Go to most recent revision | Details | Last modification | View Log | RSS feed
Rev | Author | Line No. | Line |
---|---|---|---|
5201 | serge | 1 | ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; |
2 | ;; ;; |
||
3 | ;; Copyright (C) KolibriOS team 2004-2014. All rights reserved. ;; |
||
4 | ;; Distributed under terms of the GNU General Public License ;; |
||
5 | ;; ;; |
||
6 | ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;; |
||
7 | |||
8 | $Revision: 5130 $ |
||
9 | |||
10 | ; Initializes MTRRs. |
||
11 | proc init_mtrr |
||
12 | |||
13 | cmp [BOOT_VARS+BOOT_MTRR], byte 2 |
||
14 | je .exit |
||
15 | |||
16 | bt [cpu_caps], CAPS_MTRR |
||
17 | jnc .exit |
||
18 | |||
19 | call mtrr_reconfigure |
||
20 | stdcall set_mtrr, [LFBAddress], 0x1000000, MEM_WC |
||
21 | |||
22 | .exit: |
||
23 | ret |
||
24 | endp |
||
25 | |||
26 | ; Helper procedure for mtrr_reconfigure and set_mtrr, |
||
27 | ; called before changes in MTRRs. |
||
28 | proc mtrr_begin_change |
||
29 | mov eax, cr0 |
||
30 | or eax, 0x60000000 ;disable caching |
||
31 | mov cr0, eax |
||
32 | wbinvd ;invalidate cache |
||
33 | ret |
||
34 | endp |
||
35 | |||
36 | ; Helper procedure for mtrr_reconfigure and set_mtrr, |
||
37 | ; called after changes in MTRRs. |
||
38 | proc mtrr_end_change |
||
39 | wbinvd ;again invalidate |
||
40 | mov eax, cr0 |
||
41 | and eax, not 0x60000000 |
||
42 | mov cr0, eax ; enable caching |
||
43 | ret |
||
44 | endp |
||
45 | |||
46 | ; Some limits to number of structures located in the stack. |
||
47 | MAX_USEFUL_MTRRS = 16 |
||
48 | MAX_RANGES = 16 |
||
49 | |||
50 | ; mtrr_reconfigure keeps a list of MEM_WB ranges. |
||
51 | ; This structure describes one item in the list. |
||
52 | struct mtrr_range |
||
53 | next dd ? ; next item |
||
54 | start dq ? ; first byte |
||
55 | length dq ? ; length in bytes |
||
56 | ends |
||
57 | |||
58 | uglobal |
||
59 | align 4 |
||
60 | num_variable_mtrrs dd 0 ; number of variable-range MTRRs |
||
61 | endg |
||
62 | |||
63 | ; Helper procedure for MTRR initialization. |
||
64 | ; Takes MTRR configured by BIOS and tries to recongifure them |
||
65 | ; in order to allow non-UC data at top of 4G memory. |
||
66 | ; Example: if low part of physical memory is 3.5G = 0xE0000000 bytes wide, |
||
67 | ; BIOS can configure two MTRRs so that the first MTRR describes [0, 4G) as WB |
||
68 | ; and the second MTRR describes [3.5G, 4G) as UC; |
||
69 | ; WB+UC=UC, so the resulting memory map would be as needed, |
||
70 | ; but in this configuration our attempts to map LFB at (say) 0xE8000000 as WC |
||
71 | ; would be ignored, WB+UC+WC is still UC. |
||
72 | ; So we must keep top of 4G memory not covered by MTRRs, |
||
73 | ; using three WB MTRRs [0,2G) + [2G,3G) + [3G,3.5G), |
||
74 | ; this gives the same memory map, but allows to add further entries. |
||
75 | ; See mtrrtest.asm for detailed input/output from real hardware+BIOS. |
||
76 | proc mtrr_reconfigure |
||
77 | push ebp ; we're called from init_LFB, and it feels hurt when ebp is destroyed |
||
78 | ; 1. Prepare local variables. |
||
79 | ; 1a. Create list of MAX_RANGES free (aka not yet allocated) ranges. |
||
80 | xor eax, eax |
||
81 | lea ecx, [eax+MAX_RANGES] |
||
82 | .init_ranges: |
||
83 | sub esp, sizeof.mtrr_range - 4 |
||
84 | push eax |
||
85 | mov eax, esp |
||
86 | dec ecx |
||
87 | jnz .init_ranges |
||
88 | mov eax, esp |
||
89 | ; 1b. Fill individual local variables. |
||
90 | xor edx, edx |
||
91 | sub esp, MAX_USEFUL_MTRRS * 16 ; .mtrrs |
||
92 | push edx ; .mtrrs_end |
||
93 | push edx ; .num_used_mtrrs |
||
94 | push eax ; .first_free_range |
||
95 | push edx ; .first_range: no ranges yet |
||
96 | mov cl, [cpu_phys_addr_width] |
||
97 | or eax, -1 |
||
98 | shl eax, cl ; note: this uses cl&31 = cl-32, not the entire cl |
||
99 | push eax ; .phys_reserved_mask |
||
100 | virtual at esp |
||
101 | .phys_reserved_mask dd ? |
||
102 | .first_range dd ? |
||
103 | .first_free_range dd ? |
||
104 | .num_used_mtrrs dd ? |
||
105 | .mtrrs_end dd ? |
||
106 | .mtrrs rq MAX_USEFUL_MTRRS * 2 |
||
107 | .local_vars_size = $ - esp |
||
108 | end virtual |
||
109 | |||
110 | ; 2. Get the number of variable-range MTRRs from MTRRCAP register. |
||
111 | ; Abort if zero. |
||
112 | mov ecx, 0xFE |
||
113 | rdmsr |
||
114 | test al, al |
||
115 | jz .abort |
||
116 | mov byte [num_variable_mtrrs], al |
||
117 | ; 3. Validate MTRR_DEF_TYPE register. |
