pgd_alloc and [cd]tors
Is there any real use in having a ctor/dtor for the pgd cache? Given that all pgd allocation happens via pgd_alloc/pgd_free, why not just fold the [cd]tor in?
I'm asking because Xen wants pgd[3] to be unshared in the PAE case, and it looks to me like the easiest way to handle that is by making pgd_alloc/free pv-ops and doing the appropriate thing in the Xen code. Would need to sort out the vmalloc mapping replication as well, but I think that's pretty straightforward.
J
Quoted 3 lines · * Jeremy Fitzhardinge ([email protected]) wrote:
Is there any real use in having a ctor/dtor for the pgd cache? Given that all pgd allocation happens via pgd_alloc/pgd_free, why not just fold the [cd]tor in?
I don't see their value either.
Quoted 5 lines
I'm asking because Xen wants pgd[3] to be unshared in the PAE case, and it looks to me like the easiest way to handle that is by making pgd_alloc/free pv-ops and doing the appropriate thing in the Xen code. Would need to sort out the vmalloc mapping replication as well, but I think that's pretty straightforward.
Here's an old patch from Zach that seems to address the issue at least in part. It from back when we were doing subarch, but I think you can see the attempt.
commit edb15f3f5e925d649b4dd28b0b430bea1a4c8fb4
Author: Zachary Amsden <[email protected]> Date: Wed Aug 24 14:43:14 2005 -0700
[PATCH] Add pagetable allocation notifiers
Hooks are provided for the mach-XXX subarchitecture at the time prior to a page being used as a page table at all levels, for PAE and non-PAE kernels. Note that in PAE mode, multiple PDP roots may exist on the same page with other data, so the root must be shadowed instead. This is not a performance issue, since PAE only uses 4 top level PDPEs.
The hooks are:
SetPagePTE(ppn) - indicates that a given physical page is going to be used as a page table.
SetPagePDE(ppn) - indicates that a given physical page is going to be used as a page directory.
ClearPageXXX(ppn) - indicates that the physical page is now done being used as a certain type of page.
These hooks can be used in two ways; for shadow mode, they serve as requests to pre-allocate and deallocate shadow page tables, and for direct page table mode, they serve as write protect/unprotect requests.
arch/i386/mm/init.c +3 -0
@@ -59,6 +59,7 @@ static pmd_t * __init one_md_table_init(#ifdef CONFIG_X86_PAE pmd_table = (pmd_t *)alloc_bootmem_low_pages (PAGE_SIZE ); +SetPagePDE (virt_to_page (pmd_table));set_pgd (pgd,__pgd (__pa (pmd_table) | _PAGE_PRESENT)); pud =pud_offset (pgd,0 );if (pmd_table !=pmd_offset (pud,0 ))
@@ -79,6 +80,7 @@ static pte_t * __init one_page_table_ini {if (pmd_none (*pmd)) {pte_t *page_table = (pte_t *)alloc_bootmem_low_pages (PAGE_SIZE ); +SetPagePTE (virt_to_page (page_table));set_pmd (pmd,__pmd (__pa (page_table) | _PAGE_TABLE));if (page_table !=pte_offset_kernel (pmd,0 ))BUG ();
@@ -312,6 +314,7 @@ static void __init pagetable_init (void)#ifdef CONFIG_X86_PAE int i;/* Init entries of the first-level page table to the zero page */ +SetPagePDE (virt_to_page (empty_zero_page));for (i =0 ; i <PTRS_PER_PGD ; i++)set_pgd (pgd_base + i,__pgd (__pa (empty_zero_page) | _PAGE_PRESENT));#endif
arch/i386/mm/pageattr.c +2 -0
