net: atm: remove the ambassador driver

The driver for ATM Ambassador devices spews build warnings on
microblaze. The virt_to_bus() calls discard the volatile keyword.
The right thing to do would be to migrate this driver to a modern
DMA API but it seems unlikely anyone is actually using it.
There had been no fixes or functional changes here since
the git era begun.

In fact it sounds like the FW loading was broken from 2008
'til 2012 - see commit fcdc90b025e6 ("atm: forever loop loading
ambassador firmware").

Let's remove this driver, there isn't much changing in the APIs,
if users come forward we can apologize and revert.

Link: https://lore.kernel.org/all/20220321144013.440d7fc0@kicinski-fedora-pc1c0hjn.dhcp.thefacebook.com/
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>

4 files changed, 0 insertions, 3074 deletions

diff --git a/drivers/atm/Kconfig b/drivers/atm/Kconfig
index b9370bbca828..7be08e24955c 100644
--- a/drivers/atm/Kconfig
+++ b/drivers/atm/Kconfig
@@ -244,31 +244,6 @@ config ATM_IDT77252_USE_SUNI
 	depends on ATM_IDT77252
 	default y
 
-config ATM_AMBASSADOR
-	tristate "Madge Ambassador (Collage PCI 155 Server)"
-	depends on PCI && VIRT_TO_BUS
-	select BITREVERSE
-	help
-	  This is a driver for ATMizer based ATM card produced by Madge
-	  Networks Ltd. Say Y (or M to compile as a module named ambassador)
-	  here if you have one of these cards.
-
-config ATM_AMBASSADOR_DEBUG
-	bool "Enable debugging messages"
-	depends on ATM_AMBASSADOR
-	help
-	  Somewhat useful debugging messages are available. The choice of
-	  messages is controlled by a bitmap.  This may be specified as a
-	  module argument (kernel command line argument as well?), changed
-	  dynamically using an ioctl (not yet) or changed by sending the
-	  string "Dxxxx" to VCI 1023 (where x is a hex digit).  See the file
-	  <file:drivers/atm/ambassador.h> for the meanings of the bits in the
-	  mask.
-
-	  When active, these messages can have a significant impact on the
-	  speed of the driver, and the size of your syslog files! When
-	  inactive, they will have only a modest impact on performance.
-
 config ATM_HORIZON
 	tristate "Madge Horizon [Ultra] (Collage PCI 25 and Collage PCI 155 Client)"
 	depends on PCI && VIRT_TO_BUS
diff --git a/drivers/atm/Makefile b/drivers/atm/Makefile
index aa191616a72e..99ecbc280643 100644
--- a/drivers/atm/Makefile
+++ b/drivers/atm/Makefile
@@ -7,7 +7,6 @@ fore_200e-y	:= fore200e.o
 
 obj-$(CONFIG_ATM_ZATM)		+= zatm.o uPD98402.o
 obj-$(CONFIG_ATM_NICSTAR)	+= nicstar.o
-obj-$(CONFIG_ATM_AMBASSADOR)	+= ambassador.o
 obj-$(CONFIG_ATM_HORIZON)	+= horizon.o
 obj-$(CONFIG_ATM_IA)		+= iphase.o suni.o
 obj-$(CONFIG_ATM_FORE200E)	+= fore_200e.o
diff --git a/drivers/atm/ambassador.c b/drivers/atm/ambassador.c
deleted file mode 100644
index c039b8a4fefe..000000000000
--- a/drivers/atm/ambassador.c
+++ /dev/null
@@ -1,2400 +0,0 @@
-// SPDX-License-Identifier: GPL-2.0-or-later
-/*
-  Madge Ambassador ATM Adapter driver.
-  Copyright (C) 1995-1999  Madge Networks Ltd.
-
-*/
-
-/* * dedicated to the memory of Graham Gordon 1971-1998 * */
-
-#include <linux/module.h>
-#include <linux/types.h>
-#include <linux/pci.h>
-#include <linux/kernel.h>
-#include <linux/init.h>
-#include <linux/ioport.h>
-#include <linux/atmdev.h>
-#include <linux/delay.h>
-#include <linux/interrupt.h>
-#include <linux/poison.h>
-#include <linux/bitrev.h>
-#include <linux/mutex.h>
-#include <linux/firmware.h>
-#include <linux/ihex.h>
-#include <linux/slab.h>
-
-#include <linux/atomic.h>
-#include <asm/io.h>
-#include <asm/byteorder.h>
-
-#include "ambassador.h"
-
-#define maintainer_string "Giuliano Procida at Madge Networks <gprocida@madge.com>"
-#define description_string "Madge ATM Ambassador driver"
-#define version_string "1.2.4"
-
-static inline void __init show_version (void) {
-  printk ("%s version %s\n", description_string, version_string);
-}
-
-/*
-  
-  Theory of Operation
-  
-  I Hardware, detection, initialisation and shutdown.