||
118 | mov ecx, 0x2FF |
||
119 | rdmsr |
||
120 | ; If BIOS has not initialized variable-range MTRRs, fallback to step 7. |
||
121 | test ah, 8 |
||
122 | jz .fill_ranges_from_memory_map |
||
123 | ; If the default memory type (not covered by MTRRs) is not UC, |
||
124 | ; then probably BIOS did something strange, so it is better to exit immediately |
||
125 | ; hoping for the best. |
||
126 | cmp al, MEM_UC |
||
127 | jnz .abort |
||
128 | ; 4. Validate all variable-range MTRRs |
||
129 | ; and copy configured MTRRs to the local array [.mtrrs]. |
||
130 | ; 4a. Prepare for the loop over existing variable-range MTRRs. |
||
131 | mov ecx, 0x200 |
||
132 | lea edi, [.mtrrs] |
||
133 | .get_used_mtrrs_loop: |
||
134 | ; 4b. For every MTRR, read PHYSBASEn and PHYSMASKn. |
||
135 | ; In PHYSBASEn, clear upper bits and copy to ebp:ebx. |
||
136 | rdmsr |
||
137 | or edx, [.phys_reserved_mask] |
||
138 | xor edx, [.phys_reserved_mask] |
||
139 | mov ebp, edx |
||
140 | mov ebx, eax |
||
141 | inc ecx |
||
142 | ; If PHYSMASKn is not active, ignore this MTRR. |
||
143 | rdmsr |
||
144 | inc ecx |
||
145 | test ah, 8 |
||
146 | jz .get_used_mtrrs_next |
||
147 | ; 4c. For every active MTRR, check that number of local entries is not too large. |
||
148 | inc [.num_used_mtrrs] |
||
149 | cmp [.num_used_mtrrs], MAX_USEFUL_MTRRS |
||
150 | ja .abort |
||
151 | ; 4d. For every active MTRR, store PHYSBASEn with upper bits cleared. |
||
152 | ; This contains the MTRR base and the memory type in low byte. |
||
153 | mov [edi], ebx |
||
154 | mov [edi+4], ebp |
||
155 | ; 4e. For every active MTRR, check that the range is continuous: |
||
156 | ; PHYSMASKn with upper bits set must be negated power of two, and |
||
157 | ; low bits of PHYSBASEn must be zeroes: |
||
158 | ; PHYSMASKn = 1...10...0, |
||
159 | ; PHYSBASEn = x...x0...0, |
||
160 | ; this defines a continuous range from x...x0...0 to x...x1...1, |
||
161 | ; length = 10...0 = negated PHYSMASKn. |
||
162 | ; Store length in the local array. |
||
163 | and eax, not 0xFFF |
||
164 | or edx, [.phys_reserved_mask] |
||
165 | mov dword [edi+8], 0 |
||
166 | mov dword [edi+12], 0 |
||
167 | sub [edi+8], eax |
||
168 | sbb [edi+12], edx |
||
169 | ; (x and -x) is the maximum power of two that divides x. |
||
170 | ; Condition for powers of two: (x and -x) equals x. |
||
171 | and eax, [edi+8] |
||
172 | and edx, [edi+12] |
||
173 | cmp eax, [edi+8] |
||
174 | jnz .abort |
||
175 | cmp edx, [edi+12] |
||
176 | jnz .abort |
||
177 | sub eax, 1 |
||
178 | sbb edx, 0 |
||
179 | and eax, not 0xFFF |
||
180 | and eax, ebx |
||
181 | jnz .abort |
||
182 | and edx, ebp |
||
183 | jnz .abort |
||
184 | ; 4f. For every active MTRR, validate memory type: it must be either WB or UC. |
||
185 | add edi, 16 |
||
186 | cmp bl, MEM_UC |
||
187 | jz .get_used_mtrrs_next |
||
188 | cmp bl, MEM_WB |
||
189 | jnz .abort |
||
190 | .get_used_mtrrs_next: |
||
191 | ; 4g. Repeat the loop at 4b-4f for all [num_variable_mtrrs] entries. |
||
192 | mov eax, [num_variable_mtrrs] |
||
193 | lea eax, [0x200+eax*2] |
||
194 | cmp ecx, eax |
||
195 | jb .get_used_mtrrs_loop |
||
196 | ; 4h. If no active MTRRs were detected, fallback to step 7. |
||
197 | cmp [.num_used_mtrrs], 0 |
||
198 | jz .fill_ranges_from_memory_map |
||
199 | mov [.mtrrs_end], edi |
||
200 | ; 5. Generate sorted list of ranges marked as WB. |
||
201 | ; 5a. Prepare for the loop over configured MTRRs filled at step 4. |
||
202 | lea ecx, [.mtrrs] |
||
203 | .fill_wb_ranges: |
||
204 | ; 5b. Ignore non-WB MTRRs. |
||
205 | mov ebx, [ecx] |
||
206 | cmp bl, MEM_WB |
||
207 | jnz .next_wb_range |
||
208 | mov ebp, [ecx+4] |
||
209 | and ebx, not 0xFFF ; clear memory type and reserved bits |
||
210 | ; ebp:ebx = start of the range described by the current MTRR. |
||
211 | ; 5c. Find the first existing range containing a point greater than ebp:ebx. |
||
212 | lea esi, [.first_range] |
||
213 | .find_range_wb: |
||
214 | ; If there is no next range or start of the next range is greater than ebp:ebx, |
||
215 | ; exit the loop to 5d. |
||
216 | mov edi, [esi] |
||
217 | test edi, edi |
||
218 | jz .found_place_wb |
||
219 | mov eax, ebx |
||
220 | mov edx, ebp |
||
221 | sub eax, dword [edi+mtrr_range.start] |
||
222 | sbb edx, dword [edi+mtrr_range.start+4] |
||
223 | jb .found_place_wb |
||
224 | ; Otherwise, if end of the next range is greater than or equal to ebp:ebx, |
