1 | // SPDX-License-Identifier: GPL-2.0 |
2 | #define DEBUG |
3 | |
4 | #include <linux/wait.h> |
5 | #include <linux/ptrace.h> |
6 | |
7 | #include <asm/spu.h> |
8 | #include <asm/spu_priv1.h> |
9 | #include <asm/io.h> |
10 | #include <asm/unistd.h> |
11 | |
12 | #include "spufs.h" |
13 | |
14 | /* interrupt-level stop callback function. */ |
15 | void spufs_stop_callback(struct spu *spu, int irq) |
16 | { |
17 | struct spu_context *ctx = spu->ctx; |
18 | |
19 | /* |
20 | * It should be impossible to preempt a context while an exception |
21 | * is being processed, since the context switch code is specially |
22 | * coded to deal with interrupts ... But, just in case, sanity check |
23 | * the context pointer. It is OK to return doing nothing since |
24 | * the exception will be regenerated when the context is resumed. |
25 | */ |
26 | if (ctx) { |
27 | /* Copy exception arguments into module specific structure */ |
28 | switch(irq) { |
29 | case 0 : |
30 | ctx->csa.class_0_pending = spu->class_0_pending; |
31 | ctx->csa.class_0_dar = spu->class_0_dar; |
32 | break; |
33 | case 1 : |
34 | ctx->csa.class_1_dsisr = spu->class_1_dsisr; |
35 | ctx->csa.class_1_dar = spu->class_1_dar; |
36 | break; |
37 | case 2 : |
38 | break; |
39 | } |
40 | |
41 | /* ensure that the exception status has hit memory before a |
42 | * thread waiting on the context's stop queue is woken */ |
43 | smp_wmb(); |
44 | |
45 | wake_up_all(&ctx->stop_wq); |
46 | } |
47 | } |
48 | |
49 | int spu_stopped(struct spu_context *ctx, u32 *stat) |
50 | { |
51 | u64 dsisr; |
52 | u32 stopped; |
53 | |
54 | stopped = SPU_STATUS_INVALID_INSTR | SPU_STATUS_SINGLE_STEP | |
55 | SPU_STATUS_STOPPED_BY_HALT | SPU_STATUS_STOPPED_BY_STOP; |
56 | |
57 | top: |
58 | *stat = ctx->ops->status_read(ctx); |
59 | if (*stat & stopped) { |
60 | /* |
61 | * If the spu hasn't finished stopping, we need to |
62 | * re-read the register to get the stopped value. |
63 | */ |
64 | if (*stat & SPU_STATUS_RUNNING) |
65 | goto top; |
66 | return 1; |
67 | } |
68 | |
69 | if (test_bit(SPU_SCHED_NOTIFY_ACTIVE, &ctx->sched_flags)) |
70 | return 1; |
71 | |
72 | dsisr = ctx->csa.class_1_dsisr; |
73 | if (dsisr & (MFC_DSISR_PTE_NOT_FOUND | MFC_DSISR_ACCESS_DENIED)) |
74 | return 1; |
75 | |
76 | if (ctx->csa.class_0_pending) |
77 | return 1; |
78 | |
79 | return 0; |
80 | } |
81 | |
82 | static int spu_setup_isolated(struct spu_context *ctx) |
83 | { |
84 | int ret; |
85 | u64 __iomem *mfc_cntl; |
86 | u64 sr1; |
87 | u32 status; |
88 | unsigned long timeout; |
89 | const u32 status_loading = SPU_STATUS_RUNNING |
90 | | SPU_STATUS_ISOLATED_STATE | SPU_STATUS_ISOLATED_LOAD_STATUS; |
91 | |
92 | ret = -ENODEV; |
93 | if (!isolated_loader) |
94 | goto out; |
95 | |
96 | /* |
97 | * We need to exclude userspace access to the context. |
98 | * |
99 | * To protect against memory access we invalidate all ptes |
100 | * and make sure the pagefault handlers block on the mutex. |
101 | */ |
102 | spu_unmap_mappings(ctx); |
103 | |
104 | mfc_cntl = &ctx->spu->priv2->mfc_control_RW; |
105 | |
106 | /* purge the MFC DMA queue to ensure no spurious accesses before we |
107 | * enter kernel mode */ |
108 | timeout = jiffies + HZ; |
109 | out_be64(mfc_cntl, MFC_CNTL_PURGE_DMA_REQUEST); |
110 | while ((in_be64(mfc_cntl) & MFC_CNTL_PURGE_DMA_STATUS_MASK) |
111 | != MFC_CNTL_PURGE_DMA_COMPLETE) { |
112 | if (time_after(jiffies, timeout)) { |
113 | printk(KERN_ERR "%s: timeout flushing MFC DMA queue\n" , |