-  
-  1. Supported Hardware
-  
-  This driver is for the PCI ATMizer-based Ambassador card (except
-  very early versions). It is not suitable for the similar EISA "TR7"
-  card. Commercially, both cards are known as Collage Server ATM
-  adapters.
-  
-  The loader supports image transfer to the card, image start and few
-  other miscellaneous commands.
-  
-  Only AAL5 is supported with vpi = 0 and vci in the range 0 to 1023.
-  
-  The cards are big-endian.
-  
-  2. Detection
-  
-  Standard PCI stuff, the early cards are detected and rejected.
-  
-  3. Initialisation
-  
-  The cards are reset and the self-test results are checked. The
-  microcode image is then transferred and started. This waits for a
-  pointer to a descriptor containing details of the host-based queues
-  and buffers and various parameters etc. Once they are processed
-  normal operations may begin. The BIA is read using a microcode
-  command.
-  
-  4. Shutdown
-  
-  This may be accomplished either by a card reset or via the microcode
-  shutdown command. Further investigation required.
-  
-  5. Persistent state
-  
-  The card reset does not affect PCI configuration (good) or the
-  contents of several other "shared run-time registers" (bad) which
-  include doorbell and interrupt control as well as EEPROM and PCI
-  control. The driver must be careful when modifying these registers
-  not to touch bits it does not use and to undo any changes at exit.
-  
-  II Driver software
-  
-  0. Generalities
-  
-  The adapter is quite intelligent (fast) and has a simple interface
-  (few features). VPI is always zero, 1024 VCIs are supported. There
-  is limited cell rate support. UBR channels can be capped and ABR
-  (explicit rate, but not EFCI) is supported. There is no CBR or VBR
-  support.
-  
-  1. Driver <-> Adapter Communication
-  
-  Apart from the basic loader commands, the driver communicates
-  through three entities: the command queue (CQ), the transmit queue
-  pair (TXQ) and the receive queue pairs (RXQ). These three entities
-  are set up by the host and passed to the microcode just after it has
-  been started.
-  
-  All queues are host-based circular queues. They are contiguous and
-  (due to hardware limitations) have some restrictions as to their
-  locations in (bus) memory. They are of the "full means the same as
-  empty so don't do that" variety since the adapter uses pointers
-  internally.
-  
-  The queue pairs work as follows: one queue is for supply to the
-  adapter, items in it are pending and are owned by the adapter; the
-  other is the queue for return from the adapter, items in it have
-  been dealt with by the adapter. The host adds items to the supply
-  (TX descriptors and free RX buffer descriptors) and removes items
-  from the return (TX and RX completions). The adapter deals with out
-  of order completions.
-  
-  Interrupts (card to host) and the doorbell (host to card) are used
-  for signalling.
-  
-  1. CQ
-  
-  This is to communicate "open VC", "close VC", "get stats" etc. to
-  the adapter. At most one command is retired every millisecond by the
-  card. There is no out of order completion or notification. The
-  driver needs to check the return code of the command, waiting as
-  appropriate.
-  
-  2. TXQ
-  
-  TX supply items are of variable length (scatter gather support) and
-  so the queue items are (more or less) pointers to the real thing.
-  Each TX supply item contains a unique, host-supplied handle (the skb
-  bus address seems most sensible as this works for Alphas as well,
-  there is no need to do any endian conversions on the handles).
-  
-  TX return items consist of just the handles above.
-  
-  3. RXQ (up to 4 of these with different lengths and buffer sizes)
-  
-  RX supply items consist of a unique, host-supplied handle (the skb
-  bus address again) and a pointer to the buffer data area.
-  
-  RX return items consist of the handle above, the VC, length and a
-  status word. This just screams "oh so easy" doesn't it?
-
-  Note on RX pool sizes:
-   
-  Each pool should have enough buffers to handle a back-to-back stream
-  of minimum sized frames on a single VC. For example:
-  
-    frame spacing = 3us (about right)
-    
-    delay = IRQ lat + RX handling + RX buffer replenish = 20 (us)  (a guess)
-    
-    min number of buffers for one VC = 1 + delay/spacing (buffers)
-
-    delay/spacing = latency = (20+2)/3 = 7 (buffers)  (rounding up)
-    
-  The 20us delay assumes that there is no need to sleep; if we need to
-  sleep to get buffers we are going to drop frames anyway.
-  
-  In fact, each pool should have enough buffers to support the
-  simultaneous reassembly of a separate frame on each VC and cope with
-  the case in which frames complete in round robin cell fashion on
-  each VC.
-  
-  Only one frame can complete at each cell arrival, so if "n" VCs are
-  open, the worst case is to have them all complete frames together
-  followed by all starting new frames together.
-  
-    desired number of buffers = n + delay/spacing
-    
-  These are the extreme requirements, however, they are "n+k" for some
-  "k" so we have only the constant to choose. This is the argument
-  rx_lats which current defaults to 7.
-  
-  Actually, "n ? n+k : 0" is better and this is what is implemented,
-  subject to the limit given by the pool size.
