powerpc/pseries: add RTAS work area allocator

Various pseries-specific RTAS functions take a temporary "work area"
parameter - a buffer in memory accessible to RTAS. Typically such
functions are passed the statically allocated rtas_data_buf buffer as
the argument. This buffer is protected by a global spinlock. So users
of rtas_data_buf cannot perform sleeping operations while accessing
the buffer.

Most RTAS functions that have a work area parameter can return a
status (-2/990x) that indicates that the caller should retry. Before
retrying, the caller may need to reschedule or sleep (see
rtas_busy_delay() for details). This combination of factors
leads to uncomfortable constructions like this:

do {
spin_lock(&rtas_data_buf_lock);
rc = rtas_call(token, __pa(rtas_data_buf, ...);
if (rc == 0) {
/* parse or copy out rtas_data_buf contents */
}
spin_unlock(&rtas_data_buf_lock);
} while (rtas_busy_delay(rc));

Another unfortunately common way of handling this is for callers to
blithely ignore the possibility of a -2/990x status and hope for the
best.

If users were allowed to perform blocking operations while owning a
work area, the programming model would become less tedious and
error-prone. Users could schedule away, sleep, or perform other
blocking operations without having to release and re-acquire
resources.

We could continue to use a single work area buffer, and convert
rtas_data_buf_lock to a mutex. But that would impose an unnecessarily
coarse serialization on all users. As awkward as the current design
is, it prevents longer running operations that need to repeatedly use
rtas_data_buf from blocking the progress of others.

There are more considerations. One is that while 4KB is fine for all
current in-kernel uses, some RTAS calls can take much smaller buffers,
and some (VPD, platform dumps) would likely benefit from larger
ones. Another is that at least one RTAS function (ibm,get-vpd)
has *two* work area parameters. And finally, we should expect the
number of work area users in the kernel to increase over time as we
introduce lockdown-compatible ABIs to replace less safe use cases
based on sys_rtas/librtas.

So a special-purpose allocator for RTAS work area buffers seems worth
trying.

Properties:

* The backing memory for the allocator is reserved early in boot in
order to satisfy RTAS addressing requirements, and then managed with
genalloc.
* Allocations can block, but they never fail (mempool-like).
* Prioritizes first-come, first-serve fairness over throughput.
* Early boot allocations before the allocator has been initialized are
served via an internal static buffer.

Intended to replace rtas_data_buf. New code that needs RTAS work area
buffers should prefer this API.

Signed-off-by: Nathan Lynch <[e-mail]>
Signed-off-by: Michael Ellerman <[e-mail]>
Link: https://lore.kernel.org/r/[e-mail]

