| File: | dev/pci/drm/i915/gt/intel_reset.c |
| Warning: | line 801, column 3 Value stored to 'vma_offset' is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | // SPDX-License-Identifier: MIT |
| 2 | /* |
| 3 | * Copyright © 2008-2018 Intel Corporation |
| 4 | */ |
| 5 | |
| 6 | #include <linux/sched/mm.h> |
| 7 | #include <linux/stop_machine.h> |
| 8 | #include <linux/string_helpers.h> |
| 9 | |
| 10 | #include "display/intel_display.h" |
| 11 | #include "display/intel_overlay.h" |
| 12 | |
| 13 | #include "gem/i915_gem_context.h" |
| 14 | |
| 15 | #include "gt/intel_gt_regs.h" |
| 16 | |
| 17 | #include "i915_drv.h" |
| 18 | #include "i915_file_private.h" |
| 19 | #include "i915_gpu_error.h" |
| 20 | #include "i915_irq.h" |
| 21 | #include "intel_breadcrumbs.h" |
| 22 | #include "intel_engine_pm.h" |
| 23 | #include "intel_engine_regs.h" |
| 24 | #include "intel_gt.h" |
| 25 | #include "intel_gt_pm.h" |
| 26 | #include "intel_gt_requests.h" |
| 27 | #include "intel_mchbar_regs.h" |
| 28 | #include "intel_pci_config.h" |
| 29 | #include "intel_reset.h" |
| 30 | |
| 31 | #include "uc/intel_guc.h" |
| 32 | |
| 33 | #define RESET_MAX_RETRIES3 3 |
| 34 | |
| 35 | /* XXX How to handle concurrent GGTT updates using tiling registers? */ |
| 36 | #define RESET_UNDER_STOP_MACHINE0 0 |
| 37 | |
| 38 | static void rmw_set_fw(struct intel_uncore *uncore, i915_reg_t reg, u32 set) |
| 39 | { |
| 40 | intel_uncore_rmw_fw(uncore, reg, 0, set); |
| 41 | } |
| 42 | |
| 43 | static void rmw_clear_fw(struct intel_uncore *uncore, i915_reg_t reg, u32 clr) |
| 44 | { |
| 45 | intel_uncore_rmw_fw(uncore, reg, clr, 0); |
| 46 | } |
| 47 | |
| 48 | static void client_mark_guilty(struct i915_gem_context *ctx, bool_Bool banned) |
| 49 | { |
| 50 | struct drm_i915_file_private *file_priv = ctx->file_priv; |
| 51 | unsigned long prev_hang; |
| 52 | unsigned int score; |
| 53 | |
| 54 | if (IS_ERR_OR_NULL(file_priv)) |
| 55 | return; |
| 56 | |
| 57 | score = 0; |
| 58 | if (banned) |
| 59 | score = I915_CLIENT_SCORE_CONTEXT_BAN3; |
| 60 | |
| 61 | prev_hang = xchg(&file_priv->hang_timestamp, jiffies)__sync_lock_test_and_set(&file_priv->hang_timestamp, jiffies ); |
| 62 | if (time_before(jiffies, prev_hang + I915_CLIENT_FAST_HANG_JIFFIES)time_after(prev_hang + (60 * hz),jiffies)) |
| 63 | score += I915_CLIENT_SCORE_HANG_FAST1; |
| 64 | |
| 65 | if (score) { |
| 66 | atomic_add(score, &file_priv->ban_score)__sync_fetch_and_add(&file_priv->ban_score, score); |
| 67 | |
| 68 | drm_dbg(&ctx->i915->drm,__drm_dev_dbg(((void *)0), (&ctx->i915->drm) ? (& ctx->i915->drm)->dev : ((void *)0), 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; })) |
| 69 | "client %s: gained %u ban score, now %u\n",__drm_dev_dbg(((void *)0), (&ctx->i915->drm) ? (& ctx->i915->drm)->dev : ((void *)0), 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; })) |
| 70 | ctx->name, score,__drm_dev_dbg(((void *)0), (&ctx->i915->drm) ? (& ctx->i915->drm)->dev : ((void *)0), 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; })) |
| 71 | atomic_read(&file_priv->ban_score))__drm_dev_dbg(((void *)0), (&ctx->i915->drm) ? (& ctx->i915->drm)->dev : ((void *)0), 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; })); |
| 72 | } |
| 73 | } |
| 74 | |
| 75 | static bool_Bool mark_guilty(struct i915_request *rq) |
| 76 | { |
| 77 | struct i915_gem_context *ctx; |
| 78 | unsigned long prev_hang; |
| 79 | bool_Bool banned; |
| 80 | int i; |
| 81 | |
| 82 | if (intel_context_is_closed(rq->context)) |
| 83 | return true1; |
| 84 | |
| 85 | rcu_read_lock(); |
| 86 | ctx = rcu_dereference(rq->context->gem_context)(rq->context->gem_context); |
| 87 | if (ctx && !kref_get_unless_zero(&ctx->ref)) |
| 88 | ctx = NULL((void *)0); |
| 89 | rcu_read_unlock(); |
| 90 | if (!ctx) |
| 91 | return intel_context_is_banned(rq->context); |
| 92 | |
| 93 | atomic_inc(&ctx->guilty_count)__sync_fetch_and_add(&ctx->guilty_count, 1); |
| 94 | |
| 95 | /* Cool contexts are too cool to be banned! (Used for reset testing.) */ |
| 96 | if (!i915_gem_context_is_bannable(ctx)) { |
| 97 | banned = false0; |
| 98 | goto out; |
| 99 | } |
| 100 | |
| 101 | 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) |
| 102 | "%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) |
| 103 | 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); |
| 104 | |
| 105 | /* Record the timestamp for the last N hangs */ |
| 106 | prev_hang = ctx->hang_timestamp[0]; |
| 107 | for (i = 0; i < ARRAY_SIZE(ctx->hang_timestamp)(sizeof((ctx->hang_timestamp)) / sizeof((ctx->hang_timestamp )[0])) - 1; i++) |
| 108 | ctx->hang_timestamp[i] = ctx->hang_timestamp[i + 1]; |
| 109 | ctx->hang_timestamp[i] = jiffies; |
| 110 | |
| 111 | /* If we have hung N+1 times in rapid succession, we ban the context! */ |
| 112 | banned = !i915_gem_context_is_recoverable(ctx); |
| 113 | if (time_before(jiffies, prev_hang + CONTEXT_FAST_HANG_JIFFIES)time_after(prev_hang + (120 * hz),jiffies)) |
| 114 | banned = true1; |
| 115 | if (banned) |
| 116 | drm_dbg(&ctx->i915->drm, "context %s: guilty %d, banned\n",__drm_dev_dbg(((void *)0), (&ctx->i915->drm) ? (& ctx->i915->drm)->dev : ((void *)0), 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 ; })) |
| 117 | ctx->name, atomic_read(&ctx->guilty_count))__drm_dev_dbg(((void *)0), (&ctx->i915->drm) ? (& ctx->i915->drm)->dev : ((void *)0), 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 ; })); |
| 118 | |
| 119 | client_mark_guilty(ctx, banned); |
| 120 | |
| 121 | out: |
| 122 | i915_gem_context_put(ctx); |
| 123 | return banned; |
| 124 | } |
| 125 | |
| 126 | static void mark_innocent(struct i915_request *rq) |
| 127 | { |
| 128 | struct i915_gem_context *ctx; |
| 129 | |
| 130 | rcu_read_lock(); |
| 131 | ctx = rcu_dereference(rq->context->gem_context)(rq->context->gem_context); |
| 132 | if (ctx) |
| 133 | atomic_inc(&ctx->active_count)__sync_fetch_and_add(&ctx->active_count, 1); |
| 134 | rcu_read_unlock(); |
| 135 | } |
| 136 | |
| 137 | void __i915_request_reset(struct i915_request *rq, bool_Bool guilty) |
| 138 | { |
| 139 | bool_Bool banned = false0; |
| 140 | |
| 141 | RQ_TRACE(rq, "guilty? %s\n", str_yes_no(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); |
| 142 | GEM_BUG_ON(__i915_request_is_complete(rq))((void)0); |
| 143 | |
| 144 | rcu_read_lock(); /* protect the GEM context */ |
| 145 | if (guilty) { |
| 146 | i915_request_set_error_once(rq, -EIO5); |
| 147 | __i915_request_skip(rq); |
| 148 | banned = mark_guilty(rq); |
| 149 | } else { |
| 150 | i915_request_set_error_once(rq, -EAGAIN35); |
| 151 | mark_innocent(rq); |
| 152 | } |
| 153 | rcu_read_unlock(); |
| 154 | |
| 155 | if (banned) |
| 156 | intel_context_ban(rq->context, rq); |
| 157 | } |
| 158 | |
| 159 | static bool_Bool i915_in_reset(struct pci_dev *pdev) |
| 160 | { |
| 161 | u8 gdrst; |
| 162 | |
| 163 | pci_read_config_byte(pdev, I915_GDRST0xc0, &gdrst); |
| 164 | return gdrst & GRDOM_RESET_STATUS(1 << 1); |
| 165 | } |
| 166 | |
| 167 | static int i915_do_reset(struct intel_gt *gt, |
| 168 | intel_engine_mask_t engine_mask, |
| 169 | unsigned int retry) |
| 170 | { |
| 171 | struct pci_dev *pdev = gt->i915->drm.pdev; |
| 172 | int err; |
| 173 | |
| 174 | /* Assert reset for at least 20 usec, and wait for acknowledgement. */ |
| 175 | pci_write_config_byte(pdev, I915_GDRST0xc0, GRDOM_RESET_ENABLE(1 << 0)); |
| 176 | udelay(50); |
| 177 | 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)) { preempt_disable(); 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)) preempt_enable(); __asm volatile("" : : : "memory"); if (((i915_in_reset(pdev)))) { ret = 0; break ; } if (now - base >= timeout) { ret = -60; break; } cpu_relax (); if (!(1)) { preempt_disable(); 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; } ); }); |
| 178 | |
| 179 | /* Clear the reset request. */ |
| 180 | pci_write_config_byte(pdev, I915_GDRST0xc0, 0); |
| 181 | udelay(50); |
| 182 | if (!err) |
| 183 | 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)) { preempt_disable(); 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)) preempt_enable(); __asm volatile("" : : : "memory"); if (((!i915_in_reset(pdev)))) { ret = 0; break ; } if (now - base >= timeout) { ret = -60; break; } cpu_relax (); if (!(1)) { preempt_disable(); 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; } ); }); |
| 184 | |
| 185 | return err; |
| 186 | } |
| 187 | |
| 188 | static bool_Bool g4x_reset_complete(struct pci_dev *pdev) |
| 189 | { |
| 190 | u8 gdrst; |
| 191 | |
| 192 | pci_read_config_byte(pdev, I915_GDRST0xc0, &gdrst); |
| 193 | return (gdrst & GRDOM_RESET_ENABLE(1 << 0)) == 0; |
| 194 | } |
| 195 | |
| 196 | static int g33_do_reset(struct intel_gt *gt, |
| 197 | intel_engine_mask_t engine_mask, |
| 198 | unsigned int retry) |
| 199 | { |
| 200 | struct pci_dev *pdev = gt->i915->drm.pdev; |
| 201 | |
| 202 | pci_write_config_byte(pdev, I915_GDRST0xc0, GRDOM_RESET_ENABLE(1 << 0)); |
