Re: [PATCH v2 1/5] mm: pagewalk: add the ability to install PTEs

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 10/20/24 18:20, Lorenzo Stoakes wrote:
> The existing generic pagewalk logic permits the walking of page tables,
> invoking callbacks at individual page table levels via user-provided
> mm_walk_ops callbacks.
> 
> This is useful for traversing existing page table entries, but precludes
> the ability to establish new ones.
> 
> Existing mechanism for performing a walk which also installs page table
> entries if necessary are heavily duplicated throughout the kernel, each
> with semantic differences from one another and largely unavailable for use
> elsewhere.
> 
> Rather than add yet another implementation, we extend the generic pagewalk
> logic to enable the installation of page table entries by adding a new
> install_pte() callback in mm_walk_ops. If this is specified, then upon
> encountering a missing page table entry, we allocate and install a new one
> and continue the traversal.
> 
> If a THP huge page is encountered, we make use of existing logic to split
> it. Then once we reach the PTE level, we invoke the install_pte() callback
> which provides a PTE entry to install. We do not support hugetlb at this
> stage.
> 
> If this function returns an error, or an allocation fails during the
> operation, we abort the operation altogether. It is up to the caller to
> deal appropriately with partially populated page table ranges.
> 
> If install_pte() is defined, the semantics of pte_entry() change - this
> callback is then only invoked if the entry already exists. This is a useful
> property, as it allows a caller to handle existing PTEs while installing
> new ones where necessary in the specified range.
> 
> If install_pte() is not defined, then there is no functional difference to
> this patch, so all existing logic will work precisely as it did before.
> 
> As we only permit the installation of PTEs where a mapping does not already
> exist there is no need for TLB management, however we do invoke
> update_mmu_cache() for architectures which require manual maintenance of
> mappings for other CPUs.
> 
> We explicitly do not allow the existing page walk API to expose this
> feature as it is dangerous and intended for internal mm use only. Therefore
> we provide a new walk_page_range_mm() function exposed only to
> mm/internal.h.
> 
> Reviewed-by: Jann Horn <jannh@xxxxxxxxxx>
> Signed-off-by: Lorenzo Stoakes <lorenzo.stoakes@xxxxxxxxxx>

<snip>

>  /*
>   * We want to know the real level where a entry is located ignoring any
>   * folding of levels which may be happening. For example if p4d is folded then
> @@ -29,9 +34,23 @@ static int walk_pte_range_inner(pte_t *pte, unsigned long addr,
>  	int err = 0;
>  
>  	for (;;) {
> -		err = ops->pte_entry(pte, addr, addr + PAGE_SIZE, walk);
> -		if (err)
> -		       break;
> +		if (ops->install_pte && pte_none(ptep_get(pte))) {
> +			pte_t new_pte;
> +
> +			err = ops->install_pte(addr, addr + PAGE_SIZE, &new_pte,
> +					       walk);
> +			if (err)
> +				break;
> +
> +			set_pte_at(walk->mm, addr, pte, new_pte);

While the guard pages install ptes unconditionally, maybe some install_pte
handler implementation would sometimes want to skip, should ve define an
error code that means its skipped and just continue instead of set_pte_at()?
Or leave it until such handler appears.

> +			/* Non-present before, so for arches that need it. */
> +			if (!WARN_ON_ONCE(walk->no_vma))
> +				update_mmu_cache(walk->vma, addr, pte);
> +		} else {
> +			err = ops->pte_entry(pte, addr, addr + PAGE_SIZE, walk);
> +			if (err)
> +				break;
> +		}
>  		if (addr >= end - PAGE_SIZE)
>  			break;
>  		addr += PAGE_SIZE;
> @@ -89,11 +108,14 @@ static int walk_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
>  again:
>  		next = pmd_addr_end(addr, end);
>  		if (pmd_none(*pmd)) {
> -			if (ops->pte_hole)
> +			if (ops->install_pte)
> +				err = __pte_alloc(walk->mm, pmd);
> +			else if (ops->pte_hole)
>  				err = ops->pte_hole(addr, next, depth, walk);
>  			if (err)
>  				break;
> -			continue;
> +			if (!ops->install_pte)
> +				continue;
>  		}
>  
>  		walk->action = ACTION_SUBTREE;
> @@ -116,7 +138,7 @@ static int walk_pmd_range(pud_t *pud, unsigned long addr, unsigned long end,
>  		 */
>  		if ((!walk->vma && (pmd_leaf(*pmd) || !pmd_present(*pmd))) ||
>  		    walk->action == ACTION_CONTINUE ||
> -		    !(ops->pte_entry))
> +		    !(ops->pte_entry || ops->install_pte))
>  			continue;

BTW, I find it hard to read this condition even before your patch, oh well.
But if I read it correctly, does it mean we're going to split a pmd-mapped
THP if we have a install_pte handler? But is that really necessary if the
pmd splitting results in all ptes populated, and thus the install_pte
handler can't do anything with any pte anyway?

>  		if (walk->vma)
> @@ -148,11 +170,14 @@ static int walk_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
>   again:
>  		next = pud_addr_end(addr, end);
>  		if (pud_none(*pud)) {
> -			if (ops->pte_hole)
> +			if (ops->install_pte)
> +				err = __pmd_alloc(walk->mm, pud, addr);
> +			else if (ops->pte_hole)
>  				err = ops->pte_hole(addr, next, depth, walk);
>  			if (err)
>  				break;
> -			continue;
> +			if (!ops->install_pte)
> +				continue;
>  		}
>  
>  		walk->action = ACTION_SUBTREE;
> @@ -167,7 +192,7 @@ static int walk_pud_range(p4d_t *p4d, unsigned long addr, unsigned long end,
>  
>  		if ((!walk->vma && (pud_leaf(*pud) || !pud_present(*pud))) ||
>  		    walk->action == ACTION_CONTINUE ||
> -		    !(ops->pmd_entry || ops->pte_entry))
> +		    !(ops->pmd_entry || ops->pte_entry || ops->install_pte))
>  			continue;

Ditto?





[Index of Archives]     [Linux Kernel]     [Kernel Newbies]     [x86 Platform Driver]     [Netdev]     [Linux Wireless]     [Netfilter]     [Bugtraq]     [Linux Filesystems]     [Yosemite Discussion]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Samba]     [Device Mapper]

  Powered by Linux