xref: /titanic_41/usr/src/uts/common/vm/seg_kpm.c (revision 8eea8e29cc4374d1ee24c25a07f45af132db3499)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License, Version 1.0 only
6  * (the "License").  You may not use this file except in compliance
7  * with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 /*
30  * Kernel Physical Mapping (kpm) segment driver (segkpm).
31  *
32  * This driver delivers along with the hat_kpm* interfaces an alternative
33  * mechanism for kernel mappings within the 64-bit Solaris operating system,
34  * which allows the mapping of all physical memory into the kernel address
35  * space at once. This is feasible in 64 bit kernels, e.g. for Ultrasparc II
36  * and beyond processors, since the available VA range is much larger than
37  * possible physical memory. Momentarily all physical memory is supported,
38  * that is represented by the list of memory segments (memsegs).
39  *
40  * Segkpm mappings have also very low overhead and large pages are used
41  * (when possible) to minimize the TLB and TSB footprint. It is also
42  * extentable for other than Sparc architectures (e.g. AMD64). Main
43  * advantage is the avoidance of the TLB-shootdown X-calls, which are
44  * normally needed when a kernel (global) mapping has to be removed.
45  *
46  * First example of a kernel facility that uses the segkpm mapping scheme
47  * is seg_map, where it is used as an alternative to hat_memload().
48  * See also hat layer for more information about the hat_kpm* routines.
49  * The kpm facilty can be turned off at boot time (e.g. /etc/system).
50  */
51 
52 #include <sys/types.h>
53 #include <sys/param.h>
54 #include <sys/sysmacros.h>
55 #include <sys/systm.h>
56 #include <sys/vnode.h>
57 #include <sys/cmn_err.h>
58 #include <sys/debug.h>
59 #include <sys/thread.h>
60 #include <sys/cpuvar.h>
61 #include <sys/bitmap.h>
62 #include <sys/atomic.h>
63 
64 #include <vm/seg_kmem.h>
65 #include <vm/seg_kpm.h>
66 #include <vm/hat.h>
67 #include <vm/as.h>
68 #include <vm/seg.h>
69 #include <vm/page.h>
70 
71 /*
72  * Global kpm controls.
73  * See also platform and mmu specific controls.
74  *
75  * kpm_enable -- global on/off switch for segkpm.
76  * . Set by default on 64bit platforms that have kpm support.
77  * . Will be disabled from platform layer if not supported.
78  * . Can be disabled via /etc/system.
79  *
80  * kpm_smallpages -- use only regular/system pagesize for kpm mappings.
81  * . Can be useful for critical debugging of kpm clients.
82  * . Set to zero by default for platforms that support kpm large pages.
83  *   The use of kpm large pages reduces the footprint of kpm meta data
84  *   and has all the other advantages of using large pages (e.g TLB
85  *   miss reduction).
86  * . Set by default for platforms that don't support kpm large pages or
87  *   where large pages cannot be used for other reasons (e.g. there are
88  *   only few full associative TLB entries available for large pages).
89  *
90  * segmap_kpm -- separate on/off switch for segmap using segkpm:
91  * . Set by default.
92  * . Will be disabled when kpm_enable is zero.
93  * . Will be disabled when MAXBSIZE != PAGESIZE.
94  * . Can be disabled via /etc/system.
95  *
96  */
97 int kpm_enable = 1;
98 int kpm_smallpages = 0;
99 int segmap_kpm = 1;
100 
101 /*
102  * Private seg op routines.
