mini-os: mm: use a generic page table walker
v2 cover mini-os: mm: use a generic page table walker
Instead of open coding a page table walk multiple times, use a generic page table walker instead.
This new page table walker will be used later for kexec support, too.
Changes in V2:
- addressed comments
Juergen Gross (3):
mini-os: mm: introduce generic page table walk function
mini-os: mm: switch need_pgt() to use walk_pt()
mini-os: mm: convert set_readonly() to use walk_pt()
arch/x86/mm.c | 353 +++++++++++++++++++++++++++++--------------------- 1 file changed, 204 insertions(+), 149 deletions(-)
Signature · 1 line
2.43.0
PATCH v2 1/3 mini-os: mm: introduce generic page table walk function
In x86 mm code there are multiple instances of page table walks for different purposes.
Introduce a generic page table walker being able to cover the current use cases. It will be used for other cases in future, too.
The page table walker needs some per-level data, so add a table for that data. Merge it with the already existing pt_prot[] array.
Rewrite get_pgt() to use the new walker.
---
V2:
- add idx_from_va_lvl() helper (Samuel Thibault)
--- arch/x86/mm.c | 157 +++++++++++++++++++++++++++++++++++++------------- 1 file changed, 118 insertions(+), 39 deletions(-)
arch/x86/mm.c +118 -39
@@ -125,20 +125,30 @@ void arch_mm_preinit(void *p) }#endif +static const struct { +unsigned int shift; +unsigned int entries; +pgentry_t prot; +} ptdata[PAGETABLE_LEVELS +1 ] = { + {0 ,0 ,0 }, + {L1_PAGETABLE_SHIFT ,L1_PAGETABLE_ENTRIES ,L1_PROT }, + {L2_PAGETABLE_SHIFT ,L2_PAGETABLE_ENTRIES ,L2_PROT }, + {L3_PAGETABLE_SHIFT ,L3_PAGETABLE_ENTRIES ,L3_PROT }, +#if defined(__x86_64__) + {L4_PAGETABLE_SHIFT ,L4_PAGETABLE_ENTRIES ,L4_PROT }, +#endif +}; + +static inline unsigned int idx_from_va_lvl (unsigned long va,unsigned int lvl) +{ +return (va >> ptdata[lvl].shift) & (ptdata[lvl].entries -1 ); +} +/* * Make pt_pfn a new 'level' page table frame and hook it into the page * table at offset in previous level MFN (pref_l_mfn). pt_pfn is a guest * PFN. */ -static pgentry_t pt_prot[PAGETABLE_LEVELS ] = { -L1_PROT , -L2_PROT , -L3_PROT , -#if defined(__x86_64__) -L4_PROT , -#endif -}; -static void new_pt_frame (unsigned long *pt_pfn,unsigned long prev_l_mfn,unsigned long offset,unsigned long level) {
@@ -170,7 +180,7 @@ static void new_pt_frame(unsigned long *pt_pfn, unsigned long prev_l_mfn, mmu_updates[0 ].ptr = (tab[l2_table_offset (pt_page)] &PAGE_MASK ) +sizeof (pgentry_t) *l1_table_offset (pt_page); mmu_updates[0 ].val = (pgentry_t )pfn_to_mfn (*pt_pfn) <<PAGE_SHIFT | - (pt_prot[level -1 ] & ~_PAGE_RW); + (ptdata[level].prot & ~_PAGE_RW);if ( (rc =HYPERVISOR_mmu_update (mmu_updates,1 ,NULL ,DOMID_SELF )) <0 ) {
@@ -183,7 +193,7 @@ static void new_pt_frame(unsigned long *pt_pfn, unsigned long prev_l_mfn, mmu_updates[0 ].ptr = ((pgentry_t )prev_l_mfn <<PAGE_SHIFT ) +sizeof (pgentry_t) * offset; mmu_updates[0 ].val = (pgentry_t )pfn_to_mfn (*pt_pfn) <<PAGE_SHIFT | - pt_prot[level]; + ptdata[level +1 ].prot;if ( (rc =HYPERVISOR_mmu_update (mmu_updates,1 ,NULL ,DOMID_SELF )) <0 ) {
@@ -192,7 +202,7 @@ static void new_pt_frame(unsigned long *pt_pfn, unsigned long prev_l_mfn, }#else tab =mfn_to_virt (prev_l_mfn); - tab[offset] = (*pt_pfn <<PAGE_SHIFT ) | pt_prot[level]; + tab[offset] = (*pt_pfn <<PAGE_SHIFT ) | ptdata[level +1 ].prot;#endif *pt_pfn +=1 ;
