1 // SPDX-License-Identifier: GPL-2.0-only
2 
3 #define pr_fmt(fmt)	"rtas-work-area: " fmt
4 
5 #include <linux/genalloc.h>
6 #include <linux/log2.h>
7 #include <linux/kernel.h>
8 #include <linux/memblock.h>
9 #include <linux/mempool.h>
10 #include <linux/minmax.h>
11 #include <linux/mutex.h>
12 #include <linux/numa.h>
13 #include <linux/sizes.h>
14 #include <linux/wait.h>
15 
16 #include <asm/machdep.h>
17 #include <asm/rtas-work-area.h>
18 #include <asm/rtas.h>
19 
20 enum {
21 	/*
22 	 * Ensure the pool is page-aligned.
23 	 */
24 	RTAS_WORK_AREA_ARENA_ALIGN = PAGE_SIZE,
25 	/*
26 	 * Don't let a single allocation claim the whole arena.
27 	 */
28 	RTAS_WORK_AREA_ARENA_SZ = RTAS_WORK_AREA_MAX_ALLOC_SZ * 2,
29 	/*
30 	 * The smallest known work area size is for ibm,get-vpd's
31 	 * location code argument, which is limited to 79 characters
32 	 * plus 1 nul terminator.
33 	 *
34 	 * PAPR+ 7.3.20 ibm,get-vpd RTAS Call
35 	 * PAPR+ 12.3.2.4 Converged Location Code Rules - Length Restrictions
36 	 */
37 	RTAS_WORK_AREA_MIN_ALLOC_SZ = roundup_pow_of_two(80),
38 };
39 
40 static struct {
41 	struct gen_pool *gen_pool;
42 	char *arena;
43 	struct mutex mutex; /* serializes allocations */
44 	struct wait_queue_head wqh;
45 	mempool_t descriptor_pool;
46 	bool available;
47 } rwa_state = {
48 	.mutex = __MUTEX_INITIALIZER(rwa_state.mutex),
49 	.wqh = __WAIT_QUEUE_HEAD_INITIALIZER(rwa_state.wqh),
50 };
51 
52 /*
53  * A single work area buffer and descriptor to serve requests early in
54  * boot before the allocator is fully initialized. We know 4KB is the
55  * most any boot time user needs (they all call ibm,get-system-parameter).
56  */
57 static bool early_work_area_in_use __initdata;
58 static char early_work_area_buf[SZ_4K] __initdata __aligned(SZ_4K);
59 static struct rtas_work_area early_work_area __initdata = {
60 	.buf = early_work_area_buf,
61 	.size = sizeof(early_work_area_buf),
62 };
63 
64 
rtas_work_area_alloc_early(size_t size)65 static struct rtas_work_area * __init rtas_work_area_alloc_early(size_t size)
66 {
67 	WARN_ON(size > early_work_area.size);
68 	WARN_ON(early_work_area_in_use);
69 	early_work_area_in_use = true;
70 	memset(early_work_area.buf, 0, early_work_area.size);
71 	return &early_work_area;
72 }
73 
rtas_work_area_free_early(struct rtas_work_area * work_area)74 static void __init rtas_work_area_free_early(struct rtas_work_area *work_area)
75 {
76 	WARN_ON(work_area != &early_work_area);
77 	WARN_ON(!early_work_area_in_use);
78 	early_work_area_in_use = false;
79 }
80 
__rtas_work_area_alloc(size_t size)81 struct rtas_work_area * __ref __rtas_work_area_alloc(size_t size)
82 {
83 	struct rtas_work_area *area;
84 	unsigned long addr;
85 
86 	might_sleep();
87 
88 	/*
89 	 * The rtas_work_area_alloc() wrapper enforces this at build
90 	 * time. Requests that exceed the arena size will block
91 	 * indefinitely.
92 	 */
93 	WARN_ON(size > RTAS_WORK_AREA_MAX_ALLOC_SZ);
94 
95 	if (!rwa_state.available)
96 		return rtas_work_area_alloc_early(size);
97 	/*
98 	 * To ensure FCFS behavior and prevent a high rate of smaller
99 	 * requests from starving larger ones, use the mutex to queue
100 	 * allocations.