-  
-  4. Driver locking
-  
-  Simple spinlocks are used around the TX and RX queue mechanisms.
-  Anyone with a faster, working method is welcome to implement it.
-  
-  The adapter command queue is protected with a spinlock. We always
-  wait for commands to complete.
-  
-  A more complex form of locking is used around parts of the VC open
-  and close functions. There are three reasons for a lock: 1. we need
-  to do atomic rate reservation and release (not used yet), 2. Opening
-  sometimes involves two adapter commands which must not be separated
-  by another command on the same VC, 3. the changes to RX pool size
-  must be atomic. The lock needs to work over context switches, so we
-  use a semaphore.
-  
-  III Hardware Features and Microcode Bugs
-  
-  1. Byte Ordering
-  
-  *%^"$&%^$*&^"$(%^$#&^%$(&#%$*(&^#%!"!"!*!
-  
-  2. Memory access
-  
-  All structures that are not accessed using DMA must be 4-byte
-  aligned (not a problem) and must not cross 4MB boundaries.
-  
-  There is a DMA memory hole at E0000000-E00000FF (groan).
-  
-  TX fragments (DMA read) must not cross 4MB boundaries (would be 16MB
-  but for a hardware bug).
-  
-  RX buffers (DMA write) must not cross 16MB boundaries and must
-  include spare trailing bytes up to the next 4-byte boundary; they
-  will be written with rubbish.
-  
-  The PLX likes to prefetch; if reading up to 4 u32 past the end of
-  each TX fragment is not a problem, then TX can be made to go a
-  little faster by passing a flag at init that disables a prefetch
-  workaround. We do not pass this flag. (new microcode only)
-  
-  Now we:
-  . Note that alloc_skb rounds up size to a 16byte boundary.  
-  . Ensure all areas do not traverse 4MB boundaries.
-  . Ensure all areas do not start at a E00000xx bus address.
-  (I cannot be certain, but this may always hold with Linux)
-  . Make all failures cause a loud message.
-  . Discard non-conforming SKBs (causes TX failure or RX fill delay).
-  . Discard non-conforming TX fragment descriptors (the TX fails).
-  In the future we could:
-  . Allow RX areas that traverse 4MB (but not 16MB) boundaries.
-  . Segment TX areas into some/more fragments, when necessary.
-  . Relax checks for non-DMA items (ignore hole).
-  . Give scatter-gather (iovec) requirements using ???. (?)
-  
-  3. VC close is broken (only for new microcode)
-  
-  The VC close adapter microcode command fails to do anything if any
-  frames have been received on the VC but none have been transmitted.
-  Frames continue to be reassembled and passed (with IRQ) to the
-  driver.
-  
-  IV To Do List
-  
-  . Fix bugs!
-  
-  . Timer code may be broken.
-  
-  . Deal with buggy VC close (somehow) in microcode 12.
-  
-  . Handle interrupted and/or non-blocking writes - is this a job for
-    the protocol layer?
-  
-  . Add code to break up TX fragments when they span 4MB boundaries.
-  
-  . Add SUNI phy layer (need to know where SUNI lives on card).
-  
-  . Implement a tx_alloc fn to (a) satisfy TX alignment etc. and (b)
-    leave extra headroom space for Ambassador TX descriptors.
-  
-  . Understand these elements of struct atm_vcc: recvq (proto?),
-    sleep, callback, listenq, backlog_quota, reply and user_back.
-  
-  . Adjust TX/RX skb allocation to favour IP with LANE/CLIP (configurable).
-  
-  . Impose a TX-pending limit (2?) on each VC, help avoid TX q overflow.
-  
-  . Decide whether RX buffer recycling is or can be made completely safe;
-    turn it back on. It looks like Werner is going to axe this.
-  
-  . Implement QoS changes on open VCs (involves extracting parts of VC open
-    and close into separate functions and using them to make changes).
-  
-  . Hack on command queue so that someone can issue multiple commands and wait
-    on the last one (OR only "no-op" or "wait" commands are waited for).
-  
-  . Eliminate need for while-schedule around do_command.