||
225 | ; exit the loop to 5e. |
||
226 | mov esi, edi |
||
227 | sub eax, dword [edi+mtrr_range.length] |
||
228 | sbb edx, dword [edi+mtrr_range.length+4] |
||
229 | jb .expand_wb |
||
230 | or eax, edx |
||
231 | jnz .find_range_wb |
||
232 | jmp .expand_wb |
||
233 | .found_place_wb: |
||
234 | ; 5d. ebp:ebx is not within any existing range. |
||
235 | ; Insert a new range between esi and edi. |
||
236 | ; (Later, during 5e, it can be merged with the following ranges.) |
||
237 | mov eax, [.first_free_range] |
||
238 | test eax, eax |
||
239 | jz .abort |
||
240 | mov [esi], eax |
||
241 | mov edx, [eax+mtrr_range.next] |
||
242 | mov [.first_free_range], edx |
||
243 | mov dword [eax+mtrr_range.start], ebx |
||
244 | mov dword [eax+mtrr_range.start+4], ebp |
||
245 | ; Don't fill [eax+mtrr_range.next] and [eax+mtrr_range.length] yet, |
||
246 | ; they will be calculated including merges at step 5e. |
||
247 | mov esi, edi |
||
248 | mov edi, eax |
||
249 | .expand_wb: |
||
250 | ; 5e. The range at edi contains ebp:ebx, and esi points to the first range |
||
251 | ; to be checked for merge: esi=edi if ebp:ebx was found in an existing range, |
||
252 | ; esi is next after edi if a new range with ebp:ebx was created. |
||
253 | ; Merge it with following ranges while start of the next range is not greater |
||
254 | ; than the end of the new range. |
||
255 | add ebx, [ecx+8] |
||
256 | adc ebp, [ecx+12] |
||
257 | ; ebp:ebx = end of the range described by the current MTRR. |
||
258 | .expand_wb_loop: |
||
259 | ; If there is no next range or start of the next range is greater than ebp:ebx, |
||
260 | ; exit the loop to 5g. |
||
261 | test esi, esi |
||
262 | jz .expand_wb_done |
||
263 | mov eax, ebx |
||
264 | mov edx, ebp |
||
265 | sub eax, dword [esi+mtrr_range.start] |
||
266 | sbb edx, dword [esi+mtrr_range.start+4] |
||
267 | jb .expand_wb_done |
||
268 | ; Otherwise, if end of the next range is greater than or equal to ebp:ebx, |
||
269 | ; exit the loop to 5f. |
||
270 | sub eax, dword [esi+mtrr_range.length] |
||
271 | sbb edx, dword [esi+mtrr_range.length+4] |
||
272 | jb .expand_wb_last |
||
273 | ; Otherwise, the current range is completely within the new range. |
||
274 | ; Free it and continue the loop. |
||
275 | mov edx, [esi+mtrr_range.next] |
||
276 | cmp esi, edi |
||
277 | jz @f |
||
278 | mov eax, [.first_free_range] |
||
279 | mov [esi+mtrr_range.next], eax |
||
280 | mov [.first_free_range], esi |
||
281 | @@: |
||
282 | mov esi, edx |
||
283 | jmp .expand_wb_loop |
||
284 | .expand_wb_last: |
||
285 | ; 5f. Start of the new range is inside range described by esi, |
||
286 | ; end of the new range is inside range described by edi. |
||
287 | ; If esi is equal to edi, the new range is completely within |
||
288 | ; an existing range, so proceed to the next range. |
||
289 | cmp esi, edi |
||
290 | jz .next_wb_range |
||
291 | ; Otherwise, set end of interval at esi to end of interval at edi |
||
292 | ; and free range described by edi. |
||
293 | mov ebx, dword [esi+mtrr_range.start] |
||
294 | mov ebp, dword [esi+mtrr_range.start+4] |
||
295 | add ebx, dword [esi+mtrr_range.length] |
||
296 | adc ebp, dword [esi+mtrr_range.length+4] |
||
297 | mov edx, [esi+mtrr_range.next] |
||
298 | mov eax, [.first_free_range] |
||
299 | mov [esi+mtrr_range.next], eax |
||
300 | mov [.first_free_range], esi |
||
301 | mov esi, edx |
||
302 | .expand_wb_done: |
||
303 | ; 5g. We have found the next range (maybe 0) after merging and |
||
304 | ; the new end of range (maybe ebp:ebx from the new range |
||
305 | ; or end of another existing interval calculated at step 5f). |
||
306 | ; Write them to range at edi. |
||
307 | mov [edi+mtrr_range.next], esi |
||
308 | sub ebx, dword [edi+mtrr_range.start] |
||
309 | sbb ebp, dword [edi+mtrr_range.start+4] |
||
310 | mov dword [edi+mtrr_range.length], ebx |
||
311 | mov dword [edi+mtrr_range.length+4], ebp |
||
312 | .next_wb_range: |
||
313 | ; 5h. Continue the loop 5b-5g over all configured MTRRs. |
||
314 | add ecx, 16 |
||
315 | cmp ecx, [.mtrrs_end] |
||
316 | jb .fill_wb_ranges |
||
317 | ; 6. Exclude all ranges marked as UC. |
||
318 | ; 6a. Prepare for the loop over configured MTRRs filled at step 4. |
||
319 | lea ecx, [.mtrrs] |
||
320 | .fill_uc_ranges: |
||
321 | ; 6b. Ignore non-UC MTRRs. |
||
322 | mov ebx, [ecx] |
||
