| File: | dev/pci/drm/i915/gt/intel_reset.c |
| Warning: | line 711, column 3 Value stored to 'vma_offset' is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | /* |
| 2 | * SPDX-License-Identifier: MIT |
| 3 | * |
| 4 | * Copyright © 2008-2018 Intel Corporation |
| 5 | */ |
| 6 | |
| 7 | #include <linux/sched/mm.h> |
| 8 | #include <linux/stop_machine.h> |
| 9 | |
| 10 | #include "display/intel_display_types.h" |
| 11 | #include "display/intel_overlay.h" |
| 12 | |
| 13 | #include "gem/i915_gem_context.h" |
| 14 | |
| 15 | #include "i915_drv.h" |
| 16 | #include "i915_gpu_error.h" |
| 17 | #include "i915_irq.h" |
| 18 | #include "intel_breadcrumbs.h" |
| 19 | #include "intel_engine_pm.h" |
| 20 | #include "intel_gt.h" |
| 21 | #include "intel_gt_pm.h" |
| 22 | #include "intel_reset.h" |
| 23 | |
| 24 | #include "uc/intel_guc.h" |
| 25 | #include "uc/intel_guc_submission.h" |
| 26 | |
| 27 | #define RESET_MAX_RETRIES3 3 |
| 28 | |
| 29 | /* XXX How to handle concurrent GGTT updates using tiling registers? */ |
| 30 | #define RESET_UNDER_STOP_MACHINE0 0 |
| 31 | |
| 32 | static void rmw_set_fw(struct intel_uncore *uncore, i915_reg_t reg, u32 set) |
| 33 | { |
| 34 | intel_uncore_rmw_fw(uncore, reg, 0, set); |
| 35 | } |
| 36 | |
| 37 | static void rmw_clear_fw(struct intel_uncore *uncore, i915_reg_t reg, u32 clr) |
| 38 | { |
| 39 | intel_uncore_rmw_fw(uncore, reg, clr, 0); |
| 40 | } |
| 41 | |
| 42 | static void engine_skip_context(struct i915_request *rq) |
| 43 | { |
| 44 | struct intel_engine_cs *engine = rq->engine; |
| 45 | struct intel_context *hung_ctx = rq->context; |
| 46 | |
| 47 | if (!i915_request_is_active(rq)) |
| 48 | return; |
| 49 | |
| 50 | lockdep_assert_held(&engine->active.lock)do { (void)(&engine->active.lock); } while(0); |
| 51 | list_for_each_entry_continue(rq, &engine->active.requests, sched.link)for (rq = ({ const __typeof( ((__typeof(*rq) *)0)->sched.link ) *__mptr = ((rq)->sched.link.next); (__typeof(*rq) *)( ( char *)__mptr - __builtin_offsetof(__typeof(*rq), sched.link) );}); &rq->sched.link != (&engine->active.requests ); rq = ({ const __typeof( ((__typeof(*rq) *)0)->sched.link ) *__mptr = (rq->sched.link.next); (__typeof(*rq) *)( (char *)__mptr - __builtin_offsetof(__typeof(*rq), sched.link) );} )) |
| 52 | if (rq->context == hung_ctx) { |
| 53 | i915_request_set_error_once(rq, -EIO5); |
| 54 | __i915_request_skip(rq); |
| 55 | } |
| 56 | } |
| 57 | |
| 58 | static void client_mark_guilty(struct i915_gem_context *ctx, bool_Bool banned) |
| 59 | { |
| 60 | struct drm_i915_file_private *file_priv = ctx->file_priv; |
| 61 | unsigned long prev_hang; |
| 62 | unsigned int score; |
| 63 | |
| 64 | if (IS_ERR_OR_NULL(file_priv)) |
| 65 | return; |
| 66 | |
| 67 | score = 0; |
| 68 | if (banned) |
| 69 | score = I915_CLIENT_SCORE_CONTEXT_BAN3; |
| 70 | |
| 71 | prev_hang = xchg(&file_priv->hang_timestamp, jiffies)__sync_lock_test_and_set(&file_priv->hang_timestamp, jiffies ); |
| 72 | if (time_before(jiffies, prev_hang + I915_CLIENT_FAST_HANG_JIFFIES)((long)(jiffies) - (long)(prev_hang + (60 * hz)) < 0)) |
| 73 | score += I915_CLIENT_SCORE_HANG_FAST1; |
| 74 | |
| 75 | if (score) { |
| 76 | atomic_add(score, &file_priv->ban_score)__sync_fetch_and_add(&file_priv->ban_score, score); |
| 77 | |
| 78 | drm_dbg(&ctx->i915->drm,drm_dev_dbg((&ctx->i915->drm)->dev, DRM_UT_DRIVER , "client %s: gained %u ban score, now %u\n", ctx->name, score , ({ typeof(*(&file_priv->ban_score)) __tmp = *(volatile typeof(*(&file_priv->ban_score)) *)&(*(&file_priv ->ban_score)); membar_datadep_consumer(); __tmp; })) |
| 79 | "client %s: gained %u ban score, now %u\n",drm_dev_dbg((&ctx->i915->drm)->dev, DRM_UT_DRIVER , "client %s: gained %u ban score, now %u\n", ctx->name, score , ({ typeof(*(&file_priv->ban_score)) __tmp = *(volatile typeof(*(&file_priv->ban_score)) *)&(*(&file_priv ->ban_score)); membar_datadep_consumer(); __tmp; })) |
| 80 | ctx->name, score,drm_dev_dbg((&ctx->i915->drm)->dev, DRM_UT_DRIVER , "client %s: gained %u ban score, now %u\n", ctx->name, score , ({ typeof(*(&file_priv->ban_score)) __tmp = *(volatile typeof(*(&file_priv->ban_score)) *)&(*(&file_priv ->ban_score)); membar_datadep_consumer(); __tmp; })) |
| 81 | atomic_read(&file_priv->ban_score))drm_dev_dbg((&ctx->i915->drm)->dev, DRM_UT_DRIVER , "client %s: gained %u ban score, now %u\n", ctx->name, score , ({ typeof(*(&file_priv->ban_score)) __tmp = *(volatile typeof(*(&file_priv->ban_score)) *)&(*(&file_priv ->ban_score)); membar_datadep_consumer(); __tmp; })); |
| 82 | } |
| 83 | } |
| 84 | |
| 85 | static bool_Bool mark_guilty(struct i915_request *rq) |
| 86 | { |
| 87 | struct i915_gem_context *ctx; |
| 88 | unsigned long prev_hang; |
| 89 | bool_Bool banned; |
| 90 | int i; |
| 91 | |
| 92 | if (intel_context_is_closed(rq->context)) { |
| 93 | intel_context_set_banned(rq->context); |
| 94 | return true1; |
| 95 | } |
| 96 | |
| 97 | rcu_read_lock(); |
| 98 | ctx = rcu_dereference(rq->context->gem_context)(rq->context->gem_context); |
| 99 | if (ctx && !kref_get_unless_zero(&ctx->ref)) |
| 100 | ctx = NULL((void *)0); |
| 101 | rcu_read_unlock(); |
| 102 | if (!ctx) |
| 103 | return intel_context_is_banned(rq->context); |
| 104 | |
| 105 | atomic_inc(&ctx->guilty_count)__sync_fetch_and_add(&ctx->guilty_count, 1); |
| 106 | |
| 107 | /* Cool contexts are too cool to be banned! (Used for reset testing.) */ |
| 108 | if (!i915_gem_context_is_bannable(ctx)) { |
| 109 | banned = false0; |
| 110 | goto out; |
| 111 | } |
| 112 | |
| 113 | drm_notice(&ctx->i915->drm,printf("drm:pid%d:%s *NOTICE* " "[drm] " "%s context reset due to GPU hang\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , ctx-> name) |
| 114 | "%s context reset due to GPU hang\n",printf("drm:pid%d:%s *NOTICE* " "[drm] " "%s context reset due to GPU hang\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , ctx-> name) |
| 115 | ctx->name)printf("drm:pid%d:%s *NOTICE* " "[drm] " "%s context reset due to GPU hang\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , ctx-> name); |
| 116 | |
| 117 | /* Record the timestamp for the last N hangs */ |
| 118 | prev_hang = ctx->hang_timestamp[0]; |
| 119 | for (i = 0; i < ARRAY_SIZE(ctx->hang_timestamp)(sizeof((ctx->hang_timestamp)) / sizeof((ctx->hang_timestamp )[0])) - 1; i++) |
| 120 | ctx->hang_timestamp[i] = ctx->hang_timestamp[i + 1]; |
| 121 | ctx->hang_timestamp[i] = jiffies; |
| 122 | |
| 123 | /* If we have hung N+1 times in rapid succession, we ban the context! */ |
| 124 | banned = !i915_gem_context_is_recoverable(ctx); |
| 125 | if (time_before(jiffies, prev_hang + CONTEXT_FAST_HANG_JIFFIES)((long)(jiffies) - (long)(prev_hang + (120 * hz)) < 0)) |
| 126 | banned = true1; |
| 127 | if (banned) { |
| 128 | drm_dbg(&ctx->i915->drm, "context %s: guilty %d, banned\n",drm_dev_dbg((&ctx->i915->drm)->dev, DRM_UT_DRIVER , "context %s: guilty %d, banned\n", ctx->name, ({ typeof( *(&ctx->guilty_count)) __tmp = *(volatile typeof(*(& ctx->guilty_count)) *)&(*(&ctx->guilty_count)); membar_datadep_consumer(); __tmp; })) |
| 129 | ctx->name, atomic_read(&ctx->guilty_count))drm_dev_dbg((&ctx->i915->drm)->dev, DRM_UT_DRIVER , "context %s: guilty %d, banned\n", ctx->name, ({ typeof( *(&ctx->guilty_count)) __tmp = *(volatile typeof(*(& ctx->guilty_count)) *)&(*(&ctx->guilty_count)); membar_datadep_consumer(); __tmp; })); |
| 130 | intel_context_set_banned(rq->context); |
| 131 | } |
| 132 | |
| 133 | client_mark_guilty(ctx, banned); |
| 134 | |
| 135 | out: |
| 136 | i915_gem_context_put(ctx); |
| 137 | return banned; |
| 138 | } |
| 139 | |
| 140 | static void mark_innocent(struct i915_request *rq) |
| 141 | { |
| 142 | struct i915_gem_context *ctx; |
| 143 | |
| 144 | rcu_read_lock(); |
| 145 | ctx = rcu_dereference(rq->context->gem_context)(rq->context->gem_context); |
| 146 | if (ctx) |
| 147 | atomic_inc(&ctx->active_count)__sync_fetch_and_add(&ctx->active_count, 1); |
| 148 | rcu_read_unlock(); |
| 149 | } |
| 150 | |
| 151 | void __i915_request_reset(struct i915_request *rq, bool_Bool guilty) |
| 152 | { |
| 153 | RQ_TRACE(rq, "guilty? %s\n", yesno(guilty))do { const struct i915_request *rq__ = (rq); do { const struct intel_engine_cs *e__ __attribute__((__unused__)) = (rq__-> engine); do { } while (0); } while (0); } while (0); |
| 154 | |
| 155 | GEM_BUG_ON(i915_request_completed(rq))((void)0); |
| 156 | |
| 157 | rcu_read_lock(); /* protect the GEM context */ |
| 158 | if (guilty) { |
| 159 | i915_request_set_error_once(rq, -EIO5); |
| 160 | __i915_request_skip(rq); |
| 161 | if (mark_guilty(rq)) |
| 162 | engine_skip_context(rq); |
| 163 | } else { |
| 164 | i915_request_set_error_once(rq, -EAGAIN35); |
| 165 | mark_innocent(rq); |
| 166 | } |
| 167 | rcu_read_unlock(); |
| 168 | } |
| 169 | |
| 170 | static bool_Bool i915_in_reset(struct pci_dev *pdev) |
| 171 | { |
| 172 | u8 gdrst; |
| 173 | |
| 174 | pci_read_config_byte(pdev, I915_GDRST0xc0, &gdrst); |
| 175 | return gdrst & GRDOM_RESET_STATUS(1 << 1); |
| 176 | } |
| 177 | |
| 178 | static int i915_do_reset(struct intel_gt *gt, |
| 179 | intel_engine_mask_t engine_mask, |
| 180 | unsigned int retry) |
| 181 | { |
| 182 | struct pci_dev *pdev = gt->i915->drm.pdev; |
| 183 | int err; |
| 184 | |
| 185 | /* Assert reset for at least 20 usec, and wait for acknowledgement. */ |
| 186 | pci_write_config_byte(pdev, I915_GDRST0xc0, GRDOM_RESET_ENABLE(1 << 0)); |
| 187 | udelay(50); |
