xref: /linux/drivers/misc/cxl/fault.c (revision e58e871becec2d3b04ed91c0c16fe8deac9c9dfa)
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/workqueue.h>
11 #include <linux/sched/signal.h>
12 #include <linux/sched/mm.h>
13 #include <linux/pid.h>
14 #include <linux/mm.h>
15 #include <linux/moduleparam.h>
16 
17 #undef MODULE_PARAM_PREFIX
18 #define MODULE_PARAM_PREFIX "cxl" "."
19 #include <asm/current.h>
20 #include <asm/copro.h>
21 #include <asm/mmu.h>
22 
23 #include "cxl.h"
24 #include "trace.h"
25 
26 static bool sste_matches(struct cxl_sste *sste, struct copro_slb *slb)
27 {
28 	return ((sste->vsid_data == cpu_to_be64(slb->vsid)) &&
29 		(sste->esid_data == cpu_to_be64(slb->esid)));
30 }
31 
32 /*
33  * This finds a free SSTE for the given SLB, or returns NULL if it's already in
34  * the segment table.
35  */
36 static struct cxl_sste* find_free_sste(struct cxl_context *ctx,
37 				       struct copro_slb *slb)
38 {
39 	struct cxl_sste *primary, *sste, *ret = NULL;
40 	unsigned int mask = (ctx->sst_size >> 7) - 1; /* SSTP0[SegTableSize] */
41 	unsigned int entry;
42 	unsigned int hash;
43 
44 	if (slb->vsid & SLB_VSID_B_1T)
45 		hash = (slb->esid >> SID_SHIFT_1T) & mask;
46 	else /* 256M */
47 		hash = (slb->esid >> SID_SHIFT) & mask;
48 
49 	primary = ctx->sstp + (hash << 3);
50 
51 	for (entry = 0, sste = primary; entry < 8; entry++, sste++) {
52 		if (!ret && !(be64_to_cpu(sste->esid_data) & SLB_ESID_V))
53 			ret = sste;
54 		if (sste_matches(sste, slb))
55 			return NULL;
56 	}
57 	if (ret)
58 		return ret;
59 
60 	/* Nothing free, select an entry to cast out */
61 	ret = primary + ctx->sst_lru;
62 	ctx->sst_lru = (ctx->sst_lru + 1) & 0x7;
63 
64 	return ret;
65 }
66 
67 static void cxl_load_segment(struct cxl_context *ctx, struct copro_slb *slb)
68 {
69 	/* mask is the group index, we search primary and secondary here. */
70 	struct cxl_sste *sste;
71 	unsigned long flags;
72 
73 	spin_lock_irqsave(&ctx->sste_lock, flags);
74 	sste = find_free_sste(ctx, slb);
75 	if (!sste)
76 		goto out_unlock;
77 
78 	pr_devel("CXL Populating SST[%li]: %#llx %#llx\n",
79 			sste - ctx->sstp, slb->vsid, slb->esid);
80 	trace_cxl_ste_write(ctx, sste - ctx->sstp, slb->esid, slb->vsid);
81 
82 	sste->vsid_data = cpu_to_be64(slb->vsid);
83 	sste->esid_data = cpu_to_be64(slb->esid);
84 out_unlock:
85 	spin_unlock_irqrestore(&ctx->sste_lock, flags);
86 }
87 
88 static int cxl_fault_segment(struct cxl_context *ctx, struct mm_struct *mm,
89 			     u64 ea)
90 {
91 	struct copro_slb slb = {0,0};
92 	int rc;
93 
94 	if (!(rc = copro_calculate_slb(mm, ea, &slb))) {
95 		cxl_load_segment(ctx, &slb);
96 	}
97 
98 	return rc;
99 }
100 
101 static void cxl_ack_ae(struct cxl_context *ctx)
102 {
103 	unsigned long flags;
104 
105 	cxl_ops->ack_irq(ctx, CXL_PSL_TFC_An_AE, 0);
106 
107 	spin_lock_irqsave(&ctx->lock, flags);
108 	ctx->pending_fault = true;
109 	ctx->fault_addr = ctx->dar;
110 	ctx->fault_dsisr = ctx->dsisr;
111 	spin_unlock_irqrestore(&ctx->lock, flags);
112 
113 	wake_up_all(&ctx->wq);
114 }
115 
