[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index] Re: [PATCH v4 14/21] x86: introduce helper for recording degree of contiguity in page tables
On Mon, Apr 25, 2022 at 10:41:23AM +0200, Jan Beulich wrote: > This is a re-usable helper (kind of a template) which gets introduced > without users so that the individual subsequent patches introducing such > users can get committed independently of one another. > > See the comment at the top of the new file. To demonstrate the effect, > if a page table had just 16 entries, this would be the set of markers > for a page table with fully contiguous mappings: > > index 0 1 2 3 4 5 6 7 8 9 A B C D E F > marker 4 0 1 0 2 0 1 0 3 0 1 0 2 0 1 0 > > "Contiguous" here means not only present entries with successively > increasing MFNs, each one suitably aligned for its slot, but also a > respective number of all non-present entries. > > Signed-off-by: Jan Beulich <jbeulich@xxxxxxxx> > --- > v3: Rename function and header. Introduce IS_CONTIG(). > v2: New. > > --- /dev/null > +++ b/xen/arch/x86/include/asm/pt-contig-markers.h > @@ -0,0 +1,105 @@ > +#ifndef __ASM_X86_PT_CONTIG_MARKERS_H > +#define __ASM_X86_PT_CONTIG_MARKERS_H > + > +/* > + * Short of having function templates in C, the function defined below is > + * intended to be used by multiple parties interested in recording the > + * degree of contiguity in mappings by a single page table. > + * > + * Scheme: Every entry records the order of contiguous successive entries, > + * up to the maximum order covered by that entry (which is the number of > + * clear low bits in its index, with entry 0 being the exception using > + * the base-2 logarithm of the number of entries in a single page table). > + * While a few entries need touching upon update, knowing whether the > + * table is fully contiguous (and can hence be replaced by a higher level > + * leaf entry) is then possible by simply looking at entry 0's marker. > + * > + * Prereqs: > + * - CONTIG_MASK needs to be #define-d, to a value having at least 4 > + * contiguous bits (ignored by hardware), before including this file, > + * - page tables to be passed here need to be initialized with correct > + * markers. Not sure it's very relevant, but might we worth adding that: - Null entries must have the PTE zeroed except for the CONTIG_MASK region in order to be considered as inactive. > + */ > + > +#include <xen/bitops.h> > +#include <xen/lib.h> > +#include <xen/page-size.h> > + > +/* This is the same for all anticipated users, so doesn't need passing in. */ > +#define CONTIG_LEVEL_SHIFT 9 > +#define CONTIG_NR (1 << CONTIG_LEVEL_SHIFT) > + > +#define GET_MARKER(e) MASK_EXTR(e, CONTIG_MASK) > +#define SET_MARKER(e, m) \ > + ((void)((e) = ((e) & ~CONTIG_MASK) | MASK_INSR(m, CONTIG_MASK))) > + > +#define IS_CONTIG(kind, pt, i, idx, shift, b) \ > + ((kind) == PTE_kind_leaf \ > + ? (((pt)[i] ^ (pt)[idx]) & ~CONTIG_MASK) == (1ULL << ((b) + (shift))) \ > + : !((pt)[i] & ~CONTIG_MASK)) > + > +enum PTE_kind { > + PTE_kind_null, > + PTE_kind_leaf, > + PTE_kind_table, > +}; > + > +static bool pt_update_contig_markers(uint64_t *pt, unsigned int idx, > + unsigned int level, enum PTE_kind kind) > +{ > + unsigned int b, i = idx; > + unsigned int shift = (level - 1) * CONTIG_LEVEL_SHIFT + PAGE_SHIFT; > + > + ASSERT(idx < CONTIG_NR); > + ASSERT(!(pt[idx] & CONTIG_MASK)); > + > + /* Step 1: Reduce markers in lower numbered entries. */ > + while ( i ) > + { > + b = find_first_set_bit(i); > + i &= ~(1U << b); > + if ( GET_MARKER(pt[i]) > b ) > + SET_MARKER(pt[i], b); Can't you exit early when you find an entry that already has the to-be-set contiguous marker <= b, as lower numbered entries will then also be <= b'? Ie: if ( GET_MARKER(pt[i]) <= b ) break; else SET_MARKER(pt[i], b); Thanks, Roger.
|
Lists.xenproject.org is hosted with RackSpace, monitoring our |