1/*
2 * Copyright (C) 2001-2006 Silicon Graphics, Inc.  All rights
3 * reserved.
4 *
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.
8 */
9
10/*
11 * SN Platform Special Memory (mspec) Support
12 *
13 * This driver exports the SN special memory (mspec) facility to user
14 * processes.
15 * There are three types of memory made available thru this driver:
16 * fetchops, uncached and cached.
17 *
18 * Fetchops are atomic memory operations that are implemented in the
19 * memory controller on SGI SN hardware.
20 *
21 * Uncached are used for memory write combining feature of the ia64
22 * cpu.
23 *
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.
31 */
32
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>
40#include <linux/mm.h>
41#include <linux/fs.h>
42#include <linux/vmalloc.h>
43#include <linux/string.h>
44#include <linux/slab.h>
45#include <linux/numa.h>
46#include <asm/page.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>
58
59
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"
65
66/*
67 * Page types allocated by the device.
68 */
69enum mspec_page_type {
70	MSPEC_FETCHOP = 1,
71	MSPEC_CACHED,
72	MSPEC_UNCACHED
73};
74
75#ifdef CONFIG_SGI_SN
76static int is_sn2;
77#else
78#define is_sn2		0
79#endif
80
81/*
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.
87 *
88 * The refcnt is incremented atomically because mm->mmap_sem does not
89 * protect in fork case where multiple tasks share the vma_data.
90 */
91struct 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	int flags;		/* See VMD_xxx below. */
97	unsigned long vm_start;	/* Original (unsplit) base. */
98	unsigned long vm_end;	/* Original (unsplit) end. */
99	unsigned long maddr[0];	/* Array of MSPEC addresses. */
100};
101
102#define VMD_VMALLOCED 0x1	/* vmalloc'd rather than kmalloc'd */
103
104/* used on shub2 to clear FOP cache in the HUB */
105static unsigned long scratch_page[MAX_NUMNODES];
106#define SH2_AMO_CACHE_ENTRIES	4
107
108static inline int
109mspec_zero_block(unsigned long addr, int len)
110{
111	int status;
112
113	if (is_sn2) {
114		if (is_shub2()) {
115			int nid;
116			void *p;
117			int i;
118
119			nid = nasid_to_cnodeid(get_node_number(__pa(addr)));
120			p = (void *)TO_AMO(scratch_page[nid]);
121
122			for (i=0; i < SH2_AMO_CACHE_ENTRIES; i++) {
123				FETCHOP_LOAD_OP(p, FETCHOP_LOAD);
124				p += FETCHOP_VAR_SIZE;
125			}
126		}
127
128		status = bte_copy(0, addr & ~__IA64_UNCACHED_OFFSET, len,
129				  BTE_WACQUIRE | BTE_ZERO_FILL, NULL);
130	} else {
131		memset((char *) addr, 0, len);
132		status = 0;
133	}
134	return status;
135}
136
137/*
138 * mspec_open
139 *
140 * Called when a device mapping is created by a means other than mmap
141 * (via fork, munmap, etc.).  Increments the reference count on the
142 * underlying mspec data so it is not freed prematurely.
143 */
144static void
145mspec_open(struct vm_area_struct *vma)
146{
147	struct vma_data *vdata;
148
149	vdata = vma->vm_private_data;
150	atomic_inc(&vdata->refcnt);
151}
152
153/*
154 * mspec_close
155 *
156 * Called when unmapping a device mapping. Frees all mspec pages
157 * belonging to all the vma's sharing this vma_data structure.
158 */
159static void
160mspec_close(struct vm_area_struct *vma)
161{
162	struct vma_data *vdata;
163	int index, last_index;
164	unsigned long my_page;
165
166	vdata = vma->vm_private_data;
167
168	if (!atomic_dec_and_test(&vdata->refcnt))
169		return;
170
171	last_index = (vdata->vm_end - vdata->vm_start) >> PAGE_SHIFT;
172	for (index = 0; index < last_index; index++) {
173		if (vdata->maddr[index] == 0)
174			continue;
175		/*
176		 * Clear the page before sticking it back
177		 * into the pool.