-  
-*/
-
-static void do_housekeeping (struct timer_list *t);
-/********** globals **********/
-
-static unsigned short debug = 0;
-static unsigned int cmds = 8;
-static unsigned int txs = 32;
-static unsigned int rxs[NUM_RX_POOLS] = { 64, 64, 64, 64 };
-static unsigned int rxs_bs[NUM_RX_POOLS] = { 4080, 12240, 36720, 65535 };
-static unsigned int rx_lats = 7;
-static unsigned char pci_lat = 0;
-
-static const unsigned long onegigmask = -1 << 30;
-
-/********** access to adapter **********/
-
-static inline void wr_plain (const amb_dev * dev, size_t addr, u32 data) {
-  PRINTD (DBG_FLOW|DBG_REGS, "wr: %08zx <- %08x", addr, data);
-#ifdef AMB_MMIO
-  dev->membase[addr / sizeof(u32)] = data;
-#else
-  outl (data, dev->iobase + addr);
-#endif
-}
-
-static inline u32 rd_plain (const amb_dev * dev, size_t addr) {
-#ifdef AMB_MMIO
-  u32 data = dev->membase[addr / sizeof(u32)];
-#else
-  u32 data = inl (dev->iobase + addr);
-#endif
-  PRINTD (DBG_FLOW|DBG_REGS, "rd: %08zx -> %08x", addr, data);
-  return data;
-}
-
-static inline void wr_mem (const amb_dev * dev, size_t addr, u32 data) {
-  __be32 be = cpu_to_be32 (data);
-  PRINTD (DBG_FLOW|DBG_REGS, "wr: %08zx <- %08x b[%08x]", addr, data, be);
-#ifdef AMB_MMIO
-  dev->membase[addr / sizeof(u32)] = be;
-#else
-  outl (be, dev->iobase + addr);
-#endif
-}
-
-static inline u32 rd_mem (const amb_dev * dev, size_t addr) {
-#ifdef AMB_MMIO
-  __be32 be = dev->membase[addr / sizeof(u32)];
-#else
-  __be32 be = inl (dev->iobase + addr);
-#endif
-  u32 data = be32_to_cpu (be);
-  PRINTD (DBG_FLOW|DBG_REGS, "rd: %08zx -> %08x b[%08x]", addr, data, be);
-  return data;
-}
-
-/********** dump routines **********/
-
-static inline void dump_registers (const amb_dev * dev) {
-#ifdef DEBUG_AMBASSADOR
-  if (debug & DBG_REGS) {
-    size_t i;
-    PRINTD (DBG_REGS, "reading PLX control: ");
-    for (i = 0x00; i < 0x30; i += sizeof(u32))
-      rd_mem (dev, i);
-    PRINTD (DBG_REGS, "reading mailboxes: ");
-    for (i = 0x40; i < 0x60; i += sizeof(u32))
-      rd_mem (dev, i);
-    PRINTD (DBG_REGS, "reading doorb irqev irqen reset:");
-    for (i = 0x60; i < 0x70; i += sizeof(u32))
-      rd_mem (dev, i);
-  }
-#else
-  (void) dev;
-#endif
-  return;
-}
-
-static inline void dump_loader_block (volatile loader_block * lb) {
-#ifdef DEBUG_AMBASSADOR
-  unsigned int i;
-  PRINTDB (DBG_LOAD, "lb @ %p; res: %d, cmd: %d, pay:",
-	   lb, be32_to_cpu (lb->result), be32_to_cpu (lb->command));
-  for (i = 0; i < MAX_COMMAND_DATA; ++i)
-    PRINTDM (DBG_LOAD, " %08x", be32_to_cpu (lb->payload.data[i]));
-  PRINTDE (DBG_LOAD, ", vld: %08x", be32_to_cpu (lb->valid));
-#else
-  (void) lb;
-#endif
-  return;
-}
-
-static inline void dump_command (command * cmd) {
-#ifdef DEBUG_AMBASSADOR
-  unsigned int i;
-  PRINTDB (DBG_CMD, "cmd @ %p, req: %08x, pars:",
-	   cmd, /*be32_to_cpu*/ (cmd->request));
-  for (i = 0; i < 3; ++i)
-    PRINTDM (DBG_CMD, " %08x", /*be32_to_cpu*/ (cmd->args.par[i]));
-  PRINTDE (DBG_CMD, "");
-#else
-  (void) cmd;
-#endif
-  return;
-}
-
-static inline void dump_skb (char * prefix, unsigned int vc, struct sk_buff * skb) {
-#ifdef DEBUG_AMBASSADOR
-  unsigned int i;
-  unsigned char * data = skb->data;
-  PRINTDB (DBG_DATA, "%s(%u) ", prefix, vc);
-  for (i=0; i<skb->len && i < 256;i++)
-    PRINTDM (DBG_DATA, "%02x ", data[i]);
-  PRINTDE (DBG_DATA,"");
-#else
-  (void) prefix;
-  (void) vc;
-  (void) skb;
-#endif
-  return;
-}
-
-/********** check memory areas for use by Ambassador **********/
-
-/* see limitations under Hardware Features */
-
-static int check_area (void * start, size_t length) {
-  // assumes length > 0
-  const u32 fourmegmask = -1 << 22;
-  const u32 twofivesixmask = -1 << 8;
-  const u32 starthole = 0xE0000000;
-  u32 startaddress = virt_to_bus (start);
