Re: [PATCH v4] ring-buffer: Do not check si_mem_available() before deferred init completes
From: Amit Machhiwal
Date: Wed Nov 18 2026 - 13:25:32 EST
On 2026/11/17 04:28 PM, Steven Rostedt wrote:
> On Fri, 23 Oct 2026 20:24:59 +0530
> Amit Machhiwal <amachhiw@xxxxxxxxxxxxx> wrote:
>
> > During early boot (early_trace_init() called from start_kernel()),
> > tracing allocates initial ring buffers (temp_buffer, array_buffer, and
> > snapshot_buffer) for CPU 0. Later, as secondary CPUs come online during
> > smp_init(), each CPU's ring buffer is allocated via trace_rb_cpu_prepare().
> >
> > Before attempting page allocation, __rb_allocate_pages() performs a
> > heuristic check using si_mem_available() to return early with -ENOMEM if
> > memory appears insufficient.
> >
> > However, on kernels built with CONFIG_DEFERRED_STRUCT_PAGE_INIT=y,
> > NR_FREE_PAGES only reflects the initial non-deferred pool across the entire
> > period from early_trace_init() through smp_init(). defer_init() leaves
> > only a single section per node (e.g. 16 MiB with 64 KB pages) initialized
> > up-front, and on kernels with large static binary footprints (debug
> > configurations enabling PAGE_OWNER, DEBUG_PAGEALLOC, KFENCE, or SLUB_DEBUG),
> > early core initialisations (SLUB caches, vmalloc, static ftrace records)
> > consume virtually all managed pages in this pool.
> >
> > si_mem_available() therefore returns a false negative (often 0-1 pages)
> > across both SYSTEM_BOOTING (for the initial CPU 0 buffers) and
> > SYSTEM_SCHEDULING (for secondary CPU buffers via trace_rb_cpu_prepare()),
> > causing __rb_allocate_pages() to prematurely abort with -ENOMEM.
> >
> > This failure is false: the buddy allocator is designed to replenish
> > ZONE_NORMAL on demand via deferred_grow_zone() when an actual allocation
> > is attempted. Checking si_mem_available() before calling alloc_pages_node()
> > short-circuits this on-demand growth entirely.
> >
> > deferred_grow_zone() is only permanently disabled after page_alloc_init_late()
> > completes (it spawns per-node kthreads and blocks until done), which happens
> > inside kernel_init_freeable() — before system_state is set to
> > SYSTEM_FREEING_INITMEM. Skip the si_mem_available() check until
> > SYSTEM_FREEING_INITMEM, at which point all deferred pages are initialized
> > and si_mem_available() is accurate.
> >
> > Use IS_ENABLED(CONFIG_DEFERRED_STRUCT_PAGE_INIT) to set i = LONG_MAX during
> > this window, unconditionally bypassing the threshold check. On kernels
> > without deferred init IS_ENABLED() compiles the branch away and the original
> > check is preserved unchanged.
>
> As Vincent mentioned, the above is just word salad. Please make it human
> readable. That is, state what the actual issue is and how it's fixed and
> remove all the AI fluff.
>
> Also note, just because we no longer require an "Assisted-by" tag, it
> should still be mentioned how the patch was created. I want to know.
Hi Steven,
To answer how this was created: I investigated the issue, carried out the RCA
and debug runs on a Power11 KVM setup, and authored the patch logic. When
writing the commit message, in a (clearly failed) attempt to explain the
details thoroughly, I ended up introducing the wordiness and fluff. Later I did
use an LLM for a quick rewording and spell checking, which only made it worse I
guess.
For v5, I have completely rewritten the commit message myself to be concise
and to the point: stating the actual issue (deferred page init leaves
NR_FREE_PAGES near zero, making si_mem_available() under-report), the
spurious -ENOMEM failure during early_trace_init() and smp_init(), and the
fix mechanism.
<snip>
> > i = si_mem_available();
> > + if (IS_ENABLED(CONFIG_DEFERRED_STRUCT_PAGE_INIT) &&
> > + system_state < SYSTEM_FREEING_INITMEM)
> > + i = LONG_MAX;
>
> Make the above an if/else statement.
>
> if (IS_ENABLED(CONFIG_DEFERRED_STRUCT_PAGE_INIT) &&
> unlikely(system_state < SYSTEM_FREEING_INITMEM))
> i = LONG_MAX;
> else
> i = si_mem_available();
Sure, this makes more sense and avoids calling si_mem_available() when it is not
reliable anyway. I have incorporated this in v5.
Thanks,
Amit