On 20/06/2024 19:03, Alexandre Ghiti wrote:
On 19/06/2024 05:40, Andy Chiu wrote:On Tue, Jun 18, 2024 at 9:40 PM Alexandre Ghiti <alexgh...@rivosinc.com> wrote:Hi Andy, On Tue, Jun 18, 2024 at 2:48 PM Andy Chiu <andy.c...@sifive.com> wrote:On Tue, Jun 18, 2024 at 8:02 PM Alexandre Ghiti <a...@ghiti.fr> wrote:Hi Conor, On 17/06/2024 15:23, Alexandre Ghiti wrote:Hi Conor, Sorry for the delay here. On 13/06/2024 09:48, Conor Dooley wrote:On Thu, May 23, 2024 at 01:51:34PM +0200, Alexandre Ghiti wrote:So, turns out that this patch is not sufficient. I've seen some moreCommit c97bf629963e ("riscv: Fix text patching when IPI are used")converted ftrace_make_nop() to use patch_insn_write() which does not emit any icache flush relying entirely on __ftrace_modify_code() to dothat.But we missed that ftrace_make_nop() was called very early directlywhenconverting mcount calls into nops (actually on riscv it converts 2Bnops emitted by the compiler into 4B nops).This caused crashes on multiple HW as reported by Conor and Björn since the booting core could have half-patched instructions in its icache which would trigger an illegal instruction trap: fix this by emitting alocal flush icache when early patching nops. Fixes: c97bf629963e ("riscv: Fix text patching when IPI are used") Signed-off-by: Alexandre Ghiti <alexgh...@rivosinc.com> --- arch/riscv/include/asm/cacheflush.h | 6 ++++++ arch/riscv/kernel/ftrace.c | 3 +++ 2 files changed, 9 insertions(+) diff --git a/arch/riscv/include/asm/cacheflush.h b/arch/riscv/include/asm/cacheflush.h index dd8d07146116..ce79c558a4c8 100644 --- a/arch/riscv/include/asm/cacheflush.h +++ b/arch/riscv/include/asm/cacheflush.h @@ -13,6 +13,12 @@ static inline void local_flush_icache_all(void) asm volatile ("fence.i" ::: "memory"); }+static inline void local_flush_icache_range(unsigned long start,+ unsigned long end) +{ + local_flush_icache_all(); +} + #define PG_dcache_clean PG_arch_1 static inline void flush_dcache_folio(struct folio *folio)diff --git a/arch/riscv/kernel/ftrace.c b/arch/riscv/kernel/ftrace.cindex 4f4987a6d83d..32e7c401dfb4 100644 --- a/arch/riscv/kernel/ftrace.c +++ b/arch/riscv/kernel/ftrace.c @@ -120,6 +120,9 @@ int ftrace_init_nop(struct module *mod, struct dyn_ftrace *rec) out = ftrace_make_nop(mod, rec, MCOUNT_ADDR); mutex_unlock(&text_mutex);crashes, seemingly due to initialising nops on this mutex_unlock().Palmer suggested moving the if (!mod) ... into the lock, which fixedthe problem for me.Ok, it makes sense, I completely missed that. I'll send a fix for thatshortly so that it can be merged in rc5. Thanks, AlexSo I digged a bit more and I'm afraid that the only way to make surethis issue does not happen elsewhere is to flush the icache right afterthe patching. We actually can't wait to batch the icache flush since along the way, we may call a function that has just been patched (the issue that you encountered here).Trying to provide my thoughts, please let me know if I missed anything. I think what Conor suggested is safe for init_nop, as it would be called only when there is only one core (booting) and at the loading stage of kernel modules. In the first case we just have to make sure there is no patchable entry before the core executes fence.i. The second case is unconditionally safe because there is no read-side of the race. It is a bit tricky for the generic (runtime) case of ftrace code patching, but that is not because of the batch fence.i maintenance. As long as there exists a patchable entry for the stopping thread, it is possible for them to execute on a partially patched instruction. A solution for this is again to prevent any patchable entry in the stop_machine's stopping thread. Another solution is to apply the atomic ftrace patching series which aims to get rid of the race.Yeah but my worry is that we can't make sure that functions called between the patching and the fence.i are not the ones that just get patched. At least today, patch_insn_write() etc should be marked asIIUC, the compiler does not generate a patchable entry for patch_insn_write, and so do all functions in patch.c. The entire file is not compiled with patchable entry because the flag is removed in riscv's Makefile. Please let me know if I misunderstand it.noinstr. But even then, we call core functions that could be patchable (I went through all and it *seems* we are safe *for now*, but this is very fragile).Yes, functions in the "else" clause, atomic_read() etc, are safe for now. However, it is impossible to fix as long as there exists a patchable entry along the way. This is the problem that we cannot patch 2 instructions with a concurrent read-side. On the other hand, the path of ftrace_modify_all_code may experience the batch fence.i maintenance issue, and can be fixed via a local fence.i. This is because the path is executed by only one core. There does not exist a concurrent write-side. As a result, we just need to make sure it leaves each patch-site with a local fence.i to make sure code is up-to-date.Then what do you think we should do to fix Conor's issue right now: should we mark the riscv specific functions as noinstr, cross fingers that the core functions are not (and don't become) patchable and wait for your atomic patching? Or should we flush as soon as possible as I proposed above (which to me fixes the issue but hurts performance)?Either way works for me. IMHO, the fix should include: 1. move fence.i before mutex_unlock in init_nop 2. do local fence.i in __ftrace_modify_call 3. make sure the else clause does not happen to have a patchable entryOK I have just checked again and I agree with you on the whole fix :) I may add some comments for 3 so that it is very clear.