-  u32 lastaddress = startaddress+length-1;
-  if ((startaddress ^ lastaddress) & fourmegmask ||
-      (startaddress & twofivesixmask) == starthole) {
-    PRINTK (KERN_ERR, "check_area failure: [%x,%x] - mail maintainer!",
-	    startaddress, lastaddress);
-    return -1;
-  } else {
-    return 0;
-  }
-}
-
-/********** free an skb (as per ATM device driver documentation) **********/
-
-static void amb_kfree_skb (struct sk_buff * skb) {
-  if (ATM_SKB(skb)->vcc->pop) {
-    ATM_SKB(skb)->vcc->pop (ATM_SKB(skb)->vcc, skb);
-  } else {
-    dev_kfree_skb_any (skb);
-  }
-}
-
-/********** TX completion **********/
-
-static void tx_complete (amb_dev * dev, tx_out * tx) {
-  tx_simple * tx_descr = bus_to_virt (tx->handle);
-  struct sk_buff * skb = tx_descr->skb;
-  
-  PRINTD (DBG_FLOW|DBG_TX, "tx_complete %p %p", dev, tx);
-  
-  // VC layer stats
-  atomic_inc(&ATM_SKB(skb)->vcc->stats->tx);
-  
-  // free the descriptor
-  kfree (tx_descr);
-  
-  // free the skb
-  amb_kfree_skb (skb);
-  
-  dev->stats.tx_ok++;
-  return;
-}
-
-/********** RX completion **********/
-
-static void rx_complete (amb_dev * dev, rx_out * rx) {
-  struct sk_buff * skb = bus_to_virt (rx->handle);
-  u16 vc = be16_to_cpu (rx->vc);
-  // unused: u16 lec_id = be16_to_cpu (rx->lec_id);
-  u16 status = be16_to_cpu (rx->status);
-  u16 rx_len = be16_to_cpu (rx->length);
-  
-  PRINTD (DBG_FLOW|DBG_RX, "rx_complete %p %p (len=%hu)", dev, rx, rx_len);
-  
-  // XXX move this in and add to VC stats ???
-  if (!status) {
-    struct atm_vcc * atm_vcc = dev->rxer[vc];
-    dev->stats.rx.ok++;
-    
-    if (atm_vcc) {
-      
-      if (rx_len <= atm_vcc->qos.rxtp.max_sdu) {
-	
-	if (atm_charge (atm_vcc, skb->truesize)) {
-	  
-	  // prepare socket buffer
-	  ATM_SKB(skb)->vcc = atm_vcc;
-	  skb_put (skb, rx_len);
-	  
-	  dump_skb ("<<<", vc, skb);
-	  
-	  // VC layer stats
-	  atomic_inc(&atm_vcc->stats->rx);
-	  __net_timestamp(skb);
-	  // end of our responsibility
-	  atm_vcc->push (atm_vcc, skb);
-	  return;
-	  
-	} else {
-	  // someone fix this (message), please!
-	  PRINTD (DBG_INFO|DBG_RX, "dropped thanks to atm_charge (vc %hu, truesize %u)", vc, skb->truesize);
-	  // drop stats incremented in atm_charge
-	}
-	
-      } else {
-      	PRINTK (KERN_INFO, "dropped over-size frame");
-	// should we count this?
-	atomic_inc(&atm_vcc->stats->rx_drop);
-      }
-      
-    } else {
-      PRINTD (DBG_WARN|DBG_RX, "got frame but RX closed for channel %hu", vc);
-      // this is an adapter bug, only in new version of microcode
-    }
-    
-  } else {
-    dev->stats.rx.error++;
-    if (status & CRC_ERR)
-      dev->stats.rx.badcrc++;
-    if (status & LEN_ERR)
-      dev->stats.rx.toolong++;
-    if (status & ABORT_ERR)
-      dev->stats.rx.aborted++;
-    if (status & UNUSED_ERR)
-      dev->stats.rx.unused++;
-  }
-  
-  dev_kfree_skb_any (skb);
-  return;
-}
-
-/*
-  
-  Note on queue handling.
-  
-  Here "give" and "take" refer to queue entries and a queue (pair)
-  rather than frames to or from the host or adapter. Empty frame
-  buffers are given to the RX queue pair and returned unused or
-  containing RX frames. TX frames (well, pointers to TX fragment
-  lists) are given to the TX queue pair, completions are returned.
-  
-*/
-
-/********** command queue **********/
-
-// I really don't like this, but it's the best I can do at the moment
-
-// also, the callers are responsible for byte order as the microcode
-// sometimes does 16-bit accesses (yuk yuk yuk)
-
-static int command_do (amb_dev * dev, command * cmd) {
-  amb_cq * cq = &dev->cq;
-  volatile amb_cq_ptrs * ptrs = &cq->ptrs;
-  command * my_slot;
-  
-  PRINTD (DBG_FLOW|DBG_CMD, "command_do %p", dev);
-  
-  if (test_bit (dead, &dev->flags))
-    return 0;
-  
-  spin_lock (&cq->lock);
-  
-  // if not full...
