2 * Copyright (C) 2001-2006 Silicon Graphics, Inc. All rights
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms of version 2 of the GNU General Public License
7 * as published by the Free Software Foundation.
11 * SN Platform Special Memory (mspec) Support
13 * This driver exports the SN special memory (mspec) facility to user
15 * There are three types of memory made available thru this driver:
16 * fetchops, uncached and cached.
18 * Fetchops are atomic memory operations that are implemented in the
19 * memory controller on SGI SN hardware.
21 * Uncached are used for memory write combining feature of the ia64
24 * Cached are used for areas of memory that are used as cached addresses
25 * on our partition and used as uncached addresses from other partitions.
26 * Due to a design constraint of the SN2 Shub, you can not have processors
27 * on the same FSB perform both a cached and uncached reference to the
28 * same cache line. These special memory cached regions prevent the
29 * kernel from ever dropping in a TLB entry and therefore prevent the
30 * processor from ever speculating a cache line from this page.
33 #include <linux/types.h>
34 #include <linux/kernel.h>
35 #include <linux/module.h>
36 #include <linux/init.h>
37 #include <linux/errno.h>
38 #include <linux/miscdevice.h>
39 #include <linux/spinlock.h>
42 #include <linux/vmalloc.h>
43 #include <linux/string.h>
44 #include <linux/slab.h>
45 #include <linux/numa.h>
47 #include <asm/pgtable.h>
48 #include <linux/atomic.h>
49 #include <asm/tlbflush.h>
50 #include <asm/uncached.h>
51 #include <asm/sn/addrs.h>
52 #include <asm/sn/arch.h>
53 #include <asm/sn/mspec.h>
54 #include <asm/sn/sn_cpuid.h>
55 #include <asm/sn/io.h>
56 #include <asm/sn/bte.h>
57 #include <asm/sn/shubio.h>
60 #define FETCHOP_ID "SGI Fetchop,"
61 #define CACHED_ID "Cached,"
62 #define UNCACHED_ID "Uncached"
63 #define REVISION "4.0"
64 #define MSPEC_BASENAME "mspec"
67 * Page types allocated by the device.
69 enum mspec_page_type {
82 * One of these structures is allocated when an mspec region is mmaped. The
83 * structure is pointed to by the vma->vm_private_data field in the vma struct.
84 * This structure is used to record the addresses of the mspec pages.
85 * This structure is shared by all vma's that are split off from the
86 * original vma when split_vma()'s are done.
88 * The refcnt is incremented atomically because mm->mmap_sem does not
89 * protect in fork case where multiple tasks share the vma_data.
92 atomic_t refcnt; /* Number of vmas sharing the data. */
93 spinlock_t lock; /* Serialize access to this structure. */
94 int count; /* Number of pages allocated. */
95 enum mspec_page_type type; /* Type of pages allocated. */
96 unsigned long vm_start; /* Original (unsplit) base. */
97 unsigned long vm_end; /* Original (unsplit) end. */
98 unsigned long maddr[0]; /* Array of MSPEC addresses. */
101 /* used on shub2 to clear FOP cache in the HUB */
102 static unsigned long scratch_page[MAX_NUMNODES];
103 #define SH2_AMO_CACHE_ENTRIES 4
106 mspec_zero_block(unsigned long addr, int len)
116 nid = nasid_to_cnodeid(get_node_number(__pa(addr)));
117 p = (void *)TO_AMO(scratch_page[nid]);
119 for (i=0; i < SH2_AMO_CACHE_ENTRIES; i++) {
120 FETCHOP_LOAD_OP(p, FETCHOP_LOAD);
121 p += FETCHOP_VAR_SIZE;
125 status = bte_copy(0, addr & ~__IA64_UNCACHED_OFFSET, len,
126 BTE_WACQUIRE | BTE_ZERO_FILL, NULL);
128 memset((char *) addr, 0, len);
137 * Called when a device mapping is created by a means other than mmap
138 * (via fork, munmap, etc.). Increments the reference count on the
139 * underlying mspec data so it is not freed prematurely.
142 mspec_open(struct vm_area_struct *vma)
144 struct vma_data *vdata;
146 vdata = vma->vm_private_data;
147 atomic_inc(&vdata->refcnt);
153 * Called when unmapping a device mapping. Frees all mspec pages
154 * belonging to all the vma's sharing this vma_data structure.
157 mspec_close(struct vm_area_struct *vma)
159 struct vma_data *vdata;
160 int index, last_index;
161 unsigned long my_page;
163 vdata = vma->vm_private_data;
165 if (!atomic_dec_and_test(&vdata->refcnt))
168 last_index = (vdata->vm_end - vdata->vm_start) >> PAGE_SHIFT;
169 for (index = 0; index < last_index; index++) {
170 if (vdata->maddr[index] == 0)
173 * Clear the page before sticking it back
176 my_page = vdata->maddr[index];
177 vdata->maddr[index] = 0;
178 if (!mspec_zero_block(my_page, PAGE_SIZE))
179 uncached_free_page(my_page, 1);
181 printk(KERN_WARNING "mspec_close(): "
182 "failed to zero page %ld\n", my_page);
191 * Creates a mspec page and maps it to user space.
