diff --git a/artiq/firmware/ksupport/api.rs b/artiq/firmware/ksupport/api.rs
index 1cf3ce884..55e43edd1 100644
--- a/artiq/firmware/ksupport/api.rs
+++ b/artiq/firmware/ksupport/api.rs
@@ -112,11 +112,8 @@ static mut API: &'static [(&'static str, *const ())] = &[
     api!(dma_retrieve = ::dma_retrieve),
     api!(dma_playback = ::dma_playback),
 
-    api!(drtio_get_channel_state = ::rtio::drtio_dbg::get_channel_state),
-    api!(drtio_reset_channel_state = ::rtio::drtio_dbg::reset_channel_state),
-    api!(drtio_get_fifo_space = ::rtio::drtio_dbg::get_fifo_space),
     api!(drtio_get_packet_counts = ::rtio::drtio_dbg::get_packet_counts),
-    api!(drtio_get_fifo_space_req_count = ::rtio::drtio_dbg::get_fifo_space_req_count),
+    api!(drtio_get_buffer_space_req_count = ::rtio::drtio_dbg::get_buffer_space_req_count),
 
     api!(i2c_start = ::nrt_bus::i2c::start),
     api!(i2c_restart = ::nrt_bus::i2c::restart),
diff --git a/artiq/firmware/ksupport/rtio.rs b/artiq/firmware/ksupport/rtio.rs
index ebc86b9f5..e03b2ea4c 100644
--- a/artiq/firmware/ksupport/rtio.rs
+++ b/artiq/firmware/ksupport/rtio.rs
@@ -161,24 +161,6 @@ pub mod drtio_dbg {
     use ::recv;
     use kernel_proto::*;
 
-
-    #[repr(C)]
-    pub struct ChannelState(i32, i64);
-
-    pub extern fn get_channel_state(channel: i32) -> ChannelState {
-        send(&DrtioChannelStateRequest { channel: channel as u32 });
-        recv!(&DrtioChannelStateReply { fifo_space, last_timestamp }
-              => ChannelState(fifo_space as i32, last_timestamp as i64))
-    }
-
-    pub extern fn reset_channel_state(channel: i32) {
-        send(&DrtioResetChannelStateRequest { channel: channel as u32 })
-    }
-
-    pub extern fn get_fifo_space(channel: i32) {
-        send(&DrtioGetFifoSpaceRequest { channel: channel as u32 })
-    }
-
     #[repr(C)]
     pub struct PacketCounts(i32, i32);
 
@@ -188,9 +170,9 @@ pub mod drtio_dbg {
               => PacketCounts(tx_cnt as i32, rx_cnt as i32))
     }
 
-    pub extern fn get_fifo_space_req_count(linkno: i32) -> i32 {
-        send(&DrtioFifoSpaceReqCountRequest { linkno: linkno as u8 });
-        recv!(&DrtioFifoSpaceReqCountReply { cnt }
+    pub extern fn get_buffer_space_req_count(linkno: i32) -> i32 {
+        send(&DrtioBufferSpaceReqCountRequest { linkno: linkno as u8 });
+        recv!(&DrtioBufferSpaceReqCountReply { cnt }
               => cnt as i32)
     }
 }
diff --git a/artiq/firmware/libdrtioaux/lib.rs b/artiq/firmware/libdrtioaux/lib.rs
index ece63f37e..51ff8faa8 100644
--- a/artiq/firmware/libdrtioaux/lib.rs
+++ b/artiq/firmware/libdrtioaux/lib.rs
@@ -21,6 +21,7 @@ pub enum Packet {
 
     RtioErrorRequest,
     RtioNoErrorReply,
+    RtioErrorSequenceErrorReply,
     RtioErrorCollisionReply,
     RtioErrorBusyReply,
 
@@ -55,8 +56,9 @@ impl Packet {
 
             0x20 => Packet::RtioErrorRequest,
             0x21 => Packet::RtioNoErrorReply,
-            0x22 => Packet::RtioErrorCollisionReply,
-            0x23 => Packet::RtioErrorBusyReply,
+            0x22 => Packet::RtioErrorSequenceErrorReply,
+            0x23 => Packet::RtioErrorCollisionReply,
+            0x24 => Packet::RtioErrorBusyReply,
 
             0x40 => Packet::MonitorRequest {
                 channel: read_u16(reader)?,
@@ -145,8 +147,9 @@ impl Packet {
 
             Packet::RtioErrorRequest => write_u8(writer, 0x20)?,
             Packet::RtioNoErrorReply => write_u8(writer, 0x21)?,
-            Packet::RtioErrorCollisionReply => write_u8(writer, 0x22)?,
-            Packet::RtioErrorBusyReply => write_u8(writer, 0x23)?,
+            Packet::RtioErrorSequenceErrorReply => write_u8(writer, 0x22)?,
+            Packet::RtioErrorCollisionReply => write_u8(writer, 0x23)?,
+            Packet::RtioErrorBusyReply => write_u8(writer, 0x24)?,
 
             Packet::MonitorRequest { channel, probe } => {
                 write_u8(writer, 0x40)?;
@@ -278,7 +281,7 @@ pub mod hw {
         unsafe {
             if (board::csr::DRTIO[linkidx].aux_rx_present_read)() == 1 {
                 let length = (board::csr::DRTIO[linkidx].aux_rx_length_read)();
-                let base = board::mem::DRTIO_AUX[linkidx].base + board::mem::DRTIO_AUX[linkidx].size/2; 
+                let base = board::mem::DRTIO_AUX[linkidx].base + board::mem::DRTIO_AUX[linkidx].size/2;
                 let sl = slice::from_raw_parts(base as *mut u8, length as usize);
                 Some(RxBuffer(linkno, sl))
             } else {
diff --git a/artiq/firmware/libproto/kernel_proto.rs b/artiq/firmware/libproto/kernel_proto.rs
index c59306cb8..06e5473d0 100644
--- a/artiq/firmware/libproto/kernel_proto.rs
+++ b/artiq/firmware/libproto/kernel_proto.rs
@@ -46,14 +46,10 @@ pub enum Message<'a> {
         duration: u64
     },
 
-    DrtioChannelStateRequest { channel: u32 },
-    DrtioChannelStateReply { fifo_space: u16, last_timestamp: u64 },
-    DrtioResetChannelStateRequest { channel: u32 },
-    DrtioGetFifoSpaceRequest { channel: u32 },
     DrtioPacketCountRequest { linkno: u8 },
     DrtioPacketCountReply { tx_cnt: u32, rx_cnt: u32 },
-    DrtioFifoSpaceReqCountRequest { linkno: u8 },
-    DrtioFifoSpaceReqCountReply { cnt: u32 },
+    DrtioBufferSpaceReqCountRequest { linkno: u8 },
+    DrtioBufferSpaceReqCountReply { cnt: u32 },
 
     RunFinished,
     RunException {
diff --git a/artiq/firmware/runtime/kern_hwreq.rs b/artiq/firmware/runtime/kern_hwreq.rs
index a9c67178b..402a8590c 100644
--- a/artiq/firmware/runtime/kern_hwreq.rs
+++ b/artiq/firmware/runtime/kern_hwreq.rs
@@ -323,26 +323,13 @@ pub fn process_kern_hwreq(io: &Io, request: &kern::Message) -> io::Result<bool>
             kern_acknowledge()
         }
 
-        &kern::DrtioChannelStateRequest { channel } => {
-            let (fifo_space, last_timestamp) = rtio_mgt::drtio_dbg::get_channel_state(channel);
-            kern_send(io, &kern::DrtioChannelStateReply { fifo_space: fifo_space,
-                                                          last_timestamp: last_timestamp })
-        }
-        &kern::DrtioResetChannelStateRequest { channel } => {
-            rtio_mgt::drtio_dbg::reset_channel_state(channel);
-            kern_acknowledge()
-        }
-        &kern::DrtioGetFifoSpaceRequest { channel } => {
-            rtio_mgt::drtio_dbg::get_fifo_space(channel);
-            kern_acknowledge()
-        }
         &kern::DrtioPacketCountRequest { linkno } => {
             let (tx_cnt, rx_cnt) = rtio_mgt::drtio_dbg::get_packet_counts(linkno);
             kern_send(io, &kern::DrtioPacketCountReply { tx_cnt: tx_cnt, rx_cnt: rx_cnt })
         }
-        &kern::DrtioFifoSpaceReqCountRequest { linkno } => {
-            let cnt = rtio_mgt::drtio_dbg::get_fifo_space_req_count(linkno);
-            kern_send(io, &kern::DrtioFifoSpaceReqCountReply { cnt: cnt })
+        &kern::DrtioBufferSpaceReqCountRequest { linkno } => {
+            let cnt = rtio_mgt::drtio_dbg::get_buffer_space_req_count(linkno);
+            kern_send(io, &kern::DrtioBufferSpaceReqCountReply { cnt: cnt })
         }
 
