[v9,6/8] common/domain: add a domain context record for shared_info...

Commit Message

Paul Durrant Sept. 24, 2020, 1:10 p.m. UTC

From: Paul Durrant <[e-mail]>

... and update xen-domctx to dump some information describing the record.

NOTE: Processing of the content during restore is currently limited to
      PV domains, and matches processing of the PV-only SHARED_INFO record
      done by libxc. All content is, however, saved such that restore
      processing can be modified in future without requiring a new record
      format.

Signed-off-by: Paul Durrant <[e-mail]>
---
Cc: Ian Jackson <[e-mail]>
Cc: Wei Liu <[e-mail]>
Cc: Andrew Cooper <[e-mail]>
Cc: George Dunlap <[e-mail]>
Cc: Jan Beulich <[e-mail]>
Cc: Julien Grall <[e-mail]>
Cc: Stefano Stabellini <[e-mail]>

v9:
 - Use macros to make the code less verbose
 - Add missing check for allocation failure

v8:
 - Incorporate zero-ing out of shared info fields that would be done in
   processing of SHARED_INFO from older stream versions

v7:
 - Only restore vcpu_info and arch sub-structures for PV domains, to match
   processing of SHARED_INFO in xc_sr_restore_x86_pv.c
 - Use additional option to domain_load_end() to ignore the record for
   HVM domains

v6:
 - Only save compat_shared_info buffer if has_32bit_shinfo is set
 - Validate flags field in load handler

v5:
 - Addressed comments from Julien

v4:
 - Addressed comments from Jan

v3:
 - Actually dump some of the content of shared_info

v2:
 - Drop the header change to define a 'Xen' page size and instead use a
   variable length struct now that the framework makes this is feasible
 - Guard use of 'has_32bit_shinfo' in common code with CONFIG_COMPAT
---
 tools/misc/xen-domctx.c   |  78 ++++++++++++++++++++++++++++
 xen/common/domain.c       | 105 ++++++++++++++++++++++++++++++++++++++
 xen/include/public/save.h |  13 ++++-
 3 files changed, 195 insertions(+), 1 deletion(-)

Comments

Jan Beulich Sept. 25, 2020, 12:44 p.m. UTC | #1

On 24.09.2020 15:10, Paul Durrant wrote:
> From: Paul Durrant <[e-mail]>
> 
> ... and update xen-domctx to dump some information describing the record.
> 
> NOTE: Processing of the content during restore is currently limited to
>       PV domains, and matches processing of the PV-only SHARED_INFO record
>       done by libxc. All content is, however, saved such that restore
>       processing can be modified in future without requiring a new record
>       format.
> 
> Signed-off-by: Paul Durrant <[e-mail]>

Reviewed-by: Jan Beulich <[e-mail]>

Wei Liu Sept. 30, 2020, 2:42 p.m. UTC | #2

On Thu, Sep 24, 2020 at 02:10:28PM +0100, Paul Durrant wrote:
> From: Paul Durrant <[e-mail]>
> 
> ... and update xen-domctx to dump some information describing the record.
> 
> NOTE: Processing of the content during restore is currently limited to
>       PV domains, and matches processing of the PV-only SHARED_INFO record
>       done by libxc. All content is, however, saved such that restore
>       processing can be modified in future without requiring a new record
>       format.
> 
> Signed-off-by: Paul Durrant <[e-mail]>

