On 2023/3/3 19:22, Alexander Lobakin wrote: > From: Yunsheng Lin <linyunsheng@xxxxxxxxxx> > Date: Thu, 2 Mar 2023 10:30:13 +0800 > >> On 2023/3/2 0:03, Alexander Lobakin wrote: >>> __xdp_build_skb_from_frame() state(d): >>> >>> /* Until page_pool get SKB return path, release DMA here */ >>> >>> Page Pool got skb pages recycling in April 2021, but missed this >>> function. > > [...] > >> We both rely on both skb->pp_recycle and page->pp_magic to decide >> the page is really from page pool. So there was a few corner case >> problem when we are sharing a page for different skb in the driver >> level or calling skb_clone() or skb_try_coalesce(). >> see: >> https://github.com/torvalds/linux/commit/2cc3aeb5ecccec0d266813172fcd82b4b5fa5803 >> https://lore.kernel.org/netdev/MW5PR15MB51214C0513DB08A3607FBC1FBDE19@xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx/t/ >> https://lore.kernel.org/netdev/167475990764.1934330.11960904198087757911.stgit@localhost.localdomain/ > > And they are fixed :D > No drivers currently which use Page Pool mix PP pages with non-PP. And The wireless adapter which use Page Pool *does* mix PP pages with non-PP, see below discussion: https://lore.kernel.org/netdev/156f3e120bd0757133cb6bc11b76889637b5e0a6.camel@xxxxxxxxx/ > it's impossible to trigger try_coalesce() or so at least on cpumap path > since we're only creating skbs at that moment, they don't come from > anywhere else. > >> >> As the 'struct xdp_frame' also use 'struct skb_shared_info' which is >> sharable, see xdp_get_shared_info_from_frame(). >> >> For now xdpf_clone() does not seems to handling frag page yet, >> so it should be fine for now. > > xdpf_clone() clones a frame to a new full page and doesn't copy its > skb_shared_info. > >> >> IMHO we should find a way to use per-page marker, instead of both >> per-skb and per-page markers, in order to avoid the above problem >> for xdp if xdp has a similar processing as skb, as suggested by Eric. >> >> https://lore.kernel.org/netdev/CANn89iKgZU4Q+THXupzZi4hETuKuCOvOB=iHpp5JzQTNv_Fg_A@xxxxxxxxxxxxxx/ > > As Jesper already pointed out, not having a quick way to check whether > we have to check ::pp_magic at all can decrease performance. So it's > rather a shortcut. When we are freeing a page by updating the _refcount, I think we are already touching the cache of ::pp_magic. Anyway, I am not sure checking ::pp_magic is correct when a page will be passing between different subsystem and back to the network stack eventually, checking ::pp_magic may not be correct if this happens. Another way is to use the bottom two bits in bv_page, see: https://www.spinics.net/lists/netdev/msg874099.html > >> >>> >>> /* Allow SKB to reuse area used by xdp_frame */ >>> xdp_scrub_frame(xdpf); >>> > > Thanks, > Olek > . >
