xref: /linux/drivers/misc/cxl/context.c (revision c895f6f703ad7dd2f99e751d9884b0aa5d0eea25)
1 /*
2  * Copyright 2014 IBM Corp.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public License
6  * as published by the Free Software Foundation; either version
7  * 2 of the License, or (at your option) any later version.
8  */
9 
10 #include <linux/module.h>
11 #include <linux/kernel.h>
12 #include <linux/bitmap.h>
13 #include <linux/sched.h>
14 #include <linux/pid.h>
15 #include <linux/fs.h>
16 #include <linux/mm.h>
17 #include <linux/debugfs.h>
18 #include <linux/slab.h>
19 #include <linux/idr.h>
20 #include <linux/sched/mm.h>
21 #include <linux/mmu_context.h>
22 #include <asm/cputable.h>
23 #include <asm/current.h>
24 #include <asm/copro.h>
25 
26 #include "cxl.h"
27 
28 /*
29  * Allocates space for a CXL context.
30  */
31 struct cxl_context *cxl_context_alloc(void)
32 {
33 	return kzalloc(sizeof(struct cxl_context), GFP_KERNEL);
34 }
35 
36 /*
37  * Initialises a CXL context.
38  */
39 int cxl_context_init(struct cxl_context *ctx, struct cxl_afu *afu, bool master)
40 {
41 	int i;
42 
43 	ctx->afu = afu;
44 	ctx->master = master;
45 	ctx->pid = NULL; /* Set in start work ioctl */
46 	mutex_init(&ctx->mapping_lock);
47 	ctx->mapping = NULL;
48 
49 	if (cxl_is_power8()) {
50 		spin_lock_init(&ctx->sste_lock);
51 
52 		/*
53 		 * Allocate the segment table before we put it in the IDR so that we
54 		 * can always access it when dereferenced from IDR. For the same
55 		 * reason, the segment table is only destroyed after the context is
56 		 * removed from the IDR.  Access to this in the IOCTL is protected by
57 		 * Linux filesytem symantics (can't IOCTL until open is complete).
58 		 */
59 		i = cxl_alloc_sst(ctx);
60 		if (i)
61 			return i;
62 	}
63 
64 	INIT_WORK(&ctx->fault_work, cxl_handle_fault);
65 
66 	init_waitqueue_head(&ctx->wq);
67 	spin_lock_init(&ctx->lock);
68 
69 	ctx->irq_bitmap = NULL;
70 	ctx->pending_irq = false;
71 	ctx->pending_fault = false;
72 	ctx->pending_afu_err = false;
73 
74 	INIT_LIST_HEAD(&ctx->irq_names);
75 	INIT_LIST_HEAD(&ctx->extra_irq_contexts);
76 
77 	/*
78 	 * When we have to destroy all contexts in cxl_context_detach_all() we
79 	 * end up with afu_release_irqs() called from inside a
80 	 * idr_for_each_entry(). Hence we need to make sure that anything
81 	 * dereferenced from this IDR is ok before we allocate the IDR here.
82 	 * This clears out the IRQ ranges to ensure this.
83 	 */
84 	for (i = 0; i < CXL_IRQ_RANGES; i++)
85 		ctx->irqs.range[i] = 0;
86 
87 	mutex_init(&ctx->status_mutex);
88 
89 	ctx->status = OPENED;
90 
91 	/*
92 	 * Allocating IDR! We better make sure everything's setup that
93 	 * dereferences from it.
94 	 */
95 	mutex_lock(&afu->contexts_lock);
96 	idr_preload(GFP_KERNEL);
97 	i = idr_alloc(&ctx->afu->contexts_idr, ctx, ctx->afu->adapter->min_pe,
98 		      ctx->afu->num_procs, GFP_NOWAIT);
99 	idr_preload_end();
100 	mutex_unlock(&afu->contexts_lock);
101 	if (i < 0)
102 		return i;
103 
104 	ctx->pe = i;
105 	if (cpu_has_feature(CPU_FTR_HVMODE)) {
106 		ctx->elem = &ctx->afu->native->spa[i];
107 		ctx->external_pe = ctx->pe;
108 	} else {
109 		ctx->external_pe = -1; /* assigned when attaching */
110 	}
111 	ctx->pe_inserted = false;
112 
113 	/*
114 	 * take a ref on the afu so that it stays alive at-least till
115 	 * this context is reclaimed inside reclaim_ctx.
