Re: [PATCH] selftests/vm: Fix test for virtual address range mapping for arm64

[Michal SuchÃnek]

On Mon, 15 May 2017 09:31:49 +0530
Anshuman Khandual <khandual@xxxxxxxxxxxxxxxxxx> wrote:

> On 05/10/2017 12:30 AM, Michal Suchanek wrote:
> > Arm64 has 256TB address space so fix the test to pass on Arm as
> > well.
> > 
> > Also remove unneeded numaif include.
> > 
> > Signed-off-by: Michal Suchanek <msuchanek@xxxxxxx>
> > ---
> >  tools/testing/selftests/vm/virtual_address_range.c | 36
> > ++++++++++++++++------ 1 file changed, 27 insertions(+), 9
> > deletions(-)
> > 
> > diff --git a/tools/testing/selftests/vm/virtual_address_range.c
> > b/tools/testing/selftests/vm/virtual_address_range.c index
> > 3b02aa6..ff6628f 100644 ---
> > a/tools/testing/selftests/vm/virtual_address_range.c +++
> > b/tools/testing/selftests/vm/virtual_address_range.c @@ -10,7 +10,6
> > @@ #include <string.h>
> >  #include <unistd.h>
> >  #include <errno.h>
> > -#include <numaif.h>
> >  #include <sys/mman.h>
> >  #include <sys/time.h>
> > 
> > @@ -32,15 +31,34 @@
> >   * different areas one below 128TB and one above 128TB
> >   * till it reaches 512TB. One with size 128TB and the
> >   * other being 384TB.
> > + *
> > + * On Arm64 the address space is 256TB and no high mappings
> > + * are supported so far. Presumably support can be added in
> > + * the future.
> >   */
> > +
> >  #define NR_CHUNKS_128TB   8192UL /* Number of 16GB chunks for
> > 128TB */ -#define NR_CHUNKS_384TB  24576UL /* Number of 16GB chunks
> > for 384TB */ +#define NR_CHUNKS_256TB   (NR_CHUNKS_128TB * 2UL)
> > +#define NR_CHUNKS_384TB   (NR_CHUNKS_128TB * 3UL)
> > 
> >  #define ADDR_MARK_128TB  (1UL << 47) /* First address beyond 128TB
> > */ +#define ADDR_MARK_256TB  (1UL << 48) /* First address beyond
> > 256TB */ +
> > +#ifdef __aarch64__
> > +#define HIGH_ADDR_MARK  ADDR_MARK_256TB
> > +#define HIGH_ADDR_SHIFT 49
> > +#define NR_CHUNKS_LOW   NR_CHUNKS_256TB
> > +#define NR_CHUNKS_HIGH  NR_CHUNKS_256TB
> > +#else
> > +#define HIGH_ADDR_MARK  ADDR_MARK_128TB
> > +#define HIGH_ADDR_SHIFT 48
> > +#define NR_CHUNKS_LOW   NR_CHUNKS_128TB
> > +#define NR_CHUNKS_HIGH  NR_CHUNKS_384TB
> > +#endif
> > 
> >  static char *hind_addr(void)
> >  {
> > -	int bits = 48 + rand() % 15;
> > +	int bits = HIGH_ADDR_SHIFT + rand() % 15;  
> 
> The randomization is upto 63 bits. Hence if HIGH_ADDR_SHIFT is 49
> instead of 48 then it should be rand() % 14 in that case.

I was wondering if this needs fixing but did not come up with any way
how this magic number is related to number of bits in the address.
Thanks for pointing it out.

> 
> > 
> >  	return (char *) (1UL << bits);
> >  }
> > @@ -50,14 +68,14 @@ static int validate_addr(char *ptr, int
> > high_addr) unsigned long addr = (unsigned long) ptr;
> > 
> >  	if (high_addr) {
> > -		if (addr < ADDR_MARK_128TB) {
> > +		if (addr < HIGH_ADDR_MARK) {
> >  			printf("Bad address %lx\n", addr);
> >  			return 1;
> >  		}
> >  		return 0;
> >  	}
> > 
> > -	if (addr > ADDR_MARK_128TB) {
> > +	if (addr > HIGH_ADDR_MARK) {
> >  		printf("Bad address %lx\n", addr);
> >  		return 1;
> >  	}
> > @@ -79,12 +97,12 @@ static int validate_lower_address_hint(void)
> > 
> >  int main(int argc, char *argv[])
> >  {
> > -	char *ptr[NR_CHUNKS_128TB];
> > -	char *hptr[NR_CHUNKS_384TB];
> > +	char *ptr[NR_CHUNKS_LOW];
> > +	char *hptr[NR_CHUNKS_HIGH];
> >  	char *hint;
> >  	unsigned long i, lchunks, hchunks;
> > 
> > -	for (i = 0; i < NR_CHUNKS_128TB; i++) {
> > +	for (i = 0; i < NR_CHUNKS_LOW; i++) {
> >  		ptr[i] = mmap(NULL, MAP_CHUNK_SIZE, PROT_READ |
> > PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
> > 
> > @@ -99,7 +117,7 @@ int main(int argc, char *argv[])
> >  	}
> >  	lchunks = i;
> > 
> > -	for (i = 0; i < NR_CHUNKS_384TB; i++) {
> > +	for (i = 0; i < NR_CHUNKS_HIGH; i++) {  
> 
> If ARM64 does not have address space beyond 256TB, all map requests
> beyond 256TB (which is being attempted in this second for loop) will
> fail. 

Same as the second loop does on x86_64 with l5 paging not (fully)
implemented, yes.

> But does the arch support the hint based mechanism like powerpc
> and x86 to allocate beyond certain point ? 

It does not.

> The split in the alocation
> (represented by two for loops) is because of the fact that below
> HIGH_ADDR_MARK hint is not required and its required above it till
> the end of the total VA space. In this case,
> 
> > +#define HIGH_ADDR_MARK  ADDR_MARK_256TB
> > +#define HIGH_ADDR_SHIFT 49
> > +#define NR_CHUNKS_LOW   NR_CHUNKS_256TB
> > +#define NR_CHUNKS_HIGH  NR_CHUNKS_256TB  
> 
> both the for loops will be attempting below 256TB one with and one
> without the hint mechanism. But all the validations will fail
> for the second for loop (where hint will be passed beyond 256TB)
> as the address will not be allocated beyond 256TB where it is
> capped.
> 

AFAICT the allocation fails completely in the second loop on archs that
don't implement the 'highmem' or what you want to call it which
currently includes x86_64 making the test pass. So in this patch I move
the 'highmem' to range where nothing is allowed on arm64 and extending
the 'lowmem' to the actual allowed range so the test passes there as
well. The second loop is superfluous but harmless and overall this
gives a variable split test in case the split lands on different
address on some arch in the future.

Thanks

Michal