On Wed, Jan 8, 2025 at 7:50 AM Nhat Pham nphamcs@gmail.com wrote:
On Wed, Jan 8, 2025 at 10:36 PM Yosry Ahmed yosryahmed@google.com wrote:
Oh, I was not talking about my proposed diff, but the existing logic that allocates the requests and buffers in the hotplug callbacks instead of just using alloc_percpu() to allocate them once for each possible CPU. I was wondering if there are actual setups where this matters and a significant amount of memory is being saved. Otherwise we should simplify things and just rip out the hotplug callbacks.
My vote is for ripping the hotplug callbacks (eventually) :) In addition to the discrepancy in the number of possible and online CPUs, we also need a relatively smaller memory size for the discrepancy to matter, no? Systems with hundreds of CPUs (hopefully) should have hundreds of GBs worth of memory available (if not more).
Anyhow, we can just go with the diff you sent for now (and for past kernels). Seems simple enough, and wouldn't get in the way of the eventual hotplug logic removal (if you decide to pursue it).
Yeah I sent that out just now: https://lore.kernel.org/lkml/20250108161529.3193825-1-yosryahmed@google.com/
I will look into sending an RFC to rip out the hotplug callbacks later.