrant@xxxxxxxxxx>]

> -----Original Message-----
> From: Andrew Cooper [mailto:andrew.cooper3@xxxxxxxxxx]
> Sent: 21 June 2017 16:13
> To: Xen-devel <xen-devel@xxxxxxxxxxxxx>
> Cc: Andrew Cooper <Andrew.Cooper3@xxxxxxxxxx>; Jan Beulich
> <JBeulich@xxxxxxxx>; Paul Durrant <Paul.Durrant@xxxxxxxxxx>; Razvan
> Cojocaru <rcojocaru@xxxxxxxxxxxxxxx>; Mihai Donțu
> <mdontu@xxxxxxxxxxxxxxx>
> Subject: [PATCH 6/6] x86/hvm: Implement hvmemul_write() using real
> mappings
> 
> An access which crosses a page boundary is performed atomically by x86
> hardware, albeit with a severe performance penalty.  An important corner
> case
> is when a straddled access hits two pages which differ in whether a
> translation exists, or in net access rights.
> 
> The use of hvm_copy*() in hvmemul_write() is problematic, because it
> performs
> a translation then completes the partial write, before moving onto the next
> translation.
> 
> If an individual emulated write straddles two pages, the first of which is
> writable, and the second of which is not, the first half of the write will
> complete before #PF is raised from the second half.
> 
> This results in guest state corruption as a side effect of emulation, which
> has been observed to cause windows to crash while under introspection.
> 
> Introduce the hvmemul_{,un}map_linear_addr() helpers, which translate an
> entire contents of a linear access, and vmap() the underlying frames to
> provide a contiguous virtual mapping for the emulator to use.  This is the
> same mechanism as used by the shadow emulation code.
> 
> This will catch any translation issues and abort the emulation before any
> modifications occur.
> 
> Signed-off-by: Andrew Cooper <andrew.cooper3@xxxxxxxxxx>
> ---
> CC: Jan Beulich <JBeulich@xxxxxxxx>
> CC: Paul Durrant <paul.durrant@xxxxxxxxxx>
> CC: Razvan Cojocaru <rcojocaru@xxxxxxxxxxxxxxx>
> CC: Mihai Donțu <mdontu@xxxxxxxxxxxxxxx>
> 
> While the maximum size of linear mapping is capped at 1 page, the logic in
> the
> helpers is written to work properly as hvmemul_ctxt->mfn[] gets longer,
> specifically with XSAVE instruction emulation in mind.
> 
> This has only had light testing so far.
> ---
>  xen/arch/x86/hvm/emulate.c        | 179
> ++++++++++++++++++++++++++++++++++----
>  xen/include/asm-x86/hvm/emulate.h |   7 ++
>  2 files changed, 169 insertions(+), 17 deletions(-)
> 
> diff --git a/xen/arch/x86/hvm/emulate.c b/xen/arch/x86/hvm/emulate.c
> index 384ad0b..804bea5 100644
> --- a/xen/arch/x86/hvm/emulate.c
> +++ b/xen/arch/x86/hvm/emulate.c
> @@ -498,6 +498,159 @@ static int hvmemul_do_mmio_addr(paddr_t
> mmio_gpa,
>  }
> 
>  /*
> + * Map the frame(s) covering an individual linear access, for writeable
> + * access.  May return NULL for MMIO, or ERR_PTR(~X86EMUL_*) for other
> errors
> + * including ERR_PTR(~X86EMUL_OKAY) for write-discard mappings.
> + *
> + * In debug builds, map() checks that each slot in hvmemul_ctxt->mfn[] is
> + * clean before use, and poisions unused slots with INVALID_MFN.
> + */
> +static void *hvmemul_map_linear_addr(
> +    unsigned long linear, unsigned int bytes, uint32_t pfec,
> +    struct hvm_emulate_ctxt *hvmemul_ctxt)
> +{
> +    struct vcpu *curr = current;
> +    void *err, *mapping;
> +
> +    /* First and final gfns which need mapping. */
> +    unsigned long frame = linear >> PAGE_SHIFT, first = frame;
> +    unsigned long final = (linear + bytes - !!bytes) >> PAGE_SHIFT;

Do we need to worry about linear + bytes overflowing here?

Also, is this ever legitimately called with bytes == 0?