         &kern::I2cStartRequest { busno } => {
diff --git a/artiq/firmware/runtime/rtio_mgt.rs b/artiq/firmware/runtime/rtio_mgt.rs
index e7578d92c..c411295f4 100644
--- a/artiq/firmware/runtime/rtio_mgt.rs
+++ b/artiq/firmware/runtime/rtio_mgt.rs
@@ -73,21 +73,11 @@ pub mod drtio {
         unsafe {
             (csr::DRTIO[linkidx].reset_write)(1);
             while (csr::DRTIO[linkidx].o_wait_read)() == 1 {}
-        }
-        // TODO: determine actual number of remote FIFOs
-        for channel in 0..16 {
-            unsafe {
-                (csr::DRTIO[linkidx].chan_sel_override_write)(channel);
-                (csr::DRTIO[linkidx].chan_sel_override_en_write)(1);
 
-                (csr::DRTIO[linkidx].o_reset_channel_status_write)(1);
-                (csr::DRTIO[linkidx].o_get_fifo_space_write)(1);
-                while (csr::DRTIO[linkidx].o_wait_read)() == 1 {}
-                info!("[LINK#{}] FIFO space on channel {} is {}",
-                    linkno, channel, (csr::DRTIO[linkidx].o_dbg_fifo_space_read)());
-
-                (csr::DRTIO[linkidx].chan_sel_override_en_write)(0);
-            }
+            (csr::DRTIO[linkidx].o_get_buffer_space_write)(1);
+            while (csr::DRTIO[linkidx].o_wait_read)() == 1 {}
+            info!("[LINK#{}] buffer space is {}",
+                linkno, (csr::DRTIO[linkidx].o_dbg_buffer_space_read)());
         }
     }
 
@@ -130,7 +120,7 @@ pub mod drtio {
                 error!("[LINK#{}] received truncated packet", linkno);
             }
             if errors & 4 != 0 {
-                error!("[LINK#{}] timeout attempting to get remote FIFO space", linkno);
+                error!("[LINK#{}] timeout attempting to get remote buffer space", linkno);
             }
         }
     }
@@ -139,6 +129,8 @@ pub mod drtio {
         drtioaux::hw::send_link(linkno, &drtioaux::Packet::RtioErrorRequest).unwrap();
         match drtioaux::hw::recv_timeout_link(linkno, None) {
             Ok(drtioaux::Packet::RtioNoErrorReply) => (),
+            Ok(drtioaux::Packet::RtioErrorSequenceErrorReply) =>
+                error!("[LINK#{}] RTIO sequence error", linkno),
             Ok(drtioaux::Packet::RtioErrorCollisionReply) =>
                 error!("[LINK#{}] RTIO collision", linkno),
             Ok(drtioaux::Packet::RtioErrorBusyReply) =>
@@ -251,42 +243,6 @@ pub fn init_core() {
 pub mod drtio_dbg {
     use board::csr;
 
-    // TODO: routing
-    pub fn get_channel_state(channel: u32) -> (u16, u64) {
-        let linkno = ((channel >> 16) - 1) as usize;
-        let node_channel = channel as u16;
-        unsafe {
-            (csr::DRTIO[linkno].chan_sel_override_write)(node_channel as u16);
-            (csr::DRTIO[linkno].chan_sel_override_en_write)(1);
-            let fifo_space = (csr::DRTIO[linkno].o_dbg_fifo_space_read)();
-            let last_timestamp = (csr::DRTIO[linkno].o_dbg_last_timestamp_read)();
-            (csr::DRTIO[linkno].chan_sel_override_en_write)(0);
-            (fifo_space, last_timestamp)
-        }
-    }
-
-    pub fn reset_channel_state(channel: u32) {
-        let linkno = ((channel >> 16) - 1) as usize;
-        let node_channel = channel as u16;
-        unsafe {
-            (csr::DRTIO[linkno].chan_sel_override_write)(node_channel);
-            (csr::DRTIO[linkno].chan_sel_override_en_write)(1);
-            (csr::DRTIO[linkno].o_reset_channel_status_write)(1);
-            (csr::DRTIO[linkno].chan_sel_override_en_write)(0);
-        }
-    }
-
-    pub fn get_fifo_space(channel: u32) {
-        let linkno = ((channel >> 16) - 1) as usize;
-        let node_channel = channel as u16;
-        unsafe {
-            (csr::DRTIO[linkno].chan_sel_override_write)(node_channel);
-            (csr::DRTIO[linkno].chan_sel_override_en_write)(1);
-            (csr::DRTIO[linkno].o_get_fifo_space_write)(1);
-            (csr::DRTIO[linkno].chan_sel_override_en_write)(0);
-        }
-    }
-
     pub fn get_packet_counts(linkno: u8) -> (u32, u32) {
         let linkno = linkno as usize;
         unsafe {
@@ -296,23 +252,17 @@ pub mod drtio_dbg {
         }
     }
 
-    pub fn get_fifo_space_req_count(linkno: u8) -> u32 {
+    pub fn get_buffer_space_req_count(linkno: u8) -> u32 {
         let linkno = linkno as usize;
         unsafe {
-            (csr::DRTIO[linkno].o_dbg_fifo_space_req_cnt_read)()
+            (csr::DRTIO[linkno].o_dbg_buffer_space_req_cnt_read)()
         }
     }
 }
 
 #[cfg(not(has_drtio))]
 pub mod drtio_dbg {
-    pub fn get_channel_state(_channel: u32) -> (u16, u64) { (0, 0) }
-
-    pub fn reset_channel_state(_channel: u32) {}
-
-    pub fn get_fifo_space(_channel: u32) {}
-
     pub fn get_packet_counts(_linkno: u8) -> (u32, u32) { (0, 0) }
 
-    pub fn get_fifo_space_req_count(_linkno: u8) -> u32 { 0 }
+    pub fn get_buffer_space_req_count(_linkno: u8) -> u32 { 0 }
 }
diff --git a/artiq/firmware/satman/lib.rs b/artiq/firmware/satman/lib.rs
index 5ffb0d6a2..4df701b4d 100644
--- a/artiq/firmware/satman/lib.rs
+++ b/artiq/firmware/satman/lib.rs
@@ -26,13 +26,19 @@ fn process_aux_packet(p: &drtioaux::Packet) {
                 unsafe {
                     (board::csr::DRTIO[0].rtio_error_write)(1);
                 }
-                drtioaux::hw::send_link(0, &drtioaux::Packet::RtioErrorCollisionReply).unwrap();
+                drtioaux::hw::send_link(0, &drtioaux::Packet::RtioErrorSequenceErrorReply).unwrap();
             } else if errors & 2 != 0 {
                 unsafe {
                     (board::csr::DRTIO[0].rtio_error_write)(2);
                 }
+                drtioaux::hw::send_link(0, &drtioaux::Packet::RtioErrorCollisionReply).unwrap();
+            } else if errors & 4 != 0 {
+                unsafe {
+                    (board::csr::DRTIO[0].rtio_error_write)(4);
+                }
                 drtioaux::hw::send_link(0, &drtioaux::Packet::RtioErrorBusyReply).unwrap();
-            } else {
+            }
+            else {
                 drtioaux::hw::send_link(0, &drtioaux::Packet::RtioNoErrorReply).unwrap();
             }
         }
@@ -153,9 +159,6 @@ fn process_errors() {
     if errors & 8 != 0 {
         error!("write overflow");
     }
-    if errors & 16 != 0 {
-        error!("write sequence error");
-    }
 }
 
 
diff --git a/artiq/gateware/drtio/core.py b/artiq/gateware/drtio/core.py
index c3da1cb51..f2ecce9e4 100644
--- a/artiq/gateware/drtio/core.py
+++ b/artiq/gateware/drtio/core.py
@@ -3,10 +3,12 @@ from types import SimpleNamespace
 from migen import *
 from migen.genlib.cdc import ElasticBuffer
 
