| Message ID | 1490880938-90331-1-git-send-email-agraf@suse.de |
|---|---|
| State | New |
| Headers | show |
On Do, 2017-03-30 at 15:35 +0200, Alexander Graf wrote: > When driving QEMU from the outside, we have basically no chance to > determine how quickly the guest OS picks up key events, so we usually > have to limit ourselves to very slow keyboard presses to make sure > the guest always has enough chance to pick them up. > > This patch adds trace events for when the guest polls for HID keyboard > events. That way we can be reasonably safely assume that the guest > handled that one key event and can type the next. Hmm, I'm wondering why you are tracing the event details? Input layer has tracepoints for this already, so why duplicate this here? Especially as you have to do quite some effort to store the data for tracing? For queue management (in automated testing I assume?) a simple "queue empty" trace point should do the trick, no? Or maybe a "queue has $i of $n slots filled atm" trace points, if you wanna know how much you can pass in without loosing events due to the queue being full. cheers, Gerd
On 30 March 2017 at 14:35, Alexander Graf <agraf@suse.de> wrote: > When driving QEMU from the outside, we have basically no chance to > determine how quickly the guest OS picks up key events, so we usually > have to limit ourselves to very slow keyboard presses to make sure > the guest always has enough chance to pick them up. > > This patch adds trace events for when the guest polls for HID keyboard > events. That way we can be reasonably safely assume that the guest > handled that one key event and can type the next. This feels like the wrong way to solve the problem to me. Maybe we should have something like the network and UART device models do where they can say "my queue is full, I cannot accept any more data from the outside world" ? thanks -- PMM
On 03/30/2017 04:00 PM, Gerd Hoffmann wrote: > On Do, 2017-03-30 at 15:35 +0200, Alexander Graf wrote: >> When driving QEMU from the outside, we have basically no chance to >> determine how quickly the guest OS picks up key events, so we usually >> have to limit ourselves to very slow keyboard presses to make sure >> the guest always has enough chance to pick them up. >> >> This patch adds trace events for when the guest polls for HID keyboard >> events. That way we can be reasonably safely assume that the guest >> handled that one key event and can type the next. > Hmm, I'm wondering why you are tracing the event details? Input layer > has tracepoints for this already, so why duplicate this here? > Especially as you have to do quite some effort to store the data for > tracing? Because we need to know when the events are off the queue, not pushed onto the queue. > For queue management (in automated testing I assume?) a simple "queue > empty" trace point should do the trick, no? Or maybe a "queue has $i of > $n slots filled atm" trace points, if you wanna know how much you can > pass in without loosing events due to the queue being full. That's a very good point. In fact, that would make things much much easier for the other side. If we just know "queue is drained", we know we can send new events. Alex
On 03/30/2017 04:10 PM, Peter Maydell wrote: > On 30 March 2017 at 14:35, Alexander Graf <agraf@suse.de> wrote: >> When driving QEMU from the outside, we have basically no chance to >> determine how quickly the guest OS picks up key events, so we usually >> have to limit ourselves to very slow keyboard presses to make sure >> the guest always has enough chance to pick them up. >> >> This patch adds trace events for when the guest polls for HID keyboard >> events. That way we can be reasonably safely assume that the guest >> handled that one key event and can type the next. > This feels like the wrong way