| 203 | 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)) { preempt_disable(); 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)) preempt_enable(); __asm volatile("" : : : "memory"); if (((g4x_reset_complete(pdev)))) { ret = 0 ; break; } if (now - base >= timeout) { ret = -60; break; } cpu_relax(); if (!(1)) { preempt_disable(); 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; } ); }); |
| 204 | } |
| 205 | |
| 206 | static int g4x_do_reset(struct intel_gt *gt, |
| 207 | intel_engine_mask_t engine_mask, |
| 208 | unsigned int retry) |
| 209 | { |
| 210 | struct pci_dev *pdev = gt->i915->drm.pdev; |
| 211 | struct intel_uncore *uncore = gt->uncore; |
| 212 | int ret; |
| 213 | |
| 214 | /* WaVcpClkGateDisableForMediaReset:ctg,elk */ |
| 215 | rmw_set_fw(uncore, VDECCLK_GATE_D((const i915_reg_t){ .reg = (0x620C) }), VCP_UNIT_CLOCK_GATE_DISABLE(1 << 4)); |
| 216 | intel_uncore_posting_read_fw(uncore, VDECCLK_GATE_D)((void)__raw_uncore_read32(uncore, ((const i915_reg_t){ .reg = (0x620C) }))); |
| 217 | |
| 218 | pci_write_config_byte(pdev, I915_GDRST0xc0, |
| 219 | GRDOM_MEDIA(3 << 2) | GRDOM_RESET_ENABLE(1 << 0)); |
| 220 | 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)) { preempt_disable(); 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)) preempt_enable(); __asm volatile("" : : : "memory"); if (((g4x_reset_complete(pdev)))) { ret = 0 ; break; } if (now - base >= timeout) { ret = -60; break; } cpu_relax(); if (!(1)) { preempt_disable(); 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; } ); }); |
| 221 | if (ret) { |
| 222 | GT_TRACE(gt, "Wait for media reset failed\n")do { const struct intel_gt *gt__ __attribute__((__unused__)) = (gt); do { } while (0); } while (0); |
| 223 | goto out; |
| 224 | } |
| 225 | |
| 226 | pci_write_config_byte(pdev, I915_GDRST0xc0, |
| 227 | GRDOM_RENDER(1 << 2) | GRDOM_RESET_ENABLE(1 << 0)); |
| 228 | 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)) { preempt_disable(); 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)) preempt_enable(); __asm volatile("" : : : "memory"); if (((g4x_reset_complete(pdev)))) { ret = 0 ; break; } if (now - base >= timeout) { ret = -60; break; } cpu_relax(); if (!(1)) { preempt_disable(); 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; } ); }); |
| 229 | if (ret) { |
| 230 | GT_TRACE(gt, "Wait for render reset failed\n")do { const struct intel_gt *gt__ __attribute__((__unused__)) = (gt); do { } while (0); } while (0); |
| 231 | goto out; |
| 232 | } |
| 233 | |
| 234 | out: |
| 235 | pci_write_config_byte(pdev, I915_GDRST0xc0, 0); |
| 236 | |
| 237 | rmw_clear_fw(uncore, VDECCLK_GATE_D((const i915_reg_t){ .reg = (0x620C) }), VCP_UNIT_CLOCK_GATE_DISABLE(1 << 4)); |
| 238 | intel_uncore_posting_read_fw(uncore, VDECCLK_GATE_D)((void)__raw_uncore_read32(uncore, ((const i915_reg_t){ .reg = (0x620C) }))); |
| 239 | |
| 240 | return ret; |
| 241 | } |
| 242 | |
| 243 | static int ilk_do_reset(struct intel_gt *gt, intel_engine_mask_t engine_mask, |
| 244 | unsigned int retry) |
| 245 | { |
| 246 | struct intel_uncore *uncore = gt->uncore; |
| 247 | int ret; |
| 248 | |
| 249 | intel_uncore_write_fw(uncore, ILK_GDSR,__raw_uncore_write32(uncore, ((const i915_reg_t){ .reg = (0x10000 + 0x2ca4) }), (1 << 1) | (1 << 0)) |
| 250 | ILK_GRDOM_RENDER | ILK_GRDOM_RESET_ENABLE)__raw_uncore_write32(uncore, ((const i915_reg_t){ .reg = (0x10000 + 0x2ca4) }), (1 << 1) | (1 << 0)); |
| 251 | ret = __intel_wait_for_register_fw(uncore, ILK_GDSR((const i915_reg_t){ .reg = (0x10000 + 0x2ca4) }), |
| 252 | ILK_GRDOM_RESET_ENABLE(1 << 0), 0, |
| 253 | 5000, 0, |
| 254 | NULL((void *)0)); |
| 255 | if (ret) { |
| 256 | GT_TRACE(gt, "Wait for render reset failed\n")do { const struct intel_gt *gt__ __attribute__((__unused__)) = (gt); do { } while (0); } while (0); |
| 257 | goto out; |
| 258 | } |
| 259 | |
| 260 | intel_uncore_write_fw(uncore, ILK_GDSR,__raw_uncore_write32(uncore, ((const i915_reg_t){ .reg = (0x10000 + 0x2ca4) }), (3 << 1) | (1 << 0)) |
| 261 | ILK_GRDOM_MEDIA | ILK_GRDOM_RESET_ENABLE)__raw_uncore_write32(uncore, ((const i915_reg_t){ .reg = (0x10000 + 0x2ca4) }), (3 << 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 | GT_TRACE(gt, "Wait for media reset failed\n")do { const struct intel_gt *gt__ __attribute__((__unused__)) = (gt); do { } while (0); } while (0); |
| 268 | goto out; |
| 269 | } |
| 270 | |
| 271 | out: |
| 272 | intel_uncore_write_fw(uncore, ILK_GDSR, 0)__raw_uncore_write32(uncore, ((const i915_reg_t){ .reg = (0x10000 + 0x2ca4) }), 0); |
| 273 | intel_uncore_posting_read_fw(uncore, ILK_GDSR)((void)__raw_uncore_read32(uncore, ((const i915_reg_t){ .reg = (0x10000 + 0x2ca4) }))); |
| 274 | return ret; |
| 275 | } |
| 276 | |
| 277 | /* Reset the hardware domains (GENX_GRDOM_*) specified by mask */ |
| 278 | static int gen6_hw_domain_reset(struct intel_gt *gt, u32 hw_domain_mask) |
| 279 | { |
| 280 | struct intel_uncore *uncore = gt->uncore; |
| 281 | int loops = 2; |
| 282 | int err; |
| 283 | |
| 284 | /* |
| 285 | * GEN6_GDRST is not in the gt power well, no need to check |
| 286 | * for fifo space for the write or forcewake the chip for |
| 287 | * the read |
| 288 | */ |
| 289 | do { |
| 290 | intel_uncore_write_fw(uncore, GEN6_GDRST, hw_domain_mask)__raw_uncore_write32(uncore, ((const i915_reg_t){ .reg = (0x941c ) }), hw_domain_mask); |
| 291 | |
| 292 | /* |
| 293 | * Wait for the device to ack the reset requests. |
| 294 | * |
| 295 | * On some platforms, e.g. Jasperlake, we see that the |
| 296 | * engine register state is not cleared until shortly after |
| 297 | * GDRST reports completion, causing a failure as we try |
| 298 | * to immediately resume while the internal state is still |
| 299 | * in flux. If we immediately repeat the reset, the second |
| 300 | * reset appears to serialise with the first, and since |
| 301 | * it is a no-op, the registers should retain their reset |
| 302 | * value. However, there is still a concern that upon |
| 303 | * leaving the second reset, the internal engine state |
| 304 | * is still in flux and not ready for resuming. |
| 305 | */ |
| 306 | err = __intel_wait_for_register_fw(uncore, GEN6_GDRST((const i915_reg_t){ .reg = (0x941c) }), |
| 307 | hw_domain_mask, 0, |
| 308 | 2000, 0, |
| 309 | NULL((void *)0)); |
| 310 | } while (err == 0 && --loops); |
| 311 | if (err) |
| 312 | GT_TRACE(gt,do { const struct intel_gt *gt__ __attribute__((__unused__)) = (gt); do { } while (0); } while (0) |
| 313 | "Wait for 0x%08x engines reset failed\n",do { const struct intel_gt *gt__ __attribute__((__unused__)) = (gt); do { } while (0); } while (0) |
| 314 | hw_domain_mask)do { const struct intel_gt *gt__ __attribute__((__unused__)) = (gt); do { } while (0); } while (0); |
| 315 | |
| 316 | /* |
| 317 | * As we have observed that the engine state is still volatile |
| 318 | * after GDRST is acked, impose a small delay to let everything settle. |
| 319 | */ |
| 320 | udelay(50); |
| 321 | |
| 322 | return err; |
| 323 | } |
| 324 | |
| 325 | static int __gen6_reset_engines(struct intel_gt *gt, |
| 326 | intel_engine_mask_t engine_mask, |
| 327 | unsigned int retry) |
| 328 | { |
| 329 | struct intel_engine_cs *engine; |
| 330 | u32 hw_mask; |
| 331 | |
| 332 | if (engine_mask == ALL_ENGINES((intel_engine_mask_t)~0ul)) { |
| 333 | hw_mask = GEN6_GRDOM_FULL(1 << 0); |
| 334 | } else { |
| 335 | intel_engine_mask_t tmp; |
| 336 | |
| 337 | hw_mask = 0; |
| 338 | 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;) { |
| 339 | hw_mask |= engine->reset_domain; |
| 340 | } |
| 341 | } |
| 342 | |
| 343 | return gen6_hw_domain_reset(gt, hw_mask); |
| 344 | } |
| 345 | |
| 346 | static int gen6_reset_engines(struct intel_gt *gt, |
| 347 | intel_engine_mask_t engine_mask, |
| 348 | unsigned int retry) |
| 349 | { |
| 350 | unsigned long flags; |
| 351 | int ret; |
| 352 | |
| 353 | spin_lock_irqsave(>->uncore->lock, flags)do { flags = 0; mtx_enter(>->uncore->lock); } while (0); |
| 354 | ret = __gen6_reset_engines(gt, engine_mask, retry); |
| 355 | spin_unlock_irqrestore(>->uncore->lock, flags)do { (void)(flags); mtx_leave(>->uncore->lock); } while (0); |
| 356 | |
| 357 | return ret; |
| 358 | } |
| 359 | |
| 360 | static struct intel_engine_cs *find_sfc_paired_vecs_engine(struct intel_engine_cs *engine) |
| 361 | { |
| 362 | int vecs_id; |
| 363 | |
| 364 | GEM_BUG_ON(engine->class != VIDEO_DECODE_CLASS)((void)0); |
| 365 | |
| 366 | vecs_id = _VECS((engine->instance) / 2)(VECS0 + ((engine->instance) / 2)); |
| 367 | |
| 368 | return engine->gt->engine[vecs_id]; |
| 369 | } |
| 370 | |
| 371 | struct sfc_lock_data { |
| 372 | i915_reg_t lock_reg; |
| 373 | i915_reg_t ack_reg; |
| 374 | i915_reg_t usage_reg; |
| 375 | u32 lock_bit; |
| 376 | u32 ack_bit; |
| 377 | u32 usage_bit; |
| 378 | u32 reset_bit; |
| 379 | }; |
| 380 | |
| 381 | static void get_sfc_forced_lock_data(struct intel_engine_cs *engine, |
| 382 | struct sfc_lock_data *sfc_lock) |
| 383 | { |
| 384 | switch (engine->class) { |
| 385 | default: |
| 386 | MISSING_CASE(engine->class)({ int __ret = !!(1); if (__ret) printf("Missing case (%s == %ld)\n" , "engine->class", (long)(engine->class)); __builtin_expect (!!(__ret), 0); }); |
| 387 | fallthroughdo {} while (0); |
| 388 | case VIDEO_DECODE_CLASS1: |