@@ -202,6 +212,82 @@ static void new_pt_frame(unsigned long *pt_pfn, unsigned long prev_l_mfn,static mmu_update_t mmu_updates[L1_PAGETABLE_ENTRIES +1 ];#endif +/* +* Walk recursively through all PTEs calling a specified function. The function +* is allowed to change the PTE, the walker will follow the new value. +* The walk will cover the virtual address range [from_va .. to_va]. +* The supplied function will be called with the following parameters: +* va: base virtual address of the area covered by the current PTE +* lvl: page table level of the PTE (1 = lowest level, PAGETABLE_LEVELS = +* PTE in page table addressed by %cr3) +* is_leaf: true if PTE doesn't address another page table (it is either at +* level 1, or invalid, or has its PSE bit set) +* pte: address of the PTE +* par: parameter, passed to walk_pt() by caller +* Return value of func() being non-zero will terminate walk_pt(), walk_pt() +* will return that value in this case, zero else. +*/ +static int walk_pt (unsigned long from_va,unsigned long to_va, +int (func )(unsigned long va,unsigned int lvl, +bool is_leaf,pgentry_t *pte,void *par), +void *par) +{ +unsigned int lvl =PAGETABLE_LEVELS ; +unsigned int ptindex[PAGETABLE_LEVELS +1 ]; +unsigned long va =round_pgdown (from_va); +unsigned long va_lvl; +pgentry_t *tab[PAGETABLE_LEVELS +1 ]; +pgentry_t *pte; +bool is_leaf; +int ret; + +/* Start at top level page table. */ + tab[lvl] = pt_base; + ptindex[lvl] =idx_from_va_lvl (va, lvl); + +while ( va < (to_va | (PAGE_SIZE -1 )) ) + { + pte = tab[lvl] + ptindex[lvl]; + is_leaf = (lvl ==L1_FRAME ) || (*pte & _PAGE_PSE) || + !(*pte & _PAGE_PRESENT); + va_lvl = va & ~((1UL << ptdata[lvl].shift) -1 ); + ret =func (va_lvl, lvl, is_leaf, pte, par); +if ( ret ) +return ret; + +/* PTE might have been modified by func(), reevaluate leaf state. */ + is_leaf = (lvl ==L1_FRAME ) || (*pte & _PAGE_PSE) || + !(*pte & _PAGE_PRESENT); + +if ( is_leaf ) + { +/* Reached a leaf PTE. Advance to next page. */ + va +=1UL << ptdata[lvl].shift; + ptindex[lvl]++; + +/* Check for the need to traverse up again. */ +while ( ptindex[lvl] == ptdata[lvl].entries ) + { +/* End of virtual address space? */ +if ( lvl ==PAGETABLE_LEVELS ) +return 0 ; +/* Reached end of current page table, one level up. */ + lvl++; + ptindex[lvl]++; + } + } +else + { +/* Not a leaf, walk one level down. */ + lvl--; + tab[lvl] =mfn_to_virt (pte_to_mfn (*pte)); + ptindex[lvl] =idx_from_va_lvl (va, lvl); + } + } + +return 0 ; +} +/* * Build the initial pagetable. */
@@ -407,36 +493,29 @@ static void set_readonly(void *text, void *etext)/* * get the PTE for virtual address va if it exists. Otherwise NULL. */ -static pgentry_t *get_pgt (unsigned long va) +static int get_pgt_func (unsigned long va,unsigned int lvl,bool is_leaf, +pgentry_t *pte,void *par) { -unsigned long mfn; -pgentry_t *tab; -unsigned offset; +pgentry_t **result; - tab = pt_base; - mfn =virt_to_mfn (pt_base); +if ( !(*pte & _PAGE_PRESENT) && lvl >L1_FRAME ) +return -1 ; -#if defined(__x86_64__) - offset =l4_table_offset (va); -if ( !(tab[offset] & _PAGE_PRESENT) ) -return NULL ; - mfn =pte_to_mfn (tab[offset]); - tab =mfn_to_virt (mfn); -#endif - offset =l3_table_offset (va); -if ( !(tab[offset] & _PAGE_PRESENT) ) -return NULL ; - mfn =pte_to_mfn (tab[offset]); - tab =mfn_to_virt (mfn); - offset =l2_table_offset (va); -if ( !(tab[offset] & _PAGE_PRESENT) ) -return NULL ; -if ( tab[offset] & _PAGE_PSE ) -return &tab[offset]; - mfn =pte_to_mfn (tab[offset]); - tab =mfn_to_virt (mfn); - offset =l1_table_offset (va); -return &tab[offset]; +if ( lvl >L1_FRAME && !(*pte & _PAGE_PSE) ) +return 0 ; + + result = par; + *result = pte; + +return 0 ; +} + +static pgentry_t *get_pgt (unsigned long va) +{ +pgentry_t *tab =NULL ; + +walk_pt (va, va, get_pgt_func, &tab); +return tab; }
Signature · 1 line
2.43.0
Re: [PATCH v2 1/3] mini-os: mm: introduce generic page table walk function
Juergen Gross, le mar. 13 août 2024 15:41:56 +0200, a ecrit:
Quoted 6 lines
In x86 mm code there are multiple instances of page table walks for different purposes.