323 | cmp bl, MEM_UC |
||
324 | jnz .next_uc_range |
||
325 | mov ebp, [ecx+4] |
||
326 | and ebx, not 0xFFF ; clear memory type and reserved bits |
||
327 | ; ebp:ebx = start of the range described by the current MTRR. |
||
328 | lea esi, [.first_range] |
||
329 | ; 6c. Find the first existing range containing a point greater than ebp:ebx. |
||
330 | .find_range_uc: |
||
331 | ; If there is no next range, ignore this MTRR, |
||
332 | ; exit the loop and continue to next MTRR. |
||
333 | mov edi, [esi] |
||
334 | test edi, edi |
||
335 | jz .next_uc_range |
||
336 | ; If start of the next range is greater than or equal to ebp:ebx, |
||
337 | ; exit the loop to 6e. |
||
338 | mov eax, dword [edi+mtrr_range.start] |
||
339 | mov edx, dword [edi+mtrr_range.start+4] |
||
340 | sub eax, ebx |
||
341 | sbb edx, ebp |
||
342 | jnb .truncate_uc |
||
343 | ; Otherwise, continue the loop if end of the next range is less than ebp:ebx, |
||
344 | ; exit the loop to 6d otherwise. |
||
345 | mov esi, edi |
||
346 | add eax, dword [edi+mtrr_range.length] |
||
347 | adc edx, dword [edi+mtrr_range.length+4] |
||
348 | jnb .find_range_uc |
||
349 | ; 6d. ebp:ebx is inside (or at end of) an existing range. |
||
350 | ; Split the range. (The second range, maybe containing completely within UC-range, |
||
351 | ; maybe of zero length, can be removed at step 6e, if needed.) |
||
352 | mov edi, [.first_free_range] |
||
353 | test edi, edi |
||
354 | jz .abort |
||
355 | mov dword [edi+mtrr_range.start], ebx |
||
356 | mov dword [edi+mtrr_range.start+4], ebp |
||
357 | mov dword [edi+mtrr_range.length], eax |
||
358 | mov dword [edi+mtrr_range.length+4], edx |
||
359 | mov eax, [edi+mtrr_range.next] |
||
360 | mov [.first_free_range], eax |
||
361 | mov eax, [esi+mtrr_range.next] |
||
362 | mov [edi+mtrr_range.next], eax |
||
363 | ; don't change [esi+mtrr_range.next] yet, it will be filled at step 6e |
||
364 | mov eax, ebx |
||
365 | mov edx, ebp |
||
366 | sub eax, dword [esi+mtrr_range.start] |
||
367 | sbb edx, dword [esi+mtrr_range.start+4] |
||
368 | mov dword [esi+mtrr_range.length], eax |
||
369 | mov dword [esi+mtrr_range.length+4], edx |
||
370 | .truncate_uc: |
||
371 | ; 6e. edi is the first range after ebp:ebx, check it and next ranges |
||
372 | ; for intersection with the new range, truncate heads. |
||
373 | add ebx, [ecx+8] |
||
374 | adc ebp, [ecx+12] |
||
375 | ; ebp:ebx = end of the range described by the current MTRR. |
||
376 | .truncate_uc_loop: |
||
377 | ; If start of the next range is greater than ebp:ebx, |
||
378 | ; exit the loop to 6g. |
||
379 | mov eax, ebx |
||
380 | mov edx, ebp |
||
381 | sub eax, dword [edi+mtrr_range.start] |
||
382 | sbb edx, dword [edi+mtrr_range.start+4] |
||
383 | jb .truncate_uc_done |
||
384 | ; Otherwise, if end of the next range is greater than ebp:ebx, |
||
385 | ; exit the loop to 6f. |
||
386 | sub eax, dword [edi+mtrr_range.length] |
||
387 | sbb edx, dword [edi+mtrr_range.length+4] |
||
388 | jb .truncate_uc_last |
||
389 | ; Otherwise, the current range is completely within the new range. |
||
390 | ; Free it and continue the loop if there is a next range. |
||
391 | ; If that was a last range, exit the loop to 6g. |
||
392 | mov edx, [edi+mtrr_range.next] |
||
393 | mov eax, [.first_free_range] |
||
394 | mov [.first_free_range], edi |
||
395 | mov [edi+mtrr_range.next], eax |
||
396 | mov edi, edx |
||
397 | test edi, edi |
||
398 | jnz .truncate_uc_loop |
||
399 | jmp .truncate_uc_done |
||
400 | .truncate_uc_last: |
||
401 | ; 6f. The range at edi partially intersects with the UC-range described by MTRR. |
||
402 | ; Truncate it from the head. |
||
403 | mov dword [edi+mtrr_range.start], ebx |
||
404 | mov dword [edi+mtrr_range.start+4], ebp |
||
405 | neg eax |
||
406 | adc edx, 0 |
||
407 | neg edx |
||
408 | mov dword [edi+mtrr_range.length], eax |
||
409 | mov dword [edi+mtrr_range.length+4], edx |
||
410 | .truncate_uc_done: |
||
411 | ; 6g. We have found the next range (maybe 0) after intersection. |
||
412 | ; Write it to [esi+mtrr_range.next]. |
||
413 | mov [esi+mtrr_range.next], edi |
||
414 | .next_uc_range: |
||
415 | ; 6h. Continue the loop 6b-6g over all configured MTRRs. |
||
416 | add ecx, 16 |
||
417 | cmp ecx, [.mtrrs_end] |
||
418 | jb .fill_uc_ranges |
||
419 | ; Sanity check: if there are no ranges after steps 5-6, |
||
420 | ; fallback to step 7. Otherwise, go to 8. |