> +
> +    /*
> +     * mfn points to the next free slot.  All used slots have a page 
> reference
> +     * held on them.
> +     */
> +    mfn_t *mfn = &hvmemul_ctxt->mfn[0];
> +
> +    /*
> +     * The caller has no legitimate reason for trying a zero-byte write, but
> +     * final is calculate to fail safe in release builds.
> +     *
> +     * The maximum write size depends on the number of adjacent mfns[]
> which
> +     * can be vmap()'d, accouting for possible misalignment within the 
> region.
> +     * The higher level emulation callers are responsible for ensuring that
> +     * mfns[] is large enough for the requested write size.
> +     */
> +    if ( bytes == 0 ||
> +         final - first > ARRAY_SIZE(hvmemul_ctxt->mfn) - 1 )
> +    {

I guess not, so why the weird looking calculation for final? It's value will 
not be used when bytes == 0.

> +        ASSERT_UNREACHABLE();
> +        goto unhandleable;
> +    }
> +
> +    do {
> +        enum hvm_translation_result res;
> +        struct page_info *page;
> +        pagefault_info_t pfinfo;
> +        p2m_type_t p2mt;
> +
> +        res = hvm_translate_get_page(curr, frame << PAGE_SHIFT, true, pfec,
> +                                     &pfinfo, &page, NULL, &p2mt);
> +
> +        switch ( res )
> +        {
> +        case HVMTRANS_okay:
> +            break;
> +
> +        case HVMTRANS_bad_linear_to_gfn:
> +            x86_emul_pagefault(pfinfo.ec, pfinfo.linear, 
> &hvmemul_ctxt->ctxt);
> +            err = ERR_PTR(~(long)X86EMUL_EXCEPTION);
> +            goto out;
> +
> +        case HVMTRANS_bad_gfn_to_mfn:
> +            err = NULL;
> +            goto out;
> +
> +        case HVMTRANS_gfn_paged_out:
> +        case HVMTRANS_gfn_shared:
> +            err = ERR_PTR(~(long)X86EMUL_RETRY);
> +            goto out;
> +
> +        default:
> +            goto unhandleable;
> +        }
> +
> +        /* Error checking.  Confirm that the current slot is clean. */
> +        ASSERT(mfn_x(*mfn) == 0);
> +
> +        *mfn++ = _mfn(page_to_mfn(page));
> +        frame++;
> +
> +        if ( p2m_is_discard_write(p2mt) )
> +        {
> +            err = ERR_PTR(~(long)X86EMUL_OKAY);
> +            goto out;
> +        }
> +
> +    } while ( frame < final );
> +
> +    /* Entire access within a single frame? */
> +    if ( first == final )
> +        mapping = map_domain_page(hvmemul_ctxt->mfn[0]) + (linear &
> ~PAGE_MASK);
> +    /* Multiple frames? Need to vmap(). */
> +    else if ( (mapping = vmap(hvmemul_ctxt->mfn,
> +                              mfn - hvmemul_ctxt->mfn)) == NULL )
> +        goto unhandleable;
> +
> +#ifndef NDEBUG /* Poision unused mfn[]s with INVALID_MFN. */
> +    while ( mfn < hvmemul_ctxt->mfn + ARRAY_SIZE(hvmemul_ctxt->mfn) )
> +    {
> +        ASSERT(mfn_x(*mfn) == 0);
> +        *mfn++ = INVALID_MFN;
> +    }
> +#endif
> +
> +    return mapping;
> +
> + unhandleable:
> +    err = ERR_PTR(~(long)X86EMUL_UNHANDLEABLE);
> +
> + out:
> +    /* Drop all held references. */
> +    while ( mfn > hvmemul_ctxt->mfn )
> +        put_page(mfn_to_page(mfn_x(*mfn--)));
> +
> +    return err;
> +}
> +
> +static void hvmemul_unmap_linear_addr(
> +    void *mapping, unsigned long linear, unsigned int bytes,
> +    struct hvm_emulate_ctxt *hvmemul_ctxt)
> +{
> +    struct domain *currd = current->domain;
> +    unsigned long frame = linear >> PAGE_SHIFT;
> +    unsigned long final = (linear + bytes - !!bytes) >> PAGE_SHIFT;
> +    mfn_t *mfn = &hvmemul_ctxt->mfn[0];
> +
> +    ASSERT(bytes > 0);

Why not return if bytes == 0? I know it's not a legitimate call but in a 
non-debug build it would result in unmap_domain_page() being called below.