+from artiq.gateware.rtio.sed.core import *
+from artiq.gateware.rtio.input_collector import *
 from artiq.gateware.drtio import (link_layer, aux_controller,
-                                  rt_packet_satellite, rt_ios_satellite,
+                                  rt_packet_satellite, rt_iobuffer_satellite,
                                   rt_errors_satellite,
-                                  rt_packet_master, rt_controller_master) 
+                                  rt_packet_master, rt_controller_master)
 
 
 class ChannelInterface:
@@ -49,7 +51,8 @@ class GenericRXSynchronizer(Module):
 
 
 class DRTIOSatellite(Module):
-    def __init__(self, chanif, channels, rx_synchronizer=None, fine_ts_width=3, full_ts_width=63):
+    def __init__(self, chanif, channels, rx_synchronizer=None, fine_ts_width=3,
+                 lane_count=8, fifo_depth=128):
         if rx_synchronizer is None:
             rx_synchronizer = GenericRXSynchronizer()
             self.submodules += rx_synchronizer
@@ -77,11 +80,29 @@ class DRTIOSatellite(Module):
         self.submodules.rt_packet = ClockDomainsRenamer("rtio")(
             rt_packet_satellite.RTPacketSatellite(link_layer_sync))
 
-        self.submodules.ios = rt_ios_satellite.IOS(
-            self.rt_packet, channels, fine_ts_width, full_ts_width)
+        coarse_ts = Signal(64 - fine_ts_width)
+        self.sync.rtio += \
+            If(self.rt_packet.tsc_load,
+                coarse_ts.eq(self.rt_packet.tsc_load_value)
+            ).Else(
+                coarse_ts.eq(coarse_ts + 1)
+            )
+        self.comb += self.rt_packet.cri.counter.eq(coarse_ts << fine_ts_width)
+
+        self.submodules.outputs = ClockDomainsRenamer("rio")(
+            SED(channels, fine_ts_width, "sync",
+                lane_count=lane_count, fifo_depth=fifo_depth,
+                report_buffer_space=True, interface=self.rt_packet.cri))
+        self.comb += self.outputs.coarse_timestamp.eq(coarse_ts)
+        self.sync += self.outputs.minimum_coarse_timestamp.eq(coarse_ts + 16)
+
+        self.submodules.inputs = ClockDomainsRenamer("rio")(
+            InputCollector(channels, fine_ts_width, "sync",
+                           interface=self.rt_packet.cri))
+        self.comb += self.inputs.coarse_timestamp.eq(coarse_ts)
 
         self.submodules.rt_errors = rt_errors_satellite.RTErrorsSatellite(
-            self.rt_packet, self.ios)
+            self.rt_packet, self.outputs)
 
         self.clock_domains.cd_rio = ClockDomain()
         self.clock_domains.cd_rio_phy = ClockDomain()
diff --git a/artiq/gateware/drtio/rt_controller_master.py b/artiq/gateware/drtio/rt_controller_master.py
index b7ecfd5a3..12683db01 100644
--- a/artiq/gateware/drtio/rt_controller_master.py
+++ b/artiq/gateware/drtio/rt_controller_master.py
@@ -15,9 +15,6 @@ class _CSRs(AutoCSR):
     def __init__(self):
         self.protocol_error = CSR(3)
 
-        self.chan_sel_override = CSRStorage(16)
-        self.chan_sel_override_en = CSRStorage()
-
         self.tsc_correction = CSRStorage(64)
         self.set_time = CSR()
         self.underflow_margin = CSRStorage(16, reset=200)
@@ -25,11 +22,9 @@ class _CSRs(AutoCSR):
         self.reset = CSR()
         self.reset_phy = CSR()
 
-        self.o_get_fifo_space = CSR()
-        self.o_dbg_fifo_space = CSRStatus(16)
-        self.o_dbg_last_timestamp = CSRStatus(64)
-        self.o_dbg_fifo_space_req_cnt = CSRStatus(32)
-        self.o_reset_channel_status = CSR()
+        self.o_get_buffer_space = CSR()
+        self.o_dbg_buffer_space = CSRStatus(16)
+        self.o_dbg_buffer_space_req_cnt = CSRStatus(32)
         self.o_wait = CSRStatus()
 
 
@@ -59,27 +54,20 @@ class RTController(Module):
         # protocol errors
         err_unknown_packet_type = Signal()
         err_packet_truncated = Signal()
-        signal_fifo_space_timeout = Signal()
-        err_fifo_space_timeout = Signal()
+        signal_buffer_space_timeout = Signal()
+        err_buffer_space_timeout = Signal()
         self.sync.sys_with_rst += [
             If(self.csrs.protocol_error.re,
                 If(self.csrs.protocol_error.r[0], err_unknown_packet_type.eq(0)),
                 If(self.csrs.protocol_error.r[1], err_packet_truncated.eq(0)),
-                If(self.csrs.protocol_error.r[2], err_fifo_space_timeout.eq(0))
+                If(self.csrs.protocol_error.r[2], err_buffer_space_timeout.eq(0))
             ),
             If(rt_packet.err_unknown_packet_type, err_unknown_packet_type.eq(1)),
             If(rt_packet.err_packet_truncated, err_packet_truncated.eq(1)),
-            If(signal_fifo_space_timeout, err_fifo_space_timeout.eq(1))
+            If(signal_buffer_space_timeout, err_buffer_space_timeout.eq(1))
         ]
         self.comb += self.csrs.protocol_error.w.eq(
-            Cat(err_unknown_packet_type, err_packet_truncated, err_fifo_space_timeout))
-
-        # channel selection
-        chan_sel = Signal(16)
-        self.comb += chan_sel.eq(
-            Mux(self.csrs.chan_sel_override_en.storage,
-                self.csrs.chan_sel_override.storage,
-                self.cri.chan_sel[:16]))
+            Cat(err_unknown_packet_type, err_packet_truncated, err_buffer_space_timeout))
 
         # master RTIO counter and counter synchronization
         self.submodules.counter = RTIOCounter(64-fine_ts_width)
@@ -122,26 +110,16 @@ class RTController(Module):
         self.comb += self.cd_rtio_with_rst.clk.eq(ClockSignal("rtio"))
         self.specials += AsyncResetSynchronizer(self.cd_rtio_with_rst, local_reset)
 
-        # remote channel status cache
-        fifo_spaces_mem = Memory(16, channel_count)
-        fifo_spaces = fifo_spaces_mem.get_port(write_capable=True)
-        self.specials += fifo_spaces_mem, fifo_spaces
-        last_timestamps_mem = Memory(64, channel_count)
-        last_timestamps = last_timestamps_mem.get_port(write_capable=True)
-        self.specials += last_timestamps_mem, last_timestamps
-
         # common packet fields
-        rt_packet_fifo_request = Signal()
+        chan_sel = self.cri.chan_sel[:16]
+        rt_packet_buffer_request = Signal()
         rt_packet_read_request = Signal()
         self.comb += [
-            fifo_spaces.adr.eq(chan_sel),
-            last_timestamps.adr.eq(chan_sel),
-            last_timestamps.dat_w.eq(self.cri.timestamp),
             rt_packet.sr_channel.eq(chan_sel),
             rt_packet.sr_address.eq(self.cri.o_address),
             rt_packet.sr_data.eq(self.cri.o_data),
             rt_packet.sr_timestamp.eq(self.cri.timestamp),
-            If(rt_packet_fifo_request,
+            If(rt_packet_buffer_request,
                 rt_packet.sr_notwrite.eq(1),
                 rt_packet.sr_address.eq(0)
             ),
@@ -154,30 +132,28 @@ class RTController(Module):
         # output status
         o_status_wait = Signal()
         o_status_underflow = Signal()
-        o_status_sequence_error = Signal()
         self.comb += [
             self.cri.o_status.eq(Cat(
-                o_status_wait, o_status_underflow, o_status_sequence_error)),
+                o_status_wait, o_status_underflow)),
             self.csrs.o_wait.status.eq(o_status_wait)
         ]
-        o_sequence_error_set = Signal()
         o_underflow_set = Signal()
         self.sync.sys_with_rst += [
             If(self.cri.cmd == cri.commands["write"],
-                o_status_underflow.eq(0),
-                o_status_sequence_error.eq(0),
+                o_status_underflow.eq(0)
             ),
-            If(o_underflow_set, o_status_underflow.eq(1)),
-            If(o_sequence_error_set, o_status_sequence_error.eq(1))
+            If(o_underflow_set, o_status_underflow.eq(1))
         ]
 
         timeout_counter = WaitTimer(8191)
         self.submodules += timeout_counter
 
-        cond_sequence_error = self.cri.timestamp < last_timestamps.dat_r
-        cond_underflow = ((self.cri.timestamp[fine_ts_width:]
+        cond_underflow = Signal()
+        self.comb += cond_underflow.eq((self.cri.timestamp[fine_ts_width:]
                            - self.csrs.underflow_margin.storage[fine_ts_width:]) < self.counter.value_sys)
 
