Skip to content

Commit

Permalink
MIPS: Change definition of cpu_relax() for Loongson-3
Browse files Browse the repository at this point in the history
Linux expects that if a CPU modifies a memory location, then that
modification will eventually become visible to other CPUs in the system.

Loongson 3 CPUs include a Store Fill Buffer (SFB) which sits between a
core & its L1 data cache, queueing memory accesses & allowing for faster
forwarding of data from pending stores to younger loads from the core.
Unfortunately the SFB prioritizes loads such that a continuous stream of
loads may cause a pending write to be buffered indefinitely. This is
problematic if we end up with 2 CPUs which each perform a store that the
other polls for - one or both CPUs may end up with their stores buffered
in the SFB, never reaching cache due to the continuous reads from the
poll loop. Such a deadlock condition has been observed whilst running
qspinlock code.

This patch changes the definition of cpu_relax() to smp_mb() for
Loongson-3, forcing a flush of the SFB on SMP systems which will cause
any pending writes to make it as far as the L1 caches where they will
become visible to other CPUs. If the kernel is not compiled for SMP
support, this will expand to a barrier() as before.

This workaround matches that currently implemented for ARM when
CONFIG_ARM_ERRATA_754327=y, which was introduced by commit 534be1d
("ARM: 6194/1: change definition of cpu_relax() for ARM11MPCore").

Although the workaround is only required when the Loongson 3 SFB
functionality is enabled, and we only began explicitly enabling that
functionality in v4.7 with commit 1e820da ("MIPS: Loongson-3:
Introduce CONFIG_LOONGSON3_ENHANCEMENT"), existing or future firmware
may enable the SFB which means we may need the workaround backported to
earlier kernels too.

[[email protected]:
  - Reword commit message & comment.
  - Limit stable backport to v3.15+ where we support Loongson 3 CPUs.]

Signed-off-by: Huacai Chen <[email protected]>
Signed-off-by: Paul Burton <[email protected]>
References: 534be1d ("ARM: 6194/1: change definition of cpu_relax() for ARM11MPCore")
References: 1e820da ("MIPS: Loongson-3: Introduce CONFIG_LOONGSON3_ENHANCEMENT")
Patchwork: https://patchwork.linux-mips.org/patch/19830/
Cc: Ralf Baechle <[email protected]>
Cc: James Hogan <[email protected]>
Cc: [email protected]
Cc: Fuxin Zhang <[email protected]>
Cc: Zhangjin Wu <[email protected]>
Cc: Huacai Chen <[email protected]>
Cc: [email protected] # v3.15+
chenhuacai authored and paulburton committed Jul 23, 2018

Verified

This commit was signed with the committer’s verified signature.
qier222 qier222
1 parent 7886542 commit a307188
Showing 1 changed file with 13 additions and 0 deletions.
13 changes: 13 additions & 0 deletions arch/mips/include/asm/processor.h
Original file line number Diff line number Diff line change
@@ -386,7 +386,20 @@ unsigned long get_wchan(struct task_struct *p);
#define KSTK_ESP(tsk) (task_pt_regs(tsk)->regs[29])
#define KSTK_STATUS(tsk) (task_pt_regs(tsk)->cp0_status)

#ifdef CONFIG_CPU_LOONGSON3
/*
* Loongson-3's SFB (Store-Fill-Buffer) may buffer writes indefinitely when a
* tight read loop is executed, because reads take priority over writes & the
* hardware (incorrectly) doesn't ensure that writes will eventually occur.
*
* Since spin loops of any kind should have a cpu_relax() in them, force an SFB
* flush from cpu_relax() such that any pending writes will become visible as
* expected.
*/
#define cpu_relax() smp_mb()
#else
#define cpu_relax() barrier()
#endif

/*
* Return_address is a replacement for __builtin_return_address(count)

0 comments on commit a307188

Please sign in to comment.