||
421 | cmp [.first_range], 0 |
||
422 | jnz .ranges_ok |
||
423 | .fill_ranges_from_memory_map: |
||
424 | ; 7. BIOS has not configured variable-range MTRRs. |
||
425 | ; Create one range from 0 to [MEM_AMOUNT]. |
||
426 | mov eax, [.first_free_range] |
||
427 | mov edx, [eax+mtrr_range.next] |
||
428 | mov [.first_free_range], edx |
||
429 | mov [.first_range], eax |
||
430 | xor edx, edx |
||
431 | mov [eax+mtrr_range.next], edx |
||
432 | mov dword [eax+mtrr_range.start], edx |
||
433 | mov dword [eax+mtrr_range.start+4], edx |
||
434 | mov ecx, [MEM_AMOUNT] |
||
435 | mov dword [eax+mtrr_range.length], ecx |
||
436 | mov dword [eax+mtrr_range.length+4], edx |
||
437 | .ranges_ok: |
||
438 | ; 8. We have calculated list of WB-ranges. |
||
439 | ; Now we should calculate a list of MTRRs so that |
||
440 | ; * every MTRR describes a range with length = power of 2 and start that is aligned, |
||
441 | ; * every MTRR can be WB or UC |
||
442 | ; * (sum of all WB ranges) minus (sum of all UC ranges) equals the calculated list |
||
443 | ; * top of 4G memory must not be covered by any ranges |
||
444 | ; Example: range [0,0xBC000000) can be converted to |
||
445 | ; [0,0x80000000)+[0x80000000,0xC0000000)-[0xBC000000,0xC0000000) |
||
446 | ; WB +WB -UC |
||
447 | ; but not to [0,0x100000000)-[0xC0000000,0x100000000)-[0xBC000000,0xC0000000). |
||
448 | ; 8a. Check that list of ranges is [0,something) plus, optionally, [4G,something). |
||
449 | ; This holds in practice (see mtrrtest.asm for real-life examples) |
||
450 | ; and significantly simplifies the code: ranges are independent, start of range |
||
451 | ; is almost always aligned (the only exception >4G upper memory can be easily covered), |
||
452 | ; there is no need to consider adding holes before start of range, only |
||
453 | ; append them to end of range. |
||
454 | xor eax, eax |
||
455 | mov edi, [.first_range] |
||
456 | cmp dword [edi+mtrr_range.start], eax |
||
457 | jnz .abort |
||
458 | cmp dword [edi+mtrr_range.start+4], eax |
||
459 | jnz .abort |
||
460 | cmp dword [edi+mtrr_range.length+4], eax |
||
461 | jnz .abort |
||
462 | mov edx, [edi+mtrr_range.next] |
||
463 | test edx, edx |
||
464 | jz @f |
||
465 | cmp dword [edx+mtrr_range.start], eax |
||
466 | jnz .abort |
||
467 | cmp dword [edx+mtrr_range.start+4], 1 |
||
468 | jnz .abort |
||
469 | cmp [edx+mtrr_range.next], eax |
||
470 | jnz .abort |
||
471 | @@: |
||
472 | ; 8b. Initialize: no MTRRs filled. |
||
473 | mov [.num_used_mtrrs], eax |
||
474 | lea esi, [.mtrrs] |
||
475 | .range2mtrr_loop: |
||
476 | ; 8c. If we are dealing with upper-memory range (after 4G) |
||
477 | ; with length > start, create one WB MTRR with [start,2*start), |
||
478 | ; reset start to 2*start and return to this step. |
||
479 | ; Example: [4G,24G) -> [4G,8G) {returning} + [8G,16G) {returning} |
||
480 | ; + [16G,24G) {advancing to ?}. |
||
481 | mov eax, dword [edi+mtrr_range.length+4] |
||
482 | test eax, eax |
||
483 | jz .less4G |
||
484 | mov edx, dword [edi+mtrr_range.start+4] |
||
485 | cmp eax, edx |
||
486 | jb .start_aligned |
||
487 | inc [.num_used_mtrrs] |
||
488 | cmp [.num_used_mtrrs], MAX_USEFUL_MTRRS |
||
489 | ja .abort |
||
490 | mov dword [esi], MEM_WB |
||
491 | mov dword [esi+4], edx |
||
492 | mov dword [esi+8], 0 |
||
493 | mov dword [esi+12], edx |
||
494 | add esi, 16 |
||
495 | add dword [edi+mtrr_range.start+4], edx |
||
496 | sub dword [edi+mtrr_range.length+4], edx |
||
497 | jnz .range2mtrr_loop |
||
498 | cmp dword [edi+mtrr_range.length], 0 |
||
499 | jz .range2mtrr_next |
||
500 | .less4G: |
||
501 | ; 8d. If we are dealing with low-memory range (before 4G) |
||
502 | ; and appending a maximal-size hole would create a range covering top of 4G, |
||
503 | ; create a maximal-size WB range and return to this step. |
||
504 | ; Example: for [0,0xBC000000) the following steps would consider |
||
505 | ; variants [0,0x80000000)+(another range to be splitted) and |
||
506 | ; [0,0x100000000)-(another range to be splitted); we forbid the last variant, |
||
507 | ; so the first variant must be used. |
||
508 | bsr ecx, dword [edi+mtrr_range.length] |
||
509 | xor edx, edx |
||
510 | inc edx |
||
511 | shl edx, cl |
||
512 | lea eax, [edx*2] |
||
513 | add eax, dword [edi+mtrr_range.start] |
||
514 | jnz .start_aligned |
||
515 | inc [.num_used_mtrrs] |
||
516 | cmp [.num_used_mtrrs], MAX_USEFUL_MTRRS |
