[RFCv2,12/15] xen/arm: add Persistent Map (PMAP) infrastructure

Commit Message

Julien Grall April 25, 2021, 8:13 p.m. UTC

From: Wei Liu <[e-mail]>

The basic idea is like Persistent Kernel Map (PKMAP) in Linux. We
pre-populate all the relevant page tables before the system is fully
set up.

We will need it on Arm in order to rework the arm64 version of
xenheap_setup_mappings() as we may need to use pages allocated from
the boot allocator before they are effectively mapped.

This infrastructure is not lock-protected therefore can only be used
before smpboot. After smpboot, map_domain_page() has to be used.

This is based on the x86 version [1] that was originally implemented
by Wei Liu.

Take the opportunity to switch the parameter attr from unsigned to
unsigned int.

[1] <[e-mail]>

Signed-off-by: Wei Liu <[e-mail]>
Signed-off-by: Hongyan Xia <[e-mail]>
[julien: Adapted for Arm]
Signed-off-by: Julien Grall <[e-mail]>

---
    Changes in v2:
        - New patch

Cc: Jan Beulich <[e-mail]>
Cc: Wei Liu <[e-mail]>
Cc: Andrew Cooper <[e-mail]>
Cc: Roger Pau Monné <[e-mail]>
Cc: Hongian Xia <[e-mail]>

This is mostly a copy of the PMAP infrastructure currently discussed
on x86. The only difference is how the page-tables are updated.

I think we want to consider to provide a common infrastructure. But
I haven't done it yet to gather feedback on the overall series
first.
---
 xen/arch/arm/Makefile        |   1 +
 xen/arch/arm/mm.c            |   7 +--
 xen/arch/arm/pmap.c          | 101 +++++++++++++++++++++++++++++++++++
 xen/include/asm-arm/config.h |   2 +
 xen/include/asm-arm/lpae.h   |   8 +++
 xen/include/asm-arm/pmap.h   |  10 ++++
 6 files changed, 123 insertions(+), 6 deletions(-)
 create mode 100644 xen/arch/arm/pmap.c
 create mode 100644 xen/include/asm-arm/pmap.h

Comments

Xia, Hongyan April 26, 2021, 9:41 a.m. UTC | #1

On Sun, 2021-04-25 at 21:13 +0100, Julien Grall wrote:
> From: Wei Liu <[e-mail]>
> 
> The basic idea is like Persistent Kernel Map (PKMAP) in Linux. We
> pre-populate all the relevant page tables before the system is fully
> set up.
> 
> We will need it on Arm in order to rework the arm64 version of
> xenheap_setup_mappings() as we may need to use pages allocated from
> the boot allocator before they are effectively mapped.
> 
> This infrastructure is not lock-protected therefore can only be used
> before smpboot. After smpboot, map_domain_page() has to be used.
> 
> This is based on the x86 version [1] that was originally implemented
> by Wei Liu.
> 
> Take the opportunity to switch the parameter attr from unsigned to
> unsigned int.
> 
> [1] <
> [e-mail]
> >
> 
> Signed-off-by: Wei Liu <[e-mail]>
> Signed-off-by: Hongyan Xia <[e-mail]>
> [julien: Adapted for Arm]
> Signed-off-by: Julien Grall <[e-mail]>

[...]

> diff --git a/xen/arch/arm/pmap.c b/xen/arch/arm/pmap.c
> new file mode 100644
> index 000000000000..702b1bde982d
> --- /dev/null
> +++ b/xen/arch/arm/pmap.c
> @@ -0,0 +1,101 @@
> +#include <xen/init.h>
> +#include <xen/mm.h>
> +
> +#include <asm/bitops.h>
> +#include <asm/flushtlb.h>
> +#include <asm/pmap.h>
> +
> +/*
> + * To be able to use FIXMAP_PMAP_BEGIN.
> + * XXX: move fixmap definition in a separate header
> + */
> +#include <xen/acpi.h>
> +
> +/*
> + * Simple mapping infrastructure to map / unmap pages in fixed map.
> + * This is used to set up the page table for mapcache, which is used
> + * by map domain page infrastructure.
> + *
> + * This structure is not protected by any locks, so it must not be
> used after
> + * smp bring-up.
> + */
> +
> +/* Bitmap to track which slot is used */
> +static unsigned long __initdata inuse;
> +
> +/* XXX: Find an header to declare it */
> +extern lpae_t xen_fixmap[LPAE_ENTRIES];
> +
> +void *__init pmap_map(mfn_t mfn)
> +{
> +    unsigned long flags;
> +    unsigned int idx;
> +    vaddr_t linear;
> +    unsigned int slot;
> +    lpae_t *entry, pte;
> +
> +    BUILD_BUG_ON(sizeof(inuse) * BITS_PER_LONG < NUM_FIX_PMAP);

This seems wrong to me. It should multiply with something like
BITS_PER_BYTE.