178		 */
179		my_page = vdata->maddr[index];
180		vdata->maddr[index] = 0;
181		if (!mspec_zero_block(my_page, PAGE_SIZE))
182			uncached_free_page(my_page, 1);
183		else
184			printk(KERN_WARNING "mspec_close(): "
185			       "failed to zero page %ld\n", my_page);
186	}
187
188	if (vdata->flags & VMD_VMALLOCED)
189		vfree(vdata);
190	else
191		kfree(vdata);
192}
193
194/*
195 * mspec_fault
196 *
197 * Creates a mspec page and maps it to user space.
198 */
199static int
200mspec_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
201{
202	unsigned long paddr, maddr;
203	unsigned long pfn;
204	pgoff_t index = vmf->pgoff;
205	struct vma_data *vdata = vma->vm_private_data;
206
207	maddr = (volatile unsigned long) vdata->maddr[index];
208	if (maddr == 0) {
209		maddr = uncached_alloc_page(numa_node_id(), 1);
210		if (maddr == 0)
211			return VM_FAULT_OOM;
212
213		spin_lock(&vdata->lock);
214		if (vdata->maddr[index] == 0) {
215			vdata->count++;
216			vdata->maddr[index] = maddr;
217		} else {
218			uncached_free_page(maddr, 1);
219			maddr = vdata->maddr[index];
220		}
221		spin_unlock(&vdata->lock);
222	}
223
224	if (vdata->type == MSPEC_FETCHOP)
225		paddr = TO_AMO(maddr);
226	else
227		paddr = maddr & ~__IA64_UNCACHED_OFFSET;
228
229	pfn = paddr >> PAGE_SHIFT;
230
231	/*
232	 * vm_insert_pfn can fail with -EBUSY, but in that case it will
233	 * be because another thread has installed the pte first, so it
234	 * is no problem.
235	 */
236	vm_insert_pfn(vma, (unsigned long)vmf->virtual_address, pfn);
237
238	return VM_FAULT_NOPAGE;
239}
240
241static const struct vm_operations_struct mspec_vm_ops = {
242	.open = mspec_open,
243	.close = mspec_close,
244	.fault = mspec_fault,
245};
246
247/*
248 * mspec_mmap
249 *
250 * Called when mmapping the device.  Initializes the vma with a fault handler
251 * and private data structure necessary to allocate, track, and free the
252 * underlying pages.
253 */
254static int
255mspec_mmap(struct file *file, struct vm_area_struct *vma,
256					enum mspec_page_type type)
257{
258	struct vma_data *vdata;
259	int pages, vdata_size, flags = 0;
260
261	if (vma->vm_pgoff != 0)
262		return -EINVAL;
263
264	if ((vma->vm_flags & VM_SHARED) == 0)
265		return -EINVAL;
266
267	if ((vma->vm_flags & VM_WRITE) == 0)
268		return -EPERM;
269
270	pages = vma_pages(vma);
271	vdata_size = sizeof(struct vma_data) + pages * sizeof(long);
272	if (vdata_size <= PAGE_SIZE)
273		vdata = kzalloc(vdata_size, GFP_KERNEL);
274	else {
275		vdata = vzalloc(vdata_size);
276		flags = VMD_VMALLOCED;
277	}
278	if (!vdata)
279		return -ENOMEM;
280
281	vdata->vm_start = vma->vm_start;
282	vdata->vm_end = vma->vm_end;
283	vdata->flags = flags;
284	vdata->type = type;
285	spin_lock_init(&vdata->lock);
286	atomic_set(&vdata->refcnt, 1);
287	vma->vm_private_data = vdata;
288
289	vma->vm_flags |= VM_IO | VM_PFNMAP | VM_DONTEXPAND | VM_DONTDUMP;
290	if (vdata->type == MSPEC_FETCHOP || vdata->type == MSPEC_UNCACHED)
291		vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
292	vma->vm_ops = &mspec_vm_ops;
293
294	return 0;
295}
296
297static int
298fetchop_mmap(struct file *file, struct vm_area_struct *vma)
299{
300	return mspec_mmap(file, vma, MSPEC_FETCHOP);
301}
302
303static int
304cached_mmap(struct file *file, struct vm_area_struct *vma)
305{
306	return mspec_mmap(file, vma, MSPEC_CACHED);
307}
308
309static int
310uncached_mmap(struct file *file, struct vm_area_struct *vma)
311{
312	return mspec_mmap(file, vma, MSPEC_UNCACHED);
313}
314
315static const struct file_operations fetchop_fops = {
316	.owner = THIS_MODULE,
317	.mmap = fetchop_mmap,
318	.llseek = noop_llseek,
319};
320
321static struct miscdevice fetchop_miscdev = {
322	.minor = MISC_DYNAMIC_MINOR,
323	.name = "sgi_fetchop",
324	.fops = &fetchop_fops
325};
326
327static const struct file_operations cached_fops = {
328	.owner = THIS_MODULE,
329	.mmap = cached_mmap,
330	.llseek = noop_llseek,
331};
332
333static struct miscdevice cached_miscdev = {
334	.minor = MISC_DYNAMIC_MINOR,
335	.name = "mspec_cached",
336	.fops = &cached_fops
337};
338
339static const struct file_operations uncached_fops = {
340	.owner = THIS_MODULE,
341	.mmap = uncached_mmap,
342	.llseek = noop_llseek,
343};
344
345static struct miscdevice uncached_miscdev = {
346	.minor = MISC_DYNAMIC_MINOR,
347	.name = "mspec_uncached",
348	.fops = &uncached_fops
349};
350
351/*
352 * mspec_init
353 *
354 * Called at boot time to initialize the mspec facility.