||
517 | ja .abort |
||
518 | mov eax, dword [edi+mtrr_range.start] |
||
519 | mov dword [esi], eax |
||
520 | or dword [esi], MEM_WB |
||
521 | mov dword [esi+4], 0 |
||
522 | mov dword [esi+8], edx |
||
523 | mov dword [esi+12], 0 |
||
524 | add esi, 16 |
||
525 | add dword [edi+mtrr_range.start], edx |
||
526 | sub dword [edi+mtrr_range.length], edx |
||
527 | jnz .less4G |
||
528 | jmp .range2mtrr_next |
||
529 | .start_aligned: |
||
530 | ; Start is aligned for any allowed length, maximum-size hole is allowed. |
||
531 | ; Select the best MTRR configuration for one range. |
||
532 | ; length=...101101 |
||
533 | ; Without hole at the end, we need one WB MTRR for every 1-bit in length: |
||
534 | ; length=...100000 + ...001000 + ...000100 + ...000001 |
||
535 | ; We can also append one hole at the end so that one 0-bit (selected by us) |
||
536 | ; becomes 1 and all lower bits become 0 for WB-range: |
||
537 | ; length=...110000 - (...00010 + ...00001) |
||
538 | ; In this way, we need one WB MTRR for every 1-bit higher than the selected bit, |
||
539 | ; one WB MTRR for the selected bit, one UC MTRR for every 0-bit between |
||
540 | ; the selected bit and lowest 1-bit (they become 1-bits after negation) |
||
541 | ; and one UC MTRR for lowest 1-bit. |
||
542 | ; So we need to select 0-bit with the maximal difference |
||
543 | ; (number of 0-bits) - (number of 1-bits) between selected and lowest 1-bit, |
||
544 | ; this equals the gain from using a hole. If the difference is negative for |
||
545 | ; all 0-bits, don't append hole. |
||
546 | ; Note that lowest 1-bit is not included when counting, but selected 0-bit is. |
||
547 | ; 8e. Find the optimal bit position for hole. |
||
548 | ; eax = current difference, ebx = best difference, |
||
549 | ; ecx = hole bit position, edx = current bit position. |
||
550 | xor eax, eax |
||
551 | xor ebx, ebx |
||
552 | xor ecx, ecx |
||
553 | bsf edx, dword [edi+mtrr_range.length] |
||
554 | jnz @f |
||
555 | bsf edx, dword [edi+mtrr_range.length+4] |
||
556 | add edx, 32 |
||
557 | @@: |
||
558 | push edx ; save position of lowest 1-bit for step 8f |
||
559 | .calc_stat: |
||
560 | inc edx |
||
561 | cmp edx, 64 |
||
562 | jae .stat_done |
||
563 | inc eax ; increment difference in hope for 1-bit |
||
564 | ; Note: bt conveniently works with both .length and .length+4, |
||
565 | ; depending on whether edx>=32. |
||
566 | bt dword [edi+mtrr_range.length], edx |
||
567 | jc .calc_stat |
||
568 | dec eax ; hope was wrong, decrement difference to correct 'inc' |
||
569 | dec eax ; and again, now getting the real difference |
||
570 | cmp eax, ebx |
||
571 | jle .calc_stat |
||
572 | mov ebx, eax |
||
573 | mov ecx, edx |
||
574 | jmp .calc_stat |
||
575 | .stat_done: |
||
576 | ; 8f. If we decided to create a hole, flip all bits between lowest and selected. |
||
577 | pop edx ; restore position of lowest 1-bit saved at step 8e |
||
578 | test ecx, ecx |
||
579 | jz .fill_hi_init |
||
580 | @@: |
||
581 | inc edx |
||
582 | cmp edx, ecx |
||
583 | ja .fill_hi_init |
||
584 | btc dword [edi+mtrr_range.length], edx |
||
585 | jmp @b |
||
586 | .fill_hi_init: |
||
587 | ; 8g. Create MTRR ranges corresponding to upper 32 bits. |
||
588 | sub ecx, 32 |
||
589 | .fill_hi_loop: |
||
590 | bsr edx, dword [edi+mtrr_range.length+4] |
||
591 | jz .fill_hi_done |
||
592 | inc [.num_used_mtrrs] |
||
593 | cmp [.num_used_mtrrs], MAX_USEFUL_MTRRS |
||
594 | ja .abort |
||
595 | mov eax, dword [edi+mtrr_range.start] |
||
596 | mov [esi], eax |
||
597 | mov eax, dword [edi+mtrr_range.start+4] |
||
598 | mov [esi+4], eax |
||
599 | xor eax, eax |
||
600 | mov [esi+8], eax |
||
601 | bts eax, edx |
||
602 | mov [esi+12], eax |
||
603 | cmp edx, ecx |
||
604 | jl .fill_hi_uc |
||
605 | or dword [esi], MEM_WB |
||
606 | add dword [edi+mtrr_range.start+4], eax |
||
607 | jmp @f |
||
608 | .fill_hi_uc: |
||
609 | sub dword [esi+4], eax |
||
610 | sub dword [edi+mtrr_range.start+4], eax |
||
611 | @@: |
||
612 | add esi, 16 |
||
613 | sub dword [edi+mtrr_range.length], eax |
||
614 | jmp .fill_hi_loop |
||
615 | .fill_hi_done: |
||
616 | ; 8h. Create MTRR ranges corresponding to lower 32 bits. |
||
617 | add ecx, 32 |
||
618 | .fill_lo_loop: |
||
619 | bsr edx, dword [edi+mtrr_range.length] |
||
620 | jz .range2mtrr_next |
||
621 | inc [.num_used_mtrrs] |
||
622 | cmp [.num_used_mtrrs], MAX_USEFUL_MTRRS |
