xref: /titanic_50/usr/src/uts/i86pc/sys/rootnex.h (revision c2e7b48d563d0e56b74d853118918af352e75cbb)
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 #ifndef	_SYS_ROOTNEX_H
28 #define	_SYS_ROOTNEX_H
29 
30 #pragma ident	"%Z%%M%	%I%	%E% SMI"
31 
32 /*
33  * x86 root nexus implementation specific state
34  */
35 
36 #include <sys/types.h>
37 #include <sys/conf.h>
38 #include <sys/modctl.h>
39 #include <sys/sunddi.h>
40 
41 #ifdef	__cplusplus
42 extern "C" {
43 #endif
44 
45 
46 /* size of buffer used for ctlop reportdev */
47 #define	REPORTDEV_BUFSIZE	1024
48 
49 /* min and max interrupt vectors */
50 #define	VEC_MIN			1
51 #define	VEC_MAX			255
52 
53 /* atomic increment/decrement to keep track of outstanding binds, etc */
54 #define	ROOTNEX_PROF_INC(addr)		atomic_inc_64(addr)
55 #define	ROOTNEX_PROF_DEC(addr)		atomic_add_64(addr, -1)
56 
57 /* set in dmac_type to signify that this cookie uses the copy buffer */
58 #define	ROOTNEX_USES_COPYBUF		0x80000000
59 
60 /*
61  * integer or boolean property name and value. A few static rootnex properties
62  * are created during rootnex attach from an array of rootnex_intprop_t..
63  */
64 typedef struct rootnex_intprop_s {
65 	char	*prop_name;
66 	int	prop_value;
67 } rootnex_intprop_t;
68 
69 /*
70  * sgl related information which is visible to rootnex_get_sgl(). Trying to
71  * isolate get_sgl() as much as possible so it can be easily replaced.
72  */
73 typedef struct rootnex_sglinfo_s {
74 	/*
75 	 * These are passed into rootnex_get_sgl().
76 	 *
77 	 * si_min_addr - the minimum physical address
78 	 * si_max_addr - the maximum physical address
79 	 * si_max_cookie_size - the maximum size of a physically contiguous
80 	 *    piece of memory that we can handle in a sgl.
81 	 * si_segmask - segment mask to determine if we cross a segment boundary
82 	 * si_max_pages - max number of pages this sgl could occupy (which
83 	 *    is also the maximum number of cookies we might see.
84 	 */
85 	uint64_t	si_min_addr;
86 	uint64_t	si_max_addr;
87 	uint64_t	si_max_cookie_size;
88 	uint64_t	si_segmask;
89 	uint_t		si_max_pages;
90 
91 	/*
92 	 * these are returned by rootnex_get_sgl()
93 	 *
94 	 * si_copybuf_req - amount of copy buffer needed by the buffer.
95 	 * si_buf_offset - The initial offset into the first page of the buffer.
96 	 *    It's set in get sgl and used in the bind slow path to help
97 	 *    calculate the current page index & offset from the current offset
98 	 *    which is relative to the start of the buffer.
99 	 * si_asp - address space of buffer passed in.
100 	 * si_sgl_size - The actual number of cookies in the sgl. This does
101 	 *    not reflect and sharing that we might do on window boundaries.
102 	 */
103 	size_t		si_copybuf_req;
104 	off_t		si_buf_offset;
105 	struct as	*si_asp;
106 	uint_t		si_sgl_size;
107 } rootnex_sglinfo_t;
108 
109 /*
110  * When we have to use the copy buffer, we allocate one of these structures per
111  * buffer page to track which pages need the copy buffer, what the kernel
112  * virtual address is (which the device can't reach), and what the copy buffer
113  * virtual address is (where the device dma's to/from). For 32-bit kernels,
114  * since we can't use seg kpm, we also need to keep the page_t around and state
115  * if we've currently mapped in the page into KVA space for buffers which don't
116  * have kva already and when we have multiple windows because we used up all our
117  * copy buffer space.
118  */
119 typedef struct rootnex_pgmap_s {
120 	boolean_t	pm_uses_copybuf;
121 #if !defined(__amd64)
122 	boolean_t	pm_mapped;
123 	page_t		*pm_pp;
124 	caddr_t		pm_vaddr;
125 #endif
126 	caddr_t		pm_kaddr;
127 	caddr_t		pm_cbaddr;
128 } rootnex_pgmap_t;
129 
130 /*
131  * We only need to trim a buffer when we have multiple windows. Each window has
132  * trim state. We might have trimmed the end of the previous window, leaving the
133  * first cookie of this window trimmed[tr_trim_first] (which basically means we
134  * won't start with a new cookie), or we might need to trim the end of the
135  * current window [tr_trim_last] (which basically means we won't end with a
136  * complete cookie). We keep the same state for the first & last cookie in a
137  * window (a window can have one or more cookies). However, when we trim the
138  * last cookie, we keep a pointer to the last cookie in the trim state since we
139  * only need this info when we trim. The pointer to the first cookie in the
140  * window is in the window state since we need to know what the first cookie in
141  * the window is in various places.