Introduce a generic page table walker being able to cover the current use cases. It will be used for other cases in future, too.
Quoted 6 lines
The page table walker needs some per-level data, so add a table for that data. Merge it with the already existing pt_prot[] array.
Rewrite
get_pgt()to use the new walker.Signed-off-by Juergen Gross <[email protected]>
Quoted 221 lines
---
V2:
- addidx_from_va_lvl()helper (Samuel Thibault)--- arch/x86/mm.c | 157 +++++++++++++++++++++++++++++++++++++------------- 1 file changed, 118 insertions(+), 39 deletions(-)arch/x86/mm.c +118
-39@@ -125,20 +125,30 @@ void arch_mm_preinit(void *p) }#endif +static const struct { +unsigned int shift; +unsigned int entries; +pgentry_t prot; +} ptdata[PAGETABLE_LEVELS +1 ] = { + {0 ,0 ,0 }, + {L1_PAGETABLE_SHIFT ,L1_PAGETABLE_ENTRIES ,L1_PROT }, + {L2_PAGETABLE_SHIFT ,L2_PAGETABLE_ENTRIES ,L2_PROT }, + {L3_PAGETABLE_SHIFT ,L3_PAGETABLE_ENTRIES ,L3_PROT }, +#if defined(__x86_64__) + {L4_PAGETABLE_SHIFT ,L4_PAGETABLE_ENTRIES ,L4_PROT }, +#endif +}; + +static inline unsigned int idx_from_va_lvl (unsigned long va,unsigned int lvl) +{ +return (va >> ptdata[lvl].shift) & (ptdata[lvl].entries -1 ); +} +/* * Make pt_pfn a new 'level' page table frame and hook it into the page * table at offset in previous level MFN (pref_l_mfn). pt_pfn is a guest * PFN. */ -static pgentry_t pt_prot[PAGETABLE_LEVELS ] = { -L1_PROT , -L2_PROT , -L3_PROT , -#if defined(__x86_64__) -L4_PROT , -#endif -}; -static void new_pt_frame (unsigned long *pt_pfn,unsigned long prev_l_mfn,unsigned long offset,unsigned long level) {@@ -170,7 +180,7 @@ static void new_pt_frame(unsigned long *pt_pfn, unsigned long prev_l_mfn, mmu_updates[0 ].ptr = (tab[l2_table_offset (pt_page)] &PAGE_MASK ) +sizeof (pgentry_t) *l1_table_offset (pt_page); mmu_updates[0 ].val = (pgentry_t )pfn_to_mfn (*pt_pfn) <<PAGE_SHIFT | - (pt_prot[level -1 ] & ~_PAGE_RW); + (ptdata[level].prot & ~_PAGE_RW);if ( (rc =HYPERVISOR_mmu_update (mmu_updates,1 ,NULL ,DOMID_SELF )) <0 ) {@@ -183,7 +193,7 @@ static void new_pt_frame(unsigned long *pt_pfn, unsigned long prev_l_mfn, mmu_updates[0 ].ptr = ((pgentry_t )prev_l_mfn <<PAGE_SHIFT ) +sizeof (pgentry_t) * offset; mmu_updates[0 ].val = (pgentry_t )pfn_to_mfn (*pt_pfn) <<PAGE_SHIFT | - pt_prot[level]; + ptdata[level +1 ].prot;if ( (rc =HYPERVISOR_mmu_update (mmu_updates,1 ,NULL ,DOMID_SELF )) <0 ) {@@ -192,7 +202,7 @@ static void new_pt_frame(unsigned long *pt_pfn, unsigned long prev_l_mfn, }#else tab =mfn_to_virt (prev_l_mfn); - tab[offset] = (*pt_pfn <<PAGE_SHIFT ) | pt_prot[level]; + tab[offset] = (*pt_pfn <<PAGE_SHIFT ) | ptdata[level +1 ].prot;#endif *pt_pfn +=1 ;@@ -202,6 +212,82 @@ static void new_pt_frame(unsigned long *pt_pfn, unsigned long prev_l_mfn,static mmu_update_t mmu_updates[L1_PAGETABLE_ENTRIES +1 ];#endif +/* +* Walk recursively through all PTEs calling a specified function. The function +* is allowed to change the PTE, the walker will follow the new value. +* The walk will cover the virtual address range [from_va .. to_va]. +* The supplied function will be called with the following parameters: +* va: base virtual address of the area covered by the current PTE +* lvl: page table level of the PTE (1 = lowest level, PAGETABLE_LEVELS = +* PTE in page table addressed by %cr3) +* is_leaf: true if PTE doesn't address another page table (it is either at +* level 1, or invalid, or has its PSE bit set) +* pte: address of the PTE +* par: parameter, passed to walk_pt() by caller +* Return value of