||
623 | ja .abort |
||
624 | mov eax, dword [edi+mtrr_range.start] |
||
625 | mov [esi], eax |
||
626 | mov eax, dword [edi+mtrr_range.start+4] |
||
627 | mov [esi+4], eax |
||
628 | xor eax, eax |
||
629 | mov [esi+12], eax |
||
630 | bts eax, edx |
||
631 | mov [esi+8], eax |
||
632 | cmp edx, ecx |
||
633 | jl .fill_lo_uc |
||
634 | or dword [esi], MEM_WB |
||
635 | add dword [edi+mtrr_range.start], eax |
||
636 | jmp @f |
||
637 | .fill_lo_uc: |
||
638 | sub dword [esi], eax |
||
639 | sub dword [edi+mtrr_range.start], eax |
||
640 | @@: |
||
641 | add esi, 16 |
||
642 | sub dword [edi+mtrr_range.length], eax |
||
643 | jmp .fill_lo_loop |
||
644 | .range2mtrr_next: |
||
645 | ; 8i. Repeat the loop at 8c-8h for all ranges. |
||
646 | mov edi, [edi+mtrr_range.next] |
||
647 | test edi, edi |
||
648 | jnz .range2mtrr_loop |
||
649 | ; 9. We have calculated needed MTRRs, now setup them in the CPU. |
||
650 | ; 9a. Abort if number of MTRRs is too large. |
||
651 | mov eax, [num_variable_mtrrs] |
||
652 | cmp [.num_used_mtrrs], eax |
||
653 | ja .abort |
||
654 | |||
655 | ; 9b. Prepare for changes. |
||
656 | call mtrr_begin_change |
||
657 | |||
658 | ; 9c. Prepare for loop over MTRRs. |
||
659 | lea esi, [.mtrrs] |
||
660 | mov ecx, 0x200 |
||
661 | @@: |
||
662 | ; 9d. For every MTRR, copy PHYSBASEn as is: step 8 has configured |
||
663 | ; start value and type bits as needed. |
||
664 | mov eax, [esi] |
||
665 | mov edx, [esi+4] |
||
666 | wrmsr |
||
667 | inc ecx |
||
668 | ; 9e. For every MTRR, calculate PHYSMASKn = -(length) or 0x800 |
||
669 | ; with upper bits cleared, 0x800 = MTRR is valid. |
||
670 | xor eax, eax |
||
671 | xor edx, edx |
||
672 | sub eax, [esi+8] |
||
673 | sbb edx, [esi+12] |
||
674 | or eax, 0x800 |
||
675 | or edx, [.phys_reserved_mask] |
||
676 | xor edx, [.phys_reserved_mask] |
||
677 | wrmsr |
||
678 | inc ecx |
||
679 | ; 9f. Continue steps 9d and 9e for all MTRRs calculated at step 8. |
||
680 | add esi, 16 |
||
681 | dec [.num_used_mtrrs] |
||
682 | jnz @b |
||
683 | ; 9g. Zero other MTRRs. |
||
684 | xor eax, eax |
||
685 | xor edx, edx |
||
686 | mov ebx, [num_variable_mtrrs] |
||
687 | lea ebx, [0x200+ebx*2] |
||
688 | @@: |
||
689 | cmp ecx, ebx |
||
690 | jae @f |
||
691 | wrmsr |
||
692 | inc ecx |
||
693 | wrmsr |
||
694 | inc ecx |
||
695 | jmp @b |
||
696 | @@: |
||
697 | |||
698 | ; 9i. Configure MTRR_DEF_TYPE. |
||
699 | mov ecx, 0x2FF |
||
700 | rdmsr |
||
701 | or ah, 8 ; enable variable-ranges MTRR |
||
702 | and al, 0xF0; default memtype = UC |
||
703 | wrmsr |
||
704 | |||
705 | ; 9j. Changes are done. |
||
706 | call mtrr_end_change |
||
707 | |||
708 | .abort: |
||
709 | add esp, .local_vars_size + MAX_RANGES * sizeof.mtrr_range |
||
710 | pop ebp |
||
711 | ret |
||
712 | endp |
||
713 | |||
714 | ; Allocate&set one MTRR for given range. |
||
715 | ; size must be power of 2 that divides base. |
||
716 | proc set_mtrr stdcall, base:dword,size:dword,mem_type:dword |
||
717 | ; find unused register |
||
718 | mov ecx, 0x201 |
||
719 | .scan: |
||
720 | rdmsr |
||
721 | dec ecx |
||
722 | test ah, 8 |
||
723 | jz .found |
||
724 | rdmsr |
||
725 | test edx, edx |
||
726 | jnz @f |
||
727 | and eax, not 0xFFF ; clear reserved bits |
||
728 | cmp eax, [base] |
||
729 | jz .ret |
||
730 | @@: |
||
731 | add ecx, 3 |
||
732 | mov eax, [num_variable_mtrrs] |
||
733 | lea eax, [0x200+eax*2] |
||
734 | cmp ecx, eax |
||
735 | jb .scan |
||
736 | ; no free registers, ignore the call |
||
737 | .ret: |
||
738 | ret |
||
739 | .found: |
||
740 | ; found, write values |
||
741 | call mtrr_begin_change |
||
742 | xor edx, edx |
||
743 | mov eax, [base] |
||
744 | or eax, [mem_type] |
||
745 | wrmsr |
||
746 | |||
747 | mov al, [cpu_phys_addr_width] |
||
748 | xor edx, edx |
||
749 | bts edx, eax |
||
750 | xor eax, eax |
||
751 | sub eax, [size] |
||
752 | sbb edx, 0 |
||
753 | or eax, 0x800 |
||
754 | inc ecx |
||
755 | wrmsr |
||
756 | call mtrr_end_change |
||
757 | ret |
||
758 | endp |
||
759 | |||
760 | ; Helper procedure for mtrr_validate. |
||
761 | ; Calculates memory type for given address according to variable-range MTRRs. |
||
762 | ; Assumes that MTRRs are enabled. |
||
763 | ; in: ebx = 32-bit physical address |
||
764 | ; out: eax = memory type for ebx |
||
765 | proc mtrr_get_real_type |
||
766 | ; 1. Initialize: we have not yet found any MTRRs covering ebx. |