116 	 */
117 	cxl_afu_get(afu);
118 	return 0;
119 }
120 
121 void cxl_context_set_mapping(struct cxl_context *ctx,
122 			struct address_space *mapping)
123 {
124 	mutex_lock(&ctx->mapping_lock);
125 	ctx->mapping = mapping;
126 	mutex_unlock(&ctx->mapping_lock);
127 }
128 
129 static int cxl_mmap_fault(struct vm_fault *vmf)
130 {
131 	struct vm_area_struct *vma = vmf->vma;
132 	struct cxl_context *ctx = vma->vm_file->private_data;
133 	u64 area, offset;
134 
135 	offset = vmf->pgoff << PAGE_SHIFT;
136 
137 	pr_devel("%s: pe: %i address: 0x%lx offset: 0x%llx\n",
138 			__func__, ctx->pe, vmf->address, offset);
139 
140 	if (ctx->afu->current_mode == CXL_MODE_DEDICATED) {
141 		area = ctx->afu->psn_phys;
142 		if (offset >= ctx->afu->adapter->ps_size)
143 			return VM_FAULT_SIGBUS;
144 	} else {
145 		area = ctx->psn_phys;
146 		if (offset >= ctx->psn_size)
147 			return VM_FAULT_SIGBUS;
148 	}
149 
150 	mutex_lock(&ctx->status_mutex);
151 
152 	if (ctx->status != STARTED) {
153 		mutex_unlock(&ctx->status_mutex);
154 		pr_devel("%s: Context not started, failing problem state access\n", __func__);
155 		if (ctx->mmio_err_ff) {
156 			if (!ctx->ff_page) {
157 				ctx->ff_page = alloc_page(GFP_USER);
158 				if (!ctx->ff_page)
159 					return VM_FAULT_OOM;
160 				memset(page_address(ctx->ff_page), 0xff, PAGE_SIZE);
161 			}
162 			get_page(ctx->ff_page);
163 			vmf->page = ctx->ff_page;
164 			vma->vm_page_prot = pgprot_cached(vma->vm_page_prot);
165 			return 0;
166 		}
167 		return VM_FAULT_SIGBUS;
168 	}
169 
170 	vm_insert_pfn(vma, vmf->address, (area + offset) >> PAGE_SHIFT);
171 
172 	mutex_unlock(&ctx->status_mutex);
173 
174 	return VM_FAULT_NOPAGE;
175 }
176 
177 static const struct vm_operations_struct cxl_mmap_vmops = {
178 	.fault = cxl_mmap_fault,
179 };
180 
181 /*
182  * Map a per-context mmio space into the given vma.
183  */
184 int cxl_context_iomap(struct cxl_context *ctx, struct vm_area_struct *vma)
185 {
186 	u64 start = vma->vm_pgoff << PAGE_SHIFT;
187 	u64 len = vma->vm_end - vma->vm_start;
188 
189 	if (ctx->afu->current_mode == CXL_MODE_DEDICATED) {
190 		if (start + len > ctx->afu->adapter->ps_size)
191 			return -EINVAL;
192 
193 		if (cxl_is_power9()) {
194 			/*
195 			 * Make sure there is a valid problem state
196 			 * area space for this AFU.
197 			 */
198 			if (ctx->master && !ctx->afu->psa) {
199 				pr_devel("AFU doesn't support mmio space\n");
200 				return -EINVAL;
201 			}
202 
203 			/* Can't mmap until the AFU is enabled */
204 			if (!ctx->afu->enabled)
205 				return -EBUSY;
206 		}
207 	} else {
208 		if (start + len > ctx->psn_size)
209 			return -EINVAL;
210 
211 		/* Make sure there is a valid per process space for this AFU */
212 		if ((ctx->master && !ctx->afu->psa) || (!ctx->afu->pp_psa)) {
213 			pr_devel("AFU doesn't support mmio space\n");
214 			return -EINVAL;
215 		}
216 
217 		/* Can't mmap until the AFU is enabled */
218 		if (!ctx->afu->enabled)
219 			return -EBUSY;
220 	}
221 
222 	pr_devel("%s: mmio physical: %llx pe: %i master:%i\n", __func__,
223 		 ctx->psn_phys, ctx->pe , ctx->master);
224 
225 	vma->vm_flags |= VM_IO | VM_PFNMAP;
226 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
227 	vma->vm_ops = &cxl_mmap_vmops;
228 	return 0;
229 }
230 
231 /*
232  * Detach a context from the hardware. This disables interrupts and doesn't
233  * return until all outstanding interrupts for this context have completed. The
234  * hardware should no longer access *ctx after this has returned.
