1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3  * Copyright 2014 IBM Corp.
4  */
5 
6 #include <linux/module.h>
7 #include <linux/kernel.h>
8 #include <linux/bitmap.h>
9 #include <linux/sched.h>
10 #include <linux/pid.h>
11 #include <linux/fs.h>
12 #include <linux/mm.h>
13 #include <linux/debugfs.h>
14 #include <linux/slab.h>
15 #include <linux/idr.h>
16 #include <linux/sched/mm.h>
17 #include <linux/mmu_context.h>
18 #include <asm/cputable.h>
19 #include <asm/current.h>
20 #include <asm/copro.h>
21 
22 #include "cxl.h"
23 
24 /*
25  * Allocates space for a CXL context.
26  */
cxl_context_alloc(void)27 struct cxl_context *cxl_context_alloc(void)
28 {
29 	return kzalloc(sizeof(struct cxl_context), GFP_KERNEL);
30 }
31 
32 /*
33  * Initialises a CXL context.
34  */
cxl_context_init(struct cxl_context * ctx,struct cxl_afu * afu,bool master)35 int cxl_context_init(struct cxl_context *ctx, struct cxl_afu *afu, bool master)
36 {
37 	int i;
38 
39 	ctx->afu = afu;
40 	ctx->master = master;
41 	ctx->pid = NULL; /* Set in start work ioctl */
42 	mutex_init(&ctx->mapping_lock);
43 	ctx->mapping = NULL;
44 	ctx->tidr = 0;
45 	ctx->assign_tidr = false;
46 
47 	if (cxl_is_power8()) {
48 		spin_lock_init(&ctx->sste_lock);
49 
50 		/*
51 		 * Allocate the segment table before we put it in the IDR so that we
52 		 * can always access it when dereferenced from IDR. For the same
53 		 * reason, the segment table is only destroyed after the context is
54 		 * removed from the IDR.  Access to this in the IOCTL is protected by
55 		 * Linux filesystem semantics (can't IOCTL until open is complete).
56 		 */
57 		i = cxl_alloc_sst(ctx);
58 		if (i)
59 			return i;
60 	}
61 
62 	INIT_WORK(&ctx->fault_work, cxl_handle_fault);
63 
64 	init_waitqueue_head(&ctx->wq);
65 	spin_lock_init(&ctx->lock);
66 
67 	ctx->irq_bitmap = NULL;
68 	ctx->pending_irq = false;
69 	ctx->pending_fault = false;
70 	ctx->pending_afu_err = false;
71 
72 	INIT_LIST_HEAD(&ctx->irq_names);
73 
74 	/*
75 	 * When we have to destroy all contexts in cxl_context_detach_all() we
76 	 * end up with afu_release_irqs() called from inside a
77 	 * idr_for_each_entry(). Hence we need to make sure that anything
78 	 * dereferenced from this IDR is ok before we allocate the IDR here.
79 	 * This clears out the IRQ ranges to ensure this.
80 	 */
81 	for (i = 0; i < CXL_IRQ_RANGES; i++)
82 		ctx->irqs.range[i] = 0;
83 
84 	mutex_init(&ctx->status_mutex);
85 
86 	ctx->status = OPENED;
87 
88 	/*
89 	 * Allocating IDR! We better make sure everything's setup that
90 	 * dereferences from it.
91 	 */
92 	mutex_lock(&afu->contexts_lock);
93 	idr_preload(GFP_KERNEL);
94 	i = idr_alloc(&ctx->afu->contexts_idr, ctx, 0,
95 		      ctx->afu->num_procs, GFP_NOWAIT);
96 	idr_preload_end();
97 	mutex_unlock(&afu->contexts_lock);
98 	if (i < 0)
99 		return i;
100 
101 	ctx->pe = i;
102 	if (cpu_has_feature(CPU_FTR_HVMODE)) {
103 		ctx->elem = &ctx->afu->native->spa[i];
104 		ctx->external_pe = ctx->pe;
105 	} else {
106 		ctx->external_pe = -1; /* assigned when attaching */
107 	}
108 	ctx->pe_inserted = false;
109 
110 	/*
111 	 * take a ref on the afu so that it stays alive at-least till
112 	 * this context is reclaimed inside reclaim_ctx.