| 188 | err = wait_for_atomic(i915_in_reset(pdev), 50)({ extern char _ctassert[(!(!__builtin_constant_p((50) * 1000 ))) ? 1 : -1 ] __attribute__((__unused__)); extern char _ctassert [(!(((50) * 1000) > 50000)) ? 1 : -1 ] __attribute__((__unused__ )); ({ int cpu, ret, timeout = (((50) * 1000)) * 1000; u64 base ; do { } while (0); if (!(1)) { ; cpu = (({struct cpu_info *__ci ; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof (struct cpu_info, ci_self))); __ci;})->ci_cpuid); } base = local_clock(); for (;;) { u64 now = local_clock(); if (!(1)) ; __asm volatile("" : : : "memory"); if (((i915_in_reset(pdev )))) { ret = 0; break; } if (now - base >= timeout) { ret = -60; break; } cpu_relax(); if (!(1)) { ; if (__builtin_expect (!!(cpu != (({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self ))); __ci;})->ci_cpuid)), 0)) { timeout -= now - base; cpu = (({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self ))); __ci;})->ci_cpuid); base = local_clock(); } } } ret; } ); }); |
| 189 | |
| 190 | /* Clear the reset request. */ |
| 191 | pci_write_config_byte(pdev, I915_GDRST0xc0, 0); |
| 192 | udelay(50); |
| 193 | if (!err) |
| 194 | err = wait_for_atomic(!i915_in_reset(pdev), 50)({ extern char _ctassert[(!(!__builtin_constant_p((50) * 1000 ))) ? 1 : -1 ] __attribute__((__unused__)); extern char _ctassert [(!(((50) * 1000) > 50000)) ? 1 : -1 ] __attribute__((__unused__ )); ({ int cpu, ret, timeout = (((50) * 1000)) * 1000; u64 base ; do { } while (0); if (!(1)) { ; cpu = (({struct cpu_info *__ci ; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof (struct cpu_info, ci_self))); __ci;})->ci_cpuid); } base = local_clock(); for (;;) { u64 now = local_clock(); if (!(1)) ; __asm volatile("" : : : "memory"); if (((!i915_in_reset(pdev )))) { ret = 0; break; } if (now - base >= timeout) { ret = -60; break; } cpu_relax(); if (!(1)) { ; if (__builtin_expect (!!(cpu != (({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self ))); __ci;})->ci_cpuid)), 0)) { timeout -= now - base; cpu = (({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self ))); __ci;})->ci_cpuid); base = local_clock(); } } } ret; } ); }); |
| 195 | |
| 196 | return err; |
| 197 | } |
| 198 | |
| 199 | static bool_Bool g4x_reset_complete(struct pci_dev *pdev) |
| 200 | { |
| 201 | u8 gdrst; |
| 202 | |
| 203 | pci_read_config_byte(pdev, I915_GDRST0xc0, &gdrst); |
| 204 | return (gdrst & GRDOM_RESET_ENABLE(1 << 0)) == 0; |
| 205 | } |
| 206 | |
| 207 | static int g33_do_reset(struct intel_gt *gt, |
| 208 | intel_engine_mask_t engine_mask, |
| 209 | unsigned int retry) |
| 210 | { |
| 211 | struct pci_dev *pdev = gt->i915->drm.pdev; |
| 212 | |
| 213 | pci_write_config_byte(pdev, I915_GDRST0xc0, GRDOM_RESET_ENABLE(1 << 0)); |
| 214 | return wait_for_atomic(g4x_reset_complete(pdev), 50)({ extern char _ctassert[(!(!__builtin_constant_p((50) * 1000 ))) ? 1 : -1 ] __attribute__((__unused__)); extern char _ctassert [(!(((50) * 1000) > 50000)) ? 1 : -1 ] __attribute__((__unused__ )); ({ int cpu, ret, timeout = (((50) * 1000)) * 1000; u64 base ; do { } while (0); if (!(1)) { ; cpu = (({struct cpu_info *__ci ; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof (struct cpu_info, ci_self))); __ci;})->ci_cpuid); } base = local_clock(); for (;;) { u64 now = local_clock(); if (!(1)) ; __asm volatile("" : : : "memory"); if (((g4x_reset_complete (pdev)))) { ret = 0; break; } if (now - base >= timeout) { ret = -60; break; } cpu_relax(); if (!(1)) { ; if (__builtin_expect (!!(cpu != (({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self ))); __ci;})->ci_cpuid)), 0)) { timeout -= now - base; cpu = (({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self ))); __ci;})->ci_cpuid); base = local_clock(); } } } ret; } ); }); |
| 215 | } |
| 216 | |
| 217 | static int g4x_do_reset(struct intel_gt *gt, |
| 218 | intel_engine_mask_t engine_mask, |
| 219 | unsigned int retry) |
| 220 | { |
| 221 | struct pci_dev *pdev = gt->i915->drm.pdev; |
| 222 | struct intel_uncore *uncore = gt->uncore; |
| 223 | int ret; |
| 224 | |
| 225 | /* WaVcpClkGateDisableForMediaReset:ctg,elk */ |
| 226 | rmw_set_fw(uncore, VDECCLK_GATE_D((const i915_reg_t){ .reg = (0x620C) }), VCP_UNIT_CLOCK_GATE_DISABLE(1 << 4)); |
| 227 | intel_uncore_posting_read_fw(uncore, VDECCLK_GATE_D)((void)__raw_uncore_read32(uncore, ((const i915_reg_t){ .reg = (0x620C) }))); |
| 228 | |
| 229 | pci_write_config_byte(pdev, I915_GDRST0xc0, |
| 230 | GRDOM_MEDIA(3 << 2) | GRDOM_RESET_ENABLE(1 << 0)); |
| 231 | ret = wait_for_atomic(g4x_reset_complete(pdev), 50)({ extern char _ctassert[(!(!__builtin_constant_p((50) * 1000 ))) ? 1 : -1 ] __attribute__((__unused__)); extern char _ctassert [(!(((50) * 1000) > 50000)) ? 1 : -1 ] __attribute__((__unused__ )); ({ int cpu, ret, timeout = (((50) * 1000)) * 1000; u64 base ; do { } while (0); if (!(1)) { ; cpu = (({struct cpu_info *__ci ; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof (struct cpu_info, ci_self))); __ci;})->ci_cpuid); } base = local_clock(); for (;;) { u64 now = local_clock(); if (!(1)) ; __asm volatile("" : : : "memory"); if (((g4x_reset_complete (pdev)))) { ret = 0; break; } if (now - base >= timeout) { ret = -60; break; } cpu_relax(); if (!(1)) { ; if (__builtin_expect (!!(cpu != (({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self ))); __ci;})->ci_cpuid)), 0)) { timeout -= now - base; cpu = (({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self ))); __ci;})->ci_cpuid); base = local_clock(); } } } ret; } ); }); |
| 232 | if (ret) { |
| 233 | drm_dbg(>->i915->drm, "Wait for media reset failed\n")drm_dev_dbg((>->i915->drm)->dev, DRM_UT_DRIVER , "Wait for media reset failed\n"); |
| 234 | goto out; |
| 235 | } |
| 236 | |
| 237 | pci_write_config_byte(pdev, I915_GDRST0xc0, |
| 238 | GRDOM_RENDER(1 << 2) | GRDOM_RESET_ENABLE(1 << 0)); |
| 239 | ret = wait_for_atomic(g4x_reset_complete(pdev), 50)({ extern char _ctassert[(!(!__builtin_constant_p((50) * 1000 ))) ? 1 : -1 ] __attribute__((__unused__)); extern char _ctassert [(!(((50) * 1000) > 50000)) ? 1 : -1 ] __attribute__((__unused__ )); ({ int cpu, ret, timeout = (((50) * 1000)) * 1000; u64 base ; do { } while (0); if (!(1)) { ; cpu = (({struct cpu_info *__ci ; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof (struct cpu_info, ci_self))); __ci;})->ci_cpuid); } base = local_clock(); for (;;) { u64 now = local_clock(); if (!(1)) ; __asm volatile("" : : : "memory"); if (((g4x_reset_complete (pdev)))) { ret = 0; break; } if (now - base >= timeout) { ret = -60; break; } cpu_relax(); if (!(1)) { ; if (__builtin_expect (!!(cpu != (({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self ))); __ci;})->ci_cpuid)), 0)) { timeout -= now - base; cpu = (({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self ))); __ci;})->ci_cpuid); base = local_clock(); } } } ret; } ); }); |
| 240 | if (ret) { |
| 241 | drm_dbg(>->i915->drm, "Wait for render reset failed\n")drm_dev_dbg((>->i915->drm)->dev, DRM_UT_DRIVER , "Wait for render reset failed\n"); |
| 242 | goto out; |
| 243 | } |
| 244 | |
| 245 | out: |
| 246 | pci_write_config_byte(pdev, I915_GDRST0xc0, 0); |
| 247 | |
| 248 | rmw_clear_fw(uncore, VDECCLK_GATE_D((const i915_reg_t){ .reg = (0x620C) }), VCP_UNIT_CLOCK_GATE_DISABLE(1 << 4)); |
| 249 | intel_uncore_posting_read_fw(uncore, VDECCLK_GATE_D)((void)__raw_uncore_read32(uncore, ((const i915_reg_t){ .reg = (0x620C) }))); |
| 250 | |
| 251 | return ret; |
| 252 | } |
| 253 | |
| 254 | static int ilk_do_reset(struct intel_gt *gt, intel_engine_mask_t engine_mask, |
| 255 | unsigned int retry) |
| 256 | { |
| 257 | struct intel_uncore *uncore = gt->uncore; |
| 258 | int ret; |
| 259 | |
| 260 | intel_uncore_write_fw(uncore, ILK_GDSR,__raw_uncore_write32(uncore, ((const i915_reg_t){ .reg = (0x10000 + 0x2ca4) }), (1 << 1) | (1 << 0)) |
| 261 | ILK_GRDOM_RENDER | ILK_GRDOM_RESET_ENABLE)__raw_uncore_write32(uncore, ((const i915_reg_t){ .reg = (0x10000 + 0x2ca4) }), (1 << 1) | (1 << 0)); |
| 262 | ret = __intel_wait_for_register_fw(uncore, ILK_GDSR((const i915_reg_t){ .reg = (0x10000 + 0x2ca4) }), |
| 263 | ILK_GRDOM_RESET_ENABLE(1 << 0), 0, |
| 264 | 5000, 0, |
| 265 | NULL((void *)0)); |
| 266 | if (ret) { |
| 267 | drm_dbg(>->i915->drm, "Wait for render reset failed\n")drm_dev_dbg((>->i915->drm)->dev, DRM_UT_DRIVER , "Wait for render reset failed\n"); |
| 268 | goto out; |
| 269 | } |
| 270 | |
| 271 | intel_uncore_write_fw(uncore, ILK_GDSR,__raw_uncore_write32(uncore, ((const i915_reg_t){ .reg = (0x10000 + 0x2ca4) }), (3 << 1) | (1 << 0)) |
| 272 | ILK_GRDOM_MEDIA | ILK_GRDOM_RESET_ENABLE)__raw_uncore_write32(uncore, ((const i915_reg_t){ .reg = (0x10000 + 0x2ca4) }), (3 << 1) | (1 << 0)); |
| 273 | ret = __intel_wait_for_register_fw(uncore, ILK_GDSR((const i915_reg_t){ .reg = (0x10000 + 0x2ca4) }), |
| 274 | ILK_GRDOM_RESET_ENABLE(1 << 0), 0, |
| 275 | 5000, 0, |
| 276 | NULL((void *)0)); |
| 277 | if (ret) { |
| 278 | drm_dbg(>->i915->drm, "Wait for media reset failed\n")drm_dev_dbg((>->i915->drm)->dev, DRM_UT_DRIVER , "Wait for media reset failed\n"); |
| 279 | goto out; |
| 280 | } |
| 281 | |
| 282 | out: |
| 283 | intel_uncore_write_fw(uncore, ILK_GDSR, 0)__raw_uncore_write32(uncore, ((const i915_reg_t){ .reg = (0x10000 + 0x2ca4) }), 0); |
| 284 | intel_uncore_posting_read_fw(uncore, ILK_GDSR)((void)__raw_uncore_read32(uncore, ((const i915_reg_t){ .reg = (0x10000 + 0x2ca4) }))); |
| 285 | return ret; |
| 286 | } |
| 287 | |
| 288 | /* Reset the hardware domains (GENX_GRDOM_*) specified by mask */ |
| 289 | static int gen6_hw_domain_reset(struct intel_gt *gt, u32 hw_domain_mask) |
| 290 | { |
| 291 | struct intel_uncore *uncore = gt->uncore; |
| 292 | int err; |
| 293 | |
| 294 | /* |
| 295 | * GEN6_GDRST is not in the gt power well, no need to check |