Acked-by: Wei Liu <[e-mail]>

Andrew Cooper Oct. 5, 2020, 10:39 a.m. UTC | #3

On 24/09/2020 14:10, Paul Durrant wrote:
> diff --git a/tools/misc/xen-domctx.c b/tools/misc/xen-domctx.c
> index 243325dfce..6ead7ea89d 100644
> --- a/tools/misc/xen-domctx.c
> +++ b/tools/misc/xen-domctx.c
> @@ -31,6 +31,7 @@
>  #include <errno.h>
>  
>  #include <xenctrl.h>
> +#include <xen-tools/libs.h>
>  #include <xen/xen.h>
>  #include <xen/domctl.h>
>  #include <xen/save.h>
> @@ -61,6 +62,82 @@ static void dump_header(void)
>  
>  }
>  
> +static void print_binary(const char *prefix, const void *val, size_t size,
> +                         const char *suffix)
> +{
> +    printf("%s", prefix);
> +
> +    while ( size-- )
> +    {
> +        uint8_t octet = *(const uint8_t *)val++;
> +        unsigned int i;
> +
> +        for ( i = 0; i < 8; i++ )
> +        {
> +            printf("%u", octet & 1);
> +            octet >>= 1;
> +        }
> +    }
> +
> +    printf("%s", suffix);
> +}
> +
> +static void dump_shared_info(void)
> +{
> +    DOMAIN_SAVE_TYPE(SHARED_INFO) *s;
> +    bool has_32bit_shinfo;
> +    shared_info_any_t *info;
> +    unsigned int i, n;
> +
> +    GET_PTR(s);
> +    has_32bit_shinfo = s->flags & DOMAIN_SAVE_32BIT_SHINFO;
> +
> +    printf("    SHARED_INFO: has_32bit_shinfo: %s buffer_size: %u\n",
> +           has_32bit_shinfo ? "true" : "false", s->buffer_size);
> +
> +    info = (shared_info_any_t *)s->buffer;
> +
> +#define GET_FIELD_PTR(_f)            \
> +    (has_32bit_shinfo ?              \
> +     (const void *)&(info->x32._f) : \
> +     (const void *)&(info->x64._f))
> +#define GET_FIELD_SIZE(_f) \
> +    (has_32bit_shinfo ? sizeof(info->x32._f) : sizeof(info->x64._f))
> +#define GET_FIELD(_f) \
> +    (has_32bit_shinfo ? info->x32._f : info->x64._f)
> +
> +    n = has_32bit_shinfo ?
> +        ARRAY_SIZE(info->x32.evtchn_pending) :
> +        ARRAY_SIZE(info->x64.evtchn_pending);
> +
> +    for ( i = 0; i < n; i++ )
> +    {
> +        const char *prefix = !i ?
> +            "                 evtchn_pending: " :
> +            "                                 ";
> +
> +        print_binary(prefix, GET_FIELD_PTR(evtchn_pending[0]),
> +                 GET_FIELD_SIZE(evtchn_pending[0]), "\n");
> +    }
> +
> +    for ( i = 0; i < n; i++ )
> +    {
> +        const char *prefix = !i ?
> +            "                    evtchn_mask: " :
> +            "                                 ";
> +
> +        print_binary(prefix, GET_FIELD_PTR(evtchn_mask[0]),
> +                 GET_FIELD_SIZE(evtchn_mask[0]), "\n");
> +    }

What about domains using FIFO?  This is meaningless for them.

> +
> +    printf("                 wc: version: %u sec: %u nsec: %u\n",
> +           GET_FIELD(wc_version), GET_FIELD(wc_sec), GET_FIELD(wc_nsec));

wc_sec_hi is also a rather critical field in this calculation.

> +
> +#undef GET_FIELD
> +#undef GET_FIELD_SIZE
> +#undef GET_FIELD_PTR
> +}
> +
>  static void dump_end(void)
>  {
>      DOMAIN_SAVE_TYPE(END) *e;
> @@ -173,6 +250,7 @@ int main(int argc, char **argv)
>              switch (desc->typecode)
>              {
>              case DOMAIN_SAVE_CODE(HEADER): dump_header(); break;
> +            case DOMAIN_SAVE_CODE(SHARED_INFO): dump_shared_info(); break;
>              case DOMAIN_SAVE_CODE(END): dump_end(); break;
>              default:
>                  printf("Unknown type %u: skipping\n", desc->typecode);
> diff --git a/xen/common/domain.c b/xen/common/domain.c
> index 8cfa2e0b6b..6709f9c79e 100644
> --- a/xen/common/domain.c
> +++ b/xen/common/domain.c
> @@ -33,6 +33,7 @@
>  #include <xen/xenoprof.h>
>  #include <xen/irq.h>
>  #include <xen/argo.h>
> +#include <xen/save.h>
>  #include <asm/debugger.h>
>  #include <asm/p2m.h>
>  #include <asm/processor.h>
> @@ -1657,6 +1658,110 @@ int continue_hypercall_on_cpu(
>      return 0;
>  }
>  
> +static int save_shared_info(const struct domain *d, struct domain_context *c,
> +                            bool dry_run)
> +{
> +    struct domain_shared_info_context ctxt = {
> +#ifdef CONFIG_COMPAT
> +        .flags = has_32bit_shinfo(d) ? DOMAIN_SAVE_32BIT_SHINFO : 0,
> +        .buffer_size = has_32bit_shinfo(d) ?
> +                       sizeof(struct compat_shared_info) :
> +                       sizeof(struct shared_info),
> +#else
> +        .buffer_size = sizeof(struct shared_info),
> +#endif
> +    };
> +    size_t hdr_size = offsetof(typeof(ctxt), buffer);
> +    int rc;
> +
> +    rc = DOMAIN_SAVE_BEGIN(SHARED_INFO, c, 0);
> +    if ( rc )
> +        return rc;
> +
> +    rc = domain_save_data(c, &ctxt, hdr_size);
> +    if ( rc )
> +        return rc;
> +
> +    rc = domain_save_data(c, d->shared_info, ctxt.buffer_size);
> +    if ( rc )
> +        return rc;
> +
> +    return domain_save_end(c);
> +}
> +
> +static int load_shared_info(struct domain *d, struct domain_context *c)
> +{
> +    struct domain_shared_info_context ctxt;
> +    size_t hdr_size = offsetof(typeof(ctxt), buffer);
> +    unsigned int i;
> +    int rc;
> +
> +    rc = DOMAIN_LOAD_BEGIN(SHARED_INFO, c, &i);
> +    if ( rc )
> +        return rc;
> +
> +    if ( i ) /* expect only a single instance */
> +        return -ENXIO;
> +
> +    rc = domain_load_data(c, &ctxt, hdr_size);
> +    if ( rc )
> +        return rc;
> +
> +    if ( ctxt.buffer_size > sizeof(shared_info_t) ||
> +         (ctxt.flags & ~DOMAIN_SAVE_32BIT_SHINFO) )
> +        return -EINVAL;
> +
> +    if ( ctxt.flags & DOMAIN_SAVE_32BIT_SHINFO )
> +    {
> +#ifdef CONFIG_COMPAT
> +        has_32bit_shinfo(d) = true;

