Li RongQing wrote: > when changes the rx/tx ring to 4096, rx/tx_bi needs an order > 5 pages, and allocation maybe fail due to memory fragmentation > so switch to kvzalloc
Hi Li, Thanks for your patches, but I think you're solving a problem that isn't a problem (besides that the warning is being printed.) After all, the driver either gets the memory that it needed via the kernel waiting (since we didn't set GFP_NOWAIT), or ENOMEM gets returned to the user. If your kernel is so close to OOM that it's having this problem aren't you going to have other issues? Anyway.... This driver goes to great lengths to not make any changes to the existing queues if an allocation fails via ethtool (in fact this is code I authored). Maybe a better option is to just set __GFP_NOWARN on the kcalloc call? Then if there is a memory allocation error we'll just reflect it to user space and let the user retry, with no noisy kernel message. For performance reasons, having actually contiguous regions from kcalloc should help performance vs kvcalloc virtually contiguous regions. I'd prefer that you don't solve the problem this way. How about just marking the allocations as GFP_RETRY_MAYFAIL and GFP_NOWARN? The same goes for the iavf patch. Jesse