| 389 | sfc_lock->lock_reg = GEN11_VCS_SFC_FORCED_LOCK(engine->mmio_base)((const i915_reg_t){ .reg = ((engine->mmio_base) + 0x88c) } ); |
| 390 | sfc_lock->lock_bit = GEN11_VCS_SFC_FORCED_LOCK_BIT(1 << 0); |
| 391 | |
| 392 | sfc_lock->ack_reg = GEN11_VCS_SFC_LOCK_STATUS(engine->mmio_base)((const i915_reg_t){ .reg = ((engine->mmio_base) + 0x890) } ); |
| 393 | sfc_lock->ack_bit = GEN11_VCS_SFC_LOCK_ACK_BIT(1 << 1); |
| 394 | |
| 395 | sfc_lock->usage_reg = GEN11_VCS_SFC_LOCK_STATUS(engine->mmio_base)((const i915_reg_t){ .reg = ((engine->mmio_base) + 0x890) } ); |
| 396 | sfc_lock->usage_bit = GEN11_VCS_SFC_USAGE_BIT(1 << 0); |
| 397 | sfc_lock->reset_bit = GEN11_VCS_SFC_RESET_BIT(engine->instance)(((u32)((1UL << (17)) + 0)) << ((engine->instance ) >> 1)); |
| 398 | |
| 399 | break; |
| 400 | case VIDEO_ENHANCEMENT_CLASS2: |
| 401 | sfc_lock->lock_reg = GEN11_VECS_SFC_FORCED_LOCK(engine->mmio_base)((const i915_reg_t){ .reg = ((engine->mmio_base) + 0x201c) }); |
| 402 | sfc_lock->lock_bit = GEN11_VECS_SFC_FORCED_LOCK_BIT(1 << 0); |
| 403 | |
| 404 | sfc_lock->ack_reg = GEN11_VECS_SFC_LOCK_ACK(engine->mmio_base)((const i915_reg_t){ .reg = ((engine->mmio_base) + 0x2018) }); |
| 405 | sfc_lock->ack_bit = GEN11_VECS_SFC_LOCK_ACK_BIT(1 << 0); |
| 406 | |
| 407 | sfc_lock->usage_reg = GEN11_VECS_SFC_USAGE(engine->mmio_base)((const i915_reg_t){ .reg = ((engine->mmio_base) + 0x2014) }); |
| 408 | sfc_lock->usage_bit = GEN11_VECS_SFC_USAGE_BIT(1 << 0); |
| 409 | sfc_lock->reset_bit = GEN11_VECS_SFC_RESET_BIT(engine->instance)(((u32)((1UL << (17)) + 0)) << (engine->instance )); |
| 410 | |
| 411 | break; |
| 412 | } |
| 413 | } |
| 414 | |
| 415 | static int gen11_lock_sfc(struct intel_engine_cs *engine, |
| 416 | u32 *reset_mask, |
| 417 | u32 *unlock_mask) |
| 418 | { |
| 419 | struct intel_uncore *uncore = engine->uncore; |
| 420 | u8 vdbox_sfc_access = engine->gt->info.vdbox_sfc_access; |
| 421 | struct sfc_lock_data sfc_lock; |
| 422 | bool_Bool lock_obtained, lock_to_other = false0; |
| 423 | int ret; |
| 424 | |
| 425 | switch (engine->class) { |
| 426 | case VIDEO_DECODE_CLASS1: |
| 427 | if ((BIT(engine->instance)(1UL << (engine->instance)) & vdbox_sfc_access) == 0) |
| 428 | return 0; |
| 429 | |
| 430 | fallthroughdo {} while (0); |
| 431 | case VIDEO_ENHANCEMENT_CLASS2: |
| 432 | get_sfc_forced_lock_data(engine, &sfc_lock); |
| 433 | |
| 434 | break; |
| 435 | default: |
| 436 | return 0; |
| 437 | } |
| 438 | |
| 439 | if (!(intel_uncore_read_fw(uncore, sfc_lock.usage_reg)__raw_uncore_read32(uncore, sfc_lock.usage_reg) & sfc_lock.usage_bit)) { |
| 440 | struct intel_engine_cs *paired_vecs; |
| 441 | |
| 442 | if (engine->class != VIDEO_DECODE_CLASS1 || |
| 443 | GRAPHICS_VER(engine->i915)((&(engine->i915)->__runtime)->graphics.ip.ver) != 12) |
| 444 | return 0; |
| 445 | |
| 446 | /* |
| 447 | * Wa_14010733141 |
| 448 | * |
| 449 | * If the VCS-MFX isn't using the SFC, we also need to check |
| 450 | * whether VCS-HCP is using it. If so, we need to issue a *VE* |
| 451 | * forced lock on the VE engine that shares the same SFC. |
| 452 | */ |
| 453 | if (!(intel_uncore_read_fw(uncore,__raw_uncore_read32(uncore, ((const i915_reg_t){ .reg = ((engine ->mmio_base) + 0x2914) })) |
| 454 | GEN12_HCP_SFC_LOCK_STATUS(engine->mmio_base))__raw_uncore_read32(uncore, ((const i915_reg_t){ .reg = ((engine ->mmio_base) + 0x2914) })) & |
| 455 | GEN12_HCP_SFC_USAGE_BIT((u32)((1UL << (0)) + 0)))) |
| 456 | return 0; |
| 457 | |
| 458 | paired_vecs = find_sfc_paired_vecs_engine(engine); |
| 459 | get_sfc_forced_lock_data(paired_vecs, &sfc_lock); |
| 460 | lock_to_other = true1; |
| 461 | *unlock_mask |= paired_vecs->mask; |
| 462 | } else { |
| 463 | *unlock_mask |= engine->mask; |
| 464 | } |
| 465 | |
| 466 | /* |
| 467 | * If the engine is using an SFC, tell the engine that a software reset |
| 468 | * is going to happen. The engine will then try to force lock the SFC. |
| 469 | * If SFC ends up being locked to the engine we want to reset, we have |
| 470 | * to reset it as well (we will unlock it once the reset sequence is |
| 471 | * completed). |
| 472 | */ |
| 473 | rmw_set_fw(uncore, sfc_lock.lock_reg, sfc_lock.lock_bit); |
| 474 | |
| 475 | ret = __intel_wait_for_register_fw(uncore, |
| 476 | sfc_lock.ack_reg, |
| 477 | sfc_lock.ack_bit, |
| 478 | sfc_lock.ack_bit, |
| 479 | 1000, 0, NULL((void *)0)); |
| 480 | |
| 481 | /* |
| 482 | * Was the SFC released while we were trying to lock it? |
| 483 | * |
| 484 | * We should reset both the engine and the SFC if: |
| 485 | * - We were locking the SFC to this engine and the lock succeeded |
| 486 | * OR |
| 487 | * - We were locking the SFC to a different engine (Wa_14010733141) |
| 488 | * but the SFC was released before the lock was obtained. |
| 489 | * |
| 490 | * Otherwise we need only reset the engine by itself and we can |
| 491 | * leave the SFC alone. |
| 492 | */ |
| 493 | lock_obtained = (intel_uncore_read_fw(uncore, sfc_lock.usage_reg)__raw_uncore_read32(uncore, sfc_lock.usage_reg) & |
| 494 | sfc_lock.usage_bit) != 0; |
| 495 | if (lock_obtained == lock_to_other) |
| 496 | return 0; |
| 497 | |
| 498 | if (ret) { |
| 499 | ENGINE_TRACE(engine, "Wait for SFC forced lock ack failed\n")do { const struct intel_engine_cs *e__ __attribute__((__unused__ )) = (engine); do { } while (0); } while (0); |
| 500 | return ret; |
| 501 | } |
| 502 | |
| 503 | *reset_mask |= sfc_lock.reset_bit; |
| 504 | return 0; |
| 505 | } |
| 506 | |
| 507 | static void gen11_unlock_sfc(struct intel_engine_cs *engine) |
| 508 | { |
| 509 | struct intel_uncore *uncore = engine->uncore; |
| 510 | u8 vdbox_sfc_access = engine->gt->info.vdbox_sfc_access; |
| 511 | struct sfc_lock_data sfc_lock = {}; |
| 512 | |
| 513 | if (engine->class != VIDEO_DECODE_CLASS1 && |
| 514 | engine->class != VIDEO_ENHANCEMENT_CLASS2) |
| 515 | return; |
| 516 | |
| 517 | if (engine->class == VIDEO_DECODE_CLASS1 && |
| 518 | (BIT(engine->instance)(1UL << (engine->instance)) & vdbox_sfc_access) == 0) |
| 519 | return; |
| 520 | |
| 521 | get_sfc_forced_lock_data(engine, &sfc_lock); |
| 522 | |
| 523 | rmw_clear_fw(uncore, sfc_lock.lock_reg, sfc_lock.lock_bit); |
| 524 | } |
| 525 | |
| 526 | static int __gen11_reset_engines(struct intel_gt *gt, |
| 527 | intel_engine_mask_t engine_mask, |
| 528 | unsigned int retry) |
| 529 | { |
| 530 | struct intel_engine_cs *engine; |
| 531 | intel_engine_mask_t tmp; |
| 532 | u32 reset_mask, unlock_mask = 0; |
| 533 | int ret; |
| 534 | |
| 535 | if (engine_mask == ALL_ENGINES((intel_engine_mask_t)~0ul)) { |
| 536 | reset_mask = GEN11_GRDOM_FULL(1 << 0); |
| 537 | } else { |
| 538 | reset_mask = 0; |
| 539 | 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;) { |
| 540 | reset_mask |= engine->reset_domain; |
| 541 | ret = gen11_lock_sfc(engine, &reset_mask, &unlock_mask); |
| 542 | if (ret) |
| 543 | goto sfc_unlock; |
| 544 | } |
| 545 | } |
| 546 | |
| 547 | ret = gen6_hw_domain_reset(gt, reset_mask); |
| 548 | |
| 549 | sfc_unlock: |
| 550 | /* |
| 551 | * We unlock the SFC based on the lock status and not the result of |
| 552 | * gen11_lock_sfc to make sure that we clean properly if something |
| 553 | * wrong happened during the lock (e.g. lock acquired after timeout |
| 554 | * expiration). |
| 555 | * |
| 556 | * Due to Wa_14010733141, we may have locked an SFC to an engine that |
| 557 | * wasn't being reset. So instead of calling gen11_unlock_sfc() |
| 558 | * on engine_mask, we instead call it on the mask of engines that our |
| 559 | * gen11_lock_sfc() calls told us actually had locks attempted. |
| 560 | */ |
| 561 | for_each_engine_masked(engine, gt, unlock_mask, tmp)for ((tmp) = (unlock_mask) & (gt)->info.engine_mask; ( tmp) ? ((engine) = (gt)->engine[({ int __idx = ffs(tmp) - 1 ; tmp &= ~(1UL << (__idx)); __idx; })]), 1 : 0;) |
| 562 | gen11_unlock_sfc(engine); |
| 563 | |
| 564 | return ret; |
| 565 | } |
| 566 | |
| 567 | static int gen8_engine_reset_prepare(struct intel_engine_cs *engine) |
| 568 | { |
| 569 | struct intel_uncore *uncore = engine->uncore; |
| 570 | const i915_reg_t reg = RING_RESET_CTL(engine->mmio_base)((const i915_reg_t){ .reg = ((engine->mmio_base) + 0xd0) } ); |
| 571 | u32 request, mask, ack; |
| 572 | int ret; |
| 573 | |
| 574 | if (I915_SELFTEST_ONLY(should_fail(&engine->reset_timeout, 1))0) |
| 575 | return -ETIMEDOUT60; |
| 576 | |
| 577 | ack = intel_uncore_read_fw(uncore, reg)__raw_uncore_read32(uncore, reg); |
| 578 | if (ack & RESET_CTL_CAT_ERROR((u32)((1UL << (2)) + 0))) { |
| 579 | /* |
| 580 | * For catastrophic errors, ready-for-reset sequence |
| 581 | * needs to be bypassed: HAS#396813 |
| 582 | */ |
| 583 | request = RESET_CTL_CAT_ERROR((u32)((1UL << (2)) + 0)); |
| 584 | mask = RESET_CTL_CAT_ERROR((u32)((1UL << (2)) + 0)); |
| 585 | |
| 586 | /* Catastrophic errors need to be cleared by HW */ |
| 587 | ack = 0; |
| 588 | } else if (!(ack & RESET_CTL_READY_TO_RESET((u32)((1UL << (1)) + 0)))) { |
| 589 | request = RESET_CTL_REQUEST_RESET((u32)((1UL << (0)) + 0)); |
| 590 | mask = RESET_CTL_READY_TO_RESET((u32)((1UL << (1)) + 0)); |
| 591 | ack = RESET_CTL_READY_TO_RESET((u32)((1UL << (1)) + 0)); |