func() being non-zero will terminate walk_pt(), walk_pt() +* will return that value in this case, zero else. +*/ +static int walk_pt (unsigned long from_va,unsigned long to_va, +int (func )(unsigned long va,unsigned int lvl, +bool is_leaf,pgentry_t *pte,void *par), +void *par) +{ +unsigned int lvl =PAGETABLE_LEVELS ; +unsigned int ptindex[PAGETABLE_LEVELS +1 ]; +unsigned long va =round_pgdown (from_va); +unsigned long va_lvl; +pgentry_t *tab[PAGETABLE_LEVELS +1 ]; +pgentry_t *pte; +bool is_leaf; +int ret; + +/* Start at top level page table. */ + tab[lvl] = pt_base; + ptindex[lvl] =idx_from_va_lvl (va, lvl); + +while ( va < (to_va | (PAGE_SIZE -1 )) ) + { + pte = tab[lvl] + ptindex[lvl]; + is_leaf = (lvl ==L1_FRAME ) || (*pte & _PAGE_PSE) || + !(*pte & _PAGE_PRESENT); + va_lvl = va & ~((1UL << ptdata[lvl].shift) -1 ); + ret =func (va_lvl, lvl, is_leaf, pte, par); +if ( ret ) +return ret; + +/* PTE might have been modified by func(), reevaluate leaf state. */ + is_leaf = (lvl ==L1_FRAME ) || (*pte & _PAGE_PSE) || + !(*pte & _PAGE_PRESENT); + +if ( is_leaf ) + { +/* Reached a leaf PTE. Advance to next page. */ + va +=1UL << ptdata[lvl].shift; + ptindex[lvl]++; + +/* Check for the need to traverse up again. */ +while ( ptindex[lvl] == ptdata[lvl].entries ) + { +/* End of virtual address space? */ +if ( lvl ==PAGETABLE_LEVELS ) +return 0 ; +/* Reached end of current page table, one level up. */ + lvl++; + ptindex[lvl]++; + } + } +else + { +/* Not a leaf, walk one level down. */ + lvl--; + tab[lvl] =mfn_to_virt (pte_to_mfn (*pte)); + ptindex[lvl] =idx_from_va_lvl (va, lvl); + } + } + +return 0 ; +} +/* * Build the initial pagetable. */ @@ -407,36 +493,29 @@ static void set_readonly(void *text, void *etext)/* * get the PTE for virtual address va if it exists. Otherwise NULL. */ -static pgentry_t *get_pgt (unsigned long va) +static int get_pgt_func (unsigned long va,unsigned int lvl,bool is_leaf, +pgentry_t *pte,void *par) { -unsigned long mfn; -pgentry_t *tab; -unsigned offset; +pgentry_t **result; - tab = pt_base; - mfn =virt_to_mfn (pt_base); +if ( !(*pte & _PAGE_PRESENT) && lvl >L1_FRAME ) +return -1 ; -#if defined(__x86_64__) - offset =l4_table_offset (va); -if ( !(tab[offset] & _PAGE_PRESENT) ) -return NULL ; - mfn =pte_to_mfn (tab[offset]); - tab =mfn_to_virt (mfn); -#endif - offset =l3_table_offset (va); -if ( !(tab[offset] & _PAGE_PRESENT) ) -return NULL ; - mfn =pte_to_mfn (tab[offset]); - tab =mfn_to_virt (mfn); - offset =l2_table_offset (va); -if ( !(tab[offset] & _PAGE_PRESENT) ) -return NULL ; -if ( tab[offset] & _PAGE_PSE ) -return &tab[offset]; - mfn =pte_to_mfn (tab[offset]); - tab =mfn_to_virt (mfn); - offset =l1_table_offset (va); -return &tab[offset]; +if ( lvl >L1_FRAME && !(*pte & _PAGE_PSE) ) +return 0 ; + + result = par; + *result = pte; + +return 0 ; +} + +static pgentry_t *get_pgt (unsigned long va) +{ +pgentry_t *tab =NULL ; + +walk_pt (va, va, get_pgt_func, &tab); +return tab; }Signature · 1 line
2.43.0
Signature · 5 lines
Samuel <N> je déteste import <N> parce que lorsque tu fais du python et que tu oublies le #!/bin/env python et que tu mets le fichier exécutable <N> import est exécuté -+- #ens-mim - pourquoi mon script python change le curseur de la souris ?! -+-
Re: [PATCH v2 1/3] mini-os: mm: introduce generic page table walk function
Quoted 7 lines · Samuel Thibault
Juergen Gross, le mar. 13 août 2024 15:41:56 +0200, a ecrit:
Quoted 6 lines
In x86 mm code there are multiple instances of page table walks for different purposes.