d->arch.has_32bit_shinfo

> +#else
> +        return -EINVAL;
> +#endif
> +    }
> +
> +    if ( is_pv_domain(d) )
> +    {
> +        shared_info_t *shinfo = xmalloc(shared_info_t);
> +
> +        if ( !shinfo )
> +            return -ENOMEM;
> +
> +        rc = domain_load_data(c, shinfo, sizeof(*shinfo));
> +        if ( rc )
> +            goto out;

There's no need for a memory allocation, or to double buffer this data. 
You can memcpy() straight out of the context record.

> +
> +        memcpy(&shared_info(d, vcpu_info), &__shared_info(d, shinfo, vcpu_info),
> +               sizeof(shared_info(d, vcpu_info)));
> +        memcpy(&shared_info(d, arch), &__shared_info(d, shinfo, arch),
> +               sizeof(shared_info(d, arch)));
> +
> +        memset(&shared_info(d, evtchn_pending), 0,
> +               sizeof(shared_info(d, evtchn_pending)));
> +        memset(&shared_info(d, evtchn_mask), 0xff,
> +               sizeof(shared_info(d, evtchn_mask)));
> +
> +        shared_info(d, arch.pfn_to_mfn_frame_list_list) = 0;
> +        for ( i = 0; i < XEN_LEGACY_MAX_VCPUS; i++ )
> +            shared_info(d, vcpu_info[i].evtchn_pending_sel) = 0;

What is the plan for transparent migrate here?  While this is ok for
regular migrate, its definitely not for transparent.

> +
> +        rc = domain_load_end(c, false);
> +
> +    out:
> +        xfree(shinfo);
> +    }
> +    else
> +        /*
> +         * No modifications to shared_info are required for restoring non-PV
> +         * domains.
> +         */
> +        rc = domain_load_end(c, true);
> +
> +    return rc;
> +}
> +
> +DOMAIN_REGISTER_SAVE_LOAD(SHARED_INFO, save_shared_info, load_shared_info);
> +
>  /*
>   * Local variables:
>   * mode: C
> diff --git a/xen/include/public/save.h b/xen/include/public/save.h
> index 551dbbddb8..0e855a4b97 100644
> --- a/xen/include/public/save.h
> +++ b/xen/include/public/save.h
> @@ -82,7 +82,18 @@ struct domain_save_header {
>  };
>  DECLARE_DOMAIN_SAVE_TYPE(HEADER, 1, struct domain_save_header);
>  
> -#define DOMAIN_SAVE_CODE_MAX 1
> +struct domain_shared_info_context {
> +    uint32_t flags;
> +
> +#define DOMAIN_SAVE_32BIT_SHINFO 0x00000001
> +
> +    uint32_t buffer_size;

This struct is already wrapped with a header including a size which
encompasses buffer.

Multiple overlapping size fields is an easy way to memory corruption,
because it causes ambiguity as to which one is right.

