2 * Copyright 2014 IBM Corp.
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.
10 #include <linux/interrupt.h>
11 #include <linux/workqueue.h>
12 #include <linux/sched.h>
13 #include <linux/wait.h>
14 #include <linux/slab.h>
15 #include <linux/pid.h>
16 #include <asm/cputable.h>
17 #include <misc/cxl-base.h>
22 static int afu_irq_range_start(void)
24 if (cpu_has_feature(CPU_FTR_HVMODE))
29 static irqreturn_t schedule_cxl_fault(struct cxl_context *ctx, u64 dsisr, u64 dar)
33 schedule_work(&ctx->fault_work);
37 irqreturn_t cxl_irq(int irq, struct cxl_context *ctx, struct cxl_irq_info *irq_info)
41 dsisr = irq_info->dsisr;
44 trace_cxl_psl_irq(ctx, irq, dsisr, dar);
46 pr_devel("CXL interrupt %i for afu pe: %i DSISR: %#llx DAR: %#llx\n", irq, ctx->pe, dsisr, dar);
48 if (dsisr & CXL_PSL_DSISR_An_DS) {
50 * We don't inherently need to sleep to handle this, but we do
51 * need to get a ref to the task's mm, which we can't do from
52 * irq context without the potential for a deadlock since it
53 * takes the task_lock. An alternate option would be to keep a
54 * reference to the task's mm the entire time it has cxl open,
55 * but to do that we need to solve the issue where we hold a
56 * ref to the mm, but the mm can hold a ref to the fd after an
57 * mmap preventing anything from being cleaned up.
59 pr_devel("Scheduling segment miss handling for later pe: %i\n", ctx->pe);
60 return schedule_cxl_fault(ctx, dsisr, dar);
63 if (dsisr & CXL_PSL_DSISR_An_M)
64 pr_devel("CXL interrupt: PTE not found\n");
65 if (dsisr & CXL_PSL_DSISR_An_P)
66 pr_devel("CXL interrupt: Storage protection violation\n");
67 if (dsisr & CXL_PSL_DSISR_An_A)
68 pr_devel("CXL interrupt: AFU lock access to write through or cache inhibited storage\n");
69 if (dsisr & CXL_PSL_DSISR_An_S)
70 pr_devel("CXL interrupt: Access was afu_wr or afu_zero\n");
71 if (dsisr & CXL_PSL_DSISR_An_K)
72 pr_devel("CXL interrupt: Access not permitted by virtual page class key protection\n");
74 if (dsisr & CXL_PSL_DSISR_An_DM) {
76 * In some cases we might be able to handle the fault
77 * immediately if hash_page would succeed, but we still need
78 * the task's mm, which as above we can't get without a lock
80 pr_devel("Scheduling page fault handling for later pe: %i\n", ctx->pe);
81 return schedule_cxl_fault(ctx, dsisr, dar);
83 if (dsisr & CXL_PSL_DSISR_An_ST)
84 WARN(1, "CXL interrupt: Segment Table PTE not found\n");
85 if (dsisr & CXL_PSL_DSISR_An_UR)
86 pr_devel("CXL interrupt: AURP PTE not found\n");
87 if (dsisr & CXL_PSL_DSISR_An_PE)
88 return cxl_ops->handle_psl_slice_error(ctx, dsisr,
90 if (dsisr & CXL_PSL_DSISR_An_AE) {
91 pr_devel("CXL interrupt: AFU Error 0x%016llx\n", irq_info->afu_err);
93 if (ctx->pending_afu_err) {
95 * This shouldn't happen - the PSL treats these errors
96 * as fatal and will have reset the AFU, so there's not
97 * much point buffering multiple AFU errors.
98 * OTOH if we DO ever see a storm of these come in it's
99 * probably best that we log them somewhere:
101 dev_err_ratelimited(&ctx->afu->dev, "CXL AFU Error "
102 "undelivered to pe %i: 0x%016llx\n",
103 ctx->pe, irq_info->afu_err);
105 spin_lock(&ctx->lock);
106 ctx->afu_err = irq_info->afu_err;
107 ctx->pending_afu_err = 1;
108 spin_unlock(&ctx->lock);
110 wake_up_all(&ctx->wq);
113 cxl_ops->ack_irq(ctx, CXL_PSL_TFC_An_A, 0);
116 if (dsisr & CXL_PSL_DSISR_An_OC)
117 pr_devel("CXL interrupt: OS Context Warning\n");
119 WARN(1, "Unhandled CXL PSL IRQ\n");
123 static irqreturn_t cxl_irq_afu(int irq, void *data)
125 struct cxl_context *ctx = data;
126 irq_hw_number_t hwirq = irqd_to_hwirq(irq_get_irq_data(irq));
127 int irq_off, afu_irq = 0;
132 * Look for the interrupt number.