114 | __func__); |
115 | ret = -EIO; |
116 | goto out; |
117 | } |
118 | cond_resched(); |
119 | } |
120 | |
121 | /* clear purge status */ |
122 | out_be64(mfc_cntl, 0); |
123 | |
124 | /* put the SPE in kernel mode to allow access to the loader */ |
125 | sr1 = spu_mfc_sr1_get(ctx->spu); |
126 | sr1 &= ~MFC_STATE1_PROBLEM_STATE_MASK; |
127 | spu_mfc_sr1_set(ctx->spu, sr1); |
128 | |
129 | /* start the loader */ |
130 | ctx->ops->signal1_write(ctx, (unsigned long)isolated_loader >> 32); |
131 | ctx->ops->signal2_write(ctx, |
132 | (unsigned long)isolated_loader & 0xffffffff); |
133 | |
134 | ctx->ops->runcntl_write(ctx, |
135 | SPU_RUNCNTL_RUNNABLE | SPU_RUNCNTL_ISOLATE); |
136 | |
137 | ret = 0; |
138 | timeout = jiffies + HZ; |
139 | while (((status = ctx->ops->status_read(ctx)) & status_loading) == |
140 | status_loading) { |
141 | if (time_after(jiffies, timeout)) { |
142 | printk(KERN_ERR "%s: timeout waiting for loader\n" , |
143 | __func__); |
144 | ret = -EIO; |
145 | goto out_drop_priv; |
146 | } |
147 | cond_resched(); |
148 | } |
149 | |
150 | if (!(status & SPU_STATUS_RUNNING)) { |
151 | /* If isolated LOAD has failed: run SPU, we will get a stop-and |
152 | * signal later. */ |
153 | pr_debug("%s: isolated LOAD failed\n" , __func__); |
154 | ctx->ops->runcntl_write(ctx, SPU_RUNCNTL_RUNNABLE); |
155 | ret = -EACCES; |
156 | goto out_drop_priv; |
157 | } |
158 | |
159 | if (!(status & SPU_STATUS_ISOLATED_STATE)) { |
160 | /* This isn't allowed by the CBEA, but check anyway */ |
161 | pr_debug("%s: SPU fell out of isolated mode?\n" , __func__); |
162 | ctx->ops->runcntl_write(ctx, SPU_RUNCNTL_STOP); |
163 | ret = -EINVAL; |
164 | goto out_drop_priv; |
165 | } |
166 | |
167 | out_drop_priv: |
168 | /* Finished accessing the loader. Drop kernel mode */ |
169 | sr1 |= MFC_STATE1_PROBLEM_STATE_MASK; |
170 | spu_mfc_sr1_set(ctx->spu, sr1); |
171 | |
172 | out: |
173 | return ret; |
174 | } |
175 | |
176 | static int spu_run_init(struct spu_context *ctx, u32 *npc) |
177 | { |
178 | unsigned long runcntl = SPU_RUNCNTL_RUNNABLE; |
179 | int ret; |
180 | |
181 | spuctx_switch_state(ctx, new_state: SPU_UTIL_SYSTEM); |
182 | |
183 | /* |
184 | * NOSCHED is synchronous scheduling with respect to the caller. |
185 | * The caller waits for the context to be loaded. |
186 | */ |
187 | if (ctx->flags & SPU_CREATE_NOSCHED) { |
188 | if (ctx->state == SPU_STATE_SAVED) { |
189 | ret = spu_activate(ctx, flags: 0); |
190 | if (ret) |
191 | return ret; |
192 | } |
193 | } |
194 | |
195 | /* |
196 | * Apply special setup as required. |
197 | */ |
198 | if (ctx->flags & SPU_CREATE_ISOLATE) { |
199 | if (!(ctx->ops->status_read(ctx) & SPU_STATUS_ISOLATED_STATE)) { |
200 | ret = spu_setup_isolated(ctx); |
201 | if (ret) |
202 | return ret; |
203 | } |
204 | |
205 | /* |
206 | * If userspace has set the runcntrl register (eg, to |
207 | * issue an isolated exit), we need to re-set it here |
208 | */ |
209 | runcntl = ctx->ops->runcntl_read(ctx) & |
210 | (SPU_RUNCNTL_RUNNABLE | SPU_RUNCNTL_ISOLATE); |
211 | if (runcntl == 0) |
212 | runcntl = SPU_RUNCNTL_RUNNABLE; |
213 | } else { |
214 | unsigned long privcntl; |
215 | |
216 | if (test_thread_flag(TIF_SINGLESTEP)) |
217 | privcntl = SPU_PRIVCNTL_MODE_SINGLE_STEP; |
218 | else |
219 | privcntl = SPU_PRIVCNTL_MODE_NORMAL; |
220 | |
221 | ctx->ops->privcntl_write(ctx, privcntl); |
222 | ctx->ops->npc_write(ctx, *npc); |
223 | } |
224 | |