~Andrew

Paul Durrant Oct. 7, 2020, 12:03 p.m. UTC | #4

> -----Original Message-----
> From: Andrew Cooper <[e-mail]>
> Sent: 05 October 2020 11:40
> To: Paul Durrant <[e-mail]>; [e-mail]
> Cc: Paul Durrant <[e-mail]>; Ian Jackson <[e-mail]>; Wei Liu <[e-mail]>;
> George Dunlap <[e-mail]>; Jan Beulich <[e-mail]>; Julien Grall
> <[e-mail]>; Stefano Stabellini <[e-mail]>
> Subject: Re: [PATCH v9 6/8] common/domain: add a domain context record for shared_info...
> 
> On 24/09/2020 14:10, Paul Durrant wrote:
> > diff --git a/tools/misc/xen-domctx.c b/tools/misc/xen-domctx.c
> > index 243325dfce..6ead7ea89d 100644
> > --- a/tools/misc/xen-domctx.c
> > +++ b/tools/misc/xen-domctx.c
> > @@ -31,6 +31,7 @@
> >  #include <errno.h>
> >
> >  #include <xenctrl.h>
> > +#include <xen-tools/libs.h>
> >  #include <xen/xen.h>
> >  #include <xen/domctl.h>
> >  #include <xen/save.h>
> > @@ -61,6 +62,82 @@ static void dump_header(void)
> >
> >  }
> >
> > +static void print_binary(const char *prefix, const void *val, size_t size,
> > +                         const char *suffix)
> > +{
> > +    printf("%s", prefix);
> > +
> > +    while ( size-- )
> > +    {
> > +        uint8_t octet = *(const uint8_t *)val++;
> > +        unsigned int i;
> > +
> > +        for ( i = 0; i < 8; i++ )
> > +        {
> > +            printf("%u", octet & 1);
> > +            octet >>= 1;
> > +        }
> > +    }
> > +
> > +    printf("%s", suffix);
> > +}
> > +
> > +static void dump_shared_info(void)
> > +{
> > +    DOMAIN_SAVE_TYPE(SHARED_INFO) *s;
> > +    bool has_32bit_shinfo;
> > +    shared_info_any_t *info;
> > +    unsigned int i, n;
> > +
> > +    GET_PTR(s);
> > +    has_32bit_shinfo = s->flags & DOMAIN_SAVE_32BIT_SHINFO;
> > +
> > +    printf("    SHARED_INFO: has_32bit_shinfo: %s buffer_size: %u\n",
> > +           has_32bit_shinfo ? "true" : "false", s->buffer_size);
> > +
> > +    info = (shared_info_any_t *)s->buffer;
> > +
> > +#define GET_FIELD_PTR(_f)            \
> > +    (has_32bit_shinfo ?              \
> > +     (const void *)&(info->x32._f) : \
> > +     (const void *)&(info->x64._f))
> > +#define GET_FIELD_SIZE(_f) \
> > +    (has_32bit_shinfo ? sizeof(info->x32._f) : sizeof(info->x64._f))
> > +#define GET_FIELD(_f) \
> > +    (has_32bit_shinfo ? info->x32._f : info->x64._f)
> > +
> > +    n = has_32bit_shinfo ?
> > +        ARRAY_SIZE(info->x32.evtchn_pending) :
> > +        ARRAY_SIZE(info->x64.evtchn_pending);
> > +
> > +    for ( i = 0; i < n; i++ )
> > +    {
> > +        const char *prefix = !i ?
> > +            "                 evtchn_pending: " :
> > +            "                                 ";
> > +
> > +        print_binary(prefix, GET_FIELD_PTR(evtchn_pending[0]),
> > +                 GET_FIELD_SIZE(evtchn_pending[0]), "\n");
> > +    }
> > +
> > +    for ( i = 0; i < n; i++ )
> > +    {
> > +        const char *prefix = !i ?
> > +            "                    evtchn_mask: " :
> > +            "                                 ";
> > +
> > +        print_binary(prefix, GET_FIELD_PTR(evtchn_mask[0]),
> > +                 GET_FIELD_SIZE(evtchn_mask[0]), "\n");
> > +    }
> 
> What about domains using FIFO?  This is meaningless for them.
> 

Indeed, but this is essentially a debug tool so I'd rather it just dumped everything that might be useful.

> > +
> > +    printf("                 wc: version: %u sec: %u nsec: %u\n",
> > +           GET_FIELD(wc_version), GET_FIELD(wc_sec), GET_FIELD(wc_nsec));
> 
> wc_sec_hi is also a rather critical field in this calculation.	
> 

Ok.