-  if (cq->pending < cq->maximum) {
-    // remember my slot for later
-    my_slot = ptrs->in;
-    PRINTD (DBG_CMD, "command in slot %p", my_slot);
-    
-    dump_command (cmd);
-    
-    // copy command in
-    *ptrs->in = *cmd;
-    cq->pending++;
-    ptrs->in = NEXTQ (ptrs->in, ptrs->start, ptrs->limit);
-    
-    // mail the command
-    wr_mem (dev, offsetof(amb_mem, mb.adapter.cmd_address), virt_to_bus (ptrs->in));
-    
-    if (cq->pending > cq->high)
-      cq->high = cq->pending;
-    spin_unlock (&cq->lock);
-    
-    // these comments were in a while-loop before, msleep removes the loop
-    // go to sleep
-    // PRINTD (DBG_CMD, "wait: sleeping %lu for command", timeout);
-    msleep(cq->pending);
-    
-    // wait for my slot to be reached (all waiters are here or above, until...)
-    while (ptrs->out != my_slot) {
-      PRINTD (DBG_CMD, "wait: command slot (now at %p)", ptrs->out);
-      set_current_state(TASK_UNINTERRUPTIBLE);
-      schedule();
-    }
-    
-    // wait on my slot (... one gets to its slot, and... )
-    while (ptrs->out->request != cpu_to_be32 (SRB_COMPLETE)) {
-      PRINTD (DBG_CMD, "wait: command slot completion");
-      set_current_state(TASK_UNINTERRUPTIBLE);
-      schedule();
-    }
-    
-    PRINTD (DBG_CMD, "command complete");
-    // update queue (... moves the queue along to the next slot)
-    spin_lock (&cq->lock);
-    cq->pending--;
-    // copy command out
-    *cmd = *ptrs->out;
-    ptrs->out = NEXTQ (ptrs->out, ptrs->start, ptrs->limit);
-    spin_unlock (&cq->lock);
-    
-    return 0;
-  } else {
-    cq->filled++;
-    spin_unlock (&cq->lock);
-    return -EAGAIN;
-  }
-  
-}
-
-/********** TX queue pair **********/
-
-static int tx_give (amb_dev * dev, tx_in * tx) {
-  amb_txq * txq = &dev->txq;
-  unsigned long flags;
-  
-  PRINTD (DBG_FLOW|DBG_TX, "tx_give %p", dev);
-
-  if (test_bit (dead, &dev->flags))
-    return 0;
-  
-  spin_lock_irqsave (&txq->lock, flags);
-  
-  if (txq->pending < txq->maximum) {
-    PRINTD (DBG_TX, "TX in slot %p", txq->in.ptr);
-
-    *txq->in.ptr = *tx;
-    txq->pending++;
-    txq->in.ptr = NEXTQ (txq->in.ptr, txq->in.start, txq->in.limit);
-    // hand over the TX and ring the bell
-    wr_mem (dev, offsetof(amb_mem, mb.adapter.tx_address), virt_to_bus (txq->in.ptr));
-    wr_mem (dev, offsetof(amb_mem, doorbell), TX_FRAME);
-    
-    if (txq->pending > txq->high)
-      txq->high = txq->pending;
-    spin_unlock_irqrestore (&txq->lock, flags);
-    return 0;
-  } else {
-    txq->filled++;
-    spin_unlock_irqrestore (&txq->lock, flags);
-    return -EAGAIN;
-  }
-}
-
-static int tx_take (amb_dev * dev) {
-  amb_txq * txq = &dev->txq;
-  unsigned long flags;
-  
-  PRINTD (DBG_FLOW|DBG_TX, "tx_take %p", dev);
-  
-  spin_lock_irqsave (&txq->lock, flags);
-  
-  if (txq->pending && txq->out.ptr->handle) {
-    // deal with TX completion
-    tx_complete (dev, txq->out.ptr);
-    // mark unused again
-    txq->out.ptr->handle = 0;
-    // remove item
-    txq->pending--;
-    txq->out.ptr = NEXTQ (txq->out.ptr, txq->out.start, txq->out.limit);
-    
-    spin_unlock_irqrestore (&txq->lock, flags);
-    return 0;
-  } else {
-    
-    spin_unlock_irqrestore (&txq->lock, flags);
-    return -1;
-  }
-}
-
-/********** RX queue pairs **********/
-
-static int rx_give (amb_dev * dev, rx_in * rx, unsigned char pool) {
-  amb_rxq * rxq = &dev->rxq[pool];
-  unsigned long flags;
-  
-  PRINTD (DBG_FLOW|DBG_RX, "rx_give %p[%hu]", dev, pool);
-  
-  spin_lock_irqsave (&rxq->lock, flags);
-  
-  if (rxq->pending < rxq->maximum) {
-    PRINTD (DBG_RX, "RX in slot %p", rxq->in.ptr);
-
-    *rxq->in.ptr = *rx;
-    rxq->pending++;
-    rxq->in.ptr = NEXTQ (rxq->in.ptr, rxq->in.start, rxq->in.limit);
-    // hand over the RX buffer
-    wr_mem (dev, offsetof(amb_mem, mb.adapter.rx_address[pool]), virt_to_bus (rxq->in.ptr));
