just realized that it's broken
PATCH 5/5 just realized that it's broken
This patch was moving the cpumask out of struct page_info into the page itself, but I just realized that this it not valid, since the purpose of the mask is to avoid a later TLB flush, hence the base assumption is that the page might still be in the TLB of some CPU, thus making it possible for a mis- behaved guest to still write to that page.
Sorry for the mis-numbering,
Jan
Fortunately I think it can be got rid of and probably at little or no performance cost. I will take a look maybe next week.
-- Keir
Quoted 15 lines · Jan Beulich
This patch was moving the cpumask out of struct
page_infointo the page itself, but I just realized that this it not valid, since the purpose of the mask
is to avoid a later TLB flush, hence the base assumption is that the page might still be in the TLB of some CPU, thus making it possible for a mis- behaved guest to still write to that page.Sorry for the mis-numbering,
Jan_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel
Quoted 9 lines
Quoted 1 line
Quoted 1 line
"Jan Beulich" <[email protected]> 30.01.09 11:48 >>>
This patch was moving the cpumask out of struct
page_infointo the page itself, but I just realized that this it not valid, since the purpose of the mask is to avoid a later TLB flush, hence the base assumption is that the page might still be in the TLB of some CPU, thus making it possible for a mis- behaved guest to still write to that page.Sorry for the mis-numbering,
Jan
Actually, after some more consideration I think the patch is actually correct (minus an issue that already has been in place for a while anyway): All access a domain may continue have to a page it's freeing (through stale TLB entries) is read/execute. This is because when the type count of a page (i.e. in particular a writeable one) a (domain-)global TLB flush is performed anyway. Allowing the freeing domain (as well as the one subsequently allocating) to read the cpumask does not present a problem in my opinion.
The issue mentioned above that I think current code has even without that patch (though I may easily have missed some aspect here) is that there is no enforcement of an immediate TLB flush when a writeable page's type count drops to zero - instead the flush is deferred until the end of the current operation or batch. With the removal of the use of the per-domain lock in these code paths it is no longer valid to defer the flush this much - it must be carried out before the page in question gets unlocked.
And I would think that invalidate_shadow_ldt() has a similar issue, plus it seems questionable that it does a local flush when acting on the current vCPU, but a global one for all others.
Jan
*******************************************************
Move the (potentially large) cpumask field of free pages into the page itself, thus making sizeof(struct page_info) independent of the configured number of CPUs, which avoids penalizing systems with few CPUs just because the hypervisor is able to support many.
xen/common/page_alloc.c +3 -3
@@ -376,7 +376,7 @@ static struct page_info *alloc_heap_pageBUG_ON (pg[i].count_info !=0 );/* Add in any extra CPUs that need flushing because of this page. */ -cpus_andnot (extra_cpus_mask, pg[i].u.free.cpumask, mask); +cpus_andnot (extra_cpus_mask,PFN_CPUMASK (&pg[i]), mask);tlbflush_filter (extra_cpus_mask, pg[i].tlbflush_timestamp);cpus_or (mask, mask, extra_cpus_mask);
@@ -425,11 +425,11 @@ static void free_heap_pages(if ( (d =page_get_owner (&pg[i])) !=NULL ) { pg[i].tlbflush_timestamp =tlbflush_current_time (); - pg[i].u.free.cpumask = d->domain_dirty_cpumask; +PFN_CPUMASK (&pg[i]) = d->domain_dirty_cpumask; }else { -cpus_clear (pg[i].u.free.cpumask); +cpus_clear (PFN_CPUMASK (&pg[i])); } }
xen/include/asm-ia64/mm.h +4 -1
@@ -35,8 +35,11 @@ typedef unsigned long page_flags_t;* Every architecture must ensure the following: * 1. 'struct page_info' contains a 'struct list_head list'. * 2. Provide a PFN_ORDER() macro for accessing the order of a free page. +* 3. Provide a PFN_CPUMASK() macro for accessing the mask of possibly-tainted +* TLBs. */ -#define PFN_ORDER (_pfn) ((_pfn)->u.free.order) +#define PFN_ORDER (_pfn) ((_pfn)->u.free.order) +#define PFN_CPUMASK (_pfn) ((_pfn)->u.free.cpumask)#define PRtype_info "016lx"
xen/include/asm-x86/mm.h +12 -0
@@ -14,8 +14,18 @@* Every architecture must ensure the following: * 1. 'struct page_info' contains a 'struct page_list_entry list'. * 2. Provide a PFN_ORDER() macro for accessing the order of a free page. +* 3. Provide a PFN_CPUMASK() macro for accessing the mask of possibly-tainted +* TLBs. */ #define PFN_ORDER (_pfn) ((_pfn)->v.free.order) +#ifdef __i386__ +# define PFN_CPUMASK (_pfn) ((_pfn)->u.free.cpumask) +#else +# define PFN_CPUMASK (_pfn) (*(cpumask_t *)page_to_virt(_pfn)) +# if NR_CPUS >PAGE_SIZE *8 +# error Cannot handle this many CPUs. +# endif +#endif /* * This definition is solely for the use in struct page_info (and
@@ -72,8 +82,10 @@ struct page_info/* Page is on a free list: ((count_info & PGC_count_mask) == 0). */ struct { +#ifdef __i386__/* Mask of possibly-tainted TLBs. */ cpumask_t cpumask; +#endif } free; } u;
Quoted 8 lines · Jan Beulich
The issue mentioned above that I think current code has even without that patch (though I may easily have missed some aspect here) is that there is no enforcement of an immediate TLB flush when a writeable page's type count drops to zero - instead the flush is deferred until the end of the current operation or batch. With the removal of the use of the per-domain lock in these code paths it is no longer valid to defer the flush this much - it must be carried out before the page in question gets unlocked.
Oh, good point. We don't need the synchronous flush in free_page_type() any more as we do not rely on correctness of the linear map any more (we use it, but then verify sufficient correctness via get_page()/lock_page(), so now a guest can only shoot itself in the foot).
I'll revert back to the lazy scheme we used to use. Which of course means cpumask-in-the-page will not be safe. I will kill it off by other means.
For invalidate_shadow_ldt(), the LDT mappings are per-VCPU so only the local CPU needs flushing. The reason for flush_tlb_mask() is I think a conservative attempt at flushing the correct TLB if we are not the VCPU in question. I would think it is better to use v->vcpu_dirty_cpumask, which maybe didn't exist when i_s_l() was last looked at. I shall take a look.
-- Keir