Introduce a generic page table walker being able to cover the current use cases. It will be used for other cases in future, too.
Quoted 8 lines
Quoted 6 lines
The page table walker needs some per-level data, so add a table for that data. Merge it with the already existing pt_prot[] array.
Rewrite
get_pgt()to use the new walker.Signed-off-by Juergen Gross <[email protected]>Reviewed-by SAmuel Thibault <[email protected]>
Samuel - I've taken the liberty to convert the odd upper-case A.
Jürgen - looking at the mini-os short-log is quite odd, and increasingly so. In the mini-os repo surely all commits are about mini-os; the subject prefixes are kind of pointless. To nevertheless identify such patches on xen-devel@, may I suggest to use (taking this patch as an example) "[PATCH MINI-OS] mm: introduce generic page table walk function" instead (or, of course, whatever variation thereof you may like better)?
Jan
Re: [PATCH v2 1/3] mini-os: mm: introduce generic page table walk function
Quoted 17 lines · Jan Beulich
Quoted 7 lines · Samuel Thibault
Juergen Gross, le mar. 13 août 2024 15:41:56 +0200, a ecrit:
Quoted 6 lines
In x86 mm code there are multiple instances of page table walks for different purposes.
Introduce a generic page table walker being able to cover the current use cases. It will be used for other cases in future, too.
Quoted 8 lines
Quoted 6 lines
The page table walker needs some per-level data, so add a table for that data. Merge it with the already existing pt_prot[] array.
Rewrite
get_pgt()to use the new walker.Signed-off-by Juergen Gross <[email protected]>Reviewed-by SAmuel Thibault <[email protected]>
Quoted 8 lines
Samuel - I've taken the liberty to convert the odd upper-case A.
Jürgen - looking at the mini-os short-log is quite odd, and increasingly so. In the mini-os repo surely all commits are about mini-os; the subject prefixes are kind of pointless. To nevertheless identify such patches on xen-devel@, may I suggest to use (taking this patch as an example) "[PATCH MINI-OS] mm: introduce generic page table walk function" instead (or, of course, whatever variation thereof you may like better)?
Okay, fine with me.
Juergen
Re: [PATCH v2 1/3] mini-os: mm: introduce generic page table walk function
In reply to Jan BeulichJan Beulich, le mar. 20 août 2024 16:17:26 +0200, a ecrit:
Quoted 11 lines · Samuel Thibault
Quoted 10 lines
Juergen Gross, le mar. 13 août 2024 15:41:56 +0200, a ecrit:
Quoted 9 lines
In x86 mm code there are multiple instances of page table walks for different purposes.
Introduce a generic page table walker being able to cover the current use cases. It will be used for other cases in future, too.
The page table walker needs some per-level data, so add a table for that data. Merge it with the already existing pt_prot[] array.
Quoted 7 lines
Quoted 5 lines
Quoted 3 lines
Rewrite
get_pgt()to use the new walker.Signed-off-by Juergen Gross <[email protected]>Reviewed-by SAmuel Thibault <[email protected]>Samuel - I've taken the liberty to convert the odd upper-case A.
Oh, sure, thanks :)
Samuel
PATCH v2 2/3 mini-os: mm: switch need_pgt() to use walk_pt()
In reply to Juergen GrossInstead of open coding a page table walk, use walk_pt() in need_pgt().