| 296 | * for fifo space for the write or forcewake the chip for |
| 297 | * the read |
| 298 | */ |
| 299 | intel_uncore_write_fw(uncore, GEN6_GDRST, hw_domain_mask)__raw_uncore_write32(uncore, ((const i915_reg_t){ .reg = (0x941c ) }), hw_domain_mask); |
| 300 | |
| 301 | /* Wait for the device to ack the reset requests */ |
| 302 | err = __intel_wait_for_register_fw(uncore, |
| 303 | GEN6_GDRST((const i915_reg_t){ .reg = (0x941c) }), hw_domain_mask, 0, |
| 304 | 500, 0, |
| 305 | NULL((void *)0)); |
| 306 | if (err) |
| 307 | drm_dbg(>->i915->drm,drm_dev_dbg((>->i915->drm)->dev, DRM_UT_DRIVER , "Wait for 0x%08x engines reset failed\n", hw_domain_mask) |
| 308 | "Wait for 0x%08x engines reset failed\n",drm_dev_dbg((>->i915->drm)->dev, DRM_UT_DRIVER , "Wait for 0x%08x engines reset failed\n", hw_domain_mask) |
| 309 | hw_domain_mask)drm_dev_dbg((>->i915->drm)->dev, DRM_UT_DRIVER , "Wait for 0x%08x engines reset failed\n", hw_domain_mask); |
| 310 | |
| 311 | return err; |
| 312 | } |
| 313 | |
| 314 | static int gen6_reset_engines(struct intel_gt *gt, |
| 315 | intel_engine_mask_t engine_mask, |
| 316 | unsigned int retry) |
| 317 | { |
| 318 | static const u32 hw_engine_mask[] = { |
| 319 | [RCS0] = GEN6_GRDOM_RENDER(1 << 1), |
| 320 | [BCS0] = GEN6_GRDOM_BLT(1 << 3), |
| 321 | [VCS0] = GEN6_GRDOM_MEDIA(1 << 2), |
| 322 | [VCS1] = GEN8_GRDOM_MEDIA2(1 << 7), |
| 323 | [VECS0] = GEN6_GRDOM_VECS(1 << 4), |
| 324 | }; |
| 325 | struct intel_engine_cs *engine; |
| 326 | u32 hw_mask; |
| 327 | |
| 328 | if (engine_mask == ALL_ENGINES((intel_engine_mask_t)~0ul)) { |
| 329 | hw_mask = GEN6_GRDOM_FULL(1 << 0); |
| 330 | } else { |
| 331 | intel_engine_mask_t tmp; |
| 332 | |
| 333 | hw_mask = 0; |
| 334 | for_each_engine_masked(engine, gt, engine_mask, tmp)for ((tmp) = (engine_mask) & (gt)->info.engine_mask; ( tmp) ? ((engine) = (gt)->engine[({ int __idx = ffs(tmp) - 1 ; tmp &= ~(1UL << (__idx)); __idx; })]), 1 : 0;) { |
| 335 | GEM_BUG_ON(engine->id >= ARRAY_SIZE(hw_engine_mask))((void)0); |
| 336 | hw_mask |= hw_engine_mask[engine->id]; |
| 337 | } |
| 338 | } |
| 339 | |
| 340 | return gen6_hw_domain_reset(gt, hw_mask); |
| 341 | } |
| 342 | |
| 343 | static int gen11_lock_sfc(struct intel_engine_cs *engine, u32 *hw_mask) |
| 344 | { |
| 345 | struct intel_uncore *uncore = engine->uncore; |
| 346 | u8 vdbox_sfc_access = engine->gt->info.vdbox_sfc_access; |
| 347 | i915_reg_t sfc_forced_lock, sfc_forced_lock_ack; |
| 348 | u32 sfc_forced_lock_bit, sfc_forced_lock_ack_bit; |
| 349 | i915_reg_t sfc_usage; |
| 350 | u32 sfc_usage_bit; |
| 351 | u32 sfc_reset_bit; |
| 352 | int ret; |
| 353 | |
| 354 | switch (engine->class) { |
| 355 | case VIDEO_DECODE_CLASS1: |
| 356 | if ((BIT(engine->instance)(1UL << (engine->instance)) & vdbox_sfc_access) == 0) |
| 357 | return 0; |
| 358 | |
| 359 | sfc_forced_lock = GEN11_VCS_SFC_FORCED_LOCK(engine)((const i915_reg_t){ .reg = ((engine)->mmio_base + 0x88C) } ); |
| 360 | sfc_forced_lock_bit = GEN11_VCS_SFC_FORCED_LOCK_BIT(1 << 0); |
| 361 | |
| 362 | sfc_forced_lock_ack = GEN11_VCS_SFC_LOCK_STATUS(engine)((const i915_reg_t){ .reg = ((engine)->mmio_base + 0x890) } ); |
| 363 | sfc_forced_lock_ack_bit = GEN11_VCS_SFC_LOCK_ACK_BIT(1 << 1); |
| 364 | |
| 365 | sfc_usage = GEN11_VCS_SFC_LOCK_STATUS(engine)((const i915_reg_t){ .reg = ((engine)->mmio_base + 0x890) } ); |
| 366 | sfc_usage_bit = GEN11_VCS_SFC_USAGE_BIT(1 << 0); |
| 367 | sfc_reset_bit = GEN11_VCS_SFC_RESET_BIT(engine->instance)((1 << 17) << ((engine->instance) >> 1)); |
| 368 | break; |
| 369 | |
| 370 | case VIDEO_ENHANCEMENT_CLASS2: |
| 371 | sfc_forced_lock = GEN11_VECS_SFC_FORCED_LOCK(engine)((const i915_reg_t){ .reg = ((engine)->mmio_base + 0x201C) }); |
| 372 | sfc_forced_lock_bit = GEN11_VECS_SFC_FORCED_LOCK_BIT(1 << 0); |
| 373 | |
| 374 | sfc_forced_lock_ack = GEN11_VECS_SFC_LOCK_ACK(engine)((const i915_reg_t){ .reg = ((engine)->mmio_base + 0x2018) }); |
| 375 | sfc_forced_lock_ack_bit = GEN11_VECS_SFC_LOCK_ACK_BIT(1 << 0); |
| 376 | |
| 377 | sfc_usage = GEN11_VECS_SFC_USAGE(engine)((const i915_reg_t){ .reg = ((engine)->mmio_base + 0x2014) }); |
| 378 | sfc_usage_bit = GEN11_VECS_SFC_USAGE_BIT(1 << 0); |
| 379 | sfc_reset_bit = GEN11_VECS_SFC_RESET_BIT(engine->instance)((1 << 17) << (engine->instance)); |
| 380 | break; |
| 381 | |
| 382 | default: |
| 383 | return 0; |
| 384 | } |
| 385 | |
| 386 | /* |
| 387 | * If the engine is using a SFC, tell the engine that a software reset |
| 388 | * is going to happen. The engine will then try to force lock the SFC. |
| 389 | * If SFC ends up being locked to the engine we want to reset, we have |
| 390 | * to reset it as well (we will unlock it once the reset sequence is |
| 391 | * completed). |
| 392 | */ |
| 393 | if (!(intel_uncore_read_fw(uncore, sfc_usage)__raw_uncore_read32(uncore, sfc_usage) & sfc_usage_bit)) |
| 394 | return 0; |
| 395 | |
| 396 | rmw_set_fw(uncore, sfc_forced_lock, sfc_forced_lock_bit); |
| 397 | |
| 398 | ret = __intel_wait_for_register_fw(uncore, |
| 399 | sfc_forced_lock_ack, |
| 400 | sfc_forced_lock_ack_bit, |
| 401 | sfc_forced_lock_ack_bit, |
| 402 | 1000, 0, NULL((void *)0)); |
| 403 | |
| 404 | /* Was the SFC released while we were trying to lock it? */ |
| 405 | if (!(intel_uncore_read_fw(uncore, sfc_usage)__raw_uncore_read32(uncore, sfc_usage) & sfc_usage_bit)) |
| 406 | return 0; |
| 407 | |
| 408 | if (ret) { |
| 409 | drm_dbg(&engine->i915->drm,drm_dev_dbg((&engine->i915->drm)->dev, DRM_UT_DRIVER , "Wait for SFC forced lock ack failed\n") |
| 410 | "Wait for SFC forced lock ack failed\n")drm_dev_dbg((&engine->i915->drm)->dev, DRM_UT_DRIVER , "Wait for SFC forced lock ack failed\n"); |
| 411 | return ret; |
| 412 | } |
| 413 | |
| 414 | *hw_mask |= sfc_reset_bit; |
| 415 | return 0; |
| 416 | } |
| 417 | |
| 418 | static void gen11_unlock_sfc(struct intel_engine_cs *engine) |
| 419 | { |
| 420 | struct intel_uncore *uncore = engine->uncore; |
| 421 | u8 vdbox_sfc_access = engine->gt->info.vdbox_sfc_access; |
| 422 | i915_reg_t sfc_forced_lock; |
| 423 | u32 sfc_forced_lock_bit; |
| 424 | |
| 425 | switch (engine->class) { |
| 426 | case VIDEO_DECODE_CLASS1: |
| 427 | if ((BIT(engine->instance)(1UL << (engine->instance)) & vdbox_sfc_access) == 0) |
| 428 | return; |
| 429 | |
| 430 | sfc_forced_lock = GEN11_VCS_SFC_FORCED_LOCK(engine)((const i915_reg_t){ .reg = ((engine)->mmio_base + 0x88C) } ); |
| 431 | sfc_forced_lock_bit = GEN11_VCS_SFC_FORCED_LOCK_BIT(1 << 0); |
| 432 | break; |
| 433 | |
| 434 | case VIDEO_ENHANCEMENT_CLASS2: |
| 435 | sfc_forced_lock = GEN11_VECS_SFC_FORCED_LOCK(engine)((const i915_reg_t){ .reg = ((engine)->mmio_base + 0x201C) }); |
| 436 | sfc_forced_lock_bit = GEN11_VECS_SFC_FORCED_LOCK_BIT(1 << 0); |
| 437 | break; |
| 438 | |
| 439 | default: |
| 440 | return; |
| 441 | } |
| 442 | |
| 443 | rmw_clear_fw(uncore, sfc_forced_lock, sfc_forced_lock_bit); |
| 444 | } |
| 445 | |
| 446 | static int gen11_reset_engines(struct intel_gt *gt, |
| 447 | intel_engine_mask_t engine_mask, |
| 448 | unsigned int retry) |
| 449 | { |
| 450 | static const u32 hw_engine_mask[] = { |
| 451 | [RCS0] = GEN11_GRDOM_RENDER(1 << 1), |
| 452 | [BCS0] = GEN11_GRDOM_BLT(1 << 2), |
| 453 | [VCS0] = GEN11_GRDOM_MEDIA(1 << 5), |
| 454 | [VCS1] = GEN11_GRDOM_MEDIA2(1 << 6), |
| 455 | [VCS2] = GEN11_GRDOM_MEDIA3(1 << 7), |
| 456 | [VCS3] = GEN11_GRDOM_MEDIA4(1 << 8), |
| 457 | [VECS0] = GEN11_GRDOM_VECS(1 << 13), |
| 458 | [VECS1] = GEN11_GRDOM_VECS2(1 << 14), |
| 459 | }; |
| 460 | struct intel_engine_cs *engine; |
| 461 | intel_engine_mask_t tmp; |
| 462 | u32 hw_mask; |
| 463 | int ret; |
| 464 | |
| 465 | if (engine_mask == ALL_ENGINES((intel_engine_mask_t)~0ul)) { |
| 466 | hw_mask = GEN11_GRDOM_FULL(1 << 0); |
| 467 | } else { |
| 468 | hw_mask = 0; |
| 469 | for_each_engine_masked(engine, gt, engine_mask, tmp)for ((tmp) = (engine_mask) & (gt)->info.engine_mask; ( tmp) ? ((engine) = (gt)->engine[({ int __idx = ffs(tmp) - 1 ; tmp &= ~(1UL << (__idx)); __idx; })]), 1 : 0;) { |
| 470 | GEM_BUG_ON(engine->id >= ARRAY_SIZE(hw_engine_mask))((void)0); |
| 471 | hw_mask |= hw_engine_mask[engine->id]; |
| 472 | ret = gen11_lock_sfc(engine, &hw_mask); |
| 473 | if (ret) |
| 474 | goto sfc_unlock; |
| 475 | } |
| 476 | } |
| 477 | |
| 478 | ret = gen6_hw_domain_reset(gt, hw_mask); |
| 479 | |
| 480 | sfc_unlock: |
| 481 | /* |
| 482 | * We unlock the SFC based on the lock status and not the result of |
| 483 | * gen11_lock_sfc to make sure that we clean properly if something |
| 484 | * wrong happened during the lock (e.g. lock acquired after timeout |
| 485 | * expiration). |
| 486 | */ |
| 487 | if (engine_mask != ALL_ENGINES((intel_engine_mask_t)~0ul)) |
| 488 | for_each_engine_masked(engine, gt, engine_mask, tmp)for ((tmp) = (engine_mask) & (gt)->info.engine_mask; ( tmp) ? ((engine) = (gt)->engine[({ int __idx = ffs(tmp) - 1 ; tmp &= ~(1UL << (__idx)); __idx; })]), 1 : 0;) |
| 489 | gen11_unlock_sfc(engine); |
| 490 | |
| 491 | return ret; |
| 492 | } |
| 493 | |
| 494 | static int gen8_engine_reset_prepare(struct intel_engine_cs *engine) |
| 495 | { |
| 496 | struct intel_uncore *uncore = engine->uncore; |
| 497 | const i915_reg_t reg = RING_RESET_CTL(engine->mmio_base)((const i915_reg_t){ .reg = ((engine->mmio_base) + 0xd0) } ); |
| 498 | u32 request, mask, ack; |
| 499 | int ret; |
| 500 | |
| 501 | ack = intel_uncore_read_fw(uncore, reg)__raw_uncore_read32(uncore, reg); |
| 502 | if (ack & RESET_CTL_CAT_ERROR((u32)((1UL << (2)) + 0))) { |
| 503 | /* |
| 504 | * For catastrophic errors, ready-for-reset sequence |
| 505 | * needs to be bypassed: HAS#396813 |