I noticed this line was already present before the Arm version so
probably my fault :(, which also needs to be fixed.

> +
> +    ASSERT(system_state < SYS_STATE_smp_boot);
> +
> +    local_irq_save(flags);
> +
> +    idx = find_first_zero_bit(&inuse, NUM_FIX_PMAP);
> +    if ( idx == NUM_FIX_PMAP )
> +        panic("Out of PMAP slots\n");
> +
> +    __set_bit(idx, &inuse);
> +
> +    slot = idx + FIXMAP_PMAP_BEGIN;
> +    ASSERT(slot >= FIXMAP_PMAP_BEGIN && slot <= FIXMAP_PMAP_END);
> +

From here...

> +    linear = FIXMAP_ADDR(slot);
> +    /*
> +     * We cannot use set_fixmap() here. We use PMAP when there is no
> direct map,
> +     * so map_pages_to_xen() called by set_fixmap() needs to map
> pages on
> +     * demand, which then calls pmap() again, resulting in a loop.
> Modify the
> +     * PTEs directly instead. The same is true for pmap_unmap().
> +     */
> +    entry = &xen_fixmap[third_table_offset(linear)];
> +
> +    ASSERT(!lpae_is_valid(*entry));
> +
> +    pte = mfn_to_xen_entry(mfn, PAGE_HYPERVISOR_RW);
> +    pte.pt.table = 1;
> +    write_pte(entry, pte);
> +

...to here, I wonder if we can move this chunk into arch (like void
*arch_write_pmap_slot(slot)). Such an arch function hides how fixmap is
handled and how page table entry is written behind arch, and the rest
can just be common.

> +    local_irq_restore(flags);
> +
> +    return (void *)linear;
> +}
> +
> +void __init pmap_unmap(const void *p)
> +{
> +    unsigned long flags;
> +    unsigned int idx;
> +    lpae_t *entry;
> +    lpae_t pte = { 0 };
> +    unsigned int slot = third_table_offset((vaddr_t)p);
> +
> +    ASSERT(system_state < SYS_STATE_smp_boot);
> +    ASSERT(slot >= FIXMAP_PMAP_BEGIN && slot <= FIXMAP_PMAP_END);
> +
> +    idx = slot - FIXMAP_PMAP_BEGIN;
> +    local_irq_save(flags);
> +
> +    __clear_bit(idx, &inuse);
> +    entry = &xen_fixmap[third_table_offset((vaddr_t)p)];
> +    write_pte(entry, pte);
> +    flush_xen_tlb_range_va_local((vaddr_t)p, PAGE_SIZE);

and the same for the above, something like arch_clear_pmap(void *) and
the rest into common.

From a quick glance, I don't think x86 and Arm share any useful TLB
flush helpers? So the TLB flush probably should be behind arch as well.

> +
> +    local_irq_restore(flags);
> +}
> +
> +/*
> + * Local variables:
> + * mode: C
> + * c-file-style: "BSD"
> + * c-basic-offset: 4
> + * indent-tabs-mode: nil
> + * End:
> + */

[...]

> diff --git a/xen/include/asm-arm/pmap.h b/xen/include/asm-arm/pmap.h
> new file mode 100644
> index 000000000000..8e1dce93f8e4
> --- /dev/null
> +++ b/xen/include/asm-arm/pmap.h
> @@ -0,0 +1,10 @@
> +#ifndef __ASM_PMAP_H__
> +#define __ARM_PMAP_H__

This line doesn't seem to match the #ifndef, but if the functions are
moved to common, this header can be moved to common as well.

Hongyan

Wei Liu April 28, 2021, 9:47 p.m. UTC | #2

On Sun, Apr 25, 2021 at 09:13:15PM +0100, Julien Grall wrote:
> From: Wei Liu <[e-mail]>
> 
> The basic idea is like Persistent Kernel Map (PKMAP) in Linux. We
> pre-populate all the relevant page tables before the system is fully
> set up.
> 
> We will need it on Arm in order to rework the arm64 version of
> xenheap_setup_mappings() as we may need to use pages allocated from
> the boot allocator before they are effectively mapped.
> 
> This infrastructure is not lock-protected therefore can only be used
> before smpboot. After smpboot, map_domain_page() has to be used.
> 
> This is based on the x86 version [1] that was originally implemented
> by Wei Liu.
> 
> Take the opportunity to switch the parameter attr from unsigned to
> unsigned int.
> 
> [1] <[e-mail]>
> 
> Signed-off-by: Wei Liu <[e-mail]>
> Signed-off-by: Hongyan Xia <[e-mail]>
> [julien: Adapted for Arm]
> Signed-off-by: Julien Grall <[e-mail]>
> 
> ---
>     Changes in v2:
>         - New patch
> 
> Cc: Jan Beulich <[e-mail]>
> Cc: Wei Liu <[e-mail]>
> Cc: Andrew Cooper <[e-mail]>
> Cc: Roger Pau Monné <[e-mail]>
> Cc: Hongian Xia <[e-mail]>
> 
> This is mostly a copy of the PMAP infrastructure currently discussed
> on x86. The only difference is how the page-tables are updated.
> 
> I think we want to consider to provide a common infrastructure. But
> I haven't done it yet to gather feedback on the overall series
> first.

