hmmm,
Sure it is applicable here, but while most function using this
strategy would be happy to have some part of the job done, this one
won't:
for instance: receiving N packets always makes sense, even if M>N
packet are waiting to be received.
In this case, getting half of an anwer does not really make sense...
But I see your point regarding allocation.
Didn't we say that libC was always supported (hence malloc).
There is no need for speed here...
Christophe
On 22 February 2017 at 23:38, Bill Fischofer <bill.fischo...@linaro.org> wrote:
>
>
> On Wed, Feb 22, 2017 at 6:55 AM, Christophe Milard
> <christophe.mil...@linaro.org> wrote:
>>
>> Implementation of the device query function for the linux-gen ODP.
>>
>> Signed-off-by: Christophe Milard <christophe.mil...@linaro.org>
>> ---
>> platform/linux-generic/drv_driver.c | 37
>> +++++++++++++++++++++++++++++++++++++
>> 1 file changed, 37 insertions(+)
>>
>> diff --git a/platform/linux-generic/drv_driver.c
>> b/platform/linux-generic/drv_driver.c
>> index 48a90a2..517a3c6 100644
>> --- a/platform/linux-generic/drv_driver.c
>> +++ b/platform/linux-generic/drv_driver.c
>> @@ -376,6 +376,43 @@ static void device_destroy_terminate(odpdrv_device_t
>> drv_device)
>> _odp_ishm_pool_free(list_elt_pool, device);
>> }
>>
>> +odpdrv_device_t *odpdrv_device_query(odpdrv_enumr_t enumr, const char
>> *address)
>> +{
>> + _odpdrv_device_t *dev;
>> + odpdrv_device_t *res;
>> + int index = 0;
>> +
>> + int size = sizeof(odpdrv_device_t); /* for the
>> ODPDRV_DEVICE_INVALID */
>> +
>> + /* parse the list of device a first time to determine the size of
>> + * the memory to be allocated:
>> + */
>> + dev_list_read_lock();
>> + dev = device_lst.head;
>> + while (dev) {
>> + if ((dev->param.enumerator == enumr) &&
>> + ((address == NULL) ||
>> + (strcmp(dev->param.address, address) == 0)))
>> + size += sizeof(odpdrv_device_t);
>> + dev = dev->next;
>> + }
>> +
>> + /* then fill the list: */
>> + res = (odpdrv_device_t *)malloc(size);
>> + dev = device_lst.head;
>> + while (dev) {
>> + if ((dev->param.enumerator == enumr) &&
>> + ((address == NULL) ||
>> + (strcmp(dev->param.address, address) == 0)))
>> + res[index++] = (odpdrv_device_t)dev;
>> + dev = dev->next;
>> + }
>> + dev_list_read_unlock();
>> + res[index++] = ODPDRV_DEVICE_INVALID;
>> +
>> + return res; /* must be freed by caller! */
>
>
> Most other ODP APIs that return a variable-number of return parameters have
> the caller supply a return array and size and the routine fills that in and
> returns the number of elements returned. Why is that model not suitable
> here? The concern about doing malloc() calls within ODP code is that
> constrains the caller as to the memory model being used. If the caller
> supplies the return array then it can obtain that memory from wherever it
> wishes.
>
>>
>> +}
>> +
>> odpdrv_devio_t odpdrv_devio_register(odpdrv_devio_param_t *param)
>> {
>> ODP_ERR("NOT Supported yet! Driver %s Registration!\n.",
>> --
>> 2.7.4
>>
>