103  */
104 faultcode_t segkpm_fault(struct hat *hat, struct seg *seg, caddr_t addr,
105 			size_t len, enum fault_type type, enum seg_rw rw);
106 static void	segkpm_dump(struct seg *);
107 static void	segkpm_badop(void);
108 static int	segkpm_notsup(void);
109 
110 #define	SEGKPM_BADOP(t)	(t(*)())segkpm_badop
111 #define	SEGKPM_NOTSUP	(int(*)())segkpm_notsup
112 
113 static struct seg_ops segkpm_ops = {
114 	SEGKPM_BADOP(int),	/* dup */
115 	SEGKPM_BADOP(int),	/* unmap */
116 	SEGKPM_BADOP(void),	/* free */
117 	segkpm_fault,
118 	SEGKPM_BADOP(int),	/* faulta */
119 	SEGKPM_BADOP(int),	/* setprot */
120 	SEGKPM_BADOP(int),	/* checkprot */
121 	SEGKPM_BADOP(int),	/* kluster */
122 	SEGKPM_BADOP(size_t),	/* swapout */
123 	SEGKPM_BADOP(int),	/* sync */
124 	SEGKPM_BADOP(size_t),	/* incore */
125 	SEGKPM_BADOP(int),	/* lockop */
126 	SEGKPM_BADOP(int),	/* getprot */
127 	SEGKPM_BADOP(u_offset_t), /* getoffset */
128 	SEGKPM_BADOP(int),	/* gettype */
129 	SEGKPM_BADOP(int),	/* getvp */
130 	SEGKPM_BADOP(int),	/* advise */
131 	segkpm_dump,		/* dump */
132 	SEGKPM_NOTSUP,		/* pagelock */
133 	SEGKPM_BADOP(int),	/* setpgsz */
134 	SEGKPM_BADOP(int),	/* getmemid */
135 };
136 
137 /*
138  * kpm_pgsz and kpm_pgshft are set by platform layer.
139  */
140 size_t		kpm_pgsz;	/* kpm page size */
141 uint_t		kpm_pgshft;	/* kpm page shift */
142 u_offset_t	kpm_pgoff;	/* kpm page offset mask */
143 uint_t		kpmp2pshft;	/* kpm page to page shift */
144 pgcnt_t		kpmpnpgs;	/* how many pages per kpm page */
145 
146 
147 #ifdef	SEGKPM_SUPPORT
148 
149 int
150 segkpm_create(struct seg *seg, void *argsp)
151 {
152 	struct segkpm_data *skd;
153 	struct segkpm_crargs *b = (struct segkpm_crargs *)argsp;
154 	ushort_t *p;
155 	int i, j;
156 
157 	ASSERT(seg->s_as && RW_WRITE_HELD(&seg->s_as->a_lock));
158 	ASSERT(btokpmp(seg->s_size) >= 1 &&
159 		kpmpageoff((uintptr_t)seg->s_base) == 0 &&
160 		kpmpageoff((uintptr_t)seg->s_base + seg->s_size) == 0);
161 
162 	skd = kmem_zalloc(sizeof (struct segkpm_data), KM_SLEEP);
163 
164 	seg->s_data = (void *)skd;
165 	seg->s_ops = &segkpm_ops;
166 	skd->skd_prot = b->prot;
167 
168 	/*
169 	 * (1) Segkpm virtual addresses are based on physical adresses.
170 	 * From this and in opposite to other segment drivers it is
171 	 * often required to allocate a page first to be able to
172 	 * calculate the final segkpm virtual address.
173 	 * (2) Page  allocation is done by calling page_create_va(),
174 	 * one important input argument is a virtual address (also
175 	 * expressed by the "va" in the function name). This function
176 	 * is highly optimized to select the right page for an optimal
177 	 * processor and platform support (e.g. virtual addressed
178 	 * caches (VAC), physical addressed caches, NUMA).
179 	 *
180 	 * Because of (1) the approach is to generate a faked virtual
181 	 * address for calling page_create_va(). In order to exploit
182 	 * the abilities of (2), especially to utilize the cache
183 	 * hierarchy (3) and to avoid VAC alias conflicts (4) the
184 	 * selection has to be done carefully. For each virtual color
185 	 * a separate counter is provided (4). The count values are
186 	 * used for the utilization of all cache lines (3) and are
187 	 * corresponding to the cache bins.