+1. This infrastructure should be common code with arch hooks where
appropriate.

Wei.

Stefano Stabellini May 14, 2021, 11:25 p.m. UTC | #3

On Sun, 25 Apr 2021, Julien Grall wrote:
> From: Wei Liu <[e-mail]>
> 
> The basic idea is like Persistent Kernel Map (PKMAP) in Linux. We
> pre-populate all the relevant page tables before the system is fully
> set up.
> 
> We will need it on Arm in order to rework the arm64 version of
> xenheap_setup_mappings() as we may need to use pages allocated from
> the boot allocator before they are effectively mapped.
> 
> This infrastructure is not lock-protected therefore can only be used
> before smpboot. After smpboot, map_domain_page() has to be used.
> 
> This is based on the x86 version [1] that was originally implemented
> by Wei Liu.
> 
> Take the opportunity to switch the parameter attr from unsigned to
> unsigned int.
> 
> [1] <[e-mail]>
> 
> Signed-off-by: Wei Liu <[e-mail]>
> Signed-off-by: Hongyan Xia <[e-mail]>
> [julien: Adapted for Arm]
> Signed-off-by: Julien Grall <[e-mail]>
> 
> ---
>     Changes in v2:
>         - New patch
> 
> Cc: Jan Beulich <[e-mail]>
> Cc: Wei Liu <[e-mail]>
> Cc: Andrew Cooper <[e-mail]>
> Cc: Roger Pau Monné <[e-mail]>
> Cc: Hongian Xia <[e-mail]>
> 
> This is mostly a copy of the PMAP infrastructure currently discussed
> on x86. The only difference is how the page-tables are updated.
> 
> I think we want to consider to provide a common infrastructure. But
> I haven't done it yet to gather feedback on the overall series
> first.
> ---
>  xen/arch/arm/Makefile        |   1 +
>  xen/arch/arm/mm.c            |   7 +--
>  xen/arch/arm/pmap.c          | 101 +++++++++++++++++++++++++++++++++++
>  xen/include/asm-arm/config.h |   2 +
>  xen/include/asm-arm/lpae.h   |   8 +++
>  xen/include/asm-arm/pmap.h   |  10 ++++
>  6 files changed, 123 insertions(+), 6 deletions(-)
>  create mode 100644 xen/arch/arm/pmap.c
>  create mode 100644 xen/include/asm-arm/pmap.h
> 
> diff --git a/xen/arch/arm/Makefile b/xen/arch/arm/Makefile
> index ca75f1040dcc..2196ada24941 100644
> --- a/xen/arch/arm/Makefile
> +++ b/xen/arch/arm/Makefile
> @@ -40,6 +40,7 @@ obj-y += percpu.o
>  obj-y += platform.o
>  obj-y += platform_hypercall.o
>  obj-y += physdev.o
> +obj-y += pmap.init.o
>  obj-y += processor.o
>  obj-y += psci.o
>  obj-y += setup.o
> diff --git a/xen/arch/arm/mm.c b/xen/arch/arm/mm.c
> index d090fdfd5994..5e713b599611 100644
> --- a/xen/arch/arm/mm.c
> +++ b/xen/arch/arm/mm.c
> @@ -288,12 +288,7 @@ void dump_hyp_walk(vaddr_t addr)
>      dump_pt_walk(ttbr, addr, HYP_PT_ROOT_LEVEL, 1);
>  }
>  
> -/*
> - * Standard entry type that we'll use to build Xen's own pagetables.
> - * We put the same permissions at every level, because they're ignored
> - * by the walker in non-leaf entries.
> - */
> -static inline lpae_t mfn_to_xen_entry(mfn_t mfn, unsigned attr)
> +lpae_t mfn_to_xen_entry(mfn_t mfn, unsigned int attr)
>  {
>      lpae_t e = (lpae_t) {
>          .pt = {
> diff --git a/xen/arch/arm/pmap.c b/xen/arch/arm/pmap.c
> new file mode 100644
> index 000000000000..702b1bde982d
> --- /dev/null
> +++ b/xen/arch/arm/pmap.c
> @@ -0,0 +1,101 @@
> +#include <xen/init.h>
> +#include <xen/mm.h>
> +
> +#include <asm/bitops.h>
> +#include <asm/flushtlb.h>
> +#include <asm/pmap.h>
> +
> +/*
> + * To be able to use FIXMAP_PMAP_BEGIN.
> + * XXX: move fixmap definition in a separate header
> + */

yes please :-)