arch/powerpc/include/asm/rtas-work-area.h

···
+/* SPDX-License-Identifier: GPL-2.0-only */
+#ifndef _ASM_POWERPC_RTAS_WORK_AREA_H
+#define _ASM_POWERPC_RTAS_WORK_AREA_H
+
+#include <linux/build_bug.h>
+#include <linux/sizes.h>
+#include <linux/types.h>
+
+#include <asm/page.h>
+
+/**
+ * struct rtas_work_area - RTAS work area descriptor.
+ *
+ * Descriptor for a "work area" in PAPR terminology that satisfies
+ * RTAS addressing requirements.
+ */
+struct rtas_work_area {
+	/* private: Use the APIs provided below. */
+	char *buf;
+	size_t size;
+};
+
+enum {
+	/* Maximum allocation size, enforced at build time. */
+	RTAS_WORK_AREA_MAX_ALLOC_SZ = SZ_128K,
+};
+
+/**
+ * rtas_work_area_alloc() - Acquire a work area of the requested size.
+ * @size_: Allocation size. Must be compile-time constant and not more
+ *         than %RTAS_WORK_AREA_MAX_ALLOC_SZ.
+ *
+ * Allocate a buffer suitable for passing to RTAS functions that have
+ * a memory address parameter, often (but not always) referred to as a
+ * "work area" in PAPR. Although callers are allowed to block while
+ * holding a work area, the amount of memory reserved for this purpose
+ * is limited, and allocations should be short-lived. A good guideline
+ * is to release any allocated work area before returning from a
+ * system call.
+ *
+ * This function does not fail. It blocks until the allocation
+ * succeeds. To prevent deadlocks, callers are discouraged from
+ * allocating more than one work area simultaneously in a single task
+ * context.
+ *
+ * Context: This function may sleep.
+ * Return: A &struct rtas_work_area descriptor for the allocated work area.
+ */
+#define rtas_work_area_alloc(size_) ({				\
+	static_assert(__builtin_constant_p(size_));		\
+	static_assert((size_) > 0);				\
+	static_assert((size_) <= RTAS_WORK_AREA_MAX_ALLOC_SZ);	\
+	__rtas_work_area_alloc(size_);				\
+})
+
+/*
+ * Do not call __rtas_work_area_alloc() directly. Use
+ * rtas_work_area_alloc().
+ */
+struct rtas_work_area *__rtas_work_area_alloc(size_t size);
+
+/**
+ * rtas_work_area_free() - Release a work area.
+ * @area: Work area descriptor as returned from rtas_work_area_alloc().
+ *
+ * Return a work area buffer to the pool.
+ */
+void rtas_work_area_free(struct rtas_work_area *area);
+
+static inline char *rtas_work_area_raw_buf(const struct rtas_work_area *area)
+{
+	return area->buf;
+}
+
+static inline size_t rtas_work_area_size(const struct rtas_work_area *area)
+{
+	return area->size;
+}
+
+static inline phys_addr_t rtas_work_area_phys(const struct rtas_work_area *area)
+{
+	return __pa(area->buf);
+}
+
+/*
+ * Early setup for the work area allocator. Call from
+ * rtas_initialize() only.
+ */
+
+#ifdef CONFIG_PPC_PSERIES
+void rtas_work_area_reserve_arena(phys_addr_t limit);
+#else /* CONFIG_PPC_PSERIES */
+static inline void rtas_work_area_reserve_arena(phys_addr_t limit) {}
+#endif /* CONFIG_PPC_PSERIES */
+
+#endif /* _ASM_POWERPC_RTAS_WORK_AREA_H */

arch/powerpc/kernel/rtas.c

···
 #include <asm/machdep.h>
 #include <asm/mmu.h>
 #include <asm/page.h>
+#include <asm/rtas-work-area.h>
 #include <asm/rtas.h>
 #include <asm/time.h>
 #include <asm/trace.h>
···
 #endif
 	ibm_open_errinjct_token = rtas_token("ibm,open-errinjct");
 	ibm_errinjct_token = rtas_token("ibm,errinjct");
+
+	rtas_work_area_reserve_arena(rtas_region);
 }
 
 int __init early_init_dt_scan_rtas(unsigned long node,

arch/powerpc/platforms/pseries/Makefile

···
 ccflags-$(CONFIG_PPC_PSERIES_DEBUG)	+= -DDEBUG
 
 obj-y			:= lpar.o hvCall.o nvram.o reconfig.o \
-			   of_helpers.o \
+			   of_helpers.o rtas-work-area.o \
 			   setup.o iommu.o event_sources.o ras.o \
 			   firmware.o power.o dlpar.o mobility.o rng.o \
 			   pci.o pci_dlpar.o eeh_pseries.o msi.o \