116 static int cxl_handle_segment_miss(struct cxl_context *ctx,
117 				   struct mm_struct *mm, u64 ea)
118 {
119 	int rc;
120 
121 	pr_devel("CXL interrupt: Segment fault pe: %i ea: %#llx\n", ctx->pe, ea);
122 	trace_cxl_ste_miss(ctx, ea);
123 
124 	if ((rc = cxl_fault_segment(ctx, mm, ea)))
125 		cxl_ack_ae(ctx);
126 	else {
127 
128 		mb(); /* Order seg table write to TFC MMIO write */
129 		cxl_ops->ack_irq(ctx, CXL_PSL_TFC_An_R, 0);
130 	}
131 
132 	return IRQ_HANDLED;
133 }
134 
135 static void cxl_handle_page_fault(struct cxl_context *ctx,
136 				  struct mm_struct *mm, u64 dsisr, u64 dar)
137 {
138 	unsigned flt = 0;
139 	int result;
140 	unsigned long access, flags, inv_flags = 0;
141 
142 	trace_cxl_pte_miss(ctx, dsisr, dar);
143 
144 	if ((result = copro_handle_mm_fault(mm, dar, dsisr, &flt))) {
145 		pr_devel("copro_handle_mm_fault failed: %#x\n", result);
146 		return cxl_ack_ae(ctx);
147 	}
148 
149 	if (!radix_enabled()) {
150 		/*
151 		 * update_mmu_cache() will not have loaded the hash since current->trap
152 		 * is not a 0x400 or 0x300, so just call hash_page_mm() here.
153 		 */
154 		access = _PAGE_PRESENT | _PAGE_READ;
155 		if (dsisr & CXL_PSL_DSISR_An_S)
156 			access |= _PAGE_WRITE;
157 
158 		access |= _PAGE_PRIVILEGED;
159 		if ((!ctx->kernel) || (REGION_ID(dar) == USER_REGION_ID))
160 			access &= ~_PAGE_PRIVILEGED;
161 
162 		if (dsisr & DSISR_NOHPTE)
163 			inv_flags |= HPTE_NOHPTE_UPDATE;
164 
165 		local_irq_save(flags);
166 		hash_page_mm(mm, dar, access, 0x300, inv_flags);
167 		local_irq_restore(flags);
168 	}
169 	pr_devel("Page fault successfully handled for pe: %i!\n", ctx->pe);
170 	cxl_ops->ack_irq(ctx, CXL_PSL_TFC_An_R, 0);
171 }
172 
173 /*
174  * Returns the mm_struct corresponding to the context ctx.
175  * mm_users == 0, the context may be in the process of being closed.
176  */
177 static struct mm_struct *get_mem_context(struct cxl_context *ctx)
178 {
179 	if (ctx->mm == NULL)
180 		return NULL;
181 
182 	if (!atomic_inc_not_zero(&ctx->mm->mm_users))
183 		return NULL;
184 
185 	return ctx->mm;
186 }
187 
188 static bool cxl_is_segment_miss(struct cxl_context *ctx, u64 dsisr)
189 {
190 	if ((cxl_is_psl8(ctx->afu)) && (dsisr & CXL_PSL_DSISR_An_DS))
191 		return true;
192 
193 	return false;
194 }
195 
196 static bool cxl_is_page_fault(struct cxl_context *ctx, u64 dsisr)
197 {
198 	if ((cxl_is_psl8(ctx->afu)) && (dsisr & CXL_PSL_DSISR_An_DM))
199 		return true;
200 
201 	if ((cxl_is_psl9(ctx->afu)) &&
202 	   ((dsisr & CXL_PSL9_DSISR_An_CO_MASK) &
203 		(CXL_PSL9_DSISR_An_PF_SLR | CXL_PSL9_DSISR_An_PF_RGC |
204 		 CXL_PSL9_DSISR_An_PF_RGP | CXL_PSL9_DSISR_An_PF_HRH |
205 		 CXL_PSL9_DSISR_An_PF_STEG)))
206 		return true;
207 
208 	return false;
209 }
210 
211 void cxl_handle_fault(struct work_struct *fault_work)
212 {
213 	struct cxl_context *ctx =
214 		container_of(fault_work, struct cxl_context, fault_work);
215 	u64 dsisr = ctx->dsisr;
216 	u64 dar = ctx->dar;
217 	struct mm_struct *mm = NULL;
218 
219 	if (cpu_has_feature(CPU_FTR_HVMODE)) {