+        buffer_space = Signal(16)
+
         # input status
         i_status_wait_event = Signal()
         i_status_overflow = Signal()
@@ -208,56 +184,51 @@ class RTController(Module):
 
         fsm.act("IDLE",
             If(self.cri.cmd == cri.commands["write"],
-                If(cond_sequence_error,
-                    o_sequence_error_set.eq(1)
-                ).Elif(cond_underflow,
+                If(cond_underflow,
                     o_underflow_set.eq(1)
                 ).Else(
                     NextState("WRITE")
                 )
             ),
             If(self.cri.cmd == cri.commands["read"], NextState("READ")),
-            If(self.csrs.o_get_fifo_space.re, NextState("GET_FIFO_SPACE"))
+            If(self.csrs.o_get_buffer_space.re, NextState("GET_BUFFER_SPACE"))
         )
         fsm.act("WRITE",
             o_status_wait.eq(1),
             rt_packet.sr_stb.eq(1),
             If(rt_packet.sr_ack,
-                fifo_spaces.we.eq(1),
-                fifo_spaces.dat_w.eq(fifo_spaces.dat_r - 1),
-                last_timestamps.we.eq(1),
-                If(fifo_spaces.dat_r <= 1,
-                    NextState("GET_FIFO_SPACE")
+                NextValue(buffer_space, buffer_space - 1),
+                If(buffer_space <= 1,
+                    NextState("GET_BUFFER_SPACE")
                 ).Else(
                     NextState("IDLE")
                 )
             )
         )
-        fsm.act("GET_FIFO_SPACE",
+        fsm.act("GET_BUFFER_SPACE",
             o_status_wait.eq(1),
-            rt_packet.fifo_space_not_ack.eq(1),
-            rt_packet_fifo_request.eq(1),
+            rt_packet.buffer_space_not_ack.eq(1),
+            rt_packet_buffer_request.eq(1),
             rt_packet.sr_stb.eq(1),
             If(rt_packet.sr_ack,
-                NextState("GET_FIFO_SPACE_REPLY")
+                NextState("GET_BUFFER_SPACE_REPLY")
             )
         )
-        fsm.act("GET_FIFO_SPACE_REPLY",
+        fsm.act("GET_BUFFER_SPACE_REPLY",
             o_status_wait.eq(1),
-            fifo_spaces.dat_w.eq(rt_packet.fifo_space),
-            fifo_spaces.we.eq(1),
-            rt_packet.fifo_space_not_ack.eq(1),
-            If(rt_packet.fifo_space_not,
-                If(rt_packet.fifo_space != 0,
+            NextValue(buffer_space, rt_packet.buffer_space),
+            rt_packet.buffer_space_not_ack.eq(1),
+            If(rt_packet.buffer_space_not,
+               If(rt_packet.buffer_space != 0,
                     NextState("IDLE")
                 ).Else(
-                    NextState("GET_FIFO_SPACE")
+                    NextState("GET_BUFFER_SPACE")
                 )
             ),
             timeout_counter.wait.eq(1),
             If(timeout_counter.done,
-                signal_fifo_space_timeout.eq(1),
-                NextState("GET_FIFO_SPACE")
+                signal_buffer_space_timeout.eq(1),
+                NextState("GET_BUFFER_SPACE")
             )
         )
         fsm.act("READ",
@@ -278,21 +249,12 @@ class RTController(Module):
             )
         )
 
-        # channel state access
-        self.comb += [
-            self.csrs.o_dbg_fifo_space.status.eq(fifo_spaces.dat_r),
-            self.csrs.o_dbg_last_timestamp.status.eq(last_timestamps.dat_r),
-            If(self.csrs.o_reset_channel_status.re,
-                fifo_spaces.dat_w.eq(0),
-                fifo_spaces.we.eq(1),
-                last_timestamps.dat_w.eq(0),
-                last_timestamps.we.eq(1)
-            )
-        ]
+        # debug CSRs
+        self.comb += self.csrs.o_dbg_buffer_space.status.eq(buffer_space),
         self.sync += \
-            If((rt_packet.sr_stb & rt_packet.sr_ack & rt_packet_fifo_request),
-                self.csrs.o_dbg_fifo_space_req_cnt.status.eq(
-                    self.csrs.o_dbg_fifo_space_req_cnt.status + 1)
+            If((rt_packet.sr_stb & rt_packet.sr_ack & rt_packet_buffer_request),
+               self.csrs.o_dbg_buffer_space_req_cnt.status.eq(
+                   self.csrs.o_dbg_buffer_space_req_cnt.status + 1)
             )
 
     def get_csrs(self):
diff --git a/artiq/gateware/drtio/rt_errors_satellite.py b/artiq/gateware/drtio/rt_errors_satellite.py
index e93a7bc9c..173dd7dcf 100644
--- a/artiq/gateware/drtio/rt_errors_satellite.py
+++ b/artiq/gateware/drtio/rt_errors_satellite.py
@@ -7,31 +7,48 @@ from artiq.gateware.rtio.cdc import BlindTransfer
 
 
 class RTErrorsSatellite(Module, AutoCSR):
-    def __init__(self, rt_packet, ios):
-        self.protocol_error = CSR(5)
-        self.rtio_error = CSR(2)
+    def __init__(self, rt_packet, outputs):
+        self.protocol_error = CSR(4)
+        self.rtio_error = CSR(3)
 
         def error_csr(csr, *sources):
-            for n, source in enumerate(sources):
+            for n, (source, detect_edges) in enumerate(sources):
+                assert isinstance(source, Signal)
                 pending = Signal(related=source)
                 xfer = BlindTransfer(odomain="sys")
                 self.submodules += xfer
-                self.comb += xfer.i.eq(source)
+                if detect_edges:
+                    source_r = Signal()
+                    self.sync.rio += source_r.eq(source)
+                    self.comb += xfer.i.eq(source & source_r)
+                else:
+                    self.comb += xfer.i.eq(source)
                 self.sync += [
                     If(csr.re & csr.r[n], pending.eq(0)),
                     If(xfer.o, pending.eq(1))
                 ]
                 self.comb += csr.w[n].eq(pending)
 
