eal: allow start two dpdk with no-huge option
Checks
Commit Message
From: Kignelman OUATTARA <thiery.ouattara@outscale.com>
in last version (v18.02), we was using no-huge option to
start 2 dpdk instances simultanusly (for testing purpose).
but since v18.05 when we start 2 instances:
- the first dpdk app start normaly
- the 2nd can't start because LOCK_EX option is set in
flock(fd, LOCK_EX | LOCK_NB).
So i did this patch to change LOCK_EX to LOCK_SH if no-huge
option is set.
Signed-off-by: Kignelman OUATTARA <thiery.ouattara@outscale.com>
---
lib/librte_eal/common/eal_common_fbarray.c | 6 +++++-
1 file changed, 5 insertions(+), 1 deletion(-)
Comments
On 10-Jul-18 3:54 PM, thiery.ouattara@outscale.com wrote:
> From: Kignelman OUATTARA <thiery.ouattara@outscale.com>
>
> in last version (v18.02), we was using no-huge option to
> start 2 dpdk instances simultanusly (for testing purpose).
>
> but since v18.05 when we start 2 instances:
> - the first dpdk app start normaly
> - the 2nd can't start because LOCK_EX option is set in
> flock(fd, LOCK_EX | LOCK_NB).
>
> So i did this patch to change LOCK_EX to LOCK_SH if no-huge
> option is set.
>
> Signed-off-by: Kignelman OUATTARA <thiery.ouattara@outscale.com>
> ---
Hi Kignelman,
I don't think this is safe to do. Even though hugepage memory is not
used, the --no-huge mode still stores page segments in fbarrays, so
while you would be able to *run* DPDK in such a scenario, the second
process would corrupt the memory of the first.
As a proper alternative, i would suggest looking at my --in-memory mode
patchset:
http://patches.dpdk.org/patch/40582/
http://patches.dpdk.org/patch/40583/
http://patches.dpdk.org/patch/40585/
http://patches.dpdk.org/patch/40584/
http://patches.dpdk.org/patch/40587/
http://patches.dpdk.org/patch/40586/
http://patches.dpdk.org/patch/40588/
http://patches.dpdk.org/patch/40590/
http://patches.dpdk.org/patch/40589/
This will solve the problem at its source.
2018-07-10 17:00 GMT+02:00 Burakov, Anatoly <anatoly.burakov@intel.com>:
> On 10-Jul-18 3:54 PM, thiery.ouattara@outscale.com wrote:
>
>> From: Kignelman OUATTARA <thiery.ouattara@outscale.com>
>>
>> in last version (v18.02), we was using no-huge option to
>> start 2 dpdk instances simultanusly (for testing purpose).
>>
>> but since v18.05 when we start 2 instances:
>> - the first dpdk app start normaly
>> - the 2nd can't start because LOCK_EX option is set in
>> flock(fd, LOCK_EX | LOCK_NB).
>>
>> So i did this patch to change LOCK_EX to LOCK_SH if no-huge
>> option is set.
>>
>> Signed-off-by: Kignelman OUATTARA <thiery.ouattara@outscale.com>
>> ---
>>
>
> Hi Kignelman,
>
> I don't think this is safe to do. Even though hugepage memory is not used,
> the --no-huge mode still stores page segments in fbarrays, so while you
> would be able to *run* DPDK in such a scenario, the second process would
> corrupt the memory of the first.
>
> As a proper alternative, i would suggest looking at my --in-memory mode
> patchset:
>
> http://patches.dpdk.org/patch/40582/
> http://patches.dpdk.org/patch/40583/
> http://patches.dpdk.org/patch/40585/
> http://patches.dpdk.org/patch/40584/
> http://patches.dpdk.org/patch/40587/
> http://patches.dpdk.org/patch/40586/
> http://patches.dpdk.org/patch/40588/
> http://patches.dpdk.org/patch/40590/
> http://patches.dpdk.org/patch/40589/
>
> This will solve the problem at its source.
>
> --
> Thanks,
> Anatoly
>
Hi Anatoly,
Thanks for your reply, i will test with your patches
Thanks,
Kignelman
@@ -414,6 +414,7 @@ struct used_mask {
struct used_mask *msk;
void *data = NULL;
int fd = -1;
+ int lock_opt = LOCK_EX;
if (arr == NULL) {
rte_errno = EINVAL;
@@ -436,6 +437,9 @@ struct used_mask {
eal_get_fbarray_path(path, sizeof(path), name);
+ if (internal_config.no_hugetlbfs)
+ lock_opt = LOCK_SH;
+
/*
* Each fbarray is unique to process namespace, i.e. the filename
* depends on process prefix. Try to take out a lock and see if we
@@ -447,7 +451,7 @@ struct used_mask {
path, strerror(errno));
rte_errno = errno;
goto fail;
- } else if (flock(fd, LOCK_EX | LOCK_NB)) {
+ } else if (flock(fd, lock_opt | LOCK_NB)) {
RTE_LOG(DEBUG, EAL, "%s(): couldn't lock %s: %s\n", __func__,
path, strerror(errno));
rte_errno = EBUSY;