On Fri, Sep 23, 2022 at 11:51 AM Anirudh Venkataramanan <anirudh.venkataramanan@xxxxxxxxx> wrote: > > On 9/23/2022 8:31 AM, Alexander Duyck wrote: > > On Thu, Sep 22, 2022 at 3:38 PM Anirudh Venkataramanan > > <anirudh.venkataramanan@xxxxxxxxx> wrote: > >> > >> On 9/22/2022 1:58 PM, Alexander Duyck wrote: > >>> On Thu, Sep 22, 2022 at 1:07 PM Anirudh Venkataramanan > >>> <anirudh.venkataramanan@xxxxxxxxx> wrote: > >>>> > >>>> > >>>> Following Fabio's patches, I made similar changes for e1000/e1000e and > >>>> submitted them to IWL [1]. > >>>> > >>>> Yesterday, Ira Weiny pointed me to some feedback from Dave Hansen on the > >>>> use of page_address() [2]. My understanding of this feedback is that > >>>> it's safer to use kmap_local_page() instead of page_address(), because > >>>> you don't always know how the underlying page was allocated. > >>>> > >>>> This approach (of using kmap_local_page() instead of page_address()) > >>>> makes sense to me. Any reason not to go this way? > >>>> > >>>> [1] > >>>> > >>>> https://patchwork.ozlabs.org/project/intel-wired-lan/patch/20220919180949.388785-1-anirudh.venkataramanan@xxxxxxxxx/ > >>>> > >>>> https://patchwork.ozlabs.org/project/intel-wired-lan/patch/20220919180949.388785-2-anirudh.venkataramanan@xxxxxxxxx/ > >>>> > >>>> [2] > >>>> https://lore.kernel.org/lkml/5d667258-b58b-3d28-3609-e7914c99b31b@xxxxxxxxx/ > >>>> > >>>> Ani > >>> > >>> For the two patches you referenced the driver is the one allocating > >>> the pages. So in such a case the page_address should be acceptable. > >>> Specifically we are falling into alloc_page(GFP_ATOMIC) which should > >>> fall into the first case that Dave Hansen called out. > >> > >> Right. However, I did run into a case in the chelsio inline crypto > >> driver where it seems like the pages are allocated outside the driver. > >> In such cases, kmap_local_page() would be the right approach, as the > >> driver can't make assumptions on how the page was allocated. > > > > Right, but that is comparing apples and oranges. As I said for Tx it > > would make sense, but since we are doing the allocations for Rx that > > isn't the case so we don't need it. > > > >> ... and this makes me wonder why not just use kmap_local_page() even in > >> cases where the page allocation was done in the driver. IMO, this is > >> simpler because > >> > >> a) you don't have to care how a page was allocated. kmap_local_page() > >> will create a temporary mapping if required, if not it just becomes a > >> wrapper to page_address(). > >> > >> b) should a future patch change the allocation to be from highmem, you > >> don't have to change a bunch of page_address() calls to be > >> kmap_local_page(). > >> > >> Is using page_address() directly beneficial in some way? > > > > By that argument why don't we just leave the code alone and keep using > > kmap? I am pretty certain that is the logic that had us using kmap in > > the first place since it also dumps us with page_address in most cases > > and we didn't care much about the other architectures. > > Well, my understanding is that kmap_local_page() doesn't have the > overheads kmap() has, and that alone is reason enough to replace kmap() > and kmap_atomic() with kmap_local_page() where possible. It has less overhead, but there is still some pretty significant code involved. Basically in the cases where it can't bail out and just call page_address it will call __kmap_local_page_prot(), https://elixir.bootlin.com/linux/v6.0-rc4/source/mm/highmem.c#L517. > > If you look at > > the kmap_local_page() it just adds an extra step or two to calling > > page_address(). In this case it is adding extra complication to > > something that isn't needed which is the reason why we are going > > through this in the first place. If we are going to pull the bandage I > > suggest we might as well just go all the way and not take a half-step > > since we don't actually need kmap or its related calls for this. > > I don't really see this as "pulling the kmap() bandage", but a "use a > more appropriate kmap function if you can" type situation. My concern is that it is more of a half step in the case of the e1000/e1000e drivers. We likely should have fixed this some time ago when I had rewritten the Rx path for the igb and ixgbe drivers, but I just didn't get around to it because if I messed with other areas it would have required more validation. I'd rather not carry around the extra code or function calls if we don't need it. > FWIW, I am not against using page_address(). Just wanted to hash this > out and get to a conclusion before I made new changes. > > Ani I gathered as much based on your other conversation. This is essentially the module-local case you had referred to in which the page is allocated and used within the module so there is no need to be concerned about it possibly being a highmem page. Thanks, - Alex