| 506 | */ |
| 507 | request = RESET_CTL_CAT_ERROR((u32)((1UL << (2)) + 0)); |
| 508 | mask = RESET_CTL_CAT_ERROR((u32)((1UL << (2)) + 0)); |
| 509 | |
| 510 | /* Catastrophic errors need to be cleared by HW */ |
| 511 | ack = 0; |
| 512 | } else if (!(ack & RESET_CTL_READY_TO_RESET((u32)((1UL << (1)) + 0)))) { |
| 513 | request = RESET_CTL_REQUEST_RESET((u32)((1UL << (0)) + 0)); |
| 514 | mask = RESET_CTL_READY_TO_RESET((u32)((1UL << (1)) + 0)); |
| 515 | ack = RESET_CTL_READY_TO_RESET((u32)((1UL << (1)) + 0)); |
| 516 | } else { |
| 517 | return 0; |
| 518 | } |
| 519 | |
| 520 | intel_uncore_write_fw(uncore, reg, _MASKED_BIT_ENABLE(request))__raw_uncore_write32(uncore, reg, ({ typeof(request) _a = (request ); ({ if (__builtin_constant_p(_a)) do { } while (0); if (__builtin_constant_p (_a)) do { } while (0); if (__builtin_constant_p(_a) && __builtin_constant_p(_a)) do { } while (0); ((_a) << 16 | (_a)); }); })); |
| 521 | ret = __intel_wait_for_register_fw(uncore, reg, mask, ack, |
| 522 | 700, 0, NULL((void *)0)); |
| 523 | if (ret) |
| 524 | drm_err(&engine->i915->drm,printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "%s reset request timed out: {request: %08x, RESET_CTL: %08x}\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , engine ->name, request, __raw_uncore_read32(uncore, reg)) |
| 525 | "%s reset request timed out: {request: %08x, RESET_CTL: %08x}\n",printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "%s reset request timed out: {request: %08x, RESET_CTL: %08x}\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , engine ->name, request, __raw_uncore_read32(uncore, reg)) |
| 526 | engine->name, request,printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "%s reset request timed out: {request: %08x, RESET_CTL: %08x}\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , engine ->name, request, __raw_uncore_read32(uncore, reg)) |
| 527 | intel_uncore_read_fw(uncore, reg))printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "%s reset request timed out: {request: %08x, RESET_CTL: %08x}\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , engine ->name, request, __raw_uncore_read32(uncore, reg)); |
| 528 | |
| 529 | return ret; |
| 530 | } |
| 531 | |
| 532 | static void gen8_engine_reset_cancel(struct intel_engine_cs *engine) |
| 533 | { |
| 534 | intel_uncore_write_fw(engine->uncore,__raw_uncore_write32(engine->uncore, ((const i915_reg_t){ . reg = ((engine->mmio_base) + 0xd0) }), (({ if (__builtin_constant_p ((((u32)((1UL << (0)) + 0))))) do { } while (0); if (__builtin_constant_p (0)) do { } while (0); if (__builtin_constant_p((((u32)((1UL << (0)) + 0)))) && __builtin_constant_p(0)) do { } while (0); (((((u32)((1UL << (0)) + 0)))) << 16 | (0)) ; }))) |
| 535 | RING_RESET_CTL(engine->mmio_base),__raw_uncore_write32(engine->uncore, ((const i915_reg_t){ . reg = ((engine->mmio_base) + 0xd0) }), (({ if (__builtin_constant_p ((((u32)((1UL << (0)) + 0))))) do { } while (0); if (__builtin_constant_p (0)) do { } while (0); if (__builtin_constant_p((((u32)((1UL << (0)) + 0)))) && __builtin_constant_p(0)) do { } while (0); (((((u32)((1UL << (0)) + 0)))) << 16 | (0)) ; }))) |
| 536 | _MASKED_BIT_DISABLE(RESET_CTL_REQUEST_RESET))__raw_uncore_write32(engine->uncore, ((const i915_reg_t){ . reg = ((engine->mmio_base) + 0xd0) }), (({ if (__builtin_constant_p ((((u32)((1UL << (0)) + 0))))) do { } while (0); if (__builtin_constant_p (0)) do { } while (0); if (__builtin_constant_p((((u32)((1UL << (0)) + 0)))) && __builtin_constant_p(0)) do { } while (0); (((((u32)((1UL << (0)) + 0)))) << 16 | (0)) ; }))); |
| 537 | } |
| 538 | |
| 539 | static int gen8_reset_engines(struct intel_gt *gt, |
| 540 | intel_engine_mask_t engine_mask, |
| 541 | unsigned int retry) |
| 542 | { |
| 543 | struct intel_engine_cs *engine; |
| 544 | const bool_Bool reset_non_ready = retry >= 1; |
| 545 | intel_engine_mask_t tmp; |
| 546 | int ret; |
| 547 | |
| 548 | for_each_engine_masked(engine, gt, engine_mask, tmp)for ((tmp) = (engine_mask) & (gt)->info.engine_mask; ( tmp) ? ((engine) = (gt)->engine[({ int __idx = ffs(tmp) - 1 ; tmp &= ~(1UL << (__idx)); __idx; })]), 1 : 0;) { |
| 549 | ret = gen8_engine_reset_prepare(engine); |
| 550 | if (ret && !reset_non_ready) |
| 551 | goto skip_reset; |
| 552 | |
| 553 | /* |
| 554 | * If this is not the first failed attempt to prepare, |
| 555 | * we decide to proceed anyway. |
| 556 | * |
| 557 | * By doing so we risk context corruption and with |
| 558 | * some gens (kbl), possible system hang if reset |
| 559 | * happens during active bb execution. |
| 560 | * |
| 561 | * We rather take context corruption instead of |
| 562 | * failed reset with a wedged driver/gpu. And |
| 563 | * active bb execution case should be covered by |
| 564 | * stop_engines() we have before the reset. |
| 565 | */ |
| 566 | } |
| 567 | |
| 568 | if (INTEL_GEN(gt->i915)((&(gt->i915)->__info)->gen) >= 11) |
| 569 | ret = gen11_reset_engines(gt, engine_mask, retry); |
| 570 | else |
| 571 | ret = gen6_reset_engines(gt, engine_mask, retry); |
| 572 | |
| 573 | skip_reset: |
| 574 | for_each_engine_masked(engine, gt, engine_mask, tmp)for ((tmp) = (engine_mask) & (gt)->info.engine_mask; ( tmp) ? ((engine) = (gt)->engine[({ int __idx = ffs(tmp) - 1 ; tmp &= ~(1UL << (__idx)); __idx; })]), 1 : 0;) |
| 575 | gen8_engine_reset_cancel(engine); |
| 576 | |
| 577 | return ret; |
| 578 | } |
| 579 | |
| 580 | static int mock_reset(struct intel_gt *gt, |
| 581 | intel_engine_mask_t mask, |
| 582 | unsigned int retry) |
| 583 | { |
| 584 | return 0; |
| 585 | } |
| 586 | |
| 587 | typedef int (*reset_func)(struct intel_gt *, |
| 588 | intel_engine_mask_t engine_mask, |
| 589 | unsigned int retry); |
| 590 | |
| 591 | static reset_func intel_get_gpu_reset(const struct intel_gt *gt) |
| 592 | { |
| 593 | struct drm_i915_privateinteldrm_softc *i915 = gt->i915; |
| 594 | |
| 595 | if (is_mock_gt(gt)) |
| 596 | return mock_reset; |
| 597 | else if (INTEL_GEN(i915)((&(i915)->__info)->gen) >= 8) |
| 598 | return gen8_reset_engines; |
| 599 | else if (INTEL_GEN(i915)((&(i915)->__info)->gen) >= 6) |
| 600 | return gen6_reset_engines; |
| 601 | else if (INTEL_GEN(i915)((&(i915)->__info)->gen) >= 5) |
| 602 | return ilk_do_reset; |
| 603 | else if (IS_G4X(i915)(IS_PLATFORM(i915, INTEL_G45) || IS_PLATFORM(i915, INTEL_GM45 ))) |
| 604 | return g4x_do_reset; |
| 605 | else if (IS_G33(i915)IS_PLATFORM(i915, INTEL_G33) || IS_PINEVIEW(i915)IS_PLATFORM(i915, INTEL_PINEVIEW)) |
| 606 | return g33_do_reset; |
| 607 | else if (INTEL_GEN(i915)((&(i915)->__info)->gen) >= 3) |
| 608 | return i915_do_reset; |
| 609 | else |
| 610 | return NULL((void *)0); |
| 611 | } |
| 612 | |
| 613 | int __intel_gt_reset(struct intel_gt *gt, intel_engine_mask_t engine_mask) |
| 614 | { |
| 615 | const int retries = engine_mask == ALL_ENGINES((intel_engine_mask_t)~0ul) ? RESET_MAX_RETRIES3 : 1; |
| 616 | reset_func reset; |
| 617 | int ret = -ETIMEDOUT60; |
| 618 | int retry; |
| 619 | |
| 620 | reset = intel_get_gpu_reset(gt); |
| 621 | if (!reset) |
| 622 | return -ENODEV19; |
| 623 | |
| 624 | /* |
| 625 | * If the power well sleeps during the reset, the reset |
| 626 | * request may be dropped and never completes (causing -EIO). |
| 627 | */ |
| 628 | intel_uncore_forcewake_get(gt->uncore, FORCEWAKE_ALL); |
| 629 | for (retry = 0; ret == -ETIMEDOUT60 && retry < retries; retry++) { |
| 630 | GT_TRACE(gt, "engine_mask=%x\n", engine_mask)do { const struct intel_gt *gt__ __attribute__((__unused__)) = (gt); do { } while (0); } while (0); |
| 631 | preempt_disable(); |
| 632 | ret = reset(gt, engine_mask, retry); |
| 633 | preempt_enable(); |
| 634 | } |
| 635 | intel_uncore_forcewake_put(gt->uncore, FORCEWAKE_ALL); |
| 636 | |
| 637 | return ret; |
| 638 | } |
| 639 | |
| 640 | bool_Bool intel_has_gpu_reset(const struct intel_gt *gt) |
| 641 | { |
| 642 | if (!gt->i915->params.reset) |
| 643 | return NULL((void *)0); |
| 644 | |
| 645 | return intel_get_gpu_reset(gt); |
| 646 | } |
| 647 | |
| 648 | bool_Bool intel_has_reset_engine(const struct intel_gt *gt) |
| 649 | { |
| 650 | if (gt->i915->params.reset < 2) |
| 651 | return false0; |
| 652 | |
| 653 | return INTEL_INFO(gt->i915)(&(gt->i915)->__info)->has_reset_engine; |
| 654 | } |
| 655 | |
| 656 | int intel_reset_guc(struct intel_gt *gt) |
| 657 | { |
| 658 | u32 guc_domain = |
| 659 | INTEL_GEN(gt->i915)((&(gt->i915)->__info)->gen) >= 11 ? GEN11_GRDOM_GUC(1 << 3) : GEN9_GRDOM_GUC(1 << 5); |
| 660 | int ret; |
| 661 | |
| 662 | GEM_BUG_ON(!HAS_GT_UC(gt->i915))((void)0); |
| 663 | |
| 664 | intel_uncore_forcewake_get(gt->uncore, FORCEWAKE_ALL); |
| 665 | ret = gen6_hw_domain_reset(gt, guc_domain); |
| 666 | intel_uncore_forcewake_put(gt->uncore, FORCEWAKE_ALL); |
| 667 | |
| 668 | return ret; |
| 669 | } |
| 670 | |
| 671 | /* |
| 672 | * Ensure irq handler finishes, and not run again. |
| 673 | * Also return the active request so that we only search for it once. |
| 674 | */ |
| 675 | static void reset_prepare_engine(struct intel_engine_cs *engine) |
| 676 | { |
| 677 | /* |
| 678 | * During the reset sequence, we must prevent the engine from |
| 679 | * entering RC6. As the context state is undefined until we restart |
| 680 | * the engine, if it does enter RC6 during the reset, the state |
| 681 | * written to the powercontext is undefined and so we may lose |
| 682 | * GPU state upon resume, i.e. fail to restart after a reset. |
| 683 | */ |
| 684 | intel_uncore_forcewake_get(engine->uncore, FORCEWAKE_ALL); |
| 685 | if (engine->reset.prepare) |
| 686 | engine->reset.prepare(engine); |
