From 5febf432df1cfa5b25d99b54c32103fafbdd0eb9 Mon Sep 17 00:00:00 2001 From: Miguel Ojeda Date: Sun, 16 Aug 2026 15:32:33 +0200 Subject: kbuild: rust: keep Rust objects out of Clang LTO with inline helpers Under `CONFIG_LTO_CLANG` + `CONFIG_RUST_INLINE_HELPERS`, one may hit `objtool` errors such as: vmlinux.o: error: objtool: _R..._3Gsp4boot+0xd6a: can't find jump dest instruction at .text._R..._3Gsp4boot+0x1dfd The reason is that in such builds, the Clang invocation that compiles the combined Rust plus helpers bitcode emits LLVM bitcode (again) -- the final code generation happens in the linker's LTO step, which the `-mllvm` trap options passed to Clang do not reach. This, in turn, means that unreachable traps are missing, and the impossible paths do not merely fallthrough to the next symbol, but past the end of their own section, since LTO builds place each function in its own section. Thus filter `CC_FLAGS_LTO` out of the Clang invocation, so that it always emits machine code directly, with the traps in place. Assisted-by: LLM Cc: Gary Guo Cc: Boqun Feng Cc: Alice Ryhl Cc: Matthew Maurer Cc: Josh Poimboeuf Cc: Peter Zijlstra Cc: stable@vger.kernel.org Fixes: 3a2486cc1da5 ("kbuild: rust: provide an option to inline C helpers into Rust") Acked-by: Gary Guo Link: https://patch.msgid.link/20260816133233.197500-2-ojeda@kernel.org Signed-off-by: Miguel Ojeda --- drivers/gpu/host1x/fence.c | 154 +++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 154 insertions(+) create mode 100644 drivers/gpu/host1x/fence.c (limited to 'drivers/gpu/host1x/fence.c') diff --git a/drivers/gpu/host1x/fence.c b/drivers/gpu/host1x/fence.c new file mode 100644 index 000000000..b9a7d0bf9 --- /dev/null +++ b/drivers/gpu/host1x/fence.c @@ -0,0 +1,154 @@ +// SPDX-License-Identifier: GPL-2.0-only +/* + * Syncpoint dma_fence implementation + * + * Copyright (c) 2020, NVIDIA Corporation. + */ + +#include +#include +#include +#include +#include + +#include "fence.h" +#include "intr.h" +#include "syncpt.h" + +static const char *host1x_syncpt_fence_get_driver_name(struct dma_fence *f) +{ + return "host1x"; +} + +static const char *host1x_syncpt_fence_get_timeline_name(struct dma_fence *f) +{ + return "syncpoint"; +} + +static struct host1x_syncpt_fence *to_host1x_fence(struct dma_fence *f) +{ + return container_of(f, struct host1x_syncpt_fence, base); +} + +static bool host1x_syncpt_fence_enable_signaling(struct dma_fence *f) +{ + struct host1x_syncpt_fence *sf = to_host1x_fence(f); + + if (host1x_syncpt_is_expired(sf->sp, sf->threshold)) + return false; + + /* Reference for interrupt path. */ + dma_fence_get(f); + + /* + * The dma_fence framework requires the fence driver to keep a + * reference to any fences for which 'enable_signaling' has been + * called (and that have not been signalled). + * + * We cannot currently always guarantee that all fences get signalled + * or cancelled. As such, for such situations, set up a timeout, so + * that long-lasting fences will get reaped eventually. + */ + if (sf->timeout) { + /* Reference for timeout path. */ + dma_fence_get(f); + schedule_delayed_work(&sf->timeout_work, msecs_to_jiffies(30000)); + } + + host1x_intr_add_fence_locked(sf->sp->host, sf); + + /* + * The fence may get signalled at any time after the above call, + * so we need to initialize all state used by signalling + * before it. + */ + + return true; +} + +static const struct dma_fence_ops host1x_syncpt_fence_ops = { + .get_driver_name = host1x_syncpt_fence_get_driver_name, + .get_timeline_name = host1x_syncpt_fence_get_timeline_name, + .enable_signaling = host1x_syncpt_fence_enable_signaling, +}; + +void host1x_fence_signal(struct host1x_syncpt_fence *f) +{ + if (atomic_xchg(&f->signaling, 1)) { + /* + * Already on timeout path, but we removed the fence before + * timeout path could, so drop interrupt path reference. + */ + dma_fence_put(&f->base); + return; + } + + if (f->timeout && cancel_delayed_work(&f->timeout_work)) { + /* + * We know that the timeout path will not be entered. + * Safe to drop the timeout path's reference now. + */ + dma_fence_put(&f->base); + } + + dma_fence_signal_locked(&f->base); + dma_fence_put(&f->base); +} + +static void do_fence_timeout(struct work_struct *work) +{ + struct delayed_work *dwork = (struct delayed_work *)work; + struct host1x_syncpt_fence *f = + container_of(dwork, struct host1x_syncpt_fence, timeout_work); + + if (atomic_xchg(&f->signaling, 1)) { + /* Already on interrupt path, drop timeout path reference if any. */ + if (f->timeout) + dma_fence_put(&f->base); + return; + } + + if (host1x_intr_remove_fence(f->sp->host, f)) { + /* + * Managed to remove fence from queue, so it's safe to drop + * the interrupt path's reference. + */ + dma_fence_put(&f->base); + } + + dma_fence_set_error(&f->base, -ETIMEDOUT); + dma_fence_signal(&f->base); + if (f->timeout) + dma_fence_put(&f->base); +} + +struct dma_fence *host1x_fence_create(struct host1x_syncpt *sp, u32 threshold, + bool timeout) +{ + struct host1x_syncpt_fence *fence; + + fence = kzalloc_obj(*fence); + if (!fence) + return ERR_PTR(-ENOMEM); + + fence->sp = sp; + fence->threshold = threshold; + fence->timeout = timeout; + + dma_fence_init(&fence->base, &host1x_syncpt_fence_ops, &sp->fences.lock, + dma_fence_context_alloc(1), 0); + + INIT_DELAYED_WORK(&fence->timeout_work, do_fence_timeout); + + return &fence->base; +} +EXPORT_SYMBOL(host1x_fence_create); + +void host1x_fence_cancel(struct dma_fence *f) +{ + struct host1x_syncpt_fence *sf = to_host1x_fence(f); + + schedule_delayed_work(&sf->timeout_work, 0); + flush_delayed_work(&sf->timeout_work); +} +EXPORT_SYMBOL(host1x_fence_cancel); -- cgit v1.3.1