> > +
> > +#undef GET_FIELD
> > +#undef GET_FIELD_SIZE
> > +#undef GET_FIELD_PTR
> > +}
> > +
> >  static void dump_end(void)
> >  {
> >      DOMAIN_SAVE_TYPE(END) *e;
> > @@ -173,6 +250,7 @@ int main(int argc, char **argv)
> >              switch (desc->typecode)
> >              {
> >              case DOMAIN_SAVE_CODE(HEADER): dump_header(); break;
> > +            case DOMAIN_SAVE_CODE(SHARED_INFO): dump_shared_info(); break;
> >              case DOMAIN_SAVE_CODE(END): dump_end(); break;
> >              default:
> >                  printf("Unknown type %u: skipping\n", desc->typecode);
> > diff --git a/xen/common/domain.c b/xen/common/domain.c
> > index 8cfa2e0b6b..6709f9c79e 100644
> > --- a/xen/common/domain.c
> > +++ b/xen/common/domain.c
> > @@ -33,6 +33,7 @@
> >  #include <xen/xenoprof.h>
> >  #include <xen/irq.h>
> >  #include <xen/argo.h>
> > +#include <xen/save.h>
> >  #include <asm/debugger.h>
> >  #include <asm/p2m.h>
> >  #include <asm/processor.h>
> > @@ -1657,6 +1658,110 @@ int continue_hypercall_on_cpu(
> >      return 0;
> >  }
> >
> > +static int save_shared_info(const struct domain *d, struct domain_context *c,
> > +                            bool dry_run)
> > +{
> > +    struct domain_shared_info_context ctxt = {
> > +#ifdef CONFIG_COMPAT
> > +        .flags = has_32bit_shinfo(d) ? DOMAIN_SAVE_32BIT_SHINFO : 0,
> > +        .buffer_size = has_32bit_shinfo(d) ?
> > +                       sizeof(struct compat_shared_info) :
> > +                       sizeof(struct shared_info),
> > +#else
> > +        .buffer_size = sizeof(struct shared_info),
> > +#endif
> > +    };
> > +    size_t hdr_size = offsetof(typeof(ctxt), buffer);
> > +    int rc;
> > +
> > +    rc = DOMAIN_SAVE_BEGIN(SHARED_INFO, c, 0);
> > +    if ( rc )
> > +        return rc;
> > +
> > +    rc = domain_save_data(c, &ctxt, hdr_size);
> > +    if ( rc )
> > +        return rc;
> > +
> > +    rc = domain_save_data(c, d->shared_info, ctxt.buffer_size);
> > +    if ( rc )
> > +        return rc;
> > +
> > +    return domain_save_end(c);
> > +}
> > +
> > +static int load_shared_info(struct domain *d, struct domain_context *c)
> > +{
> > +    struct domain_shared_info_context ctxt;
> > +    size_t hdr_size = offsetof(typeof(ctxt), buffer);
> > +    unsigned int i;
> > +    int rc;
> > +
> > +    rc = DOMAIN_LOAD_BEGIN(SHARED_INFO, c, &i);
> > +    if ( rc )
> > +        return rc;
> > +
> > +    if ( i ) /* expect only a single instance */
> > +        return -ENXIO;
> > +
> > +    rc = domain_load_data(c, &ctxt, hdr_size);
> > +    if ( rc )
> > +        return rc;
> > +
> > +    if ( ctxt.buffer_size > sizeof(shared_info_t) ||
> > +         (ctxt.flags & ~DOMAIN_SAVE_32BIT_SHINFO) )
> > +        return -EINVAL;
> > +
> > +    if ( ctxt.flags & DOMAIN_SAVE_32BIT_SHINFO )
> > +    {
> > +#ifdef CONFIG_COMPAT
> > +        has_32bit_shinfo(d) = true;
> 
> d->arch.has_32bit_shinfo
> 

If you'd prefer, ok.

> > +#else
> > +        return -EINVAL;
> > +#endif
> > +    }
> > +
> > +    if ( is_pv_domain(d) )
> > +    {
> > +        shared_info_t *shinfo = xmalloc(shared_info_t);
> > +
> > +        if ( !shinfo )
> > +            return -ENOMEM;
> > +
> > +        rc = domain_load_data(c, shinfo, sizeof(*shinfo));
> > +        if ( rc )
> > +            goto out;
> 
> There's no need for a memory allocation, or to double buffer this data.
> You can memcpy() straight out of the context record.
> 

That would mean re-working the way that domain_load_data() works. I'd really rather not.