| 687 | } |
| 688 | |
| 689 | static void revoke_mmaps(struct intel_gt *gt) |
| 690 | { |
| 691 | int i; |
| 692 | |
| 693 | for (i = 0; i < gt->ggtt->num_fences; i++) { |
| 694 | struct drm_vma_offset_node *node; |
| 695 | struct i915_vma *vma; |
| 696 | u64 vma_offset; |
| 697 | |
| 698 | vma = READ_ONCE(gt->ggtt->fence_regs[i].vma)({ typeof(gt->ggtt->fence_regs[i].vma) __tmp = *(volatile typeof(gt->ggtt->fence_regs[i].vma) *)&(gt->ggtt ->fence_regs[i].vma); membar_datadep_consumer(); __tmp; }); |
| 699 | if (!vma) |
| 700 | continue; |
| 701 | |
| 702 | if (!i915_vma_has_userfault(vma)) |
| 703 | continue; |
| 704 | |
| 705 | GEM_BUG_ON(vma->fence != >->ggtt->fence_regs[i])((void)0); |
| 706 | |
| 707 | if (!vma->mmo) |
| 708 | continue; |
| 709 | |
| 710 | node = &vma->mmo->vma_node; |
| 711 | vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT12; |
Value stored to 'vma_offset' is never read | |
| 712 | |
| 713 | #ifdef __linux__ |
| 714 | unmap_mapping_range(gt->i915->drm.anon_inode->i_mapping, |
| 715 | drm_vma_node_offset_addr(node) + vma_offset, |
| 716 | vma->size, |
| 717 | 1); |
| 718 | #else |
| 719 | { |
| 720 | struct drm_i915_privateinteldrm_softc *dev_priv = vma->obj->base.dev->dev_private; |
| 721 | struct vm_page *pg; |
| 722 | |
| 723 | for (pg = &dev_priv->pgs[atop(vma->node.start)((vma->node.start) >> 12)]; |
| 724 | pg != &dev_priv->pgs[atop(vma->node.start + vma->size)((vma->node.start + vma->size) >> 12)]; |
| 725 | pg++) |
| 726 | pmap_page_protect(pg, PROT_NONE0x00); |
| 727 | } |
| 728 | #endif |
| 729 | } |
| 730 | } |
| 731 | |
| 732 | static intel_engine_mask_t reset_prepare(struct intel_gt *gt) |
| 733 | { |
| 734 | struct intel_engine_cs *engine; |
| 735 | intel_engine_mask_t awake = 0; |
| 736 | enum intel_engine_id id; |
| 737 | |
| 738 | for_each_engine(engine, gt, id)for ((id) = 0; (id) < I915_NUM_ENGINES; (id)++) if (!((engine ) = (gt)->engine[(id)])) {} else { |
| 739 | if (intel_engine_pm_get_if_awake(engine)) |
| 740 | awake |= engine->mask; |
| 741 | reset_prepare_engine(engine); |
| 742 | } |
| 743 | |
| 744 | intel_uc_reset_prepare(>->uc); |
| 745 | |
| 746 | return awake; |
| 747 | } |
| 748 | |
| 749 | static void gt_revoke(struct intel_gt *gt) |
| 750 | { |
| 751 | revoke_mmaps(gt); |
| 752 | } |
| 753 | |
| 754 | static int gt_reset(struct intel_gt *gt, intel_engine_mask_t stalled_mask) |
| 755 | { |
| 756 | struct intel_engine_cs *engine; |
| 757 | enum intel_engine_id id; |
| 758 | int err; |
| 759 | |
| 760 | /* |
| 761 | * Everything depends on having the GTT running, so we need to start |
| 762 | * there. |
| 763 | */ |
| 764 | err = i915_ggtt_enable_hw(gt->i915); |
| 765 | if (err) |
| 766 | return err; |
| 767 | |
| 768 | for_each_engine(engine, gt, id)for ((id) = 0; (id) < I915_NUM_ENGINES; (id)++) if (!((engine ) = (gt)->engine[(id)])) {} else |
| 769 | __intel_engine_reset(engine, stalled_mask & engine->mask); |
| 770 | |
| 771 | intel_ggtt_restore_fences(gt->ggtt); |
| 772 | |
| 773 | return err; |
| 774 | } |
| 775 | |
| 776 | static void reset_finish_engine(struct intel_engine_cs *engine) |
| 777 | { |
| 778 | if (engine->reset.finish) |
| 779 | engine->reset.finish(engine); |
| 780 | intel_uncore_forcewake_put(engine->uncore, FORCEWAKE_ALL); |
| 781 | |
| 782 | intel_engine_signal_breadcrumbs(engine); |
| 783 | } |
| 784 | |
| 785 | static void reset_finish(struct intel_gt *gt, intel_engine_mask_t awake) |
| 786 | { |
| 787 | struct intel_engine_cs *engine; |
| 788 | enum intel_engine_id id; |
| 789 | |
| 790 | for_each_engine(engine, gt, id)for ((id) = 0; (id) < I915_NUM_ENGINES; (id)++) if (!((engine ) = (gt)->engine[(id)])) {} else { |
| 791 | reset_finish_engine(engine); |
| 792 | if (awake & engine->mask) |
| 793 | intel_engine_pm_put(engine); |
| 794 | } |
| 795 | } |
| 796 | |
| 797 | static void nop_submit_request(struct i915_request *request) |
| 798 | { |
| 799 | struct intel_engine_cs *engine = request->engine; |
| 800 | unsigned long flags; |
| 801 | |
| 802 | RQ_TRACE(request, "-EIO\n")do { const struct i915_request *rq__ = (request); do { const struct intel_engine_cs *e__ __attribute__((__unused__)) = (rq__-> engine); do { } while (0); } while (0); } while (0); |
| 803 | i915_request_set_error_once(request, -EIO5); |
| 804 | |
| 805 | spin_lock_irqsave(&engine->active.lock, flags)do { flags = 0; mtx_enter(&engine->active.lock); } while (0); |
| 806 | __i915_request_submit(request); |
| 807 | i915_request_mark_complete(request); |
| 808 | spin_unlock_irqrestore(&engine->active.lock, flags)do { (void)(flags); mtx_leave(&engine->active.lock); } while (0); |
| 809 | |
| 810 | intel_engine_signal_breadcrumbs(engine); |
| 811 | } |
| 812 | |
| 813 | static void __intel_gt_set_wedged(struct intel_gt *gt) |
| 814 | { |
| 815 | struct intel_engine_cs *engine; |
| 816 | intel_engine_mask_t awake; |
| 817 | enum intel_engine_id id; |
| 818 | |
| 819 | if (test_bit(I915_WEDGED(64 - 1), >->reset.flags)) |
| 820 | return; |
| 821 | |
| 822 | GT_TRACE(gt, "start\n")do { const struct intel_gt *gt__ __attribute__((__unused__)) = (gt); do { } while (0); } while (0); |
| 823 | |
| 824 | /* |
| 825 | * First, stop submission to hw, but do not yet complete requests by |
| 826 | * rolling the global seqno forward (since this would complete requests |
| 827 | * for which we haven't set the fence error to EIO yet). |
| 828 | */ |
| 829 | awake = reset_prepare(gt); |
| 830 | |
| 831 | /* Even if the GPU reset fails, it should still stop the engines */ |
| 832 | if (!INTEL_INFO(gt->i915)(&(gt->i915)->__info)->gpu_reset_clobbers_display) |
| 833 | __intel_gt_reset(gt, ALL_ENGINES((intel_engine_mask_t)~0ul)); |
| 834 | |
| 835 | for_each_engine(engine, gt, id)for ((id) = 0; (id) < I915_NUM_ENGINES; (id)++) if (!((engine ) = (gt)->engine[(id)])) {} else |
| 836 | engine->submit_request = nop_submit_request; |
| 837 | |
| 838 | /* |
| 839 | * Make sure no request can slip through without getting completed by |
| 840 | * either this call here to intel_engine_write_global_seqno, or the one |
| 841 | * in nop_submit_request. |
| 842 | */ |
| 843 | synchronize_rcu_expedited(); |
| 844 | set_bit(I915_WEDGED(64 - 1), >->reset.flags); |
| 845 | |
| 846 | /* Mark all executing requests as skipped */ |
| 847 | for_each_engine(engine, gt, id)for ((id) = 0; (id) < I915_NUM_ENGINES; (id)++) if (!((engine ) = (gt)->engine[(id)])) {} else |
| 848 | if (engine->reset.cancel) |
| 849 | engine->reset.cancel(engine); |
| 850 | |
| 851 | reset_finish(gt, awake); |
| 852 | |
| 853 | GT_TRACE(gt, "end\n")do { const struct intel_gt *gt__ __attribute__((__unused__)) = (gt); do { } while (0); } while (0); |
| 854 | } |
| 855 | |
| 856 | void intel_gt_set_wedged(struct intel_gt *gt) |
| 857 | { |
| 858 | intel_wakeref_t wakeref; |
| 859 | |
| 860 | if (test_bit(I915_WEDGED(64 - 1), >->reset.flags)) |
| 861 | return; |
| 862 | |
| 863 | wakeref = intel_runtime_pm_get(gt->uncore->rpm); |
| 864 | mutex_lock(>->reset.mutex)rw_enter_write(>->reset.mutex); |
| 865 | |
| 866 | if (GEM_SHOW_DEBUG()(0)) { |
| 867 | struct drm_printer p = drm_debug_printer(__func__); |
| 868 | struct intel_engine_cs *engine; |
| 869 | enum intel_engine_id id; |
| 870 | |
| 871 | drm_printf(&p, "called from %pS\n", (void *)_RET_IP___builtin_return_address(0)); |
| 872 | for_each_engine(engine, gt, id)for ((id) = 0; (id) < I915_NUM_ENGINES; (id)++) if (!((engine ) = (gt)->engine[(id)])) {} else { |
| 873 | if (intel_engine_is_idle(engine)) |
| 874 | continue; |
| 875 | |
| 876 | intel_engine_dump(engine, &p, "%s\n", engine->name); |
| 877 | } |
| 878 | } |
| 879 | |
| 880 | __intel_gt_set_wedged(gt); |
| 881 | |
| 882 | mutex_unlock(>->reset.mutex)rw_exit_write(>->reset.mutex); |
| 883 | intel_runtime_pm_put(gt->uncore->rpm, wakeref); |
| 884 | } |
| 885 | |
| 886 | static bool_Bool __intel_gt_unset_wedged(struct intel_gt *gt) |
| 887 | { |
| 888 | struct intel_gt_timelines *timelines = >->timelines; |
| 889 | struct intel_timeline *tl; |
| 890 | bool_Bool ok; |
| 891 | |
| 892 | if (!test_bit(I915_WEDGED(64 - 1), >->reset.flags)) |
| 893 | return true1; |
| 894 | |
| 895 | /* Never fully initialised, recovery impossible */ |
| 896 | if (intel_gt_has_unrecoverable_error(gt)) |
| 897 | return false0; |
| 898 | |
| 899 | GT_TRACE(gt, "start\n")do { const struct intel_gt *gt__ __attribute__((__unused__)) = (gt); do { } while (0); } while (0); |
| 900 | |
| 901 | /* |
| 902 | * Before unwedging, make sure that all pending operations |
| 903 | * are flushed and errored out - we may have requests waiting upon |
| 904 | * third party fences. We marked all inflight requests as EIO, and |
| 905 | * every execbuf since returned EIO, for consistency we want all |
| 906 | * the currently pending requests to also be marked as EIO, which |
| 907 | * is done inside our nop_submit_request - and so we must wait. |
| 908 | * |
| 909 | * No more can be submitted until we reset the wedged bit. |
| 910 | */ |
| 911 | spin_lock(&timelines->lock)mtx_enter(&timelines->lock); |
| 912 | list_for_each_entry(tl, &timelines->active_list, link)for (tl = ({ const __typeof( ((__typeof(*tl) *)0)->link ) * __mptr = ((&timelines->active_list)->next); (__typeof (*tl) *)( (char *)__mptr - __builtin_offsetof(__typeof(*tl), link ) );}); &tl->link != (&timelines->active_list); tl = ({ const __typeof( ((__typeof(*tl) *)0)->link ) *__mptr = (tl->link.next); (__typeof(*tl) *)( (char *)__mptr - __builtin_offsetof (__typeof(*tl), link) );})) { |
| 913 | struct dma_fence *fence; |
| 914 | |
| 915 | fence = i915_active_fence_get(&tl->last_request); |