142  *
143  * If we do trim a cookie, we save away the physical address and size of the
144  * cookie so that we can over write the cookie when we switch windows (the
145  * space for a cookie which is in two windows is shared between the windows.
146  * We keep around the same information for the last page in a window.
147  *
148  * if we happened to trim on a page that uses the copy buffer, and that page
149  * is also in the middle of a window boundary because we have filled up the
150  * copy buffer, we need to remember the copy buffer address for both windows
151  * since the same page will have different copy buffer addresses in the two
152  * windows. We need to due the same for kaddr in the 32-bit kernel since we
153  * have a limited kva space which we map to.
154  */
155 typedef struct rootnex_trim_s {
156 	boolean_t		tr_trim_first;
157 	boolean_t		tr_trim_last;
158 	ddi_dma_cookie_t	*tr_last_cookie;
159 	uint64_t		tr_first_paddr;
160 	uint64_t		tr_last_paddr;
161 	size_t			tr_first_size;
162 	size_t			tr_last_size;
163 
164 	boolean_t		tr_first_copybuf_win;
165 	boolean_t		tr_last_copybuf_win;
166 	uint_t			tr_first_pidx;
167 	uint_t			tr_last_pidx;
168 	caddr_t			tr_first_cbaddr;
169 	caddr_t			tr_last_cbaddr;
170 #if !defined(__amd64)
171 	caddr_t			tr_first_kaddr;
172 	caddr_t			tr_last_kaddr;
173 #endif
174 } rootnex_trim_t;
175 
176 /*
177  * per window state. A bound DMA handle can have multiple windows. Each window
178  * will have the following state. We track if this window needs to sync,
179  * the offset into the buffer where the window starts, the size of the window.
180  * a pointer to the first cookie in the window, the number of cookies in the
181  * window, and the trim state for the window. For the 32-bit kernel, we keep
182  * track of if we need to remap the copy buffer when we switch to a this window
183  */
184 typedef struct rootnex_window_s {
185 	boolean_t		wd_dosync;
186 	uint_t			wd_cookie_cnt;
187 	off_t			wd_offset;
188 	size_t			wd_size;
189 	ddi_dma_cookie_t	*wd_first_cookie;
190 	rootnex_trim_t		wd_trim;
191 #if !defined(__amd64)
192 	boolean_t		wd_remap_copybuf;
193 #endif
194 } rootnex_window_t;
195 
196 /* per dma handle private state */
197 typedef struct rootnex_dma_s {
198 	/*
199 	 * sgl related state used to build and describe the sgl.
200 	 *
201 	 * dp_partial_required - used in the bind slow path to identify if we
202 	 *    need to do a partial mapping or not.
203 	 * dp_trim_required - used in the bind slow path to identify if we
204 	 *    need to trim when switching to a new window. This should only be
205 	 *    set when partial is set.
206 	 * dp_granularity_power_2 - set in alloc handle and used in bind slow
207 	 *    path to determine if we & or % to calculate the trim.
208 	 * dp_dma - copy of dma "object" passed in during bind
209 	 * dp_maxxfer - trimmed dma_attr_maxxfer so that it is a whole
210 	 *    multiple of granularity
211 	 * dp_sglinfo - See rootnex_sglinfo_t above.
212 	 */
213 	boolean_t		dp_partial_required;
214 	boolean_t		dp_trim_required;
215 	boolean_t		dp_granularity_power_2;
216 	uint64_t		dp_maxxfer;
217 	ddi_dma_obj_t		dp_dma;
218 	rootnex_sglinfo_t	dp_sglinfo;
219 
220 	/*
221 	 * Copy buffer related state
222 	 *
223 	 * dp_copybuf_size - the actual size of the copy buffer that we are
224 	 *    using. This can be smaller that dp_copybuf_req, i.e. bind size >
225 	 *    max copy buffer size.
226 	 * dp_cbaddr - kernel address of copy buffer. Used to determine where
227 	 *    where to copy to/from.
228 	 * dp_cbsize - the "real" size returned from the copy buffer alloc.
229 	 *    Set in the copybuf alloc and used to free copybuf.
230 	 * dp_pgmap - page map used in sync to determine which pages in the
231 	 *    buffer use the copy buffer and what addresses to use to copy to/
232 	 *    from.
233 	 * dp_cb_remaping - status if this bind causes us to have to remap
234 	 *    the copybuf when switching to new windows. This is only used in
235 	 *    the 32-bit kernel since we use seg kpm in the 64-bit kernel for
236 	 *    this case.