> +#include <xen/acpi.h>
> +
> +/*
> + * Simple mapping infrastructure to map / unmap pages in fixed map.
> + * This is used to set up the page table for mapcache, which is used
> + * by map domain page infrastructure.
> + *
> + * This structure is not protected by any locks, so it must not be used after
> + * smp bring-up.
> + */
> +
> +/* Bitmap to track which slot is used */
> +static unsigned long __initdata inuse;
> +
> +/* XXX: Find an header to declare it */

yep


> +extern lpae_t xen_fixmap[LPAE_ENTRIES];
> +
> +void *__init pmap_map(mfn_t mfn)
> +{
> +    unsigned long flags;
> +    unsigned int idx;
> +    vaddr_t linear;
> +    unsigned int slot;
> +    lpae_t *entry, pte;
> +
> +    BUILD_BUG_ON(sizeof(inuse) * BITS_PER_LONG < NUM_FIX_PMAP);
> +
> +    ASSERT(system_state < SYS_STATE_smp_boot);

One small concern here is that we have been using SYS_STATE_early_boot
to switch implementation of things like xen_map_table. Between
SYS_STATE_early_boot and SYS_STATE_smp_boot there is SYS_STATE_boot.

I guess I am wondering if instead of three potentially different mapping
functions (<= SYS_STATE_early_boot, < SYS_STATE_smp_boot, >=
SYS_STATE_smp_boot) we can get away with only two?



> +    local_irq_save(flags);
> +
> +    idx = find_first_zero_bit(&inuse, NUM_FIX_PMAP);
> +    if ( idx == NUM_FIX_PMAP )
> +        panic("Out of PMAP slots\n");
> +
> +    __set_bit(idx, &inuse);
> +
> +    slot = idx + FIXMAP_PMAP_BEGIN;
> +    ASSERT(slot >= FIXMAP_PMAP_BEGIN && slot <= FIXMAP_PMAP_END);
> +
> +    linear = FIXMAP_ADDR(slot);
> +    /*
> +     * We cannot use set_fixmap() here. We use PMAP when there is no direct map,
> +     * so map_pages_to_xen() called by set_fixmap() needs to map pages on
> +     * demand, which then calls pmap() again, resulting in a loop. Modify the
> +     * PTEs directly instead. The same is true for pmap_unmap().
> +     */
> +    entry = &xen_fixmap[third_table_offset(linear)];
> +
> +    ASSERT(!lpae_is_valid(*entry));
> +
> +    pte = mfn_to_xen_entry(mfn, PAGE_HYPERVISOR_RW);
> +    pte.pt.table = 1;
> +    write_pte(entry, pte);
> +
> +    local_irq_restore(flags);
> +
> +    return (void *)linear;
> +}
> +
> +void __init pmap_unmap(const void *p)
> +{
> +    unsigned long flags;
> +    unsigned int idx;
> +    lpae_t *entry;
> +    lpae_t pte = { 0 };
> +    unsigned int slot = third_table_offset((vaddr_t)p);
> +
> +    ASSERT(system_state < SYS_STATE_smp_boot);
> +    ASSERT(slot >= FIXMAP_PMAP_BEGIN && slot <= FIXMAP_PMAP_END);
> +
> +    idx = slot - FIXMAP_PMAP_BEGIN;
> +    local_irq_save(flags);
> +
> +    __clear_bit(idx, &inuse);
> +    entry = &xen_fixmap[third_table_offset((vaddr_t)p)];
> +    write_pte(entry, pte);
> +    flush_xen_tlb_range_va_local((vaddr_t)p, PAGE_SIZE);
> +
> +    local_irq_restore(flags);
> +}
> +
> +/*
> + * Local variables:
> + * mode: C
> + * c-file-style: "BSD"
> + * c-basic-offset: 4
> + * indent-tabs-mode: nil
> + * End:
> + */
> diff --git a/xen/include/asm-arm/config.h b/xen/include/asm-arm/config.h
> index 85d4a510ce8a..35050855b6e1 100644
> --- a/xen/include/asm-arm/config.h
> +++ b/xen/include/asm-arm/config.h
> @@ -180,6 +180,8 @@
>  #define FIXMAP_MISC     1  /* Ephemeral mappings of hardware */
>  #define FIXMAP_ACPI_BEGIN  2  /* Start mappings of ACPI tables */
>  #define FIXMAP_ACPI_END    (FIXMAP_ACPI_BEGIN + NUM_FIXMAP_ACPI_PAGES - 1)  /* End mappings of ACPI tables */
> +#define FIXMAP_PMAP_BEGIN (FIXMAP_ACPI_END + 1)
> +#define FIXMAP_PMAP_END (FIXMAP_PMAP_BEGIN + NUM_FIX_PMAP - 1)
>  
>  #define NR_hypercalls 64
>  
> diff --git a/xen/include/asm-arm/lpae.h b/xen/include/asm-arm/lpae.h
> index 310f5225e056..81fd482ab2ce 100644
> --- a/xen/include/asm-arm/lpae.h
> +++ b/xen/include/asm-arm/lpae.h
> @@ -4,6 +4,7 @@
>  #ifndef __ASSEMBLY__
>  
>  #include <xen/page-defs.h>
> +#include <xen/mm-frame.h>
>  
>  /*
>   * WARNING!  Unlike the x86 pagetable code, where l1 is the lowest level and
> @@ -168,6 +169,13 @@ static inline bool lpae_is_superpage(lpae_t pte, unsigned int level)
>          third_table_offset(addr)            \
>      }
>  
> +/*
> + * Standard entry type that we'll use to build Xen's own pagetables.
> + * We put the same permissions at every level, because they're ignored
> + * by the walker in non-leaf entries.
> + */
> +lpae_t mfn_to_xen_entry(mfn_t mfn, unsigned int attr);
> +
>  #endif /* __ASSEMBLY__ */
>  
>  /*
> diff --git a/xen/include/asm-arm/pmap.h b/xen/include/asm-arm/pmap.h
> new file mode 100644
> index 000000000000..8e1dce93f8e4
> --- /dev/null
> +++ b/xen/include/asm-arm/pmap.h
> @@ -0,0 +1,10 @@
> +#ifndef __ASM_PMAP_H__
> +#define __ARM_PMAP_H__

