Re: [dpdk-dev] [PATCH v2] eal: add madvise to avoid dump memory

["Burakov, Anatoly" <anatoly.burakov@intel.com>]

On 23-Apr-20 4:43 PM, Li Feng wrote:
> Avoid dump all mapped memory to a core dump file when crash.
> Otherwise it will very large and it's hard to analyze with gdb.
> 
> In my test, it will dump 128GiB memory to a core dump file when integrated
> to spdk with default configuration.

Suggested rewording:

Currently, even though memory is mapped with PROT_NONE, this does not 
cause it to be excluded from core dumps. This is counter-productive, 
because in a lot of cases, this memory will go unused (e.g. when the 
memory subsystem preallocates VA space but hasn't yet mapped physical 
pages into it).

Use `madvise()` call with MADV_DONTDUMP parameter to exclude the 
unmapped memory from being dumped.

> 
> Signed-off-by: Li Feng <fengli@smartx.com>
> ---
>   lib/librte_eal/common/eal_common_memory.c | 14 ++++++++++++++
>   1 file changed, 14 insertions(+)
> 
> diff --git a/lib/librte_eal/common/eal_common_memory.c b/lib/librte_eal/common/eal_common_memory.c
> index cc7d54e0c..2d9564b28 100644
> --- a/lib/librte_eal/common/eal_common_memory.c
> +++ b/lib/librte_eal/common/eal_common_memory.c
> @@ -177,6 +177,20 @@ eal_get_virtual_area(void *requested_addr, size_t *size,
>   		after_len = RTE_PTR_DIFF(map_end, aligned_end);
>   		if (after_len > 0)
>   			munmap(aligned_end, after_len);
> +
> +		/*
> +		 * Exclude this pages from a core dump.
> +		 */
> +		if (madvise(aligned_addr, *size, MADV_DONTDUMP) != 0)
> +			RTE_LOG(WARNING, EAL, "Madvise with MADV_DONTDUMP failed: %s\n",
> +				strerror(errno)); > +	} else {
> +		/*
> +		 * Exclude this pages from a core dump.
> +		 */
> +		if (madvise(mapped_addr, map_sz, MADV_DONTDUMP) != 0)
> +			RTE_LOG(WARNING, EAL, "Madvise with MADV_DONTDUMP failed: %s\n",
> +				strerror(errno));
>   	}
>   
>   	return aligned_addr;
> 

For the contents of this patch,

Acked-by: Anatoly Burakov <anatoly.burakov@intel.com>

However, even though this is good to have, after some more thought, i 
believe the fix is incomplete, because this is not the only place we're 
reserving anonymous memory. We're also doing so in 
`eal_memalloc.c:free_seg()`, so an `madvise()` call should also be added 
there.

@David, now that i think of it, the PROT_NONE patch also was incomplete, 
as we only set PROT_NONE to memory that's initially reserved, but not 
when it's unmapped and returned back to the pool of anonymous memory. 
So, eal_memalloc.c should also remap anonymous memory with PROT_NONE.

@Li Feng, would you be so kind as to provide a patch replacing PROT_READ 
with PROT_NONE in eal_memalloc.c as well? Thank you very much!

-- 
Thanks,
Anatoly