Re: [PATCH v4 6/6] xen/ppc: mm-radix: Allocate all paging structures at runtime

[Jan Beulich <jbeulich@xxxxxxxx>]

On 12.04.2024 05:55, Shawn Anastasio wrote:
> In the initial mm-radix implementation, the in-memory partition and
> process tables required to configure the MMU, as well as the page tables
> themselves were all allocated statically since the boot allocator was
> not yet available.
> 
> Now that it is, allocate these structures at runtime and bump the size
> of the Process Table to its maximum supported value (on POWER9).
> 
> Signed-off-by: Shawn Anastasio <sanastasio@xxxxxxxxxxxxxxxxxxxxx>
> ---
> Changes in v4:
>   - use mfn_add in initial_page_alloc()
>   - zero pages returned by initial_page_alloc()
> 
>  xen/arch/ppc/mm-radix.c | 231 +++++++++++++++++++++-------------------
>  1 file changed, 123 insertions(+), 108 deletions(-)
> 
> diff --git a/xen/arch/ppc/mm-radix.c b/xen/arch/ppc/mm-radix.c
> index ab5a10695c..8f4bfa87c0 100644
> --- a/xen/arch/ppc/mm-radix.c
> +++ b/xen/arch/ppc/mm-radix.c
> @@ -21,69 +21,105 @@ void enable_mmu(void);
>  #define radix_dprintk(...)
>  #endif
> 
> -#define INITIAL_LVL1_PD_COUNT      1
> -#define INITIAL_LVL2_LVL3_PD_COUNT 2
> -#define INITIAL_LVL4_PT_COUNT      256
> -
> -static size_t __initdata initial_lvl1_pd_pool_used;
> -static struct lvl1_pd initial_lvl1_pd_pool[INITIAL_LVL1_PD_COUNT];
> -
> -static size_t __initdata initial_lvl2_lvl3_pd_pool_used;
> -static struct lvl2_pd initial_lvl2_lvl3_pd_pool[INITIAL_LVL2_LVL3_PD_COUNT];
> -
> -static size_t __initdata initial_lvl4_pt_pool_used;
> -static struct lvl4_pt initial_lvl4_pt_pool[INITIAL_LVL4_PT_COUNT];
> -
> -/* Only reserve minimum Partition and Process tables  */
>  #define PATB_SIZE_LOG2 16 /* Only supported partition table size on POWER9 */
>  #define PATB_SIZE      (1UL << PATB_SIZE_LOG2)
> -#define PRTB_SIZE_LOG2 12
> +#define PRTB_SIZE_LOG2 24 /* Maximum process table size on POWER9 */
>  #define PRTB_SIZE      (1UL << PRTB_SIZE_LOG2)
> 
> -static struct patb_entry
> -    __aligned(PATB_SIZE) initial_patb[PATB_SIZE / sizeof(struct patb_entry)];
> +static struct patb_entry *initial_patb;
> +static struct prtb_entry *initial_prtb;
> 
> -static struct prtb_entry
> -    __aligned(PRTB_SIZE) initial_prtb[PRTB_SIZE / sizeof(struct prtb_entry)];
> +static mfn_t __initdata min_alloc_mfn = {-1};
> +static mfn_t __initdata max_alloc_mfn = {0};
> 
> -static __init struct lvl1_pd *lvl1_pd_pool_alloc(void)
> +/*
> + * A thin wrapper for alloc_boot_pages that keeps track of the maximum and
> + * minimum mfns that have been allocated. This information is used by
> + * setup_initial_mapping to include the allocated pages in the initial
> + * page mapping.

Plus everything in between. Together with the further comment below I'm
afraid I still think that the constraints (and the justification for them
being tolerable at least for the time being) aren't made sufficiently
clear (perhaps in another code comment next to the respective two
variables' definitions).