225 | ctx->ops->runcntl_write(ctx, runcntl); |
226 | |
227 | if (ctx->flags & SPU_CREATE_NOSCHED) { |
228 | spuctx_switch_state(ctx, new_state: SPU_UTIL_USER); |
229 | } else { |
230 | |
231 | if (ctx->state == SPU_STATE_SAVED) { |
232 | ret = spu_activate(ctx, flags: 0); |
233 | if (ret) |
234 | return ret; |
235 | } else { |
236 | spuctx_switch_state(ctx, new_state: SPU_UTIL_USER); |
237 | } |
238 | } |
239 | |
240 | set_bit(nr: SPU_SCHED_SPU_RUN, addr: &ctx->sched_flags); |
241 | return 0; |
242 | } |
243 | |
244 | static int spu_run_fini(struct spu_context *ctx, u32 *npc, |
245 | u32 *status) |
246 | { |
247 | int ret = 0; |
248 | |
249 | spu_del_from_rq(ctx); |
250 | |
251 | *status = ctx->ops->status_read(ctx); |
252 | *npc = ctx->ops->npc_read(ctx); |
253 | |
254 | spuctx_switch_state(ctx, new_state: SPU_UTIL_IDLE_LOADED); |
255 | clear_bit(nr: SPU_SCHED_SPU_RUN, addr: &ctx->sched_flags); |
256 | spu_switch_log_notify(NULL, ctx, type: SWITCH_LOG_EXIT, val: *status); |
257 | spu_release(ctx); |
258 | |
259 | if (signal_pending(current)) |
260 | ret = -ERESTARTSYS; |
261 | |
262 | return ret; |
263 | } |
264 | |
265 | /* |
266 | * SPU syscall restarting is tricky because we violate the basic |
267 | * assumption that the signal handler is running on the interrupted |
268 | * thread. Here instead, the handler runs on PowerPC user space code, |
269 | * while the syscall was called from the SPU. |
270 | * This means we can only do a very rough approximation of POSIX |
271 | * signal semantics. |
272 | */ |
273 | static int spu_handle_restartsys(struct spu_context *ctx, long *spu_ret, |
274 | unsigned int *npc) |
275 | { |
276 | int ret; |
277 | |
278 | switch (*spu_ret) { |
279 | case -ERESTARTSYS: |
280 | case -ERESTARTNOINTR: |
281 | /* |
282 | * Enter the regular syscall restarting for |
283 | * sys_spu_run, then restart the SPU syscall |
284 | * callback. |
285 | */ |
286 | *npc -= 8; |
287 | ret = -ERESTARTSYS; |
288 | break; |
289 | case -ERESTARTNOHAND: |
290 | case -ERESTART_RESTARTBLOCK: |
291 | /* |
292 | * Restart block is too hard for now, just return -EINTR |
293 | * to the SPU. |
294 | * ERESTARTNOHAND comes from sys_pause, we also return |
295 | * -EINTR from there. |
296 | * Assume that we need to be restarted ourselves though. |
297 | */ |
298 | *spu_ret = -EINTR; |
299 | ret = -ERESTARTSYS; |
300 | break; |
301 | default: |
302 | printk(KERN_WARNING "%s: unexpected return code %ld\n" , |
303 | __func__, *spu_ret); |
304 | ret = 0; |
305 | } |
306 | return ret; |
307 | } |
308 | |
309 | static int spu_process_callback(struct spu_context *ctx) |
310 | { |
311 | struct spu_syscall_block s; |
312 | u32 ls_pointer, npc; |
313 | void __iomem *ls; |
314 | long spu_ret; |
315 | int ret; |
316 | |
317 | /* get syscall block from local store */ |
318 | npc = ctx->ops->npc_read(ctx) & ~3; |
319 | ls = (void __iomem *)ctx->ops->get_ls(ctx); |
320 | ls_pointer = in_be32(ls + npc); |
321 | if (ls_pointer > (LS_SIZE - sizeof(s))) |
322 | return -EFAULT; |
323 | memcpy_fromio(&s, ls + ls_pointer, sizeof(s)); |
324 | |
325 | /* do actual syscall without pinning the spu */ |
326 | ret = 0; |
327 | spu_ret = -ENOSYS; |
328 | npc += 4; |
329 | |
330 | if (s.nr_ret < NR_syscalls) { |
331 | spu_release(ctx); |
332 | /* do actual system call from here */ |
333 | spu_ret = spu_sys_callback(&s); |
334 | if (spu_ret <= -ERESTARTSYS) { |
335 | ret = spu_handle_restartsys(ctx, spu_ret: &spu_ret, npc: &npc); |
336 | } |