---
V2:
- add comment and ASSERT() (Samuel Thibault)
--- arch/x86/mm.c | 72 +++++++++++++++++++++------------------------------ 1 file changed, 30 insertions(+), 42 deletions(-)
arch/x86/mm.c +30 -42
@@ -523,57 +523,45 @@ static pgentry_t *get_pgt(unsigned long va)* return a valid PTE for a given virtual address. If PTE does not exist, * allocate page-table pages. */ -pgentry_t *need_pgt (unsigned long va) +static int need_pgt_func (unsigned long va,unsigned int lvl,bool is_leaf, +pgentry_t *pte,void *par) { +pgentry_t **result = par;unsigned long pt_mfn; -pgentry_t *tab;unsigned long pt_pfn; -unsigned offset; +unsigned int idx; - tab = pt_base; - pt_mfn =virt_to_mfn (pt_base); +if ( !is_leaf ) +return 0 ; -#if defined(__x86_64__) - offset =l4_table_offset (va); -if ( !(tab[offset] & _PAGE_PRESENT) ) - { - pt_pfn =virt_to_pfn (alloc_page ()); -if ( !pt_pfn ) -return NULL ; -new_pt_frame (&pt_pfn, pt_mfn, offset,L3_FRAME ); - } -ASSERT (tab[offset] & _PAGE_PRESENT); - pt_mfn =pte_to_mfn (tab[offset]); - tab =mfn_to_virt (pt_mfn); -#endif - offset =l3_table_offset (va); -if ( !(tab[offset] & _PAGE_PRESENT) ) - { - pt_pfn =virt_to_pfn (alloc_page ()); -if ( !pt_pfn ) -return NULL ; -new_pt_frame (&pt_pfn, pt_mfn, offset,L2_FRAME ); - } -ASSERT (tab[offset] & _PAGE_PRESENT); - pt_mfn =pte_to_mfn (tab[offset]); - tab =mfn_to_virt (pt_mfn); - offset =l2_table_offset (va); -if ( !(tab[offset] & _PAGE_PRESENT) ) +if ( lvl ==L1_FRAME || (*pte & _PAGE_PRESENT) ) { - pt_pfn =virt_to_pfn (alloc_page ()); -if ( !pt_pfn ) -return NULL ; -new_pt_frame (&pt_pfn, pt_mfn, offset,L1_FRAME ); +/* +* The PTE is not addressing a page table (is_leaf is true). If we are +* either at the lowest level or we have a valid large page, we don't +* need to allocate a page table. +*/ +ASSERT (lvl ==L1_FRAME || (*pte & _PAGE_PSE)); + *result = pte; +return 1 ; } -ASSERT (tab[offset] & _PAGE_PRESENT); -if ( tab[offset] & _PAGE_PSE ) -return &tab[offset]; - pt_mfn =pte_to_mfn (tab[offset]); - tab =mfn_to_virt (pt_mfn); + pt_mfn =virt_to_mfn (pte); + pt_pfn =virt_to_pfn (alloc_page ()); +if ( !pt_pfn ) +return -1 ; + idx =idx_from_va_lvl (va, lvl); +new_pt_frame (&pt_pfn, pt_mfn, idx, lvl -1 ); - offset =l1_table_offset (va); -return &tab[offset]; +return 0 ; +} + +pgentry_t *need_pgt (unsigned long va) +{ +pgentry_t *tab =NULL ; + +walk_pt (va, va, need_pgt_func, &tab); +return tab; }EXPORT_SYMBOL (need_pgt);
Signature · 1 line
2.43.0
Re: [PATCH v2 2/3] mini-os: mm: switch need_pgt() to use walk_pt()
Juergen Gross, le mar. 13 août 2024 15:41:57 +0200, a ecrit:
Quoted 3 lines
Instead of open coding a page table walk, use
walk_pt()inneed_pgt().Signed-off-by Juergen Gross <[email protected]>
Quoted 102 lines
---
V2:
- add comment andASSERT()(Samuel Thibault)--- arch/x86/mm.c | 72 +++++++++++++++++++++------------------------------ 1 file changed, 30 insertions(+), 42 deletions(-)arch/x86/mm.c +30
-42@@ -523,57 +523,45 @@ static pgentry_t *get_pgt(unsigned long va)* return a valid PTE for a given virtual address. If PTE does not exist, * allocate page-table pages. */ -pgentry_t *need_pgt (unsigned long va) +static int need_pgt_func (unsigned long va,unsigned int lvl,bool is_leaf, +pgentry_t *pte,void *par) { +pgentry_t **result = par;unsigned long pt_mfn; -pgentry_t *tab;unsigned long pt_pfn; -unsigned offset; +unsigned int idx; - tab = pt_base; - pt_mfn =virt_to_mfn (pt_base); +if ( !is_leaf ) +return 0 ; -#if defined(__x86_64__) - offset =l4_table_offset (va); -if ( !(tab[offset] & _PAGE_PRESENT) ) - { - pt_pfn =virt_to_pfn (alloc_page ()); -if ( !pt_pfn ) -return NULL ; -new_pt_frame (&pt_pfn, pt_mfn, offset,L3_FRAME ); - } -ASSERT (tab[offset] & _PAGE_PRESENT); - pt_mfn =pte_to_mfn (tab[offset]); - tab =mfn_to_virt (pt_mfn); -#endif - offset =l3_table_offset (va); -if ( !