@@ -52,6 +52,7 @@ static struct page *split_large_page(uns address =__pa (address); addr = address &LARGE_PAGE_MASK ; pbase = (pte_t *)page_address (base); +SetPagePTE (virt_to_page (pbase));for (i =0 ; i <PTRS_PER_PTE ; i++, addr +=PAGE_SIZE ) {set_pte (&pbase[i],pfn_pte (addr >>PAGE_SHIFT , addr == address ? prot :PAGE_KERNEL ));
@@ -146,6 +147,7 @@ __change_page_attr(struct page *page, pgBUG_ON (!page_count (kpte_page));if (cpu_has_pse && (page_count (kpte_page) ==1 )) { +ClearPagePTE (virt_to_page (kpte));list_add (&kpte_page->lru, &df_list);revert_page (kpte_page, address); }
arch/i386/mm/pgtable.c +9 -2
@@ -209,6 +209,7 @@ void pgd_ctor(void *pgd, kmem_cache_t *cif (PTRS_PER_PMD ==1 ) {memset (pgd,0 ,USER_PTRS_PER_PGD *sizeof (pgd_t)); +SetPagePDE (virt_to_page (pgd));spin_lock_irqsave (&pgd_lock, flags); }
@@ -227,6 +228,7 @@ void pgd_dtor(void *pgd, kmem_cache_t *c {unsigned long flags;/* can be called from interrupt context */ +ClearPagePDE (virt_to_page (pgd));spin_lock_irqsave (&pgd_lock, flags);pgd_list_del (pgd);spin_unlock_irqrestore (&pgd_lock, flags);
@@ -244,13 +246,16 @@ pgd_t *pgd_alloc(struct mm_struct *mm)pmd_t *pmd =kmem_cache_alloc (pmd_cache,GFP_KERNEL );if (!pmd) goto out_oom; +SetPagePDE (virt_to_page (pmd));set_pgd (&pgd[i],__pgd (1 +__pa (pmd))); }return pgd; out_oom: -for (i--; i >=0 ; i--) +for (i--; i >=0 ; i--) { +ClearPagePDE (pfn_to_page (pgd_val (pgd[i]) >>PAGE_SHIFT ));kmem_cache_free (pmd_cache, (void *)__va (pgd_val (pgd[i])-1 )); + }kmem_cache_free (pgd_cache, pgd);return NULL ; }
@@ -261,8 +266,10 @@ void pgd_free(pgd_t *pgd)/* in the PAE case user pgd entries are overwritten before usage */ if (PTRS_PER_PMD >1 ) -for (i =0 ; i <USER_PTRS_PER_PGD ; ++i) +for (i =0 ; i <USER_PTRS_PER_PGD ; ++i) { +ClearPagePDE (pfn_to_page (pgd_val (pgd[i]) >>PAGE_SHIFT ));kmem_cache_free (pmd_cache, (void *)__va (pgd_val (pgd[i])-1 )); + }/* in the non-PAE case, free_pgtables() clears user pgd entries */ kmem_cache_free (pgd_cache, pgd); }
include/asm-i386/mach-default/mach_pgalloc.h +34 -1
@@ -1,7 +1,40 @@ +/* +* Copyright (C) 2005, VMware, Inc. +* +* All rights reserved. +* +* This program is free software; you can redistribute it and/or modify +* it under the terms of the GNU General Public License as published by +* the Free Software Foundation; either version 2 of the License, or +* (at your option) any later version. +* +* This program is distributed in the hope that it will be useful, but +* WITHOUT ANY WARRANTY; without even the implied warranty of +* MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE, GOOD TITLE or +* NON INFRINGEMENT. See the GNU General Public License for more +* details. +* +* You should have received a copy of the GNU General Public License +* along with this program; if not, write to the Free Software +* Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA. +* +* Send feedback to [email protected] +* +*/ #ifndef __ASM_MACH_PGALLOC_H#define __ASM_MACH_PGALLOC_H +/* +* These hooks allow the hypervisor to be notified about page type +* transition events. +*/ + +#define SetPagePTE (_page) +#define ClearPagePTE (_page) +#define SetPagePDE (_page) +#define ClearPagePDE (_page) +#define SetPagesLDT (_va, _pages)#define ClearPagesLDT (_va, _pages) -#endif +#endif /* _ASM_MACH_PGALLOC_H */
