Re: [PATCH 2/2] xen/swiotlb: don't destroy contiguous region in all cases

[Jan Beulich <jbeulich@xxxxxxxx>]

On 11.02.2025 13:04, Juergen Gross wrote:
> In case xen_swiotlb_alloc_coherent() needed to create a contiguous
> region only for other reason than the memory not being compliant with
> the device's DMA mask, there is no reason why this contiguous region
> should be destroyed by xen_swiotlb_free_coherent() later. Destroying
> this region should be done only, if the memory of the region was
> allocated with more stringent placement requirements than the memory
> it did replace.

I'm not convinced of this: Even the mere property of being contiguous
may already be enough to warrant freeing when possible. The hypervisor
may not have that many contiguous areas available. The bigger the
chunk, the more important to give it back once no longer needed in
this shape.

Plus also take into account how Xen behaves here: It specifically tries
to hold back, during boot, lower addressed memory to later satisfy such
requests. Hence even if you don't ask for address restricted memory,
you may get back such. You'd need to compare input and output addresses,
not input addresses and requested restriction to alleviate this.

> --- a/arch/x86/xen/mmu_pv.c
> +++ b/arch/x86/xen/mmu_pv.c
> @@ -2208,19 +2208,22 @@ void __init xen_init_mmu_ops(void)
>  static unsigned long discontig_frames[1<<MAX_CONTIG_ORDER];
>  
>  #define VOID_PTE (mfn_pte(0, __pgprot(0)))
> -static void xen_zap_pfn_range(unsigned long vaddr, unsigned int order,
> -                             unsigned long *in_frames,
> -                             unsigned long *out_frames)
> +static int xen_zap_pfn_range(unsigned long vaddr, unsigned int order,
> +                          unsigned long *in_frames,
> +                          unsigned long *out_frames)
>  {
>       int i;
> +     u64 address_bits = 0;

First I was inclined to suggest to use paddr_t here, but ...

>       struct multicall_space mcs;
>  
>       xen_mc_batch();
>       for (i = 0; i < (1UL<<order); i++, vaddr += PAGE_SIZE) {
>               mcs = __xen_mc_entry(0);
>  
> -             if (in_frames)
> +             if (in_frames) {
>                       in_frames[i] = virt_to_mfn((void *)vaddr);
> +                     address_bits |= in_frames[i] << PAGE_SHIFT;

... why do a shift on every loop iteration when you can ...

> +             }
>  
>               MULTI_update_va_mapping(mcs.mc, vaddr, VOID_PTE, 0);
>               __set_phys_to_machine(virt_to_pfn((void *)vaddr), 
> INVALID_P2M_ENTRY);
> @@ -2229,6 +2232,8 @@ static void xen_zap_pfn_range(unsigned long vaddr, 
> unsigned int order,
>                       out_frames[i] = virt_to_pfn((void *)vaddr);
>       }
>       xen_mc_issue(0);
> +
> +     return fls64(address_bits);

... simply add in PAGE_SHIFT here, once?

> @@ -2321,7 +2326,8 @@ static int xen_exchange_memory(unsigned long 
> extents_in, unsigned int order_in,
>  
>  int xen_create_contiguous_region(phys_addr_t pstart, unsigned int order,
>                                unsigned int address_bits,
> -                              dma_addr_t *dma_handle)
> +                              dma_addr_t *dma_handle,
> +                              unsigned int *address_bits_in)
>  {
>       unsigned long *in_frames = discontig_frames, out_frame;
>       unsigned long  flags;
> @@ -2336,7 +2342,7 @@ int xen_create_contiguous_region(phys_addr_t pstart, 
> unsigned int order,
>       spin_lock_irqsave(&xen_reservation_lock, flags);
>  
>       /* 1. Zap current PTEs, remembering MFNs. */
> -     xen_zap_pfn_range(vstart, order, in_frames, NULL);
> +     *address_bits_in = xen_zap_pfn_range(vstart, order, in_frames, NULL);

Nit: Converting plain int to unsigned int, when there's no real reason
to do any conversion. Since xen_zap_pfn_range() can't return a negative
value for the caller caring about the return value (yet more obviously
so with the suggested adjustment, and then true for both callers), the
function could easily return unsigned int.

Jan