355 */
356static int __init
357mspec_init(void)
358{
359	int ret;
360	int nid;
361
362	/*
363	 * The fetchop device only works on SN2 hardware, uncached and cached
364	 * memory drivers should both be valid on all ia64 hardware
365	 */
366#ifdef CONFIG_SGI_SN
367	if (ia64_platform_is("sn2")) {
368		is_sn2 = 1;
369		if (is_shub2()) {
370			ret = -ENOMEM;
371			for_each_node_state(nid, N_ONLINE) {
372				int actual_nid;
373				int nasid;
374				unsigned long phys;
375
376				scratch_page[nid] = uncached_alloc_page(nid, 1);
377				if (scratch_page[nid] == 0)
378					goto free_scratch_pages;
379				phys = __pa(scratch_page[nid]);
380				nasid = get_node_number(phys);
381				actual_nid = nasid_to_cnodeid(nasid);
382				if (actual_nid != nid)
383					goto free_scratch_pages;
384			}
385		}
386
387		ret = misc_register(&fetchop_miscdev);
388		if (ret) {
389			printk(KERN_ERR
390			       "%s: failed to register device %i\n",
391			       FETCHOP_ID, ret);
392			goto free_scratch_pages;
393		}
394	}
395#endif
396	ret = misc_register(&cached_miscdev);
397	if (ret) {
398		printk(KERN_ERR "%s: failed to register device %i\n",
399		       CACHED_ID, ret);
400		if (is_sn2)
401			misc_deregister(&fetchop_miscdev);
402		goto free_scratch_pages;
403	}
404	ret = misc_register(&uncached_miscdev);
405	if (ret) {
406		printk(KERN_ERR "%s: failed to register device %i\n",
407		       UNCACHED_ID, ret);
408		misc_deregister(&cached_miscdev);
409		if (is_sn2)
410			misc_deregister(&fetchop_miscdev);
411		goto free_scratch_pages;
412	}
413
414	printk(KERN_INFO "%s %s initialized devices: %s %s %s\n",
415	       MSPEC_BASENAME, REVISION, is_sn2 ? FETCHOP_ID : "",
416	       CACHED_ID, UNCACHED_ID);
417
418	return 0;
419
420 free_scratch_pages:
421	for_each_node(nid) {
422		if (scratch_page[nid] != 0)
423			uncached_free_page(scratch_page[nid], 1);
424	}
425	return ret;
426}
427
428static void __exit
429mspec_exit(void)
430{
431	int nid;
432
433	misc_deregister(&uncached_miscdev);
434	misc_deregister(&cached_miscdev);
435	if (is_sn2) {
436		misc_deregister(&fetchop_miscdev);
437
438		for_each_node(nid) {
439			if (scratch_page[nid] != 0)
440				uncached_free_page(scratch_page[nid], 1);
441		}
442	}
443}
444
445module_init(mspec_init);
446module_exit(mspec_exit);
447
448MODULE_AUTHOR("Silicon Graphics, Inc. <linux-altix@sgi.com>");
449MODULE_DESCRIPTION("Driver for SGI SN special memory operations");
450MODULE_LICENSE("GPL");
451