| 916 | if (!fence) |
| 917 | continue; |
| 918 | |
| 919 | spin_unlock(&timelines->lock)mtx_leave(&timelines->lock); |
| 920 | |
| 921 | /* |
| 922 | * All internal dependencies (i915_requests) will have |
| 923 | * been flushed by the set-wedge, but we may be stuck waiting |
| 924 | * for external fences. These should all be capped to 10s |
| 925 | * (I915_FENCE_TIMEOUT) so this wait should not be unbounded |
| 926 | * in the worst case. |
| 927 | */ |
| 928 | dma_fence_default_wait(fence, false0, MAX_SCHEDULE_TIMEOUT(0x7fffffff)); |
| 929 | dma_fence_put(fence); |
| 930 | |
| 931 | /* Restart iteration after droping lock */ |
| 932 | spin_lock(&timelines->lock)mtx_enter(&timelines->lock); |
| 933 | tl = list_entry(&timelines->active_list, typeof(*tl), link)({ const __typeof( ((typeof(*tl) *)0)->link ) *__mptr = (& timelines->active_list); (typeof(*tl) *)( (char *)__mptr - __builtin_offsetof(typeof(*tl), link) );}); |
| 934 | } |
| 935 | spin_unlock(&timelines->lock)mtx_leave(&timelines->lock); |
| 936 | |
| 937 | /* We must reset pending GPU events before restoring our submission */ |
| 938 | ok = !HAS_EXECLISTS(gt->i915)((&(gt->i915)->__info)->has_logical_ring_contexts ); /* XXX better agnosticism desired */ |
| 939 | if (!INTEL_INFO(gt->i915)(&(gt->i915)->__info)->gpu_reset_clobbers_display) |
| 940 | ok = __intel_gt_reset(gt, ALL_ENGINES((intel_engine_mask_t)~0ul)) == 0; |
| 941 | if (!ok) { |
| 942 | /* |
| 943 | * Warn CI about the unrecoverable wedged condition. |
| 944 | * Time for a reboot. |
| 945 | */ |
| 946 | add_taint_for_CI(gt->i915, TAINT_WARN1); |
| 947 | return false0; |
| 948 | } |
| 949 | |
| 950 | /* |
| 951 | * Undo nop_submit_request. We prevent all new i915 requests from |
| 952 | * being queued (by disallowing execbuf whilst wedged) so having |
| 953 | * waited for all active requests above, we know the system is idle |
| 954 | * and do not have to worry about a thread being inside |
| 955 | * engine->submit_request() as we swap over. So unlike installing |
| 956 | * the nop_submit_request on reset, we can do this from normal |
| 957 | * context and do not require stop_machine(). |
| 958 | */ |
| 959 | intel_engines_reset_default_submission(gt); |
| 960 | |
| 961 | GT_TRACE(gt, "end\n")do { const struct intel_gt *gt__ __attribute__((__unused__)) = (gt); do { } while (0); } while (0); |
| 962 | |
| 963 | smp_mb__before_atomic()do { } while (0); /* complete takeover before enabling execbuf */ |
| 964 | clear_bit(I915_WEDGED(64 - 1), >->reset.flags); |
| 965 | |
| 966 | return true1; |
| 967 | } |
| 968 | |
| 969 | bool_Bool intel_gt_unset_wedged(struct intel_gt *gt) |
| 970 | { |
| 971 | bool_Bool result; |
| 972 | |
| 973 | mutex_lock(>->reset.mutex)rw_enter_write(>->reset.mutex); |
| 974 | result = __intel_gt_unset_wedged(gt); |
| 975 | mutex_unlock(>->reset.mutex)rw_exit_write(>->reset.mutex); |
| 976 | |
| 977 | return result; |
| 978 | } |
| 979 | |
| 980 | static int do_reset(struct intel_gt *gt, intel_engine_mask_t stalled_mask) |
| 981 | { |
| 982 | int err, i; |
| 983 | |
| 984 | gt_revoke(gt); |
| 985 | |
| 986 | err = __intel_gt_reset(gt, ALL_ENGINES((intel_engine_mask_t)~0ul)); |
| 987 | for (i = 0; err && i < RESET_MAX_RETRIES3; i++) { |
| 988 | drm_msleep(10 * (i + 1))mdelay(10 * (i + 1)); |
| 989 | err = __intel_gt_reset(gt, ALL_ENGINES((intel_engine_mask_t)~0ul)); |
| 990 | } |
| 991 | if (err) |
| 992 | return err; |
| 993 | |
| 994 | return gt_reset(gt, stalled_mask); |
| 995 | } |
| 996 | |
| 997 | static int resume(struct intel_gt *gt) |
| 998 | { |
| 999 | struct intel_engine_cs *engine; |
| 1000 | enum intel_engine_id id; |
| 1001 | int ret; |
| 1002 | |
| 1003 | for_each_engine(engine, gt, id)for ((id) = 0; (id) < I915_NUM_ENGINES; (id)++) if (!((engine ) = (gt)->engine[(id)])) {} else { |
| 1004 | ret = intel_engine_resume(engine); |
| 1005 | if (ret) |
| 1006 | return ret; |
| 1007 | } |
| 1008 | |
| 1009 | return 0; |
| 1010 | } |
| 1011 | |
| 1012 | /** |
| 1013 | * intel_gt_reset - reset chip after a hang |
| 1014 | * @gt: #intel_gt to reset |
| 1015 | * @stalled_mask: mask of the stalled engines with the guilty requests |
| 1016 | * @reason: user error message for why we are resetting |
| 1017 | * |
| 1018 | * Reset the chip. Useful if a hang is detected. Marks the device as wedged |
| 1019 | * on failure. |
| 1020 | * |
| 1021 | * Procedure is fairly simple: |
| 1022 | * - reset the chip using the reset reg |
| 1023 | * - re-init context state |
| 1024 | * - re-init hardware status page |
| 1025 | * - re-init ring buffer |
| 1026 | * - re-init interrupt state |
| 1027 | * - re-init display |
| 1028 | */ |
| 1029 | void intel_gt_reset(struct intel_gt *gt, |
| 1030 | intel_engine_mask_t stalled_mask, |
| 1031 | const char *reason) |
| 1032 | { |
| 1033 | intel_engine_mask_t awake; |
| 1034 | int ret; |
| 1035 | |
| 1036 | GT_TRACE(gt, "flags=%lx\n", gt->reset.flags)do { const struct intel_gt *gt__ __attribute__((__unused__)) = (gt); do { } while (0); } while (0); |
| 1037 | |
| 1038 | might_sleep()assertwaitok(); |
| 1039 | GEM_BUG_ON(!test_bit(I915_RESET_BACKOFF, >->reset.flags))((void)0); |
| 1040 | mutex_lock(>->reset.mutex)rw_enter_write(>->reset.mutex); |
| 1041 | |
| 1042 | /* Clear any previous failed attempts at recovery. Time to try again. */ |
| 1043 | if (!__intel_gt_unset_wedged(gt)) |
| 1044 | goto unlock; |
| 1045 | |
| 1046 | if (reason) |
| 1047 | drm_notice(>->i915->drm,printf("drm:pid%d:%s *NOTICE* " "[drm] " "Resetting chip for %s\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , reason ) |
| 1048 | "Resetting chip for %s\n", reason)printf("drm:pid%d:%s *NOTICE* " "[drm] " "Resetting chip for %s\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , reason ); |
| 1049 | atomic_inc(>->i915->gpu_error.reset_count)__sync_fetch_and_add(>->i915->gpu_error.reset_count , 1); |
| 1050 | |
| 1051 | awake = reset_prepare(gt); |
| 1052 | |
| 1053 | if (!intel_has_gpu_reset(gt)) { |
| 1054 | if (gt->i915->params.reset) |
| 1055 | drm_err(>->i915->drm, "GPU reset not supported\n")printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "GPU reset not supported\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__); |
| 1056 | else |
| 1057 | drm_dbg(>->i915->drm, "GPU reset disabled\n")drm_dev_dbg((>->i915->drm)->dev, DRM_UT_DRIVER , "GPU reset disabled\n"); |
| 1058 | goto error; |
| 1059 | } |
| 1060 | |
| 1061 | if (INTEL_INFO(gt->i915)(&(gt->i915)->__info)->gpu_reset_clobbers_display) |
| 1062 | intel_runtime_pm_disable_interrupts(gt->i915); |
| 1063 | |
| 1064 | if (do_reset(gt, stalled_mask)) { |
| 1065 | drm_err(>->i915->drm, "Failed to reset chip\n")printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "Failed to reset chip\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__); |
| 1066 | goto taint; |
| 1067 | } |
| 1068 | |
| 1069 | if (INTEL_INFO(gt->i915)(&(gt->i915)->__info)->gpu_reset_clobbers_display) |
| 1070 | intel_runtime_pm_enable_interrupts(gt->i915); |
| 1071 | |
| 1072 | intel_overlay_reset(gt->i915); |
| 1073 | |
| 1074 | /* |
| 1075 | * Next we need to restore the context, but we don't use those |
| 1076 | * yet either... |
| 1077 | * |
| 1078 | * Ring buffer needs to be re-initialized in the KMS case, or if X |
| 1079 | * was running at the time of the reset (i.e. we weren't VT |
| 1080 | * switched away). |
| 1081 | */ |
| 1082 | ret = intel_gt_init_hw(gt); |
| 1083 | if (ret) { |
| 1084 | drm_err(>->i915->drm,printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "Failed to initialise HW following reset (%d)\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , ret) |
| 1085 | "Failed to initialise HW following reset (%d)\n",printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "Failed to initialise HW following reset (%d)\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , ret) |
| 1086 | ret)printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "Failed to initialise HW following reset (%d)\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , ret); |
| 1087 | goto taint; |
| 1088 | } |
| 1089 | |
| 1090 | ret = resume(gt); |
| 1091 | if (ret) |
| 1092 | goto taint; |
| 1093 | |
| 1094 | finish: |
| 1095 | reset_finish(gt, awake); |
| 1096 | unlock: |
| 1097 | mutex_unlock(>->reset.mutex)rw_exit_write(>->reset.mutex); |
| 1098 | return; |
| 1099 | |
| 1100 | taint: |
| 1101 | /* |
| 1102 | * History tells us that if we cannot reset the GPU now, we |
| 1103 | * never will. This then impacts everything that is run |
| 1104 | * subsequently. On failing the reset, we mark the driver |
| 1105 | * as wedged, preventing further execution on the GPU. |
| 1106 | * We also want to go one step further and add a taint to the |
| 1107 | * kernel so that any subsequent faults can be traced back to |
| 1108 | * this failure. This is important for CI, where if the |
| 1109 | * GPU/driver fails we would like to reboot and restart testing |
| 1110 | * rather than continue on into oblivion. For everyone else, |
| 1111 | * the system should still plod along, but they have been warned! |
| 1112 | */ |
| 1113 | add_taint_for_CI(gt->i915, TAINT_WARN1); |
| 1114 | error: |
| 1115 | __intel_gt_set_wedged(gt); |
| 1116 | goto finish; |
| 1117 | } |
| 1118 | |
| 1119 | static inline int intel_gt_reset_engine(struct intel_engine_cs *engine) |
| 1120 | { |
| 1121 | return __intel_gt_reset(engine->gt, engine->mask); |
| 1122 | } |
| 1123 | |
| 1124 | /** |
| 1125 | * intel_engine_reset - reset GPU engine to recover from a hang |
| 1126 | * @engine: engine to reset |
| 1127 | * @msg: reason for GPU reset; or NULL for no drm_notice() |
