1// SPDX-License-Identifier: GPL-2.0-only
  2/*
  3 * Copyright (C) 2001-2008 Silicon Graphics, Inc.  All rights reserved.
  4 *
  5 * A simple uncached page allocator using the generic allocator. This
  6 * allocator first utilizes the spare (spill) pages found in the EFI
  7 * memmap and will then start converting cached pages to uncached ones
  8 * at a granule at a time. Node awareness is implemented by having a
  9 * pool of pages per node.
 10 */
 11
 12#include <linux/types.h>
 13#include <linux/kernel.h>
 14#include <linux/module.h>
 15#include <linux/init.h>
 16#include <linux/errno.h>
 17#include <linux/string.h>
 18#include <linux/efi.h>
 19#include <linux/nmi.h>
 20#include <linux/genalloc.h>
 21#include <linux/gfp.h>
 22#include <linux/pgtable.h>
 23#include <asm/efi.h>
 24#include <asm/page.h>
 25#include <asm/pal.h>
 26#include <linux/atomic.h>
 27#include <asm/tlbflush.h>
 28
 29struct uncached_pool {
 30	struct gen_pool *pool;
 31	struct mutex add_chunk_mutex;	/* serialize adding a converted chunk */
 32	int nchunks_added;		/* #of converted chunks added to pool */
 33	atomic_t status;		/* smp called function's return status*/
 34};
 35
 36#define MAX_CONVERTED_CHUNKS_PER_NODE	2
 37
 38struct uncached_pool uncached_pools[MAX_NUMNODES];
 39
 40
 41static void uncached_ipi_visibility(void *data)
 42{
 43	int status;
 44	struct uncached_pool *uc_pool = (struct uncached_pool *)data;
 45
 46	status = ia64_pal_prefetch_visibility(PAL_VISIBILITY_PHYSICAL);
 47	if ((status != PAL_VISIBILITY_OK) &&
 48	    (status != PAL_VISIBILITY_OK_REMOTE_NEEDED))
 49		atomic_inc(&uc_pool->status);
 50}
 51
 52
 53static void uncached_ipi_mc_drain(void *data)
 54{
 55	int status;
 56	struct uncached_pool *uc_pool = (struct uncached_pool *)data;
 57
 58	status = ia64_pal_mc_drain();
 59	if (status != PAL_STATUS_SUCCESS)
 60		atomic_inc(&uc_pool->status);
 61}
 62
 63
 64/*
 65 * Add a new chunk of uncached memory pages to the specified pool.
 66 *
 67 * @pool: pool to add new chunk of uncached memory to
 68 * @nid: node id of node to allocate memory from, or -1
 69 *
 70 * This is accomplished by first allocating a granule of cached memory pages
 71 * and then converting them to uncached memory pages.
 72 */
 73static int uncached_add_chunk(struct uncached_pool *uc_pool, int nid)
 74{
 75	struct page *page;
 76	int status, i, nchunks_added = uc_pool->nchunks_added;
 77	unsigned long c_addr, uc_addr;
 78
 79	if (mutex_lock_interruptible(&uc_pool->add_chunk_mutex) != 0)
 80		return -1;	/* interrupted by a signal */
 81
 82	if (uc_pool->nchunks_added > nchunks_added) {
 83		/* someone added a new chunk while we were waiting */
 84		mutex_unlock(&uc_pool->add_chunk_mutex);
 85		return 0;
 86	}
 87
 88	if (uc_pool->nchunks_added >= MAX_CONVERTED_CHUNKS_PER_NODE) {
 89		mutex_unlock(&uc_pool->add_chunk_mutex);
 90		return -1;
 91	}
 92
 93	/* attempt to allocate a granule's worth of cached memory pages */
 94
 95	page = __alloc_pages_node(nid,
 96				GFP_KERNEL | __GFP_ZERO | __GFP_THISNODE,
 97				IA64_GRANULE_SHIFT-PAGE_SHIFT);
 98	if (!page) {
 99		mutex_unlock(&uc_pool->add_chunk_mutex);
100		return -1;
101	}
102
103	/* convert the memory pages from cached to uncached */
104
105	c_addr = (unsigned long)page_address(page);
106	uc_addr = c_addr - PAGE_OFFSET + __IA64_UNCACHED_OFFSET;
107
108	/*
109	 * There's a small race here where it's possible for someone to
110	 * access the page through /dev/mem halfway through the conversion
111	 * to uncached - not sure it's really worth bothering about
112	 */
113	for (i = 0; i < (IA64_GRANULE_SIZE / PAGE_SIZE); i++)
114		SetPageUncached(&page[i]);
115
116	flush_tlb_kernel_range(uc_addr, uc_addr + IA64_GRANULE_SIZE);
117
118	status = ia64_pal_prefetch_visibility(PAL_VISIBILITY_PHYSICAL);
119	if (status == PAL_VISIBILITY_OK_REMOTE_NEEDED) {
120		atomic_set(&uc_pool->status, 0);
121		smp_call_function(uncached_ipi_visibility, uc_pool, 1);
122		if (atomic_read(&uc_pool->status))
123			goto failed;
124	} else if (status != PAL_VISIBILITY_OK)
125		goto failed;
126
127	preempt_disable();
128
129	flush_icache_range(uc_addr, uc_addr + IA64_GRANULE_SIZE);
130
131	/* flush the just introduced uncached translation from the TLB */
132	local_flush_tlb_all();
133
134	preempt_enable();
135
136	status = ia64_pal_mc_drain();
137	if (status != PAL_STATUS_SUCCESS)
138		goto failed;
139	atomic_set(&uc_pool->status, 0);
140	smp_call_function(uncached_ipi_mc_drain, uc_pool, 1);
141	if (atomic_read(&uc_pool->status))
142		goto failed;
143
144	/*
145	 * The chunk of memory pages has been converted to uncached so now we
146	 * can add it to the pool.