ASM/ARM


> +/* Large enough for mapping 5 levels of page tables with some headroom */
> +#define NUM_FIX_PMAP 8
> +
> +void *pmap_map(mfn_t mfn);
> +void pmap_unmap(const void *p);

> +#endif	/* __ASM_PMAP_H__ */

Julien Grall May 15, 2021, 8:54 a.m. UTC | #4

Hi Stefano,

On 15/05/2021 00:25, Stefano Stabellini wrote:
> On Sun, 25 Apr 2021, Julien Grall wrote:
>> From: Wei Liu <[e-mail]>
>> +extern lpae_t xen_fixmap[LPAE_ENTRIES];
>> +
>> +void *__init pmap_map(mfn_t mfn)
>> +{
>> +    unsigned long flags;
>> +    unsigned int idx;
>> +    vaddr_t linear;
>> +    unsigned int slot;
>> +    lpae_t *entry, pte;
>> +
>> +    BUILD_BUG_ON(sizeof(inuse) * BITS_PER_LONG < NUM_FIX_PMAP);
>> +
>> +    ASSERT(system_state < SYS_STATE_smp_boot);
> 
> One small concern here is that we have been using SYS_STATE_early_boot
> to switch implementation of things like xen_map_table. Between
> SYS_STATE_early_boot and SYS_STATE_smp_boot there is SYS_STATE_boot.
> 
> I guess I am wondering if instead of three potentially different mapping
> functions (<= SYS_STATE_early_boot, < SYS_STATE_smp_boot, >=
> SYS_STATE_smp_boot) we can get away with only two?

This is more flexible than the existing method to map memory when state 
== SYS_STATE_early_boot. If you look at the next patch (#13), you will 
see that there will be only two method to map memory.

[...]

>> diff --git a/xen/include/asm-arm/pmap.h b/xen/include/asm-arm/pmap.h
>> new file mode 100644
>> index 000000000000..8e1dce93f8e4
>> --- /dev/null
>> +++ b/xen/include/asm-arm/pmap.h
>> @@ -0,0 +1,10 @@
>> +#ifndef __ASM_PMAP_H__
>> +#define __ARM_PMAP_H__
> 
> ASM/ARM

I will fix it.

> 
> 
>> +/* Large enough for mapping 5 levels of page tables with some headroom */
>> +#define NUM_FIX_PMAP 8
>> +
>> +void *pmap_map(mfn_t mfn);
>> +void pmap_unmap(const void *p);
> 
>> +#endif	/* __ASM_PMAP_H__ */

