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/system.h>
48 #include <asm/pgtable.h>
49 #include <linux/atomic.h>
50 #include <asm/tlbflush.h>
51 #include <asm/uncached.h>
52 #include <asm/sn/addrs.h>
53 #include <asm/sn/arch.h>
54 #include <asm/sn/mspec.h>
55 #include <asm/sn/sn_cpuid.h>
56 #include <asm/sn/io.h>
57 #include <asm/sn/bte.h>
58 #include <asm/sn/shubio.h>
61 #define FETCHOP_ID "SGI Fetchop,"
62 #define CACHED_ID "Cached,"
63 #define UNCACHED_ID "Uncached"
64 #define REVISION "4.0"
65 #define MSPEC_BASENAME "mspec"
68 * Page types allocated by the device.
70 enum mspec_page_type
{
83 * One of these structures is allocated when an mspec region is mmaped. The
84 * structure is pointed to by the vma->vm_private_data field in the vma struct.
85 * This structure is used to record the addresses of the mspec pages.
86 * This structure is shared by all vma's that are split off from the
87 * original vma when split_vma()'s are done.
89 * The refcnt is incremented atomically because mm->mmap_sem does not
90 * protect in fork case where multiple tasks share the vma_data.
93 atomic_t refcnt
; /* Number of vmas sharing the data. */
94 spinlock_t lock
; /* Serialize access to this structure. */
95 int count
; /* Number of pages allocated. */
96 enum mspec_page_type type
; /* Type of pages allocated. */
97 int flags
; /* See VMD_xxx below. */
98 unsigned long vm_start
; /* Original (unsplit) base. */
99 unsigned long vm_end
; /* Original (unsplit) end. */
100 unsigned long maddr
[0]; /* Array of MSPEC addresses. */
103 #define VMD_VMALLOCED 0x1 /* vmalloc'd rather than kmalloc'd */
105 /* used on shub2 to clear FOP cache in the HUB */
106 static unsigned long scratch_page
[MAX_NUMNODES
];
107 #define SH2_AMO_CACHE_ENTRIES 4
110 mspec_zero_block(unsigned long addr
, int len
)
120 nid
= nasid_to_cnodeid(get_node_number(__pa(addr
)));
121 p
= (void *)TO_AMO(scratch_page
[nid
]);
123 for (i
=0; i
< SH2_AMO_CACHE_ENTRIES
; i
++) {
124 FETCHOP_LOAD_OP(p
, FETCHOP_LOAD
);
125 p
+= FETCHOP_VAR_SIZE
;
129 status
= bte_copy(0, addr
& ~__IA64_UNCACHED_OFFSET
, len
,
130 BTE_WACQUIRE
| BTE_ZERO_FILL
, NULL
);
132 memset((char *) addr
, 0, len
);
141 * Called when a device mapping is created by a means other than mmap
142 * (via fork, munmap, etc.). Increments the reference count on the
143 * underlying mspec data so it is not freed prematurely.
146 mspec_open(struct vm_area_struct
*vma
)
148 struct vma_data
*vdata
;
150 vdata
= vma
->vm_private_data
;
151 atomic_inc(&vdata
->refcnt
);
157 * Called when unmapping a device mapping. Frees all mspec pages
158 * belonging to all the vma's sharing this vma_data structure.
161 mspec_close(struct vm_area_struct
*vma
)
163 struct vma_data
*vdata
;
164 int index
, last_index
;
165 unsigned long my_page
;
167 vdata
= vma
->vm_private_data
;
169 if (!atomic_dec_and_test(&vdata
->refcnt
))
172 last_index
= (vdata
->vm_end
- vdata
->vm_start
) >> PAGE_SHIFT
;
173 for (index
= 0; index
< last_index
; index
++) {
174 if (vdata
->maddr
[index
] == 0)
177 * Clear the page before sticking it back
180 my_page
= vdata
->maddr
[index
];
181 vdata
->maddr
[index
] = 0;
182 if (!mspec_zero_block(my_page
, PAGE_SIZE
))
183 uncached_free_page(my_page
, 1);
185 printk(KERN_WARNING
"mspec_close(): "
186 "failed to zero page %ld\n", my_page
);
189 if (vdata
->flags
& VMD_VMALLOCED
)
198 * Creates a mspec page and maps it to user space.