> + * Additionally, allocated pages are zeroed before return.
> + */
> +static mfn_t __init initial_page_alloc(unsigned long nr_pfns,
> +                                       unsigned long pfn_align)
>  {
> -    if ( initial_lvl1_pd_pool_used >= INITIAL_LVL1_PD_COUNT )
> -    {
> -        early_printk("Ran out of space for LVL1 PD!\n");
> -        die();
> -    }
> +    mfn_t mfn_first, mfn_last;
> 
> -    return &initial_lvl1_pd_pool[initial_lvl1_pd_pool_used++];
> -}
> +    mfn_first = alloc_boot_pages(nr_pfns, pfn_align);
> +    mfn_last = mfn_add(mfn_first, nr_pfns - 1);
> 
> -static __init struct lvl2_pd *lvl2_pd_pool_alloc(void)
> -{
> -    if ( initial_lvl2_lvl3_pd_pool_used >= INITIAL_LVL2_LVL3_PD_COUNT )
> -    {
> -        early_printk("Ran out of space for LVL2/3 PD!\n");
> -        die();
> -    }
> +    min_alloc_mfn = _mfn(min(mfn_x(min_alloc_mfn), mfn_x(mfn_first)));
> +    max_alloc_mfn = _mfn(max(mfn_x(max_alloc_mfn), mfn_x(mfn_last)));
> 
> -    return &initial_lvl2_lvl3_pd_pool[initial_lvl2_lvl3_pd_pool_used++];
> +    memset(__va(mfn_to_maddr(mfn_first)), 0, nr_pfns << PAGE_SHIFT);
> +
> +    return mfn_first;
>  }
> 
> -static __init struct lvl3_pd *lvl3_pd_pool_alloc(void)
> +static __init void *initial_pd_pt_alloc(void)
>  {
> -    BUILD_BUG_ON(sizeof(struct lvl3_pd) != sizeof(struct lvl2_pd));
> +    BUILD_BUG_ON(sizeof(struct lvl1_pd) > PAGE_SIZE);
> +    BUILD_BUG_ON(sizeof(struct lvl2_pd) > PAGE_SIZE);
> +    BUILD_BUG_ON(sizeof(struct lvl3_pd) > PAGE_SIZE);
> +    BUILD_BUG_ON(sizeof(struct lvl4_pt) > PAGE_SIZE);
> 
> -    return (struct lvl3_pd *) lvl2_pd_pool_alloc();
> +    return __va(mfn_to_maddr(initial_page_alloc(1, 1)));
>  }
> 
> -static __init struct lvl4_pt *lvl4_pt_pool_alloc(void)
> +static void map_page_initial(struct lvl1_pd *lvl1, vaddr_t virt, paddr_t 
> phys,
> +                             unsigned long flags)
>  {
> -    if ( initial_lvl4_pt_pool_used >= INITIAL_LVL4_PT_COUNT )
> +    struct lvl2_pd *lvl2;
> +    struct lvl3_pd *lvl3;
> +    struct lvl4_pt *lvl4;
> +    pde_t *pde;
> +    pte_t *pte;
> +
> +    /* Allocate LVL 2 PD if necessary */
> +    pde = pt_entry(lvl1, virt);
> +    if ( !pde_is_valid(*pde) )
>      {
> -        early_printk("Ran out of space for LVL4 PT!\n");
> -        die();
> +        lvl2 = initial_pd_pt_alloc();
> +        *pde = paddr_to_pde(__pa(lvl2), PDE_VALID,
> +                            XEN_PT_ENTRIES_LOG2_LVL_2);
> +    }
> +    else
> +        lvl2 = __va(pde_to_paddr(*pde));
> +
> +    /* Allocate LVL 3 PD if necessary */
> +    pde = pt_entry(lvl2, virt);
> +    if ( !pde_is_valid(*pde) )
> +    {
> +        lvl3 = initial_pd_pt_alloc();
> +        *pde = paddr_to_pde(__pa(lvl3), PDE_VALID,
> +                            XEN_PT_ENTRIES_LOG2_LVL_3);
> +    }
> +    else
> +        lvl3 = __va(pde_to_paddr(*pde));
> +
> +    /* Allocate LVL 4 PT if necessary */
> +    pde = pt_entry(lvl3, virt);
> +    if ( !pde_is_valid(*pde) )
> +    {
> +        lvl4 = initial_pd_pt_alloc();
> +        *pde = paddr_to_pde(__pa(lvl4), PDE_VALID,
> +                            XEN_PT_ENTRIES_LOG2_LVL_4);
>      }
> +    else
> +        lvl4 = __va(pde_to_paddr(*pde));
> 
> -    return &initial_lvl4_pt_pool[initial_lvl4_pt_pool_used++];
> +    /* Finally, create PTE in LVL 4 PT */
> +    pte = pt_entry(lvl4, virt);
> +    if ( !pte_is_valid(*pte) )
> +    {
> +        radix_dprintk("%016lx being mapped to %016lx\n", phys, virt);
> +        *pte = paddr_to_pte(phys, flags);
> +    }
> +    else
> +    {
> +        early_printk("BUG: Tried to create PTE for already-mapped page!");
> +        die();
> +    }
>  }
> 
>  static void __init setup_initial_mapping(struct lvl1_pd *lvl1,
> @@ -92,6 +128,7 @@ static void __init setup_initial_mapping(struct lvl1_pd 
> *lvl1,
>                                           paddr_t phys_base)
>  {
>      uint64_t page_addr;
> +    mfn_t previous_max_alloc_mfn;
> 
>      if ( map_start & ~PAGE_MASK )
>      {
