Re: [PATCH] net: bonding: Fix transmit load balancing in balance-alb mode if specified by sysfs

[Nikolay Aleksandrov <nikolay@cumulusnetworks.com>]

On 08/09/17 05:06, Kosuke Tatsukawa wrote:
> Hi,
> 
>> On  7.09.2017 01:47, Kosuke Tatsukawa wrote:
>>> Commit cbf5ecb30560 ("net: bonding: Fix transmit load balancing in
>>> balance-alb mode") tried to fix transmit dynamic load balancing in
>>> balance-alb mode, which wasn't working after commit 8b426dc54cf4
>>> ("bonding: remove hardcoded value").
>>>
>>> It turned out that my previous patch only fixed the case when
>>> balance-alb was specified as bonding module parameter, and not when
>>> balance-alb mode was set using /sys/class/net/*/bonding/mode (the most
>>> common usage).  In the latter case, tlb_dynamic_lb was set up according
>>> to the default mode of the bonding interface, which happens to be
>>> balance-rr.
>>>
>>> This additional patch addresses this issue by setting up tlb_dynamic_lb
>>> to 1 if "mode" is set to balance-alb through the sysfs interface.
>>>
>>> I didn't add code to change tlb_balance_lb back to the default value for
>>> other modes, because "mode" is usually set up only once during
>>> initialization, and it's not worthwhile to change the static variable
>>> bonding_defaults in bond_main.c to a global variable just for this
>>> purpose.
>>>
>>> Commit 8b426dc54cf4 also changes the value of tlb_dynamic_lb for
>>> balance-tlb mode if it is set up using the sysfs interface.  I didn't
>>> change that behavior, because the value of tlb_balance_lb can be changed
>>> using the sysfs interface for balance-tlb, and I didn't like changing
>>> the default value back and forth for balance-tlb.
>>>
>>> As for balance-alb, /sys/class/net/*/bonding/tlb_balance_lb cannot be
>>> written to.  However, I think balance-alb with tlb_dynamic_lb set to 0
>>> is not an intended usage, so there is little use making it writable at
>>> this moment.
>>>
>>> Fixes: 8b426dc54cf4 ("bonding: remove hardcoded value")
>>> Reported-by: Reinis Rozitis <r@roze.lv>
>>> Signed-off-by: Kosuke Tatsukawa <tatsu@ab.jp.nec.com>
>>> Cc: stable@vger.kernel.org  # v4.12+
>>> ---
>>>  drivers/net/bonding/bond_options.c |    3 +++
>>>  1 files changed, 3 insertions(+), 0 deletions(-)
>>>
>>
>> I don't believe this to be the right solution, hardcoding it like this
>> changes user-visible behaviour. The issue is that if someone configured
>> it to be 0 in tlb mode, suddenly it will become 1 and will silently
>> override their config if they switch the mode to alb. Also it robs users
>> from their choice.
>>
>> If you think this should be settable in ALB mode, then IMO you should
>> edit tlb_dynamic_lb option's unsuppmodes and allow it to be set in ALB.
>> That would also be consistent with how it's handled in TLB mode.
> 
> No, I don't think tlb_dynamic_lb should be settable in balance-alb at
> this point.  All the current commits regarding tlb_dynamic_lb are for
> balance-tlb mode, so I don't think balance-alb with tlb_dynamic_lb set
> to 0 is an intended usage.
> 
> 
>> Going back and looking at your previous fix I'd argue that it is also
>> wrong, you should've removed the mode check altogether to return the
>> original behaviour where the dynamic_lb is set to 1 (enabled) by
>> default and then ALB mode would've had it, of course that would've left
>> the case of setting it to 0 in TLB mode and switching to ALB, but that
>> is a different issue.
> 
> Maybe balance-alb shouldn't be dependent on tlb_dynamic_lb.
> tlb_dynamic_lb is referenced in the following functions.
> 
>  + bond_do_alb_xmit()  -- Used by both balance-tlb and balance-alb
>  + bond_tlb_xmit()  -- Only used by balance-tlb
>  + bond_open()  -- Used by both balance-tlb and balance-alb
>  + bond_check_params()  -- Used during module initialization
>  + bond_fill_info()  -- Used to get/set value
>  + bond_option_tlb_dynamic_lb_set()  -- Used to get/set value
>  + bonding_show_tlb_dynamic_lb()  -- Used to get/set value
>  + bond_is_nondyn_tlb()  -- Only referenced if balance-tlb mode
> 
> The following untested patch adds code to make balance-alb work as if
> tlb_dynamic_lb==1 for the functions which affect balance-alb mode.  It
> also reverts my previous patch.
> 
> What do you think about this approach?
> ---
> Kosuke TATSUKAWA  | 1st Platform Software Division
>                   | NEC Solution Innovators
>                   | tatsu@ab.jp.nec.com
> 

Logically the approach looks good, that being said it adds unnecessary tests in
the fast path, why not just something like the patch below ? That only leaves the
problem if it is zeroed in TLB and switched to ALB mode, and that is a one line
fix to unsuppmodes just allow it to be set for that specific case. The below
returns the default behaviour before the commit in your Fixes tag.


diff --git a/drivers/net/bonding/bond_main.c b/drivers/net/bonding/bond_main.c
index fc63992ab0e0..c99dc59d729b 100644
--- a/drivers/net/bonding/bond_main.c
+++ b/drivers/net/bonding/bond_main.c
@@ -4289,7 +4289,7 @@ static int bond_check_params(struct bond_params *params)
 	int bond_mode	= BOND_MODE_ROUNDROBIN;
 	int xmit_hashtype = BOND_XMIT_POLICY_LAYER2;
 	int lacp_fast = 0;
-	int tlb_dynamic_lb = 0;
+	int tlb_dynamic_lb;
 
 	/* Convert string parameters. */
 	if (mode) {
@@ -4601,16 +4601,13 @@ static int bond_check_params(struct bond_params *params)
 	}
 	ad_user_port_key = valptr->value;
 
-	if ((bond_mode == BOND_MODE_TLB) || (bond_mode == BOND_MODE_ALB)) {
-		bond_opt_initstr(&newval, "default");
-		valptr = bond_opt_parse(bond_opt_get(BOND_OPT_TLB_DYNAMIC_LB),
-					&newval);
-		if (!valptr) {
-			pr_err("Error: No tlb_dynamic_lb default value");
-			return -EINVAL;
-		}
-		tlb_dynamic_lb = valptr->value;
+	bond_opt_initstr(&newval, "default");
+	valptr = bond_opt_parse(bond_opt_get(BOND_OPT_TLB_DYNAMIC_LB), &newval);
+	if (!valptr) {
+		pr_err("Error: No tlb_dynamic_lb default value");
+		return -EINVAL;
 	}
+	tlb_dynamic_lb = valptr->value;
 
 	if (lp_interval == 0) {
 		pr_warn("Warning: ip_interval must be between 1 and %d, so it was reset to %d\n",