> > +
> > +        memcpy(&shared_info(d, vcpu_info), &__shared_info(d, shinfo, vcpu_info),
> > +               sizeof(shared_info(d, vcpu_info)));
> > +        memcpy(&shared_info(d, arch), &__shared_info(d, shinfo, arch),
> > +               sizeof(shared_info(d, arch)));
> > +
> > +        memset(&shared_info(d, evtchn_pending), 0,
> > +               sizeof(shared_info(d, evtchn_pending)));
> > +        memset(&shared_info(d, evtchn_mask), 0xff,
> > +               sizeof(shared_info(d, evtchn_mask)));
> > +
> > +        shared_info(d, arch.pfn_to_mfn_frame_list_list) = 0;
> > +        for ( i = 0; i < XEN_LEGACY_MAX_VCPUS; i++ )
> > +            shared_info(d, vcpu_info[i].evtchn_pending_sel) = 0;
> 
> What is the plan for transparent migrate here?  While this is ok for
> regular migrate, its definitely not for transparent.
> 

Quite true, as evidenced that this is inside 'if ( is_pv_domain(d) )'. It is not yet clear how much of the shared info we need for transparent migrate. It may be nothing.

> > +
> > +        rc = domain_load_end(c, false);
> > +
> > +    out:
> > +        xfree(shinfo);
> > +    }
> > +    else
> > +        /*
> > +         * No modifications to shared_info are required for restoring non-PV
> > +         * domains.
> > +         */
> > +        rc = domain_load_end(c, true);
> > +
> > +    return rc;
> > +}
> > +
> > +DOMAIN_REGISTER_SAVE_LOAD(SHARED_INFO, save_shared_info, load_shared_info);
> > +
> >  /*
> >   * Local variables:
> >   * mode: C
> > diff --git a/xen/include/public/save.h b/xen/include/public/save.h
> > index 551dbbddb8..0e855a4b97 100644
> > --- a/xen/include/public/save.h
> > +++ b/xen/include/public/save.h
> > @@ -82,7 +82,18 @@ struct domain_save_header {
> >  };
> >  DECLARE_DOMAIN_SAVE_TYPE(HEADER, 1, struct domain_save_header);
> >
> > -#define DOMAIN_SAVE_CODE_MAX 1
> > +struct domain_shared_info_context {
> > +    uint32_t flags;
> > +
> > +#define DOMAIN_SAVE_32BIT_SHINFO 0x00000001
> > +
> > +    uint32_t buffer_size;
> 
> This struct is already wrapped with a header including a size which
> encompasses buffer.
> 
> Multiple overlapping size fields is an easy way to memory corruption,
> because it causes ambiguity as to which one is right.
> 

The record size currently includes padding. I'm re-working that in v10 and so this size can be dropped.

  Paul

Jan Beulich Oct. 13, 2020, 11:49 a.m. UTC | #5

On 05.10.2020 12:39, Andrew Cooper wrote:
> On 24/09/2020 14:10, Paul Durrant wrote:
>> +static int load_shared_info(struct domain *d, struct domain_context *c)
>> +{
>> +    struct domain_shared_info_context ctxt;
>> +    size_t hdr_size = offsetof(typeof(ctxt), buffer);
>> +    unsigned int i;
>> +    int rc;
>> +
>> +    rc = DOMAIN_LOAD_BEGIN(SHARED_INFO, c, &i);
>> +    if ( rc )
>> +        return rc;
>> +
>> +    if ( i ) /* expect only a single instance */
>> +        return -ENXIO;
>> +
>> +    rc = domain_load_data(c, &ctxt, hdr_size);
>> +    if ( rc )
>> +        return rc;
>> +
>> +    if ( ctxt.buffer_size > sizeof(shared_info_t) ||
>> +         (ctxt.flags & ~DOMAIN_SAVE_32BIT_SHINFO) )
>> +        return -EINVAL;
>> +
>> +    if ( ctxt.flags & DOMAIN_SAVE_32BIT_SHINFO )
>> +    {
>> +#ifdef CONFIG_COMPAT
>> +        has_32bit_shinfo(d) = true;
> 
> d->arch.has_32bit_shinfo

But this is common code, i.e. using d->arch directly is a layering
violation. I know your dislike of lvalues disguised by function-
like macros, but what do you do?