And after another check, a few other callsites need to emit a sfence.i like ftrace_make_call() and ftrace_make_nop(), which in the end, amounts to adding a sfence.i right after the patching is done. So after a private discussion with Andy, we both agreed that the initial solution I posted was the safest. Admittedly, the impact on performance is unknown, so any input on this side would be appreciated.
@Conor: I'm sending this as a proper patch if you can give it a try so that we can merge that in rc6.
Thanks, Alex
Thanks Andy! AlexThanks, AndyThanks, AlexI don't know how much it will impact the performance but I guess it will.Unless someone has a better idea (I added Andy and Puranjay in cc), hereis the patch that implements this, can you give it a try? Thanks diff --git a/arch/riscv/kernel/ftrace.c b/arch/riscv/kernel/ftrace.c index 87cbd86576b2..4b95c574fd04 100644 --- a/arch/riscv/kernel/ftrace.c +++ b/arch/riscv/kernel/ftrace.c @@ -120,9 +120,6 @@ int ftrace_init_nop(struct module *mod, struct dyn_ftrace *rec) out = ftrace_make_nop(mod, rec, MCOUNT_ADDR); mutex_unlock(&text_mutex); - if (!mod) - local_flush_icache_range(rec->ip, rec->ip + MCOUNT_INSN_SIZE); - return out; } @@ -156,9 +153,9 @@ static int __ftrace_modify_code(void *data) } else {while (atomic_read(¶m->cpu_count) <= num_online_cpus())cpu_relax(); - } - local_flush_icache_all(); + local_flush_icache_all(); + } return 0; } diff --git a/arch/riscv/kernel/patch.c b/arch/riscv/kernel/patch.c index 4007563fb607..ab03732d06c4 100644 --- a/arch/riscv/kernel/patch.c +++ b/arch/riscv/kernel/patch.c@@ -89,6 +89,14 @@ static int __patch_insn_set(void *addr, u8 c, size_t len)memset(waddr, c, len); + /* + * We could have just patched a function that is about to be + * called so make sure we don't execute partially patched + * instructions by flushing the icache as soon as possible. + */ + local_flush_icache_range((unsigned long)waddr, + (unsigned long)waddr + len); + patch_unmap(FIX_TEXT_POKE0); if (across_pages) @@ -135,6 +143,14 @@ static int __patch_insn_write(void *addr, const void *insn, size_t len) ret = copy_to_kernel_nofault(waddr, insn, len); + /* + * We could have just patched a function that is about to be + * called so make sure we don't execute partially patched + * instructions by flushing the icache as soon as possible. + */ + local_flush_icache_range((unsigned long)waddr, + (unsigned long)waddr + len); + patch_unmap(FIX_TEXT_POKE0); if (across_pages)@@ -189,9 +205,6 @@ int patch_text_set_nosync(void *addr, u8 c, size_t len)ret = patch_insn_set(tp, c, len); - if (!ret)- flush_icache_range((uintptr_t)tp, (uintptr_t)tp + len);- return ret; } NOKPROBE_SYMBOL(patch_text_set_nosync);@@ -224,9 +237,6 @@ int patch_text_nosync(void *addr, const void *insns,size_t len) ret = patch_insn_write(tp, insns, len); - if (!ret)- flush_icache_range((uintptr_t) tp, (uintptr_t) tp + len);- return ret; } NOKPROBE_SYMBOL(patch_text_nosync); @@ -253,9 +263,9 @@ static int patch_text_cb(void *data) } else {while (atomic_read(&patch->cpu_count) <= num_online_cpus())cpu_relax(); - } - local_flush_icache_all(); + local_flush_icache_all(); + } return ret; }+ if (!mod)+ local_flush_icache_range(rec->ip, rec->ip + MCOUNT_INSN_SIZE);+ return out; } -- 2.39.2 _______________________________________________ linux-riscv mailing list linux-ri...@lists.infradead.org http://lists.infradead.org/mailman/listinfo/linux-riscv_______________________________________________ linux-riscv mailing list linux-ri...@lists.infradead.org http://lists.infradead.org/mailman/listinfo/linux-riscvCheers, Andy_______________________________________________ linux-riscv mailing list linux-ri...@lists.infradead.org http://lists.infradead.org/mailman/listinfo/linux-riscv