| 592 | } else { |
| 593 | return 0; |
| 594 | } |
| 595 | |
| 596 | 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)); }); })); |
| 597 | ret = __intel_wait_for_register_fw(uncore, reg, mask, ack, |
| 598 | 700, 0, NULL((void *)0)); |
| 599 | if (ret) |
| 600 | 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)) |
| 601 | "%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)) |
| 602 | 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)) |
| 603 | 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)); |
| 604 | |
| 605 | return ret; |
| 606 | } |
| 607 | |
| 608 | static void gen8_engine_reset_cancel(struct intel_engine_cs *engine) |
| 609 | { |
| 610 | 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)) ; }))) |
| 611 | 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)) ; }))) |
| 612 | _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)) ; }))); |
| 613 | } |
| 614 | |
| 615 | static int gen8_reset_engines(struct intel_gt *gt, |
| 616 | intel_engine_mask_t engine_mask, |
| 617 | unsigned int retry) |
| 618 | { |
| 619 | struct intel_engine_cs *engine; |
| 620 | const bool_Bool reset_non_ready = retry >= 1; |
| 621 | intel_engine_mask_t tmp; |
| 622 | unsigned long flags; |
| 623 | int ret; |
| 624 | |
| 625 | spin_lock_irqsave(>->uncore->lock, flags)do { flags = 0; mtx_enter(>->uncore->lock); } while (0); |
| 626 | |
| 627 | 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;) { |
| 628 | ret = gen8_engine_reset_prepare(engine); |
| 629 | if (ret && !reset_non_ready) |
| 630 | goto skip_reset; |
| 631 | |
| 632 | /* |
| 633 | * If this is not the first failed attempt to prepare, |
| 634 | * we decide to proceed anyway. |
| 635 | * |
| 636 | * By doing so we risk context corruption and with |
| 637 | * some gens (kbl), possible system hang if reset |
| 638 | * happens during active bb execution. |
| 639 | * |
| 640 | * We rather take context corruption instead of |
| 641 | * failed reset with a wedged driver/gpu. And |
| 642 | * active bb execution case should be covered by |
| 643 | * stop_engines() we have before the reset. |
| 644 | */ |
| 645 | } |
| 646 | |
| 647 | /* |
| 648 | * Wa_22011100796:dg2, whenever Full soft reset is required, |
| 649 | * reset all individual engines firstly, and then do a full soft reset. |
| 650 | * |
| 651 | * This is best effort, so ignore any error from the initial reset. |
| 652 | */ |
| 653 | if (IS_DG2(gt->i915)IS_PLATFORM(gt->i915, INTEL_DG2) && engine_mask == ALL_ENGINES((intel_engine_mask_t)~0ul)) |
| 654 | __gen11_reset_engines(gt, gt->info.engine_mask, 0); |
| 655 | |
| 656 | if (GRAPHICS_VER(gt->i915)((&(gt->i915)->__runtime)->graphics.ip.ver) >= 11) |
| 657 | ret = __gen11_reset_engines(gt, engine_mask, retry); |
| 658 | else |
| 659 | ret = __gen6_reset_engines(gt, engine_mask, retry); |
| 660 | |
| 661 | skip_reset: |
| 662 | 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;) |
| 663 | gen8_engine_reset_cancel(engine); |
| 664 | |
| 665 | spin_unlock_irqrestore(>->uncore->lock, flags)do { (void)(flags); mtx_leave(>->uncore->lock); } while (0); |
| 666 | |
| 667 | return ret; |
| 668 | } |
| 669 | |
| 670 | static int mock_reset(struct intel_gt *gt, |
| 671 | intel_engine_mask_t mask, |
| 672 | unsigned int retry) |
| 673 | { |
| 674 | return 0; |
| 675 | } |
| 676 | |
| 677 | typedef int (*reset_func)(struct intel_gt *, |
| 678 | intel_engine_mask_t engine_mask, |
| 679 | unsigned int retry); |
| 680 | |
| 681 | static reset_func intel_get_gpu_reset(const struct intel_gt *gt) |
| 682 | { |
| 683 | struct drm_i915_privateinteldrm_softc *i915 = gt->i915; |
| 684 | |
| 685 | if (is_mock_gt(gt)) |
| 686 | return mock_reset; |
| 687 | else if (GRAPHICS_VER(i915)((&(i915)->__runtime)->graphics.ip.ver) >= 8) |
| 688 | return gen8_reset_engines; |
| 689 | else if (GRAPHICS_VER(i915)((&(i915)->__runtime)->graphics.ip.ver) >= 6) |
| 690 | return gen6_reset_engines; |
| 691 | else if (GRAPHICS_VER(i915)((&(i915)->__runtime)->graphics.ip.ver) >= 5) |
| 692 | return ilk_do_reset; |
| 693 | else if (IS_G4X(i915)(IS_PLATFORM(i915, INTEL_G45) || IS_PLATFORM(i915, INTEL_GM45 ))) |
| 694 | return g4x_do_reset; |
| 695 | else if (IS_G33(i915)IS_PLATFORM(i915, INTEL_G33) || IS_PINEVIEW(i915)IS_PLATFORM(i915, INTEL_PINEVIEW)) |
| 696 | return g33_do_reset; |
| 697 | else if (GRAPHICS_VER(i915)((&(i915)->__runtime)->graphics.ip.ver) >= 3) |
| 698 | return i915_do_reset; |
| 699 | else |
| 700 | return NULL((void *)0); |
| 701 | } |
| 702 | |
| 703 | int __intel_gt_reset(struct intel_gt *gt, intel_engine_mask_t engine_mask) |
| 704 | { |
| 705 | const int retries = engine_mask == ALL_ENGINES((intel_engine_mask_t)~0ul) ? RESET_MAX_RETRIES3 : 1; |
| 706 | reset_func reset; |
| 707 | int ret = -ETIMEDOUT60; |
| 708 | int retry; |
| 709 | |
| 710 | reset = intel_get_gpu_reset(gt); |
| 711 | if (!reset) |
| 712 | return -ENODEV19; |
| 713 | |
| 714 | /* |
| 715 | * If the power well sleeps during the reset, the reset |
| 716 | * request may be dropped and never completes (causing -EIO). |
| 717 | */ |
| 718 | intel_uncore_forcewake_get(gt->uncore, FORCEWAKE_ALL); |
| 719 | for (retry = 0; ret == -ETIMEDOUT60 && retry < retries; retry++) { |
| 720 | GT_TRACE(gt, "engine_mask=%x\n", engine_mask)do { const struct intel_gt *gt__ __attribute__((__unused__)) = (gt); do { } while (0); } while (0); |
| 721 | preempt_disable(); |
| 722 | ret = reset(gt, engine_mask, retry); |
| 723 | preempt_enable(); |
| 724 | } |
| 725 | intel_uncore_forcewake_put(gt->uncore, FORCEWAKE_ALL); |
| 726 | |
| 727 | return ret; |
| 728 | } |
| 729 | |
| 730 | bool_Bool intel_has_gpu_reset(const struct intel_gt *gt) |
| 731 | { |
| 732 | if (!gt->i915->params.reset) |
| 733 | return NULL((void *)0); |
| 734 | |
| 735 | return intel_get_gpu_reset(gt); |
| 736 | } |
| 737 | |
| 738 | bool_Bool intel_has_reset_engine(const struct intel_gt *gt) |
| 739 | { |
| 740 | if (gt->i915->params.reset < 2) |
| 741 | return false0; |
| 742 | |
| 743 | return INTEL_INFO(gt->i915)(&(gt->i915)->__info)->has_reset_engine; |
| 744 | } |
| 745 | |
| 746 | int intel_reset_guc(struct intel_gt *gt) |
| 747 | { |
| 748 | u32 guc_domain = |
| 749 | GRAPHICS_VER(gt->i915)((&(gt->i915)->__runtime)->graphics.ip.ver) >= 11 ? GEN11_GRDOM_GUC((u32)((1UL << (3)) + 0)) : GEN9_GRDOM_GUC(1 << 5); |
| 750 | int ret; |
| 751 | |
| 752 | GEM_BUG_ON(!HAS_GT_UC(gt->i915))((void)0); |
| 753 | |
| 754 | intel_uncore_forcewake_get(gt->uncore, FORCEWAKE_ALL); |
| 755 | ret = gen6_hw_domain_reset(gt, guc_domain); |
| 756 | intel_uncore_forcewake_put(gt->uncore, FORCEWAKE_ALL); |
| 757 | |
| 758 | return ret; |
| 759 | } |
| 760 | |
| 761 | /* |
| 762 | * Ensure irq handler finishes, and not run again. |
| 763 | * Also return the active request so that we only search for it once. |
| 764 | */ |
| 765 | static void reset_prepare_engine(struct intel_engine_cs *engine) |
| 766 | { |
| 767 | /* |
| 768 | * During the reset sequence, we must prevent the engine from |
| 769 | * entering RC6. As the context state is undefined until we restart |
| 770 | * the engine, if it does enter RC6 during the reset, the state |
| 771 | * written to the powercontext is undefined and so we may lose |
| 772 | * GPU state upon resume, i.e. fail to restart after a reset. |
| 773 | */ |
| 774 | intel_uncore_forcewake_get(engine->uncore, FORCEWAKE_ALL); |
| 775 | if (engine->reset.prepare) |
| 776 | engine->reset.prepare(engine); |
| 777 | } |
| 778 | |
| 779 | static void revoke_mmaps(struct intel_gt *gt) |
| 780 | { |
| 781 | int i; |
| 782 | |
| 783 | for (i = 0; i < gt->ggtt->num_fences; i++) { |
| 784 | struct drm_vma_offset_node *node; |
| 785 | struct i915_vma *vma; |
| 786 | u64 vma_offset; |
| 787 | |
| 788 | 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; }); |
| 789 | if (!vma) |
| 790 | continue; |
| 791 | |
| 792 | if (!i915_vma_has_userfault(vma)) |
| 793 | continue; |
| 794 | |
| 795 | GEM_BUG_ON(vma->fence != >->ggtt->fence_regs[i])((void)0); |
| 796 | |
| 797 | if (!vma->mmo) |
| 798 | continue; |
| 799 | |
| 800 | node = &vma->mmo->vma_node; |
| 801 | vma_offset = vma->gtt_view.partial.offset << PAGE_SHIFT12; |
Value stored to 'vma_offset' is never read | |
| 802 | |
| 803 | #ifdef __linux__ |
| 804 | unmap_mapping_range(gt->i915->drm.anon_inode->i_mapping, |
| 805 | drm_vma_node_offset_addr(node) + vma_offset, |
| 806 | vma->size, |
| 807 | 1); |
| 808 | #else |
| 809 | { |
| 810 | struct drm_i915_privateinteldrm_softc *dev_priv = vma->obj->base.dev->dev_private; |
| 811 | struct vm_page *pg; |
| 812 | |
| 813 | for (pg = &dev_priv->pgs[atop(vma->node.start)((vma->node.start) >> 12)]; |
| 814 | pg != &dev_priv->pgs[atop(vma->node.start + vma->size)((vma->node.start + vma->size) >> 12)]; |
| 815 | pg++) |
| 816 | pmap_page_protect(pg, PROT_NONE0x00); |
| 817 | } |
| 818 | #endif |
| 819 | } |
| 820 | } |
| 821 | |
| 822 | static intel_engine_mask_t reset_prepare(struct intel_gt *gt) |
| 823 | { |
| 824 | struct intel_engine_cs *engine; |
| 825 | intel_engine_mask_t awake = 0; |
| 826 | enum intel_engine_id id; |
| 827 | |