> @@ -105,81 +142,47 @@ static void __init setup_initial_mapping(struct lvl1_pd 
> *lvl1,
>          die();
>      }
> 
> +    /* Identity map Xen itself */
>      for ( page_addr = map_start; page_addr < map_end; page_addr += PAGE_SIZE 
> )
>      {
> -        struct lvl2_pd *lvl2;
> -        struct lvl3_pd *lvl3;
> -        struct lvl4_pt *lvl4;
> -        pde_t *pde;
> -        pte_t *pte;
> -
> -        /* Allocate LVL 2 PD if necessary */
> -        pde = pt_entry(lvl1, page_addr);
> -        if ( !pde_is_valid(*pde) )
> -        {
> -            lvl2 = lvl2_pd_pool_alloc();
> -            *pde = paddr_to_pde(__pa(lvl2), PDE_VALID,
> -                                XEN_PT_ENTRIES_LOG2_LVL_2);
> -        }
> -        else
> -            lvl2 = __va(pde_to_paddr(*pde));
> +        unsigned long flags;
> 
> -        /* Allocate LVL 3 PD if necessary */
> -        pde = pt_entry(lvl2, page_addr);
> -        if ( !pde_is_valid(*pde) )
> +        if ( is_kernel_text(page_addr) || is_kernel_inittext(page_addr) )
>          {
> -            lvl3 = lvl3_pd_pool_alloc();
> -            *pde = paddr_to_pde(__pa(lvl3), PDE_VALID,
> -                                XEN_PT_ENTRIES_LOG2_LVL_3);
> +            radix_dprintk("%016lx being marked as TEXT (RX)\n", page_addr);
> +            flags = PTE_XEN_RX;
>          }
> -        else
> -            lvl3 = __va(pde_to_paddr(*pde));
> -
> -        /* Allocate LVL 4 PT if necessary */
> -        pde = pt_entry(lvl3, page_addr);
> -        if ( !pde_is_valid(*pde) )
> +        else if ( is_kernel_rodata(page_addr) )
>          {
> -            lvl4 = lvl4_pt_pool_alloc();
> -            *pde = paddr_to_pde(__pa(lvl4), PDE_VALID,
> -                                XEN_PT_ENTRIES_LOG2_LVL_4);
> +            radix_dprintk("%016lx being marked as RODATA (RO)\n", page_addr);
> +            flags = PTE_XEN_RO;
>          }
>          else
> -            lvl4 = __va(pde_to_paddr(*pde));
> -
> -        /* Finally, create PTE in LVL 4 PT */
> -        pte = pt_entry(lvl4, page_addr);
> -        if ( !pte_is_valid(*pte) )
>          {
> -            unsigned long paddr = (page_addr - map_start) + phys_base;
> -            unsigned long flags;
> -
> -            radix_dprintk("%016lx being mapped to %016lx\n", paddr, 
> page_addr);
> -            if ( is_kernel_text(page_addr) || is_kernel_inittext(page_addr) )
> -            {
> -                radix_dprintk("%016lx being marked as TEXT (RX)\n", 
> page_addr);
> -                flags = PTE_XEN_RX;
> -            }
> -            else if ( is_kernel_rodata(page_addr) )
> -            {
> -                radix_dprintk("%016lx being marked as RODATA (RO)\n", 
> page_addr);
> -                flags = PTE_XEN_RO;
> -            }
> -            else
> -            {
> -                radix_dprintk("%016lx being marked as DEFAULT (RW)\n", 
> page_addr);
> -                flags = PTE_XEN_RW;
> -            }
> -
> -            *pte = paddr_to_pte(paddr, flags);
> -            radix_dprintk("%016lx is the result of PTE map\n",
> -                paddr_to_pte(paddr, flags).pte);
> -        }
> -        else
> -        {
> -            early_printk("BUG: Tried to create PTE for already-mapped 
> page!");
> -            die();
> +            radix_dprintk("%016lx being marked as DEFAULT (RW)\n", 
> page_addr);
> +            flags = PTE_XEN_RW;
>          }
> +
> +        map_page_initial(lvl1, page_addr, (page_addr - map_start) + 
> phys_base, flags);
> +    }
> +
> +    previous_max_alloc_mfn = max_alloc_mfn;
> +
> +    /*
> +     * Identity map all pages we've allocated for paging structures. This act
> +     * itself will allocate more pages, so continue until we've mapped from
> +     * `max_alloc_mfn` down to `min_alloc_mfn`. This assumes that the heap 
> grows
> +     * downwards, which matches the behavior of alloc_boot_pages.

... as long as you only ever allocate individual pages (which looks
to be the case for such incremental allocations, but that imo doesn't
go without saying).

Jan