237 	 * dp_kva - kernel heap arena vmem space for mapping to buffers which
238 	 *    we don't have a kernel VA to bcopy to/from. This is only used in
239 	 *    the 32-bit kernel since we use seg kpm in the 64-bit kernel for
240 	 *    this case.
241 	 */
242 	size_t			dp_copybuf_size;
243 	caddr_t			dp_cbaddr;
244 	size_t			dp_cbsize;
245 	rootnex_pgmap_t		*dp_pgmap;
246 #if !defined(__amd64)
247 	boolean_t		dp_cb_remaping;
248 	caddr_t			dp_kva;
249 #endif
250 
251 	/*
252 	 * window related state. The pointer to the window state array which may
253 	 * be a pointer into the pre allocated state, or we may have had to
254 	 * allocate the window array on the fly because it wouldn't fit. If
255 	 * we allocate it, we'll use dp_need_to_free_window and dp_window_size
256 	 * during cleanup. dp_current_win keeps track of the current window.
257 	 * dp_max_win is the maximum number of windows we could have.
258 	 */
259 	uint_t			dp_current_win;
260 	rootnex_window_t	*dp_window;
261 	boolean_t		dp_need_to_free_window;
262 	uint_t			dp_window_size;
263 	uint_t			dp_max_win;
264 
265 	/* dip of driver which "owns" handle. set to rdip in alloc_handle() */
266 	dev_info_t		*dp_dip;
267 
268 	/*
269 	 * dp_mutex and dp_inuse are only used to see if a driver is trying to
270 	 * bind to an already bound dma handle. dp_mutex only used for dp_inuse
271 	 */
272 	kmutex_t		dp_mutex;
273 	boolean_t		dp_inuse;
274 
275 	/*
276 	 * cookie related state. The pointer to the cookies (dp_cookies) may
277 	 * be a pointer into the pre allocated state, or we may have had to
278 	 * allocate the cookie array on the fly because it wouldn't fit. If
279 	 * we allocate it, we'll use dp_need_to_free_cookie and dp_cookie_size
280 	 * during cleanup. dp_current_cookie is only used in the obsoleted
281 	 * interfaces to determine when we've used up all the cookies in a
282 	 * window during nextseg()..
283 	 */
284 	size_t			dp_cookie_size;
285 	ddi_dma_cookie_t	*dp_cookies;
286 	boolean_t		dp_need_to_free_cookie;
287 	uint_t			dp_current_cookie; /* for obsoleted I/Fs */
288 
289 	/*
290 	 * pre allocated space for the bind state, allocated during alloc
291 	 * handle. For a lot of devices, this will save us from having to do
292 	 * kmem_alloc's during the bind most of the time. kmem_alloc's can be
293 	 * expensive on x86 when the cpu count goes up since xcalls are
294 	 * expensive on x86.
295 	 */
296 	uchar_t			*dp_prealloc_buffer;
297 } rootnex_dma_t;
298 
299 /*
300  * profile/performance counters. Most things will be dtrace probes, but there
301  * are a couple of things we want to keep track all the time. We track the
302  * total number of active handles and binds (i.e. an alloc without a free or
303  * a bind without an unbind) since rootnex attach. We also track the total
304  * number of binds which have failed since rootnex attach.
305  */
306 typedef enum {
307 	ROOTNEX_CNT_ACTIVE_HDLS = 0,
308 	ROOTNEX_CNT_ACTIVE_BINDS = 1,
309 	ROOTNEX_CNT_ALLOC_FAIL = 2,
310 	ROOTNEX_CNT_BIND_FAIL = 3,
311 	ROOTNEX_CNT_SYNC_FAIL = 4,
312 	ROOTNEX_CNT_GETWIN_FAIL = 5,
313 
314 	/* This one must be last */
315 	ROOTNEX_CNT_LAST
316 } rootnex_cnt_t;
317 
318 /*
319  * global driver state.
320  *   r_dmahdl_cache - dma_handle kmem_cache
321  *   r_dvma_call_list_id - ddi_set_callback() id
322  *   r_peekpoke_mutex - serialize peeks and pokes.
323  *   r_dip - rootnex dip
324  *   r_reserved_msg_printed - ctlops reserve message threshold
325  *   r_counters - profile/performance counters
326  */
327 typedef struct rootnex_state_s {
328 	uint_t		r_prealloc_cookies;
329 	uint_t		r_prealloc_size;
330 	kmem_cache_t	*r_dmahdl_cache;
331 	uintptr_t	r_dvma_call_list_id;
332 	kmutex_t	r_peekpoke_mutex;
333 	dev_info_t	*r_dip;
334 	boolean_t	r_reserved_msg_printed;
335 	uint64_t	r_counters[ROOTNEX_CNT_LAST];
336 } rootnex_state_t;
337 
338 
339 #ifdef	__cplusplus
340 }
341 #endif
342 
343 #endif	/* _SYS_ROOTNEX_H */
344