||
767 | push 0 |
||
768 | mov ecx, 0x201 |
||
769 | .mtrr_loop: |
||
770 | ; 2. For every MTRR, check whether it is valid; if not, continue to the next MTRR. |
||
771 | rdmsr |
||
772 | dec ecx |
||
773 | test ah, 8 |
||
774 | jz .next |
||
775 | ; 3. For every valid MTRR, check whether (ebx and PHYSMASKn) == PHYSBASEn, |
||
776 | ; excluding low 12 bits. |
||
777 | and eax, ebx |
||
778 | push eax |
||
779 | rdmsr |
||
780 | test edx, edx |
||
781 | pop edx |
||
782 | jnz .next |
||
783 | xor edx, eax |
||
784 | and edx, not 0xFFF |
||
785 | jnz .next |
||
786 | ; 4. If so, set the bit corresponding to memory type defined by this MTRR. |
||
787 | and eax, 7 |
||
788 | bts [esp], eax |
||
789 | .next: |
||
790 | ; 5. Continue loop at 2-4 for all variable-range MTRRs. |
||
791 | add ecx, 3 |
||
792 | mov eax, [num_variable_mtrrs] |
||
793 | lea eax, [0x200+eax*2] |
||
794 | cmp ecx, eax |
||
795 | jb .mtrr_loop |
||
796 | ; 6. If no MTRRs cover address in ebx, use default MTRR type from MTRR_DEF_CAP. |
||
797 | pop edx |
||
798 | test edx, edx |
||
799 | jz .default |
||
800 | ; 7. Find&clear 1-bit in edx. |
||
801 | bsf eax, edx |
||
802 | btr edx, eax |
||
803 | ; 8. If there was only one 1-bit, then all MTRRs are consistent, return that bit. |
||
804 | test edx, edx |
||
805 | jz .nothing |
||
806 | ; Otherwise, return MEM_UC (e.g. WB+UC is UC). |
||
807 | xor eax, eax |
||
808 | .nothing: |
||
809 | ret |
||
810 | .default: |
||
811 | mov ecx, 0x2FF |
||
812 | rdmsr |
||
813 | movzx eax, al |
||
814 | ret |
||
815 | endp |
||
816 | |||
817 | ; If MTRRs are configured improperly, this is not obvious to the user; |
||
818 | ; everything works, but the performance can be horrible. |
||
819 | ; Try to detect this and let the user know that the low performance |
||
820 | ; is caused by some problem and is not a global property of the system. |
||
821 | ; Let's hope he would report it to developers... |
||
822 | proc mtrr_validate |
||
823 | ; 1. If MTRRs are not supported, they cannot be configured improperly. |
||
824 | ; Note: VirtualBox claims MTRR support in cpuid, but emulates MTRRCAP=0, |
||
825 | ; which is efficiently equivalent to absent MTRRs. |
||
826 | ; So check [num_variable_mtrrs] instead of CAPS_MTRR in [cpu_caps]. |
||
827 | cmp [num_variable_mtrrs], 0 |
||
828 | jz .exit |
||
829 | ; 2. If variable-range MTRRs are not configured, this is a problem. |
||
830 | mov ecx, 0x2FF |
||
831 | rdmsr |
||
832 | test ah, 8 |
||
833 | jz .fail |
||
834 | ; 3. Get the memory type for address somewhere inside working memory. |
||
835 | ; It must be write-back. |
||
836 | mov ebx, 0x27FFFF |
||
837 | call mtrr_get_real_type |
||
838 | cmp al, MEM_WB |
||
839 | jnz .fail |
||
840 | ; 4. If we're using a mode with LFB, |
||
841 | ; get the memory type for last pixel of the framebuffer. |
||
842 | ; It must be write-combined. |
||
843 | test word [SCR_MODE], 0x4000 |
||
844 | jz .exit |
||
845 | mov eax, [_display.pitch] |
||
846 | mul [_display.height] |
||
847 | dec eax |
||
848 | ; LFB is mapped to virtual address LFB_BASE, |
||
849 | ; it uses global pages if supported by CPU. |
||
850 | mov ebx, [sys_proc+PROC.pdt_0+(LFB_BASE shr 20)] |
||
851 | test ebx, PG_LARGE |
||
852 | jnz @f |
||
853 | mov ebx, [page_tabs+(LFB_BASE shr 10)] |
||
854 | @@: |
||
855 | and ebx, not 0xFFF |
||
856 | add ebx, eax |
||
857 | call mtrr_get_real_type |
||
858 | cmp al, MEM_WC |
||
859 | jz .exit |
||
860 | ; 5. The check at step 4 fails on Bochs: |
||
861 | ; Bochs BIOS configures MTRRs in a strange way not respecting [cpu_phys_addr_width], |
||
862 | ; so mtrr_reconfigure avoids to touch anything. |
||
863 | ; However, Bochs core ignores MTRRs (keeping them only for rdmsr/wrmsr), |
||
864 | ; so we don't care about proper setting for Bochs. |
||
865 | ; Use northbridge PCI id to detect Bochs: it emulates either i440fx or i430fx |
||
866 | ; depending on configuration file. |
||
867 | mov eax, [pcidev_list.fd] |
||
868 | cmp eax, pcidev_list ; sanity check: fail if no PCI devices |
||
869 | jz .fail |
||
870 | cmp [eax+PCIDEV.vendor_device_id], 0x12378086 |
||
871 | jz .exit |
||
872 | cmp [eax+PCIDEV.vendor_device_id], 0x01228086 |
||
873 | jnz .fail |
||
874 | .exit: |
||
875 | ret |
||
876 | .fail: |
||
877 | mov ebx, mtrr_user_message |
||
878 | mov ebp, notifyapp |
||
879 | call fs_execute_from_sysdir_param |
||
880 | ret |
||
881 | endp |