147	 */
148	status = gen_pool_add(uc_pool->pool, uc_addr, IA64_GRANULE_SIZE, nid);
149	if (status)
150		goto failed;
151
152	uc_pool->nchunks_added++;
153	mutex_unlock(&uc_pool->add_chunk_mutex);
154	return 0;
155
156	/* failed to convert or add the chunk so give it back to the kernel */
157failed:
158	for (i = 0; i < (IA64_GRANULE_SIZE / PAGE_SIZE); i++)
159		ClearPageUncached(&page[i]);
160
161	free_pages(c_addr, IA64_GRANULE_SHIFT-PAGE_SHIFT);
162	mutex_unlock(&uc_pool->add_chunk_mutex);
163	return -1;
164}
165
166
167/*
168 * uncached_alloc_page
169 *
170 * @starting_nid: node id of node to start with, or -1
171 * @n_pages: number of contiguous pages to allocate
172 *
173 * Allocate the specified number of contiguous uncached pages on the
174 * requested node. If not enough contiguous uncached pages are available
175 * on the requested node, roundrobin starting with the next higher node.
176 */
177unsigned long uncached_alloc_page(int starting_nid, int n_pages)
178{
179	unsigned long uc_addr;
180	struct uncached_pool *uc_pool;
181	int nid;
182
183	if (unlikely(starting_nid >= MAX_NUMNODES))
184		return 0;
185
186	if (starting_nid < 0)
187		starting_nid = numa_node_id();
188	nid = starting_nid;
189
190	do {
191		if (!node_state(nid, N_HIGH_MEMORY))
192			continue;
193		uc_pool = &uncached_pools[nid];
194		if (uc_pool->pool == NULL)
195			continue;
196		do {
197			uc_addr = gen_pool_alloc(uc_pool->pool,
198						 n_pages * PAGE_SIZE);
199			if (uc_addr != 0)
200				return uc_addr;
201		} while (uncached_add_chunk(uc_pool, nid) == 0);
202
203	} while ((nid = (nid + 1) % MAX_NUMNODES) != starting_nid);
204
205	return 0;
206}
207EXPORT_SYMBOL(uncached_alloc_page);
208
209
210/*
211 * uncached_free_page
212 *
213 * @uc_addr: uncached address of first page to free
214 * @n_pages: number of contiguous pages to free
215 *
216 * Free the specified number of uncached pages.
217 */
218void uncached_free_page(unsigned long uc_addr, int n_pages)
219{
220	int nid = paddr_to_nid(uc_addr - __IA64_UNCACHED_OFFSET);
221	struct gen_pool *pool = uncached_pools[nid].pool;
222
223	if (unlikely(pool == NULL))
224		return;
225
226	if ((uc_addr & (0XFUL << 60)) != __IA64_UNCACHED_OFFSET)
227		panic("uncached_free_page invalid address %lx\n", uc_addr);
228
229	gen_pool_free(pool, uc_addr, n_pages * PAGE_SIZE);
230}
231EXPORT_SYMBOL(uncached_free_page);
232
233
234/*
235 * uncached_build_memmap,
236 *
237 * @uc_start: uncached starting address of a chunk of uncached memory
238 * @uc_end: uncached ending address of a chunk of uncached memory
239 * @arg: ignored, (NULL argument passed in on call to efi_memmap_walk_uc())
240 *
241 * Called at boot time to build a map of pages that can be used for
242 * memory special operations.
243 */
244static int __init uncached_build_memmap(u64 uc_start, u64 uc_end, void *arg)
245{
246	int nid = paddr_to_nid(uc_start - __IA64_UNCACHED_OFFSET);
247	struct gen_pool *pool = uncached_pools[nid].pool;
248	size_t size = uc_end - uc_start;
249
250	touch_softlockup_watchdog();
251
252	if (pool != NULL) {
253		memset((char *)uc_start, 0, size);
254		(void) gen_pool_add(pool, uc_start, size, nid);
255	}
256	return 0;
257}
258
259
260static int __init uncached_init(void)
261{
262	int nid;
263
264	for_each_online_node(nid) {
265		uncached_pools[nid].pool = gen_pool_create(PAGE_SHIFT, nid);
266		mutex_init(&uncached_pools[nid].add_chunk_mutex);
267	}
268
269	efi_memmap_walk_uc(uncached_build_memmap, NULL);
270	return 0;
271}
272
273__initcall(uncached_init);