| 1128 | * |
| 1129 | * Reset a specific GPU engine. Useful if a hang is detected. |
| 1130 | * Returns zero on successful reset or otherwise an error code. |
| 1131 | * |
| 1132 | * Procedure is: |
| 1133 | * - identifies the request that caused the hang and it is dropped |
| 1134 | * - reset engine (which will force the engine to idle) |
| 1135 | * - re-init/configure engine |
| 1136 | */ |
| 1137 | int intel_engine_reset(struct intel_engine_cs *engine, const char *msg) |
| 1138 | { |
| 1139 | struct intel_gt *gt = engine->gt; |
| 1140 | bool_Bool uses_guc = intel_engine_in_guc_submission_mode(engine); |
| 1141 | int ret; |
| 1142 | |
| 1143 | ENGINE_TRACE(engine, "flags=%lx\n", gt->reset.flags)do { const struct intel_engine_cs *e__ __attribute__((__unused__ )) = (engine); do { } while (0); } while (0); |
| 1144 | GEM_BUG_ON(!test_bit(I915_RESET_ENGINE + engine->id, >->reset.flags))((void)0); |
| 1145 | |
| 1146 | if (!intel_engine_pm_get_if_awake(engine)) |
| 1147 | return 0; |
| 1148 | |
| 1149 | reset_prepare_engine(engine); |
| 1150 | |
| 1151 | if (msg) |
| 1152 | drm_notice(&engine->i915->drm,printf("drm:pid%d:%s *NOTICE* " "[drm] " "Resetting %s for %s\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , engine ->name, msg) |
| 1153 | "Resetting %s for %s\n", engine->name, msg)printf("drm:pid%d:%s *NOTICE* " "[drm] " "Resetting %s for %s\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , engine ->name, msg); |
| 1154 | atomic_inc(&engine->i915->gpu_error.reset_engine_count[engine->uabi_class])__sync_fetch_and_add(&engine->i915->gpu_error.reset_engine_count [engine->uabi_class], 1); |
| 1155 | |
| 1156 | if (!uses_guc) |
| 1157 | ret = intel_gt_reset_engine(engine); |
| 1158 | else |
| 1159 | ret = intel_guc_reset_engine(&engine->gt->uc.guc, engine); |
| 1160 | if (ret) { |
| 1161 | /* If we fail here, we expect to fallback to a global reset */ |
| 1162 | drm_dbg(>->i915->drm, "%sFailed to reset %s, ret=%d\n",drm_dev_dbg((>->i915->drm)->dev, DRM_UT_DRIVER , "%sFailed to reset %s, ret=%d\n", uses_guc ? "GuC " : "", engine ->name, ret) |
| 1163 | uses_guc ? "GuC " : "", engine->name, ret)drm_dev_dbg((>->i915->drm)->dev, DRM_UT_DRIVER , "%sFailed to reset %s, ret=%d\n", uses_guc ? "GuC " : "", engine ->name, ret); |
| 1164 | goto out; |
| 1165 | } |
| 1166 | |
| 1167 | /* |
| 1168 | * The request that caused the hang is stuck on elsp, we know the |
| 1169 | * active request and can drop it, adjust head to skip the offending |
| 1170 | * request to resume executing remaining requests in the queue. |
| 1171 | */ |
| 1172 | __intel_engine_reset(engine, true1); |
| 1173 | |
| 1174 | /* |
| 1175 | * The engine and its registers (and workarounds in case of render) |
| 1176 | * have been reset to their default values. Follow the init_ring |
| 1177 | * process to program RING_MODE, HWSP and re-enable submission. |
| 1178 | */ |
| 1179 | ret = intel_engine_resume(engine); |
| 1180 | |
| 1181 | out: |
| 1182 | intel_engine_cancel_stop_cs(engine); |
| 1183 | reset_finish_engine(engine); |
| 1184 | intel_engine_pm_put_async(engine); |
| 1185 | return ret; |
| 1186 | } |
| 1187 | |
| 1188 | static void intel_gt_reset_global(struct intel_gt *gt, |
| 1189 | u32 engine_mask, |
| 1190 | const char *reason) |
| 1191 | { |
| 1192 | #ifdef notyet |
| 1193 | struct kobject *kobj = >->i915->drm.primary->kdev->kobj; |
| 1194 | char *error_event[] = { I915_ERROR_UEVENT"ERROR" "=1", NULL((void *)0) }; |
| 1195 | char *reset_event[] = { I915_RESET_UEVENT"RESET" "=1", NULL((void *)0) }; |
| 1196 | char *reset_done_event[] = { I915_ERROR_UEVENT"ERROR" "=0", NULL((void *)0) }; |
| 1197 | #endif |
| 1198 | struct intel_wedge_me w; |
| 1199 | |
| 1200 | kobject_uevent_env(kobj, KOBJ_CHANGE, error_event); |
| 1201 | |
| 1202 | drm_dbg(>->i915->drm, "resetting chip, engines=%x\n", engine_mask)drm_dev_dbg((>->i915->drm)->dev, DRM_UT_DRIVER , "resetting chip, engines=%x\n", engine_mask); |
| 1203 | kobject_uevent_env(kobj, KOBJ_CHANGE, reset_event); |
| 1204 | |
| 1205 | /* Use a watchdog to ensure that our reset completes */ |
| 1206 | intel_wedge_on_timeout(&w, gt, 5 * HZ)for (__intel_init_wedge((&w), (gt), (5 * hz), __func__); ( &w)->gt; __intel_fini_wedge((&w))) { |
| 1207 | intel_prepare_reset(gt->i915); |
| 1208 | |
| 1209 | /* Flush everyone using a resource about to be clobbered */ |
| 1210 | synchronize_srcu_expedited(>->reset.backoff_srcu); |
| 1211 | |
| 1212 | intel_gt_reset(gt, engine_mask, reason); |
| 1213 | |
| 1214 | intel_finish_reset(gt->i915); |
| 1215 | } |
| 1216 | |
| 1217 | if (!test_bit(I915_WEDGED(64 - 1), >->reset.flags)) |
| 1218 | kobject_uevent_env(kobj, KOBJ_CHANGE, reset_done_event); |
| 1219 | } |
| 1220 | |
| 1221 | /** |
| 1222 | * intel_gt_handle_error - handle a gpu error |
| 1223 | * @gt: the intel_gt |
| 1224 | * @engine_mask: mask representing engines that are hung |
| 1225 | * @flags: control flags |
| 1226 | * @fmt: Error message format string |
| 1227 | * |
| 1228 | * Do some basic checking of register state at error time and |
| 1229 | * dump it to the syslog. Also call i915_capture_error_state() to make |
| 1230 | * sure we get a record and make it available in debugfs. Fire a uevent |
| 1231 | * so userspace knows something bad happened (should trigger collection |
| 1232 | * of a ring dump etc.). |
| 1233 | */ |
| 1234 | void intel_gt_handle_error(struct intel_gt *gt, |
| 1235 | intel_engine_mask_t engine_mask, |
| 1236 | unsigned long flags, |
| 1237 | const char *fmt, ...) |
| 1238 | { |
| 1239 | struct intel_engine_cs *engine; |
| 1240 | intel_wakeref_t wakeref; |
| 1241 | intel_engine_mask_t tmp; |
| 1242 | char error_msg[80]; |
| 1243 | char *msg = NULL((void *)0); |
| 1244 | |
| 1245 | if (fmt) { |
| 1246 | va_list args; |
| 1247 | |
| 1248 | va_start(args, fmt)__builtin_va_start((args), fmt); |
| 1249 | vscnprintf(error_msg, sizeof(error_msg), fmt, args); |
| 1250 | va_end(args)__builtin_va_end((args)); |
| 1251 | |
| 1252 | msg = error_msg; |
| 1253 | } |
| 1254 | |
| 1255 | /* |
| 1256 | * In most cases it's guaranteed that we get here with an RPM |
| 1257 | * reference held, for example because there is a pending GPU |
| 1258 | * request that won't finish until the reset is done. This |
| 1259 | * isn't the case at least when we get here by doing a |
| 1260 | * simulated reset via debugfs, so get an RPM reference. |
| 1261 | */ |
| 1262 | wakeref = intel_runtime_pm_get(gt->uncore->rpm); |
| 1263 | |
| 1264 | engine_mask &= gt->info.engine_mask; |
| 1265 | |
| 1266 | if (flags & I915_ERROR_CAPTURE(1UL << (0))) { |
| 1267 | i915_capture_error_state(gt->i915); |
| 1268 | intel_gt_clear_error_registers(gt, engine_mask); |
| 1269 | } |
| 1270 | |
| 1271 | /* |
| 1272 | * Try engine reset when available. We fall back to full reset if |
| 1273 | * single reset fails. |
| 1274 | */ |
| 1275 | if (intel_has_reset_engine(gt) && !intel_gt_is_wedged(gt)) { |
| 1276 | for_each_engine_masked(engine, gt, engine_mask, tmp)for ((tmp) = (engine_mask) & (gt)->info.engine_mask; ( tmp) ? ((engine) = (gt)->engine[({ int __idx = ffs(tmp) - 1 ; tmp &= ~(1UL << (__idx)); __idx; })]), 1 : 0;) { |
| 1277 | BUILD_BUG_ON(I915_RESET_MODESET >= I915_RESET_ENGINE)extern char _ctassert[(!(1 >= 2)) ? 1 : -1 ] __attribute__ ((__unused__)); |
| 1278 | if (test_and_set_bit(I915_RESET_ENGINE2 + engine->id, |
| 1279 | >->reset.flags)) |
| 1280 | continue; |
| 1281 | |
| 1282 | if (intel_engine_reset(engine, msg) == 0) |
| 1283 | engine_mask &= ~engine->mask; |
| 1284 | |
| 1285 | clear_and_wake_up_bit(I915_RESET_ENGINE2 + engine->id, |
| 1286 | >->reset.flags); |
| 1287 | } |
| 1288 | } |
| 1289 | |
| 1290 | if (!engine_mask) |
| 1291 | goto out; |
| 1292 | |
| 1293 | /* Full reset needs the mutex, stop any other user trying to do so. */ |
| 1294 | if (test_and_set_bit(I915_RESET_BACKOFF0, >->reset.flags)) { |
| 1295 | wait_event(gt->reset.queue,do { if (!(!test_bit(0, >->reset.flags))) ({ long ret = 0; do { int __error; unsigned long deadline; ((!cold) ? (void )0 : __assert("diagnostic ", "/usr/src/sys/dev/pci/drm/i915/gt/intel_reset.c" , 1296, "!cold")); mtx_enter(&sch_mtx); deadline = jiffies + ret; __error = msleep(>->reset.queue, &sch_mtx , 0, "drmweti", ret); ret = deadline - jiffies; if (__error == -1 || __error == 4) { ret = -4; mtx_leave(&sch_mtx); break ; } if ((0) > 0 && (ret <= 0 || __error == 35)) { mtx_leave(&sch_mtx); ret = ((!test_bit(0, >-> reset.flags))) ? 1 : 0; break; } mtx_leave(&sch_mtx); } while (ret > 0 && !(!test_bit(0, >->reset.flags ))); ret; }); } while (0) |
| 1296 | !test_bit(I915_RESET_BACKOFF, >->reset.flags))do { if (!(!test_bit(0, >->reset.flags))) ({ long ret = 0; do { int __error; unsigned long deadline; ((!cold) ? (void )0 : __assert("diagnostic ", "/usr/src/sys/dev/pci/drm/i915/gt/intel_reset.c" , 1296, "!cold")); mtx_enter(&sch_mtx); deadline = jiffies + ret; __error = msleep(>->reset.queue, &sch_mtx , 0, "drmweti", ret); ret = deadline - jiffies; if (__error == -1 || __error == 4) { ret = -4; mtx_leave(&sch_mtx); break ; } if ((0) > 0 && (ret <= 0 || __error == 35)) { mtx_leave(&sch_mtx); ret = ((!test_bit(0, >-> reset.flags))) ? 1 : 0; break; } mtx_leave(&sch_mtx); } while (ret > 0 && !(!test_bit(0, >->reset.flags ))); ret; }); } while (0); |
| 1297 | goto out; /* piggy-back on the other reset */ |
| 1298 | } |
| 1299 | |
| 1300 | /* Make sure i915_reset_trylock() sees the I915_RESET_BACKOFF */ |
| 1301 | synchronize_rcu_expedited(); |
| 1302 | |
| 1303 | /* Prevent any other reset-engine attempt. */ |
| 1304 | for_each_engine(engine, gt, tmp)for ((tmp) = 0; (tmp) < I915_NUM_ENGINES; (tmp)++) if (!(( engine) = (gt)->engine[(tmp)])) {} else { |