113 	 */
114 	cxl_afu_get(afu);
115 	return 0;
116 }
117 
cxl_context_set_mapping(struct cxl_context * ctx,struct address_space * mapping)118 void cxl_context_set_mapping(struct cxl_context *ctx,
119 			struct address_space *mapping)
120 {
121 	mutex_lock(&ctx->mapping_lock);
122 	ctx->mapping = mapping;
123 	mutex_unlock(&ctx->mapping_lock);
124 }
125 
cxl_mmap_fault(struct vm_fault * vmf)126 static vm_fault_t cxl_mmap_fault(struct vm_fault *vmf)
127 {
128 	struct vm_area_struct *vma = vmf->vma;
129 	struct cxl_context *ctx = vma->vm_file->private_data;
130 	u64 area, offset;
131 	vm_fault_t ret;
132 
133 	offset = vmf->pgoff << PAGE_SHIFT;
134 
135 	pr_devel("%s: pe: %i address: 0x%lx offset: 0x%llx\n",
136 			__func__, ctx->pe, vmf->address, offset);
137 
138 	if (ctx->afu->current_mode == CXL_MODE_DEDICATED) {
139 		area = ctx->afu->psn_phys;
140 		if (offset >= ctx->afu->adapter->ps_size)
141 			return VM_FAULT_SIGBUS;
142 	} else {
143 		area = ctx->psn_phys;
144 		if (offset >= ctx->psn_size)
145 			return VM_FAULT_SIGBUS;
146 	}
147 
148 	mutex_lock(&ctx->status_mutex);
149 
150 	if (ctx->status != STARTED) {
151 		mutex_unlock(&ctx->status_mutex);
152 		pr_devel("%s: Context not started, failing problem state access\n", __func__);
153 		if (ctx->mmio_err_ff) {
154 			if (!ctx->ff_page) {
155 				ctx->ff_page = alloc_page(GFP_USER);
156 				if (!ctx->ff_page)
157 					return VM_FAULT_OOM;
158 				memset(page_address(ctx->ff_page), 0xff, PAGE_SIZE);
159 			}
160 			get_page(ctx->ff_page);
161 			vmf->page = ctx->ff_page;
162 			vma->vm_page_prot = pgprot_cached(vma->vm_page_prot);
163 			return 0;
164 		}
165 		return VM_FAULT_SIGBUS;
166 	}
167 
168 	ret = vmf_insert_pfn(vma, vmf->address, (area + offset) >> PAGE_SHIFT);
169 
170 	mutex_unlock(&ctx->status_mutex);
171 
172 	return ret;
173 }
174 
175 static const struct vm_operations_struct cxl_mmap_vmops = {
176 	.fault = cxl_mmap_fault,
177 };
178 
179 /*
180  * Map a per-context mmio space into the given vma.
181  */
cxl_context_iomap(struct cxl_context * ctx,struct vm_area_struct * vma)182 int cxl_context_iomap(struct cxl_context *ctx, struct vm_area_struct *vma)
183 {
184 	u64 start = vma->vm_pgoff << PAGE_SHIFT;
185 	u64 len = vma->vm_end - vma->vm_start;
186 
187 	if (ctx->afu->current_mode == CXL_MODE_DEDICATED) {
188 		if (start + len > ctx->afu->adapter->ps_size)
189 			return -EINVAL;
190 
191 		if (cxl_is_power9()) {
192 			/*
193 			 * Make sure there is a valid problem state
194 			 * area space for this AFU.
195 			 */
196 			if (ctx->master && !ctx->afu->psa) {
197 				pr_devel("AFU doesn't support mmio space\n");
198 				return -EINVAL;
199 			}
200 
201 			/* Can't mmap until the AFU is enabled */
202 			if (!ctx->afu->enabled)
203 				return -EBUSY;
204 		}
205 	} else {
206 		if (start + len > ctx->psn_size)
207 			return -EINVAL;
208 
209 		/* Make sure there is a valid per process space for this AFU */
210 		if ((ctx->master && !ctx->afu->psa) || (!ctx->afu->pp_psa)) {
211 			pr_devel("AFU doesn't support mmio space\n");
212 			return -EINVAL;
213 		}
214 
215 		/* Can't mmap until the AFU is enabled */
216 		if (!ctx->afu->enabled)
217 			return -EBUSY;
218 	}
219 
220 	pr_devel("%s: mmio physical: %llx pe: %i master:%i\n", __func__,
221 		 ctx->psn_phys, ctx->pe , ctx->master);
222 
223 	vma->vm_flags |= VM_IO | VM_PFNMAP;
224 	vma->vm_page_prot = pgprot_noncached(vma->vm_page_prot);
225 	vma->vm_ops = &cxl_mmap_vmops;
226 	return 0;
227 }
228 
229 /*
230  * Detach a context from the hardware. This disables interrupts and doesn't
231  * return until all outstanding interrupts for this context have completed. The
232  * hardware should no longer access *ctx after this has returned.
233  */
__detach_context(struct cxl_context * ctx)234 int __detach_context(struct cxl_context *ctx)
235 {
236 	enum cxl_context_status status;
237 
238 	mutex_lock(&ctx->status_mutex);
239 	status = ctx->status;
240 	ctx->status = CLOSED;
241 	mutex_unlock(&ctx->status_mutex);
242 	if (status != STARTED)
243 		return -EBUSY;
244 
245 	/* Only warn if we detached while the link was OK.
