On 2021/4/1 21:08, Leon Romanovsky wrote:
On Thu, Apr 01, 2021 at 03:32:19PM +0800, Weihang Li wrote:
From: Lang Cheng <chenglang@xxxxxxxxxx>
Fix error of cmd's context number calculation algorithm to enable all of
32 cmd entries and support 32 concurrent accesses.
Signed-off-by: Lang Cheng <chenglang@xxxxxxxxxx>
Signed-off-by: Weihang Li <liweihang@xxxxxxxxxx>
---
drivers/infiniband/hw/hns/hns_roce_cmd.c | 62 ++++++++++++-----------------
drivers/infiniband/hw/hns/hns_roce_device.h | 6 +--
2 files changed, 27 insertions(+), 41 deletions(-)
<...>
- WARN_ON(cmd->free_head < 0);
- context = &cmd->context[cmd->free_head];
- context->token += cmd->token_mask + 1;
- cmd->free_head = context->next;
+
+ do {
+ context = &cmd->context[cmd->free_head];
+ cmd->free_head = context->next;
+ } while (context->busy);
+
+ context->busy = 1;
This "busy" flag after do-while together with release in __hns_roce_cmd_mbox_wait()
is interesting thing. Are you sure that it won't loop forever here?
When initializing resources in hns_roce_cmd_use_events(), ensure that
the number of semaphores is consistent with the depth of context[].
int hns_roce_cmd_use_events( )
{
hr_cmd->context = kcalloc(hr_cmd->max_cmds, ...);
sema_init(&hr_cmd->event_sem, hr_cmd->max_cmds);
}
Then, when someone gets the event_sem in hns_roce_cmd_mbox_wait(), it
means that there must be a not busy context.
Thanks.
Thanks
.
null