arch/powerpc/platforms/pseries/rtas-work-area.c

···
+// SPDX-License-Identifier: GPL-2.0-only
+
+#define pr_fmt(fmt)	"rtas-work-area: " fmt
+
+#include <linux/genalloc.h>
+#include <linux/log2.h>
+#include <linux/kernel.h>
+#include <linux/memblock.h>
+#include <linux/mempool.h>
+#include <linux/minmax.h>
+#include <linux/mutex.h>
+#include <linux/numa.h>
+#include <linux/sizes.h>
+#include <linux/wait.h>
+
+#include <asm/machdep.h>
+#include <asm/rtas-work-area.h>
+#include <asm/rtas.h>
+
+enum {
+	/*
+	 * Ensure the pool is page-aligned.
+	 */
+	RTAS_WORK_AREA_ARENA_ALIGN = PAGE_SIZE,
+	/*
+	 * Don't let a single allocation claim the whole arena.
+	 */
+	RTAS_WORK_AREA_ARENA_SZ = RTAS_WORK_AREA_MAX_ALLOC_SZ * 2,
+	/*
+	 * The smallest known work area size is for ibm,get-vpd's
+	 * location code argument, which is limited to 79 characters
+	 * plus 1 nul terminator.
+	 *
+	 * PAPR+ 7.3.20 ibm,get-vpd RTAS Call
+	 * PAPR+ 12.3.2.4 Converged Location Code Rules - Length Restrictions
+	 */
+	RTAS_WORK_AREA_MIN_ALLOC_SZ = roundup_pow_of_two(80),
+};
+
+static struct {
+	struct gen_pool *gen_pool;
+	char *arena;
+	struct mutex mutex; /* serializes allocations */
+	struct wait_queue_head wqh;
+	mempool_t descriptor_pool;
+	bool available;
+} rwa_state = {
+	.mutex = __MUTEX_INITIALIZER(rwa_state.mutex),
+	.wqh = __WAIT_QUEUE_HEAD_INITIALIZER(rwa_state.wqh),
+};
+
+/*
+ * A single work area buffer and descriptor to serve requests early in
+ * boot before the allocator is fully initialized. We know 4KB is the
+ * most any boot time user needs (they all call ibm,get-system-parameter).
+ */
+static bool early_work_area_in_use __initdata;
+static char early_work_area_buf[SZ_4K] __initdata __aligned(SZ_4K);
+static struct rtas_work_area early_work_area __initdata = {
+	.buf = early_work_area_buf,
+	.size = sizeof(early_work_area_buf),
+};
+
+
+static struct rtas_work_area * __init rtas_work_area_alloc_early(size_t size)
+{
+	WARN_ON(size > early_work_area.size);
+	WARN_ON(early_work_area_in_use);
+	early_work_area_in_use = true;
+	memset(early_work_area.buf, 0, early_work_area.size);
+	return &early_work_area;
+}
+
+static void __init rtas_work_area_free_early(struct rtas_work_area *work_area)
+{
+	WARN_ON(work_area != &early_work_area);
+	WARN_ON(!early_work_area_in_use);
+	early_work_area_in_use = false;
+}
+
+struct rtas_work_area * __ref __rtas_work_area_alloc(size_t size)
+{
+	struct rtas_work_area *area;
+	unsigned long addr;
+
+	might_sleep();
+
+	/*
+	 * The rtas_work_area_alloc() wrapper enforces this at build
+	 * time. Requests that exceed the arena size will block
+	 * indefinitely.
+	 */
+	WARN_ON(size > RTAS_WORK_AREA_MAX_ALLOC_SZ);
+
+	if (!rwa_state.available)
+		return rtas_work_area_alloc_early(size);
+	/*
+	 * To ensure FCFS behavior and prevent a high rate of smaller
+	 * requests from starving larger ones, use the mutex to queue
+	 * allocations.
+	 */
+	mutex_lock(&rwa_state.mutex);
+	wait_event(rwa_state.wqh,
+		   (addr = gen_pool_alloc(rwa_state.gen_pool, size)) != 0);
