Le 26/08/2024 à 11:15, Przemek Kitszel a écrit :
On 8/23/24 11:10, Dan Carpenter wrote:
On Fri, Aug 23, 2024 at 08:23:29AM +0200, Christophe JAILLET wrote:
In idpf_add_del_mac_filters(), filters are chunked up into multiple
messages to avoid sending a control queue message buffer that is too
large.
Each chunk has up to IDPF_NUM_FILTERS_PER_MSG entries. So except for the
last iteration which can be smaller, space for exactly
IDPF_NUM_FILTERS_PER_MSG entries is allocated.
There is no need to free and reallocate a smaller array just for the
last
iteration.
This slightly simplifies the code and avoid an (unlikely) memory
allocation
failure.
Thanks, that is indeed an improvement.
Signed-off-by: Christophe JAILLET <christophe.jaillet@xxxxxxxxxx>
---
drivers/net/ethernet/intel/idpf/idpf_virtchnl.c | 7 +++++--
1 file changed, 5 insertions(+), 2 deletions(-)
diff --git a/drivers/net/ethernet/intel/idpf/idpf_virtchnl.c
b/drivers/net/ethernet/intel/idpf/idpf_virtchnl.c
index 70986e12da28..b6f4b58e1094 100644
--- a/drivers/net/ethernet/intel/idpf/idpf_virtchnl.c
+++ b/drivers/net/ethernet/intel/idpf/idpf_virtchnl.c
@@ -3669,12 +3669,15 @@ int idpf_add_del_mac_filters(struct
idpf_vport *vport,
entries_size = sizeof(struct virtchnl2_mac_addr) *
num_entries;
buf_size = struct_size(ma_list, mac_addr_list, num_entries);
- if (!ma_list || num_entries != IDPF_NUM_FILTERS_PER_MSG) {
- kfree(ma_list);
+ if (!ma_list) {
ma_list = kzalloc(buf_size, GFP_ATOMIC);
if (!ma_list)
return -ENOMEM;
} else {
+ /* ma_list was allocated in the first iteration
+ * so IDPF_NUM_FILTERS_PER_MSG entries are
+ * available
+ */
memset(ma_list, 0, buf_size);
}
It would be even nicer to move the ma_list allocation outside the loop:
buf_size = struct_size(ma_list, mac_addr_list,
IDPF_NUM_FILTERS_PER_MSG);
ma_list = kmalloc(buf_size, GFP_ATOMIC);
good point
I've opened whole function for inspection and it asks for even more,
as of now, we allocate an array in atomic context, just to have a copy
of some stuff from the spinlock-protected list.
It would be good to have allocation as pointed by Dan prior to iteration
and fill it on the fly, sending when new message would not fit plus once
at the end. Sending procedure is safe to be called under a spinlock.
If I understand correctly, you propose to remove the initial copy in
mac_addr and hold &vport_config->mac_filter_list_lock till the end of
the function?
That's it?
There is a wait_for_completion_timeout() in idpf_vc_xn_exec() and the
default time-out is IDPF_VC_XN_DEFAULT_TIMEOUT_MSEC (60 * 1000)
So, should an issue occurs, and the time out run till the end, we could
hold the 'mac_filter_list_lock' spinlock for up to 60 seconds?
Is that ok?
And if in asynch update mode, idpf_mac_filter_async_handler() also takes
&vport_config->mac_filter_list_lock;. Could we dead-lock?
So, I'm not sure to understand what you propose, or the code in
idpf_add_del_mac_filters() and co.
CCing author; CCing Olek to ask if there are already some refactors that
would conflict with this.
I'll way a few days for these feedbacks and send a v2.
CJ
regards,
dan carpenter