-        # The master is normally responsible for avoiding output overflows,
-        # output underflows, and sequence errors.
-        # Error reports here are only for diagnosing internal ARTIQ bugs.
-        error_csr(self.protocol_error, 
-                  rt_packet.unknown_packet_type,
-                  rt_packet.packet_truncated,
-                  ios.write_underflow,
-                  ios.write_overflow,
-                  ios.write_sequence_error)
+
+        # The master is normally responsible for avoiding output overflows
+        # and output underflows. The error reports here are only for diagnosing
+        # internal ARTIQ bugs.
+        underflow = Signal()
+        overflow = Signal()
+        sequence_error = Signal()
+        self.comb += [
+            underflow.eq(outputs.cri.o_status[1]),
+            overflow.eq(outputs.cri.o_status[0]),
+            sequence_error.eq(outputs.cri.o_status[2])
+        ]
+        error_csr(self.protocol_error,
+                  (rt_packet.unknown_packet_type, False),
+                  (rt_packet.packet_truncated, False),
+                  (underflow, True),
+                  (overflow, True)
+        )
         error_csr(self.rtio_error,
-                  ios.collision,
-                  ios.busy)
+                  (sequence_error, True),
+                  (outputs.collision, False),
+                  (outputs.busy, False)
+        )
diff --git a/artiq/gateware/drtio/rt_ios_satellite.py b/artiq/gateware/drtio/rt_ios_satellite.py
deleted file mode 100644
index f9fc9096e..000000000
--- a/artiq/gateware/drtio/rt_ios_satellite.py
+++ /dev/null
@@ -1,246 +0,0 @@
-"""Real-time I/O scheduler for satellites"""
-
-from migen import *
-from migen.genlib.fifo import SyncFIFOBuffered
-from migen.genlib.record import *
-
-from artiq.gateware.rtio import rtlink
-
-
-class IOS(Module):
-    def __init__(self, rt_packet, channels, max_fine_ts_width, full_ts_width):
-        self.write_underflow = Signal()
-        self.write_overflow = Signal()
-        self.write_sequence_error = Signal()
-        self.collision = Signal()
-        self.busy = Signal()
-
-        self.rt_packet = rt_packet
-        self.max_fine_ts_width = max_fine_ts_width
-
-        self.tsc = Signal(full_ts_width - max_fine_ts_width)
-        self.sync.rtio += \
-            If(rt_packet.tsc_load,
-                self.tsc.eq(rt_packet.tsc_load_value)
-            ).Else(
-                self.tsc.eq(self.tsc + 1)
-            )
-        self.comb += rt_packet.tsc_input.eq(self.tsc)
-
-        self.sync.rio += [
-            self.write_underflow.eq(0),
-            self.write_overflow.eq(0),
-            self.collision.eq(0),
-            self.busy.eq(0)
-        ]
-        for n, channel in enumerate(channels):
-            self.add_output(n, channel)
-            self.add_input(n, channel)
-
-    def add_output(self, n, channel):
-        rt_packet = self.rt_packet
-        max_fine_ts_width = self.max_fine_ts_width
-
-        interface = channel.interface.o
-        data_width = rtlink.get_data_width(interface)
-        address_width = rtlink.get_address_width(interface)
-        fine_ts_width = rtlink.get_fine_ts_width(interface)
-        assert fine_ts_width <= max_fine_ts_width
-
-        we = Signal()
-        self.comb += we.eq(rt_packet.write_stb
-                           & (rt_packet.write_channel == n))
-        write_timestamp = rt_packet.write_timestamp[max_fine_ts_width-fine_ts_width:]
-        write_timestamp_coarse = rt_packet.write_timestamp[max_fine_ts_width:]
-        write_timestamp_fine = rt_packet.write_timestamp[max_fine_ts_width-fine_ts_width:max_fine_ts_width]
-
-        # latency compensation
-        if interface.delay:
-            tsc_comp = Signal.like(self.tsc)
-            self.sync.rtio += tsc_comp.eq(self.tsc - interface.delay + 1)
-        else:
-            tsc_comp = self.tsc
-
-        # FIFO
-        ev_layout = []
-        if data_width:
-            ev_layout.append(("data", data_width))
-        if address_width:
-            ev_layout.append(("address", address_width))
-        ev_layout.append(("timestamp", len(self.tsc) + fine_ts_width))
-
-        fifo = ClockDomainsRenamer("rio")(
-            SyncFIFOBuffered(layout_len(ev_layout), channel.ofifo_depth))
-        self.submodules += fifo
-        fifo_in = Record(ev_layout)
-        fifo_out = Record(ev_layout)
-        self.comb += [
-            fifo.din.eq(fifo_in.raw_bits()),
-            fifo_out.raw_bits().eq(fifo.dout)
-        ]
-
-        # Buffer
-        buf_pending = Signal()
-        buf = Record(ev_layout)
-        buf_just_written = Signal()
-
-        # Special cases
-        replace = Signal()
-        sequence_error = Signal()
-        collision = Signal()
-        any_error = Signal()
-        if interface.enable_replace:
-            # Note: replace may be asserted at the same time as collision
-            # when addresses are different. In that case, it is a collision.
-            self.sync.rio += replace.eq(write_timestamp == buf.timestamp)
-        # Detect sequence errors on coarse timestamps only
-        # so that they are mutually exclusive with collision errors.
-        self.sync.rio += sequence_error.eq(write_timestamp_coarse < buf.timestamp[fine_ts_width:])
-        if interface.enable_replace:
-            if address_width:
-                different_addresses = rt_packet.write_address != buf.address
-            else:
-                different_addresses = 0
-            if fine_ts_width:
-                self.sync.rio += collision.eq(
-                    (write_timestamp_coarse == buf.timestamp[fine_ts_width:])
-                    & ((write_timestamp_fine != buf.timestamp[:fine_ts_width])
-                       |different_addresses))
-            else:
-                self.sync.rio += collision.eq(
-                    (write_timestamp == buf.timestamp) & different_addresses)
-        else:
-            self.sync.rio += collision.eq(
-                write_timestamp_coarse == buf.timestamp[fine_ts_width:])
-        self.comb += any_error.eq(sequence_error | collision)
-        self.sync.rio += [
-            If(we & sequence_error, self.write_sequence_error.eq(1)),
-            If(we & collision, self.collision.eq(1))
-        ]
-
-        # Buffer read and FIFO write
-        self.comb += fifo_in.eq(buf)
-        in_guard_time = Signal()
-        self.comb += in_guard_time.eq(
-            buf.timestamp[fine_ts_width:] < tsc_comp + 4)
-        self.sync.rio += If(in_guard_time, buf_pending.eq(0))
-        report_underflow = Signal()
-        self.comb += \
-            If(buf_pending,
-                If(in_guard_time,
-                    If(buf_just_written,
-                        report_underflow.eq(1)
-                    ).Else(
-                        fifo.we.eq(1)
-                    )
-                ),
-                If(we & ~replace & ~any_error,
-                   fifo.we.eq(1)
-                )
-            )
-        self.sync.rio += If(report_underflow, self.write_underflow.eq(1))
-
-        # Buffer write
-        # Must come after read to handle concurrent read+write properly
-        self.sync.rio += [
-            buf_just_written.eq(0),
-            If(we & ~any_error,
-                buf_just_written.eq(1),
-                buf_pending.eq(1),
-                buf.timestamp.eq(write_timestamp),
-                buf.data.eq(rt_packet.write_data) if data_width else [],
-                buf.address.eq(rt_packet.write_address) if address_width else [],
-            ),
-            If(we & ~fifo.writable, self.write_overflow.eq(1))
-        ]
-
-        # FIFO level
-        self.sync.rio += \
-            If(rt_packet.fifo_space_update &
-               (rt_packet.fifo_space_channel == n),
-                rt_packet.fifo_space.eq(channel.ofifo_depth - fifo.level))
-
-        # FIFO read
-        self.sync.rio += [
-            fifo.re.eq(0),
-            interface.stb.eq(0),
-            If(fifo.readable &
-               (fifo_out.timestamp[fine_ts_width:] == tsc_comp),
-                fifo.re.eq(1),
-                interface.stb.eq(1)
-            )
-        ]
-        if data_width:
-            self.sync.rio += interface.data.eq(fifo_out.data)
-        if address_width:
-            self.sync.rio += interface.address.eq(fifo_out.address)
-        if fine_ts_width:
-            self.sync.rio += interface.fine_ts.eq(fifo_out.timestamp[:fine_ts_width])
-
-        self.sync.rio += If(interface.stb & interface.busy, self.busy.eq(1))
-
-    def add_input(self, n, channel):
-        rt_packet = self.rt_packet
-
-        interface = channel.interface.i
-        if interface is None:
-            return
-        data_width = rtlink.get_data_width(interface)
-        fine_ts_width = rtlink.get_fine_ts_width(interface)
-
-        selected = Signal()
-        self.comb += selected.eq(rt_packet.read_channel == n)
-
-        # latency compensation
-        if interface.delay:
-            tsc_comp = Signal.like(self.tsc)
-            self.sync.rtio += tsc_comp.eq(self.tsc - interface.delay + 1)
-        else:
-            tsc_comp = self.tsc
-
-        # FIFO
-        ev_layout = []
-        if data_width:
-            ev_layout.append(("data", data_width))
-        if interface.timestamped:
-            ev_layout.append(("timestamp", len(self.tsc) + fine_ts_width))
-
-        fifo = ClockDomainsRenamer("rio")(
-            SyncFIFOBuffered(layout_len(ev_layout), channel.ififo_depth))
-        self.submodules += fifo
-        fifo_in = Record(ev_layout)
-        fifo_out = Record(ev_layout)
-        self.comb += [
-            fifo.din.eq(fifo_in.raw_bits()),
-            fifo_out.raw_bits().eq(fifo.dout)
-        ]
-
-        # FIFO write
-        if data_width:
-            self.comb += fifo_in.data.eq(interface.data)
-        if interface.timestamped:
-            if fine_ts_width:
-                full_ts = Cat(interface.fine_ts, tsc_comp)
-            else:
-                full_ts = tsc_comp
-            self.comb += fifo_in.timestamp.eq(full_ts)
-        self.comb += fifo.we.eq(interface.stb)
-
-        overflow = Signal()
-        self.comb += If(selected, rt_packet.read_overflow.eq(overflow))
-        self.sync.rio += [
-            If(selected & rt_packet.read_overflow_ack, overflow.eq(0)),
-            If(fifo.we & ~fifo.writable, overflow.eq(1))
-        ]
-
-        # FIFO read
-        if data_width:
-            self.comb += If(selected, rt_packet.read_data.eq(fifo_out.data))
-        if interface.timestamped:
-            self.comb += If(selected, rt_packet.read_timestamp.eq(fifo_out.timestamp))
-        self.comb += [
-            If(selected,
-                rt_packet.read_readable.eq(fifo.readable),
-                fifo.re.eq(rt_packet.read_consume)
-            )
-        ]
diff --git a/artiq/gateware/drtio/rt_packet_master.py b/artiq/gateware/drtio/rt_packet_master.py
index a32004ebd..94dea3944 100644
--- a/artiq/gateware/drtio/rt_packet_master.py
+++ b/artiq/gateware/drtio/rt_packet_master.py
@@ -66,12 +66,12 @@ class RTPacketMaster(Module):
 
