| Message ID | 1508845602-33508-1-git-send-email-maninder1.s@samsung.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Tue, Oct 24, 2017 at 05:16:42PM +0530, Maninder Singh wrote: > This patch stops unwinding backtrace in case of below 2 cases. > > (Issue observed while porting stackdepot on ARM, duplicate > entries created in stackdepot > reference patch for workaround in stackdepot:- > https://lkml.org/lkml/2017/10/11/353 > ). > > 1. If address belongs to irq/exception code, ignore it. > save_stack+0x40/0xec > __set_page_owner+0x2c/0x64 > .... > .... > __handle_domain_irq+0x9c/0x130 > gic_handle_irq+0x40/0x80 > __irq_usr+0x4c/0x60 > 0xb6507818 > ^^^^^^^ However, we _do_ want to trace through an IRQ taken in SVC mode, but you completely remove all that code. So, not taking this patch, sorry.
Hi Russell, >On Tue, Oct 24, 2017 at 05:16:42PM +0530, Maninder Singh wrote: >> This patch stops unwinding backtrace in case of below 2 cases. >> >> (Issue observed while porting stackdepot on ARM, duplicate >> entries created in stackdepot >> reference patch for workaround in stackdepot:- >> https://lkml.org/lkml/2017/10/11/353 >> ). >> >> 1. If address belongs to irq/exception code, ignore it. >> save_stack+0x40/0xec >> __set_page_owner+0x2c/0x64 >> .... >> .... >> __handle_domain_irq+0x9c/0x130 >> gic_handle_irq+0x40/0x80 >> __irq_usr+0x4c/0x60 >> 0xb6507818 >> ^^^^^^^ > >However, we _do_ want to trace through an IRQ taken in SVC mode, but you >completely remove all that code. So, not taking this patch, sorry. OK. Thanks. But can we add some marker to distinguish before and after interrupt context frames. so that we can remove interrupt that frames from stackdepot, because due to interrupt allocations, we end up with so many stackdepot entries and which results in more memory consumption by stackdepot. Something like below: __set_page_owner+0x2c/0x64 ... ... __handle_domain_irq+0x9c/0x130 gic_handle_irq+0x40/0x80 __irq_usr+0x4c/0x60 0xFFFFFFFF (marker) 0xb6507818 .... Thus in stackdepot we can save before 0xFFFFFFF. or can you provide any suggestion for the same if we can try for?
On Wed, Nov 15, 2017 at 11:09:22AM +0000, Maninder Singh wrote: > Hi Russell, > > >On Tue, Oct 24, 2017 at 05:16:42PM +0530, Maninder Singh wrote: > >> This patch stops unwinding backtrace in case of below 2 cases. > >> > >> (Issue observed while porting stackdepot on ARM, duplicate > >> entries created in stackdepot > >> reference patch for workaround in stackdepot:- > >> https://lkml.org/lkml/2017/10/11/353 > >> ). > >> > >> 1. If address belongs to irq/exception code, ignore it. > >> save_stack+0x40/0xec > >> __set_page_owner+0x2c/0x64 > >> .... > >> .... > >> __handle_domain_irq+0x9c/0x130 > >> gic_handle_irq+0x40/0x80 > >> __irq_usr+0x4c/0x60 > >> 0xb6507818 > >> ^^^^^^^ > > > >However, we _do_ want to trace through an IRQ taken in SVC mode, but you > >completely remove all that code. So, not taking this patch, sorry. > > OK. Thanks. > > But can we add some marker to distinguish before and after interrupt context frames. > so that we can remove interrupt that frames from stackdepot, because due to interrupt allocations, > we end up with so many stackdepot entries and which results in more memory consumption by stackdepot. Notice that I said "taken in SVC mode" not "taken in user mode". Your example below is for a user mode interrupt. > Something like below: > __set_page_owner+0x2c/0x64 > ... > ... > __handle_domain_irq+0x9c/0x130 > gic_handle_irq+0x40/0x80 > __irq_usr+0x4c/0x60 > 0xFFFFFFFF (marker) > 0xb6507818 > .... > > > Thus in stackdepot we can save before 0xFFFFFFF. > or can you provide any suggestion for the same if we can try for?
diff --git a/arch/arm/kernel/stacktrace.c b/arch/arm/kernel/stacktrace.c index 3a2fa20..80e953e 100644 --- a/arch/arm/kernel/stacktrace.c +++ b/arch/arm/kernel/stacktrace.c @@ -82,6 +82,9 @@ static int save_trace(struct stackframe *frame, void *d) return 0; } + if (!__kernel_text_address(addr)) + return 1; + trace->entries[trace->nr_entries++] = addr; if (trace->nr_entries >= trace->max_entries) @@ -98,12 +101,8 @@ static int save_trace(struct stackframe *frame, void *d) data->last_pc = frame->pc; if (!in_exception_text(addr)) return 0; - - regs = (struct pt_regs *)frame->sp; - - trace->entries[trace->nr_entries++] = regs->ARM_pc; - - return trace->nr_entries >= trace->max_entries; + else + return 1; } /* This must be noinline to so that our skip calculation works correctly */