Re: [dpdk-dev] [PATCH] mem: unmap unneeded space

[Bruce Richardson <bruce.richardson@intel.com>]

On Thu, Apr 19, 2018 at 05:35:25PM +0100, Anatoly Burakov wrote:
> When we ask to reserve virtual areas, we usually include
> alignment in the mapping size, and that memory ends up
> being wasted. Wasting a gigabyte of VA space while trying to
> reserve one gigabyte is pretty expensive on 32-bit, so after
> we're done mapping, unmap unneeded space.
> 
> Signed-off-by: Anatoly Burakov <anatoly.burakov@intel.com>
> ---
>  lib/librte_eal/common/eal_common_memory.c | 23 +++++++++++++++++++++--
>  1 file changed, 21 insertions(+), 2 deletions(-)
> 
> diff --git a/lib/librte_eal/common/eal_common_memory.c b/lib/librte_eal/common/eal_common_memory.c
> index 24a9ed5..8dd026a 100644
> --- a/lib/librte_eal/common/eal_common_memory.c
> +++ b/lib/librte_eal/common/eal_common_memory.c
> @@ -75,8 +75,13 @@ eal_get_virtual_area(void *requested_addr, size_t *size,
>  
>  	do {
>  		map_sz = no_align ? *size : *size + page_sz;
> +		if (map_sz > SIZE_MAX) {
> +			RTE_LOG(ERR, EAL, "Map size too big\n");
> +			rte_errno = E2BIG;
> +			return NULL;
> +		}
>  
> -		mapped_addr = mmap(requested_addr, map_sz, PROT_READ,
> +		mapped_addr = mmap(requested_addr, (size_t)map_sz, PROT_READ,
>  				mmap_flags, -1, 0);
>  		if (mapped_addr == MAP_FAILED && allow_shrink)
>  			*size -= page_sz;
> @@ -113,8 +118,22 @@ eal_get_virtual_area(void *requested_addr, size_t *size,
>  		RTE_LOG(WARNING, EAL, "   This may cause issues with mapping memory into secondary processes\n");
>  	}
>  
> -	if (unmap)
> +	if (unmap) {
>  		munmap(mapped_addr, map_sz);
> +	} else if (!no_align) {
> +		void *unmap_pre, *unmap_post, *map_end;
> +		size_t pre_len, post_len;
> +
> +		/* unmap all of the extra space */
> +		unmap_pre = mapped_addr;
> +		pre_len = RTE_PTR_DIFF(aligned_addr, mapped_addr);
> +		map_end = RTE_PTR_ADD(mapped_addr, (size_t)map_sz);
> +		unmap_post = RTE_PTR_ADD(aligned_addr, *size);
> +		post_len = RTE_PTR_DIFF(map_end, unmap_post);
> +
> +		munmap(unmap_pre, pre_len);
> +		munmap(unmap_post, post_len);

I suggest you take parts of the commit log and add them as a comment into
the code just after the else if.

Also, for readability, I think it would improve things to put the
munmap(unmap_pre, pre_len) immediately after the calculation of pre_len,
then a black line and then the rest of the code for clearing the post_len
block. I found it harder to decipher trying to keep track of all the
variables simultaneously in the two blocks. Splitting them into a
pre-blocka and a post-block should help here.

/Bruce