         # standard request interface
         #
-        # notwrite=1 address=0  FIFO space request <channel>
+        # notwrite=1 address=0  buffer space request
         # notwrite=1 address=1  read request <channel, timestamp>
         #
         # optimized for write throughput
         # requests are performed on the DRTIO link preserving their order of issue
-        # this is important for FIFO space requests, which have to be ordered
+        # this is important for buffer space requests, which have to be ordered
         # wrt writes.
         self.sr_stb = Signal()
         self.sr_ack = Signal()
@@ -81,10 +81,10 @@ class RTPacketMaster(Module):
         self.sr_address = Signal(16)
         self.sr_data = Signal(512)
 
-        # fifo space reply interface
-        self.fifo_space_not = Signal()
-        self.fifo_space_not_ack = Signal()
-        self.fifo_space = Signal(16)
+        # buffer space reply interface
+        self.buffer_space_not = Signal()
+        self.buffer_space_not_ack = Signal()
+        self.buffer_space = Signal(16)
 
         # read reply interface
         self.read_not = Signal()
@@ -209,11 +209,11 @@ class RTPacketMaster(Module):
             )
 
         # CDC
-        fifo_space_not = Signal()
-        fifo_space = Signal(16)
+        buffer_space_not = Signal()
+        buffer_space = Signal(16)
         self.submodules += _CrossDomainNotification("rtio_rx",
-            fifo_space_not, fifo_space,
-            self.fifo_space_not, self.fifo_space_not_ack, self.fifo_space)
+            buffer_space_not, buffer_space,
+            self.buffer_space_not, self.buffer_space_not_ack, self.buffer_space)
 
         set_time_stb = Signal()
         set_time_ack = Signal()
@@ -274,7 +274,7 @@ class RTPacketMaster(Module):
             If(sr_buf_readable,
                 If(sr_notwrite,
                     Case(sr_address[0], {
-                        0: NextState("FIFO_SPACE"),
+                        0: NextState("BUFFER_SPACE"),
                         1: NextState("READ")
                     }),
                 ).Else(
@@ -316,8 +316,8 @@ class RTPacketMaster(Module):
                 NextState("IDLE")
             )
         )
-        tx_fsm.act("FIFO_SPACE",
-            tx_dp.send("fifo_space_request", channel=sr_channel),
+        tx_fsm.act("BUFFER_SPACE",
+            tx_dp.send("buffer_space_request"),
             If(tx_dp.packet_last,
                 sr_buf_re.eq(1),
                 NextState("IDLE")
@@ -369,7 +369,7 @@ class RTPacketMaster(Module):
                 If(rx_dp.packet_last,
                     Case(rx_dp.packet_type, {
                         rx_plm.types["echo_reply"]: echo_received_now.eq(1),
-                        rx_plm.types["fifo_space_reply"]: NextState("FIFO_SPACE"),
+                        rx_plm.types["buffer_space_reply"]: NextState("BUFFER_SPACE"),
                         rx_plm.types["read_reply"]: NextState("READ_REPLY"),
                         rx_plm.types["read_reply_noevent"]: NextState("READ_REPLY_NOEVENT"),
                         "default": err_unknown_packet_type.i.eq(1)
@@ -382,9 +382,9 @@ class RTPacketMaster(Module):
                 err_packet_truncated.i.eq(1)
             )
         )
-        rx_fsm.act("FIFO_SPACE",
-            fifo_space_not.eq(1),
-            fifo_space.eq(rx_dp.packet_as["fifo_space_reply"].space),
+        rx_fsm.act("BUFFER_SPACE",
+            buffer_space_not.eq(1),
+            buffer_space.eq(rx_dp.packet_as["buffer_space_reply"].space),
             NextState("INPUT")
         )
         rx_fsm.act("READ_REPLY",
diff --git a/artiq/gateware/drtio/rt_packet_satellite.py b/artiq/gateware/drtio/rt_packet_satellite.py
index 249f8ca27..47b9a1d55 100644
--- a/artiq/gateware/drtio/rt_packet_satellite.py
+++ b/artiq/gateware/drtio/rt_packet_satellite.py
@@ -3,6 +3,7 @@
 from migen import *
 from migen.genlib.fsm import *
 
+from artiq.gateware.rtio import cri
 from artiq.gateware.drtio.rt_serializer import *
 
 
@@ -13,30 +14,11 @@ class RTPacketSatellite(Module):
 
         self.tsc_load = Signal()
         self.tsc_load_value = Signal(64)
-        self.tsc_input = Signal(64)
 
         self.reset = Signal(reset=1)
         self.reset_phy = Signal(reset=1)
 
-        self.fifo_space_channel = Signal(16)
-        self.fifo_space_update = Signal()
-        self.fifo_space = Signal(16)
-
-        # write parameters are stable one cycle before stb is asserted,
-        # and when stb is asserted.
-        self.write_stb = Signal()
-        self.write_timestamp = Signal(64)
-        self.write_channel = Signal(16)
-        self.write_address = Signal(16)
-        self.write_data = Signal(512)
-
-        self.read_channel = Signal(16)
-        self.read_readable = Signal()
-        self.read_consume = Signal()
-        self.read_data = Signal(32)
-        self.read_timestamp = Signal(64)
-        self.read_overflow = Signal()
-        self.read_overflow_ack = Signal()
+        self.cri = cri.Interface()
 
         # # #
 
@@ -69,27 +51,49 @@ class RTPacketSatellite(Module):
 
         # RX->TX
         echo_req = Signal()
-        fifo_space_set = Signal()
-        fifo_space_req = Signal()
-        fifo_space_ack = Signal()
+        buffer_space_set = Signal()
+        buffer_space_req = Signal()
+        buffer_space_ack = Signal()
         self.sync += [
-            If(fifo_space_ack, fifo_space_req.eq(0)),
-            If(fifo_space_set, fifo_space_req.eq(1)),
+            If(buffer_space_ack, buffer_space_req.eq(0)),
+            If(buffer_space_set, buffer_space_req.eq(1)),
+        ]
+
+        buffer_space_update = Signal()
+        buffer_space = Signal(16)
+        self.sync += If(buffer_space_update, buffer_space.eq(self.cri.o_buffer_space))
+
+        load_read_request = Signal()
+        clear_read_request = Signal()
+        read_request_pending = Signal()
+        self.sync += [
+            If(clear_read_request | self.reset,
+                read_request_pending.eq(0)
+            ),
+            If(load_read_request,
+                read_request_pending.eq(1),
+            )
         ]
 
         # RX FSM
+        read = Signal()
         self.comb += [
             self.tsc_load_value.eq(
                 rx_dp.packet_as["set_time"].timestamp),
-            self.fifo_space_channel.eq(
-                rx_dp.packet_as["fifo_space_request"].channel),
-            self.write_timestamp.eq(
-                rx_dp.packet_as["write"].timestamp),
-            self.write_channel.eq(
-                rx_dp.packet_as["write"].channel),
-            self.write_address.eq(
+            If(load_read_request | read_request_pending,
+                self.cri.chan_sel.eq(
+                    rx_dp.packet_as["read_request"].channel),
+                self.cri.timestamp.eq(
+                    rx_dp.packet_as["read_request"].timeout)
+            ).Else(
+                self.cri.chan_sel.eq(
+                    rx_dp.packet_as["write"].channel),
+                self.cri.timestamp.eq(
+                    rx_dp.packet_as["write"].timestamp)
+            ),
+            self.cri.o_address.eq(
                 rx_dp.packet_as["write"].address),
-            self.write_data.eq(
+            self.cri.o_data.eq(
                 Cat(rx_dp.packet_as["write"].short_data, write_data_buffer)),
         ]
 
@@ -100,26 +104,6 @@ class RTPacketSatellite(Module):
             self.reset_phy.eq(reset_phy)
         ]
 
-        load_read_request = Signal()
-        clear_read_request = Signal()
-        read_request_pending = Signal()
-        read_request_time_limit = Signal(64)
-        read_request_timeout = Signal()
-        read_request_wait = Signal()  # 1 cycle latency channel→(data,overflow) and time_limit→timeout
-        self.sync += [
-            If(clear_read_request | self.reset,
-                read_request_pending.eq(0)
-            ),
-            read_request_wait.eq(0),
-            If(load_read_request,
-                read_request_pending.eq(1),
-                read_request_wait.eq(1),
-                self.read_channel.eq(rx_dp.packet_as["read_request"].channel),
-                read_request_time_limit.eq(rx_dp.packet_as["read_request"].timeout)
-            ),
-            read_request_timeout.eq(self.tsc_input >= read_request_time_limit),
-        ]
-
         rx_fsm = FSM(reset_state="INPUT")
         self.submodules += rx_fsm
 