(tab[offset] & _PAGE_PRESENT) ) - { - pt_pfn =virt_to_pfn (alloc_page ()); -if ( !pt_pfn ) -return NULL ; -new_pt_frame (&pt_pfn, pt_mfn, offset,L2_FRAME ); - } -ASSERT (tab[offset] & _PAGE_PRESENT); - pt_mfn =pte_to_mfn (tab[offset]); - tab =mfn_to_virt (pt_mfn); - offset =l2_table_offset (va); -if ( !(tab[offset] & _PAGE_PRESENT) ) +if ( lvl ==L1_FRAME || (*pte & _PAGE_PRESENT) ) { - pt_pfn =virt_to_pfn (alloc_page ()); -if ( !pt_pfn ) -return NULL ; -new_pt_frame (&pt_pfn, pt_mfn, offset,L1_FRAME ); +/* +* The PTE is not addressing a page table (is_leaf is true). If we are +* either at the lowest level or we have a valid large page, we don't +* need to allocate a page table. +*/ +ASSERT (lvl ==L1_FRAME || (*pte & _PAGE_PSE)); + *result = pte; +return 1 ; } -ASSERT (tab[offset] & _PAGE_PRESENT); -if ( tab[offset] & _PAGE_PSE ) -return &tab[offset]; - pt_mfn =pte_to_mfn (tab[offset]); - tab =mfn_to_virt (pt_mfn); + pt_mfn =virt_to_mfn (pte); + pt_pfn =virt_to_pfn (alloc_page ()); +if ( !pt_pfn ) +return -1 ; + idx =idx_from_va_lvl (va, lvl); +new_pt_frame (&pt_pfn, pt_mfn, idx, lvl -1 ); - offset =l1_table_offset (va); -return &tab[offset]; +return 0 ; +} + +pgentry_t *need_pgt (unsigned long va) +{ +pgentry_t *tab =NULL ; + +walk_pt (va, va, need_pgt_func, &tab); +return tab; }EXPORT_SYMBOL (need_pgt);Signature · 1 line
2.43.0
Signature · 5 lines
Samuel <N> je déteste import <N> parce que lorsque tu fais du python et que tu oublies le #!/bin/env python et que tu mets le fichier exécutable <N> import est exécuté -+- #ens-mim - pourquoi mon script python change le curseur de la souris ?! -+-
PATCH v2 3/3 mini-os: mm: convert set_readonly() to use walk_pt()
In reply to Juergen GrossInstead of having another copy of a page table walk in set_readonly(), just use walk_pt().
As it will be needed later anyway, split out the TLB flushing into a dedicated function.
---
V2:
- clear count after doing an mmu_update call (Samuel Thibault)
- do final mmu_update call from set_readonly() if needed (Samuel Thibault)
--- arch/x86/mm.c | 124 +++++++++++++++++++++++--------------------------- 1 file changed, 56 insertions(+), 68 deletions(-)
arch/x86/mm.c +56 -68
@@ -402,92 +402,80 @@ static void build_pagetable(unsigned long *start_pfn, unsigned long *max_pfn)* Mark portion of the address space read only. */ extern struct shared_info shared_info; -static void set_readonly (void *text,void *etext) -{ -unsigned long start_address = - ((unsigned long ) text +PAGE_SIZE -1 ) &PAGE_MASK ; -unsigned long end_address = (unsigned long ) etext; -pgentry_t *tab = pt_base, page; -unsigned long mfn =pfn_to_mfn (virt_to_pfn (pt_base)); -unsigned long offset; -unsigned long page_size =PAGE_SIZE ; + +struct set_readonly_par { +unsigned long etext;#ifdef CONFIG_PARAVIRT -int count =0 ; -int rc; +unsigned int count;#endif +}; -printk ("setting %p-%p readonly\n" , text, etext); +static int set_readonly_func (unsigned long va,unsigned int lvl,bool is_leaf, +pgentry_t *pte,void *par) +{ +struct set_readonly_par *ro = par; -while ( start_address + page_size <= end_address ) - { - tab = pt_base; - mfn =pfn_to_mfn (virt_to_pfn (pt_base)); +if ( !is_leaf ) +return 0 ; -#if defined(__x86_64__) - offset =l4_table_offset (start_address); - page = tab[offset]; - mfn =pte_to_mfn (page); - tab =to_virt (mfn_to_pfn (mfn) <<PAGE_SHIFT ); -#endif - offset =l3_table_offset (start_address); - page = tab[offset]; - mfn =pte_to_mfn (page); - tab =to_virt (mfn_to_pfn (mfn) <<PAGE_SHIFT ); - offset =l2_table_offset (start_address); -if ( !