Cheers,

Julien Grall July 3, 2021, 5:57 p.m. UTC | #5

Hi Hongyan,

On 26/04/2021 10:41, Xia, Hongyan wrote:
> On Sun, 2021-04-25 at 21:13 +0100, Julien Grall wrote:
>> From: Wei Liu <[e-mail]>
>>
>> The basic idea is like Persistent Kernel Map (PKMAP) in Linux. We
>> pre-populate all the relevant page tables before the system is fully
>> set up.
>>
>> We will need it on Arm in order to rework the arm64 version of
>> xenheap_setup_mappings() as we may need to use pages allocated from
>> the boot allocator before they are effectively mapped.
>>
>> This infrastructure is not lock-protected therefore can only be used
>> before smpboot. After smpboot, map_domain_page() has to be used.
>>
>> This is based on the x86 version [1] that was originally implemented
>> by Wei Liu.
>>
>> Take the opportunity to switch the parameter attr from unsigned to
>> unsigned int.
>>
>> [1] <
>> [e-mail]
>>>
>>
>> Signed-off-by: Wei Liu <[e-mail]>
>> Signed-off-by: Hongyan Xia <[e-mail]>
>> [julien: Adapted for Arm]
>> Signed-off-by: Julien Grall <[e-mail]>
> 
> [...]
> 
>> diff --git a/xen/arch/arm/pmap.c b/xen/arch/arm/pmap.c
>> new file mode 100644
>> index 000000000000..702b1bde982d
>> --- /dev/null
>> +++ b/xen/arch/arm/pmap.c
>> @@ -0,0 +1,101 @@
>> +#include <xen/init.h>
>> +#include <xen/mm.h>
>> +
>> +#include <asm/bitops.h>
>> +#include <asm/flushtlb.h>
>> +#include <asm/pmap.h>
>> +
>> +/*
>> + * To be able to use FIXMAP_PMAP_BEGIN.
>> + * XXX: move fixmap definition in a separate header
>> + */
>> +#include <xen/acpi.h>
>> +
>> +/*
>> + * Simple mapping infrastructure to map / unmap pages in fixed map.
>> + * This is used to set up the page table for mapcache, which is used
>> + * by map domain page infrastructure.
>> + *
>> + * This structure is not protected by any locks, so it must not be
>> used after
>> + * smp bring-up.
>> + */
>> +
>> +/* Bitmap to track which slot is used */
>> +static unsigned long __initdata inuse;
>> +
>> +/* XXX: Find an header to declare it */
>> +extern lpae_t xen_fixmap[LPAE_ENTRIES];
>> +
>> +void *__init pmap_map(mfn_t mfn)
>> +{
>> +    unsigned long flags;
>> +    unsigned int idx;
>> +    vaddr_t linear;
>> +    unsigned int slot;
>> +    lpae_t *entry, pte;
>> +
>> +    BUILD_BUG_ON(sizeof(inuse) * BITS_PER_LONG < NUM_FIX_PMAP);
> 
> This seems wrong to me. It should multiply with something like
> BITS_PER_BYTE.

Good spot! I have updated my tree.

> 
> I noticed this line was already present before the Arm version so
> probably my fault :(, which also needs to be fixed.

This should be taken care as the next version will create the pmap in 
common code :).

> 
>> +
>> +    ASSERT(system_state < SYS_STATE_smp_boot);
>> +
>> +    local_irq_save(flags);
>> +
>> +    idx = find_first_zero_bit(&inuse, NUM_FIX_PMAP);
>> +    if ( idx == NUM_FIX_PMAP )
>> +        panic("Out of PMAP slots\n");
>> +
>> +    __set_bit(idx, &inuse);
>> +
>> +    slot = idx + FIXMAP_PMAP_BEGIN;
>> +    ASSERT(slot >= FIXMAP_PMAP_BEGIN && slot <= FIXMAP_PMAP_END);
>> +
> 
>  From here...
> 
>> +    linear = FIXMAP_ADDR(slot);
>> +    /*
>> +     * We cannot use set_fixmap() here. We use PMAP when there is no
>> direct map,
>> +     * so map_pages_to_xen() called by set_fixmap() needs to map
>> pages on
>> +     * demand, which then calls pmap() again, resulting in a loop.
>> Modify the
>> +     * PTEs directly instead. The same is true for pmap_unmap().
>> +     */
>> +    entry = &xen_fixmap[third_table_offset(linear)];
>> +
>> +    ASSERT(!lpae_is_valid(*entry));
>> +
>> +    pte = mfn_to_xen_entry(mfn, PAGE_HYPERVISOR_RW);
>> +    pte.pt.table = 1;
>> +    write_pte(entry, pte);
>> +
> 
> ...to here, I wonder if we can move this chunk into arch (like void
> *arch_write_pmap_slot(slot)). Such an arch function hides how fixmap is
> handled and how page table entry is written behind arch, and the rest
> can just be common.

This is similar to what I had in mind. Let me give a try for the next 
version.