Jan

Patch

diff --git a/tools/misc/xen-domctx.c b/tools/misc/xen-domctx.c
index 243325dfce..6ead7ea89d 100644
--- a/tools/misc/xen-domctx.c
+++ b/tools/misc/xen-domctx.c
@@ -31,6 +31,7 @@ 
 #include <errno.h>
 
 #include <xenctrl.h>
+#include <xen-tools/libs.h>
 #include <xen/xen.h>
 #include <xen/domctl.h>
 #include <xen/save.h>
@@ -61,6 +62,82 @@  static void dump_header(void)
 
 }
 
+static void print_binary(const char *prefix, const void *val, size_t size,
+                         const char *suffix)
+{
+    printf("%s", prefix);
+
+    while ( size-- )
+    {
+        uint8_t octet = *(const uint8_t *)val++;
+        unsigned int i;
+
+        for ( i = 0; i < 8; i++ )
+        {
+            printf("%u", octet & 1);
+            octet >>= 1;
+        }
+    }
+
+    printf("%s", suffix);
+}
+
+static void dump_shared_info(void)
+{
+    DOMAIN_SAVE_TYPE(SHARED_INFO) *s;
+    bool has_32bit_shinfo;
+    shared_info_any_t *info;
+    unsigned int i, n;
+
+    GET_PTR(s);
+    has_32bit_shinfo = s->flags & DOMAIN_SAVE_32BIT_SHINFO;
+
+    printf("    SHARED_INFO: has_32bit_shinfo: %s buffer_size: %u\n",
+           has_32bit_shinfo ? "true" : "false", s->buffer_size);
+
+    info = (shared_info_any_t *)s->buffer;
+
+#define GET_FIELD_PTR(_f)            \
+    (has_32bit_shinfo ?              \
+     (const void *)&(info->x32._f) : \
+     (const void *)&(info->x64._f))
+#define GET_FIELD_SIZE(_f) \
+    (has_32bit_shinfo ? sizeof(info->x32._f) : sizeof(info->x64._f))
+#define GET_FIELD(_f) \
+    (has_32bit_shinfo ? info->x32._f : info->x64._f)
+
+    n = has_32bit_shinfo ?
+        ARRAY_SIZE(info->x32.evtchn_pending) :
+        ARRAY_SIZE(info->x64.evtchn_pending);
+
+    for ( i = 0; i < n; i++ )
+    {
+        const char *prefix = !i ?
+            "                 evtchn_pending: " :
+            "                                 ";
+
+        print_binary(prefix, GET_FIELD_PTR(evtchn_pending[0]),
+                 GET_FIELD_SIZE(evtchn_pending[0]), "\n");
+    }
+
+    for ( i = 0; i < n; i++ )
+    {
+        const char *prefix = !i ?
+            "                    evtchn_mask: " :
+            "                                 ";
+
+        print_binary(prefix, GET_FIELD_PTR(evtchn_mask[0]),
+                 GET_FIELD_SIZE(evtchn_mask[0]), "\n");
+    }
+
+    printf("                 wc: version: %u sec: %u nsec: %u\n",
+           GET_FIELD(wc_version), GET_FIELD(wc_sec), GET_FIELD(wc_nsec));
+
+#undef GET_FIELD
+#undef GET_FIELD_SIZE
+#undef GET_FIELD_PTR
+}
+
 static void dump_end(void)
 {
     DOMAIN_SAVE_TYPE(END) *e;
@@ -173,6 +250,7 @@  int main(int argc, char **argv)
             switch (desc->typecode)
             {
             case DOMAIN_SAVE_CODE(HEADER): dump_header(); break;
+            case DOMAIN_SAVE_CODE(SHARED_INFO): dump_shared_info(); break;
             case DOMAIN_SAVE_CODE(END): dump_end(); break;
             default:
                 printf("Unknown type %u: skipping\n", desc->typecode);
diff --git a/xen/common/domain.c b/xen/common/domain.c
index 8cfa2e0b6b..6709f9c79e 100644
--- a/xen/common/domain.c
+++ b/xen/common/domain.c
@@ -33,6 +33,7 @@ 
 #include <xen/xenoprof.h>
 #include <xen/irq.h>
 #include <xen/argo.h>
+#include <xen/save.h>
 #include <asm/debugger.h>
 #include <asm/p2m.h>
 #include <asm/processor.h>
@@ -1657,6 +1658,110 @@  int continue_hypercall_on_cpu(
     return 0;
 }
 