337 | mutex_lock(&ctx->state_mutex); |
338 | if (ret == -ERESTARTSYS) |
339 | return ret; |
340 | } |
341 | |
342 | /* need to re-get the ls, as it may have changed when we released the |
343 | * spu */ |
344 | ls = (void __iomem *)ctx->ops->get_ls(ctx); |
345 | |
346 | /* write result, jump over indirect pointer */ |
347 | memcpy_toio(ls + ls_pointer, &spu_ret, sizeof(spu_ret)); |
348 | ctx->ops->npc_write(ctx, npc); |
349 | ctx->ops->runcntl_write(ctx, SPU_RUNCNTL_RUNNABLE); |
350 | return ret; |
351 | } |
352 | |
353 | long spufs_run_spu(struct spu_context *ctx, u32 *npc, u32 *event) |
354 | { |
355 | int ret; |
356 | u32 status; |
357 | |
358 | if (mutex_lock_interruptible(&ctx->run_mutex)) |
359 | return -ERESTARTSYS; |
360 | |
361 | ctx->event_return = 0; |
362 | |
363 | ret = spu_acquire(ctx); |
364 | if (ret) |
365 | goto out_unlock; |
366 | |
367 | spu_enable_spu(ctx); |
368 | |
369 | spu_update_sched_info(ctx); |
370 | |
371 | ret = spu_run_init(ctx, npc); |
372 | if (ret) { |
373 | spu_release(ctx); |
374 | goto out; |
375 | } |
376 | |
377 | do { |
378 | ret = spufs_wait(ctx->stop_wq, spu_stopped(ctx, &status)); |
379 | if (unlikely(ret)) { |
380 | /* |
381 | * This is nasty: we need the state_mutex for all the |
382 | * bookkeeping even if the syscall was interrupted by |
383 | * a signal. ewww. |
384 | */ |
385 | mutex_lock(&ctx->state_mutex); |
386 | break; |
387 | } |
388 | if (unlikely(test_and_clear_bit(SPU_SCHED_NOTIFY_ACTIVE, |
389 | &ctx->sched_flags))) { |
390 | if (!(status & SPU_STATUS_STOPPED_BY_STOP)) |
391 | continue; |
392 | } |
393 | |
394 | spuctx_switch_state(ctx, SPU_UTIL_SYSTEM); |
395 | |
396 | if ((status & SPU_STATUS_STOPPED_BY_STOP) && |
397 | (status >> SPU_STOP_STATUS_SHIFT == 0x2104)) { |
398 | ret = spu_process_callback(ctx); |
399 | if (ret) |
400 | break; |
401 | status &= ~SPU_STATUS_STOPPED_BY_STOP; |
402 | } |
403 | ret = spufs_handle_class1(ctx); |
404 | if (ret) |
405 | break; |
406 | |
407 | ret = spufs_handle_class0(ctx); |
408 | if (ret) |
409 | break; |
410 | |
411 | if (signal_pending(current)) |
412 | ret = -ERESTARTSYS; |
413 | } while (!ret && !(status & (SPU_STATUS_STOPPED_BY_STOP | |
414 | SPU_STATUS_STOPPED_BY_HALT | |
415 | SPU_STATUS_SINGLE_STEP))); |
416 | |
417 | spu_disable_spu(ctx); |
418 | ret = spu_run_fini(ctx, npc, status: &status); |
419 | spu_yield(ctx); |
420 | |
421 | if ((status & SPU_STATUS_STOPPED_BY_STOP) && |
422 | (((status >> SPU_STOP_STATUS_SHIFT) & 0x3f00) == 0x2100)) |
423 | ctx->stats.libassist++; |
424 | |
425 | if ((ret == 0) || |
426 | ((ret == -ERESTARTSYS) && |
427 | ((status & SPU_STATUS_STOPPED_BY_HALT) || |
428 | (status & SPU_STATUS_SINGLE_STEP) || |
429 | ((status & SPU_STATUS_STOPPED_BY_STOP) && |
430 | (status >> SPU_STOP_STATUS_SHIFT != 0x2104))))) |
431 | ret = status; |
432 | |
433 | /* Note: we don't need to force_sig SIGTRAP on single-step |
434 | * since we have TIF_SINGLESTEP set, thus the kernel will do |
435 | * it upon return from the syscall anyway. |
436 | */ |
437 | if (unlikely(status & SPU_STATUS_SINGLE_STEP)) |
438 | ret = -ERESTARTSYS; |
439 | |
440 | else if (unlikely((status & SPU_STATUS_STOPPED_BY_STOP) |
441 | && (status >> SPU_STOP_STATUS_SHIFT) == 0x3fff)) { |
442 | force_sig(SIGTRAP); |
443 | ret = -ERESTARTSYS; |
444 | } |
445 | |
446 | out: |
447 | *event = ctx->event_return; |
448 | out_unlock: |
449 | mutex_unlock(lock: &ctx->run_mutex); |
450 | return ret; |
451 | } |
452 | |