sched/cache: Keep nr_pref_llc_running in the runnable domain, to fix LLC mis-scheduling buggrafted

alb_break_llc() decides whether to break LLC preference during active
load balance. It does so by testing that every runnable fair task on the
source rq prefers its LLC:

	env->src_rq->nr_pref_llc_running == env->src_rq->cfs.h_nr_runnable

But the two counters cover different sets. nr_pref_llc_running is updated
in account_llc_enqueue()/account_llc_dequeue(), next to cfs_rq->nr_queued,
so it follows queued tasks. h_nr_runnable is updated in set_delayed()/
clear_delayed() and drops delay-dequeued tasks.

So under DELAY_DEQUEUE, a preferring task that goes to sleep stays counted
in nr_pref_llc_running while h_nr_runnable falls. The equality then breaks,
alb_break_llc() returns false, and active balance is free to pull a task
off its preferred LLC. Active balance only moves runnable tasks, and this
is the only LLC check it consults: once the stopper runs, LBF_ACTIVE_LB
skips the per-task test in can_migrate_task(). The runnable set is the one
we want.

Fix it on the counter side. A task should be counted in
nr_pref_llc_running exactly while it is both queued on its preferred LLC
(pref_llc_queued) and runnable (!sched_delayed). Define that membership
once in task_pref_llc_runnable(), and adjust the counter only through
pref_llc_running_inc()/pref_llc_running_dec() from the four sites that
change either input: account_llc_enqueue(), account_llc_dequeue(),
set_delayed() and clear_delayed(). Gating every update on the same
predicate keeps the delay, wake and dequeue paths from double-counting
or underflowing; see the comments at those sites for the ordering.

nr_llc_running and sd->llc_counts are not touched and stay on queued
semantics.

Fixes: 714059f79ff0 ("sched/cache: Handle moving single tasks to/from their preferred LLC")
Closes: https://lore.kernel.org/lkml/20260827135000.735138-1-zhanxusheng@xiaomi.com/
Reported-by: Zhan Xusheng <zhanxusheng@xiaomi.com>
Suggested-by: Chen Yu <yu.c.chen@intel.com>
Signed-off-by: Tim Chen <tim.c.chen@linux.intel.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Reviewed-by: Kayra Cizmeci <kayracizmeci@gmail.com>
Cc: <stable@kernel.org> # v7.2.x
Link: https://patch.msgid.link/06af61afedac32e6477f57feb4d658f6c411c3af.1790035273.git.tim.c.chen@linux.intel.com