> 
>> +    local_irq_restore(flags);
>> +
>> +    return (void *)linear;
>> +}
>> +
>> +void __init pmap_unmap(const void *p)
>> +{
>> +    unsigned long flags;
>> +    unsigned int idx;
>> +    lpae_t *entry;
>> +    lpae_t pte = { 0 };
>> +    unsigned int slot = third_table_offset((vaddr_t)p);
>> +
>> +    ASSERT(system_state < SYS_STATE_smp_boot);
>> +    ASSERT(slot >= FIXMAP_PMAP_BEGIN && slot <= FIXMAP_PMAP_END);
>> +
>> +    idx = slot - FIXMAP_PMAP_BEGIN;
>> +    local_irq_save(flags);
>> +
>> +    __clear_bit(idx, &inuse);
>> +    entry = &xen_fixmap[third_table_offset((vaddr_t)p)];
>> +    write_pte(entry, pte);
>> +    flush_xen_tlb_range_va_local((vaddr_t)p, PAGE_SIZE);
> 
> and the same for the above, something like arch_clear_pmap(void *) and
> the rest into common.
> 
>  From a quick glance, I don't think x86 and Arm share any useful TLB
> flush helpers? So the TLB flush probably should be behind arch as well.

We could potential define flush_tlb_one_local() on Arm. But, I am not 
sure this is worth it because the page table manipulation is mainly 
happening in arch code so far. Although, this might change in the future.

> 
>> +
>> +    local_irq_restore(flags);
>> +}
>> +
>> +/*
>> + * Local variables:
>> + * mode: C
>> + * c-file-style: "BSD"
>> + * c-basic-offset: 4
>> + * indent-tabs-mode: nil
>> + * End:
>> + */
> 
> [...]
> 
>> diff --git a/xen/include/asm-arm/pmap.h b/xen/include/asm-arm/pmap.h
>> new file mode 100644
>> index 000000000000..8e1dce93f8e4
>> --- /dev/null
>> +++ b/xen/include/asm-arm/pmap.h
>> @@ -0,0 +1,10 @@
>> +#ifndef __ASM_PMAP_H__
>> +#define __ARM_PMAP_H__
> 
> This line doesn't seem to match the #ifndef, but if the functions are
> moved to common, this header can be moved to common as well.

Stefano pointed out the same. I have fixed it in my tree but I will 
likely rename the guard as the header will be moved in common.

Cheers,

Patch

diff --git a/xen/arch/arm/Makefile b/xen/arch/arm/Makefile
index ca75f1040dcc..2196ada24941 100644
--- a/xen/arch/arm/Makefile
+++ b/xen/arch/arm/Makefile
@@ -40,6 +40,7 @@  obj-y += percpu.o
 obj-y += platform.o
 obj-y += platform_hypercall.o
 obj-y += physdev.o
+obj-y += pmap.init.o
 obj-y += processor.o
 obj-y += psci.o
 obj-y += setup.o
diff --git a/xen/arch/arm/mm.c b/xen/arch/arm/mm.c
index d090fdfd5994..5e713b599611 100644
--- a/xen/arch/arm/mm.c
+++ b/xen/arch/arm/mm.c
@@ -288,12 +288,7 @@  void dump_hyp_walk(vaddr_t addr)
     dump_pt_walk(ttbr, addr, HYP_PT_ROOT_LEVEL, 1);
 }
 