(tab[offset] & _PAGE_PSE) ) - { - page = tab[offset]; - mfn =pte_to_mfn (page); - tab =to_virt (mfn_to_pfn (mfn) <<PAGE_SHIFT ); +if ( va + (1UL << ptdata[lvl].shift) > ro->etext ) +return 1 ; - offset =l1_table_offset (start_address); - } +if ( va == (unsigned long )&shared_info ) + { +printk ("skipped %lx\n" , va); +return 0 ; + } -if ( start_address != (unsigned long )&shared_info ) - {#ifdef CONFIG_PARAVIRT - mmu_updates[count].ptr = - ((pgentry_t )mfn <<PAGE_SHIFT ) +sizeof (pgentry_t) * offset; - mmu_updates[count].val = tab[offset] & ~_PAGE_RW; - count++; + mmu_updates[ro->count].ptr =virt_to_mach (pte); + mmu_updates[ro->count].val = *pte & ~_PAGE_RW; + ro->count++; + +if ( ro->count ==L1_PAGETABLE_ENTRIES ) + { + ro->count =0 ; +if (HYPERVISOR_mmu_update (mmu_updates, ro->count,NULL , +DOMID_SELF ) <0 ) +BUG (); + }#else - tab[offset] &= ~_PAGE_RW; + *pte &= ~_PAGE_RW;#endif - } -else -printk ("skipped %lx\n" , start_address); - start_address += page_size; +return 0 ; +}#ifdef CONFIG_PARAVIRT -if ( count ==L1_PAGETABLE_ENTRIES || - start_address + page_size > end_address ) - { - rc =HYPERVISOR_mmu_update (mmu_updates, count,NULL ,DOMID_SELF ); -if ( rc <0 ) - { -printk ("ERROR: set_readonly(): PTE could not be updated\n" ); -do_exit (); - } - count =0 ; - } +static void tlb_flush (void ) +{ +mmuext_op_t op = { .cmd =MMUEXT_TLB_FLUSH_ALL }; +int count; + +HYPERVISOR_mmuext_op (&op,1 , &count,DOMID_SELF ); +}#else -if ( start_address == (1UL <<L2_PAGETABLE_SHIFT ) ) - page_size =1UL <<L2_PAGETABLE_SHIFT ; +static void tlb_flush (void ) +{ +write_cr3 ((unsigned long )pt_base); +}#endif - } + +static void set_readonly (void *text,void *etext) +{ +struct set_readonly_par setro = { .etext = (unsigned long )etext }; +unsigned long start_address =PAGE_ALIGN ((unsigned long )text); + +printk ("setting %p-%p readonly\n" , text, etext); +walk_pt (start_address, setro.etext, set_readonly_func, &setro);#ifdef CONFIG_PARAVIRT - { -mmuext_op_t op = { - .cmd =MMUEXT_TLB_FLUSH_ALL , - }; -int count; -HYPERVISOR_mmuext_op (&op,1 , &count,DOMID_SELF ); - } -#else -write_cr3 ((unsigned long )pt_base); +if ( setro.count && +HYPERVISOR_mmu_update (mmu_updates, setro.count,NULL ,DOMID_SELF ) <0 ) +BUG ();#endif + +tlb_flush (); }/*
Signature · 1 line
2.43.0
Re: [PATCH v2 3/3] mini-os: mm: convert set_readonly() to use walk_pt()
Juergen Gross, le mar. 13 août 2024 15:41:58 +0200, a ecrit:
arch/x86/mm.c static void build_pagetable(unsigned long *start_pfn, unsigned long *max_pfn) · line 439
Quoted 5 lines
arch/x86/mm.c +5
-0+if ( ro->count ==L1_PAGETABLE_ENTRIES ) + { + ro->count =0 ; +if (HYPERVISOR_mmu_update (mmu_updates, ro->count,NULL , +DOMID_SELF ) <0 )
You need to set ro->count *after* calling mmu_update.
Apart from this,
Samuel
Re: [PATCH v2 3/3] mini-os: mm: convert set_readonly() to use walk_pt()
Quoted 8 lines · Samuel Thibault
Juergen Gross, le mar. 13 août 2024 15:41:58 +0200, a ecrit:
arch/x86/mm.c static void build_pagetable(unsigned long *start_pfn, unsigned long *max_pfn) · line 439
Quoted 5 lines
arch/x86/mm.c +5
-0+if ( ro->count ==L1_PAGETABLE_ENTRIES ) + { + ro->count =0 ; +if (HYPERVISOR_mmu_update (mmu_updates, ro->count,NULL , +DOMID_SELF ) <0 )You need to set
ro->count*after* callingmmu_update.
Happy to move the line while committing, so long as Jürgen agrees.
Jan
Quoted 6 lines
Apart from this,
Reviewed-by Samuel Thibault <[email protected]>Samuel
Re: [PATCH v2 3/3] mini-os: mm: convert set_readonly() to use walk_pt()
Quoted 11 lines · Jan Beulich
Quoted 8 lines · Samuel Thibault
Juergen Gross, le mar. 13 août 2024 15:41:58 +0200, a ecrit:
arch/x86/mm.c static void build_pagetable(unsigned long *start_pfn, unsigned long *max_pfn) · line 439
Quoted 5 lines
arch/x86/mm.c +5
-0+if ( ro->count ==L1_PAGETABLE_ENTRIES ) + { + ro->count =0 ; +if (HYPERVISOR_mmu_update (mmu_updates, ro->count,NULL , +DOMID_SELF ) <0 )You need to set
ro->count*after* callingmmu_update.Happy to move the line while committing, so long as Jürgen agrees.
Oh yes, of course I do.
Thanks,
Juergen