| 1305 | while (test_and_set_bit(I915_RESET_ENGINE2 + engine->id, |
| 1306 | >->reset.flags)) |
| 1307 | wait_on_bit(>->reset.flags, |
| 1308 | I915_RESET_ENGINE2 + engine->id, |
| 1309 | TASK_UNINTERRUPTIBLE0); |
| 1310 | } |
| 1311 | |
| 1312 | intel_gt_reset_global(gt, engine_mask, msg); |
| 1313 | |
| 1314 | for_each_engine(engine, gt, tmp)for ((tmp) = 0; (tmp) < I915_NUM_ENGINES; (tmp)++) if (!(( engine) = (gt)->engine[(tmp)])) {} else |
| 1315 | clear_bit_unlock(I915_RESET_ENGINE2 + engine->id, |
| 1316 | >->reset.flags); |
| 1317 | clear_bit_unlock(I915_RESET_BACKOFF0, >->reset.flags); |
| 1318 | smp_mb__after_atomic()do { } while (0); |
| 1319 | wake_up_all(>->reset.queue)wake_up(>->reset.queue); |
| 1320 | |
| 1321 | out: |
| 1322 | intel_runtime_pm_put(gt->uncore->rpm, wakeref); |
| 1323 | } |
| 1324 | |
| 1325 | int intel_gt_reset_trylock(struct intel_gt *gt, int *srcu) |
| 1326 | { |
| 1327 | might_lock(>->reset.backoff_srcu); |
| 1328 | might_sleep()assertwaitok(); |
| 1329 | |
| 1330 | rcu_read_lock(); |
| 1331 | while (test_bit(I915_RESET_BACKOFF0, >->reset.flags)) { |
| 1332 | rcu_read_unlock(); |
| 1333 | |
| 1334 | if (wait_event_interruptible(gt->reset.queue,({ int __ret = 0; if (!(!test_bit(0, >->reset.flags) )) __ret = ({ long ret = 0; do { int __error; unsigned long deadline ; ((!cold) ? (void)0 : __assert("diagnostic ", "/usr/src/sys/dev/pci/drm/i915/gt/intel_reset.c" , 1336, "!cold")); mtx_enter(&sch_mtx); deadline = jiffies + ret; __error = msleep(>->reset.queue, &sch_mtx , 0x100, "drmweti", ret); ret = deadline - jiffies; if (__error == -1 || __error == 4) { ret = -4; mtx_leave(&sch_mtx); break ; } if ((0) > 0 && (ret <= 0 || __error == 35)) { mtx_leave(&sch_mtx); ret = ((!test_bit(0, >-> reset.flags))) ? 1 : 0; break; } mtx_leave(&sch_mtx); } while (ret > 0 && !(!test_bit(0, >->reset.flags ))); ret; }); __ret; }) |
| 1335 | !test_bit(I915_RESET_BACKOFF,({ int __ret = 0; if (!(!test_bit(0, >->reset.flags) )) __ret = ({ long ret = 0; do { int __error; unsigned long deadline ; ((!cold) ? (void)0 : __assert("diagnostic ", "/usr/src/sys/dev/pci/drm/i915/gt/intel_reset.c" , 1336, "!cold")); mtx_enter(&sch_mtx); deadline = jiffies + ret; __error = msleep(>->reset.queue, &sch_mtx , 0x100, "drmweti", ret); ret = deadline - jiffies; if (__error == -1 || __error == 4) { ret = -4; mtx_leave(&sch_mtx); break ; } if ((0) > 0 && (ret <= 0 || __error == 35)) { mtx_leave(&sch_mtx); ret = ((!test_bit(0, >-> reset.flags))) ? 1 : 0; break; } mtx_leave(&sch_mtx); } while (ret > 0 && !(!test_bit(0, >->reset.flags ))); ret; }); __ret; }) |
| 1336 | >->reset.flags))({ int __ret = 0; if (!(!test_bit(0, >->reset.flags) )) __ret = ({ long ret = 0; do { int __error; unsigned long deadline ; ((!cold) ? (void)0 : __assert("diagnostic ", "/usr/src/sys/dev/pci/drm/i915/gt/intel_reset.c" , 1336, "!cold")); mtx_enter(&sch_mtx); deadline = jiffies + ret; __error = msleep(>->reset.queue, &sch_mtx , 0x100, "drmweti", ret); ret = deadline - jiffies; if (__error == -1 || __error == 4) { ret = -4; mtx_leave(&sch_mtx); break ; } if ((0) > 0 && (ret <= 0 || __error == 35)) { mtx_leave(&sch_mtx); ret = ((!test_bit(0, >-> reset.flags))) ? 1 : 0; break; } mtx_leave(&sch_mtx); } while (ret > 0 && !(!test_bit(0, >->reset.flags ))); ret; }); __ret; })) |
| 1337 | return -EINTR4; |
| 1338 | |
| 1339 | rcu_read_lock(); |
| 1340 | } |
| 1341 | *srcu = srcu_read_lock(>->reset.backoff_srcu)0; |
| 1342 | rcu_read_unlock(); |
| 1343 | |
| 1344 | return 0; |
| 1345 | } |
| 1346 | |
| 1347 | void intel_gt_reset_unlock(struct intel_gt *gt, int tag) |
| 1348 | __releases(>->reset.backoff_srcu) |
| 1349 | { |
| 1350 | srcu_read_unlock(>->reset.backoff_srcu, tag); |
| 1351 | } |
| 1352 | |
| 1353 | int intel_gt_terminally_wedged(struct intel_gt *gt) |
| 1354 | { |
| 1355 | might_sleep()assertwaitok(); |
| 1356 | |
| 1357 | if (!intel_gt_is_wedged(gt)) |
| 1358 | return 0; |
| 1359 | |
| 1360 | if (intel_gt_has_unrecoverable_error(gt)) |
| 1361 | return -EIO5; |
| 1362 | |
| 1363 | /* Reset still in progress? Maybe we will recover? */ |
| 1364 | if (wait_event_interruptible(gt->reset.queue,({ int __ret = 0; if (!(!test_bit(0, >->reset.flags) )) __ret = ({ long ret = 0; do { int __error; unsigned long deadline ; ((!cold) ? (void)0 : __assert("diagnostic ", "/usr/src/sys/dev/pci/drm/i915/gt/intel_reset.c" , 1366, "!cold")); mtx_enter(&sch_mtx); deadline = jiffies + ret; __error = msleep(>->reset.queue, &sch_mtx , 0x100, "drmweti", ret); ret = deadline - jiffies; if (__error == -1 || __error == 4) { ret = -4; mtx_leave(&sch_mtx); break ; } if ((0) > 0 && (ret <= 0 || __error == 35)) { mtx_leave(&sch_mtx); ret = ((!test_bit(0, >-> reset.flags))) ? 1 : 0; break; } mtx_leave(&sch_mtx); } while (ret > 0 && !(!test_bit(0, >->reset.flags ))); ret; }); __ret; }) |
| 1365 | !test_bit(I915_RESET_BACKOFF,({ int __ret = 0; if (!(!test_bit(0, >->reset.flags) )) __ret = ({ long ret = 0; do { int __error; unsigned long deadline ; ((!cold) ? (void)0 : __assert("diagnostic ", "/usr/src/sys/dev/pci/drm/i915/gt/intel_reset.c" , 1366, "!cold")); mtx_enter(&sch_mtx); deadline = jiffies + ret; __error = msleep(>->reset.queue, &sch_mtx , 0x100, "drmweti", ret); ret = deadline - jiffies; if (__error == -1 || __error == 4) { ret = -4; mtx_leave(&sch_mtx); break ; } if ((0) > 0 && (ret <= 0 || __error == 35)) { mtx_leave(&sch_mtx); ret = ((!test_bit(0, >-> reset.flags))) ? 1 : 0; break; } mtx_leave(&sch_mtx); } while (ret > 0 && !(!test_bit(0, >->reset.flags ))); ret; }); __ret; }) |
| 1366 | >->reset.flags))({ int __ret = 0; if (!(!test_bit(0, >->reset.flags) )) __ret = ({ long ret = 0; do { int __error; unsigned long deadline ; ((!cold) ? (void)0 : __assert("diagnostic ", "/usr/src/sys/dev/pci/drm/i915/gt/intel_reset.c" , 1366, "!cold")); mtx_enter(&sch_mtx); deadline = jiffies + ret; __error = msleep(>->reset.queue, &sch_mtx , 0x100, "drmweti", ret); ret = deadline - jiffies; if (__error == -1 || __error == 4) { ret = -4; mtx_leave(&sch_mtx); break ; } if ((0) > 0 && (ret <= 0 || __error == 35)) { mtx_leave(&sch_mtx); ret = ((!test_bit(0, >-> reset.flags))) ? 1 : 0; break; } mtx_leave(&sch_mtx); } while (ret > 0 && !(!test_bit(0, >->reset.flags ))); ret; }); __ret; })) |
| 1367 | return -EINTR4; |
| 1368 | |
| 1369 | return intel_gt_is_wedged(gt) ? -EIO5 : 0; |
| 1370 | } |
| 1371 | |
| 1372 | void intel_gt_set_wedged_on_init(struct intel_gt *gt) |
| 1373 | { |
| 1374 | BUILD_BUG_ON(I915_RESET_ENGINE + I915_NUM_ENGINES >extern char _ctassert[(!(2 + I915_NUM_ENGINES > (64 - 3))) ? 1 : -1 ] __attribute__((__unused__)) |
| 1375 | I915_WEDGED_ON_INIT)extern char _ctassert[(!(2 + I915_NUM_ENGINES > (64 - 3))) ? 1 : -1 ] __attribute__((__unused__)); |
| 1376 | intel_gt_set_wedged(gt); |
| 1377 | set_bit(I915_WEDGED_ON_INIT(64 - 3), >->reset.flags); |
| 1378 | |
| 1379 | /* Wedged on init is non-recoverable */ |
| 1380 | add_taint_for_CI(gt->i915, TAINT_WARN1); |
| 1381 | } |
| 1382 | |
| 1383 | void intel_gt_set_wedged_on_fini(struct intel_gt *gt) |
| 1384 | { |
| 1385 | intel_gt_set_wedged(gt); |
| 1386 | set_bit(I915_WEDGED_ON_FINI(64 - 2), >->reset.flags); |
| 1387 | } |
| 1388 | |
| 1389 | void intel_gt_init_reset(struct intel_gt *gt) |
| 1390 | { |
| 1391 | init_waitqueue_head(>->reset.queue); |
| 1392 | rw_init(>->reset.mutex, "gtres")_rw_init_flags(>->reset.mutex, "gtres", 0, ((void *) 0)); |
| 1393 | init_srcu_struct(>->reset.backoff_srcu); |
| 1394 | |
| 1395 | /* no GPU until we are ready! */ |
| 1396 | __set_bit(I915_WEDGED(64 - 1), >->reset.flags); |
| 1397 | } |
| 1398 | |
| 1399 | void intel_gt_fini_reset(struct intel_gt *gt) |
| 1400 | { |
| 1401 | cleanup_srcu_struct(>->reset.backoff_srcu); |
| 1402 | } |
| 1403 | |
| 1404 | static void intel_wedge_me(struct work_struct *work) |
| 1405 | { |
| 1406 | struct intel_wedge_me *w = container_of(work, typeof(*w), work.work)({ const __typeof( ((typeof(*w) *)0)->work.work ) *__mptr = (work); (typeof(*w) *)( (char *)__mptr - __builtin_offsetof( typeof(*w), work.work) );}); |
| 1407 | |
| 1408 | drm_err(&w->gt->i915->drm,printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "%s timed out, cancelling all in-flight rendering.\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , w-> name) |
| 1409 | "%s timed out, cancelling all in-flight rendering.\n",printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "%s timed out, cancelling all in-flight rendering.\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , w-> name) |
| 1410 | w->name)printf("drm:pid%d:%s *ERROR* " "[drm] " "*ERROR* " "%s timed out, cancelling all in-flight rendering.\n" , ({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc->p_p->ps_pid, __func__ , w-> name); |
| 1411 | intel_gt_set_wedged(w->gt); |
| 1412 | } |
| 1413 | |
| 1414 | void __intel_init_wedge(struct intel_wedge_me *w, |
| 1415 | struct intel_gt *gt, |
| 1416 | long timeout, |
| 1417 | const char *name) |
| 1418 | { |
| 1419 | w->gt = gt; |
| 1420 | w->name = name; |
| 1421 | |
| 1422 | INIT_DELAYED_WORK_ONSTACK(&w->work, intel_wedge_me); |
| 1423 | schedule_delayed_work(&w->work, timeout); |
| 1424 | } |
| 1425 | |
| 1426 | void __intel_fini_wedge(struct intel_wedge_me *w) |
| 1427 | { |
| 1428 | cancel_delayed_work_sync(&w->work); |
| 1429 | destroy_delayed_work_on_stack(&w->work); |
| 1430 | w->gt = NULL((void *)0); |
| 1431 | } |
| 1432 | |
| 1433 | #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)0 |
| 1434 | #include "selftest_reset.c" |
| 1435 | #include "selftest_hangcheck.c" |
| 1436 | #endif |