[PATCH v15 03/10] x86/shadow: 1-bit-disable doesn't need to detach old tables
Jan Beulich <[email protected]> Tue, 28 Jul 2026 16:21:22 +0200
| Newsgroups | gmane.comp.emulators.xen.devel |
|---|---|
| Message-ID | <[email protected]> |
Just ahead of the loop being modified sh_new_mode() is called, which in turn calls sh_update_paging_modes() and thus sh_detach_old_tables(). To retain the intended effect of not leaving any shadows in place, avoid re-establishing a new top-level shadow from sh_update_paging_modes(). Signed-off-by: Jan Beulich <[email protected]> Acked-by: Roger Pau Monné <[email protected]> --- This doesn't need to be part of the XSA, but it eliminates one point of concern wrt preemption checking. However, in the context of "x86/mm: make more of log-dirty mode enable/disable preemptable" I started wondering whether things don't want doing differently. 1-bit-enables technically don't need to purge all shadows: - log-dirty enable only needs to get rid of all L1 entries; this could be done via hash lookup (of all L1/FL1 shadows) instead of shadow_blow_tables(), - test enable doesn't need any purging; it may be the intention though that it does a certain amount of purging, - force enable doesn't need to do any purging either. We could therefore pass a new boolean through sh_new_mode() to sh_update_paging_modes() to suppress the call to sh_detach_old_tables() for some (all?) of the 1-bit enables. 1-bit disables (log-dirty as well as test) match the above, but of course they will need to (at least) detach all tables when shadow mode is being turned off altogether (i.e. the very invocation that's being deleted here). As done in the other patch, the amount of work to do by the (then possibly retained) sh_detach_old_tables() here could be bounded by "preparatory" purging (in a properly preemptable way) in shadow_one_bit_disable() or its callers. Using shadow_blow_tables() for that purpose is probably preferable anyway, as sh_detach_old_tables() alone won't get rid of pinned shadows. --- v15: Re-base. v13: New. --- a/xen/arch/x86/mm/shadow/common.c +++ b/xen/arch/x86/mm/shadow/common.c @@ -1982,7 +1982,8 @@ static void sh_update_paging_modes(struc } #endif /* OOS */ - v->arch.paging.mode->update_cr3(v, false); + if ( paging_mode_enabled(d) ) + v->arch.paging.mode->update_cr3(v, false); } /* @@ -2459,8 +2460,6 @@ static int shadow_one_bit_disable(struct d->arch.paging.free_pages, d->arch.paging.p2m_pages); for_each_vcpu(d, v) { - if ( v->arch.paging.mode ) - sh_detach_old_tables(v); if ( !(v->arch.flags & TF_kernel_mode) ) make_cr3(v, pagetable_get_mfn(v->arch.guest_table_user)); else