101 	 */
102 	mutex_lock(&rwa_state.mutex);
103 	wait_event(rwa_state.wqh,
104 		   (addr = gen_pool_alloc(rwa_state.gen_pool, size)) != 0);
105 	mutex_unlock(&rwa_state.mutex);
106 
107 	area = mempool_alloc(&rwa_state.descriptor_pool, GFP_KERNEL);
108 	area->buf = (char *)addr;
109 	area->size = size;
110 
111 	return area;
112 }
113 
rtas_work_area_free(struct rtas_work_area * area)114 void __ref rtas_work_area_free(struct rtas_work_area *area)
115 {
116 	if (!rwa_state.available) {
117 		rtas_work_area_free_early(area);
118 		return;
119 	}
120 
121 	gen_pool_free(rwa_state.gen_pool, (unsigned long)area->buf, area->size);
122 	mempool_free(area, &rwa_state.descriptor_pool);
123 	wake_up(&rwa_state.wqh);
124 }
125 
126 /*
127  * Initialization of the work area allocator happens in two parts. To
128  * reliably reserve an arena that satisfies RTAS addressing
129  * requirements, we must perform a memblock allocation early,
130  * immmediately after RTAS instantiation. Then we have to wait until
131  * the slab allocator is up before setting up the descriptor mempool
132  * and adding the arena to a gen_pool.
133  */
rtas_work_area_allocator_init(void)134 static __init int rtas_work_area_allocator_init(void)
135 {
136 	const unsigned int order = ilog2(RTAS_WORK_AREA_MIN_ALLOC_SZ);
137 	const phys_addr_t pa_start = __pa(rwa_state.arena);
138 	const phys_addr_t pa_end = pa_start + RTAS_WORK_AREA_ARENA_SZ - 1;
139 	struct gen_pool *pool;
140 	const int nid = NUMA_NO_NODE;
141 	int err;
142 
143 	err = -ENOMEM;
144 	if (!rwa_state.arena)
145 		goto err_out;
146 
147 	pool = gen_pool_create(order, nid);
148 	if (!pool)
149 		goto err_out;
150 	/*
151 	 * All RTAS functions that consume work areas are OK with
152 	 * natural alignment, when they have alignment requirements at
153 	 * all.
154 	 */
155 	gen_pool_set_algo(pool, gen_pool_first_fit_order_align, NULL);
156 
157 	err = gen_pool_add(pool, (unsigned long)rwa_state.arena,
158 			   RTAS_WORK_AREA_ARENA_SZ, nid);
159 	if (err)
160 		goto err_destroy;
161 
162 	err = mempool_init_kmalloc_pool(&rwa_state.descriptor_pool, 1,
163 					sizeof(struct rtas_work_area));
164 	if (err)
165 		goto err_destroy;
166 
167 	rwa_state.gen_pool = pool;
168 	rwa_state.available = true;
169 
170 	pr_debug("arena [%pa-%pa] (%uK), min/max alloc sizes %u/%u\n",
171 		 &pa_start, &pa_end,
172 		 RTAS_WORK_AREA_ARENA_SZ / SZ_1K,
173 		 RTAS_WORK_AREA_MIN_ALLOC_SZ,
174 		 RTAS_WORK_AREA_MAX_ALLOC_SZ);
175 
176 	return 0;
177 
178 err_destroy:
179 	gen_pool_destroy(pool);
180 err_out:
181 	return err;
182 }
183 machine_arch_initcall(pseries, rtas_work_area_allocator_init);
184 
185 /**
186  * rtas_work_area_reserve_arena() - Reserve memory suitable for RTAS work areas.
187  */
rtas_work_area_reserve_arena(const phys_addr_t limit)188 void __init rtas_work_area_reserve_arena(const phys_addr_t limit)
189 {
190 	const phys_addr_t align = RTAS_WORK_AREA_ARENA_ALIGN;
191 	const phys_addr_t size = RTAS_WORK_AREA_ARENA_SZ;
192 	const phys_addr_t min = MEMBLOCK_LOW_LIMIT;
193 	const int nid = NUMA_NO_NODE;
194 
195 	/*
196 	 * Too early for a machine_is(pseries) check. But PAPR
197 	 * effectively mandates that ibm,get-system-parameter is
198 	 * present:
199 	 *
200 	 * R1–7.3.16–1. All platforms must support the System
201 	 * Parameters option.
202 	 *
203 	 * So set up the arena if we find that, with a fallback to
204 	 * ibm,configure-connector, just in case.
205 	 */
206 	if (rtas_function_implemented(RTAS_FN_IBM_GET_SYSTEM_PARAMETER) ||
207 	    rtas_function_implemented(RTAS_FN_IBM_CONFIGURE_CONNECTOR))
208 		rwa_state.arena = memblock_alloc_try_nid(size, align, min, limit, nid);
209 }
210