> On Nov 26, 2025, at 8:08 PM, Jason Wang <[email protected]> wrote: > > On Thu, Nov 27, 2025 at 3:48 AM Jon Kohler <[email protected]> wrote: >> >> >>> On Nov 26, 2025, at 5:25 AM, Arnd Bergmann <[email protected]> wrote: >>> >>> On Wed, Nov 26, 2025, at 07:04, Jason Wang wrote: >>>> On Wed, Nov 26, 2025 at 3:45 AM Jon Kohler <[email protected]> wrote: >>>>>> On Nov 19, 2025, at 8:57 PM, Jason Wang <[email protected]> wrote: >>>>>> On Tue, Nov 18, 2025 at 1:35 AM Jon Kohler <[email protected]> wrote: >>>>> Same deal goes for __put_user() vs put_user by way of commit >>>>> e3aa6243434f ("ARM: 8795/1: spectre-v1.1: use put_user() for >>>>> __put_user()”) >>>>> >>>>> Looking at arch/arm/mm/Kconfig, there are a variety of scenarios >>>>> where CONFIG_CPU_SPECTRE will be enabled automagically. Looking at >>>>> commit 252309adc81f ("ARM: Make CONFIG_CPU_V7 valid for 32bit ARMv8 >>>>> implementations") >>>>> it says that "ARMv8 is a superset of ARMv7", so I’d guess that just >>>>> about everything ARM would include this by default? >>> >>> I think the more relevant commit is for 64-bit Arm here, but this does >>> the same thing, see 84624087dd7e ("arm64: uaccess: Don't bother >>> eliding access_ok checks in __{get, put}_user"). >> >> Ah! Right, this is definitely the important bit, as it makes it >> crystal clear that these are exactly the same thing. The current >> code is: >> #define get_user __get_user >> #define put_user __put_user >> >> So, this patch changing from __* to regular versions is a no-op >> on arm side of the house, yea? >> >>> I would think that if we change the __get_user() to get_user() >>> in this driver, the same should be done for the >>> __copy_{from,to}_user(), which similarly skips the access_ok() >>> check but not the PAN/SMAP handling. >> >> Perhaps, thats a good call out. I’d file that under one battle >> at a time. Let’s get get/put user dusted first, then go down >> that road? >> >>> In general, the access_ok()/__get_user()/__copy_from_user() >>> pattern isn't really helpful any more, as Linus already >>> explained. I can't tell from the vhost driver code whether >>> we can just drop the access_ok() here and use the plain >>> get_user()/copy_from_user(), or if it makes sense to move >>> to the newer user_access_begin()/unsafe_get_user()/ >>> unsafe_copy_from_user()/user_access_end() and try optimize >>> out a few PAN/SMAP flips in the process. > > Right, according to my testing in the past, PAN/SMAP is a killer for > small packet performance (PPS).
For sure, every little bit helps along that path > >> >> In general, I think there are a few spots where we might be >> able to optimize (vhost_get_vq_desc perhaps?) as that gets >> called quite a bit and IIRC there are at least two flips >> in there that perhaps we could elide to one? An investigation >> for another day I think. > > Did you mean trying to read descriptors in a batch, that would be > better and with IN_ORDER it would be even faster as a single (at most > two) copy_from_user() might work (without the need to use > user_access_begin()/user_access_end(). Yep. I haven’t fully thought through it, just a drive-by idea from looking at code for the recent work I’ve been doing, just scratching my head thinking there *must* be something we can do better there. Basically on the get rx/tx bufs path as well as the vhost_add_used_and_signal_n path, I think we could cluster together some of the get/put users and copy to/from’s. Would take some massaging, but I think there is something there. >> >> Anyhow, with this info - Jason - is there anything else you >> can think of that we want to double click on? > > Nope. > > Thanks Ok thanks. Perhaps we can land this in next and let it soak in, though, knock on wood, I don’t think there will be fallout (famous last words!) ?