133 * On bare-metal, we know range 0 only contains the PSL
134 * interrupt so we could start counting at range 1 and initialize
136 * In a guest, range 0 also contains AFU interrupts, so it must
137 * be counted for. Therefore we initialize afu_irq at 0 to take into
138 * account the PSL interrupt.
140 * For code-readability, it just seems easier to go over all
141 * the ranges on bare-metal and guest. The end result is the same.
143 for (r = 0; r < CXL_IRQ_RANGES; r++) {
144 irq_off = hwirq - ctx->irqs.offset[r];
145 range = ctx->irqs.range[r];
146 if (irq_off >= 0 && irq_off < range) {
152 if (unlikely(r >= CXL_IRQ_RANGES)) {
153 WARN(1, "Received AFU IRQ out of range for pe %i (virq %i hwirq %lx)\n",
154 ctx->pe, irq, hwirq);
158 trace_cxl_afu_irq(ctx, afu_irq, irq, hwirq);
159 pr_devel("Received AFU interrupt %i for pe: %i (virq %i hwirq %lx)\n",
160 afu_irq, ctx->pe, irq, hwirq);
162 if (unlikely(!ctx->irq_bitmap)) {
163 WARN(1, "Received AFU IRQ for context with no IRQ bitmap\n");
166 spin_lock(&ctx->lock);
167 set_bit(afu_irq - 1, ctx->irq_bitmap);
168 ctx->pending_irq = true;
169 spin_unlock(&ctx->lock);
171 wake_up_all(&ctx->wq);
176 unsigned int cxl_map_irq(struct cxl *adapter, irq_hw_number_t hwirq,
177 irq_handler_t handler, void *cookie, const char *name)
183 virq = irq_create_mapping(NULL, hwirq);
185 dev_warn(&adapter->dev, "cxl_map_irq: irq_create_mapping failed\n");
189 if (cxl_ops->setup_irq)
190 cxl_ops->setup_irq(adapter, hwirq, virq);
192 pr_devel("hwirq %#lx mapped to virq %u\n", hwirq, virq);
194 result = request_irq(virq, handler, 0, name, cookie);
196 dev_warn(&adapter->dev, "cxl_map_irq: request_irq failed: %i\n", result);
203 void cxl_unmap_irq(unsigned int virq, void *cookie)
205 free_irq(virq, cookie);
208 int cxl_register_one_irq(struct cxl *adapter,
209 irq_handler_t handler,
211 irq_hw_number_t *dest_hwirq,
212 unsigned int *dest_virq,
217 if ((hwirq = cxl_ops->alloc_one_irq(adapter)) < 0)
220 if (!(virq = cxl_map_irq(adapter, hwirq, handler, cookie, name)))
229 cxl_ops->release_one_irq(adapter, hwirq);
233 void afu_irq_name_free(struct cxl_context *ctx)
235 struct cxl_irq_name *irq_name, *tmp;
237 list_for_each_entry_safe(irq_name, tmp, &ctx->irq_names, list) {
238 kfree(irq_name->name);
239 list_del(&irq_name->list);
244 int afu_allocate_irqs(struct cxl_context *ctx, u32 count)
247 struct cxl_irq_name *irq_name;
251 * In native mode, range 0 is reserved for the multiplexed
252 * PSL interrupt. It has been allocated when the AFU was initialized.
254 * In a guest, the PSL interrupt is not mutliplexed, but per-context,
255 * and is the first interrupt from range 0. It still needs to be
256 * allocated, so bump the count by one.
258 if (cpu_has_feature(CPU_FTR_HVMODE))
261 alloc_count = count + 1;
263 if ((rc = cxl_ops->alloc_irq_ranges(&ctx->irqs, ctx->afu->adapter,
267 if (cpu_has_feature(CPU_FTR_HVMODE)) {
268 /* Multiplexed PSL Interrupt */
269 ctx->irqs.offset[0] = ctx->afu->native->psl_hwirq;
270 ctx->irqs.range[0] = 1;
273 ctx->irq_count = count;
274 ctx->irq_bitmap = kcalloc(BITS_TO_LONGS(count),
275 sizeof(*ctx->irq_bitmap), GFP_KERNEL);
276 if (!ctx->irq_bitmap)
280 * Allocate names first. If any fail, bail out before allocating
281 * actual hardware IRQs.