201 mspec_fault(struct vm_area_struct
*vma
, struct vm_fault
*vmf
)
203 unsigned long paddr
, maddr
;
205 pgoff_t index
= vmf
->pgoff
;
206 struct vma_data
*vdata
= vma
->vm_private_data
;
208 maddr
= (volatile unsigned long) vdata
->maddr
[index
];
210 maddr
= uncached_alloc_page(numa_node_id(), 1);
214 spin_lock(&vdata
->lock
);
215 if (vdata
->maddr
[index
] == 0) {
217 vdata
->maddr
[index
] = maddr
;
219 uncached_free_page(maddr
, 1);
220 maddr
= vdata
->maddr
[index
];
222 spin_unlock(&vdata
->lock
);
225 if (vdata
->type
== MSPEC_FETCHOP
)
226 paddr
= TO_AMO(maddr
);
228 paddr
= maddr
& ~__IA64_UNCACHED_OFFSET
;
230 pfn
= paddr
>> PAGE_SHIFT
;
233 * vm_insert_pfn can fail with -EBUSY, but in that case it will
234 * be because another thread has installed the pte first, so it
237 vm_insert_pfn(vma
, (unsigned long)vmf
->virtual_address
, pfn
);
239 return VM_FAULT_NOPAGE
;
242 static const struct vm_operations_struct mspec_vm_ops
= {
244 .close
= mspec_close
,
245 .fault
= mspec_fault
,
251 * Called when mmapping the device. Initializes the vma with a fault handler
252 * and private data structure necessary to allocate, track, and free the
256 mspec_mmap(struct file
*file
, struct vm_area_struct
*vma
,
257 enum mspec_page_type type
)
259 struct vma_data
*vdata
;
260 int pages
, vdata_size
, flags
= 0;
262 if (vma
->vm_pgoff
!= 0)
265 if ((vma
->vm_flags
& VM_SHARED
) == 0)
268 if ((vma
->vm_flags
& VM_WRITE
) == 0)
271 pages
= (vma
->vm_end
- vma
->vm_start
) >> PAGE_SHIFT
;
272 vdata_size
= sizeof(struct vma_data
) + pages
* sizeof(long);
273 if (vdata_size
<= PAGE_SIZE
)
274 vdata
= kzalloc(vdata_size
, GFP_KERNEL
);
276 vdata
= vzalloc(vdata_size
);
277 flags
= VMD_VMALLOCED
;
282 vdata
->vm_start
= vma
->vm_start
;
283 vdata
->vm_end
= vma
->vm_end
;
284 vdata
->flags
= flags
;
286 spin_lock_init(&vdata
->lock
);
287 vdata
->refcnt
= ATOMIC_INIT(1);
288 vma
->vm_private_data
= vdata
;
290 vma
->vm_flags
|= (VM_IO
| VM_RESERVED
| VM_PFNMAP
| VM_DONTEXPAND
);
291 if (vdata
->type
== MSPEC_FETCHOP
|| vdata
->type
== MSPEC_UNCACHED
)
292 vma
->vm_page_prot
= pgprot_noncached(vma
->vm_page_prot
);
293 vma
->vm_ops
= &mspec_vm_ops
;
299 fetchop_mmap(struct file
*file
, struct vm_area_struct
*vma
)
301 return mspec_mmap(file
, vma
, MSPEC_FETCHOP
);
305 cached_mmap(struct file
*file
, struct vm_area_struct
*vma
)
307 return mspec_mmap(file
, vma
, MSPEC_CACHED
);
311 uncached_mmap(struct file
*file
, struct vm_area_struct
*vma
)
313 return mspec_mmap(file
, vma
, MSPEC_UNCACHED
);
316 static const struct file_operations fetchop_fops
= {
317 .owner
= THIS_MODULE
,
318 .mmap
= fetchop_mmap
,
319 .llseek
= noop_llseek
,
322 static struct miscdevice fetchop_miscdev
= {
323 .minor
= MISC_DYNAMIC_MINOR
,
324 .name
= "sgi_fetchop",
325 .fops
= &fetchop_fops
328 static const struct file_operations cached_fops
= {
329 .owner
= THIS_MODULE
,
331 .llseek
= noop_llseek
,
334 static struct miscdevice cached_miscdev
= {
335 .minor
= MISC_DYNAMIC_MINOR
,
336 .name
= "mspec_cached",
340 static const struct file_operations uncached_fops
= {
341 .owner
= THIS_MODULE
,
342 .mmap
= uncached_mmap
,
343 .llseek
= noop_llseek
,
346 static struct miscdevice uncached_miscdev
= {
347 .minor
= MISC_DYNAMIC_MINOR
,
348 .name
= "mspec_uncached",
349 .fops
= &uncached_fops
355 * Called at boot time to initialize the mspec facility.
364 * The fetchop device only works on SN2 hardware, uncached and cached
365 * memory drivers should both be valid on all ia64 hardware
368 if (ia64_platform_is("sn2")) {
372 for_each_node_state(nid
, N_ONLINE
) {
377 scratch_page
[nid
] = uncached_alloc_page(nid
, 1);
378 if (scratch_page
[nid
] == 0)
379 goto free_scratch_pages
;
380 phys
= __pa(scratch_page
[nid
]);
381 nasid
= get_node_number(phys
);
382 actual_nid
= nasid_to_cnodeid(nasid
);
383 if (actual_nid
!= nid
)
384 goto free_scratch_pages
;
388 ret
= misc_register(&fetchop_miscdev
);
391 "%s: failed to register device %i\n",
393 goto free_scratch_pages
;
397 ret
= misc_register(&cached_miscdev
);
399 printk(KERN_ERR
"%s: failed to register device %i\n",
402 misc_deregister(&fetchop_miscdev
);
403 goto free_scratch_pages
;
405 ret
= misc_register(&uncached_miscdev
);
407 printk(KERN_ERR
"%s: failed to register device %i\n",
409 misc_deregister(&cached_miscdev
);
411 misc_deregister(&fetchop_miscdev
);
412 goto free_scratch_pages
;
415 printk(KERN_INFO
"%s %s initialized devices: %s %s %s\n",
416 MSPEC_BASENAME
, REVISION
, is_sn2
? FETCHOP_ID
: "",
417 CACHED_ID
, UNCACHED_ID
);
423 if (scratch_page
[nid
] != 0)
424 uncached_free_page(scratch_page
[nid
], 1);
434 misc_deregister(&uncached_miscdev
);
435 misc_deregister(&cached_miscdev
);
437 misc_deregister(&fetchop_miscdev
);
440 if (scratch_page
[nid
] != 0)
441 uncached_free_page(scratch_page
[nid
], 1);
446 module_init(mspec_init
);
447 module_exit(mspec_exit
);
449 MODULE_AUTHOR("Silicon Graphics, Inc. <linux-altix@sgi.com>");
450 MODULE_DESCRIPTION("Driver for SGI SN special memory operations");
451 MODULE_LICENSE("GPL");