include/asm-i386/pgalloc.h +16 -4
@@ -5,14 +5,22 @@#include <asm/fixmap.h> #include <linux/threads.h> #include <linux/mm.h> /* for struct page */ +#include <mach_pgalloc.h> -#define pmd_populate_kernel (mm, pmd, pte) \ - set_pmd(pmd, __pmd(_PAGE_TABLE + __pa(pte))) +#define pmd_populate_kernel (mm, pmd, pte) \ +do { \ + SetPagePTE(virt_to_page(pte)); \ + set_pmd(pmd, __pmd(_PAGE_TABLE + __pa(pte))); \ +} while (0)#define pmd_populate (mm, pmd, pte) \ +do { \ + SetPagePTE(pte); \ set_pmd(pmd, __pmd(_PAGE_TABLE + \ ((unsigned long long)page_to_pfn(pte) << \ - (unsigned long long) PAGE_SHIFT))) + (unsigned long long) PAGE_SHIFT))); \ +} while (0) +/* * Allocate and free page tables. */
@@ -33,7 +41,11 @@ static inline void pte_free(struct page } -#define __pte_free_tlb (tlb,pte) tlb_remove_page((tlb),(pte)) +#define __pte_free_tlb (tlb,pte) \ +do { \ + tlb_remove_page((tlb),(pte)); \ + ClearPagePTE(pte); \ +} while (0)#ifdef CONFIG_X86_PAE /*
Quoted 10 lines · Jeremy Fitzhardinge
Is there any real use in having a ctor/dtor for the pgd cache? Given that all pgd allocation happens via pgd_alloc/pgd_free, why not just fold the [cd]tor in?
I'm asking because Xen wants pgd[3] to be unshared in the PAE case, and it looks to me like the easiest way to handle that is by making pgd_alloc/free pv-ops and doing the appropriate thing in the Xen code. Would need to sort out the vmalloc mapping replication as well, but I think that's pretty straightforward.
Actually, vmalloc mapping replication is not quite straightforward - you have to replicate it under locks and track all pgds using pgd_list if you support large pages.
Zach
Quoted 3 lines · Zachary Amsden
Actually, vmalloc mapping replication is not quite straightforward - you have to replicate it under locks and track all pgds using
pgd_listif you support large pages.
But that machinery already exists. It would continue to work under PAE wouldn't it?
J
Quoted 10 lines · Jeremy Fitzhardinge
Quoted 4 lines · Zachary Amsden
Actually, vmalloc mapping replication is not quite straightforward - you have to replicate it under locks and track all pgds using
pgd_listif you support large pages.But that machinery already exists. It would continue to work under PAE wouldn't it?
No, it doesn't exist under PAE, since PAE uses a common pgd[3] therefore the init_mm pgd never needs changing for large page / split page PD transitions. The code is there for non-PAE, but compiled out of the PAE version.
Zach
Quoted 4 lines · Zachary Amsden
No, it doesn't exist under PAE, since PAE uses a common pgd[3] therefore the
init_mmpgd never needs changing for large page / split page PD transitions. The code is there for non-PAE, but compiled out of the PAE version.
Yep, but it would work if it were restored, right. I'm adding SHARED_KERNEL_PMD to replace all the tests on PTRS_PER_PMD, and a shared_kernel_pmd flag to pvops.
J
Quoted 4 lines · Zachary Amsden
No, it doesn't exist under PAE, since PAE uses a common pgd[3] therefore the
init_mmpgd never needs changing for large page / split page PD transitions. The code is there for non-PAE, but compiled out of the PAE version.
That is fixable ;)
cheers,
Gerd
Signature · 2 lines
Gerd Hoffmann <[email protected]> http://www.suse.de/~kraxel/julika-dora.jpeg