283 for (r = afu_irq_range_start(); r < CXL_IRQ_RANGES; r++) {
284 for (i = 0; i < ctx->irqs.range[r]; i++) {
285 irq_name = kmalloc(sizeof(struct cxl_irq_name),
289 irq_name->name = kasprintf(GFP_KERNEL, "cxl-%s-pe%i-%i",
290 dev_name(&ctx->afu->dev),
292 if (!irq_name->name) {
296 /* Add to tail so next look get the correct order */
297 list_add_tail(&irq_name->list, &ctx->irq_names);
304 cxl_ops->release_irq_ranges(&ctx->irqs, ctx->afu->adapter);
305 afu_irq_name_free(ctx);
309 static void afu_register_hwirqs(struct cxl_context *ctx)
311 irq_hw_number_t hwirq;
312 struct cxl_irq_name *irq_name;
314 irqreturn_t (*handler)(int irq, void *data);
316 /* We've allocated all memory now, so let's do the irq allocations */
317 irq_name = list_first_entry(&ctx->irq_names, struct cxl_irq_name, list);
318 for (r = afu_irq_range_start(); r < CXL_IRQ_RANGES; r++) {
319 hwirq = ctx->irqs.offset[r];
320 for (i = 0; i < ctx->irqs.range[r]; hwirq++, i++) {
321 if (r == 0 && i == 0)
323 * The very first interrupt of range 0 is
324 * always the PSL interrupt, but we only
325 * need to connect a handler for guests,
326 * because there's one PSL interrupt per
328 * On bare-metal, the PSL interrupt is
329 * multiplexed and was setup when the AFU
332 handler = cxl_ops->psl_interrupt;
334 handler = cxl_irq_afu;
335 cxl_map_irq(ctx->afu->adapter, hwirq, handler, ctx,
337 irq_name = list_next_entry(irq_name, list);
342 int afu_register_irqs(struct cxl_context *ctx, u32 count)
346 rc = afu_allocate_irqs(ctx, count);
350 afu_register_hwirqs(ctx);
354 void afu_release_irqs(struct cxl_context *ctx, void *cookie)
356 irq_hw_number_t hwirq;
360 for (r = afu_irq_range_start(); r < CXL_IRQ_RANGES; r++) {
361 hwirq = ctx->irqs.offset[r];
362 for (i = 0; i < ctx->irqs.range[r]; hwirq++, i++) {
363 virq = irq_find_mapping(NULL, hwirq);
365 cxl_unmap_irq(virq, cookie);
369 afu_irq_name_free(ctx);
370 cxl_ops->release_irq_ranges(&ctx->irqs, ctx->afu->adapter);
375 void cxl_afu_decode_psl_serr(struct cxl_afu *afu, u64 serr)
378 "PSL Slice error received. Check AFU for root cause.\n");
379 dev_crit(&afu->dev, "PSL_SERR_An: 0x%016llx\n", serr);
380 if (serr & CXL_PSL_SERR_An_afuto)
381 dev_crit(&afu->dev, "AFU MMIO Timeout\n");
382 if (serr & CXL_PSL_SERR_An_afudis)
384 "MMIO targeted Accelerator that was not enabled\n");
385 if (serr & CXL_PSL_SERR_An_afuov)
386 dev_crit(&afu->dev, "AFU CTAG Overflow\n");
387 if (serr & CXL_PSL_SERR_An_badsrc)
388 dev_crit(&afu->dev, "Bad Interrupt Source\n");
389 if (serr & CXL_PSL_SERR_An_badctx)
390 dev_crit(&afu->dev, "Bad Context Handle\n");
391 if (serr & CXL_PSL_SERR_An_llcmdis)
392 dev_crit(&afu->dev, "LLCMD to Disabled AFU\n");
393 if (serr & CXL_PSL_SERR_An_llcmdto)
394 dev_crit(&afu->dev, "LLCMD Timeout to AFU\n");
395 if (serr & CXL_PSL_SERR_An_afupar)
396 dev_crit(&afu->dev, "AFU MMIO Parity Error\n");
397 if (serr & CXL_PSL_SERR_An_afudup)
398 dev_crit(&afu->dev, "AFU MMIO Duplicate CTAG Error\n");
399 if (serr & CXL_PSL_SERR_An_AE)
401 "AFU asserted JDONE with JERROR in AFU Directed Mode\n");