@@ -138,7 +122,7 @@ class RTPacketSatellite(Module):
                         rx_plm.types["set_time"]: NextState("SET_TIME"),
                         rx_plm.types["reset"]: NextState("RESET"),
                         rx_plm.types["write"]: NextState("WRITE"),
-                        rx_plm.types["fifo_space_request"]: NextState("FIFO_SPACE"),
+                        rx_plm.types["buffer_space_request"]: NextState("BUFFER_SPACE"),
                         rx_plm.types["read_request"]: NextState("READ_REQUEST"),
                         "default": self.unknown_packet_type.eq(1)
                     })
@@ -165,7 +149,7 @@ class RTPacketSatellite(Module):
 
         rx_fsm.act("WRITE",
             If(write_data_buffer_cnt == rx_dp.packet_as["write"].extra_data_cnt,
-                self.write_stb.eq(1),
+                self.cri.cmd.eq(cri.commands["write"]),
                 NextState("INPUT")
             ).Else(
                 write_data_buffer_load.eq(1),
@@ -175,14 +159,15 @@ class RTPacketSatellite(Module):
                 )
             )
         )
-        rx_fsm.act("FIFO_SPACE",
-            fifo_space_set.eq(1),
-            self.fifo_space_update.eq(1),
+        rx_fsm.act("BUFFER_SPACE",
+            buffer_space_set.eq(1),
+            buffer_space_update.eq(1),
             NextState("INPUT")
         )
 
         rx_fsm.act("READ_REQUEST",
             load_read_request.eq(1),
+            self.cri.cmd.eq(cri.commands["read"]),
             NextState("INPUT")
         )
 
@@ -192,11 +177,11 @@ class RTPacketSatellite(Module):
 
         tx_fsm.act("IDLE",
             If(echo_req, NextState("ECHO")),
-            If(fifo_space_req, NextState("FIFO_SPACE")),
-            If(~read_request_wait & read_request_pending,
-                If(read_request_timeout, NextState("READ_TIMEOUT")),
-                If(self.read_overflow, NextState("READ_OVERFLOW")),
-                If(self.read_readable, NextState("READ"))
+            If(buffer_space_req, NextState("BUFFER_SPACE")),
+            If(read_request_pending,
+                If(~self.cri.i_status[2], NextState("READ")),
+                If(self.cri.i_status[0], NextState("READ_TIMEOUT")),
+                If(self.cri.i_status[1], NextState("READ_OVERFLOW"))
             )
         )
 
@@ -205,9 +190,9 @@ class RTPacketSatellite(Module):
             If(tx_dp.packet_last, NextState("IDLE"))
         )
 
-        tx_fsm.act("FIFO_SPACE",
-            fifo_space_ack.eq(1),
-            tx_dp.send("fifo_space_reply", space=self.fifo_space),
+        tx_fsm.act("BUFFER_SPACE",
+            buffer_space_ack.eq(1),
+            tx_dp.send("buffer_space_reply", space=buffer_space),
             If(tx_dp.packet_last, NextState("IDLE"))
         )
 
@@ -220,17 +205,15 @@ class RTPacketSatellite(Module):
             tx_dp.send("read_reply_noevent", overflow=1),
             clear_read_request.eq(1),
             If(tx_dp.packet_last,
-                self.read_overflow_ack.eq(1),
                 NextState("IDLE")
             )
         )
         tx_fsm.act("READ",
             tx_dp.send("read_reply",
-                       timestamp=self.read_timestamp,
-                       data=self.read_data),
+                       timestamp=self.cri.i_timestamp,
+                       data=self.cri.i_data),
             clear_read_request.eq(1),
             If(tx_dp.packet_last,
-                self.read_consume.eq(1),
                 NextState("IDLE")
             )
         )
diff --git a/artiq/gateware/drtio/rt_serializer.py b/artiq/gateware/drtio/rt_serializer.py
index df1e2e5c6..8ba5668fc 100644
--- a/artiq/gateware/drtio/rt_serializer.py
+++ b/artiq/gateware/drtio/rt_serializer.py
@@ -18,7 +18,7 @@ class PacketLayoutManager:
         self.layouts = dict()
         self.types = dict()
         self.type_names = dict()
-    
+
     def add_type(self, name, *fields, pad=True):
         type_n = len(self.types)
         self.types[name] = type_n
@@ -54,7 +54,7 @@ def get_m2s_layouts(alignment):
                           ("address", 16),
                           ("extra_data_cnt", 8),
                           ("short_data", short_data_len))
-    plm.add_type("fifo_space_request", ("channel", 16))
+    plm.add_type("buffer_space_request")
 
     plm.add_type("read_request", ("channel", 16), ("timeout", 64))
 
@@ -66,7 +66,7 @@ def get_s2m_layouts(alignment):
 
     plm.add_type("echo_reply")
 
-    plm.add_type("fifo_space_reply", ("space", 16))
+    plm.add_type("buffer_space_reply", ("space", 16))
 
     plm.add_type("read_reply", ("timestamp", 64), ("data", 32))
     plm.add_type("read_reply_noevent", ("overflow", 1))  # overflow=0→timeout
@@ -110,7 +110,7 @@ class ReceiveDatapath(Module):
         packet_buffer_count = Signal(max=w_in_packet+1)
         self.sync += \
             If(self.packet_buffer_load,
-                Case(packet_buffer_count, 
+                Case(packet_buffer_count,
                      {i: packet_buffer[i*ws:(i+1)*ws].eq(self.data_r)
                       for i in range(w_in_packet)}),
                 packet_buffer_count.eq(packet_buffer_count + 1)
diff --git a/artiq/gateware/test/drtio/test_full_stack.py b/artiq/gateware/test/drtio/test_full_stack.py
index 59e5b2897..3b1281005 100644
--- a/artiq/gateware/test/drtio/test_full_stack.py
+++ b/artiq/gateware/test/drtio/test_full_stack.py
@@ -52,7 +52,8 @@ class DUT(Module):
         self.ttl1 = Signal()
         self.transceivers = DummyTransceiverPair(nwords)
 
-        self.submodules.master = DRTIOMaster(self.transceivers.alice)
+        self.submodules.master = DRTIOMaster(self.transceivers.alice,
+                                             fine_ts_width=0)
         self.submodules.master_ki = rtio.KernelInitiator(self.master.cri)
 
         rx_synchronizer = DummyRXSynchronizer()
@@ -60,132 +61,166 @@ class DUT(Module):
         self.submodules.phy1 = ttl_simple.Output(self.ttl1)
         self.submodules.phy2 = SimpleIOPHY(512, 32)  # test wide output data
         rtio_channels = [
-            rtio.Channel.from_phy(self.phy0, ofifo_depth=4),
-            rtio.Channel.from_phy(self.phy1, ofifo_depth=4),
-            rtio.Channel.from_phy(self.phy2, ofifo_depth=4),
+            rtio.Channel.from_phy(self.phy0),
+            rtio.Channel.from_phy(self.phy1),
+            rtio.Channel.from_phy(self.phy2),
         ]
         self.submodules.satellite = DRTIOSatellite(
-            self.transceivers.bob, rtio_channels, rx_synchronizer)
-        
+            self.transceivers.bob, rtio_channels, rx_synchronizer,
+            lane_count=4, fifo_depth=8, fine_ts_width=0)
+
+
+class OutputsTestbench:
+    def __init__(self):
+        self.dut = DUT(2)
+        self.now = 0
+
+    def init(self):
+        yield from self.dut.master.rt_controller.csrs.underflow_margin.write(100)
+        while not (yield from self.dut.master.link_layer.link_status.read()):
+            yield
+        yield from self.get_buffer_space()
+
+    def get_buffer_space(self):
+        csrs = self.dut.master.rt_controller.csrs
+        yield from csrs.o_get_buffer_space.write(1)
+        yield
+        while (yield from csrs.o_wait.read()):
+            yield
+        r = (yield from csrs.o_dbg_buffer_space.read())
+        return r
+
+    def delay(self, dt):
+        self.now += dt
+
+    def sync(self):
+        t = self.now + 15
+        while (yield self.dut.master.cri.counter) < t:
+            yield
+
+    def write(self, channel, data):
+        kcsrs = self.dut.master_ki
+        yield from kcsrs.chan_sel.write(channel)
+        yield from kcsrs.timestamp.write(self.now)
+        yield from kcsrs.o_data.write(data)
+        yield from kcsrs.o_we.write(1)
+        yield
+        status = 1
+        wlen = 0
+        while status:
+            status = yield from kcsrs.o_status.read()
+            if status & 2:
+                raise RTIOUnderflow
+            if status & 4:
+                raise RTIOSequenceError
+            yield
+            wlen += 1
+        return wlen
+
+    @passive
+    def check_ttls(self, ttl_changes):
+        cycle = 0
+        old_ttls = [0, 0]
+        while True:
+            ttls = [(yield self.dut.ttl0), (yield self.dut.ttl1)]
+            for n, (old_ttl, ttl) in enumerate(zip(old_ttls, ttls)):
+                if ttl != old_ttl:
+                    ttl_changes.append((cycle, n))
+            old_ttls = ttls
+            yield
+            cycle += 1
+
 
 class TestFullStack(unittest.TestCase):
     clocks = {"sys": 8, "rtio": 5, "rtio_rx": 5,
               "rio": 5, "rio_phy": 5,
               "sys_with_rst": 8, "rtio_with_rst": 5}
 