220 		if (cxl_p2n_read(ctx->afu, CXL_PSL_DSISR_An) != dsisr ||
221 		    cxl_p2n_read(ctx->afu, CXL_PSL_DAR_An) != dar ||
222 		    cxl_p2n_read(ctx->afu, CXL_PSL_PEHandle_An) != ctx->pe) {
223 			/* Most likely explanation is harmless - a dedicated
224 			 * process has detached and these were cleared by the
225 			 * PSL purge, but warn about it just in case
226 			 */
227 			dev_notice(&ctx->afu->dev, "cxl_handle_fault: Translation fault regs changed\n");
228 			return;
229 		}
230 	}
231 
232 	/* Early return if the context is being / has been detached */
233 	if (ctx->status == CLOSED) {
234 		cxl_ack_ae(ctx);
235 		return;
236 	}
237 
238 	pr_devel("CXL BOTTOM HALF handling fault for afu pe: %i. "
239 		"DSISR: %#llx DAR: %#llx\n", ctx->pe, dsisr, dar);
240 
241 	if (!ctx->kernel) {
242 
243 		mm = get_mem_context(ctx);
244 		if (mm == NULL) {
245 			pr_devel("%s: unable to get mm for pe=%d pid=%i\n",
246 				 __func__, ctx->pe, pid_nr(ctx->pid));
247 			cxl_ack_ae(ctx);
248 			return;
249 		} else {
250 			pr_devel("Handling page fault for pe=%d pid=%i\n",
251 				 ctx->pe, pid_nr(ctx->pid));
252 		}
253 	}
254 
255 	if (cxl_is_segment_miss(ctx, dsisr))
256 		cxl_handle_segment_miss(ctx, mm, dar);
257 	else if (cxl_is_page_fault(ctx, dsisr))
258 		cxl_handle_page_fault(ctx, mm, dsisr, dar);
259 	else
260 		WARN(1, "cxl_handle_fault has nothing to handle\n");
261 
262 	if (mm)
263 		mmput(mm);
264 }
265 
266 static void cxl_prefault_one(struct cxl_context *ctx, u64 ea)
267 {
268 	struct mm_struct *mm;
269 
270 	mm = get_mem_context(ctx);
271 	if (mm == NULL) {
272 		pr_devel("cxl_prefault_one unable to get mm %i\n",
273 			 pid_nr(ctx->pid));
274 		return;
275 	}
276 
277 	cxl_fault_segment(ctx, mm, ea);
278 
279 	mmput(mm);
280 }
281 
282 static u64 next_segment(u64 ea, u64 vsid)
283 {
284 	if (vsid & SLB_VSID_B_1T)
285 		ea |= (1ULL << 40) - 1;
286 	else
287 		ea |= (1ULL << 28) - 1;
288 
289 	return ea + 1;
290 }
291 
292 static void cxl_prefault_vma(struct cxl_context *ctx)
293 {
294 	u64 ea, last_esid = 0;
295 	struct copro_slb slb;
296 	struct vm_area_struct *vma;
297 	int rc;
298 	struct mm_struct *mm;
299 
300 	mm = get_mem_context(ctx);
301 	if (mm == NULL) {
302 		pr_devel("cxl_prefault_vm unable to get mm %i\n",
303 			 pid_nr(ctx->pid));
304 		return;
305 	}
306 
307 	down_read(&mm->mmap_sem);
308 	for (vma = mm->mmap; vma; vma = vma->vm_next) {
309 		for (ea = vma->vm_start; ea < vma->vm_end;
310 				ea = next_segment(ea, slb.vsid)) {
311 			rc = copro_calculate_slb(mm, ea, &slb);
312 			if (rc)
313 				continue;
314 
315 			if (last_esid == slb.esid)
316 				continue;
317 
318 			cxl_load_segment(ctx, &slb);
319 			last_esid = slb.esid;
320 		}
321 	}
322 	up_read(&mm->mmap_sem);
323 
324 	mmput(mm);
325 }
326 
327 void cxl_prefault(struct cxl_context *ctx, u64 wed)
328 {
329 	switch (ctx->afu->prefault_mode) {
330 	case CXL_PREFAULT_WED:
331 		cxl_prefault_one(ctx, wed);
332 		break;
333 	case CXL_PREFAULT_ALL:
334 		cxl_prefault_vma(ctx);
335 		break;
336 	default:
337 		break;
338 	}
339 }
340