-/*
- * Standard entry type that we'll use to build Xen's own pagetables.
- * We put the same permissions at every level, because they're ignored
- * by the walker in non-leaf entries.
- */
-static inline lpae_t mfn_to_xen_entry(mfn_t mfn, unsigned attr)
+lpae_t mfn_to_xen_entry(mfn_t mfn, unsigned int attr)
 {
     lpae_t e = (lpae_t) {
         .pt = {
diff --git a/xen/arch/arm/pmap.c b/xen/arch/arm/pmap.c
new file mode 100644
index 000000000000..702b1bde982d
--- /dev/null
+++ b/xen/arch/arm/pmap.c
@@ -0,0 +1,101 @@ 
+#include <xen/init.h>
+#include <xen/mm.h>
+
+#include <asm/bitops.h>
+#include <asm/flushtlb.h>
+#include <asm/pmap.h>
+
+/*
+ * To be able to use FIXMAP_PMAP_BEGIN.
+ * XXX: move fixmap definition in a separate header
+ */
+#include <xen/acpi.h>
+
+/*
+ * Simple mapping infrastructure to map / unmap pages in fixed map.
+ * This is used to set up the page table for mapcache, which is used
+ * by map domain page infrastructure.
+ *
+ * This structure is not protected by any locks, so it must not be used after
+ * smp bring-up.
+ */
+
+/* Bitmap to track which slot is used */
+static unsigned long __initdata inuse;
+
+/* XXX: Find an header to declare it */
+extern lpae_t xen_fixmap[LPAE_ENTRIES];
+
+void *__init pmap_map(mfn_t mfn)
+{
+    unsigned long flags;
+    unsigned int idx;
+    vaddr_t linear;
+    unsigned int slot;
+    lpae_t *entry, pte;
+
+    BUILD_BUG_ON(sizeof(inuse) * BITS_PER_LONG < NUM_FIX_PMAP);
+
+    ASSERT(system_state < SYS_STATE_smp_boot);
+
+    local_irq_save(flags);
+
+    idx = find_first_zero_bit(&inuse, NUM_FIX_PMAP);
+    if ( idx == NUM_FIX_PMAP )
+        panic("Out of PMAP slots\n");
+
+    __set_bit(idx, &inuse);
+
+    slot = idx + FIXMAP_PMAP_BEGIN;
+    ASSERT(slot >= FIXMAP_PMAP_BEGIN && slot <= FIXMAP_PMAP_END);
+
+    linear = FIXMAP_ADDR(slot);
+    /*
+     * We cannot use set_fixmap() here. We use PMAP when there is no direct map,
+     * so map_pages_to_xen() called by set_fixmap() needs to map pages on
+     * demand, which then calls pmap() again, resulting in a loop. Modify the
+     * PTEs directly instead. The same is true for pmap_unmap().
+     */
+    entry = &xen_fixmap[third_table_offset(linear)];
+
+    ASSERT(!lpae_is_valid(*entry));
+
+    pte = mfn_to_xen_entry(mfn, PAGE_HYPERVISOR_RW);
+    pte.pt.table = 1;
+    write_pte(entry, pte);
+
+    local_irq_restore(flags);
+
+    return (void *)linear;
+}
+
+void __init pmap_unmap(const void *p)
+{
+    unsigned long flags;
+    unsigned int idx;
+    lpae_t *entry;
+    lpae_t pte = { 0 };
+    unsigned int slot = third_table_offset((vaddr_t)p);
+
+    ASSERT(system_state < SYS_STATE_smp_boot);
+    ASSERT(slot >= FIXMAP_PMAP_BEGIN && slot <= FIXMAP_PMAP_END);
+
+    idx = slot - FIXMAP_PMAP_BEGIN;
+    local_irq_save(flags);
+
+    __clear_bit(idx, &inuse);
+    entry = &xen_fixmap[third_table_offset((vaddr_t)p)];
+    write_pte(entry, pte);
+    flush_xen_tlb_range_va_local((vaddr_t)p, PAGE_SIZE);
+
+    local_irq_restore(flags);
+}
+
+/*
+ * Local variables:
+ * mode: C
+ * c-file-style: "BSD"
+ * c-basic-offset: 4
+ * indent-tabs-mode: nil
+ * End:
+ */
diff --git a/xen/include/asm-arm/config.h b/xen/include/asm-arm/config.h
index 85d4a510ce8a..35050855b6e1 100644
--- a/xen/include/asm-arm/config.h
+++ b/xen/include/asm-arm/config.h
@@ -180,6 +180,8 @@ 
 #define FIXMAP_MISC     1  /* Ephemeral mappings of hardware */
 #define FIXMAP_ACPI_BEGIN  2  /* Start mappings of ACPI tables */
 #define FIXMAP_ACPI_END    (FIXMAP_ACPI_BEGIN + NUM_FIXMAP_ACPI_PAGES - 1)  /* End mappings of ACPI tables */
+#define FIXMAP_PMAP_BEGIN (FIXMAP_ACPI_END + 1)
+#define FIXMAP_PMAP_END (FIXMAP_PMAP_BEGIN + NUM_FIX_PMAP - 1)
 
 #define NR_hypercalls 64
 
diff --git a/xen/include/asm-arm/lpae.h b/xen/include/asm-arm/lpae.h
index 310f5225e056..81fd482ab2ce 100644
--- a/xen/include/asm-arm/lpae.h
+++ b/xen/include/asm-arm/lpae.h
@@ -4,6 +4,7 @@ 
 #ifndef __ASSEMBLY__
 
 #include <xen/page-defs.h>
+#include <xen/mm-frame.h>
 
 /*
  * WARNING!  Unlike the x86 pagetable code, where l1 is the lowest level and
@@ -168,6 +169,13 @@  static inline bool lpae_is_superpage(lpae_t pte, unsigned int level)
         third_table_offset(addr)            \
     }
 
+/*
+ * Standard entry type that we'll use to build Xen's own pagetables.
+ * We put the same permissions at every level, because they're ignored
+ * by the walker in non-leaf entries.
+ */
+lpae_t mfn_to_xen_entry(mfn_t mfn, unsigned int attr);
+
 #endif /* __ASSEMBLY__ */
 
 /*
diff --git a/xen/include/asm-arm/pmap.h b/xen/include/asm-arm/pmap.h
new file mode 100644
index 000000000000..8e1dce93f8e4
--- /dev/null
+++ b/xen/include/asm-arm/pmap.h
@@ -0,0 +1,10 @@ 
+#ifndef __ASM_PMAP_H__
+#define __ARM_PMAP_H__
+
+/* Large enough for mapping 5 levels of page tables with some headroom */
+#define NUM_FIX_PMAP 8
+
+void *pmap_map(mfn_t mfn);
+void pmap_unmap(const void *p);
+
+#endif	/* __ASM_PMAP_H__ */