+static int save_shared_info(const struct domain *d, struct domain_context *c,
+                            bool dry_run)
+{
+    struct domain_shared_info_context ctxt = {
+#ifdef CONFIG_COMPAT
+        .flags = has_32bit_shinfo(d) ? DOMAIN_SAVE_32BIT_SHINFO : 0,
+        .buffer_size = has_32bit_shinfo(d) ?
+                       sizeof(struct compat_shared_info) :
+                       sizeof(struct shared_info),
+#else
+        .buffer_size = sizeof(struct shared_info),
+#endif
+    };
+    size_t hdr_size = offsetof(typeof(ctxt), buffer);
+    int rc;
+
+    rc = DOMAIN_SAVE_BEGIN(SHARED_INFO, c, 0);
+    if ( rc )
+        return rc;
+
+    rc = domain_save_data(c, &ctxt, hdr_size);
+    if ( rc )
+        return rc;
+
+    rc = domain_save_data(c, d->shared_info, ctxt.buffer_size);
+    if ( rc )
+        return rc;
+
+    return domain_save_end(c);
+}
+
+static int load_shared_info(struct domain *d, struct domain_context *c)
+{
+    struct domain_shared_info_context ctxt;
+    size_t hdr_size = offsetof(typeof(ctxt), buffer);
+    unsigned int i;
+    int rc;
+
+    rc = DOMAIN_LOAD_BEGIN(SHARED_INFO, c, &i);
+    if ( rc )
+        return rc;
+
+    if ( i ) /* expect only a single instance */
+        return -ENXIO;
+
+    rc = domain_load_data(c, &ctxt, hdr_size);
+    if ( rc )
+        return rc;
+
+    if ( ctxt.buffer_size > sizeof(shared_info_t) ||
+         (ctxt.flags & ~DOMAIN_SAVE_32BIT_SHINFO) )
+        return -EINVAL;
+
+    if ( ctxt.flags & DOMAIN_SAVE_32BIT_SHINFO )
+    {
+#ifdef CONFIG_COMPAT
+        has_32bit_shinfo(d) = true;
+#else
+        return -EINVAL;
+#endif
+    }
+
+    if ( is_pv_domain(d) )
+    {
+        shared_info_t *shinfo = xmalloc(shared_info_t);
+
+        if ( !shinfo )
+            return -ENOMEM;
+
+        rc = domain_load_data(c, shinfo, sizeof(*shinfo));
+        if ( rc )
+            goto out;
+
+        memcpy(&shared_info(d, vcpu_info), &__shared_info(d, shinfo, vcpu_info),
+               sizeof(shared_info(d, vcpu_info)));
+        memcpy(&shared_info(d, arch), &__shared_info(d, shinfo, arch),
+               sizeof(shared_info(d, arch)));
+
+        memset(&shared_info(d, evtchn_pending), 0,
+               sizeof(shared_info(d, evtchn_pending)));
+        memset(&shared_info(d, evtchn_mask), 0xff,
+               sizeof(shared_info(d, evtchn_mask)));
+
+        shared_info(d, arch.pfn_to_mfn_frame_list_list) = 0;
+        for ( i = 0; i < XEN_LEGACY_MAX_VCPUS; i++ )
+            shared_info(d, vcpu_info[i].evtchn_pending_sel) = 0;
+
+        rc = domain_load_end(c, false);
+
+    out:
+        xfree(shinfo);
+    }
+    else
+        /*
+         * No modifications to shared_info are required for restoring non-PV
+         * domains.
+         */
+        rc = domain_load_end(c, true);
+
+    return rc;
+}
+
+DOMAIN_REGISTER_SAVE_LOAD(SHARED_INFO, save_shared_info, load_shared_info);
+
 /*
  * Local variables:
  * mode: C
diff --git a/xen/include/public/save.h b/xen/include/public/save.h
index 551dbbddb8..0e855a4b97 100644
--- a/xen/include/public/save.h
+++ b/xen/include/public/save.h
@@ -82,7 +82,18 @@  struct domain_save_header {
 };
 DECLARE_DOMAIN_SAVE_TYPE(HEADER, 1, struct domain_save_header);
 
-#define DOMAIN_SAVE_CODE_MAX 1
+struct domain_shared_info_context {
+    uint32_t flags;
+
+#define DOMAIN_SAVE_32BIT_SHINFO 0x00000001
+
+    uint32_t buffer_size;
+    uint8_t buffer[XEN_FLEX_ARRAY_DIM]; /* Implementation specific size */
+};
+
+DECLARE_DOMAIN_SAVE_TYPE(SHARED_INFO, 2, struct domain_shared_info_context);
+
+#define DOMAIN_SAVE_CODE_MAX 2
 
 #endif /* defined(__XEN__) || defined(__XEN_TOOLS__) */