246 	 * If detach fails when hw is down, we don't care.
247 	 */
248 	WARN_ON(cxl_ops->detach_process(ctx) &&
249 		cxl_ops->link_ok(ctx->afu->adapter, ctx->afu));
250 	flush_work(&ctx->fault_work); /* Only needed for dedicated process */
251 
252 	/*
253 	 * Wait until no further interrupts are presented by the PSL
254 	 * for this context.
255 	 */
256 	if (cxl_ops->irq_wait)
257 		cxl_ops->irq_wait(ctx);
258 
259 	/* release the reference to the group leader and mm handling pid */
260 	put_pid(ctx->pid);
261 
262 	cxl_ctx_put();
263 
264 	/* Decrease the attached context count on the adapter */
265 	cxl_adapter_context_put(ctx->afu->adapter);
266 
267 	/* Decrease the mm count on the context */
268 	cxl_context_mm_count_put(ctx);
269 	if (ctx->mm)
270 		mm_context_remove_copro(ctx->mm);
271 	ctx->mm = NULL;
272 
273 	return 0;
274 }
275 
276 /*
277  * Detach the given context from the AFU. This doesn't actually
278  * free the context but it should stop the context running in hardware
279  * (ie. prevent this context from generating any further interrupts
280  * so that it can be freed).
281  */
cxl_context_detach(struct cxl_context * ctx)282 void cxl_context_detach(struct cxl_context *ctx)
283 {
284 	int rc;
285 
286 	rc = __detach_context(ctx);
287 	if (rc)
288 		return;
289 
290 	afu_release_irqs(ctx, ctx);
291 	wake_up_all(&ctx->wq);
292 }
293 
294 /*
295  * Detach all contexts on the given AFU.
296  */
cxl_context_detach_all(struct cxl_afu * afu)297 void cxl_context_detach_all(struct cxl_afu *afu)
298 {
299 	struct cxl_context *ctx;
300 	int tmp;
301 
302 	mutex_lock(&afu->contexts_lock);
303 	idr_for_each_entry(&afu->contexts_idr, ctx, tmp) {
304 		/*
305 		 * Anything done in here needs to be setup before the IDR is
306 		 * created and torn down after the IDR removed
307 		 */
308 		cxl_context_detach(ctx);
309 
310 		/*
311 		 * We are force detaching - remove any active PSA mappings so
312 		 * userspace cannot interfere with the card if it comes back.
313 		 * Easiest way to exercise this is to unbind and rebind the
314 		 * driver via sysfs while it is in use.
315 		 */
316 		mutex_lock(&ctx->mapping_lock);
317 		if (ctx->mapping)
318 			unmap_mapping_range(ctx->mapping, 0, 0, 1);
319 		mutex_unlock(&ctx->mapping_lock);
320 	}
321 	mutex_unlock(&afu->contexts_lock);
322 }
323 
reclaim_ctx(struct rcu_head * rcu)324 static void reclaim_ctx(struct rcu_head *rcu)
325 {
326 	struct cxl_context *ctx = container_of(rcu, struct cxl_context, rcu);
327 
328 	if (cxl_is_power8())
329 		free_page((u64)ctx->sstp);
330 	if (ctx->ff_page)
331 		__free_page(ctx->ff_page);
332 	ctx->sstp = NULL;
333 
334 	bitmap_free(ctx->irq_bitmap);
335 
336 	/* Drop ref to the afu device taken during cxl_context_init */
337 	cxl_afu_put(ctx->afu);
338 
339 	kfree(ctx);
340 }
341 
cxl_context_free(struct cxl_context * ctx)342 void cxl_context_free(struct cxl_context *ctx)
343 {
344 	if (ctx->kernelapi && ctx->mapping)
345 		cxl_release_mapping(ctx);
346 	mutex_lock(&ctx->afu->contexts_lock);
347 	idr_remove(&ctx->afu->contexts_idr, ctx->pe);
348 	mutex_unlock(&ctx->afu->contexts_lock);
349 	call_rcu(&ctx->rcu, reclaim_ctx);
350 }
351 
cxl_context_mm_count_get(struct cxl_context * ctx)352 void cxl_context_mm_count_get(struct cxl_context *ctx)
353 {
354 	if (ctx->mm)
355 		mmgrab(ctx->mm);
356 }
357 
cxl_context_mm_count_put(struct cxl_context * ctx)358 void cxl_context_mm_count_put(struct cxl_context *ctx)
359 {
360 	if (ctx->mm)
361 		mmdrop(ctx->mm);
362 }
363