-    def test_outputs(self):
-        dut = DUT(2)
-        kcsrs = dut.master_ki
-        csrs = dut.master.rt_controller.csrs
-        mgr = dut.master.rt_manager
-        saterr = dut.satellite.rt_errors
-
+    def test_pulses(self):
+        tb = OutputsTestbench()
         ttl_changes = []
         correct_ttl_changes = [
-            # from test_pulses
-            (203, 0),
             (208, 0),
-            (208, 1),
-            (214, 1),
-
-            # from test_fifo_space
-            (414, 0),
-            (454, 0),
-            (494, 0),
-            (534, 0),
-            (574, 0),
-            (614, 0)
+            (213, 0),
+            (213, 1),
+            (219, 1),
         ]
 
-        now = 0
-        def delay(dt):
-            nonlocal now
-            now += dt
+        def test():
+            yield from tb.init()
+            tb.delay(200)
+            yield from tb.write(0, 1)
+            tb.delay(5)
+            yield from tb.write(0, 0)
+            yield from tb.write(1, 1)
+            tb.delay(6)
+            yield from tb.write(1, 0)
+            yield from tb.sync()
 
-        def get_fifo_space(channel):
-            yield from csrs.chan_sel_override_en.write(1)
-            yield from csrs.chan_sel_override.write(channel)
-            yield from csrs.o_get_fifo_space.write(1)
-            yield
-            while (yield from csrs.o_wait.read()):
-                yield
-            r = (yield from csrs.o_dbg_fifo_space.read())
-            yield from csrs.chan_sel_override_en.write(0)
-            return r
+        run_simulation(tb.dut,
+            {"sys": test(), "rtio": tb.check_ttls(ttl_changes)}, self.clocks)
+        self.assertEqual(ttl_changes, correct_ttl_changes)
 
-        def write(channel, data):
-            yield from kcsrs.chan_sel.write(channel)
-            yield from kcsrs.timestamp.write(now)
-            yield from kcsrs.o_data.write(data)
-            yield from kcsrs.o_we.write(1)
-            yield
-            status = 1
-            wlen = 0
-            while status:
-                status = yield from kcsrs.o_status.read()
-                if status & 2:
-                    raise RTIOUnderflow
-                if status & 4:
-                    raise RTIOSequenceError
-                yield
-                wlen += 1
-            return wlen
+    def test_underflow(self):
+        tb = OutputsTestbench()
 
-        def test_init():
-            yield from get_fifo_space(0)
-            yield from get_fifo_space(1)
-
-        def test_underflow():
+        def test():
+            yield from tb.init()
             with self.assertRaises(RTIOUnderflow):
-                yield from write(0, 0)
+                yield from tb.write(0, 0)
 
-        def test_pulses():
-            delay(200*8)
-            yield from write(0, 1)
-            delay(5*8)
-            yield from write(0, 1)
-            yield from write(0, 0)  # replace
-            yield from write(1, 1)
-            delay(6*8)
-            yield from write(1, 0)
+        run_simulation(tb.dut, {"sys": test()}, self.clocks)
 
-        def test_sequence_error():
-            delay(-200*8)
-            with self.assertRaises(RTIOSequenceError):
-                yield from write(0, 1)
-            delay(200*8)
+    def test_large_data(self):
+        tb = OutputsTestbench()
 
-        def test_large_data():
+        def test():
+            yield from tb.init()
             correct_large_data = random.Random(0).randrange(2**512-1)
-            self.assertNotEqual((yield dut.phy2.received_data), correct_large_data)
-            delay(10*8)
-            yield from write(2, correct_large_data)
-            for i in range(45):
-                yield
-            self.assertEqual((yield dut.phy2.received_data), correct_large_data)
+            self.assertNotEqual((yield tb.dut.phy2.received_data), correct_large_data)
+            tb.delay(200)
+            yield from tb.write(2, correct_large_data)
+            yield from tb.sync()
+            self.assertEqual((yield tb.dut.phy2.received_data), correct_large_data)
 
-        def test_fifo_space():
-            delay(200*8)
+        run_simulation(tb.dut, {"sys": test()}, self.clocks)
+
+    def test_buffer_space(self):
+        tb = OutputsTestbench()
+        ttl_changes = []
+        correct_ttl_changes = [(258 + 40*i, 0) for i in range(10)]
+
+        def test():
+            yield from tb.init()
+            tb.delay(250)
             max_wlen = 0
-            for _ in range(3):
-                wlen = yield from write(0, 1)
+            for i in range(10):
+                wlen = yield from tb.write(0, (i + 1) % 2)
                 max_wlen = max(max_wlen, wlen)
-                delay(40*8)
-                wlen = yield from write(0, 0)
-                max_wlen = max(max_wlen, wlen)
-                delay(40*8)
-            # check that some writes caused FIFO space requests
+                tb.delay(40)
+            # check that some writes caused buffer space requests
             self.assertGreater(max_wlen, 5)
+            yield from tb.sync()
 
-        def test_tsc_error():
+        run_simulation(tb.dut,
+            {"sys": test(), "rtio": tb.check_ttls(ttl_changes)}, self.clocks)
+        self.assertEqual(ttl_changes, correct_ttl_changes)
+
+    def test_tsc_error(self):
+        tb = OutputsTestbench()
+
+        def test():
+            saterr = tb.dut.satellite.rt_errors
+            csrs = tb.dut.master.rt_controller.csrs
+            yield from tb.init()
             errors = yield from saterr.protocol_error.read()
             self.assertEqual(errors, 0)
             yield from csrs.tsc_correction.write(100000000)
             yield from csrs.set_time.write(1)
             for i in range(15):
                yield
-            delay(10000*8)
-            yield from write(0, 1)
+            tb.delay(10000)
+            yield from tb.write(0, 1)
             for i in range(12):
                yield
             errors = yield from saterr.protocol_error.read()
@@ -195,39 +230,7 @@ class TestFullStack(unittest.TestCase):
             errors = yield from saterr.protocol_error.read()
             self.assertEqual(errors, 0)
 
-        def wait_ttl_events():
-            while len(ttl_changes) < len(correct_ttl_changes):
-                yield
-
-        def test():
-            while not (yield from dut.master.link_layer.link_status.read()):
-                yield
-
-            yield from test_init()
-            yield from test_underflow()
-            yield from test_pulses()
-            yield from test_sequence_error()
-            yield from test_fifo_space()
-            yield from test_large_data()
-            yield from test_tsc_error()
-            yield from wait_ttl_events()
-
-        @passive
-        def check_ttls():
-            cycle = 0
-            old_ttls = [0, 0]
-            while True:
-                ttls = [(yield dut.ttl0), (yield dut.ttl1)]
-                for n, (old_ttl, ttl) in enumerate(zip(old_ttls, ttls)):
-                    if ttl != old_ttl:
-                        ttl_changes.append((cycle, n))
-                old_ttls = ttls
-                yield
-                cycle += 1
-
-        run_simulation(dut,
-            {"sys": test(), "rtio": check_ttls()}, self.clocks)
-        self.assertEqual(ttl_changes, correct_ttl_changes)
+        run_simulation(tb.dut, {"sys": test()}, self.clocks)
 
     def test_inputs(self):
         dut = DUT(2)
@@ -250,8 +253,7 @@ class TestFullStack(unittest.TestCase):
                     (yield from kcsrs.i_timestamp.read()))
 
         def test():
-            # wait for link layer ready
-            for i in range(5):
+            while not (yield from dut.master.link_layer.link_status.read()):
                 yield
 
             i1 = yield from get_input(10)
@@ -269,7 +271,7 @@ class TestFullStack(unittest.TestCase):
             yield dut.phy2.rtlink.i.stb.eq(0)
 
         run_simulation(dut,
-            {"sys": test(), "rtio": generate_input()}, self.clocks, vcd_name="foo.vcd")
+            {"sys": test(), "rtio": generate_input()}, self.clocks)
 
     def test_echo(self):
         dut = DUT(2)