-    
-    spin_unlock_irqrestore (&rxq->lock, flags);
-    return 0;
-  } else {
-    spin_unlock_irqrestore (&rxq->lock, flags);
-    return -1;
-  }
-}
-
-static int rx_take (amb_dev * dev, unsigned char pool) {
-  amb_rxq * rxq = &dev->rxq[pool];
-  unsigned long flags;
-  
-  PRINTD (DBG_FLOW|DBG_RX, "rx_take %p[%hu]", dev, pool);
-  
-  spin_lock_irqsave (&rxq->lock, flags);
-  
-  if (rxq->pending && (rxq->out.ptr->status || rxq->out.ptr->length)) {
-    // deal with RX completion
-    rx_complete (dev, rxq->out.ptr);
-    // mark unused again
-    rxq->out.ptr->status = 0;
-    rxq->out.ptr->length = 0;
-    // remove item
-    rxq->pending--;
-    rxq->out.ptr = NEXTQ (rxq->out.ptr, rxq->out.start, rxq->out.limit);
-    
-    if (rxq->pending < rxq->low)
-      rxq->low = rxq->pending;
-    spin_unlock_irqrestore (&rxq->lock, flags);
-    return 0;
-  } else {
-    if (!rxq->pending && rxq->buffers_wanted)
-      rxq->emptied++;
-    spin_unlock_irqrestore (&rxq->lock, flags);
-    return -1;
-  }
-}
-
-/********** RX Pool handling **********/
-
-/* pre: buffers_wanted = 0, post: pending = 0 */
-static void drain_rx_pool (amb_dev * dev, unsigned char pool) {
-  amb_rxq * rxq = &dev->rxq[pool];
-  
-  PRINTD (DBG_FLOW|DBG_POOL, "drain_rx_pool %p %hu", dev, pool);
-  
-  if (test_bit (dead, &dev->flags))
-    return;
-  
-  /* we are not quite like the fill pool routines as we cannot just
-     remove one buffer, we have to remove all of them, but we might as
-     well pretend... */
-  if (rxq->pending > rxq->buffers_wanted) {
-    command cmd;
-    cmd.request = cpu_to_be32 (SRB_FLUSH_BUFFER_Q);
-    cmd.args.flush.flags = cpu_to_be32 (pool << SRB_POOL_SHIFT);
-    while (command_do (dev, &cmd))
-      schedule();
-    /* the pool may also be emptied via the interrupt handler */
-    while (rxq->pending > rxq->buffers_wanted)
-      if (rx_take (dev, pool))
-	schedule();
-  }
-  
-  return;
-}
-
-static void drain_rx_pools (amb_dev * dev) {
-  unsigned char pool;
-  
-  PRINTD (DBG_FLOW|DBG_POOL, "drain_rx_pools %p", dev);
-  
-  for (pool = 0; pool < NUM_RX_POOLS; ++pool)
-    drain_rx_pool (dev, pool);
-}
-
-static void fill_rx_pool (amb_dev * dev, unsigned char pool,
-                                 gfp_t priority)
-{
-  rx_in rx;
-  amb_rxq * rxq;
-  
-  PRINTD (DBG_FLOW|DBG_POOL, "fill_rx_pool %p %hu %x", dev, pool, priority);
-  
-  if (test_bit (dead, &dev->flags))
-    return;
-  
-  rxq = &dev->rxq[pool];
-  while (rxq->pending < rxq->maximum && rxq->pending < rxq->buffers_wanted) {
-    
-    struct sk_buff * skb = alloc_skb (rxq->buffer_size, priority);
-    if (!skb) {
-      PRINTD (DBG_SKB|DBG_POOL, "failed to allocate skb for RX pool %hu", pool);
-      return;
-    }
-    if (check_area (skb->data, skb->truesize)) {
-      dev_kfree_skb_any (skb);
-      return;
-    }
-    // cast needed as there is no %? for pointer differences
-    PRINTD (DBG_SKB, "allocated skb at %p, head %p, area %li",
-	    skb, skb->head, (long) skb_end_offset(skb));
-    rx.handle = virt_to_bus (skb);
-    rx.host_address = cpu_to_be32 (virt_to_bus (skb->data));
-    if (rx_give (dev, &rx, pool))
-      dev_kfree_skb_any (skb);
-    
-  }
-  
-  return;
-}
-
-// top up all RX pools
-static void fill_rx_pools (amb_dev * dev) {
-  unsigned char pool;
-  
-  PRINTD (DBG_FLOW|DBG_POOL, "fill_rx_pools %p", dev);
-  
-  for (pool = 0; pool < NUM_RX_POOLS; ++pool)
-    fill_rx_pool (dev, pool, GFP_ATOMIC);
-  
-  return;
-}
-
-/********** enable host interrupts **********/
-
-static void interrupts_on (amb_dev * dev) {
-  wr_plain (dev, offsetof(amb_mem, interrupt_control),
-	    rd_plain (dev, offsetof(amb_mem, interrupt_control))
-	    | AMB_INTERRUPT_BITS);
-}
-
-/********** disable host interrupts **********/
-
-static void interrupts_off (amb_dev * dev) {
-  wr_plain (dev, offsetof(amb_mem, interrupt_control),
