On 21/09/11 15:44, David Vrabel wrote:
> On 15/09/11 22:37, Jeremy Fitzhardinge wrote:
>> On 09/15/2011 05:40 AM, David Vrabel wrote:
>>> This set of pages avoids the need to call vmalloc_sync_all() when
>>> mapping foreign pages. Two new functions are adding for mapping
>>> foreign pages onto RAM pages instead of vmalloc space.
>>>
>>> [...]
>>
>> But that said, if you want to allocate virtual addresses for mapping
>> granted pages, then alloc_vm_area() is the right thing to use. You need
>> to make sure you touch the pages from within the kernel before passing
>> those addresses to a hypercall to make sure the mappings are established
>> within the current task (possibly within a no-preempt region to
>> guarantee that nothing odd happens). Or alternatively, you could switch
>> the current pagetable to init_mm for the hypercall (if it isn't already)
>> - since that's the reference pagetable - assuming you're not passing
>> usermode virtual addresses to the hypercall.
>
> Making alloc_vm_area() fill in a array of pte_t *'s and passing the pointer
> to the PTE (using the GNTMAP_contains_pte flag) in the hypercall should
> allow Xen to update the PTEs in init_mm directly.
>
> However, I'm not sure what to do about ia64 where GNTMAP_contains_pte
> is not supported. Any ideas?
>
> Here's the untested patch I have so far.
This (sort of) works with some hacks/changes (described below).
> diff --git a/arch/x86/xen/grant-table.c b/arch/x86/xen/grant-table.c
> index 49ba9b5..77348e8 100644
> --- a/arch/x86/xen/grant-table.c
> +++ b/arch/x86/xen/grant-table.c
> @@ -71,7 +71,7 @@ int arch_gnttab_map_shared(unsigned long *frames, unsigned
> long nr_gframes,
>
> if (shared == NULL) {
> struct vm_struct *area =
> - alloc_vm_area(PAGE_SIZE * max_nr_gframes);
> + alloc_vm_area(PAGE_SIZE * max_nr_gframes, NULL);
> BUG_ON(area == NULL);
> shared = area->addr;
> *__shared = shared;
> diff --git a/drivers/xen/xenbus/xenbus_client.c
> b/drivers/xen/xenbus/xenbus_client.c
> index cdacf92..75eb179 100644
> --- a/drivers/xen/xenbus/xenbus_client.c
> +++ b/drivers/xen/xenbus/xenbus_client.c
> @@ -435,19 +435,20 @@ EXPORT_SYMBOL_GPL(xenbus_free_evtchn);
> int xenbus_map_ring_valloc(struct xenbus_device *dev, int gnt_ref, void
> **vaddr)
> {
> struct gnttab_map_grant_ref op = {
> - .flags = GNTMAP_host_map,
> + .flags = GNTMAP_host_map | GNTMAP_contains_pte,
> .ref = gnt_ref,
> .dom = dev->otherend_id,
> };
> struct vm_struct *area;
> + pte_t *pte;
alloc_vm_area() is allocating two pages instead of the one we ask for.
I hacked this with
pte_t *pte[2];
> *vaddr = NULL;
>
> - area = alloc_vm_area(PAGE_SIZE);
> + area = alloc_vm_area(PAGE_SIZE, &pte);
> if (!area)
> return -ENOMEM;
>
> - op.host_addr = (unsigned long)area->addr;
> + op.host_addr = (unsigned long)pte;
With the GNTMAP_contains_pte flags this needs a machine address here so:
op.host_addr = arbitrary_virt_to_machine(pte).maddr;
> if (HYPERVISOR_grant_table_op(GNTTABOP_map_grant_ref, &op, 1))
> BUG();
Also need to do something with xenbus_map_ring_vfree() which doesn't work.
> diff --git a/include/linux/vmalloc.h b/include/linux/vmalloc.h
> index 9332e52..1a77252 100644
> --- a/include/linux/vmalloc.h
> +++ b/include/linux/vmalloc.h
> @@ -118,7 +118,7 @@ unmap_kernel_range(unsigned long addr, unsigned long size)
> #endif
>
> /* Allocate/destroy a 'vmalloc' VM area. */
> -extern struct vm_struct *alloc_vm_area(size_t size);
> +extern struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes);
> extern void free_vm_area(struct vm_struct *area);
>
> /* for /dev/kmem */
> diff --git a/mm/vmalloc.c b/mm/vmalloc.c
> index 5016f19..786b4f6 100644
> --- a/mm/vmalloc.c
> +++ b/mm/vmalloc.c
> @@ -2105,23 +2105,30 @@ void __attribute__((weak)) vmalloc_sync_all(void)
>
> static int f(pte_t *pte, pgtable_t table, unsigned long addr, void *data)
> {
> - /* apply_to_page_range() does all the hard work. */
> + pte_t ***p = data;
> +
> + if (p) {
> + *(*p) = pte;
> + (*p)++;
> + }
> return 0;
> }
>
> /**
> * alloc_vm_area - allocate a range of kernel address space
> * @size: size of the area
> + * @ptes: returns the PTEs for the address space
> *
> * Returns: NULL on failure, vm_struct on success
> *
> * This function reserves a range of kernel address space, and
> * allocates pagetables to map that range. No actual mappings
> - * are created. If the kernel address space is not shared
> - * between processes, it syncs the pagetable across all
> - * processes.
> + * are created.
> + *
> + * If @ptes is non-NULL, pointers to the PTEs (in init_mm)
> + * allocated for the VM area are returned.
> */
> -struct vm_struct *alloc_vm_area(size_t size)
> +struct vm_struct *alloc_vm_area(size_t size, pte_t **ptes)
> {
> struct vm_struct *area;
>
> @@ -2135,19 +2142,11 @@ struct vm_struct *alloc_vm_area(size_t size)
> * of kernel virtual address space and mapped into init_mm.
> */
> if (apply_to_page_range(&init_mm, (unsigned long)area->addr,
> - area->size, f, NULL)) {
> + area->size, f, ptes ? &ptes : NULL)) {
> free_vm_area(area);
> return NULL;
> }
>
> - /*
> - * If the allocated address space is passed to a hypercall
> - * before being used then we cannot rely on a page fault to
> - * trigger an update of the page tables. So sync all the page
> - * tables here.
> - */
> - vmalloc_sync_all();
> -
> return area;
> }
> EXPORT_SYMBOL_GPL(alloc_vm_area);
_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel
|
Home