| 828 | /* For GuC mode, ensure submission is disabled before stopping ring */ |
| 829 | intel_uc_reset_prepare(>->uc); |
| 830 | |
| 831 | for_each_engine(engine, gt, id)for ((id) = 0; (id) < I915_NUM_ENGINES; (id)++) if (!((engine ) = (gt)->engine[(id)])) {} else { |
| 832 | if (intel_engine_pm_get_if_awake(engine)) |
| 833 | awake |= engine->mask; |
| 834 | reset_prepare_engine(engine); |
| 835 | } |
| 836 | |
| 837 | return awake; |
| 838 | } |
| 839 | |
| 840 | static void gt_revoke(struct intel_gt *gt) |
| 841 | { |
| 842 | revoke_mmaps(gt); |
| 843 | } |
| 844 | |
| 845 | static int gt_reset(struct intel_gt *gt, intel_engine_mask_t stalled_mask) |
| 846 | { |
| 847 | struct intel_engine_cs *engine; |
| 848 | enum intel_engine_id id; |
| 849 | int err; |
| 850 | |
| 851 | /* |
| 852 | * Everything depends on having the GTT running, so we need to start |
| 853 | * there. |
| 854 | */ |
| 855 | err = i915_ggtt_enable_hw(gt->i915); |
| 856 | if (err) |
| 857 | return err; |
| 858 | |
| 859 | local_bh_disable(); |
| 860 | for_each_engine(engine, gt, id)for ((id) = 0; (id) < I915_NUM_ENGINES; (id)++) if (!((engine ) = (gt)->engine[(id)])) {} else |
| 861 | __intel_engine_reset(engine, stalled_mask & engine->mask); |
| 862 | local_bh_enable(); |
| 863 | |
| 864 | intel_uc_reset(>->uc, ALL_ENGINES((intel_engine_mask_t)~0ul)); |
| 865 | |
| 866 | intel_ggtt_restore_fences(gt->ggtt); |
| 867 | |
| 868 | return err; |
| 869 | } |
| 870 | |
| 871 | static void reset_finish_engine(struct intel_engine_cs *engine) |
| 872 | { |
| 873 | if (engine->reset.finish) |
| 874 | engine->reset.finish(engine); |
| 875 | intel_uncore_forcewake_put(engine->uncore, FORCEWAKE_ALL); |
| 876 | |
| 877 | intel_engine_signal_breadcrumbs(engine); |
| 878 | } |
| 879 | |
| 880 | static void reset_finish(struct intel_gt *gt, intel_engine_mask_t awake) |
| 881 | { |
| 882 | struct intel_engine_cs *engine; |
| 883 | enum intel_engine_id id; |
| 884 | |
| 885 | for_each_engine(engine, gt, id)for ((id) = 0; (id) < I915_NUM_ENGINES; (id)++) if (!((engine ) = (gt)->engine[(id)])) {} else { |
| 886 | reset_finish_engine(engine); |
| 887 | if (awake & engine->mask) |
| 888 | intel_engine_pm_put(engine); |
| 889 | } |
| 890 | |
| 891 | intel_uc_reset_finish(>->uc); |
| 892 | } |
| 893 | |
| 894 | static void nop_submit_request(struct i915_request *request) |
| 895 | { |
| 896 | 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); |
| 897 | |
| 898 | request = i915_request_mark_eio(request); |
| 899 | if (request) { |
| 900 | i915_request_submit(request); |
| 901 | intel_engine_signal_breadcrumbs(request->engine); |
| 902 | |
| 903 | i915_request_put(request); |
| 904 | } |
| 905 | } |
| 906 | |
| 907 | static void __intel_gt_set_wedged(struct intel_gt *gt) |
| 908 | { |
| 909 | struct intel_engine_cs *engine; |
| 910 | intel_engine_mask_t awake; |
| 911 | enum intel_engine_id id; |
| 912 | |
| 913 | if (test_bit(I915_WEDGED(64 - 1), >->reset.flags)) |
| 914 | return; |
| 915 | |
| 916 | GT_TRACE(gt, "start\n")do { const struct intel_gt *gt__ __attribute__((__unused__)) = (gt); do { } while (0); } while (0); |
| 917 | |
| 918 | /* |
| 919 | * First, stop submission to hw, but do not yet complete requests by |
| 920 | * rolling the global seqno forward (since this would complete requests |
| 921 | * for which we haven't set the fence error to EIO yet). |
| 922 | */ |
| 923 | awake = reset_prepare(gt); |
| 924 | |
| 925 | /* Even if the GPU reset fails, it should still stop the engines */ |
| 926 | if (!INTEL_INFO(gt->i915)(&(gt->i915)->__info)->gpu_reset_clobbers_display) |
| 927 | __intel_gt_reset(gt, ALL_ENGINES((intel_engine_mask_t)~0ul)); |
| 928 | |
| 929 | for_each_engine(engine, gt, id)for ((id) = 0; (id) < I915_NUM_ENGINES; (id)++) if (!((engine ) = (gt)->engine[(id)])) {} else |
| 930 | engine->submit_request = nop_submit_request; |
| 931 | |
| 932 | /* |
| 933 | * Make sure no request can slip through without getting completed by |
| 934 | * either this call here to intel_engine_write_global_seqno, or the one |
| 935 | * in nop_submit_request. |
| 936 | */ |
| 937 | synchronize_rcu_expedited(); |
| 938 | set_bit(I915_WEDGED(64 - 1), >->reset.flags); |
| 939 | |
| 940 | /* Mark all executing requests as skipped */ |
| 941 | local_bh_disable(); |
| 942 | for_each_engine(engine, gt, id)for ((id) = 0; (id) < I915_NUM_ENGINES; (id)++) if (!((engine ) = (gt)->engine[(id)])) {} else |
| 943 | if (engine->reset.cancel) |
| 944 | engine->reset.cancel(engine); |
| 945 | intel_uc_cancel_requests(>->uc); |
| 946 | local_bh_enable(); |
| 947 | |
| 948 | reset_finish(gt, awake); |
| 949 | |
| 950 | GT_TRACE(gt, "end\n")do { const struct intel_gt *gt__ __attribute__((__unused__)) = (gt); do { } while (0); } while (0); |
| 951 | } |
| 952 | |
| 953 | void intel_gt_set_wedged(struct intel_gt *gt) |
| 954 | { |
| 955 | intel_wakeref_t wakeref; |
| 956 | |
| 957 | if (test_bit(I915_WEDGED(64 - 1), >->reset.flags)) |
| 958 | return; |
| 959 | |
| 960 | wakeref = intel_runtime_pm_get(gt->uncore->rpm); |
| 961 | mutex_lock(>->reset.mutex)rw_enter_write(>->reset.mutex); |
| 962 | |
| 963 | if (GEM_SHOW_DEBUG()(0)) { |
| 964 | struct drm_printer p = drm_debug_printer(__func__); |
| 965 | struct intel_engine_cs *engine; |
| 966 | enum intel_engine_id id; |
| 967 | |
| 968 | drm_printf(&p, "called from %pS\n", (void *)_RET_IP___builtin_return_address(0)); |
| 969 | for_each_engine(engine, gt, id)for ((id) = 0; (id) < I915_NUM_ENGINES; (id)++) if (!((engine ) = (gt)->engine[(id)])) {} else { |
| 970 | if (intel_engine_is_idle(engine)) |
| 971 | continue; |
| 972 | |
| 973 | intel_engine_dump(engine, &p, "%s\n", engine->name); |
| 974 | } |
| 975 | } |
| 976 | |
| 977 | __intel_gt_set_wedged(gt); |
| 978 | |
| 979 | mutex_unlock(>->reset.mutex)rw_exit_write(>->reset.mutex); |
| 980 | intel_runtime_pm_put(gt->uncore->rpm, wakeref); |
| 981 | } |
| 982 | |
| 983 | static bool_Bool __intel_gt_unset_wedged(struct intel_gt *gt) |
| 984 | { |
| 985 | struct intel_gt_timelines *timelines = >->timelines; |
| 986 | struct intel_timeline *tl; |
| 987 | bool_Bool ok; |
| 988 | |
| 989 | if (!test_bit(I915_WEDGED(64 - 1), >->reset.flags)) |
| 990 | return true1; |
| 991 | |
| 992 | /* Never fully initialised, recovery impossible */ |
| 993 | if (intel_gt_has_unrecoverable_error(gt)) |
| 994 | return false0; |
| 995 | |
| 996 | GT_TRACE(gt, "start\n")do { const struct intel_gt *gt__ __attribute__((__unused__)) = (gt); do { } while (0); } while (0); |
| 997 | |
| 998 | /* |
| 999 | * Before unwedging, make sure that all pending operations |
| 1000 | * are flushed and errored out - we may have requests waiting upon |
| 1001 | * third party fences. We marked all inflight requests as EIO, and |
| 1002 | * every execbuf since returned EIO, for consistency we want all |
| 1003 | * the currently pending requests to also be marked as EIO, which |
| 1004 | * is done inside our nop_submit_request - and so we must wait. |
| 1005 | * |
| 1006 | * No more can be submitted until we reset the wedged bit. |
| 1007 | */ |
| 1008 | spin_lock(&timelines->lock)mtx_enter(&timelines->lock); |
| 1009 | 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) );})) { |
| 1010 | struct dma_fence *fence; |
| 1011 | |
| 1012 | fence = i915_active_fence_get(&tl->last_request); |
| 1013 | if (!fence) |
| 1014 | continue; |
| 1015 | |
| 1016 | spin_unlock(&timelines->lock)mtx_leave(&timelines->lock); |
| 1017 | |
| 1018 | /* |
| 1019 | * All internal dependencies (i915_requests) will have |
| 1020 | * been flushed by the set-wedge, but we may be stuck waiting |
| 1021 | * for external fences. These should all be capped to 10s |
| 1022 | * (I915_FENCE_TIMEOUT) so this wait should not be unbounded |
| 1023 | * in the worst case. |
| 1024 | */ |
| 1025 | dma_fence_default_wait(fence, false0, MAX_SCHEDULE_TIMEOUT(0x7fffffff)); |
| 1026 | dma_fence_put(fence); |
| 1027 | |
| 1028 | /* Restart iteration after droping lock */ |
| 1029 | spin_lock(&timelines->lock)mtx_enter(&timelines->lock); |
| 1030 | 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) );}); |
| 1031 | } |
| 1032 | spin_unlock(&timelines->lock)mtx_leave(&timelines->lock); |
| 1033 | |
| 1034 | /* We must reset pending GPU events before restoring our submission */ |
| 1035 | ok = !HAS_EXECLISTS(gt->i915)((&(gt->i915)->__info)->has_logical_ring_contexts ); /* XXX better agnosticism desired */ |
| 1036 | if (!INTEL_INFO(gt->i915)(&(gt->i915)->__info)->gpu_reset_clobbers_display) |
| 1037 | ok = __intel_gt_reset(gt, ALL_ENGINES((intel_engine_mask_t)~0ul)) == 0; |
| 1038 | if (!ok) { |
| 1039 | /* |
| 1040 | * Warn CI about the unrecoverable wedged condition. |
| 1041 | * Time for a reboot. |
| 1042 | */ |
| 1043 | add_taint_for_CI(gt->i915, TAINT_WARN1); |
| 1044 | return false0; |
| 1045 | } |
| 1046 | |
| 1047 | /* |
| 1048 | * Undo nop_submit_request. We prevent all new i915 requests from |
| 1049 | * being queued (by disallowing execbuf whilst wedged) so having |
| 1050 | * waited for all active requests above, we know the system is idle |
| 1051 | * and do not have to worry about a thread being inside |