235  */
236 int __detach_context(struct cxl_context *ctx)
237 {
238 	enum cxl_context_status status;
239 
240 	mutex_lock(&ctx->status_mutex);
241 	status = ctx->status;
242 	ctx->status = CLOSED;
243 	mutex_unlock(&ctx->status_mutex);
244 	if (status != STARTED)
245 		return -EBUSY;
246 
247 	/* Only warn if we detached while the link was OK.
248 	 * If detach fails when hw is down, we don't care.
249 	 */
250 	WARN_ON(cxl_ops->detach_process(ctx) &&
251 		cxl_ops->link_ok(ctx->afu->adapter, ctx->afu));
252 	flush_work(&ctx->fault_work); /* Only needed for dedicated process */
253 
254 	/*
255 	 * Wait until no further interrupts are presented by the PSL
256 	 * for this context.
257 	 */
258 	if (cxl_ops->irq_wait)
259 		cxl_ops->irq_wait(ctx);
260 
261 	/* release the reference to the group leader and mm handling pid */
262 	put_pid(ctx->pid);
263 
264 	cxl_ctx_put();
265 
266 	/* Decrease the attached context count on the adapter */
267 	cxl_adapter_context_put(ctx->afu->adapter);
268 
269 	/* Decrease the mm count on the context */
270 	cxl_context_mm_count_put(ctx);
271 	if (ctx->mm)
272 		mm_context_remove_copro(ctx->mm);
273 	ctx->mm = NULL;
274 
275 	return 0;
276 }
277 
278 /*
279  * Detach the given context from the AFU. This doesn't actually
280  * free the context but it should stop the context running in hardware
281  * (ie. prevent this context from generating any further interrupts
282  * so that it can be freed).
283  */
284 void cxl_context_detach(struct cxl_context *ctx)
285 {
286 	int rc;
287 
288 	rc = __detach_context(ctx);
289 	if (rc)
290 		return;
291 
292 	afu_release_irqs(ctx, ctx);
293 	wake_up_all(&ctx->wq);
294 }
295 
296 /*
297  * Detach all contexts on the given AFU.
298  */
299 void cxl_context_detach_all(struct cxl_afu *afu)
300 {
301 	struct cxl_context *ctx;
302 	int tmp;
303 
304 	mutex_lock(&afu->contexts_lock);
305 	idr_for_each_entry(&afu->contexts_idr, ctx, tmp) {
306 		/*
307 		 * Anything done in here needs to be setup before the IDR is
308 		 * created and torn down after the IDR removed
309 		 */
310 		cxl_context_detach(ctx);
311 
312 		/*
313 		 * We are force detaching - remove any active PSA mappings so
314 		 * userspace cannot interfere with the card if it comes back.
315 		 * Easiest way to exercise this is to unbind and rebind the
316 		 * driver via sysfs while it is in use.
317 		 */
318 		mutex_lock(&ctx->mapping_lock);
319 		if (ctx->mapping)
320 			unmap_mapping_range(ctx->mapping, 0, 0, 1);
321 		mutex_unlock(&ctx->mapping_lock);
322 	}
323 	mutex_unlock(&afu->contexts_lock);
324 }
325 
326 static void reclaim_ctx(struct rcu_head *rcu)
327 {
328 	struct cxl_context *ctx = container_of(rcu, struct cxl_context, rcu);
329 
330 	if (cxl_is_power8())
331 		free_page((u64)ctx->sstp);
332 	if (ctx->ff_page)
333 		__free_page(ctx->ff_page);
334 	ctx->sstp = NULL;
335 
336 	kfree(ctx->irq_bitmap);
337 
338 	/* Drop ref to the afu device taken during cxl_context_init */
339 	cxl_afu_put(ctx->afu);
340 
341 	kfree(ctx);
342 }
343 
344 void cxl_context_free(struct cxl_context *ctx)
345 {
346 	if (ctx->kernelapi && ctx->mapping)
347 		cxl_release_mapping(ctx);
348 	mutex_lock(&ctx->afu->contexts_lock);
349 	idr_remove(&ctx->afu->contexts_idr, ctx->pe);
350 	mutex_unlock(&ctx->afu->contexts_lock);
351 	call_rcu(&ctx->rcu, reclaim_ctx);
352 }
353 
354 void cxl_context_mm_count_get(struct cxl_context *ctx)
355 {
356 	if (ctx->mm)
357 		atomic_inc(&ctx->mm->mm_count);
358 }
359 
360 void cxl_context_mm_count_put(struct cxl_context *ctx)
361 {
362 	if (ctx->mm)
363 		mmdrop(ctx->mm);
364 }
365