+	mutex_unlock(&rwa_state.mutex);
+
+	area = mempool_alloc(&rwa_state.descriptor_pool, GFP_KERNEL);
+	area->buf = (char *)addr;
+	area->size = size;
+
+	return area;
+}
+
+void __ref rtas_work_area_free(struct rtas_work_area *area)
+{
+	if (!rwa_state.available) {
+		rtas_work_area_free_early(area);
+		return;
+	}
+
+	gen_pool_free(rwa_state.gen_pool, (unsigned long)area->buf, area->size);
+	mempool_free(area, &rwa_state.descriptor_pool);
+	wake_up(&rwa_state.wqh);
+}
+
+/*
+ * Initialization of the work area allocator happens in two parts. To
+ * reliably reserve an arena that satisfies RTAS addressing
+ * requirements, we must perform a memblock allocation early,
+ * immmediately after RTAS instantiation. Then we have to wait until
+ * the slab allocator is up before setting up the descriptor mempool
+ * and adding the arena to a gen_pool.
+ */
+static __init int rtas_work_area_allocator_init(void)
+{
+	const unsigned int order = ilog2(RTAS_WORK_AREA_MIN_ALLOC_SZ);
+	const phys_addr_t pa_start = __pa(rwa_state.arena);
+	const phys_addr_t pa_end = pa_start + RTAS_WORK_AREA_ARENA_SZ - 1;
+	struct gen_pool *pool;
+	const int nid = NUMA_NO_NODE;
+	int err;
+
+	err = -ENOMEM;
+	if (!rwa_state.arena)
+		goto err_out;
+
+	pool = gen_pool_create(order, nid);
+	if (!pool)
+		goto err_out;
+	/*
+	 * All RTAS functions that consume work areas are OK with
+	 * natural alignment, when they have alignment requirements at
+	 * all.
+	 */
+	gen_pool_set_algo(pool, gen_pool_first_fit_order_align, NULL);
+
+	err = gen_pool_add(pool, (unsigned long)rwa_state.arena,
+			   RTAS_WORK_AREA_ARENA_SZ, nid);
+	if (err)
+		goto err_destroy;
+
+	err = mempool_init_kmalloc_pool(&rwa_state.descriptor_pool, 1,
+					sizeof(struct rtas_work_area));
+	if (err)
+		goto err_destroy;
+
+	rwa_state.gen_pool = pool;
+	rwa_state.available = true;
+
+	pr_debug("arena [%pa-%pa] (%uK), min/max alloc sizes %u/%u\n",
+		 &pa_start, &pa_end,
+		 RTAS_WORK_AREA_ARENA_SZ / SZ_1K,
+		 RTAS_WORK_AREA_MIN_ALLOC_SZ,
+		 RTAS_WORK_AREA_MAX_ALLOC_SZ);
+
+	return 0;
+
+err_destroy:
+	gen_pool_destroy(pool);
+err_out:
+	return err;
+}
+machine_arch_initcall(pseries, rtas_work_area_allocator_init);
+
+/**
+ * rtas_work_area_reserve_arena() - Reserve memory suitable for RTAS work areas.
+ */
+void __init rtas_work_area_reserve_arena(const phys_addr_t limit)
+{
+	const phys_addr_t align = RTAS_WORK_AREA_ARENA_ALIGN;
+	const phys_addr_t size = RTAS_WORK_AREA_ARENA_SZ;
+	const phys_addr_t min = MEMBLOCK_LOW_LIMIT;
+	const int nid = NUMA_NO_NODE;
+
+	/*
+	 * Too early for a machine_is(pseries) check. But PAPR
+	 * effectively mandates that ibm,get-system-parameter is
+	 * present:
+	 *
+	 * R1–7.3.16–1. All platforms must support the System
+	 * Parameters option.
+	 *
+	 * So set up the arena if we find that, with a fallback to
+	 * ibm,configure-connector, just in case.
+	 */
+	if (rtas_service_present("ibm,get-system-parameter") ||
+	    rtas_service_present("ibm,configure-connector"))
+		rwa_state.arena = memblock_alloc_try_nid(size, align, min, limit, nid);
+}