| 1052 | * engine->submit_request() as we swap over. So unlike installing |
| 1053 | * the nop_submit_request on reset, we can do this from normal |
| 1054 | * context and do not require stop_machine(). |
| 1055 | */ |
| 1056 | intel_engines_reset_default_submission(gt); |
| 1057 | |
| 1058 | GT_TRACE(gt, "end\n")do { const struct intel_gt *gt__ __attribute__((__unused__)) = (gt); do { } while (0); } while (0); |
| 1059 | |
| 1060 | smp_mb__before_atomic()do { } while (0); /* complete takeover before enabling execbuf */ |
| 1061 | clear_bit(I915_WEDGED(64 - 1), >->reset.flags); |
| 1062 | |
| 1063 | return true1; |
| 1064 | } |
| 1065 | |
| 1066 | bool_Bool intel_gt_unset_wedged(struct intel_gt *gt) |
| 1067 | { |
| 1068 | bool_Bool result; |
| 1069 | |
| 1070 | mutex_lock(>->reset.mutex)rw_enter_write(>->reset.mutex); |
| 1071 | result = __intel_gt_unset_wedged(gt); |
| 1072 | mutex_unlock(>->reset.mutex)rw_exit_write(>->reset.mutex); |
| 1073 | |
| 1074 | return result; |
| 1075 | } |
| 1076 | |
| 1077 | static int do_reset(struct intel_gt *gt, intel_engine_mask_t stalled_mask) |
| 1078 | { |
| 1079 | int err, i; |
| 1080 | |
| 1081 | err = __intel_gt_reset(gt, ALL_ENGINES((intel_engine_mask_t)~0ul)); |
| 1082 | for (i = 0; err && i < RESET_MAX_RETRIES3; i++) { |
| 1083 | drm_msleep(10 * (i + 1))mdelay(10 * (i + 1)); |
| 1084 | err = __intel_gt_reset(gt, ALL_ENGINES((intel_engine_mask_t)~0ul)); |
| 1085 | } |
| 1086 | if (err) |
| 1087 | return err; |
| 1088 | |
| 1089 | return gt_reset(gt, stalled_mask); |
| 1090 | } |
| 1091 | |
| 1092 | static int resume(struct intel_gt *gt) |
| 1093 | { |
| 1094 | struct intel_engine_cs *engine; |
| 1095 | enum intel_engine_id id; |
| 1096 | int ret; |
| 1097 | |
| 1098 | for_each_engine(engine, gt, id)for ((id) = 0; (id) < I915_NUM_ENGINES; (id)++) if (!((engine ) = (gt)->engine[(id)])) {} else { |
| 1099 | ret = intel_engine_resume(engine); |
| 1100 | if (ret) |
| 1101 | return ret; |
| 1102 | } |
| 1103 | |
| 1104 | return 0; |
| 1105 | } |
| 1106 | |
| 1107 | /** |
| 1108 | * intel_gt_reset - reset chip after a hang |
| 1109 | * @gt: #intel_gt to reset |
| 1110 | * @stalled_mask: mask of the stalled engines with the guilty requests |
| 1111 | * @reason: user error message for why we are resetting |
| 1112 | * |
| 1113 | * Reset the chip. Useful if a hang is detected. Marks the device as wedged |
| 1114 | * on failure. |
| 1115 | * |
| 1116 | * Procedure is fairly simple: |
| 1117 | * - reset the chip using the reset reg |
| 1118 | * - re-init context state |
| 1119 | * - re-init hardware status page |
| 1120 | * - re-init ring buffer |
| 1121 | * - re-init interrupt state |
| 1122 | * - re-init display |
| 1123 | */ |
| 1124 | void intel_gt_reset(struct intel_gt *gt, |
| 1125 | intel_engine_mask_t stalled_mask, |
| 1126 | const char *reason) |
| 1127 | { |
| 1128 | intel_engine_mask_t awake; |
| 1129 | int ret; |
| 1130 | |
| 1131 | GT_TRACE(gt, "flags=%lx\n", gt->reset.flags)do { const struct intel_gt *gt__ __attribute__((__unused__)) = (gt); do { } while (0); } while (0); |
| 1132 | |
| 1133 | might_sleep()assertwaitok(); |
| 1134 | GEM_BUG_ON(!test_bit(I915_RESET_BACKOFF, >->reset.flags))((void)0); |
| 1135 | |
| 1136 | /* |
| 1137 | * FIXME: Revoking cpu mmap ptes cannot be done from a dma_fence |
| 1138 | * critical section like gpu reset. |
| 1139 | */ |
| 1140 | gt_revoke(gt); |
| 1141 | |
| 1142 | mutex_lock(>->reset.mutex)rw_enter_write(>->reset.mutex); |
| 1143 | |
| 1144 | /* Clear any previous failed attempts at recovery. Time to try again. */ |
| 1145 | if (!__intel_gt_unset_wedged(gt)) |
| 1146 | goto unlock; |
| 1147 | |
| 1148 | if (reason) |
| 1149 | 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 ) |
| 1150 | "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 ); |
| 1151 | atomic_inc(>->i915->gpu_error.reset_count)__sync_fetch_and_add(>->i915->gpu_error.reset_count , 1); |
| 1152 | |
| 1153 | awake = reset_prepare(gt); |
| 1154 | |
| 1155 | if (!intel_has_gpu_reset(gt)) { |
| 1156 | if (gt->i915->params.reset) |
| 1157 | 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__); |
| 1158 | else |
| 1159 | drm_dbg(>->i915->drm, "GPU reset disabled\n")__drm_dev_dbg(((void *)0), (>->i915->drm) ? (& gt->i915->drm)->dev : ((void *)0), DRM_UT_DRIVER, "GPU reset disabled\n" ); |
| 1160 | goto error; |
| 1161 | } |
| 1162 | |
| 1163 | if (INTEL_INFO(gt->i915)(&(gt->i915)->__info)->gpu_reset_clobbers_display) |
| 1164 | intel_runtime_pm_disable_interrupts(gt->i915); |
| 1165 | |
| 1166 | if (do_reset(gt, stalled_mask)) { |
| 1167 | 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__); |
| 1168 | goto taint; |
| 1169 | } |
| 1170 | |
| 1171 | if (INTEL_INFO(gt->i915)(&(gt->i915)->__info)->gpu_reset_clobbers_display) |
| 1172 | intel_runtime_pm_enable_interrupts(gt->i915); |
| 1173 | |
| 1174 | intel_overlay_reset(gt->i915); |
| 1175 | |
| 1176 | /* |
| 1177 | * Next we need to restore the context, but we don't use those |
| 1178 | * yet either... |
| 1179 | * |
| 1180 | * Ring buffer needs to be re-initialized in the KMS case, or if X |
| 1181 | * was running at the time of the reset (i.e. we weren't VT |
| 1182 | * switched away). |
| 1183 | */ |
| 1184 | ret = intel_gt_init_hw(gt); |
| 1185 | if (ret) { |
| 1186 | 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) |
| 1187 | "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) |
| 1188 | 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); |
| 1189 | goto taint; |
| 1190 | } |
| 1191 | |
| 1192 | ret = resume(gt); |
| 1193 | if (ret) |
| 1194 | goto taint; |
| 1195 | |
| 1196 | finish: |
| 1197 | reset_finish(gt, awake); |
| 1198 | unlock: |
| 1199 | mutex_unlock(>->reset.mutex)rw_exit_write(>->reset.mutex); |
| 1200 | return; |
| 1201 | |
| 1202 | taint: |
| 1203 | /* |
| 1204 | * History tells us that if we cannot reset the GPU now, we |
| 1205 | * never will. This then impacts everything that is run |
| 1206 | * subsequently. On failing the reset, we mark the driver |
| 1207 | * as wedged, preventing further execution on the GPU. |
| 1208 | * We also want to go one step further and add a taint to the |
| 1209 | * kernel so that any subsequent faults can be traced back to |
| 1210 | * this failure. This is important for CI, where if the |
| 1211 | * GPU/driver fails we would like to reboot and restart testing |
| 1212 | * rather than continue on into oblivion. For everyone else, |
| 1213 | * the system should still plod along, but they have been warned! |
| 1214 | */ |
| 1215 | add_taint_for_CI(gt->i915, TAINT_WARN1); |
| 1216 | error: |
| 1217 | __intel_gt_set_wedged(gt); |
| 1218 | goto finish; |
| 1219 | } |
| 1220 | |
| 1221 | static int intel_gt_reset_engine(struct intel_engine_cs *engine) |
| 1222 | { |
| 1223 | return __intel_gt_reset(engine->gt, engine->mask); |
| 1224 | } |
| 1225 | |
| 1226 | int __intel_engine_reset_bh(struct intel_engine_cs *engine, const char *msg) |
| 1227 | { |
| 1228 | struct intel_gt *gt = engine->gt; |
| 1229 | int ret; |
| 1230 | |
| 1231 | ENGINE_TRACE(engine, "flags=%lx\n", gt->reset.flags)do { const struct intel_engine_cs *e__ __attribute__((__unused__ )) = (engine); do { } while (0); } while (0); |
| 1232 | GEM_BUG_ON(!test_bit(I915_RESET_ENGINE + engine->id, >->reset.flags))((void)0); |
| 1233 | |
| 1234 | if (intel_engine_uses_guc(engine)) |
| 1235 | return -ENODEV19; |
| 1236 | |
| 1237 | if (!intel_engine_pm_get_if_awake(engine)) |
| 1238 | return 0; |
| 1239 | |
| 1240 | reset_prepare_engine(engine); |
| 1241 | |
| 1242 | if (msg) |
| 1243 | 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) |
| 1244 | "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); |
| 1245 | 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); |
| 1246 | |
| 1247 | ret = intel_gt_reset_engine(engine); |
| 1248 | if (ret) { |
| 1249 | /* If we fail here, we expect to fallback to a global reset */ |
| 1250 | ENGINE_TRACE(engine, "Failed to reset %s, err: %d\n", engine->name, ret)do { const struct intel_engine_cs *e__ __attribute__((__unused__ )) = (engine); do { } while (0); } while (0); |
| 1251 | goto out; |
| 1252 | } |
| 1253 | |
| 1254 | /* |
| 1255 | * The request that caused the hang is stuck on elsp, we know the |
| 1256 | * active request and can drop it, adjust head to skip the offending |
| 1257 | * request to resume executing remaining requests in the queue. |
| 1258 | */ |
| 1259 | __intel_engine_reset(engine, true1); |
| 1260 | |
| 1261 | /* |
| 1262 | * The engine and its registers (and workarounds in case of render) |
| 1263 | * have been reset to their default values. Follow the init_ring |
| 1264 | * process to program RING_MODE, HWSP and re-enable submission. |
| 1265 | */ |
| 1266 | ret = intel_engine_resume(engine); |
| 1267 | |
| 1268 | out: |
| 1269 | intel_engine_cancel_stop_cs(engine); |
| 1270 | reset_finish_engine(engine); |
| 1271 | intel_engine_pm_put_async(engine); |
| 1272 | return ret; |
| 1273 | } |
| 1274 | |
| 1275 | /** |
| 1276 | * intel_engine_reset - reset GPU engine to recover from a hang |
| 1277 | * @engine: engine to reset |
| 1278 | * @msg: reason for GPU reset; or NULL for no drm_notice() |