  Paul

> +
> +    if ( frame == final )
> +        unmap_domain_page(mapping);
> +    else
> +        vunmap(mapping);
> +
> +    do
> +    {
> +        ASSERT(mfn_valid(*mfn));
> +        paging_mark_dirty(currd, *mfn);
> +        put_page(mfn_to_page(mfn_x(*mfn)));
> +
> +        frame++;
> +        *mfn++ = _mfn(0); /* Clean slot for map()'s error checking. */
> +
> +    } while ( frame < final );
> +
> +
> +#ifndef NDEBUG /* Check (and clean) all unused mfns. */
> +    while ( mfn < hvmemul_ctxt->mfn + ARRAY_SIZE(hvmemul_ctxt->mfn) )
> +    {
> +        ASSERT(mfn_eq(*mfn, INVALID_MFN));
> +        *mfn++ = _mfn(0);
> +    }
> +#endif
> +}
> +
> +/*
>   * Convert addr from linear to physical form, valid over the range
>   * [addr, addr + *reps * bytes_per_rep]. *reps is adjusted according to
>   * the valid computed range. It is always >0 when X86EMUL_OKAY is
> returned.
> @@ -987,11 +1140,11 @@ static int hvmemul_write(
>      struct hvm_emulate_ctxt *hvmemul_ctxt =
>          container_of(ctxt, struct hvm_emulate_ctxt, ctxt);
>      struct vcpu *curr = current;
> -    pagefault_info_t pfinfo;
>      unsigned long addr, reps = 1;
>      uint32_t pfec = PFEC_page_present | PFEC_write_access;
>      struct hvm_vcpu_io *vio = &curr->arch.hvm_vcpu.hvm_io;
>      int rc;
> +    void *mapping;
> 
>      if ( is_x86_system_segment(seg) )
>          pfec |= PFEC_implicit;
> @@ -1007,23 +1160,15 @@ static int hvmemul_write(
>           (vio->mmio_gla == (addr & PAGE_MASK)) )
>          return hvmemul_linear_mmio_write(addr, bytes, p_data, pfec,
> hvmemul_ctxt, 1);
> 
> -    rc = hvm_copy_to_guest_linear(addr, p_data, bytes, pfec, &pfinfo);
> -
> -    switch ( rc )
> -    {
> -    case HVMTRANS_okay:
> -        break;
> -    case HVMTRANS_bad_linear_to_gfn:
> -        x86_emul_pagefault(pfinfo.ec, pfinfo.linear, &hvmemul_ctxt->ctxt);
> -        return X86EMUL_EXCEPTION;
> -    case HVMTRANS_bad_gfn_to_mfn:
> +    mapping = hvmemul_map_linear_addr(addr, bytes, pfec,
> hvmemul_ctxt);
> +    if ( IS_ERR(mapping) )
> +        return ~PTR_ERR(mapping);
> +    else if ( !mapping )
>          return hvmemul_linear_mmio_write(addr, bytes, p_data, pfec,
> hvmemul_ctxt, 0);
> -    case HVMTRANS_gfn_paged_out:
> -    case HVMTRANS_gfn_shared:
> -        return X86EMUL_RETRY;
> -    default:
> -        return X86EMUL_UNHANDLEABLE;
> -    }
> +
> +    memcpy(mapping, p_data, bytes);
> +
> +    hvmemul_unmap_linear_addr(mapping, addr, bytes, hvmemul_ctxt);
> 
>      return X86EMUL_OKAY;
>  }
> diff --git a/xen/include/asm-x86/hvm/emulate.h b/xen/include/asm-
> x86/hvm/emulate.h
> index 8864775..65efd4e 100644
> --- a/xen/include/asm-x86/hvm/emulate.h
> +++ b/xen/include/asm-x86/hvm/emulate.h
> @@ -37,6 +37,13 @@ struct hvm_emulate_ctxt {
>      unsigned long seg_reg_accessed;
>      unsigned long seg_reg_dirty;
> 
> +    /*
> +     * MFNs behind temporary mappings in the write callback.  The length is
> +     * arbitrary, and can be increased if writes longer than PAGE_SIZE are
> +     * needed.
> +     */
> +    mfn_t mfn[2];
> +
>      uint32_t intr_shadow;
> 
>      bool_t set_context;
> --
> 2.1.4

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
https://lists.xen.org/xen-devel