On Mon, Dec 19, 2022 at 04:41:16PM +0900, Leesoo Ahn wrote: > > On 22. 12. 18. 17:55, Greg KH wrote: > > On Sun, Dec 18, 2022 at 01:18:51AM +0900, Leesoo Ahn wrote: > > > The current source pushes skb into dev->done queue by calling > > > skb_queue_tail() and then, call skb_dequeue() to pop for rx_cleanup state > > > to free urb and skb next in usbnet_bh(). > > > It wastes CPU resource with extra instructions. Instead, use return values > > > jumping to rx_cleanup case directly to free them. Therefore calling > > > skb_queue_tail() and skb_dequeue() is not necessary. > > > > > > The follows are just showing difference between calling skb_queue_tail() > > > and using return values jumping to rx_cleanup state directly in usbnet_bh() > > > in Arm64 instructions with perf tool. > > > > > > ----------- calling skb_queue_tail() ----------- > > > │ if (!(dev->driver_info->flags & FLAG_RX_ASSEMBLE)) > > > 7.58 │248: ldr x0, [x20, #16] > > > 2.46 │24c: ldr w0, [x0, #8] > > > 1.64 │250: ↑ tbnz w0, #14, 16c > > > │ dev->net->stats.rx_errors++; > > > 0.57 │254: ldr x1, [x20, #184] > > > 1.64 │258: ldr x0, [x1, #336] > > > 2.65 │25c: add x0, x0, #0x1 > > > │260: str x0, [x1, #336] > > > │ skb_queue_tail(&dev->done, skb); > > > 0.38 │264: mov x1, x19 > > > │268: mov x0, x21 > > > 2.27 │26c: → bl skb_queue_tail > > > 0.57 │270: ↑ b 44 // branch to call skb_dequeue() > > > > > > ----------- jumping to rx_cleanup state ----------- > > > │ if (!(dev->driver_info->flags & FLAG_RX_ASSEMBLE)) > > > 1.69 │25c: ldr x0, [x21, #16] > > > 4.78 │260: ldr w0, [x0, #8] > > > 3.28 │264: ↑ tbnz w0, #14, e4 // jump to 'rx_cleanup' state > > > │ dev->net->stats.rx_errors++; > > > 0.09 │268: ldr x1, [x21, #184] > > > 2.72 │26c: ldr x0, [x1, #336] > > > 3.37 │270: add x0, x0, #0x1 > > > 0.09 │274: str x0, [x1, #336] > > > 0.66 │278: ↑ b e4 // branch to 'rx_cleanup' state > > Interesting, but does this even really matter given the slow speed of > > the USB hardware? > > It doesn't if USB hardware has slow speed but in software view, it's still > worth avoiding calling skb_queue_tail() and skb_dequeue() which work with > spinlock, if possible. But can you actually measure that in either CPU load or in increased transfer speeds? thanks, greg k-h