to solve the problem to me. > Maybe we should have something like the network and > UART device models do where they can say "my queue is > full, I cannot accept any more data from the outside > world" ? That's really hard to realize unfortunately. If events come in from VNC, how do you stall new events from arriving? Would you stop reading the VNC stream? Would you allocate random buffers until you manage to finally push data onto the HID queue again? Keep in mind that HID events might deliberately not be interpreted. OVMF doesn't have to have a tablet driver usually for example, so your tablet queue will fill up as soon as you push any mouse movement information in. Alex
diff --git a/hw/input/hid.c b/hw/input/hid.c index fa9cc4c..b68f597 100644 --- a/hw/input/hid.c +++ b/hw/input/hid.c @@ -228,7 +228,7 @@ static void hid_keyboard_event(DeviceState *dev, QemuConsole *src, { HIDState *hs = (HIDState *)dev; int scancodes[3], i, count; - int slot; + int slot = -1; InputKeyEvent *key = evt->u.key.data; count = qemu_input_key_value_to_scancode(key->key, @@ -241,6 +241,13 @@ static void hid_keyboard_event(DeviceState *dev, QemuConsole *src, for (i = 0; i < count; i++) { slot = (hs->head + hs->n) & QUEUE_MASK; hs->n++; hs->kbd.keycodes[slot] = scancodes[i]; + hs->kbd.poll_trigger[slot] = NULL; + } + + if (slot != -1) { + hs->kbd.poll_trigger[slot] = g_new(HIDPollTrigger, 1); + hs->kbd.poll_trigger[slot]->key = *key->key; + hs->kbd.poll_trigger[slot]->down = key->down; } hs->event(hs); } @@ -256,6 +263,23 @@ static void hid_keyboard_process_keycode(HIDState *hs) slot = hs->head & QUEUE_MASK; QUEUE_INCR(hs->head); hs->n--; keycode = hs->kbd.keycodes[slot]; + /* Trace polled key events */ + if (hs->kbd.poll_trigger[slot]) { + HIDPollTrigger *t = hs->kbd.poll_trigger[slot]; + switch (t->key.type) { + case KEY_VALUE_KIND_NUMBER: + trace_hid_kbd_key_nr_polled(t->key.u.number.data, t->down); + break; + case KEY_VALUE_KIND_QCODE: + trace_hid_kbd_key_qt_polled(t->key.u.qcode.data, t->down); + break; + default: + break; + } + g_free(t); + hs->kbd.poll_trigger[slot] = NULL; + } + key = keycode & 0x7f; index = key | ((hs->kbd.modifiers & (1 << 8)) >> 1); hid_code = hid_usage_keys[index]; diff --git a/hw/input/trace-events b/hw/input/trace-events index f3bfbed..7a6b5ff 100644 --- a/hw/input/trace-events +++ b/hw/input/trace-events @@ -24,6 +24,8 @@ milkymist_softusb_pulse_irq(void) "Pulse IRQ" # hw/input/hid.c hid_kbd_queue_full(void) "queue full" +hid_kbd_key_nr_polled(uint64_t value, bool down) "val %#"PRIx64" down %d" +hid_kbd_key_qt_polled(uint64_t value, bool down) "val %#"PRIx64" down %d" # hw/input/virtio virtio_input_queue_full(void) "queue full" diff --git a/include/hw/input/hid.h b/include/hw/input/hid.h index 2127c7c..ff9541e 100644 --- a/include/hw/input/hid.h +++ b/include/hw/input/hid.h @@ -25,8 +25,14 @@ typedef struct HIDMouseState { int mouse_grabbed; } HIDMouseState; +typedef struct HIDPollTrigger { + KeyValue key; + bool down; +} HIDPollTrigger; + typedef struct HIDKeyboardState { uint32_t keycodes[QUEUE_LENGTH]; + HIDPollTrigger *poll_trigger[QUEUE_LENGTH]; uint16_t modifiers; uint8_t leds; uint8_t key[16];
When driving QEMU from the outside, we have basically no chance to determine how quickly the guest OS picks up key events, so we usually have to limit ourselves to very slow keyboard presses to make sure the guest always has enough chance to pick them up. This patch adds trace events for when the guest polls for HID keyboard events. That way we can be reasonably safely assume that the guest handled that one key event and can type the next. Signed-off-by: Alexander Graf <agraf@suse.de> --- hw/input/hid.c | 26 +++++++++++++++++++++++++- hw/input/trace-events | 2 ++ include/hw/input/hid.h | 6 ++++++ 3 files changed, 33 insertions(+), 1 deletion(-)