On 03/18/2019 02:37 PM, Waiman Long wrote: > On 03/17/2019 02:27 PM, Manfred Spraul wrote: >> Hi Waiman, >> >> On 2/28/19 7:47 PM, Waiman Long wrote: >>> For ipcmni_extend mode, the sequence number space is only 7 bits. So >>> the chance of id reuse is relatively high compared with the non-extended >>> mode. >>> >>> To alleviate this id reuse problem, the id allocation will be done >>> cyclically to cycle through all the 24-bit id space before wrapping >>> around when in ipcmni_extend mode. This may cause the use of more memory >>> in term of the number of xa_nodes allocated as well as potentially more >>> cachelines used as the xa_nodes may be spread more sparsely in this >>> case. >>> >>> There is probably a slight memory and performance cost in doing cyclic >>> id allocation. For applications that really need more than 32k unique >>> IPC >>> identifiers, this is a small price to pay to avoid the id reuse problem. >> Have you measured it? >> >> I have observed -3% for semop() for a 4 level radix tree compared to a >> 1-level radix tree, and I'm a bit reluctant to accept that. >> Especially as the percentage will increase if the syscall overhead >> goes down again (-> less spectre impact). >> > It is both Spectre (retpoline) and Meltdown (PTI). PTI is not needed in > AMD CPU and so you may see a bit higher slowdown. The use of idr_replace() in your previous patch may also slow the code path a bit to reduce the performance difference that you saw. This is actually my main concern with using idr_replace() as suggested by Matthew, but I am OK to use it if people think absolute correctness is more important. Cheers, Longman
