Re: [dpdk-dev] [PATCH] eal: don't advertise a physical address when no hugepages

[santosh <santosh.shukla@caviumnetworks.com>]

Hi Olivier,

On Friday 23 June 2017 10:38 PM, Jan Blunck wrote:

> On Fri, Jun 23, 2017 at 10:11 AM, Olivier Matz <olivier.matz@6wind.com> wrote:
>> Hi Jan,
>>
>> On Sat, 10 Jun 2017 10:31:22 +0200, Jan Blunck <jblunck@infradead.org> wrote:
>>> On Fri, Jun 9, 2017 at 10:29 AM, Olivier Matz <olivier.matz@6wind.com> wrote:
>>>> When populating a mempool with a virtual memory area, the mempool
>>>> library expects to be able to get the physical address of each page.
>>>>
>>>> When started with --no-huge, the physical addresses may not be available
>>>> because the pages are not locked in memory. It sometimes returns
>>>> RTE_BAD_PHYS_ADDR, which makes the mempool_populate() function to fail.
>>>>
>>>> This was working before the commit cdc242f260e7 ("eal/linux: support
>>>> running as unprivileged user"), because rte_mem_virt2phy() was returning
>>>> 0 instead of RTE_BAD_PHYS_ADDR, which was seen as a valid physical
>>>> address.
>>>>
>>>> Since --no-huge is a debug function that breaks the support of physical
>>>> drivers, always set physical addresses to RTE_BAD_PHYS_ADDR in memzones
>>>> or in rte_mem_virt2phy(), and ensure that mempool won't complain in that
>>>> case.
>>>>
>>>> Fixes: cdc242f260e7 ("eal/linux: support running as unprivileged user")
>>>>
>>>> CC: stable@dpdk.org
>>>> Signed-off-by: Olivier Matz <olivier.matz@6wind.com>
>>>> ---
>>>>  lib/librte_eal/common/eal_common_memzone.c | 5 ++++-
>>>>  lib/librte_eal/linuxapp/eal/eal_memory.c   | 7 +++++++
>>>>  lib/librte_mempool/rte_mempool.c           | 2 +-
>>>>  3 files changed, 12 insertions(+), 2 deletions(-)
>>>>
>>>> diff --git a/lib/librte_eal/common/eal_common_memzone.c b/lib/librte_eal/common/eal_common_memzone.c
>>>> index 3026e36b8..c465c8fc2 100644
>>>> --- a/lib/librte_eal/common/eal_common_memzone.c
>>>> +++ b/lib/librte_eal/common/eal_common_memzone.c
>>>> @@ -251,7 +251,10 @@ memzone_reserve_aligned_thread_unsafe(const char *name, size_t len,
>>>>
>>>>         mcfg->memzone_cnt++;
>>>>         snprintf(mz->name, sizeof(mz->name), "%s", name);
>>>> -       mz->phys_addr = rte_malloc_virt2phy(mz_addr);
>>>> +       if (rte_eal_has_hugepages())
>>>> +               mz->phys_addr = rte_malloc_virt2phy(mz_addr);
>>>> +       else
>>>> +               mz->phys_addr = RTE_BAD_PHYS_ADDR;
>>> Since you set phys_addrs_available to false rte_malloc_virt2phy()
>>> anyway returns RTE_BAD_PHYS_ADDR so I believe the conditional isn't
>>> necessary here.
>>>
>>> Rest of the patch looks good to me.
>> The variable phys_addrs_available only impacts rte_mem_virt2phy().
>> Here, for memzones allocation, rte_malloc_virt2phy() is used, and
>> it gets its physical address by retrieving it from the memseg structure.
>>
>> With the full patch, "dump_memzone" displays something like:
>>   Zone 0: name:<rte_eth_dev_data>, phys:0xffffffffffffffff, len:0x30100, [...]
>>   ...
>>
>> If I strip the memzone part, it displays:
>>   Zone 0: name:<rte_eth_dev_data>, phys:0x7fe382c62640, len:0x30100, [...]
>>   ...
>>
>> So I think we should either keep the patch as is, or change the memseg
>> and malloc part like this (it's maybe better):
>>
>>   --- a/lib/librte_eal/common/rte_malloc.c
>>   +++ b/lib/librte_eal/common/rte_malloc.c
>>   @@ -254,5 +254,7 @@ rte_malloc_virt2phy(const void *addr)
>>           const struct malloc_elem *elem = malloc_elem_from_data(addr);
>>           if (elem == NULL)
>>                   return 0;
>>   +       if (elem->ms->phys_addr == RTE_BAD_PHYS_ADDR)
>>   +               return RTE_BAD_PHYS_ADDR;
>>           return elem->ms->phys_addr + ((uintptr_t)addr - (uintptr_t)elem->ms->addr);
>>    }
>>   diff --git a/lib/librte_eal/linuxapp/eal/eal_memory.c b/lib/librte_eal/linuxapp/eal/eal_memory.c
>>   index 1c99852..2a401ca 100644
>>   --- a/lib/librte_eal/linuxapp/eal/eal_memory.c
>>   +++ b/lib/librte_eal/linuxapp/eal/eal_memory.c
>>   @@ -973,7 +973,7 @@ rte_eal_hugepage_init(void)
>>                                           strerror(errno));
>>                           return -1;
>>                   }
>>   -               mcfg->memseg[0].phys_addr = (phys_addr_t)(uintptr_t)addr;
>>   +               mcfg->memseg[0].phys_addr = RTE_BAD_PHYS_ADDR;
>>                   mcfg->memseg[0].addr = addr;
>>                   mcfg->memseg[0].hugepage_sz = RTE_PGSIZE_4K;
>>                   mcfg->memseg[0].len = internal_config.memory;
>>
>>
>> Let me know what you are ok with this and I'll send a v2.
>>
> This approach looks better to me.
>
> hanks,
> Jan

Approach LGTM, though small comment:
I think we also need to fix error return description for API rte_malloc_virt2phy.
It says 'NULL' on error. It should be 0 or RTE_BAD_PHYS_ADDR.

In fact, we should remove '0' as error return and keep RTE_BAD_PHYS_ADDR as the error value.
If so then change may look like:

if (elem == NULL || elem->ms->phys_addr == RTE_BAD_PHYS_ADDR)
    return RTE_BAD_PHYS_ADDR;

Provided that return value '0' considered as error value in current code.

Having said that, few drivers using rte_malloc_virt2phy without an error check.
I guess now they must check return value before using phys_addr_t.