[CRIU] [PATCH 10/18] x86: cpu -- Add support for extended xsave area

[Dmitry Safonov]

2018-07-19 13:47 GMT+01:00 Cyrill Gorcunov <gorcunov at gmail.com>:
> cpu extensions (such as avx-512) require bigger xsave
> area to keep fpu registers set, so we allocate a page
> per process to keep them all. On checkpoint we parse
> runtime fpu features and dump them into an image and
> do reverse on restore procedure.
>
> Signed-off-by: Cyrill Gorcunov <gorcunov at gmail.com>

Reviewed-by: Dmitry Safonov <0x7f454c46 at gmail.com>

> @@ -17,6 +17,21 @@ int save_task_regs(void *x, user_regs_struct_t *regs, user_fpregs_struct_t *fpre
>
>  #define assign_reg(dst, src, e)                do { dst->e = (__typeof__(dst->e))src.e; } while (0)
>  #define assign_array(dst, src, e)      memcpy(dst->e, &src.e, sizeof(src.e))
> +#define assign_xsave(feature, xsave, member, area)                                     \
> +       do {                                                                            \
> +               if (compel_fpu_has_feature(feature)) {                                  \
> +                       uint32_t off = compel_fpu_feature_offset(feature);              \
> +                       void *from = &area[off];                                        \
> +                       size_t size = pb_repeated_size(xsave, member);                  \
> +                       size_t xsize = (size_t)compel_fpu_feature_size(feature);        \
> +                       if (xsize != size) {                                            \
> +                               pr_err("%s reported %zu bytes (expecting %zu)\n",       \
> +                                       # feature, xsize, size);                        \
> +                               return -1;                                              \
> +                       }                                                               \
> +                       memcpy(xsave->member, from, size);                              \
> +               }                                                                       \
> +       } while (0)

Lovely helper, sadly I also can't find an easy way to convert it into inline
function.


> @@ -113,6 +141,62 @@ static void alloc_tls(ThreadInfoX86 *ti, void **mempool)
>         }
>  }
>
> +static int alloc_xsave_extends(UserX86XsaveEntry *xsave)
> +{
> +       if (compel_fpu_has_feature(XFEATURE_YMM)) {
> +               xsave->n_ymmh_space     = 64;
> +               xsave->ymmh_space       = xzalloc(pb_repeated_size(xsave, ymmh_space));
> +               if (!xsave->ymmh_space)
> +                       goto err;
> +       }
> +
> +       if (compel_fpu_has_feature(XFEATURE_BNDREGS)) {
> +               xsave->n_bndreg_state   = 4 * 2;
> +               xsave->bndreg_state     = xzalloc(pb_repeated_size(xsave, bndreg_state));
> +               if (!xsave->bndreg_state)
> +                       goto err;
> +       }
> +
> +       if (compel_fpu_has_feature(XFEATURE_BNDCSR)) {
> +               xsave->n_bndcsr_state   = 2;
> +               xsave->bndcsr_state     = xzalloc(pb_repeated_size(xsave, bndcsr_state));
> +               if (!xsave->bndcsr_state)
> +                       goto err;
> +       }
> +
> +       if (compel_fpu_has_feature(XFEATURE_OPMASK)) {
> +               xsave->n_opmask_reg     = 8;
> +               xsave->opmask_reg       = xzalloc(pb_repeated_size(xsave, opmask_reg));
> +               if (!xsave->opmask_reg)
> +                       goto err;
> +       }
> +
> +       if (compel_fpu_has_feature(XFEATURE_ZMM_Hi256)) {
> +               xsave->n_zmm_upper      = 16 * 4;
> +               xsave->zmm_upper        = xzalloc(pb_repeated_size(xsave, zmm_upper));
> +               if (!xsave->zmm_upper)
> +                       goto err;
> +       }
> +
> +       if (compel_fpu_has_feature(XFEATURE_Hi16_ZMM)) {
> +               xsave->n_hi16_zmm       = 16 * 8;
> +               xsave->hi16_zmm         = xzalloc(pb_repeated_size(xsave, hi16_zmm));
> +               if (!xsave->hi16_zmm)
> +                       goto err;
> +       }
> +
> +       if (compel_fpu_has_feature(XFEATURE_PKRU)) {
> +               xsave->n_pkru           = 2;
> +               xsave->pkru             = xzalloc(pb_repeated_size(xsave, pkru));
> +               if (!xsave->pkru)
> +                       goto err;
> +       }
> +
> +       return 0;
> +err:
> +       return -1;

Probably, Coverity will be unhappy about non-freeing them on error-path.

Thanks,
             Dmitry