194 mspec_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
196 unsigned long paddr, maddr;
198 pgoff_t index = vmf->pgoff;
199 struct vma_data *vdata = vma->vm_private_data;
201 maddr = (volatile unsigned long) vdata->maddr[index];
203 maddr = uncached_alloc_page(numa_node_id(), 1);
207 spin_lock(&vdata->lock);
208 if (vdata->maddr[index] == 0) {
210 vdata->maddr[index] = maddr;
212 uncached_free_page(maddr, 1);
213 maddr = vdata->maddr[index];
215 spin_unlock(&vdata->lock);
218 if (vdata->type == MSPEC_FETCHOP)
219 paddr = TO_AMO(maddr);
221 paddr = maddr & ~__IA64_UNCACHED_OFFSET;
223 pfn = paddr >> PAGE_SHIFT;
226 * vm_insert_pfn can fail with -EBUSY, but in that case it will
227 * be because another thread has installed the pte first, so it
230 vm_insert_pfn(vma, (unsigned long)vmf->virtual_address, pfn);
232 return VM_FAULT_NOPAGE;
235 static const struct vm_operations_struct mspec_vm_ops = {
237 .close = mspec_close,
238 .fault = mspec_fault,
244 * Called when mmapping the device. Initializes the vma with a fault handler
245 * and private data structure necessary to allocate, track, and free the
249 mspec_mmap(struct file *file, struct vm_area_struct *vma,
250 enum mspec_page_type type)
252 struct vma_data *vdata;
253 int pages, vdata_size;
255 if (vma->vm_pgoff != 0)
258 if ((vma->vm_flags & VM_SHARED) == 0)
261 if ((vma->vm_flags & VM_WRITE) == 0)
264 pages = vma_pages(vma);
265 vdata_size = sizeof(struct vma_data) + pages * sizeof(long);
266 if (vdata_size <= PAGE_SIZE)
267 vdata = kzalloc(vdata_size, GFP_KERNEL);
269 vdata = vzalloc(vdata_size);
273 vdata->vm_start = vma->vm_start;
274 vdata->vm_end = vma->vm_end;
276 spin_lock_init(&vdata->lock);
277 atomic_set(&vdata->refcnt, 1);
278 vma->vm_private_data = vdata;
280 vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
281 if (vdata->type == MSPEC_FETCHOP || vdata->type == MSPEC_UNCACHED)
282 vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
283 vma->vm_ops = &mspec_vm_ops;
289 fetchop_mmap(struct file *file, struct vm_area_struct *vma)
291 return mspec_mmap(file, vma, MSPEC_FETCHOP);
295 cached_mmap(struct file *file, struct vm_area_struct *vma)
297 return mspec_mmap(file, vma, MSPEC_CACHED);
301 uncached_mmap(struct file *file, struct vm_area_struct *vma)
303 return mspec_mmap(file, vma, MSPEC_UNCACHED);
306 static const struct file_operations fetchop_fops = {
307 .owner = THIS_MODULE,
308 .mmap = fetchop_mmap,
309 .llseek = noop_llseek,
312 static struct miscdevice fetchop_miscdev = {
313 .minor = MISC_DYNAMIC_MINOR,
314 .name = "sgi_fetchop",
315 .fops = &fetchop_fops
318 static const struct file_operations cached_fops = {
319 .owner = THIS_MODULE,
321 .llseek = noop_llseek,
324 static struct miscdevice cached_miscdev = {
325 .minor = MISC_DYNAMIC_MINOR,
326 .name = "mspec_cached",
330 static const struct file_operations uncached_fops = {
331 .owner = THIS_MODULE,
332 .mmap = uncached_mmap,
333 .llseek = noop_llseek,
336 static struct miscdevice uncached_miscdev = {
337 .minor = MISC_DYNAMIC_MINOR,
338 .name = "mspec_uncached",
339 .fops = &uncached_fops
345 * Called at boot time to initialize the mspec facility.
354 * The fetchop device only works on SN2 hardware, uncached and cached
355 * memory drivers should both be valid on all ia64 hardware
358 if (ia64_platform_is("sn2")) {
362 for_each_node_state(nid, N_ONLINE) {
367 scratch_page[nid] = uncached_alloc_page(nid, 1);
368 if (scratch_page[nid] == 0)
369 goto free_scratch_pages;
370 phys = __pa(scratch_page[nid]);
371 nasid = get_node_number(phys);
372 actual_nid = nasid_to_cnodeid(nasid);
373 if (actual_nid != nid)
374 goto free_scratch_pages;
378 ret = misc_register(&fetchop_miscdev);
381 "%s: failed to register device %i\n",
383 goto free_scratch_pages;
387 ret = misc_register(&cached_miscdev);
389 printk(KERN_ERR "%s: failed to register device %i\n",
392 misc_deregister(&fetchop_miscdev);
393 goto free_scratch_pages;
395 ret = misc_register(&uncached_miscdev);
397 printk(KERN_ERR "%s: failed to register device %i\n",
399 misc_deregister(&cached_miscdev);
401 misc_deregister(&fetchop_miscdev);
402 goto free_scratch_pages;
405 printk(KERN_INFO "%s %s initialized devices: %s %s %s\n",
406 MSPEC_BASENAME, REVISION, is_sn2 ? FETCHOP_ID : "",
407 CACHED_ID, UNCACHED_ID);
413 if (scratch_page[nid] != 0)
414 uncached_free_page(scratch_page[nid], 1);
424 misc_deregister(&uncached_miscdev);
425 misc_deregister(&cached_miscdev);
427 misc_deregister(&fetchop_miscdev);
430 if (scratch_page[nid] != 0)
431 uncached_free_page(scratch_page[nid], 1);
436 module_init(mspec_init);
437 module_exit(mspec_exit);
440 MODULE_DESCRIPTION("Driver for SGI SN special memory operations");
441 MODULE_LICENSE("GPL");