tscdeadline_latency: Some fixes of hangs
v2 cover tscdeadline_latency: Some fixes of hangs
The previous version only contain patch 1 but when at it I noticed more ways to hang it... so let's fix them all up...
v2:
- use safe_halt() as suggested [Sean]
- added two more patches
Peter Xu (3):
tscdeadline_latency: Check condition first before loop
tscdeadline_latency: Limit size
tscdeadline_latency: Stop timer when reach max loop
x86/tscdeadline_latency.c | 27 +++++++++++++++++---------- 1 file changed, 17 insertions(+), 10 deletions(-)
Signature · 1 line
2.21.0
PATCH v2 1/3 tscdeadline_latency: Check condition first before loop
This patch fixes a tscdeadline_latency hang when specifying a very small breakmax value. It's easily reproduced on my host when set breakmax to e.g. 10 TSC clocks.
The problem is test_tsc_deadline_timer() can be very slow because we've got printf() in there. So when reach the main loop we might have already triggered the IRQ handler for multiple times and we might have triggered the hitmax condition which will turn IRQ off. Then with no IRQ that first HLT instruction can last forever.
Fix this by don't enable irq and use safe_halt() as suggested by Sean Christopherson.
--- x86/tscdeadline_latency.c | 5 +---- 1 file changed, 1 insertion(+), 4 deletions(-)
x86/tscdeadline_latency.c +1 -4
@@ -70,8 +70,6 @@ static void tsc_deadline_timer_isr(isr_regs_t *regs)static void start_tsc_deadline_timer (void ) {handle_irq (TSC_DEADLINE_TIMER_VECTOR , tsc_deadline_timer_isr); -irq_enable (); -wrmsr (MSR_IA32_TSCDEADLINE ,rdmsr (MSR_IA32_TSC )+delta); asmvolatile ("nop" ); }
@@ -116,10 +114,9 @@ int main(int argc, char **argv) breakmax = argc <=3 ?0 :atol (argv[3 ]);printf ("breakmax=%d\n" , breakmax);test_tsc_deadline_timer (); -irq_enable ();do { - asmvolatile ("hlt" ); +safe_halt (); }while (!hitmax && table_idx < size);for (i =0 ; i < table_idx; i++) {
Signature · 1 line
2.21.0
PATCH v2 2/3 tscdeadline_latency: Limit size
In reply to Peter XuWhen size>TABLE_SIZE it can hang as well (say, 10001), because when we get TABLE_SIZE IRQs we'll stop increasing table_idx, then we'll never be able to get out of the main loop.
--- x86/tscdeadline_latency.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-)
x86/tscdeadline_latency.c +5 -1
@@ -112,7 +112,11 @@ int main(int argc, char **argv) delta = argc <=1 ?200000 :atol (argv[1 ]); size = argc <=2 ?TABLE_SIZE :atol (argv[2 ]); breakmax = argc <=3 ?0 :atol (argv[3 ]); -printf ("breakmax=%d\n" , breakmax); + +if (size >=TABLE_SIZE ) + size =TABLE_SIZE ; +printf ("delta=%d, size=%d, breakmax=%d\n" , delta, size, breakmax); +test_tsc_deadline_timer (); do {
Signature · 1 line
2.21.0
PATCH v2 3/3 tscdeadline_latency: Stop timer when reach max loop
In reply to Peter XuThis fixes another possible hang of this test when delay is configured too small (e.g., 2000) so that the IRQ handler will trigger forever.
Let's stop the timer if we've collected enough data so that the program can quit always.
--- x86/tscdeadline_latency.c | 16 +++++++++++----- 1 file changed, 11 insertions(+), 5 deletions(-)
x86/tscdeadline_latency.c +11 -5
@@ -50,18 +50,24 @@ int breakmax = 0;static void tsc_deadline_timer_isr (isr_regs_t *regs) { -u64 now =rdtsc (); +u64 now =rdtsc (), delay; ++tdt_count; -if (table_idx <TABLE_SIZE && tdt_count >1 ) - table[table_idx++] = now - exptime; +if (tdt_count ==1 ) +/* Skip the 1st IRQ */ + goto setup_next; -if (breakmax && tdt_count >1 && (now - exptime) > breakmax) { - hitmax =1 ; + table[table_idx++] = delay = now - exptime; + +/* Stop if either we finished collection or hit max */ +if ((breakmax && delay > breakmax) || table_idx >=TABLE_SIZE ) { +if (breakmax) + hitmax =1 ;apic_write (APIC_EOI ,0 );return ; } +setup_next: exptime = now+delta;wrmsr (MSR_IA32_TSCDEADLINE , now+delta);apic_write (APIC_EOI ,0 );
Signature · 1 line
2.21.0