| 1279 | * |
| 1280 | * Reset a specific GPU engine. Useful if a hang is detected. |
| 1281 | * Returns zero on successful reset or otherwise an error code. |
| 1282 | * |
| 1283 | * Procedure is: |
| 1284 | * - identifies the request that caused the hang and it is dropped |
| 1285 | * - reset engine (which will force the engine to idle) |
| 1286 | * - re-init/configure engine |
| 1287 | */ |
| 1288 | int intel_engine_reset(struct intel_engine_cs *engine, const char *msg) |
| 1289 | { |
| 1290 | int err; |
| 1291 | |
| 1292 | local_bh_disable(); |
| 1293 | err = __intel_engine_reset_bh(engine, msg); |
| 1294 | local_bh_enable(); |
| 1295 | |
| 1296 | return err; |
| 1297 | } |
| 1298 | |
| 1299 | static void intel_gt_reset_global(struct intel_gt *gt, |
| 1300 | u32 engine_mask, |
| 1301 | const char *reason) |
| 1302 | { |
| 1303 | #ifdef notyet |
| 1304 | struct kobject *kobj = >->i915->drm.primary->kdev->kobj; |
| 1305 | char *error_event[] = { I915_ERROR_UEVENT"ERROR" "=1", NULL((void *)0) }; |
| 1306 | char *reset_event[] = { I915_RESET_UEVENT"RESET" "=1", NULL((void *)0) }; |
| 1307 | char *reset_done_event[] = { I915_ERROR_UEVENT"ERROR" "=0", NULL((void *)0) }; |
| 1308 | #endif |
| 1309 | struct intel_wedge_me w; |
| 1310 | |
| 1311 | kobject_uevent_env(kobj, KOBJ_CHANGE, error_event); |
| 1312 | |
| 1313 | GT_TRACE(gt, "resetting chip, engines=%x\n", engine_mask)do { const struct intel_gt *gt__ __attribute__((__unused__)) = (gt); do { } while (0); } while (0); |
| 1314 | kobject_uevent_env(kobj, KOBJ_CHANGE, reset_event); |
| 1315 | |
| 1316 | /* Use a watchdog to ensure that our reset completes */ |
| 1317 | intel_wedge_on_timeout(&w, gt, 5 * HZ)for (__intel_init_wedge((&w), (gt), (5 * hz), __func__); ( &w)->gt; __intel_fini_wedge((&w))) { |
| 1318 | intel_display_prepare_reset(gt->i915); |
| 1319 | |
| 1320 | intel_gt_reset(gt, engine_mask, reason); |
| 1321 | |
| 1322 | intel_display_finish_reset(gt->i915); |
| 1323 | } |
| 1324 | |
| 1325 | if (!test_bit(I915_WEDGED(64 - 1), >->reset.flags)) |
| 1326 | kobject_uevent_env(kobj, KOBJ_CHANGE, reset_done_event); |
| 1327 | } |
| 1328 | |
| 1329 | /** |
| 1330 | * intel_gt_handle_error - handle a gpu error |
| 1331 | * @gt: the intel_gt |
| 1332 | * @engine_mask: mask representing engines that are hung |
| 1333 | * @flags: control flags |
| 1334 | * @fmt: Error message format string |
| 1335 | * |
| 1336 | * Do some basic checking of register state at error time and |
| 1337 | * dump it to the syslog. Also call i915_capture_error_state() to make |
| 1338 | * sure we get a record and make it available in debugfs. Fire a uevent |
| 1339 | * so userspace knows something bad happened (should trigger collection |
| 1340 | * of a ring dump etc.). |
| 1341 | */ |
| 1342 | void intel_gt_handle_error(struct intel_gt *gt, |
| 1343 | intel_engine_mask_t engine_mask, |
| 1344 | unsigned long flags, |
| 1345 | const char *fmt, ...) |
| 1346 | { |
| 1347 | struct intel_engine_cs *engine; |
| 1348 | intel_wakeref_t wakeref; |
| 1349 | intel_engine_mask_t tmp; |
| 1350 | char error_msg[80]; |
| 1351 | char *msg = NULL((void *)0); |
| 1352 | |
| 1353 | if (fmt) { |
| 1354 | va_list args; |
| 1355 | |
| 1356 | va_start(args, fmt)__builtin_va_start((args), fmt); |
| 1357 | vscnprintf(error_msg, sizeof(error_msg), fmt, args); |
| 1358 | va_end(args)__builtin_va_end((args)); |
| 1359 | |
| 1360 | msg = error_msg; |
| 1361 | } |
| 1362 | |
| 1363 | /* |
| 1364 | * In most cases it's guaranteed that we get here with an RPM |
| 1365 | * reference held, for example because there is a pending GPU |
| 1366 | * request that won't finish until the reset is done. This |
| 1367 | * isn't the case at least when we get here by doing a |
| 1368 | * simulated reset via debugfs, so get an RPM reference. |
| 1369 | */ |
| 1370 | wakeref = intel_runtime_pm_get(gt->uncore->rpm); |
| 1371 | |
| 1372 | engine_mask &= gt->info.engine_mask; |
| 1373 | |
| 1374 | if (flags & I915_ERROR_CAPTURE(1UL << (0))) { |
| 1375 | i915_capture_error_state(gt, engine_mask, CORE_DUMP_FLAG_NONE0x0); |
| 1376 | intel_gt_clear_error_registers(gt, engine_mask); |
| 1377 | } |
| 1378 | |
| 1379 | /* |
| 1380 | * Try engine reset when available. We fall back to full reset if |
| 1381 | * single reset fails. |
| 1382 | */ |
| 1383 | if (!intel_uc_uses_guc_submission(>->uc) && |
| 1384 | intel_has_reset_engine(gt) && !intel_gt_is_wedged(gt)) { |
| 1385 | local_bh_disable(); |
| 1386 | 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;) { |
| 1387 | BUILD_BUG_ON(I915_RESET_MODESET >= I915_RESET_ENGINE)extern char _ctassert[(!(1 >= 2)) ? 1 : -1 ] __attribute__ ((__unused__)); |
| 1388 | if (test_and_set_bit(I915_RESET_ENGINE2 + engine->id, |
| 1389 | >->reset.flags)) |
| 1390 | continue; |
| 1391 | |
| 1392 | if (__intel_engine_reset_bh(engine, msg) == 0) |
| 1393 | engine_mask &= ~engine->mask; |
| 1394 | |
| 1395 | clear_and_wake_up_bit(I915_RESET_ENGINE2 + engine->id, |
| 1396 | >->reset.flags); |
| 1397 | } |
| 1398 | local_bh_enable(); |
| 1399 | } |
| 1400 | |
| 1401 | if (!engine_mask) |
| 1402 | goto out; |
| 1403 | |
| 1404 | /* Full reset needs the mutex, stop any other user trying to do so. */ |
| 1405 | if (test_and_set_bit(I915_RESET_BACKOFF0, >->reset.flags)) { |
| 1406 | wait_event(gt->reset.queue,do { if (!(!test_bit(0, >->reset.flags))) ({ long __ret = 0; struct wait_queue_entry __wq_entry; init_wait_entry(& __wq_entry, 0); do { int __error, __wait; unsigned long deadline ; ((!cold) ? (void)0 : __assert("diagnostic ", "/usr/src/sys/dev/pci/drm/i915/gt/intel_reset.c" , 1407, "!cold")); prepare_to_wait(>->reset.queue, & __wq_entry, 0); deadline = jiffies + __ret; __wait = !(!test_bit (0, >->reset.flags)); __error = sleep_finish(__ret, __wait ); if ((0) > 0) __ret = deadline - jiffies; if (__error == -1 || __error == 4) { __ret = -4; break; } if ((0) > 0 && (__ret <= 0 || __error == 35)) { __ret = ((!test_bit(0, & gt->reset.flags))) ? 1 : 0; break; } } while (__ret > 0 && !(!test_bit(0, >->reset.flags))); finish_wait (>->reset.queue, &__wq_entry); __ret; }); } while (0) |
| 1407 | !test_bit(I915_RESET_BACKOFF, >->reset.flags))do { if (!(!test_bit(0, >->reset.flags))) ({ long __ret = 0; struct wait_queue_entry __wq_entry; init_wait_entry(& __wq_entry, 0); do { int __error, __wait; unsigned long deadline ; ((!cold) ? (void)0 : __assert("diagnostic ", "/usr/src/sys/dev/pci/drm/i915/gt/intel_reset.c" , 1407, "!cold")); prepare_to_wait(>->reset.queue, & __wq_entry, 0); deadline = jiffies + __ret; __wait = !(!test_bit (0, >->reset.flags)); __error = sleep_finish(__ret, __wait ); if ((0) > 0) __ret = deadline - jiffies; if (__error == -1 || __error == 4) { __ret = -4; break; } if ((0) > 0 && (__ret <= 0 || __error == 35)) { __ret = ((!test_bit(0, & gt->reset.flags))) ? 1 : 0; break; } } while (__ret > 0 && !(!test_bit(0, >->reset.flags))); finish_wait (>->reset.queue, &__wq_entry); __ret; }); } while (0); |
| 1408 | goto out; /* piggy-back on the other reset */ |
| 1409 | } |
| 1410 | |
| 1411 | /* Make sure i915_reset_trylock() sees the I915_RESET_BACKOFF */ |
| 1412 | synchronize_rcu_expedited(); |
| 1413 | |
| 1414 | /* |
| 1415 | * Prevent any other reset-engine attempt. We don't do this for GuC |
| 1416 | * submission the GuC owns the per-engine reset, not the i915. |
| 1417 | */ |
| 1418 | if (!intel_uc_uses_guc_submission(>->uc)) { |
| 1419 | for_each_engine(engine, gt, tmp)for ((tmp) = 0; (tmp) < I915_NUM_ENGINES; (tmp)++) if (!(( engine) = (gt)->engine[(tmp)])) {} else { |
| 1420 | while (test_and_set_bit(I915_RESET_ENGINE2 + engine->id, |
| 1421 | >->reset.flags)) |
| 1422 | wait_on_bit(>->reset.flags, |
| 1423 | I915_RESET_ENGINE2 + engine->id, |
| 1424 | TASK_UNINTERRUPTIBLE0); |
| 1425 | } |
| 1426 | } |
| 1427 | |
| 1428 | /* Flush everyone using a resource about to be clobbered */ |
| 1429 | synchronize_srcu_expedited(>->reset.backoff_srcu); |
| 1430 | |
| 1431 | intel_gt_reset_global(gt, engine_mask, msg); |
| 1432 | |
| 1433 | if (!intel_uc_uses_guc_submission(>->uc)) { |
| 1434 | for_each_engine(engine, gt, tmp)for ((tmp) = 0; (tmp) < I915_NUM_ENGINES; (tmp)++) if (!(( engine) = (gt)->engine[(tmp)])) {} else |
| 1435 | clear_bit_unlock(I915_RESET_ENGINE2 + engine->id, |
| 1436 | >->reset.flags); |
| 1437 | } |
| 1438 | clear_bit_unlock(I915_RESET_BACKOFF0, >->reset.flags); |
| 1439 | smp_mb__after_atomic()do { } while (0); |
| 1440 | wake_up_all(>->reset.queue)wake_up(>->reset.queue); |
| 1441 | |
| 1442 | out: |
| 1443 | intel_runtime_pm_put(gt->uncore->rpm, wakeref); |
| 1444 | } |
| 1445 | |
| 1446 | int intel_gt_reset_trylock(struct intel_gt *gt, int *srcu) |
| 1447 | { |
| 1448 | might_lock(>->reset.backoff_srcu); |
| 1449 | might_sleep()assertwaitok(); |
| 1450 | |
| 1451 | rcu_read_lock(); |
| 1452 | while (test_bit(I915_RESET_BACKOFF0, >->reset.flags)) { |
| 1453 | rcu_read_unlock(); |
| 1454 | |