-	    rd_plain (dev, offsetof(amb_mem, interrupt_control))
-	    &~ AMB_INTERRUPT_BITS);
-}
-
-/********** interrupt handling **********/
-
-static irqreturn_t interrupt_handler(int irq, void *dev_id) {
-  amb_dev * dev = dev_id;
-  
-  PRINTD (DBG_IRQ|DBG_FLOW, "interrupt_handler: %p", dev_id);
-  
-  {
-    u32 interrupt = rd_plain (dev, offsetof(amb_mem, interrupt));
-  
-    // for us or someone else sharing the same interrupt
-    if (!interrupt) {
-      PRINTD (DBG_IRQ, "irq not for me: %d", irq);
-      return IRQ_NONE;
-    }
-    
-    // definitely for us
-    PRINTD (DBG_IRQ, "FYI: interrupt was %08x", interrupt);
-    wr_plain (dev, offsetof(amb_mem, interrupt), -1);
-  }
-  
-  {
-    unsigned int irq_work = 0;
-    unsigned char pool;
-    for (pool = 0; pool < NUM_RX_POOLS; ++pool)
-      while (!rx_take (dev, pool))
-	++irq_work;
-    while (!tx_take (dev))
-      ++irq_work;
-  
-    if (irq_work) {
-      fill_rx_pools (dev);
-
-      PRINTD (DBG_IRQ, "work done: %u", irq_work);
-    } else {
-      PRINTD (DBG_IRQ|DBG_WARN, "no work done");
-    }
-  }
-  
-  PRINTD (DBG_IRQ|DBG_FLOW, "interrupt_handler done: %p", dev_id);
-  return IRQ_HANDLED;
-}
-
-/********** make rate (not quite as much fun as Horizon) **********/
-
-static int make_rate (unsigned int rate, rounding r,
-		      u16 * bits, unsigned int * actual) {
-  unsigned char exp = -1; // hush gcc
-  unsigned int man = -1;  // hush gcc
-  
-  PRINTD (DBG_FLOW|DBG_QOS, "make_rate %u", rate);
-  
-  // rates in cells per second, ITU format (nasty 16-bit floating-point)
-  // given 5-bit e and 9-bit m:
-  // rate = EITHER (1+m/2^9)*2^e    OR 0
-  // bits = EITHER 1<<14 | e<<9 | m OR 0
-  // (bit 15 is "reserved", bit 14 "non-zero")
-  // smallest rate is 0 (special representation)
-  // largest rate is (1+511/512)*2^31 = 4290772992 (< 2^32-1)
-  // smallest non-zero rate is (1+0/512)*2^0 = 1 (> 0)
-  // simple algorithm:
-  // find position of top bit, this gives e
-  // remove top bit and shift (rounding if feeling clever) by 9-e
-  
-  // ucode bug: please don't set bit 14! so 0 rate not representable
-  
-  if (rate > 0xffc00000U) {
-    // larger than largest representable rate
-    
-    if (r == round_up) {
-	return -EINVAL;
-    } else {
-      exp = 31;
-      man = 511;
-    }
-    
-  } else if (rate) {
-    // representable rate
-    
-    exp = 31;
-    man = rate;
-    
-    // invariant: rate = man*2^(exp-31)
-    while (!(man & (1<<31))) {
-      exp = exp - 1;
-      man = man<<1;
-    }
-    
-    // man has top bit set
-    // rate = (2^31+(man-2^31))*2^(exp-31)
-    // rate = (1+(man-2^31)/2^31)*2^exp
-    man = man<<1;
-    man &= 0xffffffffU; // a nop on 32-bit systems
-    // rate = (1+man/2^32)*2^exp
-    
-    // exp is in the range 0 to 31, man is in the range 0 to 2^32-1
-    // time to lose significance... we want m in the range 0 to 2^9-1
-    // rounding presents a minor problem... we first decide which way
-    // we are rounding (based on given rounding direction and possibly
-    // the bits of the mantissa that are to be discarded).
-    
-    switch (r) {
-      case round_down: {
-	// just truncate
-	man = man>>(32-9);
-	break;
-      }
-      case round_up: {
-	// check all bits that we are discarding
-	if (man & (~0U>>9)) {
-	  man = (man>>(32-9)) + 1;
-	  if (man == (1<<9)) {
-	    // no need to check for round up outside of range
-	    man = 0;
-	    exp += 1;
-	  }
-	} else {
-	  man = (man>>(32-9));
-	}
-	break;
-      }
-      case round_nearest: {
-	// check msb that we are discarding
-	if (man & (1<<(32-9-1))) {
-	  man = (man>>(32-9)) + 1;
-	  if (man == (1<<9)) {
-	    // no need to check for round up outside of range
-	    man = 0;
-	    exp += 1;
-	  }
-	} else {
-	  man = (man>>(32-9));
-	}
-	break;
-      }
-    }
-    
-  } else {
-    // zero rate - not representable
-    
-    if (r == r