Messages in this thread | | | Date | Thu, 05 Jun 2008 09:51:55 +0100 | From | "Jan Beulich" <> | Subject | operation ordering during pgd_alloc/pgd_free |
| |
At present, pgd_ctor() adds a new pgd to pgd_list solely based on !SHARED_KERNEL_PMD. For PAE && !SHARED_KERNEL_PMD (i.e. Xen) this doesn't seem correct, as the pgd is still empty, which will confuse vmalloc_sync_all(). So in this case, list insertion should only happen at the end of pgd_prepopulate_pmd().
Likewise, pgd_free() calls pgd_mop_up_pmds() *before* pgd_dtor(), with the former zeroing pgd entries as it goes and only the latter removing the pgd from the list. Just as above this can confuse vmalloc_sync_all(), so here I would think that the two calls should just be swapped. However, if they get swapped, careful inspection of the interaction with save/restore will be needed - XenSource's Linux tree has a comment specifically to that effect:
/* * After this the pgd should not be pinned for the duration of this * function's execution. We should never sleep and thus never race: * 1. User pmds will not become write-protected under our feet due * to a concurrent mm_pin_all(). * 2. The machine addresses in PGD entries will not become invalid * due to a concurrent save/restore. */
Since that tree doesn't support preemption, this is perhaps fine, but likely going to cause problems in the (preemptable) pv-ops code.
The issue with vmalloc_sync_all() would even go unnoticed, since the patch to unify the pgd_list mechanism with x86-64 removed the BUG_ON() that was meant to trigger on issues like this.
But maybe I'm missing something?
Jan
| |