| 1455 | if (wait_event_interruptible(gt->reset.queue,({ int __ret = 0; if (!(!test_bit(0, >->reset.flags) )) __ret = ({ long __ret = 0; struct wait_queue_entry __wq_entry ; init_wait_entry(&__wq_entry, 0); do { int __error, __wait ; unsigned long deadline; ((!cold) ? (void)0 : __assert("diagnostic " , "/usr/src/sys/dev/pci/drm/i915/gt/intel_reset.c", 1457, "!cold" )); prepare_to_wait(>->reset.queue, &__wq_entry, 0x100); deadline = jiffies + __ret; __wait = !(!test_bit(0, & gt->reset.flags)); __error = sleep_finish(__ret, __wait); if ((0) > 0) __ret = deadline - jiffies; if (__error == -1 || __error == 4) { __ret = -4; break; } if ((0) > 0 && (__ret <= 0 || __error == 35)) { __ret = ((!test_bit(0, & gt->reset.flags))) ? 1 : 0; break; } } while (__ret > 0 && !(!test_bit(0, >->reset.flags))); finish_wait (>->reset.queue, &__wq_entry); __ret; }); __ret; }) |
| 1456 | !test_bit(I915_RESET_BACKOFF,({ int __ret = 0; if (!(!test_bit(0, >->reset.flags) )) __ret = ({ long __ret = 0; struct wait_queue_entry __wq_entry ; init_wait_entry(&__wq_entry, 0); do { int __error, __wait ; unsigned long deadline; ((!cold) ? (void)0 : __assert("diagnostic " , "/usr/src/sys/dev/pci/drm/i915/gt/intel_reset.c", 1457, "!cold" )); prepare_to_wait(>->reset.queue, &__wq_entry, 0x100); deadline = jiffies + __ret; __wait = !(!test_bit(0, & gt->reset.flags)); __error = sleep_finish(__ret, __wait); if ((0) > 0) __ret = deadline - jiffies; if (__error == -1 || __error == 4) { __ret = -4; break; } if ((0) > 0 && (__ret <= 0 || __error == 35)) { __ret = ((!test_bit(0, & gt->reset.flags))) ? 1 : 0; break; } } while (__ret > 0 && !(!test_bit(0, >->reset.flags))); finish_wait (>->reset.queue, &__wq_entry); __ret; }); __ret; }) |
| 1457 | >->reset.flags))({ int __ret = 0; if (!(!test_bit(0, >->reset.flags) )) __ret = ({ long __ret = 0; struct wait_queue_entry __wq_entry ; init_wait_entry(&__wq_entry, 0); do { int __error, __wait ; unsigned long deadline; ((!cold) ? (void)0 : __assert("diagnostic " , "/usr/src/sys/dev/pci/drm/i915/gt/intel_reset.c", 1457, "!cold" )); prepare_to_wait(>->reset.queue, &__wq_entry, 0x100); deadline = jiffies + __ret; __wait = !(!test_bit(0, & gt->reset.flags)); __error = sleep_finish(__ret, __wait); if ((0) > 0) __ret = deadline - jiffies; if (__error == -1 || __error == 4) { __ret = -4; break; } if ((0) > 0 && (__ret <= 0 || __error == 35)) { __ret = ((!test_bit(0, & gt->reset.flags))) ? 1 : 0; break; } } while (__ret > 0 && !(!test_bit(0, >->reset.flags))); finish_wait (>->reset.queue, &__wq_entry); __ret; }); __ret; })) |
| 1458 | return -EINTR4; |
| 1459 | |
| 1460 | rcu_read_lock(); |
| 1461 | } |
| 1462 | *srcu = srcu_read_lock(>->reset.backoff_srcu)0; |
| 1463 | rcu_read_unlock(); |
| 1464 | |
| 1465 | return 0; |
| 1466 | } |
| 1467 | |
| 1468 | void intel_gt_reset_unlock(struct intel_gt *gt, int tag) |
| 1469 | __releases(>->reset.backoff_srcu) |
| 1470 | { |
| 1471 | srcu_read_unlock(>->reset.backoff_srcu, tag); |
| 1472 | } |
| 1473 | |
| 1474 | int intel_gt_terminally_wedged(struct intel_gt *gt) |
| 1475 | { |
| 1476 | might_sleep()assertwaitok(); |
| 1477 | |
| 1478 | if (!intel_gt_is_wedged(gt)) |
| 1479 | return 0; |
| 1480 | |
| 1481 | if (intel_gt_has_unrecoverable_error(gt)) |
| 1482 | return -EIO5; |
| 1483 | |
| 1484 | /* Reset still in progress? Maybe we will recover? */ |
| 1485 | if (wait_event_interruptible(gt->reset.queue,({ int __ret = 0; if (!(!test_bit(0, >->reset.flags) )) __ret = ({ long __ret = 0; struct wait_queue_entry __wq_entry ; init_wait_entry(&__wq_entry, 0); do { int __error, __wait ; unsigned long deadline; ((!cold) ? (void)0 : __assert("diagnostic " , "/usr/src/sys/dev/pci/drm/i915/gt/intel_reset.c", 1487, "!cold" )); prepare_to_wait(>->reset.queue, &__wq_entry, 0x100); deadline = jiffies + __ret; __wait = !(!test_bit(0, & gt->reset.flags)); __error = sleep_finish(__ret, __wait); if ((0) > 0) __ret = deadline - jiffies; if (__error == -1 || __error == 4) { __ret = -4; break; } if ((0) > 0 && (__ret <= 0 || __error == 35)) { __ret = ((!test_bit(0, & gt->reset.flags))) ? 1 : 0; break; } } while (__ret > 0 && !(!test_bit(0, >->reset.flags))); finish_wait (>->reset.queue, &__wq_entry); __ret; }); __ret; }) |
| 1486 | !test_bit(I915_RESET_BACKOFF,({ int __ret = 0; if (!(!test_bit(0, >->reset.flags) )) __ret = ({ long __ret = 0; struct wait_queue_entry __wq_entry ; init_wait_entry(&__wq_entry, 0); do { int __error, __wait ; unsigned long deadline; ((!cold) ? (void)0 : __assert("diagnostic " , "/usr/src/sys/dev/pci/drm/i915/gt/intel_reset.c", 1487, "!cold" )); prepare_to_wait(>->reset.queue, &__wq_entry, 0x100); deadline = jiffies + __ret; __wait = !(!test_bit(0, & gt->reset.flags)); __error = sleep_finish(__ret, __wait); if ((0) > 0) __ret = deadline - jiffies; if (__error == -1 || __error == 4) { __ret = -4; break; } if ((0) > 0 && (__ret <= 0 || __error == 35)) { __ret = ((!test_bit(0, & gt->reset.flags))) ? 1 : 0; break; } } while (__ret > 0 && !(!test_bit(0, >->reset.flags))); finish_wait (>->reset.queue, &__wq_entry); __ret; }); __ret; }) |
| 1487 | >->reset.flags))({ int __ret = 0; if (!(!test_bit(0, >->reset.flags) )) __ret = ({ long __ret = 0; struct wait_queue_entry __wq_entry ; init_wait_entry(&__wq_entry, 0); do { int __error, __wait ; unsigned long deadline; ((!cold) ? (void)0 : __assert("diagnostic " , "/usr/src/sys/dev/pci/drm/i915/gt/intel_reset.c", 1487, "!cold" )); prepare_to_wait(>->reset.queue, &__wq_entry, 0x100); deadline = jiffies + __ret; __wait = !(!test_bit(0, & gt->reset.flags)); __error = sleep_finish(__ret, __wait); if ((0) > 0) __ret = deadline - jiffies; if (__error == -1 || __error == 4) { __ret = -4; break; } if ((0) > 0 && (__ret <= 0 || __error == 35)) { __ret = ((!test_bit(0, & gt->reset.flags))) ? 1 : 0; break; } } while (__ret > 0 && !(!test_bit(0, >->reset.flags))); finish_wait (>->reset.queue, &__wq_entry); __ret; }); __ret; })) |
| 1488 | return -EINTR4; |
| 1489 | |
| 1490 | return intel_gt_is_wedged(gt) ? -EIO5 : 0; |
| 1491 | } |
| 1492 | |
| 1493 | void intel_gt_set_wedged_on_init(struct intel_gt *gt) |
| 1494 | { |
| 1495 | BUILD_BUG_ON(I915_RESET_ENGINE + I915_NUM_ENGINES >extern char _ctassert[(!(2 + I915_NUM_ENGINES > (64 - 3))) ? 1 : -1 ] __attribute__((__unused__)) |
| 1496 | I915_WEDGED_ON_INIT)extern char _ctassert[(!(2 + I915_NUM_ENGINES > (64 - 3))) ? 1 : -1 ] __attribute__((__unused__)); |
| 1497 | intel_gt_set_wedged(gt); |
| 1498 | i915_disable_error_state(gt->i915, -ENODEV19); |
| 1499 | set_bit(I915_WEDGED_ON_INIT(64 - 3), >->reset.flags); |
| 1500 | |
| 1501 | /* Wedged on init is non-recoverable */ |
| 1502 | add_taint_for_CI(gt->i915, TAINT_WARN1); |
| 1503 | } |
| 1504 | |
| 1505 | void intel_gt_set_wedged_on_fini(struct intel_gt *gt) |
| 1506 | { |
| 1507 | intel_gt_set_wedged(gt); |
| 1508 | i915_disable_error_state(gt->i915, -ENODEV19); |
| 1509 | set_bit(I915_WEDGED_ON_FINI(64 - 2), >->reset.flags); |
| 1510 | intel_gt_retire_requests(gt); /* cleanup any wedged requests */ |
| 1511 | } |
| 1512 | |
| 1513 | void intel_gt_init_reset(struct intel_gt *gt) |
| 1514 | { |
| 1515 | init_waitqueue_head(>->reset.queue); |
| 1516 | rw_init(>->reset.mutex, "gtres")_rw_init_flags(>->reset.mutex, "gtres", 0, ((void *) 0)); |
| 1517 | init_srcu_struct(>->reset.backoff_srcu); |
| 1518 | |
| 1519 | /* |
| 1520 | * While undesirable to wait inside the shrinker, complain anyway. |
| 1521 | * |
| 1522 | * If we have to wait during shrinking, we guarantee forward progress |
| 1523 | * by forcing the reset. Therefore during the reset we must not |
| 1524 | * re-enter the shrinker. By declaring that we take the reset mutex |
| 1525 | * within the shrinker, we forbid ourselves from performing any |
| 1526 | * fs-reclaim or taking related locks during reset. |
| 1527 | */ |
| 1528 | i915_gem_shrinker_taints_mutex(gt->i915, >->reset.mutex); |
| 1529 | |
| 1530 | /* no GPU until we are ready! */ |
| 1531 | __set_bit(I915_WEDGED(64 - 1), >->reset.flags); |
| 1532 | } |
| 1533 | |
| 1534 | void intel_gt_fini_reset(struct intel_gt *gt) |
| 1535 | { |
| 1536 | cleanup_srcu_struct(>->reset.backoff_srcu); |
| 1537 | } |
| 1538 | |
| 1539 | static void intel_wedge_me(struct work_struct *work) |
| 1540 | { |
| 1541 | 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) );}); |
| 1542 | |
| 1543 | 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) |
| 1544 | "%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) |
| 1545 | 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); |
| 1546 | intel_gt_set_wedged(w->gt); |
| 1547 | } |
| 1548 | |
| 1549 | void __intel_init_wedge(struct intel_wedge_me *w, |
| 1550 | struct intel_gt *gt, |
| 1551 | long timeout, |
| 1552 | const char *name) |
| 1553 | { |
| 1554 | w->gt = gt; |
| 1555 | w->name = name; |
| 1556 | |
| 1557 | INIT_DELAYED_WORK_ONSTACK(&w->work, intel_wedge_me); |
| 1558 | schedule_delayed_work(&w->work, timeout); |
| 1559 | } |
| 1560 | |
| 1561 | void __intel_fini_wedge(struct intel_wedge_me *w) |
| 1562 | { |
| 1563 | cancel_delayed_work_sync(&w->work); |
| 1564 | destroy_delayed_work_on_stack(&w->work); |
| 1565 | w->gt = NULL((void *)0); |
| 1566 | } |
| 1567 | |
| 1568 | #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)0 |
| 1569 | #include "selftest_reset.c" |
| 1570 | #include "selftest_hangcheck.c" |
| 1571 | #endif |