188 	 */
189 	skd->skd_nvcolors = b->nvcolors;
190 
191 	p = skd->skd_va_select =
192 		kmem_zalloc(NCPU * b->nvcolors * sizeof (ushort_t), KM_SLEEP);
193 
194 	for (i = 0; i < NCPU; i++)
195 		for (j = 0; j < b->nvcolors; j++, p++)
196 			*p = j;
197 
198 	return (0);
199 }
200 
201 /*
202  * This routine is called via a machine specific fault handling
203  * routine.
204  */
205 /* ARGSUSED */
206 faultcode_t
207 segkpm_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len,
208 	enum fault_type type, enum seg_rw rw)
209 {
210 	faultcode_t error;
211 
212 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
213 
214 	error = hat_kpm_fault(hat, addr);
215 
216 	return (error);
217 }
218 
219 #define	addr_to_vcolor(addr, vcolors) \
220 	((int)(((uintptr_t)(addr) & ((vcolors << PAGESHIFT) - 1)) >> PAGESHIFT))
221 
222 /*
223  * Create a virtual address that can be used for invocations of
224  * page_create_va. Goal is to utilize the cache hierarchy (round
225  * robin bins) and to select the right color for virtual indexed
226  * caches. It isn't exact since we also increment the bin counter
227  * when the caller uses VOP_GETPAGE and gets a hit in the page
228  * cache, but we keep the bins turning for cache distribution
229  * (see also segkpm_create block comment).
230  */
231 caddr_t
232 segkpm_create_va(u_offset_t off)
233 {
234 	int vcolor;
235 	ushort_t *p;
236 	struct segkpm_data *skd = (struct segkpm_data *)segkpm->s_data;
237 	int nvcolors = skd->skd_nvcolors;
238 	caddr_t	va;
239 
240 	vcolor = (nvcolors > 1) ? addr_to_vcolor(off, nvcolors) : 0;
241 	p = &skd->skd_va_select[(CPU->cpu_id * nvcolors) + vcolor];
242 	va = (caddr_t)ptob(*p);
243 
244 	atomic_add_16(p, nvcolors);
245 
246 	return (va);
247 }
248 
249 /*
250  * Unload mapping if the instance has an active kpm mapping.
251  */
252 void
253 segkpm_mapout_validkpme(struct kpme *kpme)
254 {
255 	caddr_t vaddr;
256 	page_t *pp;
257 
258 retry:
259 	if ((pp = kpme->kpe_page) == NULL) {
260 		return;
261 	}
262 
263 	if (page_lock(pp, SE_SHARED, (kmutex_t *)NULL, P_RECLAIM) == 0)
264 		goto retry;
265 
266 	/*
267 	 * Check if segkpm mapping is not unloaded in the meantime
268 	 */
269 	if (kpme->kpe_page == NULL) {
270 		page_unlock(pp);
271 		return;
272 	}
273 
274 	vaddr = hat_kpm_page2va(pp, 1);
275 	hat_kpm_mapout(pp, kpme, vaddr);
276 	page_unlock(pp);
277 }
278 
279 static void
280 segkpm_badop()
281 {
282 	panic("segkpm_badop");
283 }
284 
285 #else	/* SEGKPM_SUPPORT */
286 
287 /* segkpm stubs */
288 
289 /*ARGSUSED*/
290 int segkpm_create(struct seg *seg, void *argsp) { return (0); }
291 
292 /* ARGSUSED */
293 faultcode_t
294 segkpm_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len,
295 	enum fault_type type, enum seg_rw rw)
296 {
297 	return ((faultcode_t)0);
298 }
299 
300 /* ARGSUSED */
301 caddr_t segkpm_create_va(u_offset_t off) { return (NULL); }
302 
303 /* ARGSUSED */
304 void segkpm_mapout_validkpme(struct kpme *kpme) {}
305 
306 static void
307 segkpm_badop() {}
308 
309 #endif	/* SEGKPM_SUPPORT */
310 
311 static int
312 segkpm_notsup()
313 {
314 	return (ENOTSUP);
315 }
316 
317 /*
318  * segkpm pages are not dumped, so we just return
319  */
320 /*ARGSUSED*/
321 static void
322 segkpm_dump(struct seg *seg)
323 {}
324