1 files changed, 104 insertions, 0 deletions

diff --git a/Documentation/networking/netmem.rst b/Documentation/networking/netmem.rst
new file mode 100644
index 000000000..217869d11
--- /dev/null
+++ b/Documentation/networking/netmem.rst
@@ -0,0 +1,104 @@
+.. SPDX-License-Identifier: GPL-2.0
+
+==================================
+Netmem Support for Network Drivers
+==================================
+
+This document outlines the requirements for network drivers to support netmem,
+an abstract memory type that enables features like device memory TCP. By
+supporting netmem, drivers can work with various underlying memory types
+with little to no modification.
+
+Benefits of Netmem :
+
+* Flexibility: Netmem can be backed by different memory types (e.g., struct
+  page, DMA-buf), allowing drivers to support various use cases such as device
+  memory TCP.
+* Future-proof: Drivers with netmem support are ready for upcoming
+  features that rely on it.
+* Simplified Development: Drivers interact with a consistent API,
+  regardless of the underlying memory implementation.
+
+Driver RX Requirements
+======================
+
+1. The driver must support page_pool.
+
+2. The driver must support the tcp-data-split ethtool option.
+
+3. The driver must use the page_pool netmem APIs for payload memory. The netmem
+   APIs currently 1-to-1 correspond with page APIs. Conversion to netmem should
+   be achievable by switching the page APIs to netmem APIs and tracking memory
+   via netmem_refs in the driver rather than struct page * :
+
+   - page_pool_alloc -> page_pool_alloc_netmem
+   - page_pool_get_dma_addr -> page_pool_get_dma_addr_netmem
+   - page_pool_put_page -> page_pool_put_netmem
+
+   Not all page APIs have netmem equivalents at the moment. If your driver
+   relies on a missing netmem API, feel free to add and propose to netdev@, or
+   reach out to the maintainers and/or almasrymina@google.com for help adding
+   the netmem API.
+
+4. The driver must use the following PP_FLAGS:
+
+   - PP_FLAG_DMA_MAP: netmem is not dma-mappable by the driver. The driver
+     must delegate the dma mapping to the page_pool, which knows when
+     dma-mapping is (or is not) appropriate.
+   - PP_FLAG_DMA_SYNC_DEV: netmem dma addr is not necessarily dma-syncable
+     by the driver. The driver must delegate the dma syncing to the page_pool,
+     which knows when dma-syncing is (or is not) appropriate.
+   - PP_FLAG_ALLOW_UNREADABLE_NETMEM. The driver must specify this flag iff
+     tcp-data-split is enabled.
+
+5. The driver must not assume the netmem is readable and/or backed by pages.
+   The netmem returned by the page_pool may be unreadable, in which case
+   netmem_address() will return NULL. The driver must correctly handle
+   unreadable netmem, i.e. don't attempt to handle its contents when
+   netmem_address() is NULL.
+
+   Ideally, drivers should not have to check the underlying netmem type via
+   helpers like netmem_is_net_iov() or convert the netmem to any of its
+   underlying types via netmem_to_page() or netmem_to_net_iov(). In most cases,
+   netmem or page_pool helpers that abstract this complexity are provided
+   (and more can be added).
+
+6. The driver must use page_pool_dma_sync_netmem_for_cpu() in lieu of
+   dma_sync_single_range_for_cpu(). For some memory providers, dma_syncing for
+   CPU will be done by the page_pool, for others (particularly dmabuf memory
+   provider), dma syncing for CPU is the responsibility of the userspace using
+   dmabuf APIs. The driver must delegate the entire dma-syncing operation to
+   the page_pool which will do it correctly.
+
+7. Avoid implementing driver-specific recycling on top of the page_pool. Drivers
+   cannot hold onto a struct page to do their own recycling as the netmem may
+   not be backed by a struct page. However, you may hold onto a page_pool
+   reference with page_pool_fragment_netmem() or page_pool_ref_netmem() for
+   that purpose, but be mindful that some netmem types might have longer
+   circulation times, such as when userspace holds a reference in zerocopy
+   scenarios.
+
+Driver TX Requirements
+======================
+
+1. The Driver must not pass the netmem dma_addr to any of the dma-mapping APIs
+   directly. This is because netmem dma_addrs may come from a source like
+   dma-buf that is not compatible with the dma-mapping APIs.
+
+   Helpers like netmem_dma_unmap_page_attrs() & netmem_dma_unmap_addr_set()
+   should be used in lieu of dma_unmap_page[_attrs](), dma_unmap_addr_set().
+   The netmem variants will handle netmem dma_addrs correctly regardless of the
+   source, delegating to the dma-mapping APIs when appropriate.
+
+   Not all dma-mapping APIs have netmem equivalents at the moment. If your
+   driver relies on a missing netmem API, feel free to add and propose to
+   netdev@, or reach out to the maintainers and/or almasrymina@google.com for
+   help adding the netmem API.
+
+2. Driver should declare support by setting `netdev->netmem_tx` to the
+   appropriate mode:
+
+   - `NETMEM_TX_DMA`: for physical devices that perform DMA.
+
+   - `NETMEM_TX_NO_DMA`: for virtual or passthrough devices that do
+     not DMA, but still support handling of netmem-backed skbs.